diff options
Diffstat (limited to 'media/libvpx/config/linux')
28 files changed, 8211 insertions, 0 deletions
diff --git a/media/libvpx/config/linux/arm/vp8_rtcd.h b/media/libvpx/config/linux/arm/vp8_rtcd.h new file mode 100644 index 0000000000..be54a85cde --- /dev/null +++ b/media/libvpx/config/linux/arm/vp8_rtcd.h @@ -0,0 +1,265 @@ +// This file is generated. Do not edit. +#ifndef VP8_RTCD_H_ +#define VP8_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +/* + * VP8 + */ + +struct blockd; +struct macroblockd; +struct loop_filter_info; + +/* Encoder forward decls */ +struct block; +struct macroblock; +struct variance_vtable; +union int_mv; +struct yv12_buffer_config; + +#ifdef __cplusplus +extern "C" { +#endif + +void vp8_bilinear_predict16x16_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict16x16_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_bilinear_predict16x16)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_bilinear_predict4x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict4x4_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_bilinear_predict4x4)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_bilinear_predict8x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict8x4_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_bilinear_predict8x4)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_bilinear_predict8x8_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict8x8_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_bilinear_predict8x8)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +int vp8_block_error_c(short *coeff, short *dqcoeff); +#define vp8_block_error vp8_block_error_c + +void vp8_copy32xn_c(const unsigned char *src_ptr, int src_stride, unsigned char *dst_ptr, int dst_stride, int height); +#define vp8_copy32xn vp8_copy32xn_c + +void vp8_copy_mem16x16_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +void vp8_copy_mem16x16_neon(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +RTCD_EXTERN void (*vp8_copy_mem16x16)(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); + +void vp8_copy_mem8x4_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +void vp8_copy_mem8x4_neon(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +RTCD_EXTERN void (*vp8_copy_mem8x4)(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); + +void vp8_copy_mem8x8_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +void vp8_copy_mem8x8_neon(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +RTCD_EXTERN void (*vp8_copy_mem8x8)(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); + +void vp8_dc_only_idct_add_c(short input_dc, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +void vp8_dc_only_idct_add_neon(short input_dc, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +RTCD_EXTERN void (*vp8_dc_only_idct_add)(short input_dc, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); + +int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +int vp8_denoiser_filter_neon(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +RTCD_EXTERN int (*vp8_denoiser_filter)(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); + +int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +int vp8_denoiser_filter_uv_neon(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +RTCD_EXTERN int (*vp8_denoiser_filter_uv)(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); + +void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *dest, int stride); +void vp8_dequant_idct_add_neon(short *input, short *dq, unsigned char *dest, int stride); +RTCD_EXTERN void (*vp8_dequant_idct_add)(short *input, short *dq, unsigned char *dest, int stride); + +void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs); +void vp8_dequant_idct_add_uv_block_neon(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs); +RTCD_EXTERN void (*vp8_dequant_idct_add_uv_block)(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs); + +void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs); +void vp8_dequant_idct_add_y_block_neon(short *q, short *dq, unsigned char *dst, int stride, char *eobs); +RTCD_EXTERN void (*vp8_dequant_idct_add_y_block)(short *q, short *dq, unsigned char *dst, int stride, char *eobs); + +void vp8_dequantize_b_c(struct blockd*, short *DQC); +void vp8_dequantize_b_neon(struct blockd*, short *DQC); +RTCD_EXTERN void (*vp8_dequantize_b)(struct blockd*, short *DQC); + +int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); +#define vp8_diamond_search_sad vp8_diamond_search_sad_c + +void vp8_fast_quantize_b_c(struct block *, struct blockd *); +void vp8_fast_quantize_b_neon(struct block *, struct blockd *); +RTCD_EXTERN void (*vp8_fast_quantize_b)(struct block *, struct blockd *); + +void vp8_loop_filter_bh_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_bh_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +RTCD_EXTERN void (*vp8_loop_filter_bh)(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); + +void vp8_loop_filter_bv_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_bv_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +RTCD_EXTERN void (*vp8_loop_filter_bv)(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); + +void vp8_loop_filter_mbh_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_mbh_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +RTCD_EXTERN void (*vp8_loop_filter_mbh)(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); + +void vp8_loop_filter_mbv_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_mbv_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +RTCD_EXTERN void (*vp8_loop_filter_mbv)(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); + +void vp8_loop_filter_bhs_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_bhs_neon(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +RTCD_EXTERN void (*vp8_loop_filter_simple_bh)(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); + +void vp8_loop_filter_bvs_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_bvs_neon(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +RTCD_EXTERN void (*vp8_loop_filter_simple_bv)(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); + +void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_mbhs_neon(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +RTCD_EXTERN void (*vp8_loop_filter_simple_mbh)(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); + +void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_mbvs_neon(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +RTCD_EXTERN void (*vp8_loop_filter_simple_mbv)(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); + +int vp8_mbblock_error_c(struct macroblock *mb, int dc); +#define vp8_mbblock_error vp8_mbblock_error_c + +int vp8_mbuverror_c(struct macroblock *mb); +#define vp8_mbuverror vp8_mbuverror_c + +int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int error_per_bit, int search_range, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); +#define vp8_refining_search_sad vp8_refining_search_sad_c + +void vp8_regular_quantize_b_c(struct block *, struct blockd *); +#define vp8_regular_quantize_b vp8_regular_quantize_b_c + +void vp8_short_fdct4x4_c(short *input, short *output, int pitch); +void vp8_short_fdct4x4_neon(short *input, short *output, int pitch); +RTCD_EXTERN void (*vp8_short_fdct4x4)(short *input, short *output, int pitch); + +void vp8_short_fdct8x4_c(short *input, short *output, int pitch); +void vp8_short_fdct8x4_neon(short *input, short *output, int pitch); +RTCD_EXTERN void (*vp8_short_fdct8x4)(short *input, short *output, int pitch); + +void vp8_short_idct4x4llm_c(short *input, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +void vp8_short_idct4x4llm_neon(short *input, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +RTCD_EXTERN void (*vp8_short_idct4x4llm)(short *input, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); + +void vp8_short_inv_walsh4x4_c(short *input, short *mb_dqcoeff); +void vp8_short_inv_walsh4x4_neon(short *input, short *mb_dqcoeff); +RTCD_EXTERN void (*vp8_short_inv_walsh4x4)(short *input, short *mb_dqcoeff); + +void vp8_short_inv_walsh4x4_1_c(short *input, short *mb_dqcoeff); +#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c + +void vp8_short_walsh4x4_c(short *input, short *output, int pitch); +void vp8_short_walsh4x4_neon(short *input, short *output, int pitch); +RTCD_EXTERN void (*vp8_short_walsh4x4)(short *input, short *output, int pitch); + +void vp8_sixtap_predict16x16_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict16x16_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_sixtap_predict16x16)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_sixtap_predict4x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict4x4_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_sixtap_predict4x4)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_sixtap_predict8x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict8x4_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_sixtap_predict8x4)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_sixtap_predict8x8_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict8x8_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_sixtap_predict8x8)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_rtcd(void); + +#include "vpx_config.h" + +#ifdef RTCD_C +#include "vpx_ports/arm.h" +static void setup_rtcd_internal(void) +{ + int flags = arm_cpu_caps(); + + (void)flags; + + vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_c; + if (flags & HAS_NEON) vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_neon; + vp8_bilinear_predict4x4 = vp8_bilinear_predict4x4_c; + if (flags & HAS_NEON) vp8_bilinear_predict4x4 = vp8_bilinear_predict4x4_neon; + vp8_bilinear_predict8x4 = vp8_bilinear_predict8x4_c; + if (flags & HAS_NEON) vp8_bilinear_predict8x4 = vp8_bilinear_predict8x4_neon; + vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_c; + if (flags & HAS_NEON) vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_neon; + vp8_copy_mem16x16 = vp8_copy_mem16x16_c; + if (flags & HAS_NEON) vp8_copy_mem16x16 = vp8_copy_mem16x16_neon; + vp8_copy_mem8x4 = vp8_copy_mem8x4_c; + if (flags & HAS_NEON) vp8_copy_mem8x4 = vp8_copy_mem8x4_neon; + vp8_copy_mem8x8 = vp8_copy_mem8x8_c; + if (flags & HAS_NEON) vp8_copy_mem8x8 = vp8_copy_mem8x8_neon; + vp8_dc_only_idct_add = vp8_dc_only_idct_add_c; + if (flags & HAS_NEON) vp8_dc_only_idct_add = vp8_dc_only_idct_add_neon; + vp8_denoiser_filter = vp8_denoiser_filter_c; + if (flags & HAS_NEON) vp8_denoiser_filter = vp8_denoiser_filter_neon; + vp8_denoiser_filter_uv = vp8_denoiser_filter_uv_c; + if (flags & HAS_NEON) vp8_denoiser_filter_uv = vp8_denoiser_filter_uv_neon; + vp8_dequant_idct_add = vp8_dequant_idct_add_c; + if (flags & HAS_NEON) vp8_dequant_idct_add = vp8_dequant_idct_add_neon; + vp8_dequant_idct_add_uv_block = vp8_dequant_idct_add_uv_block_c; + if (flags & HAS_NEON) vp8_dequant_idct_add_uv_block = vp8_dequant_idct_add_uv_block_neon; + vp8_dequant_idct_add_y_block = vp8_dequant_idct_add_y_block_c; + if (flags & HAS_NEON) vp8_dequant_idct_add_y_block = vp8_dequant_idct_add_y_block_neon; + vp8_dequantize_b = vp8_dequantize_b_c; + if (flags & HAS_NEON) vp8_dequantize_b = vp8_dequantize_b_neon; + vp8_fast_quantize_b = vp8_fast_quantize_b_c; + if (flags & HAS_NEON) vp8_fast_quantize_b = vp8_fast_quantize_b_neon; + vp8_loop_filter_bh = vp8_loop_filter_bh_c; + if (flags & HAS_NEON) vp8_loop_filter_bh = vp8_loop_filter_bh_neon; + vp8_loop_filter_bv = vp8_loop_filter_bv_c; + if (flags & HAS_NEON) vp8_loop_filter_bv = vp8_loop_filter_bv_neon; + vp8_loop_filter_mbh = vp8_loop_filter_mbh_c; + if (flags & HAS_NEON) vp8_loop_filter_mbh = vp8_loop_filter_mbh_neon; + vp8_loop_filter_mbv = vp8_loop_filter_mbv_c; + if (flags & HAS_NEON) vp8_loop_filter_mbv = vp8_loop_filter_mbv_neon; + vp8_loop_filter_simple_bh = vp8_loop_filter_bhs_c; + if (flags & HAS_NEON) vp8_loop_filter_simple_bh = vp8_loop_filter_bhs_neon; + vp8_loop_filter_simple_bv = vp8_loop_filter_bvs_c; + if (flags & HAS_NEON) vp8_loop_filter_simple_bv = vp8_loop_filter_bvs_neon; + vp8_loop_filter_simple_mbh = vp8_loop_filter_simple_horizontal_edge_c; + if (flags & HAS_NEON) vp8_loop_filter_simple_mbh = vp8_loop_filter_mbhs_neon; + vp8_loop_filter_simple_mbv = vp8_loop_filter_simple_vertical_edge_c; + if (flags & HAS_NEON) vp8_loop_filter_simple_mbv = vp8_loop_filter_mbvs_neon; + vp8_short_fdct4x4 = vp8_short_fdct4x4_c; + if (flags & HAS_NEON) vp8_short_fdct4x4 = vp8_short_fdct4x4_neon; + vp8_short_fdct8x4 = vp8_short_fdct8x4_c; + if (flags & HAS_NEON) vp8_short_fdct8x4 = vp8_short_fdct8x4_neon; + vp8_short_idct4x4llm = vp8_short_idct4x4llm_c; + if (flags & HAS_NEON) vp8_short_idct4x4llm = vp8_short_idct4x4llm_neon; + vp8_short_inv_walsh4x4 = vp8_short_inv_walsh4x4_c; + if (flags & HAS_NEON) vp8_short_inv_walsh4x4 = vp8_short_inv_walsh4x4_neon; + vp8_short_walsh4x4 = vp8_short_walsh4x4_c; + if (flags & HAS_NEON) vp8_short_walsh4x4 = vp8_short_walsh4x4_neon; + vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_c; + if (flags & HAS_NEON) vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_neon; + vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_c; + if (flags & HAS_NEON) vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_neon; + vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_c; + if (flags & HAS_NEON) vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_neon; + vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_c; + if (flags & HAS_NEON) vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_neon; +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/arm/vp9_rtcd.h b/media/libvpx/config/linux/arm/vp9_rtcd.h new file mode 100644 index 0000000000..46fcf338fb --- /dev/null +++ b/media/libvpx/config/linux/arm/vp9_rtcd.h @@ -0,0 +1,130 @@ +// This file is generated. Do not edit. +#ifndef VP9_RTCD_H_ +#define VP9_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +/* + * VP9 + */ + +#include "vpx/vpx_integer.h" +#include "vp9/common/vp9_common.h" +#include "vp9/common/vp9_enums.h" +#include "vp9/common/vp9_filter.h" + +struct macroblockd; + +/* Encoder forward decls */ +struct macroblock; +struct macroblock_plane; +struct vp9_sad_table; +struct ScanOrder; +struct search_site_config; +struct mv; +union int_mv; +struct yv12_buffer_config; + +#ifdef __cplusplus +extern "C" { +#endif + +int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz); +int64_t vp9_block_error_neon(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz); +RTCD_EXTERN int64_t (*vp9_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz); + +int64_t vp9_block_error_fp_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size); +int64_t vp9_block_error_fp_neon(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size); +RTCD_EXTERN int64_t (*vp9_block_error_fp)(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size); + +int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, uint32_t start_mv_sad, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_sad_table *sad_fn_ptr, const struct mv *center_mv); +int vp9_diamond_search_sad_neon(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, uint32_t start_mv_sad, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_sad_table *sad_fn_ptr, const struct mv *center_mv); +RTCD_EXTERN int (*vp9_diamond_search_sad)(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, uint32_t start_mv_sad, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_sad_table *sad_fn_ptr, const struct mv *center_mv); + +void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type); +void vp9_fht16x16_neon(const int16_t *input, tran_low_t *output, int stride, int tx_type); +RTCD_EXTERN void (*vp9_fht16x16)(const int16_t *input, tran_low_t *output, int stride, int tx_type); + +void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type); +void vp9_fht4x4_neon(const int16_t *input, tran_low_t *output, int stride, int tx_type); +RTCD_EXTERN void (*vp9_fht4x4)(const int16_t *input, tran_low_t *output, int stride, int tx_type); + +void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type); +void vp9_fht8x8_neon(const int16_t *input, tran_low_t *output, int stride, int tx_type); +RTCD_EXTERN void (*vp9_fht8x8)(const int16_t *input, tran_low_t *output, int stride, int tx_type); + +void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride); +#define vp9_fwht4x4 vp9_fwht4x4_c + +void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +void vp9_iht16x16_256_add_neon(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +RTCD_EXTERN void (*vp9_iht16x16_256_add)(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); + +void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +void vp9_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +RTCD_EXTERN void (*vp9_iht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); + +void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +void vp9_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +RTCD_EXTERN void (*vp9_iht8x8_64_add)(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); + +void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +RTCD_EXTERN void (*vp9_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); + +void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +RTCD_EXTERN void (*vp9_quantize_fp_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); + +void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst, INTERP_FILTER filter_type, int phase_scaler); +void vp9_scale_and_extend_frame_neon(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst, INTERP_FILTER filter_type, int phase_scaler); +RTCD_EXTERN void (*vp9_scale_and_extend_frame)(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst, INTERP_FILTER filter_type, int phase_scaler); + +void vp9_rtcd(void); + +#include "vpx_config.h" + +#ifdef RTCD_C +#include "vpx_ports/arm.h" +static void setup_rtcd_internal(void) +{ + int flags = arm_cpu_caps(); + + (void)flags; + + vp9_block_error = vp9_block_error_c; + if (flags & HAS_NEON) vp9_block_error = vp9_block_error_neon; + vp9_block_error_fp = vp9_block_error_fp_c; + if (flags & HAS_NEON) vp9_block_error_fp = vp9_block_error_fp_neon; + vp9_diamond_search_sad = vp9_diamond_search_sad_c; + if (flags & HAS_NEON) vp9_diamond_search_sad = vp9_diamond_search_sad_neon; + vp9_fht16x16 = vp9_fht16x16_c; + if (flags & HAS_NEON) vp9_fht16x16 = vp9_fht16x16_neon; + vp9_fht4x4 = vp9_fht4x4_c; + if (flags & HAS_NEON) vp9_fht4x4 = vp9_fht4x4_neon; + vp9_fht8x8 = vp9_fht8x8_c; + if (flags & HAS_NEON) vp9_fht8x8 = vp9_fht8x8_neon; + vp9_iht16x16_256_add = vp9_iht16x16_256_add_c; + if (flags & HAS_NEON) vp9_iht16x16_256_add = vp9_iht16x16_256_add_neon; + vp9_iht4x4_16_add = vp9_iht4x4_16_add_c; + if (flags & HAS_NEON) vp9_iht4x4_16_add = vp9_iht4x4_16_add_neon; + vp9_iht8x8_64_add = vp9_iht8x8_64_add_c; + if (flags & HAS_NEON) vp9_iht8x8_64_add = vp9_iht8x8_64_add_neon; + vp9_quantize_fp = vp9_quantize_fp_c; + if (flags & HAS_NEON) vp9_quantize_fp = vp9_quantize_fp_neon; + vp9_quantize_fp_32x32 = vp9_quantize_fp_32x32_c; + if (flags & HAS_NEON) vp9_quantize_fp_32x32 = vp9_quantize_fp_32x32_neon; + vp9_scale_and_extend_frame = vp9_scale_and_extend_frame_c; + if (flags & HAS_NEON) vp9_scale_and_extend_frame = vp9_scale_and_extend_frame_neon; +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/arm/vpx_config.asm b/media/libvpx/config/linux/arm/vpx_config.asm new file mode 100644 index 0000000000..ee43d0f922 --- /dev/null +++ b/media/libvpx/config/linux/arm/vpx_config.asm @@ -0,0 +1,97 @@ +@ This file was created from a .asm file +@ using the ads2gas.pl script. +.syntax unified +.equ VPX_ARCH_ARM , 1 +.equ VPX_ARCH_AARCH64 , 0 +.equ VPX_ARCH_MIPS , 0 +.equ VPX_ARCH_X86 , 0 +.equ VPX_ARCH_X86_64 , 0 +.equ VPX_ARCH_PPC , 0 +.equ VPX_ARCH_LOONGARCH , 0 +.equ HAVE_NEON_ASM , 1 +.equ HAVE_NEON , 1 +.equ HAVE_NEON_DOTPROD , 0 +.equ HAVE_NEON_I8MM , 0 +.equ HAVE_SVE , 0 +.equ HAVE_MIPS32 , 0 +.equ HAVE_DSPR2 , 0 +.equ HAVE_MSA , 0 +.equ HAVE_MIPS64 , 0 +.equ HAVE_MMX , 0 +.equ HAVE_SSE , 0 +.equ HAVE_SSE2 , 0 +.equ HAVE_SSE3 , 0 +.equ HAVE_SSSE3 , 0 +.equ HAVE_SSE4_1 , 0 +.equ HAVE_AVX , 0 +.equ HAVE_AVX2 , 0 +.equ HAVE_AVX512 , 0 +.equ HAVE_VSX , 0 +.equ HAVE_MMI , 0 +.equ HAVE_LSX , 0 +.equ HAVE_LASX , 0 +.equ HAVE_VPX_PORTS , 1 +.equ HAVE_PTHREAD_H , 1 +.equ CONFIG_DEPENDENCY_TRACKING , 1 +.equ CONFIG_EXTERNAL_BUILD , 1 +.equ CONFIG_INSTALL_DOCS , 0 +.equ CONFIG_INSTALL_BINS , 1 +.equ CONFIG_INSTALL_LIBS , 1 +.equ CONFIG_INSTALL_SRCS , 0 +.equ CONFIG_DEBUG , 0 +.equ CONFIG_GPROF , 0 +.equ CONFIG_GCOV , 0 +.equ CONFIG_RVCT , 0 +.equ CONFIG_GCC , 1 +.equ CONFIG_MSVS , 0 +.equ CONFIG_PIC , 1 +.equ CONFIG_BIG_ENDIAN , 0 +.equ CONFIG_CODEC_SRCS , 0 +.equ CONFIG_DEBUG_LIBS , 0 +.equ CONFIG_DEQUANT_TOKENS , 0 +.equ CONFIG_DC_RECON , 0 +.equ CONFIG_RUNTIME_CPU_DETECT , 1 +.equ CONFIG_POSTPROC , 0 +.equ CONFIG_VP9_POSTPROC , 0 +.equ CONFIG_MULTITHREAD , 1 +.equ CONFIG_INTERNAL_STATS , 0 +.equ CONFIG_VP8_ENCODER , 1 +.equ CONFIG_VP8_DECODER , 1 +.equ CONFIG_VP9_ENCODER , 1 +.equ CONFIG_VP9_DECODER , 1 +.equ CONFIG_VP8 , 1 +.equ CONFIG_VP9 , 1 +.equ CONFIG_ENCODERS , 1 +.equ CONFIG_DECODERS , 1 +.equ CONFIG_STATIC_MSVCRT , 0 +.equ CONFIG_SPATIAL_RESAMPLING , 1 +.equ CONFIG_REALTIME_ONLY , 1 +.equ CONFIG_ONTHEFLY_BITPACKING , 0 +.equ CONFIG_ERROR_CONCEALMENT , 0 +.equ CONFIG_SHARED , 0 +.equ CONFIG_STATIC , 1 +.equ CONFIG_SMALL , 0 +.equ CONFIG_POSTPROC_VISUALIZER , 0 +.equ CONFIG_OS_SUPPORT , 1 +.equ CONFIG_UNIT_TESTS , 0 +.equ CONFIG_WEBM_IO , 1 +.equ CONFIG_LIBYUV , 1 +.equ CONFIG_DECODE_PERF_TESTS , 0 +.equ CONFIG_ENCODE_PERF_TESTS , 0 +.equ CONFIG_MULTI_RES_ENCODING , 1 +.equ CONFIG_TEMPORAL_DENOISING , 1 +.equ CONFIG_VP9_TEMPORAL_DENOISING , 0 +.equ CONFIG_COEFFICIENT_RANGE_CHECKING , 0 +.equ CONFIG_VP9_HIGHBITDEPTH , 0 +.equ CONFIG_BETTER_HW_COMPATIBILITY , 0 +.equ CONFIG_EXPERIMENTAL , 0 +.equ CONFIG_SIZE_LIMIT , 1 +.equ CONFIG_ALWAYS_ADJUST_BPM , 0 +.equ CONFIG_BITSTREAM_DEBUG , 0 +.equ CONFIG_MISMATCH_DEBUG , 0 +.equ CONFIG_FP_MB_STATS , 0 +.equ CONFIG_EMULATE_HARDWARE , 0 +.equ CONFIG_NON_GREEDY_MV , 0 +.equ CONFIG_RATE_CTRL , 0 +.equ CONFIG_COLLECT_COMPONENT_TIMING , 0 + .section .note.GNU-stack,"",%progbits diff --git a/media/libvpx/config/linux/arm/vpx_config.c b/media/libvpx/config/linux/arm/vpx_config.c new file mode 100644 index 0000000000..c885d910c0 --- /dev/null +++ b/media/libvpx/config/linux/arm/vpx_config.c @@ -0,0 +1,10 @@ +/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */ +/* */ +/* Use of this source code is governed by a BSD-style license */ +/* that can be found in the LICENSE file in the root of the source */ +/* tree. An additional intellectual property rights grant can be found */ +/* in the file PATENTS. All contributing project authors may */ +/* be found in the AUTHORS file in the root of the source tree. */ +#include "vpx/vpx_codec.h" +static const char* const cfg = "--target=armv7-linux-gcc --enable-external-build --disable-examples --disable-install-docs --disable-unit-tests --enable-multi-res-encoding --size-limit=8192x4608 --enable-pic --disable-avx512 --enable-runtime-cpu-detect --enable-realtime-only"; +const char *vpx_codec_build_config(void) {return cfg;} diff --git a/media/libvpx/config/linux/arm/vpx_config.h b/media/libvpx/config/linux/arm/vpx_config.h new file mode 100644 index 0000000000..bfd2c04e07 --- /dev/null +++ b/media/libvpx/config/linux/arm/vpx_config.h @@ -0,0 +1,108 @@ +/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */ +/* */ +/* Use of this source code is governed by a BSD-style license */ +/* that can be found in the LICENSE file in the root of the source */ +/* tree. An additional intellectual property rights grant can be found */ +/* in the file PATENTS. All contributing project authors may */ +/* be found in the AUTHORS file in the root of the source tree. */ +/* This file automatically generated by configure. Do not edit! */ +#ifndef VPX_CONFIG_H +#define VPX_CONFIG_H +#define RESTRICT +#define INLINE inline +#define VPX_ARCH_ARM 1 +#define VPX_ARCH_AARCH64 0 +#define VPX_ARCH_MIPS 0 +#define VPX_ARCH_X86 0 +#define VPX_ARCH_X86_64 0 +#define VPX_ARCH_PPC 0 +#define VPX_ARCH_LOONGARCH 0 +#define HAVE_NEON_ASM 1 +#define HAVE_NEON 1 +#define HAVE_NEON_DOTPROD 0 +#define HAVE_NEON_I8MM 0 +#define HAVE_SVE 0 +#define HAVE_MIPS32 0 +#define HAVE_DSPR2 0 +#define HAVE_MSA 0 +#define HAVE_MIPS64 0 +#define HAVE_MMX 0 +#define HAVE_SSE 0 +#define HAVE_SSE2 0 +#define HAVE_SSE3 0 +#define HAVE_SSSE3 0 +#define HAVE_SSE4_1 0 +#define HAVE_AVX 0 +#define HAVE_AVX2 0 +#define HAVE_AVX512 0 +#define HAVE_VSX 0 +#define HAVE_MMI 0 +#define HAVE_LSX 0 +#define HAVE_LASX 0 +#define HAVE_VPX_PORTS 1 +#define HAVE_PTHREAD_H 1 +#define CONFIG_DEPENDENCY_TRACKING 1 +#define CONFIG_EXTERNAL_BUILD 1 +#define CONFIG_INSTALL_DOCS 0 +#define CONFIG_INSTALL_BINS 1 +#define CONFIG_INSTALL_LIBS 1 +#define CONFIG_INSTALL_SRCS 0 +#define CONFIG_DEBUG 0 +#define CONFIG_GPROF 0 +#define CONFIG_GCOV 0 +#define CONFIG_RVCT 0 +#define CONFIG_GCC 1 +#define CONFIG_MSVS 0 +#define CONFIG_PIC 1 +#define CONFIG_BIG_ENDIAN 0 +#define CONFIG_CODEC_SRCS 0 +#define CONFIG_DEBUG_LIBS 0 +#define CONFIG_DEQUANT_TOKENS 0 +#define CONFIG_DC_RECON 0 +#define CONFIG_RUNTIME_CPU_DETECT 1 +#define CONFIG_POSTPROC 0 +#define CONFIG_VP9_POSTPROC 0 +#define CONFIG_MULTITHREAD 1 +#define CONFIG_INTERNAL_STATS 0 +#define CONFIG_VP8_ENCODER 1 +#define CONFIG_VP8_DECODER 1 +#define CONFIG_VP9_ENCODER 1 +#define CONFIG_VP9_DECODER 1 +#define CONFIG_VP8 1 +#define CONFIG_VP9 1 +#define CONFIG_ENCODERS 1 +#define CONFIG_DECODERS 1 +#define CONFIG_STATIC_MSVCRT 0 +#define CONFIG_SPATIAL_RESAMPLING 1 +#define CONFIG_REALTIME_ONLY 1 +#define CONFIG_ONTHEFLY_BITPACKING 0 +#define CONFIG_ERROR_CONCEALMENT 0 +#define CONFIG_SHARED 0 +#define CONFIG_STATIC 1 +#define CONFIG_SMALL 0 +#define CONFIG_POSTPROC_VISUALIZER 0 +#define CONFIG_OS_SUPPORT 1 +#define CONFIG_UNIT_TESTS 0 +#define CONFIG_WEBM_IO 1 +#define CONFIG_LIBYUV 1 +#define CONFIG_DECODE_PERF_TESTS 0 +#define CONFIG_ENCODE_PERF_TESTS 0 +#define CONFIG_MULTI_RES_ENCODING 1 +#define CONFIG_TEMPORAL_DENOISING 1 +#define CONFIG_VP9_TEMPORAL_DENOISING 0 +#define CONFIG_COEFFICIENT_RANGE_CHECKING 0 +#define CONFIG_VP9_HIGHBITDEPTH 0 +#define CONFIG_BETTER_HW_COMPATIBILITY 0 +#define CONFIG_EXPERIMENTAL 0 +#define CONFIG_SIZE_LIMIT 1 +#define CONFIG_ALWAYS_ADJUST_BPM 0 +#define CONFIG_BITSTREAM_DEBUG 0 +#define CONFIG_MISMATCH_DEBUG 0 +#define CONFIG_FP_MB_STATS 0 +#define CONFIG_EMULATE_HARDWARE 0 +#define CONFIG_NON_GREEDY_MV 0 +#define CONFIG_RATE_CTRL 0 +#define CONFIG_COLLECT_COMPONENT_TIMING 0 +#define DECODE_WIDTH_LIMIT 8192 +#define DECODE_HEIGHT_LIMIT 4608 +#endif /* VPX_CONFIG_H */ diff --git a/media/libvpx/config/linux/arm/vpx_dsp_rtcd.h b/media/libvpx/config/linux/arm/vpx_dsp_rtcd.h new file mode 100644 index 0000000000..fbe85df601 --- /dev/null +++ b/media/libvpx/config/linux/arm/vpx_dsp_rtcd.h @@ -0,0 +1,1415 @@ +// This file is generated. Do not edit. +#ifndef VPX_DSP_RTCD_H_ +#define VPX_DSP_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +/* + * DSP + */ + +#include "vpx/vpx_integer.h" +#include "vpx_dsp/vpx_dsp_common.h" +#include "vpx_dsp/vpx_filter.h" +#if CONFIG_VP9_ENCODER + struct macroblock_plane; + struct ScanOrder; +#endif + + +#ifdef __cplusplus +extern "C" { +#endif + +unsigned int vpx_avg_4x4_c(const uint8_t *, int p); +unsigned int vpx_avg_4x4_neon(const uint8_t *, int p); +RTCD_EXTERN unsigned int (*vpx_avg_4x4)(const uint8_t *, int p); + +unsigned int vpx_avg_8x8_c(const uint8_t *, int p); +unsigned int vpx_avg_8x8_neon(const uint8_t *, int p); +RTCD_EXTERN unsigned int (*vpx_avg_8x8)(const uint8_t *, int p); + +void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride); +void vpx_comp_avg_pred_neon(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride); +RTCD_EXTERN void (*vpx_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride); + +void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_avg)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_avg_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_vert_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_avg_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve_avg_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve_avg)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve_copy_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d117_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d117_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d117_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d117_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d117_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d117_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d117_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d117_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d135_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d135_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d135_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d135_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d135_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d135_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d135_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d153_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d153_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d153_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d153_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d207_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d207_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d207_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d207_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d45_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d45_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d45_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d45_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c + +void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d63_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d63_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d63_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d63_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c + +void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_128_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_128_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_128_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_128_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_left_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_left_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_left_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_left_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_top_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_top_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_top_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_top_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct16x16_neon(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct16x16)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct16x16_1_neon(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct16x16_1)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_neon(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct32x32)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_1_neon(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct32x32_1)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_rd_neon(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct32x32_rd)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct4x4_neon(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct4x4)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct4x4_1_neon(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct4x4_1)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct8x8_neon(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct8x8)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct8x8_1_neon(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct8x8_1)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_get16x16var_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +void vpx_get16x16var_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +RTCD_EXTERN void (*vpx_get16x16var)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); + +unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride); +unsigned int vpx_get4x4sse_cs_neon(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_get4x4sse_cs)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride); + +void vpx_get8x8var_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +void vpx_get8x8var_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +RTCD_EXTERN void (*vpx_get8x8var)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); + +unsigned int vpx_get_mb_ss_c(const int16_t *); +#define vpx_get_mb_ss vpx_get_mb_ss_c + +void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_h_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_h_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_h_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_h_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_16x16_neon(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +RTCD_EXTERN void (*vpx_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); + +void vpx_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_32x32_neon(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +RTCD_EXTERN void (*vpx_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); + +void vpx_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_8x8_neon(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +RTCD_EXTERN void (*vpx_hadamard_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); + +void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c + +void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_10_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct16x16_10_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_1_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct16x16_1_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_256_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct16x16_256_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct16x16_38_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_38_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct16x16_38_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_1024_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct32x32_1024_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_135_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct32x32_135_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_1_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct32x32_1_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_34_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct32x32_34_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct4x4_16_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct4x4_16_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct4x4_1_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct4x4_1_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_12_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct8x8_12_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_1_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct8x8_1_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_64_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct8x8_64_add)(const tran_low_t *input, uint8_t *dest, int stride); + +int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width); +int16_t