diff options
author | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-05-15 03:35:49 +0000 |
---|---|---|
committer | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-05-15 03:35:49 +0000 |
commit | d8bbc7858622b6d9c278469aab701ca0b609cddf (patch) | |
tree | eff41dc61d9f714852212739e6b3738b82a2af87 /third_party/aom/av1/common/arm | |
parent | Releasing progress-linux version 125.0.3-1~progress7.99u1. (diff) | |
download | firefox-d8bbc7858622b6d9c278469aab701ca0b609cddf.tar.xz firefox-d8bbc7858622b6d9c278469aab701ca0b609cddf.zip |
Merging upstream version 126.0.
Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>
Diffstat (limited to 'third_party/aom/av1/common/arm')
12 files changed, 4070 insertions, 425 deletions
diff --git a/third_party/aom/av1/common/arm/highbd_compound_convolve_neon.c b/third_party/aom/av1/common/arm/highbd_compound_convolve_neon.c index fc03a2ee04..9247ded6bf 100644 --- a/third_party/aom/av1/common/arm/highbd_compound_convolve_neon.c +++ b/third_party/aom/av1/common/arm/highbd_compound_convolve_neon.c @@ -20,266 +20,9 @@ #include "aom_ports/mem.h" #include "av1/common/convolve.h" #include "av1/common/filter.h" +#include "av1/common/arm/highbd_compound_convolve_neon.h" #include "av1/common/arm/highbd_convolve_neon.h" -#define ROUND_SHIFT 2 * FILTER_BITS - ROUND0_BITS - COMPOUND_ROUND1_BITS - -static INLINE void highbd_12_comp_avg_neon(const uint16_t *src_ptr, - int src_stride, uint16_t *dst_ptr, - int dst_stride, int w, int h, - ConvolveParams *conv_params, - const int offset, const int bd) { - CONV_BUF_TYPE *ref_ptr = conv_params->dst; - const int ref_stride = conv_params->dst_stride; - const uint16x4_t offset_vec = vdup_n_u16(offset); - const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); - - if (w == 4) { - do { - const uint16x4_t src = vld1_u16(src_ptr); - const uint16x4_t ref = vld1_u16(ref_ptr); - - uint16x4_t avg = vhadd_u16(src, ref); - int32x4_t d0 = vreinterpretq_s32_u32(vsubl_u16(avg, offset_vec)); - - uint16x4_t d0_u16 = vqrshrun_n_s32(d0, ROUND_SHIFT - 2); - d0_u16 = vmin_u16(d0_u16, vget_low_u16(max)); - - vst1_u16(dst_ptr, d0_u16); - - src_ptr += src_stride; - ref_ptr += ref_stride; - dst_ptr += dst_stride; - } while (--h != 0); - } else { - do { - int width = w; - const uint16_t *src = src_ptr; - const uint16_t *ref = ref_ptr; - uint16_t *dst = dst_ptr; - do { - const uint16x8_t s = vld1q_u16(src); - const uint16x8_t r = vld1q_u16(ref); - - uint16x8_t avg = vhaddq_u16(s, r); - int32x4_t d0_lo = - vreinterpretq_s32_u32(vsubl_u16(vget_low_u16(avg), offset_vec)); - int32x4_t d0_hi = - vreinterpretq_s32_u32(vsubl_u16(vget_high_u16(avg), offset_vec)); - - uint16x8_t d0 = vcombine_u16(vqrshrun_n_s32(d0_lo, ROUND_SHIFT - 2), - vqrshrun_n_s32(d0_hi, ROUND_SHIFT - 2)); - d0 = vminq_u16(d0, max); - vst1q_u16(dst, d0); - - src += 8; - ref += 8; - dst += 8; - width -= 8; - } while (width != 0); - - src_ptr += src_stride; - ref_ptr += ref_stride; - dst_ptr += dst_stride; - } while (--h != 0); - } -} - -static INLINE void highbd_comp_avg_neon(const uint16_t *src_ptr, int src_stride, - uint16_t *dst_ptr, int dst_stride, - int w, int h, - ConvolveParams *conv_params, - const int offset, const int bd) { - CONV_BUF_TYPE *ref_ptr = conv_params->dst; - const int ref_stride = conv_params->dst_stride; - const uint16x4_t offset_vec = vdup_n_u16(offset); - const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); - - if (w == 4) { - do { - const uint16x4_t src = vld1_u16(src_ptr); - const uint16x4_t ref = vld1_u16(ref_ptr); - - uint16x4_t avg = vhadd_u16(src, ref); - int32x4_t d0 = vreinterpretq_s32_u32(vsubl_u16(avg, offset_vec)); - - uint16x4_t d0_u16 = vqrshrun_n_s32(d0, ROUND_SHIFT); - d0_u16 = vmin_u16(d0_u16, vget_low_u16(max)); - - vst1_u16(dst_ptr, d0_u16); - - src_ptr += src_stride; - ref_ptr += ref_stride; - dst_ptr += dst_stride; - } while (--h != 0); - } else { - do { - int width = w; - const uint16_t *src = src_ptr; - const uint16_t *ref = ref_ptr; - uint16_t *dst = dst_ptr; - do { - const uint16x8_t s = vld1q_u16(src); - const uint16x8_t r = vld1q_u16(ref); - - uint16x8_t avg = vhaddq_u16(s, r); - int32x4_t d0_lo = - vreinterpretq_s32_u32(vsubl_u16(vget_low_u16(avg), offset_vec)); - int32x4_t d0_hi = - vreinterpretq_s32_u32(vsubl_u16(vget_high_u16(avg), offset_vec)); - - uint16x8_t d0 = vcombine_u16(vqrshrun_n_s32(d0_lo, ROUND_SHIFT), - vqrshrun_n_s32(d0_hi, ROUND_SHIFT)); - d0 = vminq_u16(d0, max); - vst1q_u16(dst, d0); - - src += 8; - ref += 8; - dst += 8; - width -= 8; - } while (width != 0); - - src_ptr += src_stride; - ref_ptr += ref_stride; - dst_ptr += dst_stride; - } while (--h != 0); - } -} - -static INLINE void highbd_12_dist_wtd_comp_avg_neon( - const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, - int w, int h, ConvolveParams *conv_params, const int offset, const int bd) { - CONV_BUF_TYPE *ref_ptr = conv_params->dst; - const int ref_stride = conv_params->dst_stride; - const uint32x4_t offset_vec = vdupq_n_u32(offset); - const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); - uint16x4_t fwd_offset = vdup_n_u16(conv_params->fwd_offset); - uint16x4_t bck_offset = vdup_n_u16(conv_params->bck_offset); - - // Weighted averaging - if (w == 4) { - do { - const uint16x4_t src = vld1_u16(src_ptr); - const uint16x4_t ref = vld1_u16(ref_ptr); - - uint32x4_t wtd_avg = vmull_u16(ref, fwd_offset); - wtd_avg = vmlal_u16(wtd_avg, src, bck_offset); - wtd_avg = vshrq_n_u32(wtd_avg, DIST_PRECISION_BITS); - int32x4_t d0 = vreinterpretq_s32_u32(vsubq_u32(wtd_avg, offset_vec)); - - uint16x4_t d0_u16 = vqrshrun_n_s32(d0, ROUND_SHIFT - 2); - d0_u16 = vmin_u16(d0_u16, vget_low_u16(max)); - - vst1_u16(dst_ptr, d0_u16); - - src_ptr += src_stride; - dst_ptr += dst_stride; - ref_ptr += ref_stride; - } while (--h != 0); - } else { - do { - int width = w; - const uint16_t *src = src_ptr; - const uint16_t *ref = ref_ptr; - uint16_t *dst = dst_ptr; - do { - const uint16x8_t s = vld1q_u16(src); - const uint16x8_t r = vld1q_u16(ref); - - uint32x4_t wtd_avg0 = vmull_u16(vget_low_u16(r), fwd_offset); - wtd_avg0 = vmlal_u16(wtd_avg0, vget_low_u16(s), bck_offset); - wtd_avg0 = vshrq_n_u32(wtd_avg0, DIST_PRECISION_BITS); - int32x4_t d0 = vreinterpretq_s32_u32(vsubq_u32(wtd_avg0, offset_vec)); - - uint32x4_t wtd_avg1 = vmull_u16(vget_high_u16(r), fwd_offset); - wtd_avg1 = vmlal_u16(wtd_avg1, vget_high_u16(s), bck_offset); - wtd_avg1 = vshrq_n_u32(wtd_avg1, DIST_PRECISION_BITS); - int32x4_t d1 = vreinterpretq_s32_u32(vsubq_u32(wtd_avg1, offset_vec)); - - uint16x8_t d01 = vcombine_u16(vqrshrun_n_s32(d0, ROUND_SHIFT - 2), - vqrshrun_n_s32(d1, ROUND_SHIFT - 2)); - d01 = vminq_u16(d01, max); - vst1q_u16(dst, d01); - - src += 8; - ref += 8; - dst += 8; - width -= 8; - } while (width != 0); - src_ptr += src_stride; - dst_ptr += dst_stride; - ref_ptr += ref_stride; - } while (--h != 0); - } -} - -static INLINE void highbd_dist_wtd_comp_avg_neon( - const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, - int w, int h, ConvolveParams *conv_params, const int offset, const int bd) { - CONV_BUF_TYPE *ref_ptr = conv_params->dst; - const int ref_stride = conv_params->dst_stride; - const uint32x4_t offset_vec = vdupq_n_u32(offset); - const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); - uint16x4_t fwd_offset = vdup_n_u16(conv_params->fwd_offset); - uint16x4_t bck_offset = vdup_n_u16(conv_params->bck_offset); - - // Weighted averaging - if (w == 4) { - do { - const uint16x4_t src = vld1_u16(src_ptr); - const uint16x4_t ref = vld1_u16(ref_ptr); - - uint32x4_t wtd_avg = vmull_u16(ref, fwd_offset); - wtd_avg = vmlal_u16(wtd_avg, src, bck_offset); - wtd_avg = vshrq_n_u32(wtd_avg, DIST_PRECISION_BITS); - int32x4_t d0 = vreinterpretq_s32_u32(vsubq_u32(wtd_avg, offset_vec)); - - uint16x4_t d0_u16 = vqrshrun_n_s32(d0, ROUND_SHIFT); - d0_u16 = vmin_u16(d0_u16, vget_low_u16(max)); - - vst1_u16(dst_ptr, d0_u16); - - src_ptr += src_stride; - dst_ptr += dst_stride; - ref_ptr += ref_stride; - } while (--h != 0); - } else { - do { - int width = w; - const uint16_t *src = src_ptr; - const uint16_t *ref = ref_ptr; - uint16_t *dst = dst_ptr; - do { - const uint16x8_t s = vld1q_u16(src); - const uint16x8_t r = vld1q_u16(ref); - - uint32x4_t wtd_avg0 = vmull_u16(vget_low_u16(r), fwd_offset); - wtd_avg0 = vmlal_u16(wtd_avg0, vget_low_u16(s), bck_offset); - wtd_avg0 = vshrq_n_u32(wtd_avg0, DIST_PRECISION_BITS); - int32x4_t d0 = vreinterpretq_s32_u32(vsubq_u32(wtd_avg0, offset_vec)); - - uint32x4_t wtd_avg1 = vmull_u16(vget_high_u16(r), fwd_offset); - wtd_avg1 = vmlal_u16(wtd_avg1, vget_high_u16(s), bck_offset); - wtd_avg1 = vshrq_n_u32(wtd_avg1, DIST_PRECISION_BITS); - int32x4_t d1 = vreinterpretq_s32_u32(vsubq_u32(wtd_avg1, offset_vec)); - - uint16x8_t d01 = vcombine_u16(vqrshrun_n_s32(d0, ROUND_SHIFT), - vqrshrun_n_s32(d1, ROUND_SHIFT)); - d01 = vminq_u16(d01, max); - vst1q_u16(dst, d01); - - src += 8; - ref += 8; - dst += 8; - width -= 8; - } while (width != 0); - src_ptr += src_stride; - dst_ptr += dst_stride; - ref_ptr += ref_stride; - } while (--h != 0); - } -} - static INLINE uint16x4_t highbd_12_convolve6_4( const int16x4_t s0, const int16x4_t s1, const int16x4_t s2, const int16x4_t s3, const int16x4_t s4, const int16x4_t s5, @@ -743,9 +486,6 @@ void av1_highbd_dist_wtd_convolve_x_neon( const int im_stride = MAX_SB_SIZE; const int horiz_offset = filter_params_x->taps / 2 - 1; assert(FILTER_BITS == COMPOUND_ROUND1_BITS); - const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0; - const int offset_avg = (1 << (offset_bits - conv_params->round_1)) + - (1 << (offset_bits - conv_params->round_1 - 1)); const int offset_convolve = (1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS)) + (1 << (bd + FILTER_BITS - 1)); @@ -768,10 +508,10 @@ void av1_highbd_dist_wtd_convolve_x_neon( } if (conv_params->use_dist_wtd_comp_avg) { highbd_12_dist_wtd_comp_avg_neon(im_block, im_stride, dst, dst_stride, - w, h, conv_params, offset_avg, bd); + w, h, conv_params); } else { highbd_12_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, h, - conv_params, offset_avg, bd); + conv_params); } } else { if (x_filter_taps <= 6 && w != 4) { @@ -795,10 +535,10 @@ void av1_highbd_dist_wtd_convolve_x_neon( } if (conv_params->use_dist_wtd_comp_avg) { highbd_dist_wtd_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, - h, conv_params, offset_avg, bd); + h, conv_params, bd); } else { highbd_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, h, - conv_params, offset_avg, bd); + conv_params, bd); } } else { if (x_filter_taps <= 6 && w != 4) { @@ -971,6 +711,212 @@ static INLINE void highbd_dist_wtd_convolve_y_6tap_neon( } } +static INLINE uint16x4_t highbd_12_convolve4_4( + const int16x4_t s0, const int16x4_t s1, const int16x4_t s2, + const int16x4_t s3, const int16x4_t filter, const int32x4_t offset) { + int32x4_t sum = vmlal_lane_s16(offset, s0, filter, 0); + sum = vmlal_lane_s16(sum, s1, filter, 1); + sum = vmlal_lane_s16(sum, s2, filter, 2); + sum = vmlal_lane_s16(sum, s3, filter, 3); + + return vqshrun_n_s32(sum, ROUND0_BITS + 2); +} + +static INLINE uint16x8_t highbd_12_convolve4_8( + const int16x8_t s0, const int16x8_t s1, const int16x8_t s2, + const int16x8_t s3, const int16x4_t filter, const int32x4_t offset) { + int32x4_t sum0 = vmlal_lane_s16(offset, vget_low_s16(s0), filter, 0); + sum0 = vmlal_lane_s16(sum0, vget_low_s16(s1), filter, 1); + sum0 = vmlal_lane_s16(sum0, vget_low_s16(s2), filter, 2); + sum0 = vmlal_lane_s16(sum0, vget_low_s16(s3), filter, 3); + + int32x4_t sum1 = vmlal_lane_s16(offset, vget_high_s16(s0), filter, 0); + sum1 = vmlal_lane_s16(sum1, vget_high_s16(s1), filter, 1); + sum1 = vmlal_lane_s16(sum1, vget_high_s16(s2), filter, 2); + sum1 = vmlal_lane_s16(sum1, vget_high_s16(s3), filter, 3); + + return vcombine_u16(vqshrun_n_s32(sum0, ROUND0_BITS + 2), + vqshrun_n_s32(sum1, ROUND0_BITS + 2)); +} + +static INLINE void highbd_12_dist_wtd_convolve_y_4tap_neon( + const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, + int w, int h, const int16_t *y_filter_ptr, const int offset) { + const int16x4_t y_filter = vld1_s16(y_filter_ptr + 2); + const int32x4_t offset_vec = vdupq_n_s32(offset); + + if (w == 4) { + const int16_t *s = (const int16_t *)src_ptr; + uint16_t *d = dst_ptr; + + int16x4_t s0, s1, s2; + load_s16_4x3(s, src_stride, &s0, &s1, &s2); + s += 3 * src_stride; + + do { + int16x4_t s3, s4, s5, s6; + load_s16_4x4(s, src_stride, &s3, &s4, &s5, &s6); + + uint16x4_t d0 = + highbd_12_convolve4_4(s0, s1, s2, s3, y_filter, offset_vec); + uint16x4_t d1 = + highbd_12_convolve4_4(s1, s2, s3, s4, y_filter, offset_vec); + uint16x4_t d2 = + highbd_12_convolve4_4(s2, s3, s4, s5, y_filter, offset_vec); + uint16x4_t d3 = + highbd_12_convolve4_4(s3, s4, s5, s6, y_filter, offset_vec); + + store_u16_4x4(d, dst_stride, d0, d1, d2, d3); + + s0 = s4; + s1 = s5; + s2 = s6; + + s += 4 * src_stride; + d += 4 * dst_stride; + h -= 4; + } while (h != 0); + } else { + do { + int height = h; + const int16_t *s = (const int16_t *)src_ptr; + uint16_t *d = dst_ptr; + + int16x8_t s0, s1, s2; + load_s16_8x3(s, src_stride, &s0, &s1, &s2); + s += 3 * src_stride; + + do { + int16x8_t s3, s4, s5, s6; + load_s16_8x4(s, src_stride, &s3, &s4, &s5, &s6); + + uint16x8_t d0 = + highbd_12_convolve4_8(s0, s1, s2, s3, y_filter, offset_vec); + uint16x8_t d1 = + highbd_12_convolve4_8(s1, s2, s3, s4, y_filter, offset_vec); + uint16x8_t d2 = + highbd_12_convolve4_8(s2, s3, s4, s5, y_filter, offset_vec); + uint16x8_t d3 = + highbd_12_convolve4_8(s3, s4, s5, s6, y_filter, offset_vec); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s0 = s4; + s1 = s5; + s2 = s6; + + s += 4 * src_stride; + d += 4 * dst_stride; + height -= 4; + } while (height != 0); + src_ptr += 8; + dst_ptr += 8; + w -= 8; + } while (w != 0); + } +} + +static INLINE uint16x4_t highbd_convolve4_4( + const int16x4_t s0, const int16x4_t s1, const int16x4_t s2, + const int16x4_t s3, const int16x4_t filter, const int32x4_t offset) { + int32x4_t sum = vmlal_lane_s16(offset, s0, filter, 0); + sum = vmlal_lane_s16(sum, s1, filter, 1); + sum = vmlal_lane_s16(sum, s2, filter, 2); + sum = vmlal_lane_s16(sum, s3, filter, 3); + + return vqshrun_n_s32(sum, ROUND0_BITS); +} + +static INLINE uint16x8_t highbd_convolve4_8( + const int16x8_t s0, const int16x8_t s1, const int16x8_t s2, + const int16x8_t s3, const int16x4_t filter, const int32x4_t offset) { + int32x4_t sum0 = vmlal_lane_s16(offset, vget_low_s16(s0), filter, 0); + sum0 = vmlal_lane_s16(sum0, vget_low_s16(s1), filter, 1); + sum0 = vmlal_lane_s16(sum0, vget_low_s16(s2), filter, 2); + sum0 = vmlal_lane_s16(sum0, vget_low_s16(s3), filter, 3); + + int32x4_t sum1 = vmlal_lane_s16(offset, vget_high_s16(s0), filter, 0); + sum1 = vmlal_lane_s16(sum1, vget_high_s16(s1), filter, 1); + sum1 = vmlal_lane_s16(sum1, vget_high_s16(s2), filter, 2); + sum1 = vmlal_lane_s16(sum1, vget_high_s16(s3), filter, 3); + + return vcombine_u16(vqshrun_n_s32(sum0, ROUND0_BITS), + vqshrun_n_s32(sum1, ROUND0_BITS)); +} + +static INLINE void highbd_dist_wtd_convolve_y_4tap_neon( + const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, + int w, int h, const int16_t *y_filter_ptr, const int offset) { + const int16x4_t y_filter = vld1_s16(y_filter_ptr + 2); + const int32x4_t offset_vec = vdupq_n_s32(offset); + + if (w == 4) { + const int16_t *s = (const int16_t *)src_ptr; + uint16_t *d = dst_ptr; + + int16x4_t s0, s1, s2; + load_s16_4x3(s, src_stride, &s0, &s1, &s2); + s += 3 * src_stride; + + do { + int16x4_t s3, s4, s5, s6; + load_s16_4x4(s, src_stride, &s3, &s4, &s5, &s6); + + uint16x4_t d0 = highbd_convolve4_4(s0, s1, s2, s3, y_filter, offset_vec); + uint16x4_t d1 = highbd_convolve4_4(s1, s2, s3, s4, y_filter, offset_vec); + uint16x4_t d2 = highbd_convolve4_4(s2, s3, s4, s5, y_filter, offset_vec); + uint16x4_t d3 = highbd_convolve4_4(s3, s4, s5, s6, y_filter, offset_vec); + + store_u16_4x4(d, dst_stride, d0, d1, d2, d3); + + s0 = s4; + s1 = s5; + s2 = s6; + + s += 4 * src_stride; + d += 4 * dst_stride; + h -= 4; + } while (h != 0); + } else { + do { + int height = h; + const int16_t *s = (const int16_t *)src_ptr; + uint16_t *d = dst_ptr; + + int16x8_t s0, s1, s2; + load_s16_8x3(s, src_stride, &s0, &s1, &s2); + s += 3 * src_stride; + + do { + int16x8_t s3, s4, s5, s6; + load_s16_8x4(s, src_stride, &s3, &s4, &s5, &s6); + + uint16x8_t d0 = + highbd_convolve4_8(s0, s1, s2, s3, y_filter, offset_vec); + uint16x8_t d1 = + highbd_convolve4_8(s1, s2, s3, s4, y_filter, offset_vec); + uint16x8_t d2 = + highbd_convolve4_8(s2, s3, s4, s5, y_filter, offset_vec); + uint16x8_t d3 = + highbd_convolve4_8(s3, s4, s5, s6, y_filter, offset_vec); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s0 = s4; + s1 = s5; + s2 = s6; + + s += 4 * src_stride; + d += 4 * dst_stride; + height -= 4; + } while (height != 0); + src_ptr += 8; + dst_ptr += 8; + w -= 8; + } while (w != 0); + } +} + static INLINE void highbd_12_dist_wtd_convolve_y_8tap_neon( const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, int w, int h, const int16_t *y_filter_ptr, const int offset) { @@ -1148,9 +1094,6 @@ void av1_highbd_dist_wtd_convolve_y_neon( const int im_stride = MAX_SB_SIZE; const int vert_offset = filter_params_y->taps / 2 - 1; assert(FILTER_BITS == COMPOUND_ROUND1_BITS); - const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0; - const int round_offset_avg = (1 << (offset_bits - conv_params->round_1)) + - (1 << (offset_bits - conv_params->round_1 - 1)); const int round_offset_conv = (1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS)) + (1 << (bd + FILTER_BITS - 1)); @@ -1162,7 +1105,11 @@ void av1_highbd_dist_wtd_convolve_y_neon( if (bd == 12) { if (conv_params->do_average) { - if (y_filter_taps <= 6) { + if (y_filter_taps <= 4) { + highbd_12_dist_wtd_convolve_y_4tap_neon( + src + 2 * src_stride, src_stride, im_block, im_stride, w, h, + y_filter_ptr, round_offset_conv); + } else if (y_filter_taps == 6) { highbd_12_dist_wtd_convolve_y_6tap_neon( src + src_stride, src_stride, im_block, im_stride, w, h, y_filter_ptr, round_offset_conv); @@ -1173,14 +1120,17 @@ void av1_highbd_dist_wtd_convolve_y_neon( } if (conv_params->use_dist_wtd_comp_avg) { highbd_12_dist_wtd_comp_avg_neon(im_block, im_stride, dst, dst_stride, - w, h, conv_params, round_offset_avg, - bd); + w, h, conv_params); } else { highbd_12_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, h, - conv_params, round_offset_avg, bd); + conv_params); } } else { - if (y_filter_taps <= 6) { + if (y_filter_taps <= 4) { + highbd_12_dist_wtd_convolve_y_4tap_neon( + src + 2 * src_stride, src_stride, dst16, dst16_stride, w, h, + y_filter_ptr, round_offset_conv); + } else if (y_filter_taps == 6) { highbd_12_dist_wtd_convolve_y_6tap_neon( src + src_stride, src_stride, dst16, dst16_stride, w, h, y_filter_ptr, round_offset_conv); @@ -1192,7 +1142,11 @@ void av1_highbd_dist_wtd_convolve_y_neon( } } else { if (conv_params->do_average) { - if (y_filter_taps <= 6) { + if (y_filter_taps <= 4) { + highbd_dist_wtd_convolve_y_4tap_neon(src + 2 * src_stride, src_stride, + im_block, im_stride, w, h, + y_filter_ptr, round_offset_conv); + } else if (y_filter_taps == 6) { highbd_dist_wtd_convolve_y_6tap_neon(src + src_stride, src_stride, im_block, im_stride, w, h, y_filter_ptr, round_offset_conv); @@ -1203,13 +1157,17 @@ void av1_highbd_dist_wtd_convolve_y_neon( } if (conv_params->use_dist_wtd_comp_avg) { highbd_dist_wtd_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, - h, conv_params, round_offset_avg, bd); + h, conv_params, bd); } else { highbd_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, h, - conv_params, round_offset_avg, bd); + conv_params, bd); } } else { - if (y_filter_taps <= 6) { + if (y_filter_taps <= 4) { + highbd_dist_wtd_convolve_y_4tap_neon(src + 2 * src_stride, src_stride, + dst16, dst16_stride, w, h, + y_filter_ptr, round_offset_conv); + } else if (y_filter_taps == 6) { highbd_dist_wtd_convolve_y_6tap_neon(src + src_stride, src_stride, dst16, dst16_stride, w, h, y_filter_ptr, round_offset_conv); @@ -1285,18 +1243,18 @@ void av1_highbd_dist_wtd_convolve_2d_copy_neon(const uint16_t *src, if (conv_params->use_dist_wtd_comp_avg) { if (bd == 12) { highbd_12_dist_wtd_comp_avg_neon(im_block, im_stride, dst, dst_stride, - w, h, conv_params, round_offset, bd); + w, h, conv_params); } else { highbd_dist_wtd_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, - h, conv_params, round_offset, bd); + h, conv_params, bd); } } else { if (bd == 12) { highbd_12_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, h, - conv_params, round_offset, bd); + conv_params); } else { highbd_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, h, - conv_params, round_offset, bd); + conv_params, bd); } } } @@ -1949,9 +1907,6 @@ void av1_highbd_dist_wtd_convolve_2d_neon( (1 << (bd + FILTER_BITS - 1)) + (1 << (conv_params->round_0 - 1)); const int y_offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0; const int round_offset_conv_y = (1 << y_offset_bits); - const int round_offset_avg = - ((1 << (y_offset_bits - conv_params->round_1)) + - (1 << (y_offset_bits - conv_params->round_1 - 1))); const uint16_t *src_ptr = src - vert_offset * src_stride - horiz_offset; @@ -2012,19 +1967,18 @@ void av1_highbd_dist_wtd_convolve_2d_neon( if (conv_params->use_dist_wtd_comp_avg) { if (bd == 12) { highbd_12_dist_wtd_comp_avg_neon(im_block2, im_stride, dst, dst_stride, - w, h, conv_params, round_offset_avg, - bd); + w, h, conv_params); } else { highbd_dist_wtd_comp_avg_neon(im_block2, im_stride, dst, dst_stride, w, - h, conv_params, round_offset_avg, bd); + h, conv_params, bd); } } else { if (bd == 12) { highbd_12_comp_avg_neon(im_block2, im_stride, dst, dst_stride, w, h, - conv_params, round_offset_avg, bd); + conv_params); } else { highbd_comp_avg_neon(im_block2, im_stride, dst, dst_stride, w, h, - conv_params, round_offset_avg, bd); + conv_params, bd); } } } diff --git a/third_party/aom/av1/common/arm/highbd_compound_convolve_neon.h b/third_party/aom/av1/common/arm/highbd_compound_convolve_neon.h new file mode 100644 index 0000000000..c9344f3adf --- /dev/null +++ b/third_party/aom/av1/common/arm/highbd_compound_convolve_neon.h @@ -0,0 +1,293 @@ +/* + * Copyright (c) 2024, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <assert.h> +#include <arm_neon.h> + +#include "config/aom_config.h" +#include "config/av1_rtcd.