diff options
Diffstat (limited to 'media/libvpx/libvpx/vp9/common/arm')
7 files changed, 1667 insertions, 0 deletions
diff --git a/media/libvpx/libvpx/vp9/common/arm/neon/vp9_highbd_iht16x16_add_neon.c b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_highbd_iht16x16_add_neon.c new file mode 100644 index 0000000000..aeb7e49c10 --- /dev/null +++ b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_highbd_iht16x16_add_neon.c @@ -0,0 +1,446 @@ +/* + * Copyright (c) 2018 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include <arm_neon.h> + +#include "./vpx_dsp_rtcd.h" +#include "vp9/common/vp9_enums.h" +#include "vp9/common/arm/neon/vp9_iht_neon.h" +#include "vpx_dsp/arm/highbd_idct_neon.h" +#include "vpx_dsp/arm/idct_neon.h" +#include "vpx_dsp/arm/transpose_neon.h" +#include "vpx_dsp/inv_txfm.h" + +// Use macros to make sure argument lane is passed in as an constant integer. + +#define vmull_lane_s32_dual(in, c, lane, out) \ + do { \ + out[0].val[0] = vmull_lane_s32(vget_low_s32(in.val[0]), c, lane); \ + out[0].val[1] = vmull_lane_s32(vget_low_s32(in.val[1]), c, lane); \ + out[1].val[0] = vmull_lane_s32(vget_high_s32(in.val[0]), c, lane); \ + out[1].val[1] = vmull_lane_s32(vget_high_s32(in.val[1]), c, lane); \ + } while (0) + +#define vmlal_lane_s32_dual(in, c, lane, out) \ + do { \ + out[0].val[0] = \ + vmlal_lane_s32(out[0].val[0], vget_low_s32(in.val[0]), c, lane); \ + out[0].val[1] = \ + vmlal_lane_s32(out[0].val[1], vget_low_s32(in.val[1]), c, lane); \ + out[1].val[0] = \ + vmlal_lane_s32(out[1].val[0], vget_high_s32(in.val[0]), c, lane); \ + out[1].val[1] = \ + vmlal_lane_s32(out[1].val[1], vget_high_s32(in.val[1]), c, lane); \ + } while (0) + +#define vmlsl_lane_s32_dual(in, c, lane, out) \ + do { \ + out[0].val[0] = \ + vmlsl_lane_s32(out[0].val[0], vget_low_s32(in.val[0]), c, lane); \ + out[0].val[1] = \ + vmlsl_lane_s32(out[0].val[1], vget_low_s32(in.val[1]), c, lane); \ + out[1].val[0] = \ + vmlsl_lane_s32(out[1].val[0], vget_high_s32(in.val[0]), c, lane); \ + out[1].val[1] = \ + vmlsl_lane_s32(out[1].val[1], vget_high_s32(in.val[1]), c, lane); \ + } while (0) + +static INLINE int32x4x2_t +highbd_dct_const_round_shift_low_8(const int64x2x2_t *const in) { + int32x4x2_t out; + out.val[0] = vcombine_s32(vrshrn_n_s64(in[0].val[0], DCT_CONST_BITS), + vrshrn_n_s64(in[1].val[0], DCT_CONST_BITS)); + out.val[1] = vcombine_s32(vrshrn_n_s64(in[0].val[1], DCT_CONST_BITS), + vrshrn_n_s64(in[1].val[1], DCT_CONST_BITS)); + return out; +} + +#define highbd_iadst_half_butterfly(in, c, lane, out) \ + do { \ + int64x2x2_t _t[2]; \ + vmull_lane_s32_dual(in, c, lane, _t); \ + out = highbd_dct_const_round_shift_low_8(_t); \ + } while (0) + +#define highbd_iadst_butterfly(in0, in1, c, lane0, lane1, s0, s1) \ + do { \ + vmull_lane_s32_dual(in0, c, lane0, s0); \ + vmull_lane_s32_dual(in0, c, lane1, s1); \ + vmlal_lane_s32_dual(in1, c, lane1, s0); \ + vmlsl_lane_s32_dual(in1, c, lane0, s1); \ + } while (0) + +static INLINE int32x4x2_t vaddq_s32_dual(const int32x4x2_t in0, + const int32x4x2_t in1) { + int32x4x2_t out; + out.val[0] = vaddq_s32(in0.val[0], in1.val[0]); + out.val[1] = vaddq_s32(in0.val[1], in1.val[1]); + return out; +} + +static INLINE int64x2x2_t vaddq_s64_dual(const int64x2x2_t in0, + const int64x2x2_t in1) { + int64x2x2_t out; + out.val[0] = vaddq_s64(in0.val[0], in1.val[0]); + out.val[1] = vaddq_s64(in0.val[1], in1.val[1]); + return out; +} + +static INLINE int32x4x2_t vsubq_s32_dual(const int32x4x2_t in0, + const int32x4x2_t in1) { + int32x4x2_t out; + out.val[0] = vsubq_s32(in0.val[0], in1.val[0]); + out.val[1] = vsubq_s32(in0.val[1], in1.val[1]); + return out; +} + +static INLINE int64x2x2_t vsubq_s64_dual(const int64x2x2_t in0, + const int64x2x2_t in1) { + int64x2x2_t out; + out.val[0] = vsubq_s64(in0.val[0], in1.val[0]); + out.val[1] = vsubq_s64(in0.val[1], in1.val[1]); + return out; +} + +static INLINE int32x4x2_t vcombine_s32_dual(const int32x2x2_t in0, + const int32x2x2_t in1) { + int32x4x2_t out; + out.val[0] = vcombine_s32(in0.val[0], in1.val[0]); + out.val[1] = vcombine_s32(in0.val[1], in1.val[1]); + return out; +} + +static INLINE int32x4x2_t highbd_add_dct_const_round_shift_low_8( + const int64x2x2_t *const in0, const int64x2x2_t *const in1) { + const int64x2x2_t sum_lo = vaddq_s64_dual(in0[0], in1[0]); + const int64x2x2_t sum_hi = vaddq_s64_dual(in0[1], in1[1]); + int32x2x2_t out_lo, out_hi; + + out_lo.val[0] = vrshrn_n_s64(sum_lo.val[0], DCT_CONST_BITS); + out_lo.val[1] = vrshrn_n_s64(sum_lo.val[1], DCT_CONST_BITS); + out_hi.val[0] = vrshrn_n_s64(sum_hi.val[0], DCT_CONST_BITS); + out_hi.val[1] = vrshrn_n_s64(sum_hi.val[1], DCT_CONST_BITS); + return vcombine_s32_dual(out_lo, out_hi); +} + +static INLINE int32x4x2_t highbd_sub_dct_const_round_shift_low_8( + const int64x2x2_t *const in0, const int64x2x2_t *const in1) { + const int64x2x2_t sub_lo = vsubq_s64_dual(in0[0], in1[0]); + const int64x2x2_t sub_hi = vsubq_s64_dual(in0[1], in1[1]); + int32x2x2_t out_lo, out_hi; + + out_lo.val[0] = vrshrn_n_s64(sub_lo.val[0], DCT_CONST_BITS); + out_lo.val[1] = vrshrn_n_s64(sub_lo.val[1], DCT_CONST_BITS); + out_hi.val[0] = vrshrn_n_s64(sub_hi.val[0], DCT_CONST_BITS); + out_hi.val[1] = vrshrn_n_s64(sub_hi.val[1], DCT_CONST_BITS); + return vcombine_s32_dual(out_lo, out_hi); +} + +static INLINE int32x4x2_t vnegq_s32_dual(const int32x4x2_t in) { + int32x4x2_t out; + out.val[0] = vnegq_s32(in.val[0]); + out.val[1] = vnegq_s32(in.val[1]); + return out; +} + +static void highbd_iadst16_neon(const int32_t *input, int32_t *output, + uint16_t *dest, const int stride, + const int bd) { + const int32x4_t c_1_31_5_27 = + create_s32x4_neon(cospi_1_64, cospi_31_64, cospi_5_64, cospi_27_64); + const int32x4_t c_9_23_13_19 = + create_s32x4_neon(cospi_9_64, cospi_23_64, cospi_13_64, cospi_19_64); + const int32x4_t c_17_15_21_11 = + create_s32x4_neon(cospi_17_64, cospi_15_64, cospi_21_64, cospi_11_64); + const int32x4_t c_25_7_29_3 = + create_s32x4_neon(cospi_25_64, cospi_7_64, cospi_29_64, cospi_3_64); + const int32x4_t c_4_28_20_12 = + create_s32x4_neon(cospi_4_64, cospi_28_64, cospi_20_64, cospi_12_64); + const int32x4_t c_16_n16_8_24 = + create_s32x4_neon(cospi_16_64, -cospi_16_64, cospi_8_64, cospi_24_64); + int32x4x2_t in[16], out[16]; + int32x4x2_t x[16], t[12]; + int64x2x2_t s0[2], s1[2], s2[2], s3[2], s4[2], s5[2], s6[2], s7[2]; + int64x2x2_t s8[2], s9[2], s10[2], s11[2], s12[2], s13[2], s14[2], s15[2]; + + // Load input (16x8) + in[0].val[0] = vld1q_s32(input); + in[0].val[1] = vld1q_s32(input + 4); + input += 8; + in[8].val[0] = vld1q_s32(input); + in[8].val[1] = vld1q_s32(input + 4); + input += 8; + in[1].val[0] = vld1q_s32(input); + in[1].val[1] = vld1q_s32(input + 4); + input += 8; + in[9].val[0] = vld1q_s32(input); + in[9].val[1] = vld1q_s32(input + 4); + input += 8; + in[2].val[0] = vld1q_s32(input); + in[2].val[1] = vld1q_s32(input + 4); + input += 8; + in[10].val[0] = vld1q_s32(input); + in[10].val[1] = vld1q_s32(input + 4); + input += 8; + in[3].val[0] = vld1q_s32(input); + in[3].val[1] = vld1q_s32(input + 4); + input += 8; + in[11].val[0] = vld1q_s32(input); + in[11].val[1] = vld1q_s32(input + 4); + input += 8; + in[4].val[0] = vld1q_s32(input); + in[4].val[1] = vld1q_s32(input + 4); + input += 8; + in[12].val[0] = vld1q_s32(input); + in[12].val[1] = vld1q_s32(input + 4); + input += 8; + in[5].val[0] = vld1q_s32(input); + in[5].val[1] = vld1q_s32(input + 4); + input += 8; + in[13].val[0] = vld1q_s32(input); + in[13].val[1] = vld1q_s32(input + 4); + input += 8; + in[6].val[0] = vld1q_s32(input); + in[6].val[1] = vld1q_s32(input + 4); + input += 8; + in[14].val[0] = vld1q_s32(input); + in[14].val[1] = vld1q_s32(input + 4); + input += 8; + in[7].val[0] = vld1q_s32(input); + in[7].val[1] = vld1q_s32(input + 4); + input += 8; + in[15].val[0] = vld1q_s32(input); + in[15].val[1] = vld1q_s32(input + 4); + + // Transpose + transpose_s32_8x8(&in[0], &in[1], &in[2], &in[3], &in[4], &in[5], &in[6], + &in[7]); + transpose_s32_8x8(&in[8], &in[9], &in[10], &in[11], &in[12], &in[13], &in[14], + &in[15]); + + x[0] = in[15]; + x[1] = in[0]; + x[2] = in[13]; + x[3] = in[2]; + x[4] = in[11]; + x[5] = in[4]; + x[6] = in[9]; + x[7] = in[6]; + x[8] = in[7]; + x[9] = in[8]; + x[10] = in[5]; + x[11] = in[10]; + x[12] = in[3]; + x[13] = in[12]; + x[14] = in[1]; + x[15] = in[14]; + + // stage 1 + highbd_iadst_butterfly(x[0], x[1], vget_low_s32(c_1_31_5_27), 0, 1, s0, s1); + highbd_iadst_butterfly(x[2], x[3], vget_high_s32(c_1_31_5_27), 0, 1, s2, s3); + highbd_iadst_butterfly(x[4], x[5], vget_low_s32(c_9_23_13_19), 0, 1, s4, s5); + highbd_iadst_butterfly(x[6], x[7], vget_high_s32(c_9_23_13_19), 0, 1, s6, s7); + highbd_iadst_butterfly(x[8], x[9], vget_low_s32(c_17_15_21_11), 0, 1, s8, s9); + highbd_iadst_butterfly(x[10], x[11], vget_high_s32(c_17_15_21_11), 0, 1, s10, + s11); + highbd_iadst_butterfly(x[12], x[13], vget_low_s32(c_25_7_29_3), 0, 1, s12, + s13); + highbd_iadst_butterfly(x[14], x[15], vget_high_s32(c_25_7_29_3), 0, 1, s14, + s15); + + x[0] = highbd_add_dct_const_round_shift_low_8(s0, s8); + x[1] = highbd_add_dct_const_round_shift_low_8(s1, s9); + x[2] = highbd_add_dct_const_round_shift_low_8(s2, s10); + x[3] = highbd_add_dct_const_round_shift_low_8(s3, s11); + x[4] = highbd_add_dct_const_round_shift_low_8(s4, s12); + x[5] = highbd_add_dct_const_round_shift_low_8(s5, s13); + x[6] = highbd_add_dct_const_round_shift_low_8(s6, s14); + x[7] = highbd_add_dct_const_round_shift_low_8(s7, s15); + x[8] = highbd_sub_dct_const_round_shift_low_8(s0, s8); + x[9] = highbd_sub_dct_const_round_shift_low_8(s1, s9); + x[10] = highbd_sub_dct_const_round_shift_low_8(s2, s10); + x[11] = highbd_sub_dct_const_round_shift_low_8(s3, s11); + x[12] = highbd_sub_dct_const_round_shift_low_8(s4, s12); + x[13] = highbd_sub_dct_const_round_shift_low_8(s5, s13); + x[14] = highbd_sub_dct_const_round_shift_low_8(s6, s14); + x[15] = highbd_sub_dct_const_round_shift_low_8(s7, s15); + + // stage 2 + t[0] = x[0]; + t[1] = x[1]; + t[2] = x[2]; + t[3] = x[3]; + t[4] = x[4]; + t[5] = x[5]; + t[6] = x[6]; + t[7] = x[7]; + highbd_iadst_butterfly(x[8], x[9], vget_low_s32(c_4_28_20_12), 0, 1, s8, s9); + highbd_iadst_butterfly(x[10], x[11], vget_high_s32(c_4_28_20_12), 0, 1, s10, + s11); + highbd_iadst_butterfly(x[13], x[12], vget_low_s32(c_4_28_20_12), 1, 0, s13, + s12); + highbd_iadst_butterfly(x[15], x[14], vget_high_s32(c_4_28_20_12), 1, 0, s15, + s14); + + x[0] = vaddq_s32_dual(t[0], t[4]); + x[1] = vaddq_s32_dual(t[1], t[5]); + x[2] = vaddq_s32_dual(t[2], t[6]); + x[3] = vaddq_s32_dual(t[3], t[7]); + x[4] = vsubq_s32_dual(t[0], t[4]); + x[5] = vsubq_s32_dual(t[1], t[5]); + x[6] = vsubq_s32_dual(t[2], t[6]); + x[7] = vsubq_s32_dual(t[3], t[7]); + x[8] = highbd_add_dct_const_round_shift_low_8(s8, s12); + x[9] = highbd_add_dct_const_round_shift_low_8(s9, s13); + x[10] = highbd_add_dct_const_round_shift_low_8(s10, s14); + x[11] = highbd_add_dct_const_round_shift_low_8(s11, s15); + x[12] = highbd_sub_dct_const_round_shift_low_8(s8, s12); + x[13] = highbd_sub_dct_const_round_shift_low_8(s9, s13); + x[14] = highbd_sub_dct_const_round_shift_low_8(s10, s14); + x[15] = highbd_sub_dct_const_round_shift_low_8(s11, s15); + + // stage 3 + t[0] = x[0]; + t[1] = x[1]; + t[2] = x[2]; + t[3] = x[3]; + highbd_iadst_butterfly(x[4], x[5], vget_high_s32(c_16_n16_8_24), 0, 1, s4, + s5); + highbd_iadst_butterfly(x[7], x[6], vget_high_s32(c_16_n16_8_24), 1, 0, s7, + s6); + t[8] = x[8]; + t[9] = x[9]; + t[10] = x[10]; + t[11] = x[11]; + highbd_iadst_butterfly(x[12], x[13], vget_high_s32(c_16_n16_8_24), 0, 1, s12, + s13); + highbd_iadst_butterfly(x[15], x[14], vget_high_s32(c_16_n16_8_24), 1, 0, s15, + s14); + + x[0] = vaddq_s32_dual(t[0], t[2]); + x[1] = vaddq_s32_dual(t[1], t[3]); + x[2] = vsubq_s32_dual(t[0], t[2]); + x[3] = vsubq_s32_dual(t[1], t[3]); + x[4] = highbd_add_dct_const_round_shift_low_8(s4, s6); + x[5] = highbd_add_dct_const_round_shift_low_8(s5, s7); + x[6] = highbd_sub_dct_const_round_shift_low_8(s4, s6); + x[7] = highbd_sub_dct_const_round_shift_low_8(s5, s7); + x[8] = vaddq_s32_dual(t[8], t[10]); + x[9] = vaddq_s32_dual(t[9], t[11]); + x[10] = vsubq_s32_dual(t[8], t[10]); + x[11] = vsubq_s32_dual(t[9], t[11]); + x[12] = highbd_add_dct_const_round_shift_low_8(s12, s14); + x[13] = highbd_add_dct_const_round_shift_low_8(s13, s15); + x[14] = highbd_sub_dct_const_round_shift_low_8(s12, s14); + x[15] = highbd_sub_dct_const_round_shift_low_8(s13, s15); + + // stage 4 + { + const int32x4x2_t sum = vaddq_s32_dual(x[2], x[3]); + const int32x4x2_t sub = vsubq_s32_dual(x[2], x[3]); + highbd_iadst_half_butterfly(sum, vget_low_s32(c_16_n16_8_24), 1, x[2]); + highbd_iadst_half_butterfly(sub, vget_low_s32(c_16_n16_8_24), 0, x[3]); + } + { + const int32x4x2_t sum = vaddq_s32_dual(x[7], x[6]); + const int32x4x2_t sub = vsubq_s32_dual(x[7], x[6]); + highbd_iadst_half_butterfly(sum, vget_low_s32(c_16_n16_8_24), 0, x[6]); + highbd_iadst_half_butterfly(sub, vget_low_s32(c_16_n16_8_24), 0, x[7]); + } + { + const int32x4x2_t sum = vaddq_s32_dual(x[11], x[10]); + const int32x4x2_t sub = vsubq_s32_dual(x[11], x[10]); + highbd_iadst_half_butterfly(sum, vget_low_s32(c_16_n16_8_24), 0, x[10]); + highbd_iadst_half_butterfly(sub, vget_low_s32(c_16_n16_8_24), 0, x[11]); + } + { + const int32x4x2_t sum = vaddq_s32_dual(x[14], x[15]); + const int32x4x2_t sub = vsubq_s32_dual(x[14], x[15]); + highbd_iadst_half_butterfly(sum, vget_low_s32(c_16_n16_8_24), 1, x[14]); + highbd_iadst_half_butterfly(sub, vget_low_s32(c_16_n16_8_24), 0, x[15]); + } + + out[0] = x[0]; + out[1] = vnegq_s32_dual(x[8]); + out[2] = x[12]; + out[3] = vnegq_s32_dual(x[4]); + out[4] = x[6]; + out[5] = x[14]; + out[6] = x[10]; + out[7] = x[2]; + out[8] = x[3]; + out[9] = x[11]; + out[10] = x[15]; + out[11] = x[7]; + out[12] = x[5]; + out[13] = vnegq_s32_dual(x[13]); + out[14] = x[9]; + out[15] = vnegq_s32_dual(x[1]); + + if (output) { + highbd_idct16x16_store_pass1(out, output); + } else { + highbd_idct16x16_add_store(out, dest, stride, bd); + } +} + +typedef void (*highbd_iht_1d)(const int32_t *input, int32_t *output, + uint16_t *dest, const int stride, const int bd); + +typedef struct { + highbd_iht_1d cols, rows; // vertical and horizontal +} highbd_iht_2d; + +void vp9_highbd_iht16x16_256_add_neon(const tran_low_t *input, uint16_t *dest, + int stride, int tx_type, int bd) { + if (bd == 8) { + static const iht_2d IHT_16[] = { + { vpx_idct16x16_256_add_half1d, + vpx_idct16x16_256_add_half1d }, // DCT_DCT = 0 + { vpx_iadst16x16_256_add_half1d, + vpx_idct16x16_256_add_half1d }, // ADST_DCT = 1 + { vpx_idct16x16_256_add_half1d, + vpx_iadst16x16_256_add_half1d }, // DCT_ADST = 2 + { vpx_iadst16x16_256_add_half1d, + vpx_iadst16x16_256_add_half1d } // ADST_ADST = 3 + }; + const iht_2d ht = IHT_16[tx_type]; + int16_t row_output[16 * 16]; + + // pass 1 + ht.rows(input, row_output, dest, stride, 1); // upper 8 rows + ht.rows(input + 8 * 16, row_output + 8, dest, stride, 1); // lower 8 rows + + // pass 2 + ht.cols(row_output, NULL, dest, stride, 1); // left 8 columns + ht.cols(row_output + 16 * 8, NULL, dest + 8, stride, 1); // right 8 columns + } else { + static const highbd_iht_2d IHT_16[] = { + { vpx_highbd_idct16x16_256_add_half1d, + vpx_highbd_idct16x16_256_add_half1d }, // DCT_DCT = 0 + { highbd_iadst16_neon, + vpx_highbd_idct16x16_256_add_half1d }, // ADST_DCT = 1 + { vpx_highbd_idct16x16_256_add_half1d, + highbd_iadst16_neon }, // DCT_ADST = 2 + { highbd_iadst16_neon, highbd_iadst16_neon } // ADST_ADST = 3 + }; + const highbd_iht_2d ht = IHT_16[tx_type]; + int32_t row_output[16 * 16]; + + // pass 1 + ht.rows(input, row_output, dest, stride, bd); // upper 8 rows + ht.rows(input + 8 * 16, row_output + 8, dest, stride, bd); // lower 8 rows + + // pass 2 + ht.cols(row_output, NULL, dest, stride, bd); // left 8 columns + ht.cols(row_output + 8 * 16, NULL, dest + 8, stride, + bd); // right 8 columns + } +} diff --git a/media/libvpx/libvpx/vp9/common/arm/neon/vp9_highbd_iht4x4_add_neon.c b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_highbd_iht4x4_add_neon.c new file mode 100644 index 0000000000..52c4f1937d --- /dev/null +++ b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_highbd_iht4x4_add_neon.c @@ -0,0 +1,181 @@ +/* + * Copyright (c) 2018 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include <arm_neon.h> +#include <assert.h> + +#include "./vp9_rtcd.h" +#include "./vpx_config.h" +#include "vp9/common/vp9_common.h" +#include "vp9/common/arm/neon/vp9_iht_neon.h" +#include "vpx_dsp/arm/highbd_idct_neon.h" +#include "vpx_dsp/arm/idct_neon.h" +#include "vpx_dsp/arm/mem_neon.h" +#include "vpx_dsp/txfm_common.h" + +static INLINE void highbd_iadst4(int32x4_t *const io) { + const int32_t sinpis[4] = { sinpi_1_9, sinpi_2_9, sinpi_3_9, sinpi_4_9 }; + const int32x4_t sinpi = vld1q_s32(sinpis); + int64x2x2_t s[7], t[4]; + int32x4_t s7; + + s[0].val[0] = vmull_lane_s32(vget_low_s32(io[0]), vget_low_s32(sinpi), 0); + s[0].val[1] = vmull_lane_s32(vget_high_s32(io[0]), vget_low_s32(sinpi), 0); + s[1].val[0] = vmull_lane_s32(vget_low_s32(io[0]), vget_low_s32(sinpi), 1); + s[1].val[1] = vmull_lane_s32(vget_high_s32(io[0]), vget_low_s32(sinpi), 1); + s[2].val[0] = vmull_lane_s32(vget_low_s32(io[1]), vget_high_s32(sinpi), 0); + s[2].val[1] = vmull_lane_s32(vget_high_s32(io[1]), vget_high_s32(sinpi), 0); + s[3].val[0] = vmull_lane_s32(vget_low_s32(io[2]), vget_high_s32(sinpi), 1); + s[3].val[1] = vmull_lane_s32(vget_high_s32(io[2]), vget_high_s32(sinpi), 1); + s[4].val[0] = vmull_lane_s32(vget_low_s32(io[2]), vget_low_s32(sinpi), 0); + s[4].val[1] = vmull_lane_s32(vget_high_s32(io[2]), vget_low_s32(sinpi), 0); + s[5].val[0] = vmull_lane_s32(vget_low_s32(io[3]), vget_low_s32(sinpi), 1); + s[5].val[1] = vmull_lane_s32(vget_high_s32(io[3]), vget_low_s32(sinpi), 1); + s[6].val[0] = vmull_lane_s32(vget_low_s32(io[3]), vget_high_s32(sinpi), 1); + s[6].val[1] = vmull_lane_s32(vget_high_s32(io[3]), vget_high_s32(sinpi), 1); + s7 = vsubq_s32(io[0], io[2]); + s7 = vaddq_s32(s7, io[3]); + + s[0].val[0] = vaddq_s64(s[0].val[0], s[3].val[0]); + s[0].val[1] = vaddq_s64(s[0].val[1], s[3].val[1]); + s[0].val[0] = vaddq_s64(s[0].val[0], s[5].val[0]); + s[0].val[1] = vaddq_s64(s[0].val[1], s[5].val[1]); + s[1].val[0] = vsubq_s64(s[1].val[0], s[4].val[0]); + s[1].val[1] = vsubq_s64(s[1].val[1], s[4].val[1]); + s[1].val[0] = vsubq_s64(s[1].val[0], s[6].val[0]); + s[1].val[1] = vsubq_s64(s[1].val[1], s[6].val[1]); + s[3] = s[2]; + s[2].val[0] = vmull_lane_s32(vget_low_s32(s7), vget_high_s32(sinpi), 0); + s[2].val[1] = vmull_lane_s32(vget_high_s32(s7), vget_high_s32(sinpi), 0); + + t[0].val[0] = vaddq_s64(s[0].val[0], s[3].val[0]); + t[0].val[1] = vaddq_s64(s[0].val[1], s[3].val[1]); + t[1].val[0] = vaddq_s64(s[1].val[0], s[3].val[0]); + t[1].val[1] = vaddq_s64(s[1].val[1], s[3].val[1]); + t[2] = s[2]; + t[3].val[0] = vaddq_s64(s[0].val[0], s[1].val[0]); + t[3].val[1] = vaddq_s64(s[0].val[1], s[1].val[1]); + t[3].val[0] = vsubq_s64(t[3].val[0], s[3].val[0]); + t[3].val[1] = vsubq_s64(t[3].val[1], s[3].val[1]); + io[0] = vcombine_s32(vrshrn_n_s64(t[0].val[0], DCT_CONST_BITS), + vrshrn_n_s64(t[0].val[1], DCT_CONST_BITS)); + io[1] = vcombine_s32(vrshrn_n_s64(t[1].val[0], DCT_CONST_BITS), + vrshrn_n_s64(t[1].val[1], DCT_CONST_BITS)); + io[2] = vcombine_s32(vrshrn_n_s64(t[2].val[0], DCT_CONST_BITS), + vrshrn_n_s64(t[2].val[1], DCT_CONST_BITS)); + io[3] = vcombine_s32(vrshrn_n_s64(t[3].val[0], DCT_CONST_BITS), + vrshrn_n_s64(t[3].val[1], DCT_CONST_BITS)); +} + +void vp9_highbd_iht4x4_16_add_neon(const tran_low_t *input, uint16_t *dest, + int stride, int tx_type, int bd) { + const int16x8_t max = vdupq_n_s16((1 << bd) - 1); + int16x8_t a[2]; + int32x4_t c[4]; + + c[0] = vld1q_s32(input); + c[1] = vld1q_s32(input + 4); + c[2] = vld1q_s32(input + 8); + c[3] = vld1q_s32(input + 12); + + if (bd == 8) { + a[0] = vcombine_s16(vmovn_s32(c[0]), vmovn_s32(c[1])); + a[1] = vcombine_s16(vmovn_s32(c[2]), vmovn_s32(c[3])); + transpose_s16_4x4q(&a[0], &a[1]); + + switch (tx_type) { + case DCT_DCT: + idct4x4_16_kernel_bd8(a); + a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1])); + transpose_s16_4x4q(&a[0], &a[1]); + idct4x4_16_kernel_bd8(a); + a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1])); + break; + + case ADST_DCT: + idct4x4_16_kernel_bd8(a); + a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1])); + transpose_s16_4x4q(&a[0], &a[1]); + iadst4(a); + break; + + case DCT_ADST: + iadst4(a); + transpose_s16_4x4q(&a[0], &a[1]); + idct4x4_16_kernel_bd8(a); + a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1])); + break; + + default: + assert(tx_type == ADST_ADST); + iadst4(a); + transpose_s16_4x4q(&a[0], &a[1]); + iadst4(a); + break; + } + a[0] = vrshrq_n_s16(a[0], 4); + a[1] = vrshrq_n_s16(a[1], 4); + } else { + switch (tx_type) { + case DCT_DCT: { + const int32x4_t cospis = vld1q_s32(kCospi32); + + if (bd == 10) { + idct4x4_16_kernel_bd10(cospis, c); + idct4x4_16_kernel_bd10(cospis, c); + } else { + idct4x4_16_kernel_bd12(cospis, c); + idct4x4_16_kernel_bd12(cospis, c); + } + break; + } + + case ADST_DCT: { + const int32x4_t cospis = vld1q_s32(kCospi32); + + if (bd == 10) { + idct4x4_16_kernel_bd10(cospis, c); + } else { + idct4x4_16_kernel_bd12(cospis, c); + } + transpose_s32_4x4(&c[0], &c[1], &c[2], &c[3]); + highbd_iadst4(c); + break; + } + + case DCT_ADST: { + const int32x4_t cospis = vld1q_s32(kCospi32); + + transpose_s32_4x4(&c[0], &c[1], &c[2], &c[3]); + highbd_iadst4(c); + if (bd == 10) { + idct4x4_16_kernel_bd10(cospis, c); + } else { + idct4x4_16_kernel_bd12(cospis, c); + } + break; + } + + default: { + assert(tx_type == ADST_ADST); + transpose_s32_4x4(&c[0], &c[1], &c[2], &c[3]); + highbd_iadst4(c); + transpose_s32_4x4(&c[0], &c[1], &c[2], &c[3]); + highbd_iadst4(c); + break; + } + } + a[0] = vcombine_s16(vqrshrn_n_s32(c[0], 4), vqrshrn_n_s32(c[1], 4)); + a[1] = vcombine_s16(vqrshrn_n_s32(c[2], 4), vqrshrn_n_s32(c[3], 4)); + } + + highbd_idct4x4_1_add_kernel1(&dest, stride, a[0], max); + highbd_idct4x4_1_add_kernel1(&dest, stride, a[1], max); +} diff --git a/media/libvpx/libvpx/vp9/common/arm/neon/vp9_highbd_iht8x8_add_neon.c b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_highbd_iht8x8_add_neon.c new file mode 100644 index 0000000000..2232c6841c --- /dev/null +++ b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_highbd_iht8x8_add_neon.c @@ -0,0 +1,345 @@ +/* + * Copyright (c) 2018 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include <arm_neon.h> + +#include "./vpx_dsp_rtcd.h" +#include "vp9/common/vp9_enums.h" +#include "vp9/common/arm/neon/vp9_iht_neon.h" +#include "vpx_dsp/arm/highbd_idct_neon.h" +#include "vpx_dsp/arm/idct_neon.h" +#include "vpx_dsp/arm/transpose_neon.h" +#include "vpx_dsp/inv_txfm.h" + +static INLINE void highbd_iadst_half_butterfly_neon(int32x4_t *const x, + const int32x2_t c) { + const int32x4_t sum = vaddq_s32(x[0], x[1]); + const int32x4_t sub = vsubq_s32(x[0], x[1]); + const int64x2_t t0_lo = vmull_lane_s32(vget_low_s32(sum), c, 0); + const int64x2_t t1_lo = vmull_lane_s32(vget_low_s32(sub), c, 0); + const int64x2_t t0_hi = vmull_lane_s32(vget_high_s32(sum), c, 0); + const int64x2_t t1_hi = vmull_lane_s32(vget_high_s32(sub), c, 0); + const int32x2_t out0_lo = vrshrn_n_s64(t0_lo, DCT_CONST_BITS); + const int32x2_t out1_lo = vrshrn_n_s64(t1_lo, DCT_CONST_BITS); + const int32x2_t out0_hi = vrshrn_n_s64(t0_hi, DCT_CONST_BITS); + const int32x2_t out1_hi = vrshrn_n_s64(t1_hi, DCT_CONST_BITS); + + x[0] = vcombine_s32(out0_lo, out0_hi); + x[1] = vcombine_s32(out1_lo, out1_hi); +} + +static INLINE void highbd_iadst_butterfly_lane_0_1_neon(const int32x4_t in0, + const int32x4_t in1, + const int32x2_t c, + int64x2_t *const s0, + int64x2_t *const s1) { + const int64x2_t t0_lo = vmull_lane_s32(vget_low_s32(in0), c, 0); + const int64x2_t t1_lo = vmull_lane_s32(vget_low_s32(in0), c, 1); + const int64x2_t t0_hi = vmull_lane_s32(vget_high_s32(in0), c, 0); + const int64x2_t t1_hi = vmull_lane_s32(vget_high_s32(in0), c, 1); + + s0[0] = vmlal_lane_s32(t0_lo, vget_low_s32(in1), c, 1); + s1[0] = vmlsl_lane_s32(t1_lo, vget_low_s32(in1), c, 0); + s0[1] = vmlal_lane_s32(t0_hi, vget_high_s32(in1), c, 1); + s1[1] = vmlsl_lane_s32(t1_hi, vget_high_s32(in1), c, 0); +} + +static INLINE void highbd_iadst_butterfly_lane_1_0_neon(const int32x4_t in0, + const int32x4_t in1, + const int32x2_t c, + int64x2_t *const s0, + int64x2_t *const s1) { + const int64x2_t t0_lo = vmull_lane_s32(vget_low_s32(in0), c, 1); + const int64x2_t t1_lo = vmull_lane_s32(vget_low_s32(in0), c, 0); + const int64x2_t t0_hi = vmull_lane_s32(vget_high_s32(in0), c, 1); + const int64x2_t t1_hi = vmull_lane_s32(vget_high_s32(in0), c, 0); + + s0[0] = vmlal_lane_s32(t0_lo, vget_low_s32(in1), c, 0); + s1[0] = vmlsl_lane_s32(t1_lo, vget_low_s32(in1), c, 1); + s0[1] = vmlal_lane_s32(t0_hi, vget_high_s32(in1), c, 0); + s1[1] = vmlsl_lane_s32(t1_hi, vget_high_s32(in1), c, 1); +} + +static INLINE int32x4_t highbd_add_dct_const_round_shift_low_8( + const int64x2_t *const in0, const int64x2_t *const in1) { + const int64x2_t sum_lo = vaddq_s64(in0[0], in1[0]); + const int64x2_t sum_hi = vaddq_s64(in0[1], in1[1]); + const int32x2_t out_lo = vrshrn_n_s64(sum_lo, DCT_CONST_BITS); + const int32x2_t out_hi = vrshrn_n_s64(sum_hi, DCT_CONST_BITS); + return vcombine_s32(out_lo, out_hi); +} + +static INLINE int32x4_t highbd_sub_dct_const_round_shift_low_8( + const int64x2_t *const in0, const int64x2_t *const in1) { + const int64x2_t sub_lo = vsubq_s64(in0[0], in1[0]); + const int64x2_t sub_hi = vsubq_s64(in0[1], in1[1]); + const int32x2_t out_lo = vrshrn_n_s64(sub_lo, DCT_CONST_BITS); + const int32x2_t out_hi = vrshrn_n_s64(sub_hi, DCT_CONST_BITS); + return vcombine_s32(out_lo, out_hi); +} + +static INLINE void highbd_iadst8(int32x4_t *const io0, int32x4_t *const io1, + int32x4_t *const io2, int32x4_t *const io3, + int32x4_t *const io4, int32x4_t *const io5, + int32x4_t *const io6, int32x4_t *const io7) { + const int32x4_t c0 = + create_s32x4_neon(cospi_2_64, cospi_30_64, cospi_10_64, cospi_22_64); + const int32x4_t c1 = + create_s32x4_neon(cospi_18_64, cospi_14_64, cospi_26_64, cospi_6_64); + const int32x4_t c2 = + create_s32x4_neon(cospi_16_64, 0, cospi_8_64, cospi_24_64); + int32x4_t x[8], t[4]; + int64x2_t s[8][2]; + + x[0] = *io7; + x[1] = *io0; + x[2] = *io5; + x[3] = *io2; + x[4] = *io3; + x[5] = *io4; + x[6] = *io1; + x[7] = *io6; + + // stage 1 + highbd_iadst_butterfly_lane_0_1_neon(x[0], x[1], vget_low_s32(c0), s[0], + s[1]); + highbd_iadst_butterfly_lane_0_1_neon(x[2], x[3], vget_high_s32(c0), s[2], + s[3]); + highbd_iadst_butterfly_lane_0_1_neon(x[4], x[5], vget_low_s32(c1), s[4], + s[5]); + highbd_iadst_butterfly_lane_0_1_neon(x[6], x[7], vget_high_s32(c1), s[6], + s[7]); + + x[0] = highbd_add_dct_const_round_shift_low_8(s[0], s[4]); + x[1] = highbd_add_dct_const_round_shift_low_8(s[1], s[5]); + x[2] = highbd_add_dct_const_round_shift_low_8(s[2], s[6]); + x[3] = highbd_add_dct_const_round_shift_low_8(s[3], s[7]); + x[4] = highbd_sub_dct_const_round_shift_low_8(s[0], s[4]); + x[5] = highbd_sub_dct_const_round_shift_low_8(s[1], s[5]); + x[6] = highbd_sub_dct_const_round_shift_low_8(s[2], s[6]); + x[7] = highbd_sub_dct_const_round_shift_low_8(s[3], s[7]); + + // stage 2 + t[0] = x[0]; + t[1] = x[1]; + t[2] = x[2]; + t[3] = x[3]; + highbd_iadst_butterfly_lane_0_1_neon(x[4], x[5], vget_high_s32(c2), s[4], + s[5]); + highbd_iadst_butterfly_lane_1_0_neon(x[7], x[6], vget_high_s32(c2), s[7], + s[6]); + + x[0] = vaddq_s32(t[0], t[2]); + x[1] = vaddq_s32(t[1], t[3]); + x[2] = vsubq_s32(t[0], t[2]); + x[3] = vsubq_s32(t[1], t[3]); + x[4] = highbd_add_dct_const_round_shift_low_8(s[4], s[6]); + x[5] = highbd_add_dct_const_round_shift_low_8(s[5], s[7]); + x[6] = highbd_sub_dct_const_round_shift_low_8(s[4], s[6]); + x[7] = highbd_sub_dct_const_round_shift_low_8(s[5], s[7]); + + // stage 3 + highbd_iadst_half_butterfly_neon(x + 2, vget_low_s32(c2)); + highbd_iadst_half_butterfly_neon(x + 6, vget_low_s32(c2)); + + *io0 = x[0]; + *io1 = vnegq_s32(x[4]); + *io2 = x[6]; + *io3 = vnegq_s32(x[2]); + *io4 = x[3]; + *io5 = vnegq_s32(x[7]); + *io6 = x[5]; + *io7 = vnegq_s32(x[1]); +} + +void vp9_highbd_iht8x8_64_add_neon(const tran_low_t *input, uint16_t *dest, + int stride, int tx_type, int bd) { + int32x4_t a[16]; + int16x8_t c[8]; + + a[0] = vld1q_s32(input); + a[1] = vld1q_s32(input + 4); + a[2] = vld1q_s32(input + 8); + a[3] = vld1q_s32(input + 12); + a[4] = vld1q_s32(input + 16); + a[5] = vld1q_s32(input + 20); + a[6] = vld1q_s32(input + 24); + a[7] = vld1q_s32(input + 28); + a[8] = vld1q_s32(input + 32); + a[9] = vld1q_s32(input + 36); + a[10] = vld1q_s32(input + 40); + a[11] = vld1q_s32(input + 44); + a[12] = vld1q_s32(input + 48); + a[13] = vld1q_s32(input + 52); + a[14] = vld1q_s32(input + 56); + a[15] = vld1q_s32(input + 60); + + if (bd == 8) { + c[0] = vcombine_s16(vmovn_s32(a[0]), vmovn_s32(a[1])); + c[1] = vcombine_s16(vmovn_s32(a[2]), vmovn_s32(a[3])); + c[2] = vcombine_s16(vmovn_s32(a[4]), vmovn_s32(a[5])); + c[3] = vcombine_s16(vmovn_s32(a[6]), vmovn_s32(a[7])); + c[4] = vcombine_s16(vmovn_s32(a[8]), vmovn_s32(a[9])); + c[5] = vcombine_s16(vmovn_s32(a[10]), vmovn_s32(a[11])); + c[6] = vcombine_s16(vmovn_s32(a[12]), vmovn_s32(a[13])); + c[7] = vcombine_s16(vmovn_s32(a[14]), vmovn_s32(a[15])); + + switch (tx_type) { + case DCT_DCT: { + const int16x8_t cospis = vld1q_s16(kCospi); + const int16x4_t cospis0 = vget_low_s16(cospis); // cospi 0, 8, 16, 24 + const int16x4_t cospis1 = vget_high_s16(cospis); // cospi 4, 12, 20, 28 + + idct8x8_64_1d_bd8(cospis0, cospis1, c); + idct8x8_64_1d_bd8(cospis0, cospis1, c); + break; + } + + case ADST_DCT: { + const int16x8_t cospis = vld1q_s16(kCospi); + const int16x4_t cospis0 = vget_low_s16(cospis); // cospi 0, 8, 16, 24 + const int16x4_t cospis1 = vget_high_s16(cospis); // cospi 4, 12, 20, 28 + + idct8x8_64_1d_bd8(cospis0, cospis1, c); + transpose_s16_8x8(&c[0], &c[1], &c[2], &c[3], &c[4], &c[5], &c[6], + &c[7]); + iadst8(c); + break; + } + + case DCT_ADST: { + const int16x8_t cospis = vld1q_s16(kCospi); + const int16x4_t cospis0 = vget_low_s16(cospis); // cospi 0, 8, 16, 24 + const int16x4_t cospis1 = vget_high_s16(cospis); // cospi 4, 12, 20, 28 + + transpose_s16_8x8(&c[0], &c[1], &c[2], &c[3], &c[4], &c[5], &c[6], + &c[7]); + iadst8(c); + idct8x8_64_1d_bd8(cospis0, cospis1, c); + break; + } + + default: { + transpose_s16_8x8(&c[0], &c[1], &c[2], &c[3], &c[4], &c[5], &c[6], + &c[7]); + iadst8(c); + transpose_s16_8x8(&c[0], &c[1], &c[2], &c[3], &c[4], &c[5], &c[6], + &c[7]); + iadst8(c); + break; + } + } + + c[0] = vrshrq_n_s16(c[0], 5); + c[1] = vrshrq_n_s16(c[1], 5); + c[2] = vrshrq_n_s16(c[2], 5); + c[3] = vrshrq_n_s16(c[3], 5); + c[4] = vrshrq_n_s16(c[4], 5); + c[5] = vrshrq_n_s16(c[5], 5); + c[6] = vrshrq_n_s16(c[6], 5); + c[7] = vrshrq_n_s16(c[7], 5); + } else { + switch (tx_type) { + case DCT_DCT: { + const int32x4_t cospis0 = vld1q_s32(kCospi32); // cospi 0, 8, 16, 24 + const int32x4_t cospis1 = + vld1q_s32(kCospi32 + 4); // cospi 4, 12, 20, 28 + + if (bd == 10) { + idct8x8_64_half1d_bd10(cospis0, cospis1, &a[0], &a[1], &a[2], &a[3], + &a[4], &a[5], &a[6], &a[7]); + idct8x8_64_half1d_bd10(cospis0, cospis1, &a[8], &a[9], &a[10], &a[11], + &a[12], &a[13], &a[14], &a[15]); + idct8x8_64_half1d_bd10(cospis0, cospis1, &a[0], &a[8], &a[1], &a[9], + &a[2], &a[10], &a[3], &a[11]); + idct8x8_64_half1d_bd10(cospis0, cospis1, &a[4], &a[12], &a[5], &a[13], + &a[6], &a[14], &a[7], &a[15]); + } else { + idct8x8_64_half1d_bd12(cospis0, cospis1, &a[0], &a[1], &a[2], &a[3], + &a[4], &a[5], &a[6], &a[7]); + idct8x8_64_half1d_bd12(cospis0, cospis1, &a[8], &a[9], &a[10], &a[11], + &a[12], &a[13], &a[14], &a[15]); + idct8x8_64_half1d_bd12(cospis0, cospis1, &a[0], &a[8], &a[1], &a[9], + &a[2], &a[10], &a[3], &a[11]); + idct8x8_64_half1d_bd12(cospis0, cospis1, &a[4], &a[12], &a[5], &a[13], + &a[6], &a[14], &a[7], &a[15]); + } + break; + } + + case ADST_DCT: { + const int32x4_t cospis0 = vld1q_s32(kCospi32); // cospi 0, 8, 16, 24 + const int32x4_t cospis1 = + vld1q_s32(kCospi32 + 4); // cospi 4, 12, 20, 28 + + idct8x8_64_half1d_bd12(cospis0, cospis1, &a[0], &a[1], &a[2], &a[3], + &a[4], &a[5], &a[6], &a[7]); + idct8x8_64_half1d_bd12(cospis0, cospis1, &a[8], &a[9], &a[10], &a[11], + &a[12], &a[13], &a[14], &a[15]); + transpose_s32_8x4(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3], + &a[11]); + highbd_iadst8(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3], &a[11]); + transpose_s32_8x4(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7], + &a[15]); + highbd_iadst8(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7], + &a[15]); + break; + } + + case DCT_ADST: { + const int32x4_t cospis0 = vld1q_s32(kCospi32); // cospi 0, 8, 16, 24 + const int32x4_t cospis1 = + vld1q_s32(kCospi32 + 4); // cospi 4, 12, 20, 28 + + transpose_s32_8x4(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], + &a[7]); + highbd_iadst8(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], &a[7]); + transpose_s32_8x4(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13], &a[14], + &a[15]); + highbd_iadst8(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13], &a[14], + &a[15]); + idct8x8_64_half1d_bd12(cospis0, cospis1, &a[0], &a[8], &a[1], &a[9], + &a[2], &a[10], &a[3], &a[11]); + idct8x8_64_half1d_bd12(cospis0, cospis1, &a[4], &a[12], &a[5], &a[13], + &a[6], &a[14], &a[7], &a[15]); + break; + } + + default: { + assert(tx_type == ADST_ADST); + transpose_s32_8x4(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], + &a[7]); + highbd_iadst8(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], &a[7]); + transpose_s32_8x4(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13], &a[14], + &a[15]); + highbd_iadst8(&a[8], &a[9], &a[10], &a[11], &a[12], &a[13], &a[14], + &a[15]); + transpose_s32_8x4(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3], + &a[11]); + highbd_iadst8(&a[0], &a[8], &a[1], &a[9], &a[2], &a[10], &a[3], &a[11]); + transpose_s32_8x4(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7], + &a[15]); + highbd_iadst8(&a[4], &a[12], &a[5], &a[13], &a[6], &a[14], &a[7], + &a[15]); + break; + } + } + + c[0] = vcombine_s16(vrshrn_n_s32(a[0], 5), vrshrn_n_s32(a[4], 5)); + c[1] = vcombine_s16(vrshrn_n_s32(a[8], 5), vrshrn_n_s32(a[12], 5)); + c[2] = vcombine_s16(vrshrn_n_s32(a[1], 5), vrshrn_n_s32(a[5], 5)); + c[3] = vcombine_s16(vrshrn_n_s32(a[9], 5), vrshrn_n_s32(a[13], 5)); + c[4] = vcombine_s16(vrshrn_n_s32(a[2], 5), vrshrn_n_s32(a[6], 5)); + c[5] = vcombine_s16(vrshrn_n_s32(a[10], 5), vrshrn_n_s32(a[14], 5)); + c[6] = vcombine_s16(vrshrn_n_s32(a[3], 5), vrshrn_n_s32(a[7], 5)); + c[7] = vcombine_s16(vrshrn_n_s32(a[11], 5), vrshrn_n_s32(a[15], 5)); + } + highbd_add8x8(c, dest, stride, bd); +} diff --git a/media/libvpx/libvpx/vp9/common/arm/neon/vp9_iht16x16_add_neon.c b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_iht16x16_add_neon.c new file mode 100644 index 0000000000..db72ff1161 --- /dev/null +++ b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_iht16x16_add_neon.c @@ -0,0 +1,279 @@ +/* + * Copyright (c) 2018 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include <arm_neon.h> +#include <assert.h> + +#include "./vp9_rtcd.h" +#include "./vpx_config.h" +#include "vp9/common/vp9_common.h" +#include "vp9/common/arm/neon/vp9_iht_neon.h" +#include "vpx_dsp/arm/idct_neon.h" +#include "vpx_dsp/arm/mem_neon.h" +#include "vpx_dsp/arm/transpose_neon.h" + +void vpx_iadst16x16_256_add_half1d(const void *const input, int16_t *output, + void *const dest, const int stride, + const int highbd_flag) { + int16x8_t in[16], out[16]; + const int16x4_t c_1_31_5_27 = + create_s16x4_neon(cospi_1_64, cospi_31_64, cospi_5_64, cospi_27_64); + const int16x4_t c_9_23_13_19 = + create_s16x4_neon(cospi_9_64, cospi_23_64, cospi_13_64, cospi_19_64); + const int16x4_t c_17_15_21_11 = + create_s16x4_neon(cospi_17_64, cospi_15_64, cospi_21_64, cospi_11_64); + const int16x4_t