vpx_int_pro_col_neon(const uint8_t *ref, const int width); +RTCD_EXTERN int16_t (*vpx_int_pro_col)(const uint8_t *ref, const int width); + +void vpx_int_pro_row_c(int16_t hbuf[16], const uint8_t *ref, const int ref_stride, const int height); +void vpx_int_pro_row_neon(int16_t hbuf[16], const uint8_t *ref, const int ref_stride, const int height); +RTCD_EXTERN void (*vpx_int_pro_row)(int16_t hbuf[16], const uint8_t *ref, const int ref_stride, const int height); + +void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_c + +void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c + +void vpx_lpf_horizontal_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_16_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_horizontal_16)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_horizontal_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_16_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_horizontal_16_dual)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_4_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_horizontal_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_horizontal_4_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*vpx_lpf_horizontal_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_8_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_horizontal_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_horizontal_8_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*vpx_lpf_horizontal_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_16_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_vertical_16)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_16_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_vertical_16_dual)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_4_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_vertical_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_vertical_4_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*vpx_lpf_vertical_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_8_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_vertical_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_vertical_8_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*vpx_lpf_vertical_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max); +void vpx_minmax_8x8_neon(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max); +RTCD_EXTERN void (*vpx_minmax_8x8)(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max); + +unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +RTCD_EXTERN void (*vpx_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); + +void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_32x32_neon(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +RTCD_EXTERN void (*vpx_quantize_b_32x32)(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); + +unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad4x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad4x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad4x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad4x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad4x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad4x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad8x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad8x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad8x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad8x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad8x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad8x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad8x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad8x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_4x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_4x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_8x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_8x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +int vpx_satd_c(const int16_t *coeff, int length); +int vpx_satd_neon(const int16_t *coeff, int length); +RTCD_EXTERN int (*vpx_satd)(const int16_t *coeff, int length); + +void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_scaled_2d_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c + +void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c + +void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c + +void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_horiz vpx_scaled_horiz_c + +void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_vert vpx_scaled_vert_c + +int64_t vpx_sse_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); +int64_t vpx_sse_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); +RTCD_EXTERN int64_t (*vpx_sse)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); + +uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x16_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x32_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x8_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x16_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x32_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x64_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance4x4_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance4x8_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x32_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x64_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x16_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x4_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x8_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x16_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance16x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x32_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance16x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x8_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance16x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x16_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance32x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x32_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance32x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x64_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance32x64)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance4x4_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance4x4)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance4x8_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance4x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x32_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance64x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x64_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance64x64)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x16_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance8x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x4_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance8x4)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x8_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance8x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride); +void vpx_subtract_block_neon(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride); +RTCD_EXTERN void (*vpx_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride); + +uint64_t vpx_sum_squares_2d_i16_c(const int16_t *src, int stride, int size); +uint64_t vpx_sum_squares_2d_i16_neon(const int16_t *src, int stride, int size); +RTCD_EXTERN uint64_t (*vpx_sum_squares_2d_i16)(const int16_t *src, int stride, int size); + +void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_tm_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_tm_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_tm_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_tm_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_v_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_v_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_v_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_v_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance4x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c + +int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl); +int vpx_vector_var_neon(const int16_t *ref, const int16_t *src, const int bwl); +RTCD_EXTERN int (*vpx_vector_var)(const int16_t *ref, const int16_t *src, const int bwl); + +void vpx_dsp_rtcd(void); + +#include "vpx_config.h" + +#ifdef RTCD_C +#include "vpx_ports/arm.h" +static void setup_rtcd_internal(void) +{ + int flags = arm_cpu_caps(); + + (void)flags; + + vpx_avg_4x4 = vpx_avg_4x4_c; + if (flags & HAS_NEON) vpx_avg_4x4 = vpx_avg_4x4_neon; + vpx_avg_8x8 = vpx_avg_8x8_c; + if (flags & HAS_NEON) vpx_avg_8x8 = vpx_avg_8x8_neon; + vpx_comp_avg_pred = vpx_comp_avg_pred_c; + if (flags & HAS_NEON) vpx_comp_avg_pred = vpx_comp_avg_pred_neon; + vpx_convolve8 = vpx_convolve8_c; + if (flags & HAS_NEON) vpx_convolve8 = vpx_convolve8_neon; + vpx_convolve8_avg = vpx_convolve8_avg_c; + if (flags & HAS_NEON) vpx_convolve8_avg = vpx_convolve8_avg_neon; + vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_c; + if (flags & HAS_NEON) vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_neon; + vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_c; + if (flags & HAS_NEON) vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_neon; + vpx_convolve8_horiz = vpx_convolve8_horiz_c; + if (flags & HAS_NEON) vpx_convolve8_horiz = vpx_convolve8_horiz_neon; + vpx_convolve8_vert = vpx_convolve8_vert_c; + if (flags & HAS_NEON) vpx_convolve8_vert = vpx_convolve8_vert_neon; + vpx_convolve_avg = vpx_convolve_avg_c; + if (flags & HAS_NEON) vpx_convolve_avg = vpx_convolve_avg_neon; + vpx_convolve_copy = vpx_convolve_copy_c; + if (flags & HAS_NEON) vpx_convolve_copy = vpx_convolve_copy_neon; + vpx_d117_predictor_16x16 = vpx_d117_predictor_16x16_c; + if (flags & HAS_NEON) vpx_d117_predictor_16x16 = vpx_d117_predictor_16x16_neon; + vpx_d117_predictor_32x32 = vpx_d117_predictor_32x32_c; + if (flags & HAS_NEON) vpx_d117_predictor_32x32 = vpx_d117_predictor_32x32_neon; + vpx_d117_predictor_4x4 = vpx_d117_predictor_4x4_c; + if (flags & HAS_NEON) vpx_d117_predictor_4x4 = vpx_d117_predictor_4x4_neon; + vpx_d117_predictor_8x8 = vpx_d117_predictor_8x8_c; + if (flags & HAS_NEON) vpx_d117_predictor_8x8 = vpx_d117_predictor_8x8_neon; + vpx_d135_predictor_16x16 = vpx_d135_predictor_16x16_c; + if (flags & HAS_NEON) vpx_d135_predictor_16x16 = vpx_d135_predictor_16x16_neon; + vpx_d135_predictor_32x32 = vpx_d135_predictor_32x32_c; + if (flags & HAS_NEON) vpx_d135_predictor_32x32 = vpx_d135_predictor_32x32_neon; + vpx_d135_predictor_4x4 = vpx_d135_predictor_4x4_c; + if (flags & HAS_NEON) vpx_d135_predictor_4x4 = vpx_d135_predictor_4x4_neon; + vpx_d135_predictor_8x8 = vpx_d135_predictor_8x8_c; + if (flags & HAS_NEON) vpx_d135_predictor_8x8 = vpx_d135_predictor_8x8_neon; + vpx_d153_predictor_16x16 = vpx_d153_predictor_16x16_c; + if (flags & HAS_NEON) vpx_d153_predictor_16x16 = vpx_d153_predictor_16x16_neon; + vpx_d153_predictor_32x32 = vpx_d153_predictor_32x32_c; + if (flags & HAS_NEON) vpx_d153_predictor_32x32 = vpx_d153_predictor_32x32_neon; + vpx_d153_predictor_4x4 = vpx_d153_predictor_4x4_c; + if (flags & HAS_NEON) vpx_d153_predictor_4x4 = vpx_d153_predictor_4x4_neon; + vpx_d153_predictor_8x8 = vpx_d153_predictor_8x8_c; + if (flags & HAS_NEON) vpx_d153_predictor_8x8 = vpx_d153_predictor_8x8_neon; + vpx_d207_predictor_16x16 = vpx_d207_predictor_16x16_c; + if (flags & HAS_NEON) vpx_d207_predictor_16x16 = vpx_d207_predictor_16x16_neon; + vpx_d207_predictor_32x32 = vpx_d207_predictor_32x32_c; + if (flags & HAS_NEON) vpx_d207_predictor_32x32 = vpx_d207_predictor_32x32_neon; + vpx_d207_predictor_4x4 = vpx_d207_predictor_4x4_c; + if (flags & HAS_NEON) vpx_d207_predictor_4x4 = vpx_d207_predictor_4x4_neon; + vpx_d207_predictor_8x8 = vpx_d207_predictor_8x8_c; + if (flags & HAS_NEON) vpx_d207_predictor_8x8 = vpx_d207_predictor_8x8_neon; + vpx_d45_predictor_16x16 = vpx_d45_predictor_16x16_c; + if (flags & HAS_NEON) vpx_d45_predictor_16x16 = vpx_d45_predictor_16x16_neon; + vpx_d45_predictor_32x32 = vpx_d45_predictor_32x32_c; + if (flags & HAS_NEON) vpx_d45_predictor_32x32 = vpx_d45_predictor_32x32_neon; + vpx_d45_predictor_4x4 = vpx_d45_predictor_4x4_c; + if (flags & HAS_NEON) vpx_d45_predictor_4x4 = vpx_d45_predictor_4x4_neon; + vpx_d45_predictor_8x8 = vpx_d45_predictor_8x8_c; + if (flags & HAS_NEON) vpx_d45_predictor_8x8 = vpx_d45_predictor_8x8_neon; + vpx_d63_predictor_16x16 = vpx_d63_predictor_16x16_c; + if (flags & HAS_NEON) vpx_d63_predictor_16x16 = vpx_d63_predictor_16x16_neon; + vpx_d63_predictor_32x32 = vpx_d63_predictor_32x32_c; + if (flags & HAS_NEON) vpx_d63_predictor_32x32 = vpx_d63_predictor_32x32_neon; + vpx_d63_predictor_4x4 = vpx_d63_predictor_4x4_c; + if (flags & HAS_NEON) vpx_d63_predictor_4x4 = vpx_d63_predictor_4x4_neon; + vpx_d63_predictor_8x8 = vpx_d63_predictor_8x8_c; + if (flags & HAS_NEON) vpx_d63_predictor_8x8 = vpx_d63_predictor_8x8_neon; + vpx_dc_128_predictor_16x16 = vpx_dc_128_predictor_16x16_c; + if (flags & HAS_NEON) vpx_dc_128_predictor_16x16 = vpx_dc_128_predictor_16x16_neon; + vpx_dc_128_predictor_32x32 = vpx_dc_128_predictor_32x32_c; + if (flags & HAS_NEON) vpx_dc_128_predictor_32x32 = vpx_dc_128_predictor_32x32_neon; + vpx_dc_128_predictor_4x4 = vpx_dc_128_predictor_4x4_c; + if (flags & HAS_NEON) vpx_dc_128_predictor_4x4 = vpx_dc_128_predictor_4x4_neon; + vpx_dc_128_predictor_8x8 = vpx_dc_128_predictor_8x8_c; + if (flags & HAS_NEON) vpx_dc_128_predictor_8x8 = vpx_dc_128_predictor_8x8_neon; + vpx_dc_left_predictor_16x16 = vpx_dc_left_predictor_16x16_c; + if (flags & HAS_NEON) vpx_dc_left_predictor_16x16 = vpx_dc_left_predictor_16x16_neon; + vpx_dc_left_predictor_32x32 = vpx_dc_left_predictor_32x32_c; + if (flags & HAS_NEON) vpx_dc_left_predictor_32x32 = vpx_dc_left_predictor_32x32_neon; + vpx_dc_left_predictor_4x4 = vpx_dc_left_predictor_4x4_c; + if (flags & HAS_NEON) vpx_dc_left_predictor_4x4 = vpx_dc_left_predictor_4x4_neon; + vpx_dc_left_predictor_8x8 = vpx_dc_left_predictor_8x8_c; + if (flags & HAS_NEON) vpx_dc_left_predictor_8x8 = vpx_dc_left_predictor_8x8_neon; + vpx_dc_predictor_16x16 = vpx_dc_predictor_16x16_c; + if (flags & HAS_NEON) vpx_dc_predictor_16x16 = vpx_dc_predictor_16x16_neon; + vpx_dc_predictor_32x32 = vpx_dc_predictor_32x32_c; + if (flags & HAS_NEON) vpx_dc_predictor_32x32 = vpx_dc_predictor_32x32_neon; + vpx_dc_predictor_4x4 = vpx_dc_predictor_4x4_c; + if (flags & HAS_NEON) vpx_dc_predictor_4x4 = vpx_dc_predictor_4x4_neon; + vpx_dc_predictor_8x8 = vpx_dc_predictor_8x8_c; + if (flags & HAS_NEON) vpx_dc_predictor_8x8 = vpx_dc_predictor_8x8_neon; + vpx_dc_top_predictor_16x16 = vpx_dc_top_predictor_16x16_c; + if (flags & HAS_NEON) vpx_dc_top_predictor_16x16 = vpx_dc_top_predictor_16x16_neon; + vpx_dc_top_predictor_32x32 = vpx_dc_top_predictor_32x32_c; + if (flags & HAS_NEON) vpx_dc_top_predictor_32x32 = vpx_dc_top_predictor_32x32_neon; + vpx_dc_top_predictor_4x4 = vpx_dc_top_predictor_4x4_c; + if (flags & HAS_NEON) vpx_dc_top_predictor_4x4 = vpx_dc_top_predictor_4x4_neon; + vpx_dc_top_predictor_8x8 = vpx_dc_top_predictor_8x8_c; + if (flags & HAS_NEON) vpx_dc_top_predictor_8x8 = vpx_dc_top_predictor_8x8_neon; + vpx_fdct16x16 = vpx_fdct16x16_c; + if (flags & HAS_NEON) vpx_fdct16x16 = vpx_fdct16x16_neon; + vpx_fdct16x16_1 = vpx_fdct16x16_1_c; + if (flags & HAS_NEON) vpx_fdct16x16_1 = vpx_fdct16x16_1_neon; + vpx_fdct32x32 = vpx_fdct32x32_c; + if (flags & HAS_NEON) vpx_fdct32x32 = vpx_fdct32x32_neon; + vpx_fdct32x32_1 = vpx_fdct32x32_1_c; + if (flags & HAS_NEON) vpx_fdct32x32_1 = vpx_fdct32x32_1_neon; + vpx_fdct32x32_rd = vpx_fdct32x32_rd_c; + if (flags & HAS_NEON) vpx_fdct32x32_rd = vpx_fdct32x32_rd_neon; + vpx_fdct4x4 = vpx_fdct4x4_c; + if (flags & HAS_NEON) vpx_fdct4x4 = vpx_fdct4x4_neon; + vpx_fdct4x4_1 = vpx_fdct4x4_1_c; + if (flags & HAS_NEON) vpx_fdct4x4_1 = vpx_fdct4x4_1_neon; + vpx_fdct8x8 = vpx_fdct8x8_c; + if (flags & HAS_NEON) vpx_fdct8x8 = vpx_fdct8x8_neon; + vpx_fdct8x8_1 = vpx_fdct8x8_1_c; + if (flags & HAS_NEON) vpx_fdct8x8_1 = vpx_fdct8x8_1_neon; + vpx_get16x16var = vpx_get16x16var_c; + if (flags & HAS_NEON) vpx_get16x16var = vpx_get16x16var_neon; + vpx_get4x4sse_cs = vpx_get4x4sse_cs_c; + if (flags & HAS_NEON) vpx_get4x4sse_cs = vpx_get4x4sse_cs_neon; + vpx_get8x8var = vpx_get8x8var_c; + if (flags & HAS_NEON) vpx_get8x8var = vpx_get8x8var_neon; + vpx_h_predictor_16x16 = vpx_h_predictor_16x16_c; + if (flags & HAS_NEON) vpx_h_predictor_16x16 = vpx_h_predictor_16x16_neon; + vpx_h_predictor_32x32 = vpx_h_predictor_32x32_c; + if (flags & HAS_NEON) vpx_h_predictor_32x32 = vpx_h_predictor_32x32_neon; + vpx_h_predictor_4x4 = vpx_h_predictor_4x4_c; + if (flags & HAS_NEON) vpx_h_predictor_4x4 = vpx_h_predictor_4x4_neon; + vpx_h_predictor_8x8 = vpx_h_predictor_8x8_c; + if (flags & HAS_NEON) vpx_h_predictor_8x8 = vpx_h_predictor_8x8_neon; + vpx_hadamard_16x16 = vpx_hadamard_16x16_c; + if (flags & HAS_NEON) vpx_hadamard_16x16 = vpx_hadamard_16x16_neon; + vpx_hadamard_32x32 = vpx_hadamard_32x32_c; + if (flags & HAS_NEON) vpx_hadamard_32x32 = vpx_hadamard_32x32_neon; + vpx_hadamard_8x8 = vpx_hadamard_8x8_c; + if (flags & HAS_NEON) vpx_hadamard_8x8 = vpx_hadamard_8x8_neon; + vpx_idct16x16_10_add = vpx_idct16x16_10_add_c; + if (flags & HAS_NEON) vpx_idct16x16_10_add = vpx_idct16x16_10_add_neon; + vpx_idct16x16_1_add = vpx_idct16x16_1_add_c; + if (flags & HAS_NEON) vpx_idct16x16_1_add = vpx_idct16x16_1_add_neon; + vpx_idct16x16_256_add = vpx_idct16x16_256_add_c; + if (flags & HAS_NEON) vpx_idct16x16_256_add = vpx_idct16x16_256_add_neon; + vpx_idct16x16_38_add = vpx_idct16x16_38_add_c; + if (flags & HAS_NEON) vpx_idct16x16_38_add = vpx_idct16x16_38_add_neon; + vpx_idct32x32_1024_add = vpx_idct32x32_1024_add_c; + if (flags & HAS_NEON) vpx_idct32x32_1024_add = vpx_idct32x32_1024_add_neon; + vpx_idct32x32_135_add = vpx_idct32x32_135_add_c; + if (flags & HAS_NEON) vpx_idct32x32_135_add = vpx_idct32x32_135_add_neon; + vpx_idct32x32_1_add = vpx_idct32x32_1_add_c; + if (flags & HAS_NEON) vpx_idct32x32_1_add = vpx_idct32x32_1_add_neon; + vpx_idct32x32_34_add = vpx_idct32x32_34_add_c; + if (flags & HAS_NEON) vpx_idct32x32_34_add = vpx_idct32x32_34_add_neon; + vpx_idct4x4_16_add = vpx_idct4x4_16_add_c; + if (flags & HAS_NEON) vpx_idct4x4_16_add = vpx_idct4x4_16_add_neon; + vpx_idct4x4_1_add = vpx_idct4x4_1_add_c; + if (flags & HAS_NEON) vpx_idct4x4_1_add = vpx_idct4x4_1_add_neon; + vpx_idct8x8_12_add = vpx_idct8x8_12_add_c; + if (flags & HAS_NEON) vpx_idct8x8_12_add = vpx_idct8x8_12_add_neon; + vpx_idct8x8_1_add = vpx_idct8x8_1_add_c; + if (flags & HAS_NEON) vpx_idct8x8_1_add = vpx_idct8x8_1_add_neon; + vpx_idct8x8_64_add = vpx_idct8x8_64_add_c; + if (flags & HAS_NEON) vpx_idct8x8_64_add = vpx_idct8x8_64_add_neon; + vpx_int_pro_col = vpx_int_pro_col_c; + if (flags & HAS_NEON) vpx_int_pro_col = vpx_int_pro_col_neon; + vpx_int_pro_row = vpx_int_pro_row_c; + if (flags & HAS_NEON) vpx_int_pro_row = vpx_int_pro_row_neon; + vpx_lpf_horizontal_16 = vpx_lpf_horizontal_16_c; + if (flags & HAS_NEON) vpx_lpf_horizontal_16 = vpx_lpf_horizontal_16_neon; + vpx_lpf_horizontal_16_dual = vpx_lpf_horizontal_16_dual_c; + if (flags & HAS_NEON) vpx_lpf_horizontal_16_dual = vpx_lpf_horizontal_16_dual_neon; + vpx_lpf_horizontal_4 = vpx_lpf_horizontal_4_c; + if (flags & HAS_NEON) vpx_lpf_horizontal_4 = vpx_lpf_horizontal_4_neon; + vpx_lpf_horizontal_4_dual = vpx_lpf_horizontal_4_dual_c; + if (flags & HAS_NEON) vpx_lpf_horizontal_4_dual = vpx_lpf_horizontal_4_dual_neon; + vpx_lpf_horizontal_8 = vpx_lpf_horizontal_8_c; + if (flags & HAS_NEON) vpx_lpf_horizontal_8 = vpx_lpf_horizontal_8_neon; + vpx_lpf_horizontal_8_dual = vpx_lpf_horizontal_8_dual_c; + if (flags & HAS_NEON) vpx_lpf_horizontal_8_dual = vpx_lpf_horizontal_8_dual_neon; + vpx_lpf_vertical_16 = vpx_lpf_vertical_16_c; + if (flags & HAS_NEON) vpx_lpf_vertical_16 = vpx_lpf_vertical_16_neon; + vpx_lpf_vertical_16_dual = vpx_lpf_vertical_16_dual_c; + if (flags & HAS_NEON) vpx_lpf_vertical_16_dual = vpx_lpf_vertical_16_dual_neon; + vpx_lpf_vertical_4 = vpx_lpf_vertical_4_c; + if (flags & HAS_NEON) vpx_lpf_vertical_4 = vpx_lpf_vertical_4_neon; + vpx_lpf_vertical_4_dual = vpx_lpf_vertical_4_dual_c; + if (flags & HAS_NEON) vpx_lpf_vertical_4_dual = vpx_lpf_vertical_4_dual_neon; + vpx_lpf_vertical_8 = vpx_lpf_vertical_8_c; + if (flags & HAS_NEON) vpx_lpf_vertical_8 = vpx_lpf_vertical_8_neon; + vpx_lpf_vertical_8_dual = vpx_lpf_vertical_8_dual_c; + if (flags & HAS_NEON) vpx_lpf_vertical_8_dual = vpx_lpf_vertical_8_dual_neon; + vpx_minmax_8x8 = vpx_minmax_8x8_c; + if (flags & HAS_NEON) vpx_minmax_8x8 = vpx_minmax_8x8_neon; + vpx_mse16x16 = vpx_mse16x16_c; + if (flags & HAS_NEON) vpx_mse16x16 = vpx_mse16x16_neon; + vpx_mse16x8 = vpx_mse16x8_c; + if (flags & HAS_NEON) vpx_mse16x8 = vpx_mse16x8_neon; + vpx_mse8x16 = vpx_mse8x16_c; + if (flags & HAS_NEON) vpx_mse8x16 = vpx_mse8x16_neon; + vpx_mse8x8 = vpx_mse8x8_c; + if (flags & HAS_NEON) vpx_mse8x8 = vpx_mse8x8_neon; + vpx_quantize_b = vpx_quantize_b_c; + if (flags & HAS_NEON) vpx_quantize_b = vpx_quantize_b_neon; + vpx_quantize_b_32x32 = vpx_quantize_b_32x32_c; + if (flags & HAS_NEON) vpx_quantize_b_32x32 = vpx_quantize_b_32x32_neon; + vpx_sad16x16 = vpx_sad16x16_c; + if (flags & HAS_NEON) vpx_sad16x16 = vpx_sad16x16_neon; + vpx_sad16x16_avg = vpx_sad16x16_avg_c; + if (flags & HAS_NEON) vpx_sad16x16_avg = vpx_sad16x16_avg_neon; + vpx_sad16x16x4d = vpx_sad16x16x4d_c; + if (flags & HAS_NEON) vpx_sad16x16x4d = vpx_sad16x16x4d_neon; + vpx_sad16x32 = vpx_sad16x32_c; + if (flags & HAS_NEON) vpx_sad16x32 = vpx_sad16x32_neon; + vpx_sad16x32_avg = vpx_sad16x32_avg_c; + if (flags & HAS_NEON) vpx_sad16x32_avg = vpx_sad16x32_avg_neon; + vpx_sad16x32x4d = vpx_sad16x32x4d_c; + if (flags & HAS_NEON) vpx_sad16x32x4d = vpx_sad16x32x4d_neon; + vpx_sad16x8 = vpx_sad16x8_c; + if (flags & HAS_NEON) vpx_sad16x8 = vpx_sad16x8_neon; + vpx_sad16x8_avg = vpx_sad16x8_avg_c; + if (flags & HAS_NEON) vpx_sad16x8_avg = vpx_sad16x8_avg_neon; + vpx_sad16x8x4d = vpx_sad16x8x4d_c; + if (flags & HAS_NEON) vpx_sad16x8x4d = vpx_sad16x8x4d_neon; + vpx_sad32x16 = vpx_sad32x16_c; + if (flags & HAS_NEON) vpx_sad32x16 = vpx_sad32x16_neon; + vpx_sad32x16_avg = vpx_sad32x16_avg_c; + if (flags & HAS_NEON) vpx_sad32x16_avg = vpx_sad32x16_avg_neon; + vpx_sad32x16x4d = vpx_sad32x16x4d_c; + if (flags & HAS_NEON) vpx_sad32x16x4d = vpx_sad32x16x4d_neon; + vpx_sad32x32 = vpx_sad32x32_c; + if (flags & HAS_NEON) vpx_sad32x32 = vpx_sad32x32_neon; + vpx_sad32x32_avg = vpx_sad32x32_avg_c; + if (flags & HAS_NEON) vpx_sad32x32_avg = vpx_sad32x32_avg_neon; + vpx_sad32x32x4d = vpx_sad32x32x4d_c; + if (flags & HAS_NEON) vpx_sad32x32x4d = vpx_sad32x32x4d_neon; + vpx_sad32x64 = vpx_sad32x64_c; + if (flags & HAS_NEON) vpx_sad32x64 = vpx_sad32x64_neon; + vpx_sad32x64_avg = vpx_sad32x64_avg_c; + if (flags & HAS_NEON) vpx_sad32x64_avg = vpx_sad32x64_avg_neon; + vpx_sad32x64x4d = vpx_sad32x64x4d_c; + if (flags & HAS_NEON) vpx_sad32x64x4d = vpx_sad32x64x4d_neon; + vpx_sad4x4 = vpx_sad4x4_c; + if (flags & HAS_NEON) vpx_sad4x4 = vpx_sad4x4_neon; + vpx_sad4x4_avg = vpx_sad4x4_avg_c; + if (flags & HAS_NEON) vpx_sad4x4_avg = vpx_sad4x4_avg_neon; + vpx_sad4x4x4d = vpx_sad4x4x4d_c; + if (flags & HAS_NEON) vpx_sad4x4x4d = vpx_sad4x4x4d_neon; + vpx_sad4x8 = vpx_sad4x8_c; + if (flags & HAS_NEON) vpx_sad4x8 = vpx_sad4x8_neon; + vpx_sad4x8_avg = vpx_sad4x8_avg_c; + if (flags & HAS_NEON) vpx_sad4x8_avg = vpx_sad4x8_avg_neon; + vpx_sad4x8x4d = vpx_sad4x8x4d_c; + if (flags & HAS_NEON) vpx_sad4x8x4d = vpx_sad4x8x4d_neon; + vpx_sad64x32 = vpx_sad64x32_c; + if (flags & HAS_NEON) vpx_sad64x32 = vpx_sad64x32_neon; + vpx_sad64x32_avg = vpx_sad64x32_avg_c; + if (flags & HAS_NEON) vpx_sad64x32_avg = vpx_sad64x32_avg_neon; + vpx_sad64x32x4d = vpx_sad64x32x4d_c; + if (flags & HAS_NEON) vpx_sad64x32x4d = vpx_sad64x32x4d_neon; + vpx_sad64x64 = vpx_sad64x64_c; + if (flags & HAS_NEON) vpx_sad64x64 = vpx_sad64x64_neon; + vpx_sad64x64_avg = vpx_sad64x64_avg_c; + if (flags & HAS_NEON) vpx_sad64x64_avg = vpx_sad64x64_avg_neon; + vpx_sad64x64x4d = vpx_sad64x64x4d_c; + if (flags & HAS_NEON) vpx_sad64x64x4d = vpx_sad64x64x4d_neon; + vpx_sad8x16 = vpx_sad8x16_c; + if (flags & HAS_NEON) vpx_sad8x16 = vpx_sad8x16_neon; + vpx_sad8x16_avg = vpx_sad8x16_avg_c; + if (flags & HAS_NEON) vpx_sad8x16_avg = vpx_sad8x16_avg_neon; + vpx_sad8x16x4d = vpx_sad8x16x4d_c; + if (flags & HAS_NEON) vpx_sad8x16x4d = vpx_sad8x16x4d_neon; + vpx_sad8x4 = vpx_sad8x4_c; + if (flags & HAS_NEON) vpx_sad8x4 = vpx_sad8x4_neon; + vpx_sad8x4_avg = vpx_sad8x4_avg_c; + if (flags & HAS_NEON) vpx_sad8x4_avg = vpx_sad8x4_avg_neon; + vpx_sad8x4x4d = vpx_sad8x4x4d_c; + if (flags & HAS_NEON) vpx_sad8x4x4d = vpx_sad8x4x4d_neon; + vpx_sad8x8 = vpx_sad8x8_c; + if (flags & HAS_NEON) vpx_sad8x8 = vpx_sad8x8_neon; + vpx_sad8x8_avg = vpx_sad8x8_avg_c; + if (flags & HAS_NEON) vpx_sad8x8_avg = vpx_sad8x8_avg_neon; + vpx_sad8x8x4d = vpx_sad8x8x4d_c; + if (flags & HAS_NEON) vpx_sad8x8x4d = vpx_sad8x8x4d_neon; + vpx_sad_skip_16x16 = vpx_sad_skip_16x16_c; + if (flags & HAS_NEON) vpx_sad_skip_16x16 = vpx_sad_skip_16x16_neon; + vpx_sad_skip_16x16x4d = vpx_sad_skip_16x16x4d_c; + if (flags & HAS_NEON) vpx_sad_skip_16x16x4d = vpx_sad_skip_16x16x4d_neon; + vpx_sad_skip_16x32 = vpx_sad_skip_16x32_c; + if (flags & HAS_NEON) vpx_sad_skip_16x32 = vpx_sad_skip_16x32_neon; + vpx_sad_skip_16x32x4d = vpx_sad_skip_16x32x4d_c; + if (flags & HAS_NEON) vpx_sad_skip_16x32x4d = vpx_sad_skip_16x32x4d_neon; + vpx_sad_skip_16x8 = vpx_sad_skip_16x8_c; + if (flags & HAS_NEON) vpx_sad_skip_16x8 = vpx_sad_skip_16x8_neon; + vpx_sad_skip_16x8x4d = vpx_sad_skip_16x8x4d_c; + if (flags & HAS_NEON) vpx_sad_skip_16x8x4d = vpx_sad_skip_16x8x4d_neon; + vpx_sad_skip_32x16 = vpx_sad_skip_32x16_c; + if (flags & HAS_NEON) vpx_sad_skip_32x16 = vpx_sad_skip_32x16_neon; + vpx_sad_skip_32x16x4d = vpx_sad_skip_32x16x4d_c; + if (flags & HAS_NEON) vpx_sad_skip_32x16x4d = vpx_sad_skip_32x16x4d_neon; + vpx_sad_skip_32x32 = vpx_sad_skip_32x32_c; + if (flags & HAS_NEON) vpx_sad_skip_32x32 = vpx_sad_skip_32x32_neon; + vpx_sad_skip_32x32x4d = vpx_sad_skip_32x32x4d_c; + if (flags & HAS_NEON) vpx_sad_skip_32x32x4d = vpx_sad_skip_32x32x4d_neon; + vpx_sad_skip_32x64 = vpx_sad_skip_32x64_c; + if (flags & HAS_NEON) vpx_sad_skip_32x64 = vpx_sad_skip_32x64_neon; + vpx_sad_skip_32x64x4d = vpx_sad_skip_32x64x4d_c; + if (flags & HAS_NEON) vpx_sad_skip_32x64x4d = vpx_sad_skip_32x64x4d_neon; + vpx_sad_skip_4x4 = vpx_sad_skip_4x4_c; + if (flags & HAS_NEON) vpx_sad_skip_4x4 = vpx_sad_skip_4x4_neon; + vpx_sad_skip_4x4x4d = vpx_sad_skip_4x4x4d_c; + if (flags & HAS_NEON) vpx_sad_skip_4x4x4d = vpx_sad_skip_4x4x4d_neon; + vpx_sad_skip_4x8 = vpx_sad_skip_4x8_c; + if (flags & HAS_NEON) vpx_sad_skip_4x8 = vpx_sad_skip_4x8_neon; + vpx_sad_skip_4x8x4d = vpx_sad_skip_4x8x4d_c; + if (flags & HAS_NEON) vpx_sad_skip_4x8x4d = vpx_sad_skip_4x8x4d_neon; + vpx_sad_skip_64x32 = vpx_sad_skip_64x32_c; + if (flags & HAS_NEON) vpx_sad_skip_64x32 = vpx_sad_skip_64x32_neon; + vpx_sad_skip_64x32x4d = vpx_sad_skip_64x32x4d_c; + if (flags & HAS_NEON) vpx_sad_skip_64x32x4d = vpx_sad_skip_64x32x4d_neon; + vpx_sad_skip_64x64 = vpx_sad_skip_64x64_c; + if (flags & HAS_NEON) vpx_sad_skip_64x64 = vpx_sad_skip_64x64_neon; + vpx_sad_skip_64x64x4d = vpx_sad_skip_64x64x4d_c; + if (flags & HAS_NEON) vpx_sad_skip_64x64x4d = vpx_sad_skip_64x64x4d_neon; + vpx_sad_skip_8x16 = vpx_sad_skip_8x16_c; + if (flags & HAS_NEON) vpx_sad_skip_8x16 = vpx_sad_skip_8x16_neon; + vpx_sad_skip_8x16x4d = vpx_sad_skip_8x16x4d_c; + if (flags & HAS_NEON) vpx_sad_skip_8x16x4d = vpx_sad_skip_8x16x4d_neon; + vpx_sad_skip_8x4 = vpx_sad_skip_8x4_c; + if (flags & HAS_NEON) vpx_sad_skip_8x4 = vpx_sad_skip_8x4_neon; + vpx_sad_skip_8x4x4d = vpx_sad_skip_8x4x4d_c; + if (flags & HAS_NEON) vpx_sad_skip_8x4x4d = vpx_sad_skip_8x4x4d_neon; + vpx_sad_skip_8x8 = vpx_sad_skip_8x8_c; + if (flags & HAS_NEON) vpx_sad_skip_8x8 = vpx_sad_skip_8x8_neon; + vpx_sad_skip_8x8x4d = vpx_sad_skip_8x8x4d_c; + if (flags & HAS_NEON) vpx_sad_skip_8x8x4d = vpx_sad_skip_8x8x4d_neon; + vpx_satd = vpx_satd_c; + if (flags & HAS_NEON) vpx_satd = vpx_satd_neon; + vpx_scaled_2d = vpx_scaled_2d_c; + if (flags & HAS_NEON) vpx_scaled_2d = vpx_scaled_2d_neon; + vpx_sse = vpx_sse_c; + if (flags & HAS_NEON) vpx_sse = vpx_sse_neon; + vpx_sub_pixel_avg_variance16x16 = vpx_sub_pixel_avg_variance16x16_c; + if (flags & HAS_NEON) vpx_sub_pixel_avg_variance16x16 = vpx_sub_pixel_avg_variance16x16_neon; + vpx_sub_pixel_avg_variance16x32 = vpx_sub_pixel_avg_variance16x32_c; + if (flags & HAS_NEON) vpx_sub_pixel_avg_variance16x32 = vpx_sub_pixel_avg_variance16x32_neon; + vpx_sub_pixel_avg_variance16x8 = vpx_sub_pixel_avg_variance16x8_c; + if (flags & HAS_NEON) vpx_sub_pixel_avg_variance16x8 = vpx_sub_pixel_avg_variance16x8_neon; + vpx_sub_pixel_avg_variance32x16 = vpx_sub_pixel_avg_variance32x16_c; + if (flags & HAS_NEON) vpx_sub_pixel_avg_variance32x16 = vpx_sub_pixel_avg_variance32x16_neon; + vpx_sub_pixel_avg_variance32x32 = vpx_sub_pixel_avg_variance32x32_c; + if (flags & HAS_NEON) vpx_sub_pixel_avg_variance32x32 = vpx_sub_pixel_avg_variance32x32_neon; + vpx_sub_pixel_avg_variance32x64 = vpx_sub_pixel_avg_variance32x64_c; + if (flags & HAS_NEON) vpx_sub_pixel_avg_variance32x64 = vpx_sub_pixel_avg_variance32x64_neon; + vpx_sub_pixel_avg_variance4x4 = vpx_sub_pixel_avg_variance4x4_c; + if (flags & HAS_NEON) vpx_sub_pixel_avg_variance4x4 = vpx_sub_pixel_avg_variance4x4_neon; + vpx_sub_pixel_avg_variance4x8 = vpx_sub_pixel_avg_variance4x8_c; + if (flags & HAS_NEON) vpx_sub_pixel_avg_variance4x8 = vpx_sub_pixel_avg_variance4x8_neon; + vpx_sub_pixel_avg_variance64x32 = vpx_sub_pixel_avg_variance64x32_c; + if (flags & HAS_NEON) vpx_sub_pixel_avg_variance64x32 = vpx_sub_pixel_avg_variance64x32_neon; + vpx_sub_pixel_avg_variance64x64 = vpx_sub_pixel_avg_variance64x64_c; + if (flags & HAS_NEON) vpx_sub_pixel_avg_variance64x64 = vpx_sub_pixel_avg_variance64x64_neon; + vpx_sub_pixel_avg_variance8x16 = vpx_sub_pixel_avg_variance8x16_c; + if (flags & HAS_NEON) vpx_sub_pixel_avg_variance8x16 = vpx_sub_pixel_avg_variance8x16_neon; + vpx_sub_pixel_avg_variance8x4 = vpx_sub_pixel_avg_variance8x4_c; + if (flags & HAS_NEON) vpx_sub_pixel_avg_variance8x4 = vpx_sub_pixel_avg_variance8x4_neon; + vpx_sub_pixel_avg_variance8x8 = vpx_sub_pixel_avg_variance8x8_c; + if (flags & HAS_NEON) vpx_sub_pixel_avg_variance8x8 = vpx_sub_pixel_avg_variance8x8_neon; + vpx_sub_pixel_variance16x16 = vpx_sub_pixel_variance16x16_c; + if (flags & HAS_NEON) vpx_sub_pixel_variance16x16 = vpx_sub_pixel_variance16x16_neon; + vpx_sub_pixel_variance16x32 = vpx_sub_pixel_variance16x32_c; + if (flags & HAS_NEON) vpx_sub_pixel_variance16x32 = vpx_sub_pixel_variance16x32_neon; + vpx_sub_pixel_variance16x8 = vpx_sub_pixel_variance16x8_c; + if (flags & HAS_NEON) vpx_sub_pixel_variance16x8 = vpx_sub_pixel_variance16x8_neon; + vpx_sub_pixel_variance32x16 = vpx_sub_pixel_variance32x16_c; + if (flags & HAS_NEON) vpx_sub_pixel_variance32x16 = vpx_sub_pixel_variance32x16_neon; + vpx_sub_pixel_variance32x32 = vpx_sub_pixel_variance32x32_c; + if (flags & HAS_NEON) vpx_sub_pixel_variance32x32 = vpx_sub_pixel_variance32x32_neon; + vpx_sub_pixel_variance32x64 = vpx_sub_pixel_variance32x64_c; + if (flags & HAS_NEON) vpx_sub_pixel_variance32x64 = vpx_sub_pixel_variance32x64_neon; + vpx_sub_pixel_variance4x4 = vpx_sub_pixel_variance4x4_c; + if (flags & HAS_NEON) vpx_sub_pixel_variance4x4 = vpx_sub_pixel_variance4x4_neon; + vpx_sub_pixel_variance4x8 = vpx_sub_pixel_variance4x8_c; + if (flags & HAS_NEON) vpx_sub_pixel_variance4x8 = vpx_sub_pixel_variance4x8_neon; + vpx_sub_pixel_variance64x32 = vpx_sub_pixel_variance64x32_c; + if (flags & HAS_NEON) vpx_sub_pixel_variance64x32 = vpx_sub_pixel_variance64x32_neon; + vpx_sub_pixel_variance64x64 = vpx_sub_pixel_variance64x64_c; + if (flags & HAS_NEON) vpx_sub_pixel_variance64x64 = vpx_sub_pixel_variance64x64_neon; + vpx_sub_pixel_variance8x16 = vpx_sub_pixel_variance8x16_c; + if (flags & HAS_NEON) vpx_sub_pixel_variance8x16 = vpx_sub_pixel_variance8x16_neon; + vpx_sub_pixel_variance8x4 = vpx_sub_pixel_variance8x4_c; + if (flags & HAS_NEON) vpx_sub_pixel_variance8x4 = vpx_sub_pixel_variance8x4_neon; + vpx_sub_pixel_variance8x8 = vpx_sub_pixel_variance8x8_c; + if (flags & HAS_NEON) vpx_sub_pixel_variance8x8 = vpx_sub_pixel_variance8x8_neon; + vpx_subtract_block = vpx_subtract_block_c; + if (flags & HAS_NEON) vpx_subtract_block = vpx_subtract_block_neon; + vpx_sum_squares_2d_i16 = vpx_sum_squares_2d_i16_c; + if (flags & HAS_NEON) vpx_sum_squares_2d_i16 = vpx_sum_squares_2d_i16_neon; + vpx_tm_predictor_16x16 = vpx_tm_predictor_16x16_c; + if (flags & HAS_NEON) vpx_tm_predictor_16x16 = vpx_tm_predictor_16x16_neon; + vpx_tm_predictor_32x32 = vpx_tm_predictor_32x32_c; + if (flags & HAS_NEON) vpx_tm_predictor_32x32 = vpx_tm_predictor_32x32_neon; + vpx_tm_predictor_4x4 = vpx_tm_predictor_4x4_c; + if (flags & HAS_NEON) vpx_tm_predictor_4x4 = vpx_tm_predictor_4x4_neon; + vpx_tm_predictor_8x8 = vpx_tm_predictor_8x8_c; + if (flags & HAS_NEON) vpx_tm_predictor_8x8 = vpx_tm_predictor_8x8_neon; + vpx_v_predictor_16x16 = vpx_v_predictor_16x16_c; + if (flags & HAS_NEON) vpx_v_predictor_16x16 = vpx_v_predictor_16x16_neon; + vpx_v_predictor_32x32 = vpx_v_predictor_32x32_c; + if (flags & HAS_NEON) vpx_v_predictor_32x32 = vpx_v_predictor_32x32_neon; + vpx_v_predictor_4x4 = vpx_v_predictor_4x4_c; + if (flags & HAS_NEON) vpx_v_predictor_4x4 = vpx_v_predictor_4x4_neon; + vpx_v_predictor_8x8 = vpx_v_predictor_8x8_c; + if (flags & HAS_NEON) vpx_v_predictor_8x8 = vpx_v_predictor_8x8_neon; + vpx_variance16x16 = vpx_variance16x16_c; + if (flags & HAS_NEON) vpx_variance16x16 = vpx_variance16x16_neon; + vpx_variance16x32 = vpx_variance16x32_c; + if (flags & HAS_NEON) vpx_variance16x32 = vpx_variance16x32_neon; + vpx_variance16x8 = vpx_variance16x8_c; + if (flags & HAS_NEON) vpx_variance16x8 = vpx_variance16x8_neon; + vpx_variance32x16 = vpx_variance32x16_c; + if (flags & HAS_NEON) vpx_variance32x16 = vpx_variance32x16_neon; + vpx_variance32x32 = vpx_variance32x32_c; + if (flags & HAS_NEON) vpx_variance32x32 = vpx_variance32x32_neon; + vpx_variance32x64 = vpx_variance32x64_c; + if (flags & HAS_NEON) vpx_variance32x64 = vpx_variance32x64_neon; + vpx_variance4x4 = vpx_variance4x4_c; + if (flags & HAS_NEON) vpx_variance4x4 = vpx_variance4x4_neon; + vpx_variance4x8 = vpx_variance4x8_c; + if (flags & HAS_NEON) vpx_variance4x8 = vpx_variance4x8_neon; + vpx_variance64x32 = vpx_variance64x32_c; + if (flags & HAS_NEON) vpx_variance64x32 = vpx_variance64x32_neon; + vpx_variance64x64 = vpx_variance64x64_c; + if (flags & HAS_NEON) vpx_variance64x64 = vpx_variance64x64_neon; + vpx_variance8x16 = vpx_variance8x16_c; + if (flags & HAS_NEON) vpx_variance8x16 = vpx_variance8x16_neon; + vpx_variance8x4 = vpx_variance8x4_c; + if (flags & HAS_NEON) vpx_variance8x4 = vpx_variance8x4_neon; + vpx_variance8x8 = vpx_variance8x8_c; + if (flags & HAS_NEON) vpx_variance8x8 = vpx_variance8x8_neon; + vpx_vector_var = vpx_vector_var_c; + if (flags & HAS_NEON) vpx_vector_var = vpx_vector_var_neon; +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/arm/vpx_scale_rtcd.h b/media/libvpx/config/linux/arm/vpx_scale_rtcd.h new file mode 100644 index 0000000000..b371368275 --- /dev/null +++ b/media/libvpx/config/linux/arm/vpx_scale_rtcd.h @@ -0,0 +1,75 @@ +// This file is generated. Do not edit. +#ifndef VPX_SCALE_RTCD_H_ +#define VPX_SCALE_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +struct yv12_buffer_config; + +#ifdef __cplusplus +extern "C" { +#endif + +void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width); +#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c + +void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width); +#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c + +void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width); +#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c + +void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c + +void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c + +void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c + +void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c + +void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc); +#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c + +void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf); +#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c + +void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf); +#define vpx_extend_frame_borders vpx_extend_frame_borders_c + +void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf); +#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c + +void vpx_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc); +#define vpx_yv12_copy_frame vpx_yv12_copy_frame_c + +void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc); +#define vpx_yv12_copy_y vpx_yv12_copy_y_c + +void vpx_scale_rtcd(void); + +#include "vpx_config.h" + +#ifdef RTCD_C +#include "vpx_ports/arm.h" +static void setup_rtcd_internal(void) +{ + int flags = arm_cpu_caps(); + + (void)flags; + +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/arm64/vp8_rtcd.h b/media/libvpx/config/linux/arm64/vp8_rtcd.h new file mode 100644 index 0000000000..d204ef751e --- /dev/null +++ b/media/libvpx/config/linux/arm64/vp8_rtcd.h @@ -0,0 +1,201 @@ +// This file is generated. Do not edit. +#ifndef VP8_RTCD_H_ +#define VP8_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +/* + * VP8 + */ + +struct blockd; +struct macroblockd; +struct loop_filter_info; + +/* Encoder forward decls */ +struct block; +struct macroblock; +struct variance_vtable; +union int_mv; +struct yv12_buffer_config; + +#ifdef __cplusplus +extern "C" { +#endif + +void vp8_bilinear_predict16x16_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict16x16_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +#define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_neon + +void vp8_bilinear_predict4x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict4x4_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_neon + +void vp8_bilinear_predict8x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict8x4_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_neon + +void vp8_bilinear_predict8x8_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict8x8_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +#define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_neon + +int vp8_block_error_c(short *coeff, short *dqcoeff); +#define vp8_block_error vp8_block_error_c + +void vp8_copy32xn_c(const unsigned char *src_ptr, int src_stride, unsigned char *dst_ptr, int dst_stride, int height); +#define vp8_copy32xn vp8_copy32xn_c + +void vp8_copy_mem16x16_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +void vp8_copy_mem16x16_neon(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +#define vp8_copy_mem16x16 vp8_copy_mem16x16_neon + +void vp8_copy_mem8x4_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +void vp8_copy_mem8x4_neon(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +#define vp8_copy_mem8x4 vp8_copy_mem8x4_neon + +void vp8_copy_mem8x8_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +void vp8_copy_mem8x8_neon(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +#define vp8_copy_mem8x8 vp8_copy_mem8x8_neon + +void vp8_dc_only_idct_add_c(short input_dc, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +void vp8_dc_only_idct_add_neon(short input_dc, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +#define vp8_dc_only_idct_add vp8_dc_only_idct_add_neon + +int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +int vp8_denoiser_filter_neon(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +#define vp8_denoiser_filter vp8_denoiser_filter_neon + +int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +int vp8_denoiser_filter_uv_neon(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_neon + +void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *dest, int stride); +void vp8_dequant_idct_add_neon(short *input, short *dq, unsigned char *dest, int stride); +#define vp8_dequant_idct_add vp8_dequant_idct_add_neon + +void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs); +void vp8_dequant_idct_add_uv_block_neon(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs); +#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_neon + +void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs); +void vp8_dequant_idct_add_y_block_neon(short *q, short *dq, unsigned char *dst, int stride, char *eobs); +#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_neon + +void vp8_dequantize_b_c(struct blockd*, short *DQC); +void vp8_dequantize_b_neon(struct blockd*, short *DQC); +#define vp8_dequantize_b vp8_dequantize_b_neon + +int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); +#define vp8_diamond_search_sad vp8_diamond_search_sad_c + +void vp8_fast_quantize_b_c(struct block *, struct blockd *); +void vp8_fast_quantize_b_neon(struct block *, struct blockd *); +#define vp8_fast_quantize_b vp8_fast_quantize_b_neon + +void vp8_loop_filter_bh_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_bh_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +#define vp8_loop_filter_bh vp8_loop_filter_bh_neon + +void vp8_loop_filter_bv_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_bv_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +#define vp8_loop_filter_bv vp8_loop_filter_bv_neon + +void vp8_loop_filter_mbh_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_mbh_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +#define vp8_loop_filter_mbh vp8_loop_filter_mbh_neon + +void vp8_loop_filter_mbv_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_mbv_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +#define vp8_loop_filter_mbv vp8_loop_filter_mbv_neon + +void vp8_loop_filter_bhs_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_bhs_neon(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_neon + +void vp8_loop_filter_bvs_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_bvs_neon(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_neon + +void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_mbhs_neon(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +#define vp8_loop_filter_simple_mbh vp8_loop_filter_mbhs_neon + +void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_mbvs_neon(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +#define vp8_loop_filter_simple_mbv vp8_loop_filter_mbvs_neon + +int vp8_mbblock_error_c(struct macroblock *mb, int dc); +#define vp8_mbblock_error vp8_mbblock_error_c + +int vp8_mbuverror_c(struct macroblock *mb); +#define vp8_mbuverror vp8_mbuverror_c + +int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int error_per_bit, int search_range, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); +#define vp8_refining_search_sad vp8_refining_search_sad_c + +void vp8_regular_quantize_b_c(struct block *, struct blockd *); +#define vp8_regular_quantize_b vp8_regular_quantize_b_c + +void vp8_short_fdct4x4_c(short *input, short *output, int pitch); +void vp8_short_fdct4x4_neon(short *input, short *output, int pitch); +#define vp8_short_fdct4x4 vp8_short_fdct4x4_neon + +void vp8_short_fdct8x4_c(short *input, short *output, int pitch); +void vp8_short_fdct8x4_neon(short *input, short *output, int pitch); +#define vp8_short_fdct8x4 vp8_short_fdct8x4_neon + +void vp8_short_idct4x4llm_c(short *input, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +void vp8_short_idct4x4llm_neon(short *input, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +#define vp8_short_idct4x4llm vp8_short_idct4x4llm_neon + +void vp8_short_inv_walsh4x4_c(short *input, short *mb_dqcoeff); +void vp8_short_inv_walsh4x4_neon(short *input, short *mb_dqcoeff); +#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_neon + +void vp8_short_inv_walsh4x4_1_c(short *input, short *mb_dqcoeff); +#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c + +void vp8_short_walsh4x4_c(short *input, short *output, int pitch); +void vp8_short_walsh4x4_neon(short *input, short *output, int pitch); +#define vp8_short_walsh4x4 vp8_short_walsh4x4_neon + +void vp8_sixtap_predict16x16_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict16x16_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +#define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_neon + +void vp8_sixtap_predict4x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict4x4_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_neon + +void vp8_sixtap_predict8x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict8x4_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +#define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_neon + +void vp8_sixtap_predict8x8_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict8x8_neon(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +#define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_neon + +void vp8_rtcd(void); + +#include "vpx_config.h" + +#ifdef RTCD_C +#include "vpx_ports/arm.h" +static void setup_rtcd_internal(void) +{ + int flags = arm_cpu_caps(); + + (void)flags; + +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/arm64/vp9_rtcd.h b/media/libvpx/config/linux/arm64/vp9_rtcd.h new file mode 100644 index 0000000000..738de4f9f4 --- /dev/null +++ b/media/libvpx/config/linux/arm64/vp9_rtcd.h @@ -0,0 +1,106 @@ +// This file is generated. Do not edit. +#ifndef VP9_RTCD_H_ +#define VP9_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +/* + * VP9 + */ + +#include "vpx/vpx_integer.h" +#include "vp9/common/vp9_common.h" +#include "vp9/common/vp9_enums.h" +#include "vp9/common/vp9_filter.h" + +struct macroblockd; + +/* Encoder forward decls */ +struct macroblock; +struct macroblock_plane; +struct vp9_sad_table; +struct ScanOrder; +struct search_site_config; +struct mv; +union int_mv; +struct yv12_buffer_config; + +#ifdef __cplusplus +extern "C" { +#endif + +int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz); +int64_t vp9_block_error_neon(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz); +#define vp9_block_error vp9_block_error_neon + +int64_t vp9_block_error_fp_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size); +int64_t vp9_block_error_fp_neon(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size); +#define vp9_block_error_fp vp9_block_error_fp_neon + +int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, uint32_t start_mv_sad, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_sad_table *sad_fn_ptr, const struct mv *center_mv); +int vp9_diamond_search_sad_neon(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, uint32_t start_mv_sad, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_sad_table *sad_fn_ptr, const struct mv *center_mv); +#define vp9_diamond_search_sad vp9_diamond_search_sad_neon + +void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type); +void vp9_fht16x16_neon(const int16_t *input, tran_low_t *output, int stride, int tx_type); +#define vp9_fht16x16 vp9_fht16x16_neon + +void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type); +void vp9_fht4x4_neon(const int16_t *input, tran_low_t *output, int stride, int tx_type); +#define vp9_fht4x4 vp9_fht4x4_neon + +void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type); +void vp9_fht8x8_neon(const int16_t *input, tran_low_t *output, int stride, int tx_type); +#define vp9_fht8x8 vp9_fht8x8_neon + +void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride); +#define vp9_fwht4x4 vp9_fwht4x4_c + +void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +void vp9_iht16x16_256_add_neon(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +#define vp9_iht16x16_256_add vp9_iht16x16_256_add_neon + +void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +void vp9_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +#define vp9_iht4x4_16_add vp9_iht4x4_16_add_neon + +void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +void vp9_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +#define vp9_iht8x8_64_add vp9_iht8x8_64_add_neon + +void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +#define vp9_quantize_fp vp9_quantize_fp_neon + +void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_neon + +void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst, INTERP_FILTER filter_type, int phase_scaler); +void vp9_scale_and_extend_frame_neon(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst, INTERP_FILTER filter_type, int phase_scaler); +#define vp9_scale_and_extend_frame vp9_scale_and_extend_frame_neon + +void vp9_rtcd(void); + +#include "vpx_config.h" + +#ifdef RTCD_C +#include "vpx_ports/arm.h" +static void setup_rtcd_internal(void) +{ + int flags = arm_cpu_caps(); + + (void)flags; + +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/arm64/vpx_config.asm b/media/libvpx/config/linux/arm64/vpx_config.asm new file mode 100644 index 0000000000..499c16202c --- /dev/null +++ b/media/libvpx/config/linux/arm64/vpx_config.asm @@ -0,0 +1,97 @@ +@ This file was created from a .asm file +@ using the ads2gas.pl script. +.syntax unified +.equ VPX_ARCH_ARM , 1 +.equ VPX_ARCH_AARCH64 , 1 +.equ VPX_ARCH_MIPS , 0 +.equ VPX_ARCH_X86 , 0 +.equ VPX_ARCH_X86_64 , 0 +.equ VPX_ARCH_PPC , 0 +.equ VPX_ARCH_LOONGARCH , 0 +.equ HAVE_NEON_ASM , 0 +.equ HAVE_NEON , 1 +.equ HAVE_NEON_DOTPROD , 1 +.equ HAVE_NEON_I8MM , 1 +.equ HAVE_SVE , 1 +.equ HAVE_MIPS32 , 0 +.equ HAVE_DSPR2 , 0 +.equ HAVE_MSA , 0 +.equ HAVE_MIPS64 , 0 +.equ HAVE_MMX , 0 +.equ HAVE_SSE , 0 +.equ HAVE_SSE2 , 0 +.equ HAVE_SSE3 , 0 +.equ HAVE_SSSE3 , 0 +.equ HAVE_SSE4_1 , 0 +.equ HAVE_AVX , 0 +.equ HAVE_AVX2 , 0 +.equ HAVE_AVX512 , 0 +.equ HAVE_VSX , 0 +.equ HAVE_MMI , 0 +.equ HAVE_LSX , 0 +.equ HAVE_LASX , 0 +.equ HAVE_VPX_PORTS , 1 +.equ HAVE_PTHREAD_H , 1 +.equ CONFIG_DEPENDENCY_TRACKING , 1 +.equ CONFIG_EXTERNAL_BUILD , 1 +.equ CONFIG_INSTALL_DOCS , 0 +.equ CONFIG_INSTALL_BINS , 1 +.equ CONFIG_INSTALL_LIBS , 1 +.equ CONFIG_INSTALL_SRCS , 0 +.equ CONFIG_DEBUG , 0 +.equ CONFIG_GPROF , 0 +.equ CONFIG_GCOV , 0 +.equ CONFIG_RVCT , 0 +.equ CONFIG_GCC , 1 +.equ CONFIG_MSVS , 0 +.equ CONFIG_PIC , 1 +.equ CONFIG_BIG_ENDIAN , 0 +.equ CONFIG_CODEC_SRCS , 0 +.equ CONFIG_DEBUG_LIBS , 0 +.equ CONFIG_DEQUANT_TOKENS , 0 +.equ CONFIG_DC_RECON , 0 +.equ CONFIG_RUNTIME_CPU_DETECT , 1 +.equ CONFIG_POSTPROC , 0 +.equ CONFIG_VP9_POSTPROC , 0 +.equ CONFIG_MULTITHREAD , 1 +.equ CONFIG_INTERNAL_STATS , 0 +.equ CONFIG_VP8_ENCODER , 1 +.equ CONFIG_VP8_DECODER , 1 +.equ CONFIG_VP9_ENCODER , 1 +.equ CONFIG_VP9_DECODER , 1 +.equ CONFIG_VP8 , 1 +.equ CONFIG_VP9 , 1 +.equ CONFIG_ENCODERS , 1 +.equ CONFIG_DECODERS , 1 +.equ CONFIG_STATIC_MSVCRT , 0 +.equ CONFIG_SPATIAL_RESAMPLING , 1 +.equ CONFIG_REALTIME_ONLY , 1 +.equ CONFIG_ONTHEFLY_BITPACKING , 0 +.equ CONFIG_ERROR_CONCEALMENT , 0 +.equ CONFIG_SHARED , 0 +.equ CONFIG_STATIC , 1 +.equ CONFIG_SMALL , 0 +.equ CONFIG_POSTPROC_VISUALIZER , 0 +.equ CONFIG_OS_SUPPORT , 1 +.equ CONFIG_UNIT_TESTS , 0 +.equ CONFIG_WEBM_IO , 1 +.equ CONFIG_LIBYUV , 1 +.equ CONFIG_DECODE_PERF_TESTS , 0 +.equ CONFIG_ENCODE_PERF_TESTS , 0 +.equ CONFIG_MULTI_RES_ENCODING , 1 +.equ CONFIG_TEMPORAL_DENOISING , 1 +.equ CONFIG_VP9_TEMPORAL_DENOISING , 0 +.equ CONFIG_COEFFICIENT_RANGE_CHECKING , 0 +.equ CONFIG_VP9_HIGHBITDEPTH , 0 +.equ CONFIG_BETTER_HW_COMPATIBILITY , 0 +.equ CONFIG_EXPERIMENTAL , 0 +.equ CONFIG_SIZE_LIMIT , 1 +.equ CONFIG_ALWAYS_ADJUST_BPM , 0 +.equ CONFIG_BITSTREAM_DEBUG , 0 +.equ CONFIG_MISMATCH_DEBUG , 0 +.equ CONFIG_FP_MB_STATS , 0 +.equ CONFIG_EMULATE_HARDWARE , 0 +.equ CONFIG_NON_GREEDY_MV , 0 +.equ CONFIG_RATE_CTRL , 0 +.equ CONFIG_COLLECT_COMPONENT_TIMING , 0 + .section .note.GNU-stack,"",%progbits diff --git a/media/libvpx/config/linux/arm64/vpx_config.c b/media/libvpx/config/linux/arm64/vpx_config.c new file mode 100644 index 0000000000..74baa0689c --- /dev/null +++ b/media/libvpx/config/linux/arm64/vpx_config.c @@ -0,0 +1,10 @@ +/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */ +/* */ +/* Use of this source code is governed by a BSD-style license */ +/* that can be found in the LICENSE file in the root of the source */ +/* tree. An additional intellectual property rights grant can be found */ +/* in the file PATENTS. All contributing project authors may */ +/* be found in the AUTHORS file in the root of the source tree. */ +#include "vpx/vpx_codec.h" +static const char* const cfg = "--target=arm64-linux-gcc --enable-external-build --disable-examples --disable-install-docs --disable-unit-tests --enable-multi-res-encoding --size-limit=8192x4608 --enable-pic --disable-avx512 --enable-realtime-only"; +const char *vpx_codec_build_config(void) {return cfg;} diff --git a/media/libvpx/config/linux/arm64/vpx_config.h b/media/libvpx/config/linux/arm64/vpx_config.h new file mode 100644 index 0000000000..3c5f2e33ca --- /dev/null +++ b/media/libvpx/config/linux/arm64/vpx_config.h @@ -0,0 +1,108 @@ +/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */ +/* */ +/* Use of this source code is governed by a BSD-style license */ +/* that can be found in the LICENSE file in the root of the source */ +/* tree. An additional intellectual property rights grant can be found */ +/* in the file PATENTS. All contributing project authors may */ +/* be found in the AUTHORS file in the root of the source tree. */ +/* This file automatically generated by configure. Do not edit! */ +#ifndef VPX_CONFIG_H +#define VPX_CONFIG_H +#define RESTRICT +#define INLINE inline +#define VPX_ARCH_ARM 1 +#define VPX_ARCH_AARCH64 1 +#define VPX_ARCH_MIPS 0 +#define VPX_ARCH_X86 0 +#define VPX_ARCH_X86_64 0 +#define VPX_ARCH_PPC 0 +#define VPX_ARCH_LOONGARCH 0 +#define HAVE_NEON_ASM 0 +#define HAVE_NEON 1 +#define HAVE_NEON_DOTPROD 1 +#define HAVE_NEON_I8MM 1 +#define HAVE_SVE 1 +#define HAVE_MIPS32 0 +#define HAVE_DSPR2 0 +#define HAVE_MSA 0 +#define HAVE_MIPS64 0 +#define HAVE_MMX 0 +#define HAVE_SSE 0 +#define HAVE_SSE2 0 +#define HAVE_SSE3 0 +#define HAVE_SSSE3 0 +#define HAVE_SSE4_1 0 +#define HAVE_AVX 0 +#define HAVE_AVX2 0 +#define HAVE_AVX512 0 +#define HAVE_VSX 0 +#define HAVE_MMI 0 +#define HAVE_LSX 0 +#define HAVE_LASX 0 +#define HAVE_VPX_PORTS 1 +#define HAVE_PTHREAD_H 1 +#define CONFIG_DEPENDENCY_TRACKING 1 +#define CONFIG_EXTERNAL_BUILD 1 +#define CONFIG_INSTALL_DOCS 0 +#define CONFIG_INSTALL_BINS 1 +#define CONFIG_INSTALL_LIBS 1 +#define CONFIG_INSTALL_SRCS 0 +#define CONFIG_DEBUG 0 +#define CONFIG_GPROF 0 +#define CONFIG_GCOV 0 +#define CONFIG_RVCT 0 +#define CONFIG_GCC 1 +#define CONFIG_MSVS 0 +#define CONFIG_PIC 1 +#define CONFIG_BIG_ENDIAN 0 +#define CONFIG_CODEC_SRCS 0 +#define CONFIG_DEBUG_LIBS 0 +#define CONFIG_DEQUANT_TOKENS 0 +#define CONFIG_DC_RECON 0 +#define CONFIG_RUNTIME_CPU_DETECT 1 +#define CONFIG_POSTPROC 0 +#define CONFIG_VP9_POSTPROC 0 +#define CONFIG_MULTITHREAD 1 +#define CONFIG_INTERNAL_STATS 0 +#define CONFIG_VP8_ENCODER 1 +#define CONFIG_VP8_DECODER 1 +#define CONFIG_VP9_ENCODER 1 +#define CONFIG_VP9_DECODER 1 +#define CONFIG_VP8 1 +#define CONFIG_VP9 1 +#define CONFIG_ENCODERS 1 +#define CONFIG_DECODERS 1 +#define CONFIG_STATIC_MSVCRT 0 +#define CONFIG_SPATIAL_RESAMPLING 1 +#define CONFIG_REALTIME_ONLY 1 +#define CONFIG_ONTHEFLY_BITPACKING 0 +#define CONFIG_ERROR_CONCEALMENT 0 +#define CONFIG_SHARED 0 +#define CONFIG_STATIC 1 +#define CONFIG_SMALL 0 +#define CONFIG_POSTPROC_VISUALIZER 0 +#define CONFIG_OS_SUPPORT 1 +#define CONFIG_UNIT_TESTS 0 +#define CONFIG_WEBM_IO 1 +#define CONFIG_LIBYUV 1 +#define CONFIG_DECODE_PERF_TESTS 0 +#define CONFIG_ENCODE_PERF_TESTS 0 +#define CONFIG_MULTI_RES_ENCODING 1 +#define CONFIG_TEMPORAL_DENOISING 1 +#define CONFIG_VP9_TEMPORAL_DENOISING 0 +#define CONFIG_COEFFICIENT_RANGE_CHECKING 0 +#define CONFIG_VP9_HIGHBITDEPTH 0 +#define CONFIG_BETTER_HW_COMPATIBILITY 0 +#define CONFIG_EXPERIMENTAL 0 +#define CONFIG_SIZE_LIMIT 1 +#define CONFIG_ALWAYS_ADJUST_BPM 0 +#define CONFIG_BITSTREAM_DEBUG 0 +#define CONFIG_MISMATCH_DEBUG 0 +#define CONFIG_FP_MB_STATS 0 +#define CONFIG_EMULATE_HARDWARE 0 +#define CONFIG_NON_GREEDY_MV 0 +#define CONFIG_RATE_CTRL 0 +#define CONFIG_COLLECT_COMPONENT_TIMING 0 +#define DECODE_WIDTH_LIMIT 8192 +#define DECODE_HEIGHT_LIMIT 4608 +#endif /* VPX_CONFIG_H */ diff --git a/media/libvpx/config/linux/arm64/vpx_dsp_rtcd.h b/media/libvpx/config/linux/arm64/vpx_dsp_rtcd.h new file mode 100644 index 0000000000..5a9b05ca14 --- /dev/null +++ b/media/libvpx/config/linux/arm64/vpx_dsp_rtcd.h @@ -0,0 +1,1184 @@ +// This file is generated. Do not edit. +#ifndef VPX_DSP_RTCD_H_ +#define VPX_DSP_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +/* + * DSP + */ + +#include "vpx/vpx_integer.h" +#include "vpx_dsp/vpx_dsp_common.h" +#include "vpx_dsp/vpx_filter.h" +#if CONFIG_VP9_ENCODER + struct macroblock_plane; + struct ScanOrder; +#endif + + +#ifdef __cplusplus +extern "C" { +#endif + +unsigned int vpx_avg_4x4_c(const uint8_t *, int p); +unsigned int vpx_avg_4x4_neon(const uint8_t *, int p); +#define vpx_avg_4x4 vpx_avg_4x4_neon + +unsigned int vpx_avg_8x8_c(const uint8_t *, int p); +unsigned int vpx_avg_8x8_neon(const uint8_t *, int p); +#define vpx_avg_8x8 vpx_avg_8x8_neon + +void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride); +void vpx_comp_avg_pred_neon(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride); +#define vpx_comp_avg_pred vpx_comp_avg_pred_neon + +void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_avg)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_horiz_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_horiz_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_avg_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_vert_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_vert_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_avg_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_horiz_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_horiz_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_vert_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve_avg_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_convolve_avg vpx_convolve_avg_neon + +void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve_copy_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_convolve_copy vpx_convolve_copy_neon + +void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d117_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_neon + +void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d117_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_neon + +void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d117_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_neon + +void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d117_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_neon + +void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d135_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_neon + +void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d135_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_neon + +void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_neon + +void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d135_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_neon + +void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_neon + +void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_neon + +void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_neon + +void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_neon + +void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_neon + +void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_neon + +void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_neon + +void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_neon + +void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d45_predictor_16x16 vpx_d45_predictor_16x16_neon + +void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d45_predictor_32x32 vpx_d45_predictor_32x32_neon + +void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d45_predictor_4x4 vpx_d45_predictor_4x4_neon + +void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d45_predictor_8x8 vpx_d45_predictor_8x8_neon + +void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c + +void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_neon + +void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_neon + +void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_neon + +void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_neon + +void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c + +void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_128_predictor_16x16 vpx_dc_128_predictor_16x16_neon + +void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_128_predictor_32x32 vpx_dc_128_predictor_32x32_neon + +void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_128_predictor_4x4 vpx_dc_128_predictor_4x4_neon + +void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_128_predictor_8x8 vpx_dc_128_predictor_8x8_neon + +void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_left_predictor_16x16 vpx_dc_left_predictor_16x16_neon + +void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_left_predictor_32x32 vpx_dc_left_predictor_32x32_neon + +void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_left_predictor_4x4 vpx_dc_left_predictor_4x4_neon + +void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_left_predictor_8x8 vpx_dc_left_predictor_8x8_neon + +void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_predictor_16x16 vpx_dc_predictor_16x16_neon + +void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_predictor_32x32 vpx_dc_predictor_32x32_neon + +void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_predictor_4x4 vpx_dc_predictor_4x4_neon + +void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_predictor_8x8 vpx_dc_predictor_8x8_neon + +void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_top_predictor_16x16 vpx_dc_top_predictor_16x16_neon + +void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_top_predictor_32x32 vpx_dc_top_predictor_32x32_neon + +void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_top_predictor_4x4 vpx_dc_top_predictor_4x4_neon + +void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_top_predictor_8x8 vpx_dc_top_predictor_8x8_neon + +void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct16x16_neon(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct16x16 vpx_fdct16x16_neon + +void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct16x16_1_neon(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct16x16_1 vpx_fdct16x16_1_neon + +void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_neon(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct32x32 vpx_fdct32x32_neon + +void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_1_neon(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct32x32_1 vpx_fdct32x32_1_neon + +void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_rd_neon(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct32x32_rd vpx_fdct32x32_rd_neon + +void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct4x4_neon(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct4x4 vpx_fdct4x4_neon + +void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct4x4_1_neon(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct4x4_1 vpx_fdct4x4_1_neon + +void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct8x8_neon(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct8x8 vpx_fdct8x8_neon + +void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct8x8_1_neon(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct8x8_1 vpx_fdct8x8_1_neon + +void vpx_get16x16var_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +void vpx_get16x16var_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +void vpx_get16x16var_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +RTCD_EXTERN void (*vpx_get16x16var)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); + +unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride); +unsigned int vpx_get4x4sse_cs_neon(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride); +unsigned int vpx_get4x4sse_cs_neon_dotprod(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_get4x4sse_cs)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride); + +void vpx_get8x8var_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +void vpx_get8x8var_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +void vpx_get8x8var_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +RTCD_EXTERN void (*vpx_get8x8var)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); + +unsigned int vpx_get_mb_ss_c(const int16_t *); +#define vpx_get_mb_ss vpx_get_mb_ss_c + +void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_h_predictor_16x16 vpx_h_predictor_16x16_neon + +void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_h_predictor_32x32 vpx_h_predictor_32x32_neon + +void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_h_predictor_4x4 vpx_h_predictor_4x4_neon + +void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_h_predictor_8x8 vpx_h_predictor_8x8_neon + +void vpx_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_16x16_neon(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +#define vpx_hadamard_16x16 vpx_hadamard_16x16_neon + +void vpx_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_32x32_neon(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +#define vpx_hadamard_32x32 vpx_hadamard_32x32_neon + +void vpx_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_8x8_neon(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +#define vpx_hadamard_8x8 vpx_hadamard_8x8_neon + +void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c + +void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_10_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct16x16_10_add vpx_idct16x16_10_add_neon + +void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_1_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct16x16_1_add vpx_idct16x16_1_add_neon + +void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_256_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct16x16_256_add vpx_idct16x16_256_add_neon + +void vpx_idct16x16_38_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_38_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct16x16_38_add vpx_idct16x16_38_add_neon + +void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_1024_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct32x32_1024_add vpx_idct32x32_1024_add_neon + +void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_135_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct32x32_135_add vpx_idct32x32_135_add_neon + +void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_1_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct32x32_1_add vpx_idct32x32_1_add_neon + +void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_34_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct32x32_34_add vpx_idct32x32_34_add_neon + +void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct4x4_16_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct4x4_16_add vpx_idct4x4_16_add_neon + +void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct4x4_1_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct4x4_1_add vpx_idct4x4_1_add_neon + +void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_12_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct8x8_12_add vpx_idct8x8_12_add_neon + +void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_1_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct8x8_1_add vpx_idct8x8_1_add_neon + +void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_64_add_neon(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct8x8_64_add vpx_idct8x8_64_add_neon + +int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width); +int16_t vpx_int_pro_col_neon(const uint8_t *ref, const int width); +#define vpx_int_pro_col vpx_int_pro_col_neon + +void vpx_int_pro_row_c(int16_t hbuf[16], const uint8_t *ref, const int ref_stride, const int height); +void vpx_int_pro_row_neon(int16_t hbuf[16], const uint8_t *ref, const int ref_stride, const int height); +#define vpx_int_pro_row vpx_int_pro_row_neon + +void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_c + +void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c + +void vpx_lpf_horizontal_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_16_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_horizontal_16 vpx_lpf_horizontal_16_neon + +void vpx_lpf_horizontal_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_16_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_horizontal_16_dual vpx_lpf_horizontal_16_dual_neon + +void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_4_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_neon + +void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_horizontal_4_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +#define vpx_lpf_horizontal_4_dual vpx_lpf_horizontal_4_dual_neon + +void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_8_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_horizontal_8 vpx_lpf_horizontal_8_neon + +void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_horizontal_8_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +#define vpx_lpf_horizontal_8_dual vpx_lpf_horizontal_8_dual_neon + +void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_16_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_vertical_16 vpx_lpf_vertical_16_neon + +void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_16_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_vertical_16_dual vpx_lpf_vertical_16_dual_neon + +void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_4_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_vertical_4 vpx_lpf_vertical_4_neon + +void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_vertical_4_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +#define vpx_lpf_vertical_4_dual vpx_lpf_vertical_4_dual_neon + +void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_8_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_vertical_8 vpx_lpf_vertical_8_neon + +void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_vertical_8_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +#define vpx_lpf_vertical_8_dual vpx_lpf_vertical_8_dual_neon + +void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max); +void vpx_minmax_8x8_neon(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max); +#define vpx_minmax_8x8 vpx_minmax_8x8_neon + +unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x16_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x8_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse8x16_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse8x8_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +#define vpx_quantize_b vpx_quantize_b_neon + +void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_32x32_neon(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +#define vpx_quantize_b_32x32 vpx_quantize_b_32x32_neon + +unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x16_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x16_avg_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x16x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x32_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x32_avg_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x32x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x8_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x8_avg_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x8x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x16_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x16_avg_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x16x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x32_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x32_avg_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x32x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x64_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x64_avg_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x64x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad4x4 vpx_sad4x4_neon + +unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad4x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +#define vpx_sad4x4_avg vpx_sad4x4_avg_neon + +void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad4x4x4d vpx_sad4x4x4d_neon + +unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad4x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad4x8 vpx_sad4x8_neon + +unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad4x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +#define vpx_sad4x8_avg vpx_sad4x8_avg_neon + +void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad4x8x4d vpx_sad4x8x4d_neon + +unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x32_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x32_avg_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad64x32x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x64_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x64_avg_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad64x64x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad8x16 vpx_sad8x16_neon + +unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad8x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +#define vpx_sad8x16_avg vpx_sad8x16_avg_neon + +void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad8x16x4d vpx_sad8x16x4d_neon + +unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad8x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad8x4 vpx_sad8x4_neon + +unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad8x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +#define vpx_sad8x4_avg vpx_sad8x4_avg_neon + +void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad8x4x4d vpx_sad8x4x4d_neon + +unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad8x8 vpx_sad8x8_neon + +unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad8x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +#define vpx_sad8x8_avg vpx_sad8x8_avg_neon + +void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad8x8x4d vpx_sad8x8x4d_neon + +unsigned int vpx_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x16_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x16x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x32_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x32x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x8_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x8x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x16_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x16x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x32_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x32x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x64_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x64x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_4x4 vpx_sad_skip_4x4_neon + +void vpx_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_4x4x4d vpx_sad_skip_4x4x4d_neon + +unsigned int vpx_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_4x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_4x8 vpx_sad_skip_4x8_neon + +void vpx_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_4x8x4d vpx_sad_skip_4x8x4d_neon + +unsigned int vpx_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x32_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x32x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x64_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x64x4d_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_8x16 vpx_sad_skip_8x16_neon + +void vpx_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_8x16x4d vpx_sad_skip_8x16x4d_neon + +unsigned int vpx_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_8x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_8x4 vpx_sad_skip_8x4_neon + +void vpx_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_8x4x4d vpx_sad_skip_8x4x4d_neon + +unsigned int vpx_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_8x8 vpx_sad_skip_8x8_neon + +void vpx_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_8x8x4d vpx_sad_skip_8x8x4d_neon + +int vpx_satd_c(const int16_t *coeff, int length); +int vpx_satd_neon(const int16_t *coeff, int length); +#define vpx_satd vpx_satd_neon + +void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_scaled_2d_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_2d vpx_scaled_2d_neon + +void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c + +void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c + +void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c + +void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_horiz vpx_scaled_horiz_c + +void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_vert vpx_scaled_vert_c + +int64_t vpx_sse_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); +int64_t vpx_sse_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); +int64_t vpx_sse_neon_dotprod(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); +RTCD_EXTERN int64_t (*vpx_sse)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); + +uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x16_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +#define vpx_sub_pixel_avg_variance16x16 vpx_sub_pixel_avg_variance16x16_neon + +uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x32_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +#define vpx_sub_pixel_avg_variance16x32 vpx_sub_pixel_avg_variance16x32_neon + +uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x8_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +#define vpx_sub_pixel_avg_variance16x8 vpx_sub_pixel_avg_variance16x8_neon + +uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x16_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +#define vpx_sub_pixel_avg_variance32x16 vpx_sub_pixel_avg_variance32x16_neon + +uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x32_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +#define vpx_sub_pixel_avg_variance32x32 vpx_sub_pixel_avg_variance32x32_neon + +uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x64_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +#define vpx_sub_pixel_avg_variance32x64 vpx_sub_pixel_avg_variance32x64_neon + +uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance4x4_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +#define vpx_sub_pixel_avg_variance4x4 vpx_sub_pixel_avg_variance4x4_neon + +uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance4x8_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +#define vpx_sub_pixel_avg_variance4x8 vpx_sub_pixel_avg_variance4x8_neon + +uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x32_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +#define vpx_sub_pixel_avg_variance64x32 vpx_sub_pixel_avg_variance64x32_neon + +uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x64_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +#define vpx_sub_pixel_avg_variance64x64 vpx_sub_pixel_avg_variance64x64_neon + +uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x16_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +#define vpx_sub_pixel_avg_variance8x16 vpx_sub_pixel_avg_variance8x16_neon + +uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x4_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +#define vpx_sub_pixel_avg_variance8x4 vpx_sub_pixel_avg_variance8x4_neon + +uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x8_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +#define vpx_sub_pixel_avg_variance8x8 vpx_sub_pixel_avg_variance8x8_neon + +uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x16_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +#define vpx_sub_pixel_variance16x16 vpx_sub_pixel_variance16x16_neon + +uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x32_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +#define vpx_sub_pixel_variance16x32 vpx_sub_pixel_variance16x32_neon + +uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x8_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +#define vpx_sub_pixel_variance16x8 vpx_sub_pixel_variance16x8_neon + +uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x16_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +#define vpx_sub_pixel_variance32x16 vpx_sub_pixel_variance32x16_neon + +uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x32_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +#define vpx_sub_pixel_variance32x32 vpx_sub_pixel_variance32x32_neon + +uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x64_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +#define vpx_sub_pixel_variance32x64 vpx_sub_pixel_variance32x64_neon + +uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance4x4_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +#define vpx_sub_pixel_variance4x4 vpx_sub_pixel_variance4x4_neon + +uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance4x8_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +#define vpx_sub_pixel_variance4x8 vpx_sub_pixel_variance4x8_neon + +uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x32_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +#define vpx_sub_pixel_variance64x32 vpx_sub_pixel_variance64x32_neon + +uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x64_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +#define vpx_sub_pixel_variance64x64 vpx_sub_pixel_variance64x64_neon + +uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x16_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +#define vpx_sub_pixel_variance8x16 vpx_sub_pixel_variance8x16_neon + +uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x4_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +#define vpx_sub_pixel_variance8x4 vpx_sub_pixel_variance8x4_neon + +uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x8_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +#define vpx_sub_pixel_variance8x8 vpx_sub_pixel_variance8x8_neon + +void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride); +void vpx_subtract_block_neon(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride); +#define vpx_subtract_block vpx_subtract_block_neon + +uint64_t vpx_sum_squares_2d_i16_c(const int16_t *src, int stride, int size); +uint64_t vpx_sum_squares_2d_i16_neon(const int16_t *src, int stride, int size); +#define vpx_sum_squares_2d_i16 vpx_sum_squares_2d_i16_neon + +void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_tm_predictor_16x16 vpx_tm_predictor_16x16_neon + +void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_tm_predictor_32x32 vpx_tm_predictor_32x32_neon + +void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_tm_predictor_4x4 vpx_tm_predictor_4x4_neon + +void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_tm_predictor_8x8 vpx_tm_predictor_8x8_neon + +void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_v_predictor_16x16 vpx_v_predictor_16x16_neon + +void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_v_predictor_32x32 vpx_v_predictor_32x32_neon + +void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_v_predictor_4x4 vpx_v_predictor_4x4_neon + +void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_v_predictor_8x8 vpx_v_predictor_8x8_neon + +unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x16_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x32_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x8_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x16_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x32_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x64_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance4x4_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance4x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance4x8_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x32_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x64_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x16_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x4_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x8_neon_dotprod(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c + +int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl); +int vpx_vector_var_neon(const int16_t *ref, const int16_t *src, const int bwl); +#define vpx_vector_var vpx_vector_var_neon + +void vpx_dsp_rtcd(void); + +#include "vpx_config.h" + +#ifdef RTCD_C +#include "vpx_ports/arm.h" +static void setup_rtcd_internal(void) +{ + int flags = arm_cpu_caps(); + + (void)flags; + + vpx_convolve8 = vpx_convolve8_neon; + if (flags & HAS_NEON_DOTPROD) vpx_convolve8 = vpx_convolve8_neon_dotprod; + if (flags & HAS_NEON_I8MM) vpx_convolve8 = vpx_convolve8_neon_i8mm; + vpx_convolve8_avg = vpx_convolve8_avg_neon; + if (flags & HAS_NEON_DOTPROD) vpx_convolve8_avg = vpx_convolve8_avg_neon_dotprod; + if (flags & HAS_NEON_I8MM) vpx_convolve8_avg = vpx_convolve8_avg_neon_i8mm; + vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_neon; + if (flags & HAS_NEON_DOTPROD) vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_neon_dotprod; + if (flags & HAS_NEON_I8MM) vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_neon_i8mm; + vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_neon; + if (flags & HAS_NEON_DOTPROD) vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_neon_dotprod; + if (flags & HAS_NEON_I8MM) vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_neon_i8mm; + vpx_convolve8_horiz = vpx_convolve8_horiz_neon; + if (flags & HAS_NEON_DOTPROD) vpx_convolve8_horiz = vpx_convolve8_horiz_neon_dotprod; + if (flags & HAS_NEON_I8MM) vpx_convolve8_horiz = vpx_convolve8_horiz_neon_i8mm; + vpx_convolve8_vert = vpx_convolve8_vert_neon; + if (flags & HAS_NEON_DOTPROD) vpx_convolve8_vert = vpx_convolve8_vert_neon_dotprod; + if (flags & HAS_NEON_I8MM) vpx_convolve8_vert = vpx_convolve8_vert_neon_i8mm; + vpx_get16x16var = vpx_get16x16var_neon; + if (flags & HAS_NEON_DOTPROD) vpx_get16x16var = vpx_get16x16var_neon_dotprod; + vpx_get4x4sse_cs = vpx_get4x4sse_cs_neon; + if (flags & HAS_NEON_DOTPROD) vpx_get4x4sse_cs = vpx_get4x4sse_cs_neon_dotprod; + vpx_get8x8var = vpx_get8x8var_neon; + if (flags & HAS_NEON_DOTPROD) vpx_get8x8var = vpx_get8x8var_neon_dotprod; + vpx_mse16x16 = vpx_mse16x16_neon; + if (flags & HAS_NEON_DOTPROD) vpx_mse16x16 = vpx_mse16x16_neon_dotprod; + vpx_mse16x8 = vpx_mse16x8_neon; + if (flags & HAS_NEON_DOTPROD) vpx_mse16x8 = vpx_mse16x8_neon_dotprod; + vpx_mse8x16 = vpx_mse8x16_neon; + if (flags & HAS_NEON_DOTPROD) vpx_mse8x16 = vpx_mse8x16_neon_dotprod; + vpx_mse8x8 = vpx_mse8x8_neon; + if (flags & HAS_NEON_DOTPROD) vpx_mse8x8 = vpx_mse8x8_neon_dotprod; + vpx_sad16x16 = vpx_sad16x16_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad16x16 = vpx_sad16x16_neon_dotprod; + vpx_sad16x16_avg = vpx_sad16x16_avg_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad16x16_avg = vpx_sad16x16_avg_neon_dotprod; + vpx_sad16x16x4d = vpx_sad16x16x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad16x16x4d = vpx_sad16x16x4d_neon_dotprod; + vpx_sad16x32 = vpx_sad16x32_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad16x32 = vpx_sad16x32_neon_dotprod; + vpx_sad16x32_avg = vpx_sad16x32_avg_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad16x32_avg = vpx_sad16x32_avg_neon_dotprod; + vpx_sad16x32x4d = vpx_sad16x32x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad16x32x4d = vpx_sad16x32x4d_neon_dotprod; + vpx_sad16x8 = vpx_sad16x8_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad16x8 = vpx_sad16x8_neon_dotprod; + vpx_sad16x8_avg = vpx_sad16x8_avg_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad16x8_avg = vpx_sad16x8_avg_neon_dotprod; + vpx_sad16x8x4d = vpx_sad16x8x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad16x8x4d = vpx_sad16x8x4d_neon_dotprod; + vpx_sad32x16 = vpx_sad32x16_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad32x16 = vpx_sad32x16_neon_dotprod; + vpx_sad32x16_avg = vpx_sad32x16_avg_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad32x16_avg = vpx_sad32x16_avg_neon_dotprod; + vpx_sad32x16x4d = vpx_sad32x16x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad32x16x4d = vpx_sad32x16x4d_neon_dotprod; + vpx_sad32x32 = vpx_sad32x32_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad32x32 = vpx_sad32x32_neon_dotprod; + vpx_sad32x32_avg = vpx_sad32x32_avg_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad32x32_avg = vpx_sad32x32_avg_neon_dotprod; + vpx_sad32x32x4d = vpx_sad32x32x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad32x32x4d = vpx_sad32x32x4d_neon_dotprod; + vpx_sad32x64 = vpx_sad32x64_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad32x64 = vpx_sad32x64_neon_dotprod; + vpx_sad32x64_avg = vpx_sad32x64_avg_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad32x64_avg = vpx_sad32x64_avg_neon_dotprod; + vpx_sad32x64x4d = vpx_sad32x64x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad32x64x4d = vpx_sad32x64x4d_neon_dotprod; + vpx_sad64x32 = vpx_sad64x32_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad64x32 = vpx_sad64x32_neon_dotprod; + vpx_sad64x32_avg = vpx_sad64x32_avg_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad64x32_avg = vpx_sad64x32_avg_neon_dotprod; + vpx_sad64x32x4d = vpx_sad64x32x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad64x32x4d = vpx_sad64x32x4d_neon_dotprod; + vpx_sad64x64 = vpx_sad64x64_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad64x64 = vpx_sad64x64_neon_dotprod; + vpx_sad64x64_avg = vpx_sad64x64_avg_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad64x64_avg = vpx_sad64x64_avg_neon_dotprod; + vpx_sad64x64x4d = vpx_sad64x64x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad64x64x4d = vpx_sad64x64x4d_neon_dotprod; + vpx_sad_skip_16x16 = vpx_sad_skip_16x16_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_16x16 = vpx_sad_skip_16x16_neon_dotprod; + vpx_sad_skip_16x16x4d = vpx_sad_skip_16x16x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_16x16x4d = vpx_sad_skip_16x16x4d_neon_dotprod; + vpx_sad_skip_16x32 = vpx_sad_skip_16x32_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_16x32 = vpx_sad_skip_16x32_neon_dotprod; + vpx_sad_skip_16x32x4d = vpx_sad_skip_16x32x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_16x32x4d = vpx_sad_skip_16x32x4d_neon_dotprod; + vpx_sad_skip_16x8 = vpx_sad_skip_16x8_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_16x8 = vpx_sad_skip_16x8_neon_dotprod; + vpx_sad_skip_16x8x4d = vpx_sad_skip_16x8x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_16x8x4d = vpx_sad_skip_16x8x4d_neon_dotprod; + vpx_sad_skip_32x16 = vpx_sad_skip_32x16_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_32x16 = vpx_sad_skip_32x16_neon_dotprod; + vpx_sad_skip_32x16x4d = vpx_sad_skip_32x16x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_32x16x4d = vpx_sad_skip_32x16x4d_neon_dotprod; + vpx_sad_skip_32x32 = vpx_sad_skip_32x32_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_32x32 = vpx_sad_skip_32x32_neon_dotprod; + vpx_sad_skip_32x32x4d = vpx_sad_skip_32x32x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_32x32x4d = vpx_sad_skip_32x32x4d_neon_dotprod; + vpx_sad_skip_32x64 = vpx_sad_skip_32x64_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_32x64 = vpx_sad_skip_32x64_neon_dotprod; + vpx_sad_skip_32x64x4d = vpx_sad_skip_32x64x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_32x64x4d = vpx_sad_skip_32x64x4d_neon_dotprod; + vpx_sad_skip_64x32 = vpx_sad_skip_64x32_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_64x32 = vpx_sad_skip_64x32_neon_dotprod; + vpx_sad_skip_64x32x4d = vpx_sad_skip_64x32x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_64x32x4d = vpx_sad_skip_64x32x4d_neon_dotprod; + vpx_sad_skip_64x64 = vpx_sad_skip_64x64_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_64x64 = vpx_sad_skip_64x64_neon_dotprod; + vpx_sad_skip_64x64x4d = vpx_sad_skip_64x64x4d_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sad_skip_64x64x4d = vpx_sad_skip_64x64x4d_neon_dotprod; + vpx_sse = vpx_sse_neon; + if (flags & HAS_NEON_DOTPROD) vpx_sse = vpx_sse_neon_dotprod; + vpx_variance16x16 = vpx_variance16x16_neon; + if (flags & HAS_NEON_DOTPROD) vpx_variance16x16 = vpx_variance16x16_neon_dotprod; + vpx_variance16x32 = vpx_variance16x32_neon; + if (flags & HAS_NEON_DOTPROD) vpx_variance16x32 = vpx_variance16x32_neon_dotprod; + vpx_variance16x8 = vpx_variance16x8_neon; + if (flags & HAS_NEON_DOTPROD) vpx_variance16x8 = vpx_variance16x8_neon_dotprod; + vpx_variance32x16 = vpx_variance32x16_neon; + if (flags & HAS_NEON_DOTPROD) vpx_variance32x16 = vpx_variance32x16_neon_dotprod; + vpx_variance32x32 = vpx_variance32x32_neon; + if (flags & HAS_NEON_DOTPROD) vpx_variance32x32 = vpx_variance32x32_neon_dotprod; + vpx_variance32x64 = vpx_variance32x64_neon; + if (flags & HAS_NEON_DOTPROD) vpx_variance32x64 = vpx_variance32x64_neon_dotprod; + vpx_variance4x4 = vpx_variance4x4_neon; + if (flags & HAS_NEON_DOTPROD) vpx_variance4x4 = vpx_variance4x4_neon_dotprod; + vpx_variance4x8 = vpx_variance4x8_neon; + if (flags & HAS_NEON_DOTPROD) vpx_variance4x8 = vpx_variance4x8_neon_dotprod; + vpx_variance64x32 = vpx_variance64x32_neon; + if (flags & HAS_NEON_DOTPROD) vpx_variance64x32 = vpx_variance64x32_neon_dotprod; + vpx_variance64x64 = vpx_variance64x64_neon; + if (flags & HAS_NEON_DOTPROD) vpx_variance64x64 = vpx_variance64x64_neon_dotprod; + vpx_variance8x16 = vpx_variance8x16_neon; + if (flags & HAS_NEON_DOTPROD) vpx_variance8x16 = vpx_variance8x16_neon_dotprod; + vpx_variance8x4 = vpx_variance8x4_neon; + if (flags & HAS_NEON_DOTPROD) vpx_variance8x4 = vpx_variance8x4_neon_dotprod; + vpx_variance8x8 = vpx_variance8x8_neon; + if (flags & HAS_NEON_DOTPROD) vpx_variance8x8 = vpx_variance8x8_neon_dotprod; +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/arm64/vpx_scale_rtcd.h b/media/libvpx/config/linux/arm64/vpx_scale_rtcd.h new file mode 100644 index 0000000000..b371368275 --- /dev/null +++ b/media/libvpx/config/linux/arm64/vpx_scale_rtcd.h @@ -0,0 +1,75 @@ +// This file is generated. Do not edit. +#ifndef VPX_SCALE_RTCD_H_ +#define VPX_SCALE_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +struct yv12_buffer_config; + +#ifdef __cplusplus +extern "C" { +#endif + +void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width); +#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c + +void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width); +#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c + +void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width); +#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c + +void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c + +void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c + +void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c + +void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c + +void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc); +#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c + +void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf); +#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c + +void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf); +#define vpx_extend_frame_borders vpx_extend_frame_borders_c + +void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf); +#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c + +void vpx_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc); +#define vpx_yv12_copy_frame vpx_yv12_copy_frame_c + +void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc); +#define vpx_yv12_copy_y vpx_yv12_copy_y_c + +void vpx_scale_rtcd(void); + +#include "vpx_config.h" + +#ifdef RTCD_C +#include "vpx_ports/arm.h" +static void setup_rtcd_internal(void) +{ + int flags = arm_cpu_caps(); + + (void)flags; + +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/ia32/vp8_rtcd.h b/media/libvpx/config/linux/ia32/vp8_rtcd.h new file mode 100644 index 0000000000..7a23227e4d --- /dev/null +++ b/media/libvpx/config/linux/ia32/vp8_rtcd.h @@ -0,0 +1,323 @@ +// This file is generated. Do not edit. +#ifndef VP8_RTCD_H_ +#define VP8_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +/* + * VP8 + */ + +struct blockd; +struct macroblockd; +struct loop_filter_info; + +/* Encoder forward decls */ +struct block; +struct macroblock; +struct variance_vtable; +union int_mv; +struct yv12_buffer_config; + +#ifdef __cplusplus +extern "C" { +#endif + +void vp8_bilinear_predict16x16_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict16x16_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict16x16_ssse3(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_bilinear_predict16x16)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_bilinear_predict4x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict4x4_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_bilinear_predict4x4)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_bilinear_predict8x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict8x4_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_bilinear_predict8x4)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_bilinear_predict8x8_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict8x8_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict8x8_ssse3(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_bilinear_predict8x8)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +int vp8_block_error_c(short *coeff, short *dqcoeff); +int vp8_block_error_sse2(short *coeff, short *dqcoeff); +RTCD_EXTERN int (*vp8_block_error)(short *coeff, short *dqcoeff); + +void vp8_copy32xn_c(const unsigned char *src_ptr, int src_stride, unsigned char *dst_ptr, int dst_stride, int height); +void vp8_copy32xn_sse2(const unsigned char *src_ptr, int src_stride, unsigned char *dst_ptr, int dst_stride, int height); +void vp8_copy32xn_sse3(const unsigned char *src_ptr, int src_stride, unsigned char *dst_ptr, int dst_stride, int height); +RTCD_EXTERN void (*vp8_copy32xn)(const unsigned char *src_ptr, int src_stride, unsigned char *dst_ptr, int dst_stride, int height); + +void vp8_copy_mem16x16_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +void vp8_copy_mem16x16_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +RTCD_EXTERN void (*vp8_copy_mem16x16)(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); + +void vp8_copy_mem8x4_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +void vp8_copy_mem8x4_mmx(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +RTCD_EXTERN void (*vp8_copy_mem8x4)(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); + +void vp8_copy_mem8x8_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +void vp8_copy_mem8x8_mmx(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +RTCD_EXTERN void (*vp8_copy_mem8x8)(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); + +void vp8_dc_only_idct_add_c(short input_dc, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +void vp8_dc_only_idct_add_mmx(short input_dc, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +RTCD_EXTERN void (*vp8_dc_only_idct_add)(short input_dc, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); + +int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +RTCD_EXTERN int (*vp8_denoiser_filter)(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); + +int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +int vp8_denoiser_filter_uv_sse2(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +RTCD_EXTERN int (*vp8_denoiser_filter_uv)(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); + +void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *dest, int stride); +void vp8_dequant_idct_add_mmx(short *input, short *dq, unsigned char *dest, int stride); +RTCD_EXTERN void (*vp8_dequant_idct_add)(short *input, short *dq, unsigned char *dest, int stride); + +void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs); +void vp8_dequant_idct_add_uv_block_sse2(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs); +RTCD_EXTERN void (*vp8_dequant_idct_add_uv_block)(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs); + +void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs); +void vp8_dequant_idct_add_y_block_sse2(short *q, short *dq, unsigned char *dst, int stride, char *eobs); +RTCD_EXTERN void (*vp8_dequant_idct_add_y_block)(short *q, short *dq, unsigned char *dst, int stride, char *eobs); + +void vp8_dequantize_b_c(struct blockd*, short *DQC); +void vp8_dequantize_b_mmx(struct blockd*, short *DQC); +RTCD_EXTERN void (*vp8_dequantize_b)(struct blockd*, short *DQC); + +int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); +int vp8_diamond_search_sadx4(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); +RTCD_EXTERN int (*vp8_diamond_search_sad)(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); + +void vp8_fast_quantize_b_c(struct block *, struct blockd *); +void vp8_fast_quantize_b_sse2(struct block *, struct blockd *); +void vp8_fast_quantize_b_ssse3(struct block *, struct blockd *); +RTCD_EXTERN void (*vp8_fast_quantize_b)(struct block *, struct blockd *); + +void vp8_filter_by_weight16x16_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight); +void vp8_filter_by_weight16x16_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight); +RTCD_EXTERN void (*vp8_filter_by_weight16x16)(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight); + +void vp8_filter_by_weight4x4_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight); +#define vp8_filter_by_weight4x4 vp8_filter_by_weight4x4_c + +void vp8_filter_by_weight8x8_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight); +void vp8_filter_by_weight8x8_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight); +RTCD_EXTERN void (*vp8_filter_by_weight8x8)(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight); + +void vp8_loop_filter_bh_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_bh_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +RTCD_EXTERN void (*vp8_loop_filter_bh)(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); + +void vp8_loop_filter_bv_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_bv_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +RTCD_EXTERN void (*vp8_loop_filter_bv)(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); + +void vp8_loop_filter_mbh_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_mbh_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +RTCD_EXTERN void (*vp8_loop_filter_mbh)(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); + +void vp8_loop_filter_mbv_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_mbv_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +RTCD_EXTERN void (*vp8_loop_filter_mbv)(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); + +void vp8_loop_filter_bhs_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_bhs_sse2(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +RTCD_EXTERN void (*vp8_loop_filter_simple_bh)(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); + +void vp8_loop_filter_bvs_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_bvs_sse2(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +RTCD_EXTERN void (*vp8_loop_filter_simple_bv)(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); + +void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_simple_horizontal_edge_sse2(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +RTCD_EXTERN void (*vp8_loop_filter_simple_mbh)(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); + +void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_simple_vertical_edge_sse2(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +RTCD_EXTERN void (*vp8_loop_filter_simple_mbv)(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); + +int vp8_mbblock_error_c(struct macroblock *mb, int dc); +int vp8_mbblock_error_sse2(struct macroblock *mb, int dc); +RTCD_EXTERN int (*vp8_mbblock_error)(struct macroblock *mb, int dc); + +int vp8_mbuverror_c(struct macroblock *mb); +int vp8_mbuverror_sse2(struct macroblock *mb); +RTCD_EXTERN int (*vp8_mbuverror)(struct macroblock *mb); + +int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int error_per_bit, int search_range, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); +int vp8_refining_search_sadx4(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int error_per_bit, int search_range, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); +RTCD_EXTERN int (*vp8_refining_search_sad)(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int error_per_bit, int search_range, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); + +void vp8_regular_quantize_b_c(struct block *, struct blockd *); +void vp8_regular_quantize_b_sse2(struct block *, struct blockd *); +void vp8_regular_quantize_b_sse4_1(struct block *, struct blockd *); +RTCD_EXTERN void (*vp8_regular_quantize_b)(struct block *, struct blockd *); + +void vp8_short_fdct4x4_c(short *input, short *output, int pitch); +void vp8_short_fdct4x4_sse2(short *input, short *output, int pitch); +RTCD_EXTERN void (*vp8_short_fdct4x4)(short *input, short *output, int pitch); + +void vp8_short_fdct8x4_c(short *input, short *output, int pitch); +void vp8_short_fdct8x4_sse2(short *input, short *output, int pitch); +RTCD_EXTERN void (*vp8_short_fdct8x4)(short *input, short *output, int pitch); + +void vp8_short_idct4x4llm_c(short *input, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +void vp8_short_idct4x4llm_mmx(short *input, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +RTCD_EXTERN void (*vp8_short_idct4x4llm)(short *input, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); + +void vp8_short_inv_walsh4x4_c(short *input, short *mb_dqcoeff); +void vp8_short_inv_walsh4x4_sse2(short *input, short *mb_dqcoeff); +RTCD_EXTERN void (*vp8_short_inv_walsh4x4)(short *input, short *mb_dqcoeff); + +void vp8_short_inv_walsh4x4_1_c(short *input, short *mb_dqcoeff); +#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c + +void vp8_short_walsh4x4_c(short *input, short *output, int pitch); +void vp8_short_walsh4x4_sse2(short *input, short *output, int pitch); +RTCD_EXTERN void (*vp8_short_walsh4x4)(short *input, short *output, int pitch); + +void vp8_sixtap_predict16x16_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict16x16_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict16x16_ssse3(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_sixtap_predict16x16)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_sixtap_predict4x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict4x4_mmx(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict4x4_ssse3(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_sixtap_predict4x4)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_sixtap_predict8x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict8x4_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict8x4_ssse3(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_sixtap_predict8x4)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_sixtap_predict8x8_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict8x8_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict8x8_ssse3(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_sixtap_predict8x8)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_temporal_filter_apply_c(unsigned