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/arm/mem_neon.h" +#include "aom_ports/mem.h" + +#define ROUND_SHIFT 2 * FILTER_BITS - ROUND0_BITS - COMPOUND_ROUND1_BITS + +static INLINE void highbd_12_comp_avg_neon(const uint16_t *src_ptr, + int src_stride, uint16_t *dst_ptr, + int dst_stride, int w, int h, + ConvolveParams *conv_params) { + const int offset_bits = 12 + 2 * FILTER_BITS - ROUND0_BITS - 2; + const int offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + + (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); + + CONV_BUF_TYPE *ref_ptr = conv_params->dst; + const int ref_stride = conv_params->dst_stride; + const uint16x4_t offset_vec = vdup_n_u16((uint16_t)offset); + const uint16x8_t max = vdupq_n_u16((1 << 12) - 1); + + if (w == 4) { + do { + const uint16x4_t src = vld1_u16(src_ptr); + const uint16x4_t ref = vld1_u16(ref_ptr); + + uint16x4_t avg = vhadd_u16(src, ref); + int32x4_t d0 = vreinterpretq_s32_u32(vsubl_u16(avg, offset_vec)); + + uint16x4_t d0_u16 = vqrshrun_n_s32(d0, ROUND_SHIFT - 2); + d0_u16 = vmin_u16(d0_u16, vget_low_u16(max)); + + vst1_u16(dst_ptr, d0_u16); + + src_ptr += src_stride; + ref_ptr += ref_stride; + dst_ptr += dst_stride; + } while (--h != 0); + } else { + do { + int width = w; + const uint16_t *src = src_ptr; + const uint16_t *ref = ref_ptr; + uint16_t *dst = dst_ptr; + do { + const uint16x8_t s = vld1q_u16(src); + const uint16x8_t r = vld1q_u16(ref); + + uint16x8_t avg = vhaddq_u16(s, r); + int32x4_t d0_lo = + vreinterpretq_s32_u32(vsubl_u16(vget_low_u16(avg), offset_vec)); + int32x4_t d0_hi = + vreinterpretq_s32_u32(vsubl_u16(vget_high_u16(avg), offset_vec)); + + uint16x8_t d0 = vcombine_u16(vqrshrun_n_s32(d0_lo, ROUND_SHIFT - 2), + vqrshrun_n_s32(d0_hi, ROUND_SHIFT - 2)); + d0 = vminq_u16(d0, max); + vst1q_u16(dst, d0); + + src += 8; + ref += 8; + dst += 8; + width -= 8; + } while (width != 0); + + src_ptr += src_stride; + ref_ptr += ref_stride; + dst_ptr += dst_stride; + } while (--h != 0); + } +} + +static INLINE void highbd_comp_avg_neon(const uint16_t *src_ptr, int src_stride, + uint16_t *dst_ptr, int dst_stride, + int w, int h, + ConvolveParams *conv_params, + const int bd) { + const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; + const int offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + + (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); + + CONV_BUF_TYPE *ref_ptr = conv_params->dst; + const int ref_stride = conv_params->dst_stride; + const uint16x4_t offset_vec = vdup_n_u16((uint16_t)offset); + const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); + + if (w == 4) { + do { + const uint16x4_t src = vld1_u16(src_ptr); + const uint16x4_t ref = vld1_u16(ref_ptr); + + uint16x4_t avg = vhadd_u16(src, ref); + int32x4_t d0 = vreinterpretq_s32_u32(vsubl_u16(avg, offset_vec)); + + uint16x4_t d0_u16 = vqrshrun_n_s32(d0, ROUND_SHIFT); + d0_u16 = vmin_u16(d0_u16, vget_low_u16(max)); + + vst1_u16(dst_ptr, d0_u16); + + src_ptr += src_stride; + ref_ptr += ref_stride; + dst_ptr += dst_stride; + } while (--h != 0); + } else { + do { + int width = w; + const uint16_t *src = src_ptr; + const uint16_t *ref = ref_ptr; + uint16_t *dst = dst_ptr; + do { + const uint16x8_t s = vld1q_u16(src); + const uint16x8_t r = vld1q_u16(ref); + + uint16x8_t avg = vhaddq_u16(s, r); + int32x4_t d0_lo = + vreinterpretq_s32_u32(vsubl_u16(vget_low_u16(avg), offset_vec)); + int32x4_t d0_hi = + vreinterpretq_s32_u32(vsubl_u16(vget_high_u16(avg), offset_vec)); + + uint16x8_t d0 = vcombine_u16(vqrshrun_n_s32(d0_lo, ROUND_SHIFT), + vqrshrun_n_s32(d0_hi, ROUND_SHIFT)); + d0 = vminq_u16(d0, max); + vst1q_u16(dst, d0); + + src += 8; + ref += 8; + dst += 8; + width -= 8; + } while (width != 0); + + src_ptr += src_stride; + ref_ptr += ref_stride; + dst_ptr += dst_stride; + } while (--h != 0); + } +} + +static INLINE void highbd_12_dist_wtd_comp_avg_neon( + const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, + int w, int h, ConvolveParams *conv_params) { + const int offset_bits = 12 + 2 * FILTER_BITS - ROUND0_BITS - 2; + const int offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + + (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); + + CONV_BUF_TYPE *ref_ptr = conv_params->dst; + const int ref_stride = conv_params->dst_stride; + const uint32x4_t offset_vec = vdupq_n_u32(offset); + const uint16x8_t max = vdupq_n_u16((1 << 12) - 1); + uint16x4_t fwd_offset = vdup_n_u16(conv_params->fwd_offset); + uint16x4_t bck_offset = vdup_n_u16(conv_params->bck_offset); + + // Weighted averaging + if (w == 4) { + do { + const uint16x4_t src = vld1_u16(src_ptr); + const uint16x4_t ref = vld1_u16(ref_ptr); + + uint32x4_t wtd_avg = vmull_u16(ref, fwd_offset); + wtd_avg = vmlal_u16(wtd_avg, src, bck_offset); + wtd_avg = vshrq_n_u32(wtd_avg, DIST_PRECISION_BITS); + int32x4_t d0 = vreinterpretq_s32_u32(vsubq_u32(wtd_avg, offset_vec)); + + uint16x4_t d0_u16 = vqrshrun_n_s32(d0, ROUND_SHIFT - 2); + d0_u16 = vmin_u16(d0_u16, vget_low_u16(max)); + + vst1_u16(dst_ptr, d0_u16); + + src_ptr += src_stride; + dst_ptr += dst_stride; + ref_ptr += ref_stride; + } while (--h != 0); + } else { + do { + int width = w; + const uint16_t *src = src_ptr; + const uint16_t *ref = ref_ptr; + uint16_t *dst = dst_ptr; + do { + const uint16x8_t s = vld1q_u16(src); + const uint16x8_t r = vld1q_u16(ref); + + uint32x4_t wtd_avg0 = vmull_u16(vget_low_u16(r), fwd_offset); + wtd_avg0 = vmlal_u16(wtd_avg0, vget_low_u16(s), bck_offset); + wtd_avg0 = vshrq_n_u32(wtd_avg0, DIST_PRECISION_BITS); + int32x4_t d0 = vreinterpretq_s32_u32(vsubq_u32(wtd_avg0, offset_vec)); + + uint32x4_t wtd_avg1 = vmull_u16(vget_high_u16(r), fwd_offset); + wtd_avg1 = vmlal_u16(wtd_avg1, vget_high_u16(s), bck_offset); + wtd_avg1 = vshrq_n_u32(wtd_avg1, DIST_PRECISION_BITS); + int32x4_t d1 = vreinterpretq_s32_u32(vsubq_u32(wtd_avg1, offset_vec)); + + uint16x8_t d01 = vcombine_u16(vqrshrun_n_s32(d0, ROUND_SHIFT - 2), + vqrshrun_n_s32(d1, ROUND_SHIFT - 2)); + d01 = vminq_u16(d01, max); + vst1q_u16(dst, d01); + + src += 8; + ref += 8; + dst += 8; + width -= 8; + } while (width != 0); + src_ptr += src_stride; + dst_ptr += dst_stride; + ref_ptr += ref_stride; + } while (--h != 0); + } +} + +static INLINE void highbd_dist_wtd_comp_avg_neon( + const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, + int w, int h, ConvolveParams *conv_params, const int bd) { + const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; + const int offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + + (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); + + CONV_BUF_TYPE *ref_ptr = conv_params->dst; + const int ref_stride = conv_params->dst_stride; + const uint32x4_t offset_vec = vdupq_n_u32(offset); + const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); + uint16x4_t fwd_offset = vdup_n_u16(conv_params->fwd_offset); + uint16x4_t bck_offset = vdup_n_u16(conv_params->bck_offset); + + // Weighted averaging + if (w == 4) { + do { + const uint16x4_t src = vld1_u16(src_ptr); + const uint16x4_t ref = vld1_u16(ref_ptr); + + uint32x4_t wtd_avg = vmull_u16(ref, fwd_offset); + wtd_avg = vmlal_u16(wtd_avg, src, bck_offset); + wtd_avg = vshrq_n_u32(wtd_avg, DIST_PRECISION_BITS); + int32x4_t d0 = vreinterpretq_s32_u32(vsubq_u32(wtd_avg, offset_vec)); + + uint16x4_t d0_u16 = vqrshrun_n_s32(d0, ROUND_SHIFT); + d0_u16 = vmin_u16(d0_u16, vget_low_u16(max)); + + vst1_u16(dst_ptr, d0_u16); + + src_ptr += src_stride; + dst_ptr += dst_stride; + ref_ptr += ref_stride; + } while (--h != 0); + } else { + do { + int width = w; + const uint16_t *src = src_ptr; + const uint16_t *ref = ref_ptr; + uint16_t *dst = dst_ptr; + do { + const uint16x8_t s = vld1q_u16(src); + const uint16x8_t r = vld1q_u16(ref); + + uint32x4_t wtd_avg0 = vmull_u16(vget_low_u16(r), fwd_offset); + wtd_avg0 = vmlal_u16(wtd_avg0, vget_low_u16(s), bck_offset); + wtd_avg0 = vshrq_n_u32(wtd_avg0, DIST_PRECISION_BITS); + int32x4_t d0 = vreinterpretq_s32_u32(vsubq_u32(wtd_avg0, offset_vec)); + + uint32x4_t wtd_avg1 = vmull_u16(vget_high_u16(r), fwd_offset); + wtd_avg1 = vmlal_u16(wtd_avg1, vget_high_u16(s), bck_offset); + wtd_avg1 = vshrq_n_u32(wtd_avg1, DIST_PRECISION_BITS); + int32x4_t d1 = vreinterpretq_s32_u32(vsubq_u32(wtd_avg1, offset_vec)); + + uint16x8_t d01 = vcombine_u16(vqrshrun_n_s32(d0, ROUND_SHIFT), + vqrshrun_n_s32(d1, ROUND_SHIFT)); + d01 = vminq_u16(d01, max); + vst1q_u16(dst, d01); + + src += 8; + ref += 8; + dst += 8; + width -= 8; + } while (width != 0); + src_ptr += src_stride; + dst_ptr += dst_stride; + ref_ptr += ref_stride; + } while (--h != 0); + } +} diff --git a/third_party/aom/av1/common/arm/highbd_compound_convolve_sve2.c b/third_party/aom/av1/common/arm/highbd_compound_convolve_sve2.c new file mode 100644 index 0000000000..1d6c9b4faf --- /dev/null +++ b/third_party/aom/av1/common/arm/highbd_compound_convolve_sve2.c @@ -0,0 +1,1555 @@ +/* + * Copyright (c) 2024, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <assert.h> +#include <arm_neon.h> + +#include "config/aom_config.h" +#include "config/av1_rtcd.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/arm/aom_neon_sve_bridge.h" +#include "aom_dsp/arm/aom_neon_sve2_bridge.h" +#include "aom_dsp/arm/mem_neon.h" +#include "aom_ports/mem.h" +#include "av1/common/convolve.h" +#include "av1/common/filter.h" +#include "av1/common/filter.h" +#include "av1/common/arm/highbd_compound_convolve_neon.h" +#include "av1/common/arm/highbd_convolve_neon.h" +#include "av1/common/arm/highbd_convolve_sve2.h" + +DECLARE_ALIGNED(16, static const uint16_t, kDotProdTbl[32]) = { + 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6, + 4, 5, 6, 7, 5, 6, 7, 0, 6, 7, 0, 1, 7, 0, 1, 2, +}; + +static INLINE uint16x8_t highbd_12_convolve8_8_x(int16x8_t s0[8], + int16x8_t filter, + int64x2_t offset) { + int64x2_t sum[8]; + sum[0] = aom_sdotq_s16(offset, s0[0], filter); + sum[1] = aom_sdotq_s16(offset, s0[1], filter); + sum[2] = aom_sdotq_s16(offset, s0[2], filter); + sum[3] = aom_sdotq_s16(offset, s0[3], filter); + sum[4] = aom_sdotq_s16(offset, s0[4], filter); + sum[5] = aom_sdotq_s16(offset, s0[5], filter); + sum[6] = aom_sdotq_s16(offset, s0[6], filter); + sum[7] = aom_sdotq_s16(offset, s0[7], filter); + + sum[0] = vpaddq_s64(sum[0], sum[1]); + sum[2] = vpaddq_s64(sum[2], sum[3]); + sum[4] = vpaddq_s64(sum[4], sum[5]); + sum[6] = vpaddq_s64(sum[6], sum[7]); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum[0]), vmovn_s64(sum[2])); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum[4]), vmovn_s64(sum[6])); + + return vcombine_u16(vqrshrun_n_s32(sum0123, ROUND0_BITS + 2), + vqrshrun_n_s32(sum4567, ROUND0_BITS + 2)); +} + +static INLINE void highbd_12_dist_wtd_convolve_x_8tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *x_filter_ptr) { + const int64x1_t offset_vec = + vcreate_s64((1 << (12 + FILTER_BITS)) + (1 << (12 + FILTER_BITS - 1))); + const int64x2_t offset_lo = vcombine_s64(offset_vec, vdup_n_s64(0)); + + const int16x8_t filter = vld1q_s16(x_filter_ptr); + + do { + const int16_t *s = (const int16_t *)src; + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0[8], s1[8], s2[8], s3[8]; + load_s16_8x8(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3], + &s0[4], &s0[5], &s0[6], &s0[7]); + load_s16_8x8(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3], + &s1[4], &s1[5], &s1[6], &s1[7]); + load_s16_8x8(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3], + &s2[4], &s2[5], &s2[6], &s2[7]); + load_s16_8x8(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3], + &s3[4], &s3[5], &s3[6], &s3[7]); + + uint16x8_t d0 = highbd_12_convolve8_8_x(s0, filter, offset_lo); + uint16x8_t d1 = highbd_12_convolve8_8_x(s1, filter, offset_lo); + uint16x8_t d2 = highbd_12_convolve8_8_x(s2, filter, offset_lo); + uint16x8_t d3 = highbd_12_convolve8_8_x(s3, filter, offset_lo); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); +} + +static INLINE uint16x8_t highbd_convolve8_8_x(int16x8_t s0[8], int16x8_t filter, + int64x2_t offset) { + int64x2_t sum[8]; + sum[0] = aom_sdotq_s16(offset, s0[0], filter); + sum[1] = aom_sdotq_s16(offset, s0[1], filter); + sum[2] = aom_sdotq_s16(offset, s0[2], filter); + sum[3] = aom_sdotq_s16(offset, s0[3], filter); + sum[4] = aom_sdotq_s16(offset, s0[4], filter); + sum[5] = aom_sdotq_s16(offset, s0[5], filter); + sum[6] = aom_sdotq_s16(offset, s0[6], filter); + sum[7] = aom_sdotq_s16(offset, s0[7], filter); + + sum[0] = vpaddq_s64(sum[0], sum[1]); + sum[2] = vpaddq_s64(sum[2], sum[3]); + sum[4] = vpaddq_s64(sum[4], sum[5]); + sum[6] = vpaddq_s64(sum[6], sum[7]); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum[0]), vmovn_s64(sum[2])); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum[4]), vmovn_s64(sum[6])); + + return vcombine_u16(vqrshrun_n_s32(sum0123, ROUND0_BITS), + vqrshrun_n_s32(sum4567, ROUND0_BITS)); +} + +static INLINE void highbd_dist_wtd_convolve_x_8tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *x_filter_ptr, const int bd) { + const int64x1_t offset_vec = + vcreate_s64((1 << (bd + FILTER_BITS)) + (1 << (bd + FILTER_BITS - 1))); + const int64x2_t offset_lo = vcombine_s64(offset_vec, vdup_n_s64(0)); + + const int16x8_t filter = vld1q_s16(x_filter_ptr); + + do { + const int16_t *s = (const int16_t *)src; + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0[8], s1[8], s2[8], s3[8]; + load_s16_8x8(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3], + &s0[4], &s0[5], &s0[6], &s0[7]); + load_s16_8x8(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3], + &s1[4], &s1[5], &s1[6], &s1[7]); + load_s16_8x8(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3], + &s2[4], &s2[5], &s2[6], &s2[7]); + load_s16_8x8(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3], + &s3[4], &s3[5], &s3[6], &s3[7]); + + uint16x8_t d0 = highbd_convolve8_8_x(s0, filter, offset_lo); + uint16x8_t d1 = highbd_convolve8_8_x(s1, filter, offset_lo); + uint16x8_t d2 = highbd_convolve8_8_x(s2, filter, offset_lo); + uint16x8_t d3 = highbd_convolve8_8_x(s3, filter, offset_lo); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); +} + +// clang-format off +DECLARE_ALIGNED(16, static const uint16_t, kDeinterleaveTbl[8]) = { + 0, 2, 4, 6, 1, 3, 5, 7, +}; +// clang-format on + +static INLINE uint16x4_t highbd_12_convolve4_4_x(int16x8_t s0, int16x8_t filter, + int64x2_t offset, + uint16x8x2_t permute_tbl) { + int16x8_t permuted_samples0 = aom_tbl_s16(s0, permute_tbl.val[0]); + int16x8_t permuted_samples1 = aom_tbl_s16(s0, permute_tbl.val[1]); + + int64x2_t sum01 = aom_svdot_lane_s16(offset, permuted_samples0, filter, 0); + int64x2_t sum23 = aom_svdot_lane_s16(offset, permuted_samples1, filter, 0); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + + return vqrshrun_n_s32(sum0123, ROUND0_BITS + 2); +} + +static INLINE uint16x8_t highbd_12_convolve4_8_x(int16x8_t s0[4], + int16x8_t filter, + int64x2_t offset, + uint16x8_t tbl) { + int64x2_t sum04 = aom_svdot_lane_s16(offset, s0[0], filter, 0); + int64x2_t sum15 = aom_svdot_lane_s16(offset, s0[1], filter, 0); + int64x2_t sum26 = aom_svdot_lane_s16(offset, s0[2], filter, 0); + int64x2_t sum37 = aom_svdot_lane_s16(offset, s0[3], filter, 0); + + int32x4_t sum0415 = vcombine_s32(vmovn_s64(sum04), vmovn_s64(sum15)); + int32x4_t sum2637 = vcombine_s32(vmovn_s64(sum26), vmovn_s64(sum37)); + + uint16x8_t res = vcombine_u16(vqrshrun_n_s32(sum0415, ROUND0_BITS + 2), + vqrshrun_n_s32(sum2637, ROUND0_BITS + 2)); + return aom_tbl_u16(res, tbl); +} + +static INLINE void highbd_12_dist_wtd_convolve_x_4tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *x_filter_ptr) { + const int64x2_t offset = + vdupq_n_s64((1 << (12 + FILTER_BITS)) + (1 << (12 + FILTER_BITS - 1))); + + const int16x4_t x_filter = vld1_s16(x_filter_ptr + 2); + const int16x8_t filter = vcombine_s16(x_filter, vdup_n_s16(0)); + + if (width == 4) { + uint16x8x2_t permute_tbl = vld1q_u16_x2(kDotProdTbl); + + const int16_t *s = (const int16_t *)(src); + + do { + int16x8_t s0, s1, s2, s3; + load_s16_8x4(s, src_stride, &s0, &s1, &s2, &s3); + + uint16x4_t d0 = highbd_12_convolve4_4_x(s0, filter, offset, permute_tbl); + uint16x4_t d1 = highbd_12_convolve4_4_x(s1, filter, offset, permute_tbl); + uint16x4_t d2 = highbd_12_convolve4_4_x(s2, filter, offset, permute_tbl); + uint16x4_t d3 = highbd_12_convolve4_4_x(s3, filter, offset, permute_tbl); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } else { + uint16x8_t idx = vld1q_u16(kDeinterleaveTbl); + + do { + const int16_t *s = (const int16_t *)(src); + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0[4], s1[4], s2[4], s3[4]; + load_s16_8x4(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3]); + load_s16_8x4(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3]); + load_s16_8x4(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3]); + load_s16_8x4(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3]); + + uint16x8_t d0 = highbd_12_convolve4_8_x(s0, filter, offset, idx); + uint16x8_t d1 = highbd_12_convolve4_8_x(s1, filter, offset, idx); + uint16x8_t d2 = highbd_12_convolve4_8_x(s2, filter, offset, idx); + uint16x8_t d3 = highbd_12_convolve4_8_x(s3, filter, offset, idx); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } +} + +static INLINE uint16x4_t highbd_convolve4_4_x(int16x8_t s0, int16x8_t filter, + int64x2_t offset, + uint16x8x2_t permute_tbl) { + int16x8_t permuted_samples0 = aom_tbl_s16(s0, permute_tbl.val[0]); + int16x8_t permuted_samples1 = aom_tbl_s16(s0, permute_tbl.val[1]); + + int64x2_t sum01 = aom_svdot_lane_s16(offset, permuted_samples0, filter, 0); + int64x2_t sum23 = aom_svdot_lane_s16(offset, permuted_samples1, filter, 0); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + + return vqrshrun_n_s32(sum0123, ROUND0_BITS); +} + +static INLINE uint16x8_t highbd_convolve4_8_x(int16x8_t s0[4], int16x8_t filter, + int64x2_t offset, + uint16x8_t tbl) { + int64x2_t sum04 = aom_svdot_lane_s16(offset, s0[0], filter, 0); + int64x2_t sum15 = aom_svdot_lane_s16(offset, s0[1], filter, 0); + int64x2_t sum26 = aom_svdot_lane_s16(offset, s0[2], filter, 0); + int64x2_t sum37 = aom_svdot_lane_s16(offset, s0[3], filter, 0); + + int32x4_t sum0415 = vcombine_s32(vmovn_s64(sum04), vmovn_s64(sum15)); + int32x4_t sum2637 = vcombine_s32(vmovn_s64(sum26), vmovn_s64(sum37)); + + uint16x8_t res = vcombine_u16(vqrshrun_n_s32(sum0415, ROUND0_BITS), + vqrshrun_n_s32(sum2637, ROUND0_BITS)); + return aom_tbl_u16(res, tbl); +} + +static INLINE void highbd_dist_wtd_convolve_x_4tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *x_filter_ptr, const int bd) { + const int64x2_t offset = + vdupq_n_s64((1 << (bd + FILTER_BITS)) + (1 << (bd + FILTER_BITS - 1))); + + const int16x4_t x_filter = vld1_s16(x_filter_ptr + 2); + const int16x8_t filter = vcombine_s16(x_filter, vdup_n_s16(0)); + + if (width == 4) { + uint16x8x2_t permute_tbl = vld1q_u16_x2(kDotProdTbl); + + const int16_t *s = (const int16_t *)(src); + + do { + int16x8_t s0, s1, s2, s3; + load_s16_8x4(s, src_stride, &s0, &s1, &s2, &s3); + + uint16x4_t d0 = highbd_convolve4_4_x(s0, filter, offset, permute_tbl); + uint16x4_t d1 = highbd_convolve4_4_x(s1, filter, offset, permute_tbl); + uint16x4_t d2 = highbd_convolve4_4_x(s2, filter, offset, permute_tbl); + uint16x4_t d3 = highbd_convolve4_4_x(s3, filter, offset, permute_tbl); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } else { + uint16x8_t idx = vld1q_u16(kDeinterleaveTbl); + + do { + const int16_t *s = (const int16_t *)(src); + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0[4], s1[4], s2[4], s3[4]; + load_s16_8x4(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3]); + load_s16_8x4(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3]); + load_s16_8x4(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3]); + load_s16_8x4(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3]); + + uint16x8_t d0 = highbd_convolve4_8_x(s0, filter, offset, idx); + uint16x8_t d1 = highbd_convolve4_8_x(s1, filter, offset, idx); + uint16x8_t d2 = highbd_convolve4_8_x(s2, filter, offset, idx); + uint16x8_t d3 = highbd_convolve4_8_x(s3, filter, offset, idx); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } +} + +void av1_highbd_dist_wtd_convolve_x_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, + int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, + ConvolveParams *conv_params, int bd) { + DECLARE_ALIGNED(16, uint16_t, + im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]); + CONV_BUF_TYPE *dst16 = conv_params->dst; + const int x_filter_taps = get_filter_tap(filter_params_x, subpel_x_qn); + + if (x_filter_taps == 6) { + av1_highbd_dist_wtd_convolve_x_neon(src, src_stride, dst, dst_stride, w, h, + filter_params_x, subpel_x_qn, + conv_params, bd); + return; + } + + int dst16_stride = conv_params->dst_stride; + const int im_stride = MAX_SB_SIZE; + const int horiz_offset = filter_params_x->taps / 2 - 1; + assert(FILTER_BITS == COMPOUND_ROUND1_BITS); + + const int16_t *x_filter_ptr = av1_get_interp_filter_subpel_kernel( + filter_params_x, subpel_x_qn & SUBPEL_MASK); + + src -= horiz_offset; + + if (bd == 12) { + if (conv_params->do_average) { + if (x_filter_taps <= 4) { + highbd_12_dist_wtd_convolve_x_4tap_sve2(src + 2, src_stride, im_block, + im_stride, w, h, x_filter_ptr); + } else { + highbd_12_dist_wtd_convolve_x_8tap_sve2(src, src_stride, im_block, + im_stride, w, h, x_filter_ptr); + } + + if (conv_params->use_dist_wtd_comp_avg) { + highbd_12_dist_wtd_comp_avg_neon(im_block, im_stride, dst, dst_stride, + w, h, conv_params); + + } else { + highbd_12_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, h, + conv_params); + } + } else { + if (x_filter_taps <= 4) { + highbd_12_dist_wtd_convolve_x_4tap_sve2( + src + 2, src_stride, dst16, dst16_stride, w, h, x_filter_ptr); + } else { + highbd_12_dist_wtd_convolve_x_8tap_sve2( + src, src_stride, dst16, dst16_stride, w, h, x_filter_ptr); + } + } + } else { + if (conv_params->do_average) { + if (x_filter_taps <= 4) { + highbd_dist_wtd_convolve_x_4tap_sve2(src + 2, src_stride, im_block, + im_stride, w, h, x_filter_ptr, bd); + } else { + highbd_dist_wtd_convolve_x_8tap_sve2(src, src_stride, im_block, + im_stride, w, h, x_filter_ptr, bd); + } + + if (conv_params->use_dist_wtd_comp_avg) { + highbd_dist_wtd_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, + h, conv_params, bd); + } else { + highbd_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, h, + conv_params, bd); + } + } else { + if (x_filter_taps <= 4) { + highbd_dist_wtd_convolve_x_4tap_sve2( + src + 2, src_stride, dst16, dst16_stride, w, h, x_filter_ptr, bd); + } else { + highbd_dist_wtd_convolve_x_8tap_sve2( + src, src_stride, dst16, dst16_stride, w, h, x_filter_ptr, bd); + } + } + } +} + +static INLINE uint16x4_t highbd_12_convolve8_4_y(int16x8_t samples_lo[2], + int16x8_t samples_hi[2], + int16x8_t filter, + int64x2_t offset) { + int64x2_t sum01 = aom_svdot_lane_s16(offset, samples_lo[0], filter, 0); + sum01 = aom_svdot_lane_s16(sum01, samples_hi[0], filter, 1); + + int64x2_t sum23 = aom_svdot_lane_s16(offset, samples_lo[1], filter, 0); + sum23 = aom_svdot_lane_s16(sum23, samples_hi[1], filter, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + + return vqrshrun_n_s32(sum0123, ROUND0_BITS + 2); +} + +static INLINE uint16x8_t highbd_12_convolve8_8_y(int16x8_t samples_lo[4], + int16x8_t samples_hi[4], + int16x8_t filter, + int64x2_t offset) { + int64x2_t sum01 = aom_svdot_lane_s16(offset, samples_lo[0], filter, 0); + sum01 = aom_svdot_lane_s16(sum01, samples_hi[0], filter, 1); + + int64x2_t sum23 = aom_svdot_lane_s16(offset, samples_lo[1], filter, 0); + sum23 = aom_svdot_lane_s16(sum23, samples_hi[1], filter, 1); + + int64x2_t sum45 = aom_svdot_lane_s16(offset, samples_lo[2], filter, 0); + sum45 = aom_svdot_lane_s16(sum45, samples_hi[2], filter, 1); + + int64x2_t sum67 = aom_svdot_lane_s16(offset, samples_lo[3], filter, 0); + sum67 = aom_svdot_lane_s16(sum67, samples_hi[3], filter, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum45), vmovn_s64(sum67)); + + return vcombine_u16(vqrshrun_n_s32(sum0123, ROUND0_BITS + 2), + vqrshrun_n_s32(sum4567, ROUND0_BITS + 2)); +} + +static INLINE void highbd_12_dist_wtd_convolve_y_8tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *y_filter_ptr) { + const int64x2_t offset = + vdupq_n_s64((1 << (12 + FILTER_BITS)) + (1 << (12 + FILTER_BITS - 1))); + const int16x8_t y_filter = vld1q_s16(y_filter_ptr); + + uint16x8x3_t merge_block_tbl = vld1q_u16_x3(kDotProdMergeBlockTbl); + // Scale indices by size of the true vector length to avoid reading from an + // 'undefined' portion of a vector on a system with SVE vectors > 128-bit. + uint16x8_t correction0 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000000000000ULL)); + merge_block_tbl.val[0] = vaddq_u16(merge_block_tbl.val[0], correction0); + uint16x8_t correction1 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100000000ULL)); + merge_block_tbl.val[1] = vaddq_u16(merge_block_tbl.val[1], correction1); + + uint16x8_t correction2 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100010000ULL)); + merge_block_tbl.val[2] = vaddq_u16(merge_block_tbl.val[2], correction2); + + if (width == 4) { + int16_t *s = (int16_t *)src; + int16x4_t s0, s1, s2, s3, s4, s5, s6; + load_s16_4x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); + s += 7 * src_stride; + + // This operation combines a conventional transpose and the sample permute + // required before computing the dot product. + int16x8_t s0123[2], s1234[2], s2345[2], s3456[2]; + transpose_concat_4x4(s0, s1, s2, s3, s0123); + transpose_concat_4x4(s1, s2, s3, s4, s1234); + transpose_concat_4x4(s2, s3, s4, s5, s2345); + transpose_concat_4x4(s3, s4, s5, s6, s3456); + + do { + int16x4_t s7, s8, s9, s10; + load_s16_4x4(s, src_stride, &s7, &s8, &s9, &s10); + + int16x8_t s4567[2], s5678[2], s6789[2], s789A[2]; + // Transpose and shuffle the 4 lines that were loaded. + transpose_concat_4x4(s7, s8, s9, s10, s789A); + + // Merge new data into block from previous iteration. + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[0], s4567); + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[1], s5678); + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[2], s6789); + + uint16x4_t d0 = highbd_12_convolve8_4_y(s0123, s4567, y_filter, offset); + uint16x4_t d1 = highbd_12_convolve8_4_y(s1234, s5678, y_filter, offset); + uint16x4_t d2 = highbd_12_convolve8_4_y(s2345, s6789, y_filter, offset); + uint16x4_t d3 = highbd_12_convolve8_4_y(s3456, s789A, y_filter, offset); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. + s0123[0] = s4567[0]; + s0123[1] = s4567[1]; + s1234[0] = s5678[0]; + s1234[1] = s5678[1]; + s2345[0] = s6789[0]; + s2345[1] = s6789[1]; + s3456[0] = s789A[0]; + s3456[1] = s789A[1]; + + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } else { + do { + int h = height; + int16_t *s = (int16_t *)src; + uint16_t *d = dst; + + int16x8_t s0, s1, s2, s3, s4, s5, s6; + load_s16_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); + s += 7 * src_stride; + + // This operation combines a conventional transpose and the sample permute + // required before computing the dot product. + int16x8_t s0123[4], s1234[4], s2345[4], s3456[4]; + transpose_concat_8x4(s0, s1, s2, s3, s0123); + transpose_concat_8x4(s1, s2, s3, s4, s1234); + transpose_concat_8x4(s2, s3, s4, s5, s2345); + transpose_concat_8x4(s3, s4, s5, s6, s3456); + + do { + int16x8_t s7, s8, s9, s10; + load_s16_8x4(s, src_stride, &s7, &s8, &s9, &s10); + int16x8_t s4567[4], s5678[4], s6789[4], s789A[4]; + + // Transpose and shuffle the 4 lines that were loaded. + transpose_concat_8x4(s7, s8, s9, s10, s789A); + + // Merge new data into block from previous iteration. + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[0], s4567); + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[1], s5678); + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[2], s6789); + + uint16x8_t d0 = highbd_12_convolve8_8_y(s0123, s4567, y_filter, offset); + uint16x8_t d1 = highbd_12_convolve8_8_y(s1234, s5678, y_filter, offset); + uint16x8_t d2 = highbd_12_convolve8_8_y(s2345, s6789, y_filter, offset); + uint16x8_t d3 = highbd_12_convolve8_8_y(s3456, s789A, y_filter, offset); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. + s0123[0] = s4567[0]; + s0123[1] = s4567[1]; + s0123[2] = s4567[2]; + s0123[3] = s4567[3]; + s1234[0] = s5678[0]; + s1234[1] = s5678[1]; + s1234[2] = s5678[2]; + s1234[3] = s5678[3]; + s2345[0] = s6789[0]; + s2345[1] = s6789[1]; + s2345[2] = s6789[2]; + s2345[3] = s6789[3]; + s3456[0] = s789A[0]; + s3456[1] = s789A[1]; + s3456[2] = s789A[2]; + s3456[3] = s789A[3]; + + s += 4 * src_stride; + d += 4 * dst_stride; + h -= 4; + } while (h != 0); + src += 8; + dst += 8; + width -= 8; + } while (width != 0); + } +} + +static INLINE uint16x4_t highbd_convolve8_4_y(int16x8_t samples_lo[2], + int16x8_t samples_hi[2], + int16x8_t filter, + int64x2_t offset) { + int64x2_t sum01 = aom_svdot_lane_s16(offset, samples_lo[0], filter, 0); + sum01 = aom_svdot_lane_s16(sum01, samples_hi[0], filter, 1); + + int64x2_t sum23 = aom_svdot_lane_s16(offset, samples_lo[1], filter, 0); + sum23 = aom_svdot_lane_s16(sum23, samples_hi[1], filter, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + + return vqrshrun_n_s32(sum0123, ROUND0_BITS); +} + +static INLINE uint16x8_t highbd_convolve8_8_y(int16x8_t samples_lo[4], + int16x8_t samples_hi[4], + int16x8_t filter, + int64x2_t offset) { + int64x2_t sum01 = aom_svdot_lane_s16(offset, samples_lo[0], filter, 0); + sum01 = aom_svdot_lane_s16(sum01, samples_hi[0], filter, 1); + + int64x2_t sum23 = aom_svdot_lane_s16(offset, samples_lo[1], filter, 0); + sum23 = aom_svdot_lane_s16(sum23, samples_hi[1], filter, 1); + + int64x2_t sum45 = aom_svdot_lane_s16(offset, samples_lo[2], filter, 0); + sum45 = aom_svdot_lane_s16(sum45, samples_hi[2], filter, 1); + + int64x2_t sum67 = aom_svdot_lane_s16(offset, samples_lo[3], filter, 0); + sum67 = aom_svdot_lane_s16(sum67, samples_hi[3], filter, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum45), vmovn_s64(sum67)); + + return vcombine_u16(vqrshrun_n_s32(sum0123, ROUND0_BITS), + vqrshrun_n_s32(sum4567, ROUND0_BITS)); +} + +static INLINE void highbd_dist_wtd_convolve_y_8tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *y_filter_ptr, const int bd) { + const int64x2_t offset = + vdupq_n_s64((1 << (bd + FILTER_BITS)) + (1 << (bd + FILTER_BITS - 1))); + const int16x8_t y_filter = vld1q_s16(y_filter_ptr); + + uint16x8x3_t merge_block_tbl = vld1q_u16_x3(kDotProdMergeBlockTbl); + // Scale indices by size of the true vector length to avoid reading from an + // 'undefined' portion of a vector on a system with SVE vectors > 128-bit. + uint16x8_t correction0 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000000000000ULL)); + merge_block_tbl.val[0] = vaddq_u16(merge_block_tbl.val[0], correction0); + uint16x8_t correction1 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100000000ULL)); + merge_block_tbl.val[1] = vaddq_u16(merge_block_tbl.val[1], correction1); + + uint16x8_t correction2 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100010000ULL)); + merge_block_tbl.val[2] = vaddq_u16(merge_block_tbl.val[2], correction2); + + if (width == 4) { + int16_t *s = (int16_t *)src; + int16x4_t s0, s1, s2, s3, s4, s5, s6; + load_s16_4x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); + s += 7 * src_stride; + + // This operation combines a conventional transpose and the sample permute + // required before computing the dot product. + int16x8_t s0123[2], s1234[2], s2345[2], s3456[2]; + transpose_concat_4x4(s0, s1, s2, s3, s0123); + transpose_concat_4x4(s1, s2, s3, s4, s1234); + transpose_concat_4x4(s2, s3, s4, s5, s2345); + transpose_concat_4x4(s3, s4, s5, s6, s3456); + + do { + int16x4_t s7, s8, s9, s10; + load_s16_4x4(s, src_stride, &s7, &s8, &s9, &s10); + + int16x8_t s4567[2], s5678[2], s6789[2], s789A[2]; + // Transpose and shuffle the 4 lines that were loaded. + transpose_concat_4x4(s7, s8, s9, s10, s789A); + + // Merge new data into block from previous iteration. + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[0], s4567); + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[1], s5678); + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[2], s6789); + + uint16x4_t d0 = highbd_convolve8_4_y(s0123, s4567, y_filter, offset); + uint16x4_t d1 = highbd_convolve8_4_y(s1234, s5678, y_filter, offset); + uint16x4_t d2 = highbd_convolve8_4_y(s2345, s6789, y_filter, offset); + uint16x4_t d3 = highbd_convolve8_4_y(s3456, s789A, y_filter, offset); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. + s0123[0] = s4567[0]; + s0123[1] = s4567[1]; + s1234[0] = s5678[0]; + s1234[1] = s5678[1]; + s2345[0] = s6789[0]; + s2345[1] = s6789[1]; + s3456[0] = s789A[0]; + s3456[1] = s789A[1]; + + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } else { + do { + int h = height; + int16_t *s = (int16_t *)src; + uint16_t *d = dst; + + int16x8_t s0, s1, s2, s3, s4, s5, s6; + load_s16_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); + s += 7 * src_stride; + + // This operation combines a conventional transpose and the sample permute + // required before computing the dot product. + int16x8_t s0123[4], s1234[4], s2345[4], s3456[4]; + transpose_concat_8x4(s0, s1, s2, s3, s0123); + transpose_concat_8x4(s1, s2, s3, s4, s1234); + transpose_concat_8x4(s2, s3, s4, s5, s2345); + transpose_concat_8x4(s3, s4, s5, s6, s3456); + + do { + int16x8_t s7, s8, s9, s10; + load_s16_8x4(s, src_stride, &s7, &s8, &s9, &s10); + int16x8_t s4567[4], s5678[4], s6789[4], s789A[4]; + + // Transpose and shuffle the 4 lines that were loaded. + transpose_concat_8x4(s7, s8, s9, s10, s789A); + + // Merge new data into block from previous iteration. + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[0], s4567); + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[1], s5678); + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[2], s6789); + + uint16x8_t d0 = highbd_convolve8_8_y(s0123, s4567, y_filter, offset); + uint16x8_t d1 = highbd_convolve8_8_y(s1234, s5678, y_filter, offset); + uint16x8_t d2 = highbd_convolve8_8_y(s2345, s6789, y_filter, offset); + uint16x8_t d3 = highbd_convolve8_8_y(s3456, s789A, y_filter, offset); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. + s0123[0] = s4567[0]; + s0123[1] = s4567[1]; + s0123[2] = s4567[2]; + s0123[3] = s4567[3]; + s1234[0] = s5678[0]; + s1234[1] = s5678[1]; + s1234[2] = s5678[2]; + s1234[3] = s5678[3]; + s2345[0] = s6789[0]; + s2345[1] = s6789[1]; + s2345[2] = s6789[2]; + s2345[3] = s6789[3]; + s3456[0] = s789A[0]; + s3456[1] = s789A[1]; + s3456[2] = s789A[2]; + s3456[3] = s789A[3]; + + s += 4 * src_stride; + d += 4 * dst_stride; + h -= 4; + } while (h != 0); + src += 8; + dst += 8; + width -= 8; + } while (width != 0); + } +} + +void av1_highbd_dist_wtd_convolve_y_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, + int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, + ConvolveParams *conv_params, int bd) { + DECLARE_ALIGNED(16, uint16_t, + im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]); + CONV_BUF_TYPE *dst16 = conv_params->dst; + const int y_filter_taps = get_filter_tap(filter_params_y, subpel_y_qn); + + if (y_filter_taps != 8) { + av1_highbd_dist_wtd_convolve_y_neon(src, src_stride, dst, dst_stride, w, h, + filter_params_y, subpel_y_qn, + conv_params, bd); + return; + } + + int dst16_stride = conv_params->dst_stride; + const int im_stride = MAX_SB_SIZE; + const int vert_offset = filter_params_y->taps / 2 - 1; + assert(FILTER_BITS == COMPOUND_ROUND1_BITS); + + const int16_t *y_filter_ptr = av1_get_interp_filter_subpel_kernel( + filter_params_y, subpel_y_qn & SUBPEL_MASK); + + src -= vert_offset * src_stride; + + if (bd == 12) { + if (conv_params->do_average) { + highbd_12_dist_wtd_convolve_y_8tap_sve2(src, src_stride, im_block, + im_stride, w, h, y_filter_ptr); + if (conv_params->use_dist_wtd_comp_avg) { + highbd_12_dist_wtd_comp_avg_neon(im_block, im_stride, dst, dst_stride, + w, h, conv_params); + } else { + highbd_12_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, h, + conv_params); + } + } else { + highbd_12_dist_wtd_convolve_y_8tap_sve2(src, src_stride, dst16, + dst16_stride, w, h, y_filter_ptr); + } + } else { + if (conv_params->do_average) { + highbd_dist_wtd_convolve_y_8tap_sve2(src, src_stride, im_block, im_stride, + w, h, y_filter_ptr, bd); + if (conv_params->use_dist_wtd_comp_avg) { + highbd_dist_wtd_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, + h, conv_params, bd); + } else { + highbd_comp_avg_neon(im_block, im_stride, dst, dst_stride, w, h, + conv_params, bd); + } + } else { + highbd_dist_wtd_convolve_y_8tap_sve2(src, src_stride, dst16, dst16_stride, + w, h, y_filter_ptr, bd); + } + } +} + +static INLINE void highbd_12_dist_wtd_convolve_2d_horiz_8tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *x_filter_ptr) { + const int64x2_t offset = vdupq_n_s64(1 << (12 + FILTER_BITS - 2)); + const int16x8_t filter = vld1q_s16(x_filter_ptr); + + // We are only doing 8-tap and 4-tap vertical convolutions, therefore we know + // that im_h % 4 = 3, so we can do the loop across the whole block 4 rows at + // a time and then process the last 3 rows separately. + + do { + const int16_t *s = (const int16_t *)src; + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0[8], s1[8], s2[8], s3[8]; + load_s16_8x8(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3], + &s0[4], &s0[5], &s0[6], &s0[7]); + load_s16_8x8(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3], + &s1[4], &s1[5], &s1[6], &s1[7]); + load_s16_8x8(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3], + &s2[4], &s2[5], &s2[6], &s2[7]); + load_s16_8x8(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3], + &s3[4], &s3[5], &s3[6], &s3[7]); + + uint16x8_t d0 = highbd_12_convolve8_8_x(s0, filter, offset); + uint16x8_t d1 = highbd_12_convolve8_8_x(s1, filter, offset); + uint16x8_t d2 = highbd_12_convolve8_8_x(s2, filter, offset); + uint16x8_t d3 = highbd_12_convolve8_8_x(s3, filter, offset); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height > 4); + + // Process final 3 rows. + const int16_t *s = (const int16_t *)src; + do { + int16x8_t s0[8], s1[8], s2[8]; + load_s16_8x8(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3], &s0[4], + &s0[5], &s0[6], &s0[7]); + load_s16_8x8(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3], &s1[4], + &s1[5], &s1[6], &s1[7]); + load_s16_8x8(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3], &s2[4], + &s2[5], &s2[6], &s2[7]); + + uint16x8_t d0 = highbd_12_convolve8_8_x(s0, filter, offset); + uint16x8_t d1 = highbd_12_convolve8_8_x(s1, filter, offset); + uint16x8_t d2 = highbd_12_convolve8_8_x(s2, filter, offset); + + store_u16_8x3(dst, dst_stride, d0, d1, d2); + s += 8; + dst += 8; + width -= 8; + } while (width != 0); +} + +static INLINE void highbd_dist_wtd_convolve_2d_horiz_8tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *x_filter_ptr, const int bd) { + const int64x2_t offset = vdupq_n_s64(1 << (bd + FILTER_BITS - 2)); + const int16x8_t filter = vld1q_s16(x_filter_ptr); + + // We are only doing 8-tap and 4-tap vertical convolutions, therefore we know + // that im_h % 4 = 3, so we can do the loop across the whole block 4 rows at + // a time and then process the last 3 rows separately. + + do { + const int16_t *s = (const int16_t *)src; + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0[8], s1[8], s2[8], s3[8]; + load_s16_8x8(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3], + &s0[4], &s0[5], &s0[6], &s0[7]); + load_s16_8x8(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3], + &s1[4], &s1[5], &s1[6], &s1[7]); + load_s16_8x8(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3], + &s2[4], &s2[5], &s2[6], &s2[7]); + load_s16_8x8(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3], + &s3[4], &s3[5], &s3[6], &s3[7]); + + uint16x8_t d0 = highbd_convolve8_8_x(s0, filter, offset); + uint16x8_t d1 = highbd_convolve8_8_x(s1, filter, offset); + uint16x8_t d2 = highbd_convolve8_8_x(s2, filter, offset); + uint16x8_t d3 = highbd_convolve8_8_x(s3, filter, offset); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height > 4); + + // Process final 3 rows. + const int16_t *s = (const int16_t *)src; + do { + int16x8_t s0[8], s1[8], s2[8]; + load_s16_8x8(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3], &s0[4], + &s0[5], &s0[6], &s0[7]); + load_s16_8x8(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3], &s1[4], + &s1[5], &s1[6], &s1[7]); + load_s16_8x8(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3], &s2[4], + &s2[5], &s2[6], &s2[7]); + + uint16x8_t d0 = highbd_convolve8_8_x(s0, filter, offset); + uint16x8_t d1 = highbd_convolve8_8_x(s1, filter, offset); + uint16x8_t d2 = highbd_convolve8_8_x(s2, filter, offset); + + store_u16_8x3(dst, dst_stride, d0, d1, d2); + s += 8; + dst += 8; + width -= 8; + } while (width != 0); +} + +static INLINE void highbd_12_dist_wtd_convolve_2d_horiz_4tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *x_filter_ptr) { + const int64x2_t offset = vdupq_n_s64(1 << (12 + FILTER_BITS - 1)); + const int16x4_t x_filter = vld1_s16(x_filter_ptr + 2); + const int16x8_t filter = vcombine_s16(x_filter, vdup_n_s16(0)); + + // We are only doing 8-tap and 4-tap vertical convolutions, therefore we know + // that im_h % 4 = 3, so we can do the loop across the whole block 4 rows at + // a time and then process the last 3 rows separately. + + if (width == 4) { + uint16x8x2_t permute_tbl = vld1q_u16_x2(kDotProdTbl); + + const int16_t *s = (const int16_t *)(src); + + do { + int16x8_t s0, s1, s2, s3; + load_s16_8x4(s, src_stride, &s0, &s1, &s2, &s3); + + uint16x4_t d0 = highbd_12_convolve4_4_x(s0, filter, offset, permute_tbl); + uint16x4_t d1 = highbd_12_convolve4_4_x(s1, filter, offset, permute_tbl); + uint16x4_t d2 = highbd_12_convolve4_4_x(s2, filter, offset, permute_tbl); + uint16x4_t d3 = highbd_12_convolve4_4_x(s3, filter, offset, permute_tbl); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height > 4); + + // Process final 3 rows. + int16x8_t s0, s1, s2; + load_s16_8x3(s, src_stride, &s0, &s1, &s2); + + uint16x4_t d0 = highbd_12_convolve4_4_x(s0, filter, offset, permute_tbl); + uint16x4_t d1 = highbd_12_convolve4_4_x(s1, filter, offset, permute_tbl); + uint16x4_t d2 = highbd_12_convolve4_4_x(s2, filter, offset, permute_tbl); + + store_u16_4x3(dst, dst_stride, d0, d1, d2); + + } else { + uint16x8_t idx = vld1q_u16(kDeinterleaveTbl); + + do { + const int16_t *s = (const int16_t *)(src); + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0[4], s1[4], s2[4], s3[4]; + load_s16_8x4(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3]); + load_s16_8x4(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3]); + load_s16_8x4(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3]); + load_s16_8x4(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3]); + + uint16x8_t d0 = highbd_12_convolve4_8_x(s0, filter, offset, idx); + uint16x8_t d1 = highbd_12_convolve4_8_x(s1, filter, offset, idx); + uint16x8_t d2 = highbd_12_convolve4_8_x(s2, filter, offset, idx); + uint16x8_t d3 = highbd_12_convolve4_8_x(s3, filter, offset, idx); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height > 4); + + // Process final 3 rows. + const int16_t *s = (const int16_t *)(src); + + do { + int16x8_t s0[4], s1[4], s2[4]; + load_s16_8x4(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3]); + load_s16_8x4(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3]); + load_s16_8x4(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3]); + + uint16x8_t d0 = highbd_12_convolve4_8_x(s0, filter, offset, idx); + uint16x8_t d1 = highbd_12_convolve4_8_x(s1, filter, offset, idx); + uint16x8_t d2 = highbd_12_convolve4_8_x(s2, filter, offset, idx); + + store_u16_8x3(dst, dst_stride, d0, d1, d2); + + s += 8; + dst += 8; + width -= 8; + } while (width != 0); + } +} + +static INLINE void highbd_dist_wtd_convolve_2d_horiz_4tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *x_filter_ptr, const int bd) { + const int64x2_t offset = vdupq_n_s64(1 << (bd + FILTER_BITS - 1)); + const int16x4_t x_filter = vld1_s16(x_filter_ptr + 2); + const int16x8_t filter = vcombine_s16(x_filter, vdup_n_s16(0)); + + // We are only doing 8-tap and 4-tap vertical convolutions, therefore we know + // that im_h % 4 = 3, so we can do the loop across the whole block 4 rows at + // a time and then process the last 3 rows separately. + + if (width == 4) { + uint16x8x2_t permute_tbl = vld1q_u16_x2(kDotProdTbl); + + const int16_t *s = (const int16_t *)(src); + + do { + int16x8_t s0, s1, s2, s3; + load_s16_8x4(s, src_stride, &s0, &s1, &s2, &s3); + + uint16x4_t d0 = highbd_convolve4_4_x(s0, filter, offset, permute_tbl); + uint16x4_t d1 = highbd_convolve4_4_x(s1, filter, offset, permute_tbl); + uint16x4_t d2 = highbd_convolve4_4_x(s2, filter, offset, permute_tbl); + uint16x4_t d3 = highbd_convolve4_4_x(s3, filter, offset, permute_tbl); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height > 4); + + // Process final 3 rows. + int16x8_t s0, s1, s2; + load_s16_8x3(s, src_stride, &s0, &s1, &s2); + + uint16x4_t d0 = highbd_convolve4_4_x(s0, filter, offset, permute_tbl); + uint16x4_t d1 = highbd_convolve4_4_x(s1, filter, offset, permute_tbl); + uint16x4_t d2 = highbd_convolve4_4_x(s2, filter, offset, permute_tbl); + + store_u16_4x3(dst, dst_stride, d0, d1, d2); + } else { + uint16x8_t idx = vld1q_u16(kDeinterleaveTbl); + + do { + const int16_t *s = (const int16_t *)(src); + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0[4], s1[4], s2[4], s3[4]; + load_s16_8x4(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3]); + load_s16_8x4(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3]); + load_s16_8x4(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3]); + load_s16_8x4(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3]); + + uint16x8_t d0 = highbd_convolve4_8_x(s0, filter, offset, idx); + uint16x8_t d1 = highbd_convolve4_8_x(s1, filter, offset, idx); + uint16x8_t d2 = highbd_convolve4_8_x(s2, filter, offset, idx); + uint16x8_t d3 = highbd_convolve4_8_x(s3, filter, offset, idx); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height > 4); + + // Process final 3 rows. + const int16_t *s = (const int16_t *)(src); + + do { + int16x8_t s0[4], s1[4], s2[4]; + load_s16_8x4(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3]); + load_s16_8x4(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3]); + load_s16_8x4(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3]); + + uint16x8_t d0 = highbd_convolve4_8_x(s0, filter, offset, idx); + uint16x8_t d1 = highbd_convolve4_8_x(s1, filter, offset, idx); + uint16x8_t d2 = highbd_convolve4_8_x(s2, filter, offset, idx); + + store_u16_8x3(dst, dst_stride, d0, d1, d2); + + s += 8; + dst += 8; + width -= 8; + } while (width != 0); + } +} + +static INLINE uint16x4_t highbd_convolve8_4_2d_v(int16x8_t samples_lo[2], + int16x8_t samples_hi[2], + int16x8_t filter, + int64x2_t offset) { + int64x2_t sum01 = aom_svdot_lane_s16(offset, samples_lo[0], filter, 0); + sum01 = aom_svdot_lane_s16(sum01, samples_hi[0], filter, 1); + + int64x2_t sum23 = aom_svdot_lane_s16(offset, samples_lo[1], filter, 0); + sum23 = aom_svdot_lane_s16(sum23, samples_hi[1], filter, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + + return vqrshrun_n_s32(sum0123, COMPOUND_ROUND1_BITS); +} + +static INLINE uint16x8_t highbd_convolve8_8_2d_v(int16x8_t samples_lo[4], + int16x8_t samples_hi[4], + int16x8_t filter, + int64x2_t offset) { + int64x2_t sum01 = aom_svdot_lane_s16(offset, samples_lo[0], filter, 0); + sum01 = aom_svdot_lane_s16(sum01, samples_hi[0], filter, 1); + + int64x2_t sum23 = aom_svdot_lane_s16(offset, samples_lo[1], filter, 0); + sum23 = aom_svdot_lane_s16(sum23, samples_hi[1], filter, 1); + + int64x2_t sum45 = aom_svdot_lane_s16(offset, samples_lo[2], filter, 0); + sum45 = aom_svdot_lane_s16(sum45, samples_hi[2], filter, 1); + + int64x2_t sum67 = aom_svdot_lane_s16(offset, samples_lo[3], filter, 0); + sum67 = aom_svdot_lane_s16(sum67, samples_hi[3], filter, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum45), vmovn_s64(sum67)); + + return vcombine_u16(vqrshrun_n_s32(sum0123, COMPOUND_ROUND1_BITS), + vqrshrun_n_s32(sum4567, COMPOUND_ROUND1_BITS)); +} + +static INLINE void highbd_dist_wtd_convolve_2d_vert_8tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *y_filter_ptr, int offset) { + const int16x8_t y_filter = vld1q_s16(y_filter_ptr); + const int64x2_t offset_s64 = vdupq_n_s64(offset); + + uint16x8x3_t merge_block_tbl = vld1q_u16_x3(kDotProdMergeBlockTbl); + // Scale indices by size of the true vector length to avoid reading from an + // 'undefined' portion of a vector on a system with SVE vectors > 128-bit. + uint16x8_t correction0 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000000000000ULL)); + merge_block_tbl.val[0] = vaddq_u16(merge_block_tbl.val[0], correction0); + + uint16x8_t correction1 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100000000ULL)); + merge_block_tbl.val[1] = vaddq_u16(merge_block_tbl.val[1], correction1); + + uint16x8_t correction2 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100010000ULL)); + merge_block_tbl.val[2] = vaddq_u16(merge_block_tbl.val[2], correction2); + + if (width == 4) { + int16_t *s = (int16_t *)src; + int16x4_t s0, s1, s2, s3, s4, s5, s6; + load_s16_4x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); + s += 7 * src_stride; + + // This operation combines a conventional transpose and the sample permute + // required before computing the dot product. + int16x8_t s0123[2], s1234[2], s2345[2], s3456[2]; + transpose_concat_4x4(s0, s1, s2, s3, s0123); + transpose_concat_4x4(s1, s2, s3, s4, s1234); + transpose_concat_4x4(s2, s3, s4, s5, s2345); + transpose_concat_4x4(s3, s4, s5, s6, s3456); + + do { + int16x4_t s7, s8, s9, s10; + load_s16_4x4(s, src_stride, &s7, &s8, &s9, &s10); + + int16x8_t s4567[2], s5678[2], s6789[2], s789A[2]; + // Transpose and shuffle the 4 lines that were loaded. + transpose_concat_4x4(s7, s8, s9, s10, s789A); + + // Merge new data into block from previous iteration. + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[0], s4567); + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[1], s5678); + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[2], s6789); + + uint16x4_t d0 = + highbd_convolve8_4_2d_v(s0123, s4567, y_filter, offset_s64); + uint16x4_t d1 = + highbd_convolve8_4_2d_v(s1234, s5678, y_filter, offset_s64); + uint16x4_t d2 = + highbd_convolve8_4_2d_v(s2345, s6789, y_filter, offset_s64); + uint16x4_t d3 = + highbd_convolve8_4_2d_v(s3456, s789A, y_filter, offset_s64); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. + s0123[0] = s4567[0]; + s0123[1] = s4567[1]; + s1234[0] = s5678[0]; + s1234[1] = s5678[1]; + s2345[0] = s6789[0]; + s2345[1] = s6789[1]; + s3456[0] = s789A[0]; + s3456[1] = s789A[1]; + + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } else { + do { + int h = height; + int16_t *s = (int16_t *)src; + uint16_t *d = dst; + + int16x8_t s0, s1, s2, s3, s4, s5, s6; + load_s16_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); + s += 7 * src_stride; + + // This operation combines a conventional transpose and the sample permute + // required before computing the dot product. + int16x8_t s0123[4], s1234[4], s2345[4], s3456[4]; + transpose_concat_8x4(s0, s1, s2, s3, s0123); + transpose_concat_8x4(s1, s2, s3, s4, s1234); + transpose_concat_8x4(s2, s3, s4, s5, s2345); + transpose_concat_8x4(s3, s4, s5, s6, s3456); + + do { + int16x8_t s7, s8, s9, s10; + load_s16_8x4(s, src_stride, &s7, &s8, &s9, &s10); + int16x8_t s4567[4], s5678[4], s6789[4], s789A[4]; + + // Transpose and shuffle the 4 lines that were loaded. + transpose_concat_8x4(s7, s8, s9, s10, s789A); + + // Merge new data into block from previous iteration. + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[0], s4567); + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[1], s5678); + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[2], s6789); + + uint16x8_t d0 = + highbd_convolve8_8_2d_v(s0123, s4567, y_filter, offset_s64); + uint16x8_t d1 = + highbd_convolve8_8_2d_v(s1234, s5678, y_filter, offset_s64); + uint16x8_t d2 = + highbd_convolve8_8_2d_v(s2345, s6789, y_filter, offset_s64); + uint16x8_t d3 = + highbd_convolve8_8_2d_v(s3456, s789A, y_filter, offset_s64); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. + s0123[0] = s4567[0]; + s0123[1] = s4567[1]; + s0123[2] = s4567[2]; + s0123[3] = s4567[3]; + s1234[0] = s5678[0]; + s1234[1] = s5678[1]; + s1234[2] = s5678[2]; + s1234[3] = s5678[3]; + s2345[0] = s6789[0]; + s2345[1] = s6789[1]; + s2345[2] = s6789[2]; + s2345[3] = s6789[3]; + s3456[0] = s789A[0]; + s3456[1] = s789A[1]; + s3456[2] = s789A[2]; + s3456[3] = s789A[3]; + + s += 4 * src_stride; + d += 4 * dst_stride; + h -= 4; + } while (h != 0); + src += 8; + dst += 8; + width -= 8; + } while (width != 0); + } +} + +static INLINE uint16x4_t highbd_convolve4_4_2d_v( + const int16x4_t s0, const int16x4_t s1, const int16x4_t s2, + const int16x4_t s3, const int16x4_t filter, const int32x4_t offset) { + int32x4_t sum = vmlal_lane_s16(offset, s0, filter, 0); + sum = vmlal_lane_s16(sum, s1, filter, 1); + sum = vmlal_lane_s16(sum, s2, filter, 2); + sum = vmlal_lane_s16(sum, s3, filter, 3); + + return vqrshrun_n_s32(sum, COMPOUND_ROUND1_BITS); +} + +static INLINE uint16x8_t highbd_convolve4_8_2d_v( + const int16x8_t s0, const int16x8_t s1, const int16x8_t s2, + const int16x8_t s3, const int16x4_t filter, const int32x4_t offset) { + int32x4_t sum0 = vmlal_lane_s16(offset, vget_low_s16(s0), filter, 0); + sum0 = vmlal_lane_s16(sum0, vget_low_s16(s1), filter, 1); + sum0 = vmlal_lane_s16(sum0, vget_low_s16(s2), filter, 2); + sum0 = vmlal_lane_s16(sum0, vget_low_s16(s3), filter, 3); + + int32x4_t sum1 = vmlal_lane_s16(offset, vget_high_s16(s0), filter, 0); + sum1 = vmlal_lane_s16(sum1, vget_high_s16(s1), filter, 1); + sum1 = vmlal_lane_s16(sum1, vget_high_s16(s2), filter, 2); + sum1 = vmlal_lane_s16(sum1, vget_high_s16(s3), filter, 3); + + return vcombine_u16(vqrshrun_n_s32(sum0, COMPOUND_ROUND1_BITS), + vqrshrun_n_s32(sum1, COMPOUND_ROUND1_BITS)); +} + +static INLINE void highbd_dist_wtd_convolve_2d_vert_4tap_neon( + const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, + int w, int h, const int16_t *y_filter_ptr, const int offset) { + const int16x4_t y_filter = vld1_s16(y_filter_ptr + 2); + const int32x4_t offset_vec = vdupq_n_s32(offset); + + if (w == 4) { + const int16_t *s = (const int16_t *)src_ptr; + uint16_t *d = dst_ptr; + + int16x4_t s0, s1, s2; + load_s16_4x3(s, src_stride, &s0, &s1, &s2); + s += 3 * src_stride; + + do { + int16x4_t s3, s4, s5, s6; + load_s16_4x4(s, src_stride, &s3, &s4, &s5, &s6); + + uint16x4_t d0 = + highbd_convolve4_4_2d_v(s0, s1, s2, s3, y_filter, offset_vec); + uint16x4_t d1 = + highbd_convolve4_4_2d_v(s1, s2, s3, s4, y_filter, offset_vec); + uint16x4_t d2 = + highbd_convolve4_4_2d_v(s2, s3, s4, s5, y_filter, offset_vec); + uint16x4_t d3 = + highbd_convolve4_4_2d_v(s3, s4, s5, s6, y_filter, offset_vec); + + store_u16_4x4(d, dst_stride, d0, d1, d2, d3); + + s0 = s4; + s1 = s5; + s2 = s6; + + s += 4 * src_stride; + d += 4 * dst_stride; + h -= 4; + } while (h != 0); + } else { + do { + int height = h; + const int16_t *s = (const int16_t *)src_ptr; + uint16_t *d = dst_ptr; + + int16x8_t s0, s1, s2; + load_s16_8x3(s, src_stride, &s0, &s1, &s2); + s += 3 * src_stride; + + do { + int16x8_t s3, s4, s5, s6; + load_s16_8x4(s, src_stride, &s3, &s4, &s5, &s6); + + uint16x8_t d0 = + highbd_convolve4_8_2d_v(s0, s1, s2, s3, y_filter, offset_vec); + uint16x8_t d1 = + highbd_convolve4_8_2d_v(s1, s2, s3, s4, y_filter, offset_vec); + uint16x8_t d2 = + highbd_convolve4_8_2d_v(s2, s3, s4, s5, y_filter, offset_vec); + uint16x8_t d3 = + highbd_convolve4_8_2d_v(s3, s4, s5, s6, y_filter, offset_vec); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s0 = s4; + s1 = s5; + s2 = s6; + + s += 4 * src_stride; + d += 4 * dst_stride; + height -= 4; + } while (height != 0); + src_ptr += 8; + dst_ptr += 8; + w -= 8; + } while (w != 0); + } +} + +void av1_highbd_dist_wtd_convolve_2d_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, + int h, const InterpFilterParams *filter_params_x, + const InterpFilterParams *filter_params_y, const int subpel_x_qn, + const int subpel_y_qn, ConvolveParams *conv_params, int bd) { + DECLARE_ALIGNED(16, uint16_t, + im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]); + DECLARE_ALIGNED(16, uint16_t, + im_block2[(MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]); + + CONV_BUF_TYPE *dst16 = conv_params->dst; + int dst16_stride = conv_params->dst_stride; + const int x_filter_taps = get_filter_tap(filter_params_x, subpel_x_qn); + const int clamped_x_taps = x_filter_taps < 4 ? 4 : x_filter_taps; + + const int y_filter_taps = get_filter_tap(filter_params_y, subpel_y_qn); + const int clamped_y_taps = y_filter_taps < 4 ? 4 : y_filter_taps; + + if (x_filter_taps == 6 || y_filter_taps == 6) { + av1_highbd_dist_wtd_convolve_2d_neon( + src, src_stride, dst, dst_stride, w, h, filter_params_x, + filter_params_y, subpel_x_qn, subpel_y_qn, conv_params, bd); + return; + } + + const int im_h = h + clamped_y_taps - 1; + const int im_stride = MAX_SB_SIZE; + const int vert_offset = clamped_y_taps / 2 - 1; + const int horiz_offset = clamped_x_taps / 2 - 1; + const int y_offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0; + const int round_offset_conv_y = (1 << y_offset_bits); + + const uint16_t *src_ptr = src - vert_offset * src_stride - horiz_offset; + + const int16_t *x_filter_ptr = av1_get_interp_filter_subpel_kernel( + filter_params_x, subpel_x_qn & SUBPEL_MASK); + const int16_t *y_filter_ptr = av1_get_interp_filter_subpel_kernel( + filter_params_y, subpel_y_qn & SUBPEL_MASK); + + if (bd == 12) { + if (x_filter_taps <= 4) { + highbd_12_dist_wtd_convolve_2d_horiz_4tap_sve2( + src_ptr, src_stride, im_block, im_stride, w, im_h, x_filter_ptr); + } else { + highbd_12_dist_wtd_convolve_2d_horiz_8tap_sve2( + src_ptr, src_stride, im_block, im_stride, w, im_h, x_filter_ptr); + } + } else { + if (x_filter_taps <= 4) { + highbd_dist_wtd_convolve_2d_horiz_4tap_sve2( + src_ptr, src_stride, im_block, im_stride, w, im_h, x_filter_ptr, bd); + } else { + highbd_dist_wtd_convolve_2d_horiz_8tap_sve2( + src_ptr, src_stride, im_block, im_stride, w, im_h, x_filter_ptr, bd); + } + } + + if (conv_params->do_average) { + if (y_filter_taps <= 4) { + highbd_dist_wtd_convolve_2d_vert_4tap_neon(im_block, im_stride, im_block2, + im_stride, w, h, y_filter_ptr, + round_offset_conv_y); + } else { + highbd_dist_wtd_convolve_2d_vert_8tap_sve2(im_block, im_stride, im_block2, + im_stride, w, h, y_filter_ptr, + round_offset_conv_y); + } + if (conv_params->use_dist_wtd_comp_avg) { + if (bd == 12) { + highbd_12_dist_wtd_comp_avg_neon(im_block2, im_stride, dst, dst_stride, + w, h, conv_params); + + } else { + highbd_dist_wtd_comp_avg_neon(im_block2, im_stride, dst, dst_stride, w, + h, conv_params, bd); + } + } else { + if (bd == 12) { + highbd_12_comp_avg_neon(im_block2, im_stride, dst, dst_stride, w, h, + conv_params); + + } else { + highbd_comp_avg_neon(im_block2, im_stride, dst, dst_stride, w, h, + conv_params, bd); + } + } + } else { + if (y_filter_taps <= 4) { + highbd_dist_wtd_convolve_2d_vert_4tap_neon( + im_block, im_stride, dst16, dst16_stride, w, h, y_filter_ptr, + round_offset_conv_y); + } else { + highbd_dist_wtd_convolve_2d_vert_8tap_sve2( + im_block, im_stride, dst16, dst16_stride, w, h, y_filter_ptr, + round_offset_conv_y); + } + } +} diff --git a/third_party/aom/av1/common/arm/highbd_convolve_sve2.c b/third_party/aom/av1/common/arm/highbd_convolve_sve2.c new file mode 100644 index 0000000000..82eb12fcea --- /dev/null +++ b/third_party/aom/av1/common/arm/highbd_convolve_sve2.c @@ -0,0 +1,1720 @@ +/* + * Copyright (c) 2024, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <assert.h> +#include <arm_neon.h> + +#include "config/aom_config.h" +#include "config/av1_rtcd.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/arm/aom_neon_sve_bridge.h" +#include "aom_dsp/arm/aom_neon_sve2_bridge.h" +#include "aom_dsp/arm/mem_neon.h" +#include "aom_ports/mem.h" +#include "av1/common/convolve.h" +#include "av1/common/filter.h" +#include "av1/common/arm/highbd_convolve_sve2.h" + +DECLARE_ALIGNED(16, static const uint16_t, kDotProdTbl[32]) = { + 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6, + 4, 5, 6, 7, 5, 6, 7, 0, 6, 7, 0, 1, 7, 0, 1, 2, +}; + +static INLINE uint16x4_t convolve12_4_x( + int16x8_t s0, int16x8_t s1, int16x8_t filter_0_7, int16x8_t filter_4_11, + const int64x2_t offset, uint16x8x4_t permute_tbl, uint16x4_t max) { + int16x8_t permuted_samples[6]; + permuted_samples[0] = aom_tbl_s16(s0, permute_tbl.val[0]); + permuted_samples[1] = aom_tbl_s16(s0, permute_tbl.val[1]); + permuted_samples[2] = aom_tbl2_s16(s0, s1, permute_tbl.val[2]); + permuted_samples[3] = aom_tbl2_s16(s0, s1, permute_tbl.val[3]); + permuted_samples[4] = aom_tbl_s16(s1, permute_tbl.val[0]); + permuted_samples[5] = aom_tbl_s16(s1, permute_tbl.val[1]); + + int64x2_t sum01 = + aom_svdot_lane_s16(offset, permuted_samples[0], filter_0_7, 0); + sum01 = aom_svdot_lane_s16(sum01, permuted_samples[2], filter_0_7, 1); + sum01 = aom_svdot_lane_s16(sum01, permuted_samples[4], filter_4_11, 1); + + int64x2_t sum23 = + aom_svdot_lane_s16(offset, permuted_samples[1], filter_0_7, 0); + sum23 = aom_svdot_lane_s16(sum23, permuted_samples[3], filter_0_7, 1); + sum23 = aom_svdot_lane_s16(sum23, permuted_samples[5], filter_4_11, 1); + + int32x4_t res0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + uint16x4_t res = vqrshrun_n_s32(res0123, FILTER_BITS); + + return vmin_u16(res, max); +} + +static INLINE uint16x8_t convolve12_8_x(int16x8_t s0, int16x8_t s1, + int16x8_t s2, int16x8_t filter_0_7, + int16x8_t filter_4_11, int64x2_t offset, + uint16x8x4_t permute_tbl, + uint16x8_t max) { + int16x8_t permuted_samples[8]; + permuted_samples[0] = aom_tbl_s16(s0, permute_tbl.val[0]); + permuted_samples[1] = aom_tbl_s16(s0, permute_tbl.val[1]); + permuted_samples[2] = aom_tbl2_s16(s0, s1, permute_tbl.val[2]); + permuted_samples[3] = aom_tbl2_s16(s0, s1, permute_tbl.val[3]); + permuted_samples[4] = aom_tbl_s16(s1, permute_tbl.val[0]); + permuted_samples[5] = aom_tbl_s16(s1, permute_tbl.val[1]); + permuted_samples[6] = aom_tbl2_s16(s1, s2, permute_tbl.val[2]); + permuted_samples[7] = aom_tbl2_s16(s1, s2, permute_tbl.val[3]); + + int64x2_t sum01 = + aom_svdot_lane_s16(offset, permuted_samples[0], filter_0_7, 0); + sum01 = aom_svdot_lane_s16(sum01, permuted_samples[2], filter_0_7, 1); + sum01 = aom_svdot_lane_s16(sum01, permuted_samples[4], filter_4_11, 1); + + int64x2_t sum23 = + aom_svdot_lane_s16(offset, permuted_samples[1], filter_0_7, 0); + sum23 = aom_svdot_lane_s16(sum23, permuted_samples[3], filter_0_7, 1); + sum23 = aom_svdot_lane_s16(sum23, permuted_samples[5], filter_4_11, 1); + + int64x2_t sum45 = + aom_svdot_lane_s16(offset, permuted_samples[2], filter_0_7, 0); + sum45 = aom_svdot_lane_s16(sum45, permuted_samples[4], filter_0_7, 1); + sum45 = aom_svdot_lane_s16(sum45, permuted_samples[6], filter_4_11, 1); + + int64x2_t sum67 = + aom_svdot_lane_s16(offset, permuted_samples[3], filter_0_7, 0); + sum67 = aom_svdot_lane_s16(sum67, permuted_samples[5], filter_0_7, 1); + sum67 = aom_svdot_lane_s16(sum67, permuted_samples[7], filter_4_11, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum45), vmovn_s64(sum67)); + + uint16x8_t res = vcombine_u16(vqrshrun_n_s32(sum0123, FILTER_BITS), + vqrshrun_n_s32(sum4567, FILTER_BITS)); + + return vminq_u16(res, max); +} + +static INLINE void highbd_convolve_x_sr_12tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *y_filter_ptr, + ConvolveParams *conv_params, int bd) { + // This shim allows to do only one rounding shift instead of two. + const int64x2_t offset = vdupq_n_s64(1 << (conv_params->round_0 - 1)); + + const int16x8_t y_filter_0_7 = vld1q_s16(y_filter_ptr); + const int16x8_t y_filter_4_11 = vld1q_s16(y_filter_ptr + 4); + + uint16x8x4_t permute_tbl = vld1q_u16_x4(kDotProdTbl); + // Scale indices by size of the true vector length to avoid reading from an + // 'undefined' portion of a vector on a system with SVE vectors > 128-bit. + uint16x8_t correction0 = vreinterpretq_u16_u64(vcombine_u64( + vdup_n_u64(0), vdup_n_u64(svcnth() * 0x0001000000000000ULL))); + permute_tbl.val[2] = vaddq_u16(permute_tbl.val[2], correction0); + + uint16x8_t correction1 = vreinterpretq_u16_u64( + vcombine_u64(vdup_n_u64(svcnth() * 0x0001000100000000ULL), + vdup_n_u64(svcnth() * 0x0001000100010000ULL))); + permute_tbl.val[3] = vaddq_u16(permute_tbl.val[3], correction1); + + if (width == 4) { + const uint16x4_t max = vdup_n_u16((1 << bd) - 1); + const int16_t *s = (const int16_t *)src; + + do { + int16x8_t s0, s1, s2, s3, s4, s5, s6, s7; + load_s16_8x4(s, src_stride, &s0, &s2, &s4, &s6); + load_s16_8x4(s + 8, src_stride, &s1, &s3, &s5, &s7); + + uint16x4_t d0 = convolve12_4_x(s0, s1, y_filter_0_7, y_filter_4_11, + offset, permute_tbl, max); + uint16x4_t d1 = convolve12_4_x(s2, s3, y_filter_0_7, y_filter_4_11, + offset, permute_tbl, max); + uint16x4_t d2 = convolve12_4_x(s4, s5, y_filter_0_7, y_filter_4_11, + offset, permute_tbl, max); + uint16x4_t d3 = convolve12_4_x(s6, s7, y_filter_0_7, y_filter_4_11, + offset, permute_tbl, max); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } else { + const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); + + do { + const int16_t *s = (const int16_t *)src; + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11; + load_s16_8x4(s, src_stride, &s0, &s3, &s6, &s9); + load_s16_8x4(s + 8, src_stride, &s1, &s4, &s7, &s10); + load_s16_8x4(s + 16, src_stride, &s2, &s5, &s8, &s11); + + uint16x8_t d0 = convolve12_8_x(s0, s1, s2, y_filter_0_7, y_filter_4_11, + offset, permute_tbl, max); + uint16x8_t d1 = convolve12_8_x(s3, s4, s5, y_filter_0_7, y_filter_4_11, + offset, permute_tbl, max); + uint16x8_t d2 = convolve12_8_x(s6, s7, s8, y_filter_0_7, y_filter_4_11, + offset, permute_tbl, max); + uint16x8_t d3 = convolve12_8_x(s9, s10, s11, y_filter_0_7, + y_filter_4_11, offset, permute_tbl, max); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } +} + +static INLINE uint16x8_t convolve8_8_x(int16x8_t s0[8], int16x8_t filter, + int64x2_t offset, uint16x8_t max) { + int64x2_t sum[8]; + sum[0] = aom_sdotq_s16(offset, s0[0], filter); + sum[1] = aom_sdotq_s16(offset, s0[1], filter); + sum[2] = aom_sdotq_s16(offset, s0[2], filter); + sum[3] = aom_sdotq_s16(offset, s0[3], filter); + sum[4] = aom_sdotq_s16(offset, s0[4], filter); + sum[5] = aom_sdotq_s16(offset, s0[5], filter); + sum[6] = aom_sdotq_s16(offset, s0[6], filter); + sum[7] = aom_sdotq_s16(offset, s0[7], filter); + + sum[0] = vpaddq_s64(sum[0], sum[1]); + sum[2] = vpaddq_s64(sum[2], sum[3]); + sum[4] = vpaddq_s64(sum[4], sum[5]); + sum[6] = vpaddq_s64(sum[6], sum[7]); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum[0]), vmovn_s64(sum[2])); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum[4]), vmovn_s64(sum[6])); + + uint16x8_t res = vcombine_u16(vqrshrun_n_s32(sum0123, FILTER_BITS), + vqrshrun_n_s32(sum4567, FILTER_BITS)); + + return vminq_u16(res, max); +} + +static INLINE void highbd_convolve_x_sr_8tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *y_filter_ptr, + ConvolveParams *conv_params, int bd) { + const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); + // This shim allows to do only one rounding shift instead of two. + const int64_t offset = 1 << (conv_params->round_0 - 1); + const int64x2_t offset_lo = vcombine_s64((int64x1_t)(offset), vdup_n_s64(0)); + + const int16x8_t filter = vld1q_s16(y_filter_ptr); + + do { + const int16_t *s = (const int16_t *)src; + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0[8], s1[8], s2[8], s3[8]; + load_s16_8x8(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3], + &s0[4], &s0[5], &s0[6], &s0[7]); + load_s16_8x8(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3], + &s1[4], &s1[5], &s1[6], &s1[7]); + load_s16_8x8(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3], + &s2[4], &s2[5], &s2[6], &s2[7]); + load_s16_8x8(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3], + &s3[4], &s3[5], &s3[6], &s3[7]); + + uint16x8_t d0 = convolve8_8_x(s0, filter, offset_lo, max); + uint16x8_t d1 = convolve8_8_x(s1, filter, offset_lo, max); + uint16x8_t d2 = convolve8_8_x(s2, filter, offset_lo, max); + uint16x8_t d3 = convolve8_8_x(s3, filter, offset_lo, max); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); +} + +// clang-format off +DECLARE_ALIGNED(16, static const uint16_t, kDeinterleaveTbl[8]) = { + 0, 2, 4, 6, 1, 3, 5, 7, +}; +// clang-format on + +static INLINE uint16x4_t convolve4_4_x(int16x8_t s0, int16x8_t filter, + int64x2_t offset, + uint16x8x2_t permute_tbl, + uint16x4_t max) { + int16x8_t permuted_samples0 = aom_tbl_s16(s0, permute_tbl.val[0]); + int16x8_t permuted_samples1 = aom_tbl_s16(s0, permute_tbl.val[1]); + + int64x2_t sum01 = aom_svdot_lane_s16(offset, permuted_samples0, filter, 0); + int64x2_t sum23 = aom_svdot_lane_s16(offset, permuted_samples1, filter, 0); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + uint16x4_t res = vqrshrun_n_s32(sum0123, FILTER_BITS); + + return vmin_u16(res, max); +} + +static INLINE uint16x8_t convolve4_8_x(int16x8_t s0[4], int16x8_t filter, + int64x2_t offset, uint16x8_t tbl, + uint16x8_t max) { + int64x2_t sum04 = aom_svdot_lane_s16(offset, s0[0], filter, 0); + int64x2_t sum15 = aom_svdot_lane_s16(offset, s0[1], filter, 0); + int64x2_t sum26 = aom_svdot_lane_s16(offset, s0[2], filter, 0); + int64x2_t sum37 = aom_svdot_lane_s16(offset, s0[3], filter, 0); + + int32x4_t sum0415 = vcombine_s32(vmovn_s64(sum04), vmovn_s64(sum15)); + int32x4_t sum2637 = vcombine_s32(vmovn_s64(sum26), vmovn_s64(sum37)); + + uint16x8_t res = vcombine_u16(vqrshrun_n_s32(sum0415, FILTER_BITS), + vqrshrun_n_s32(sum2637, FILTER_BITS)); + res = aom_tbl_u16(res, tbl); + + return vminq_u16(res, max); +} + +static INLINE void highbd_convolve_x_sr_4tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *x_filter_ptr, + ConvolveParams *conv_params, int bd) { + // This shim allows to do only one rounding shift instead of two. + const int64x2_t offset = vdupq_n_s64(1 << (conv_params->round_0 - 1)); + + const int16x4_t x_filter = vld1_s16(x_filter_ptr + 2); + const int16x8_t filter = vcombine_s16(x_filter, vdup_n_s16(0)); + + if (width == 4) { + const uint16x4_t max = vdup_n_u16((1 << bd) - 1); + uint16x8x2_t permute_tbl = vld1q_u16_x2(kDotProdTbl); + + const int16_t *s = (const int16_t *)(src); + + do { + int16x8_t s0, s1, s2, s3; + load_s16_8x4(s, src_stride, &s0, &s1, &s2, &s3); + + uint16x4_t d0 = convolve4_4_x(s0, filter, offset, permute_tbl, max); + uint16x4_t d1 = convolve4_4_x(s1, filter, offset, permute_tbl, max); + uint16x4_t d2 = convolve4_4_x(s2, filter, offset, permute_tbl, max); + uint16x4_t d3 = convolve4_4_x(s3, filter, offset, permute_tbl, max); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } else { + const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); + uint16x8_t idx = vld1q_u16(kDeinterleaveTbl); + + do { + const int16_t *s = (const int16_t *)(src); + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0[4], s1[4], s2[4], s3[4]; + load_s16_8x4(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3]); + load_s16_8x4(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3]); + load_s16_8x4(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3]); + load_s16_8x4(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3]); + + uint16x8_t d0 = convolve4_8_x(s0, filter, offset, idx, max); + uint16x8_t d1 = convolve4_8_x(s1, filter, offset, idx, max); + uint16x8_t d2 = convolve4_8_x(s2, filter, offset, idx, max); + uint16x8_t d3 = convolve4_8_x(s3, filter, offset, idx, max); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } +} + +void av1_highbd_convolve_x_sr_sve2(const uint16_t *src, int src_stride, + uint16_t *dst, int dst_stride, int w, int h, + const InterpFilterParams *filter_params_x, + const int subpel_x_qn, + ConvolveParams *conv_params, int bd) { + if (w == 2 || h == 2) { + av1_highbd_convolve_x_sr_c(src, src_stride, dst, dst_stride, w, h, + filter_params_x, subpel_x_qn, conv_params, bd); + return; + } + + const int x_filter_taps = get_filter_tap(filter_params_x, subpel_x_qn); + + if (x_filter_taps == 6) { + av1_highbd_convolve_x_sr_neon(src, src_stride, dst, dst_stride, w, h, + filter_params_x, subpel_x_qn, conv_params, + bd); + return; + } + + const int horiz_offset = filter_params_x->taps / 2 - 1; + const int16_t *x_filter_ptr = av1_get_interp_filter_subpel_kernel( + filter_params_x, subpel_x_qn & SUBPEL_MASK); + + src -= horiz_offset; + + if (x_filter_taps == 12) { + highbd_convolve_x_sr_12tap_sve2(src, src_stride, dst, dst_stride, w, h, + x_filter_ptr, conv_params, bd); + return; + } + + if (x_filter_taps == 8) { + highbd_convolve_x_sr_8tap_sve2(src, src_stride, dst, dst_stride, w, h, + x_filter_ptr, conv_params, bd); + return; + } + + highbd_convolve_x_sr_4tap_sve2(src + 2, src_stride, dst, dst_stride, w, h, + x_filter_ptr, conv_params, bd); +} + +static INLINE uint16x4_t highbd_convolve12_4_y(int16x8_t s0[2], int16x8_t s1[2], + int16x8_t s2[2], + int16x8_t filter_0_7, + int16x8_t filter_4_11, + uint16x4_t max) { + int64x2_t sum[2]; + + sum[0] = aom_svdot_lane_s16(vdupq_n_s64(0), s0[0], filter_0_7, 0); + sum[0] = aom_svdot_lane_s16(sum[0], s1[0], filter_0_7, 1); + sum[0] = aom_svdot_lane_s16(sum[0], s2[0], filter_4_11, 1); + + sum[1] = aom_svdot_lane_s16(vdupq_n_s64(0), s0[1], filter_0_7, 0); + sum[1] = aom_svdot_lane_s16(sum[1], s1[1], filter_0_7, 1); + sum[1] = aom_svdot_lane_s16(sum[1], s2[1], filter_4_11, 1); + + int32x4_t res_s32 = vcombine_s32(vmovn_s64(sum[0]), vmovn_s64(sum[1])); + + uint16x4_t res = vqrshrun_n_s32(res_s32, FILTER_BITS); + + return vmin_u16(res, max); +} + +static INLINE void highbd_convolve_y_sr_12tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *y_filter_ptr, int bd) { + const int16x8_t y_filter_0_7 = vld1q_s16(y_filter_ptr); + const int16x8_t y_filter_4_11 = vld1q_s16(y_filter_ptr + 4); + + uint16x8x3_t merge_block_tbl = vld1q_u16_x3(kDotProdMergeBlockTbl); + // Scale indices by size of the true vector length to avoid reading from an + // 'undefined' portion of a vector on a system with SVE vectors > 128-bit. + uint16x8_t correction0 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000000000000ULL)); + merge_block_tbl.val[0] = vaddq_u16(merge_block_tbl.val[0], correction0); + + uint16x8_t correction1 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100000000ULL)); + merge_block_tbl.val[1] = vaddq_u16(merge_block_tbl.val[1], correction1); + + uint16x8_t correction2 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100010000ULL)); + merge_block_tbl.val[2] = vaddq_u16(merge_block_tbl.val[2], correction2); + + const uint16x4_t max = vdup_n_u16((1 << bd) - 1); + + do { + int16_t *s = (int16_t *)src; + uint16_t *d = dst; + int h = height; + + int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, sA; + load_s16_4x11(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7, &s8, + &s9, &sA); + s += 11 * src_stride; + + int16x8_t s0123[2], s1234[2], s2345[2], s3456[2], s4567[2], s5678[2], + s6789[2], s789A[2]; + transpose_concat_4x4(s0, s1, s2, s3, s0123); + transpose_concat_4x4(s1, s2, s3, s4, s1234); + transpose_concat_4x4(s2, s3, s4, s5, s2345); + transpose_concat_4x4(s3, s4, s5, s6, s3456); + transpose_concat_4x4(s4, s5, s6, s7, s4567); + transpose_concat_4x4(s5, s6, s7, s8, s5678); + transpose_concat_4x4(s6, s7, s8, s9, s6789); + transpose_concat_4x4(s7, s8, s9, sA, s789A); + + do { + int16x4_t sB, sC, sD, sE; + load_s16_4x4(s, src_stride, &sB, &sC, &sD, &sE); + + int16x8_t s89AB[2], s9ABC[2], sABCD[2], sBCDE[2]; + transpose_concat_4x4(sB, sC, sD, sE, sBCDE); + + // Use the above transpose and reuse data from the previous loop to get + // the rest. + aom_tbl2x2_s16(s789A, sBCDE, merge_block_tbl.val[0], s89AB); + aom_tbl2x2_s16(s789A, sBCDE, merge_block_tbl.val[1], s9ABC); + aom_tbl2x2_s16(s789A, sBCDE, merge_block_tbl.val[2], sABCD); + + uint16x4_t d0 = highbd_convolve12_4_y(s0123, s4567, s89AB, y_filter_0_7, + y_filter_4_11, max); + uint16x4_t d1 = highbd_convolve12_4_y(s1234, s5678, s9ABC, y_filter_0_7, + y_filter_4_11, max); + uint16x4_t d2 = highbd_convolve12_4_y(s2345, s6789, sABCD, y_filter_0_7, + y_filter_4_11, max); + uint16x4_t d3 = highbd_convolve12_4_y(s3456, s789A, sBCDE, y_filter_0_7, + y_filter_4_11, max); + + store_u16_4x4(d, dst_stride, d0, d1, d2, d3); + + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. + s0123[0] = s4567[0]; + s0123[1] = s4567[1]; + s1234[0] = s5678[0]; + s1234[1] = s5678[1]; + s2345[0] = s6789[0]; + s2345[1] = s6789[1]; + s3456[0] = s789A[0]; + s3456[1] = s789A[1]; + s4567[0] = s89AB[0]; + s4567[1] = s89AB[1]; + s5678[0] = s9ABC[0]; + s5678[1] = s9ABC[1]; + s6789[0] = sABCD[0]; + s6789[1] = sABCD[1]; + s789A[0] = sBCDE[0]; + s789A[1] = sBCDE[1]; + + s += 4 * src_stride; + d += 4 * dst_stride; + h -= 4; + } while (h != 0); + src += 4; + dst += 4; + width -= 4; + } while (width != 0); +} + +static INLINE uint16x4_t highbd_convolve8_4_y(int16x8_t samples_lo[2], + int16x8_t samples_hi[2], + int16x8_t filter, + uint16x4_t max) { + int64x2_t sum01 = + aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[0], filter, 0); + sum01 = aom_svdot_lane_s16(sum01, samples_hi[0], filter, 1); + + int64x2_t sum23 = + aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[1], filter, 0); + sum23 = aom_svdot_lane_s16(sum23, samples_hi[1], filter, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + uint16x4_t res = vqrshrun_n_s32(sum0123, FILTER_BITS); + return vmin_u16(res, max); +} + +static INLINE uint16x8_t highbd_convolve8_8_y(int16x8_t samples_lo[4], + int16x8_t samples_hi[4], + int16x8_t filter, + uint16x8_t max) { + int64x2_t sum01 = + aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[0], filter, 0); + sum01 = aom_svdot_lane_s16(sum01, samples_hi[0], filter, 1); + + int64x2_t sum23 = + aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[1], filter, 0); + sum23 = aom_svdot_lane_s16(sum23, samples_hi[1], filter, 1); + + int64x2_t sum45 = + aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[2], filter, 0); + sum45 = aom_svdot_lane_s16(sum45, samples_hi[2], filter, 1); + + int64x2_t sum67 = + aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[3], filter, 0); + sum67 = aom_svdot_lane_s16(sum67, samples_hi[3], filter, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum45), vmovn_s64(sum67)); + uint16x8_t res = vcombine_u16(vqrshrun_n_s32(sum0123, FILTER_BITS), + vqrshrun_n_s32(sum4567, FILTER_BITS)); + return vminq_u16(res, max); +} + +void highbd_convolve_y_sr_8tap_sve2(const uint16_t *src, ptrdiff_t src_stride, + uint16_t *dst, ptrdiff_t dst_stride, + int width, int height, + const int16_t *filter_y, int bd) { + assert(w >= 4 && h >= 4); + + const int16x8_t y_filter = vld1q_s16(filter_y); + + uint16x8x3_t merge_block_tbl = vld1q_u16_x3(kDotProdMergeBlockTbl); + // Scale indices by size of the true vector length to avoid reading from an + // 'undefined' portion of a vector on a system with SVE vectors > 128-bit. + uint16x8_t correction0 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000000000000ULL)); + merge_block_tbl.val[0] = vaddq_u16(merge_block_tbl.val[0], correction0); + + uint16x8_t correction1 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100000000ULL)); + merge_block_tbl.val[1] = vaddq_u16(merge_block_tbl.val[1], correction1); + + uint16x8_t correction2 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100010000ULL)); + merge_block_tbl.val[2] = vaddq_u16(merge_block_tbl.val[2], correction2); + + if (width == 4) { + const uint16x4_t max = vdup_n_u16((1 << bd) - 1); + int16_t *s = (int16_t *)src; + + int16x4_t s0, s1, s2, s3, s4, s5, s6; + load_s16_4x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); + s += 7 * src_stride; + + // This operation combines a conventional transpose and the sample permute + // required before computing the dot product. + int16x8_t s0123[2], s1234[2], s2345[2], s3456[2]; + transpose_concat_4x4(s0, s1, s2, s3, s0123); + transpose_concat_4x4(s1, s2, s3, s4, s1234); + transpose_concat_4x4(s2, s3, s4, s5, s2345); + transpose_concat_4x4(s3, s4, s5, s6, s3456); + + do { + int16x4_t s7, s8, s9, s10; + load_s16_4x4(s, src_stride, &s7, &s8, &s9, &s10); + + int16x8_t s4567[2], s5678[2], s6789[2], s789A[2]; + // Transpose and shuffle the 4 lines that were loaded. + transpose_concat_4x4(s7, s8, s9, s10, s789A); + + // Merge new data into block from previous iteration. + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[0], s4567); + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[1], s5678); + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[2], s6789); + + uint16x4_t d0 = highbd_convolve8_4_y(s0123, s4567, y_filter, max); + uint16x4_t d1 = highbd_convolve8_4_y(s1234, s5678, y_filter, max); + uint16x4_t d2 = highbd_convolve8_4_y(s2345, s6789, y_filter, max); + uint16x4_t d3 = highbd_convolve8_4_y(s3456, s789A, y_filter, max); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. + s0123[0] = s4567[0]; + s0123[1] = s4567[1]; + s1234[0] = s5678[0]; + s1234[1] = s5678[1]; + s2345[0] = s6789[0]; + s2345[1] = s6789[1]; + s3456[0] = s789A[0]; + s3456[1] = s789A[1]; + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } else { + const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); + + do { + int h = height; + int16_t *s = (int16_t *)src; + uint16_t *d = dst; + + int16x8_t s0, s1, s2, s3, s4, s5, s6; + load_s16_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); + s += 7 * src_stride; + + // This operation combines a conventional transpose and the sample permute + // required before computing the dot product. + int16x8_t s0123[4], s1234[4], s2345[4], s3456[4]; + transpose_concat_8x4(s0, s1, s2, s3, s0123); + transpose_concat_8x4(s1, s2, s3, s4, s1234); + transpose_concat_8x4(s2, s3, s4, s5, s2345); + transpose_concat_8x4(s3, s4, s5, s6, s3456); + + do { + int16x8_t s7, s8, s9, s10; + load_s16_8x4(s, src_stride, &s7, &s8, &s9, &s10); + + int16x8_t s4567[4], s5678[4], s6789[4], s789A[4]; + // Transpose and shuffle the 4 lines that were loaded. + transpose_concat_8x4(s7, s8, s9, s10, s789A); + + // Merge new data into block from previous iteration. + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[0], s4567); + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[1], s5678); + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[2], s6789); + + uint16x8_t d0 = highbd_convolve8_8_y(s0123, s4567, y_filter, max); + uint16x8_t d1 = highbd_convolve8_8_y(s1234, s5678, y_filter, max); + uint16x8_t d2 = highbd_convolve8_8_y(s2345, s6789, y_filter, max); + uint16x8_t d3 = highbd_convolve8_8_y(s3456, s789A, y_filter, max); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. + s0123[0] = s4567[0]; + s0123[1] = s4567[1]; + s0123[2] = s4567[2]; + s0123[3] = s4567[3]; + s1234[0] = s5678[0]; + s1234[1] = s5678[1]; + s1234[2] = s5678[2]; + s1234[3] = s5678[3]; + s2345[0] = s6789[0]; + s2345[1] = s6789[1]; + s2345[2] = s6789[2]; + s2345[3] = s6789[3]; + s3456[0] = s789A[0]; + s3456[1] = s789A[1]; + s3456[2] = s789A[2]; + s3456[3] = s789A[3]; + + s += 4 * src_stride; + d += 4 * dst_stride; + h -= 4; + } while (h != 0); + src += 8; + dst += 8; + width -= 8; + } while (width != 0); + } +} + +static INLINE uint16x4_t highbd_convolve4_4_y(int16x8_t samples[2], + int16x8_t filter, + uint16x4_t max) { + int64x2_t sum01 = aom_svdot_lane_s16(vdupq_n_s64(0), samples[0], filter, 0); + int64x2_t sum23 = aom_svdot_lane_s16(vdupq_n_s64(0), samples[1], filter, 0); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + uint16x4_t res = vqrshrun_n_s32(sum0123, FILTER_BITS); + return vmin_u16(res, max); +} + +static INLINE uint16x8_t highbd_convolve4_8_y(int16x8_t samples[4], + int16x8_t filter, + uint16x8_t max) { + int64x2_t sum01 = aom_svdot_lane_s16(vdupq_n_s64(0), samples[0], filter, 0); + int64x2_t sum23 = aom_svdot_lane_s16(vdupq_n_s64(0), samples[1], filter, 0); + int64x2_t sum45 = aom_svdot_lane_s16(vdupq_n_s64(0), samples[2], filter, 0); + int64x2_t sum67 = aom_svdot_lane_s16(vdupq_n_s64(0), samples[3], filter, 0); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum45), vmovn_s64(sum67)); + uint16x8_t res = vcombine_u16(vqrshrun_n_s32(sum0123, FILTER_BITS), + vqrshrun_n_s32(sum4567, FILTER_BITS)); + return vminq_u16(res, max); +} + +void highbd_convolve_y_sr_4tap_sve2(const uint16_t *src, ptrdiff_t src_stride, + uint16_t *dst, ptrdiff_t dst_stride, + int width, int height, + const int16_t *filter_y, int bd) { + assert(w >= 4 && h >= 4); + + const int16x8_t y_filter = + vcombine_s16(vld1_s16(filter_y + 2), vdup_n_s16(0)); + + if (width == 4) { + const uint16x4_t max = vdup_n_u16((1 << bd) - 1); + int16_t *s = (int16_t *)src; + + int16x4_t s0, s1, s2; + load_s16_4x3(s, src_stride, &s0, &s1, &s2); + s += 3 * src_stride; + + do { + int16x4_t s3, s4, s5, s6; + load_s16_4x4(s, src_stride, &s3, &s4, &s5, &s6); + + // This operation combines a conventional transpose and the sample permute + // required before computing the dot product. + int16x8_t s0123[2], s1234[2], s2345[2], s3456[2]; + transpose_concat_4x4(s0, s1, s2, s3, s0123); + transpose_concat_4x4(s1, s2, s3, s4, s1234); + transpose_concat_4x4(s2, s3, s4, s5, s2345); + transpose_concat_4x4(s3, s4, s5, s6, s3456); + + uint16x4_t d0 = highbd_convolve4_4_y(s0123, y_filter, max); + uint16x4_t d1 = highbd_convolve4_4_y(s1234, y_filter, max); + uint16x4_t d2 = highbd_convolve4_4_y(s2345, y_filter, max); + uint16x4_t d3 = highbd_convolve4_4_y(s3456, y_filter, max); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + // Shuffle everything up four rows. + s0 = s4; + s1 = s5; + s2 = s6; + + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } else { + const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); + + do { + int h = height; + int16_t *s = (int16_t *)src; + uint16_t *d = dst; + + int16x8_t s0, s1, s2; + load_s16_8x3(s, src_stride, &s0, &s1, &s2); + s += 3 * src_stride; + + do { + int16x8_t s3, s4, s5, s6; + load_s16_8x4(s, src_stride, &s3, &s4, &s5, &s6); + + // This operation combines a conventional transpose and the sample + // permute required before computing the dot product. + int16x8_t s0123[4], s1234[4], s2345[4], s3456[4]; + transpose_concat_8x4(s0, s1, s2, s3, s0123); + transpose_concat_8x4(s1, s2, s3, s4, s1234); + transpose_concat_8x4(s2, s3, s4, s5, s2345); + transpose_concat_8x4(s3, s4, s5, s6, s3456); + + uint16x8_t d0 = highbd_convolve4_8_y(s0123, y_filter, max); + uint16x8_t d1 = highbd_convolve4_8_y(s1234, y_filter, max); + uint16x8_t d2 = highbd_convolve4_8_y(s2345, y_filter, max); + uint16x8_t d3 = highbd_convolve4_8_y(s3456, y_filter, max); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + // Shuffle everything up four rows. + s0 = s4; + s1 = s5; + s2 = s6; + + s += 4 * src_stride; + d += 4 * dst_stride; + h -= 4; + } while (h != 0); + src += 8; + dst += 8; + width -= 8; + } while (width != 0); + } +} + +void av1_highbd_convolve_y_sr_sve2(const uint16_t *src, int src_stride, + uint16_t *dst, int dst_stride, int w, int h, + const InterpFilterParams *filter_params_y, + const int subpel_y_qn, int bd) { + if (w == 2 || h == 2) { + av1_highbd_convolve_y_sr_c(src, src_stride, dst, dst_stride, w, h, + filter_params_y, subpel_y_qn, bd); + return; + } + const int y_filter_taps = get_filter_tap(filter_params_y, subpel_y_qn); + + if (y_filter_taps == 6) { + av1_highbd_convolve_y_sr_neon(src, src_stride, dst, dst_stride, w, h, + filter_params_y, subpel_y_qn, bd); + return; + } + + const int vert_offset = filter_params_y->taps / 2 - 1; + const int16_t *y_filter_ptr = av1_get_interp_filter_subpel_kernel( + filter_params_y, subpel_y_qn & SUBPEL_MASK); + + src -= vert_offset * src_stride; + + if (y_filter_taps > 8) { + highbd_convolve_y_sr_12tap_sve2(src, src_stride, dst, dst_stride, w, h, + y_filter_ptr, bd); + return; + } + + if (y_filter_taps == 4) { + highbd_convolve_y_sr_4tap_sve2(src + 2 * src_stride, src_stride, dst, + dst_stride, w, h, y_filter_ptr, bd); + return; + } + + highbd_convolve_y_sr_8tap_sve2(src, src_stride, dst, dst_stride, w, h, + y_filter_ptr, bd); +} + +static INLINE uint16x4_t convolve12_4_2d_h( + int16x8_t s0, int16x8_t s1, int16x8_t filter_0_7, int16x8_t filter_4_11, + const int64x2_t offset, int32x4_t shift, uint16x8x4_t permute_tbl) { + int16x8_t permuted_samples[6]; + permuted_samples[0] = aom_tbl_s16(s0, permute_tbl.val[0]); + permuted_samples[1] = aom_tbl_s16(s0, permute_tbl.val[1]); + permuted_samples[2] = aom_tbl2_s16(s0, s1, permute_tbl.val[2]); + permuted_samples[3] = aom_tbl2_s16(s0, s1, permute_tbl.val[3]); + permuted_samples[4] = aom_tbl_s16(s1, permute_tbl.val[0]); + permuted_samples[5] = aom_tbl_s16(s1, permute_tbl.val[1]); + + int64x2_t sum01 = + aom_svdot_lane_s16(offset, permuted_samples[0], filter_0_7, 0); + sum01 = aom_svdot_lane_s16(sum01, permuted_samples[2], filter_0_7, 1); + sum01 = aom_svdot_lane_s16(sum01, permuted_samples[4], filter_4_11, 1); + + int64x2_t sum23 = + aom_svdot_lane_s16(offset, permuted_samples[1], filter_0_7, 0); + sum23 = aom_svdot_lane_s16(sum23, permuted_samples[3], filter_0_7, 1); + sum23 = aom_svdot_lane_s16(sum23, permuted_samples[5], filter_4_11, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + sum0123 = vqrshlq_s32(sum0123, shift); + return vqmovun_s32(sum0123); +} + +static INLINE uint16x8_t convolve12_8_2d_h(int16x8_t s0, int16x8_t s1, + int16x8_t s2, int16x8_t filter_0_7, + int16x8_t filter_4_11, + int64x2_t offset, int32x4_t shift, + uint16x8x4_t permute_tbl) { + int16x8_t permuted_samples[8]; + permuted_samples[0] = aom_tbl_s16(s0, permute_tbl.val[0]); + permuted_samples[1] = aom_tbl_s16(s0, permute_tbl.val[1]); + permuted_samples[2] = aom_tbl2_s16(s0, s1, permute_tbl.val[2]); + permuted_samples[3] = aom_tbl2_s16(s0, s1, permute_tbl.val[3]); + permuted_samples[4] = aom_tbl_s16(s1, permute_tbl.val[0]); + permuted_samples[5] = aom_tbl_s16(s1, permute_tbl.val[1]); + permuted_samples[6] = aom_tbl2_s16(s1, s2, permute_tbl.val[2]); + permuted_samples[7] = aom_tbl2_s16(s1, s2, permute_tbl.val[3]); + + int64x2_t sum01 = + aom_svdot_lane_s16(offset, permuted_samples[0], filter_0_7, 0); + sum01 = aom_svdot_lane_s16(sum01, permuted_samples[2], filter_0_7, 1); + sum01 = aom_svdot_lane_s16(sum01, permuted_samples[4], filter_4_11, 1); + + int64x2_t sum23 = + aom_svdot_lane_s16(offset, permuted_samples[1], filter_0_7, 0); + sum23 = aom_svdot_lane_s16(sum23, permuted_samples[3], filter_0_7, 1); + sum23 = aom_svdot_lane_s16(sum23, permuted_samples[5], filter_4_11, 1); + + int64x2_t sum45 = + aom_svdot_lane_s16(offset, permuted_samples[2], filter_0_7, 0); + sum45 = aom_svdot_lane_s16(sum45, permuted_samples[4], filter_0_7, 1); + sum45 = aom_svdot_lane_s16(sum45, permuted_samples[6], filter_4_11, 1); + + int64x2_t sum67 = + aom_svdot_lane_s16(offset, permuted_samples[3], filter_0_7, 0); + sum67 = aom_svdot_lane_s16(sum67, permuted_samples[5], filter_0_7, 1); + sum67 = aom_svdot_lane_s16(sum67, permuted_samples[7], filter_4_11, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum45), vmovn_s64(sum67)); + + sum0123 = vqrshlq_s32(sum0123, shift); + sum4567 = vqrshlq_s32(sum4567, shift); + + return vcombine_u16(vqmovun_s32(sum0123), vqmovun_s32(sum4567)); +} + +static INLINE void highbd_convolve_2d_sr_horiz_12tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *y_filter_ptr, + ConvolveParams *conv_params, const int x_offset) { + const int64x2_t offset = vdupq_n_s64(x_offset); + const int32x4_t shift = vdupq_n_s32(-conv_params->round_0); + + const int16x8_t y_filter_0_7 = vld1q_s16(y_filter_ptr); + const int16x8_t y_filter_4_11 = vld1q_s16(y_filter_ptr + 4); + + uint16x8x4_t permute_tbl = vld1q_u16_x4(kDotProdTbl); + // Scale indices by size of the true vector length to avoid reading from an + // 'undefined' portion of a vector on a system with SVE vectors > 128-bit. + uint16x8_t correction0 = vreinterpretq_u16_u64(vcombine_u64( + vdup_n_u64(0), vdup_n_u64(svcnth() * 0x0001000000000000ULL))); + permute_tbl.val[2] = vaddq_u16(permute_tbl.val[2], correction0); + + uint16x8_t correction1 = vreinterpretq_u16_u64( + vcombine_u64(vdup_n_u64(svcnth() * 0x0001000100000000ULL), + vdup_n_u64(svcnth() * 0x0001000100010000ULL))); + permute_tbl.val[3] = vaddq_u16(permute_tbl.val[3], correction1); + + if (width == 4) { + const int16_t *s = (const int16_t *)src; + + do { + int16x8_t s0, s1, s2, s3, s4, s5, s6, s7; + load_s16_8x4(s, src_stride, &s0, &s2, &s4, &s6); + load_s16_8x4(s + 8, src_stride, &s1, &s3, &s5, &s7); + + uint16x4_t d0 = convolve12_4_2d_h(s0, s1, y_filter_0_7, y_filter_4_11, + offset, shift, permute_tbl); + uint16x4_t d1 = convolve12_4_2d_h(s2, s3, y_filter_0_7, y_filter_4_11, + offset, shift, permute_tbl); + uint16x4_t d2 = convolve12_4_2d_h(s4, s5, y_filter_0_7, y_filter_4_11, + offset, shift, permute_tbl); + uint16x4_t d3 = convolve12_4_2d_h(s6, s7, y_filter_0_7, y_filter_4_11, + offset, shift, permute_tbl); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + dst += 4 * dst_stride; + s += 4 * src_stride; + height -= 4; + } while (height > 0); + } else { + do { + const int16_t *s = (const int16_t *)src; + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11; + load_s16_8x4(s, src_stride, &s0, &s3, &s6, &s9); + load_s16_8x4(s + 8, src_stride, &s1, &s4, &s7, &s10); + load_s16_8x4(s + 16, src_stride, &s2, &s5, &s8, &s11); + + uint16x8_t d0 = + convolve12_8_2d_h(s0, s1, s2, y_filter_0_7, y_filter_4_11, offset, + shift, permute_tbl); + uint16x8_t d1 = + convolve12_8_2d_h(s3, s4, s5, y_filter_0_7, y_filter_4_11, offset, + shift, permute_tbl); + uint16x8_t d2 = + convolve12_8_2d_h(s6, s7, s8, y_filter_0_7, y_filter_4_11, offset, + shift, permute_tbl); + uint16x8_t d3 = + convolve12_8_2d_h(s9, s10, s11, y_filter_0_7, y_filter_4_11, offset, + shift, permute_tbl); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height > 0); + } +} + +static INLINE uint16x8_t convolve8_8_2d_h(int16x8_t s0[8], int16x8_t filter, + int64x2_t offset, int32x4_t shift) { + int64x2_t sum[8]; + sum[0] = aom_sdotq_s16(offset, s0[0], filter); + sum[1] = aom_sdotq_s16(offset, s0[1], filter); + sum[2] = aom_sdotq_s16(offset, s0[2], filter); + sum[3] = aom_sdotq_s16(offset, s0[3], filter); + sum[4] = aom_sdotq_s16(offset, s0[4], filter); + sum[5] = aom_sdotq_s16(offset, s0[5], filter); + sum[6] = aom_sdotq_s16(offset, s0[6], filter); + sum[7] = aom_sdotq_s16(offset, s0[7], filter); + + sum[0] = vpaddq_s64(sum[0], sum[1]); + sum[2] = vpaddq_s64(sum[2], sum[3]); + sum[4] = vpaddq_s64(sum[4], sum[5]); + sum[6] = vpaddq_s64(sum[6], sum[7]); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum[0]), vmovn_s64(sum[2])); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum[4]), vmovn_s64(sum[6])); + + sum0123 = vqrshlq_s32(sum0123, shift); + sum4567 = vqrshlq_s32(sum4567, shift); + + return vcombine_u16(vqmovun_s32(sum0123), vqmovun_s32(sum4567)); +} + +static INLINE void highbd_convolve_2d_sr_horiz_8tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *y_filter_ptr, + ConvolveParams *conv_params, const int x_offset) { + const int64x2_t offset = vdupq_n_s64(x_offset); + const int64x2_t offset_lo = vcombine_s64(vget_low_s64(offset), vdup_n_s64(0)); + const int32x4_t shift = vdupq_n_s32(-conv_params->round_0); + + const int16x8_t filter = vld1q_s16(y_filter_ptr); + + do { + const int16_t *s = (const int16_t *)src; + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0[8], s1[8], s2[8], s3[8]; + load_s16_8x8(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3], + &s0[4], &s0[5], &s0[6], &s0[7]); + load_s16_8x8(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3], + &s1[4], &s1[5], &s1[6], &s1[7]); + load_s16_8x8(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3], + &s2[4], &s2[5], &s2[6], &s2[7]); + load_s16_8x8(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3], + &s3[4], &s3[5], &s3[6], &s3[7]); + + uint16x8_t d0 = convolve8_8_2d_h(s0, filter, offset_lo, shift); + uint16x8_t d1 = convolve8_8_2d_h(s1, filter, offset_lo, shift); + uint16x8_t d2 = convolve8_8_2d_h(s2, filter, offset_lo, shift); + uint16x8_t d3 = convolve8_8_2d_h(s3, filter, offset_lo, shift); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height > 0); +} + +static INLINE uint16x4_t convolve4_4_2d_h(int16x8_t s0, int16x8_t filter, + int64x2_t offset, int32x4_t shift, + uint16x8x2_t permute_tbl) { + int16x8_t permuted_samples0 = aom_tbl_s16(s0, permute_tbl.val[0]); + int16x8_t permuted_samples1 = aom_tbl_s16(s0, permute_tbl.val[1]); + + int64x2_t sum01 = aom_svdot_lane_s16(offset, permuted_samples0, filter, 0); + int64x2_t sum23 = aom_svdot_lane_s16(offset, permuted_samples1, filter, 0); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + sum0123 = vqrshlq_s32(sum0123, shift); + return vqmovun_s32(sum0123); +} + +static INLINE uint16x8_t convolve4_8_2d_h(int16x8_t s0[8], int16x8_t filter, + int64x2_t offset, int32x4_t shift, + uint16x8_t tbl) { + int64x2_t sum04 = aom_svdot_lane_s16(offset, s0[0], filter, 0); + int64x2_t sum15 = aom_svdot_lane_s16(offset, s0[1], filter, 0); + int64x2_t sum26 = aom_svdot_lane_s16(offset, s0[2], filter, 0); + int64x2_t sum37 = aom_svdot_lane_s16(offset, s0[3], filter, 0); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum04), vmovn_s64(sum15)); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum26), vmovn_s64(sum37)); + + sum0123 = vqrshlq_s32(sum0123, shift); + sum4567 = vqrshlq_s32(sum4567, shift); + + uint16x8_t res = vcombine_u16(vqmovun_s32(sum0123), vqmovun_s32(sum4567)); + return aom_tbl_u16(res, tbl); +} + +static INLINE void highbd_convolve_2d_sr_horiz_4tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *x_filter_ptr, + ConvolveParams *conv_params, const int x_offset) { + const int64x2_t offset = vdupq_n_s64(x_offset); + const int32x4_t shift = vdupq_n_s32(-conv_params->round_0); + + const int16x4_t x_filter = vld1_s16(x_filter_ptr + 2); + const int16x8_t filter = vcombine_s16(x_filter, vdup_n_s16(0)); + + if (width == 4) { + const int16_t *s = (const int16_t *)(src); + + uint16x8x2_t permute_tbl = vld1q_u16_x2(kDotProdTbl); + + do { + int16x8_t s0, s1, s2, s3; + load_s16_8x4(s, src_stride, &s0, &s1, &s2, &s3); + + uint16x4_t d0 = convolve4_4_2d_h(s0, filter, offset, shift, permute_tbl); + uint16x4_t d1 = convolve4_4_2d_h(s1, filter, offset, shift, permute_tbl); + uint16x4_t d2 = convolve4_4_2d_h(s2, filter, offset, shift, permute_tbl); + uint16x4_t d3 = convolve4_4_2d_h(s3, filter, offset, shift, permute_tbl); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height > 0); + } else { + uint16x8_t idx = vld1q_u16(kDeinterleaveTbl); + + do { + const int16_t *s = (const int16_t *)(src); + uint16_t *d = dst; + int w = width; + + do { + int16x8_t s0[8], s1[8], s2[8], s3[8]; + load_s16_8x8(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3], + &s0[4], &s0[5], &s0[6], &s0[7]); + load_s16_8x8(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3], + &s1[4], &s1[5], &s1[6], &s1[7]); + load_s16_8x8(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3], + &s2[4], &s2[5], &s2[6], &s2[7]); + load_s16_8x8(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3], + &s3[4], &s3[5], &s3[6], &s3[7]); + + uint16x8_t d0 = convolve4_8_2d_h(s0, filter, offset, shift, idx); + uint16x8_t d1 = convolve4_8_2d_h(s1, filter, offset, shift, idx); + uint16x8_t d2 = convolve4_8_2d_h(s2, filter, offset, shift, idx); + uint16x8_t d3 = convolve4_8_2d_h(s3, filter, offset, shift, idx); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + s += 8; + d += 8; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height > 0); + } +} + +static INLINE uint16x4_t highbd_convolve12_4_2d_v( + int16x8_t s0[2], int16x8_t s1[2], int16x8_t s2[2], int16x8_t filter_0_7, + int16x8_t filter_4_11, int32x4_t shift, int64x2_t offset, uint16x4_t max) { + int64x2_t sum01 = aom_svdot_lane_s16(offset, s0[0], filter_0_7, 0); + sum01 = aom_svdot_lane_s16(sum01, s1[0], filter_0_7, 1); + sum01 = aom_svdot_lane_s16(sum01, s2[0], filter_4_11, 1); + + int64x2_t sum23 = aom_svdot_lane_s16(offset, s0[1], filter_0_7, 0); + sum23 = aom_svdot_lane_s16(sum23, s1[1], filter_0_7, 1); + sum23 = aom_svdot_lane_s16(sum23, s2[1], filter_4_11, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + sum0123 = vshlq_s32(sum0123, shift); + + uint16x4_t res = vqmovun_s32(sum0123); + + return vmin_u16(res, max); +} + +static INLINE void highbd_convolve_2d_sr_vert_12tap_sve2( + const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, + int width, int height, const int16_t *y_filter_ptr, + ConvolveParams *conv_params, int bd, const int y_offset) { + const int64x2_t offset = vdupq_n_s64(y_offset); + const int32x4_t shift = vdupq_n_s32(-conv_params->round_1); + + const int16x8_t y_filter_0_7 = vld1q_s16(y_filter_ptr); + const int16x8_t y_filter_4_11 = vld1q_s16(y_filter_ptr + 4); + + uint16x8x3_t merge_block_tbl = vld1q_u16_x3(kDotProdMergeBlockTbl); + // Scale indices by size of the true vector length to avoid reading from an + // 'undefined' portion of a vector on a system with SVE vectors > 128-bit. + uint16x8_t correction0 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000000000000ULL)); + merge_block_tbl.val[0] = vaddq_u16(merge_block_tbl.val[0], correction0); + + uint16x8_t correction1 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100000000ULL)); + merge_block_tbl.val[1] = vaddq_u16(merge_block_tbl.val[1], correction1); + + uint16x8_t correction2 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100010000ULL)); + merge_block_tbl.val[2] = vaddq_u16(merge_block_tbl.val[2], correction2); + + const uint16x4_t max = vdup_n_u16((1 << bd) - 1); + + do { + int16_t *s = (int16_t *)src; + uint16_t *d = (uint16_t *)dst; + int h = height; + + int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, sA; + load_s16_4x11(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7, &s8, + &s9, &sA); + s += 11 * src_stride; + + int16x8_t s0123[2], s1234[2], s2345[2], s3456[2], s4567[2], s5678[2], + s6789[2], s789A[2]; + // This operation combines a conventional transpose and the sample permute + // required before computing the dot product. + transpose_concat_4x4(s0, s1, s2, s3, s0123); + transpose_concat_4x4(s1, s2, s3, s4, s1234); + transpose_concat_4x4(s2, s3, s4, s5, s2345); + transpose_concat_4x4(s3, s4, s5, s6, s3456); + transpose_concat_4x4(s4, s5, s6, s7, s4567); + transpose_concat_4x4(s5, s6, s7, s8, s5678); + transpose_concat_4x4(s6, s7, s8, s9, s6789); + transpose_concat_4x4(s7, s8, s9, sA, s789A); + + do { + int16x4_t sB, sC, sD, sE; + load_s16_4x4(s, src_stride, &sB, &sC, &sD, &sE); + + int16x8_t s89AB[2], s9ABC[2], sABCD[2], sBCDE[2]; + transpose_concat_4x4(sB, sC, sD, sE, sBCDE); + + // Use the above transpose and reuse data from the previous loop to get + // the rest. + aom_tbl2x2_s16(s789A, sBCDE, merge_block_tbl.val[0], s89AB); + aom_tbl2x2_s16(s789A, sBCDE, merge_block_tbl.val[1], s9ABC); + aom_tbl2x2_s16(s789A, sBCDE, merge_block_tbl.val[2], sABCD); + + uint16x4_t d0 = highbd_convolve12_4_2d_v( + s0123, s4567, s89AB, y_filter_0_7, y_filter_4_11, shift, offset, max); + uint16x4_t d1 = highbd_convolve12_4_2d_v( + s1234, s5678, s9ABC, y_filter_0_7, y_filter_4_11, shift, offset, max); + uint16x4_t d2 = highbd_convolve12_4_2d_v( + s2345, s6789, sABCD, y_filter_0_7, y_filter_4_11, shift, offset, max); + uint16x4_t d3 = highbd_convolve12_4_2d_v( + s3456, s789A, sBCDE, y_filter_0_7, y_filter_4_11, shift, offset, max); + + store_u16_4x4(d, dst_stride, d0, d1, d2, d3); + + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. + s0123[0] = s4567[0]; + s0123[1] = s4567[1]; + s1234[0] = s5678[0]; + s1234[1] = s5678[1]; + s2345[0] = s6789[0]; + s2345[1] = s6789[1]; + s3456[0] = s789A[0]; + s3456[1] = s789A[1]; + s4567[0] = s89AB[0]; + s4567[1] = s89AB[1]; + s5678[0] = s9ABC[0]; + s5678[1] = s9ABC[1]; + s6789[0] = sABCD[0]; + s6789[1] = sABCD[1]; + s789A[0] = sBCDE[0]; + s789A[1] = sBCDE[1]; + + s += 4 * src_stride; + d += 4 * dst_stride; + h -= 4; + } while (h != 0); + src += 4; + dst += 4; + width -= 4; + } while (width != 0); +} + +static INLINE uint16x4_t highbd_convolve8_4_2d_v( + int16x8_t samples_lo[2], int16x8_t samples_hi[2], int16x8_t filter, + int32x4_t shift, int64x2_t offset, uint16x4_t max) { + int64x2_t sum01 = aom_svdot_lane_s16(offset, samples_lo[0], filter, 0); + sum01 = aom_svdot_lane_s16(sum01, samples_hi[0], filter, 1); + + int64x2_t sum23 = aom_svdot_lane_s16(offset, samples_lo[1], filter, 0); + sum23 = aom_svdot_lane_s16(sum23, samples_hi[1], filter, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + sum0123 = vshlq_s32(sum0123, shift); + + uint16x4_t res = vqmovun_s32(sum0123); + return vmin_u16(res, max); +} + +static INLINE uint16x8_t highbd_convolve8_8_2d_v( + int16x8_t samples_lo[4], int16x8_t samples_hi[4], int16x8_t filter, + int32x4_t shift, int64x2_t offset, uint16x8_t max) { + int64x2_t sum01 = aom_svdot_lane_s16(offset, samples_lo[0], filter, 0); + sum01 = aom_svdot_lane_s16(sum01, samples_hi[0], filter, 1); + + int64x2_t sum23 = aom_svdot_lane_s16(offset, samples_lo[1], filter, 0); + sum23 = aom_svdot_lane_s16(sum23, samples_hi[1], filter, 1); + + int64x2_t sum45 = aom_svdot_lane_s16(offset, samples_lo[2], filter, 0); + sum45 = aom_svdot_lane_s16(sum45, samples_hi[2], filter, 1); + + int64x2_t sum67 = aom_svdot_lane_s16(offset, samples_lo[3], filter, 0); + sum67 = aom_svdot_lane_s16(sum67, samples_hi[3], filter, 1); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum45), vmovn_s64(sum67)); + + sum0123 = vshlq_s32(sum0123, shift); + sum4567 = vshlq_s32(sum4567, shift); + + uint16x8_t res = vcombine_u16(vqmovun_s32(sum0123), vqmovun_s32(sum4567)); + return vminq_u16(res, max); +} + +void highbd_convolve_2d_sr_vert_8tap_sve2(const uint16_t *src, + ptrdiff_t src_stride, uint16_t *dst, + ptrdiff_t dst_stride, int width, + int height, const int16_t *filter_y, + ConvolveParams *conv_params, int bd, + const int y_offset) { + assert(w >= 4 && h >= 4); + const int64x2_t offset = vdupq_n_s64(y_offset); + const int32x4_t shift = vdupq_n_s32(-conv_params->round_1); + const int16x8_t y_filter = vld1q_s16(filter_y); + + uint16x8x3_t merge_block_tbl = vld1q_u16_x3(kDotProdMergeBlockTbl); + // Scale indices by size of the true vector length to avoid reading from an + // 'undefined' portion of a vector on a system with SVE vectors > 128-bit. + uint16x8_t correction0 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000000000000ULL)); + merge_block_tbl.val[0] = vaddq_u16(merge_block_tbl.val[0], correction0); + + uint16x8_t correction1 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100000000ULL)); + merge_block_tbl.val[1] = vaddq_u16(merge_block_tbl.val[1], correction1); + + uint16x8_t correction2 = + vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100010000ULL)); + merge_block_tbl.val[2] = vaddq_u16(merge_block_tbl.val[2], correction2); + + if (width == 4) { + const uint16x4_t max = vdup_n_u16((1 << bd) - 1); + int16_t *s = (int16_t *)src; + + int16x4_t s0, s1, s2, s3, s4, s5, s6; + load_s16_4x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); + s += 7 * src_stride; + + // This operation combines a conventional transpose and the sample permute + // required before computing the dot product. + int16x8_t s0123[2], s1234[2], s2345[2], s3456[2]; + transpose_concat_4x4(s0, s1, s2, s3, s0123); + transpose_concat_4x4(s1, s2, s3, s4, s1234); + transpose_concat_4x4(s2, s3, s4, s5, s2345); + transpose_concat_4x4(s3, s4, s5, s6, s3456); + + do { + int16x4_t s7, s8, s9, s10; + load_s16_4x4(s, src_stride, &s7, &s8, &s9, &s10); + + int16x8_t s4567[2], s5678[2], s6789[2], s789A[2]; + // Transpose and shuffle the 4 lines that were loaded. + transpose_concat_4x4(s7, s8, s9, s10, s789A); + + // Merge new data into block from previous iteration. + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[0], s4567); + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[1], s5678); + aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[2], s6789); + + uint16x4_t d0 = + highbd_convolve8_4_2d_v(s0123, s4567, y_filter, shift, offset, max); + uint16x4_t d1 = + highbd_convolve8_4_2d_v(s1234, s5678, y_filter, shift, offset, max); + uint16x4_t d2 = + highbd_convolve8_4_2d_v(s2345, s6789, y_filter, shift, offset, max); + uint16x4_t d3 = + highbd_convolve8_4_2d_v(s3456, s789A, y_filter, shift, offset, max); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. + s0123[0] = s4567[0]; + s0123[1] = s4567[1]; + s1234[0] = s5678[0]; + s1234[1] = s5678[1]; + s2345[0] = s6789[0]; + s2345[1] = s6789[1]; + s3456[0] = s789A[0]; + s3456[1] = s789A[1]; + + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } else { + const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); + + do { + int h = height; + int16_t *s = (int16_t *)src; + uint16_t *d = dst; + + int16x8_t s0, s1, s2, s3, s4, s5, s6; + load_s16_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); + s += 7 * src_stride; + + // This operation combines a conventional transpose and the sample permute + // required before computing the dot product. + int16x8_t s0123[4], s1234[4], s2345[4], s3456[4]; + transpose_concat_8x4(s0, s1, s2, s3, s0123); + transpose_concat_8x4(s1, s2, s3, s4, s1234); + transpose_concat_8x4(s2, s3, s4, s5, s2345); + transpose_concat_8x4(s3, s4, s5, s6, s3456); + + do { + int16x8_t s7, s8, s9, s10; + load_s16_8x4(s, src_stride, &s7, &s8, &s9, &s10); + + int16x8_t s4567[4], s5678[4], s6789[4], s789A[4]; + // Transpose and shuffle the 4 lines that were loaded. + transpose_concat_8x4(s7, s8, s9, s10, s789A); + + // Merge new data into block from previous iteration. + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[0], s4567); + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[1], s5678); + aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[2], s6789); + + uint16x8_t d0 = + highbd_convolve8_8_2d_v(s0123, s4567, y_filter, shift, offset, max); + uint16x8_t d1 = + highbd_convolve8_8_2d_v(s1234, s5678, y_filter, shift, offset, max); + uint16x8_t d2 = + highbd_convolve8_8_2d_v(s2345, s6789, y_filter, shift, offset, max); + uint16x8_t d3 = + highbd_convolve8_8_2d_v(s3456, s789A, y_filter, shift, offset, max); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. + s0123[0] = s4567[0]; + s0123[1] = s4567[1]; + s0123[2] = s4567[2]; + s0123[3] = s4567[3]; + s1234[0] = s5678[0]; + s1234[1] = s5678[1]; + s1234[2] = s5678[2]; + s1234[3] = s5678[3]; + s2345[0] = s6789[0]; + s2345[1] = s6789[1]; + s2345[2] = s6789[2]; + s2345[3] = s6789[3]; + s3456[0] = s789A[0]; + s3456[1] = s789A[1]; + s3456[2] = s789A[2]; + s3456[3] = s789A[3]; + + s += 4 * src_stride; + d += 4 * dst_stride; + h -= 4; + } while (h != 0); + src += 8; + dst += 8; + width -= 8; + } while (width != 0); + } +} + +static INLINE uint16x4_t highbd_convolve4_4_2d_v(int16x8_t samples[2], + int16x8_t filter, + int32x4_t shift, + int64x2_t offset, + uint16x4_t max) { + int64x2_t sum01 = aom_svdot_lane_s16(offset, samples[0], filter, 0); + int64x2_t sum23 = aom_svdot_lane_s16(offset, samples[1], filter, 0); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + sum0123 = vshlq_s32(sum0123, shift); + + uint16x4_t res = vqmovun_s32(sum0123); + return vmin_u16(res, max); +} + +static INLINE uint16x8_t highbd_convolve4_8_2d_v(int16x8_t samples[4], + int16x8_t filter, + int32x4_t shift, + int64x2_t offset, + uint16x8_t max) { + int64x2_t sum01 = aom_svdot_lane_s16(offset, samples[0], filter, 0); + int64x2_t sum23 = aom_svdot_lane_s16(offset, samples[1], filter, 0); + int64x2_t sum45 = aom_svdot_lane_s16(offset, samples[2], filter, 0); + int64x2_t sum67 = aom_svdot_lane_s16(offset, samples[3], filter, 0); + + int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23)); + int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum45), vmovn_s64(sum67)); + + sum0123 = vshlq_s32(sum0123, shift); + sum4567 = vshlq_s32(sum4567, shift); + + uint16x8_t res = vcombine_u16(vqmovun_s32(sum0123), vqmovun_s32(sum4567)); + return vminq_u16(res, max); +} + +void highbd_convolve_2d_sr_vert_4tap_sve2(const uint16_t *src, + ptrdiff_t src_stride, uint16_t *dst, + ptrdiff_t dst_stride, int width, + int height, const int16_t *filter_y, + ConvolveParams *conv_params, int bd, + const int y_offset) { + assert(w >= 4 && h >= 4); + const int64x2_t offset = vdupq_n_s64(y_offset); + const int32x4_t shift = vdupq_n_s32(-conv_params->round_1); + + const int16x8_t y_filter = + vcombine_s16(vld1_s16(filter_y + 2), vdup_n_s16(0)); + + if (width == 4) { + const uint16x4_t max = vdup_n_u16((1 << bd) - 1); + int16_t *s = (int16_t *)(src); + + int16x4_t s0, s1, s2; + load_s16_4x3(s, src_stride, &s0, &s1, &s2); + s += 3 * src_stride; + + do { + int16x4_t s3, s4, s5, s6; + load_s16_4x4(s, src_stride, &s3, &s4, &s5, &s6); + + // This operation combines a conventional transpose and the sample permute + // required before computing the dot product. + int16x8_t s0123[2], s1234[2], s2345[2], s3456[2]; + transpose_concat_4x4(s0, s1, s2, s3, s0123); + transpose_concat_4x4(s1, s2, s3, s4, s1234); + transpose_concat_4x4(s2, s3, s4, s5, s2345); + transpose_concat_4x4(s3, s4, s5, s6, s3456); + + uint16x4_t d0 = + highbd_convolve4_4_2d_v(s0123, y_filter, shift, offset, max); + uint16x4_t d1 = + highbd_convolve4_4_2d_v(s1234, y_filter, shift, offset, max); + uint16x4_t d2 = + highbd_convolve4_4_2d_v(s2345, y_filter, shift, offset, max); + uint16x4_t d3 = + highbd_convolve4_4_2d_v(s3456, y_filter, shift, offset, max); + + store_u16_4x4(dst, dst_stride, d0, d1, d2, d3); + + // Shuffle everything up four rows. + s0 = s4; + s1 = s5; + s2 = s6; + + s += 4 * src_stride; + dst += 4 * dst_stride; + height -= 4; + } while (height != 0); + } else { + const uint16x8_t max = vdupq_n_u16((1 << bd) - 1); + + do { + int h = height; + int16_t *s = (int16_t *)(src); + uint16_t *d = dst; + + int16x8_t s0, s1, s2; + load_s16_8x3(s, src_stride, &s0, &s1, &s2); + s += 3 * src_stride; + + do { + int16x8_t s3, s4, s5, s6; + load_s16_8x4(s, src_stride, &s3, &s4, &s5, &s6); + + // This operation combines a conventional transpose and the sample + // permute required before computing the dot product. + int16x8_t s0123[4], s1234[4], s2345[4], s3456[4]; + transpose_concat_8x4(s0, s1, s2, s3, s0123); + transpose_concat_8x4(s1, s2, s3, s4, s1234); + transpose_concat_8x4(s2, s3, s4, s5, s2345); + transpose_concat_8x4(s3, s4, s5, s6, s3456); + + uint16x8_t d0 = + highbd_convolve4_8_2d_v(s0123, y_filter, shift, offset, max); + uint16x8_t d1 = + highbd_convolve4_8_2d_v(s1234, y_filter, shift, offset, max); + uint16x8_t d2 = + highbd_convolve4_8_2d_v(s2345, y_filter, shift, offset, max); + uint16x8_t d3 = + highbd_convolve4_8_2d_v(s3456, y_filter, shift, offset, max); + + store_u16_8x4(d, dst_stride, d0, d1, d2, d3); + + // Shuffle everything up four rows. + s0 = s4; + s1 = s5; + s2 = s6; + + s += 4 * src_stride; + d += 4 * dst_stride; + h -= 4; + } while (h != 0); + src += 8; + dst += 8; + width -= 8; + } while (width != 0); + } +} + +void av1_highbd_convolve_2d_sr_sve2(const uint16_t *src, int src_stride, + uint16_t *dst, int dst_stride, int w, int h, + const InterpFilterParams *filter_params_x, + const InterpFilterParams *filter_params_y, + const int subpel_x_qn, + const int subpel_y_qn, + ConvolveParams *conv_params, int bd) { + if (w == 2 || h == 2) { + av1_highbd_convolve_2d_sr_c(src, src_stride, dst, dst_stride, w, h, + filter_params_x, filter_params_y, subpel_x_qn, + subpel_y_qn, conv_params, bd); + return; + } + + DECLARE_ALIGNED(16, uint16_t, + im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]); + const int x_filter_taps = get_filter_tap(filter_params_x, subpel_x_qn); + const int y_filter_taps = get_filter_tap(filter_params_y, subpel_y_qn); + + if (x_filter_taps == 6 || y_filter_taps == 6) { + av1_highbd_convolve_2d_sr_neon(src, src_stride, dst, dst_stride, w, h, + filter_params_x, filter_params_y, + subpel_x_qn, subpel_y_qn, conv_params, bd); + return; + } + + const int clamped_x_taps = x_filter_taps < 4 ? 4 : x_filter_taps; + const int clamped_y_taps = y_filter_taps < 4 ? 4 : y_filter_taps; + + const int im_stride = MAX_SB_SIZE; + const int vert_offset = clamped_y_taps / 2 - 1; + const int horiz_offset = clamped_x_taps / 2 - 1; + const int x_offset = (1 << (bd + FILTER_BITS - 1)); + const int y_offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0; + // The extra shim of (1 << (conv_params->round_1 - 1)) allows us to do a + // simple shift left instead of a rounding saturating shift left. + const int y_offset = + (1 << (conv_params->round_1 - 1)) - (1 << (y_offset_bits - 1)); + + const uint16_t *src_ptr = src - vert_offset * src_stride - horiz_offset; + + const int16_t *x_filter_ptr = av1_get_interp_filter_subpel_kernel( + filter_params_x, subpel_x_qn & SUBPEL_MASK); + const int16_t *y_filter_ptr = av1_get_interp_filter_subpel_kernel( + filter_params_y, subpel_y_qn & SUBPEL_MASK); + const int im_h = h + clamped_y_taps - 1; + + if (x_filter_taps > 8) { + highbd_convolve_2d_sr_horiz_12tap_sve2(src_ptr, src_stride, im_block, + im_stride, w, im_h, x_filter_ptr, + conv_params, x_offset); + + highbd_convolve_2d_sr_vert_12tap_sve2(im_block, im_stride, dst, dst_stride, + w, h, y_filter_ptr, conv_params, bd, + y_offset); + return; + } + + if (x_filter_taps <= 4) { + highbd_convolve_2d_sr_horiz_4tap_sve2(src_ptr, src_stride, im_block, + im_stride, w, im_h, x_filter_ptr, + conv_params, x_offset); + } else { + highbd_convolve_2d_sr_horiz_8tap_sve2(src_ptr, src_stride, im_block, + im_stride, w, im_h, x_filter_ptr, + conv_params, x_offset); + } + + if (y_filter_taps <= 4) { + highbd_convolve_2d_sr_vert_4tap_sve2(im_block, im_stride, dst, dst_stride, + w, h, y_filter_ptr, conv_params, bd, + y_offset); + } else { + highbd_convolve_2d_sr_vert_8tap_sve2(im_block, im_stride, dst, dst_stride, + w, h, y_filter_ptr, conv_params, bd, + y_offset); + } +} diff --git a/third_party/aom/av1/common/arm/highbd_convolve_sve2.h b/third_party/aom/av1/common/arm/highbd_convolve_sve2.h new file mode 100644 index 0000000000..05e23deef4 --- /dev/null +++ b/third_party/aom/av1/common/arm/highbd_convolve_sve2.h @@ -0,0 +1,97 @@ +/* + * Copyright (c) 2023, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_AV1_COMMON_ARM_HIGHBD_CONVOLVE_SVE2_H_ +#define AOM_AV1_COMMON_ARM_HIGHBD_CONVOLVE_SVE2_H_ + +#include <arm_neon.h> + +#include "aom_dsp/arm/aom_neon_sve2_bridge.h" + +// clang-format off +DECLARE_ALIGNED(16, static const uint16_t, kDotProdMergeBlockTbl[24]) = { + // Shift left and insert new last column in transposed 4x4 block. + 1, 2, 3, 0, 5, 6, 7, 4, + // Shift left and insert two new columns in transposed 4x4 block. + 2, 3, 0, 1, 6, 7, 4, 5, + // Shift left and insert three new columns in transposed 4x4 block. + 3, 0, 1, 2, 7, 4, 5, 6, +}; +// clang-format on + +static INLINE void transpose_concat_4x4(int16x4_t s0, int16x4_t s1, + int16x4_t s2, int16x4_t s3, + int16x8_t res[2]) { + // Transpose 16-bit elements and concatenate result rows as follows: + // s0: 00, 01, 02, 03 + // s1: 10, 11, 12, 13 + // s2: 20, 21, 22, 23 + // s3: 30, 31, 32, 33 + // + // res[0]: 00 10 20 30 01 11 21 31 + // res[1]: 02 12 22 32 03 13 23 33 + + int16x8_t s0q = vcombine_s16(s0, vdup_n_s16(0)); + int16x8_t s1q = vcombine_s16(s1, vdup_n_s16(0)); + int16x8_t s2q = vcombine_s16(s2, vdup_n_s16(0)); + int16x8_t s3q = vcombine_s16(s3, vdup_n_s16(0)); + + int32x4_t s01 = vreinterpretq_s32_s16(vzip1q_s16(s0q, s1q)); + int32x4_t s23 = vreinterpretq_s32_s16(vzip1q_s16(s2q, s3q)); + + int32x4x2_t s0123 = vzipq_s32(s01, s23); + + res[0] = vreinterpretq_s16_s32(s0123.val[0]); + res[1] = vreinterpretq_s16_s32(s0123.val[1]); +} + +static INLINE void transpose_concat_8x4(int16x8_t s0, int16x8_t s1, + int16x8_t s2, int16x8_t s3, + int16x8_t res[4]) { + // Transpose 16-bit elements and concatenate result rows as follows: + // s0: 00, 01, 02, 03, 04, 05, 06, 07 + // s1: 10, 11, 12, 13, 14, 15, 16, 17 + // s2: 20, 21, 22, 23, 24, 25, 26, 27 + // s3: 30, 31, 32, 33, 34, 35, 36, 37 + // + // res[0]: 00 10 20 30 01 11 21 31 + // res[1]: 02 12 22 32 03 13 23 33 + // res[2]: 04 14 24 34 05 15 25 35 + // res[3]: 06 16 26 36 07 17 27 37 + + int16x8x2_t tr01_16 = vzipq_s16(s0, s1); + int16x8x2_t tr23_16 = vzipq_s16(s2, s3); + int32x4x2_t tr01_32 = vzipq_s32(vreinterpretq_s32_s16(tr01_16.val[0]), + vreinterpretq_s32_s16(tr23_16.val[0])); + int32x4x2_t tr23_32 = vzipq_s32(vreinterpretq_s32_s16(tr01_16.val[1]), + vreinterpretq_s32_s16(tr23_16.val[1])); + + res[0] = vreinterpretq_s16_s32(tr01_32.val[0]); + res[1] = vreinterpretq_s16_s32(tr01_32.val[1]); + res[2] = vreinterpretq_s16_s32(tr23_32.val[0]); + res[3] = vreinterpretq_s16_s32(tr23_32.val[1]); +} + +static INLINE void aom_tbl2x4_s16(int16x8_t t0[4], int16x8_t t1[4], + uint16x8_t tbl, int16x8_t res[4]) { + res[0] = aom_tbl2_s16(t0[0], t1[0], tbl); + res[1] = aom_tbl2_s16(t0[1], t1[1], tbl); + res[2] = aom_tbl2_s16(t0[2], t1[2], tbl); + res[3] = aom_tbl2_s16(t0[3], t1[3], tbl); +} + +static INLINE void aom_tbl2x2_s16(int16x8_t t0[2], int16x8_t t1[2], + uint16x8_t tbl, int16x8_t res[2]) { + res[0] = aom_tbl2_s16(t0[0], t1[0], tbl); + res[1] = aom_tbl2_s16(t0[1], t1[1], tbl); +} + +#endif // AOM_AV1_COMMON_ARM_HIGHBD_CONVOLVE_SVE2_H_ diff --git a/third_party/aom/av1/common/arm/highbd_warp_plane_neon.c b/third_party/aom/av1/common/arm/highbd_warp_plane_neon.c index c6f1e3ad92..89647bc921 100644 --- a/third_party/aom/av1/common/arm/highbd_warp_plane_neon.c +++ b/third_party/aom/av1/common/arm/highbd_warp_plane_neon.c @@ -23,8 +23,8 @@ #include "config/av1_rtcd.h" #include "highbd_warp_plane_neon.h" -static INLINE int16x8_t highbd_horizontal_filter_4x1_f4(uint16x8x2_t in, int bd, - int sx, int alpha) { +static AOM_FORCE_INLINE int16x8_t +highbd_horizontal_filter_4x1_f4(uint16x8x2_t in, int bd, int sx, int alpha) { int16x8_t f[4]; load_filters_4(f, sx, alpha); @@ -57,8 +57,8 @@ static INLINE int16x8_t highbd_horizontal_filter_4x1_f4(uint16x8x2_t in, int bd, return vcombine_s16(vmovn_s32(res), vdup_n_s16(0)); } -static INLINE int16x8_t highbd_horizontal_filter_8x1_f8(uint16x8x2_t in, int bd, - int sx, int alpha) { +static AOM_FORCE_INLINE int16x8_t +highbd_horizontal_filter_8x1_f8(uint16x8x2_t in, int bd, int sx, int alpha) { int16x8_t f[8]; load_filters_8(f, sx, alpha); @@ -111,8 +111,8 @@ static INLINE int16x8_t highbd_horizontal_filter_8x1_f8(uint16x8x2_t in, int bd, return vcombine_s16(vmovn_s32(res0), vmovn_s32(res1)); } -static INLINE int16x8_t highbd_horizontal_filter_4x1_f1(uint16x8x2_t in, int bd, - int sx) { +static AOM_FORCE_INLINE int16x8_t +highbd_horizontal_filter_4x1_f1(uint16x8x2_t in, int bd, int sx) { int16x8_t f = load_filters_1(sx); int16x8_t rv0 = vextq_s16(vreinterpretq_s16_u16(in.val[0]), @@ -144,8 +144,8 @@ static INLINE int16x8_t highbd_horizontal_filter_4x1_f1(uint16x8x2_t in, int bd, return vcombine_s16(vmovn_s32(res), vdup_n_s16(0)); } -static INLINE int16x8_t highbd_horizontal_filter_8x1_f1(uint16x8x2_t in, int bd, - int sx) { +static AOM_FORCE_INLINE int16x8_t +highbd_horizontal_filter_8x1_f1(uint16x8x2_t in, int bd, int sx) { int16x8_t f = load_filters_1(sx); int16x8_t rv0 = vextq_s16(vreinterpretq_s16_u16(in.val[0]), @@ -197,7 +197,8 @@ static INLINE int16x8_t highbd_horizontal_filter_8x1_f1(uint16x8x2_t in, int bd, return vcombine_s16(vmovn_s32(res0), vmovn_s32(res1)); } -static INLINE int32x4_t vertical_filter_4x1_f1(const int16x8_t *tmp, int sy) { +static AOM_FORCE_INLINE int32x4_t vertical_filter_4x1_f1(const int16x8_t *tmp, + int sy) { const int16x8_t f = load_filters_1(sy); const int16x4_t f0123 = vget_low_s16(f); const int16x4_t f4567 = vget_high_s16(f); @@ -213,7 +214,8 @@ static INLINE int32x4_t vertical_filter_4x1_f1(const int16x8_t *tmp, int sy) { return m0123; } -static INLINE int32x4x2_t vertical_filter_8x1_f1(const int16x8_t *tmp, int sy) { +static AOM_FORCE_INLINE int32x4x2_t vertical_filter_8x1_f1(const int16x8_t *tmp, + int sy) { const int16x8_t f = load_filters_1(sy); const int16x4_t f0123 = vget_low_s16(f); const int16x4_t f4567 = vget_high_s16(f); @@ -238,8 +240,8 @@ static INLINE int32x4x2_t vertical_filter_8x1_f1(const int16x8_t *tmp, int sy) { return (int32x4x2_t){ { m0123, m4567 } }; } -static INLINE int32x4_t vertical_filter_4x1_f4(const int16x8_t *tmp, int sy, - int gamma) { +static AOM_FORCE_INLINE int32x4_t vertical_filter_4x1_f4(const int16x8_t *tmp, + int sy, int gamma) { int16x8_t s0, s1, s2, s3; transpose_elems_s16_4x8( vget_low_s16(tmp[0]), vget_low_s16(tmp[1]), vget_low_s16(tmp[2]), @@ -262,8 +264,8 @@ static INLINE int32x4_t vertical_filter_4x1_f4(const int16x8_t *tmp, int sy, return horizontal_add_4d_s32x4(m0123); } -static INLINE int32x4x2_t vertical_filter_8x1_f8(const int16x8_t *tmp, int sy, - int gamma) { +static AOM_FORCE_INLINE int32x4x2_t vertical_filter_8x1_f8(const int16x8_t *tmp, + int sy, int gamma) { int16x8_t s0 = tmp[0]; int16x8_t s1 = tmp[1]; int16x8_t s2 = tmp[2]; diff --git a/third_party/aom/av1/common/arm/highbd_warp_plane_neon.h b/third_party/aom/av1/common/arm/highbd_warp_plane_neon.h index 3b8982898e..48af4a707b 100644 --- a/third_party/aom/av1/common/arm/highbd_warp_plane_neon.h +++ b/third_party/aom/av1/common/arm/highbd_warp_plane_neon.h @@ -23,29 +23,31 @@ #include "av1/common/warped_motion.h" #include "config/av1_rtcd.h" -static INLINE int16x8_t highbd_horizontal_filter_4x1_f4(uint16x8x2_t in, int bd, - int sx, int alpha); +static AOM_FORCE_INLINE int16x8_t +highbd_horizontal_filter_4x1_f4(uint16x8x2_t in, int bd, int sx, int alpha); -static INLINE int16x8_t highbd_horizontal_filter_8x1_f8(uint16x8x2_t in, int bd, - int sx, int alpha); +static AOM_FORCE_INLINE int16x8_t +highbd_horizontal_filter_8x1_f8(uint16x8x2_t in, int bd, int sx, int alpha); -static INLINE int16x8_t highbd_horizontal_filter_4x1_f1(uint16x8x2_t in, int bd, - int sx); +static AOM_FORCE_INLINE int16x8_t +highbd_horizontal_filter_4x1_f1(uint16x8x2_t in, int bd, int sx); -static INLINE int16x8_t highbd_horizontal_filter_8x1_f1(uint16x8x2_t in, int bd, - int sx); +static AOM_FORCE_INLINE int16x8_t +highbd_horizontal_filter_8x1_f1(uint16x8x2_t in, int bd, int sx); -static INLINE int32x4_t vertical_filter_4x1_f1(const int16x8_t *tmp, int sy); +static AOM_FORCE_INLINE int32x4_t vertical_filter_4x1_f1(const int16x8_t *tmp, + int sy); -static INLINE int32x4x2_t vertical_filter_8x1_f1(const int16x8_t *tmp, int sy); +static AOM_FORCE_INLINE int32x4x2_t vertical_filter_8x1_f1(const int16x8_t *tmp, + int sy); -static INLINE int32x4_t vertical_filter_4x1_f4(const int16x8_t *tmp, int sy, - int gamma); +static AOM_FORCE_INLINE int32x4_t vertical_filter_4x1_f4(const int16x8_t *tmp, + int sy, int gamma); -static INLINE int32x4x2_t vertical_filter_8x1_f8(const int16x8_t *tmp, int sy, - int gamma); +static AOM_FORCE_INLINE int32x4x2_t vertical_filter_8x1_f8(const int16x8_t *tmp, + int sy, int gamma); -static INLINE int16x8_t load_filters_1(int ofs) { +static AOM_FORCE_INLINE int16x8_t load_filters_1(int ofs) { const int ofs0 = ROUND_POWER_OF_TWO(ofs, WARPEDDIFF_PREC_BITS); const int16_t *base = @@ -53,7 +55,8 @@ static INLINE int16x8_t load_filters_1(int ofs) { return vld1q_s16(base + ofs0 * 8); } -static INLINE void load_filters_4(int16x8_t out[], int ofs, int stride) { +static AOM_FORCE_INLINE void load_filters_4(int16x8_t out[], int ofs, + int stride) { const int ofs0 = ROUND_POWER_OF_TWO(ofs + stride * 0, WARPEDDIFF_PREC_BITS); const int ofs1 = ROUND_POWER_OF_TWO(ofs + stride * 1, WARPEDDIFF_PREC_BITS); const int ofs2 = ROUND_POWER_OF_TWO(ofs + stride * 2, WARPEDDIFF_PREC_BITS); @@ -67,7 +70,8 @@ static INLINE void load_filters_4(int16x8_t out[], int ofs, int stride) { out[3] = vld1q_s16(base + ofs3 * 8); } -static INLINE void load_filters_8(int16x8_t out[], int ofs, int stride) { +static AOM_FORCE_INLINE void load_filters_8(int16x8_t out[], int ofs, + int stride) { const int ofs0 = ROUND_POWER_OF_TWO(ofs + stride * 0, WARPEDDIFF_PREC_BITS); const int ofs1 = ROUND_POWER_OF_TWO(ofs + stride * 1, WARPEDDIFF_PREC_BITS); const int ofs2 = ROUND_POWER_OF_TWO(ofs + stride * 2, WARPEDDIFF_PREC_BITS); @@ -89,16 +93,18 @@ static INLINE void load_filters_8(int16x8_t out[], int ofs, int stride) { out[7] = vld1q_s16(base + ofs7 * 8); } -static INLINE uint16x4_t clip_pixel_highbd_vec(int32x4_t val, int bd) { +static AOM_FORCE_INLINE uint16x4_t clip_pixel_highbd_vec(int32x4_t val, + int bd) { const int limit = (1 << bd) - 1; return vqmovun_s32(vminq_s32(val, vdupq_n_s32(limit))); } -static INLINE void warp_affine_horizontal(const uint16_t *ref, int width, - int height, int stride, int p_width, - int16_t alpha, int16_t beta, int iy4, - int sx4, int ix4, int16x8_t tmp[], - int bd) { +static AOM_FORCE_INLINE void warp_affine_horizontal(const uint16_t *ref, + int width, int height, + int stride, int p_width, + int16_t alpha, int16_t beta, + int iy4, int sx4, int ix4, + int16x8_t tmp[], int bd) { const int round0 = (bd == 12) ? ROUND0_BITS + 2 : ROUND0_BITS; if (ix4 <= -7) { @@ -197,7 +203,7 @@ static INLINE void warp_affine_horizontal(const uint16_t *ref, int width, } } -static INLINE void highbd_vertical_filter_4x1_f4( +static AOM_FORCE_INLINE void highbd_vertical_filter_4x1_f4( uint16_t *pred, int p_stride, int bd, uint16_t *dst, int dst_stride, bool is_compound, bool do_average, bool use_dist_wtd_comp_avg, int fwd, int bwd, int16_t gamma, const int16x8_t *tmp, int i, int sy, int j) { @@ -253,7 +259,7 @@ static INLINE void highbd_vertical_filter_4x1_f4( vst1_u16(dst16, res0); } -static INLINE void highbd_vertical_filter_8x1_f8( +static AOM_FORCE_INLINE void highbd_vertical_filter_8x1_f8( uint16_t *pred, int p_stride, int bd, uint16_t *dst, int dst_stride, bool is_compound, bool do_average, bool use_dist_wtd_comp_avg, int fwd, int bwd, int16_t gamma, const int16x8_t *tmp, int i, int sy, int j) { @@ -328,7 +334,7 @@ static INLINE void highbd_vertical_filter_8x1_f8( vst1_u16(dst16 + 4, res1); } -static INLINE void warp_affine_vertical( +static AOM_FORCE_INLINE void warp_affine_vertical( uint16_t *pred, int p_width, int p_height, int p_stride, int bd, uint16_t *dst, int dst_stride, bool is_compound, bool do_average, bool use_dist_wtd_comp_avg, int fwd, int bwd, int16_t gamma, int16_t delta, @@ -354,7 +360,7 @@ static INLINE void warp_affine_vertical( } } -static INLINE void highbd_warp_affine_common( +static AOM_FORCE_INLINE void highbd_warp_affine_common( const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, diff --git a/third_party/aom/av1/common/arm/highbd_warp_plane_sve.c b/third_party/aom/av1/common/arm/highbd_warp_plane_sve.c index 7a14f21846..87e033fd00 100644 --- a/third_party/aom/av1/common/arm/highbd_warp_plane_sve.c +++ b/third_party/aom/av1/common/arm/highbd_warp_plane_sve.c @@ -15,7 +15,7 @@ #include <arm_neon_sve_bridge.h> #include "aom_dsp/aom_dsp_common.h" -#include "aom_dsp/arm/dot_sve.h" +#include "aom_dsp/arm/aom_neon_sve_bridge.h" #include "aom_dsp/arm/mem_neon.h" #include "aom_dsp/arm/transpose_neon.h" #include "aom_ports/mem.h" @@ -24,8 +24,8 @@ #include "config/av1_rtcd.h" #include "highbd_warp_plane_neon.h" -static INLINE int16x8_t highbd_horizontal_filter_4x1_f4(uint16x8x2_t in, int bd, - int sx, int alpha) { +static AOM_FORCE_INLINE int16x8_t +highbd_horizontal_filter_4x1_f4(uint16x8x2_t in, int bd, int sx, int alpha) { int16x8_t f[4]; load_filters_4(f, sx, alpha); @@ -55,8 +55,8 @@ static INLINE int16x8_t highbd_horizontal_filter_4x1_f4(uint16x8x2_t in, int bd, return vcombine_s16(vmovn_s32(res), vdup_n_s16(0)); } -static INLINE int16x8_t highbd_horizontal_filter_8x1_f8(uint16x8x2_t in, int bd, - int sx, int alpha) { +static AOM_FORCE_INLINE int16x8_t +highbd_horizontal_filter_8x1_f8(uint16x8x2_t in, int bd, int sx, int alpha) { int16x8_t f[8]; load_filters_8(f, sx, alpha); @@ -103,8 +103,8 @@ static INLINE int16x8_t highbd_horizontal_filter_8x1_f8(uint16x8x2_t in, int bd, return vcombine_s16(vmovn_s32(res0), vmovn_s32(res1)); } -static INLINE int16x8_t highbd_horizontal_filter_4x1_f1(uint16x8x2_t in, int bd, - int sx) { +static AOM_FORCE_INLINE int16x8_t +highbd_horizontal_filter_4x1_f1(uint16x8x2_t in, int bd, int sx) { int16x8_t f = load_filters_1(sx); int16x8_t rv0 = vextq_s16(vreinterpretq_s16_u16(in.val[0]), @@ -133,8 +133,8 @@ static INLINE int16x8_t highbd_horizontal_filter_4x1_f1(uint16x8x2_t in, int bd, return vcombine_s16(vmovn_s32(res), vdup_n_s16(0)); } -static INLINE int16x8_t highbd_horizontal_filter_8x1_f1(uint16x8x2_t in, int bd, - int sx) { +static AOM_FORCE_INLINE int16x8_t +highbd_horizontal_filter_8x1_f1(uint16x8x2_t in, int bd, int sx) { int16x8_t f = load_filters_1(sx); int16x8_t rv0 = vextq_s16(vreinterpretq_s16_u16(in.val[0]), @@ -180,7 +180,8 @@ static INLINE int16x8_t highbd_horizontal_filter_8x1_f1(uint16x8x2_t in, int bd, return vcombine_s16(vmovn_s32(res0), vmovn_s32(res1)); } -static INLINE int32x4_t vertical_filter_4x1_f1(const int16x8_t *tmp, int sy) { +static AOM_FORCE_INLINE int32x4_t vertical_filter_4x1_f1(const int16x8_t *tmp, + int sy) { const int16x8_t f = load_filters_1(sy); const int16x4_t f0123 = vget_low_s16(f); const int16x4_t f4567 = vget_high_s16(f); @@ -197,7 +198,8 @@ static INLINE int32x4_t vertical_filter_4x1_f1(const int16x8_t *tmp, int sy) { return m0123; } -static INLINE int32x4x2_t vertical_filter_8x1_f1(const int16x8_t *tmp, int sy) { +static AOM_FORCE_INLINE int32x4x2_t vertical_filter_8x1_f1(const int16x8_t *tmp, + int sy) { const int16x8_t f = load_filters_1(sy); const int16x4_t f0123 = vget_low_s16(f); const int16x4_t f4567 = vget_high_s16(f); @@ -223,8 +225,8 @@ static INLINE int32x4x2_t vertical_filter_8x1_f1(const int16x8_t *tmp, int sy) { return (int32x4x2_t){ { m0123, m4567 } }; } -static INLINE int32x4_t vertical_filter_4x1_f4(const int16x8_t *tmp, int sy, - int gamma) { +static AOM_FORCE_INLINE int32x4_t vertical_filter_4x1_f4(const int16x8_t *tmp, + int sy, int gamma) { int16x8_t s0, s1, s2, s3; transpose_elems_s16_4x8( vget_low_s16(tmp[0]), vget_low_s16(tmp[1]), vget_low_s16(tmp[2]), @@ -244,8 +246,8 @@ static INLINE int32x4_t vertical_filter_4x1_f4(const int16x8_t *tmp, int sy, return vcombine_s32(vmovn_s64(m01), vmovn_s64(m23)); } -static INLINE int32x4x2_t vertical_filter_8x1_f8(const int16x8_t *tmp, int sy, - int gamma) { +static AOM_FORCE_INLINE int32x4x2_t vertical_filter_8x1_f8(const int16x8_t *tmp, + int sy, int gamma) { int16x8_t s0 = tmp[0]; int16x8_t s1 = tmp[1]; int16x8_t s2 = tmp[2]; diff --git a/third_party/aom/av1/common/arm/warp_plane_neon.c b/third_party/aom/av1/common/arm/warp_plane_neon.c index 4723154398..546aa2965b 100644 --- a/third_party/aom/av1/common/arm/warp_plane_neon.c +++ b/third_party/aom/av1/common/arm/warp_plane_neon.c @@ -11,8 +11,8 @@ #include "warp_plane_neon.h" -static INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, int sx, - int alpha) { +static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, + int sx, int alpha) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); // Loading the 8 filter taps @@ -39,8 +39,8 @@ static INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, int sx, return vreinterpretq_s16_u16(res); } -static INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in, int sx, - int alpha) { +static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in, + int sx, int alpha) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); // Loading the 8 filter taps @@ -75,7 +75,8 @@ static INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in, int sx, return vreinterpretq_s16_u16(res); } -static INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in, int sx) { +static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in, + int sx) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); int16x8_t f_s16 = @@ -101,7 +102,8 @@ static INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in, int sx) { return vreinterpretq_s16_u16(res); } -static INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in, int sx) { +static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in, + int sx) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); int16x8_t f_s16 = @@ -135,8 +137,8 @@ static INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in, int sx) { return vreinterpretq_s16_u16(res); } -static INLINE void vertical_filter_4x1_f1(const int16x8_t *src, int32x4_t *res, - int sy) { +static AOM_FORCE_INLINE void vertical_filter_4x1_f1(const int16x8_t *src, + int32x4_t *res, int sy) { int16x4_t s0 = vget_low_s16(src[0]); int16x4_t s1 = vget_low_s16(src[1]); int16x4_t s2 = vget_low_s16(src[2]); @@ -161,8 +163,9 @@ static INLINE void vertical_filter_4x1_f1(const int16x8_t *src, int32x4_t *res, *res = m0123; } -static INLINE void vertical_filter_4x1_f4(const int16x8_t *src, int32x4_t *res, - int sy, int gamma) { +static AOM_FORCE_INLINE void vertical_filter_4x1_f4(const int16x8_t *src, + int32x4_t *res, int sy, + int gamma) { int16x8_t s0, s1, s2, s3; transpose_elems_s16_4x8( vget_low_s16(src[0]), vget_low_s16(src[1]), vget_low_s16(src[2]), @@ -186,9 +189,10 @@ static INLINE void vertical_filter_4x1_f4(const int16x8_t *src, int32x4_t *res, *res = horizontal_add_4d_s32x4(m0123_pairs); } -static INLINE void vertical_filter_8x1_f1(const int16x8_t *src, - int32x4_t *res_low, - int32x4_t *res_high, int sy) { +static AOM_FORCE_INLINE void vertical_filter_8x1_f1(const int16x8_t *src, + int32x4_t *res_low, + int32x4_t *res_high, + int sy) { int16x8_t s0 = src[0]; int16x8_t s1 = src[1]; int16x8_t s2 = src[2]; @@ -223,10 +227,10 @@ static INLINE void vertical_filter_8x1_f1(const int16x8_t *src, *res_high = m4567; } -static INLINE void vertical_filter_8x1_f8(const int16x8_t *src, - int32x4_t *res_low, - int32x4_t *res_high, int sy, - int gamma) { +static AOM_FORCE_INLINE void vertical_filter_8x1_f8(const int16x8_t *src, + int32x4_t *res_low, + int32x4_t *res_high, int sy, + int gamma) { int16x8_t s0 = src[0]; int16x8_t s1 = src[1]; int16x8_t s2 = src[2]; diff --git a/third_party/aom/av1/common/arm/warp_plane_neon.h b/third_party/aom/av1/common/arm/warp_plane_neon.h index 5afd72f4ab..eece007ef3 100644 --- a/third_party/aom/av1/common/arm/warp_plane_neon.h +++ b/third_party/aom/av1/common/arm/warp_plane_neon.h @@ -24,32 +24,37 @@ #include "av1/common/warped_motion.h" #include "av1/common/scale.h" -static INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, int sx, - int alpha); +static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, + int sx, int alpha); -static INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in, int sx, - int alpha); +static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in, + int sx, int alpha); -static INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in, int sx); +static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in, + int sx); -static INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in, int sx); +static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in, + int sx); -static INLINE void vertical_filter_4x1_f1(const int16x8_t *src, int32x4_t *res, - int sy); +static AOM_FORCE_INLINE void vertical_filter_4x1_f1(const int16x8_t *src, + int32x4_t *res, int sy); -static INLINE void vertical_filter_4x1_f4(const int16x8_t *src, int32x4_t *res, - int sy, int gamma); +static AOM_FORCE_INLINE void vertical_filter_4x1_f4(const int16x8_t *src, + int32x4_t *res, int sy, + int gamma); -static INLINE void vertical_filter_8x1_f1(const int16x8_t *src, - int32x4_t *res_low, - int32x4_t *res_high, int sy); +static AOM_FORCE_INLINE void vertical_filter_8x1_f1(const int16x8_t *src, + int32x4_t *res_low, + int32x4_t *res_high, + int sy); -static INLINE void vertical_filter_8x1_f8(const int16x8_t *src, - int32x4_t *res_low, - int32x4_t *res_high, int sy, - int gamma); +static AOM_FORCE_INLINE void vertical_filter_8x1_f8(const int16x8_t *src, + int32x4_t *res_low, + int32x4_t *res_high, int sy, + int gamma); -static INLINE void load_filters_4(int16x8_t out[], int offset, int stride) { +static AOM_FORCE_INLINE void load_filters_4(int16x8_t out[], int offset, + int stride) { out[0] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 0 * stride) >> WARPEDDIFF_PREC_BITS))); out[1] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 1 * stride) >> @@ -60,7 +65,8 @@ static INLINE void load_filters_4(int16x8_t out[], int offset, int stride) { WARPEDDIFF_PREC_BITS))); } -static INLINE void load_filters_8(int16x8_t out[], int offset, int stride) { +static AOM_FORCE_INLINE void load_filters_8(int16x8_t out[], int offset, + int stride) { out[0] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 0 * stride) >> WARPEDDIFF_PREC_BITS))); out[1] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 1 * stride) >> @@ -79,16 +85,14 @@ static INLINE void load_filters_8(int16x8_t out[], int offset, int stride) { WARPEDDIFF_PREC_BITS))); } -static INLINE int clamp_iy(int iy, int height) { +static AOM_FORCE_INLINE int clamp_iy(int iy, int height) { return clamp(iy, 0, height - 1); } -static INLINE void warp_affine_horizontal(const uint8_t *ref, int width, - int height, int stride, int p_width, - int p_height, int16_t alpha, - int16_t beta, const int64_t x4, - const int64_t y4, const int i, - int16x8_t tmp[]) { +static AOM_FORCE_INLINE void warp_affine_horizontal( + const uint8_t *ref, int width, int height, int stride, int p_width, + int p_height, int16_t alpha, int16_t beta, const int64_t x4, + const int64_t y4, const int i, int16x8_t tmp[]) { const int bd = 8; const int reduce_bits_horiz = ROUND0_BITS; const int height_limit = AOMMIN(8, p_height - i) + 7; @@ -197,7 +201,7 @@ static INLINE void warp_affine_horizontal(const uint8_t *ref, int width, } } -static INLINE void warp_affine_vertical( +static AOM_FORCE_INLINE void warp_affine_vertical( uint8_t *pred, int p_width, int p_height, int p_stride, int is_compound, uint16_t *dst, int dst_stride, int do_average, int use_dist_wtd_comp_avg, int16_t gamma, int16_t delta, const int64_t y4, const int i, const int j, @@ -325,7 +329,7 @@ static INLINE void warp_affine_vertical( } } -static INLINE void av1_warp_affine_common( +static AOM_FORCE_INLINE void av1_warp_affine_common( const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, diff --git a/third_party/aom/av1/common/arm/warp_plane_neon_i8mm.c b/third_party/aom/av1/common/arm/warp_plane_neon_i8mm.c index 39e3ad99f4..22a1be17b5 100644 --- a/third_party/aom/av1/common/arm/warp_plane_neon_i8mm.c +++ b/third_party/aom/av1/common/arm/warp_plane_neon_i8mm.c @@ -17,8 +17,8 @@ DECLARE_ALIGNED(16, static const uint8_t, usdot_permute_idx[48]) = { 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 }; -static INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, int sx, - int alpha) { +static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, + int sx, int alpha) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); // Loading the 8 filter taps @@ -45,8 +45,8 @@ static INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, int sx, return vreinterpretq_s16_u16(res); } -static INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in, int sx, - int alpha) { +static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in, + int sx, int alpha) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); // Loading the 8 filter taps @@ -83,7 +83,8 @@ static INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in, int sx, return vreinterpretq_s16_u16(res); } -static INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in, int sx) { +static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in, + int sx) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); int16x8_t f_s16 = @@ -112,7 +113,8 @@ static INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in, int sx) { return vreinterpretq_s16_u16(res); } -static INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in, int sx) { +static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in, + int sx) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); int16x8_t f_s16 = @@ -149,8 +151,8 @@ static INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in, int sx) { return vreinterpretq_s16_u16(res); } -static INLINE void vertical_filter_4x1_f1(const int16x8_t *src, int32x4_t *res, - int sy) { +static AOM_FORCE_INLINE void vertical_filter_4x1_f1(const int16x8_t *src, + int32x4_t *res, int sy) { int16x4_t s0 = vget_low_s16(src[0]); int16x4_t s1 = vget_low_s16(src[1]); int16x4_t s2 = vget_low_s16(src[2]); @@ -175,8 +177,9 @@ static INLINE void vertical_filter_4x1_f1(const int16x8_t *src, int32x4_t *res, *res = m0123; } -static INLINE void vertical_filter_4x1_f4(const int16x8_t *src, int32x4_t *res, - int sy, int gamma) { +static AOM_FORCE_INLINE void vertical_filter_4x1_f4(const int16x8_t *src, + int32x4_t *res, int sy, + int gamma) { int16x8_t s0, s1, s2, s3; transpose_elems_s16_4x8( vget_low_s16(src[0]), vget_low_s16(src[1]), vget_low_s16(src[2]), @@ -200,9 +203,10 @@ static INLINE void vertical_filter_4x1_f4(const int16x8_t *src, int32x4_t *res, *res = horizontal_add_4d_s32x4(m0123_pairs); } -static INLINE void vertical_filter_8x1_f1(const int16x8_t *src, - int32x4_t *res_low, - int32x4_t *res_high, int sy) { +static AOM_FORCE_INLINE void vertical_filter_8x1_f1(const int16x8_t *src, + int32x4_t *res_low, + int32x4_t *res_high, + int sy) { int16x8_t s0 = src[0]; int16x8_t s1 = src[1]; int16x8_t s2 = src[2]; @@ -237,10 +241,10 @@ static INLINE void vertical_filter_8x1_f1(const int16x8_t *src, *res_high = m4567; } -static INLINE void vertical_filter_8x1_f8(const int16x8_t *src, - int32x4_t *res_low, - int32x4_t *res_high, int sy, - int gamma) { +static AOM_FORCE_INLINE void vertical_filter_8x1_f8(const int16x8_t *src, + int32x4_t *res_low, + int32x4_t *res_high, int sy, + int gamma) { int16x8_t s0 = src[0]; int16x8_t s1 = src[1]; int16x8_t s2 = src[2]; diff --git a/third_party/aom/av1/common/arm/warp_plane_sve.c b/third_party/aom/av1/common/arm/warp_plane_sve.c index 8a4bf5747b..c70b066174 100644 --- a/third_party/aom/av1/common/arm/warp_plane_sve.c +++ b/third_party/aom/av1/common/arm/warp_plane_sve.c @@ -11,7 +11,7 @@ #include <arm_neon.h> -#include "aom_dsp/arm/dot_sve.h" +#include "aom_dsp/arm/aom_neon_sve_bridge.h" #include "warp_plane_neon.h" DECLARE_ALIGNED(16, static const uint8_t, usdot_permute_idx[48]) = { @@ -20,8 +20,8 @@ DECLARE_ALIGNED(16, static const uint8_t, usdot_permute_idx[48]) = { 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 }; -static INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, int sx, - int alpha) { +static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, + int sx, int alpha) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); // Loading the 8 filter taps @@ -48,8 +48,8 @@ static INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, int sx, return vreinterpretq_s16_u16(res); } -static INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in, int sx, - int alpha) { +static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in, + int sx, int alpha) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); // Loading the 8 filter taps @@ -86,7 +86,8 @@ static INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in, int sx, return vreinterpretq_s16_u16(res); } -static INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in, int sx) { +static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in, + int sx) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); int16x8_t f_s16 = @@ -115,7 +116,8 @@ static INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in, int sx) { return vreinterpretq_s16_u16(res); } -static INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in, int sx) { +static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in, + int sx) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); int16x8_t f_s16 = @@ -152,8 +154,8 @@ static INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in, int sx) { return vreinterpretq_s16_u16(res); } -static INLINE void vertical_filter_4x1_f1(const int16x8_t *src, int32x4_t *res, - int sy) { +static AOM_FORCE_INLINE void vertical_filter_4x1_f1(const int16x8_t *src, + int32x4_t *res, int sy) { int16x4_t s0 = vget_low_s16(src[0]); int16x4_t s1 = vget_low_s16(src[1]); int16x4_t s2 = vget_low_s16(src[2]); @@ -178,8 +180,9 @@ static INLINE void vertical_filter_4x1_f1(const int16x8_t *src, int32x4_t *res, *res = m0123; } -static INLINE void vertical_filter_4x1_f4(const int16x8_t *src, int32x4_t *res, - int sy, int gamma) { +static AOM_FORCE_INLINE void vertical_filter_4x1_f4(const int16x8_t *src, + int32x4_t *res, int sy, + int gamma) { int16x8_t s0, s1, s2, s3; transpose_elems_s16_4x8( vget_low_s16(src[0]), vget_low_s16(src[1]), vget_low_s16(src[2]), @@ -200,9 +203,10 @@ static INLINE void vertical_filter_4x1_f4(const int16x8_t *src, int32x4_t *res, *res = vcombine_s32(vmovn_s64(m01), vmovn_s64(m23)); } -static INLINE void vertical_filter_8x1_f1(const int16x8_t *src, - int32x4_t *res_low, - int32x4_t *res_high, int sy) { +static AOM_FORCE_INLINE void vertical_filter_8x1_f1(const int16x8_t *src, + int32x4_t *res_low, + int32x4_t *res_high, + int sy) { int16x8_t s0 = src[0]; int16x8_t s1 = src[1]; int16x8_t s2 = src[2]; @@ -237,10 +241,10 @@ static INLINE void vertical_filter_8x1_f1(const int16x8_t *src, *res_high = m4567; } -static INLINE void vertical_filter_8x1_f8(const int16x8_t *src, - int32x4_t *res_low, - int32x4_t *res_high, int sy, - int gamma) { +static AOM_FORCE_INLINE void vertical_filter_8x1_f8(const int16x8_t *src, + int32x4_t *res_low, + int32x4_t *res_high, int sy, + int gamma) { int16x8_t s0 = src[0]; int16x8_t s1 = src[1]; int16x8_t s2 = src[2]; |