c_25_7_29_3 = + create_s16x4_neon(cospi_25_64, cospi_7_64, cospi_29_64, cospi_3_64); + const int16x4_t c_4_28_20_12 = + create_s16x4_neon(cospi_4_64, cospi_28_64, cospi_20_64, cospi_12_64); + const int16x4_t c_16_n16_8_24 = + create_s16x4_neon(cospi_16_64, -cospi_16_64, cospi_8_64, cospi_24_64); + int16x8_t x[16], t[12]; + int32x4_t s0[2], s1[2], s2[2], s3[2], s4[2], s5[2], s6[2], s7[2]; + int32x4_t s8[2], s9[2], s10[2], s11[2], s12[2], s13[2], s14[2], s15[2]; + + // Load input (16x8) + if (output) { + const tran_low_t *inputT = (const tran_low_t *)input; + in[0] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[8] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[1] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[9] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[2] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[10] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[3] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[11] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[4] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[12] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[5] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[13] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[6] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[14] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[7] = load_tran_low_to_s16q(inputT); + inputT += 8; + in[15] = load_tran_low_to_s16q(inputT); + } else { + const int16_t *inputT = (const int16_t *)input; + in[0] = vld1q_s16(inputT); + inputT += 8; + in[8] = vld1q_s16(inputT); + inputT += 8; + in[1] = vld1q_s16(inputT); + inputT += 8; + in[9] = vld1q_s16(inputT); + inputT += 8; + in[2] = vld1q_s16(inputT); + inputT += 8; + in[10] = vld1q_s16(inputT); + inputT += 8; + in[3] = vld1q_s16(inputT); + inputT += 8; + in[11] = vld1q_s16(inputT); + inputT += 8; + in[4] = vld1q_s16(inputT); + inputT += 8; + in[12] = vld1q_s16(inputT); + inputT += 8; + in[5] = vld1q_s16(inputT); + inputT += 8; + in[13] = vld1q_s16(inputT); + inputT += 8; + in[6] = vld1q_s16(inputT); + inputT += 8; + in[14] = vld1q_s16(inputT); + inputT += 8; + in[7] = vld1q_s16(inputT); + inputT += 8; + in[15] = vld1q_s16(inputT); + } + + // Transpose + transpose_s16_8x8(&in[0], &in[1], &in[2], &in[3], &in[4], &in[5], &in[6], + &in[7]); + transpose_s16_8x8(&in[8], &in[9], &in[10], &in[11], &in[12], &in[13], &in[14], + &in[15]); + + x[0] = in[15]; + x[1] = in[0]; + x[2] = in[13]; + x[3] = in[2]; + x[4] = in[11]; + x[5] = in[4]; + x[6] = in[9]; + x[7] = in[6]; + x[8] = in[7]; + x[9] = in[8]; + x[10] = in[5]; + x[11] = in[10]; + x[12] = in[3]; + x[13] = in[12]; + x[14] = in[1]; + x[15] = in[14]; + + // stage 1 + iadst_butterfly_lane_0_1_neon(x[0], x[1], c_1_31_5_27, s0, s1); + iadst_butterfly_lane_2_3_neon(x[2], x[3], c_1_31_5_27, s2, s3); + iadst_butterfly_lane_0_1_neon(x[4], x[5], c_9_23_13_19, s4, s5); + iadst_butterfly_lane_2_3_neon(x[6], x[7], c_9_23_13_19, s6, s7); + iadst_butterfly_lane_0_1_neon(x[8], x[9], c_17_15_21_11, s8, s9); + iadst_butterfly_lane_2_3_neon(x[10], x[11], c_17_15_21_11, s10, s11); + iadst_butterfly_lane_0_1_neon(x[12], x[13], c_25_7_29_3, s12, s13); + iadst_butterfly_lane_2_3_neon(x[14], x[15], c_25_7_29_3, s14, s15); + + x[0] = add_dct_const_round_shift_low_8(s0, s8); + x[1] = add_dct_const_round_shift_low_8(s1, s9); + x[2] = add_dct_const_round_shift_low_8(s2, s10); + x[3] = add_dct_const_round_shift_low_8(s3, s11); + x[4] = add_dct_const_round_shift_low_8(s4, s12); + x[5] = add_dct_const_round_shift_low_8(s5, s13); + x[6] = add_dct_const_round_shift_low_8(s6, s14); + x[7] = add_dct_const_round_shift_low_8(s7, s15); + x[8] = sub_dct_const_round_shift_low_8(s0, s8); + x[9] = sub_dct_const_round_shift_low_8(s1, s9); + x[10] = sub_dct_const_round_shift_low_8(s2, s10); + x[11] = sub_dct_const_round_shift_low_8(s3, s11); + x[12] = sub_dct_const_round_shift_low_8(s4, s12); + x[13] = sub_dct_const_round_shift_low_8(s5, s13); + x[14] = sub_dct_const_round_shift_low_8(s6, s14); + x[15] = sub_dct_const_round_shift_low_8(s7, s15); + + // stage 2 + t[0] = x[0]; + t[1] = x[1]; + t[2] = x[2]; + t[3] = x[3]; + t[4] = x[4]; + t[5] = x[5]; + t[6] = x[6]; + t[7] = x[7]; + iadst_butterfly_lane_0_1_neon(x[8], x[9], c_4_28_20_12, s8, s9); + iadst_butterfly_lane_2_3_neon(x[10], x[11], c_4_28_20_12, s10, s11); + iadst_butterfly_lane_1_0_neon(x[13], x[12], c_4_28_20_12, s13, s12); + iadst_butterfly_lane_3_2_neon(x[15], x[14], c_4_28_20_12, s15, s14); + + x[0] = vaddq_s16(t[0], t[4]); + x[1] = vaddq_s16(t[1], t[5]); + x[2] = vaddq_s16(t[2], t[6]); + x[3] = vaddq_s16(t[3], t[7]); + x[4] = vsubq_s16(t[0], t[4]); + x[5] = vsubq_s16(t[1], t[5]); + x[6] = vsubq_s16(t[2], t[6]); + x[7] = vsubq_s16(t[3], t[7]); + x[8] = add_dct_const_round_shift_low_8(s8, s12); + x[9] = add_dct_const_round_shift_low_8(s9, s13); + x[10] = add_dct_const_round_shift_low_8(s10, s14); + x[11] = add_dct_const_round_shift_low_8(s11, s15); + x[12] = sub_dct_const_round_shift_low_8(s8, s12); + x[13] = sub_dct_const_round_shift_low_8(s9, s13); + x[14] = sub_dct_const_round_shift_low_8(s10, s14); + x[15] = sub_dct_const_round_shift_low_8(s11, s15); + + // stage 3 + t[0] = x[0]; + t[1] = x[1]; + t[2] = x[2]; + t[3] = x[3]; + iadst_butterfly_lane_2_3_neon(x[4], x[5], c_16_n16_8_24, s4, s5); + iadst_butterfly_lane_3_2_neon(x[7], x[6], c_16_n16_8_24, s7, s6); + t[8] = x[8]; + t[9] = x[9]; + t[10] = x[10]; + t[11] = x[11]; + iadst_butterfly_lane_2_3_neon(x[12], x[13], c_16_n16_8_24, s12, s13); + iadst_butterfly_lane_3_2_neon(x[15], x[14], c_16_n16_8_24, s15, s14); + + x[0] = vaddq_s16(t[0], t[2]); + x[1] = vaddq_s16(t[1], t[3]); + x[2] = vsubq_s16(t[0], t[2]); + x[3] = vsubq_s16(t[1], t[3]); + x[4] = add_dct_const_round_shift_low_8(s4, s6); + x[5] = add_dct_const_round_shift_low_8(s5, s7); + x[6] = sub_dct_const_round_shift_low_8(s4, s6); + x[7] = sub_dct_const_round_shift_low_8(s5, s7); + x[8] = vaddq_s16(t[8], t[10]); + x[9] = vaddq_s16(t[9], t[11]); + x[10] = vsubq_s16(t[8], t[10]); + x[11] = vsubq_s16(t[9], t[11]); + x[12] = add_dct_const_round_shift_low_8(s12, s14); + x[13] = add_dct_const_round_shift_low_8(s13, s15); + x[14] = sub_dct_const_round_shift_low_8(s12, s14); + x[15] = sub_dct_const_round_shift_low_8(s13, s15); + + // stage 4 + iadst_half_butterfly_neg_neon(&x[3], &x[2], c_16_n16_8_24); + iadst_half_butterfly_pos_neon(&x[7], &x[6], c_16_n16_8_24); + iadst_half_butterfly_pos_neon(&x[11], &x[10], c_16_n16_8_24); + iadst_half_butterfly_neg_neon(&x[15], &x[14], c_16_n16_8_24); + + out[0] = x[0]; + out[1] = vnegq_s16(x[8]); + out[2] = x[12]; + out[3] = vnegq_s16(x[4]); + out[4] = x[6]; + out[5] = x[14]; + out[6] = x[10]; + out[7] = x[2]; + out[8] = x[3]; + out[9] = x[11]; + out[10] = x[15]; + out[11] = x[7]; + out[12] = x[5]; + out[13] = vnegq_s16(x[13]); + out[14] = x[9]; + out[15] = vnegq_s16(x[1]); + + if (output) { + idct16x16_store_pass1(out, output); + } else { + if (highbd_flag) { + idct16x16_add_store_bd8(out, dest, stride); + } else { + idct16x16_add_store(out, dest, stride); + } + } +} + +void vp9_iht16x16_256_add_neon(const tran_low_t *input, uint8_t *dest, + int stride, int tx_type) { + static const iht_2d IHT_16[] = { + { vpx_idct16x16_256_add_half1d, + vpx_idct16x16_256_add_half1d }, // DCT_DCT = 0 + { vpx_iadst16x16_256_add_half1d, + vpx_idct16x16_256_add_half1d }, // ADST_DCT = 1 + { vpx_idct16x16_256_add_half1d, + vpx_iadst16x16_256_add_half1d }, // DCT_ADST = 2 + { vpx_iadst16x16_256_add_half1d, + vpx_iadst16x16_256_add_half1d } // ADST_ADST = 3 + }; + const iht_2d ht = IHT_16[tx_type]; + int16_t row_output[16 * 16]; + + // pass 1 + ht.rows(input, row_output, dest, stride, 0); // upper 8 rows + ht.rows(input + 8 * 16, row_output + 8, dest, stride, 0); // lower 8 rows + + // pass 2 + ht.cols(row_output, NULL, dest, stride, 0); // left 8 columns + ht.cols(row_output + 16 * 8, NULL, dest + 8, stride, 0); // right 8 columns +} diff --git a/media/libvpx/libvpx/vp9/common/arm/neon/vp9_iht4x4_add_neon.c b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_iht4x4_add_neon.c new file mode 100644 index 0000000000..4f0a90f215 --- /dev/null +++ b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_iht4x4_add_neon.c @@ -0,0 +1,76 @@ +/* + * Copyright (c) 2014 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include <arm_neon.h> +#include <assert.h> + +#include "./vp9_rtcd.h" +#include "./vpx_config.h" +#include "vp9/common/vp9_common.h" +#include "vp9/common/arm/neon/vp9_iht_neon.h" +#include "vpx_dsp/arm/idct_neon.h" +#include "vpx_dsp/arm/mem_neon.h" +#include "vpx_dsp/txfm_common.h" + +void vp9_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest, int stride, + int tx_type) { + int16x8_t a[2]; + uint8x8_t s[2], d[2]; + uint16x8_t sum[2]; + + assert(!((intptr_t)dest % sizeof(uint32_t))); + assert(!(stride % sizeof(uint32_t))); + + a[0] = load_tran_low_to_s16q(input); + a[1] = load_tran_low_to_s16q(input + 8); + transpose_s16_4x4q(&a[0], &a[1]); + + switch (tx_type) { + case DCT_DCT: + idct4x4_16_kernel_bd8(a); + a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1])); + transpose_s16_4x4q(&a[0], &a[1]); + idct4x4_16_kernel_bd8(a); + a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1])); + break; + + case ADST_DCT: + idct4x4_16_kernel_bd8(a); + a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1])); + transpose_s16_4x4q(&a[0], &a[1]); + iadst4(a); + break; + + case DCT_ADST: + iadst4(a); + transpose_s16_4x4q(&a[0], &a[1]); + idct4x4_16_kernel_bd8(a); + a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1])); + break; + + default: + assert(tx_type == ADST_ADST); + iadst4(a); + transpose_s16_4x4q(&a[0], &a[1]); + iadst4(a); + break; + } + + a[0] = vrshrq_n_s16(a[0], 4); + a[1] = vrshrq_n_s16(a[1], 4); + s[0] = load_u8(dest, stride); + s[1] = load_u8(dest + 2 * stride, stride); + sum[0] = vaddw_u8(vreinterpretq_u16_s16(a[0]), s[0]); + sum[1] = vaddw_u8(vreinterpretq_u16_s16(a[1]), s[1]); + d[0] = vqmovun_s16(vreinterpretq_s16_u16(sum[0])); + d[1] = vqmovun_s16(vreinterpretq_s16_u16(sum[1])); + store_u8(dest, stride, d[0]); + store_u8(dest + 2 * stride, stride, d[1]); +} diff --git a/media/libvpx/libvpx/vp9/common/arm/neon/vp9_iht8x8_add_neon.c b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_iht8x8_add_neon.c new file mode 100644 index 0000000000..46ee632e01 --- /dev/null +++ b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_iht8x8_add_neon.c @@ -0,0 +1,68 @@ +/* + * Copyright (c) 2014 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include <arm_neon.h> +#include <assert.h> + +#include "./vp9_rtcd.h" +#include "./vpx_config.h" +#include "vp9/common/vp9_common.h" +#include "vp9/common/arm/neon/vp9_iht_neon.h" +#include "vpx_dsp/arm/idct_neon.h" +#include "vpx_dsp/arm/mem_neon.h" +#include "vpx_dsp/arm/transpose_neon.h" + +void vp9_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest, int stride, + int tx_type) { + const int16x8_t cospis = vld1q_s16(kCospi); + const int16x4_t cospis0 = vget_low_s16(cospis); // cospi 0, 8, 16, 24 + const int16x4_t cospis1 = vget_high_s16(cospis); // cospi 4, 12, 20, 28 + int16x8_t a[8]; + + a[0] = load_tran_low_to_s16q(input + 0 * 8); + a[1] = load_tran_low_to_s16q(input + 1 * 8); + a[2] = load_tran_low_to_s16q(input + 2 * 8); + a[3] = load_tran_low_to_s16q(input + 3 * 8); + a[4] = load_tran_low_to_s16q(input + 4 * 8); + a[5] = load_tran_low_to_s16q(input + 5 * 8); + a[6] = load_tran_low_to_s16q(input + 6 * 8); + a[7] = load_tran_low_to_s16q(input + 7 * 8); + + transpose_s16_8x8(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], &a[7]); + + switch (tx_type) { + case DCT_DCT: + idct8x8_64_1d_bd8_kernel(cospis0, cospis1, a); + transpose_s16_8x8(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], &a[7]); + idct8x8_64_1d_bd8_kernel(cospis0, cospis1, a); + break; + + case ADST_DCT: + idct8x8_64_1d_bd8_kernel(cospis0, cospis1, a); + transpose_s16_8x8(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], &a[7]); + iadst8(a); + break; + + case DCT_ADST: + iadst8(a); + transpose_s16_8x8(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], &a[7]); + idct8x8_64_1d_bd8_kernel(cospis0, cospis1, a); + break; + + default: + assert(tx_type == ADST_ADST); + iadst8(a); + transpose_s16_8x8(&a[0], &a[1], &a[2], &a[3], &a[4], &a[5], &a[6], &a[7]); + iadst8(a); + break; + } + + idct8x8_add8x8_neon(a, dest, stride); +} diff --git a/media/libvpx/libvpx/vp9/common/arm/neon/vp9_iht_neon.h b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_iht_neon.h new file mode 100644 index 0000000000..c64822e27c --- /dev/null +++ b/media/libvpx/libvpx/vp9/common/arm/neon/vp9_iht_neon.h @@ -0,0 +1,272 @@ +/* + * Copyright (c) 2018 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef VPX_VP9_COMMON_ARM_NEON_VP9_IHT_NEON_H_ +#define VPX_VP9_COMMON_ARM_NEON_VP9_IHT_NEON_H_ + +#include <arm_neon.h> + +#include "./vp9_rtcd.h" +#include "./vpx_config.h" +#include "vp9/common/vp9_common.h" +#include "vpx_dsp/arm/idct_neon.h" +#include "vpx_dsp/arm/mem_neon.h" +#include "vpx_dsp/txfm_common.h" + +static INLINE void iadst4(int16x8_t *const io) { + const int32x4_t c3 = vdupq_n_s32(sinpi_3_9); + int16x4_t x[4]; + int32x4_t s[8], output[4]; + const int16x4_t c = + create_s16x4_neon(sinpi_1_9, sinpi_2_9, sinpi_3_9, sinpi_4_9); + + x[0] = vget_low_s16(io[0]); + x[1] = vget_low_s16(io[1]); + x[2] = vget_high_s16(io[0]); + x[3] = vget_high_s16(io[1]); + + s[0] = vmull_lane_s16(x[0], c, 0); + s[1] = vmull_lane_s16(x[0], c, 1); + s[2] = vmull_lane_s16(x[1], c, 2); + s[3] = vmull_lane_s16(x[2], c, 3); + s[4] = vmull_lane_s16(x[2], c, 0); + s[5] = vmull_lane_s16(x[3], c, 1); + s[6] = vmull_lane_s16(x[3], c, 3); + s[7] = vaddl_s16(x[0], x[3]); + s[7] = vsubw_s16(s[7], x[2]); + + s[0] = vaddq_s32(s[0], s[3]); + s[0] = vaddq_s32(s[0], s[5]); + s[1] = vsubq_s32(s[1], s[4]); + s[1] = vsubq_s32(s[1], s[6]); + s[3] = s[2]; + s[2] = vmulq_s32(c3, s[7]); + + output[0] = vaddq_s32(s[0], s[3]); + output[1] = vaddq_s32(s[1], s[3]); + output[2] = s[2]; + output[3] = vaddq_s32(s[0], s[1]); + output[3] = vsubq_s32(output[3], s[3]); + dct_const_round_shift_low_8_dual(output, &io[0], &io[1]); +} + +static INLINE void iadst_half_butterfly_neon(int16x8_t *const x, + const int16x4_t c) { + // Don't add/sub before multiply, which will overflow in iadst8. + const int32x4_t x0_lo = vmull_lane_s16(vget_low_s16(x[0]), c, 0); + const int32x4_t x0_hi = vmull_lane_s16(vget_high_s16(x[0]), c, 0); + const int32x4_t x1_lo = vmull_lane_s16(vget_low_s16(x[1]), c, 0); + const int32x4_t x1_hi = vmull_lane_s16(vget_high_s16(x[1]), c, 0); + int32x4_t t0[2], t1[2]; + + t0[0] = vaddq_s32(x0_lo, x1_lo); + t0[1] = vaddq_s32(x0_hi, x1_hi); + t1[0] = vsubq_s32(x0_lo, x1_lo); + t1[1] = vsubq_s32(x0_hi, x1_hi); + x[0] = dct_const_round_shift_low_8(t0); + x[1] = dct_const_round_shift_low_8(t1); +} + +static INLINE void iadst_half_butterfly_neg_neon(int16x8_t *const x0, + int16x8_t *const x1, + const int16x4_t c) { + // Don't add/sub before multiply, which will overflow in iadst8. + const int32x4_t x0_lo = vmull_lane_s16(vget_low_s16(*x0), c, 1); + const int32x4_t x0_hi = vmull_lane_s16(vget_high_s16(*x0), c, 1); + const int32x4_t x1_lo = vmull_lane_s16(vget_low_s16(*x1), c, 1); + const int32x4_t x1_hi = vmull_lane_s16(vget_high_s16(*x1), c, 1); + int32x4_t t0[2], t1[2]; + + t0[0] = vaddq_s32(x0_lo, x1_lo); + t0[1] = vaddq_s32(x0_hi, x1_hi); + t1[0] = vsubq_s32(x0_lo, x1_lo); + t1[1] = vsubq_s32(x0_hi, x1_hi); + *x1 = dct_const_round_shift_low_8(t0); + *x0 = dct_const_round_shift_low_8(t1); +} + +static INLINE void iadst_half_butterfly_pos_neon(int16x8_t *const x0, + int16x8_t *const x1, + const int16x4_t c) { + // Don't add/sub before multiply, which will overflow in iadst8. + const int32x4_t x0_lo = vmull_lane_s16(vget_low_s16(*x0), c, 0); + const int32x4_t x0_hi = vmull_lane_s16(vget_high_s16(*x0), c, 0); + const int32x4_t x1_lo = vmull_lane_s16(vget_low_s16(*x1), c, 0); + const int32x4_t x1_hi = vmull_lane_s16(vget_high_s16(*x1), c, 0); + int32x4_t t0[2], t1[2]; + + t0[0] = vaddq_s32(x0_lo, x1_lo); + t0[1] = vaddq_s32(x0_hi, x1_hi); + t1[0] = vsubq_s32(x0_lo, x1_lo); + t1[1] = vsubq_s32(x0_hi, x1_hi); + *x1 = dct_const_round_shift_low_8(t0); + *x0 = dct_const_round_shift_low_8(t1); +} + +static INLINE void iadst_butterfly_lane_0_1_neon(const int16x8_t in0, + const int16x8_t in1, + const int16x4_t c, + int32x4_t *const s0, + int32x4_t *const s1) { + s0[0] = vmull_lane_s16(vget_low_s16(in0), c, 0); + s0[1] = vmull_lane_s16(vget_high_s16(in0), c, 0); + s1[0] = vmull_lane_s16(vget_low_s16(in0), c, 1); + s1[1] = vmull_lane_s16(vget_high_s16(in0), c, 1); + + s0[0] = vmlal_lane_s16(s0[0], vget_low_s16(in1), c, 1); + s0[1] = vmlal_lane_s16(s0[1], vget_high_s16(in1), c, 1); + s1[0] = vmlsl_lane_s16(s1[0], vget_low_s16(in1), c, 0); + s1[1] = vmlsl_lane_s16(s1[1], vget_high_s16(in1), c, 0); +} + +static INLINE void iadst_butterfly_lane_2_3_neon(const int16x8_t in0, + const int16x8_t in1, + const int16x4_t c, + int32x4_t *const s0, + int32x4_t *const s1) { + s0[0] = vmull_lane_s16(vget_low_s16(in0), c, 2); + s0[1] = vmull_lane_s16(vget_high_s16(in0), c, 2); + s1[0] = vmull_lane_s16(vget_low_s16(in0), c, 3); + s1[1] = vmull_lane_s16(vget_high_s16(in0), c, 3); + + s0[0] = vmlal_lane_s16(s0[0], vget_low_s16(in1), c, 3); + s0[1] = vmlal_lane_s16(s0[1], vget_high_s16(in1), c, 3); + s1[0] = vmlsl_lane_s16(s1[0], vget_low_s16(in1), c, 2); + s1[1] = vmlsl_lane_s16(s1[1], vget_high_s16(in1), c, 2); +} + +static INLINE void iadst_butterfly_lane_1_0_neon(const int16x8_t in0, + const int16x8_t in1, + const int16x4_t c, + int32x4_t *const s0, + int32x4_t *const s1) { + s0[0] = vmull_lane_s16(vget_low_s16(in0), c, 1); + s0[1] = vmull_lane_s16(vget_high_s16(in0), c, 1); + s1[0] = vmull_lane_s16(vget_low_s16(in0), c, 0); + s1[1] = vmull_lane_s16(vget_high_s16(in0), c, 0); + + s0[0] = vmlal_lane_s16(s0[0], vget_low_s16(in1), c, 0); + s0[1] = vmlal_lane_s16(s0[1], vget_high_s16(in1), c, 0); + s1[0] = vmlsl_lane_s16(s1[0], vget_low_s16(in1), c, 1); + s1[1] = vmlsl_lane_s16(s1[1], vget_high_s16(in1), c, 1); +} + +static INLINE void iadst_butterfly_lane_3_2_neon(const int16x8_t in0, + const int16x8_t in1, + const int16x4_t c, + int32x4_t *const s0, + int32x4_t *const s1) { + s0[0] = vmull_lane_s16(vget_low_s16(in0), c, 3); + s0[1] = vmull_lane_s16(vget_high_s16(in0), c, 3); + s1[0] = vmull_lane_s16(vget_low_s16(in0), c, 2); + s1[1] = vmull_lane_s16(vget_high_s16(in0), c, 2); + + s0[0] = vmlal_lane_s16(s0[0], vget_low_s16(in1), c, 2); + s0[1] = vmlal_lane_s16(s0[1], vget_high_s16(in1), c, 2); + s1[0] = vmlsl_lane_s16(s1[0], vget_low_s16(in1), c, 3); + s1[1] = vmlsl_lane_s16(s1[1], vget_high_s16(in1), c, 3); +} + +static INLINE int16x8_t add_dct_const_round_shift_low_8( + const int32x4_t *const in0, const int32x4_t *const in1) { + int32x4_t sum[2]; + + sum[0] = vaddq_s32(in0[0], in1[0]); + sum[1] = vaddq_s32(in0[1], in1[1]); + return dct_const_round_shift_low_8(sum); +} + +static INLINE int16x8_t sub_dct_const_round_shift_low_8( + const int32x4_t *const in0, const int32x4_t *const in1) { + int32x4_t sum[2]; + + sum[0] = vsubq_s32(in0[0], in1[0]); + sum[1] = vsubq_s32(in0[1], in1[1]); + return dct_const_round_shift_low_8(sum); +} + +static INLINE void iadst8(int16x8_t *const io) { + const int16x4_t c0 = + create_s16x4_neon(cospi_2_64, cospi_30_64, cospi_10_64, cospi_22_64); + const int16x4_t c1 = + create_s16x4_neon(cospi_18_64, cospi_14_64, cospi_26_64, cospi_6_64); + const int16x4_t c2 = + create_s16x4_neon(cospi_16_64, 0, cospi_8_64, cospi_24_64); + int16x8_t x[8], t[4]; + int32x4_t s0[2], s1[2], s2[2], s3[2], s4[2], s5[2], s6[2], s7[2]; + + x[0] = io[7]; + x[1] = io[0]; + x[2] = io[5]; + x[3] = io[2]; + x[4] = io[3]; + x[5] = io[4]; + x[6] = io[1]; + x[7] = io[6]; + + // stage 1 + iadst_butterfly_lane_0_1_neon(x[0], x[1], c0, s0, s1); + iadst_butterfly_lane_2_3_neon(x[2], x[3], c0, s2, s3); + iadst_butterfly_lane_0_1_neon(x[4], x[5], c1, s4, s5); + iadst_butterfly_lane_2_3_neon(x[6], x[7], c1, s6, s7); + + x[0] = add_dct_const_round_shift_low_8(s0, s4); + x[1] = add_dct_const_round_shift_low_8(s1, s5); + x[2] = add_dct_const_round_shift_low_8(s2, s6); + x[3] = add_dct_const_round_shift_low_8(s3, s7); + x[4] = sub_dct_const_round_shift_low_8(s0, s4); + x[5] = sub_dct_const_round_shift_low_8(s1, s5); + x[6] = sub_dct_const_round_shift_low_8(s2, s6); + x[7] = sub_dct_const_round_shift_low_8(s3, s7); + + // stage 2 + t[0] = x[0]; + t[1] = x[1]; + t[2] = x[2]; + t[3] = x[3]; + iadst_butterfly_lane_2_3_neon(x[4], x[5], c2, s4, s5); + iadst_butterfly_lane_3_2_neon(x[7], x[6], c2, s7, s6); + + x[0] = vaddq_s16(t[0], t[2]); + x[1] = vaddq_s16(t[1], t[3]); + x[2] = vsubq_s16(t[0], t[2]); + x[3] = vsubq_s16(t[1], t[3]); + x[4] = add_dct_const_round_shift_low_8(s4, s6); + x[5] = add_dct_const_round_shift_low_8(s5, s7); + x[6] = sub_dct_const_round_shift_low_8(s4, s6); + x[7] = sub_dct_const_round_shift_low_8(s5, s7); + + // stage 3 + iadst_half_butterfly_neon(x + 2, c2); + iadst_half_butterfly_neon(x + 6, c2); + + io[0] = x[0]; + io[1] = vnegq_s16(x[4]); + io[2] = x[6]; + io[3] = vnegq_s16(x[2]); + io[4] = x[3]; + io[5] = vnegq_s16(x[7]); + io[6] = x[5]; + io[7] = vnegq_s16(x[1]); +} + +void vpx_iadst16x16_256_add_half1d(const void *const input, int16_t *output, + void *const dest, const int stride, + const int highbd_flag); + +typedef void (*iht_1d)(const void *const input, int16_t *output, + void *const dest, const int stride, + const int highbd_flag); + +typedef struct { + iht_1d cols, rows; // vertical and horizontal +} iht_2d; + +#endif // VPX_VP9_COMMON_ARM_NEON_VP9_IHT_NEON_H_ |