char *frame1, unsigned int stride, unsigned char *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, unsigned short *count); +void vp8_temporal_filter_apply_sse2(unsigned char *frame1, unsigned int stride, unsigned char *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, unsigned short *count); +RTCD_EXTERN void (*vp8_temporal_filter_apply)(unsigned char *frame1, unsigned int stride, unsigned char *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, unsigned short *count); + +void vp8_rtcd(void); + +#ifdef RTCD_C +#include "vpx_ports/x86.h" +static void setup_rtcd_internal(void) +{ + int flags = x86_simd_caps(); + + (void)flags; + + vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_c; + if (flags & HAS_SSE2) vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_sse2; + if (flags & HAS_SSSE3) vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_ssse3; + vp8_bilinear_predict4x4 = vp8_bilinear_predict4x4_c; + if (flags & HAS_SSE2) vp8_bilinear_predict4x4 = vp8_bilinear_predict4x4_sse2; + vp8_bilinear_predict8x4 = vp8_bilinear_predict8x4_c; + if (flags & HAS_SSE2) vp8_bilinear_predict8x4 = vp8_bilinear_predict8x4_sse2; + vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_c; + if (flags & HAS_SSE2) vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_sse2; + if (flags & HAS_SSSE3) vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_ssse3; + vp8_block_error = vp8_block_error_c; + if (flags & HAS_SSE2) vp8_block_error = vp8_block_error_sse2; + vp8_copy32xn = vp8_copy32xn_c; + if (flags & HAS_SSE2) vp8_copy32xn = vp8_copy32xn_sse2; + if (flags & HAS_SSE3) vp8_copy32xn = vp8_copy32xn_sse3; + vp8_copy_mem16x16 = vp8_copy_mem16x16_c; + if (flags & HAS_SSE2) vp8_copy_mem16x16 = vp8_copy_mem16x16_sse2; + vp8_copy_mem8x4 = vp8_copy_mem8x4_c; + if (flags & HAS_MMX) vp8_copy_mem8x4 = vp8_copy_mem8x4_mmx; + vp8_copy_mem8x8 = vp8_copy_mem8x8_c; + if (flags & HAS_MMX) vp8_copy_mem8x8 = vp8_copy_mem8x8_mmx; + vp8_dc_only_idct_add = vp8_dc_only_idct_add_c; + if (flags & HAS_MMX) vp8_dc_only_idct_add = vp8_dc_only_idct_add_mmx; + vp8_denoiser_filter = vp8_denoiser_filter_c; + if (flags & HAS_SSE2) vp8_denoiser_filter = vp8_denoiser_filter_sse2; + vp8_denoiser_filter_uv = vp8_denoiser_filter_uv_c; + if (flags & HAS_SSE2) vp8_denoiser_filter_uv = vp8_denoiser_filter_uv_sse2; + vp8_dequant_idct_add = vp8_dequant_idct_add_c; + if (flags & HAS_MMX) vp8_dequant_idct_add = vp8_dequant_idct_add_mmx; + vp8_dequant_idct_add_uv_block = vp8_dequant_idct_add_uv_block_c; + if (flags & HAS_SSE2) vp8_dequant_idct_add_uv_block = vp8_dequant_idct_add_uv_block_sse2; + vp8_dequant_idct_add_y_block = vp8_dequant_idct_add_y_block_c; + if (flags & HAS_SSE2) vp8_dequant_idct_add_y_block = vp8_dequant_idct_add_y_block_sse2; + vp8_dequantize_b = vp8_dequantize_b_c; + if (flags & HAS_MMX) vp8_dequantize_b = vp8_dequantize_b_mmx; + vp8_diamond_search_sad = vp8_diamond_search_sad_c; + if (flags & HAS_SSE2) vp8_diamond_search_sad = vp8_diamond_search_sadx4; + vp8_fast_quantize_b = vp8_fast_quantize_b_c; + if (flags & HAS_SSE2) vp8_fast_quantize_b = vp8_fast_quantize_b_sse2; + if (flags & HAS_SSSE3) vp8_fast_quantize_b = vp8_fast_quantize_b_ssse3; + vp8_filter_by_weight16x16 = vp8_filter_by_weight16x16_c; + if (flags & HAS_SSE2) vp8_filter_by_weight16x16 = vp8_filter_by_weight16x16_sse2; + vp8_filter_by_weight8x8 = vp8_filter_by_weight8x8_c; + if (flags & HAS_SSE2) vp8_filter_by_weight8x8 = vp8_filter_by_weight8x8_sse2; + vp8_loop_filter_bh = vp8_loop_filter_bh_c; + if (flags & HAS_SSE2) vp8_loop_filter_bh = vp8_loop_filter_bh_sse2; + vp8_loop_filter_bv = vp8_loop_filter_bv_c; + if (flags & HAS_SSE2) vp8_loop_filter_bv = vp8_loop_filter_bv_sse2; + vp8_loop_filter_mbh = vp8_loop_filter_mbh_c; + if (flags & HAS_SSE2) vp8_loop_filter_mbh = vp8_loop_filter_mbh_sse2; + vp8_loop_filter_mbv = vp8_loop_filter_mbv_c; + if (flags & HAS_SSE2) vp8_loop_filter_mbv = vp8_loop_filter_mbv_sse2; + vp8_loop_filter_simple_bh = vp8_loop_filter_bhs_c; + if (flags & HAS_SSE2) vp8_loop_filter_simple_bh = vp8_loop_filter_bhs_sse2; + vp8_loop_filter_simple_bv = vp8_loop_filter_bvs_c; + if (flags & HAS_SSE2) vp8_loop_filter_simple_bv = vp8_loop_filter_bvs_sse2; + vp8_loop_filter_simple_mbh = vp8_loop_filter_simple_horizontal_edge_c; + if (flags & HAS_SSE2) vp8_loop_filter_simple_mbh = vp8_loop_filter_simple_horizontal_edge_sse2; + vp8_loop_filter_simple_mbv = vp8_loop_filter_simple_vertical_edge_c; + if (flags & HAS_SSE2) vp8_loop_filter_simple_mbv = vp8_loop_filter_simple_vertical_edge_sse2; + vp8_mbblock_error = vp8_mbblock_error_c; + if (flags & HAS_SSE2) vp8_mbblock_error = vp8_mbblock_error_sse2; + vp8_mbuverror = vp8_mbuverror_c; + if (flags & HAS_SSE2) vp8_mbuverror = vp8_mbuverror_sse2; + vp8_refining_search_sad = vp8_refining_search_sad_c; + if (flags & HAS_SSE2) vp8_refining_search_sad = vp8_refining_search_sadx4; + vp8_regular_quantize_b = vp8_regular_quantize_b_c; + if (flags & HAS_SSE2) vp8_regular_quantize_b = vp8_regular_quantize_b_sse2; + if (flags & HAS_SSE4_1) vp8_regular_quantize_b = vp8_regular_quantize_b_sse4_1; + vp8_short_fdct4x4 = vp8_short_fdct4x4_c; + if (flags & HAS_SSE2) vp8_short_fdct4x4 = vp8_short_fdct4x4_sse2; + vp8_short_fdct8x4 = vp8_short_fdct8x4_c; + if (flags & HAS_SSE2) vp8_short_fdct8x4 = vp8_short_fdct8x4_sse2; + vp8_short_idct4x4llm = vp8_short_idct4x4llm_c; + if (flags & HAS_MMX) vp8_short_idct4x4llm = vp8_short_idct4x4llm_mmx; + vp8_short_inv_walsh4x4 = vp8_short_inv_walsh4x4_c; + if (flags & HAS_SSE2) vp8_short_inv_walsh4x4 = vp8_short_inv_walsh4x4_sse2; + vp8_short_walsh4x4 = vp8_short_walsh4x4_c; + if (flags & HAS_SSE2) vp8_short_walsh4x4 = vp8_short_walsh4x4_sse2; + vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_c; + if (flags & HAS_SSE2) vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_sse2; + if (flags & HAS_SSSE3) vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_ssse3; + vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_c; + if (flags & HAS_MMX) vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_mmx; + if (flags & HAS_SSSE3) vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_ssse3; + vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_c; + if (flags & HAS_SSE2) vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_sse2; + if (flags & HAS_SSSE3) vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_ssse3; + vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_c; + if (flags & HAS_SSE2) vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_sse2; + if (flags & HAS_SSSE3) vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_ssse3; + vp8_temporal_filter_apply = vp8_temporal_filter_apply_c; + if (flags & HAS_SSE2) vp8_temporal_filter_apply = vp8_temporal_filter_apply_sse2; +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/ia32/vp9_rtcd.h b/media/libvpx/config/linux/ia32/vp9_rtcd.h new file mode 100644 index 0000000000..3136db471c --- /dev/null +++ b/media/libvpx/config/linux/ia32/vp9_rtcd.h @@ -0,0 +1,156 @@ +// This file is generated. Do not edit. +#ifndef VP9_RTCD_H_ +#define VP9_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +/* + * VP9 + */ + +#include "vpx/vpx_integer.h" +#include "vp9/common/vp9_common.h" +#include "vp9/common/vp9_enums.h" +#include "vp9/common/vp9_filter.h" + +struct macroblockd; + +/* Encoder forward decls */ +struct macroblock; +struct macroblock_plane; +struct vp9_sad_table; +struct ScanOrder; +struct search_site_config; +struct mv; +union int_mv; +struct yv12_buffer_config; + +#ifdef __cplusplus +extern "C" { +#endif + +void vp9_apply_temporal_filter_c(const uint8_t *y_src, int y_src_stride, const uint8_t *y_pre, int y_pre_stride, const uint8_t *u_src, const uint8_t *v_src, int uv_src_stride, const uint8_t *u_pre, const uint8_t *v_pre, int uv_pre_stride, unsigned int block_width, unsigned int block_height, int ss_x, int ss_y, int strength, const int *const blk_fw, int use_32x32, uint32_t *y_accumulator, uint16_t *y_count, uint32_t *u_accumulator, uint16_t *u_count, uint32_t *v_accumulator, uint16_t *v_count); +void vp9_apply_temporal_filter_sse4_1(const uint8_t *y_src, int y_src_stride, const uint8_t *y_pre, int y_pre_stride, const uint8_t *u_src, const uint8_t *v_src, int uv_src_stride, const uint8_t *u_pre, const uint8_t *v_pre, int uv_pre_stride, unsigned int block_width, unsigned int block_height, int ss_x, int ss_y, int strength, const int *const blk_fw, int use_32x32, uint32_t *y_accumulator, uint16_t *y_count, uint32_t *u_accumulator, uint16_t *u_count, uint32_t *v_accumulator, uint16_t *v_count); +RTCD_EXTERN void (*vp9_apply_temporal_filter)(const uint8_t *y_src, int y_src_stride, const uint8_t *y_pre, int y_pre_stride, const uint8_t *u_src, const uint8_t *v_src, int uv_src_stride, const uint8_t *u_pre, const uint8_t *v_pre, int uv_pre_stride, unsigned int block_width, unsigned int block_height, int ss_x, int ss_y, int strength, const int *const blk_fw, int use_32x32, uint32_t *y_accumulator, uint16_t *y_count, uint32_t *u_accumulator, uint16_t *u_count, uint32_t *v_accumulator, uint16_t *v_count); + +int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz); +int64_t vp9_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz); +int64_t vp9_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz); +RTCD_EXTERN int64_t (*vp9_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz); + +int64_t vp9_block_error_fp_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size); +int64_t vp9_block_error_fp_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size); +int64_t vp9_block_error_fp_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size); +RTCD_EXTERN int64_t (*vp9_block_error_fp)(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size); + +int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, uint32_t start_mv_sad, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_sad_table *sad_fn_ptr, const struct mv *center_mv); +#define vp9_diamond_search_sad vp9_diamond_search_sad_c + +void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type); +void vp9_fht16x16_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type); +RTCD_EXTERN void (*vp9_fht16x16)(const int16_t *input, tran_low_t *output, int stride, int tx_type); + +void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type); +void vp9_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type); +RTCD_EXTERN void (*vp9_fht4x4)(const int16_t *input, tran_low_t *output, int stride, int tx_type); + +void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type); +void vp9_fht8x8_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type); +RTCD_EXTERN void (*vp9_fht8x8)(const int16_t *input, tran_low_t *output, int stride, int tx_type); + +void vp9_filter_by_weight16x16_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight); +void vp9_filter_by_weight16x16_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight); +RTCD_EXTERN void (*vp9_filter_by_weight16x16)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight); + +void vp9_filter_by_weight8x8_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight); +void vp9_filter_by_weight8x8_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight); +RTCD_EXTERN void (*vp9_filter_by_weight8x8)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight); + +void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride); +void vp9_fwht4x4_sse2(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vp9_fwht4x4)(const int16_t *input, tran_low_t *output, int stride); + +void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +void vp9_iht16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +RTCD_EXTERN void (*vp9_iht16x16_256_add)(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); + +void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +void vp9_iht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +RTCD_EXTERN void (*vp9_iht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); + +void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +void vp9_iht8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +RTCD_EXTERN void (*vp9_iht8x8_64_add)(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); + +void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +RTCD_EXTERN void (*vp9_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); + +void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_32x32_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +RTCD_EXTERN void (*vp9_quantize_fp_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); + +void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst, INTERP_FILTER filter_type, int phase_scaler); +void vp9_scale_and_extend_frame_ssse3(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst, INTERP_FILTER filter_type, int phase_scaler); +RTCD_EXTERN void (*vp9_scale_and_extend_frame)(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst, INTERP_FILTER filter_type, int phase_scaler); + +void vp9_rtcd(void); + +#ifdef RTCD_C +#include "vpx_ports/x86.h" +static void setup_rtcd_internal(void) +{ + int flags = x86_simd_caps(); + + (void)flags; + + vp9_apply_temporal_filter = vp9_apply_temporal_filter_c; + if (flags & HAS_SSE4_1) vp9_apply_temporal_filter = vp9_apply_temporal_filter_sse4_1; + vp9_block_error = vp9_block_error_c; + if (flags & HAS_SSE2) vp9_block_error = vp9_block_error_sse2; + if (flags & HAS_AVX2) vp9_block_error = vp9_block_error_avx2; + vp9_block_error_fp = vp9_block_error_fp_c; + if (flags & HAS_SSE2) vp9_block_error_fp = vp9_block_error_fp_sse2; + if (flags & HAS_AVX2) vp9_block_error_fp = vp9_block_error_fp_avx2; + vp9_fht16x16 = vp9_fht16x16_c; + if (flags & HAS_SSE2) vp9_fht16x16 = vp9_fht16x16_sse2; + vp9_fht4x4 = vp9_fht4x4_c; + if (flags & HAS_SSE2) vp9_fht4x4 = vp9_fht4x4_sse2; + vp9_fht8x8 = vp9_fht8x8_c; + if (flags & HAS_SSE2) vp9_fht8x8 = vp9_fht8x8_sse2; + vp9_filter_by_weight16x16 = vp9_filter_by_weight16x16_c; + if (flags & HAS_SSE2) vp9_filter_by_weight16x16 = vp9_filter_by_weight16x16_sse2; + vp9_filter_by_weight8x8 = vp9_filter_by_weight8x8_c; + if (flags & HAS_SSE2) vp9_filter_by_weight8x8 = vp9_filter_by_weight8x8_sse2; + vp9_fwht4x4 = vp9_fwht4x4_c; + if (flags & HAS_SSE2) vp9_fwht4x4 = vp9_fwht4x4_sse2; + vp9_iht16x16_256_add = vp9_iht16x16_256_add_c; + if (flags & HAS_SSE2) vp9_iht16x16_256_add = vp9_iht16x16_256_add_sse2; + vp9_iht4x4_16_add = vp9_iht4x4_16_add_c; + if (flags & HAS_SSE2) vp9_iht4x4_16_add = vp9_iht4x4_16_add_sse2; + vp9_iht8x8_64_add = vp9_iht8x8_64_add_c; + if (flags & HAS_SSE2) vp9_iht8x8_64_add = vp9_iht8x8_64_add_sse2; + vp9_quantize_fp = vp9_quantize_fp_c; + if (flags & HAS_SSE2) vp9_quantize_fp = vp9_quantize_fp_sse2; + if (flags & HAS_SSSE3) vp9_quantize_fp = vp9_quantize_fp_ssse3; + if (flags & HAS_AVX2) vp9_quantize_fp = vp9_quantize_fp_avx2; + vp9_quantize_fp_32x32 = vp9_quantize_fp_32x32_c; + if (flags & HAS_SSSE3) vp9_quantize_fp_32x32 = vp9_quantize_fp_32x32_ssse3; + if (flags & HAS_AVX2) vp9_quantize_fp_32x32 = vp9_quantize_fp_32x32_avx2; + vp9_scale_and_extend_frame = vp9_scale_and_extend_frame_c; + if (flags & HAS_SSSE3) vp9_scale_and_extend_frame = vp9_scale_and_extend_frame_ssse3; +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/ia32/vpx_config.asm b/media/libvpx/config/linux/ia32/vpx_config.asm new file mode 100644 index 0000000000..eaa3950d37 --- /dev/null +++ b/media/libvpx/config/linux/ia32/vpx_config.asm @@ -0,0 +1,93 @@ +%define VPX_ARCH_ARM 0 +%define VPX_ARCH_AARCH64 0 +%define VPX_ARCH_MIPS 0 +%define VPX_ARCH_X86 1 +%define VPX_ARCH_X86_64 0 +%define VPX_ARCH_PPC 0 +%define VPX_ARCH_LOONGARCH 0 +%define HAVE_NEON_ASM 0 +%define HAVE_NEON 0 +%define HAVE_NEON_DOTPROD 0 +%define HAVE_NEON_I8MM 0 +%define HAVE_SVE 0 +%define HAVE_MIPS32 0 +%define HAVE_DSPR2 0 +%define HAVE_MSA 0 +%define HAVE_MIPS64 0 +%define HAVE_MMX 1 +%define HAVE_SSE 1 +%define HAVE_SSE2 1 +%define HAVE_SSE3 1 +%define HAVE_SSSE3 1 +%define HAVE_SSE4_1 1 +%define HAVE_AVX 1 +%define HAVE_AVX2 1 +%define HAVE_AVX512 0 +%define HAVE_VSX 0 +%define HAVE_MMI 0 +%define HAVE_LSX 0 +%define HAVE_LASX 0 +%define HAVE_VPX_PORTS 1 +%define HAVE_PTHREAD_H 1 +%define CONFIG_DEPENDENCY_TRACKING 1 +%define CONFIG_EXTERNAL_BUILD 1 +%define CONFIG_INSTALL_DOCS 0 +%define CONFIG_INSTALL_BINS 1 +%define CONFIG_INSTALL_LIBS 1 +%define CONFIG_INSTALL_SRCS 0 +%define CONFIG_DEBUG 0 +%define CONFIG_GPROF 0 +%define CONFIG_GCOV 0 +%define CONFIG_RVCT 0 +%define CONFIG_GCC 1 +%define CONFIG_MSVS 0 +%define CONFIG_PIC 1 +%define CONFIG_BIG_ENDIAN 0 +%define CONFIG_CODEC_SRCS 0 +%define CONFIG_DEBUG_LIBS 0 +%define CONFIG_DEQUANT_TOKENS 0 +%define CONFIG_DC_RECON 0 +%define CONFIG_RUNTIME_CPU_DETECT 1 +%define CONFIG_POSTPROC 1 +%define CONFIG_VP9_POSTPROC 1 +%define CONFIG_MULTITHREAD 1 +%define CONFIG_INTERNAL_STATS 0 +%define CONFIG_VP8_ENCODER 1 +%define CONFIG_VP8_DECODER 1 +%define CONFIG_VP9_ENCODER 1 +%define CONFIG_VP9_DECODER 1 +%define CONFIG_VP8 1 +%define CONFIG_VP9 1 +%define CONFIG_ENCODERS 1 +%define CONFIG_DECODERS 1 +%define CONFIG_STATIC_MSVCRT 0 +%define CONFIG_SPATIAL_RESAMPLING 1 +%define CONFIG_REALTIME_ONLY 0 +%define CONFIG_ONTHEFLY_BITPACKING 0 +%define CONFIG_ERROR_CONCEALMENT 0 +%define CONFIG_SHARED 0 +%define CONFIG_STATIC 1 +%define CONFIG_SMALL 0 +%define CONFIG_POSTPROC_VISUALIZER 0 +%define CONFIG_OS_SUPPORT 1 +%define CONFIG_UNIT_TESTS 0 +%define CONFIG_WEBM_IO 1 +%define CONFIG_LIBYUV 1 +%define CONFIG_DECODE_PERF_TESTS 0 +%define CONFIG_ENCODE_PERF_TESTS 0 +%define CONFIG_MULTI_RES_ENCODING 1 +%define CONFIG_TEMPORAL_DENOISING 1 +%define CONFIG_VP9_TEMPORAL_DENOISING 0 +%define CONFIG_COEFFICIENT_RANGE_CHECKING 0 +%define CONFIG_VP9_HIGHBITDEPTH 0 +%define CONFIG_BETTER_HW_COMPATIBILITY 0 +%define CONFIG_EXPERIMENTAL 0 +%define CONFIG_SIZE_LIMIT 1 +%define CONFIG_ALWAYS_ADJUST_BPM 0 +%define CONFIG_BITSTREAM_DEBUG 0 +%define CONFIG_MISMATCH_DEBUG 0 +%define CONFIG_FP_MB_STATS 0 +%define CONFIG_EMULATE_HARDWARE 0 +%define CONFIG_NON_GREEDY_MV 0 +%define CONFIG_RATE_CTRL 0 +%define CONFIG_COLLECT_COMPONENT_TIMING 0 diff --git a/media/libvpx/config/linux/ia32/vpx_config.c b/media/libvpx/config/linux/ia32/vpx_config.c new file mode 100644 index 0000000000..6805ab62a8 --- /dev/null +++ b/media/libvpx/config/linux/ia32/vpx_config.c @@ -0,0 +1,10 @@ +/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */ +/* */ +/* Use of this source code is governed by a BSD-style license */ +/* that can be found in the LICENSE file in the root of the source */ +/* tree. An additional intellectual property rights grant can be found */ +/* in the file PATENTS. All contributing project authors may */ +/* be found in the AUTHORS file in the root of the source tree. */ +#include "vpx/vpx_codec.h" +static const char* const cfg = "--target=x86-linux-gcc --enable-external-build --disable-examples --disable-install-docs --disable-unit-tests --enable-multi-res-encoding --size-limit=8192x4608 --enable-pic --disable-avx512 --enable-postproc --enable-vp9-postproc --as=yasm"; +const char *vpx_codec_build_config(void) {return cfg;} diff --git a/media/libvpx/config/linux/ia32/vpx_config.h b/media/libvpx/config/linux/ia32/vpx_config.h new file mode 100644 index 0000000000..69fd63bf02 --- /dev/null +++ b/media/libvpx/config/linux/ia32/vpx_config.h @@ -0,0 +1,108 @@ +/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */ +/* */ +/* Use of this source code is governed by a BSD-style license */ +/* that can be found in the LICENSE file in the root of the source */ +/* tree. An additional intellectual property rights grant can be found */ +/* in the file PATENTS. All contributing project authors may */ +/* be found in the AUTHORS file in the root of the source tree. */ +/* This file automatically generated by configure. Do not edit! */ +#ifndef VPX_CONFIG_H +#define VPX_CONFIG_H +#define RESTRICT +#define INLINE inline +#define VPX_ARCH_ARM 0 +#define VPX_ARCH_AARCH64 0 +#define VPX_ARCH_MIPS 0 +#define VPX_ARCH_X86 1 +#define VPX_ARCH_X86_64 0 +#define VPX_ARCH_PPC 0 +#define VPX_ARCH_LOONGARCH 0 +#define HAVE_NEON_ASM 0 +#define HAVE_NEON 0 +#define HAVE_NEON_DOTPROD 0 +#define HAVE_NEON_I8MM 0 +#define HAVE_SVE 0 +#define HAVE_MIPS32 0 +#define HAVE_DSPR2 0 +#define HAVE_MSA 0 +#define HAVE_MIPS64 0 +#define HAVE_MMX 1 +#define HAVE_SSE 1 +#define HAVE_SSE2 1 +#define HAVE_SSE3 1 +#define HAVE_SSSE3 1 +#define HAVE_SSE4_1 1 +#define HAVE_AVX 1 +#define HAVE_AVX2 1 +#define HAVE_AVX512 0 +#define HAVE_VSX 0 +#define HAVE_MMI 0 +#define HAVE_LSX 0 +#define HAVE_LASX 0 +#define HAVE_VPX_PORTS 1 +#define HAVE_PTHREAD_H 1 +#define CONFIG_DEPENDENCY_TRACKING 1 +#define CONFIG_EXTERNAL_BUILD 1 +#define CONFIG_INSTALL_DOCS 0 +#define CONFIG_INSTALL_BINS 1 +#define CONFIG_INSTALL_LIBS 1 +#define CONFIG_INSTALL_SRCS 0 +#define CONFIG_DEBUG 0 +#define CONFIG_GPROF 0 +#define CONFIG_GCOV 0 +#define CONFIG_RVCT 0 +#define CONFIG_GCC 1 +#define CONFIG_MSVS 0 +#define CONFIG_PIC 1 +#define CONFIG_BIG_ENDIAN 0 +#define CONFIG_CODEC_SRCS 0 +#define CONFIG_DEBUG_LIBS 0 +#define CONFIG_DEQUANT_TOKENS 0 +#define CONFIG_DC_RECON 0 +#define CONFIG_RUNTIME_CPU_DETECT 1 +#define CONFIG_POSTPROC 1 +#define CONFIG_VP9_POSTPROC 1 +#define CONFIG_MULTITHREAD 1 +#define CONFIG_INTERNAL_STATS 0 +#define CONFIG_VP8_ENCODER 1 +#define CONFIG_VP8_DECODER 1 +#define CONFIG_VP9_ENCODER 1 +#define CONFIG_VP9_DECODER 1 +#define CONFIG_VP8 1 +#define CONFIG_VP9 1 +#define CONFIG_ENCODERS 1 +#define CONFIG_DECODERS 1 +#define CONFIG_STATIC_MSVCRT 0 +#define CONFIG_SPATIAL_RESAMPLING 1 +#define CONFIG_REALTIME_ONLY 0 +#define CONFIG_ONTHEFLY_BITPACKING 0 +#define CONFIG_ERROR_CONCEALMENT 0 +#define CONFIG_SHARED 0 +#define CONFIG_STATIC 1 +#define CONFIG_SMALL 0 +#define CONFIG_POSTPROC_VISUALIZER 0 +#define CONFIG_OS_SUPPORT 1 +#define CONFIG_UNIT_TESTS 0 +#define CONFIG_WEBM_IO 1 +#define CONFIG_LIBYUV 1 +#define CONFIG_DECODE_PERF_TESTS 0 +#define CONFIG_ENCODE_PERF_TESTS 0 +#define CONFIG_MULTI_RES_ENCODING 1 +#define CONFIG_TEMPORAL_DENOISING 1 +#define CONFIG_VP9_TEMPORAL_DENOISING 0 +#define CONFIG_COEFFICIENT_RANGE_CHECKING 0 +#define CONFIG_VP9_HIGHBITDEPTH 0 +#define CONFIG_BETTER_HW_COMPATIBILITY 0 +#define CONFIG_EXPERIMENTAL 0 +#define CONFIG_SIZE_LIMIT 1 +#define CONFIG_ALWAYS_ADJUST_BPM 0 +#define CONFIG_BITSTREAM_DEBUG 0 +#define CONFIG_MISMATCH_DEBUG 0 +#define CONFIG_FP_MB_STATS 0 +#define CONFIG_EMULATE_HARDWARE 0 +#define CONFIG_NON_GREEDY_MV 0 +#define CONFIG_RATE_CTRL 0 +#define CONFIG_COLLECT_COMPONENT_TIMING 0 +#define DECODE_WIDTH_LIMIT 8192 +#define DECODE_HEIGHT_LIMIT 4608 +#endif /* VPX_CONFIG_H */ diff --git a/media/libvpx/config/linux/ia32/vpx_dsp_rtcd.h b/media/libvpx/config/linux/ia32/vpx_dsp_rtcd.h new file mode 100644 index 0000000000..dd688b035b --- /dev/null +++ b/media/libvpx/config/linux/ia32/vpx_dsp_rtcd.h @@ -0,0 +1,1604 @@ +// This file is generated. Do not edit. +#ifndef VPX_DSP_RTCD_H_ +#define VPX_DSP_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +/* + * DSP + */ + +#include "vpx/vpx_integer.h" +#include "vpx_dsp/vpx_dsp_common.h" +#include "vpx_dsp/vpx_filter.h" +#if CONFIG_VP9_ENCODER + struct macroblock_plane; + struct ScanOrder; +#endif + + +#ifdef __cplusplus +extern "C" { +#endif + +unsigned int vpx_avg_4x4_c(const uint8_t *, int p); +unsigned int vpx_avg_4x4_sse2(const uint8_t *, int p); +RTCD_EXTERN unsigned int (*vpx_avg_4x4)(const uint8_t *, int p); + +unsigned int vpx_avg_8x8_c(const uint8_t *, int p); +unsigned int vpx_avg_8x8_sse2(const uint8_t *, int p); +RTCD_EXTERN unsigned int (*vpx_avg_8x8)(const uint8_t *, int p); + +void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride); +void vpx_comp_avg_pred_sse2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride); +void vpx_comp_avg_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride); +RTCD_EXTERN void (*vpx_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride); + +void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_avg)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_avg_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_avg_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve_avg)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c + +void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c + +void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c + +void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c + +void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_c + +void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_c + +void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_c + +void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_c + +void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d153_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d153_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d153_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d153_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d207_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d207_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d207_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d207_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d45_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d45_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d45_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d45_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c + +void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d63_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d63_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d63_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d63_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c + +void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_128_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_128_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_128_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_128_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_left_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_left_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_left_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_left_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_top_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_top_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_top_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_dc_top_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct16x16_sse2(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct16x16_avx2(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct16x16)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct16x16_1)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_sse2(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_avx2(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct32x32)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_1_sse2(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct32x32_1)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_rd_sse2(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_rd_avx2(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct32x32_rd)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct4x4)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct4x4_1_sse2(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct4x4_1)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct8x8_sse2(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct8x8)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct8x8_1_sse2(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct8x8_1)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_get16x16var_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +void vpx_get16x16var_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +void vpx_get16x16var_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +RTCD_EXTERN void (*vpx_get16x16var)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); + +unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride); +#define vpx_get4x4sse_cs vpx_get4x4sse_cs_c + +void vpx_get8x8var_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +void vpx_get8x8var_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +RTCD_EXTERN void (*vpx_get8x8var)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); + +unsigned int vpx_get_mb_ss_c(const int16_t *); +unsigned int vpx_get_mb_ss_sse2(const int16_t *); +RTCD_EXTERN unsigned int (*vpx_get_mb_ss)(const int16_t *); + +void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_h_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_h_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_h_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_h_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +RTCD_EXTERN void (*vpx_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); + +void vpx_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_32x32_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +RTCD_EXTERN void (*vpx_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); + +void vpx_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +RTCD_EXTERN void (*vpx_hadamard_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); + +void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c + +void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct16x16_10_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct16x16_1_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_256_add_avx2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct16x16_256_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct16x16_38_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_38_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct16x16_38_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_1024_add_avx2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct32x32_1024_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_135_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_135_add_ssse3(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_135_add_avx2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct32x32_135_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct32x32_1_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_34_add_ssse3(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct32x32_34_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct4x4_16_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct4x4_1_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_12_add_ssse3(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct8x8_12_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct8x8_1_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct8x8_64_add)(const tran_low_t *input, uint8_t *dest, int stride); + +int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width); +int16_t vpx_int_pro_col_sse2(const uint8_t *ref, const int width); +RTCD_EXTERN int16_t (*vpx_int_pro_col)(const uint8_t *ref, const int width); + +void vpx_int_pro_row_c(int16_t hbuf[16], const uint8_t *ref, const int ref_stride, const int height); +void vpx_int_pro_row_sse2(int16_t hbuf[16], const uint8_t *ref, const int ref_stride, const int height); +RTCD_EXTERN void (*vpx_int_pro_row)(int16_t hbuf[16], const uint8_t *ref, const int ref_stride, const int height); + +void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_iwht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c + +void vpx_lpf_horizontal_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_16_avx2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_horizontal_16)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_horizontal_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_16_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_16_dual_avx2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_horizontal_16_dual)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_horizontal_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*vpx_lpf_horizontal_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_horizontal_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*vpx_lpf_horizontal_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_vertical_16)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_16_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_vertical_16_dual)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_vertical_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*vpx_lpf_vertical_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_vertical_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*vpx_lpf_vertical_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void vpx_mbpost_proc_across_ip_c(unsigned char *src, int pitch, int rows, int cols,int flimit); +void vpx_mbpost_proc_across_ip_sse2(unsigned char *src, int pitch, int rows, int cols,int flimit); +RTCD_EXTERN void (*vpx_mbpost_proc_across_ip)(unsigned char *src, int pitch, int rows, int cols,int flimit); + +void vpx_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int cols,int flimit); +void vpx_mbpost_proc_down_sse2(unsigned char *dst, int pitch, int rows, int cols,int flimit); +RTCD_EXTERN void (*vpx_mbpost_proc_down)(unsigned char *dst, int pitch, int rows, int cols,int flimit); + +void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max); +void vpx_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max); +RTCD_EXTERN void (*vpx_minmax_8x8)(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max); + +unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +void vpx_plane_add_noise_c(uint8_t *start, const int8_t *noise, int blackclamp, int whiteclamp, int width, int height, int pitch); +void vpx_plane_add_noise_sse2(uint8_t *start, const int8_t *noise, int blackclamp, int whiteclamp, int width, int height, int pitch); +RTCD_EXTERN void (*vpx_plane_add_noise)(uint8_t *start, const int8_t *noise, int blackclamp, int whiteclamp, int width, int height, int pitch); + +void vpx_post_proc_down_and_across_mb_row_c(unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size); +void vpx_post_proc_down_and_across_mb_row_sse2(unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size); +RTCD_EXTERN void (*vpx_post_proc_down_and_across_mb_row)(unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size); + +void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +RTCD_EXTERN void (*vpx_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); + +void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_32x32_ssse3(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_32x32_avx(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +RTCD_EXTERN void (*vpx_quantize_b_32x32)(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); + +unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad4x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad4x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad4x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad8x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad8x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad8x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad8x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_4x4 vpx_sad_skip_4x4_c + +void vpx_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_4x4x4d vpx_sad_skip_4x4x4d_c + +unsigned int vpx_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_8x4 vpx_sad_skip_8x4_c + +void vpx_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_8x4x4d vpx_sad_skip_8x4x4d_c + +unsigned int vpx_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +int vpx_satd_c(const int16_t *coeff, int length); +int vpx_satd_sse2(const int16_t *coeff, int length); +int vpx_satd_avx2(const int16_t *coeff, int length); +RTCD_EXTERN int (*vpx_satd)(const int16_t *coeff, int length); + +void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c + +void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c + +void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c + +void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_horiz vpx_scaled_horiz_c + +void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_vert vpx_scaled_vert_c + +int64_t vpx_sse_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); +int64_t vpx_sse_sse4_1(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); +int64_t vpx_sse_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); +RTCD_EXTERN int64_t (*vpx_sse)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); + +uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance16x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance16x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance16x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance32x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance32x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance32x64)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance4x4)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance4x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance64x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance64x64)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance8x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance8x4)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance8x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride); +void vpx_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride); +void vpx_subtract_block_avx2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride); +RTCD_EXTERN void (*vpx_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride); + +uint64_t vpx_sum_squares_2d_i16_c(const int16_t *src, int stride, int size); +uint64_t vpx_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int size); +RTCD_EXTERN uint64_t (*vpx_sum_squares_2d_i16)(const int16_t *src, int stride, int size); + +void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_tm_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_tm_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_tm_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_tm_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_v_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_v_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_v_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_v_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x4_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c + +int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl); +int vpx_vector_var_sse2(const int16_t *ref, const int16_t *src, const int bwl); +RTCD_EXTERN int (*vpx_vector_var)(const int16_t *ref, const int16_t *src, const int bwl); + +void vpx_dsp_rtcd(void); + +#ifdef RTCD_C +#include "vpx_ports/x86.h" +static void setup_rtcd_internal(void) +{ + int flags = x86_simd_caps(); + + (void)flags; + + vpx_avg_4x4 = vpx_avg_4x4_c; + if (flags & HAS_SSE2) vpx_avg_4x4 = vpx_avg_4x4_sse2; + vpx_avg_8x8 = vpx_avg_8x8_c; + if (flags & HAS_SSE2) vpx_avg_8x8 = vpx_avg_8x8_sse2; + vpx_comp_avg_pred = vpx_comp_avg_pred_c; + if (flags & HAS_SSE2) vpx_comp_avg_pred = vpx_comp_avg_pred_sse2; + if (flags & HAS_AVX2) vpx_comp_avg_pred = vpx_comp_avg_pred_avx2; + vpx_convolve8 = vpx_convolve8_c; + if (flags & HAS_SSE2) vpx_convolve8 = vpx_convolve8_sse2; + if (flags & HAS_SSSE3) vpx_convolve8 = vpx_convolve8_ssse3; + if (flags & HAS_AVX2) vpx_convolve8 = vpx_convolve8_avx2; + vpx_convolve8_avg = vpx_convolve8_avg_c; + if (flags & HAS_SSE2) vpx_convolve8_avg = vpx_convolve8_avg_sse2; + if (flags & HAS_SSSE3) vpx_convolve8_avg = vpx_convolve8_avg_ssse3; + if (flags & HAS_AVX2) vpx_convolve8_avg = vpx_convolve8_avg_avx2; + vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_c; + if (flags & HAS_SSE2) vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_sse2; + if (flags & HAS_SSSE3) vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_ssse3; + if (flags & HAS_AVX2) vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_avx2; + vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_c; + if (flags & HAS_SSE2) vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_sse2; + if (flags & HAS_SSSE3) vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_ssse3; + if (flags & HAS_AVX2) vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_avx2; + vpx_convolve8_horiz = vpx_convolve8_horiz_c; + if (flags & HAS_SSE2) vpx_convolve8_horiz = vpx_convolve8_horiz_sse2; + if (flags & HAS_SSSE3) vpx_convolve8_horiz = vpx_convolve8_horiz_ssse3; + if (flags & HAS_AVX2) vpx_convolve8_horiz = vpx_convolve8_horiz_avx2; + vpx_convolve8_vert = vpx_convolve8_vert_c; + if (flags & HAS_SSE2) vpx_convolve8_vert = vpx_convolve8_vert_sse2; + if (flags & HAS_SSSE3) vpx_convolve8_vert = vpx_convolve8_vert_ssse3; + if (flags & HAS_AVX2) vpx_convolve8_vert = vpx_convolve8_vert_avx2; + vpx_convolve_avg = vpx_convolve_avg_c; + if (flags & HAS_SSE2) vpx_convolve_avg = vpx_convolve_avg_sse2; + vpx_convolve_copy = vpx_convolve_copy_c; + if (flags & HAS_SSE2) vpx_convolve_copy = vpx_convolve_copy_sse2; + vpx_d153_predictor_16x16 = vpx_d153_predictor_16x16_c; + if (flags & HAS_SSSE3) vpx_d153_predictor_16x16 = vpx_d153_predictor_16x16_ssse3; + vpx_d153_predictor_32x32 = vpx_d153_predictor_32x32_c; + if (flags & HAS_SSSE3) vpx_d153_predictor_32x32 = vpx_d153_predictor_32x32_ssse3; + vpx_d153_predictor_4x4 = vpx_d153_predictor_4x4_c; + if (flags & HAS_SSSE3) vpx_d153_predictor_4x4 = vpx_d153_predictor_4x4_ssse3; + vpx_d153_predictor_8x8 = vpx_d153_predictor_8x8_c; + if (flags & HAS_SSSE3) vpx_d153_predictor_8x8 = vpx_d153_predictor_8x8_ssse3; + vpx_d207_predictor_16x16 = vpx_d207_predictor_16x16_c; + if (flags & HAS_SSSE3) vpx_d207_predictor_16x16 = vpx_d207_predictor_16x16_ssse3; + vpx_d207_predictor_32x32 = vpx_d207_predictor_32x32_c; + if (flags & HAS_SSSE3) vpx_d207_predictor_32x32 = vpx_d207_predictor_32x32_ssse3; + vpx_d207_predictor_4x4 = vpx_d207_predictor_4x4_c; + if (flags & HAS_SSE2) vpx_d207_predictor_4x4 = vpx_d207_predictor_4x4_sse2; + vpx_d207_predictor_8x8 = vpx_d207_predictor_8x8_c; + if (flags & HAS_SSSE3) vpx_d207_predictor_8x8 = vpx_d207_predictor_8x8_ssse3; + vpx_d45_predictor_16x16 = vpx_d45_predictor_16x16_c; + if (flags & HAS_SSSE3) vpx_d45_predictor_16x16 = vpx_d45_predictor_16x16_ssse3; + vpx_d45_predictor_32x32 = vpx_d45_predictor_32x32_c; + if (flags & HAS_SSSE3) vpx_d45_predictor_32x32 = vpx_d45_predictor_32x32_ssse3; + vpx_d45_predictor_4x4 = vpx_d45_predictor_4x4_c; + if (flags & HAS_SSE2) vpx_d45_predictor_4x4 = vpx_d45_predictor_4x4_sse2; + vpx_d45_predictor_8x8 = vpx_d45_predictor_8x8_c; + if (flags & HAS_SSE2) vpx_d45_predictor_8x8 = vpx_d45_predictor_8x8_sse2; + vpx_d63_predictor_16x16 = vpx_d63_predictor_16x16_c; + if (flags & HAS_SSSE3) vpx_d63_predictor_16x16 = vpx_d63_predictor_16x16_ssse3; + vpx_d63_predictor_32x32 = vpx_d63_predictor_32x32_c; + if (flags & HAS_SSSE3) vpx_d63_predictor_32x32 = vpx_d63_predictor_32x32_ssse3; + vpx_d63_predictor_4x4 = vpx_d63_predictor_4x4_c; + if (flags & HAS_SSSE3) vpx_d63_predictor_4x4 = vpx_d63_predictor_4x4_ssse3; + vpx_d63_predictor_8x8 = vpx_d63_predictor_8x8_c; + if (flags & HAS_SSSE3) vpx_d63_predictor_8x8 = vpx_d63_predictor_8x8_ssse3; + vpx_dc_128_predictor_16x16 = vpx_dc_128_predictor_16x16_c; + if (flags & HAS_SSE2) vpx_dc_128_predictor_16x16 = vpx_dc_128_predictor_16x16_sse2; + vpx_dc_128_predictor_32x32 = vpx_dc_128_predictor_32x32_c; + if (flags & HAS_SSE2) vpx_dc_128_predictor_32x32 = vpx_dc_128_predictor_32x32_sse2; + vpx_dc_128_predictor_4x4 = vpx_dc_128_predictor_4x4_c; + if (flags & HAS_SSE2) vpx_dc_128_predictor_4x4 = vpx_dc_128_predictor_4x4_sse2; + vpx_dc_128_predictor_8x8 = vpx_dc_128_predictor_8x8_c; + if (flags & HAS_SSE2) vpx_dc_128_predictor_8x8 = vpx_dc_128_predictor_8x8_sse2; + vpx_dc_left_predictor_16x16 = vpx_dc_left_predictor_16x16_c; + if (flags & HAS_SSE2) vpx_dc_left_predictor_16x16 = vpx_dc_left_predictor_16x16_sse2; + vpx_dc_left_predictor_32x32 = vpx_dc_left_predictor_32x32_c; + if (flags & HAS_SSE2) vpx_dc_left_predictor_32x32 = vpx_dc_left_predictor_32x32_sse2; + vpx_dc_left_predictor_4x4 = vpx_dc_left_predictor_4x4_c; + if (flags & HAS_SSE2) vpx_dc_left_predictor_4x4 = vpx_dc_left_predictor_4x4_sse2; + vpx_dc_left_predictor_8x8 = vpx_dc_left_predictor_8x8_c; + if (flags & HAS_SSE2) vpx_dc_left_predictor_8x8 = vpx_dc_left_predictor_8x8_sse2; + vpx_dc_predictor_16x16 = vpx_dc_predictor_16x16_c; + if (flags & HAS_SSE2) vpx_dc_predictor_16x16 = vpx_dc_predictor_16x16_sse2; + vpx_dc_predictor_32x32 = vpx_dc_predictor_32x32_c; + if (flags & HAS_SSE2) vpx_dc_predictor_32x32 = vpx_dc_predictor_32x32_sse2; + vpx_dc_predictor_4x4 = vpx_dc_predictor_4x4_c; + if (flags & HAS_SSE2) vpx_dc_predictor_4x4 = vpx_dc_predictor_4x4_sse2; + vpx_dc_predictor_8x8 = vpx_dc_predictor_8x8_c; + if (flags & HAS_SSE2) vpx_dc_predictor_8x8 = vpx_dc_predictor_8x8_sse2; + vpx_dc_top_predictor_16x16 = vpx_dc_top_predictor_16x16_c; + if (flags & HAS_SSE2) vpx_dc_top_predictor_16x16 = vpx_dc_top_predictor_16x16_sse2; + vpx_dc_top_predictor_32x32 = vpx_dc_top_predictor_32x32_c; + if (flags & HAS_SSE2) vpx_dc_top_predictor_32x32 = vpx_dc_top_predictor_32x32_sse2; + vpx_dc_top_predictor_4x4 = vpx_dc_top_predictor_4x4_c; + if (flags & HAS_SSE2) vpx_dc_top_predictor_4x4 = vpx_dc_top_predictor_4x4_sse2; + vpx_dc_top_predictor_8x8 = vpx_dc_top_predictor_8x8_c; + if (flags & HAS_SSE2) vpx_dc_top_predictor_8x8 = vpx_dc_top_predictor_8x8_sse2; + vpx_fdct16x16 = vpx_fdct16x16_c; + if (flags & HAS_SSE2) vpx_fdct16x16 = vpx_fdct16x16_sse2; + if (flags & HAS_AVX2) vpx_fdct16x16 = vpx_fdct16x16_avx2; + vpx_fdct16x16_1 = vpx_fdct16x16_1_c; + if (flags & HAS_SSE2) vpx_fdct16x16_1 = vpx_fdct16x16_1_sse2; + vpx_fdct32x32 = vpx_fdct32x32_c; + if (flags & HAS_SSE2) vpx_fdct32x32 = vpx_fdct32x32_sse2; + if (flags & HAS_AVX2) vpx_fdct32x32 = vpx_fdct32x32_avx2; + vpx_fdct32x32_1 = vpx_fdct32x32_1_c; + if (flags & HAS_SSE2) vpx_fdct32x32_1 = vpx_fdct32x32_1_sse2; + vpx_fdct32x32_rd = vpx_fdct32x32_rd_c; + if (flags & HAS_SSE2) vpx_fdct32x32_rd = vpx_fdct32x32_rd_sse2; + if (flags & HAS_AVX2) vpx_fdct32x32_rd = vpx_fdct32x32_rd_avx2; + vpx_fdct4x4 = vpx_fdct4x4_c; + if (flags & HAS_SSE2) vpx_fdct4x4 = vpx_fdct4x4_sse2; + vpx_fdct4x4_1 = vpx_fdct4x4_1_c; + if (flags & HAS_SSE2) vpx_fdct4x4_1 = vpx_fdct4x4_1_sse2; + vpx_fdct8x8 = vpx_fdct8x8_c; + if (flags & HAS_SSE2) vpx_fdct8x8 = vpx_fdct8x8_sse2; + vpx_fdct8x8_1 = vpx_fdct8x8_1_c; + if (flags & HAS_SSE2) vpx_fdct8x8_1 = vpx_fdct8x8_1_sse2; + vpx_get16x16var = vpx_get16x16var_c; + if (flags & HAS_SSE2) vpx_get16x16var = vpx_get16x16var_sse2; + if (flags & HAS_AVX2) vpx_get16x16var = vpx_get16x16var_avx2; + vpx_get8x8var = vpx_get8x8var_c; + if (flags & HAS_SSE2) vpx_get8x8var = vpx_get8x8var_sse2; + vpx_get_mb_ss = vpx_get_mb_ss_c; + if (flags & HAS_SSE2) vpx_get_mb_ss = vpx_get_mb_ss_sse2; + vpx_h_predictor_16x16 = vpx_h_predictor_16x16_c; + if (flags & HAS_SSE2) vpx_h_predictor_16x16 = vpx_h_predictor_16x16_sse2; + vpx_h_predictor_32x32 = vpx_h_predictor_32x32_c; + if (flags & HAS_SSE2) vpx_h_predictor_32x32 = vpx_h_predictor_32x32_sse2; + vpx_h_predictor_4x4 = vpx_h_predictor_4x4_c; + if (flags & HAS_SSE2) vpx_h_predictor_4x4 = vpx_h_predictor_4x4_sse2; + vpx_h_predictor_8x8 = vpx_h_predictor_8x8_c; + if (flags & HAS_SSE2) vpx_h_predictor_8x8 = vpx_h_predictor_8x8_sse2; + vpx_hadamard_16x16 = vpx_hadamard_16x16_c; + if (flags & HAS_SSE2) vpx_hadamard_16x16 = vpx_hadamard_16x16_sse2; + if (flags & HAS_AVX2) vpx_hadamard_16x16 = vpx_hadamard_16x16_avx2; + vpx_hadamard_32x32 = vpx_hadamard_32x32_c; + if (flags & HAS_SSE2) vpx_hadamard_32x32 = vpx_hadamard_32x32_sse2; + if (flags & HAS_AVX2) vpx_hadamard_32x32 = vpx_hadamard_32x32_avx2; + vpx_hadamard_8x8 = vpx_hadamard_8x8_c; + if (flags & HAS_SSE2) vpx_hadamard_8x8 = vpx_hadamard_8x8_sse2; + vpx_idct16x16_10_add = vpx_idct16x16_10_add_c; + if (flags & HAS_SSE2) vpx_idct16x16_10_add = vpx_idct16x16_10_add_sse2; + vpx_idct16x16_1_add = vpx_idct16x16_1_add_c; + if (flags & HAS_SSE2) vpx_idct16x16_1_add = vpx_idct16x16_1_add_sse2; + vpx_idct16x16_256_add = vpx_idct16x16_256_add_c; + if (flags & HAS_SSE2) vpx_idct16x16_256_add = vpx_idct16x16_256_add_sse2; + if (flags & HAS_AVX2) vpx_idct16x16_256_add = vpx_idct16x16_256_add_avx2; + vpx_idct16x16_38_add = vpx_idct16x16_38_add_c; + if (flags & HAS_SSE2) vpx_idct16x16_38_add = vpx_idct16x16_38_add_sse2; + vpx_idct32x32_1024_add = vpx_idct32x32_1024_add_c; + if (flags & HAS_SSE2) vpx_idct32x32_1024_add = vpx_idct32x32_1024_add_sse2; + if (flags & HAS_AVX2) vpx_idct32x32_1024_add = vpx_idct32x32_1024_add_avx2; + vpx_idct32x32_135_add = vpx_idct32x32_135_add_c; + if (flags & HAS_SSE2) vpx_idct32x32_135_add = vpx_idct32x32_135_add_sse2; + if (flags & HAS_SSSE3) vpx_idct32x32_135_add = vpx_idct32x32_135_add_ssse3; + if (flags & HAS_AVX2) vpx_idct32x32_135_add = vpx_idct32x32_135_add_avx2; + vpx_idct32x32_1_add = vpx_idct32x32_1_add_c; + if (flags & HAS_SSE2) vpx_idct32x32_1_add = vpx_idct32x32_1_add_sse2; + vpx_idct32x32_34_add = vpx_idct32x32_34_add_c; + if (flags & HAS_SSE2) vpx_idct32x32_34_add = vpx_idct32x32_34_add_sse2; + if (flags & HAS_SSSE3) vpx_idct32x32_34_add = vpx_idct32x32_34_add_ssse3; + vpx_idct4x4_16_add = vpx_idct4x4_16_add_c; + if (flags & HAS_SSE2) vpx_idct4x4_16_add = vpx_idct4x4_16_add_sse2; + vpx_idct4x4_1_add = vpx_idct4x4_1_add_c; + if (flags & HAS_SSE2) vpx_idct4x4_1_add = vpx_idct4x4_1_add_sse2; + vpx_idct8x8_12_add = vpx_idct8x8_12_add_c; + if (flags & HAS_SSE2) vpx_idct8x8_12_add = vpx_idct8x8_12_add_sse2; + if (flags & HAS_SSSE3) vpx_idct8x8_12_add = vpx_idct8x8_12_add_ssse3; + vpx_idct8x8_1_add = vpx_idct8x8_1_add_c; + if (flags & HAS_SSE2) vpx_idct8x8_1_add = vpx_idct8x8_1_add_sse2; + vpx_idct8x8_64_add = vpx_idct8x8_64_add_c; + if (flags & HAS_SSE2) vpx_idct8x8_64_add = vpx_idct8x8_64_add_sse2; + vpx_int_pro_col = vpx_int_pro_col_c; + if (flags & HAS_SSE2) vpx_int_pro_col = vpx_int_pro_col_sse2; + vpx_int_pro_row = vpx_int_pro_row_c; + if (flags & HAS_SSE2) vpx_int_pro_row = vpx_int_pro_row_sse2; + vpx_iwht4x4_16_add = vpx_iwht4x4_16_add_c; + if (flags & HAS_SSE2) vpx_iwht4x4_16_add = vpx_iwht4x4_16_add_sse2; + vpx_lpf_horizontal_16 = vpx_lpf_horizontal_16_c; + if (flags & HAS_SSE2) vpx_lpf_horizontal_16 = vpx_lpf_horizontal_16_sse2; + if (flags & HAS_AVX2) vpx_lpf_horizontal_16 = vpx_lpf_horizontal_16_avx2; + vpx_lpf_horizontal_16_dual = vpx_lpf_horizontal_16_dual_c; + if (flags & HAS_SSE2) vpx_lpf_horizontal_16_dual = vpx_lpf_horizontal_16_dual_sse2; + if (flags & HAS_AVX2) vpx_lpf_horizontal_16_dual = vpx_lpf_horizontal_16_dual_avx2; + vpx_lpf_horizontal_4 = vpx_lpf_horizontal_4_c; + if (flags & HAS_SSE2) vpx_lpf_horizontal_4 = vpx_lpf_horizontal_4_sse2; + vpx_lpf_horizontal_4_dual = vpx_lpf_horizontal_4_dual_c; + if (flags & HAS_SSE2) vpx_lpf_horizontal_4_dual = vpx_lpf_horizontal_4_dual_sse2; + vpx_lpf_horizontal_8 = vpx_lpf_horizontal_8_c; + if (flags & HAS_SSE2) vpx_lpf_horizontal_8 = vpx_lpf_horizontal_8_sse2; + vpx_lpf_horizontal_8_dual = vpx_lpf_horizontal_8_dual_c; + if (flags & HAS_SSE2) vpx_lpf_horizontal_8_dual = vpx_lpf_horizontal_8_dual_sse2; + vpx_lpf_vertical_16 = vpx_lpf_vertical_16_c; + if (flags & HAS_SSE2) vpx_lpf_vertical_16 = vpx_lpf_vertical_16_sse2; + vpx_lpf_vertical_16_dual = vpx_lpf_vertical_16_dual_c; + if (flags & HAS_SSE2) vpx_lpf_vertical_16_dual = vpx_lpf_vertical_16_dual_sse2; + vpx_lpf_vertical_4 = vpx_lpf_vertical_4_c; + if (flags & HAS_SSE2) vpx_lpf_vertical_4 = vpx_lpf_vertical_4_sse2; + vpx_lpf_vertical_4_dual = vpx_lpf_vertical_4_dual_c; + if (flags & HAS_SSE2) vpx_lpf_vertical_4_dual = vpx_lpf_vertical_4_dual_sse2; + vpx_lpf_vertical_8 = vpx_lpf_vertical_8_c; + if (flags & HAS_SSE2) vpx_lpf_vertical_8 = vpx_lpf_vertical_8_sse2; + vpx_lpf_vertical_8_dual = vpx_lpf_vertical_8_dual_c; + if (flags & HAS_SSE2) vpx_lpf_vertical_8_dual = vpx_lpf_vertical_8_dual_sse2; + vpx_mbpost_proc_across_ip = vpx_mbpost_proc_across_ip_c; + if (flags & HAS_SSE2) vpx_mbpost_proc_across_ip = vpx_mbpost_proc_across_ip_sse2; + vpx_mbpost_proc_down = vpx_mbpost_proc_down_c; + if (flags & HAS_SSE2) vpx_mbpost_proc_down = vpx_mbpost_proc_down_sse2; + vpx_minmax_8x8 = vpx_minmax_8x8_c; + if (flags & HAS_SSE2) vpx_minmax_8x8 = vpx_minmax_8x8_sse2; + vpx_mse16x16 = vpx_mse16x16_c; + if (flags & HAS_SSE2) vpx_mse16x16 = vpx_mse16x16_sse2; + if (flags & HAS_AVX2) vpx_mse16x16 = vpx_mse16x16_avx2; + vpx_mse16x8 = vpx_mse16x8_c; + if (flags & HAS_SSE2) vpx_mse16x8 = vpx_mse16x8_sse2; + if (flags & HAS_AVX2) vpx_mse16x8 = vpx_mse16x8_avx2; + vpx_mse8x16 = vpx_mse8x16_c; + if (flags & HAS_SSE2) vpx_mse8x16 = vpx_mse8x16_sse2; + vpx_mse8x8 = vpx_mse8x8_c; + if (flags & HAS_SSE2) vpx_mse8x8 = vpx_mse8x8_sse2; + vpx_plane_add_noise = vpx_plane_add_noise_c; + if (flags & HAS_SSE2) vpx_plane_add_noise = vpx_plane_add_noise_sse2; + vpx_post_proc_down_and_across_mb_row = vpx_post_proc_down_and_across_mb_row_c; + if (flags & HAS_SSE2) vpx_post_proc_down_and_across_mb_row = vpx_post_proc_down_and_across_mb_row_sse2; + vpx_quantize_b = vpx_quantize_b_c; + if (flags & HAS_SSE2) vpx_quantize_b = vpx_quantize_b_sse2; + if (flags & HAS_SSSE3) vpx_quantize_b = vpx_quantize_b_ssse3; + if (flags & HAS_AVX) vpx_quantize_b = vpx_quantize_b_avx; + if (flags & HAS_AVX2) vpx_quantize_b = vpx_quantize_b_avx2; + vpx_quantize_b_32x32 = vpx_quantize_b_32x32_c; + if (flags & HAS_SSSE3) vpx_quantize_b_32x32 = vpx_quantize_b_32x32_ssse3; + if (flags & HAS_AVX) vpx_quantize_b_32x32 = vpx_quantize_b_32x32_avx; + if (flags & HAS_AVX2) vpx_quantize_b_32x32 = vpx_quantize_b_32x32_avx2; + vpx_sad16x16 = vpx_sad16x16_c; + if (flags & HAS_SSE2) vpx_sad16x16 = vpx_sad16x16_sse2; + vpx_sad16x16_avg = vpx_sad16x16_avg_c; + if (flags & HAS_SSE2) vpx_sad16x16_avg = vpx_sad16x16_avg_sse2; + vpx_sad16x16x4d = vpx_sad16x16x4d_c; + if (flags & HAS_SSE2) vpx_sad16x16x4d = vpx_sad16x16x4d_sse2; + vpx_sad16x32 = vpx_sad16x32_c; + if (flags & HAS_SSE2) vpx_sad16x32 = vpx_sad16x32_sse2; + vpx_sad16x32_avg = vpx_sad16x32_avg_c; + if (flags & HAS_SSE2) vpx_sad16x32_avg = vpx_sad16x32_avg_sse2; + vpx_sad16x32x4d = vpx_sad16x32x4d_c; + if (flags & HAS_SSE2) vpx_sad16x32x4d = vpx_sad16x32x4d_sse2; + vpx_sad16x8 = vpx_sad16x8_c; + if (flags & HAS_SSE2) vpx_sad16x8 = vpx_sad16x8_sse2; + vpx_sad16x8_avg = vpx_sad16x8_avg_c; + if (flags & HAS_SSE2) vpx_sad16x8_avg = vpx_sad16x8_avg_sse2; + vpx_sad16x8x4d = vpx_sad16x8x4d_c; + if (flags & HAS_SSE2) vpx_sad16x8x4d = vpx_sad16x8x4d_sse2; + vpx_sad32x16 = vpx_sad32x16_c; + if (flags & HAS_SSE2) vpx_sad32x16 = vpx_sad32x16_sse2; + if (flags & HAS_AVX2) vpx_sad32x16 = vpx_sad32x16_avx2; + vpx_sad32x16_avg = vpx_sad32x16_avg_c; + if (flags & HAS_SSE2) vpx_sad32x16_avg = vpx_sad32x16_avg_sse2; + if (flags & HAS_AVX2) vpx_sad32x16_avg = vpx_sad32x16_avg_avx2; + vpx_sad32x16x4d = vpx_sad32x16x4d_c; + if (flags & HAS_SSE2) vpx_sad32x16x4d = vpx_sad32x16x4d_sse2; + vpx_sad32x32 = vpx_sad32x32_c; + if (flags & HAS_SSE2) vpx_sad32x32 = vpx_sad32x32_sse2; + if (flags & HAS_AVX2) vpx_sad32x32 = vpx_sad32x32_avx2; + vpx_sad32x32_avg = vpx_sad32x32_avg_c; + if (flags & HAS_SSE2) vpx_sad32x32_avg = vpx_sad32x32_avg_sse2; + if (flags & HAS_AVX2) vpx_sad32x32_avg = vpx_sad32x32_avg_avx2; + vpx_sad32x32x4d = vpx_sad32x32x4d_c; + if (flags & HAS_SSE2) vpx_sad32x32x4d = vpx_sad32x32x4d_sse2; + if (flags & HAS_AVX2) vpx_sad32x32x4d = vpx_sad32x32x4d_avx2; + vpx_sad32x64 = vpx_sad32x64_c; + if (flags & HAS_SSE2) vpx_sad32x64 = vpx_sad32x64_sse2; + if (flags & HAS_AVX2) vpx_sad32x64 = vpx_sad32x64_avx2; + vpx_sad32x64_avg = vpx_sad32x64_avg_c; + if (flags & HAS_SSE2) vpx_sad32x64_avg = vpx_sad32x64_avg_sse2; + if (flags & HAS_AVX2) vpx_sad32x64_avg = vpx_sad32x64_avg_avx2; + vpx_sad32x64x4d = vpx_sad32x64x4d_c; + if (flags & HAS_SSE2) vpx_sad32x64x4d = vpx_sad32x64x4d_sse2; + vpx_sad4x4 = vpx_sad4x4_c; + if (flags & HAS_SSE2) vpx_sad4x4 = vpx_sad4x4_sse2; + vpx_sad4x4_avg = vpx_sad4x4_avg_c; + if (flags & HAS_SSE2) vpx_sad4x4_avg = vpx_sad4x4_avg_sse2; + vpx_sad4x4x4d = vpx_sad4x4x4d_c; + if (flags & HAS_SSE2) vpx_sad4x4x4d = vpx_sad4x4x4d_sse2; + vpx_sad4x8 = vpx_sad4x8_c; + if (flags & HAS_SSE2) vpx_sad4x8 = vpx_sad4x8_sse2; + vpx_sad4x8_avg = vpx_sad4x8_avg_c; + if (flags & HAS_SSE2) vpx_sad4x8_avg = vpx_sad4x8_avg_sse2; + vpx_sad4x8x4d = vpx_sad4x8x4d_c; + if (flags & HAS_SSE2) vpx_sad4x8x4d = vpx_sad4x8x4d_sse2; + vpx_sad64x32 = vpx_sad64x32_c; + if (flags & HAS_SSE2) vpx_sad64x32 = vpx_sad64x32_sse2; + if (flags & HAS_AVX2) vpx_sad64x32 = vpx_sad64x32_avx2; + vpx_sad64x32_avg = vpx_sad64x32_avg_c; + if (flags & HAS_SSE2) vpx_sad64x32_avg = vpx_sad64x32_avg_sse2; + if (flags & HAS_AVX2) vpx_sad64x32_avg = vpx_sad64x32_avg_avx2; + vpx_sad64x32x4d = vpx_sad64x32x4d_c; + if (flags & HAS_SSE2) vpx_sad64x32x4d = vpx_sad64x32x4d_sse2; + vpx_sad64x64 = vpx_sad64x64_c; + if (flags & HAS_SSE2) vpx_sad64x64 = vpx_sad64x64_sse2; + if (flags & HAS_AVX2) vpx_sad64x64 = vpx_sad64x64_avx2; + vpx_sad64x64_avg = vpx_sad64x64_avg_c; + if (flags & HAS_SSE2) vpx_sad64x64_avg = vpx_sad64x64_avg_sse2; + if (flags & HAS_AVX2) vpx_sad64x64_avg = vpx_sad64x64_avg_avx2; + vpx_sad64x64x4d = vpx_sad64x64x4d_c; + if (flags & HAS_SSE2) vpx_sad64x64x4d = vpx_sad64x64x4d_sse2; + if (flags & HAS_AVX2) vpx_sad64x64x4d = vpx_sad64x64x4d_avx2; + vpx_sad8x16 = vpx_sad8x16_c; + if (flags & HAS_SSE2) vpx_sad8x16 = vpx_sad8x16_sse2; + vpx_sad8x16_avg = vpx_sad8x16_avg_c; + if (flags & HAS_SSE2) vpx_sad8x16_avg = vpx_sad8x16_avg_sse2; + vpx_sad8x16x4d = vpx_sad8x16x4d_c; + if (flags & HAS_SSE2) vpx_sad8x16x4d = vpx_sad8x16x4d_sse2; + vpx_sad8x4 = vpx_sad8x4_c; + if (flags & HAS_SSE2) vpx_sad8x4 = vpx_sad8x4_sse2; + vpx_sad8x4_avg = vpx_sad8x4_avg_c; + if (flags & HAS_SSE2) vpx_sad8x4_avg = vpx_sad8x4_avg_sse2; + vpx_sad8x4x4d = vpx_sad8x4x4d_c; + if (flags & HAS_SSE2) vpx_sad8x4x4d = vpx_sad8x4x4d_sse2; + vpx_sad8x8 = vpx_sad8x8_c; + if (flags & HAS_SSE2) vpx_sad8x8 = vpx_sad8x8_sse2; + vpx_sad8x8_avg = vpx_sad8x8_avg_c; + if (flags & HAS_SSE2) vpx_sad8x8_avg = vpx_sad8x8_avg_sse2; + vpx_sad8x8x4d = vpx_sad8x8x4d_c; + if (flags & HAS_SSE2) vpx_sad8x8x4d = vpx_sad8x8x4d_sse2; + vpx_sad_skip_16x16 = vpx_sad_skip_16x16_c; + if (flags & HAS_SSE2) vpx_sad_skip_16x16 = vpx_sad_skip_16x16_sse2; + vpx_sad_skip_16x16x4d = vpx_sad_skip_16x16x4d_c; + if (flags & HAS_SSE2) vpx_sad_skip_16x16x4d = vpx_sad_skip_16x16x4d_sse2; + vpx_sad_skip_16x32 = vpx_sad_skip_16x32_c; + if (flags & HAS_SSE2) vpx_sad_skip_16x32 = vpx_sad_skip_16x32_sse2; + vpx_sad_skip_16x32x4d = vpx_sad_skip_16x32x4d_c; + if (flags & HAS_SSE2) vpx_sad_skip_16x32x4d = vpx_sad_skip_16x32x4d_sse2; + vpx_sad_skip_16x8 = vpx_sad_skip_16x8_c; + if (flags & HAS_SSE2) vpx_sad_skip_16x8 = vpx_sad_skip_16x8_sse2; + vpx_sad_skip_16x8x4d = vpx_sad_skip_16x8x4d_c; + if (flags & HAS_SSE2) vpx_sad_skip_16x8x4d = vpx_sad_skip_16x8x4d_sse2; + vpx_sad_skip_32x16 = vpx_sad_skip_32x16_c; + if (flags & HAS_SSE2) vpx_sad_skip_32x16 = vpx_sad_skip_32x16_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_32x16 = vpx_sad_skip_32x16_avx2; + vpx_sad_skip_32x16x4d = vpx_sad_skip_32x16x4d_c; + if (flags & HAS_SSE2) vpx_sad_skip_32x16x4d = vpx_sad_skip_32x16x4d_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_32x16x4d = vpx_sad_skip_32x16x4d_avx2; + vpx_sad_skip_32x32 = vpx_sad_skip_32x32_c; + if (flags & HAS_SSE2) vpx_sad_skip_32x32 = vpx_sad_skip_32x32_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_32x32 = vpx_sad_skip_32x32_avx2; + vpx_sad_skip_32x32x4d = vpx_sad_skip_32x32x4d_c; + if (flags & HAS_SSE2) vpx_sad_skip_32x32x4d = vpx_sad_skip_32x32x4d_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_32x32x4d = vpx_sad_skip_32x32x4d_avx2; + vpx_sad_skip_32x64 = vpx_sad_skip_32x64_c; + if (flags & HAS_SSE2) vpx_sad_skip_32x64 = vpx_sad_skip_32x64_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_32x64 = vpx_sad_skip_32x64_avx2; + vpx_sad_skip_32x64x4d = vpx_sad_skip_32x64x4d_c; + if (flags & HAS_SSE2) vpx_sad_skip_32x64x4d = vpx_sad_skip_32x64x4d_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_32x64x4d = vpx_sad_skip_32x64x4d_avx2; + vpx_sad_skip_4x8 = vpx_sad_skip_4x8_c; + if (flags & HAS_SSE2) vpx_sad_skip_4x8 = vpx_sad_skip_4x8_sse2; + vpx_sad_skip_4x8x4d = vpx_sad_skip_4x8x4d_c; + if (flags & HAS_SSE2) vpx_sad_skip_4x8x4d = vpx_sad_skip_4x8x4d_sse2; + vpx_sad_skip_64x32 = vpx_sad_skip_64x32_c; + if (flags & HAS_SSE2) vpx_sad_skip_64x32 = vpx_sad_skip_64x32_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_64x32 = vpx_sad_skip_64x32_avx2; + vpx_sad_skip_64x32x4d = vpx_sad_skip_64x32x4d_c; + if (flags & HAS_SSE2) vpx_sad_skip_64x32x4d = vpx_sad_skip_64x32x4d_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_64x32x4d = vpx_sad_skip_64x32x4d_avx2; + vpx_sad_skip_64x64 = vpx_sad_skip_64x64_c; + if (flags & HAS_SSE2) vpx_sad_skip_64x64 = vpx_sad_skip_64x64_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_64x64 = vpx_sad_skip_64x64_avx2; + vpx_sad_skip_64x64x4d = vpx_sad_skip_64x64x4d_c; + if (flags & HAS_SSE2) vpx_sad_skip_64x64x4d = vpx_sad_skip_64x64x4d_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_64x64x4d = vpx_sad_skip_64x64x4d_avx2; + vpx_sad_skip_8x16 = vpx_sad_skip_8x16_c; + if (flags & HAS_SSE2) vpx_sad_skip_8x16 = vpx_sad_skip_8x16_sse2; + vpx_sad_skip_8x16x4d = vpx_sad_skip_8x16x4d_c; + if (flags & HAS_SSE2) vpx_sad_skip_8x16x4d = vpx_sad_skip_8x16x4d_sse2; + vpx_sad_skip_8x8 = vpx_sad_skip_8x8_c; + if (flags & HAS_SSE2) vpx_sad_skip_8x8 = vpx_sad_skip_8x8_sse2; + vpx_sad_skip_8x8x4d = vpx_sad_skip_8x8x4d_c; + if (flags & HAS_SSE2) vpx_sad_skip_8x8x4d = vpx_sad_skip_8x8x4d_sse2; + vpx_satd = vpx_satd_c; + if (flags & HAS_SSE2) vpx_satd = vpx_satd_sse2; + if (flags & HAS_AVX2) vpx_satd = vpx_satd_avx2; + vpx_scaled_2d = vpx_scaled_2d_c; + if (flags & HAS_SSSE3) vpx_scaled_2d = vpx_scaled_2d_ssse3; + vpx_sse = vpx_sse_c; + if (flags & HAS_SSE4_1) vpx_sse = vpx_sse_sse4_1; + if (flags & HAS_AVX2) vpx_sse = vpx_sse_avx2; + vpx_sub_pixel_avg_variance16x16 = vpx_sub_pixel_avg_variance16x16_c; + if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance16x16 = vpx_sub_pixel_avg_variance16x16_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance16x16 = vpx_sub_pixel_avg_variance16x16_ssse3; + vpx_sub_pixel_avg_variance16x32 = vpx_sub_pixel_avg_variance16x32_c; + if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance16x32 = vpx_sub_pixel_avg_variance16x32_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance16x32 = vpx_sub_pixel_avg_variance16x32_ssse3; + vpx_sub_pixel_avg_variance16x8 = vpx_sub_pixel_avg_variance16x8_c; + if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance16x8 = vpx_sub_pixel_avg_variance16x8_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance16x8 = vpx_sub_pixel_avg_variance16x8_ssse3; + vpx_sub_pixel_avg_variance32x16 = vpx_sub_pixel_avg_variance32x16_c; + if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance32x16 = vpx_sub_pixel_avg_variance32x16_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance32x16 = vpx_sub_pixel_avg_variance32x16_ssse3; + vpx_sub_pixel_avg_variance32x32 = vpx_sub_pixel_avg_variance32x32_c; + if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance32x32 = vpx_sub_pixel_avg_variance32x32_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance32x32 = vpx_sub_pixel_avg_variance32x32_ssse3; + if (flags & HAS_AVX2) vpx_sub_pixel_avg_variance32x32 = vpx_sub_pixel_avg_variance32x32_avx2; + vpx_sub_pixel_avg_variance32x64 = vpx_sub_pixel_avg_variance32x64_c; + if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance32x64 = vpx_sub_pixel_avg_variance32x64_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance32x64 = vpx_sub_pixel_avg_variance32x64_ssse3; + vpx_sub_pixel_avg_variance4x4 = vpx_sub_pixel_avg_variance4x4_c; + if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance4x4 = vpx_sub_pixel_avg_variance4x4_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance4x4 = vpx_sub_pixel_avg_variance4x4_ssse3; + vpx_sub_pixel_avg_variance4x8 = vpx_sub_pixel_avg_variance4x8_c; + if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance4x8 = vpx_sub_pixel_avg_variance4x8_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance4x8 = vpx_sub_pixel_avg_variance4x8_ssse3; + vpx_sub_pixel_avg_variance64x32 = vpx_sub_pixel_avg_variance64x32_c; + if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance64x32 = vpx_sub_pixel_avg_variance64x32_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance64x32 = vpx_sub_pixel_avg_variance64x32_ssse3; + vpx_sub_pixel_avg_variance64x64 = vpx_sub_pixel_avg_variance64x64_c; + if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance64x64 = vpx_sub_pixel_avg_variance64x64_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance64x64 = vpx_sub_pixel_avg_variance64x64_ssse3; + if (flags & HAS_AVX2) vpx_sub_pixel_avg_variance64x64 = vpx_sub_pixel_avg_variance64x64_avx2; + vpx_sub_pixel_avg_variance8x16 = vpx_sub_pixel_avg_variance8x16_c; + if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance8x16 = vpx_sub_pixel_avg_variance8x16_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance8x16 = vpx_sub_pixel_avg_variance8x16_ssse3; + vpx_sub_pixel_avg_variance8x4 = vpx_sub_pixel_avg_variance8x4_c; + if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance8x4 = vpx_sub_pixel_avg_variance8x4_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance8x4 = vpx_sub_pixel_avg_variance8x4_ssse3; + vpx_sub_pixel_avg_variance8x8 = vpx_sub_pixel_avg_variance8x8_c; + if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance8x8 = vpx_sub_pixel_avg_variance8x8_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance8x8 = vpx_sub_pixel_avg_variance8x8_ssse3; + vpx_sub_pixel_variance16x16 = vpx_sub_pixel_variance16x16_c; + if (flags & HAS_SSE2) vpx_sub_pixel_variance16x16 = vpx_sub_pixel_variance16x16_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance16x16 = vpx_sub_pixel_variance16x16_ssse3; + vpx_sub_pixel_variance16x32 = vpx_sub_pixel_variance16x32_c; + if (flags & HAS_SSE2) vpx_sub_pixel_variance16x32 = vpx_sub_pixel_variance16x32_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance16x32 = vpx_sub_pixel_variance16x32_ssse3; + vpx_sub_pixel_variance16x8 = vpx_sub_pixel_variance16x8_c; + if (flags & HAS_SSE2) vpx_sub_pixel_variance16x8 = vpx_sub_pixel_variance16x8_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance16x8 = vpx_sub_pixel_variance16x8_ssse3; + vpx_sub_pixel_variance32x16 = vpx_sub_pixel_variance32x16_c; + if (flags & HAS_SSE2) vpx_sub_pixel_variance32x16 = vpx_sub_pixel_variance32x16_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance32x16 = vpx_sub_pixel_variance32x16_ssse3; + vpx_sub_pixel_variance32x32 = vpx_sub_pixel_variance32x32_c; + if (flags & HAS_SSE2) vpx_sub_pixel_variance32x32 = vpx_sub_pixel_variance32x32_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance32x32 = vpx_sub_pixel_variance32x32_ssse3; + if (flags & HAS_AVX2) vpx_sub_pixel_variance32x32 = vpx_sub_pixel_variance32x32_avx2; + vpx_sub_pixel_variance32x64 = vpx_sub_pixel_variance32x64_c; + if (flags & HAS_SSE2) vpx_sub_pixel_variance32x64 = vpx_sub_pixel_variance32x64_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance32x64 = vpx_sub_pixel_variance32x64_ssse3; + vpx_sub_pixel_variance4x4 = vpx_sub_pixel_variance4x4_c; + if (flags & HAS_SSE2) vpx_sub_pixel_variance4x4 = vpx_sub_pixel_variance4x4_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance4x4 = vpx_sub_pixel_variance4x4_ssse3; + vpx_sub_pixel_variance4x8 = vpx_sub_pixel_variance4x8_c; + if (flags & HAS_SSE2) vpx_sub_pixel_variance4x8 = vpx_sub_pixel_variance4x8_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance4x8 = vpx_sub_pixel_variance4x8_ssse3; + vpx_sub_pixel_variance64x32 = vpx_sub_pixel_variance64x32_c; + if (flags & HAS_SSE2) vpx_sub_pixel_variance64x32 = vpx_sub_pixel_variance64x32_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance64x32 = vpx_sub_pixel_variance64x32_ssse3; + vpx_sub_pixel_variance64x64 = vpx_sub_pixel_variance64x64_c; + if (flags & HAS_SSE2) vpx_sub_pixel_variance64x64 = vpx_sub_pixel_variance64x64_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance64x64 = vpx_sub_pixel_variance64x64_ssse3; + if (flags & HAS_AVX2) vpx_sub_pixel_variance64x64 = vpx_sub_pixel_variance64x64_avx2; + vpx_sub_pixel_variance8x16 = vpx_sub_pixel_variance8x16_c; + if (flags & HAS_SSE2) vpx_sub_pixel_variance8x16 = vpx_sub_pixel_variance8x16_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance8x16 = vpx_sub_pixel_variance8x16_ssse3; + vpx_sub_pixel_variance8x4 = vpx_sub_pixel_variance8x4_c; + if (flags & HAS_SSE2) vpx_sub_pixel_variance8x4 = vpx_sub_pixel_variance8x4_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance8x4 = vpx_sub_pixel_variance8x4_ssse3; + vpx_sub_pixel_variance8x8 = vpx_sub_pixel_variance8x8_c; + if (flags & HAS_SSE2) vpx_sub_pixel_variance8x8 = vpx_sub_pixel_variance8x8_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance8x8 = vpx_sub_pixel_variance8x8_ssse3; + vpx_subtract_block = vpx_subtract_block_c; + if (flags & HAS_SSE2) vpx_subtract_block = vpx_subtract_block_sse2; + if (flags & HAS_AVX2) vpx_subtract_block = vpx_subtract_block_avx2; + vpx_sum_squares_2d_i16 = vpx_sum_squares_2d_i16_c; + if (flags & HAS_SSE2) vpx_sum_squares_2d_i16 = vpx_sum_squares_2d_i16_sse2; + vpx_tm_predictor_16x16 = vpx_tm_predictor_16x16_c; + if (flags & HAS_SSE2) vpx_tm_predictor_16x16 = vpx_tm_predictor_16x16_sse2; + vpx_tm_predictor_32x32 = vpx_tm_predictor_32x32_c; + if (flags & HAS_SSE2) vpx_tm_predictor_32x32 = vpx_tm_predictor_32x32_sse2; + vpx_tm_predictor_4x4 = vpx_tm_predictor_4x4_c; + if (flags & HAS_SSE2) vpx_tm_predictor_4x4 = vpx_tm_predictor_4x4_sse2; + vpx_tm_predictor_8x8 = vpx_tm_predictor_8x8_c; + if (flags & HAS_SSE2) vpx_tm_predictor_8x8 = vpx_tm_predictor_8x8_sse2; + vpx_v_predictor_16x16 = vpx_v_predictor_16x16_c; + if (flags & HAS_SSE2) vpx_v_predictor_16x16 = vpx_v_predictor_16x16_sse2; + vpx_v_predictor_32x32 = vpx_v_predictor_32x32_c; + if (flags & HAS_SSE2) vpx_v_predictor_32x32 = vpx_v_predictor_32x32_sse2; + vpx_v_predictor_4x4 = vpx_v_predictor_4x4_c; + if (flags & HAS_SSE2) vpx_v_predictor_4x4 = vpx_v_predictor_4x4_sse2; + vpx_v_predictor_8x8 = vpx_v_predictor_8x8_c; + if (flags & HAS_SSE2) vpx_v_predictor_8x8 = vpx_v_predictor_8x8_sse2; + vpx_variance16x16 = vpx_variance16x16_c; + if (flags & HAS_SSE2) vpx_variance16x16 = vpx_variance16x16_sse2; + if (flags & HAS_AVX2) vpx_variance16x16 = vpx_variance16x16_avx2; + vpx_variance16x32 = vpx_variance16x32_c; + if (flags & HAS_SSE2) vpx_variance16x32 = vpx_variance16x32_sse2; + if (flags & HAS_AVX2) vpx_variance16x32 = vpx_variance16x32_avx2; + vpx_variance16x8 = vpx_variance16x8_c; + if (flags & HAS_SSE2) vpx_variance16x8 = vpx_variance16x8_sse2; + if (flags & HAS_AVX2) vpx_variance16x8 = vpx_variance16x8_avx2; + vpx_variance32x16 = vpx_variance32x16_c; + if (flags & HAS_SSE2) vpx_variance32x16 = vpx_variance32x16_sse2; + if (flags & HAS_AVX2) vpx_variance32x16 = vpx_variance32x16_avx2; + vpx_variance32x32 = vpx_variance32x32_c; + if (flags & HAS_SSE2) vpx_variance32x32 = vpx_variance32x32_sse2; + if (flags & HAS_AVX2) vpx_variance32x32 = vpx_variance32x32_avx2; + vpx_variance32x64 = vpx_variance32x64_c; + if (flags & HAS_SSE2) vpx_variance32x64 = vpx_variance32x64_sse2; + if (flags & HAS_AVX2) vpx_variance32x64 = vpx_variance32x64_avx2; + vpx_variance4x4 = vpx_variance4x4_c; + if (flags & HAS_SSE2) vpx_variance4x4 = vpx_variance4x4_sse2; + vpx_variance4x8 = vpx_variance4x8_c; + if (flags & HAS_SSE2) vpx_variance4x8 = vpx_variance4x8_sse2; + vpx_variance64x32 = vpx_variance64x32_c; + if (flags & HAS_SSE2) vpx_variance64x32 = vpx_variance64x32_sse2; + if (flags & HAS_AVX2) vpx_variance64x32 = vpx_variance64x32_avx2; + vpx_variance64x64 = vpx_variance64x64_c; + if (flags & HAS_SSE2) vpx_variance64x64 = vpx_variance64x64_sse2; + if (flags & HAS_AVX2) vpx_variance64x64 = vpx_variance64x64_avx2; + vpx_variance8x16 = vpx_variance8x16_c; + if (flags & HAS_SSE2) vpx_variance8x16 = vpx_variance8x16_sse2; + if (flags & HAS_AVX2) vpx_variance8x16 = vpx_variance8x16_avx2; + vpx_variance8x4 = vpx_variance8x4_c; + if (flags & HAS_SSE2) vpx_variance8x4 = vpx_variance8x4_sse2; + if (flags & HAS_AVX2) vpx_variance8x4 = vpx_variance8x4_avx2; + vpx_variance8x8 = vpx_variance8x8_c; + if (flags & HAS_SSE2) vpx_variance8x8 = vpx_variance8x8_sse2; + if (flags & HAS_AVX2) vpx_variance8x8 = vpx_variance8x8_avx2; + vpx_vector_var = vpx_vector_var_c; + if (flags & HAS_SSE2) vpx_vector_var = vpx_vector_var_sse2; +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/ia32/vpx_scale_rtcd.h b/media/libvpx/config/linux/ia32/vpx_scale_rtcd.h new file mode 100644 index 0000000000..5f09104ea6 --- /dev/null +++ b/media/libvpx/config/linux/ia32/vpx_scale_rtcd.h @@ -0,0 +1,73 @@ +// This file is generated. Do not edit. +#ifndef VPX_SCALE_RTCD_H_ +#define VPX_SCALE_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +struct yv12_buffer_config; + +#ifdef __cplusplus +extern "C" { +#endif + +void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width); +#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c + +void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width); +#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c + +void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width); +#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c + +void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c + +void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c + +void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c + +void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c + +void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc); +#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c + +void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf); +#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c + +void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf); +#define vpx_extend_frame_borders vpx_extend_frame_borders_c + +void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf); +#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c + +void vpx_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc); +#define vpx_yv12_copy_frame vpx_yv12_copy_frame_c + +void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc); +#define vpx_yv12_copy_y vpx_yv12_copy_y_c + +void vpx_scale_rtcd(void); + +#ifdef RTCD_C +#include "vpx_ports/x86.h" +static void setup_rtcd_internal(void) +{ + int flags = x86_simd_caps(); + + (void)flags; + +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/x64/vp8_rtcd.h b/media/libvpx/config/linux/x64/vp8_rtcd.h new file mode 100644 index 0000000000..dc850b4fe0 --- /dev/null +++ b/media/libvpx/config/linux/x64/vp8_rtcd.h @@ -0,0 +1,248 @@ +// This file is generated. Do not edit. +#ifndef VP8_RTCD_H_ +#define VP8_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +/* + * VP8 + */ + +struct blockd; +struct macroblockd; +struct loop_filter_info; + +/* Encoder forward decls */ +struct block; +struct macroblock; +struct variance_vtable; +union int_mv; +struct yv12_buffer_config; + +#ifdef __cplusplus +extern "C" { +#endif + +void vp8_bilinear_predict16x16_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict16x16_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict16x16_ssse3(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_bilinear_predict16x16)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_bilinear_predict4x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict4x4_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_sse2 + +void vp8_bilinear_predict8x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict8x4_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_sse2 + +void vp8_bilinear_predict8x8_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict8x8_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_bilinear_predict8x8_ssse3(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_bilinear_predict8x8)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +int vp8_block_error_c(short *coeff, short *dqcoeff); +int vp8_block_error_sse2(short *coeff, short *dqcoeff); +#define vp8_block_error vp8_block_error_sse2 + +void vp8_copy32xn_c(const unsigned char *src_ptr, int src_stride, unsigned char *dst_ptr, int dst_stride, int height); +void vp8_copy32xn_sse2(const unsigned char *src_ptr, int src_stride, unsigned char *dst_ptr, int dst_stride, int height); +void vp8_copy32xn_sse3(const unsigned char *src_ptr, int src_stride, unsigned char *dst_ptr, int dst_stride, int height); +RTCD_EXTERN void (*vp8_copy32xn)(const unsigned char *src_ptr, int src_stride, unsigned char *dst_ptr, int dst_stride, int height); + +void vp8_copy_mem16x16_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +void vp8_copy_mem16x16_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +#define vp8_copy_mem16x16 vp8_copy_mem16x16_sse2 + +void vp8_copy_mem8x4_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +void vp8_copy_mem8x4_mmx(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +#define vp8_copy_mem8x4 vp8_copy_mem8x4_mmx + +void vp8_copy_mem8x8_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +void vp8_copy_mem8x8_mmx(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride); +#define vp8_copy_mem8x8 vp8_copy_mem8x8_mmx + +void vp8_dc_only_idct_add_c(short input_dc, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +void vp8_dc_only_idct_add_mmx(short input_dc, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +#define vp8_dc_only_idct_add vp8_dc_only_idct_add_mmx + +int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +#define vp8_denoiser_filter vp8_denoiser_filter_sse2 + +int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +int vp8_denoiser_filter_uv_sse2(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising); +#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_sse2 + +void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *dest, int stride); +void vp8_dequant_idct_add_mmx(short *input, short *dq, unsigned char *dest, int stride); +#define vp8_dequant_idct_add vp8_dequant_idct_add_mmx + +void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs); +void vp8_dequant_idct_add_uv_block_sse2(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs); +#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_sse2 + +void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs); +void vp8_dequant_idct_add_y_block_sse2(short *q, short *dq, unsigned char *dst, int stride, char *eobs); +#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_sse2 + +void vp8_dequantize_b_c(struct blockd*, short *DQC); +void vp8_dequantize_b_mmx(struct blockd*, short *DQC); +#define vp8_dequantize_b vp8_dequantize_b_mmx + +int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); +int vp8_diamond_search_sadx4(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); +#define vp8_diamond_search_sad vp8_diamond_search_sadx4 + +void vp8_fast_quantize_b_c(struct block *, struct blockd *); +void vp8_fast_quantize_b_sse2(struct block *, struct blockd *); +void vp8_fast_quantize_b_ssse3(struct block *, struct blockd *); +RTCD_EXTERN void (*vp8_fast_quantize_b)(struct block *, struct blockd *); + +void vp8_filter_by_weight16x16_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight); +void vp8_filter_by_weight16x16_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight); +#define vp8_filter_by_weight16x16 vp8_filter_by_weight16x16_sse2 + +void vp8_filter_by_weight4x4_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight); +#define vp8_filter_by_weight4x4 vp8_filter_by_weight4x4_c + +void vp8_filter_by_weight8x8_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight); +void vp8_filter_by_weight8x8_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight); +#define vp8_filter_by_weight8x8 vp8_filter_by_weight8x8_sse2 + +void vp8_loop_filter_bh_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_bh_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +#define vp8_loop_filter_bh vp8_loop_filter_bh_sse2 + +void vp8_loop_filter_bv_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_bv_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +#define vp8_loop_filter_bv vp8_loop_filter_bv_sse2 + +void vp8_loop_filter_mbh_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_mbh_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +#define vp8_loop_filter_mbh vp8_loop_filter_mbh_sse2 + +void vp8_loop_filter_mbv_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +void vp8_loop_filter_mbv_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi); +#define vp8_loop_filter_mbv vp8_loop_filter_mbv_sse2 + +void vp8_loop_filter_bhs_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_bhs_sse2(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_sse2 + +void vp8_loop_filter_bvs_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_bvs_sse2(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_sse2 + +void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_simple_horizontal_edge_sse2(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +#define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_sse2 + +void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +void vp8_loop_filter_simple_vertical_edge_sse2(unsigned char *y_ptr, int y_stride, const unsigned char *blimit); +#define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_sse2 + +int vp8_mbblock_error_c(struct macroblock *mb, int dc); +int vp8_mbblock_error_sse2(struct macroblock *mb, int dc); +#define vp8_mbblock_error vp8_mbblock_error_sse2 + +int vp8_mbuverror_c(struct macroblock *mb); +int vp8_mbuverror_sse2(struct macroblock *mb); +#define vp8_mbuverror vp8_mbuverror_sse2 + +int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int error_per_bit, int search_range, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); +int vp8_refining_search_sadx4(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int error_per_bit, int search_range, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv); +#define vp8_refining_search_sad vp8_refining_search_sadx4 + +void vp8_regular_quantize_b_c(struct block *, struct blockd *); +void vp8_regular_quantize_b_sse2(struct block *, struct blockd *); +void vp8_regular_quantize_b_sse4_1(struct block *, struct blockd *); +RTCD_EXTERN void (*vp8_regular_quantize_b)(struct block *, struct blockd *); + +void vp8_short_fdct4x4_c(short *input, short *output, int pitch); +void vp8_short_fdct4x4_sse2(short *input, short *output, int pitch); +#define vp8_short_fdct4x4 vp8_short_fdct4x4_sse2 + +void vp8_short_fdct8x4_c(short *input, short *output, int pitch); +void vp8_short_fdct8x4_sse2(short *input, short *output, int pitch); +#define vp8_short_fdct8x4 vp8_short_fdct8x4_sse2 + +void vp8_short_idct4x4llm_c(short *input, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +void vp8_short_idct4x4llm_mmx(short *input, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride); +#define vp8_short_idct4x4llm vp8_short_idct4x4llm_mmx + +void vp8_short_inv_walsh4x4_c(short *input, short *mb_dqcoeff); +void vp8_short_inv_walsh4x4_sse2(short *input, short *mb_dqcoeff); +#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_sse2 + +void vp8_short_inv_walsh4x4_1_c(short *input, short *mb_dqcoeff); +#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c + +void vp8_short_walsh4x4_c(short *input, short *output, int pitch); +void vp8_short_walsh4x4_sse2(short *input, short *output, int pitch); +#define vp8_short_walsh4x4 vp8_short_walsh4x4_sse2 + +void vp8_sixtap_predict16x16_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict16x16_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict16x16_ssse3(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_sixtap_predict16x16)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_sixtap_predict4x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict4x4_mmx(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict4x4_ssse3(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_sixtap_predict4x4)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_sixtap_predict8x4_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict8x4_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict8x4_ssse3(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_sixtap_predict8x4)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_sixtap_predict8x8_c(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict8x8_sse2(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +void vp8_sixtap_predict8x8_ssse3(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); +RTCD_EXTERN void (*vp8_sixtap_predict8x8)(unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch); + +void vp8_temporal_filter_apply_c(unsigned char *frame1, unsigned int stride, unsigned char *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, unsigned short *count); +void vp8_temporal_filter_apply_sse2(unsigned char *frame1, unsigned int stride, unsigned char *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, unsigned short *count); +#define vp8_temporal_filter_apply vp8_temporal_filter_apply_sse2 + +void vp8_rtcd(void); + +#ifdef RTCD_C +#include "vpx_ports/x86.h" +static void setup_rtcd_internal(void) +{ + int flags = x86_simd_caps(); + + (void)flags; + + vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_sse2; + if (flags & HAS_SSSE3) vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_ssse3; + vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_sse2; + if (flags & HAS_SSSE3) vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_ssse3; + vp8_copy32xn = vp8_copy32xn_sse2; + if (flags & HAS_SSE3) vp8_copy32xn = vp8_copy32xn_sse3; + vp8_fast_quantize_b = vp8_fast_quantize_b_sse2; + if (flags & HAS_SSSE3) vp8_fast_quantize_b = vp8_fast_quantize_b_ssse3; + vp8_regular_quantize_b = vp8_regular_quantize_b_sse2; + if (flags & HAS_SSE4_1) vp8_regular_quantize_b = vp8_regular_quantize_b_sse4_1; + vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_sse2; + if (flags & HAS_SSSE3) vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_ssse3; + vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_mmx; + if (flags & HAS_SSSE3) vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_ssse3; + vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_sse2; + if (flags & HAS_SSSE3) vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_ssse3; + vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_sse2; + if (flags & HAS_SSSE3) vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_ssse3; +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/x64/vp9_rtcd.h b/media/libvpx/config/linux/x64/vp9_rtcd.h new file mode 100644 index 0000000000..8644c3598d --- /dev/null +++ b/media/libvpx/config/linux/x64/vp9_rtcd.h @@ -0,0 +1,135 @@ +// This file is generated. Do not edit. +#ifndef VP9_RTCD_H_ +#define VP9_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +/* + * VP9 + */ + +#include "vpx/vpx_integer.h" +#include "vp9/common/vp9_common.h" +#include "vp9/common/vp9_enums.h" +#include "vp9/common/vp9_filter.h" + +struct macroblockd; + +/* Encoder forward decls */ +struct macroblock; +struct macroblock_plane; +struct vp9_sad_table; +struct ScanOrder; +struct search_site_config; +struct mv; +union int_mv; +struct yv12_buffer_config; + +#ifdef __cplusplus +extern "C" { +#endif + +void vp9_apply_temporal_filter_c(const uint8_t *y_src, int y_src_stride, const uint8_t *y_pre, int y_pre_stride, const uint8_t *u_src, const uint8_t *v_src, int uv_src_stride, const uint8_t *u_pre, const uint8_t *v_pre, int uv_pre_stride, unsigned int block_width, unsigned int block_height, int ss_x, int ss_y, int strength, const int *const blk_fw, int use_32x32, uint32_t *y_accumulator, uint16_t *y_count, uint32_t *u_accumulator, uint16_t *u_count, uint32_t *v_accumulator, uint16_t *v_count); +void vp9_apply_temporal_filter_sse4_1(const uint8_t *y_src, int y_src_stride, const uint8_t *y_pre, int y_pre_stride, const uint8_t *u_src, const uint8_t *v_src, int uv_src_stride, const uint8_t *u_pre, const uint8_t *v_pre, int uv_pre_stride, unsigned int block_width, unsigned int block_height, int ss_x, int ss_y, int strength, const int *const blk_fw, int use_32x32, uint32_t *y_accumulator, uint16_t *y_count, uint32_t *u_accumulator, uint16_t *u_count, uint32_t *v_accumulator, uint16_t *v_count); +RTCD_EXTERN void (*vp9_apply_temporal_filter)(const uint8_t *y_src, int y_src_stride, const uint8_t *y_pre, int y_pre_stride, const uint8_t *u_src, const uint8_t *v_src, int uv_src_stride, const uint8_t *u_pre, const uint8_t *v_pre, int uv_pre_stride, unsigned int block_width, unsigned int block_height, int ss_x, int ss_y, int strength, const int *const blk_fw, int use_32x32, uint32_t *y_accumulator, uint16_t *y_count, uint32_t *u_accumulator, uint16_t *u_count, uint32_t *v_accumulator, uint16_t *v_count); + +int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz); +int64_t vp9_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz); +int64_t vp9_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz); +RTCD_EXTERN int64_t (*vp9_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz); + +int64_t vp9_block_error_fp_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size); +int64_t vp9_block_error_fp_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size); +int64_t vp9_block_error_fp_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size); +RTCD_EXTERN int64_t (*vp9_block_error_fp)(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size); + +int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, uint32_t start_mv_sad, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_sad_table *sad_fn_ptr, const struct mv *center_mv); +#define vp9_diamond_search_sad vp9_diamond_search_sad_c + +void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type); +void vp9_fht16x16_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type); +#define vp9_fht16x16 vp9_fht16x16_sse2 + +void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type); +void vp9_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type); +#define vp9_fht4x4 vp9_fht4x4_sse2 + +void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type); +void vp9_fht8x8_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type); +#define vp9_fht8x8 vp9_fht8x8_sse2 + +void vp9_filter_by_weight16x16_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight); +void vp9_filter_by_weight16x16_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight); +#define vp9_filter_by_weight16x16 vp9_filter_by_weight16x16_sse2 + +void vp9_filter_by_weight8x8_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight); +void vp9_filter_by_weight8x8_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight); +#define vp9_filter_by_weight8x8 vp9_filter_by_weight8x8_sse2 + +void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride); +void vp9_fwht4x4_sse2(const int16_t *input, tran_low_t *output, int stride); +#define vp9_fwht4x4 vp9_fwht4x4_sse2 + +void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +void vp9_iht16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +#define vp9_iht16x16_256_add vp9_iht16x16_256_add_sse2 + +void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +void vp9_iht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +#define vp9_iht4x4_16_add vp9_iht4x4_16_add_sse2 + +void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +void vp9_iht8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int stride, int tx_type); +#define vp9_iht8x8_64_add vp9_iht8x8_64_add_sse2 + +void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +RTCD_EXTERN void (*vp9_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); + +void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_32x32_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vp9_quantize_fp_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +RTCD_EXTERN void (*vp9_quantize_fp_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); + +void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst, INTERP_FILTER filter_type, int phase_scaler); +void vp9_scale_and_extend_frame_ssse3(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst, INTERP_FILTER filter_type, int phase_scaler); +RTCD_EXTERN void (*vp9_scale_and_extend_frame)(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst, INTERP_FILTER filter_type, int phase_scaler); + +void vp9_rtcd(void); + +#ifdef RTCD_C +#include "vpx_ports/x86.h" +static void setup_rtcd_internal(void) +{ + int flags = x86_simd_caps(); + + (void)flags; + + vp9_apply_temporal_filter = vp9_apply_temporal_filter_c; + if (flags & HAS_SSE4_1) vp9_apply_temporal_filter = vp9_apply_temporal_filter_sse4_1; + vp9_block_error = vp9_block_error_sse2; + if (flags & HAS_AVX2) vp9_block_error = vp9_block_error_avx2; + vp9_block_error_fp = vp9_block_error_fp_sse2; + if (flags & HAS_AVX2) vp9_block_error_fp = vp9_block_error_fp_avx2; + vp9_quantize_fp = vp9_quantize_fp_sse2; + if (flags & HAS_SSSE3) vp9_quantize_fp = vp9_quantize_fp_ssse3; + if (flags & HAS_AVX2) vp9_quantize_fp = vp9_quantize_fp_avx2; + vp9_quantize_fp_32x32 = vp9_quantize_fp_32x32_c; + if (flags & HAS_SSSE3) vp9_quantize_fp_32x32 = vp9_quantize_fp_32x32_ssse3; + if (flags & HAS_AVX2) vp9_quantize_fp_32x32 = vp9_quantize_fp_32x32_avx2; + vp9_scale_and_extend_frame = vp9_scale_and_extend_frame_c; + if (flags & HAS_SSSE3) vp9_scale_and_extend_frame = vp9_scale_and_extend_frame_ssse3; +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/x64/vpx_config.asm b/media/libvpx/config/linux/x64/vpx_config.asm new file mode 100644 index 0000000000..8715768a2e --- /dev/null +++ b/media/libvpx/config/linux/x64/vpx_config.asm @@ -0,0 +1,93 @@ +%define VPX_ARCH_ARM 0 +%define VPX_ARCH_AARCH64 0 +%define VPX_ARCH_MIPS 0 +%define VPX_ARCH_X86 0 +%define VPX_ARCH_X86_64 1 +%define VPX_ARCH_PPC 0 +%define VPX_ARCH_LOONGARCH 0 +%define HAVE_NEON_ASM 0 +%define HAVE_NEON 0 +%define HAVE_NEON_DOTPROD 0 +%define HAVE_NEON_I8MM 0 +%define HAVE_SVE 0 +%define HAVE_MIPS32 0 +%define HAVE_DSPR2 0 +%define HAVE_MSA 0 +%define HAVE_MIPS64 0 +%define HAVE_MMX 1 +%define HAVE_SSE 1 +%define HAVE_SSE2 1 +%define HAVE_SSE3 1 +%define HAVE_SSSE3 1 +%define HAVE_SSE4_1 1 +%define HAVE_AVX 1 +%define HAVE_AVX2 1 +%define HAVE_AVX512 0 +%define HAVE_VSX 0 +%define HAVE_MMI 0 +%define HAVE_LSX 0 +%define HAVE_LASX 0 +%define HAVE_VPX_PORTS 1 +%define HAVE_PTHREAD_H 1 +%define CONFIG_DEPENDENCY_TRACKING 1 +%define CONFIG_EXTERNAL_BUILD 1 +%define CONFIG_INSTALL_DOCS 0 +%define CONFIG_INSTALL_BINS 1 +%define CONFIG_INSTALL_LIBS 1 +%define CONFIG_INSTALL_SRCS 0 +%define CONFIG_DEBUG 0 +%define CONFIG_GPROF 0 +%define CONFIG_GCOV 0 +%define CONFIG_RVCT 0 +%define CONFIG_GCC 1 +%define CONFIG_MSVS 0 +%define CONFIG_PIC 1 +%define CONFIG_BIG_ENDIAN 0 +%define CONFIG_CODEC_SRCS 0 +%define CONFIG_DEBUG_LIBS 0 +%define CONFIG_DEQUANT_TOKENS 0 +%define CONFIG_DC_RECON 0 +%define CONFIG_RUNTIME_CPU_DETECT 1 +%define CONFIG_POSTPROC 1 +%define CONFIG_VP9_POSTPROC 1 +%define CONFIG_MULTITHREAD 1 +%define CONFIG_INTERNAL_STATS 0 +%define CONFIG_VP8_ENCODER 1 +%define CONFIG_VP8_DECODER 1 +%define CONFIG_VP9_ENCODER 1 +%define CONFIG_VP9_DECODER 1 +%define CONFIG_VP8 1 +%define CONFIG_VP9 1 +%define CONFIG_ENCODERS 1 +%define CONFIG_DECODERS 1 +%define CONFIG_STATIC_MSVCRT 0 +%define CONFIG_SPATIAL_RESAMPLING 1 +%define CONFIG_REALTIME_ONLY 0 +%define CONFIG_ONTHEFLY_BITPACKING 0 +%define CONFIG_ERROR_CONCEALMENT 0 +%define CONFIG_SHARED 0 +%define CONFIG_STATIC 1 +%define CONFIG_SMALL 0 +%define CONFIG_POSTPROC_VISUALIZER 0 +%define CONFIG_OS_SUPPORT 1 +%define CONFIG_UNIT_TESTS 0 +%define CONFIG_WEBM_IO 1 +%define CONFIG_LIBYUV 1 +%define CONFIG_DECODE_PERF_TESTS 0 +%define CONFIG_ENCODE_PERF_TESTS 0 +%define CONFIG_MULTI_RES_ENCODING 1 +%define CONFIG_TEMPORAL_DENOISING 1 +%define CONFIG_VP9_TEMPORAL_DENOISING 0 +%define CONFIG_COEFFICIENT_RANGE_CHECKING 0 +%define CONFIG_VP9_HIGHBITDEPTH 0 +%define CONFIG_BETTER_HW_COMPATIBILITY 0 +%define CONFIG_EXPERIMENTAL 0 +%define CONFIG_SIZE_LIMIT 1 +%define CONFIG_ALWAYS_ADJUST_BPM 0 +%define CONFIG_BITSTREAM_DEBUG 0 +%define CONFIG_MISMATCH_DEBUG 0 +%define CONFIG_FP_MB_STATS 0 +%define CONFIG_EMULATE_HARDWARE 0 +%define CONFIG_NON_GREEDY_MV 0 +%define CONFIG_RATE_CTRL 0 +%define CONFIG_COLLECT_COMPONENT_TIMING 0 diff --git a/media/libvpx/config/linux/x64/vpx_config.c b/media/libvpx/config/linux/x64/vpx_config.c new file mode 100644 index 0000000000..e4dcb394c3 --- /dev/null +++ b/media/libvpx/config/linux/x64/vpx_config.c @@ -0,0 +1,10 @@ +/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */ +/* */ +/* Use of this source code is governed by a BSD-style license */ +/* that can be found in the LICENSE file in the root of the source */ +/* tree. An additional intellectual property rights grant can be found */ +/* in the file PATENTS. All contributing project authors may */ +/* be found in the AUTHORS file in the root of the source tree. */ +#include "vpx/vpx_codec.h" +static const char* const cfg = "--target=x86_64-linux-gcc --enable-external-build --disable-examples --disable-install-docs --disable-unit-tests --enable-multi-res-encoding --size-limit=8192x4608 --enable-pic --disable-avx512 --enable-postproc --enable-vp9-postproc --as=yasm"; +const char *vpx_codec_build_config(void) {return cfg;} diff --git a/media/libvpx/config/linux/x64/vpx_config.h b/media/libvpx/config/linux/x64/vpx_config.h new file mode 100644 index 0000000000..ab4439aaf4 --- /dev/null +++ b/media/libvpx/config/linux/x64/vpx_config.h @@ -0,0 +1,108 @@ +/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */ +/* */ +/* Use of this source code is governed by a BSD-style license */ +/* that can be found in the LICENSE file in the root of the source */ +/* tree. An additional intellectual property rights grant can be found */ +/* in the file PATENTS. All contributing project authors may */ +/* be found in the AUTHORS file in the root of the source tree. */ +/* This file automatically generated by configure. Do not edit! */ +#ifndef VPX_CONFIG_H +#define VPX_CONFIG_H +#define RESTRICT +#define INLINE inline +#define VPX_ARCH_ARM 0 +#define VPX_ARCH_AARCH64 0 +#define VPX_ARCH_MIPS 0 +#define VPX_ARCH_X86 0 +#define VPX_ARCH_X86_64 1 +#define VPX_ARCH_PPC 0 +#define VPX_ARCH_LOONGARCH 0 +#define HAVE_NEON_ASM 0 +#define HAVE_NEON 0 +#define HAVE_NEON_DOTPROD 0 +#define HAVE_NEON_I8MM 0 +#define HAVE_SVE 0 +#define HAVE_MIPS32 0 +#define HAVE_DSPR2 0 +#define HAVE_MSA 0 +#define HAVE_MIPS64 0 +#define HAVE_MMX 1 +#define HAVE_SSE 1 +#define HAVE_SSE2 1 +#define HAVE_SSE3 1 +#define HAVE_SSSE3 1 +#define HAVE_SSE4_1 1 +#define HAVE_AVX 1 +#define HAVE_AVX2 1 +#define HAVE_AVX512 0 +#define HAVE_VSX 0 +#define HAVE_MMI 0 +#define HAVE_LSX 0 +#define HAVE_LASX 0 +#define HAVE_VPX_PORTS 1 +#define HAVE_PTHREAD_H 1 +#define CONFIG_DEPENDENCY_TRACKING 1 +#define CONFIG_EXTERNAL_BUILD 1 +#define CONFIG_INSTALL_DOCS 0 +#define CONFIG_INSTALL_BINS 1 +#define CONFIG_INSTALL_LIBS 1 +#define CONFIG_INSTALL_SRCS 0 +#define CONFIG_DEBUG 0 +#define CONFIG_GPROF 0 +#define CONFIG_GCOV 0 +#define CONFIG_RVCT 0 +#define CONFIG_GCC 1 +#define CONFIG_MSVS 0 +#define CONFIG_PIC 1 +#define CONFIG_BIG_ENDIAN 0 +#define CONFIG_CODEC_SRCS 0 +#define CONFIG_DEBUG_LIBS 0 +#define CONFIG_DEQUANT_TOKENS 0 +#define CONFIG_DC_RECON 0 +#define CONFIG_RUNTIME_CPU_DETECT 1 +#define CONFIG_POSTPROC 1 +#define CONFIG_VP9_POSTPROC 1 +#define CONFIG_MULTITHREAD 1 +#define CONFIG_INTERNAL_STATS 0 +#define CONFIG_VP8_ENCODER 1 +#define CONFIG_VP8_DECODER 1 +#define CONFIG_VP9_ENCODER 1 +#define CONFIG_VP9_DECODER 1 +#define CONFIG_VP8 1 +#define CONFIG_VP9 1 +#define CONFIG_ENCODERS 1 +#define CONFIG_DECODERS 1 +#define CONFIG_STATIC_MSVCRT 0 +#define CONFIG_SPATIAL_RESAMPLING 1 +#define CONFIG_REALTIME_ONLY 0 +#define CONFIG_ONTHEFLY_BITPACKING 0 +#define CONFIG_ERROR_CONCEALMENT 0 +#define CONFIG_SHARED 0 +#define CONFIG_STATIC 1 +#define CONFIG_SMALL 0 +#define CONFIG_POSTPROC_VISUALIZER 0 +#define CONFIG_OS_SUPPORT 1 +#define CONFIG_UNIT_TESTS 0 +#define CONFIG_WEBM_IO 1 +#define CONFIG_LIBYUV 1 +#define CONFIG_DECODE_PERF_TESTS 0 +#define CONFIG_ENCODE_PERF_TESTS 0 +#define CONFIG_MULTI_RES_ENCODING 1 +#define CONFIG_TEMPORAL_DENOISING 1 +#define CONFIG_VP9_TEMPORAL_DENOISING 0 +#define CONFIG_COEFFICIENT_RANGE_CHECKING 0 +#define CONFIG_VP9_HIGHBITDEPTH 0 +#define CONFIG_BETTER_HW_COMPATIBILITY 0 +#define CONFIG_EXPERIMENTAL 0 +#define CONFIG_SIZE_LIMIT 1 +#define CONFIG_ALWAYS_ADJUST_BPM 0 +#define CONFIG_BITSTREAM_DEBUG 0 +#define CONFIG_MISMATCH_DEBUG 0 +#define CONFIG_FP_MB_STATS 0 +#define CONFIG_EMULATE_HARDWARE 0 +#define CONFIG_NON_GREEDY_MV 0 +#define CONFIG_RATE_CTRL 0 +#define CONFIG_COLLECT_COMPONENT_TIMING 0 +#define DECODE_WIDTH_LIMIT 8192 +#define DECODE_HEIGHT_LIMIT 4608 +#endif /* VPX_CONFIG_H */ diff --git a/media/libvpx/config/linux/x64/vpx_dsp_rtcd.h b/media/libvpx/config/linux/x64/vpx_dsp_rtcd.h new file mode 100644 index 0000000000..755e916dd1 --- /dev/null +++ b/media/libvpx/config/linux/x64/vpx_dsp_rtcd.h @@ -0,0 +1,1296 @@ +// This file is generated. Do not edit. +#ifndef VPX_DSP_RTCD_H_ +#define VPX_DSP_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +/* + * DSP + */ + +#include "vpx/vpx_integer.h" +#include "vpx_dsp/vpx_dsp_common.h" +#include "vpx_dsp/vpx_filter.h" +#if CONFIG_VP9_ENCODER + struct macroblock_plane; + struct ScanOrder; +#endif + + +#ifdef __cplusplus +extern "C" { +#endif + +unsigned int vpx_avg_4x4_c(const uint8_t *, int p); +unsigned int vpx_avg_4x4_sse2(const uint8_t *, int p); +#define vpx_avg_4x4 vpx_avg_4x4_sse2 + +unsigned int vpx_avg_8x8_c(const uint8_t *, int p); +unsigned int vpx_avg_8x8_sse2(const uint8_t *, int p); +#define vpx_avg_8x8 vpx_avg_8x8_sse2 + +void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride); +void vpx_comp_avg_pred_sse2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride); +void vpx_comp_avg_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride); +RTCD_EXTERN void (*vpx_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride); + +void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_avg)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_avg_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_avg_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_avg_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_convolve_avg vpx_convolve_avg_sse2 + +void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_convolve_copy vpx_convolve_copy_sse2 + +void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c + +void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c + +void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c + +void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c + +void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_c + +void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_c + +void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_c + +void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_c + +void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d153_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d153_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d153_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d153_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d153_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d207_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d207_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_sse2 + +void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d207_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d207_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d45_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d45_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d45_predictor_4x4 vpx_d45_predictor_4x4_sse2 + +void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d45_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d45_predictor_8x8 vpx_d45_predictor_8x8_sse2 + +void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c + +void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d63_predictor_16x16)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d63_predictor_32x32)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d63_predictor_4x4)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_d63_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*vpx_d63_predictor_8x8)(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); + +void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c + +void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_128_predictor_16x16 vpx_dc_128_predictor_16x16_sse2 + +void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_128_predictor_32x32 vpx_dc_128_predictor_32x32_sse2 + +void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_128_predictor_4x4 vpx_dc_128_predictor_4x4_sse2 + +void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_128_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_128_predictor_8x8 vpx_dc_128_predictor_8x8_sse2 + +void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_left_predictor_16x16 vpx_dc_left_predictor_16x16_sse2 + +void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_left_predictor_32x32 vpx_dc_left_predictor_32x32_sse2 + +void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_left_predictor_4x4 vpx_dc_left_predictor_4x4_sse2 + +void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_left_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_left_predictor_8x8 vpx_dc_left_predictor_8x8_sse2 + +void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_predictor_16x16 vpx_dc_predictor_16x16_sse2 + +void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_predictor_32x32 vpx_dc_predictor_32x32_sse2 + +void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_predictor_4x4 vpx_dc_predictor_4x4_sse2 + +void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_predictor_8x8 vpx_dc_predictor_8x8_sse2 + +void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_top_predictor_16x16 vpx_dc_top_predictor_16x16_sse2 + +void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_top_predictor_32x32 vpx_dc_top_predictor_32x32_sse2 + +void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_top_predictor_4x4 vpx_dc_top_predictor_4x4_sse2 + +void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_dc_top_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_dc_top_predictor_8x8 vpx_dc_top_predictor_8x8_sse2 + +void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct16x16_sse2(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct16x16_avx2(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct16x16)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct16x16_1 vpx_fdct16x16_1_sse2 + +void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_sse2(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_avx2(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct32x32)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_1_sse2(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct32x32_1 vpx_fdct32x32_1_sse2 + +void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_rd_sse2(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct32x32_rd_avx2(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct32x32_rd)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct4x4 vpx_fdct4x4_sse2 + +void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct4x4_1_sse2(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct4x4_1 vpx_fdct4x4_1_sse2 + +void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct8x8_sse2(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct8x8_ssse3(const int16_t *input, tran_low_t *output, int stride); +RTCD_EXTERN void (*vpx_fdct8x8)(const int16_t *input, tran_low_t *output, int stride); + +void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride); +void vpx_fdct8x8_1_sse2(const int16_t *input, tran_low_t *output, int stride); +#define vpx_fdct8x8_1 vpx_fdct8x8_1_sse2 + +void vpx_get16x16var_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +void vpx_get16x16var_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +void vpx_get16x16var_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +RTCD_EXTERN void (*vpx_get16x16var)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); + +unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride); +#define vpx_get4x4sse_cs vpx_get4x4sse_cs_c + +void vpx_get8x8var_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +void vpx_get8x8var_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum); +#define vpx_get8x8var vpx_get8x8var_sse2 + +unsigned int vpx_get_mb_ss_c(const int16_t *); +unsigned int vpx_get_mb_ss_sse2(const int16_t *); +#define vpx_get_mb_ss vpx_get_mb_ss_sse2 + +void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_h_predictor_16x16 vpx_h_predictor_16x16_sse2 + +void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_h_predictor_32x32 vpx_h_predictor_32x32_sse2 + +void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_h_predictor_4x4 vpx_h_predictor_4x4_sse2 + +void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_h_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_h_predictor_8x8 vpx_h_predictor_8x8_sse2 + +void vpx_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +RTCD_EXTERN void (*vpx_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); + +void vpx_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_32x32_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +RTCD_EXTERN void (*vpx_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); + +void vpx_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +void vpx_hadamard_8x8_ssse3(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); +RTCD_EXTERN void (*vpx_hadamard_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff); + +void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c + +void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct16x16_10_add vpx_idct16x16_10_add_sse2 + +void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct16x16_1_add vpx_idct16x16_1_add_sse2 + +void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_256_add_avx2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct16x16_256_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct16x16_38_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct16x16_38_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct16x16_38_add vpx_idct16x16_38_add_sse2 + +void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_1024_add_avx2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct32x32_1024_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_135_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_135_add_ssse3(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_135_add_avx2(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct32x32_135_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct32x32_1_add vpx_idct32x32_1_add_sse2 + +void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct32x32_34_add_ssse3(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct32x32_34_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct4x4_16_add vpx_idct4x4_16_add_sse2 + +void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct4x4_1_add vpx_idct4x4_1_add_sse2 + +void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_12_add_ssse3(const tran_low_t *input, uint8_t *dest, int stride); +RTCD_EXTERN void (*vpx_idct8x8_12_add)(const tran_low_t *input, uint8_t *dest, int stride); + +void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct8x8_1_add vpx_idct8x8_1_add_sse2 + +void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_idct8x8_64_add vpx_idct8x8_64_add_sse2 + +int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width); +int16_t vpx_int_pro_col_sse2(const uint8_t *ref, const int width); +#define vpx_int_pro_col vpx_int_pro_col_sse2 + +void vpx_int_pro_row_c(int16_t hbuf[16], const uint8_t *ref, const int ref_stride, const int height); +void vpx_int_pro_row_sse2(int16_t hbuf[16], const uint8_t *ref, const int ref_stride, const int height); +#define vpx_int_pro_row vpx_int_pro_row_sse2 + +void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride); +void vpx_iwht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_sse2 + +void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int stride); +#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c + +void vpx_lpf_horizontal_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_16_avx2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_horizontal_16)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_horizontal_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_16_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_16_dual_avx2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +RTCD_EXTERN void (*vpx_lpf_horizontal_16_dual)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); + +void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_sse2 + +void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +#define vpx_lpf_horizontal_4_dual vpx_lpf_horizontal_4_dual_sse2 + +void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_horizontal_8 vpx_lpf_horizontal_8_sse2 + +void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +#define vpx_lpf_horizontal_8_dual vpx_lpf_horizontal_8_dual_sse2 + +void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_vertical_16 vpx_lpf_vertical_16_sse2 + +void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_16_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_vertical_16_dual vpx_lpf_vertical_16_dual_sse2 + +void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_vertical_4 vpx_lpf_vertical_4_sse2 + +void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +#define vpx_lpf_vertical_4_dual vpx_lpf_vertical_4_dual_sse2 + +void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +void vpx_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); +#define vpx_lpf_vertical_8 vpx_lpf_vertical_8_sse2 + +void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void vpx_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +#define vpx_lpf_vertical_8_dual vpx_lpf_vertical_8_dual_sse2 + +void vpx_mbpost_proc_across_ip_c(unsigned char *src, int pitch, int rows, int cols,int flimit); +void vpx_mbpost_proc_across_ip_sse2(unsigned char *src, int pitch, int rows, int cols,int flimit); +#define vpx_mbpost_proc_across_ip vpx_mbpost_proc_across_ip_sse2 + +void vpx_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int cols,int flimit); +void vpx_mbpost_proc_down_sse2(unsigned char *dst, int pitch, int rows, int cols,int flimit); +#define vpx_mbpost_proc_down vpx_mbpost_proc_down_sse2 + +void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max); +void vpx_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max); +#define vpx_minmax_8x8 vpx_minmax_8x8_sse2 + +unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_mse16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +#define vpx_mse8x16 vpx_mse8x16_sse2 + +unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_mse8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +#define vpx_mse8x8 vpx_mse8x8_sse2 + +void vpx_plane_add_noise_c(uint8_t *start, const int8_t *noise, int blackclamp, int whiteclamp, int width, int height, int pitch); +void vpx_plane_add_noise_sse2(uint8_t *start, const int8_t *noise, int blackclamp, int whiteclamp, int width, int height, int pitch); +#define vpx_plane_add_noise vpx_plane_add_noise_sse2 + +void vpx_post_proc_down_and_across_mb_row_c(unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size); +void vpx_post_proc_down_and_across_mb_row_sse2(unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size); +#define vpx_post_proc_down_and_across_mb_row vpx_post_proc_down_and_across_mb_row_sse2 + +void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +RTCD_EXTERN void (*vpx_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); + +void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_32x32_ssse3(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_32x32_avx(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +void vpx_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); +RTCD_EXTERN void (*vpx_quantize_b_32x32)(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order); + +unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad16x16 vpx_sad16x16_sse2 + +unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +#define vpx_sad16x16_avg vpx_sad16x16_avg_sse2 + +void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad16x16x4d vpx_sad16x16x4d_sse2 + +unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad16x32 vpx_sad16x32_sse2 + +unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +#define vpx_sad16x32_avg vpx_sad16x32_avg_sse2 + +void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad16x32x4d vpx_sad16x32x4d_sse2 + +unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad16x8 vpx_sad16x8_sse2 + +unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +#define vpx_sad16x8_avg vpx_sad16x8_avg_sse2 + +void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad16x8x4d vpx_sad16x8x4d_sse2 + +unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad32x16x4d vpx_sad32x16x4d_sse2 + +unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad32x64x4d vpx_sad32x64x4d_sse2 + +unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad4x4 vpx_sad4x4_sse2 + +unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +#define vpx_sad4x4_avg vpx_sad4x4_avg_sse2 + +void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad4x4x4d vpx_sad4x4x4d_sse2 + +unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad4x8 vpx_sad4x8_sse2 + +unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +#define vpx_sad4x8_avg vpx_sad4x8_avg_sse2 + +void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad4x8x4d vpx_sad4x8x4d_sse2 + +unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad64x32x4d vpx_sad64x32x4d_sse2 + +unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +RTCD_EXTERN unsigned int (*vpx_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); + +void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad8x16 vpx_sad8x16_sse2 + +unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +#define vpx_sad8x16_avg vpx_sad8x16_avg_sse2 + +void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad8x16x4d vpx_sad8x16x4d_sse2 + +unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad8x4 vpx_sad8x4_sse2 + +unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +#define vpx_sad8x4_avg vpx_sad8x4_avg_sse2 + +void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad8x4x4d vpx_sad8x4x4d_sse2 + +unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad8x8 vpx_sad8x8_sse2 + +unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +unsigned int vpx_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred); +#define vpx_sad8x8_avg vpx_sad8x8_avg_sse2 + +void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad8x8x4d vpx_sad8x8x4d_sse2 + +unsigned int vpx_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_16x16 vpx_sad_skip_16x16_sse2 + +void vpx_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_16x16x4d vpx_sad_skip_16x16x4d_sse2 + +unsigned int vpx_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_16x32 vpx_sad_skip_16x32_sse2 + +void vpx_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_16x32x4d vpx_sad_skip_16x32x4d_sse2 + +unsigned int vpx_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_16x8 vpx_sad_skip_16x8_sse2 + +void vpx_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_16x8x4d vpx_sad_skip_16x8x4d_sse2 + +unsigned int vpx_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_4x4 vpx_sad_skip_4x4_c + +void vpx_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_4x4x4d vpx_sad_skip_4x4x4d_c + +unsigned int vpx_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_4x8 vpx_sad_skip_4x8_sse2 + +void vpx_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_4x8x4d vpx_sad_skip_4x8x4d_sse2 + +unsigned int vpx_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +RTCD_EXTERN unsigned int (*vpx_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); + +void vpx_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +RTCD_EXTERN void (*vpx_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); + +unsigned int vpx_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_8x16 vpx_sad_skip_8x16_sse2 + +void vpx_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_8x16x4d vpx_sad_skip_8x16x4d_sse2 + +unsigned int vpx_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_8x4 vpx_sad_skip_8x4_c + +void vpx_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_8x4x4d vpx_sad_skip_8x4x4d_c + +unsigned int vpx_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +unsigned int vpx_sad_skip_8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride); +#define vpx_sad_skip_8x8 vpx_sad_skip_8x8_sse2 + +void vpx_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +void vpx_sad_skip_8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]); +#define vpx_sad_skip_8x8x4d vpx_sad_skip_8x8x4d_sse2 + +int vpx_satd_c(const int16_t *coeff, int length); +int vpx_satd_sse2(const int16_t *coeff, int length); +int vpx_satd_avx2(const int16_t *coeff, int length); +RTCD_EXTERN int (*vpx_satd)(const int16_t *coeff, int length); + +void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void vpx_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*vpx_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + +void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c + +void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c + +void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c + +void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_horiz vpx_scaled_horiz_c + +void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define vpx_scaled_vert vpx_scaled_vert_c + +int64_t vpx_sse_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); +int64_t vpx_sse_sse4_1(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); +int64_t vpx_sse_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); +RTCD_EXTERN int64_t (*vpx_sse)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, int width, int height); + +uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +uint32_t vpx_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred); + +uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance16x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance16x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance16x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance32x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance32x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance32x64)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance4x4)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance4x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance64x32)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance64x64)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance8x16)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance8x4)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +uint32_t vpx_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); +RTCD_EXTERN uint32_t (*vpx_sub_pixel_variance8x8)(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse); + +void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride); +void vpx_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride); +void vpx_subtract_block_avx2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride); +RTCD_EXTERN void (*vpx_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride); + +uint64_t vpx_sum_squares_2d_i16_c(const int16_t *src, int stride, int size); +uint64_t vpx_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int size); +#define vpx_sum_squares_2d_i16 vpx_sum_squares_2d_i16_sse2 + +void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_tm_predictor_16x16 vpx_tm_predictor_16x16_sse2 + +void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_tm_predictor_32x32 vpx_tm_predictor_32x32_sse2 + +void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_tm_predictor_4x4 vpx_tm_predictor_4x4_sse2 + +void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_tm_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_tm_predictor_8x8 vpx_tm_predictor_8x8_sse2 + +void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_v_predictor_16x16 vpx_v_predictor_16x16_sse2 + +void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_v_predictor_32x32 vpx_v_predictor_32x32_sse2 + +void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_v_predictor_4x4 vpx_v_predictor_4x4_sse2 + +void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +void vpx_v_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_v_predictor_8x8 vpx_v_predictor_8x8_sse2 + +unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +#define vpx_variance4x4 vpx_variance4x4_sse2 + +unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +#define vpx_variance4x8 vpx_variance4x8_sse2 + +unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x4_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +unsigned int vpx_variance8x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); +RTCD_EXTERN unsigned int (*vpx_variance8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse); + +void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left); +#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c + +int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl); +int vpx_vector_var_sse2(const int16_t *ref, const int16_t *src, const int bwl); +#define vpx_vector_var vpx_vector_var_sse2 + +void vpx_dsp_rtcd(void); + +#ifdef RTCD_C +#include "vpx_ports/x86.h" +static void setup_rtcd_internal(void) +{ + int flags = x86_simd_caps(); + + (void)flags; + + vpx_comp_avg_pred = vpx_comp_avg_pred_sse2; + if (flags & HAS_AVX2) vpx_comp_avg_pred = vpx_comp_avg_pred_avx2; + vpx_convolve8 = vpx_convolve8_sse2; + if (flags & HAS_SSSE3) vpx_convolve8 = vpx_convolve8_ssse3; + if (flags & HAS_AVX2) vpx_convolve8 = vpx_convolve8_avx2; + vpx_convolve8_avg = vpx_convolve8_avg_sse2; + if (flags & HAS_SSSE3) vpx_convolve8_avg = vpx_convolve8_avg_ssse3; + if (flags & HAS_AVX2) vpx_convolve8_avg = vpx_convolve8_avg_avx2; + vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_sse2; + if (flags & HAS_SSSE3) vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_ssse3; + if (flags & HAS_AVX2) vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_avx2; + vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_sse2; + if (flags & HAS_SSSE3) vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_ssse3; + if (flags & HAS_AVX2) vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_avx2; + vpx_convolve8_horiz = vpx_convolve8_horiz_sse2; + if (flags & HAS_SSSE3) vpx_convolve8_horiz = vpx_convolve8_horiz_ssse3; + if (flags & HAS_AVX2) vpx_convolve8_horiz = vpx_convolve8_horiz_avx2; + vpx_convolve8_vert = vpx_convolve8_vert_sse2; + if (flags & HAS_SSSE3) vpx_convolve8_vert = vpx_convolve8_vert_ssse3; + if (flags & HAS_AVX2) vpx_convolve8_vert = vpx_convolve8_vert_avx2; + vpx_d153_predictor_16x16 = vpx_d153_predictor_16x16_c; + if (flags & HAS_SSSE3) vpx_d153_predictor_16x16 = vpx_d153_predictor_16x16_ssse3; + vpx_d153_predictor_32x32 = vpx_d153_predictor_32x32_c; + if (flags & HAS_SSSE3) vpx_d153_predictor_32x32 = vpx_d153_predictor_32x32_ssse3; + vpx_d153_predictor_4x4 = vpx_d153_predictor_4x4_c; + if (flags & HAS_SSSE3) vpx_d153_predictor_4x4 = vpx_d153_predictor_4x4_ssse3; + vpx_d153_predictor_8x8 = vpx_d153_predictor_8x8_c; + if (flags & HAS_SSSE3) vpx_d153_predictor_8x8 = vpx_d153_predictor_8x8_ssse3; + vpx_d207_predictor_16x16 = vpx_d207_predictor_16x16_c; + if (flags & HAS_SSSE3) vpx_d207_predictor_16x16 = vpx_d207_predictor_16x16_ssse3; + vpx_d207_predictor_32x32 = vpx_d207_predictor_32x32_c; + if (flags & HAS_SSSE3) vpx_d207_predictor_32x32 = vpx_d207_predictor_32x32_ssse3; + vpx_d207_predictor_8x8 = vpx_d207_predictor_8x8_c; + if (flags & HAS_SSSE3) vpx_d207_predictor_8x8 = vpx_d207_predictor_8x8_ssse3; + vpx_d45_predictor_16x16 = vpx_d45_predictor_16x16_c; + if (flags & HAS_SSSE3) vpx_d45_predictor_16x16 = vpx_d45_predictor_16x16_ssse3; + vpx_d45_predictor_32x32 = vpx_d45_predictor_32x32_c; + if (flags & HAS_SSSE3) vpx_d45_predictor_32x32 = vpx_d45_predictor_32x32_ssse3; + vpx_d63_predictor_16x16 = vpx_d63_predictor_16x16_c; + if (flags & HAS_SSSE3) vpx_d63_predictor_16x16 = vpx_d63_predictor_16x16_ssse3; + vpx_d63_predictor_32x32 = vpx_d63_predictor_32x32_c; + if (flags & HAS_SSSE3) vpx_d63_predictor_32x32 = vpx_d63_predictor_32x32_ssse3; + vpx_d63_predictor_4x4 = vpx_d63_predictor_4x4_c; + if (flags & HAS_SSSE3) vpx_d63_predictor_4x4 = vpx_d63_predictor_4x4_ssse3; + vpx_d63_predictor_8x8 = vpx_d63_predictor_8x8_c; + if (flags & HAS_SSSE3) vpx_d63_predictor_8x8 = vpx_d63_predictor_8x8_ssse3; + vpx_fdct16x16 = vpx_fdct16x16_sse2; + if (flags & HAS_AVX2) vpx_fdct16x16 = vpx_fdct16x16_avx2; + vpx_fdct32x32 = vpx_fdct32x32_sse2; + if (flags & HAS_AVX2) vpx_fdct32x32 = vpx_fdct32x32_avx2; + vpx_fdct32x32_rd = vpx_fdct32x32_rd_sse2; + if (flags & HAS_AVX2) vpx_fdct32x32_rd = vpx_fdct32x32_rd_avx2; + vpx_fdct8x8 = vpx_fdct8x8_sse2; + if (flags & HAS_SSSE3) vpx_fdct8x8 = vpx_fdct8x8_ssse3; + vpx_get16x16var = vpx_get16x16var_sse2; + if (flags & HAS_AVX2) vpx_get16x16var = vpx_get16x16var_avx2; + vpx_hadamard_16x16 = vpx_hadamard_16x16_sse2; + if (flags & HAS_AVX2) vpx_hadamard_16x16 = vpx_hadamard_16x16_avx2; + vpx_hadamard_32x32 = vpx_hadamard_32x32_sse2; + if (flags & HAS_AVX2) vpx_hadamard_32x32 = vpx_hadamard_32x32_avx2; + vpx_hadamard_8x8 = vpx_hadamard_8x8_sse2; + if (flags & HAS_SSSE3) vpx_hadamard_8x8 = vpx_hadamard_8x8_ssse3; + vpx_idct16x16_256_add = vpx_idct16x16_256_add_sse2; + if (flags & HAS_AVX2) vpx_idct16x16_256_add = vpx_idct16x16_256_add_avx2; + vpx_idct32x32_1024_add = vpx_idct32x32_1024_add_sse2; + if (flags & HAS_AVX2) vpx_idct32x32_1024_add = vpx_idct32x32_1024_add_avx2; + vpx_idct32x32_135_add = vpx_idct32x32_135_add_sse2; + if (flags & HAS_SSSE3) vpx_idct32x32_135_add = vpx_idct32x32_135_add_ssse3; + if (flags & HAS_AVX2) vpx_idct32x32_135_add = vpx_idct32x32_135_add_avx2; + vpx_idct32x32_34_add = vpx_idct32x32_34_add_sse2; + if (flags & HAS_SSSE3) vpx_idct32x32_34_add = vpx_idct32x32_34_add_ssse3; + vpx_idct8x8_12_add = vpx_idct8x8_12_add_sse2; + if (flags & HAS_SSSE3) vpx_idct8x8_12_add = vpx_idct8x8_12_add_ssse3; + vpx_lpf_horizontal_16 = vpx_lpf_horizontal_16_sse2; + if (flags & HAS_AVX2) vpx_lpf_horizontal_16 = vpx_lpf_horizontal_16_avx2; + vpx_lpf_horizontal_16_dual = vpx_lpf_horizontal_16_dual_sse2; + if (flags & HAS_AVX2) vpx_lpf_horizontal_16_dual = vpx_lpf_horizontal_16_dual_avx2; + vpx_mse16x16 = vpx_mse16x16_sse2; + if (flags & HAS_AVX2) vpx_mse16x16 = vpx_mse16x16_avx2; + vpx_mse16x8 = vpx_mse16x8_sse2; + if (flags & HAS_AVX2) vpx_mse16x8 = vpx_mse16x8_avx2; + vpx_quantize_b = vpx_quantize_b_sse2; + if (flags & HAS_SSSE3) vpx_quantize_b = vpx_quantize_b_ssse3; + if (flags & HAS_AVX) vpx_quantize_b = vpx_quantize_b_avx; + if (flags & HAS_AVX2) vpx_quantize_b = vpx_quantize_b_avx2; + vpx_quantize_b_32x32 = vpx_quantize_b_32x32_c; + if (flags & HAS_SSSE3) vpx_quantize_b_32x32 = vpx_quantize_b_32x32_ssse3; + if (flags & HAS_AVX) vpx_quantize_b_32x32 = vpx_quantize_b_32x32_avx; + if (flags & HAS_AVX2) vpx_quantize_b_32x32 = vpx_quantize_b_32x32_avx2; + vpx_sad32x16 = vpx_sad32x16_sse2; + if (flags & HAS_AVX2) vpx_sad32x16 = vpx_sad32x16_avx2; + vpx_sad32x16_avg = vpx_sad32x16_avg_sse2; + if (flags & HAS_AVX2) vpx_sad32x16_avg = vpx_sad32x16_avg_avx2; + vpx_sad32x32 = vpx_sad32x32_sse2; + if (flags & HAS_AVX2) vpx_sad32x32 = vpx_sad32x32_avx2; + vpx_sad32x32_avg = vpx_sad32x32_avg_sse2; + if (flags & HAS_AVX2) vpx_sad32x32_avg = vpx_sad32x32_avg_avx2; + vpx_sad32x32x4d = vpx_sad32x32x4d_sse2; + if (flags & HAS_AVX2) vpx_sad32x32x4d = vpx_sad32x32x4d_avx2; + vpx_sad32x64 = vpx_sad32x64_sse2; + if (flags & HAS_AVX2) vpx_sad32x64 = vpx_sad32x64_avx2; + vpx_sad32x64_avg = vpx_sad32x64_avg_sse2; + if (flags & HAS_AVX2) vpx_sad32x64_avg = vpx_sad32x64_avg_avx2; + vpx_sad64x32 = vpx_sad64x32_sse2; + if (flags & HAS_AVX2) vpx_sad64x32 = vpx_sad64x32_avx2; + vpx_sad64x32_avg = vpx_sad64x32_avg_sse2; + if (flags & HAS_AVX2) vpx_sad64x32_avg = vpx_sad64x32_avg_avx2; + vpx_sad64x64 = vpx_sad64x64_sse2; + if (flags & HAS_AVX2) vpx_sad64x64 = vpx_sad64x64_avx2; + vpx_sad64x64_avg = vpx_sad64x64_avg_sse2; + if (flags & HAS_AVX2) vpx_sad64x64_avg = vpx_sad64x64_avg_avx2; + vpx_sad64x64x4d = vpx_sad64x64x4d_sse2; + if (flags & HAS_AVX2) vpx_sad64x64x4d = vpx_sad64x64x4d_avx2; + vpx_sad_skip_32x16 = vpx_sad_skip_32x16_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_32x16 = vpx_sad_skip_32x16_avx2; + vpx_sad_skip_32x16x4d = vpx_sad_skip_32x16x4d_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_32x16x4d = vpx_sad_skip_32x16x4d_avx2; + vpx_sad_skip_32x32 = vpx_sad_skip_32x32_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_32x32 = vpx_sad_skip_32x32_avx2; + vpx_sad_skip_32x32x4d = vpx_sad_skip_32x32x4d_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_32x32x4d = vpx_sad_skip_32x32x4d_avx2; + vpx_sad_skip_32x64 = vpx_sad_skip_32x64_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_32x64 = vpx_sad_skip_32x64_avx2; + vpx_sad_skip_32x64x4d = vpx_sad_skip_32x64x4d_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_32x64x4d = vpx_sad_skip_32x64x4d_avx2; + vpx_sad_skip_64x32 = vpx_sad_skip_64x32_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_64x32 = vpx_sad_skip_64x32_avx2; + vpx_sad_skip_64x32x4d = vpx_sad_skip_64x32x4d_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_64x32x4d = vpx_sad_skip_64x32x4d_avx2; + vpx_sad_skip_64x64 = vpx_sad_skip_64x64_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_64x64 = vpx_sad_skip_64x64_avx2; + vpx_sad_skip_64x64x4d = vpx_sad_skip_64x64x4d_sse2; + if (flags & HAS_AVX2) vpx_sad_skip_64x64x4d = vpx_sad_skip_64x64x4d_avx2; + vpx_satd = vpx_satd_sse2; + if (flags & HAS_AVX2) vpx_satd = vpx_satd_avx2; + vpx_scaled_2d = vpx_scaled_2d_c; + if (flags & HAS_SSSE3) vpx_scaled_2d = vpx_scaled_2d_ssse3; + vpx_sse = vpx_sse_c; + if (flags & HAS_SSE4_1) vpx_sse = vpx_sse_sse4_1; + if (flags & HAS_AVX2) vpx_sse = vpx_sse_avx2; + vpx_sub_pixel_avg_variance16x16 = vpx_sub_pixel_avg_variance16x16_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance16x16 = vpx_sub_pixel_avg_variance16x16_ssse3; + vpx_sub_pixel_avg_variance16x32 = vpx_sub_pixel_avg_variance16x32_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance16x32 = vpx_sub_pixel_avg_variance16x32_ssse3; + vpx_sub_pixel_avg_variance16x8 = vpx_sub_pixel_avg_variance16x8_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance16x8 = vpx_sub_pixel_avg_variance16x8_ssse3; + vpx_sub_pixel_avg_variance32x16 = vpx_sub_pixel_avg_variance32x16_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance32x16 = vpx_sub_pixel_avg_variance32x16_ssse3; + vpx_sub_pixel_avg_variance32x32 = vpx_sub_pixel_avg_variance32x32_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance32x32 = vpx_sub_pixel_avg_variance32x32_ssse3; + if (flags & HAS_AVX2) vpx_sub_pixel_avg_variance32x32 = vpx_sub_pixel_avg_variance32x32_avx2; + vpx_sub_pixel_avg_variance32x64 = vpx_sub_pixel_avg_variance32x64_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance32x64 = vpx_sub_pixel_avg_variance32x64_ssse3; + vpx_sub_pixel_avg_variance4x4 = vpx_sub_pixel_avg_variance4x4_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance4x4 = vpx_sub_pixel_avg_variance4x4_ssse3; + vpx_sub_pixel_avg_variance4x8 = vpx_sub_pixel_avg_variance4x8_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance4x8 = vpx_sub_pixel_avg_variance4x8_ssse3; + vpx_sub_pixel_avg_variance64x32 = vpx_sub_pixel_avg_variance64x32_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance64x32 = vpx_sub_pixel_avg_variance64x32_ssse3; + vpx_sub_pixel_avg_variance64x64 = vpx_sub_pixel_avg_variance64x64_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance64x64 = vpx_sub_pixel_avg_variance64x64_ssse3; + if (flags & HAS_AVX2) vpx_sub_pixel_avg_variance64x64 = vpx_sub_pixel_avg_variance64x64_avx2; + vpx_sub_pixel_avg_variance8x16 = vpx_sub_pixel_avg_variance8x16_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance8x16 = vpx_sub_pixel_avg_variance8x16_ssse3; + vpx_sub_pixel_avg_variance8x4 = vpx_sub_pixel_avg_variance8x4_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance8x4 = vpx_sub_pixel_avg_variance8x4_ssse3; + vpx_sub_pixel_avg_variance8x8 = vpx_sub_pixel_avg_variance8x8_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance8x8 = vpx_sub_pixel_avg_variance8x8_ssse3; + vpx_sub_pixel_variance16x16 = vpx_sub_pixel_variance16x16_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance16x16 = vpx_sub_pixel_variance16x16_ssse3; + vpx_sub_pixel_variance16x32 = vpx_sub_pixel_variance16x32_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance16x32 = vpx_sub_pixel_variance16x32_ssse3; + vpx_sub_pixel_variance16x8 = vpx_sub_pixel_variance16x8_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance16x8 = vpx_sub_pixel_variance16x8_ssse3; + vpx_sub_pixel_variance32x16 = vpx_sub_pixel_variance32x16_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance32x16 = vpx_sub_pixel_variance32x16_ssse3; + vpx_sub_pixel_variance32x32 = vpx_sub_pixel_variance32x32_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance32x32 = vpx_sub_pixel_variance32x32_ssse3; + if (flags & HAS_AVX2) vpx_sub_pixel_variance32x32 = vpx_sub_pixel_variance32x32_avx2; + vpx_sub_pixel_variance32x64 = vpx_sub_pixel_variance32x64_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance32x64 = vpx_sub_pixel_variance32x64_ssse3; + vpx_sub_pixel_variance4x4 = vpx_sub_pixel_variance4x4_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance4x4 = vpx_sub_pixel_variance4x4_ssse3; + vpx_sub_pixel_variance4x8 = vpx_sub_pixel_variance4x8_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance4x8 = vpx_sub_pixel_variance4x8_ssse3; + vpx_sub_pixel_variance64x32 = vpx_sub_pixel_variance64x32_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance64x32 = vpx_sub_pixel_variance64x32_ssse3; + vpx_sub_pixel_variance64x64 = vpx_sub_pixel_variance64x64_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance64x64 = vpx_sub_pixel_variance64x64_ssse3; + if (flags & HAS_AVX2) vpx_sub_pixel_variance64x64 = vpx_sub_pixel_variance64x64_avx2; + vpx_sub_pixel_variance8x16 = vpx_sub_pixel_variance8x16_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance8x16 = vpx_sub_pixel_variance8x16_ssse3; + vpx_sub_pixel_variance8x4 = vpx_sub_pixel_variance8x4_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance8x4 = vpx_sub_pixel_variance8x4_ssse3; + vpx_sub_pixel_variance8x8 = vpx_sub_pixel_variance8x8_sse2; + if (flags & HAS_SSSE3) vpx_sub_pixel_variance8x8 = vpx_sub_pixel_variance8x8_ssse3; + vpx_subtract_block = vpx_subtract_block_sse2; + if (flags & HAS_AVX2) vpx_subtract_block = vpx_subtract_block_avx2; + vpx_variance16x16 = vpx_variance16x16_sse2; + if (flags & HAS_AVX2) vpx_variance16x16 = vpx_variance16x16_avx2; + vpx_variance16x32 = vpx_variance16x32_sse2; + if (flags & HAS_AVX2) vpx_variance16x32 = vpx_variance16x32_avx2; + vpx_variance16x8 = vpx_variance16x8_sse2; + if (flags & HAS_AVX2) vpx_variance16x8 = vpx_variance16x8_avx2; + vpx_variance32x16 = vpx_variance32x16_sse2; + if (flags & HAS_AVX2) vpx_variance32x16 = vpx_variance32x16_avx2; + vpx_variance32x32 = vpx_variance32x32_sse2; + if (flags & HAS_AVX2) vpx_variance32x32 = vpx_variance32x32_avx2; + vpx_variance32x64 = vpx_variance32x64_sse2; + if (flags & HAS_AVX2) vpx_variance32x64 = vpx_variance32x64_avx2; + vpx_variance64x32 = vpx_variance64x32_sse2; + if (flags & HAS_AVX2) vpx_variance64x32 = vpx_variance64x32_avx2; + vpx_variance64x64 = vpx_variance64x64_sse2; + if (flags & HAS_AVX2) vpx_variance64x64 = vpx_variance64x64_avx2; + vpx_variance8x16 = vpx_variance8x16_sse2; + if (flags & HAS_AVX2) vpx_variance8x16 = vpx_variance8x16_avx2; + vpx_variance8x4 = vpx_variance8x4_sse2; + if (flags & HAS_AVX2) vpx_variance8x4 = vpx_variance8x4_avx2; + vpx_variance8x8 = vpx_variance8x8_sse2; + if (flags & HAS_AVX2) vpx_variance8x8 = vpx_variance8x8_avx2; +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif diff --git a/media/libvpx/config/linux/x64/vpx_scale_rtcd.h b/media/libvpx/config/linux/x64/vpx_scale_rtcd.h new file mode 100644 index 0000000000..5f09104ea6 --- /dev/null +++ b/media/libvpx/config/linux/x64/vpx_scale_rtcd.h @@ -0,0 +1,73 @@ +// This file is generated. Do not edit. +#ifndef VPX_SCALE_RTCD_H_ +#define VPX_SCALE_RTCD_H_ + +#ifdef RTCD_C +#define RTCD_EXTERN +#else +#define RTCD_EXTERN extern +#endif + +struct yv12_buffer_config; + +#ifdef __cplusplus +extern "C" { +#endif + +void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width); +#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c + +void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width); +#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c + +void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width); +#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c + +void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c + +void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c + +void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c + +void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width); +#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c + +void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc); +#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c + +void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf); +#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c + +void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf); +#define vpx_extend_frame_borders vpx_extend_frame_borders_c + +void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf); +#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c + +void vpx_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc); +#define vpx_yv12_copy_frame vpx_yv12_copy_frame_c + +void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc); +#define vpx_yv12_copy_y vpx_yv12_copy_y_c + +void vpx_scale_rtcd(void); + +#ifdef RTCD_C +#include "vpx_ports/x86.h" +static void setup_rtcd_internal(void) +{ + int flags = x86_simd_caps(); + + (void)flags; + +} +#endif + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif |