diff options
Diffstat (limited to '')
28 files changed, 12641 insertions, 0 deletions
diff --git a/third_party/aom/av1/encoder/x86/av1_fwd_txfm1d_sse4.c b/third_party/aom/av1/encoder/x86/av1_fwd_txfm1d_sse4.c new file mode 100644 index 0000000000..07615543c6 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/av1_fwd_txfm1d_sse4.c @@ -0,0 +1,1217 @@ +/* + * Copyright (c) 2018, 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 "av1/encoder/x86/av1_txfm1d_sse4.h" + +void av1_fdct32_new_sse4_1(const __m128i *input, __m128i *output, + int8_t cos_bit) { + __m128i buf0[32]; + __m128i buf1[32]; + const int32_t *cospi; + // stage 0 + // stage 1 + buf1[0] = _mm_add_epi32(input[0], input[31]); + buf1[31] = _mm_sub_epi32(input[0], input[31]); + buf1[1] = _mm_add_epi32(input[1], input[30]); + buf1[30] = _mm_sub_epi32(input[1], input[30]); + buf1[2] = _mm_add_epi32(input[2], input[29]); + buf1[29] = _mm_sub_epi32(input[2], input[29]); + buf1[3] = _mm_add_epi32(input[3], input[28]); + buf1[28] = _mm_sub_epi32(input[3], input[28]); + buf1[4] = _mm_add_epi32(input[4], input[27]); + buf1[27] = _mm_sub_epi32(input[4], input[27]); + buf1[5] = _mm_add_epi32(input[5], input[26]); + buf1[26] = _mm_sub_epi32(input[5], input[26]); + buf1[6] = _mm_add_epi32(input[6], input[25]); + buf1[25] = _mm_sub_epi32(input[6], input[25]); + buf1[7] = _mm_add_epi32(input[7], input[24]); + buf1[24] = _mm_sub_epi32(input[7], input[24]); + buf1[8] = _mm_add_epi32(input[8], input[23]); + buf1[23] = _mm_sub_epi32(input[8], input[23]); + buf1[9] = _mm_add_epi32(input[9], input[22]); + buf1[22] = _mm_sub_epi32(input[9], input[22]); + buf1[10] = _mm_add_epi32(input[10], input[21]); + buf1[21] = _mm_sub_epi32(input[10], input[21]); + buf1[11] = _mm_add_epi32(input[11], input[20]); + buf1[20] = _mm_sub_epi32(input[11], input[20]); + buf1[12] = _mm_add_epi32(input[12], input[19]); + buf1[19] = _mm_sub_epi32(input[12], input[19]); + buf1[13] = _mm_add_epi32(input[13], input[18]); + buf1[18] = _mm_sub_epi32(input[13], input[18]); + buf1[14] = _mm_add_epi32(input[14], input[17]); + buf1[17] = _mm_sub_epi32(input[14], input[17]); + buf1[15] = _mm_add_epi32(input[15], input[16]); + buf1[16] = _mm_sub_epi32(input[15], input[16]); + + // stage 2 + cospi = cospi_arr(cos_bit); + buf0[0] = _mm_add_epi32(buf1[0], buf1[15]); + buf0[15] = _mm_sub_epi32(buf1[0], buf1[15]); + buf0[1] = _mm_add_epi32(buf1[1], buf1[14]); + buf0[14] = _mm_sub_epi32(buf1[1], buf1[14]); + buf0[2] = _mm_add_epi32(buf1[2], buf1[13]); + buf0[13] = _mm_sub_epi32(buf1[2], buf1[13]); + buf0[3] = _mm_add_epi32(buf1[3], buf1[12]); + buf0[12] = _mm_sub_epi32(buf1[3], buf1[12]); + buf0[4] = _mm_add_epi32(buf1[4], buf1[11]); + buf0[11] = _mm_sub_epi32(buf1[4], buf1[11]); + buf0[5] = _mm_add_epi32(buf1[5], buf1[10]); + buf0[10] = _mm_sub_epi32(buf1[5], buf1[10]); + buf0[6] = _mm_add_epi32(buf1[6], buf1[9]); + buf0[9] = _mm_sub_epi32(buf1[6], buf1[9]); + buf0[7] = _mm_add_epi32(buf1[7], buf1[8]); + buf0[8] = _mm_sub_epi32(buf1[7], buf1[8]); + buf0[16] = buf1[16]; + buf0[17] = buf1[17]; + buf0[18] = buf1[18]; + buf0[19] = buf1[19]; + btf_32_sse4_1_type0(-cospi[32], cospi[32], buf1[20], buf1[27], buf0[20], + buf0[27], cos_bit); + btf_32_sse4_1_type0(-cospi[32], cospi[32], buf1[21], buf1[26], buf0[21], + buf0[26], cos_bit); + btf_32_sse4_1_type0(-cospi[32], cospi[32], buf1[22], buf1[25], buf0[22], + buf0[25], cos_bit); + btf_32_sse4_1_type0(-cospi[32], cospi[32], buf1[23], buf1[24], buf0[23], + buf0[24], cos_bit); + buf0[28] = buf1[28]; + buf0[29] = buf1[29]; + buf0[30] = buf1[30]; + buf0[31] = buf1[31]; + + // stage 3 + cospi = cospi_arr(cos_bit); + buf1[0] = _mm_add_epi32(buf0[0], buf0[7]); + buf1[7] = _mm_sub_epi32(buf0[0], buf0[7]); + buf1[1] = _mm_add_epi32(buf0[1], buf0[6]); + buf1[6] = _mm_sub_epi32(buf0[1], buf0[6]); + buf1[2] = _mm_add_epi32(buf0[2], buf0[5]); + buf1[5] = _mm_sub_epi32(buf0[2], buf0[5]); + buf1[3] = _mm_add_epi32(buf0[3], buf0[4]); + buf1[4] = _mm_sub_epi32(buf0[3], buf0[4]); + buf1[8] = buf0[8]; + buf1[9] = buf0[9]; + btf_32_sse4_1_type0(-cospi[32], cospi[32], buf0[10], buf0[13], buf1[10], + buf1[13], cos_bit); + btf_32_sse4_1_type0(-cospi[32], cospi[32], buf0[11], buf0[12], buf1[11], + buf1[12], cos_bit); + buf1[14] = buf0[14]; + buf1[15] = buf0[15]; + buf1[16] = _mm_add_epi32(buf0[16], buf0[23]); + buf1[23] = _mm_sub_epi32(buf0[16], buf0[23]); + buf1[17] = _mm_add_epi32(buf0[17], buf0[22]); + buf1[22] = _mm_sub_epi32(buf0[17], buf0[22]); + buf1[18] = _mm_add_epi32(buf0[18], buf0[21]); + buf1[21] = _mm_sub_epi32(buf0[18], buf0[21]); + buf1[19] = _mm_add_epi32(buf0[19], buf0[20]); + buf1[20] = _mm_sub_epi32(buf0[19], buf0[20]); + buf1[24] = _mm_sub_epi32(buf0[31], buf0[24]); + buf1[31] = _mm_add_epi32(buf0[31], buf0[24]); + buf1[25] = _mm_sub_epi32(buf0[30], buf0[25]); + buf1[30] = _mm_add_epi32(buf0[30], buf0[25]); + buf1[26] = _mm_sub_epi32(buf0[29], buf0[26]); + buf1[29] = _mm_add_epi32(buf0[29], buf0[26]); + buf1[27] = _mm_sub_epi32(buf0[28], buf0[27]); + buf1[28] = _mm_add_epi32(buf0[28], buf0[27]); + + // stage 4 + cospi = cospi_arr(cos_bit); + buf0[0] = _mm_add_epi32(buf1[0], buf1[3]); + buf0[3] = _mm_sub_epi32(buf1[0], buf1[3]); + buf0[1] = _mm_add_epi32(buf1[1], buf1[2]); + buf0[2] = _mm_sub_epi32(buf1[1], buf1[2]); + buf0[4] = buf1[4]; + btf_32_sse4_1_type0(-cospi[32], cospi[32], buf1[5], buf1[6], buf0[5], buf0[6], + cos_bit); + buf0[7] = buf1[7]; + buf0[8] = _mm_add_epi32(buf1[8], buf1[11]); + buf0[11] = _mm_sub_epi32(buf1[8], buf1[11]); + buf0[9] = _mm_add_epi32(buf1[9], buf1[10]); + buf0[10] = _mm_sub_epi32(buf1[9], buf1[10]); + buf0[12] = _mm_sub_epi32(buf1[15], buf1[12]); + buf0[15] = _mm_add_epi32(buf1[15], buf1[12]); + buf0[13] = _mm_sub_epi32(buf1[14], buf1[13]); + buf0[14] = _mm_add_epi32(buf1[14], buf1[13]); + buf0[16] = buf1[16]; + buf0[17] = buf1[17]; + btf_32_sse4_1_type0(-cospi[16], cospi[48], buf1[18], buf1[29], buf0[18], + buf0[29], cos_bit); + btf_32_sse4_1_type0(-cospi[16], cospi[48], buf1[19], buf1[28], buf0[19], + buf0[28], cos_bit); + btf_32_sse4_1_type0(-cospi[48], -cospi[16], buf1[20], buf1[27], buf0[20], + buf0[27], cos_bit); + btf_32_sse4_1_type0(-cospi[48], -cospi[16], buf1[21], buf1[26], buf0[21], + buf0[26], cos_bit); + buf0[22] = buf1[22]; + buf0[23] = buf1[23]; + buf0[24] = buf1[24]; + buf0[25] = buf1[25]; + buf0[30] = buf1[30]; + buf0[31] = buf1[31]; + + // stage 5 + cospi = cospi_arr(cos_bit); + btf_32_sse4_1_type0(cospi[32], cospi[32], buf0[0], buf0[1], buf1[0], buf1[1], + cos_bit); + btf_32_sse4_1_type1(cospi[48], cospi[16], buf0[2], buf0[3], buf1[2], buf1[3], + cos_bit); + buf1[4] = _mm_add_epi32(buf0[4], buf0[5]); + buf1[5] = _mm_sub_epi32(buf0[4], buf0[5]); + buf1[6] = _mm_sub_epi32(buf0[7], buf0[6]); + buf1[7] = _mm_add_epi32(buf0[7], buf0[6]); + buf1[8] = buf0[8]; + btf_32_sse4_1_type0(-cospi[16], cospi[48], buf0[9], buf0[14], buf1[9], + buf1[14], cos_bit); + btf_32_sse4_1_type0(-cospi[48], -cospi[16], buf0[10], buf0[13], buf1[10], + buf1[13], cos_bit); + buf1[11] = buf0[11]; + buf1[12] = buf0[12]; + buf1[15] = buf0[15]; + buf1[16] = _mm_add_epi32(buf0[16], buf0[19]); + buf1[19] = _mm_sub_epi32(buf0[16], buf0[19]); + buf1[17] = _mm_add_epi32(buf0[17], buf0[18]); + buf1[18] = _mm_sub_epi32(buf0[17], buf0[18]); + buf1[20] = _mm_sub_epi32(buf0[23], buf0[20]); + buf1[23] = _mm_add_epi32(buf0[23], buf0[20]); + buf1[21] = _mm_sub_epi32(buf0[22], buf0[21]); + buf1[22] = _mm_add_epi32(buf0[22], buf0[21]); + buf1[24] = _mm_add_epi32(buf0[24], buf0[27]); + buf1[27] = _mm_sub_epi32(buf0[24], buf0[27]); + buf1[25] = _mm_add_epi32(buf0[25], buf0[26]); + buf1[26] = _mm_sub_epi32(buf0[25], buf0[26]); + buf1[28] = _mm_sub_epi32(buf0[31], buf0[28]); + buf1[31] = _mm_add_epi32(buf0[31], buf0[28]); + buf1[29] = _mm_sub_epi32(buf0[30], buf0[29]); + buf1[30] = _mm_add_epi32(buf0[30], buf0[29]); + + // stage 6 + cospi = cospi_arr(cos_bit); + buf0[0] = buf1[0]; + buf0[1] = buf1[1]; + buf0[2] = buf1[2]; + buf0[3] = buf1[3]; + btf_32_sse4_1_type1(cospi[56], cospi[8], buf1[4], buf1[7], buf0[4], buf0[7], + cos_bit); + btf_32_sse4_1_type1(cospi[24], cospi[40], buf1[5], buf1[6], buf0[5], buf0[6], + cos_bit); + buf0[8] = _mm_add_epi32(buf1[8], buf1[9]); + buf0[9] = _mm_sub_epi32(buf1[8], buf1[9]); + buf0[10] = _mm_sub_epi32(buf1[11], buf1[10]); + buf0[11] = _mm_add_epi32(buf1[11], buf1[10]); + buf0[12] = _mm_add_epi32(buf1[12], buf1[13]); + buf0[13] = _mm_sub_epi32(buf1[12], buf1[13]); + buf0[14] = _mm_sub_epi32(buf1[15], buf1[14]); + buf0[15] = _mm_add_epi32(buf1[15], buf1[14]); + buf0[16] = buf1[16]; + btf_32_sse4_1_type0(-cospi[8], cospi[56], buf1[17], buf1[30], buf0[17], + buf0[30], cos_bit); + btf_32_sse4_1_type0(-cospi[56], -cospi[8], buf1[18], buf1[29], buf0[18], + buf0[29], cos_bit); + buf0[19] = buf1[19]; + buf0[20] = buf1[20]; + btf_32_sse4_1_type0(-cospi[40], cospi[24], buf1[21], buf1[26], buf0[21], + buf0[26], cos_bit); + btf_32_sse4_1_type0(-cospi[24], -cospi[40], buf1[22], buf1[25], buf0[22], + buf0[25], cos_bit); + buf0[23] = buf1[23]; + buf0[24] = buf1[24]; + buf0[27] = buf1[27]; + buf0[28] = buf1[28]; + buf0[31] = buf1[31]; + + // stage 7 + cospi = cospi_arr(cos_bit); + buf1[0] = buf0[0]; + buf1[1] = buf0[1]; + buf1[2] = buf0[2]; + buf1[3] = buf0[3]; + buf1[4] = buf0[4]; + buf1[5] = buf0[5]; + buf1[6] = buf0[6]; + buf1[7] = buf0[7]; + btf_32_sse4_1_type1(cospi[60], cospi[4], buf0[8], buf0[15], buf1[8], buf1[15], + cos_bit); + btf_32_sse4_1_type1(cospi[28], cospi[36], buf0[9], buf0[14], buf1[9], + buf1[14], cos_bit); + btf_32_sse4_1_type1(cospi[44], cospi[20], buf0[10], buf0[13], buf1[10], + buf1[13], cos_bit); + btf_32_sse4_1_type1(cospi[12], cospi[52], buf0[11], buf0[12], buf1[11], + buf1[12], cos_bit); + buf1[16] = _mm_add_epi32(buf0[16], buf0[17]); + buf1[17] = _mm_sub_epi32(buf0[16], buf0[17]); + buf1[18] = _mm_sub_epi32(buf0[19], buf0[18]); + buf1[19] = _mm_add_epi32(buf0[19], buf0[18]); + buf1[20] = _mm_add_epi32(buf0[20], buf0[21]); + buf1[21] = _mm_sub_epi32(buf0[20], buf0[21]); + buf1[22] = _mm_sub_epi32(buf0[23], buf0[22]); + buf1[23] = _mm_add_epi32(buf0[23], buf0[22]); + buf1[24] = _mm_add_epi32(buf0[24], buf0[25]); + buf1[25] = _mm_sub_epi32(buf0[24], buf0[25]); + buf1[26] = _mm_sub_epi32(buf0[27], buf0[26]); + buf1[27] = _mm_add_epi32(buf0[27], buf0[26]); + buf1[28] = _mm_add_epi32(buf0[28], buf0[29]); + buf1[29] = _mm_sub_epi32(buf0[28], buf0[29]); + buf1[30] = _mm_sub_epi32(buf0[31], buf0[30]); + buf1[31] = _mm_add_epi32(buf0[31], buf0[30]); + + // stage 8 + cospi = cospi_arr(cos_bit); + buf0[0] = buf1[0]; + buf0[1] = buf1[1]; + buf0[2] = buf1[2]; + buf0[3] = buf1[3]; + buf0[4] = buf1[4]; + buf0[5] = buf1[5]; + buf0[6] = buf1[6]; + buf0[7] = buf1[7]; + buf0[8] = buf1[8]; + buf0[9] = buf1[9]; + buf0[10] = buf1[10]; + buf0[11] = buf1[11]; + buf0[12] = buf1[12]; + buf0[13] = buf1[13]; + buf0[14] = buf1[14]; + buf0[15] = buf1[15]; + btf_32_sse4_1_type1(cospi[62], cospi[2], buf1[16], buf1[31], buf0[16], + buf0[31], cos_bit); + btf_32_sse4_1_type1(cospi[30], cospi[34], buf1[17], buf1[30], buf0[17], + buf0[30], cos_bit); + btf_32_sse4_1_type1(cospi[46], cospi[18], buf1[18], buf1[29], buf0[18], + buf0[29], cos_bit); + btf_32_sse4_1_type1(cospi[14], cospi[50], buf1[19], buf1[28], buf0[19], + buf0[28], cos_bit); + btf_32_sse4_1_type1(cospi[54], cospi[10], buf1[20], buf1[27], buf0[20], + buf0[27], cos_bit); + btf_32_sse4_1_type1(cospi[22], cospi[42], buf1[21], buf1[26], buf0[21], + buf0[26], cos_bit); + btf_32_sse4_1_type1(cospi[38], cospi[26], buf1[22], buf1[25], buf0[22], + buf0[25], cos_bit); + btf_32_sse4_1_type1(cospi[6], cospi[58], buf1[23], buf1[24], buf0[23], + buf0[24], cos_bit); + + // stage 9 + output[0] = buf0[0]; + output[1] = buf0[16]; + output[2] = buf0[8]; + output[3] = buf0[24]; + output[4] = buf0[4]; + output[5] = buf0[20]; + output[6] = buf0[12]; + output[7] = buf0[28]; + output[8] = buf0[2]; + output[9] = buf0[18]; + output[10] = buf0[10]; + output[11] = buf0[26]; + output[12] = buf0[6]; + output[13] = buf0[22]; + output[14] = buf0[14]; + output[15] = buf0[30]; + output[16] = buf0[1]; + output[17] = buf0[17]; + output[18] = buf0[9]; + output[19] = buf0[25]; + output[20] = buf0[5]; + output[21] = buf0[21]; + output[22] = buf0[13]; + output[23] = buf0[29]; + output[24] = buf0[3]; + output[25] = buf0[19]; + output[26] = buf0[11]; + output[27] = buf0[27]; + output[28] = buf0[7]; + output[29] = buf0[23]; + output[30] = buf0[15]; + output[31] = buf0[31]; +} + +void av1_fadst4_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range) { + const int txfm_size = 4; + const int num_per_128 = 4; + const int32_t *cospi; + __m128i buf0[4]; + __m128i buf1[4]; + int col_num = txfm_size / num_per_128; + int col; + (void)stage_range; + for (col = 0; col < col_num; col++) { + // stage 0; + int32_t stage_idx = 0; + int j; + for (j = 0; j < 4; ++j) { + buf0[j] = input[j * col_num + col]; + } + + // stage 1 + stage_idx++; + buf1[0] = buf0[3]; + buf1[1] = buf0[0]; + buf1[2] = buf0[1]; + buf1[3] = buf0[2]; + + // stage 2 + stage_idx++; + + cospi = cospi_arr(cos_bit); + btf_32_sse4_1_type0(cospi[8], cospi[56], buf1[0], buf1[1], buf0[0], buf0[1], + cos_bit); + btf_32_sse4_1_type0(cospi[40], cospi[24], buf1[2], buf1[3], buf0[2], + buf0[3], cos_bit); + + // stage 3 + stage_idx++; + buf1[0] = _mm_add_epi32(buf0[0], buf0[2]); + buf1[2] = _mm_sub_epi32(buf0[0], buf0[2]); + buf1[1] = _mm_add_epi32(buf0[1], buf0[3]); + buf1[3] = _mm_sub_epi32(buf0[1], buf0[3]); + + // stage 4 + stage_idx++; + + cospi = cospi_arr(cos_bit); + buf0[0] = buf1[0]; + buf0[1] = buf1[1]; + btf_32_sse4_1_type0(cospi[32], cospi[32], buf1[2], buf1[3], buf0[2], + buf0[3], cos_bit); + + // stage 5 + stage_idx++; + buf1[0] = buf0[0]; + buf1[1] = _mm_sub_epi32(_mm_setzero_si128(), buf0[2]); + buf1[2] = buf0[3]; + buf1[3] = _mm_sub_epi32(_mm_setzero_si128(), buf0[1]); + + for (j = 0; j < 4; ++j) { + output[j * col_num + col] = buf1[j]; + } + } +} + +void av1_fdct64_new_sse4_1(const __m128i *input, __m128i *output, + int8_t cos_bit, const int instride, + const int outstride) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1)); + + __m128i cospi_m32 = _mm_set1_epi32(-cospi[32]); + __m128i cospi_p32 = _mm_set1_epi32(cospi[32]); + __m128i cospi_m16 = _mm_set1_epi32(-cospi[16]); + __m128i cospi_p48 = _mm_set1_epi32(cospi[48]); + __m128i cospi_m48 = _mm_set1_epi32(-cospi[48]); + __m128i cospi_p16 = _mm_set1_epi32(cospi[16]); + __m128i cospi_m08 = _mm_set1_epi32(-cospi[8]); + __m128i cospi_p56 = _mm_set1_epi32(cospi[56]); + __m128i cospi_m56 = _mm_set1_epi32(-cospi[56]); + __m128i cospi_m40 = _mm_set1_epi32(-cospi[40]); + __m128i cospi_p24 = _mm_set1_epi32(cospi[24]); + __m128i cospi_m24 = _mm_set1_epi32(-cospi[24]); + __m128i cospi_p08 = _mm_set1_epi32(cospi[8]); + __m128i cospi_p40 = _mm_set1_epi32(cospi[40]); + __m128i cospi_p60 = _mm_set1_epi32(cospi[60]); + __m128i cospi_p04 = _mm_set1_epi32(cospi[4]); + __m128i cospi_p28 = _mm_set1_epi32(cospi[28]); + __m128i cospi_p36 = _mm_set1_epi32(cospi[36]); + __m128i cospi_p44 = _mm_set1_epi32(cospi[44]); + __m128i cospi_p20 = _mm_set1_epi32(cospi[20]); + __m128i cospi_p12 = _mm_set1_epi32(cospi[12]); + __m128i cospi_p52 = _mm_set1_epi32(cospi[52]); + __m128i cospi_m04 = _mm_set1_epi32(-cospi[4]); + __m128i cospi_m60 = _mm_set1_epi32(-cospi[60]); + __m128i cospi_m36 = _mm_set1_epi32(-cospi[36]); + __m128i cospi_m28 = _mm_set1_epi32(-cospi[28]); + __m128i cospi_m20 = _mm_set1_epi32(-cospi[20]); + __m128i cospi_m44 = _mm_set1_epi32(-cospi[44]); + __m128i cospi_m52 = _mm_set1_epi32(-cospi[52]); + __m128i cospi_m12 = _mm_set1_epi32(-cospi[12]); + __m128i cospi_p62 = _mm_set1_epi32(cospi[62]); + __m128i cospi_p02 = _mm_set1_epi32(cospi[2]); + __m128i cospi_p30 = _mm_set1_epi32(cospi[30]); + __m128i cospi_p34 = _mm_set1_epi32(cospi[34]); + __m128i cospi_p46 = _mm_set1_epi32(cospi[46]); + __m128i cospi_p18 = _mm_set1_epi32(cospi[18]); + __m128i cospi_p14 = _mm_set1_epi32(cospi[14]); + __m128i cospi_p50 = _mm_set1_epi32(cospi[50]); + __m128i cospi_p54 = _mm_set1_epi32(cospi[54]); + __m128i cospi_p10 = _mm_set1_epi32(cospi[10]); + __m128i cospi_p22 = _mm_set1_epi32(cospi[22]); + __m128i cospi_p42 = _mm_set1_epi32(cospi[42]); + __m128i cospi_p38 = _mm_set1_epi32(cospi[38]); + __m128i cospi_p26 = _mm_set1_epi32(cospi[26]); + __m128i cospi_p06 = _mm_set1_epi32(cospi[6]); + __m128i cospi_p58 = _mm_set1_epi32(cospi[58]); + __m128i cospi_p63 = _mm_set1_epi32(cospi[63]); + __m128i cospi_p01 = _mm_set1_epi32(cospi[1]); + __m128i cospi_p31 = _mm_set1_epi32(cospi[31]); + __m128i cospi_p33 = _mm_set1_epi32(cospi[33]); + __m128i cospi_p47 = _mm_set1_epi32(cospi[47]); + __m128i cospi_p17 = _mm_set1_epi32(cospi[17]); + __m128i cospi_p15 = _mm_set1_epi32(cospi[15]); + __m128i cospi_p49 = _mm_set1_epi32(cospi[49]); + __m128i cospi_p55 = _mm_set1_epi32(cospi[55]); + __m128i cospi_p09 = _mm_set1_epi32(cospi[9]); + __m128i cospi_p23 = _mm_set1_epi32(cospi[23]); + __m128i cospi_p41 = _mm_set1_epi32(cospi[41]); + __m128i cospi_p39 = _mm_set1_epi32(cospi[39]); + __m128i cospi_p25 = _mm_set1_epi32(cospi[25]); + __m128i cospi_p07 = _mm_set1_epi32(cospi[7]); + __m128i cospi_p57 = _mm_set1_epi32(cospi[57]); + __m128i cospi_p59 = _mm_set1_epi32(cospi[59]); + __m128i cospi_p05 = _mm_set1_epi32(cospi[5]); + __m128i cospi_p27 = _mm_set1_epi32(cospi[27]); + __m128i cospi_p37 = _mm_set1_epi32(cospi[37]); + __m128i cospi_p43 = _mm_set1_epi32(cospi[43]); + __m128i cospi_p21 = _mm_set1_epi32(cospi[21]); + __m128i cospi_p11 = _mm_set1_epi32(cospi[11]); + __m128i cospi_p53 = _mm_set1_epi32(cospi[53]); + __m128i cospi_p51 = _mm_set1_epi32(cospi[51]); + __m128i cospi_p13 = _mm_set1_epi32(cospi[13]); + __m128i cospi_p19 = _mm_set1_epi32(cospi[19]); + __m128i cospi_p45 = _mm_set1_epi32(cospi[45]); + __m128i cospi_p35 = _mm_set1_epi32(cospi[35]); + __m128i cospi_p29 = _mm_set1_epi32(cospi[29]); + __m128i cospi_p03 = _mm_set1_epi32(cospi[3]); + __m128i cospi_p61 = _mm_set1_epi32(cospi[61]); + + // stage 1 + __m128i x1[64]; + x1[0] = _mm_add_epi32(input[0 * instride], input[63 * instride]); + x1[63] = _mm_sub_epi32(input[0 * instride], input[63 * instride]); + x1[1] = _mm_add_epi32(input[1 * instride], input[62 * instride]); + x1[62] = _mm_sub_epi32(input[1 * instride], input[62 * instride]); + x1[2] = _mm_add_epi32(input[2 * instride], input[61 * instride]); + x1[61] = _mm_sub_epi32(input[2 * instride], input[61 * instride]); + x1[3] = _mm_add_epi32(input[3 * instride], input[60 * instride]); + x1[60] = _mm_sub_epi32(input[3 * instride], input[60 * instride]); + x1[4] = _mm_add_epi32(input[4 * instride], input[59 * instride]); + x1[59] = _mm_sub_epi32(input[4 * instride], input[59 * instride]); + x1[5] = _mm_add_epi32(input[5 * instride], input[58 * instride]); + x1[58] = _mm_sub_epi32(input[5 * instride], input[58 * instride]); + x1[6] = _mm_add_epi32(input[6 * instride], input[57 * instride]); + x1[57] = _mm_sub_epi32(input[6 * instride], input[57 * instride]); + x1[7] = _mm_add_epi32(input[7 * instride], input[56 * instride]); + x1[56] = _mm_sub_epi32(input[7 * instride], input[56 * instride]); + x1[8] = _mm_add_epi32(input[8 * instride], input[55 * instride]); + x1[55] = _mm_sub_epi32(input[8 * instride], input[55 * instride]); + x1[9] = _mm_add_epi32(input[9 * instride], input[54 * instride]); + x1[54] = _mm_sub_epi32(input[9 * instride], input[54 * instride]); + x1[10] = _mm_add_epi32(input[10 * instride], input[53 * instride]); + x1[53] = _mm_sub_epi32(input[10 * instride], input[53 * instride]); + x1[11] = _mm_add_epi32(input[11 * instride], input[52 * instride]); + x1[52] = _mm_sub_epi32(input[11 * instride], input[52 * instride]); + x1[12] = _mm_add_epi32(input[12 * instride], input[51 * instride]); + x1[51] = _mm_sub_epi32(input[12 * instride], input[51 * instride]); + x1[13] = _mm_add_epi32(input[13 * instride], input[50 * instride]); + x1[50] = _mm_sub_epi32(input[13 * instride], input[50 * instride]); + x1[14] = _mm_add_epi32(input[14 * instride], input[49 * instride]); + x1[49] = _mm_sub_epi32(input[14 * instride], input[49 * instride]); + x1[15] = _mm_add_epi32(input[15 * instride], input[48 * instride]); + x1[48] = _mm_sub_epi32(input[15 * instride], input[48 * instride]); + x1[16] = _mm_add_epi32(input[16 * instride], input[47 * instride]); + x1[47] = _mm_sub_epi32(input[16 * instride], input[47 * instride]); + x1[17] = _mm_add_epi32(input[17 * instride], input[46 * instride]); + x1[46] = _mm_sub_epi32(input[17 * instride], input[46 * instride]); + x1[18] = _mm_add_epi32(input[18 * instride], input[45 * instride]); + x1[45] = _mm_sub_epi32(input[18 * instride], input[45 * instride]); + x1[19] = _mm_add_epi32(input[19 * instride], input[44 * instride]); + x1[44] = _mm_sub_epi32(input[19 * instride], input[44 * instride]); + x1[20] = _mm_add_epi32(input[20 * instride], input[43 * instride]); + x1[43] = _mm_sub_epi32(input[20 * instride], input[43 * instride]); + x1[21] = _mm_add_epi32(input[21 * instride], input[42 * instride]); + x1[42] = _mm_sub_epi32(input[21 * instride], input[42 * instride]); + x1[22] = _mm_add_epi32(input[22 * instride], input[41 * instride]); + x1[41] = _mm_sub_epi32(input[22 * instride], input[41 * instride]); + x1[23] = _mm_add_epi32(input[23 * instride], input[40 * instride]); + x1[40] = _mm_sub_epi32(input[23 * instride], input[40 * instride]); + x1[24] = _mm_add_epi32(input[24 * instride], input[39 * instride]); + x1[39] = _mm_sub_epi32(input[24 * instride], input[39 * instride]); + x1[25] = _mm_add_epi32(input[25 * instride], input[38 * instride]); + x1[38] = _mm_sub_epi32(input[25 * instride], input[38 * instride]); + x1[26] = _mm_add_epi32(input[26 * instride], input[37 * instride]); + x1[37] = _mm_sub_epi32(input[26 * instride], input[37 * instride]); + x1[27] = _mm_add_epi32(input[27 * instride], input[36 * instride]); + x1[36] = _mm_sub_epi32(input[27 * instride], input[36 * instride]); + x1[28] = _mm_add_epi32(input[28 * instride], input[35 * instride]); + x1[35] = _mm_sub_epi32(input[28 * instride], input[35 * instride]); + x1[29] = _mm_add_epi32(input[29 * instride], input[34 * instride]); + x1[34] = _mm_sub_epi32(input[29 * instride], input[34 * instride]); + x1[30] = _mm_add_epi32(input[30 * instride], input[33 * instride]); + x1[33] = _mm_sub_epi32(input[30 * instride], input[33 * instride]); + x1[31] = _mm_add_epi32(input[31 * instride], input[32 * instride]); + x1[32] = _mm_sub_epi32(input[31 * instride], input[32 * instride]); + + // stage 2 + __m128i x2[64]; + x2[0] = _mm_add_epi32(x1[0], x1[31]); + x2[31] = _mm_sub_epi32(x1[0], x1[31]); + x2[1] = _mm_add_epi32(x1[1], x1[30]); + x2[30] = _mm_sub_epi32(x1[1], x1[30]); + x2[2] = _mm_add_epi32(x1[2], x1[29]); + x2[29] = _mm_sub_epi32(x1[2], x1[29]); + x2[3] = _mm_add_epi32(x1[3], x1[28]); + x2[28] = _mm_sub_epi32(x1[3], x1[28]); + x2[4] = _mm_add_epi32(x1[4], x1[27]); + x2[27] = _mm_sub_epi32(x1[4], x1[27]); + x2[5] = _mm_add_epi32(x1[5], x1[26]); + x2[26] = _mm_sub_epi32(x1[5], x1[26]); + x2[6] = _mm_add_epi32(x1[6], x1[25]); + x2[25] = _mm_sub_epi32(x1[6], x1[25]); + x2[7] = _mm_add_epi32(x1[7], x1[24]); + x2[24] = _mm_sub_epi32(x1[7], x1[24]); + x2[8] = _mm_add_epi32(x1[8], x1[23]); + x2[23] = _mm_sub_epi32(x1[8], x1[23]); + x2[9] = _mm_add_epi32(x1[9], x1[22]); + x2[22] = _mm_sub_epi32(x1[9], x1[22]); + x2[10] = _mm_add_epi32(x1[10], x1[21]); + x2[21] = _mm_sub_epi32(x1[10], x1[21]); + x2[11] = _mm_add_epi32(x1[11], x1[20]); + x2[20] = _mm_sub_epi32(x1[11], x1[20]); + x2[12] = _mm_add_epi32(x1[12], x1[19]); + x2[19] = _mm_sub_epi32(x1[12], x1[19]); + x2[13] = _mm_add_epi32(x1[13], x1[18]); + x2[18] = _mm_sub_epi32(x1[13], x1[18]); + x2[14] = _mm_add_epi32(x1[14], x1[17]); + x2[17] = _mm_sub_epi32(x1[14], x1[17]); + x2[15] = _mm_add_epi32(x1[15], x1[16]); + x2[16] = _mm_sub_epi32(x1[15], x1[16]); + x2[32] = x1[32]; + x2[33] = x1[33]; + x2[34] = x1[34]; + x2[35] = x1[35]; + x2[36] = x1[36]; + x2[37] = x1[37]; + x2[38] = x1[38]; + x2[39] = x1[39]; + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x1[40], x1[55], x2[40], x2[55], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x1[41], x1[54], x2[41], x2[54], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x1[42], x1[53], x2[42], x2[53], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x1[43], x1[52], x2[43], x2[52], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x1[44], x1[51], x2[44], x2[51], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x1[45], x1[50], x2[45], x2[50], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x1[46], x1[49], x2[46], x2[49], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x1[47], x1[48], x2[47], x2[48], + __rounding, cos_bit); + x2[56] = x1[56]; + x2[57] = x1[57]; + x2[58] = x1[58]; + x2[59] = x1[59]; + x2[60] = x1[60]; + x2[61] = x1[61]; + x2[62] = x1[62]; + x2[63] = x1[63]; + + // stage 3 + __m128i x3[64]; + x3[0] = _mm_add_epi32(x2[0], x2[15]); + x3[15] = _mm_sub_epi32(x2[0], x2[15]); + x3[1] = _mm_add_epi32(x2[1], x2[14]); + x3[14] = _mm_sub_epi32(x2[1], x2[14]); + x3[2] = _mm_add_epi32(x2[2], x2[13]); + x3[13] = _mm_sub_epi32(x2[2], x2[13]); + x3[3] = _mm_add_epi32(x2[3], x2[12]); + x3[12] = _mm_sub_epi32(x2[3], x2[12]); + x3[4] = _mm_add_epi32(x2[4], x2[11]); + x3[11] = _mm_sub_epi32(x2[4], x2[11]); + x3[5] = _mm_add_epi32(x2[5], x2[10]); + x3[10] = _mm_sub_epi32(x2[5], x2[10]); + x3[6] = _mm_add_epi32(x2[6], x2[9]); + x3[9] = _mm_sub_epi32(x2[6], x2[9]); + x3[7] = _mm_add_epi32(x2[7], x2[8]); + x3[8] = _mm_sub_epi32(x2[7], x2[8]); + x3[16] = x2[16]; + x3[17] = x2[17]; + x3[18] = x2[18]; + x3[19] = x2[19]; + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x2[20], x2[27], x3[20], x3[27], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x2[21], x2[26], x3[21], x3[26], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x2[22], x2[25], x3[22], x3[25], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x2[23], x2[24], x3[23], x3[24], + __rounding, cos_bit); + x3[28] = x2[28]; + x3[29] = x2[29]; + x3[30] = x2[30]; + x3[31] = x2[31]; + x3[32] = _mm_add_epi32(x2[32], x2[47]); + x3[47] = _mm_sub_epi32(x2[32], x2[47]); + x3[33] = _mm_add_epi32(x2[33], x2[46]); + x3[46] = _mm_sub_epi32(x2[33], x2[46]); + x3[34] = _mm_add_epi32(x2[34], x2[45]); + x3[45] = _mm_sub_epi32(x2[34], x2[45]); + x3[35] = _mm_add_epi32(x2[35], x2[44]); + x3[44] = _mm_sub_epi32(x2[35], x2[44]); + x3[36] = _mm_add_epi32(x2[36], x2[43]); + x3[43] = _mm_sub_epi32(x2[36], x2[43]); + x3[37] = _mm_add_epi32(x2[37], x2[42]); + x3[42] = _mm_sub_epi32(x2[37], x2[42]); + x3[38] = _mm_add_epi32(x2[38], x2[41]); + x3[41] = _mm_sub_epi32(x2[38], x2[41]); + x3[39] = _mm_add_epi32(x2[39], x2[40]); + x3[40] = _mm_sub_epi32(x2[39], x2[40]); + x3[48] = _mm_sub_epi32(x2[63], x2[48]); + x3[63] = _mm_add_epi32(x2[63], x2[48]); + x3[49] = _mm_sub_epi32(x2[62], x2[49]); + x3[62] = _mm_add_epi32(x2[62], x2[49]); + x3[50] = _mm_sub_epi32(x2[61], x2[50]); + x3[61] = _mm_add_epi32(x2[61], x2[50]); + x3[51] = _mm_sub_epi32(x2[60], x2[51]); + x3[60] = _mm_add_epi32(x2[60], x2[51]); + x3[52] = _mm_sub_epi32(x2[59], x2[52]); + x3[59] = _mm_add_epi32(x2[59], x2[52]); + x3[53] = _mm_sub_epi32(x2[58], x2[53]); + x3[58] = _mm_add_epi32(x2[58], x2[53]); + x3[54] = _mm_sub_epi32(x2[57], x2[54]); + x3[57] = _mm_add_epi32(x2[57], x2[54]); + x3[55] = _mm_sub_epi32(x2[56], x2[55]); + x3[56] = _mm_add_epi32(x2[56], x2[55]); + + // stage 4 + __m128i x4[64]; + x4[0] = _mm_add_epi32(x3[0], x3[7]); + x4[7] = _mm_sub_epi32(x3[0], x3[7]); + x4[1] = _mm_add_epi32(x3[1], x3[6]); + x4[6] = _mm_sub_epi32(x3[1], x3[6]); + x4[2] = _mm_add_epi32(x3[2], x3[5]); + x4[5] = _mm_sub_epi32(x3[2], x3[5]); + x4[3] = _mm_add_epi32(x3[3], x3[4]); + x4[4] = _mm_sub_epi32(x3[3], x3[4]); + x4[8] = x3[8]; + x4[9] = x3[9]; + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x3[10], x3[13], x4[10], x4[13], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x3[11], x3[12], x4[11], x4[12], + __rounding, cos_bit); + x4[14] = x3[14]; + x4[15] = x3[15]; + x4[16] = _mm_add_epi32(x3[16], x3[23]); + x4[23] = _mm_sub_epi32(x3[16], x3[23]); + x4[17] = _mm_add_epi32(x3[17], x3[22]); + x4[22] = _mm_sub_epi32(x3[17], x3[22]); + x4[18] = _mm_add_epi32(x3[18], x3[21]); + x4[21] = _mm_sub_epi32(x3[18], x3[21]); + x4[19] = _mm_add_epi32(x3[19], x3[20]); + x4[20] = _mm_sub_epi32(x3[19], x3[20]); + x4[24] = _mm_sub_epi32(x3[31], x3[24]); + x4[31] = _mm_add_epi32(x3[31], x3[24]); + x4[25] = _mm_sub_epi32(x3[30], x3[25]); + x4[30] = _mm_add_epi32(x3[30], x3[25]); + x4[26] = _mm_sub_epi32(x3[29], x3[26]); + x4[29] = _mm_add_epi32(x3[29], x3[26]); + x4[27] = _mm_sub_epi32(x3[28], x3[27]); + x4[28] = _mm_add_epi32(x3[28], x3[27]); + x4[32] = x3[32]; + x4[33] = x3[33]; + x4[34] = x3[34]; + x4[35] = x3[35]; + btf_32_type0_sse4_1_new(cospi_m16, cospi_p48, x3[36], x3[59], x4[36], x4[59], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m16, cospi_p48, x3[37], x3[58], x4[37], x4[58], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m16, cospi_p48, x3[38], x3[57], x4[38], x4[57], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m16, cospi_p48, x3[39], x3[56], x4[39], x4[56], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m48, cospi_m16, x3[40], x3[55], x4[40], x4[55], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m48, cospi_m16, x3[41], x3[54], x4[41], x4[54], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m48, cospi_m16, x3[42], x3[53], x4[42], x4[53], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m48, cospi_m16, x3[43], x3[52], x4[43], x4[52], + __rounding, cos_bit); + x4[44] = x3[44]; + x4[45] = x3[45]; + x4[46] = x3[46]; + x4[47] = x3[47]; + x4[48] = x3[48]; + x4[49] = x3[49]; + x4[50] = x3[50]; + x4[51] = x3[51]; + x4[60] = x3[60]; + x4[61] = x3[61]; + x4[62] = x3[62]; + x4[63] = x3[63]; + + // stage 5 + __m128i x5[64]; + x5[0] = _mm_add_epi32(x4[0], x4[3]); + x5[3] = _mm_sub_epi32(x4[0], x4[3]); + x5[1] = _mm_add_epi32(x4[1], x4[2]); + x5[2] = _mm_sub_epi32(x4[1], x4[2]); + x5[4] = x4[4]; + btf_32_type0_sse4_1_new(cospi_m32, cospi_p32, x4[5], x4[6], x5[5], x5[6], + __rounding, cos_bit); + x5[7] = x4[7]; + x5[8] = _mm_add_epi32(x4[8], x4[11]); + x5[11] = _mm_sub_epi32(x4[8], x4[11]); + x5[9] = _mm_add_epi32(x4[9], x4[10]); + x5[10] = _mm_sub_epi32(x4[9], x4[10]); + x5[12] = _mm_sub_epi32(x4[15], x4[12]); + x5[15] = _mm_add_epi32(x4[15], x4[12]); + x5[13] = _mm_sub_epi32(x4[14], x4[13]); + x5[14] = _mm_add_epi32(x4[14], x4[13]); + x5[16] = x4[16]; + x5[17] = x4[17]; + btf_32_type0_sse4_1_new(cospi_m16, cospi_p48, x4[18], x4[29], x5[18], x5[29], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m16, cospi_p48, x4[19], x4[28], x5[19], x5[28], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m48, cospi_m16, x4[20], x4[27], x5[20], x5[27], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m48, cospi_m16, x4[21], x4[26], x5[21], x5[26], + __rounding, cos_bit); + x5[22] = x4[22]; + x5[23] = x4[23]; + x5[24] = x4[24]; + x5[25] = x4[25]; + x5[30] = x4[30]; + x5[31] = x4[31]; + x5[32] = _mm_add_epi32(x4[32], x4[39]); + x5[39] = _mm_sub_epi32(x4[32], x4[39]); + x5[33] = _mm_add_epi32(x4[33], x4[38]); + x5[38] = _mm_sub_epi32(x4[33], x4[38]); + x5[34] = _mm_add_epi32(x4[34], x4[37]); + x5[37] = _mm_sub_epi32(x4[34], x4[37]); + x5[35] = _mm_add_epi32(x4[35], x4[36]); + x5[36] = _mm_sub_epi32(x4[35], x4[36]); + x5[40] = _mm_sub_epi32(x4[47], x4[40]); + x5[47] = _mm_add_epi32(x4[47], x4[40]); + x5[41] = _mm_sub_epi32(x4[46], x4[41]); + x5[46] = _mm_add_epi32(x4[46], x4[41]); + x5[42] = _mm_sub_epi32(x4[45], x4[42]); + x5[45] = _mm_add_epi32(x4[45], x4[42]); + x5[43] = _mm_sub_epi32(x4[44], x4[43]); + x5[44] = _mm_add_epi32(x4[44], x4[43]); + x5[48] = _mm_add_epi32(x4[48], x4[55]); + x5[55] = _mm_sub_epi32(x4[48], x4[55]); + x5[49] = _mm_add_epi32(x4[49], x4[54]); + x5[54] = _mm_sub_epi32(x4[49], x4[54]); + x5[50] = _mm_add_epi32(x4[50], x4[53]); + x5[53] = _mm_sub_epi32(x4[50], x4[53]); + x5[51] = _mm_add_epi32(x4[51], x4[52]); + x5[52] = _mm_sub_epi32(x4[51], x4[52]); + x5[56] = _mm_sub_epi32(x4[63], x4[56]); + x5[63] = _mm_add_epi32(x4[63], x4[56]); + x5[57] = _mm_sub_epi32(x4[62], x4[57]); + x5[62] = _mm_add_epi32(x4[62], x4[57]); + x5[58] = _mm_sub_epi32(x4[61], x4[58]); + x5[61] = _mm_add_epi32(x4[61], x4[58]); + x5[59] = _mm_sub_epi32(x4[60], x4[59]); + x5[60] = _mm_add_epi32(x4[60], x4[59]); + + // stage 6 + __m128i x6[64]; + btf_32_type0_sse4_1_new(cospi_p32, cospi_p32, x5[0], x5[1], x6[0], x6[1], + __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p48, cospi_p16, x5[2], x5[3], x6[2], x6[3], + __rounding, cos_bit); + x6[4] = _mm_add_epi32(x5[4], x5[5]); + x6[5] = _mm_sub_epi32(x5[4], x5[5]); + x6[6] = _mm_sub_epi32(x5[7], x5[6]); + x6[7] = _mm_add_epi32(x5[7], x5[6]); + x6[8] = x5[8]; + btf_32_type0_sse4_1_new(cospi_m16, cospi_p48, x5[9], x5[14], x6[9], x6[14], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m48, cospi_m16, x5[10], x5[13], x6[10], x6[13], + __rounding, cos_bit); + x6[11] = x5[11]; + x6[12] = x5[12]; + x6[15] = x5[15]; + x6[16] = _mm_add_epi32(x5[16], x5[19]); + x6[19] = _mm_sub_epi32(x5[16], x5[19]); + x6[17] = _mm_add_epi32(x5[17], x5[18]); + x6[18] = _mm_sub_epi32(x5[17], x5[18]); + x6[20] = _mm_sub_epi32(x5[23], x5[20]); + x6[23] = _mm_add_epi32(x5[23], x5[20]); + x6[21] = _mm_sub_epi32(x5[22], x5[21]); + x6[22] = _mm_add_epi32(x5[22], x5[21]); + x6[24] = _mm_add_epi32(x5[24], x5[27]); + x6[27] = _mm_sub_epi32(x5[24], x5[27]); + x6[25] = _mm_add_epi32(x5[25], x5[26]); + x6[26] = _mm_sub_epi32(x5[25], x5[26]); + x6[28] = _mm_sub_epi32(x5[31], x5[28]); + x6[31] = _mm_add_epi32(x5[31], x5[28]); + x6[29] = _mm_sub_epi32(x5[30], x5[29]); + x6[30] = _mm_add_epi32(x5[30], x5[29]); + x6[32] = x5[32]; + x6[33] = x5[33]; + btf_32_type0_sse4_1_new(cospi_m08, cospi_p56, x5[34], x5[61], x6[34], x6[61], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m08, cospi_p56, x5[35], x5[60], x6[35], x6[60], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m56, cospi_m08, x5[36], x5[59], x6[36], x6[59], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m56, cospi_m08, x5[37], x5[58], x6[37], x6[58], + __rounding, cos_bit); + x6[38] = x5[38]; + x6[39] = x5[39]; + x6[40] = x5[40]; + x6[41] = x5[41]; + btf_32_type0_sse4_1_new(cospi_m40, cospi_p24, x5[42], x5[53], x6[42], x6[53], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m40, cospi_p24, x5[43], x5[52], x6[43], x6[52], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m24, cospi_m40, x5[44], x5[51], x6[44], x6[51], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m24, cospi_m40, x5[45], x5[50], x6[45], x6[50], + __rounding, cos_bit); + x6[46] = x5[46]; + x6[47] = x5[47]; + x6[48] = x5[48]; + x6[49] = x5[49]; + x6[54] = x5[54]; + x6[55] = x5[55]; + x6[56] = x5[56]; + x6[57] = x5[57]; + x6[62] = x5[62]; + x6[63] = x5[63]; + + // stage 7 + __m128i x7[64]; + x7[0] = x6[0]; + x7[1] = x6[1]; + x7[2] = x6[2]; + x7[3] = x6[3]; + btf_32_type1_sse4_1_new(cospi_p56, cospi_p08, x6[4], x6[7], x7[4], x7[7], + __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p24, cospi_p40, x6[5], x6[6], x7[5], x7[6], + __rounding, cos_bit); + x7[8] = _mm_add_epi32(x6[8], x6[9]); + x7[9] = _mm_sub_epi32(x6[8], x6[9]); + x7[10] = _mm_sub_epi32(x6[11], x6[10]); + x7[11] = _mm_add_epi32(x6[11], x6[10]); + x7[12] = _mm_add_epi32(x6[12], x6[13]); + x7[13] = _mm_sub_epi32(x6[12], x6[13]); + x7[14] = _mm_sub_epi32(x6[15], x6[14]); + x7[15] = _mm_add_epi32(x6[15], x6[14]); + x7[16] = x6[16]; + btf_32_type0_sse4_1_new(cospi_m08, cospi_p56, x6[17], x6[30], x7[17], x7[30], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m56, cospi_m08, x6[18], x6[29], x7[18], x7[29], + __rounding, cos_bit); + x7[19] = x6[19]; + x7[20] = x6[20]; + btf_32_type0_sse4_1_new(cospi_m40, cospi_p24, x6[21], x6[26], x7[21], x7[26], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m24, cospi_m40, x6[22], x6[25], x7[22], x7[25], + __rounding, cos_bit); + x7[23] = x6[23]; + x7[24] = x6[24]; + x7[27] = x6[27]; + x7[28] = x6[28]; + x7[31] = x6[31]; + x7[32] = _mm_add_epi32(x6[32], x6[35]); + x7[35] = _mm_sub_epi32(x6[32], x6[35]); + x7[33] = _mm_add_epi32(x6[33], x6[34]); + x7[34] = _mm_sub_epi32(x6[33], x6[34]); + x7[36] = _mm_sub_epi32(x6[39], x6[36]); + x7[39] = _mm_add_epi32(x6[39], x6[36]); + x7[37] = _mm_sub_epi32(x6[38], x6[37]); + x7[38] = _mm_add_epi32(x6[38], x6[37]); + x7[40] = _mm_add_epi32(x6[40], x6[43]); + x7[43] = _mm_sub_epi32(x6[40], x6[43]); + x7[41] = _mm_add_epi32(x6[41], x6[42]); + x7[42] = _mm_sub_epi32(x6[41], x6[42]); + x7[44] = _mm_sub_epi32(x6[47], x6[44]); + x7[47] = _mm_add_epi32(x6[47], x6[44]); + x7[45] = _mm_sub_epi32(x6[46], x6[45]); + x7[46] = _mm_add_epi32(x6[46], x6[45]); + x7[48] = _mm_add_epi32(x6[48], x6[51]); + x7[51] = _mm_sub_epi32(x6[48], x6[51]); + x7[49] = _mm_add_epi32(x6[49], x6[50]); + x7[50] = _mm_sub_epi32(x6[49], x6[50]); + x7[52] = _mm_sub_epi32(x6[55], x6[52]); + x7[55] = _mm_add_epi32(x6[55], x6[52]); + x7[53] = _mm_sub_epi32(x6[54], x6[53]); + x7[54] = _mm_add_epi32(x6[54], x6[53]); + x7[56] = _mm_add_epi32(x6[56], x6[59]); + x7[59] = _mm_sub_epi32(x6[56], x6[59]); + x7[57] = _mm_add_epi32(x6[57], x6[58]); + x7[58] = _mm_sub_epi32(x6[57], x6[58]); + x7[60] = _mm_sub_epi32(x6[63], x6[60]); + x7[63] = _mm_add_epi32(x6[63], x6[60]); + x7[61] = _mm_sub_epi32(x6[62], x6[61]); + x7[62] = _mm_add_epi32(x6[62], x6[61]); + + // stage 8 + __m128i x8[64]; + x8[0] = x7[0]; + x8[1] = x7[1]; + x8[2] = x7[2]; + x8[3] = x7[3]; + x8[4] = x7[4]; + x8[5] = x7[5]; + x8[6] = x7[6]; + x8[7] = x7[7]; + btf_32_type1_sse4_1_new(cospi_p60, cospi_p04, x7[8], x7[15], x8[8], x8[15], + __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p28, cospi_p36, x7[9], x7[14], x8[9], x8[14], + __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p44, cospi_p20, x7[10], x7[13], x8[10], x8[13], + __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p12, cospi_p52, x7[11], x7[12], x8[11], x8[12], + __rounding, cos_bit); + x8[16] = _mm_add_epi32(x7[16], x7[17]); + x8[17] = _mm_sub_epi32(x7[16], x7[17]); + x8[18] = _mm_sub_epi32(x7[19], x7[18]); + x8[19] = _mm_add_epi32(x7[19], x7[18]); + x8[20] = _mm_add_epi32(x7[20], x7[21]); + x8[21] = _mm_sub_epi32(x7[20], x7[21]); + x8[22] = _mm_sub_epi32(x7[23], x7[22]); + x8[23] = _mm_add_epi32(x7[23], x7[22]); + x8[24] = _mm_add_epi32(x7[24], x7[25]); + x8[25] = _mm_sub_epi32(x7[24], x7[25]); + x8[26] = _mm_sub_epi32(x7[27], x7[26]); + x8[27] = _mm_add_epi32(x7[27], x7[26]); + x8[28] = _mm_add_epi32(x7[28], x7[29]); + x8[29] = _mm_sub_epi32(x7[28], x7[29]); + x8[30] = _mm_sub_epi32(x7[31], x7[30]); + x8[31] = _mm_add_epi32(x7[31], x7[30]); + x8[32] = x7[32]; + btf_32_type0_sse4_1_new(cospi_m04, cospi_p60, x7[33], x7[62], x8[33], x8[62], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m60, cospi_m04, x7[34], x7[61], x8[34], x8[61], + __rounding, cos_bit); + x8[35] = x7[35]; + x8[36] = x7[36]; + btf_32_type0_sse4_1_new(cospi_m36, cospi_p28, x7[37], x7[58], x8[37], x8[58], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m28, cospi_m36, x7[38], x7[57], x8[38], x8[57], + __rounding, cos_bit); + x8[39] = x7[39]; + x8[40] = x7[40]; + btf_32_type0_sse4_1_new(cospi_m20, cospi_p44, x7[41], x7[54], x8[41], x8[54], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m44, cospi_m20, x7[42], x7[53], x8[42], x8[53], + __rounding, cos_bit); + x8[43] = x7[43]; + x8[44] = x7[44]; + btf_32_type0_sse4_1_new(cospi_m52, cospi_p12, x7[45], x7[50], x8[45], x8[50], + __rounding, cos_bit); + btf_32_type0_sse4_1_new(cospi_m12, cospi_m52, x7[46], x7[49], x8[46], x8[49], + __rounding, cos_bit); + x8[47] = x7[47]; + x8[48] = x7[48]; + x8[51] = x7[51]; + x8[52] = x7[52]; + x8[55] = x7[55]; + x8[56] = x7[56]; + x8[59] = x7[59]; + x8[60] = x7[60]; + x8[63] = x7[63]; + + // stage 9 + __m128i x9[64]; + x9[0] = x8[0]; + x9[1] = x8[1]; + x9[2] = x8[2]; + x9[3] = x8[3]; + x9[4] = x8[4]; + x9[5] = x8[5]; + x9[6] = x8[6]; + x9[7] = x8[7]; + x9[8] = x8[8]; + x9[9] = x8[9]; + x9[10] = x8[10]; + x9[11] = x8[11]; + x9[12] = x8[12]; + x9[13] = x8[13]; + x9[14] = x8[14]; + x9[15] = x8[15]; + btf_32_type1_sse4_1_new(cospi_p62, cospi_p02, x8[16], x8[31], x9[16], x9[31], + __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p30, cospi_p34, x8[17], x8[30], x9[17], x9[30], + __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p46, cospi_p18, x8[18], x8[29], x9[18], x9[29], + __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p14, cospi_p50, x8[19], x8[28], x9[19], x9[28], + __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p54, cospi_p10, x8[20], x8[27], x9[20], x9[27], + __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p22, cospi_p42, x8[21], x8[26], x9[21], x9[26], + __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p38, cospi_p26, x8[22], x8[25], x9[22], x9[25], + __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p06, cospi_p58, x8[23], x8[24], x9[23], x9[24], + __rounding, cos_bit); + x9[32] = _mm_add_epi32(x8[32], x8[33]); + x9[33] = _mm_sub_epi32(x8[32], x8[33]); + x9[34] = _mm_sub_epi32(x8[35], x8[34]); + x9[35] = _mm_add_epi32(x8[35], x8[34]); + x9[36] = _mm_add_epi32(x8[36], x8[37]); + x9[37] = _mm_sub_epi32(x8[36], x8[37]); + x9[38] = _mm_sub_epi32(x8[39], x8[38]); + x9[39] = _mm_add_epi32(x8[39], x8[38]); + x9[40] = _mm_add_epi32(x8[40], x8[41]); + x9[41] = _mm_sub_epi32(x8[40], x8[41]); + x9[42] = _mm_sub_epi32(x8[43], x8[42]); + x9[43] = _mm_add_epi32(x8[43], x8[42]); + x9[44] = _mm_add_epi32(x8[44], x8[45]); + x9[45] = _mm_sub_epi32(x8[44], x8[45]); + x9[46] = _mm_sub_epi32(x8[47], x8[46]); + x9[47] = _mm_add_epi32(x8[47], x8[46]); + x9[48] = _mm_add_epi32(x8[48], x8[49]); + x9[49] = _mm_sub_epi32(x8[48], x8[49]); + x9[50] = _mm_sub_epi32(x8[51], x8[50]); + x9[51] = _mm_add_epi32(x8[51], x8[50]); + x9[52] = _mm_add_epi32(x8[52], x8[53]); + x9[53] = _mm_sub_epi32(x8[52], x8[53]); + x9[54] = _mm_sub_epi32(x8[55], x8[54]); + x9[55] = _mm_add_epi32(x8[55], x8[54]); + x9[56] = _mm_add_epi32(x8[56], x8[57]); + x9[57] = _mm_sub_epi32(x8[56], x8[57]); + x9[58] = _mm_sub_epi32(x8[59], x8[58]); + x9[59] = _mm_add_epi32(x8[59], x8[58]); + x9[60] = _mm_add_epi32(x8[60], x8[61]); + x9[61] = _mm_sub_epi32(x8[60], x8[61]); + x9[62] = _mm_sub_epi32(x8[63], x8[62]); + x9[63] = _mm_add_epi32(x8[63], x8[62]); + + // stage 10 + __m128i x10[64]; + x10[0] = x9[0]; + x10[1] = x9[1]; + x10[2] = x9[2]; + x10[3] = x9[3]; + x10[4] = x9[4]; + x10[5] = x9[5]; + x10[6] = x9[6]; + x10[7] = x9[7]; + x10[8] = x9[8]; + x10[9] = x9[9]; + x10[10] = x9[10]; + x10[11] = x9[11]; + x10[12] = x9[12]; + x10[13] = x9[13]; + x10[14] = x9[14]; + x10[15] = x9[15]; + x10[16] = x9[16]; + x10[17] = x9[17]; + x10[18] = x9[18]; + x10[19] = x9[19]; + x10[20] = x9[20]; + x10[21] = x9[21]; + x10[22] = x9[22]; + x10[23] = x9[23]; + x10[24] = x9[24]; + x10[25] = x9[25]; + x10[26] = x9[26]; + x10[27] = x9[27]; + x10[28] = x9[28]; + x10[29] = x9[29]; + x10[30] = x9[30]; + x10[31] = x9[31]; + btf_32_type1_sse4_1_new(cospi_p63, cospi_p01, x9[32], x9[63], x10[32], + x10[63], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p31, cospi_p33, x9[33], x9[62], x10[33], + x10[62], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p47, cospi_p17, x9[34], x9[61], x10[34], + x10[61], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p15, cospi_p49, x9[35], x9[60], x10[35], + x10[60], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p55, cospi_p09, x9[36], x9[59], x10[36], + x10[59], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p23, cospi_p41, x9[37], x9[58], x10[37], + x10[58], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p39, cospi_p25, x9[38], x9[57], x10[38], + x10[57], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p07, cospi_p57, x9[39], x9[56], x10[39], + x10[56], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p59, cospi_p05, x9[40], x9[55], x10[40], + x10[55], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p27, cospi_p37, x9[41], x9[54], x10[41], + x10[54], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p43, cospi_p21, x9[42], x9[53], x10[42], + x10[53], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p11, cospi_p53, x9[43], x9[52], x10[43], + x10[52], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p51, cospi_p13, x9[44], x9[51], x10[44], + x10[51], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p19, cospi_p45, x9[45], x9[50], x10[45], + x10[50], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p35, cospi_p29, x9[46], x9[49], x10[46], + x10[49], __rounding, cos_bit); + btf_32_type1_sse4_1_new(cospi_p03, cospi_p61, x9[47], x9[48], x10[47], + x10[48], __rounding, cos_bit); + + // stage 11 + output[0 * outstride] = x10[0]; + output[1 * outstride] = x10[32]; + output[2 * outstride] = x10[16]; + output[3 * outstride] = x10[48]; + output[4 * outstride] = x10[8]; + output[5 * outstride] = x10[40]; + output[6 * outstride] = x10[24]; + output[7 * outstride] = x10[56]; + output[8 * outstride] = x10[4]; + output[9 * outstride] = x10[36]; + output[10 * outstride] = x10[20]; + output[11 * outstride] = x10[52]; + output[12 * outstride] = x10[12]; + output[13 * outstride] = x10[44]; + output[14 * outstride] = x10[28]; + output[15 * outstride] = x10[60]; + output[16 * outstride] = x10[2]; + output[17 * outstride] = x10[34]; + output[18 * outstride] = x10[18]; + output[19 * outstride] = x10[50]; + output[20 * outstride] = x10[10]; + output[21 * outstride] = x10[42]; + output[22 * outstride] = x10[26]; + output[23 * outstride] = x10[58]; + output[24 * outstride] = x10[6]; + output[25 * outstride] = x10[38]; + output[26 * outstride] = x10[22]; + output[27 * outstride] = x10[54]; + output[28 * outstride] = x10[14]; + output[29 * outstride] = x10[46]; + output[30 * outstride] = x10[30]; + output[31 * outstride] = x10[62]; + output[32 * outstride] = x10[1]; + output[33 * outstride] = x10[33]; + output[34 * outstride] = x10[17]; + output[35 * outstride] = x10[49]; + output[36 * outstride] = x10[9]; + output[37 * outstride] = x10[41]; + output[38 * outstride] = x10[25]; + output[39 * outstride] = x10[57]; + output[40 * outstride] = x10[5]; + output[41 * outstride] = x10[37]; + output[42 * outstride] = x10[21]; + output[43 * outstride] = x10[53]; + output[44 * outstride] = x10[13]; + output[45 * outstride] = x10[45]; + output[46 * outstride] = x10[29]; + output[47 * outstride] = x10[61]; + output[48 * outstride] = x10[3]; + output[49 * outstride] = x10[35]; + output[50 * outstride] = x10[19]; + output[51 * outstride] = x10[51]; + output[52 * outstride] = x10[11]; + output[53 * outstride] = x10[43]; + output[54 * outstride] = x10[27]; + output[55 * outstride] = x10[59]; + output[56 * outstride] = x10[7]; + output[57 * outstride] = x10[39]; + output[58 * outstride] = x10[23]; + output[59 * outstride] = x10[55]; + output[60 * outstride] = x10[15]; + output[61 * outstride] = x10[47]; + output[62 * outstride] = x10[31]; + output[63 * outstride] = x10[63]; +} diff --git a/third_party/aom/av1/encoder/x86/av1_fwd_txfm2d_avx2.c b/third_party/aom/av1/encoder/x86/av1_fwd_txfm2d_avx2.c new file mode 100644 index 0000000000..592462e20d --- /dev/null +++ b/third_party/aom/av1/encoder/x86/av1_fwd_txfm2d_avx2.c @@ -0,0 +1,2068 @@ +/* + * Copyright (c) 2018, 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 "config/av1_rtcd.h" + +#include "av1/common/enums.h" +#include "av1/common/av1_txfm.h" +#include "av1/encoder/x86/av1_fwd_txfm_avx2.h" +#include "av1/common/x86/av1_txfm_sse2.h" +#include "av1/encoder/av1_fwd_txfm1d_cfg.h" +#include "av1/encoder/x86/av1_txfm1d_sse4.h" +#include "av1/encoder/x86/av1_fwd_txfm_sse2.h" +#include "aom_dsp/x86/txfm_common_avx2.h" + +static INLINE void fdct16x16_new_avx2(const __m256i *input, __m256i *output, + int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m256i _r = _mm256_set1_epi32(1 << (cos_bit - 1)); + + __m256i cospi_m32_p32 = pair_set_w16_epi16(-cospi[32], cospi[32]); + __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]); + __m256i cospi_p32_m32 = pair_set_w16_epi16(cospi[32], -cospi[32]); + __m256i cospi_p48_p16 = pair_set_w16_epi16(cospi[48], cospi[16]); + __m256i cospi_m16_p48 = pair_set_w16_epi16(-cospi[16], cospi[48]); + __m256i cospi_m48_m16 = pair_set_w16_epi16(-cospi[48], -cospi[16]); + __m256i cospi_p56_p08 = pair_set_w16_epi16(cospi[56], cospi[8]); + __m256i cospi_m08_p56 = pair_set_w16_epi16(-cospi[8], cospi[56]); + __m256i cospi_p24_p40 = pair_set_w16_epi16(cospi[24], cospi[40]); + __m256i cospi_m40_p24 = pair_set_w16_epi16(-cospi[40], cospi[24]); + __m256i cospi_p60_p04 = pair_set_w16_epi16(cospi[60], cospi[4]); + __m256i cospi_m04_p60 = pair_set_w16_epi16(-cospi[4], cospi[60]); + __m256i cospi_p28_p36 = pair_set_w16_epi16(cospi[28], cospi[36]); + __m256i cospi_m36_p28 = pair_set_w16_epi16(-cospi[36], cospi[28]); + __m256i cospi_p44_p20 = pair_set_w16_epi16(cospi[44], cospi[20]); + __m256i cospi_m20_p44 = pair_set_w16_epi16(-cospi[20], cospi[44]); + __m256i cospi_p12_p52 = pair_set_w16_epi16(cospi[12], cospi[52]); + __m256i cospi_m52_p12 = pair_set_w16_epi16(-cospi[52], cospi[12]); + + // stage 1 + __m256i x1[16]; + btf_16_adds_subs_out_avx2(&x1[0], &x1[15], input[0], input[15]); + btf_16_adds_subs_out_avx2(&x1[1], &x1[14], input[1], input[14]); + btf_16_adds_subs_out_avx2(&x1[2], &x1[13], input[2], input[13]); + btf_16_adds_subs_out_avx2(&x1[3], &x1[12], input[3], input[12]); + btf_16_adds_subs_out_avx2(&x1[4], &x1[11], input[4], input[11]); + btf_16_adds_subs_out_avx2(&x1[5], &x1[10], input[5], input[10]); + btf_16_adds_subs_out_avx2(&x1[6], &x1[9], input[6], input[9]); + btf_16_adds_subs_out_avx2(&x1[7], &x1[8], input[7], input[8]); + + // stage 2 + btf_16_adds_subs_avx2(&x1[0], &x1[7]); + btf_16_adds_subs_avx2(&x1[1], &x1[6]); + btf_16_adds_subs_avx2(&x1[2], &x1[5]); + btf_16_adds_subs_avx2(&x1[3], &x1[4]); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[10], &x1[13], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[11], &x1[12], _r, cos_bit); + + // stage 3 + btf_16_adds_subs_avx2(&x1[0], &x1[3]); + btf_16_adds_subs_avx2(&x1[1], &x1[2]); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[5], &x1[6], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[8], &x1[11]); + btf_16_adds_subs_avx2(&x1[9], &x1[10]); + btf_16_adds_subs_avx2(&x1[15], &x1[12]); + btf_16_adds_subs_avx2(&x1[14], &x1[13]); + + // stage 4 + btf_16_w16_avx2(cospi_p32_p32, cospi_p32_m32, &x1[0], &x1[1], _r, cos_bit); + btf_16_w16_avx2(cospi_p48_p16, cospi_m16_p48, &x1[2], &x1[3], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[4], &x1[5]); + btf_16_adds_subs_avx2(&x1[7], &x1[6]); + btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x1[9], &x1[14], _r, cos_bit); + btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x1[10], &x1[13], _r, cos_bit); + + // stage 5 + btf_16_w16_avx2(cospi_p56_p08, cospi_m08_p56, &x1[4], &x1[7], _r, cos_bit); + btf_16_w16_avx2(cospi_p24_p40, cospi_m40_p24, &x1[5], &x1[6], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[8], &x1[9]); + btf_16_adds_subs_avx2(&x1[11], &x1[10]); + btf_16_adds_subs_avx2(&x1[12], &x1[13]); + btf_16_adds_subs_avx2(&x1[15], &x1[14]); + + // stage 6 + btf_16_w16_avx2(cospi_p60_p04, cospi_m04_p60, &x1[8], &x1[15], _r, cos_bit); + btf_16_w16_avx2(cospi_p28_p36, cospi_m36_p28, &x1[9], &x1[14], _r, cos_bit); + btf_16_w16_avx2(cospi_p44_p20, cospi_m20_p44, &x1[10], &x1[13], _r, cos_bit); + btf_16_w16_avx2(cospi_p12_p52, cospi_m52_p12, &x1[11], &x1[12], _r, cos_bit); + + // stage 7 + output[0] = x1[0]; + output[1] = x1[8]; + output[2] = x1[4]; + output[3] = x1[12]; + output[4] = x1[2]; + output[5] = x1[10]; + output[6] = x1[6]; + output[7] = x1[14]; + output[8] = x1[1]; + output[9] = x1[9]; + output[10] = x1[5]; + output[11] = x1[13]; + output[12] = x1[3]; + output[13] = x1[11]; + output[14] = x1[7]; + output[15] = x1[15]; +} + +static INLINE void fdct16x32_new_avx2(const __m256i *input, __m256i *output, + int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m256i _r = _mm256_set1_epi32(1 << (cos_bit - 1)); + + __m256i cospi_m32_p32 = pair_set_w16_epi16(-cospi[32], cospi[32]); + __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]); + __m256i cospi_m16_p48 = pair_set_w16_epi16(-cospi[16], cospi[48]); + __m256i cospi_p48_p16 = pair_set_w16_epi16(cospi[48], cospi[16]); + __m256i cospi_m48_m16 = pair_set_w16_epi16(-cospi[48], -cospi[16]); + __m256i cospi_p32_m32 = pair_set_w16_epi16(cospi[32], -cospi[32]); + __m256i cospi_p56_p08 = pair_set_w16_epi16(cospi[56], cospi[8]); + __m256i cospi_m08_p56 = pair_set_w16_epi16(-cospi[8], cospi[56]); + __m256i cospi_p24_p40 = pair_set_w16_epi16(cospi[24], cospi[40]); + __m256i cospi_m40_p24 = pair_set_w16_epi16(-cospi[40], cospi[24]); + __m256i cospi_m56_m08 = pair_set_w16_epi16(-cospi[56], -cospi[8]); + __m256i cospi_m24_m40 = pair_set_w16_epi16(-cospi[24], -cospi[40]); + __m256i cospi_p60_p04 = pair_set_w16_epi16(cospi[60], cospi[4]); + __m256i cospi_m04_p60 = pair_set_w16_epi16(-cospi[4], cospi[60]); + __m256i cospi_p28_p36 = pair_set_w16_epi16(cospi[28], cospi[36]); + __m256i cospi_m36_p28 = pair_set_w16_epi16(-cospi[36], cospi[28]); + __m256i cospi_p44_p20 = pair_set_w16_epi16(cospi[44], cospi[20]); + __m256i cospi_m20_p44 = pair_set_w16_epi16(-cospi[20], cospi[44]); + __m256i cospi_p12_p52 = pair_set_w16_epi16(cospi[12], cospi[52]); + __m256i cospi_m52_p12 = pair_set_w16_epi16(-cospi[52], cospi[12]); + __m256i cospi_p62_p02 = pair_set_w16_epi16(cospi[62], cospi[2]); + __m256i cospi_m02_p62 = pair_set_w16_epi16(-cospi[2], cospi[62]); + __m256i cospi_p30_p34 = pair_set_w16_epi16(cospi[30], cospi[34]); + __m256i cospi_m34_p30 = pair_set_w16_epi16(-cospi[34], cospi[30]); + __m256i cospi_p46_p18 = pair_set_w16_epi16(cospi[46], cospi[18]); + __m256i cospi_m18_p46 = pair_set_w16_epi16(-cospi[18], cospi[46]); + __m256i cospi_p14_p50 = pair_set_w16_epi16(cospi[14], cospi[50]); + __m256i cospi_m50_p14 = pair_set_w16_epi16(-cospi[50], cospi[14]); + __m256i cospi_p54_p10 = pair_set_w16_epi16(cospi[54], cospi[10]); + __m256i cospi_m10_p54 = pair_set_w16_epi16(-cospi[10], cospi[54]); + __m256i cospi_p22_p42 = pair_set_w16_epi16(cospi[22], cospi[42]); + __m256i cospi_m42_p22 = pair_set_w16_epi16(-cospi[42], cospi[22]); + __m256i cospi_p38_p26 = pair_set_w16_epi16(cospi[38], cospi[26]); + __m256i cospi_m26_p38 = pair_set_w16_epi16(-cospi[26], cospi[38]); + __m256i cospi_p06_p58 = pair_set_w16_epi16(cospi[6], cospi[58]); + __m256i cospi_m58_p06 = pair_set_w16_epi16(-cospi[58], cospi[6]); + + // stage 1 + __m256i x1[32]; + btf_16_adds_subs_out_avx2(&x1[0], &x1[31], input[0], input[31]); + btf_16_adds_subs_out_avx2(&x1[1], &x1[30], input[1], input[30]); + btf_16_adds_subs_out_avx2(&x1[2], &x1[29], input[2], input[29]); + btf_16_adds_subs_out_avx2(&x1[3], &x1[28], input[3], input[28]); + btf_16_adds_subs_out_avx2(&x1[4], &x1[27], input[4], input[27]); + btf_16_adds_subs_out_avx2(&x1[5], &x1[26], input[5], input[26]); + btf_16_adds_subs_out_avx2(&x1[6], &x1[25], input[6], input[25]); + btf_16_adds_subs_out_avx2(&x1[7], &x1[24], input[7], input[24]); + btf_16_adds_subs_out_avx2(&x1[8], &x1[23], input[8], input[23]); + btf_16_adds_subs_out_avx2(&x1[9], &x1[22], input[9], input[22]); + btf_16_adds_subs_out_avx2(&x1[10], &x1[21], input[10], input[21]); + btf_16_adds_subs_out_avx2(&x1[11], &x1[20], input[11], input[20]); + btf_16_adds_subs_out_avx2(&x1[12], &x1[19], input[12], input[19]); + btf_16_adds_subs_out_avx2(&x1[13], &x1[18], input[13], input[18]); + btf_16_adds_subs_out_avx2(&x1[14], &x1[17], input[14], input[17]); + btf_16_adds_subs_out_avx2(&x1[15], &x1[16], input[15], input[16]); + + // stage 2 + btf_16_adds_subs_avx2(&x1[0], &x1[15]); + btf_16_adds_subs_avx2(&x1[1], &x1[14]); + btf_16_adds_subs_avx2(&x1[2], &x1[13]); + btf_16_adds_subs_avx2(&x1[3], &x1[12]); + btf_16_adds_subs_avx2(&x1[4], &x1[11]); + btf_16_adds_subs_avx2(&x1[5], &x1[10]); + btf_16_adds_subs_avx2(&x1[6], &x1[9]); + btf_16_adds_subs_avx2(&x1[7], &x1[8]); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[20], &x1[27], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[21], &x1[26], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[22], &x1[25], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[23], &x1[24], _r, cos_bit); + + // stage 3 + btf_16_adds_subs_avx2(&x1[0], &x1[7]); + btf_16_adds_subs_avx2(&x1[1], &x1[6]); + btf_16_adds_subs_avx2(&x1[2], &x1[5]); + btf_16_adds_subs_avx2(&x1[3], &x1[4]); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[10], &x1[13], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[11], &x1[12], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[16], &x1[23]); + btf_16_adds_subs_avx2(&x1[17], &x1[22]); + btf_16_adds_subs_avx2(&x1[18], &x1[21]); + btf_16_adds_subs_avx2(&x1[19], &x1[20]); + btf_16_adds_subs_avx2(&x1[31], &x1[24]); + btf_16_adds_subs_avx2(&x1[30], &x1[25]); + btf_16_adds_subs_avx2(&x1[29], &x1[26]); + btf_16_adds_subs_avx2(&x1[28], &x1[27]); + + // stage 4 + btf_16_adds_subs_avx2(&x1[0], &x1[3]); + btf_16_adds_subs_avx2(&x1[1], &x1[2]); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[5], &x1[6], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[8], &x1[11]); + btf_16_adds_subs_avx2(&x1[9], &x1[10]); + btf_16_adds_subs_avx2(&x1[15], &x1[12]); + btf_16_adds_subs_avx2(&x1[14], &x1[13]); + btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x1[18], &x1[29], _r, cos_bit); + btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x1[19], &x1[28], _r, cos_bit); + btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x1[20], &x1[27], _r, cos_bit); + btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x1[21], &x1[26], _r, cos_bit); + + // stage 5 + btf_16_w16_avx2(cospi_p32_p32, cospi_p32_m32, &x1[0], &x1[1], _r, cos_bit); + btf_16_w16_avx2(cospi_p48_p16, cospi_m16_p48, &x1[2], &x1[3], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[4], &x1[5]); + btf_16_adds_subs_avx2(&x1[7], &x1[6]); + btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x1[9], &x1[14], _r, cos_bit); + btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x1[10], &x1[13], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[16], &x1[19]); + btf_16_adds_subs_avx2(&x1[17], &x1[18]); + btf_16_adds_subs_avx2(&x1[23], &x1[20]); + btf_16_adds_subs_avx2(&x1[22], &x1[21]); + btf_16_adds_subs_avx2(&x1[24], &x1[27]); + btf_16_adds_subs_avx2(&x1[25], &x1[26]); + btf_16_adds_subs_avx2(&x1[31], &x1[28]); + btf_16_adds_subs_avx2(&x1[30], &x1[29]); + + // stage 6 + btf_16_w16_avx2(cospi_p56_p08, cospi_m08_p56, &x1[4], &x1[7], _r, cos_bit); + btf_16_w16_avx2(cospi_p24_p40, cospi_m40_p24, &x1[5], &x1[6], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[8], &x1[9]); + btf_16_adds_subs_avx2(&x1[11], &x1[10]); + btf_16_adds_subs_avx2(&x1[12], &x1[13]); + btf_16_adds_subs_avx2(&x1[15], &x1[14]); + btf_16_w16_avx2(cospi_m08_p56, cospi_p56_p08, &x1[17], &x1[30], _r, cos_bit); + btf_16_w16_avx2(cospi_m56_m08, cospi_m08_p56, &x1[18], &x1[29], _r, cos_bit); + btf_16_w16_avx2(cospi_m40_p24, cospi_p24_p40, &x1[21], &x1[26], _r, cos_bit); + btf_16_w16_avx2(cospi_m24_m40, cospi_m40_p24, &x1[22], &x1[25], _r, cos_bit); + + // stage 7 + btf_16_w16_avx2(cospi_p60_p04, cospi_m04_p60, &x1[8], &x1[15], _r, cos_bit); + btf_16_w16_avx2(cospi_p28_p36, cospi_m36_p28, &x1[9], &x1[14], _r, cos_bit); + btf_16_w16_avx2(cospi_p44_p20, cospi_m20_p44, &x1[10], &x1[13], _r, cos_bit); + btf_16_w16_avx2(cospi_p12_p52, cospi_m52_p12, &x1[11], &x1[12], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[16], &x1[17]); + btf_16_adds_subs_avx2(&x1[19], &x1[18]); + btf_16_adds_subs_avx2(&x1[20], &x1[21]); + btf_16_adds_subs_avx2(&x1[23], &x1[22]); + btf_16_adds_subs_avx2(&x1[24], &x1[25]); + btf_16_adds_subs_avx2(&x1[27], &x1[26]); + btf_16_adds_subs_avx2(&x1[28], &x1[29]); + btf_16_adds_subs_avx2(&x1[31], &x1[30]); + + // stage 8 + btf_16_w16_avx2(cospi_p62_p02, cospi_m02_p62, &x1[16], &x1[31], _r, cos_bit); + btf_16_w16_avx2(cospi_p30_p34, cospi_m34_p30, &x1[17], &x1[30], _r, cos_bit); + btf_16_w16_avx2(cospi_p46_p18, cospi_m18_p46, &x1[18], &x1[29], _r, cos_bit); + btf_16_w16_avx2(cospi_p14_p50, cospi_m50_p14, &x1[19], &x1[28], _r, cos_bit); + btf_16_w16_avx2(cospi_p54_p10, cospi_m10_p54, &x1[20], &x1[27], _r, cos_bit); + btf_16_w16_avx2(cospi_p22_p42, cospi_m42_p22, &x1[21], &x1[26], _r, cos_bit); + btf_16_w16_avx2(cospi_p38_p26, cospi_m26_p38, &x1[22], &x1[25], _r, cos_bit); + btf_16_w16_avx2(cospi_p06_p58, cospi_m58_p06, &x1[23], &x1[24], _r, cos_bit); + + // stage 9 + output[0] = x1[0]; + output[1] = x1[16]; + output[2] = x1[8]; + output[3] = x1[24]; + output[4] = x1[4]; + output[5] = x1[20]; + output[6] = x1[12]; + output[7] = x1[28]; + output[8] = x1[2]; + output[9] = x1[18]; + output[10] = x1[10]; + output[11] = x1[26]; + output[12] = x1[6]; + output[13] = x1[22]; + output[14] = x1[14]; + output[15] = x1[30]; + output[16] = x1[1]; + output[17] = x1[17]; + output[18] = x1[9]; + output[19] = x1[25]; + output[20] = x1[5]; + output[21] = x1[21]; + output[22] = x1[13]; + output[23] = x1[29]; + output[24] = x1[3]; + output[25] = x1[19]; + output[26] = x1[11]; + output[27] = x1[27]; + output[28] = x1[7]; + output[29] = x1[23]; + output[30] = x1[15]; + output[31] = x1[31]; +} + +static INLINE void fdct16x64_new_avx2(const __m256i *input, __m256i *output, + int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m256i _r = _mm256_set1_epi32(1 << (cos_bit - 1)); + + __m256i cospi_m32_p32 = pair_set_w16_epi16(-cospi[32], cospi[32]); + __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]); + __m256i cospi_m16_p48 = pair_set_w16_epi16(-cospi[16], cospi[48]); + __m256i cospi_p48_p16 = pair_set_w16_epi16(cospi[48], cospi[16]); + __m256i cospi_m48_m16 = pair_set_w16_epi16(-cospi[48], -cospi[16]); + __m256i cospi_p32_m32 = pair_set_w16_epi16(cospi[32], -cospi[32]); + __m256i cospi_m08_p56 = pair_set_w16_epi16(-cospi[8], cospi[56]); + __m256i cospi_p56_p08 = pair_set_w16_epi16(cospi[56], cospi[8]); + __m256i cospi_m56_m08 = pair_set_w16_epi16(-cospi[56], -cospi[8]); + __m256i cospi_m40_p24 = pair_set_w16_epi16(-cospi[40], cospi[24]); + __m256i cospi_p24_p40 = pair_set_w16_epi16(cospi[24], cospi[40]); + __m256i cospi_m24_m40 = pair_set_w16_epi16(-cospi[24], -cospi[40]); + __m256i cospi_p60_p04 = pair_set_w16_epi16(cospi[60], cospi[4]); + __m256i cospi_m04_p60 = pair_set_w16_epi16(-cospi[4], cospi[60]); + __m256i cospi_p28_p36 = pair_set_w16_epi16(cospi[28], cospi[36]); + __m256i cospi_m36_p28 = pair_set_w16_epi16(-cospi[36], cospi[28]); + __m256i cospi_p44_p20 = pair_set_w16_epi16(cospi[44], cospi[20]); + __m256i cospi_m20_p44 = pair_set_w16_epi16(-cospi[20], cospi[44]); + __m256i cospi_p12_p52 = pair_set_w16_epi16(cospi[12], cospi[52]); + __m256i cospi_m52_p12 = pair_set_w16_epi16(-cospi[52], cospi[12]); + __m256i cospi_m60_m04 = pair_set_w16_epi16(-cospi[60], -cospi[4]); + __m256i cospi_m28_m36 = pair_set_w16_epi16(-cospi[28], -cospi[36]); + __m256i cospi_m44_m20 = pair_set_w16_epi16(-cospi[44], -cospi[20]); + __m256i cospi_m12_m52 = pair_set_w16_epi16(-cospi[12], -cospi[52]); + __m256i cospi_p62_p02 = pair_set_w16_epi16(cospi[62], cospi[2]); + __m256i cospi_m02_p62 = pair_set_w16_epi16(-cospi[2], cospi[62]); + __m256i cospi_p30_p34 = pair_set_w16_epi16(cospi[30], cospi[34]); + __m256i cospi_m34_p30 = pair_set_w16_epi16(-cospi[34], cospi[30]); + __m256i cospi_p46_p18 = pair_set_w16_epi16(cospi[46], cospi[18]); + __m256i cospi_m18_p46 = pair_set_w16_epi16(-cospi[18], cospi[46]); + __m256i cospi_p14_p50 = pair_set_w16_epi16(cospi[14], cospi[50]); + __m256i cospi_m50_p14 = pair_set_w16_epi16(-cospi[50], cospi[14]); + __m256i cospi_p54_p10 = pair_set_w16_epi16(cospi[54], cospi[10]); + __m256i cospi_m10_p54 = pair_set_w16_epi16(-cospi[10], cospi[54]); + __m256i cospi_p22_p42 = pair_set_w16_epi16(cospi[22], cospi[42]); + __m256i cospi_m42_p22 = pair_set_w16_epi16(-cospi[42], cospi[22]); + __m256i cospi_p38_p26 = pair_set_w16_epi16(cospi[38], cospi[26]); + __m256i cospi_m26_p38 = pair_set_w16_epi16(-cospi[26], cospi[38]); + __m256i cospi_p06_p58 = pair_set_w16_epi16(cospi[6], cospi[58]); + __m256i cospi_m58_p06 = pair_set_w16_epi16(-cospi[58], cospi[6]); + __m256i cospi_p63_p01 = pair_set_w16_epi16(cospi[63], cospi[1]); + __m256i cospi_m01_p63 = pair_set_w16_epi16(-cospi[1], cospi[63]); + __m256i cospi_p31_p33 = pair_set_w16_epi16(cospi[31], cospi[33]); + __m256i cospi_m33_p31 = pair_set_w16_epi16(-cospi[33], cospi[31]); + __m256i cospi_p47_p17 = pair_set_w16_epi16(cospi[47], cospi[17]); + __m256i cospi_m17_p47 = pair_set_w16_epi16(-cospi[17], cospi[47]); + __m256i cospi_p15_p49 = pair_set_w16_epi16(cospi[15], cospi[49]); + __m256i cospi_m49_p15 = pair_set_w16_epi16(-cospi[49], cospi[15]); + __m256i cospi_p55_p09 = pair_set_w16_epi16(cospi[55], cospi[9]); + __m256i cospi_m09_p55 = pair_set_w16_epi16(-cospi[9], cospi[55]); + __m256i cospi_p23_p41 = pair_set_w16_epi16(cospi[23], cospi[41]); + __m256i cospi_m41_p23 = pair_set_w16_epi16(-cospi[41], cospi[23]); + __m256i cospi_p39_p25 = pair_set_w16_epi16(cospi[39], cospi[25]); + __m256i cospi_m25_p39 = pair_set_w16_epi16(-cospi[25], cospi[39]); + __m256i cospi_p07_p57 = pair_set_w16_epi16(cospi[7], cospi[57]); + __m256i cospi_m57_p07 = pair_set_w16_epi16(-cospi[57], cospi[7]); + __m256i cospi_p59_p05 = pair_set_w16_epi16(cospi[59], cospi[5]); + __m256i cospi_m05_p59 = pair_set_w16_epi16(-cospi[5], cospi[59]); + __m256i cospi_p27_p37 = pair_set_w16_epi16(cospi[27], cospi[37]); + __m256i cospi_m37_p27 = pair_set_w16_epi16(-cospi[37], cospi[27]); + __m256i cospi_p43_p21 = pair_set_w16_epi16(cospi[43], cospi[21]); + __m256i cospi_m21_p43 = pair_set_w16_epi16(-cospi[21], cospi[43]); + __m256i cospi_p11_p53 = pair_set_w16_epi16(cospi[11], cospi[53]); + __m256i cospi_m53_p11 = pair_set_w16_epi16(-cospi[53], cospi[11]); + __m256i cospi_p51_p13 = pair_set_w16_epi16(cospi[51], cospi[13]); + __m256i cospi_m13_p51 = pair_set_w16_epi16(-cospi[13], cospi[51]); + __m256i cospi_p19_p45 = pair_set_w16_epi16(cospi[19], cospi[45]); + __m256i cospi_m45_p19 = pair_set_w16_epi16(-cospi[45], cospi[19]); + __m256i cospi_p35_p29 = pair_set_w16_epi16(cospi[35], cospi[29]); + __m256i cospi_m29_p35 = pair_set_w16_epi16(-cospi[29], cospi[35]); + __m256i cospi_p03_p61 = pair_set_w16_epi16(cospi[3], cospi[61]); + __m256i cospi_m61_p03 = pair_set_w16_epi16(-cospi[61], cospi[3]); + + // stage 1 + __m256i x1[64]; + btf_16_adds_subs_out_avx2(&x1[0], &x1[63], input[0], input[63]); + btf_16_adds_subs_out_avx2(&x1[1], &x1[62], input[1], input[62]); + btf_16_adds_subs_out_avx2(&x1[2], &x1[61], input[2], input[61]); + btf_16_adds_subs_out_avx2(&x1[3], &x1[60], input[3], input[60]); + btf_16_adds_subs_out_avx2(&x1[4], &x1[59], input[4], input[59]); + btf_16_adds_subs_out_avx2(&x1[5], &x1[58], input[5], input[58]); + btf_16_adds_subs_out_avx2(&x1[6], &x1[57], input[6], input[57]); + btf_16_adds_subs_out_avx2(&x1[7], &x1[56], input[7], input[56]); + btf_16_adds_subs_out_avx2(&x1[8], &x1[55], input[8], input[55]); + btf_16_adds_subs_out_avx2(&x1[9], &x1[54], input[9], input[54]); + btf_16_adds_subs_out_avx2(&x1[10], &x1[53], input[10], input[53]); + btf_16_adds_subs_out_avx2(&x1[11], &x1[52], input[11], input[52]); + btf_16_adds_subs_out_avx2(&x1[12], &x1[51], input[12], input[51]); + btf_16_adds_subs_out_avx2(&x1[13], &x1[50], input[13], input[50]); + btf_16_adds_subs_out_avx2(&x1[14], &x1[49], input[14], input[49]); + btf_16_adds_subs_out_avx2(&x1[15], &x1[48], input[15], input[48]); + btf_16_adds_subs_out_avx2(&x1[16], &x1[47], input[16], input[47]); + btf_16_adds_subs_out_avx2(&x1[17], &x1[46], input[17], input[46]); + btf_16_adds_subs_out_avx2(&x1[18], &x1[45], input[18], input[45]); + btf_16_adds_subs_out_avx2(&x1[19], &x1[44], input[19], input[44]); + btf_16_adds_subs_out_avx2(&x1[20], &x1[43], input[20], input[43]); + btf_16_adds_subs_out_avx2(&x1[21], &x1[42], input[21], input[42]); + btf_16_adds_subs_out_avx2(&x1[22], &x1[41], input[22], input[41]); + btf_16_adds_subs_out_avx2(&x1[23], &x1[40], input[23], input[40]); + btf_16_adds_subs_out_avx2(&x1[24], &x1[39], input[24], input[39]); + btf_16_adds_subs_out_avx2(&x1[25], &x1[38], input[25], input[38]); + btf_16_adds_subs_out_avx2(&x1[26], &x1[37], input[26], input[37]); + btf_16_adds_subs_out_avx2(&x1[27], &x1[36], input[27], input[36]); + btf_16_adds_subs_out_avx2(&x1[28], &x1[35], input[28], input[35]); + btf_16_adds_subs_out_avx2(&x1[29], &x1[34], input[29], input[34]); + btf_16_adds_subs_out_avx2(&x1[30], &x1[33], input[30], input[33]); + btf_16_adds_subs_out_avx2(&x1[31], &x1[32], input[31], input[32]); + + // stage 2 + btf_16_adds_subs_avx2(&x1[0], &x1[31]); + btf_16_adds_subs_avx2(&x1[1], &x1[30]); + btf_16_adds_subs_avx2(&x1[2], &x1[29]); + btf_16_adds_subs_avx2(&x1[3], &x1[28]); + btf_16_adds_subs_avx2(&x1[4], &x1[27]); + btf_16_adds_subs_avx2(&x1[5], &x1[26]); + btf_16_adds_subs_avx2(&x1[6], &x1[25]); + btf_16_adds_subs_avx2(&x1[7], &x1[24]); + btf_16_adds_subs_avx2(&x1[8], &x1[23]); + btf_16_adds_subs_avx2(&x1[9], &x1[22]); + btf_16_adds_subs_avx2(&x1[10], &x1[21]); + btf_16_adds_subs_avx2(&x1[11], &x1[20]); + btf_16_adds_subs_avx2(&x1[12], &x1[19]); + btf_16_adds_subs_avx2(&x1[13], &x1[18]); + btf_16_adds_subs_avx2(&x1[14], &x1[17]); + btf_16_adds_subs_avx2(&x1[15], &x1[16]); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[40], &x1[55], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[41], &x1[54], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[42], &x1[53], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[43], &x1[52], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[44], &x1[51], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[45], &x1[50], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[46], &x1[49], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[47], &x1[48], _r, cos_bit); + + // stage 3 + btf_16_adds_subs_avx2(&x1[0], &x1[15]); + btf_16_adds_subs_avx2(&x1[1], &x1[14]); + btf_16_adds_subs_avx2(&x1[2], &x1[13]); + btf_16_adds_subs_avx2(&x1[3], &x1[12]); + btf_16_adds_subs_avx2(&x1[4], &x1[11]); + btf_16_adds_subs_avx2(&x1[5], &x1[10]); + btf_16_adds_subs_avx2(&x1[6], &x1[9]); + btf_16_adds_subs_avx2(&x1[7], &x1[8]); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[20], &x1[27], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[21], &x1[26], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[22], &x1[25], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[23], &x1[24], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[32], &x1[47]); + btf_16_adds_subs_avx2(&x1[33], &x1[46]); + btf_16_adds_subs_avx2(&x1[34], &x1[45]); + btf_16_adds_subs_avx2(&x1[35], &x1[44]); + btf_16_adds_subs_avx2(&x1[36], &x1[43]); + btf_16_adds_subs_avx2(&x1[37], &x1[42]); + btf_16_adds_subs_avx2(&x1[38], &x1[41]); + btf_16_adds_subs_avx2(&x1[39], &x1[40]); + btf_16_adds_subs_avx2(&x1[63], &x1[48]); + btf_16_adds_subs_avx2(&x1[62], &x1[49]); + btf_16_adds_subs_avx2(&x1[61], &x1[50]); + btf_16_adds_subs_avx2(&x1[60], &x1[51]); + btf_16_adds_subs_avx2(&x1[59], &x1[52]); + btf_16_adds_subs_avx2(&x1[58], &x1[53]); + btf_16_adds_subs_avx2(&x1[57], &x1[54]); + btf_16_adds_subs_avx2(&x1[56], &x1[55]); + + // stage 4 + btf_16_adds_subs_avx2(&x1[0], &x1[7]); + btf_16_adds_subs_avx2(&x1[1], &x1[6]); + btf_16_adds_subs_avx2(&x1[2], &x1[5]); + btf_16_adds_subs_avx2(&x1[3], &x1[4]); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[10], &x1[13], _r, cos_bit); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[11], &x1[12], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[16], &x1[23]); + btf_16_adds_subs_avx2(&x1[17], &x1[22]); + btf_16_adds_subs_avx2(&x1[18], &x1[21]); + btf_16_adds_subs_avx2(&x1[19], &x1[20]); + btf_16_adds_subs_avx2(&x1[31], &x1[24]); + btf_16_adds_subs_avx2(&x1[30], &x1[25]); + btf_16_adds_subs_avx2(&x1[29], &x1[26]); + btf_16_adds_subs_avx2(&x1[28], &x1[27]); + btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x1[36], &x1[59], _r, cos_bit); + btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x1[37], &x1[58], _r, cos_bit); + btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x1[38], &x1[57], _r, cos_bit); + btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x1[39], &x1[56], _r, cos_bit); + btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x1[40], &x1[55], _r, cos_bit); + btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x1[41], &x1[54], _r, cos_bit); + btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x1[42], &x1[53], _r, cos_bit); + btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x1[43], &x1[52], _r, cos_bit); + + // stage 5 + btf_16_adds_subs_avx2(&x1[0], &x1[3]); + btf_16_adds_subs_avx2(&x1[1], &x1[2]); + btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[5], &x1[6], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[8], &x1[11]); + btf_16_adds_subs_avx2(&x1[9], &x1[10]); + btf_16_adds_subs_avx2(&x1[15], &x1[12]); + btf_16_adds_subs_avx2(&x1[14], &x1[13]); + btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x1[18], &x1[29], _r, cos_bit); + btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x1[19], &x1[28], _r, cos_bit); + btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x1[20], &x1[27], _r, cos_bit); + btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x1[21], &x1[26], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[32], &x1[39]); + btf_16_adds_subs_avx2(&x1[33], &x1[38]); + btf_16_adds_subs_avx2(&x1[34], &x1[37]); + btf_16_adds_subs_avx2(&x1[35], &x1[36]); + btf_16_adds_subs_avx2(&x1[47], &x1[40]); + btf_16_adds_subs_avx2(&x1[46], &x1[41]); + btf_16_adds_subs_avx2(&x1[45], &x1[42]); + btf_16_adds_subs_avx2(&x1[44], &x1[43]); + btf_16_adds_subs_avx2(&x1[48], &x1[55]); + btf_16_adds_subs_avx2(&x1[49], &x1[54]); + btf_16_adds_subs_avx2(&x1[50], &x1[53]); + btf_16_adds_subs_avx2(&x1[51], &x1[52]); + btf_16_adds_subs_avx2(&x1[63], &x1[56]); + btf_16_adds_subs_avx2(&x1[62], &x1[57]); + btf_16_adds_subs_avx2(&x1[61], &x1[58]); + btf_16_adds_subs_avx2(&x1[60], &x1[59]); + + // stage 6 + btf_16_w16_avx2(cospi_p32_p32, cospi_p32_m32, &x1[0], &x1[1], _r, cos_bit); + btf_16_w16_avx2(cospi_p48_p16, cospi_m16_p48, &x1[2], &x1[3], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[4], &x1[5]); + btf_16_adds_subs_avx2(&x1[7], &x1[6]); + btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x1[9], &x1[14], _r, cos_bit); + btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x1[10], &x1[13], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[16], &x1[19]); + btf_16_adds_subs_avx2(&x1[17], &x1[18]); + btf_16_adds_subs_avx2(&x1[23], &x1[20]); + btf_16_adds_subs_avx2(&x1[22], &x1[21]); + btf_16_adds_subs_avx2(&x1[24], &x1[27]); + btf_16_adds_subs_avx2(&x1[25], &x1[26]); + btf_16_adds_subs_avx2(&x1[31], &x1[28]); + btf_16_adds_subs_avx2(&x1[30], &x1[29]); + btf_16_w16_avx2(cospi_m08_p56, cospi_p56_p08, &x1[34], &x1[61], _r, cos_bit); + btf_16_w16_avx2(cospi_m08_p56, cospi_p56_p08, &x1[35], &x1[60], _r, cos_bit); + btf_16_w16_avx2(cospi_m56_m08, cospi_m08_p56, &x1[36], &x1[59], _r, cos_bit); + btf_16_w16_avx2(cospi_m56_m08, cospi_m08_p56, &x1[37], &x1[58], _r, cos_bit); + btf_16_w16_avx2(cospi_m40_p24, cospi_p24_p40, &x1[42], &x1[53], _r, cos_bit); + btf_16_w16_avx2(cospi_m40_p24, cospi_p24_p40, &x1[43], &x1[52], _r, cos_bit); + btf_16_w16_avx2(cospi_m24_m40, cospi_m40_p24, &x1[44], &x1[51], _r, cos_bit); + btf_16_w16_avx2(cospi_m24_m40, cospi_m40_p24, &x1[45], &x1[50], _r, cos_bit); + + // stage 7 + btf_16_w16_avx2(cospi_p56_p08, cospi_m08_p56, &x1[4], &x1[7], _r, cos_bit); + btf_16_w16_avx2(cospi_p24_p40, cospi_m40_p24, &x1[5], &x1[6], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[8], &x1[9]); + btf_16_adds_subs_avx2(&x1[11], &x1[10]); + btf_16_adds_subs_avx2(&x1[12], &x1[13]); + btf_16_adds_subs_avx2(&x1[15], &x1[14]); + btf_16_w16_avx2(cospi_m08_p56, cospi_p56_p08, &x1[17], &x1[30], _r, cos_bit); + btf_16_w16_avx2(cospi_m56_m08, cospi_m08_p56, &x1[18], &x1[29], _r, cos_bit); + btf_16_w16_avx2(cospi_m40_p24, cospi_p24_p40, &x1[21], &x1[26], _r, cos_bit); + btf_16_w16_avx2(cospi_m24_m40, cospi_m40_p24, &x1[22], &x1[25], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[32], &x1[35]); + btf_16_adds_subs_avx2(&x1[33], &x1[34]); + btf_16_adds_subs_avx2(&x1[39], &x1[36]); + btf_16_adds_subs_avx2(&x1[38], &x1[37]); + btf_16_adds_subs_avx2(&x1[40], &x1[43]); + btf_16_adds_subs_avx2(&x1[41], &x1[42]); + btf_16_adds_subs_avx2(&x1[47], &x1[44]); + btf_16_adds_subs_avx2(&x1[46], &x1[45]); + btf_16_adds_subs_avx2(&x1[48], &x1[51]); + btf_16_adds_subs_avx2(&x1[49], &x1[50]); + btf_16_adds_subs_avx2(&x1[55], &x1[52]); + btf_16_adds_subs_avx2(&x1[54], &x1[53]); + btf_16_adds_subs_avx2(&x1[56], &x1[59]); + btf_16_adds_subs_avx2(&x1[57], &x1[58]); + btf_16_adds_subs_avx2(&x1[63], &x1[60]); + btf_16_adds_subs_avx2(&x1[62], &x1[61]); + + // stage 8 + btf_16_w16_avx2(cospi_p60_p04, cospi_m04_p60, &x1[8], &x1[15], _r, cos_bit); + btf_16_w16_avx2(cospi_p28_p36, cospi_m36_p28, &x1[9], &x1[14], _r, cos_bit); + btf_16_w16_avx2(cospi_p44_p20, cospi_m20_p44, &x1[10], &x1[13], _r, cos_bit); + btf_16_w16_avx2(cospi_p12_p52, cospi_m52_p12, &x1[11], &x1[12], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[16], &x1[17]); + btf_16_adds_subs_avx2(&x1[19], &x1[18]); + btf_16_adds_subs_avx2(&x1[20], &x1[21]); + btf_16_adds_subs_avx2(&x1[23], &x1[22]); + btf_16_adds_subs_avx2(&x1[24], &x1[25]); + btf_16_adds_subs_avx2(&x1[27], &x1[26]); + btf_16_adds_subs_avx2(&x1[28], &x1[29]); + btf_16_adds_subs_avx2(&x1[31], &x1[30]); + btf_16_w16_avx2(cospi_m04_p60, cospi_p60_p04, &x1[33], &x1[62], _r, cos_bit); + btf_16_w16_avx2(cospi_m60_m04, cospi_m04_p60, &x1[34], &x1[61], _r, cos_bit); + btf_16_w16_avx2(cospi_m36_p28, cospi_p28_p36, &x1[37], &x1[58], _r, cos_bit); + btf_16_w16_avx2(cospi_m28_m36, cospi_m36_p28, &x1[38], &x1[57], _r, cos_bit); + btf_16_w16_avx2(cospi_m20_p44, cospi_p44_p20, &x1[41], &x1[54], _r, cos_bit); + btf_16_w16_avx2(cospi_m44_m20, cospi_m20_p44, &x1[42], &x1[53], _r, cos_bit); + btf_16_w16_avx2(cospi_m52_p12, cospi_p12_p52, &x1[45], &x1[50], _r, cos_bit); + btf_16_w16_avx2(cospi_m12_m52, cospi_m52_p12, &x1[46], &x1[49], _r, cos_bit); + + // stage 9 + btf_16_w16_avx2(cospi_p62_p02, cospi_m02_p62, &x1[16], &x1[31], _r, cos_bit); + btf_16_w16_avx2(cospi_p30_p34, cospi_m34_p30, &x1[17], &x1[30], _r, cos_bit); + btf_16_w16_avx2(cospi_p46_p18, cospi_m18_p46, &x1[18], &x1[29], _r, cos_bit); + btf_16_w16_avx2(cospi_p14_p50, cospi_m50_p14, &x1[19], &x1[28], _r, cos_bit); + btf_16_w16_avx2(cospi_p54_p10, cospi_m10_p54, &x1[20], &x1[27], _r, cos_bit); + btf_16_w16_avx2(cospi_p22_p42, cospi_m42_p22, &x1[21], &x1[26], _r, cos_bit); + btf_16_w16_avx2(cospi_p38_p26, cospi_m26_p38, &x1[22], &x1[25], _r, cos_bit); + btf_16_w16_avx2(cospi_p06_p58, cospi_m58_p06, &x1[23], &x1[24], _r, cos_bit); + btf_16_adds_subs_avx2(&x1[32], &x1[33]); + btf_16_adds_subs_avx2(&x1[35], &x1[34]); + btf_16_adds_subs_avx2(&x1[36], &x1[37]); + btf_16_adds_subs_avx2(&x1[39], &x1[38]); + btf_16_adds_subs_avx2(&x1[40], &x1[41]); + btf_16_adds_subs_avx2(&x1[43], &x1[42]); + btf_16_adds_subs_avx2(&x1[44], &x1[45]); + btf_16_adds_subs_avx2(&x1[47], &x1[46]); + btf_16_adds_subs_avx2(&x1[48], &x1[49]); + btf_16_adds_subs_avx2(&x1[51], &x1[50]); + btf_16_adds_subs_avx2(&x1[52], &x1[53]); + btf_16_adds_subs_avx2(&x1[55], &x1[54]); + btf_16_adds_subs_avx2(&x1[56], &x1[57]); + btf_16_adds_subs_avx2(&x1[59], &x1[58]); + btf_16_adds_subs_avx2(&x1[60], &x1[61]); + btf_16_adds_subs_avx2(&x1[63], &x1[62]); + + // stage 10 + btf_16_w16_avx2(cospi_p63_p01, cospi_m01_p63, &x1[32], &x1[63], _r, cos_bit); + btf_16_w16_avx2(cospi_p31_p33, cospi_m33_p31, &x1[33], &x1[62], _r, cos_bit); + btf_16_w16_avx2(cospi_p47_p17, cospi_m17_p47, &x1[34], &x1[61], _r, cos_bit); + btf_16_w16_avx2(cospi_p15_p49, cospi_m49_p15, &x1[35], &x1[60], _r, cos_bit); + btf_16_w16_avx2(cospi_p55_p09, cospi_m09_p55, &x1[36], &x1[59], _r, cos_bit); + btf_16_w16_avx2(cospi_p23_p41, cospi_m41_p23, &x1[37], &x1[58], _r, cos_bit); + btf_16_w16_avx2(cospi_p39_p25, cospi_m25_p39, &x1[38], &x1[57], _r, cos_bit); + btf_16_w16_avx2(cospi_p07_p57, cospi_m57_p07, &x1[39], &x1[56], _r, cos_bit); + btf_16_w16_avx2(cospi_p59_p05, cospi_m05_p59, &x1[40], &x1[55], _r, cos_bit); + btf_16_w16_avx2(cospi_p27_p37, cospi_m37_p27, &x1[41], &x1[54], _r, cos_bit); + btf_16_w16_avx2(cospi_p43_p21, cospi_m21_p43, &x1[42], &x1[53], _r, cos_bit); + btf_16_w16_avx2(cospi_p11_p53, cospi_m53_p11, &x1[43], &x1[52], _r, cos_bit); + btf_16_w16_avx2(cospi_p51_p13, cospi_m13_p51, &x1[44], &x1[51], _r, cos_bit); + btf_16_w16_avx2(cospi_p19_p45, cospi_m45_p19, &x1[45], &x1[50], _r, cos_bit); + btf_16_w16_avx2(cospi_p35_p29, cospi_m29_p35, &x1[46], &x1[49], _r, cos_bit); + btf_16_w16_avx2(cospi_p03_p61, cospi_m61_p03, &x1[47], &x1[48], _r, cos_bit); + + // stage 11 + output[0] = x1[0]; + output[1] = x1[32]; + output[2] = x1[16]; + output[3] = x1[48]; + output[4] = x1[8]; + output[5] = x1[40]; + output[6] = x1[24]; + output[7] = x1[56]; + output[8] = x1[4]; + output[9] = x1[36]; + output[10] = x1[20]; + output[11] = x1[52]; + output[12] = x1[12]; + output[13] = x1[44]; + output[14] = x1[28]; + output[15] = x1[60]; + output[16] = x1[2]; + output[17] = x1[34]; + output[18] = x1[18]; + output[19] = x1[50]; + output[20] = x1[10]; + output[21] = x1[42]; + output[22] = x1[26]; + output[23] = x1[58]; + output[24] = x1[6]; + output[25] = x1[38]; + output[26] = x1[22]; + output[27] = x1[54]; + output[28] = x1[14]; + output[29] = x1[46]; + output[30] = x1[30]; + output[31] = x1[62]; + output[32] = x1[1]; + output[33] = x1[33]; + output[34] = x1[17]; + output[35] = x1[49]; + output[36] = x1[9]; + output[37] = x1[41]; + output[38] = x1[25]; + output[39] = x1[57]; + output[40] = x1[5]; + output[41] = x1[37]; + output[42] = x1[21]; + output[43] = x1[53]; + output[44] = x1[13]; + output[45] = x1[45]; + output[46] = x1[29]; + output[47] = x1[61]; + output[48] = x1[3]; + output[49] = x1[35]; + output[50] = x1[19]; + output[51] = x1[51]; + output[52] = x1[11]; + output[53] = x1[43]; + output[54] = x1[27]; + output[55] = x1[59]; + output[56] = x1[7]; + output[57] = x1[39]; + output[58] = x1[23]; + output[59] = x1[55]; + output[60] = x1[15]; + output[61] = x1[47]; + output[62] = x1[31]; + output[63] = x1[63]; +} + +static INLINE void av1_fdct32_new_avx2(const __m256i *input, __m256i *output, + int8_t cos_bit) { + __m256i x1[32]; + const int32_t *cospi = cospi_arr(cos_bit); + const __m256i _r = _mm256_set1_epi32(1 << (cos_bit - 1)); + // stage 0 + // stage 1 + btf_32_add_sub_out_avx2(&x1[0], &x1[31], input[0], input[31]); + btf_32_add_sub_out_avx2(&x1[1], &x1[30], input[1], input[30]); + btf_32_add_sub_out_avx2(&x1[2], &x1[29], input[2], input[29]); + btf_32_add_sub_out_avx2(&x1[3], &x1[28], input[3], input[28]); + btf_32_add_sub_out_avx2(&x1[4], &x1[27], input[4], input[27]); + btf_32_add_sub_out_avx2(&x1[5], &x1[26], input[5], input[26]); + btf_32_add_sub_out_avx2(&x1[6], &x1[25], input[6], input[25]); + btf_32_add_sub_out_avx2(&x1[7], &x1[24], input[7], input[24]); + btf_32_add_sub_out_avx2(&x1[8], &x1[23], input[8], input[23]); + btf_32_add_sub_out_avx2(&x1[9], &x1[22], input[9], input[22]); + btf_32_add_sub_out_avx2(&x1[10], &x1[21], input[10], input[21]); + btf_32_add_sub_out_avx2(&x1[11], &x1[20], input[11], input[20]); + btf_32_add_sub_out_avx2(&x1[12], &x1[19], input[12], input[19]); + btf_32_add_sub_out_avx2(&x1[13], &x1[18], input[13], input[18]); + btf_32_add_sub_out_avx2(&x1[14], &x1[17], input[14], input[17]); + btf_32_add_sub_out_avx2(&x1[15], &x1[16], input[15], input[16]); + + // stage 2 + btf_32_add_sub_avx2(&x1[0], &x1[15]); + btf_32_add_sub_avx2(&x1[1], &x1[14]); + btf_32_add_sub_avx2(&x1[2], &x1[13]); + btf_32_add_sub_avx2(&x1[3], &x1[12]); + btf_32_add_sub_avx2(&x1[4], &x1[11]); + btf_32_add_sub_avx2(&x1[5], &x1[10]); + btf_32_add_sub_avx2(&x1[6], &x1[9]); + btf_32_add_sub_avx2(&x1[7], &x1[8]); + btf_32_avx2_type0(-cospi[32], cospi[32], &x1[20], &x1[27], _r, cos_bit); + btf_32_avx2_type0(-cospi[32], cospi[32], &x1[21], &x1[26], _r, cos_bit); + btf_32_avx2_type0(-cospi[32], cospi[32], &x1[22], &x1[25], _r, cos_bit); + btf_32_avx2_type0(-cospi[32], cospi[32], &x1[23], &x1[24], _r, cos_bit); + + // stage 3 + btf_32_add_sub_avx2(&x1[0], &x1[7]); + btf_32_add_sub_avx2(&x1[1], &x1[6]); + btf_32_add_sub_avx2(&x1[2], &x1[5]); + btf_32_add_sub_avx2(&x1[3], &x1[4]); + btf_32_avx2_type0(-cospi[32], cospi[32], &x1[10], &x1[13], _r, cos_bit); + btf_32_avx2_type0(-cospi[32], cospi[32], &x1[11], &x1[12], _r, cos_bit); + btf_32_add_sub_avx2(&x1[16], &x1[23]); + btf_32_add_sub_avx2(&x1[17], &x1[22]); + btf_32_add_sub_avx2(&x1[18], &x1[21]); + btf_32_add_sub_avx2(&x1[19], &x1[20]); + btf_32_add_sub_avx2(&x1[31], &x1[24]); + btf_32_add_sub_avx2(&x1[30], &x1[25]); + btf_32_add_sub_avx2(&x1[29], &x1[26]); + btf_32_add_sub_avx2(&x1[28], &x1[27]); + + // stage 4 + btf_32_add_sub_avx2(&x1[0], &x1[3]); + btf_32_add_sub_avx2(&x1[1], &x1[2]); + btf_32_avx2_type0(-cospi[32], cospi[32], &x1[5], &x1[6], _r, cos_bit); + btf_32_add_sub_avx2(&x1[8], &x1[11]); + btf_32_add_sub_avx2(&x1[9], &x1[10]); + btf_32_add_sub_avx2(&x1[15], &x1[12]); + btf_32_add_sub_avx2(&x1[14], &x1[13]); + btf_32_avx2_type0(-cospi[16], cospi[48], &x1[18], &x1[29], _r, cos_bit); + btf_32_avx2_type0(-cospi[16], cospi[48], &x1[19], &x1[28], _r, cos_bit); + btf_32_avx2_type0(-cospi[48], -cospi[16], &x1[20], &x1[27], _r, cos_bit); + btf_32_avx2_type0(-cospi[48], -cospi[16], &x1[21], &x1[26], _r, cos_bit); + + // stage 5 + btf_32_avx2_type0(cospi[32], cospi[32], &x1[0], &x1[1], _r, cos_bit); + btf_32_avx2_type1(cospi[48], cospi[16], &x1[2], &x1[3], _r, cos_bit); + btf_32_add_sub_avx2(&x1[4], &x1[5]); + btf_32_add_sub_avx2(&x1[7], &x1[6]); + btf_32_avx2_type0(-cospi[16], cospi[48], &x1[9], &x1[14], _r, cos_bit); + btf_32_avx2_type0(-cospi[48], -cospi[16], &x1[10], &x1[13], _r, cos_bit); + btf_32_add_sub_avx2(&x1[16], &x1[19]); + btf_32_add_sub_avx2(&x1[17], &x1[18]); + btf_32_add_sub_avx2(&x1[23], &x1[20]); + btf_32_add_sub_avx2(&x1[22], &x1[21]); + btf_32_add_sub_avx2(&x1[24], &x1[27]); + btf_32_add_sub_avx2(&x1[25], &x1[26]); + btf_32_add_sub_avx2(&x1[31], &x1[28]); + btf_32_add_sub_avx2(&x1[30], &x1[29]); + + // stage 6 + btf_32_avx2_type1(cospi[56], cospi[8], &x1[4], &x1[7], _r, cos_bit); + btf_32_avx2_type1(cospi[24], cospi[40], &x1[5], &x1[6], _r, cos_bit); + btf_32_add_sub_avx2(&x1[8], &x1[9]); + btf_32_add_sub_avx2(&x1[11], &x1[10]); + btf_32_add_sub_avx2(&x1[12], &x1[13]); + btf_32_add_sub_avx2(&x1[15], &x1[14]); + btf_32_avx2_type0(-cospi[8], cospi[56], &x1[17], &x1[30], _r, cos_bit); + btf_32_avx2_type0(-cospi[56], -cospi[8], &x1[18], &x1[29], _r, cos_bit); + btf_32_avx2_type0(-cospi[40], cospi[24], &x1[21], &x1[26], _r, cos_bit); + btf_32_avx2_type0(-cospi[24], -cospi[40], &x1[22], &x1[25], _r, cos_bit); + + // stage 7 + btf_32_avx2_type1(cospi[60], cospi[4], &x1[8], &x1[15], _r, cos_bit); + btf_32_avx2_type1(cospi[28], cospi[36], &x1[9], &x1[14], _r, cos_bit); + btf_32_avx2_type1(cospi[44], cospi[20], &x1[10], &x1[13], _r, cos_bit); + btf_32_avx2_type1(cospi[12], cospi[52], &x1[11], &x1[12], _r, cos_bit); + btf_32_add_sub_avx2(&x1[16], &x1[17]); + btf_32_add_sub_avx2(&x1[19], &x1[18]); + btf_32_add_sub_avx2(&x1[20], &x1[21]); + btf_32_add_sub_avx2(&x1[23], &x1[22]); + btf_32_add_sub_avx2(&x1[24], &x1[25]); + btf_32_add_sub_avx2(&x1[27], &x1[26]); + btf_32_add_sub_avx2(&x1[28], &x1[29]); + btf_32_add_sub_avx2(&x1[31], &x1[30]); + + // stage 8 + btf_32_avx2_type1(cospi[62], cospi[2], &x1[16], &x1[31], _r, cos_bit); + btf_32_avx2_type1(cospi[30], cospi[34], &x1[17], &x1[30], _r, cos_bit); + btf_32_avx2_type1(cospi[46], cospi[18], &x1[18], &x1[29], _r, cos_bit); + btf_32_avx2_type1(cospi[14], cospi[50], &x1[19], &x1[28], _r, cos_bit); + btf_32_avx2_type1(cospi[54], cospi[10], &x1[20], &x1[27], _r, cos_bit); + btf_32_avx2_type1(cospi[22], cospi[42], &x1[21], &x1[26], _r, cos_bit); + btf_32_avx2_type1(cospi[38], cospi[26], &x1[22], &x1[25], _r, cos_bit); + btf_32_avx2_type1(cospi[6], cospi[58], &x1[23], &x1[24], _r, cos_bit); + + // stage 9 + output[0] = x1[0]; + output[1] = x1[16]; + output[2] = x1[8]; + output[3] = x1[24]; + output[4] = x1[4]; + output[5] = x1[20]; + output[6] = x1[12]; + output[7] = x1[28]; + output[8] = x1[2]; + output[9] = x1[18]; + output[10] = x1[10]; + output[11] = x1[26]; + output[12] = x1[6]; + output[13] = x1[22]; + output[14] = x1[14]; + output[15] = x1[30]; + output[16] = x1[1]; + output[17] = x1[17]; + output[18] = x1[9]; + output[19] = x1[25]; + output[20] = x1[5]; + output[21] = x1[21]; + output[22] = x1[13]; + output[23] = x1[29]; + output[24] = x1[3]; + output[25] = x1[19]; + output[26] = x1[11]; + output[27] = x1[27]; + output[28] = x1[7]; + output[29] = x1[23]; + output[30] = x1[15]; + output[31] = x1[31]; +} + +static INLINE void av1_fdct64_new_avx2(const __m256i *input, __m256i *output, + int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m256i _r = _mm256_set1_epi32(1 << (cos_bit - 1)); + + __m256i cospi_m32 = _mm256_set1_epi32(-cospi[32]); + __m256i cospi_p32 = _mm256_set1_epi32(cospi[32]); + __m256i cospi_m16 = _mm256_set1_epi32(-cospi[16]); + __m256i cospi_p48 = _mm256_set1_epi32(cospi[48]); + __m256i cospi_m48 = _mm256_set1_epi32(-cospi[48]); + __m256i cospi_p16 = _mm256_set1_epi32(cospi[16]); + __m256i cospi_m08 = _mm256_set1_epi32(-cospi[8]); + __m256i cospi_p56 = _mm256_set1_epi32(cospi[56]); + __m256i cospi_m56 = _mm256_set1_epi32(-cospi[56]); + __m256i cospi_m40 = _mm256_set1_epi32(-cospi[40]); + __m256i cospi_p24 = _mm256_set1_epi32(cospi[24]); + __m256i cospi_m24 = _mm256_set1_epi32(-cospi[24]); + __m256i cospi_p08 = _mm256_set1_epi32(cospi[8]); + __m256i cospi_p40 = _mm256_set1_epi32(cospi[40]); + __m256i cospi_p60 = _mm256_set1_epi32(cospi[60]); + __m256i cospi_p04 = _mm256_set1_epi32(cospi[4]); + __m256i cospi_p28 = _mm256_set1_epi32(cospi[28]); + __m256i cospi_p36 = _mm256_set1_epi32(cospi[36]); + __m256i cospi_p44 = _mm256_set1_epi32(cospi[44]); + __m256i cospi_p20 = _mm256_set1_epi32(cospi[20]); + __m256i cospi_p12 = _mm256_set1_epi32(cospi[12]); + __m256i cospi_p52 = _mm256_set1_epi32(cospi[52]); + __m256i cospi_m04 = _mm256_set1_epi32(-cospi[4]); + __m256i cospi_m60 = _mm256_set1_epi32(-cospi[60]); + __m256i cospi_m36 = _mm256_set1_epi32(-cospi[36]); + __m256i cospi_m28 = _mm256_set1_epi32(-cospi[28]); + __m256i cospi_m20 = _mm256_set1_epi32(-cospi[20]); + __m256i cospi_m44 = _mm256_set1_epi32(-cospi[44]); + __m256i cospi_m52 = _mm256_set1_epi32(-cospi[52]); + __m256i cospi_m12 = _mm256_set1_epi32(-cospi[12]); + __m256i cospi_p62 = _mm256_set1_epi32(cospi[62]); + __m256i cospi_p02 = _mm256_set1_epi32(cospi[2]); + __m256i cospi_p30 = _mm256_set1_epi32(cospi[30]); + __m256i cospi_p34 = _mm256_set1_epi32(cospi[34]); + __m256i cospi_p46 = _mm256_set1_epi32(cospi[46]); + __m256i cospi_p18 = _mm256_set1_epi32(cospi[18]); + __m256i cospi_p14 = _mm256_set1_epi32(cospi[14]); + __m256i cospi_p50 = _mm256_set1_epi32(cospi[50]); + __m256i cospi_p54 = _mm256_set1_epi32(cospi[54]); + __m256i cospi_p10 = _mm256_set1_epi32(cospi[10]); + __m256i cospi_p22 = _mm256_set1_epi32(cospi[22]); + __m256i cospi_p42 = _mm256_set1_epi32(cospi[42]); + __m256i cospi_p38 = _mm256_set1_epi32(cospi[38]); + __m256i cospi_p26 = _mm256_set1_epi32(cospi[26]); + __m256i cospi_p06 = _mm256_set1_epi32(cospi[6]); + __m256i cospi_p58 = _mm256_set1_epi32(cospi[58]); + __m256i cospi_p63 = _mm256_set1_epi32(cospi[63]); + __m256i cospi_p01 = _mm256_set1_epi32(cospi[1]); + __m256i cospi_p31 = _mm256_set1_epi32(cospi[31]); + __m256i cospi_p33 = _mm256_set1_epi32(cospi[33]); + __m256i cospi_p47 = _mm256_set1_epi32(cospi[47]); + __m256i cospi_p17 = _mm256_set1_epi32(cospi[17]); + __m256i cospi_p15 = _mm256_set1_epi32(cospi[15]); + __m256i cospi_p49 = _mm256_set1_epi32(cospi[49]); + __m256i cospi_p55 = _mm256_set1_epi32(cospi[55]); + __m256i cospi_p09 = _mm256_set1_epi32(cospi[9]); + __m256i cospi_p23 = _mm256_set1_epi32(cospi[23]); + __m256i cospi_p41 = _mm256_set1_epi32(cospi[41]); + __m256i cospi_p39 = _mm256_set1_epi32(cospi[39]); + __m256i cospi_p25 = _mm256_set1_epi32(cospi[25]); + __m256i cospi_p07 = _mm256_set1_epi32(cospi[7]); + __m256i cospi_p57 = _mm256_set1_epi32(cospi[57]); + __m256i cospi_p59 = _mm256_set1_epi32(cospi[59]); + __m256i cospi_p05 = _mm256_set1_epi32(cospi[5]); + __m256i cospi_p27 = _mm256_set1_epi32(cospi[27]); + __m256i cospi_p37 = _mm256_set1_epi32(cospi[37]); + __m256i cospi_p43 = _mm256_set1_epi32(cospi[43]); + __m256i cospi_p21 = _mm256_set1_epi32(cospi[21]); + __m256i cospi_p11 = _mm256_set1_epi32(cospi[11]); + __m256i cospi_p53 = _mm256_set1_epi32(cospi[53]); + __m256i cospi_p51 = _mm256_set1_epi32(cospi[51]); + __m256i cospi_p13 = _mm256_set1_epi32(cospi[13]); + __m256i cospi_p19 = _mm256_set1_epi32(cospi[19]); + __m256i cospi_p45 = _mm256_set1_epi32(cospi[45]); + __m256i cospi_p35 = _mm256_set1_epi32(cospi[35]); + __m256i cospi_p29 = _mm256_set1_epi32(cospi[29]); + __m256i cospi_p03 = _mm256_set1_epi32(cospi[3]); + __m256i cospi_p61 = _mm256_set1_epi32(cospi[61]); + + // stage 1 + __m256i x1[64]; + btf_32_add_sub_out_avx2(&x1[0], &x1[63], input[0], input[63]); + btf_32_add_sub_out_avx2(&x1[1], &x1[62], input[1], input[62]); + btf_32_add_sub_out_avx2(&x1[2], &x1[61], input[2], input[61]); + btf_32_add_sub_out_avx2(&x1[3], &x1[60], input[3], input[60]); + btf_32_add_sub_out_avx2(&x1[4], &x1[59], input[4], input[59]); + btf_32_add_sub_out_avx2(&x1[5], &x1[58], input[5], input[58]); + btf_32_add_sub_out_avx2(&x1[6], &x1[57], input[6], input[57]); + btf_32_add_sub_out_avx2(&x1[7], &x1[56], input[7], input[56]); + btf_32_add_sub_out_avx2(&x1[8], &x1[55], input[8], input[55]); + btf_32_add_sub_out_avx2(&x1[9], &x1[54], input[9], input[54]); + btf_32_add_sub_out_avx2(&x1[10], &x1[53], input[10], input[53]); + btf_32_add_sub_out_avx2(&x1[11], &x1[52], input[11], input[52]); + btf_32_add_sub_out_avx2(&x1[12], &x1[51], input[12], input[51]); + btf_32_add_sub_out_avx2(&x1[13], &x1[50], input[13], input[50]); + btf_32_add_sub_out_avx2(&x1[14], &x1[49], input[14], input[49]); + btf_32_add_sub_out_avx2(&x1[15], &x1[48], input[15], input[48]); + btf_32_add_sub_out_avx2(&x1[16], &x1[47], input[16], input[47]); + btf_32_add_sub_out_avx2(&x1[17], &x1[46], input[17], input[46]); + btf_32_add_sub_out_avx2(&x1[18], &x1[45], input[18], input[45]); + btf_32_add_sub_out_avx2(&x1[19], &x1[44], input[19], input[44]); + btf_32_add_sub_out_avx2(&x1[20], &x1[43], input[20], input[43]); + btf_32_add_sub_out_avx2(&x1[21], &x1[42], input[21], input[42]); + btf_32_add_sub_out_avx2(&x1[22], &x1[41], input[22], input[41]); + btf_32_add_sub_out_avx2(&x1[23], &x1[40], input[23], input[40]); + btf_32_add_sub_out_avx2(&x1[24], &x1[39], input[24], input[39]); + btf_32_add_sub_out_avx2(&x1[25], &x1[38], input[25], input[38]); + btf_32_add_sub_out_avx2(&x1[26], &x1[37], input[26], input[37]); + btf_32_add_sub_out_avx2(&x1[27], &x1[36], input[27], input[36]); + btf_32_add_sub_out_avx2(&x1[28], &x1[35], input[28], input[35]); + btf_32_add_sub_out_avx2(&x1[29], &x1[34], input[29], input[34]); + btf_32_add_sub_out_avx2(&x1[30], &x1[33], input[30], input[33]); + btf_32_add_sub_out_avx2(&x1[31], &x1[32], input[31], input[32]); + + // stage 2 + btf_32_add_sub_avx2(&x1[0], &x1[31]); + btf_32_add_sub_avx2(&x1[1], &x1[30]); + btf_32_add_sub_avx2(&x1[2], &x1[29]); + btf_32_add_sub_avx2(&x1[3], &x1[28]); + btf_32_add_sub_avx2(&x1[4], &x1[27]); + btf_32_add_sub_avx2(&x1[5], &x1[26]); + btf_32_add_sub_avx2(&x1[6], &x1[25]); + btf_32_add_sub_avx2(&x1[7], &x1[24]); + btf_32_add_sub_avx2(&x1[8], &x1[23]); + btf_32_add_sub_avx2(&x1[9], &x1[22]); + btf_32_add_sub_avx2(&x1[10], &x1[21]); + btf_32_add_sub_avx2(&x1[11], &x1[20]); + btf_32_add_sub_avx2(&x1[12], &x1[19]); + btf_32_add_sub_avx2(&x1[13], &x1[18]); + btf_32_add_sub_avx2(&x1[14], &x1[17]); + btf_32_add_sub_avx2(&x1[15], &x1[16]); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[40], &x1[55], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[41], &x1[54], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[42], &x1[53], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[43], &x1[52], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[44], &x1[51], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[45], &x1[50], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[46], &x1[49], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[47], &x1[48], _r, cos_bit); + + // stage 3 + btf_32_add_sub_avx2(&x1[0], &x1[15]); + btf_32_add_sub_avx2(&x1[1], &x1[14]); + btf_32_add_sub_avx2(&x1[2], &x1[13]); + btf_32_add_sub_avx2(&x1[3], &x1[12]); + btf_32_add_sub_avx2(&x1[4], &x1[11]); + btf_32_add_sub_avx2(&x1[5], &x1[10]); + btf_32_add_sub_avx2(&x1[6], &x1[9]); + btf_32_add_sub_avx2(&x1[7], &x1[8]); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[20], &x1[27], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[21], &x1[26], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[22], &x1[25], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[23], &x1[24], _r, cos_bit); + btf_32_add_sub_avx2(&x1[32], &x1[47]); + btf_32_add_sub_avx2(&x1[33], &x1[46]); + btf_32_add_sub_avx2(&x1[34], &x1[45]); + btf_32_add_sub_avx2(&x1[35], &x1[44]); + btf_32_add_sub_avx2(&x1[36], &x1[43]); + btf_32_add_sub_avx2(&x1[37], &x1[42]); + btf_32_add_sub_avx2(&x1[38], &x1[41]); + btf_32_add_sub_avx2(&x1[39], &x1[40]); + btf_32_add_sub_avx2(&x1[63], &x1[48]); + btf_32_add_sub_avx2(&x1[62], &x1[49]); + btf_32_add_sub_avx2(&x1[61], &x1[50]); + btf_32_add_sub_avx2(&x1[60], &x1[51]); + btf_32_add_sub_avx2(&x1[59], &x1[52]); + btf_32_add_sub_avx2(&x1[58], &x1[53]); + btf_32_add_sub_avx2(&x1[57], &x1[54]); + btf_32_add_sub_avx2(&x1[56], &x1[55]); + + // stage 4 + btf_32_add_sub_avx2(&x1[0], &x1[7]); + btf_32_add_sub_avx2(&x1[1], &x1[6]); + btf_32_add_sub_avx2(&x1[2], &x1[5]); + btf_32_add_sub_avx2(&x1[3], &x1[4]); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[10], &x1[13], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[11], &x1[12], _r, cos_bit); + btf_32_add_sub_avx2(&x1[16], &x1[23]); + btf_32_add_sub_avx2(&x1[17], &x1[22]); + btf_32_add_sub_avx2(&x1[18], &x1[21]); + btf_32_add_sub_avx2(&x1[19], &x1[20]); + btf_32_add_sub_avx2(&x1[31], &x1[24]); + btf_32_add_sub_avx2(&x1[30], &x1[25]); + btf_32_add_sub_avx2(&x1[29], &x1[26]); + btf_32_add_sub_avx2(&x1[28], &x1[27]); + btf_32_avx2_type0_new(cospi_m16, cospi_p48, &x1[36], &x1[59], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m16, cospi_p48, &x1[37], &x1[58], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m16, cospi_p48, &x1[38], &x1[57], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m16, cospi_p48, &x1[39], &x1[56], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m48, cospi_m16, &x1[40], &x1[55], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m48, cospi_m16, &x1[41], &x1[54], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m48, cospi_m16, &x1[42], &x1[53], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m48, cospi_m16, &x1[43], &x1[52], _r, cos_bit); + + // stage 5 + btf_32_add_sub_avx2(&x1[0], &x1[3]); + btf_32_add_sub_avx2(&x1[1], &x1[2]); + btf_32_avx2_type0_new(cospi_m32, cospi_p32, &x1[5], &x1[6], _r, cos_bit); + btf_32_add_sub_avx2(&x1[8], &x1[11]); + btf_32_add_sub_avx2(&x1[9], &x1[10]); + btf_32_add_sub_avx2(&x1[15], &x1[12]); + btf_32_add_sub_avx2(&x1[14], &x1[13]); + btf_32_avx2_type0_new(cospi_m16, cospi_p48, &x1[18], &x1[29], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m16, cospi_p48, &x1[19], &x1[28], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m48, cospi_m16, &x1[20], &x1[27], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m48, cospi_m16, &x1[21], &x1[26], _r, cos_bit); + btf_32_add_sub_avx2(&x1[32], &x1[39]); + btf_32_add_sub_avx2(&x1[33], &x1[38]); + btf_32_add_sub_avx2(&x1[34], &x1[37]); + btf_32_add_sub_avx2(&x1[35], &x1[36]); + btf_32_add_sub_avx2(&x1[47], &x1[40]); + btf_32_add_sub_avx2(&x1[46], &x1[41]); + btf_32_add_sub_avx2(&x1[45], &x1[42]); + btf_32_add_sub_avx2(&x1[44], &x1[43]); + btf_32_add_sub_avx2(&x1[48], &x1[55]); + btf_32_add_sub_avx2(&x1[49], &x1[54]); + btf_32_add_sub_avx2(&x1[50], &x1[53]); + btf_32_add_sub_avx2(&x1[51], &x1[52]); + btf_32_add_sub_avx2(&x1[63], &x1[56]); + btf_32_add_sub_avx2(&x1[62], &x1[57]); + btf_32_add_sub_avx2(&x1[61], &x1[58]); + btf_32_add_sub_avx2(&x1[60], &x1[59]); + + // stage 6 + btf_32_avx2_type0_new(cospi_p32, cospi_p32, &x1[0], &x1[1], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p48, cospi_p16, &x1[2], &x1[3], _r, cos_bit); + btf_32_add_sub_avx2(&x1[4], &x1[5]); + btf_32_add_sub_avx2(&x1[7], &x1[6]); + btf_32_avx2_type0_new(cospi_m16, cospi_p48, &x1[9], &x1[14], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m48, cospi_m16, &x1[10], &x1[13], _r, cos_bit); + btf_32_add_sub_avx2(&x1[16], &x1[19]); + btf_32_add_sub_avx2(&x1[17], &x1[18]); + btf_32_add_sub_avx2(&x1[23], &x1[20]); + btf_32_add_sub_avx2(&x1[22], &x1[21]); + btf_32_add_sub_avx2(&x1[24], &x1[27]); + btf_32_add_sub_avx2(&x1[25], &x1[26]); + btf_32_add_sub_avx2(&x1[31], &x1[28]); + btf_32_add_sub_avx2(&x1[30], &x1[29]); + btf_32_avx2_type0_new(cospi_m08, cospi_p56, &x1[34], &x1[61], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m08, cospi_p56, &x1[35], &x1[60], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m56, cospi_m08, &x1[36], &x1[59], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m56, cospi_m08, &x1[37], &x1[58], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m40, cospi_p24, &x1[42], &x1[53], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m40, cospi_p24, &x1[43], &x1[52], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m24, cospi_m40, &x1[44], &x1[51], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m24, cospi_m40, &x1[45], &x1[50], _r, cos_bit); + + // stage 7 + btf_32_avx2_type1_new(cospi_p56, cospi_p08, &x1[4], &x1[7], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p24, cospi_p40, &x1[5], &x1[6], _r, cos_bit); + btf_32_add_sub_avx2(&x1[8], &x1[9]); + btf_32_add_sub_avx2(&x1[11], &x1[10]); + btf_32_add_sub_avx2(&x1[12], &x1[13]); + btf_32_add_sub_avx2(&x1[15], &x1[14]); + btf_32_avx2_type0_new(cospi_m08, cospi_p56, &x1[17], &x1[30], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m56, cospi_m08, &x1[18], &x1[29], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m40, cospi_p24, &x1[21], &x1[26], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m24, cospi_m40, &x1[22], &x1[25], _r, cos_bit); + btf_32_add_sub_avx2(&x1[32], &x1[35]); + btf_32_add_sub_avx2(&x1[33], &x1[34]); + btf_32_add_sub_avx2(&x1[39], &x1[36]); + btf_32_add_sub_avx2(&x1[38], &x1[37]); + btf_32_add_sub_avx2(&x1[40], &x1[43]); + btf_32_add_sub_avx2(&x1[41], &x1[42]); + btf_32_add_sub_avx2(&x1[47], &x1[44]); + btf_32_add_sub_avx2(&x1[46], &x1[45]); + btf_32_add_sub_avx2(&x1[48], &x1[51]); + btf_32_add_sub_avx2(&x1[49], &x1[50]); + btf_32_add_sub_avx2(&x1[55], &x1[52]); + btf_32_add_sub_avx2(&x1[54], &x1[53]); + btf_32_add_sub_avx2(&x1[56], &x1[59]); + btf_32_add_sub_avx2(&x1[57], &x1[58]); + btf_32_add_sub_avx2(&x1[63], &x1[60]); + btf_32_add_sub_avx2(&x1[62], &x1[61]); + + // stage 8 + btf_32_avx2_type1_new(cospi_p60, cospi_p04, &x1[8], &x1[15], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p28, cospi_p36, &x1[9], &x1[14], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p44, cospi_p20, &x1[10], &x1[13], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p12, cospi_p52, &x1[11], &x1[12], _r, cos_bit); + btf_32_add_sub_avx2(&x1[16], &x1[17]); + btf_32_add_sub_avx2(&x1[19], &x1[18]); + btf_32_add_sub_avx2(&x1[20], &x1[21]); + btf_32_add_sub_avx2(&x1[23], &x1[22]); + btf_32_add_sub_avx2(&x1[24], &x1[25]); + btf_32_add_sub_avx2(&x1[27], &x1[26]); + btf_32_add_sub_avx2(&x1[28], &x1[29]); + btf_32_add_sub_avx2(&x1[31], &x1[30]); + btf_32_avx2_type0_new(cospi_m04, cospi_p60, &x1[33], &x1[62], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m60, cospi_m04, &x1[34], &x1[61], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m36, cospi_p28, &x1[37], &x1[58], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m28, cospi_m36, &x1[38], &x1[57], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m20, cospi_p44, &x1[41], &x1[54], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m44, cospi_m20, &x1[42], &x1[53], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m52, cospi_p12, &x1[45], &x1[50], _r, cos_bit); + btf_32_avx2_type0_new(cospi_m12, cospi_m52, &x1[46], &x1[49], _r, cos_bit); + + // stage 9 + btf_32_avx2_type1_new(cospi_p62, cospi_p02, &x1[16], &x1[31], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p30, cospi_p34, &x1[17], &x1[30], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p46, cospi_p18, &x1[18], &x1[29], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p14, cospi_p50, &x1[19], &x1[28], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p54, cospi_p10, &x1[20], &x1[27], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p22, cospi_p42, &x1[21], &x1[26], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p38, cospi_p26, &x1[22], &x1[25], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p06, cospi_p58, &x1[23], &x1[24], _r, cos_bit); + btf_32_add_sub_avx2(&x1[32], &x1[33]); + btf_32_add_sub_avx2(&x1[35], &x1[34]); + btf_32_add_sub_avx2(&x1[36], &x1[37]); + btf_32_add_sub_avx2(&x1[39], &x1[38]); + btf_32_add_sub_avx2(&x1[40], &x1[41]); + btf_32_add_sub_avx2(&x1[43], &x1[42]); + btf_32_add_sub_avx2(&x1[44], &x1[45]); + btf_32_add_sub_avx2(&x1[47], &x1[46]); + btf_32_add_sub_avx2(&x1[48], &x1[49]); + btf_32_add_sub_avx2(&x1[51], &x1[50]); + btf_32_add_sub_avx2(&x1[52], &x1[53]); + btf_32_add_sub_avx2(&x1[55], &x1[54]); + btf_32_add_sub_avx2(&x1[56], &x1[57]); + btf_32_add_sub_avx2(&x1[59], &x1[58]); + btf_32_add_sub_avx2(&x1[60], &x1[61]); + btf_32_add_sub_avx2(&x1[63], &x1[62]); + + // stage 10 + btf_32_avx2_type1_new(cospi_p63, cospi_p01, &x1[32], &x1[63], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p31, cospi_p33, &x1[33], &x1[62], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p47, cospi_p17, &x1[34], &x1[61], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p15, cospi_p49, &x1[35], &x1[60], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p55, cospi_p09, &x1[36], &x1[59], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p23, cospi_p41, &x1[37], &x1[58], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p39, cospi_p25, &x1[38], &x1[57], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p07, cospi_p57, &x1[39], &x1[56], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p59, cospi_p05, &x1[40], &x1[55], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p27, cospi_p37, &x1[41], &x1[54], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p43, cospi_p21, &x1[42], &x1[53], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p11, cospi_p53, &x1[43], &x1[52], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p51, cospi_p13, &x1[44], &x1[51], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p19, cospi_p45, &x1[45], &x1[50], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p35, cospi_p29, &x1[46], &x1[49], _r, cos_bit); + btf_32_avx2_type1_new(cospi_p03, cospi_p61, &x1[47], &x1[48], _r, cos_bit); + + // stage 11 + output[0] = x1[0]; + output[1] = x1[32]; + output[2] = x1[16]; + output[3] = x1[48]; + output[4] = x1[8]; + output[5] = x1[40]; + output[6] = x1[24]; + output[7] = x1[56]; + output[8] = x1[4]; + output[9] = x1[36]; + output[10] = x1[20]; + output[11] = x1[52]; + output[12] = x1[12]; + output[13] = x1[44]; + output[14] = x1[28]; + output[15] = x1[60]; + output[16] = x1[2]; + output[17] = x1[34]; + output[18] = x1[18]; + output[19] = x1[50]; + output[20] = x1[10]; + output[21] = x1[42]; + output[22] = x1[26]; + output[23] = x1[58]; + output[24] = x1[6]; + output[25] = x1[38]; + output[26] = x1[22]; + output[27] = x1[54]; + output[28] = x1[14]; + output[29] = x1[46]; + output[30] = x1[30]; + output[31] = x1[62]; + output[32] = x1[1]; + output[33] = x1[33]; + output[34] = x1[17]; + output[35] = x1[49]; + output[36] = x1[9]; + output[37] = x1[41]; + output[38] = x1[25]; + output[39] = x1[57]; + output[40] = x1[5]; + output[41] = x1[37]; + output[42] = x1[21]; + output[43] = x1[53]; + output[44] = x1[13]; + output[45] = x1[45]; + output[46] = x1[29]; + output[47] = x1[61]; + output[48] = x1[3]; + output[49] = x1[35]; + output[50] = x1[19]; + output[51] = x1[51]; + output[52] = x1[11]; + output[53] = x1[43]; + output[54] = x1[27]; + output[55] = x1[59]; + output[56] = x1[7]; + output[57] = x1[39]; + output[58] = x1[23]; + output[59] = x1[55]; + output[60] = x1[15]; + output[61] = x1[47]; + output[62] = x1[31]; + output[63] = x1[63]; +} + +static INLINE void fadst16x16_new_avx2(const __m256i *input, __m256i *output, + int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m256i __zero = _mm256_setzero_si256(); + const __m256i _r = _mm256_set1_epi32(1 << (cos_bit - 1)); + + __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]); + __m256i cospi_p32_m32 = pair_set_w16_epi16(cospi[32], -cospi[32]); + __m256i cospi_p16_p48 = pair_set_w16_epi16(cospi[16], cospi[48]); + __m256i cospi_p48_m16 = pair_set_w16_epi16(cospi[48], -cospi[16]); + __m256i cospi_m48_p16 = pair_set_w16_epi16(-cospi[48], cospi[16]); + __m256i cospi_p08_p56 = pair_set_w16_epi16(cospi[8], cospi[56]); + __m256i cospi_p56_m08 = pair_set_w16_epi16(cospi[56], -cospi[8]); + __m256i cospi_p40_p24 = pair_set_w16_epi16(cospi[40], cospi[24]); + __m256i cospi_p24_m40 = pair_set_w16_epi16(cospi[24], -cospi[40]); + __m256i cospi_m56_p08 = pair_set_w16_epi16(-cospi[56], cospi[8]); + __m256i cospi_m24_p40 = pair_set_w16_epi16(-cospi[24], cospi[40]); + __m256i cospi_p02_p62 = pair_set_w16_epi16(cospi[2], cospi[62]); + __m256i cospi_p62_m02 = pair_set_w16_epi16(cospi[62], -cospi[2]); + __m256i cospi_p10_p54 = pair_set_w16_epi16(cospi[10], cospi[54]); + __m256i cospi_p54_m10 = pair_set_w16_epi16(cospi[54], -cospi[10]); + __m256i cospi_p18_p46 = pair_set_w16_epi16(cospi[18], cospi[46]); + __m256i cospi_p46_m18 = pair_set_w16_epi16(cospi[46], -cospi[18]); + __m256i cospi_p26_p38 = pair_set_w16_epi16(cospi[26], cospi[38]); + __m256i cospi_p38_m26 = pair_set_w16_epi16(cospi[38], -cospi[26]); + __m256i cospi_p34_p30 = pair_set_w16_epi16(cospi[34], cospi[30]); + __m256i cospi_p30_m34 = pair_set_w16_epi16(cospi[30], -cospi[34]); + __m256i cospi_p42_p22 = pair_set_w16_epi16(cospi[42], cospi[22]); + __m256i cospi_p22_m42 = pair_set_w16_epi16(cospi[22], -cospi[42]); + __m256i cospi_p50_p14 = pair_set_w16_epi16(cospi[50], cospi[14]); + __m256i cospi_p14_m50 = pair_set_w16_epi16(cospi[14], -cospi[50]); + __m256i cospi_p58_p06 = pair_set_w16_epi16(cospi[58], cospi[6]); + __m256i cospi_p06_m58 = pair_set_w16_epi16(cospi[6], -cospi[58]); + + // stage 1 + __m256i x1[16]; + x1[0] = input[0]; + x1[1] = _mm256_subs_epi16(__zero, input[15]); + x1[2] = _mm256_subs_epi16(__zero, input[7]); + x1[3] = input[8]; + x1[4] = _mm256_subs_epi16(__zero, input[3]); + x1[5] = input[12]; + x1[6] = input[4]; + x1[7] = _mm256_subs_epi16(__zero, input[11]); + x1[8] = _mm256_subs_epi16(__zero, input[1]); + x1[9] = input[14]; + x1[10] = input[6]; + x1[11] = _mm256_subs_epi16(__zero, input[9]); + x1[12] = input[2]; + x1[13] = _mm256_subs_epi16(__zero, input[13]); + x1[14] = _mm256_subs_epi16(__zero, input[5]); + x1[15] = input[10]; + + // stage 2 + btf_16_w16_avx2(cospi_p32_p32, cospi_p32_m32, &x1[2], &x1[3], _r, cos_bit); + btf_16_w16_avx2(cospi_p32_p32, cospi_p32_m32, &x1[6], &x1[7], _r, cos_bit); + btf_16_w16_avx2(cospi_p32_p32, cospi_p32_m32, &x1[10], &x1[11], _r, cos_bit); + btf_16_w16_avx2(cospi_p32_p32, cospi_p32_m32, &x1[14], &x1[15], _r, cos_bit); + + // stage 3 + btf_16_adds_subs_avx2(&x1[0], &x1[2]); + btf_16_adds_subs_avx2(&x1[1], &x1[3]); + btf_16_adds_subs_avx2(&x1[4], &x1[6]); + btf_16_adds_subs_avx2(&x1[5], &x1[7]); + btf_16_adds_subs_avx2(&x1[8], &x1[10]); + btf_16_adds_subs_avx2(&x1[9], &x1[11]); + btf_16_adds_subs_avx2(&x1[12], &x1[14]); + btf_16_adds_subs_avx2(&x1[13], &x1[15]); + + // stage 4 + btf_16_w16_avx2(cospi_p16_p48, cospi_p48_m16, &x1[4], &x1[5], _r, cos_bit); + btf_16_w16_avx2(cospi_m48_p16, cospi_p16_p48, &x1[6], &x1[7], _r, cos_bit); + btf_16_w16_avx2(cospi_p16_p48, cospi_p48_m16, &x1[12], &x1[13], _r, cos_bit); + btf_16_w16_avx2(cospi_m48_p16, cospi_p16_p48, &x1[14], &x1[15], _r, cos_bit); + + // stage 5 + btf_16_adds_subs_avx2(&x1[0], &x1[4]); + btf_16_adds_subs_avx2(&x1[1], &x1[5]); + btf_16_adds_subs_avx2(&x1[2], &x1[6]); + btf_16_adds_subs_avx2(&x1[3], &x1[7]); + btf_16_adds_subs_avx2(&x1[8], &x1[12]); + btf_16_adds_subs_avx2(&x1[9], &x1[13]); + btf_16_adds_subs_avx2(&x1[10], &x1[14]); + btf_16_adds_subs_avx2(&x1[11], &x1[15]); + + // stage 6 + btf_16_w16_avx2(cospi_p08_p56, cospi_p56_m08, &x1[8], &x1[9], _r, cos_bit); + btf_16_w16_avx2(cospi_p40_p24, cospi_p24_m40, &x1[10], &x1[11], _r, cos_bit); + btf_16_w16_avx2(cospi_m56_p08, cospi_p08_p56, &x1[12], &x1[13], _r, cos_bit); + btf_16_w16_avx2(cospi_m24_p40, cospi_p40_p24, &x1[14], &x1[15], _r, cos_bit); + + // stage 7 + btf_16_adds_subs_avx2(&x1[0], &x1[8]); + btf_16_adds_subs_avx2(&x1[1], &x1[9]); + btf_16_adds_subs_avx2(&x1[2], &x1[10]); + btf_16_adds_subs_avx2(&x1[3], &x1[11]); + btf_16_adds_subs_avx2(&x1[4], &x1[12]); + btf_16_adds_subs_avx2(&x1[5], &x1[13]); + btf_16_adds_subs_avx2(&x1[6], &x1[14]); + btf_16_adds_subs_avx2(&x1[7], &x1[15]); + + // stage 8 + btf_16_w16_avx2(cospi_p02_p62, cospi_p62_m02, &x1[0], &x1[1], _r, cos_bit); + btf_16_w16_avx2(cospi_p10_p54, cospi_p54_m10, &x1[2], &x1[3], _r, cos_bit); + btf_16_w16_avx2(cospi_p18_p46, cospi_p46_m18, &x1[4], &x1[5], _r, cos_bit); + btf_16_w16_avx2(cospi_p26_p38, cospi_p38_m26, &x1[6], &x1[7], _r, cos_bit); + btf_16_w16_avx2(cospi_p34_p30, cospi_p30_m34, &x1[8], &x1[9], _r, cos_bit); + btf_16_w16_avx2(cospi_p42_p22, cospi_p22_m42, &x1[10], &x1[11], _r, cos_bit); + btf_16_w16_avx2(cospi_p50_p14, cospi_p14_m50, &x1[12], &x1[13], _r, cos_bit); + btf_16_w16_avx2(cospi_p58_p06, cospi_p06_m58, &x1[14], &x1[15], _r, cos_bit); + + // stage 9 + output[0] = x1[1]; + output[1] = x1[14]; + output[2] = x1[3]; + output[3] = x1[12]; + output[4] = x1[5]; + output[5] = x1[10]; + output[6] = x1[7]; + output[7] = x1[8]; + output[8] = x1[9]; + output[9] = x1[6]; + output[10] = x1[11]; + output[11] = x1[4]; + output[12] = x1[13]; + output[13] = x1[2]; + output[14] = x1[15]; + output[15] = x1[0]; +} + +static INLINE __m256i scale_round_avx2(const __m256i a, const int scale) { + const __m256i scale__r = pair_set_w16_epi16(scale, 1 << (NewSqrt2Bits - 1)); + const __m256i b = _mm256_madd_epi16(a, scale__r); + return _mm256_srai_epi32(b, NewSqrt2Bits); +} + +static INLINE void fidentity16x16_new_avx2(const __m256i *input, + __m256i *output, int8_t cos_bit) { + (void)cos_bit; + const __m256i one = _mm256_set1_epi16(1); + + for (int i = 0; i < 16; ++i) { + const __m256i a_lo = _mm256_unpacklo_epi16(input[i], one); + const __m256i a_hi = _mm256_unpackhi_epi16(input[i], one); + const __m256i b_lo = scale_round_avx2(a_lo, 2 * NewSqrt2); + const __m256i b_hi = scale_round_avx2(a_hi, 2 * NewSqrt2); + output[i] = _mm256_packs_epi32(b_lo, b_hi); + } +} + +static INLINE void fidentity16x32_new_avx2(const __m256i *input, + __m256i *output, int8_t cos_bit) { + (void)cos_bit; + for (int i = 0; i < 32; ++i) { + output[i] = _mm256_slli_epi16(input[i], 2); + } +} + +static INLINE void av1_round_shift_array_32_avx2(__m256i *input, + __m256i *output, + const int size, + const int bit) { + if (bit > 0) { + int i; + for (i = 0; i < size; i++) { + output[i] = av1_round_shift_32_avx2(input[i], bit); + } + } else { + int i; + for (i = 0; i < size; i++) { + output[i] = _mm256_slli_epi32(input[i], -bit); + } + } +} + +static INLINE void av1_round_shift_rect_array_32_avx2(__m256i *input, + __m256i *output, + const int size, + const int bit) { + const __m256i sqrt2 = _mm256_set1_epi32(NewSqrt2); + if (bit > 0) { + int i; + for (i = 0; i < size; i++) { + const __m256i r0 = av1_round_shift_32_avx2(input[i], bit); + const __m256i r1 = _mm256_mullo_epi32(sqrt2, r0); + output[i] = av1_round_shift_32_avx2(r1, NewSqrt2Bits); + } + } else { + int i; + for (i = 0; i < size; i++) { + const __m256i r0 = _mm256_slli_epi32(input[i], -bit); + const __m256i r1 = _mm256_mullo_epi32(sqrt2, r0); + output[i] = av1_round_shift_32_avx2(r1, NewSqrt2Bits); + } + } +} + +static INLINE void transpose_32_8x8_avx2(int stride, const __m256i *inputA, + __m256i *output) { + __m256i temp0 = _mm256_unpacklo_epi32(inputA[0], inputA[2]); + __m256i temp1 = _mm256_unpackhi_epi32(inputA[0], inputA[2]); + __m256i temp2 = _mm256_unpacklo_epi32(inputA[1], inputA[3]); + __m256i temp3 = _mm256_unpackhi_epi32(inputA[1], inputA[3]); + __m256i temp4 = _mm256_unpacklo_epi32(inputA[4], inputA[6]); + __m256i temp5 = _mm256_unpackhi_epi32(inputA[4], inputA[6]); + __m256i temp6 = _mm256_unpacklo_epi32(inputA[5], inputA[7]); + __m256i temp7 = _mm256_unpackhi_epi32(inputA[5], inputA[7]); + + __m256i t0 = _mm256_unpacklo_epi32(temp0, temp2); + __m256i t1 = _mm256_unpackhi_epi32(temp0, temp2); + __m256i t2 = _mm256_unpacklo_epi32(temp1, temp3); + __m256i t3 = _mm256_unpackhi_epi32(temp1, temp3); + __m256i t4 = _mm256_unpacklo_epi32(temp4, temp6); + __m256i t5 = _mm256_unpackhi_epi32(temp4, temp6); + __m256i t6 = _mm256_unpacklo_epi32(temp5, temp7); + __m256i t7 = _mm256_unpackhi_epi32(temp5, temp7); + + output[0 * stride] = _mm256_permute2x128_si256(t0, t4, 0x20); + output[1 * stride] = _mm256_permute2x128_si256(t1, t5, 0x20); + output[2 * stride] = _mm256_permute2x128_si256(t2, t6, 0x20); + output[3 * stride] = _mm256_permute2x128_si256(t3, t7, 0x20); + output[4 * stride] = _mm256_permute2x128_si256(t0, t4, 0x31); + output[5 * stride] = _mm256_permute2x128_si256(t1, t5, 0x31); + output[6 * stride] = _mm256_permute2x128_si256(t2, t6, 0x31); + output[7 * stride] = _mm256_permute2x128_si256(t3, t7, 0x31); +} + +// Store 8 16 bit values. Sign extend the values. +static INLINE void store_buffer_16bit_to_32bit_w16_avx2(const __m256i *const in, + int32_t *out, + const int stride, + const int out_size) { + for (int i = 0; i < out_size; ++i) { + _mm256_store_si256((__m256i *)(out), + _mm256_cvtepi16_epi32(_mm256_castsi256_si128(in[i]))); + _mm256_store_si256( + (__m256i *)(out + 8), + _mm256_cvtepi16_epi32(_mm256_extracti128_si256(in[i], 1))); + out += stride; + } +} + +static INLINE void store_rect_16bit_to_32bit_avx2(const __m256i a, + int32_t *const b) { + const __m256i one = _mm256_set1_epi16(1); + const __m256i a_reoder = _mm256_permute4x64_epi64(a, 0xd8); + const __m256i a_lo = _mm256_unpacklo_epi16(a_reoder, one); + const __m256i a_hi = _mm256_unpackhi_epi16(a_reoder, one); + const __m256i b_lo = scale_round_avx2(a_lo, NewSqrt2); + const __m256i b_hi = scale_round_avx2(a_hi, NewSqrt2); + _mm256_store_si256((__m256i *)b, b_lo); + _mm256_store_si256((__m256i *)(b + 8), b_hi); +} + +static INLINE void store_rect_buffer_16bit_to_32bit_w16_avx2( + const __m256i *const in, int32_t *const out, const int stride, + const int out_size) { + for (int i = 0; i < out_size; ++i) { + store_rect_16bit_to_32bit_avx2(in[i], out + i * stride); + } +} + +static const transform_1d_avx2 col_txfm16x32_arr[TX_TYPES] = { + fdct16x32_new_avx2, // DCT_DCT + NULL, // ADST_DCT + NULL, // DCT_ADST + NULL, // ADST_ADST + NULL, // FLIPADST_DCT + NULL, // DCT_FLIPADST + NULL, // FLIPADST_FLIPADST + NULL, // ADST_FLIPADST + NULL, // FLIPADST_ADST + fidentity16x32_new_avx2, // IDTX + fdct16x32_new_avx2, // V_DCT + fidentity16x32_new_avx2, // H_DCT + NULL, // V_ADST + NULL, // H_ADST + NULL, // V_FLIPADST + NULL // H_FLIPADST +}; + +static const transform_1d_avx2 row_txfm16x32_arr[TX_TYPES] = { + fdct16x32_new_avx2, // DCT_DCT + NULL, // ADST_DCT + NULL, // DCT_ADST + NULL, // ADST_ADST + NULL, // FLIPADST_DCT + NULL, // DCT_FLIPADST + NULL, // FLIPADST_FLIPADST + NULL, // ADST_FLIPADST + NULL, // FLIPADST_ADST + fidentity16x32_new_avx2, // IDTX + fidentity16x32_new_avx2, // V_DCT + fdct16x32_new_avx2, // H_DCT + NULL, // V_ADST + NULL, // H_ADST + NULL, // V_FLIPADST + NULL // H_FLIPADST +}; + +static const transform_1d_avx2 col_txfm16x16_arr[TX_TYPES] = { + fdct16x16_new_avx2, // DCT_DCT + fadst16x16_new_avx2, // ADST_DCT + fdct16x16_new_avx2, // DCT_ADST + fadst16x16_new_avx2, // ADST_ADST + fadst16x16_new_avx2, // FLIPADST_DCT + fdct16x16_new_avx2, // DCT_FLIPADST + fadst16x16_new_avx2, // FLIPADST_FLIPADST + fadst16x16_new_avx2, // ADST_FLIPADST + fadst16x16_new_avx2, // FLIPADST_ADST + fidentity16x16_new_avx2, // IDTX + fdct16x16_new_avx2, // V_DCT + fidentity16x16_new_avx2, // H_DCT + fadst16x16_new_avx2, // V_ADST + fidentity16x16_new_avx2, // H_ADST + fadst16x16_new_avx2, // V_FLIPADST + fidentity16x16_new_avx2 // H_FLIPADST +}; + +static const transform_1d_avx2 row_txfm16x16_arr[TX_TYPES] = { + fdct16x16_new_avx2, // DCT_DCT + fdct16x16_new_avx2, // ADST_DCT + fadst16x16_new_avx2, // DCT_ADST + fadst16x16_new_avx2, // ADST_ADST + fdct16x16_new_avx2, // FLIPADST_DCT + fadst16x16_new_avx2, // DCT_FLIPADST + fadst16x16_new_avx2, // FLIPADST_FLIPADST + fadst16x16_new_avx2, // ADST_FLIPADST + fadst16x16_new_avx2, // FLIPADST_ADST + fidentity16x16_new_avx2, // IDTX + fidentity16x16_new_avx2, // V_DCT + fdct16x16_new_avx2, // H_DCT + fidentity16x16_new_avx2, // V_ADST + fadst16x16_new_avx2, // H_ADST + fidentity16x16_new_avx2, // V_FLIPADST + fadst16x16_new_avx2 // H_FLIPADST +}; + +static void lowbd_fwd_txfm2d_16x16_avx2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + const TX_SIZE tx_size = TX_16X16; + __m256i buf0[16], buf1[16]; + const int8_t *shift = fwd_txfm_shift_ls[tx_size]; + const int txw_idx = get_txw_idx(tx_size); + const int txh_idx = get_txh_idx(tx_size); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = tx_size_wide[tx_size]; + const int height = tx_size_high[tx_size]; + const transform_1d_avx2 col_txfm = col_txfm16x16_arr[tx_type]; + const transform_1d_avx2 row_txfm = row_txfm16x16_arr[tx_type]; + int ud_flip, lr_flip; + + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + const int32_t i = 0; + if (ud_flip) { + load_buffer_16bit_to_16bit_flip_avx2(input + 16 * i, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit_avx2(input + 16 * i, stride, buf0, height); + } + round_shift_16bit_w16_avx2(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit_w16_avx2(buf0, height, shift[1]); + transpose_16bit_16x16_avx2(buf0, buf1 + 0 * width + 16 * i); + + __m256i *buf; + if (lr_flip) { + buf = buf0; + flip_buf_avx2(buf1 + width * i, buf, width); + } else { + buf = buf1 + width * i; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit_w16_avx2(buf, width, shift[2]); + transpose_16bit_16x16_avx2(buf, buf); + store_buffer_16bit_to_32bit_w16_avx2(buf, output + 16 * width * i, width, 16); +} + +static void lowbd_fwd_txfm2d_32x32_avx2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + const TX_SIZE tx_size = TX_32X32; + __m256i buf0[32], buf1[128]; + const int8_t *shift = fwd_txfm_shift_ls[tx_size]; + const int txw_idx = get_txw_idx(tx_size); + const int txh_idx = get_txh_idx(tx_size); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = tx_size_wide[tx_size]; + const int height = tx_size_high[tx_size]; + const transform_1d_avx2 col_txfm = col_txfm16x32_arr[tx_type]; + const transform_1d_avx2 row_txfm = row_txfm16x32_arr[tx_type]; + + int ud_flip, lr_flip; + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + + for (int i = 0; i < 2; i++) { + if (ud_flip) { + load_buffer_16bit_to_16bit_flip_avx2(input + 16 * i, stride, buf0, + height); + } else { + load_buffer_16bit_to_16bit_avx2(input + 16 * i, stride, buf0, height); + } + round_shift_16bit_w16_avx2(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit_w16_avx2(buf0, height, shift[1]); + transpose_16bit_16x16_avx2(buf0 + 0 * 16, buf1 + 0 * width + 16 * i); + transpose_16bit_16x16_avx2(buf0 + 1 * 16, buf1 + 1 * width + 16 * i); + } + + for (int i = 0; i < 2; i++) { + __m256i *buf; + if (lr_flip) { + buf = buf0; + flip_buf_avx2(buf1 + width * i, buf, width); + } else { + buf = buf1 + width * i; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit_w16_avx2(buf, width, shift[2]); + transpose_16bit_16x16_avx2(buf, buf); + store_buffer_16bit_to_32bit_w16_avx2(buf, output + 16 * width * i, width, + 16); + transpose_16bit_16x16_avx2(buf + 16, buf + 16); + store_buffer_16bit_to_32bit_w16_avx2(buf + 16, output + 16 * width * i + 16, + width, 16); + } +} + +static void lowbd_fwd_txfm2d_64x64_avx2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + (void)tx_type; + assert(tx_type == DCT_DCT); + const TX_SIZE tx_size = TX_64X64; + __m256i buf0[64], buf1[256]; + const int8_t *shift = fwd_txfm_shift_ls[tx_size]; + const int txw_idx = get_txw_idx(tx_size); + const int txh_idx = get_txh_idx(tx_size); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = tx_size_wide[tx_size]; + const int height = tx_size_high[tx_size]; + const transform_1d_avx2 col_txfm = fdct16x64_new_avx2; + const int width_div16 = (width >> 4); + const int height_div16 = (height >> 4); + + for (int i = 0; i < width_div16; i++) { + load_buffer_16bit_to_16bit_avx2(input + 16 * i, stride, buf0, height); + round_shift_16bit_w16_avx2(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit_w16_avx2(buf0, height, shift[1]); + for (int j = 0; j < AOMMIN(2, height_div16); ++j) { + transpose_16bit_16x16_avx2(buf0 + j * 16, buf1 + j * width + 16 * i); + } + } + + for (int i = 0; i < AOMMIN(2, height_div16); i++) { + __m256i bufA[64]; + __m256i bufB[64]; + __m128i *buf = (__m128i *)(buf1 + width * i); + for (int j = 0; j < width; ++j) { + bufA[j] = _mm256_cvtepi16_epi32(buf[j * 2]); + bufB[j] = _mm256_cvtepi16_epi32(buf[j * 2 + 1]); + } + av1_fdct64_new_avx2(bufA, bufA, cos_bit_row); + av1_fdct64_new_avx2(bufB, bufB, cos_bit_row); + av1_round_shift_array_32_avx2(bufA, bufA, 32, -shift[2]); + av1_round_shift_array_32_avx2(bufB, bufB, 32, -shift[2]); + + int32_t *output8 = output + 16 * 32 * i; + for (int j = 0; j < 4; ++j) { + __m256i *out = (__m256i *)(output8 + 8 * j); + transpose_32_8x8_avx2(4, bufA + 8 * j, out); + transpose_32_8x8_avx2(4, bufB + 8 * j, out + 8 * 4); + } + } +} + +static void lowbd_fwd_txfm2d_16x32_avx2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + const TX_SIZE tx_size = TX_16X32; + __m256i buf0[32], buf1[32]; + const int8_t *shift = fwd_txfm_shift_ls[tx_size]; + const int txw_idx = get_txw_idx(tx_size); + const int txh_idx = get_txh_idx(tx_size); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = tx_size_wide[tx_size]; + const int height = tx_size_high[tx_size]; + const transform_1d_avx2 col_txfm = col_txfm16x32_arr[tx_type]; + const transform_1d_avx2 row_txfm = row_txfm16x16_arr[tx_type]; + + int ud_flip, lr_flip; + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + + if (ud_flip) { + load_buffer_16bit_to_16bit_flip_avx2(input, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit_avx2(input, stride, buf0, height); + } + round_shift_16bit_w16_avx2(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit_w16_avx2(buf0, height, shift[1]); + transpose_16bit_16x16_avx2(buf0, buf1); + transpose_16bit_16x16_avx2(buf0 + 16, buf1 + 16); + + for (int i = 0; i < 2; i++) { + __m256i *buf; + if (lr_flip) { + buf = buf0; + flip_buf_avx2(buf1 + width * i, buf, width); + } else { + buf = buf1 + width * i; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit_w16_avx2(buf, width, shift[2]); + transpose_16bit_16x16_avx2(buf, buf); + store_rect_buffer_16bit_to_32bit_w16_avx2(buf, output + 16 * width * i, + width, 16); + } +} + +static void lowbd_fwd_txfm2d_32x16_avx2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m256i buf0[32], buf1[64]; + const int8_t *shift = fwd_txfm_shift_ls[TX_32X16]; + const int txw_idx = get_txw_idx(TX_32X16); + const int txh_idx = get_txh_idx(TX_32X16); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 32; + const int height = 16; + const transform_1d_avx2 col_txfm = col_txfm16x16_arr[tx_type]; + const transform_1d_avx2 row_txfm = row_txfm16x32_arr[tx_type]; + + int ud_flip, lr_flip; + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + + for (int i = 0; i < 2; i++) { + if (ud_flip) { + load_buffer_16bit_to_16bit_flip_avx2(input + 16 * i, stride, buf0, + height); + } else { + load_buffer_16bit_to_16bit_avx2(input + 16 * i, stride, buf0, height); + } + round_shift_16bit_w16_avx2(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit_w16_avx2(buf0, height, shift[1]); + transpose_16bit_16x16_avx2(buf0, buf1 + 0 * width + 16 * i); + } + + __m256i *buf; + if (lr_flip) { + buf = buf0; + flip_buf_avx2(buf1, buf, width); + } else { + buf = buf1; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit_w16_avx2(buf, width, shift[2]); + transpose_16bit_16x16_avx2(buf, buf); + store_rect_buffer_16bit_to_32bit_w16_avx2(buf, output, width, 16); + + transpose_16bit_16x16_avx2(buf + 16, buf + 16); + store_rect_buffer_16bit_to_32bit_w16_avx2(buf + 16, output + 16, width, 16); +} + +static void lowbd_fwd_txfm2d_64x32_avx2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + const TX_SIZE tx_size = TX_64X32; + __m256i buf0[64], buf1[256]; + const int8_t *shift = fwd_txfm_shift_ls[tx_size]; + const int txw_idx = get_txw_idx(tx_size); + const int txh_idx = get_txh_idx(tx_size); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = tx_size_wide[tx_size]; + const int height = tx_size_high[tx_size]; + const transform_1d_avx2 col_txfm = col_txfm16x32_arr[tx_type]; + const int width_div16 = (width >> 4); + const int height_div16 = (height >> 4); + + for (int i = 0; i < width_div16; i++) { + load_buffer_16bit_to_16bit_avx2(input + 16 * i, stride, buf0, height); + round_shift_16bit_w16_avx2(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit_w16_avx2(buf0, height, shift[1]); + for (int j = 0; j < AOMMIN(4, height_div16); ++j) { + transpose_16bit_16x16_avx2(buf0 + j * 16, buf1 + j * width + 16 * i); + } + } + assert(tx_type == DCT_DCT); + for (int i = 0; i < AOMMIN(2, height_div16); i++) { + __m256i bufA[64]; + __m256i bufB[64]; + __m128i *buf = (__m128i *)(buf1 + width * i); + for (int j = 0; j < width; ++j) { + bufA[j] = _mm256_cvtepi16_epi32(buf[j * 2]); + bufB[j] = _mm256_cvtepi16_epi32(buf[j * 2 + 1]); + } + av1_fdct64_new_avx2(bufA, bufA, cos_bit_row); + av1_fdct64_new_avx2(bufB, bufB, cos_bit_row); + av1_round_shift_rect_array_32_avx2(bufA, bufA, 32, -shift[2]); + av1_round_shift_rect_array_32_avx2(bufB, bufB, 32, -shift[2]); + + int32_t *output8 = output + 16 * 32 * i; + for (int j = 0; j < 4; ++j) { + __m256i *out = (__m256i *)(output8 + 8 * j); + transpose_32_8x8_avx2(4, bufA + 8 * j, out); + transpose_32_8x8_avx2(4, bufB + 8 * j, out + 8 * 4); + } + } +} + +static void lowbd_fwd_txfm2d_32x64_avx2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + (void)tx_type; + assert(tx_type == DCT_DCT); + const TX_SIZE tx_size = TX_32X64; + __m256i buf0[64], buf1[256]; + const int8_t *shift = fwd_txfm_shift_ls[tx_size]; + const int txw_idx = get_txw_idx(tx_size); + const int txh_idx = get_txh_idx(tx_size); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = tx_size_wide[tx_size]; + const int height = tx_size_high[tx_size]; + const transform_1d_avx2 col_txfm = fdct16x64_new_avx2; + const int width_div16 = (width >> 4); + const int height_div16 = (height >> 4); + + for (int i = 0; i < width_div16; i++) { + load_buffer_16bit_to_16bit_avx2(input + 16 * i, stride, buf0, height); + round_shift_16bit_w16_avx2(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit_w16_avx2(buf0, height, shift[1]); + for (int j = 0; j < AOMMIN(2, height_div16); ++j) { + transpose_16bit_16x16_avx2(buf0 + j * 16, buf1 + j * width + 16 * i); + } + } + + for (int i = 0; i < AOMMIN(2, height_div16); i++) { + __m256i bufA[32]; + __m256i bufB[32]; + __m128i *buf = (__m128i *)(buf1 + width * i); + for (int j = 0; j < width; ++j) { + bufA[j] = _mm256_cvtepi16_epi32(buf[j * 2]); + bufB[j] = _mm256_cvtepi16_epi32(buf[j * 2 + 1]); + } + av1_fdct32_new_avx2(bufA, bufA, cos_bit_row); + av1_fdct32_new_avx2(bufB, bufB, cos_bit_row); + av1_round_shift_rect_array_32_avx2(bufA, bufA, 32, -shift[2]); + av1_round_shift_rect_array_32_avx2(bufB, bufB, 32, -shift[2]); + + int32_t *output8 = output + 16 * 32 * i; + for (int j = 0; j < 4; ++j) { + __m256i *out = (__m256i *)(output8 + 8 * j); + transpose_32_8x8_avx2(4, bufA + 8 * j, out); + transpose_32_8x8_avx2(4, bufB + 8 * j, out + 8 * 4); + } + } +} + +static void lowbd_fwd_txfm2d_16x64_avx2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + (void)tx_type; + assert(tx_type == DCT_DCT); + const TX_SIZE tx_size = TX_16X64; + __m256i buf0[64], buf1[64]; + const int8_t *shift = fwd_txfm_shift_ls[tx_size]; + const int txw_idx = get_txw_idx(tx_size); + const int txh_idx = get_txh_idx(tx_size); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = tx_size_wide[tx_size]; + const int height = tx_size_high[tx_size]; + const transform_1d_avx2 col_txfm = fdct16x64_new_avx2; + const transform_1d_avx2 row_txfm = fdct16x16_new_avx2; + const int width_div16 = (width >> 4); + const int height_div16 = (height >> 4); + + for (int i = 0; i < width_div16; i++) { + load_buffer_16bit_to_16bit_avx2(input + 16 * i, stride, buf0, height); + round_shift_16bit_w16_avx2(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit_w16_avx2(buf0, height, shift[1]); + for (int j = 0; j < height_div16; ++j) { + transpose_16bit_16x16_avx2(buf0 + j * 16, buf1 + j * width + 16 * i); + } + } + + for (int i = 0; i < AOMMIN(4, height_div16); i++) { + __m256i *buf = buf1 + width * i; + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit_w16_avx2(buf, width, shift[2]); + int32_t *output16 = output + 16 * width * i; + for (int j = 0; j < width_div16; ++j) { + __m256i *buf16 = buf + 16 * j; + transpose_16bit_16x16_avx2(buf16, buf16); + store_buffer_16bit_to_32bit_w16_avx2(buf16, output16 + 16 * j, width, 16); + } + } + // Zero out the bottom 16x32 area. + memset(output + 16 * 32, 0, 16 * 32 * sizeof(*output)); +} + +static void lowbd_fwd_txfm2d_64x16_avx2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + (void)tx_type; + assert(tx_type == DCT_DCT); + const TX_SIZE tx_size = TX_64X16; + __m256i buf0[64], buf1[64]; + const int8_t *shift = fwd_txfm_shift_ls[tx_size]; + const int txw_idx = get_txw_idx(tx_size); + const int txh_idx = get_txh_idx(tx_size); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = tx_size_wide[tx_size]; + const int height = tx_size_high[tx_size]; + const transform_1d_avx2 col_txfm = fdct16x16_new_avx2; + const transform_1d_avx2 row_txfm = fdct16x64_new_avx2; + const int width_div16 = (width >> 4); + const int height_div16 = (height >> 4); + + for (int i = 0; i < width_div16; i++) { + load_buffer_16bit_to_16bit_avx2(input + 16 * i, stride, buf0, height); + round_shift_16bit_w16_avx2(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit_w16_avx2(buf0, height, shift[1]); + for (int j = 0; j < height_div16; ++j) { + transpose_16bit_16x16_avx2(buf0 + j * 16, buf1 + j * width + 16 * i); + } + } + + for (int i = 0; i < height_div16; i++) { + __m256i *buf = buf1 + width * i; + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit_w16_avx2(buf, width, shift[2]); + int32_t *output16 = output + 16 * 32 * i; + for (int j = 0; j < 2; ++j) { + __m256i *buf16 = buf + 16 * j; + transpose_16bit_16x16_avx2(buf16, buf16); + store_buffer_16bit_to_32bit_w16_avx2(buf16, output16 + 16 * j, 32, 16); + } + } +} + +static FwdTxfm2dFunc fwd_txfm2d_func_ls[TX_SIZES_ALL] = { + av1_lowbd_fwd_txfm2d_4x4_sse2, // 4x4 transform + av1_lowbd_fwd_txfm2d_8x8_sse2, // 8x8 transform + lowbd_fwd_txfm2d_16x16_avx2, // 16x16 transform + lowbd_fwd_txfm2d_32x32_avx2, // 32x32 transform + lowbd_fwd_txfm2d_64x64_avx2, // 64x64 transform + av1_lowbd_fwd_txfm2d_4x8_sse2, // 4x8 transform + av1_lowbd_fwd_txfm2d_8x4_sse2, // 8x4 transform + av1_lowbd_fwd_txfm2d_8x16_sse2, // 8x16 transform + av1_lowbd_fwd_txfm2d_16x8_sse2, // 16x8 transform + lowbd_fwd_txfm2d_16x32_avx2, // 16x32 transform + lowbd_fwd_txfm2d_32x16_avx2, // 32x16 transform + lowbd_fwd_txfm2d_32x64_avx2, // 32x64 transform + lowbd_fwd_txfm2d_64x32_avx2, // 64x32 transform + av1_lowbd_fwd_txfm2d_4x16_sse2, // 4x16 transform + av1_lowbd_fwd_txfm2d_16x4_sse2, // 16x4 transform + av1_lowbd_fwd_txfm2d_8x32_sse2, // 8x32 transform + av1_lowbd_fwd_txfm2d_32x8_sse2, // 32x8 transform + lowbd_fwd_txfm2d_16x64_avx2, // 16x64 transform + lowbd_fwd_txfm2d_64x16_avx2, // 64x16 transform +}; + +void av1_lowbd_fwd_txfm_avx2(const int16_t *src_diff, tran_low_t *coeff, + int diff_stride, TxfmParam *txfm_param) { + FwdTxfm2dFunc fwd_txfm2d_func = fwd_txfm2d_func_ls[txfm_param->tx_size]; + if ((fwd_txfm2d_func == NULL) || + (txfm_param->lossless && txfm_param->tx_size == TX_4X4)) { + av1_lowbd_fwd_txfm_c(src_diff, coeff, diff_stride, txfm_param); + } else { + fwd_txfm2d_func(src_diff, coeff, diff_stride, txfm_param->tx_type, + txfm_param->bd); + } +} diff --git a/third_party/aom/av1/encoder/x86/av1_fwd_txfm2d_sse4.c b/third_party/aom/av1/encoder/x86/av1_fwd_txfm2d_sse4.c new file mode 100644 index 0000000000..8ec0256eb8 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/av1_fwd_txfm2d_sse4.c @@ -0,0 +1,365 @@ +/* + * Copyright (c) 2016, 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 "config/av1_rtcd.h" + +#include "av1/common/enums.h" +#include "av1/common/av1_txfm.h" +#include "av1/common/x86/av1_txfm_sse2.h" +#include "av1/common/x86/highbd_txfm_utility_sse4.h" +#include "av1/encoder/av1_fwd_txfm1d_cfg.h" +#include "av1/encoder/x86/av1_txfm1d_sse4.h" +#include "av1/encoder/x86/av1_fwd_txfm_sse2.h" + +static INLINE void int16_array_with_stride_to_int32_array_without_stride( + const int16_t *input, int stride, int32_t *output, int txfm1d_size) { + int r, c; + for (r = 0; r < txfm1d_size; r++) { + for (c = 0; c < txfm1d_size; c++) { + output[r * txfm1d_size + c] = (int32_t)input[r * stride + c]; + } + } +} + +typedef void (*TxfmFuncSSE2)(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); + +static void fdct32_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range) { + const int txfm_size = 32; + const int num_per_128 = 4; + __m128i buf0[32]; + __m128i buf1[32]; + int col_num = txfm_size / num_per_128; + int col; + (void)stage_range; + for (col = 0; col < col_num; col++) { + int j; + for (j = 0; j < 32; ++j) { + buf0[j] = input[j * col_num + col]; + } + av1_fdct32_new_sse4_1(buf0, buf1, cos_bit); + for (j = 0; j < 32; ++j) { + output[j * col_num + col] = buf1[j]; + } + } +} + +static void fdct64_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range) { + const int txfm_size = 64; + const int num_per_128 = 4; + int col_num = txfm_size / num_per_128; + (void)stage_range; + for (int col = 0; col < col_num; col++) { + av1_fdct64_new_sse4_1((input + col), (output + col), cos_bit, col_num, + col_num); + } +} + +static INLINE TxfmFuncSSE2 fwd_txfm_type_to_func(TXFM_TYPE txfm_type) { + switch (txfm_type) { + case TXFM_TYPE_DCT32: return fdct32_new_sse4_1; break; + case TXFM_TYPE_DCT64: return fdct64_new_sse4_1; break; + default: assert(0); + } + return NULL; +} + +static INLINE void fwd_txfm2d_sse4_1(const int16_t *input, int32_t *output, + const int stride, + const TXFM_2D_FLIP_CFG *cfg, + int32_t *txfm_buf) { + // TODO(sarahparker) This does not currently support rectangular transforms + // and will break without splitting txfm_size out into row and col size. + // Rectangular transforms use c code only, so it should be ok for now. + // It will be corrected when there are sse implementations for rectangular + // transforms. + assert(cfg->tx_size < TX_SIZES); + const int txfm_size = tx_size_wide[cfg->tx_size]; + const int8_t *shift = cfg->shift; + const int8_t *stage_range_col = cfg->stage_range_col; + const int8_t *stage_range_row = cfg->stage_range_row; + const int8_t cos_bit_col = cfg->cos_bit_col; + const int8_t cos_bit_row = cfg->cos_bit_row; + const TxfmFuncSSE2 txfm_func_col = fwd_txfm_type_to_func(cfg->txfm_type_col); + const TxfmFuncSSE2 txfm_func_row = fwd_txfm_type_to_func(cfg->txfm_type_row); + + __m128i *buf_128 = (__m128i *)txfm_buf; + __m128i *out_128 = (__m128i *)output; + int num_per_128 = 4; + int txfm2d_size_128 = txfm_size * txfm_size / num_per_128; + + int16_array_with_stride_to_int32_array_without_stride(input, stride, txfm_buf, + txfm_size); + av1_round_shift_array_32_sse4_1(buf_128, out_128, txfm2d_size_128, -shift[0]); + txfm_func_col(out_128, buf_128, cos_bit_col, stage_range_col); + av1_round_shift_array_32_sse4_1(buf_128, out_128, txfm2d_size_128, -shift[1]); + transpose_32(txfm_size, out_128, buf_128); + txfm_func_row(buf_128, out_128, cos_bit_row, stage_range_row); + av1_round_shift_array_32_sse4_1(out_128, buf_128, txfm2d_size_128, -shift[2]); + transpose_32(txfm_size, buf_128, out_128); +} + +static INLINE void fwd_txfm2d_64x64_sse4_1(const int16_t *input, + int32_t *output, const int stride, + const TXFM_2D_FLIP_CFG *cfg, + int32_t *txfm_buf) { + assert(cfg->tx_size < TX_SIZES); + const int txfm_size = tx_size_wide[cfg->tx_size]; + const int8_t *shift = cfg->shift; + const int8_t *stage_range_col = cfg->stage_range_col; + const int8_t cos_bit_col = cfg->cos_bit_col; + const int8_t cos_bit_row = cfg->cos_bit_row; + const TxfmFuncSSE2 txfm_func_col = fwd_txfm_type_to_func(cfg->txfm_type_col); + __m128i *buf_128 = (__m128i *)txfm_buf; + __m128i *out_128 = (__m128i *)output; + + const int num_per_128 = 4; + int txfm2d_size_128 = txfm_size * txfm_size / num_per_128; + int col_num = txfm_size / num_per_128; + + int16_array_with_stride_to_int32_array_without_stride(input, stride, output, + txfm_size); + /*col wise transform*/ + txfm_func_col(out_128, buf_128, cos_bit_col, stage_range_col); + av1_round_shift_array_32_sse4_1(buf_128, out_128, txfm2d_size_128, -shift[1]); + transpose_32(txfm_size, out_128, buf_128); + + /*row wise transform*/ + for (int col = 0; col < (col_num >> 1); col++) { + av1_fdct64_new_sse4_1((buf_128 + col), (out_128 + col), cos_bit_row, + col_num, (col_num >> 1)); + } + + txfm2d_size_128 = (col_num >> 1) * (txfm_size >> 1); + av1_round_shift_array_32_sse4_1(out_128, buf_128, txfm2d_size_128, -shift[2]); + transpose_32x32(buf_128, out_128); +} + +void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + DECLARE_ALIGNED(16, int32_t, txfm_buf[1024]); + TXFM_2D_FLIP_CFG cfg; + av1_get_fwd_txfm_cfg(tx_type, TX_32X32, &cfg); + (void)bd; + fwd_txfm2d_sse4_1(input, output, stride, &cfg, txfm_buf); +} + +void av1_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + DECLARE_ALIGNED(16, int32_t, txfm_buf[4096]); + TXFM_2D_FLIP_CFG cfg; + av1_get_fwd_txfm_cfg(tx_type, TX_64X64, &cfg); + (void)bd; + fwd_txfm2d_64x64_sse4_1(input, output, stride, &cfg, txfm_buf); +} + +static INLINE void transpose_32_4x4x2(int stride, const __m128i *inputA, + const __m128i *inputB, __m128i *output) { + __m128i temp0 = _mm_unpacklo_epi32(inputA[0], inputA[2]); + __m128i temp1 = _mm_unpackhi_epi32(inputA[0], inputA[2]); + __m128i temp2 = _mm_unpacklo_epi32(inputA[1], inputA[3]); + __m128i temp3 = _mm_unpackhi_epi32(inputA[1], inputA[3]); + + output[0 * stride] = _mm_unpacklo_epi32(temp0, temp2); + output[1 * stride] = _mm_unpackhi_epi32(temp0, temp2); + output[2 * stride] = _mm_unpacklo_epi32(temp1, temp3); + output[3 * stride] = _mm_unpackhi_epi32(temp1, temp3); + + temp0 = _mm_unpacklo_epi32(inputB[0], inputB[2]); + temp1 = _mm_unpackhi_epi32(inputB[0], inputB[2]); + temp2 = _mm_unpacklo_epi32(inputB[1], inputB[3]); + temp3 = _mm_unpackhi_epi32(inputB[1], inputB[3]); + + output[4 * stride] = _mm_unpacklo_epi32(temp0, temp2); + output[5 * stride] = _mm_unpackhi_epi32(temp0, temp2); + output[6 * stride] = _mm_unpacklo_epi32(temp1, temp3); + output[7 * stride] = _mm_unpackhi_epi32(temp1, temp3); +} + +static void lowbd_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + (void)tx_type; + assert(tx_type == DCT_DCT); + const TX_SIZE tx_size = TX_64X64; + __m128i buf0[64], buf1[512]; + const int8_t *shift = fwd_txfm_shift_ls[tx_size]; + const int txw_idx = get_txw_idx(tx_size); + const int txh_idx = get_txh_idx(tx_size); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = tx_size_wide[tx_size]; + const int height = tx_size_high[tx_size]; + const transform_1d_sse2 col_txfm = fdct8x64_new_sse2; + const int width_div8 = (width >> 3); + const int height_div8 = (height >> 3); + + for (int i = 0; i < width_div8; i++) { + load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height); + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + for (int j = 0; j < AOMMIN(4, height_div8); ++j) { + transpose_16bit_8x8(buf0 + j * 8, buf1 + j * width + 8 * i); + } + } + for (int i = 0; i < AOMMIN(4, height_div8); i++) { + __m128i bufA[64]; + __m128i bufB[64]; + __m128i *buf = buf1 + width * i; + for (int j = 0; j < width; ++j) { + bufA[j] = _mm_cvtepi16_epi32(buf[j]); + bufB[j] = _mm_cvtepi16_epi32(_mm_unpackhi_epi64(buf[j], buf[j])); + } + av1_fdct64_new_sse4_1(bufA, bufA, cos_bit_row, 1, 1); + av1_fdct64_new_sse4_1(bufB, bufB, cos_bit_row, 1, 1); + av1_round_shift_array_32_sse4_1(bufA, bufA, 32, -shift[2]); + av1_round_shift_array_32_sse4_1(bufB, bufB, 32, -shift[2]); + + int32_t *output8 = output + 8 * 32 * i; + for (int j = 0; j < width_div8; ++j) { + __m128i *out = (__m128i *)(output8 + 4 * j); + transpose_32_4x4x2(8, bufA + 4 * j, bufB + 4 * j, out); + } + } +} + +static void lowbd_fwd_txfm2d_64x32_sse4_1(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + const TX_SIZE tx_size = TX_64X32; + __m128i buf0[64], buf1[256]; + const int8_t *shift = fwd_txfm_shift_ls[tx_size]; + const int txw_idx = get_txw_idx(tx_size); + const int txh_idx = get_txh_idx(tx_size); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = tx_size_wide[tx_size]; + const int height = tx_size_high[tx_size]; + const transform_1d_sse2 col_txfm = col_txfm8x32_arr[tx_type]; + const int width_div8 = (width >> 3); + const int height_div8 = (height >> 3); + + for (int i = 0; i < width_div8; i++) { + load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height); + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + for (int j = 0; j < AOMMIN(4, height_div8); ++j) { + transpose_16bit_8x8(buf0 + j * 8, buf1 + j * width + 8 * i); + } + } + assert(tx_type == DCT_DCT); + for (int i = 0; i < AOMMIN(4, height_div8); i++) { + __m128i bufA[64]; + __m128i bufB[64]; + __m128i *buf = buf1 + width * i; + for (int j = 0; j < width; ++j) { + bufA[j] = _mm_cvtepi16_epi32(buf[j]); + bufB[j] = _mm_cvtepi16_epi32(_mm_unpackhi_epi64(buf[j], buf[j])); + } + av1_fdct64_new_sse4_1(bufA, bufA, cos_bit_row, 1, 1); + av1_fdct64_new_sse4_1(bufB, bufB, cos_bit_row, 1, 1); + av1_round_shift_rect_array_32_sse4_1(bufA, bufA, 32, -shift[2], NewSqrt2); + av1_round_shift_rect_array_32_sse4_1(bufB, bufB, 32, -shift[2], NewSqrt2); + + int32_t *output8 = output + 8 * 32 * i; + for (int j = 0; j < width_div8; ++j) { + __m128i *out = (__m128i *)(output8 + 4 * j); + transpose_32_4x4x2(8, bufA + 4 * j, bufB + 4 * j, out); + } + } +} + +static void lowbd_fwd_txfm2d_32x64_sse4_1(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + (void)tx_type; + assert(tx_type == DCT_DCT); + const TX_SIZE tx_size = TX_32X64; + __m128i buf0[64], buf1[256]; + const int8_t *shift = fwd_txfm_shift_ls[tx_size]; + const int txw_idx = get_txw_idx(tx_size); + const int txh_idx = get_txh_idx(tx_size); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = tx_size_wide[tx_size]; + const int height = tx_size_high[tx_size]; + const transform_1d_sse2 col_txfm = fdct8x64_new_sse2; + const int width_div8 = (width >> 3); + const int height_div8 = (height >> 3); + + for (int i = 0; i < width_div8; i++) { + load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height); + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + for (int j = 0; j < AOMMIN(4, height_div8); ++j) { + transpose_16bit_8x8(buf0 + j * 8, buf1 + j * width + 8 * i); + } + } + + for (int i = 0; i < AOMMIN(4, height_div8); i++) { + __m128i bufA[32]; + __m128i bufB[32]; + __m128i *buf = buf1 + width * i; + for (int j = 0; j < width; ++j) { + bufA[j] = _mm_cvtepi16_epi32(buf[j]); + bufB[j] = _mm_cvtepi16_epi32(_mm_unpackhi_epi64(buf[j], buf[j])); + } + av1_fdct32_new_sse4_1(bufA, bufA, cos_bit_row); + av1_fdct32_new_sse4_1(bufB, bufB, cos_bit_row); + av1_round_shift_rect_array_32_sse4_1(bufA, bufA, 32, -shift[2], NewSqrt2); + av1_round_shift_rect_array_32_sse4_1(bufB, bufB, 32, -shift[2], NewSqrt2); + + int32_t *output8 = output + 8 * 32 * i; + for (int j = 0; j < (32 / 4); ++j) { + __m128i *out = (__m128i *)(output8 + 4 * j); + transpose_32_4x4x2(8, bufA + 4 * j, bufB + 4 * j, out); + } + } +} + +static FwdTxfm2dFunc fwd_txfm2d_func_ls[TX_SIZES_ALL] = { + av1_lowbd_fwd_txfm2d_4x4_sse2, // 4x4 transform + av1_lowbd_fwd_txfm2d_8x8_sse2, // 8x8 transform + av1_lowbd_fwd_txfm2d_16x16_sse2, // 16x16 transform + av1_lowbd_fwd_txfm2d_32x32_sse2, // 32x32 transform + lowbd_fwd_txfm2d_64x64_sse4_1, // 64x64 transform + av1_lowbd_fwd_txfm2d_4x8_sse2, // 4x8 transform + av1_lowbd_fwd_txfm2d_8x4_sse2, // 8x4 transform + av1_lowbd_fwd_txfm2d_8x16_sse2, // 8x16 transform + av1_lowbd_fwd_txfm2d_16x8_sse2, // 16x8 transform + av1_lowbd_fwd_txfm2d_16x32_sse2, // 16x32 transform + av1_lowbd_fwd_txfm2d_32x16_sse2, // 32x16 transform + lowbd_fwd_txfm2d_32x64_sse4_1, // 32x64 transform + lowbd_fwd_txfm2d_64x32_sse4_1, // 64x32 transform + av1_lowbd_fwd_txfm2d_4x16_sse2, // 4x16 transform + av1_lowbd_fwd_txfm2d_16x4_sse2, // 16x4 transform + av1_lowbd_fwd_txfm2d_8x32_sse2, // 8x32 transform + av1_lowbd_fwd_txfm2d_32x8_sse2, // 32x8 transform + av1_lowbd_fwd_txfm2d_16x64_sse2, // 16x64 transform + av1_lowbd_fwd_txfm2d_64x16_sse2, // 64x16 transform +}; + +void av1_lowbd_fwd_txfm_sse4_1(const int16_t *src_diff, tran_low_t *coeff, + int diff_stride, TxfmParam *txfm_param) { + FwdTxfm2dFunc fwd_txfm2d_func = fwd_txfm2d_func_ls[txfm_param->tx_size]; + if ((fwd_txfm2d_func == NULL) || + (txfm_param->lossless && txfm_param->tx_size == TX_4X4)) { + av1_lowbd_fwd_txfm_c(src_diff, coeff, diff_stride, txfm_param); + } else { + fwd_txfm2d_func(src_diff, coeff, diff_stride, txfm_param->tx_type, + txfm_param->bd); + } +} diff --git a/third_party/aom/av1/encoder/x86/av1_fwd_txfm_avx2.h b/third_party/aom/av1/encoder/x86/av1_fwd_txfm_avx2.h new file mode 100644 index 0000000000..38707137c4 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/av1_fwd_txfm_avx2.h @@ -0,0 +1,103 @@ +/* + * Copyright (c) 2018, 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_ENCODER_X86_AV1_FWD_TXFM_AVX2_H_ +#define AOM_AV1_ENCODER_X86_AV1_FWD_TXFM_AVX2_H_ +#include <immintrin.h> + +static INLINE __m256i av1_round_shift_32_avx2(__m256i vec, int bit) { + __m256i tmp, round; + round = _mm256_set1_epi32(1 << (bit - 1)); + tmp = _mm256_add_epi32(vec, round); + return _mm256_srai_epi32(tmp, bit); +} + +// out0 = in0*w0 + in1*w1 +// out1 = -in1*w0 + in0*w1 +static INLINE void btf_32_avx2_type0(const int32_t w0, const int32_t w1, + __m256i *in0, __m256i *in1, + const __m256i _r, const int32_t cos_bit) { + __m256i _in0 = *in0; + __m256i _in1 = *in1; + const __m256i ww0 = _mm256_set1_epi32(w0); + const __m256i ww1 = _mm256_set1_epi32(w1); + const __m256i in0_w0 = _mm256_mullo_epi32(_in0, ww0); + const __m256i in1_w1 = _mm256_mullo_epi32(_in1, ww1); + __m256i temp0 = _mm256_add_epi32(in0_w0, in1_w1); + temp0 = _mm256_add_epi32(temp0, _r); + *in0 = _mm256_srai_epi32(temp0, cos_bit); + const __m256i in0_w1 = _mm256_mullo_epi32(_in0, ww1); + const __m256i in1_w0 = _mm256_mullo_epi32(_in1, ww0); + __m256i temp1 = _mm256_sub_epi32(in0_w1, in1_w0); + temp1 = _mm256_add_epi32(temp1, _r); + *in1 = _mm256_srai_epi32(temp1, cos_bit); +} + +static INLINE void btf_32_avx2_type1(const int32_t w0, const int32_t w1, + __m256i *in0, __m256i *in1, + const __m256i _r, const int32_t cos_bit) { + __m256i _in0 = *in0; + __m256i _in1 = *in1; + const __m256i ww0 = _mm256_set1_epi32(w0); + const __m256i ww1 = _mm256_set1_epi32(w1); + const __m256i in0_w0 = _mm256_mullo_epi32(_in0, ww0); + const __m256i in1_w1 = _mm256_mullo_epi32(_in1, ww1); + __m256i temp0 = _mm256_add_epi32(in0_w0, in1_w1); + temp0 = _mm256_add_epi32(temp0, _r); + *in0 = _mm256_srai_epi32(temp0, cos_bit); + const __m256i in0_w1 = _mm256_mullo_epi32(_in0, ww1); + const __m256i in1_w0 = _mm256_mullo_epi32(_in1, ww0); + __m256i temp1 = _mm256_sub_epi32(in1_w0, in0_w1); + temp1 = _mm256_add_epi32(temp1, _r); + *in1 = _mm256_srai_epi32(temp1, cos_bit); +} + +// out0 = in0*w0 + in1*w1 +// out1 = -in1*w0 + in0*w1 +static INLINE void btf_32_avx2_type0_new(const __m256i ww0, const __m256i ww1, + __m256i *in0, __m256i *in1, + const __m256i _r, + const int32_t cos_bit) { + __m256i _in0 = *in0; + __m256i _in1 = *in1; + const __m256i in0_w0 = _mm256_mullo_epi32(_in0, ww0); + const __m256i in1_w1 = _mm256_mullo_epi32(_in1, ww1); + __m256i temp0 = _mm256_add_epi32(in0_w0, in1_w1); + temp0 = _mm256_add_epi32(temp0, _r); + *in0 = _mm256_srai_epi32(temp0, cos_bit); + const __m256i in0_w1 = _mm256_mullo_epi32(_in0, ww1); + const __m256i in1_w0 = _mm256_mullo_epi32(_in1, ww0); + __m256i temp1 = _mm256_sub_epi32(in0_w1, in1_w0); + temp1 = _mm256_add_epi32(temp1, _r); + *in1 = _mm256_srai_epi32(temp1, cos_bit); +} + +// out0 = in0*w0 + in1*w1 +// out1 = in1*w0 - in0*w1 +static INLINE void btf_32_avx2_type1_new(const __m256i ww0, const __m256i ww1, + __m256i *in0, __m256i *in1, + const __m256i _r, + const int32_t cos_bit) { + __m256i _in0 = *in0; + __m256i _in1 = *in1; + const __m256i in0_w0 = _mm256_mullo_epi32(_in0, ww0); + const __m256i in1_w1 = _mm256_mullo_epi32(_in1, ww1); + __m256i temp0 = _mm256_add_epi32(in0_w0, in1_w1); + temp0 = _mm256_add_epi32(temp0, _r); + *in0 = _mm256_srai_epi32(temp0, cos_bit); + const __m256i in0_w1 = _mm256_mullo_epi32(_in0, ww1); + const __m256i in1_w0 = _mm256_mullo_epi32(_in1, ww0); + __m256i temp1 = _mm256_sub_epi32(in1_w0, in0_w1); + temp1 = _mm256_add_epi32(temp1, _r); + *in1 = _mm256_srai_epi32(temp1, cos_bit); +} + +#endif // AOM_AV1_ENCODER_X86_AV1_FWD_TXFM_AVX2_H_ diff --git a/third_party/aom/av1/encoder/x86/av1_fwd_txfm_sse2.c b/third_party/aom/av1/encoder/x86/av1_fwd_txfm_sse2.c new file mode 100644 index 0000000000..6aae7ce1e4 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/av1_fwd_txfm_sse2.c @@ -0,0 +1,2889 @@ +/* + * Copyright (c) 2018, 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 "av1/common/x86/av1_txfm_sse2.h" +#include "av1/encoder/av1_fwd_txfm1d_cfg.h" +#include "av1/encoder/x86/av1_fwd_txfm_sse2.h" + +// TODO(linfengz): refine fdct4x8 and fadst4x8 optimization (if possible). + +static void fdct4x4_new_sse2(const __m128i *input, __m128i *output, + int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]); + const __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]); + const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]); + const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]); + const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1)); + __m128i u[4], v[4]; + + u[0] = _mm_unpacklo_epi16(input[0], input[1]); + u[1] = _mm_unpacklo_epi16(input[3], input[2]); + + v[0] = _mm_add_epi16(u[0], u[1]); + v[1] = _mm_sub_epi16(u[0], u[1]); + + u[0] = _mm_madd_epi16(v[0], cospi_p32_p32); // 0 + u[1] = _mm_madd_epi16(v[0], cospi_p32_m32); // 2 + u[2] = _mm_madd_epi16(v[1], cospi_p16_p48); // 1 + u[3] = _mm_madd_epi16(v[1], cospi_p48_m16); // 3 + + v[0] = _mm_add_epi32(u[0], __rounding); + v[1] = _mm_add_epi32(u[1], __rounding); + v[2] = _mm_add_epi32(u[2], __rounding); + v[3] = _mm_add_epi32(u[3], __rounding); + u[0] = _mm_srai_epi32(v[0], cos_bit); + u[1] = _mm_srai_epi32(v[1], cos_bit); + u[2] = _mm_srai_epi32(v[2], cos_bit); + u[3] = _mm_srai_epi32(v[3], cos_bit); + + output[0] = _mm_packs_epi32(u[0], u[1]); + output[1] = _mm_packs_epi32(u[2], u[3]); + output[2] = _mm_srli_si128(output[0], 8); + output[3] = _mm_srli_si128(output[1], 8); +} + +static void fdct8x4_new_sse2(const __m128i *input, __m128i *output, + int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1)); + + __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]); + __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]); + __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]); + __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]); + + // stage 1 + __m128i x1[4]; + x1[0] = _mm_adds_epi16(input[0], input[3]); + x1[3] = _mm_subs_epi16(input[0], input[3]); + x1[1] = _mm_adds_epi16(input[1], input[2]); + x1[2] = _mm_subs_epi16(input[1], input[2]); + + // stage 2 + __m128i x2[4]; + btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x1[0], x1[1], x2[0], x2[1]); + btf_16_sse2(cospi_p48_p16, cospi_m16_p48, x1[2], x1[3], x2[2], x2[3]); + + // stage 3 + output[0] = x2[0]; + output[1] = x2[2]; + output[2] = x2[1]; + output[3] = x2[3]; +} + +static void fdct4x8_new_sse2(const __m128i *input, __m128i *output, + int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1)); + + __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]); + __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]); + __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]); + __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]); + __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]); + __m128i cospi_p56_p08 = pair_set_epi16(cospi[56], cospi[8]); + __m128i cospi_m08_p56 = pair_set_epi16(-cospi[8], cospi[56]); + __m128i cospi_p24_p40 = pair_set_epi16(cospi[24], cospi[40]); + __m128i cospi_m40_p24 = pair_set_epi16(-cospi[40], cospi[24]); + + // stage 1 + __m128i x1[8]; + x1[0] = _mm_adds_epi16(input[0], input[7]); + x1[7] = _mm_subs_epi16(input[0], input[7]); + x1[1] = _mm_adds_epi16(input[1], input[6]); + x1[6] = _mm_subs_epi16(input[1], input[6]); + x1[2] = _mm_adds_epi16(input[2], input[5]); + x1[5] = _mm_subs_epi16(input[2], input[5]); + x1[3] = _mm_adds_epi16(input[3], input[4]); + x1[4] = _mm_subs_epi16(input[3], input[4]); + + // stage 2 + __m128i x2[8]; + x2[0] = _mm_adds_epi16(x1[0], x1[3]); + x2[3] = _mm_subs_epi16(x1[0], x1[3]); + x2[1] = _mm_adds_epi16(x1[1], x1[2]); + x2[2] = _mm_subs_epi16(x1[1], x1[2]); + x2[4] = x1[4]; + btf_16_w4_sse2(&cospi_m32_p32, &cospi_p32_p32, __rounding, cos_bit, &x1[5], + &x1[6], &x2[5], &x2[6]); + x2[7] = x1[7]; + + // stage 3 + __m128i x3[8]; + btf_16_w4_sse2(&cospi_p32_p32, &cospi_p32_m32, __rounding, cos_bit, &x2[0], + &x2[1], &x3[0], &x3[1]); + btf_16_w4_sse2(&cospi_p48_p16, &cospi_m16_p48, __rounding, cos_bit, &x2[2], + &x2[3], &x3[2], &x3[3]); + x3[4] = _mm_adds_epi16(x2[4], x2[5]); + x3[5] = _mm_subs_epi16(x2[4], x2[5]); + x3[6] = _mm_subs_epi16(x2[7], x2[6]); + x3[7] = _mm_adds_epi16(x2[7], x2[6]); + + // stage 4 + __m128i x4[8]; + x4[0] = x3[0]; + x4[1] = x3[1]; + x4[2] = x3[2]; + x4[3] = x3[3]; + btf_16_w4_sse2(&cospi_p56_p08, &cospi_m08_p56, __rounding, cos_bit, &x3[4], + &x3[7], &x4[4], &x4[7]); + btf_16_w4_sse2(&cospi_p24_p40, &cospi_m40_p24, __rounding, cos_bit, &x3[5], + &x3[6], &x4[5], &x4[6]); + + // stage 5 + output[0] = x4[0]; + output[1] = x4[4]; + output[2] = x4[2]; + output[3] = x4[6]; + output[4] = x4[1]; + output[5] = x4[5]; + output[6] = x4[3]; + output[7] = x4[7]; +} + +static void fdct8x8_new_sse2(const __m128i *input, __m128i *output, + int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1)); + + __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]); + __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]); + __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]); + __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]); + __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]); + __m128i cospi_p56_p08 = pair_set_epi16(cospi[56], cospi[8]); + __m128i cospi_m08_p56 = pair_set_epi16(-cospi[8], cospi[56]); + __m128i cospi_p24_p40 = pair_set_epi16(cospi[24], cospi[40]); + __m128i cospi_m40_p24 = pair_set_epi16(-cospi[40], cospi[24]); + + // stage 1 + __m128i x1[8]; + x1[0] = _mm_adds_epi16(input[0], input[7]); + x1[7] = _mm_subs_epi16(input[0], input[7]); + x1[1] = _mm_adds_epi16(input[1], input[6]); + x1[6] = _mm_subs_epi16(input[1], input[6]); + x1[2] = _mm_adds_epi16(input[2], input[5]); + x1[5] = _mm_subs_epi16(input[2], input[5]); + x1[3] = _mm_adds_epi16(input[3], input[4]); + x1[4] = _mm_subs_epi16(input[3], input[4]); + + // stage 2 + __m128i x2[8]; + x2[0] = _mm_adds_epi16(x1[0], x1[3]); + x2[3] = _mm_subs_epi16(x1[0], x1[3]); + x2[1] = _mm_adds_epi16(x1[1], x1[2]); + x2[2] = _mm_subs_epi16(x1[1], x1[2]); + x2[4] = x1[4]; + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[5], x1[6], x2[5], x2[6]); + x2[7] = x1[7]; + + // stage 3 + __m128i x3[8]; + btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x2[0], x2[1], x3[0], x3[1]); + btf_16_sse2(cospi_p48_p16, cospi_m16_p48, x2[2], x2[3], x3[2], x3[3]); + x3[4] = _mm_adds_epi16(x2[4], x2[5]); + x3[5] = _mm_subs_epi16(x2[4], x2[5]); + x3[6] = _mm_subs_epi16(x2[7], x2[6]); + x3[7] = _mm_adds_epi16(x2[7], x2[6]); + + // stage 4 + __m128i x4[8]; + x4[0] = x3[0]; + x4[1] = x3[1]; + x4[2] = x3[2]; + x4[3] = x3[3]; + btf_16_sse2(cospi_p56_p08, cospi_m08_p56, x3[4], x3[7], x4[4], x4[7]); + btf_16_sse2(cospi_p24_p40, cospi_m40_p24, x3[5], x3[6], x4[5], x4[6]); + + // stage 5 + output[0] = x4[0]; + output[1] = x4[4]; + output[2] = x4[2]; + output[3] = x4[6]; + output[4] = x4[1]; + output[5] = x4[5]; + output[6] = x4[3]; + output[7] = x4[7]; +} + +static void fdct8x16_new_sse2(const __m128i *input, __m128i *output, + int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1)); + + __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]); + __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]); + __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]); + __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]); + __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]); + __m128i cospi_m48_m16 = pair_set_epi16(-cospi[48], -cospi[16]); + __m128i cospi_p56_p08 = pair_set_epi16(cospi[56], cospi[8]); + __m128i cospi_m08_p56 = pair_set_epi16(-cospi[8], cospi[56]); + __m128i cospi_p24_p40 = pair_set_epi16(cospi[24], cospi[40]); + __m128i cospi_m40_p24 = pair_set_epi16(-cospi[40], cospi[24]); + __m128i cospi_p60_p04 = pair_set_epi16(cospi[60], cospi[4]); + __m128i cospi_m04_p60 = pair_set_epi16(-cospi[4], cospi[60]); + __m128i cospi_p28_p36 = pair_set_epi16(cospi[28], cospi[36]); + __m128i cospi_m36_p28 = pair_set_epi16(-cospi[36], cospi[28]); + __m128i cospi_p44_p20 = pair_set_epi16(cospi[44], cospi[20]); + __m128i cospi_m20_p44 = pair_set_epi16(-cospi[20], cospi[44]); + __m128i cospi_p12_p52 = pair_set_epi16(cospi[12], cospi[52]); + __m128i cospi_m52_p12 = pair_set_epi16(-cospi[52], cospi[12]); + + // stage 1 + __m128i x1[16]; + x1[0] = _mm_adds_epi16(input[0], input[15]); + x1[15] = _mm_subs_epi16(input[0], input[15]); + x1[1] = _mm_adds_epi16(input[1], input[14]); + x1[14] = _mm_subs_epi16(input[1], input[14]); + x1[2] = _mm_adds_epi16(input[2], input[13]); + x1[13] = _mm_subs_epi16(input[2], input[13]); + x1[3] = _mm_adds_epi16(input[3], input[12]); + x1[12] = _mm_subs_epi16(input[3], input[12]); + x1[4] = _mm_adds_epi16(input[4], input[11]); + x1[11] = _mm_subs_epi16(input[4], input[11]); + x1[5] = _mm_adds_epi16(input[5], input[10]); + x1[10] = _mm_subs_epi16(input[5], input[10]); + x1[6] = _mm_adds_epi16(input[6], input[9]); + x1[9] = _mm_subs_epi16(input[6], input[9]); + x1[7] = _mm_adds_epi16(input[7], input[8]); + x1[8] = _mm_subs_epi16(input[7], input[8]); + + // stage 2 + __m128i x2[16]; + x2[0] = _mm_adds_epi16(x1[0], x1[7]); + x2[7] = _mm_subs_epi16(x1[0], x1[7]); + x2[1] = _mm_adds_epi16(x1[1], x1[6]); + x2[6] = _mm_subs_epi16(x1[1], x1[6]); + x2[2] = _mm_adds_epi16(x1[2], x1[5]); + x2[5] = _mm_subs_epi16(x1[2], x1[5]); + x2[3] = _mm_adds_epi16(x1[3], x1[4]); + x2[4] = _mm_subs_epi16(x1[3], x1[4]); + x2[8] = x1[8]; + x2[9] = x1[9]; + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[10], x1[13], x2[10], x2[13]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[11], x1[12], x2[11], x2[12]); + x2[14] = x1[14]; + x2[15] = x1[15]; + + // stage 3 + __m128i x3[16]; + x3[0] = _mm_adds_epi16(x2[0], x2[3]); + x3[3] = _mm_subs_epi16(x2[0], x2[3]); + x3[1] = _mm_adds_epi16(x2[1], x2[2]); + x3[2] = _mm_subs_epi16(x2[1], x2[2]); + x3[4] = x2[4]; + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x2[5], x2[6], x3[5], x3[6]); + x3[7] = x2[7]; + x3[8] = _mm_adds_epi16(x2[8], x2[11]); + x3[11] = _mm_subs_epi16(x2[8], x2[11]); + x3[9] = _mm_adds_epi16(x2[9], x2[10]); + x3[10] = _mm_subs_epi16(x2[9], x2[10]); + x3[12] = _mm_subs_epi16(x2[15], x2[12]); + x3[15] = _mm_adds_epi16(x2[15], x2[12]); + x3[13] = _mm_subs_epi16(x2[14], x2[13]); + x3[14] = _mm_adds_epi16(x2[14], x2[13]); + + // stage 4 + __m128i x4[16]; + btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x3[0], x3[1], x4[0], x4[1]); + btf_16_sse2(cospi_p48_p16, cospi_m16_p48, x3[2], x3[3], x4[2], x4[3]); + x4[4] = _mm_adds_epi16(x3[4], x3[5]); + x4[5] = _mm_subs_epi16(x3[4], x3[5]); + x4[6] = _mm_subs_epi16(x3[7], x3[6]); + x4[7] = _mm_adds_epi16(x3[7], x3[6]); + x4[8] = x3[8]; + btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x3[9], x3[14], x4[9], x4[14]); + btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x3[10], x3[13], x4[10], x4[13]); + x4[11] = x3[11]; + x4[12] = x3[12]; + x4[15] = x3[15]; + + // stage 5 + __m128i x5[16]; + x5[0] = x4[0]; + x5[1] = x4[1]; + x5[2] = x4[2]; + x5[3] = x4[3]; + btf_16_sse2(cospi_p56_p08, cospi_m08_p56, x4[4], x4[7], x5[4], x5[7]); + btf_16_sse2(cospi_p24_p40, cospi_m40_p24, x4[5], x4[6], x5[5], x5[6]); + x5[8] = _mm_adds_epi16(x4[8], x4[9]); + x5[9] = _mm_subs_epi16(x4[8], x4[9]); + x5[10] = _mm_subs_epi16(x4[11], x4[10]); + x5[11] = _mm_adds_epi16(x4[11], x4[10]); + x5[12] = _mm_adds_epi16(x4[12], x4[13]); + x5[13] = _mm_subs_epi16(x4[12], x4[13]); + x5[14] = _mm_subs_epi16(x4[15], x4[14]); + x5[15] = _mm_adds_epi16(x4[15], x4[14]); + + // stage 6 + __m128i x6[16]; + x6[0] = x5[0]; + x6[1] = x5[1]; + x6[2] = x5[2]; + x6[3] = x5[3]; + x6[4] = x5[4]; + x6[5] = x5[5]; + x6[6] = x5[6]; + x6[7] = x5[7]; + btf_16_sse2(cospi_p60_p04, cospi_m04_p60, x5[8], x5[15], x6[8], x6[15]); + btf_16_sse2(cospi_p28_p36, cospi_m36_p28, x5[9], x5[14], x6[9], x6[14]); + btf_16_sse2(cospi_p44_p20, cospi_m20_p44, x5[10], x5[13], x6[10], x6[13]); + btf_16_sse2(cospi_p12_p52, cospi_m52_p12, x5[11], x5[12], x6[11], x6[12]); + + // stage 7 + output[0] = x6[0]; + output[1] = x6[8]; + output[2] = x6[4]; + output[3] = x6[12]; + output[4] = x6[2]; + output[5] = x6[10]; + output[6] = x6[6]; + output[7] = x6[14]; + output[8] = x6[1]; + output[9] = x6[9]; + output[10] = x6[5]; + output[11] = x6[13]; + output[12] = x6[3]; + output[13] = x6[11]; + output[14] = x6[7]; + output[15] = x6[15]; +} + +void fdct8x32_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1)); + + __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]); + __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]); + __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]); + __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]); + __m128i cospi_m48_m16 = pair_set_epi16(-cospi[48], -cospi[16]); + __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]); + __m128i cospi_p56_p08 = pair_set_epi16(cospi[56], cospi[8]); + __m128i cospi_m08_p56 = pair_set_epi16(-cospi[8], cospi[56]); + __m128i cospi_p24_p40 = pair_set_epi16(cospi[24], cospi[40]); + __m128i cospi_m40_p24 = pair_set_epi16(-cospi[40], cospi[24]); + __m128i cospi_m56_m08 = pair_set_epi16(-cospi[56], -cospi[8]); + __m128i cospi_m24_m40 = pair_set_epi16(-cospi[24], -cospi[40]); + __m128i cospi_p60_p04 = pair_set_epi16(cospi[60], cospi[4]); + __m128i cospi_m04_p60 = pair_set_epi16(-cospi[4], cospi[60]); + __m128i cospi_p28_p36 = pair_set_epi16(cospi[28], cospi[36]); + __m128i cospi_m36_p28 = pair_set_epi16(-cospi[36], cospi[28]); + __m128i cospi_p44_p20 = pair_set_epi16(cospi[44], cospi[20]); + __m128i cospi_m20_p44 = pair_set_epi16(-cospi[20], cospi[44]); + __m128i cospi_p12_p52 = pair_set_epi16(cospi[12], cospi[52]); + __m128i cospi_m52_p12 = pair_set_epi16(-cospi[52], cospi[12]); + __m128i cospi_p62_p02 = pair_set_epi16(cospi[62], cospi[2]); + __m128i cospi_m02_p62 = pair_set_epi16(-cospi[2], cospi[62]); + __m128i cospi_p30_p34 = pair_set_epi16(cospi[30], cospi[34]); + __m128i cospi_m34_p30 = pair_set_epi16(-cospi[34], cospi[30]); + __m128i cospi_p46_p18 = pair_set_epi16(cospi[46], cospi[18]); + __m128i cospi_m18_p46 = pair_set_epi16(-cospi[18], cospi[46]); + __m128i cospi_p14_p50 = pair_set_epi16(cospi[14], cospi[50]); + __m128i cospi_m50_p14 = pair_set_epi16(-cospi[50], cospi[14]); + __m128i cospi_p54_p10 = pair_set_epi16(cospi[54], cospi[10]); + __m128i cospi_m10_p54 = pair_set_epi16(-cospi[10], cospi[54]); + __m128i cospi_p22_p42 = pair_set_epi16(cospi[22], cospi[42]); + __m128i cospi_m42_p22 = pair_set_epi16(-cospi[42], cospi[22]); + __m128i cospi_p38_p26 = pair_set_epi16(cospi[38], cospi[26]); + __m128i cospi_m26_p38 = pair_set_epi16(-cospi[26], cospi[38]); + __m128i cospi_p06_p58 = pair_set_epi16(cospi[6], cospi[58]); + __m128i cospi_m58_p06 = pair_set_epi16(-cospi[58], cospi[6]); + + // stage 1 + __m128i x1[32]; + x1[0] = _mm_adds_epi16(input[0], input[31]); + x1[31] = _mm_subs_epi16(input[0], input[31]); + x1[1] = _mm_adds_epi16(input[1], input[30]); + x1[30] = _mm_subs_epi16(input[1], input[30]); + x1[2] = _mm_adds_epi16(input[2], input[29]); + x1[29] = _mm_subs_epi16(input[2], input[29]); + x1[3] = _mm_adds_epi16(input[3], input[28]); + x1[28] = _mm_subs_epi16(input[3], input[28]); + x1[4] = _mm_adds_epi16(input[4], input[27]); + x1[27] = _mm_subs_epi16(input[4], input[27]); + x1[5] = _mm_adds_epi16(input[5], input[26]); + x1[26] = _mm_subs_epi16(input[5], input[26]); + x1[6] = _mm_adds_epi16(input[6], input[25]); + x1[25] = _mm_subs_epi16(input[6], input[25]); + x1[7] = _mm_adds_epi16(input[7], input[24]); + x1[24] = _mm_subs_epi16(input[7], input[24]); + x1[8] = _mm_adds_epi16(input[8], input[23]); + x1[23] = _mm_subs_epi16(input[8], input[23]); + x1[9] = _mm_adds_epi16(input[9], input[22]); + x1[22] = _mm_subs_epi16(input[9], input[22]); + x1[10] = _mm_adds_epi16(input[10], input[21]); + x1[21] = _mm_subs_epi16(input[10], input[21]); + x1[11] = _mm_adds_epi16(input[11], input[20]); + x1[20] = _mm_subs_epi16(input[11], input[20]); + x1[12] = _mm_adds_epi16(input[12], input[19]); + x1[19] = _mm_subs_epi16(input[12], input[19]); + x1[13] = _mm_adds_epi16(input[13], input[18]); + x1[18] = _mm_subs_epi16(input[13], input[18]); + x1[14] = _mm_adds_epi16(input[14], input[17]); + x1[17] = _mm_subs_epi16(input[14], input[17]); + x1[15] = _mm_adds_epi16(input[15], input[16]); + x1[16] = _mm_subs_epi16(input[15], input[16]); + + // stage 2 + __m128i x2[32]; + x2[0] = _mm_adds_epi16(x1[0], x1[15]); + x2[15] = _mm_subs_epi16(x1[0], x1[15]); + x2[1] = _mm_adds_epi16(x1[1], x1[14]); + x2[14] = _mm_subs_epi16(x1[1], x1[14]); + x2[2] = _mm_adds_epi16(x1[2], x1[13]); + x2[13] = _mm_subs_epi16(x1[2], x1[13]); + x2[3] = _mm_adds_epi16(x1[3], x1[12]); + x2[12] = _mm_subs_epi16(x1[3], x1[12]); + x2[4] = _mm_adds_epi16(x1[4], x1[11]); + x2[11] = _mm_subs_epi16(x1[4], x1[11]); + x2[5] = _mm_adds_epi16(x1[5], x1[10]); + x2[10] = _mm_subs_epi16(x1[5], x1[10]); + x2[6] = _mm_adds_epi16(x1[6], x1[9]); + x2[9] = _mm_subs_epi16(x1[6], x1[9]); + x2[7] = _mm_adds_epi16(x1[7], x1[8]); + x2[8] = _mm_subs_epi16(x1[7], x1[8]); + x2[16] = x1[16]; + x2[17] = x1[17]; + x2[18] = x1[18]; + x2[19] = x1[19]; + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[20], x1[27], x2[20], x2[27]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[21], x1[26], x2[21], x2[26]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[22], x1[25], x2[22], x2[25]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[23], x1[24], x2[23], x2[24]); + x2[28] = x1[28]; + x2[29] = x1[29]; + x2[30] = x1[30]; + x2[31] = x1[31]; + + // stage 3 + __m128i x3[32]; + x3[0] = _mm_adds_epi16(x2[0], x2[7]); + x3[7] = _mm_subs_epi16(x2[0], x2[7]); + x3[1] = _mm_adds_epi16(x2[1], x2[6]); + x3[6] = _mm_subs_epi16(x2[1], x2[6]); + x3[2] = _mm_adds_epi16(x2[2], x2[5]); + x3[5] = _mm_subs_epi16(x2[2], x2[5]); + x3[3] = _mm_adds_epi16(x2[3], x2[4]); + x3[4] = _mm_subs_epi16(x2[3], x2[4]); + x3[8] = x2[8]; + x3[9] = x2[9]; + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x2[10], x2[13], x3[10], x3[13]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x2[11], x2[12], x3[11], x3[12]); + x3[14] = x2[14]; + x3[15] = x2[15]; + x3[16] = _mm_adds_epi16(x2[16], x2[23]); + x3[23] = _mm_subs_epi16(x2[16], x2[23]); + x3[17] = _mm_adds_epi16(x2[17], x2[22]); + x3[22] = _mm_subs_epi16(x2[17], x2[22]); + x3[18] = _mm_adds_epi16(x2[18], x2[21]); + x3[21] = _mm_subs_epi16(x2[18], x2[21]); + x3[19] = _mm_adds_epi16(x2[19], x2[20]); + x3[20] = _mm_subs_epi16(x2[19], x2[20]); + x3[24] = _mm_subs_epi16(x2[31], x2[24]); + x3[31] = _mm_adds_epi16(x2[31], x2[24]); + x3[25] = _mm_subs_epi16(x2[30], x2[25]); + x3[30] = _mm_adds_epi16(x2[30], x2[25]); + x3[26] = _mm_subs_epi16(x2[29], x2[26]); + x3[29] = _mm_adds_epi16(x2[29], x2[26]); + x3[27] = _mm_subs_epi16(x2[28], x2[27]); + x3[28] = _mm_adds_epi16(x2[28], x2[27]); + + // stage 4 + __m128i x4[32]; + x4[0] = _mm_adds_epi16(x3[0], x3[3]); + x4[3] = _mm_subs_epi16(x3[0], x3[3]); + x4[1] = _mm_adds_epi16(x3[1], x3[2]); + x4[2] = _mm_subs_epi16(x3[1], x3[2]); + x4[4] = x3[4]; + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x3[5], x3[6], x4[5], x4[6]); + x4[7] = x3[7]; + x4[8] = _mm_adds_epi16(x3[8], x3[11]); + x4[11] = _mm_subs_epi16(x3[8], x3[11]); + x4[9] = _mm_adds_epi16(x3[9], x3[10]); + x4[10] = _mm_subs_epi16(x3[9], x3[10]); + x4[12] = _mm_subs_epi16(x3[15], x3[12]); + x4[15] = _mm_adds_epi16(x3[15], x3[12]); + x4[13] = _mm_subs_epi16(x3[14], x3[13]); + x4[14] = _mm_adds_epi16(x3[14], x3[13]); + x4[16] = x3[16]; + x4[17] = x3[17]; + btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x3[18], x3[29], x4[18], x4[29]); + btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x3[19], x3[28], x4[19], x4[28]); + btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x3[20], x3[27], x4[20], x4[27]); + btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x3[21], x3[26], x4[21], x4[26]); + x4[22] = x3[22]; + x4[23] = x3[23]; + x4[24] = x3[24]; + x4[25] = x3[25]; + x4[30] = x3[30]; + x4[31] = x3[31]; + + // stage 5 + __m128i x5[32]; + btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x4[0], x4[1], x5[0], x5[1]); + btf_16_sse2(cospi_p48_p16, cospi_m16_p48, x4[2], x4[3], x5[2], x5[3]); + x5[4] = _mm_adds_epi16(x4[4], x4[5]); + x5[5] = _mm_subs_epi16(x4[4], x4[5]); + x5[6] = _mm_subs_epi16(x4[7], x4[6]); + x5[7] = _mm_adds_epi16(x4[7], x4[6]); + x5[8] = x4[8]; + btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x4[9], x4[14], x5[9], x5[14]); + btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x4[10], x4[13], x5[10], x5[13]); + x5[11] = x4[11]; + x5[12] = x4[12]; + x5[15] = x4[15]; + x5[16] = _mm_adds_epi16(x4[16], x4[19]); + x5[19] = _mm_subs_epi16(x4[16], x4[19]); + x5[17] = _mm_adds_epi16(x4[17], x4[18]); + x5[18] = _mm_subs_epi16(x4[17], x4[18]); + x5[20] = _mm_subs_epi16(x4[23], x4[20]); + x5[23] = _mm_adds_epi16(x4[23], x4[20]); + x5[21] = _mm_subs_epi16(x4[22], x4[21]); + x5[22] = _mm_adds_epi16(x4[22], x4[21]); + x5[24] = _mm_adds_epi16(x4[24], x4[27]); + x5[27] = _mm_subs_epi16(x4[24], x4[27]); + x5[25] = _mm_adds_epi16(x4[25], x4[26]); + x5[26] = _mm_subs_epi16(x4[25], x4[26]); + x5[28] = _mm_subs_epi16(x4[31], x4[28]); + x5[31] = _mm_adds_epi16(x4[31], x4[28]); + x5[29] = _mm_subs_epi16(x4[30], x4[29]); + x5[30] = _mm_adds_epi16(x4[30], x4[29]); + + // stage 6 + __m128i x6[32]; + x6[0] = x5[0]; + x6[1] = x5[1]; + x6[2] = x5[2]; + x6[3] = x5[3]; + btf_16_sse2(cospi_p56_p08, cospi_m08_p56, x5[4], x5[7], x6[4], x6[7]); + btf_16_sse2(cospi_p24_p40, cospi_m40_p24, x5[5], x5[6], x6[5], x6[6]); + x6[8] = _mm_adds_epi16(x5[8], x5[9]); + x6[9] = _mm_subs_epi16(x5[8], x5[9]); + x6[10] = _mm_subs_epi16(x5[11], x5[10]); + x6[11] = _mm_adds_epi16(x5[11], x5[10]); + x6[12] = _mm_adds_epi16(x5[12], x5[13]); + x6[13] = _mm_subs_epi16(x5[12], x5[13]); + x6[14] = _mm_subs_epi16(x5[15], x5[14]); + x6[15] = _mm_adds_epi16(x5[15], x5[14]); + x6[16] = x5[16]; + btf_16_sse2(cospi_m08_p56, cospi_p56_p08, x5[17], x5[30], x6[17], x6[30]); + btf_16_sse2(cospi_m56_m08, cospi_m08_p56, x5[18], x5[29], x6[18], x6[29]); + x6[19] = x5[19]; + x6[20] = x5[20]; + btf_16_sse2(cospi_m40_p24, cospi_p24_p40, x5[21], x5[26], x6[21], x6[26]); + btf_16_sse2(cospi_m24_m40, cospi_m40_p24, x5[22], x5[25], x6[22], x6[25]); + x6[23] = x5[23]; + x6[24] = x5[24]; + x6[27] = x5[27]; + x6[28] = x5[28]; + x6[31] = x5[31]; + + // stage 7 + __m128i x7[32]; + x7[0] = x6[0]; + x7[1] = x6[1]; + x7[2] = x6[2]; + x7[3] = x6[3]; + x7[4] = x6[4]; + x7[5] = x6[5]; + x7[6] = x6[6]; + x7[7] = x6[7]; + btf_16_sse2(cospi_p60_p04, cospi_m04_p60, x6[8], x6[15], x7[8], x7[15]); + btf_16_sse2(cospi_p28_p36, cospi_m36_p28, x6[9], x6[14], x7[9], x7[14]); + btf_16_sse2(cospi_p44_p20, cospi_m20_p44, x6[10], x6[13], x7[10], x7[13]); + btf_16_sse2(cospi_p12_p52, cospi_m52_p12, x6[11], x6[12], x7[11], x7[12]); + x7[16] = _mm_adds_epi16(x6[16], x6[17]); + x7[17] = _mm_subs_epi16(x6[16], x6[17]); + x7[18] = _mm_subs_epi16(x6[19], x6[18]); + x7[19] = _mm_adds_epi16(x6[19], x6[18]); + x7[20] = _mm_adds_epi16(x6[20], x6[21]); + x7[21] = _mm_subs_epi16(x6[20], x6[21]); + x7[22] = _mm_subs_epi16(x6[23], x6[22]); + x7[23] = _mm_adds_epi16(x6[23], x6[22]); + x7[24] = _mm_adds_epi16(x6[24], x6[25]); + x7[25] = _mm_subs_epi16(x6[24], x6[25]); + x7[26] = _mm_subs_epi16(x6[27], x6[26]); + x7[27] = _mm_adds_epi16(x6[27], x6[26]); + x7[28] = _mm_adds_epi16(x6[28], x6[29]); + x7[29] = _mm_subs_epi16(x6[28], x6[29]); + x7[30] = _mm_subs_epi16(x6[31], x6[30]); + x7[31] = _mm_adds_epi16(x6[31], x6[30]); + + // stage 8 + __m128i x8[32]; + x8[0] = x7[0]; + x8[1] = x7[1]; + x8[2] = x7[2]; + x8[3] = x7[3]; + x8[4] = x7[4]; + x8[5] = x7[5]; + x8[6] = x7[6]; + x8[7] = x7[7]; + x8[8] = x7[8]; + x8[9] = x7[9]; + x8[10] = x7[10]; + x8[11] = x7[11]; + x8[12] = x7[12]; + x8[13] = x7[13]; + x8[14] = x7[14]; + x8[15] = x7[15]; + btf_16_sse2(cospi_p62_p02, cospi_m02_p62, x7[16], x7[31], x8[16], x8[31]); + btf_16_sse2(cospi_p30_p34, cospi_m34_p30, x7[17], x7[30], x8[17], x8[30]); + btf_16_sse2(cospi_p46_p18, cospi_m18_p46, x7[18], x7[29], x8[18], x8[29]); + btf_16_sse2(cospi_p14_p50, cospi_m50_p14, x7[19], x7[28], x8[19], x8[28]); + btf_16_sse2(cospi_p54_p10, cospi_m10_p54, x7[20], x7[27], x8[20], x8[27]); + btf_16_sse2(cospi_p22_p42, cospi_m42_p22, x7[21], x7[26], x8[21], x8[26]); + btf_16_sse2(cospi_p38_p26, cospi_m26_p38, x7[22], x7[25], x8[22], x8[25]); + btf_16_sse2(cospi_p06_p58, cospi_m58_p06, x7[23], x7[24], x8[23], x8[24]); + + // stage 9 + output[0] = x8[0]; + output[1] = x8[16]; + output[2] = x8[8]; + output[3] = x8[24]; + output[4] = x8[4]; + output[5] = x8[20]; + output[6] = x8[12]; + output[7] = x8[28]; + output[8] = x8[2]; + output[9] = x8[18]; + output[10] = x8[10]; + output[11] = x8[26]; + output[12] = x8[6]; + output[13] = x8[22]; + output[14] = x8[14]; + output[15] = x8[30]; + output[16] = x8[1]; + output[17] = x8[17]; + output[18] = x8[9]; + output[19] = x8[25]; + output[20] = x8[5]; + output[21] = x8[21]; + output[22] = x8[13]; + output[23] = x8[29]; + output[24] = x8[3]; + output[25] = x8[19]; + output[26] = x8[11]; + output[27] = x8[27]; + output[28] = x8[7]; + output[29] = x8[23]; + output[30] = x8[15]; + output[31] = x8[31]; +} + +void fdct8x64_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1)); + + __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]); + __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]); + __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]); + __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]); + __m128i cospi_m48_m16 = pair_set_epi16(-cospi[48], -cospi[16]); + __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]); + __m128i cospi_m08_p56 = pair_set_epi16(-cospi[8], cospi[56]); + __m128i cospi_p56_p08 = pair_set_epi16(cospi[56], cospi[8]); + __m128i cospi_m56_m08 = pair_set_epi16(-cospi[56], -cospi[8]); + __m128i cospi_m40_p24 = pair_set_epi16(-cospi[40], cospi[24]); + __m128i cospi_p24_p40 = pair_set_epi16(cospi[24], cospi[40]); + __m128i cospi_m24_m40 = pair_set_epi16(-cospi[24], -cospi[40]); + __m128i cospi_p60_p04 = pair_set_epi16(cospi[60], cospi[4]); + __m128i cospi_m04_p60 = pair_set_epi16(-cospi[4], cospi[60]); + __m128i cospi_p28_p36 = pair_set_epi16(cospi[28], cospi[36]); + __m128i cospi_m36_p28 = pair_set_epi16(-cospi[36], cospi[28]); + __m128i cospi_p44_p20 = pair_set_epi16(cospi[44], cospi[20]); + __m128i cospi_m20_p44 = pair_set_epi16(-cospi[20], cospi[44]); + __m128i cospi_p12_p52 = pair_set_epi16(cospi[12], cospi[52]); + __m128i cospi_m52_p12 = pair_set_epi16(-cospi[52], cospi[12]); + __m128i cospi_m60_m04 = pair_set_epi16(-cospi[60], -cospi[4]); + __m128i cospi_m28_m36 = pair_set_epi16(-cospi[28], -cospi[36]); + __m128i cospi_m44_m20 = pair_set_epi16(-cospi[44], -cospi[20]); + __m128i cospi_m12_m52 = pair_set_epi16(-cospi[12], -cospi[52]); + __m128i cospi_p62_p02 = pair_set_epi16(cospi[62], cospi[2]); + __m128i cospi_m02_p62 = pair_set_epi16(-cospi[2], cospi[62]); + __m128i cospi_p30_p34 = pair_set_epi16(cospi[30], cospi[34]); + __m128i cospi_m34_p30 = pair_set_epi16(-cospi[34], cospi[30]); + __m128i cospi_p46_p18 = pair_set_epi16(cospi[46], cospi[18]); + __m128i cospi_m18_p46 = pair_set_epi16(-cospi[18], cospi[46]); + __m128i cospi_p14_p50 = pair_set_epi16(cospi[14], cospi[50]); + __m128i cospi_m50_p14 = pair_set_epi16(-cospi[50], cospi[14]); + __m128i cospi_p54_p10 = pair_set_epi16(cospi[54], cospi[10]); + __m128i cospi_m10_p54 = pair_set_epi16(-cospi[10], cospi[54]); + __m128i cospi_p22_p42 = pair_set_epi16(cospi[22], cospi[42]); + __m128i cospi_m42_p22 = pair_set_epi16(-cospi[42], cospi[22]); + __m128i cospi_p38_p26 = pair_set_epi16(cospi[38], cospi[26]); + __m128i cospi_m26_p38 = pair_set_epi16(-cospi[26], cospi[38]); + __m128i cospi_p06_p58 = pair_set_epi16(cospi[6], cospi[58]); + __m128i cospi_m58_p06 = pair_set_epi16(-cospi[58], cospi[6]); + __m128i cospi_p63_p01 = pair_set_epi16(cospi[63], cospi[1]); + __m128i cospi_m01_p63 = pair_set_epi16(-cospi[1], cospi[63]); + __m128i cospi_p31_p33 = pair_set_epi16(cospi[31], cospi[33]); + __m128i cospi_m33_p31 = pair_set_epi16(-cospi[33], cospi[31]); + __m128i cospi_p47_p17 = pair_set_epi16(cospi[47], cospi[17]); + __m128i cospi_m17_p47 = pair_set_epi16(-cospi[17], cospi[47]); + __m128i cospi_p15_p49 = pair_set_epi16(cospi[15], cospi[49]); + __m128i cospi_m49_p15 = pair_set_epi16(-cospi[49], cospi[15]); + __m128i cospi_p55_p09 = pair_set_epi16(cospi[55], cospi[9]); + __m128i cospi_m09_p55 = pair_set_epi16(-cospi[9], cospi[55]); + __m128i cospi_p23_p41 = pair_set_epi16(cospi[23], cospi[41]); + __m128i cospi_m41_p23 = pair_set_epi16(-cospi[41], cospi[23]); + __m128i cospi_p39_p25 = pair_set_epi16(cospi[39], cospi[25]); + __m128i cospi_m25_p39 = pair_set_epi16(-cospi[25], cospi[39]); + __m128i cospi_p07_p57 = pair_set_epi16(cospi[7], cospi[57]); + __m128i cospi_m57_p07 = pair_set_epi16(-cospi[57], cospi[7]); + __m128i cospi_p59_p05 = pair_set_epi16(cospi[59], cospi[5]); + __m128i cospi_m05_p59 = pair_set_epi16(-cospi[5], cospi[59]); + __m128i cospi_p27_p37 = pair_set_epi16(cospi[27], cospi[37]); + __m128i cospi_m37_p27 = pair_set_epi16(-cospi[37], cospi[27]); + __m128i cospi_p43_p21 = pair_set_epi16(cospi[43], cospi[21]); + __m128i cospi_m21_p43 = pair_set_epi16(-cospi[21], cospi[43]); + __m128i cospi_p11_p53 = pair_set_epi16(cospi[11], cospi[53]); + __m128i cospi_m53_p11 = pair_set_epi16(-cospi[53], cospi[11]); + __m128i cospi_p51_p13 = pair_set_epi16(cospi[51], cospi[13]); + __m128i cospi_m13_p51 = pair_set_epi16(-cospi[13], cospi[51]); + __m128i cospi_p19_p45 = pair_set_epi16(cospi[19], cospi[45]); + __m128i cospi_m45_p19 = pair_set_epi16(-cospi[45], cospi[19]); + __m128i cospi_p35_p29 = pair_set_epi16(cospi[35], cospi[29]); + __m128i cospi_m29_p35 = pair_set_epi16(-cospi[29], cospi[35]); + __m128i cospi_p03_p61 = pair_set_epi16(cospi[3], cospi[61]); + __m128i cospi_m61_p03 = pair_set_epi16(-cospi[61], cospi[3]); + + // stage 1 + __m128i x1[64]; + x1[0] = _mm_adds_epi16(input[0], input[63]); + x1[63] = _mm_subs_epi16(input[0], input[63]); + x1[1] = _mm_adds_epi16(input[1], input[62]); + x1[62] = _mm_subs_epi16(input[1], input[62]); + x1[2] = _mm_adds_epi16(input[2], input[61]); + x1[61] = _mm_subs_epi16(input[2], input[61]); + x1[3] = _mm_adds_epi16(input[3], input[60]); + x1[60] = _mm_subs_epi16(input[3], input[60]); + x1[4] = _mm_adds_epi16(input[4], input[59]); + x1[59] = _mm_subs_epi16(input[4], input[59]); + x1[5] = _mm_adds_epi16(input[5], input[58]); + x1[58] = _mm_subs_epi16(input[5], input[58]); + x1[6] = _mm_adds_epi16(input[6], input[57]); + x1[57] = _mm_subs_epi16(input[6], input[57]); + x1[7] = _mm_adds_epi16(input[7], input[56]); + x1[56] = _mm_subs_epi16(input[7], input[56]); + x1[8] = _mm_adds_epi16(input[8], input[55]); + x1[55] = _mm_subs_epi16(input[8], input[55]); + x1[9] = _mm_adds_epi16(input[9], input[54]); + x1[54] = _mm_subs_epi16(input[9], input[54]); + x1[10] = _mm_adds_epi16(input[10], input[53]); + x1[53] = _mm_subs_epi16(input[10], input[53]); + x1[11] = _mm_adds_epi16(input[11], input[52]); + x1[52] = _mm_subs_epi16(input[11], input[52]); + x1[12] = _mm_adds_epi16(input[12], input[51]); + x1[51] = _mm_subs_epi16(input[12], input[51]); + x1[13] = _mm_adds_epi16(input[13], input[50]); + x1[50] = _mm_subs_epi16(input[13], input[50]); + x1[14] = _mm_adds_epi16(input[14], input[49]); + x1[49] = _mm_subs_epi16(input[14], input[49]); + x1[15] = _mm_adds_epi16(input[15], input[48]); + x1[48] = _mm_subs_epi16(input[15], input[48]); + x1[16] = _mm_adds_epi16(input[16], input[47]); + x1[47] = _mm_subs_epi16(input[16], input[47]); + x1[17] = _mm_adds_epi16(input[17], input[46]); + x1[46] = _mm_subs_epi16(input[17], input[46]); + x1[18] = _mm_adds_epi16(input[18], input[45]); + x1[45] = _mm_subs_epi16(input[18], input[45]); + x1[19] = _mm_adds_epi16(input[19], input[44]); + x1[44] = _mm_subs_epi16(input[19], input[44]); + x1[20] = _mm_adds_epi16(input[20], input[43]); + x1[43] = _mm_subs_epi16(input[20], input[43]); + x1[21] = _mm_adds_epi16(input[21], input[42]); + x1[42] = _mm_subs_epi16(input[21], input[42]); + x1[22] = _mm_adds_epi16(input[22], input[41]); + x1[41] = _mm_subs_epi16(input[22], input[41]); + x1[23] = _mm_adds_epi16(input[23], input[40]); + x1[40] = _mm_subs_epi16(input[23], input[40]); + x1[24] = _mm_adds_epi16(input[24], input[39]); + x1[39] = _mm_subs_epi16(input[24], input[39]); + x1[25] = _mm_adds_epi16(input[25], input[38]); + x1[38] = _mm_subs_epi16(input[25], input[38]); + x1[26] = _mm_adds_epi16(input[26], input[37]); + x1[37] = _mm_subs_epi16(input[26], input[37]); + x1[27] = _mm_adds_epi16(input[27], input[36]); + x1[36] = _mm_subs_epi16(input[27], input[36]); + x1[28] = _mm_adds_epi16(input[28], input[35]); + x1[35] = _mm_subs_epi16(input[28], input[35]); + x1[29] = _mm_adds_epi16(input[29], input[34]); + x1[34] = _mm_subs_epi16(input[29], input[34]); + x1[30] = _mm_adds_epi16(input[30], input[33]); + x1[33] = _mm_subs_epi16(input[30], input[33]); + x1[31] = _mm_adds_epi16(input[31], input[32]); + x1[32] = _mm_subs_epi16(input[31], input[32]); + + // stage 2 + __m128i x2[64]; + x2[0] = _mm_adds_epi16(x1[0], x1[31]); + x2[31] = _mm_subs_epi16(x1[0], x1[31]); + x2[1] = _mm_adds_epi16(x1[1], x1[30]); + x2[30] = _mm_subs_epi16(x1[1], x1[30]); + x2[2] = _mm_adds_epi16(x1[2], x1[29]); + x2[29] = _mm_subs_epi16(x1[2], x1[29]); + x2[3] = _mm_adds_epi16(x1[3], x1[28]); + x2[28] = _mm_subs_epi16(x1[3], x1[28]); + x2[4] = _mm_adds_epi16(x1[4], x1[27]); + x2[27] = _mm_subs_epi16(x1[4], x1[27]); + x2[5] = _mm_adds_epi16(x1[5], x1[26]); + x2[26] = _mm_subs_epi16(x1[5], x1[26]); + x2[6] = _mm_adds_epi16(x1[6], x1[25]); + x2[25] = _mm_subs_epi16(x1[6], x1[25]); + x2[7] = _mm_adds_epi16(x1[7], x1[24]); + x2[24] = _mm_subs_epi16(x1[7], x1[24]); + x2[8] = _mm_adds_epi16(x1[8], x1[23]); + x2[23] = _mm_subs_epi16(x1[8], x1[23]); + x2[9] = _mm_adds_epi16(x1[9], x1[22]); + x2[22] = _mm_subs_epi16(x1[9], x1[22]); + x2[10] = _mm_adds_epi16(x1[10], x1[21]); + x2[21] = _mm_subs_epi16(x1[10], x1[21]); + x2[11] = _mm_adds_epi16(x1[11], x1[20]); + x2[20] = _mm_subs_epi16(x1[11], x1[20]); + x2[12] = _mm_adds_epi16(x1[12], x1[19]); + x2[19] = _mm_subs_epi16(x1[12], x1[19]); + x2[13] = _mm_adds_epi16(x1[13], x1[18]); + x2[18] = _mm_subs_epi16(x1[13], x1[18]); + x2[14] = _mm_adds_epi16(x1[14], x1[17]); + x2[17] = _mm_subs_epi16(x1[14], x1[17]); + x2[15] = _mm_adds_epi16(x1[15], x1[16]); + x2[16] = _mm_subs_epi16(x1[15], x1[16]); + x2[32] = x1[32]; + x2[33] = x1[33]; + x2[34] = x1[34]; + x2[35] = x1[35]; + x2[36] = x1[36]; + x2[37] = x1[37]; + x2[38] = x1[38]; + x2[39] = x1[39]; + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[40], x1[55], x2[40], x2[55]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[41], x1[54], x2[41], x2[54]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[42], x1[53], x2[42], x2[53]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[43], x1[52], x2[43], x2[52]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[44], x1[51], x2[44], x2[51]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[45], x1[50], x2[45], x2[50]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[46], x1[49], x2[46], x2[49]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x1[47], x1[48], x2[47], x2[48]); + x2[56] = x1[56]; + x2[57] = x1[57]; + x2[58] = x1[58]; + x2[59] = x1[59]; + x2[60] = x1[60]; + x2[61] = x1[61]; + x2[62] = x1[62]; + x2[63] = x1[63]; + + // stage 3 + __m128i x3[64]; + x3[0] = _mm_adds_epi16(x2[0], x2[15]); + x3[15] = _mm_subs_epi16(x2[0], x2[15]); + x3[1] = _mm_adds_epi16(x2[1], x2[14]); + x3[14] = _mm_subs_epi16(x2[1], x2[14]); + x3[2] = _mm_adds_epi16(x2[2], x2[13]); + x3[13] = _mm_subs_epi16(x2[2], x2[13]); + x3[3] = _mm_adds_epi16(x2[3], x2[12]); + x3[12] = _mm_subs_epi16(x2[3], x2[12]); + x3[4] = _mm_adds_epi16(x2[4], x2[11]); + x3[11] = _mm_subs_epi16(x2[4], x2[11]); + x3[5] = _mm_adds_epi16(x2[5], x2[10]); + x3[10] = _mm_subs_epi16(x2[5], x2[10]); + x3[6] = _mm_adds_epi16(x2[6], x2[9]); + x3[9] = _mm_subs_epi16(x2[6], x2[9]); + x3[7] = _mm_adds_epi16(x2[7], x2[8]); + x3[8] = _mm_subs_epi16(x2[7], x2[8]); + x3[16] = x2[16]; + x3[17] = x2[17]; + x3[18] = x2[18]; + x3[19] = x2[19]; + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x2[20], x2[27], x3[20], x3[27]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x2[21], x2[26], x3[21], x3[26]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x2[22], x2[25], x3[22], x3[25]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x2[23], x2[24], x3[23], x3[24]); + x3[28] = x2[28]; + x3[29] = x2[29]; + x3[30] = x2[30]; + x3[31] = x2[31]; + x3[32] = _mm_adds_epi16(x2[32], x2[47]); + x3[47] = _mm_subs_epi16(x2[32], x2[47]); + x3[33] = _mm_adds_epi16(x2[33], x2[46]); + x3[46] = _mm_subs_epi16(x2[33], x2[46]); + x3[34] = _mm_adds_epi16(x2[34], x2[45]); + x3[45] = _mm_subs_epi16(x2[34], x2[45]); + x3[35] = _mm_adds_epi16(x2[35], x2[44]); + x3[44] = _mm_subs_epi16(x2[35], x2[44]); + x3[36] = _mm_adds_epi16(x2[36], x2[43]); + x3[43] = _mm_subs_epi16(x2[36], x2[43]); + x3[37] = _mm_adds_epi16(x2[37], x2[42]); + x3[42] = _mm_subs_epi16(x2[37], x2[42]); + x3[38] = _mm_adds_epi16(x2[38], x2[41]); + x3[41] = _mm_subs_epi16(x2[38], x2[41]); + x3[39] = _mm_adds_epi16(x2[39], x2[40]); + x3[40] = _mm_subs_epi16(x2[39], x2[40]); + x3[48] = _mm_subs_epi16(x2[63], x2[48]); + x3[63] = _mm_adds_epi16(x2[63], x2[48]); + x3[49] = _mm_subs_epi16(x2[62], x2[49]); + x3[62] = _mm_adds_epi16(x2[62], x2[49]); + x3[50] = _mm_subs_epi16(x2[61], x2[50]); + x3[61] = _mm_adds_epi16(x2[61], x2[50]); + x3[51] = _mm_subs_epi16(x2[60], x2[51]); + x3[60] = _mm_adds_epi16(x2[60], x2[51]); + x3[52] = _mm_subs_epi16(x2[59], x2[52]); + x3[59] = _mm_adds_epi16(x2[59], x2[52]); + x3[53] = _mm_subs_epi16(x2[58], x2[53]); + x3[58] = _mm_adds_epi16(x2[58], x2[53]); + x3[54] = _mm_subs_epi16(x2[57], x2[54]); + x3[57] = _mm_adds_epi16(x2[57], x2[54]); + x3[55] = _mm_subs_epi16(x2[56], x2[55]); + x3[56] = _mm_adds_epi16(x2[56], x2[55]); + + // stage 4 + __m128i x4[64]; + x4[0] = _mm_adds_epi16(x3[0], x3[7]); + x4[7] = _mm_subs_epi16(x3[0], x3[7]); + x4[1] = _mm_adds_epi16(x3[1], x3[6]); + x4[6] = _mm_subs_epi16(x3[1], x3[6]); + x4[2] = _mm_adds_epi16(x3[2], x3[5]); + x4[5] = _mm_subs_epi16(x3[2], x3[5]); + x4[3] = _mm_adds_epi16(x3[3], x3[4]); + x4[4] = _mm_subs_epi16(x3[3], x3[4]); + x4[8] = x3[8]; + x4[9] = x3[9]; + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x3[10], x3[13], x4[10], x4[13]); + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x3[11], x3[12], x4[11], x4[12]); + x4[14] = x3[14]; + x4[15] = x3[15]; + x4[16] = _mm_adds_epi16(x3[16], x3[23]); + x4[23] = _mm_subs_epi16(x3[16], x3[23]); + x4[17] = _mm_adds_epi16(x3[17], x3[22]); + x4[22] = _mm_subs_epi16(x3[17], x3[22]); + x4[18] = _mm_adds_epi16(x3[18], x3[21]); + x4[21] = _mm_subs_epi16(x3[18], x3[21]); + x4[19] = _mm_adds_epi16(x3[19], x3[20]); + x4[20] = _mm_subs_epi16(x3[19], x3[20]); + x4[24] = _mm_subs_epi16(x3[31], x3[24]); + x4[31] = _mm_adds_epi16(x3[31], x3[24]); + x4[25] = _mm_subs_epi16(x3[30], x3[25]); + x4[30] = _mm_adds_epi16(x3[30], x3[25]); + x4[26] = _mm_subs_epi16(x3[29], x3[26]); + x4[29] = _mm_adds_epi16(x3[29], x3[26]); + x4[27] = _mm_subs_epi16(x3[28], x3[27]); + x4[28] = _mm_adds_epi16(x3[28], x3[27]); + x4[32] = x3[32]; + x4[33] = x3[33]; + x4[34] = x3[34]; + x4[35] = x3[35]; + btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x3[36], x3[59], x4[36], x4[59]); + btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x3[37], x3[58], x4[37], x4[58]); + btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x3[38], x3[57], x4[38], x4[57]); + btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x3[39], x3[56], x4[39], x4[56]); + btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x3[40], x3[55], x4[40], x4[55]); + btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x3[41], x3[54], x4[41], x4[54]); + btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x3[42], x3[53], x4[42], x4[53]); + btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x3[43], x3[52], x4[43], x4[52]); + x4[44] = x3[44]; + x4[45] = x3[45]; + x4[46] = x3[46]; + x4[47] = x3[47]; + x4[48] = x3[48]; + x4[49] = x3[49]; + x4[50] = x3[50]; + x4[51] = x3[51]; + x4[60] = x3[60]; + x4[61] = x3[61]; + x4[62] = x3[62]; + x4[63] = x3[63]; + + // stage 5 + __m128i x5[64]; + x5[0] = _mm_adds_epi16(x4[0], x4[3]); + x5[3] = _mm_subs_epi16(x4[0], x4[3]); + x5[1] = _mm_adds_epi16(x4[1], x4[2]); + x5[2] = _mm_subs_epi16(x4[1], x4[2]); + x5[4] = x4[4]; + btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x4[5], x4[6], x5[5], x5[6]); + x5[7] = x4[7]; + x5[8] = _mm_adds_epi16(x4[8], x4[11]); + x5[11] = _mm_subs_epi16(x4[8], x4[11]); + x5[9] = _mm_adds_epi16(x4[9], x4[10]); + x5[10] = _mm_subs_epi16(x4[9], x4[10]); + x5[12] = _mm_subs_epi16(x4[15], x4[12]); + x5[15] = _mm_adds_epi16(x4[15], x4[12]); + x5[13] = _mm_subs_epi16(x4[14], x4[13]); + x5[14] = _mm_adds_epi16(x4[14], x4[13]); + x5[16] = x4[16]; + x5[17] = x4[17]; + btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x4[18], x4[29], x5[18], x5[29]); + btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x4[19], x4[28], x5[19], x5[28]); + btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x4[20], x4[27], x5[20], x5[27]); + btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x4[21], x4[26], x5[21], x5[26]); + x5[22] = x4[22]; + x5[23] = x4[23]; + x5[24] = x4[24]; + x5[25] = x4[25]; + x5[30] = x4[30]; + x5[31] = x4[31]; + x5[32] = _mm_adds_epi16(x4[32], x4[39]); + x5[39] = _mm_subs_epi16(x4[32], x4[39]); + x5[33] = _mm_adds_epi16(x4[33], x4[38]); + x5[38] = _mm_subs_epi16(x4[33], x4[38]); + x5[34] = _mm_adds_epi16(x4[34], x4[37]); + x5[37] = _mm_subs_epi16(x4[34], x4[37]); + x5[35] = _mm_adds_epi16(x4[35], x4[36]); + x5[36] = _mm_subs_epi16(x4[35], x4[36]); + x5[40] = _mm_subs_epi16(x4[47], x4[40]); + x5[47] = _mm_adds_epi16(x4[47], x4[40]); + x5[41] = _mm_subs_epi16(x4[46], x4[41]); + x5[46] = _mm_adds_epi16(x4[46], x4[41]); + x5[42] = _mm_subs_epi16(x4[45], x4[42]); + x5[45] = _mm_adds_epi16(x4[45], x4[42]); + x5[43] = _mm_subs_epi16(x4[44], x4[43]); + x5[44] = _mm_adds_epi16(x4[44], x4[43]); + x5[48] = _mm_adds_epi16(x4[48], x4[55]); + x5[55] = _mm_subs_epi16(x4[48], x4[55]); + x5[49] = _mm_adds_epi16(x4[49], x4[54]); + x5[54] = _mm_subs_epi16(x4[49], x4[54]); + x5[50] = _mm_adds_epi16(x4[50], x4[53]); + x5[53] = _mm_subs_epi16(x4[50], x4[53]); + x5[51] = _mm_adds_epi16(x4[51], x4[52]); + x5[52] = _mm_subs_epi16(x4[51], x4[52]); + x5[56] = _mm_subs_epi16(x4[63], x4[56]); + x5[63] = _mm_adds_epi16(x4[63], x4[56]); + x5[57] = _mm_subs_epi16(x4[62], x4[57]); + x5[62] = _mm_adds_epi16(x4[62], x4[57]); + x5[58] = _mm_subs_epi16(x4[61], x4[58]); + x5[61] = _mm_adds_epi16(x4[61], x4[58]); + x5[59] = _mm_subs_epi16(x4[60], x4[59]); + x5[60] = _mm_adds_epi16(x4[60], x4[59]); + + // stage 6 + __m128i x6[64]; + btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x5[0], x5[1], x6[0], x6[1]); + btf_16_sse2(cospi_p48_p16, cospi_m16_p48, x5[2], x5[3], x6[2], x6[3]); + x6[4] = _mm_adds_epi16(x5[4], x5[5]); + x6[5] = _mm_subs_epi16(x5[4], x5[5]); + x6[6] = _mm_subs_epi16(x5[7], x5[6]); + x6[7] = _mm_adds_epi16(x5[7], x5[6]); + x6[8] = x5[8]; + btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x5[9], x5[14], x6[9], x6[14]); + btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x5[10], x5[13], x6[10], x6[13]); + x6[11] = x5[11]; + x6[12] = x5[12]; + x6[15] = x5[15]; + x6[16] = _mm_adds_epi16(x5[16], x5[19]); + x6[19] = _mm_subs_epi16(x5[16], x5[19]); + x6[17] = _mm_adds_epi16(x5[17], x5[18]); + x6[18] = _mm_subs_epi16(x5[17], x5[18]); + x6[20] = _mm_subs_epi16(x5[23], x5[20]); + x6[23] = _mm_adds_epi16(x5[23], x5[20]); + x6[21] = _mm_subs_epi16(x5[22], x5[21]); + x6[22] = _mm_adds_epi16(x5[22], x5[21]); + x6[24] = _mm_adds_epi16(x5[24], x5[27]); + x6[27] = _mm_subs_epi16(x5[24], x5[27]); + x6[25] = _mm_adds_epi16(x5[25], x5[26]); + x6[26] = _mm_subs_epi16(x5[25], x5[26]); + x6[28] = _mm_subs_epi16(x5[31], x5[28]); + x6[31] = _mm_adds_epi16(x5[31], x5[28]); + x6[29] = _mm_subs_epi16(x5[30], x5[29]); + x6[30] = _mm_adds_epi16(x5[30], x5[29]); + x6[32] = x5[32]; + x6[33] = x5[33]; + btf_16_sse2(cospi_m08_p56, cospi_p56_p08, x5[34], x5[61], x6[34], x6[61]); + btf_16_sse2(cospi_m08_p56, cospi_p56_p08, x5[35], x5[60], x6[35], x6[60]); + btf_16_sse2(cospi_m56_m08, cospi_m08_p56, x5[36], x5[59], x6[36], x6[59]); + btf_16_sse2(cospi_m56_m08, cospi_m08_p56, x5[37], x5[58], x6[37], x6[58]); + x6[38] = x5[38]; + x6[39] = x5[39]; + x6[40] = x5[40]; + x6[41] = x5[41]; + btf_16_sse2(cospi_m40_p24, cospi_p24_p40, x5[42], x5[53], x6[42], x6[53]); + btf_16_sse2(cospi_m40_p24, cospi_p24_p40, x5[43], x5[52], x6[43], x6[52]); + btf_16_sse2(cospi_m24_m40, cospi_m40_p24, x5[44], x5[51], x6[44], x6[51]); + btf_16_sse2(cospi_m24_m40, cospi_m40_p24, x5[45], x5[50], x6[45], x6[50]); + x6[46] = x5[46]; + x6[47] = x5[47]; + x6[48] = x5[48]; + x6[49] = x5[49]; + x6[54] = x5[54]; + x6[55] = x5[55]; + x6[56] = x5[56]; + x6[57] = x5[57]; + x6[62] = x5[62]; + x6[63] = x5[63]; + + // stage 7 + __m128i x7[64]; + x7[0] = x6[0]; + x7[1] = x6[1]; + x7[2] = x6[2]; + x7[3] = x6[3]; + btf_16_sse2(cospi_p56_p08, cospi_m08_p56, x6[4], x6[7], x7[4], x7[7]); + btf_16_sse2(cospi_p24_p40, cospi_m40_p24, x6[5], x6[6], x7[5], x7[6]); + x7[8] = _mm_adds_epi16(x6[8], x6[9]); + x7[9] = _mm_subs_epi16(x6[8], x6[9]); + x7[10] = _mm_subs_epi16(x6[11], x6[10]); + x7[11] = _mm_adds_epi16(x6[11], x6[10]); + x7[12] = _mm_adds_epi16(x6[12], x6[13]); + x7[13] = _mm_subs_epi16(x6[12], x6[13]); + x7[14] = _mm_subs_epi16(x6[15], x6[14]); + x7[15] = _mm_adds_epi16(x6[15], x6[14]); + x7[16] = x6[16]; + btf_16_sse2(cospi_m08_p56, cospi_p56_p08, x6[17], x6[30], x7[17], x7[30]); + btf_16_sse2(cospi_m56_m08, cospi_m08_p56, x6[18], x6[29], x7[18], x7[29]); + x7[19] = x6[19]; + x7[20] = x6[20]; + btf_16_sse2(cospi_m40_p24, cospi_p24_p40, x6[21], x6[26], x7[21], x7[26]); + btf_16_sse2(cospi_m24_m40, cospi_m40_p24, x6[22], x6[25], x7[22], x7[25]); + x7[23] = x6[23]; + x7[24] = x6[24]; + x7[27] = x6[27]; + x7[28] = x6[28]; + x7[31] = x6[31]; + x7[32] = _mm_adds_epi16(x6[32], x6[35]); + x7[35] = _mm_subs_epi16(x6[32], x6[35]); + x7[33] = _mm_adds_epi16(x6[33], x6[34]); + x7[34] = _mm_subs_epi16(x6[33], x6[34]); + x7[36] = _mm_subs_epi16(x6[39], x6[36]); + x7[39] = _mm_adds_epi16(x6[39], x6[36]); + x7[37] = _mm_subs_epi16(x6[38], x6[37]); + x7[38] = _mm_adds_epi16(x6[38], x6[37]); + x7[40] = _mm_adds_epi16(x6[40], x6[43]); + x7[43] = _mm_subs_epi16(x6[40], x6[43]); + x7[41] = _mm_adds_epi16(x6[41], x6[42]); + x7[42] = _mm_subs_epi16(x6[41], x6[42]); + x7[44] = _mm_subs_epi16(x6[47], x6[44]); + x7[47] = _mm_adds_epi16(x6[47], x6[44]); + x7[45] = _mm_subs_epi16(x6[46], x6[45]); + x7[46] = _mm_adds_epi16(x6[46], x6[45]); + x7[48] = _mm_adds_epi16(x6[48], x6[51]); + x7[51] = _mm_subs_epi16(x6[48], x6[51]); + x7[49] = _mm_adds_epi16(x6[49], x6[50]); + x7[50] = _mm_subs_epi16(x6[49], x6[50]); + x7[52] = _mm_subs_epi16(x6[55], x6[52]); + x7[55] = _mm_adds_epi16(x6[55], x6[52]); + x7[53] = _mm_subs_epi16(x6[54], x6[53]); + x7[54] = _mm_adds_epi16(x6[54], x6[53]); + x7[56] = _mm_adds_epi16(x6[56], x6[59]); + x7[59] = _mm_subs_epi16(x6[56], x6[59]); + x7[57] = _mm_adds_epi16(x6[57], x6[58]); + x7[58] = _mm_subs_epi16(x6[57], x6[58]); + x7[60] = _mm_subs_epi16(x6[63], x6[60]); + x7[63] = _mm_adds_epi16(x6[63], x6[60]); + x7[61] = _mm_subs_epi16(x6[62], x6[61]); + x7[62] = _mm_adds_epi16(x6[62], x6[61]); + + // stage 8 + __m128i x8[64]; + x8[0] = x7[0]; + x8[1] = x7[1]; + x8[2] = x7[2]; + x8[3] = x7[3]; + x8[4] = x7[4]; + x8[5] = x7[5]; + x8[6] = x7[6]; + x8[7] = x7[7]; + btf_16_sse2(cospi_p60_p04, cospi_m04_p60, x7[8], x7[15], x8[8], x8[15]); + btf_16_sse2(cospi_p28_p36, cospi_m36_p28, x7[9], x7[14], x8[9], x8[14]); + btf_16_sse2(cospi_p44_p20, cospi_m20_p44, x7[10], x7[13], x8[10], x8[13]); + btf_16_sse2(cospi_p12_p52, cospi_m52_p12, x7[11], x7[12], x8[11], x8[12]); + x8[16] = _mm_adds_epi16(x7[16], x7[17]); + x8[17] = _mm_subs_epi16(x7[16], x7[17]); + x8[18] = _mm_subs_epi16(x7[19], x7[18]); + x8[19] = _mm_adds_epi16(x7[19], x7[18]); + x8[20] = _mm_adds_epi16(x7[20], x7[21]); + x8[21] = _mm_subs_epi16(x7[20], x7[21]); + x8[22] = _mm_subs_epi16(x7[23], x7[22]); + x8[23] = _mm_adds_epi16(x7[23], x7[22]); + x8[24] = _mm_adds_epi16(x7[24], x7[25]); + x8[25] = _mm_subs_epi16(x7[24], x7[25]); + x8[26] = _mm_subs_epi16(x7[27], x7[26]); + x8[27] = _mm_adds_epi16(x7[27], x7[26]); + x8[28] = _mm_adds_epi16(x7[28], x7[29]); + x8[29] = _mm_subs_epi16(x7[28], x7[29]); + x8[30] = _mm_subs_epi16(x7[31], x7[30]); + x8[31] = _mm_adds_epi16(x7[31], x7[30]); + x8[32] = x7[32]; + btf_16_sse2(cospi_m04_p60, cospi_p60_p04, x7[33], x7[62], x8[33], x8[62]); + btf_16_sse2(cospi_m60_m04, cospi_m04_p60, x7[34], x7[61], x8[34], x8[61]); + x8[35] = x7[35]; + x8[36] = x7[36]; + btf_16_sse2(cospi_m36_p28, cospi_p28_p36, x7[37], x7[58], x8[37], x8[58]); + btf_16_sse2(cospi_m28_m36, cospi_m36_p28, x7[38], x7[57], x8[38], x8[57]); + x8[39] = x7[39]; + x8[40] = x7[40]; + btf_16_sse2(cospi_m20_p44, cospi_p44_p20, x7[41], x7[54], x8[41], x8[54]); + btf_16_sse2(cospi_m44_m20, cospi_m20_p44, x7[42], x7[53], x8[42], x8[53]); + x8[43] = x7[43]; + x8[44] = x7[44]; + btf_16_sse2(cospi_m52_p12, cospi_p12_p52, x7[45], x7[50], x8[45], x8[50]); + btf_16_sse2(cospi_m12_m52, cospi_m52_p12, x7[46], x7[49], x8[46], x8[49]); + x8[47] = x7[47]; + x8[48] = x7[48]; + x8[51] = x7[51]; + x8[52] = x7[52]; + x8[55] = x7[55]; + x8[56] = x7[56]; + x8[59] = x7[59]; + x8[60] = x7[60]; + x8[63] = x7[63]; + + // stage 9 + __m128i x9[64]; + x9[0] = x8[0]; + x9[1] = x8[1]; + x9[2] = x8[2]; + x9[3] = x8[3]; + x9[4] = x8[4]; + x9[5] = x8[5]; + x9[6] = x8[6]; + x9[7] = x8[7]; + x9[8] = x8[8]; + x9[9] = x8[9]; + x9[10] = x8[10]; + x9[11] = x8[11]; + x9[12] = x8[12]; + x9[13] = x8[13]; + x9[14] = x8[14]; + x9[15] = x8[15]; + btf_16_sse2(cospi_p62_p02, cospi_m02_p62, x8[16], x8[31], x9[16], x9[31]); + btf_16_sse2(cospi_p30_p34, cospi_m34_p30, x8[17], x8[30], x9[17], x9[30]); + btf_16_sse2(cospi_p46_p18, cospi_m18_p46, x8[18], x8[29], x9[18], x9[29]); + btf_16_sse2(cospi_p14_p50, cospi_m50_p14, x8[19], x8[28], x9[19], x9[28]); + btf_16_sse2(cospi_p54_p10, cospi_m10_p54, x8[20], x8[27], x9[20], x9[27]); + btf_16_sse2(cospi_p22_p42, cospi_m42_p22, x8[21], x8[26], x9[21], x9[26]); + btf_16_sse2(cospi_p38_p26, cospi_m26_p38, x8[22], x8[25], x9[22], x9[25]); + btf_16_sse2(cospi_p06_p58, cospi_m58_p06, x8[23], x8[24], x9[23], x9[24]); + x9[32] = _mm_adds_epi16(x8[32], x8[33]); + x9[33] = _mm_subs_epi16(x8[32], x8[33]); + x9[34] = _mm_subs_epi16(x8[35], x8[34]); + x9[35] = _mm_adds_epi16(x8[35], x8[34]); + x9[36] = _mm_adds_epi16(x8[36], x8[37]); + x9[37] = _mm_subs_epi16(x8[36], x8[37]); + x9[38] = _mm_subs_epi16(x8[39], x8[38]); + x9[39] = _mm_adds_epi16(x8[39], x8[38]); + x9[40] = _mm_adds_epi16(x8[40], x8[41]); + x9[41] = _mm_subs_epi16(x8[40], x8[41]); + x9[42] = _mm_subs_epi16(x8[43], x8[42]); + x9[43] = _mm_adds_epi16(x8[43], x8[42]); + x9[44] = _mm_adds_epi16(x8[44], x8[45]); + x9[45] = _mm_subs_epi16(x8[44], x8[45]); + x9[46] = _mm_subs_epi16(x8[47], x8[46]); + x9[47] = _mm_adds_epi16(x8[47], x8[46]); + x9[48] = _mm_adds_epi16(x8[48], x8[49]); + x9[49] = _mm_subs_epi16(x8[48], x8[49]); + x9[50] = _mm_subs_epi16(x8[51], x8[50]); + x9[51] = _mm_adds_epi16(x8[51], x8[50]); + x9[52] = _mm_adds_epi16(x8[52], x8[53]); + x9[53] = _mm_subs_epi16(x8[52], x8[53]); + x9[54] = _mm_subs_epi16(x8[55], x8[54]); + x9[55] = _mm_adds_epi16(x8[55], x8[54]); + x9[56] = _mm_adds_epi16(x8[56], x8[57]); + x9[57] = _mm_subs_epi16(x8[56], x8[57]); + x9[58] = _mm_subs_epi16(x8[59], x8[58]); + x9[59] = _mm_adds_epi16(x8[59], x8[58]); + x9[60] = _mm_adds_epi16(x8[60], x8[61]); + x9[61] = _mm_subs_epi16(x8[60], x8[61]); + x9[62] = _mm_subs_epi16(x8[63], x8[62]); + x9[63] = _mm_adds_epi16(x8[63], x8[62]); + + // stage 10 + __m128i x10[64]; + x10[0] = x9[0]; + x10[1] = x9[1]; + x10[2] = x9[2]; + x10[3] = x9[3]; + x10[4] = x9[4]; + x10[5] = x9[5]; + x10[6] = x9[6]; + x10[7] = x9[7]; + x10[8] = x9[8]; + x10[9] = x9[9]; + x10[10] = x9[10]; + x10[11] = x9[11]; + x10[12] = x9[12]; + x10[13] = x9[13]; + x10[14] = x9[14]; + x10[15] = x9[15]; + x10[16] = x9[16]; + x10[17] = x9[17]; + x10[18] = x9[18]; + x10[19] = x9[19]; + x10[20] = x9[20]; + x10[21] = x9[21]; + x10[22] = x9[22]; + x10[23] = x9[23]; + x10[24] = x9[24]; + x10[25] = x9[25]; + x10[26] = x9[26]; + x10[27] = x9[27]; + x10[28] = x9[28]; + x10[29] = x9[29]; + x10[30] = x9[30]; + x10[31] = x9[31]; + btf_16_sse2(cospi_p63_p01, cospi_m01_p63, x9[32], x9[63], x10[32], x10[63]); + btf_16_sse2(cospi_p31_p33, cospi_m33_p31, x9[33], x9[62], x10[33], x10[62]); + btf_16_sse2(cospi_p47_p17, cospi_m17_p47, x9[34], x9[61], x10[34], x10[61]); + btf_16_sse2(cospi_p15_p49, cospi_m49_p15, x9[35], x9[60], x10[35], x10[60]); + btf_16_sse2(cospi_p55_p09, cospi_m09_p55, x9[36], x9[59], x10[36], x10[59]); + btf_16_sse2(cospi_p23_p41, cospi_m41_p23, x9[37], x9[58], x10[37], x10[58]); + btf_16_sse2(cospi_p39_p25, cospi_m25_p39, x9[38], x9[57], x10[38], x10[57]); + btf_16_sse2(cospi_p07_p57, cospi_m57_p07, x9[39], x9[56], x10[39], x10[56]); + btf_16_sse2(cospi_p59_p05, cospi_m05_p59, x9[40], x9[55], x10[40], x10[55]); + btf_16_sse2(cospi_p27_p37, cospi_m37_p27, x9[41], x9[54], x10[41], x10[54]); + btf_16_sse2(cospi_p43_p21, cospi_m21_p43, x9[42], x9[53], x10[42], x10[53]); + btf_16_sse2(cospi_p11_p53, cospi_m53_p11, x9[43], x9[52], x10[43], x10[52]); + btf_16_sse2(cospi_p51_p13, cospi_m13_p51, x9[44], x9[51], x10[44], x10[51]); + btf_16_sse2(cospi_p19_p45, cospi_m45_p19, x9[45], x9[50], x10[45], x10[50]); + btf_16_sse2(cospi_p35_p29, cospi_m29_p35, x9[46], x9[49], x10[46], x10[49]); + btf_16_sse2(cospi_p03_p61, cospi_m61_p03, x9[47], x9[48], x10[47], x10[48]); + + // stage 11 + output[0] = x10[0]; + output[1] = x10[32]; + output[2] = x10[16]; + output[3] = x10[48]; + output[4] = x10[8]; + output[5] = x10[40]; + output[6] = x10[24]; + output[7] = x10[56]; + output[8] = x10[4]; + output[9] = x10[36]; + output[10] = x10[20]; + output[11] = x10[52]; + output[12] = x10[12]; + output[13] = x10[44]; + output[14] = x10[28]; + output[15] = x10[60]; + output[16] = x10[2]; + output[17] = x10[34]; + output[18] = x10[18]; + output[19] = x10[50]; + output[20] = x10[10]; + output[21] = x10[42]; + output[22] = x10[26]; + output[23] = x10[58]; + output[24] = x10[6]; + output[25] = x10[38]; + output[26] = x10[22]; + output[27] = x10[54]; + output[28] = x10[14]; + output[29] = x10[46]; + output[30] = x10[30]; + output[31] = x10[62]; + output[32] = x10[1]; + output[33] = x10[33]; + output[34] = x10[17]; + output[35] = x10[49]; + output[36] = x10[9]; + output[37] = x10[41]; + output[38] = x10[25]; + output[39] = x10[57]; + output[40] = x10[5]; + output[41] = x10[37]; + output[42] = x10[21]; + output[43] = x10[53]; + output[44] = x10[13]; + output[45] = x10[45]; + output[46] = x10[29]; + output[47] = x10[61]; + output[48] = x10[3]; + output[49] = x10[35]; + output[50] = x10[19]; + output[51] = x10[51]; + output[52] = x10[11]; + output[53] = x10[43]; + output[54] = x10[27]; + output[55] = x10[59]; + output[56] = x10[7]; + output[57] = x10[39]; + output[58] = x10[23]; + output[59] = x10[55]; + output[60] = x10[15]; + output[61] = x10[47]; + output[62] = x10[31]; + output[63] = x10[63]; +} + +static void fadst4x4_new_sse2(const __m128i *input, __m128i *output, + int8_t cos_bit) { + const int32_t *sinpi = sinpi_arr(cos_bit); + const __m128i sinpi_p01_p02 = pair_set_epi16(sinpi[1], sinpi[2]); + const __m128i sinpi_p04_m01 = pair_set_epi16(sinpi[4], -sinpi[1]); + const __m128i sinpi_p03_p04 = pair_set_epi16(sinpi[3], sinpi[4]); + const __m128i sinpi_m03_p02 = pair_set_epi16(-sinpi[3], sinpi[2]); + const __m128i sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi[3]); + const __m128i __zero = _mm_set1_epi16(0); + const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1)); + const __m128i in7 = _mm_add_epi16(input[0], input[1]); + __m128i u[8], v[8]; + + u[0] = _mm_unpacklo_epi16(input[0], input[1]); + u[1] = _mm_unpacklo_epi16(input[2], input[3]); + u[2] = _mm_unpacklo_epi16(in7, __zero); + u[3] = _mm_unpacklo_epi16(input[2], __zero); + u[4] = _mm_unpacklo_epi16(input[3], __zero); + + v[0] = _mm_madd_epi16(u[0], sinpi_p01_p02); // s0 + s2 + v[1] = _mm_madd_epi16(u[1], sinpi_p03_p04); // s4 + s5 + v[2] = _mm_madd_epi16(u[2], sinpi_p03_p03); // x1 + v[3] = _mm_madd_epi16(u[0], sinpi_p04_m01); // s1 - s3 + v[4] = _mm_madd_epi16(u[1], sinpi_m03_p02); // -s4 + s6 + v[5] = _mm_madd_epi16(u[3], sinpi_p03_p03); // s4 + v[6] = _mm_madd_epi16(u[4], sinpi_p03_p03); + + u[0] = _mm_add_epi32(v[0], v[1]); + u[1] = _mm_sub_epi32(v[2], v[6]); + u[2] = _mm_add_epi32(v[3], v[4]); + u[3] = _mm_sub_epi32(u[2], u[0]); + u[4] = _mm_slli_epi32(v[5], 2); + u[5] = _mm_sub_epi32(u[4], v[5]); + u[6] = _mm_add_epi32(u[3], u[5]); + + v[0] = _mm_add_epi32(u[0], __rounding); + v[1] = _mm_add_epi32(u[1], __rounding); + v[2] = _mm_add_epi32(u[2], __rounding); + v[3] = _mm_add_epi32(u[6], __rounding); + + u[0] = _mm_srai_epi32(v[0], cos_bit); + u[1] = _mm_srai_epi32(v[1], cos_bit); + u[2] = _mm_srai_epi32(v[2], cos_bit); + u[3] = _mm_srai_epi32(v[3], cos_bit); + + output[0] = _mm_packs_epi32(u[0], u[2]); + output[1] = _mm_packs_epi32(u[1], u[3]); + output[2] = _mm_srli_si128(output[0], 8); + output[3] = _mm_srli_si128(output[1], 8); +} + +static void fadst4x8_new_sse2(const __m128i *input, __m128i *output, + int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m128i __zero = _mm_setzero_si128(); + const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1)); + + __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]); + __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]); + __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]); + __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]); + __m128i cospi_m48_p16 = pair_set_epi16(-cospi[48], cospi[16]); + __m128i cospi_p04_p60 = pair_set_epi16(cospi[4], cospi[60]); + __m128i cospi_p60_m04 = pair_set_epi16(cospi[60], -cospi[4]); + __m128i cospi_p20_p44 = pair_set_epi16(cospi[20], cospi[44]); + __m128i cospi_p44_m20 = pair_set_epi16(cospi[44], -cospi[20]); + __m128i cospi_p36_p28 = pair_set_epi16(cospi[36], cospi[28]); + __m128i cospi_p28_m36 = pair_set_epi16(cospi[28], -cospi[36]); + __m128i cospi_p52_p12 = pair_set_epi16(cospi[52], cospi[12]); + __m128i cospi_p12_m52 = pair_set_epi16(cospi[12], -cospi[52]); + + // stage 1 + __m128i x1[8]; + x1[0] = input[0]; + x1[1] = _mm_subs_epi16(__zero, input[7]); + x1[2] = _mm_subs_epi16(__zero, input[3]); + x1[3] = input[4]; + x1[4] = _mm_subs_epi16(__zero, input[1]); + x1[5] = input[6]; + x1[6] = input[2]; + x1[7] = _mm_subs_epi16(__zero, input[5]); + + // stage 2 + __m128i x2[8]; + x2[0] = x1[0]; + x2[1] = x1[1]; + btf_16_w4_sse2(&cospi_p32_p32, &cospi_p32_m32, __rounding, cos_bit, &x1[2], + &x1[3], &x2[2], &x2[3]); + x2[4] = x1[4]; + x2[5] = x1[5]; + btf_16_w4_sse2(&cospi_p32_p32, &cospi_p32_m32, __rounding, cos_bit, &x1[6], + &x1[7], &x2[6], &x2[7]); + + // stage 3 + __m128i x3[8]; + x3[0] = _mm_adds_epi16(x2[0], x2[2]); + x3[2] = _mm_subs_epi16(x2[0], x2[2]); + x3[1] = _mm_adds_epi16(x2[1], x2[3]); + x3[3] = _mm_subs_epi16(x2[1], x2[3]); + x3[4] = _mm_adds_epi16(x2[4], x2[6]); + x3[6] = _mm_subs_epi16(x2[4], x2[6]); + x3[5] = _mm_adds_epi16(x2[5], x2[7]); + x3[7] = _mm_subs_epi16(x2[5], x2[7]); + + // stage 4 + __m128i x4[8]; + x4[0] = x3[0]; + x4[1] = x3[1]; + x4[2] = x3[2]; + x4[3] = x3[3]; + btf_16_w4_sse2(&cospi_p16_p48, &cospi_p48_m16, __rounding, cos_bit, &x3[4], + &x3[5], &x4[4], &x4[5]); + btf_16_w4_sse2(&cospi_m48_p16, &cospi_p16_p48, __rounding, cos_bit, &x3[6], + &x3[7], &x4[6], &x4[7]); + + // stage 5 + __m128i x5[8]; + x5[0] = _mm_adds_epi16(x4[0], x4[4]); + x5[4] = _mm_subs_epi16(x4[0], x4[4]); + x5[1] = _mm_adds_epi16(x4[1], x4[5]); + x5[5] = _mm_subs_epi16(x4[1], x4[5]); + x5[2] = _mm_adds_epi16(x4[2], x4[6]); + x5[6] = _mm_subs_epi16(x4[2], x4[6]); + x5[3] = _mm_adds_epi16(x4[3], x4[7]); + x5[7] = _mm_subs_epi16(x4[3], x4[7]); + + // stage 6 + __m128i x6[8]; + btf_16_w4_sse2(&cospi_p04_p60, &cospi_p60_m04, __rounding, cos_bit, &x5[0], + &x5[1], &x6[0], &x6[1]); + btf_16_w4_sse2(&cospi_p20_p44, &cospi_p44_m20, __rounding, cos_bit, &x5[2], + &x5[3], &x6[2], &x6[3]); + btf_16_w4_sse2(&cospi_p36_p28, &cospi_p28_m36, __rounding, cos_bit, &x5[4], + &x5[5], &x6[4], &x6[5]); + btf_16_w4_sse2(&cospi_p52_p12, &cospi_p12_m52, __rounding, cos_bit, &x5[6], + &x5[7], &x6[6], &x6[7]); + + // stage 7 + output[0] = x6[1]; + output[1] = x6[6]; + output[2] = x6[3]; + output[3] = x6[4]; + output[4] = x6[5]; + output[5] = x6[2]; + output[6] = x6[7]; + output[7] = x6[0]; +} + +static void fadst8x4_new_sse2(const __m128i *input, __m128i *output, + int8_t cos_bit) { + const int32_t *sinpi = sinpi_arr(cos_bit); + const __m128i sinpi_p01_p02 = pair_set_epi16(sinpi[1], sinpi[2]); + const __m128i sinpi_p04_m01 = pair_set_epi16(sinpi[4], -sinpi[1]); + const __m128i sinpi_p03_p04 = pair_set_epi16(sinpi[3], sinpi[4]); + const __m128i sinpi_m03_p02 = pair_set_epi16(-sinpi[3], sinpi[2]); + const __m128i sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi[3]); + const __m128i __zero = _mm_set1_epi16(0); + const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1)); + const __m128i in7 = _mm_add_epi16(input[0], input[1]); + __m128i u_lo[8], u_hi[8], v_lo[8], v_hi[8]; + + u_lo[0] = _mm_unpacklo_epi16(input[0], input[1]); + u_hi[0] = _mm_unpackhi_epi16(input[0], input[1]); + u_lo[1] = _mm_unpacklo_epi16(input[2], input[3]); + u_hi[1] = _mm_unpackhi_epi16(input[2], input[3]); + u_lo[2] = _mm_unpacklo_epi16(in7, __zero); + u_hi[2] = _mm_unpackhi_epi16(in7, __zero); + u_lo[3] = _mm_unpacklo_epi16(input[2], __zero); + u_hi[3] = _mm_unpackhi_epi16(input[2], __zero); + u_lo[4] = _mm_unpacklo_epi16(input[3], __zero); + u_hi[4] = _mm_unpackhi_epi16(input[3], __zero); + + v_lo[0] = _mm_madd_epi16(u_lo[0], sinpi_p01_p02); // s0 + s2 + v_hi[0] = _mm_madd_epi16(u_hi[0], sinpi_p01_p02); // s0 + s2 + v_lo[1] = _mm_madd_epi16(u_lo[1], sinpi_p03_p04); // s4 + s5 + v_hi[1] = _mm_madd_epi16(u_hi[1], sinpi_p03_p04); // s4 + s5 + v_lo[2] = _mm_madd_epi16(u_lo[2], sinpi_p03_p03); // x1 + v_hi[2] = _mm_madd_epi16(u_hi[2], sinpi_p03_p03); // x1 + v_lo[3] = _mm_madd_epi16(u_lo[0], sinpi_p04_m01); // s1 - s3 + v_hi[3] = _mm_madd_epi16(u_hi[0], sinpi_p04_m01); // s1 - s3 + v_lo[4] = _mm_madd_epi16(u_lo[1], sinpi_m03_p02); // -s4 + s6 + v_hi[4] = _mm_madd_epi16(u_hi[1], sinpi_m03_p02); // -s4 + s6 + v_lo[5] = _mm_madd_epi16(u_lo[3], sinpi_p03_p03); // s4 + v_hi[5] = _mm_madd_epi16(u_hi[3], sinpi_p03_p03); // s4 + v_lo[6] = _mm_madd_epi16(u_lo[4], sinpi_p03_p03); + v_hi[6] = _mm_madd_epi16(u_hi[4], sinpi_p03_p03); + + u_lo[0] = _mm_add_epi32(v_lo[0], v_lo[1]); + u_hi[0] = _mm_add_epi32(v_hi[0], v_hi[1]); + u_lo[1] = _mm_sub_epi32(v_lo[2], v_lo[6]); + u_hi[1] = _mm_sub_epi32(v_hi[2], v_hi[6]); + u_lo[2] = _mm_add_epi32(v_lo[3], v_lo[4]); + u_hi[2] = _mm_add_epi32(v_hi[3], v_hi[4]); + u_lo[3] = _mm_sub_epi32(u_lo[2], u_lo[0]); + u_hi[3] = _mm_sub_epi32(u_hi[2], u_hi[0]); + u_lo[4] = _mm_slli_epi32(v_lo[5], 2); + u_hi[4] = _mm_slli_epi32(v_hi[5], 2); + u_lo[5] = _mm_sub_epi32(u_lo[4], v_lo[5]); + u_hi[5] = _mm_sub_epi32(u_hi[4], v_hi[5]); + u_lo[6] = _mm_add_epi32(u_lo[3], u_lo[5]); + u_hi[6] = _mm_add_epi32(u_hi[3], u_hi[5]); + + v_lo[0] = _mm_add_epi32(u_lo[0], __rounding); + v_hi[0] = _mm_add_epi32(u_hi[0], __rounding); + v_lo[1] = _mm_add_epi32(u_lo[1], __rounding); + v_hi[1] = _mm_add_epi32(u_hi[1], __rounding); + v_lo[2] = _mm_add_epi32(u_lo[2], __rounding); + v_hi[2] = _mm_add_epi32(u_hi[2], __rounding); + v_lo[3] = _mm_add_epi32(u_lo[6], __rounding); + v_hi[3] = _mm_add_epi32(u_hi[6], __rounding); + + u_lo[0] = _mm_srai_epi32(v_lo[0], cos_bit); + u_hi[0] = _mm_srai_epi32(v_hi[0], cos_bit); + u_lo[1] = _mm_srai_epi32(v_lo[1], cos_bit); + u_hi[1] = _mm_srai_epi32(v_hi[1], cos_bit); + u_lo[2] = _mm_srai_epi32(v_lo[2], cos_bit); + u_hi[2] = _mm_srai_epi32(v_hi[2], cos_bit); + u_lo[3] = _mm_srai_epi32(v_lo[3], cos_bit); + u_hi[3] = _mm_srai_epi32(v_hi[3], cos_bit); + + output[0] = _mm_packs_epi32(u_lo[0], u_hi[0]); + output[1] = _mm_packs_epi32(u_lo[1], u_hi[1]); + output[2] = _mm_packs_epi32(u_lo[2], u_hi[2]); + output[3] = _mm_packs_epi32(u_lo[3], u_hi[3]); +} + +static void fadst8x8_new_sse2(const __m128i *input, __m128i *output, + int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m128i __zero = _mm_setzero_si128(); + const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1)); + + __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]); + __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]); + __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]); + __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]); + __m128i cospi_m48_p16 = pair_set_epi16(-cospi[48], cospi[16]); + __m128i cospi_p04_p60 = pair_set_epi16(cospi[4], cospi[60]); + __m128i cospi_p60_m04 = pair_set_epi16(cospi[60], -cospi[4]); + __m128i cospi_p20_p44 = pair_set_epi16(cospi[20], cospi[44]); + __m128i cospi_p44_m20 = pair_set_epi16(cospi[44], -cospi[20]); + __m128i cospi_p36_p28 = pair_set_epi16(cospi[36], cospi[28]); + __m128i cospi_p28_m36 = pair_set_epi16(cospi[28], -cospi[36]); + __m128i cospi_p52_p12 = pair_set_epi16(cospi[52], cospi[12]); + __m128i cospi_p12_m52 = pair_set_epi16(cospi[12], -cospi[52]); + + // stage 1 + __m128i x1[8]; + x1[0] = input[0]; + x1[1] = _mm_subs_epi16(__zero, input[7]); + x1[2] = _mm_subs_epi16(__zero, input[3]); + x1[3] = input[4]; + x1[4] = _mm_subs_epi16(__zero, input[1]); + x1[5] = input[6]; + x1[6] = input[2]; + x1[7] = _mm_subs_epi16(__zero, input[5]); + + // stage 2 + __m128i x2[8]; + x2[0] = x1[0]; + x2[1] = x1[1]; + btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x1[2], x1[3], x2[2], x2[3]); + x2[4] = x1[4]; + x2[5] = x1[5]; + btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x1[6], x1[7], x2[6], x2[7]); + + // stage 3 + __m128i x3[8]; + x3[0] = _mm_adds_epi16(x2[0], x2[2]); + x3[2] = _mm_subs_epi16(x2[0], x2[2]); + x3[1] = _mm_adds_epi16(x2[1], x2[3]); + x3[3] = _mm_subs_epi16(x2[1], x2[3]); + x3[4] = _mm_adds_epi16(x2[4], x2[6]); + x3[6] = _mm_subs_epi16(x2[4], x2[6]); + x3[5] = _mm_adds_epi16(x2[5], x2[7]); + x3[7] = _mm_subs_epi16(x2[5], x2[7]); + + // stage 4 + __m128i x4[8]; + x4[0] = x3[0]; + x4[1] = x3[1]; + x4[2] = x3[2]; + x4[3] = x3[3]; + btf_16_sse2(cospi_p16_p48, cospi_p48_m16, x3[4], x3[5], x4[4], x4[5]); + btf_16_sse2(cospi_m48_p16, cospi_p16_p48, x3[6], x3[7], x4[6], x4[7]); + + // stage 5 + __m128i x5[8]; + x5[0] = _mm_adds_epi16(x4[0], x4[4]); + x5[4] = _mm_subs_epi16(x4[0], x4[4]); + x5[1] = _mm_adds_epi16(x4[1], x4[5]); + x5[5] = _mm_subs_epi16(x4[1], x4[5]); + x5[2] = _mm_adds_epi16(x4[2], x4[6]); + x5[6] = _mm_subs_epi16(x4[2], x4[6]); + x5[3] = _mm_adds_epi16(x4[3], x4[7]); + x5[7] = _mm_subs_epi16(x4[3], x4[7]); + + // stage 6 + __m128i x6[8]; + btf_16_sse2(cospi_p04_p60, cospi_p60_m04, x5[0], x5[1], x6[0], x6[1]); + btf_16_sse2(cospi_p20_p44, cospi_p44_m20, x5[2], x5[3], x6[2], x6[3]); + btf_16_sse2(cospi_p36_p28, cospi_p28_m36, x5[4], x5[5], x6[4], x6[5]); + btf_16_sse2(cospi_p52_p12, cospi_p12_m52, x5[6], x5[7], x6[6], x6[7]); + + // stage 7 + output[0] = x6[1]; + output[1] = x6[6]; + output[2] = x6[3]; + output[3] = x6[4]; + output[4] = x6[5]; + output[5] = x6[2]; + output[6] = x6[7]; + output[7] = x6[0]; +} + +static void fadst8x16_new_sse2(const __m128i *input, __m128i *output, + int8_t cos_bit) { + const int32_t *cospi = cospi_arr(cos_bit); + const __m128i __zero = _mm_setzero_si128(); + const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1)); + + __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]); + __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]); + __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]); + __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]); + __m128i cospi_m48_p16 = pair_set_epi16(-cospi[48], cospi[16]); + __m128i cospi_p08_p56 = pair_set_epi16(cospi[8], cospi[56]); + __m128i cospi_p56_m08 = pair_set_epi16(cospi[56], -cospi[8]); + __m128i cospi_p40_p24 = pair_set_epi16(cospi[40], cospi[24]); + __m128i cospi_p24_m40 = pair_set_epi16(cospi[24], -cospi[40]); + __m128i cospi_m56_p08 = pair_set_epi16(-cospi[56], cospi[8]); + __m128i cospi_m24_p40 = pair_set_epi16(-cospi[24], cospi[40]); + __m128i cospi_p02_p62 = pair_set_epi16(cospi[2], cospi[62]); + __m128i cospi_p62_m02 = pair_set_epi16(cospi[62], -cospi[2]); + __m128i cospi_p10_p54 = pair_set_epi16(cospi[10], cospi[54]); + __m128i cospi_p54_m10 = pair_set_epi16(cospi[54], -cospi[10]); + __m128i cospi_p18_p46 = pair_set_epi16(cospi[18], cospi[46]); + __m128i cospi_p46_m18 = pair_set_epi16(cospi[46], -cospi[18]); + __m128i cospi_p26_p38 = pair_set_epi16(cospi[26], cospi[38]); + __m128i cospi_p38_m26 = pair_set_epi16(cospi[38], -cospi[26]); + __m128i cospi_p34_p30 = pair_set_epi16(cospi[34], cospi[30]); + __m128i cospi_p30_m34 = pair_set_epi16(cospi[30], -cospi[34]); + __m128i cospi_p42_p22 = pair_set_epi16(cospi[42], cospi[22]); + __m128i cospi_p22_m42 = pair_set_epi16(cospi[22], -cospi[42]); + __m128i cospi_p50_p14 = pair_set_epi16(cospi[50], cospi[14]); + __m128i cospi_p14_m50 = pair_set_epi16(cospi[14], -cospi[50]); + __m128i cospi_p58_p06 = pair_set_epi16(cospi[58], cospi[6]); + __m128i cospi_p06_m58 = pair_set_epi16(cospi[6], -cospi[58]); + + // stage 1 + __m128i x1[16]; + x1[0] = input[0]; + x1[1] = _mm_subs_epi16(__zero, input[15]); + x1[2] = _mm_subs_epi16(__zero, input[7]); + x1[3] = input[8]; + x1[4] = _mm_subs_epi16(__zero, input[3]); + x1[5] = input[12]; + x1[6] = input[4]; + x1[7] = _mm_subs_epi16(__zero, input[11]); + x1[8] = _mm_subs_epi16(__zero, input[1]); + x1[9] = input[14]; + x1[10] = input[6]; + x1[11] = _mm_subs_epi16(__zero, input[9]); + x1[12] = input[2]; + x1[13] = _mm_subs_epi16(__zero, input[13]); + x1[14] = _mm_subs_epi16(__zero, input[5]); + x1[15] = input[10]; + + // stage 2 + __m128i x2[16]; + x2[0] = x1[0]; + x2[1] = x1[1]; + btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x1[2], x1[3], x2[2], x2[3]); + x2[4] = x1[4]; + x2[5] = x1[5]; + btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x1[6], x1[7], x2[6], x2[7]); + x2[8] = x1[8]; + x2[9] = x1[9]; + btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x1[10], x1[11], x2[10], x2[11]); + x2[12] = x1[12]; + x2[13] = x1[13]; + btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x1[14], x1[15], x2[14], x2[15]); + + // stage 3 + __m128i x3[16]; + x3[0] = _mm_adds_epi16(x2[0], x2[2]); + x3[2] = _mm_subs_epi16(x2[0], x2[2]); + x3[1] = _mm_adds_epi16(x2[1], x2[3]); + x3[3] = _mm_subs_epi16(x2[1], x2[3]); + x3[4] = _mm_adds_epi16(x2[4], x2[6]); + x3[6] = _mm_subs_epi16(x2[4], x2[6]); + x3[5] = _mm_adds_epi16(x2[5], x2[7]); + x3[7] = _mm_subs_epi16(x2[5], x2[7]); + x3[8] = _mm_adds_epi16(x2[8], x2[10]); + x3[10] = _mm_subs_epi16(x2[8], x2[10]); + x3[9] = _mm_adds_epi16(x2[9], x2[11]); + x3[11] = _mm_subs_epi16(x2[9], x2[11]); + x3[12] = _mm_adds_epi16(x2[12], x2[14]); + x3[14] = _mm_subs_epi16(x2[12], x2[14]); + x3[13] = _mm_adds_epi16(x2[13], x2[15]); + x3[15] = _mm_subs_epi16(x2[13], x2[15]); + + // stage 4 + __m128i x4[16]; + x4[0] = x3[0]; + x4[1] = x3[1]; + x4[2] = x3[2]; + x4[3] = x3[3]; + btf_16_sse2(cospi_p16_p48, cospi_p48_m16, x3[4], x3[5], x4[4], x4[5]); + btf_16_sse2(cospi_m48_p16, cospi_p16_p48, x3[6], x3[7], x4[6], x4[7]); + x4[8] = x3[8]; + x4[9] = x3[9]; + x4[10] = x3[10]; + x4[11] = x3[11]; + btf_16_sse2(cospi_p16_p48, cospi_p48_m16, x3[12], x3[13], x4[12], x4[13]); + btf_16_sse2(cospi_m48_p16, cospi_p16_p48, x3[14], x3[15], x4[14], x4[15]); + + // stage 5 + __m128i x5[16]; + x5[0] = _mm_adds_epi16(x4[0], x4[4]); + x5[4] = _mm_subs_epi16(x4[0], x4[4]); + x5[1] = _mm_adds_epi16(x4[1], x4[5]); + x5[5] = _mm_subs_epi16(x4[1], x4[5]); + x5[2] = _mm_adds_epi16(x4[2], x4[6]); + x5[6] = _mm_subs_epi16(x4[2], x4[6]); + x5[3] = _mm_adds_epi16(x4[3], x4[7]); + x5[7] = _mm_subs_epi16(x4[3], x4[7]); + x5[8] = _mm_adds_epi16(x4[8], x4[12]); + x5[12] = _mm_subs_epi16(x4[8], x4[12]); + x5[9] = _mm_adds_epi16(x4[9], x4[13]); + x5[13] = _mm_subs_epi16(x4[9], x4[13]); + x5[10] = _mm_adds_epi16(x4[10], x4[14]); + x5[14] = _mm_subs_epi16(x4[10], x4[14]); + x5[11] = _mm_adds_epi16(x4[11], x4[15]); + x5[15] = _mm_subs_epi16(x4[11], x4[15]); + + // stage 6 + __m128i x6[16]; + x6[0] = x5[0]; + x6[1] = x5[1]; + x6[2] = x5[2]; + x6[3] = x5[3]; + x6[4] = x5[4]; + x6[5] = x5[5]; + x6[6] = x5[6]; + x6[7] = x5[7]; + btf_16_sse2(cospi_p08_p56, cospi_p56_m08, x5[8], x5[9], x6[8], x6[9]); + btf_16_sse2(cospi_p40_p24, cospi_p24_m40, x5[10], x5[11], x6[10], x6[11]); + btf_16_sse2(cospi_m56_p08, cospi_p08_p56, x5[12], x5[13], x6[12], x6[13]); + btf_16_sse2(cospi_m24_p40, cospi_p40_p24, x5[14], x5[15], x6[14], x6[15]); + + // stage 7 + __m128i x7[16]; + x7[0] = _mm_adds_epi16(x6[0], x6[8]); + x7[8] = _mm_subs_epi16(x6[0], x6[8]); + x7[1] = _mm_adds_epi16(x6[1], x6[9]); + x7[9] = _mm_subs_epi16(x6[1], x6[9]); + x7[2] = _mm_adds_epi16(x6[2], x6[10]); + x7[10] = _mm_subs_epi16(x6[2], x6[10]); + x7[3] = _mm_adds_epi16(x6[3], x6[11]); + x7[11] = _mm_subs_epi16(x6[3], x6[11]); + x7[4] = _mm_adds_epi16(x6[4], x6[12]); + x7[12] = _mm_subs_epi16(x6[4], x6[12]); + x7[5] = _mm_adds_epi16(x6[5], x6[13]); + x7[13] = _mm_subs_epi16(x6[5], x6[13]); + x7[6] = _mm_adds_epi16(x6[6], x6[14]); + x7[14] = _mm_subs_epi16(x6[6], x6[14]); + x7[7] = _mm_adds_epi16(x6[7], x6[15]); + x7[15] = _mm_subs_epi16(x6[7], x6[15]); + + // stage 8 + __m128i x8[16]; + btf_16_sse2(cospi_p02_p62, cospi_p62_m02, x7[0], x7[1], x8[0], x8[1]); + btf_16_sse2(cospi_p10_p54, cospi_p54_m10, x7[2], x7[3], x8[2], x8[3]); + btf_16_sse2(cospi_p18_p46, cospi_p46_m18, x7[4], x7[5], x8[4], x8[5]); + btf_16_sse2(cospi_p26_p38, cospi_p38_m26, x7[6], x7[7], x8[6], x8[7]); + btf_16_sse2(cospi_p34_p30, cospi_p30_m34, x7[8], x7[9], x8[8], x8[9]); + btf_16_sse2(cospi_p42_p22, cospi_p22_m42, x7[10], x7[11], x8[10], x8[11]); + btf_16_sse2(cospi_p50_p14, cospi_p14_m50, x7[12], x7[13], x8[12], x8[13]); + btf_16_sse2(cospi_p58_p06, cospi_p06_m58, x7[14], x7[15], x8[14], x8[15]); + + // stage 9 + output[0] = x8[1]; + output[1] = x8[14]; + output[2] = x8[3]; + output[3] = x8[12]; + output[4] = x8[5]; + output[5] = x8[10]; + output[6] = x8[7]; + output[7] = x8[8]; + output[8] = x8[9]; + output[9] = x8[6]; + output[10] = x8[11]; + output[11] = x8[4]; + output[12] = x8[13]; + output[13] = x8[2]; + output[14] = x8[15]; + output[15] = x8[0]; +} + +static const transform_1d_sse2 col_txfm4x4_arr[TX_TYPES] = { + fdct4x4_new_sse2, // DCT_DCT + fadst4x4_new_sse2, // ADST_DCT + fdct4x4_new_sse2, // DCT_ADST + fadst4x4_new_sse2, // ADST_ADST + fadst4x4_new_sse2, // FLIPADST_DCT + fdct4x4_new_sse2, // DCT_FLIPADST + fadst4x4_new_sse2, // FLIPADST_FLIPADST + fadst4x4_new_sse2, // ADST_FLIPADST + fadst4x4_new_sse2, // FLIPADST_ADST + fidentity4x4_new_sse2, // IDTX + fdct4x4_new_sse2, // V_DCT + fidentity4x4_new_sse2, // H_DCT + fadst4x4_new_sse2, // V_ADST + fidentity4x4_new_sse2, // H_ADST + fadst4x4_new_sse2, // V_FLIPADST + fidentity4x4_new_sse2 // H_FLIPADST +}; + +static const transform_1d_sse2 row_txfm4x4_arr[TX_TYPES] = { + fdct4x4_new_sse2, // DCT_DCT + fdct4x4_new_sse2, // ADST_DCT + fadst4x4_new_sse2, // DCT_ADST + fadst4x4_new_sse2, // ADST_ADST + fdct4x4_new_sse2, // FLIPADST_DCT + fadst4x4_new_sse2, // DCT_FLIPADST + fadst4x4_new_sse2, // FLIPADST_FLIPADST + fadst4x4_new_sse2, // ADST_FLIPADST + fadst4x4_new_sse2, // FLIPADST_ADST + fidentity4x4_new_sse2, // IDTX + fidentity4x4_new_sse2, // V_DCT + fdct4x4_new_sse2, // H_DCT + fidentity4x4_new_sse2, // V_ADST + fadst4x4_new_sse2, // H_ADST + fidentity4x4_new_sse2, // V_FLIPADST + fadst4x4_new_sse2 // H_FLIPADST +}; + +static const transform_1d_sse2 col_txfm4x8_arr[TX_TYPES] = { + fdct4x8_new_sse2, // DCT_DCT + fadst4x8_new_sse2, // ADST_DCT + fdct4x8_new_sse2, // DCT_ADST + fadst4x8_new_sse2, // ADST_ADST + fadst4x8_new_sse2, // FLIPADST_DCT + fdct4x8_new_sse2, // DCT_FLIPADST + fadst4x8_new_sse2, // FLIPADST_FLIPADST + fadst4x8_new_sse2, // ADST_FLIPADST + fadst4x8_new_sse2, // FLIPADST_ADST + fidentity8x8_new_sse2, // IDTX + fdct4x8_new_sse2, // V_DCT + fidentity8x8_new_sse2, // H_DCT + fadst4x8_new_sse2, // V_ADST + fidentity8x8_new_sse2, // H_ADST + fadst4x8_new_sse2, // V_FLIPADST + fidentity8x8_new_sse2 // H_FLIPADST +}; + +static const transform_1d_sse2 row_txfm8x4_arr[TX_TYPES] = { + fdct8x4_new_sse2, // DCT_DCT + fdct8x4_new_sse2, // ADST_DCT + fadst8x4_new_sse2, // DCT_ADST + fadst8x4_new_sse2, // ADST_ADST + fdct8x4_new_sse2, // FLIPADST_DCT + fadst8x4_new_sse2, // DCT_FLIPADST + fadst8x4_new_sse2, // FLIPADST_FLIPADST + fadst8x4_new_sse2, // ADST_FLIPADST + fadst8x4_new_sse2, // FLIPADST_ADST + fidentity8x4_new_sse2, // IDTX + fidentity8x4_new_sse2, // V_DCT + fdct8x4_new_sse2, // H_DCT + fidentity8x4_new_sse2, // V_ADST + fadst8x4_new_sse2, // H_ADST + fidentity8x4_new_sse2, // V_FLIPADST + fadst8x4_new_sse2 // H_FLIPADST +}; + +static const transform_1d_sse2 col_txfm8x4_arr[TX_TYPES] = { + fdct8x4_new_sse2, // DCT_DCT + fadst8x4_new_sse2, // ADST_DCT + fdct8x4_new_sse2, // DCT_ADST + fadst8x4_new_sse2, // ADST_ADST + fadst8x4_new_sse2, // FLIPADST_DCT + fdct8x4_new_sse2, // DCT_FLIPADST + fadst8x4_new_sse2, // FLIPADST_FLIPADST + fadst8x4_new_sse2, // ADST_FLIPADST + fadst8x4_new_sse2, // FLIPADST_ADST + fidentity8x4_new_sse2, // IDTX + fdct8x4_new_sse2, // V_DCT + fidentity8x4_new_sse2, // H_DCT + fadst8x4_new_sse2, // V_ADST + fidentity8x4_new_sse2, // H_ADST + fadst8x4_new_sse2, // V_FLIPADST + fidentity8x4_new_sse2 // H_FLIPADST +}; + +static const transform_1d_sse2 row_txfm4x8_arr[TX_TYPES] = { + fdct4x8_new_sse2, // DCT_DCT + fdct4x8_new_sse2, // ADST_DCT + fadst4x8_new_sse2, // DCT_ADST + fadst4x8_new_sse2, // ADST_ADST + fdct4x8_new_sse2, // FLIPADST_DCT + fadst4x8_new_sse2, // DCT_FLIPADST + fadst4x8_new_sse2, // FLIPADST_FLIPADST + fadst4x8_new_sse2, // ADST_FLIPADST + fadst4x8_new_sse2, // FLIPADST_ADST + fidentity8x8_new_sse2, // IDTX + fidentity8x8_new_sse2, // V_DCT + fdct4x8_new_sse2, // H_DCT + fidentity8x8_new_sse2, // V_ADST + fadst4x8_new_sse2, // H_ADST + fidentity8x8_new_sse2, // V_FLIPADST + fadst4x8_new_sse2 // H_FLIPADST +}; + +static const transform_1d_sse2 col_txfm8x8_arr[TX_TYPES] = { + fdct8x8_new_sse2, // DCT_DCT + fadst8x8_new_sse2, // ADST_DCT + fdct8x8_new_sse2, // DCT_ADST + fadst8x8_new_sse2, // ADST_ADST + fadst8x8_new_sse2, // FLIPADST_DCT + fdct8x8_new_sse2, // DCT_FLIPADST + fadst8x8_new_sse2, // FLIPADST_FLIPADST + fadst8x8_new_sse2, // ADST_FLIPADST + fadst8x8_new_sse2, // FLIPADST_ADST + fidentity8x8_new_sse2, // IDTX + fdct8x8_new_sse2, // V_DCT + fidentity8x8_new_sse2, // H_DCT + fadst8x8_new_sse2, // V_ADST + fidentity8x8_new_sse2, // H_ADST + fadst8x8_new_sse2, // V_FLIPADST + fidentity8x8_new_sse2, // H_FLIPADST +}; + +static const transform_1d_sse2 row_txfm8x8_arr[TX_TYPES] = { + fdct8x8_new_sse2, // DCT_DCT + fdct8x8_new_sse2, // ADST_DCT + fadst8x8_new_sse2, // DCT_ADST + fadst8x8_new_sse2, // ADST_ADST + fdct8x8_new_sse2, // FLIPADST_DCT + fadst8x8_new_sse2, // DCT_FLIPADST + fadst8x8_new_sse2, // FLIPADST_FLIPADST + fadst8x8_new_sse2, // ADST_FLIPADST + fadst8x8_new_sse2, // FLIPADST_ADST + fidentity8x8_new_sse2, // IDTX + fidentity8x8_new_sse2, // V_DCT + fdct8x8_new_sse2, // H_DCT + fidentity8x8_new_sse2, // V_ADST + fadst8x8_new_sse2, // H_ADST + fidentity8x8_new_sse2, // V_FLIPADST + fadst8x8_new_sse2 // H_FLIPADST +}; + +static const transform_1d_sse2 col_txfm8x16_arr[TX_TYPES] = { + fdct8x16_new_sse2, // DCT_DCT + fadst8x16_new_sse2, // ADST_DCT + fdct8x16_new_sse2, // DCT_ADST + fadst8x16_new_sse2, // ADST_ADST + fadst8x16_new_sse2, // FLIPADST_DCT + fdct8x16_new_sse2, // DCT_FLIPADST + fadst8x16_new_sse2, // FLIPADST_FLIPADST + fadst8x16_new_sse2, // ADST_FLIPADST + fadst8x16_new_sse2, // FLIPADST_ADST + fidentity8x16_new_sse2, // IDTX + fdct8x16_new_sse2, // V_DCT + fidentity8x16_new_sse2, // H_DCT + fadst8x16_new_sse2, // V_ADST + fidentity8x16_new_sse2, // H_ADST + fadst8x16_new_sse2, // V_FLIPADST + fidentity8x16_new_sse2 // H_FLIPADST +}; + +static const transform_1d_sse2 row_txfm8x16_arr[TX_TYPES] = { + fdct8x16_new_sse2, // DCT_DCT + fdct8x16_new_sse2, // ADST_DCT + fadst8x16_new_sse2, // DCT_ADST + fadst8x16_new_sse2, // ADST_ADST + fdct8x16_new_sse2, // FLIPADST_DCT + fadst8x16_new_sse2, // DCT_FLIPADST + fadst8x16_new_sse2, // FLIPADST_FLIPADST + fadst8x16_new_sse2, // ADST_FLIPADST + fadst8x16_new_sse2, // FLIPADST_ADST + fidentity8x16_new_sse2, // IDTX + fidentity8x16_new_sse2, // V_DCT + fdct8x16_new_sse2, // H_DCT + fidentity8x16_new_sse2, // V_ADST + fadst8x16_new_sse2, // H_ADST + fidentity8x16_new_sse2, // V_FLIPADST + fadst8x16_new_sse2 // H_FLIPADST +}; + +static const transform_1d_sse2 row_txfm8x32_arr[TX_TYPES] = { + fdct8x32_new_sse2, // DCT_DCT + NULL, // ADST_DCT + NULL, // DCT_ADST + NULL, // ADST_ADST + NULL, // FLIPADST_DCT + NULL, // DCT_FLIPADST + NULL, // FLIPADST_FLIPADST + NULL, // ADST_FLIPADST + NULL, // FLIPADST_ADST + fidentity8x32_new_sse2, // IDTX + fidentity8x32_new_sse2, // V_DCT + fdct8x32_new_sse2, // H_DCT + NULL, // V_ADST + NULL, // H_ADST + NULL, // V_FLIPADST + NULL // H_FLIPADST +}; + +void av1_lowbd_fwd_txfm2d_4x4_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m128i buf0[4], buf1[4], *buf; + const int8_t *shift = fwd_txfm_shift_ls[TX_4X4]; + const int txw_idx = get_txw_idx(TX_4X4); + const int txh_idx = get_txh_idx(TX_4X4); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 4; + const int height = 4; + const transform_1d_sse2 col_txfm = col_txfm4x4_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm4x4_arr[tx_type]; + int ud_flip, lr_flip; + + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + if (ud_flip) { + load_buffer_16bit_to_16bit_w4_flip(input, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit_w4(input, stride, buf0, height); + } + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_4x4(buf0, buf1); + + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1, buf, width); + } else { + buf = buf1; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_4x4(buf, buf); + store_buffer_16bit_to_32bit_w4(buf, output, width, height); +} + +void av1_lowbd_fwd_txfm2d_4x8_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)stride; + (void)bd; + __m128i buf0[8], buf1[8], *buf; + const int8_t *shift = fwd_txfm_shift_ls[TX_4X8]; + const int txw_idx = get_txw_idx(TX_4X8); + const int txh_idx = get_txh_idx(TX_4X8); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 4; + const int height = 8; + const transform_1d_sse2 col_txfm = col_txfm4x8_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm8x4_arr[tx_type]; + int ud_flip, lr_flip; + + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + if (ud_flip) { + load_buffer_16bit_to_16bit_w4_flip(input, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit_w4(input, stride, buf0, height); + } + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_4x8(buf0, buf1); + + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1, buf, width); + } else { + buf = buf1; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_8x4(buf, buf); + store_rect_buffer_16bit_to_32bit_w4(buf, output, width, height); +} + +void av1_lowbd_fwd_txfm2d_4x16_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m128i buf0[16], buf1[16]; + const int8_t *shift = fwd_txfm_shift_ls[TX_4X16]; + const int txw_idx = get_txw_idx(TX_4X16); + const int txh_idx = get_txh_idx(TX_4X16); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 4; + const int height = 16; + const transform_1d_sse2 col_txfm = col_txfm8x16_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm8x4_arr[tx_type]; + int ud_flip, lr_flip; + + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + if (ud_flip) { + load_buffer_16bit_to_16bit_w4_flip(input, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit_w4(input, stride, buf0, height); + } + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_4x8(buf0, buf1); + transpose_16bit_4x8(buf0 + 8, buf1 + 8); + + for (int i = 0; i < 2; i++) { + __m128i *buf; + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1 + 8 * i, buf, width); + } else { + buf = buf1 + 8 * i; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_8x4(buf, buf); + store_buffer_16bit_to_32bit_w4(buf, output + 8 * width * i, width, 8); + } +} + +void av1_lowbd_fwd_txfm2d_8x4_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m128i buf0[8], buf1[8], *buf; + const int8_t *shift = fwd_txfm_shift_ls[TX_8X4]; + const int txw_idx = get_txw_idx(TX_8X4); + const int txh_idx = get_txh_idx(TX_8X4); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 8; + const int height = 4; + const transform_1d_sse2 col_txfm = col_txfm8x4_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm4x8_arr[tx_type]; + int ud_flip, lr_flip; + + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + if (ud_flip) + load_buffer_16bit_to_16bit_flip(input, stride, buf0, height); + else + load_buffer_16bit_to_16bit(input, stride, buf0, height); + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_8x8(buf0, buf1); + + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1, buf, width); + } else { + buf = buf1; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_8x8(buf, buf); + store_rect_buffer_16bit_to_32bit_w8(buf, output, width, height); +} + +void av1_lowbd_fwd_txfm2d_8x8_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m128i buf0[8], buf1[8], *buf; + const int8_t *shift = fwd_txfm_shift_ls[TX_8X8]; + const int txw_idx = get_txw_idx(TX_8X8); + const int txh_idx = get_txh_idx(TX_8X8); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 8; + const int height = 8; + const transform_1d_sse2 col_txfm = col_txfm8x8_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm8x8_arr[tx_type]; + int ud_flip, lr_flip; + + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + if (ud_flip) + load_buffer_16bit_to_16bit_flip(input, stride, buf0, height); + else + load_buffer_16bit_to_16bit(input, stride, buf0, height); + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_8x8(buf0, buf1); + + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1, buf, width); + } else { + buf = buf1; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_8x8(buf, buf); + store_buffer_16bit_to_32bit_w8(buf, output, width, height); +} + +void av1_lowbd_fwd_txfm2d_8x16_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m128i buf0[16], buf1[16]; + const int8_t *shift = fwd_txfm_shift_ls[TX_8X16]; + const int txw_idx = get_txw_idx(TX_8X16); + const int txh_idx = get_txh_idx(TX_8X16); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 8; + const int height = 16; + const transform_1d_sse2 col_txfm = col_txfm8x16_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm8x8_arr[tx_type]; + int ud_flip, lr_flip; + + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + if (ud_flip) { + load_buffer_16bit_to_16bit_flip(input, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit(input, stride, buf0, height); + } + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_8x8(buf0, buf1); + transpose_16bit_8x8(buf0 + 8, buf1 + 8); + + for (int i = 0; i < 2; i++) { + __m128i *buf; + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1 + width * i, buf, width); + } else { + buf = buf1 + width * i; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_8x8(buf, buf); + store_rect_buffer_16bit_to_32bit_w8(buf, output + 8 * width * i, width, 8); + } +} + +void av1_lowbd_fwd_txfm2d_8x32_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m128i buf0[32], buf1[32]; + const int8_t *shift = fwd_txfm_shift_ls[TX_8X32]; + const int txw_idx = get_txw_idx(TX_8X32); + const int txh_idx = get_txh_idx(TX_8X32); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 8; + const int height = 32; + const transform_1d_sse2 col_txfm = col_txfm8x32_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm8x8_arr[tx_type]; + int ud_flip, lr_flip; + + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + if (ud_flip) { + load_buffer_16bit_to_16bit_flip(input, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit(input, stride, buf0, height); + } + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_8x8(buf0, buf1); + transpose_16bit_8x8(buf0 + 8, buf1 + 8); + transpose_16bit_8x8(buf0 + 16, buf1 + 16); + transpose_16bit_8x8(buf0 + 24, buf1 + 24); + + for (int i = 0; i < 4; i++) { + __m128i *buf; + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1 + width * i, buf, width); + } else { + buf = buf1 + width * i; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_8x8(buf, buf); + store_buffer_16bit_to_32bit_w8(buf, output + 8 * width * i, width, 8); + } +} + +void av1_lowbd_fwd_txfm2d_16x4_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m128i buf0[16], buf1[16]; + const int8_t *shift = fwd_txfm_shift_ls[TX_16X4]; + const int txw_idx = get_txw_idx(TX_16X4); + const int txh_idx = get_txh_idx(TX_16X4); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 16; + const int height = 4; + const transform_1d_sse2 col_txfm = col_txfm8x4_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm8x16_arr[tx_type]; + __m128i *buf; + int ud_flip, lr_flip; + + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + for (int i = 0; i < 2; i++) { + if (ud_flip) { + load_buffer_16bit_to_16bit_flip(input + 8 * i, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height); + } + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_8x4(buf0, buf1 + 8 * i); + } + + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1, buf, width); + } else { + buf = buf1; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_4x8(buf, buf); + store_buffer_16bit_to_32bit_w8(buf, output, width, height); + transpose_16bit_4x8(buf + 8, buf + 8); + store_buffer_16bit_to_32bit_w8(buf + 8, output + 8, width, height); +} + +void av1_lowbd_fwd_txfm2d_16x8_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m128i buf0[16], buf1[16]; + const int8_t *shift = fwd_txfm_shift_ls[TX_16X8]; + const int txw_idx = get_txw_idx(TX_16X8); + const int txh_idx = get_txh_idx(TX_16X8); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 16; + const int height = 8; + const transform_1d_sse2 col_txfm = col_txfm8x8_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm8x16_arr[tx_type]; + __m128i *buf; + int ud_flip, lr_flip; + + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + for (int i = 0; i < 2; i++) { + if (ud_flip) { + load_buffer_16bit_to_16bit_flip(input + 8 * i, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height); + } + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_8x8(buf0, buf1 + 8 * i); + } + + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1, buf, width); + } else { + buf = buf1; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_8x8(buf, buf); + store_rect_buffer_16bit_to_32bit_w8(buf, output, width, height); + transpose_16bit_8x8(buf + 8, buf + 8); + store_rect_buffer_16bit_to_32bit_w8(buf + 8, output + 8, width, height); +} + +void av1_lowbd_fwd_txfm2d_16x16_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m128i buf0[16], buf1[32]; + const int8_t *shift = fwd_txfm_shift_ls[TX_16X16]; + const int txw_idx = get_txw_idx(TX_16X16); + const int txh_idx = get_txh_idx(TX_16X16); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 16; + const int height = 16; + const transform_1d_sse2 col_txfm = col_txfm8x16_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm8x16_arr[tx_type]; + int ud_flip, lr_flip; + + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + for (int i = 0; i < 2; i++) { + if (ud_flip) { + load_buffer_16bit_to_16bit_flip(input + 8 * i, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height); + } + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_8x8(buf0, buf1 + 0 * width + 8 * i); + transpose_16bit_8x8(buf0 + 8, buf1 + 1 * width + 8 * i); + } + + for (int i = 0; i < 2; i++) { + __m128i *buf; + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1 + width * i, buf, width); + } else { + buf = buf1 + width * i; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_8x8(buf, buf); + store_buffer_16bit_to_32bit_w8(buf, output + 8 * width * i, width, 8); + transpose_16bit_8x8(buf + 8, buf + 8); + store_buffer_16bit_to_32bit_w8(buf + 8, output + 8 * width * i + 8, width, + 8); + } +} + +void av1_lowbd_fwd_txfm2d_16x32_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m128i buf0[32], buf1[64]; + const int8_t *shift = fwd_txfm_shift_ls[TX_16X32]; + const int txw_idx = get_txw_idx(TX_16X32); + const int txh_idx = get_txh_idx(TX_16X32); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 16; + const int height = 32; + const transform_1d_sse2 col_txfm = col_txfm8x32_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm8x16_arr[tx_type]; + + if (col_txfm != NULL && row_txfm != NULL) { + int ud_flip, lr_flip; + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + + for (int i = 0; i < 2; i++) { + if (ud_flip) { + load_buffer_16bit_to_16bit_flip(input + 8 * i, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height); + } + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_8x8(buf0 + 0 * 8, buf1 + 0 * width + 8 * i); + transpose_16bit_8x8(buf0 + 1 * 8, buf1 + 1 * width + 8 * i); + transpose_16bit_8x8(buf0 + 2 * 8, buf1 + 2 * width + 8 * i); + transpose_16bit_8x8(buf0 + 3 * 8, buf1 + 3 * width + 8 * i); + } + + for (int i = 0; i < 4; i++) { + __m128i *buf; + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1 + width * i, buf, width); + } else { + buf = buf1 + width * i; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_8x8(buf, buf); + store_rect_buffer_16bit_to_32bit_w8(buf, output + 8 * width * i, width, + 8); + transpose_16bit_8x8(buf + 8, buf + 8); + store_rect_buffer_16bit_to_32bit_w8(buf + 8, output + 8 * width * i + 8, + width, 8); + } + } else { + av1_fwd_txfm2d_16x32_c(input, output, stride, tx_type, bd); + } +} + +void av1_lowbd_fwd_txfm2d_32x8_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m128i buf0[32], buf1[32]; + const int8_t *shift = fwd_txfm_shift_ls[TX_32X8]; + const int txw_idx = get_txw_idx(TX_32X8); + const int txh_idx = get_txh_idx(TX_32X8); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 32; + const int height = 8; + const transform_1d_sse2 col_txfm = col_txfm8x8_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm8x32_arr[tx_type]; + + if (col_txfm != NULL && row_txfm != NULL) { + int ud_flip, lr_flip; + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + + for (int i = 0; i < 4; i++) { + if (ud_flip) { + load_buffer_16bit_to_16bit_flip(input + 8 * i, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height); + } + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_8x8(buf0, buf1 + 0 * width + 8 * i); + } + + for (int i = 0; i < 1; i++) { + __m128i *buf; + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1 + width * i, buf, width); + } else { + buf = buf1 + width * i; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_8x8(buf, buf); + store_buffer_16bit_to_32bit_w8(buf, output + 8 * width * i, width, + height); + transpose_16bit_8x8(buf + 8, buf + 8); + store_buffer_16bit_to_32bit_w8(buf + 8, output + 8 * width * i + 8, width, + height); + transpose_16bit_8x8(buf + 16, buf + 16); + store_buffer_16bit_to_32bit_w8(buf + 16, output + 8 * width * i + 16, + width, height); + transpose_16bit_8x8(buf + 24, buf + 24); + store_buffer_16bit_to_32bit_w8(buf + 24, output + 8 * width * i + 24, + width, height); + } + } else { + av1_fwd_txfm2d_32x16_c(input, output, stride, tx_type, bd); + } +} + +void av1_lowbd_fwd_txfm2d_32x16_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m128i buf0[32], buf1[64]; + const int8_t *shift = fwd_txfm_shift_ls[TX_32X16]; + const int txw_idx = get_txw_idx(TX_32X16); + const int txh_idx = get_txh_idx(TX_32X16); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 32; + const int height = 16; + const transform_1d_sse2 col_txfm = col_txfm8x16_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm8x32_arr[tx_type]; + + if (col_txfm != NULL && row_txfm != NULL) { + int ud_flip, lr_flip; + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + + for (int i = 0; i < 4; i++) { + if (ud_flip) { + load_buffer_16bit_to_16bit_flip(input + 8 * i, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height); + } + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_8x8(buf0, buf1 + 0 * width + 8 * i); + transpose_16bit_8x8(buf0 + 8, buf1 + 1 * width + 8 * i); + } + + for (int i = 0; i < 2; i++) { + __m128i *buf; + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1 + width * i, buf, width); + } else { + buf = buf1 + width * i; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_8x8(buf, buf); + store_rect_buffer_16bit_to_32bit_w8(buf, output + 8 * width * i, width, + 8); + transpose_16bit_8x8(buf + 8, buf + 8); + store_rect_buffer_16bit_to_32bit_w8(buf + 8, output + 8 * width * i + 8, + width, 8); + transpose_16bit_8x8(buf + 16, buf + 16); + store_rect_buffer_16bit_to_32bit_w8(buf + 16, output + 8 * width * i + 16, + width, 8); + transpose_16bit_8x8(buf + 24, buf + 24); + store_rect_buffer_16bit_to_32bit_w8(buf + 24, output + 8 * width * i + 24, + width, 8); + } + } else { + av1_fwd_txfm2d_32x16_c(input, output, stride, tx_type, bd); + } +} + +void av1_lowbd_fwd_txfm2d_32x32_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + __m128i buf0[32], buf1[128]; + const int8_t *shift = fwd_txfm_shift_ls[TX_32X32]; + const int txw_idx = get_txw_idx(TX_32X32); + const int txh_idx = get_txh_idx(TX_32X32); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = 32; + const int height = 32; + const transform_1d_sse2 col_txfm = col_txfm8x32_arr[tx_type]; + const transform_1d_sse2 row_txfm = row_txfm8x32_arr[tx_type]; + + if (col_txfm != NULL && row_txfm != NULL) { + int ud_flip, lr_flip; + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + + for (int i = 0; i < 4; i++) { + if (ud_flip) { + load_buffer_16bit_to_16bit_flip(input + 8 * i, stride, buf0, height); + } else { + load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height); + } + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + transpose_16bit_8x8(buf0 + 0 * 8, buf1 + 0 * width + 8 * i); + transpose_16bit_8x8(buf0 + 1 * 8, buf1 + 1 * width + 8 * i); + transpose_16bit_8x8(buf0 + 2 * 8, buf1 + 2 * width + 8 * i); + transpose_16bit_8x8(buf0 + 3 * 8, buf1 + 3 * width + 8 * i); + } + + for (int i = 0; i < 4; i++) { + __m128i *buf; + if (lr_flip) { + buf = buf0; + flip_buf_sse2(buf1 + width * i, buf, width); + } else { + buf = buf1 + width * i; + } + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + transpose_16bit_8x8(buf, buf); + store_buffer_16bit_to_32bit_w8(buf, output + 8 * width * i, width, 8); + transpose_16bit_8x8(buf + 8, buf + 8); + store_buffer_16bit_to_32bit_w8(buf + 8, output + 8 * width * i + 8, width, + 8); + transpose_16bit_8x8(buf + 16, buf + 16); + store_buffer_16bit_to_32bit_w8(buf + 16, output + 8 * width * i + 16, + width, 8); + transpose_16bit_8x8(buf + 24, buf + 24); + store_buffer_16bit_to_32bit_w8(buf + 24, output + 8 * width * i + 24, + width, 8); + } + } else { + av1_fwd_txfm2d_32x32_c(input, output, stride, tx_type, bd); + } +} + +void av1_lowbd_fwd_txfm2d_64x16_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + (void)tx_type; + assert(tx_type == DCT_DCT); + const TX_SIZE tx_size = TX_64X16; + __m128i buf0[64], buf1[128]; + const int8_t *shift = fwd_txfm_shift_ls[tx_size]; + const int txw_idx = get_txw_idx(tx_size); + const int txh_idx = get_txh_idx(tx_size); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = tx_size_wide[tx_size]; + const int height = tx_size_high[tx_size]; + const transform_1d_sse2 col_txfm = fdct8x16_new_sse2; + const transform_1d_sse2 row_txfm = fdct8x64_new_sse2; + const int width_div8 = (width >> 3); + const int height_div8 = (height >> 3); + + for (int i = 0; i < width_div8; i++) { + load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height); + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + for (int j = 0; j < height_div8; ++j) { + transpose_16bit_8x8(buf0 + j * 8, buf1 + j * width + 8 * i); + } + } + + for (int i = 0; i < height_div8; i++) { + __m128i *buf = buf1 + width * i; + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + int32_t *output8 = output + 8 * 32 * i; + for (int j = 0; j < 4; ++j) { + __m128i *buf8 = buf + 8 * j; + transpose_16bit_8x8(buf8, buf8); + store_buffer_16bit_to_32bit_w8(buf8, output8 + 8 * j, 32, 8); + } + } +} + +void av1_lowbd_fwd_txfm2d_16x64_sse2(const int16_t *input, int32_t *output, + int stride, TX_TYPE tx_type, int bd) { + (void)bd; + (void)tx_type; + assert(tx_type == DCT_DCT); + const TX_SIZE tx_size = TX_16X64; + __m128i buf0[64], buf1[128]; + const int8_t *shift = fwd_txfm_shift_ls[tx_size]; + const int txw_idx = get_txw_idx(tx_size); + const int txh_idx = get_txh_idx(tx_size); + const int cos_bit_col = fwd_cos_bit_col[txw_idx][txh_idx]; + const int cos_bit_row = fwd_cos_bit_row[txw_idx][txh_idx]; + const int width = tx_size_wide[tx_size]; + const int height = tx_size_high[tx_size]; + const transform_1d_sse2 col_txfm = fdct8x64_new_sse2; + const transform_1d_sse2 row_txfm = fdct8x16_new_sse2; + const int width_div8 = (width >> 3); + const int height_div8 = (height >> 3); + + for (int i = 0; i < width_div8; i++) { + load_buffer_16bit_to_16bit(input + 8 * i, stride, buf0, height); + round_shift_16bit(buf0, height, shift[0]); + col_txfm(buf0, buf0, cos_bit_col); + round_shift_16bit(buf0, height, shift[1]); + for (int j = 0; j < height_div8; ++j) { + transpose_16bit_8x8(buf0 + j * 8, buf1 + j * width + 8 * i); + } + } + + for (int i = 0; i < AOMMIN(4, height_div8); i++) { + __m128i *buf = buf1 + width * i; + row_txfm(buf, buf, cos_bit_row); + round_shift_16bit(buf, width, shift[2]); + int32_t *output8 = output + 8 * width * i; + for (int j = 0; j < width_div8; ++j) { + __m128i *buf8 = buf + 8 * j; + transpose_16bit_8x8(buf8, buf8); + store_buffer_16bit_to_32bit_w8(buf8, output8 + 8 * j, width, 8); + } + } + // Zero out the bottom 16x32 area. + memset(output + 16 * 32, 0, 16 * 32 * sizeof(*output)); +} + +static FwdTxfm2dFunc fwd_txfm2d_func_ls[TX_SIZES_ALL] = { + av1_lowbd_fwd_txfm2d_4x4_sse2, // 4x4 transform + av1_lowbd_fwd_txfm2d_8x8_sse2, // 8x8 transform + av1_lowbd_fwd_txfm2d_16x16_sse2, // 16x16 transform + av1_lowbd_fwd_txfm2d_32x32_sse2, // 32x32 transform + NULL, // 64x64 transform + av1_lowbd_fwd_txfm2d_4x8_sse2, // 4x8 transform + av1_lowbd_fwd_txfm2d_8x4_sse2, // 8x4 transform + av1_lowbd_fwd_txfm2d_8x16_sse2, // 8x16 transform + av1_lowbd_fwd_txfm2d_16x8_sse2, // 16x8 transform + av1_lowbd_fwd_txfm2d_16x32_sse2, // 16x32 transform + av1_lowbd_fwd_txfm2d_32x16_sse2, // 32x16 transform + NULL, // 32x64 transform + NULL, // 64x32 transform + av1_lowbd_fwd_txfm2d_4x16_sse2, // 4x16 transform + av1_lowbd_fwd_txfm2d_16x4_sse2, // 16x4 transform + av1_lowbd_fwd_txfm2d_8x32_sse2, // 8x32 transform + av1_lowbd_fwd_txfm2d_32x8_sse2, // 32x8 transform + av1_lowbd_fwd_txfm2d_16x64_sse2, // 16x64 transform + av1_lowbd_fwd_txfm2d_64x16_sse2, // 64x16 transform +}; + +void av1_lowbd_fwd_txfm_sse2(const int16_t *src_diff, tran_low_t *coeff, + int diff_stride, TxfmParam *txfm_param) { + FwdTxfm2dFunc fwd_txfm2d_func = fwd_txfm2d_func_ls[txfm_param->tx_size]; + + if ((fwd_txfm2d_func == NULL) || + (txfm_param->lossless && txfm_param->tx_size == TX_4X4)) + av1_lowbd_fwd_txfm_c(src_diff, coeff, diff_stride, txfm_param); + else + fwd_txfm2d_func(src_diff, coeff, diff_stride, txfm_param->tx_type, + txfm_param->bd); +} diff --git a/third_party/aom/av1/encoder/x86/av1_fwd_txfm_sse2.h b/third_party/aom/av1/encoder/x86/av1_fwd_txfm_sse2.h new file mode 100644 index 0000000000..99a6b90829 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/av1_fwd_txfm_sse2.h @@ -0,0 +1,117 @@ +/* + * Copyright (c) 2018, 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_ENCODER_X86_AV1_FWD_TXFM_SSE2_H_ +#define AOM_AV1_ENCODER_X86_AV1_FWD_TXFM_SSE2_H_ + +#include <immintrin.h> + +#include "config/aom_config.h" +#include "config/av1_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/x86/transpose_sse2.h" +#include "aom_dsp/x86/txfm_common_sse2.h" + +#ifdef __cplusplus +extern "C" { +#endif + +void fdct8x32_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit); +void fdct8x64_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit); + +static INLINE void fidentity4x4_new_sse2(const __m128i *const input, + __m128i *const output, + const int8_t cos_bit) { + (void)cos_bit; + const __m128i one = _mm_set1_epi16(1); + + for (int i = 0; i < 4; ++i) { + const __m128i a = _mm_unpacklo_epi16(input[i], one); + const __m128i b = scale_round_sse2(a, NewSqrt2); + output[i] = _mm_packs_epi32(b, b); + } +} + +static INLINE void fidentity8x4_new_sse2(const __m128i *const input, + __m128i *const output, + const int8_t cos_bit) { + (void)cos_bit; + const __m128i one = _mm_set1_epi16(1); + + for (int i = 0; i < 4; ++i) { + const __m128i a_lo = _mm_unpacklo_epi16(input[i], one); + const __m128i a_hi = _mm_unpackhi_epi16(input[i], one); + const __m128i b_lo = scale_round_sse2(a_lo, NewSqrt2); + const __m128i b_hi = scale_round_sse2(a_hi, NewSqrt2); + output[i] = _mm_packs_epi32(b_lo, b_hi); + } +} + +static INLINE void fidentity8x8_new_sse2(const __m128i *input, __m128i *output, + int8_t cos_bit) { + (void)cos_bit; + + output[0] = _mm_adds_epi16(input[0], input[0]); + output[1] = _mm_adds_epi16(input[1], input[1]); + output[2] = _mm_adds_epi16(input[2], input[2]); + output[3] = _mm_adds_epi16(input[3], input[3]); + output[4] = _mm_adds_epi16(input[4], input[4]); + output[5] = _mm_adds_epi16(input[5], input[5]); + output[6] = _mm_adds_epi16(input[6], input[6]); + output[7] = _mm_adds_epi16(input[7], input[7]); +} + +static INLINE void fidentity8x16_new_sse2(const __m128i *input, __m128i *output, + int8_t cos_bit) { + (void)cos_bit; + const __m128i one = _mm_set1_epi16(1); + + for (int i = 0; i < 16; ++i) { + const __m128i a_lo = _mm_unpacklo_epi16(input[i], one); + const __m128i a_hi = _mm_unpackhi_epi16(input[i], one); + const __m128i b_lo = scale_round_sse2(a_lo, 2 * NewSqrt2); + const __m128i b_hi = scale_round_sse2(a_hi, 2 * NewSqrt2); + output[i] = _mm_packs_epi32(b_lo, b_hi); + } +} + +static INLINE void fidentity8x32_new_sse2(const __m128i *input, __m128i *output, + int8_t cos_bit) { + (void)cos_bit; + for (int i = 0; i < 32; ++i) { + output[i] = _mm_slli_epi16(input[i], 2); + } +} + +static const transform_1d_sse2 col_txfm8x32_arr[TX_TYPES] = { + fdct8x32_new_sse2, // DCT_DCT + NULL, // ADST_DCT + NULL, // DCT_ADST + NULL, // ADST_ADST + NULL, // FLIPADST_DCT + NULL, // DCT_FLIPADST + NULL, // FLIPADST_FLIPADST + NULL, // ADST_FLIPADST + NULL, // FLIPADST_ADST + fidentity8x32_new_sse2, // IDTX + fdct8x32_new_sse2, // V_DCT + fidentity8x32_new_sse2, // H_DCT + NULL, // V_ADST + NULL, // H_ADST + NULL, // V_FLIPADST + NULL // H_FLIPADST +}; + +#ifdef __cplusplus +} +#endif + +#endif // AOM_AV1_ENCODER_X86_AV1_FWD_TXFM_SSE2_H_ diff --git a/third_party/aom/av1/encoder/x86/av1_highbd_quantize_avx2.c b/third_party/aom/av1/encoder/x86/av1_highbd_quantize_avx2.c new file mode 100644 index 0000000000..b58911fcb2 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/av1_highbd_quantize_avx2.c @@ -0,0 +1,137 @@ +/* + * Copyright (c) 2017, 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 <immintrin.h> + +#include "config/av1_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/aom_dsp_common.h" + +static INLINE void init_one_qp(const __m128i *p, __m256i *qp) { + const __m128i zero = _mm_setzero_si128(); + const __m128i dc = _mm_unpacklo_epi16(*p, zero); + const __m128i ac = _mm_unpackhi_epi16(*p, zero); + *qp = _mm256_insertf128_si256(_mm256_castsi128_si256(dc), ac, 1); +} + +static INLINE void update_qp(__m256i *qp) { + qp[0] = _mm256_permute2x128_si256(qp[0], qp[0], 0x11); + qp[1] = _mm256_permute2x128_si256(qp[1], qp[1], 0x11); + qp[2] = _mm256_permute2x128_si256(qp[2], qp[2], 0x11); +} + +static INLINE void init_qp(const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *dequant_ptr, int log_scale, + __m256i *qp) { + __m128i round = _mm_loadu_si128((const __m128i *)round_ptr); + if (log_scale) { + const __m128i round_scale = _mm_set1_epi16(1 << (15 - log_scale)); + round = _mm_mulhrs_epi16(round, round_scale); + } + const __m128i quant = _mm_loadu_si128((const __m128i *)quant_ptr); + const __m128i dequant = _mm_loadu_si128((const __m128i *)dequant_ptr); + + init_one_qp(&round, &qp[0]); + init_one_qp(&quant, &qp[1]); + init_one_qp(&dequant, &qp[2]); +} + +static INLINE void quantize(const __m256i *qp, __m256i *c, + const int16_t *iscan_ptr, int log_scale, + tran_low_t *qcoeff, tran_low_t *dqcoeff, + __m256i *eob) { + const __m256i abs_coeff = _mm256_abs_epi32(*c); + __m256i q = _mm256_add_epi32(abs_coeff, qp[0]); + + __m256i q_lo = _mm256_mul_epi32(q, qp[1]); + __m256i q_hi = _mm256_srli_epi64(q, 32); + const __m256i qp_hi = _mm256_srli_epi64(qp[1], 32); + q_hi = _mm256_mul_epi32(q_hi, qp_hi); + q_lo = _mm256_srli_epi64(q_lo, 16 - log_scale); + q_hi = _mm256_srli_epi64(q_hi, 16 - log_scale); + q_hi = _mm256_slli_epi64(q_hi, 32); + q = _mm256_or_si256(q_lo, q_hi); + const __m256i abs_s = _mm256_slli_epi32(abs_coeff, 1 + log_scale); + const __m256i mask = _mm256_cmpgt_epi32(qp[2], abs_s); + q = _mm256_andnot_si256(mask, q); + + __m256i dq = _mm256_mullo_epi32(q, qp[2]); + dq = _mm256_srai_epi32(dq, log_scale); + q = _mm256_sign_epi32(q, *c); + dq = _mm256_sign_epi32(dq, *c); + + _mm256_storeu_si256((__m256i *)qcoeff, q); + _mm256_storeu_si256((__m256i *)dqcoeff, dq); + + const __m128i isc = _mm_loadu_si128((const __m128i *)iscan_ptr); + const __m128i zr = _mm_setzero_si128(); + const __m128i lo = _mm_unpacklo_epi16(isc, zr); + const __m128i hi = _mm_unpackhi_epi16(isc, zr); + const __m256i iscan = + _mm256_insertf128_si256(_mm256_castsi128_si256(lo), hi, 1); + + const __m256i zero = _mm256_setzero_si256(); + const __m256i zc = _mm256_cmpeq_epi32(dq, zero); + const __m256i nz = _mm256_cmpeq_epi32(zc, zero); + __m256i cur_eob = _mm256_sub_epi32(iscan, nz); + cur_eob = _mm256_and_si256(cur_eob, nz); + *eob = _mm256_max_epi32(cur_eob, *eob); +} + +void av1_highbd_quantize_fp_avx2( + const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, + const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan, int log_scale) { + (void)scan; + (void)zbin_ptr; + (void)quant_shift_ptr; + const unsigned int step = 8; + __m256i qp[3], coeff; + + init_qp(round_ptr, quant_ptr, dequant_ptr, log_scale, qp); + coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr); + + __m256i eob = _mm256_setzero_si256(); + quantize(qp, &coeff, iscan, log_scale, qcoeff_ptr, dqcoeff_ptr, &eob); + + coeff_ptr += step; + qcoeff_ptr += step; + dqcoeff_ptr += step; + iscan += step; + n_coeffs -= step; + + update_qp(qp); + while (n_coeffs > 0) { + coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr); + quantize(qp, &coeff, iscan, log_scale, qcoeff_ptr, dqcoeff_ptr, &eob); + + coeff_ptr += step; + qcoeff_ptr += step; + dqcoeff_ptr += step; + iscan += step; + n_coeffs -= step; + } + { + __m256i eob_s; + eob_s = _mm256_shuffle_epi32(eob, 0xe); + eob = _mm256_max_epi16(eob, eob_s); + eob_s = _mm256_shufflelo_epi16(eob, 0xe); + eob = _mm256_max_epi16(eob, eob_s); + eob_s = _mm256_shufflelo_epi16(eob, 1); + eob = _mm256_max_epi16(eob, eob_s); + const __m128i final_eob = _mm_max_epi16(_mm256_castsi256_si128(eob), + _mm256_extractf128_si256(eob, 1)); + *eob_ptr = _mm_extract_epi16(final_eob, 0); + } +} diff --git a/third_party/aom/av1/encoder/x86/av1_highbd_quantize_sse4.c b/third_party/aom/av1/encoder/x86/av1_highbd_quantize_sse4.c new file mode 100644 index 0000000000..40b3b460b6 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/av1_highbd_quantize_sse4.c @@ -0,0 +1,195 @@ +/* + * Copyright (c) 2016, 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 <smmintrin.h> +#include <stdint.h> + +#include "config/av1_rtcd.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/x86/synonyms.h" + +// Coefficient quantization phase 1 +// param[0-2] : rounding/quan/dequan constants +static INLINE void quantize_coeff_phase1(__m128i *coeff, const __m128i *param, + const int shift, const int scale, + __m128i *qcoeff, __m128i *dquan, + __m128i *sign) { + const __m128i zero = _mm_setzero_si128(); + const __m128i one = _mm_set1_epi32(1); + + *sign = _mm_cmplt_epi32(*coeff, zero); + *sign = _mm_or_si128(*sign, one); + *coeff = _mm_abs_epi32(*coeff); + + qcoeff[0] = _mm_add_epi32(*coeff, param[0]); + qcoeff[1] = _mm_unpackhi_epi32(qcoeff[0], zero); + qcoeff[0] = _mm_unpacklo_epi32(qcoeff[0], zero); + + qcoeff[0] = _mm_mul_epi32(qcoeff[0], param[1]); + qcoeff[0] = _mm_srli_epi64(qcoeff[0], shift); + dquan[0] = _mm_mul_epi32(qcoeff[0], param[2]); + dquan[0] = _mm_srli_epi64(dquan[0], scale); + const __m128i abs_s = _mm_slli_epi32(*coeff, 1 + scale); + qcoeff[2] = _mm_cmplt_epi32(abs_s, param[3]); +} + +// Coefficient quantization phase 2 +static INLINE void quantize_coeff_phase2(__m128i *qcoeff, __m128i *dquan, + const __m128i *sign, + const __m128i *param, const int shift, + const int scale, tran_low_t *qAddr, + tran_low_t *dqAddr) { + __m128i mask0L = _mm_set_epi32(-1, -1, 0, 0); + __m128i mask0H = _mm_set_epi32(0, 0, -1, -1); + + qcoeff[1] = _mm_mul_epi32(qcoeff[1], param[1]); + qcoeff[1] = _mm_srli_epi64(qcoeff[1], shift); + dquan[1] = _mm_mul_epi32(qcoeff[1], param[2]); + dquan[1] = _mm_srli_epi64(dquan[1], scale); + + // combine L&H + qcoeff[0] = _mm_shuffle_epi32(qcoeff[0], 0xd8); + qcoeff[1] = _mm_shuffle_epi32(qcoeff[1], 0x8d); + + qcoeff[0] = _mm_and_si128(qcoeff[0], mask0H); + qcoeff[1] = _mm_and_si128(qcoeff[1], mask0L); + + dquan[0] = _mm_shuffle_epi32(dquan[0], 0xd8); + dquan[1] = _mm_shuffle_epi32(dquan[1], 0x8d); + + dquan[0] = _mm_and_si128(dquan[0], mask0H); + dquan[1] = _mm_and_si128(dquan[1], mask0L); + + qcoeff[0] = _mm_or_si128(qcoeff[0], qcoeff[1]); + dquan[0] = _mm_or_si128(dquan[0], dquan[1]); + + qcoeff[0] = _mm_sign_epi32(qcoeff[0], *sign); + dquan[0] = _mm_sign_epi32(dquan[0], *sign); + qcoeff[0] = _mm_andnot_si128(qcoeff[2], qcoeff[0]); + dquan[0] = _mm_andnot_si128(qcoeff[2], dquan[0]); + _mm_storeu_si128((__m128i *)qAddr, qcoeff[0]); + _mm_storeu_si128((__m128i *)dqAddr, dquan[0]); +} + +static INLINE void find_eob(tran_low_t *qcoeff_ptr, const int16_t *iscan, + __m128i *eob) { + const __m128i zero = _mm_setzero_si128(); + __m128i mask, iscanIdx; + const __m128i q0 = _mm_loadu_si128((__m128i const *)qcoeff_ptr); + const __m128i q1 = _mm_loadu_si128((__m128i const *)(qcoeff_ptr + 4)); + __m128i nz_flag0 = _mm_cmpeq_epi32(q0, zero); + __m128i nz_flag1 = _mm_cmpeq_epi32(q1, zero); + + nz_flag0 = _mm_cmpeq_epi32(nz_flag0, zero); + nz_flag1 = _mm_cmpeq_epi32(nz_flag1, zero); + + mask = _mm_packs_epi32(nz_flag0, nz_flag1); + iscanIdx = _mm_loadu_si128((__m128i const *)iscan); + iscanIdx = _mm_sub_epi16(iscanIdx, mask); + iscanIdx = _mm_and_si128(iscanIdx, mask); + *eob = _mm_max_epi16(*eob, iscanIdx); +} + +static INLINE uint16_t get_accumulated_eob(__m128i *eob) { + __m128i eob_shuffled; + uint16_t eobValue; + eob_shuffled = _mm_shuffle_epi32(*eob, 0xe); + *eob = _mm_max_epi16(*eob, eob_shuffled); + eob_shuffled = _mm_shufflelo_epi16(*eob, 0xe); + *eob = _mm_max_epi16(*eob, eob_shuffled); + eob_shuffled = _mm_shufflelo_epi16(*eob, 0x1); + *eob = _mm_max_epi16(*eob, eob_shuffled); + eobValue = _mm_extract_epi16(*eob, 0); + return eobValue; +} + +void av1_highbd_quantize_fp_sse4_1( + const tran_low_t *coeff_ptr, intptr_t count, const int16_t *zbin_ptr, + const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan, int log_scale) { + __m128i coeff[2], qcoeff[3], dequant[2], qparam[4], coeff_sign; + __m128i eob = _mm_setzero_si128(); + const tran_low_t *src = coeff_ptr; + tran_low_t *quanAddr = qcoeff_ptr; + tran_low_t *dquanAddr = dqcoeff_ptr; + const int shift = 16 - log_scale; + const int coeff_stride = 4; + const int quan_stride = coeff_stride; + (void)zbin_ptr; + (void)quant_shift_ptr; + (void)scan; + + memset(quanAddr, 0, count * sizeof(quanAddr[0])); + memset(dquanAddr, 0, count * sizeof(dquanAddr[0])); + + coeff[0] = _mm_loadu_si128((__m128i const *)src); + const int round1 = ROUND_POWER_OF_TWO(round_ptr[1], log_scale); + const int round0 = ROUND_POWER_OF_TWO(round_ptr[0], log_scale); + + qparam[0] = _mm_set_epi32(round1, round1, round1, round0); + qparam[1] = xx_set_64_from_32i(quant_ptr[1], quant_ptr[0]); + qparam[2] = xx_set_64_from_32i(dequant_ptr[1], dequant_ptr[0]); + qparam[3] = _mm_set_epi32(dequant_ptr[1], dequant_ptr[1], dequant_ptr[1], + dequant_ptr[0]); + + // DC and first 3 AC + quantize_coeff_phase1(&coeff[0], qparam, shift, log_scale, qcoeff, dequant, + &coeff_sign); + + // update round/quan/dquan for AC + qparam[0] = _mm_unpackhi_epi64(qparam[0], qparam[0]); + qparam[1] = xx_set1_64_from_32i(quant_ptr[1]); + qparam[2] = xx_set1_64_from_32i(dequant_ptr[1]); + qparam[3] = _mm_set1_epi32(dequant_ptr[1]); + quantize_coeff_phase2(qcoeff, dequant, &coeff_sign, qparam, shift, log_scale, + quanAddr, dquanAddr); + + // next 4 AC + coeff[1] = _mm_loadu_si128((__m128i const *)(src + coeff_stride)); + quantize_coeff_phase1(&coeff[1], qparam, shift, log_scale, qcoeff, dequant, + &coeff_sign); + quantize_coeff_phase2(qcoeff, dequant, &coeff_sign, qparam, shift, log_scale, + quanAddr + quan_stride, dquanAddr + quan_stride); + + find_eob(quanAddr, iscan, &eob); + + count -= 8; + + // loop for the rest of AC + while (count > 0) { + src += coeff_stride << 1; + quanAddr += quan_stride << 1; + dquanAddr += quan_stride << 1; + iscan += quan_stride << 1; + + coeff[0] = _mm_loadu_si128((__m128i const *)src); + coeff[1] = _mm_loadu_si128((__m128i const *)(src + coeff_stride)); + + quantize_coeff_phase1(&coeff[0], qparam, shift, log_scale, qcoeff, dequant, + &coeff_sign); + quantize_coeff_phase2(qcoeff, dequant, &coeff_sign, qparam, shift, + log_scale, quanAddr, dquanAddr); + + quantize_coeff_phase1(&coeff[1], qparam, shift, log_scale, qcoeff, dequant, + &coeff_sign); + quantize_coeff_phase2(qcoeff, dequant, &coeff_sign, qparam, shift, + log_scale, quanAddr + quan_stride, + dquanAddr + quan_stride); + + find_eob(quanAddr, iscan, &eob); + + count -= 8; + } + *eob_ptr = get_accumulated_eob(&eob); +} diff --git a/third_party/aom/av1/encoder/x86/av1_quantize_avx2.c b/third_party/aom/av1/encoder/x86/av1_quantize_avx2.c new file mode 100644 index 0000000000..df22aaba7c --- /dev/null +++ b/third_party/aom/av1/encoder/x86/av1_quantize_avx2.c @@ -0,0 +1,330 @@ +/* + * Copyright (c) 2017, 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 <immintrin.h> + +#include "config/av1_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_dsp/aom_dsp_common.h" + +static INLINE void read_coeff(const tran_low_t *coeff, __m256i *c) { + if (sizeof(tran_low_t) == 4) { + const __m256i x0 = _mm256_loadu_si256((const __m256i *)coeff); + const __m256i x1 = _mm256_loadu_si256((const __m256i *)coeff + 1); + *c = _mm256_packs_epi32(x0, x1); + *c = _mm256_permute4x64_epi64(*c, 0xD8); + } else { + *c = _mm256_loadu_si256((const __m256i *)coeff); + } +} + +static INLINE void write_zero(tran_low_t *qcoeff) { + const __m256i zero = _mm256_setzero_si256(); + if (sizeof(tran_low_t) == 4) { + _mm256_storeu_si256((__m256i *)qcoeff, zero); + _mm256_storeu_si256((__m256i *)qcoeff + 1, zero); + } else { + _mm256_storeu_si256((__m256i *)qcoeff, zero); + } +} + +static INLINE void init_one_qp(const __m128i *p, __m256i *qp) { + const __m128i ac = _mm_unpackhi_epi64(*p, *p); + *qp = _mm256_insertf128_si256(_mm256_castsi128_si256(*p), ac, 1); +} + +static INLINE void init_qp(const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *dequant_ptr, int log_scale, + __m256i *thr, __m256i *qp) { + __m128i round = _mm_loadu_si128((const __m128i *)round_ptr); + const __m128i quant = _mm_loadu_si128((const __m128i *)quant_ptr); + const __m128i dequant = _mm_loadu_si128((const __m128i *)dequant_ptr); + + if (log_scale > 0) { + const __m128i rnd = _mm_set1_epi16((int16_t)1 << (log_scale - 1)); + round = _mm_add_epi16(round, rnd); + round = _mm_srai_epi16(round, log_scale); + } + + init_one_qp(&round, &qp[0]); + init_one_qp(&quant, &qp[1]); + + if (log_scale == 1) { + qp[1] = _mm256_slli_epi16(qp[1], log_scale); + } + + init_one_qp(&dequant, &qp[2]); + *thr = _mm256_srai_epi16(qp[2], 1 + log_scale); +} + +static INLINE void update_qp(int log_scale, __m256i *thr, __m256i *qp) { + qp[0] = _mm256_permute2x128_si256(qp[0], qp[0], 0x11); + qp[1] = _mm256_permute2x128_si256(qp[1], qp[1], 0x11); + qp[2] = _mm256_permute2x128_si256(qp[2], qp[2], 0x11); + *thr = _mm256_srai_epi16(qp[2], 1 + log_scale); +} + +#define store_quan(q, addr) \ + do { \ + __m256i sign_bits = _mm256_srai_epi16(q, 15); \ + __m256i y0 = _mm256_unpacklo_epi16(q, sign_bits); \ + __m256i y1 = _mm256_unpackhi_epi16(q, sign_bits); \ + __m256i x0 = _mm256_permute2x128_si256(y0, y1, 0x20); \ + __m256i x1 = _mm256_permute2x128_si256(y0, y1, 0x31); \ + _mm256_storeu_si256((__m256i *)addr, x0); \ + _mm256_storeu_si256((__m256i *)addr + 1, x1); \ + } while (0) + +#define store_two_quan(q, addr1, dq, addr2) \ + do { \ + if (sizeof(tran_low_t) == 4) { \ + store_quan(q, addr1); \ + store_quan(dq, addr2); \ + } else { \ + _mm256_storeu_si256((__m256i *)addr1, q); \ + _mm256_storeu_si256((__m256i *)addr2, dq); \ + } \ + } while (0) + +static INLINE uint16_t quant_gather_eob(__m256i eob) { + const __m128i eob_lo = _mm256_castsi256_si128(eob); + const __m128i eob_hi = _mm256_extractf128_si256(eob, 1); + __m128i eob_s = _mm_max_epi16(eob_lo, eob_hi); + eob_s = _mm_subs_epu16(_mm_set1_epi16(INT16_MAX), eob_s); + eob_s = _mm_minpos_epu16(eob_s); + return INT16_MAX - _mm_extract_epi16(eob_s, 0); +} + +static INLINE void quantize(const __m256i *thr, const __m256i *qp, __m256i *c, + const int16_t *iscan_ptr, tran_low_t *qcoeff, + tran_low_t *dqcoeff, __m256i *eob) { + const __m256i abs_coeff = _mm256_abs_epi16(*c); + __m256i mask = _mm256_cmpgt_epi16(abs_coeff, *thr); + mask = _mm256_or_si256(mask, _mm256_cmpeq_epi16(abs_coeff, *thr)); + const int nzflag = _mm256_movemask_epi8(mask); + + if (nzflag) { + __m256i q = _mm256_adds_epi16(abs_coeff, qp[0]); + q = _mm256_mulhi_epi16(q, qp[1]); + q = _mm256_sign_epi16(q, *c); + const __m256i dq = _mm256_mullo_epi16(q, qp[2]); + + store_two_quan(q, qcoeff, dq, dqcoeff); + const __m256i zero = _mm256_setzero_si256(); + const __m256i iscan = _mm256_loadu_si256((const __m256i *)iscan_ptr); + const __m256i zero_coeff = _mm256_cmpeq_epi16(dq, zero); + const __m256i nzero_coeff = _mm256_cmpeq_epi16(zero_coeff, zero); + __m256i cur_eob = _mm256_sub_epi16(iscan, nzero_coeff); + cur_eob = _mm256_and_si256(cur_eob, nzero_coeff); + *eob = _mm256_max_epi16(*eob, cur_eob); + } else { + write_zero(qcoeff); + write_zero(dqcoeff); + } +} + +void av1_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, + const int16_t *zbin_ptr, const int16_t *round_ptr, + const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, + tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, + const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan_ptr, const int16_t *iscan_ptr) { + (void)scan_ptr; + (void)zbin_ptr; + (void)quant_shift_ptr; + const unsigned int step = 16; + + __m256i qp[3]; + __m256i coeff, thr; + const int log_scale = 0; + + init_qp(round_ptr, quant_ptr, dequant_ptr, log_scale, &thr, qp); + read_coeff(coeff_ptr, &coeff); + + __m256i eob = _mm256_setzero_si256(); + quantize(&thr, qp, &coeff, iscan_ptr, qcoeff_ptr, dqcoeff_ptr, &eob); + + coeff_ptr += step; + qcoeff_ptr += step; + dqcoeff_ptr += step; + iscan_ptr += step; + n_coeffs -= step; + + update_qp(log_scale, &thr, qp); + + while (n_coeffs > 0) { + read_coeff(coeff_ptr, &coeff); + quantize(&thr, qp, &coeff, iscan_ptr, qcoeff_ptr, dqcoeff_ptr, &eob); + + coeff_ptr += step; + qcoeff_ptr += step; + dqcoeff_ptr += step; + iscan_ptr += step; + n_coeffs -= step; + } + *eob_ptr = quant_gather_eob(eob); +} + +static INLINE void quantize_32x32(const __m256i *thr, const __m256i *qp, + __m256i *c, const int16_t *iscan_ptr, + tran_low_t *qcoeff, tran_low_t *dqcoeff, + __m256i *eob) { + const __m256i abs_coeff = _mm256_abs_epi16(*c); + __m256i mask = _mm256_cmpgt_epi16(abs_coeff, *thr); + mask = _mm256_or_si256(mask, _mm256_cmpeq_epi16(abs_coeff, *thr)); + const int nzflag = _mm256_movemask_epi8(mask); + + if (nzflag) { + __m256i q = _mm256_adds_epi16(abs_coeff, qp[0]); + q = _mm256_mulhi_epu16(q, qp[1]); + + __m256i dq = _mm256_mullo_epi16(q, qp[2]); + dq = _mm256_srli_epi16(dq, 1); + + q = _mm256_sign_epi16(q, *c); + dq = _mm256_sign_epi16(dq, *c); + + store_two_quan(q, qcoeff, dq, dqcoeff); + const __m256i zero = _mm256_setzero_si256(); + const __m256i iscan = _mm256_loadu_si256((const __m256i *)iscan_ptr); + const __m256i zero_coeff = _mm256_cmpeq_epi16(dq, zero); + const __m256i nzero_coeff = _mm256_cmpeq_epi16(zero_coeff, zero); + __m256i cur_eob = _mm256_sub_epi16(iscan, nzero_coeff); + cur_eob = _mm256_and_si256(cur_eob, nzero_coeff); + *eob = _mm256_max_epi16(*eob, cur_eob); + } else { + write_zero(qcoeff); + write_zero(dqcoeff); + } +} + +void av1_quantize_fp_32x32_avx2( + const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, + const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan_ptr, const int16_t *iscan_ptr) { + (void)scan_ptr; + (void)zbin_ptr; + (void)quant_shift_ptr; + const unsigned int step = 16; + + __m256i qp[3]; + __m256i coeff, thr; + const int log_scale = 1; + + init_qp(round_ptr, quant_ptr, dequant_ptr, log_scale, &thr, qp); + read_coeff(coeff_ptr, &coeff); + + __m256i eob = _mm256_setzero_si256(); + quantize_32x32(&thr, qp, &coeff, iscan_ptr, qcoeff_ptr, dqcoeff_ptr, &eob); + + coeff_ptr += step; + qcoeff_ptr += step; + dqcoeff_ptr += step; + iscan_ptr += step; + n_coeffs -= step; + + update_qp(log_scale, &thr, qp); + + while (n_coeffs > 0) { + read_coeff(coeff_ptr, &coeff); + quantize_32x32(&thr, qp, &coeff, iscan_ptr, qcoeff_ptr, dqcoeff_ptr, &eob); + + coeff_ptr += step; + qcoeff_ptr += step; + dqcoeff_ptr += step; + iscan_ptr += step; + n_coeffs -= step; + } + *eob_ptr = quant_gather_eob(eob); +} + +static INLINE void quantize_64x64(const __m256i *thr, const __m256i *qp, + __m256i *c, const int16_t *iscan_ptr, + tran_low_t *qcoeff, tran_low_t *dqcoeff, + __m256i *eob) { + const __m256i abs_coeff = _mm256_abs_epi16(*c); + __m256i mask = _mm256_cmpgt_epi16(abs_coeff, *thr); + mask = _mm256_or_si256(mask, _mm256_cmpeq_epi16(abs_coeff, *thr)); + const int nzflag = _mm256_movemask_epi8(mask); + + if (nzflag) { + __m256i q = _mm256_adds_epi16(abs_coeff, qp[0]); + __m256i qh = _mm256_mulhi_epi16(q, qp[1]); + __m256i ql = _mm256_mullo_epi16(q, qp[1]); + qh = _mm256_slli_epi16(qh, 2); + ql = _mm256_srli_epi16(ql, 14); + q = _mm256_or_si256(qh, ql); + const __m256i dqh = _mm256_slli_epi16(_mm256_mulhi_epi16(q, qp[2]), 14); + const __m256i dql = _mm256_srli_epi16(_mm256_mullo_epi16(q, qp[2]), 2); + __m256i dq = _mm256_or_si256(dqh, dql); + + q = _mm256_sign_epi16(q, *c); + dq = _mm256_sign_epi16(dq, *c); + + store_two_quan(q, qcoeff, dq, dqcoeff); + const __m256i zero = _mm256_setzero_si256(); + const __m256i iscan = _mm256_loadu_si256((const __m256i *)iscan_ptr); + const __m256i zero_coeff = _mm256_cmpeq_epi16(dq, zero); + const __m256i nzero_coeff = _mm256_cmpeq_epi16(zero_coeff, zero); + __m256i cur_eob = _mm256_sub_epi16(iscan, nzero_coeff); + cur_eob = _mm256_and_si256(cur_eob, nzero_coeff); + *eob = _mm256_max_epi16(*eob, cur_eob); + } else { + write_zero(qcoeff); + write_zero(dqcoeff); + } +} + +void av1_quantize_fp_64x64_avx2( + const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, + const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan_ptr, const int16_t *iscan_ptr) { + (void)scan_ptr; + (void)zbin_ptr; + (void)quant_shift_ptr; + const unsigned int step = 16; + + __m256i qp[3]; + __m256i coeff, thr; + const int log_scale = 2; + + init_qp(round_ptr, quant_ptr, dequant_ptr, log_scale, &thr, qp); + read_coeff(coeff_ptr, &coeff); + + __m256i eob = _mm256_setzero_si256(); + quantize_64x64(&thr, qp, &coeff, iscan_ptr, qcoeff_ptr, dqcoeff_ptr, &eob); + + coeff_ptr += step; + qcoeff_ptr += step; + dqcoeff_ptr += step; + iscan_ptr += step; + n_coeffs -= step; + + update_qp(log_scale, &thr, qp); + + while (n_coeffs > 0) { + read_coeff(coeff_ptr, &coeff); + quantize_64x64(&thr, qp, &coeff, iscan_ptr, qcoeff_ptr, dqcoeff_ptr, &eob); + + coeff_ptr += step; + qcoeff_ptr += step; + dqcoeff_ptr += step; + iscan_ptr += step; + n_coeffs -= step; + } + *eob_ptr = quant_gather_eob(eob); +} diff --git a/third_party/aom/av1/encoder/x86/av1_quantize_sse2.c b/third_party/aom/av1/encoder/x86/av1_quantize_sse2.c new file mode 100644 index 0000000000..b07e7717f3 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/av1_quantize_sse2.c @@ -0,0 +1,189 @@ +/* + * Copyright (c) 2016, 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 <emmintrin.h> +#include <xmmintrin.h> + +#include "config/av1_rtcd.h" + +#include "aom/aom_integer.h" + +static INLINE void read_coeff(const tran_low_t *coeff, intptr_t offset, + __m128i *c0, __m128i *c1) { + const tran_low_t *addr = coeff + offset; + if (sizeof(tran_low_t) == 4) { + const __m128i x0 = _mm_load_si128((const __m128i *)addr); + const __m128i x1 = _mm_load_si128((const __m128i *)addr + 1); + const __m128i x2 = _mm_load_si128((const __m128i *)addr + 2); + const __m128i x3 = _mm_load_si128((const __m128i *)addr + 3); + *c0 = _mm_packs_epi32(x0, x1); + *c1 = _mm_packs_epi32(x2, x3); + } else { + *c0 = _mm_load_si128((const __m128i *)addr); + *c1 = _mm_load_si128((const __m128i *)addr + 1); + } +} + +static INLINE void write_qcoeff(const __m128i *qc0, const __m128i *qc1, + tran_low_t *qcoeff, intptr_t offset) { + tran_low_t *addr = qcoeff + offset; + if (sizeof(tran_low_t) == 4) { + const __m128i zero = _mm_setzero_si128(); + __m128i sign_bits = _mm_cmplt_epi16(*qc0, zero); + __m128i y0 = _mm_unpacklo_epi16(*qc0, sign_bits); + __m128i y1 = _mm_unpackhi_epi16(*qc0, sign_bits); + _mm_store_si128((__m128i *)addr, y0); + _mm_store_si128((__m128i *)addr + 1, y1); + + sign_bits = _mm_cmplt_epi16(*qc1, zero); + y0 = _mm_unpacklo_epi16(*qc1, sign_bits); + y1 = _mm_unpackhi_epi16(*qc1, sign_bits); + _mm_store_si128((__m128i *)addr + 2, y0); + _mm_store_si128((__m128i *)addr + 3, y1); + } else { + _mm_store_si128((__m128i *)addr, *qc0); + _mm_store_si128((__m128i *)addr + 1, *qc1); + } +} + +static INLINE void write_zero(tran_low_t *qcoeff, intptr_t offset) { + const __m128i zero = _mm_setzero_si128(); + tran_low_t *addr = qcoeff + offset; + if (sizeof(tran_low_t) == 4) { + _mm_store_si128((__m128i *)addr, zero); + _mm_store_si128((__m128i *)addr + 1, zero); + _mm_store_si128((__m128i *)addr + 2, zero); + _mm_store_si128((__m128i *)addr + 3, zero); + } else { + _mm_store_si128((__m128i *)addr, zero); + _mm_store_si128((__m128i *)addr + 1, zero); + } +} + +static INLINE void quantize(const int16_t *iscan_ptr, + const tran_low_t *coeff_ptr, intptr_t n_coeffs, + tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, + const __m128i *round0, const __m128i *round1, + const __m128i *quant0, const __m128i *quant1, + const __m128i *dequant0, const __m128i *dequant1, + const __m128i *thr0, const __m128i *thr1, + __m128i *eob) { + __m128i coeff0, coeff1; + // Do DC and first 15 AC + read_coeff(coeff_ptr, n_coeffs, &coeff0, &coeff1); + + // Poor man's sign extract + const __m128i coeff0_sign = _mm_srai_epi16(coeff0, 15); + const __m128i coeff1_sign = _mm_srai_epi16(coeff1, 15); + __m128i qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign); + __m128i qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign); + qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); + qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); + const __m128i mask0 = _mm_or_si128(_mm_cmpgt_epi16(qcoeff0, *thr0), + _mm_cmpeq_epi16(qcoeff0, *thr0)); + const __m128i mask1 = _mm_or_si128(_mm_cmpgt_epi16(qcoeff1, *thr1), + _mm_cmpeq_epi16(qcoeff1, *thr1)); + const int16_t nzflag = _mm_movemask_epi8(mask0) | _mm_movemask_epi8(mask1); + + if (nzflag) { + qcoeff0 = _mm_adds_epi16(qcoeff0, *round0); + qcoeff1 = _mm_adds_epi16(qcoeff1, *round1); + const __m128i qtmp0 = _mm_mulhi_epi16(qcoeff0, *quant0); + const __m128i qtmp1 = _mm_mulhi_epi16(qcoeff1, *quant1); + + // Reinsert signs + qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign); + qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign); + qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); + qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); + + write_qcoeff(&qcoeff0, &qcoeff1, qcoeff_ptr, n_coeffs); + + coeff0 = _mm_mullo_epi16(qcoeff0, *dequant0); + coeff1 = _mm_mullo_epi16(qcoeff1, *dequant1); + + write_qcoeff(&coeff0, &coeff1, dqcoeff_ptr, n_coeffs); + + const __m128i zero = _mm_setzero_si128(); + // Scan for eob + const __m128i zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); + const __m128i zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); + const __m128i nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero); + const __m128i nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero); + const __m128i iscan0 = + _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs)); + const __m128i iscan1 = + _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1); + // Add one to convert from indices to counts + const __m128i iscan0_nz = _mm_sub_epi16(iscan0, nzero_coeff0); + const __m128i iscan1_nz = _mm_sub_epi16(iscan1, nzero_coeff1); + const __m128i eob0 = _mm_and_si128(iscan0_nz, nzero_coeff0); + const __m128i eob1 = _mm_and_si128(iscan1_nz, nzero_coeff1); + const __m128i eob2 = _mm_max_epi16(eob0, eob1); + *eob = _mm_max_epi16(*eob, eob2); + } else { + write_zero(qcoeff_ptr, n_coeffs); + write_zero(dqcoeff_ptr, n_coeffs); + } +} + +void av1_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, + const int16_t *zbin_ptr, const int16_t *round_ptr, + const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, + tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, + const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan_ptr, const int16_t *iscan_ptr) { + (void)scan_ptr; + (void)zbin_ptr; + (void)quant_shift_ptr; + + coeff_ptr += n_coeffs; + iscan_ptr += n_coeffs; + qcoeff_ptr += n_coeffs; + dqcoeff_ptr += n_coeffs; + n_coeffs = -n_coeffs; + + const __m128i round0 = _mm_load_si128((const __m128i *)round_ptr); + const __m128i round1 = _mm_unpackhi_epi64(round0, round0); + const __m128i quant0 = _mm_load_si128((const __m128i *)quant_ptr); + const __m128i quant1 = _mm_unpackhi_epi64(quant0, quant0); + const __m128i dequant0 = _mm_load_si128((const __m128i *)dequant_ptr); + const __m128i dequant1 = _mm_unpackhi_epi64(dequant0, dequant0); + const __m128i thr0 = _mm_srai_epi16(dequant0, 1); + const __m128i thr1 = _mm_srai_epi16(dequant1, 1); + __m128i eob = _mm_setzero_si128(); + + quantize(iscan_ptr, coeff_ptr, n_coeffs, qcoeff_ptr, dqcoeff_ptr, &round0, + &round1, &quant0, &quant1, &dequant0, &dequant1, &thr0, &thr1, &eob); + + n_coeffs += 8 * 2; + + // AC only loop + while (n_coeffs < 0) { + quantize(iscan_ptr, coeff_ptr, n_coeffs, qcoeff_ptr, dqcoeff_ptr, &round1, + &round1, &quant1, &quant1, &dequant1, &dequant1, &thr1, &thr1, + &eob); + n_coeffs += 8 * 2; + } + + // Accumulate EOB + { + __m128i eob_shuffled; + eob_shuffled = _mm_shuffle_epi32(eob, 0xe); + eob = _mm_max_epi16(eob, eob_shuffled); + eob_shuffled = _mm_shufflelo_epi16(eob, 0xe); + eob = _mm_max_epi16(eob, eob_shuffled); + eob_shuffled = _mm_shufflelo_epi16(eob, 0x1); + eob = _mm_max_epi16(eob, eob_shuffled); + *eob_ptr = _mm_extract_epi16(eob, 1); + } +} diff --git a/third_party/aom/av1/encoder/x86/av1_quantize_ssse3_x86_64.asm b/third_party/aom/av1/encoder/x86/av1_quantize_ssse3_x86_64.asm new file mode 100644 index 0000000000..ad4ae274e2 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/av1_quantize_ssse3_x86_64.asm @@ -0,0 +1,204 @@ +; +; Copyright (c) 2016, 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. +; + +; + +%define private_prefix av1 + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA +pw_1: times 8 dw 1 + +SECTION .text + +%macro QUANTIZE_FP 2 +cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \ + shift, qcoeff, dqcoeff, dequant, \ + eob, scan, iscan + cmp dword skipm, 0 + jne .blank + + ; actual quantize loop - setup pointers, rounders, etc. + movifnidn coeffq, coeffmp + movifnidn ncoeffq, ncoeffmp + mov r2, dequantmp + movifnidn zbinq, zbinmp + movifnidn roundq, roundmp + movifnidn quantq, quantmp + mova m1, [roundq] ; m1 = round + mova m2, [quantq] ; m2 = quant +%ifidn %1, fp_32x32 + pcmpeqw m5, m5 + psrlw m5, 15 + paddw m1, m5 + psrlw m1, 1 ; m1 = (m1 + 1) / 2 +%endif + mova m3, [r2q] ; m3 = dequant + mov r3, qcoeffmp + mov r4, dqcoeffmp + mov r5, iscanmp +%ifidn %1, fp_32x32 + psllw m2, 1 +%endif + pxor m5, m5 ; m5 = dedicated zero + + lea coeffq, [ coeffq+ncoeffq*2] + lea r5q, [ r5q+ncoeffq*2] + lea r3q, [ r3q+ncoeffq*2] + lea r4q, [r4q+ncoeffq*2] + neg ncoeffq + + ; get DC and first 15 AC coeffs + mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i] + mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i] + pabsw m6, m9 ; m6 = abs(m9) + pabsw m11, m10 ; m11 = abs(m10) + pcmpeqw m7, m7 + + paddsw m6, m1 ; m6 += round + punpckhqdq m1, m1 + paddsw m11, m1 ; m11 += round + pmulhw m8, m6, m2 ; m8 = m6*q>>16 + punpckhqdq m2, m2 + pmulhw m13, m11, m2 ; m13 = m11*q>>16 + psignw m8, m9 ; m8 = reinsert sign + psignw m13, m10 ; m13 = reinsert sign + mova [r3q+ncoeffq*2+ 0], m8 + mova [r3q+ncoeffq*2+16], m13 +%ifidn %1, fp_32x32 + pabsw m8, m8 + pabsw m13, m13 +%endif + pmullw m8, m3 ; r4[i] = r3[i] * q + punpckhqdq m3, m3 + pmullw m13, m3 ; r4[i] = r3[i] * q +%ifidn %1, fp_32x32 + psrlw m8, 1 + psrlw m13, 1 + psignw m8, m9 + psignw m13, m10 + psrlw m0, m3, 2 +%else + psrlw m0, m3, 1 +%endif + mova [r4q+ncoeffq*2+ 0], m8 + mova [r4q+ncoeffq*2+16], m13 + pcmpeqw m8, m5 ; m8 = c[i] == 0 + pcmpeqw m13, m5 ; m13 = c[i] == 0 + mova m6, [ r5q+ncoeffq*2+ 0] ; m6 = scan[i] + mova m11, [ r5q+ncoeffq*2+16] ; m11 = scan[i] + psubw m6, m7 ; m6 = scan[i] + 1 + psubw m11, m7 ; m11 = scan[i] + 1 + pandn m8, m6 ; m8 = max(eob) + pandn m13, m11 ; m13 = max(eob) + pmaxsw m8, m13 + add ncoeffq, mmsize + jz .accumulate_eob + +.ac_only_loop: + mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i] + mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i] + pabsw m6, m9 ; m6 = abs(m9) + pabsw m11, m10 ; m11 = abs(m10) + + pcmpgtw m7, m6, m0 + pcmpgtw m12, m11, m0 + pmovmskb r6d, m7 + pmovmskb r2d, m12 + + or r6, r2 + jz .skip_iter + + pcmpeqw m7, m7 + + paddsw m6, m1 ; m6 += round + paddsw m11, m1 ; m11 += round + pmulhw m14, m6, m2 ; m14 = m6*q>>16 + pmulhw m13, m11, m2 ; m13 = m11*q>>16 + psignw m14, m9 ; m14 = reinsert sign + psignw m13, m10 ; m13 = reinsert sign + mova [r3q+ncoeffq*2+ 0], m14 + mova [r3q+ncoeffq*2+16], m13 +%ifidn %1, fp_32x32 + pabsw m14, m14 + pabsw m13, m13 +%endif + pmullw m14, m3 ; r4[i] = r3[i] * q + pmullw m13, m3 ; r4[i] = r3[i] * q +%ifidn %1, fp_32x32 + psrlw m14, 1 + psrlw m13, 1 + psignw m14, m9 + psignw m13, m10 +%endif + mova [r4q+ncoeffq*2+ 0], m14 + mova [r4q+ncoeffq*2+16], m13 + pcmpeqw m14, m5 ; m14 = c[i] == 0 + pcmpeqw m13, m5 ; m13 = c[i] == 0 + mova m6, [ r5q+ncoeffq*2+ 0] ; m6 = scan[i] + mova m11, [ r5q+ncoeffq*2+16] ; m11 = scan[i] + psubw m6, m7 ; m6 = scan[i] + 1 + psubw m11, m7 ; m11 = scan[i] + 1 + pandn m14, m6 ; m14 = max(eob) + pandn m13, m11 ; m13 = max(eob) + pmaxsw m8, m14 + pmaxsw m8, m13 + add ncoeffq, mmsize + jl .ac_only_loop + + jmp .accumulate_eob +.skip_iter: + mova [r3q+ncoeffq*2+ 0], m5 + mova [r3q+ncoeffq*2+16], m5 + mova [r4q+ncoeffq*2+ 0], m5 + mova [r4q+ncoeffq*2+16], m5 + add ncoeffq, mmsize + jl .ac_only_loop + +.accumulate_eob: + ; horizontally accumulate/max eobs and write into [eob] memory pointer + mov r2, eobmp + pshufd m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0x1 + pmaxsw m8, m7 + pextrw r6, m8, 0 + mov [r2], r6 + RET + + ; skip-block, i.e. just write all zeroes +.blank: + mov r0, dqcoeffmp + movifnidn ncoeffq, ncoeffmp + mov r2, qcoeffmp + mov r3, eobmp + + lea r0q, [r0q+ncoeffq*2] + lea r2q, [r2q+ncoeffq*2] + neg ncoeffq + pxor m7, m7 +.blank_loop: + mova [r0q+ncoeffq*2+ 0], m7 + mova [r0q+ncoeffq*2+16], m7 + mova [r2q+ncoeffq*2+ 0], m7 + mova [r2q+ncoeffq*2+16], m7 + add ncoeffq, mmsize + jl .blank_loop + mov word [r3q], 0 + RET +%endmacro + +INIT_XMM ssse3 +QUANTIZE_FP fp, 7 +QUANTIZE_FP fp_32x32, 7 diff --git a/third_party/aom/av1/encoder/x86/av1_ssim_opt_x86_64.asm b/third_party/aom/av1/encoder/x86/av1_ssim_opt_x86_64.asm new file mode 100644 index 0000000000..faa2a232a3 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/av1_ssim_opt_x86_64.asm @@ -0,0 +1,222 @@ +; +; Copyright (c) 2016, 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 "aom_ports/x86_abi_support.asm" + +; tabulate_ssim - sums sum_s,sum_r,sum_sq_s,sum_sq_r, sum_sxr +%macro TABULATE_SSIM 0 + paddusw xmm15, xmm3 ; sum_s + paddusw xmm14, xmm4 ; sum_r + movdqa xmm1, xmm3 + pmaddwd xmm1, xmm1 + paddd xmm13, xmm1 ; sum_sq_s + movdqa xmm2, xmm4 + pmaddwd xmm2, xmm2 + paddd xmm12, xmm2 ; sum_sq_r + pmaddwd xmm3, xmm4 + paddd xmm11, xmm3 ; sum_sxr +%endmacro + +; Sum across the register %1 starting with q words +%macro SUM_ACROSS_Q 1 + movdqa xmm2,%1 + punpckldq %1,xmm0 + punpckhdq xmm2,xmm0 + paddq %1,xmm2 + movdqa xmm2,%1 + punpcklqdq %1,xmm0 + punpckhqdq xmm2,xmm0 + paddq %1,xmm2 +%endmacro + +; Sum across the register %1 starting with q words +%macro SUM_ACROSS_W 1 + movdqa xmm1, %1 + punpcklwd %1,xmm0 + punpckhwd xmm1,xmm0 + paddd %1, xmm1 + SUM_ACROSS_Q %1 +%endmacro + +SECTION .text + +;void ssim_parms_sse2( +; unsigned char *s, +; int sp, +; unsigned char *r, +; int rp +; unsigned long *sum_s, +; unsigned long *sum_r, +; unsigned long *sum_sq_s, +; unsigned long *sum_sq_r, +; unsigned long *sum_sxr); +; +; TODO: Use parm passing through structure, probably don't need the pxors +; ( calling app will initialize to 0 ) could easily fit everything in sse2 +; without too much hastle, and can probably do better estimates with psadw +; or pavgb At this point this is just meant to be first pass for calculating +; all the parms needed for 16x16 ssim so we can play with dssim as distortion +; in mode selection code. +global sym(av1_ssim_parms_16x16_sse2) PRIVATE +sym(av1_ssim_parms_16x16_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 9 + SAVE_XMM 15 + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;s + mov rcx, arg(1) ;sp + mov rdi, arg(2) ;r + mov rax, arg(3) ;rp + + pxor xmm0, xmm0 + pxor xmm15,xmm15 ;sum_s + pxor xmm14,xmm14 ;sum_r + pxor xmm13,xmm13 ;sum_sq_s + pxor xmm12,xmm12 ;sum_sq_r + pxor xmm11,xmm11 ;sum_sxr + + mov rdx, 16 ;row counter +.NextRow: + + ;grab source and reference pixels + movdqu xmm5, [rsi] + movdqu xmm6, [rdi] + movdqa xmm3, xmm5 + movdqa xmm4, xmm6 + punpckhbw xmm3, xmm0 ; high_s + punpckhbw xmm4, xmm0 ; high_r + + TABULATE_SSIM + + movdqa xmm3, xmm5 + movdqa xmm4, xmm6 + punpcklbw xmm3, xmm0 ; low_s + punpcklbw xmm4, xmm0 ; low_r + + TABULATE_SSIM + + add rsi, rcx ; next s row + add rdi, rax ; next r row + + dec rdx ; counter + jnz .NextRow + + SUM_ACROSS_W xmm15 + SUM_ACROSS_W xmm14 + SUM_ACROSS_Q xmm13 + SUM_ACROSS_Q xmm12 + SUM_ACROSS_Q xmm11 + + mov rdi,arg(4) + movd [rdi], xmm15; + mov rdi,arg(5) + movd [rdi], xmm14; + mov rdi,arg(6) + movd [rdi], xmm13; + mov rdi,arg(7) + movd [rdi], xmm12; + mov rdi,arg(8) + movd [rdi], xmm11; + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void ssim_parms_sse2( +; unsigned char *s, +; int sp, +; unsigned char *r, +; int rp +; unsigned long *sum_s, +; unsigned long *sum_r, +; unsigned long *sum_sq_s, +; unsigned long *sum_sq_r, +; unsigned long *sum_sxr); +; +; TODO: Use parm passing through structure, probably don't need the pxors +; ( calling app will initialize to 0 ) could easily fit everything in sse2 +; without too much hastle, and can probably do better estimates with psadw +; or pavgb At this point this is just meant to be first pass for calculating +; all the parms needed for 16x16 ssim so we can play with dssim as distortion +; in mode selection code. +global sym(av1_ssim_parms_8x8_sse2) PRIVATE +sym(av1_ssim_parms_8x8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 9 + SAVE_XMM 15 + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;s + mov rcx, arg(1) ;sp + mov rdi, arg(2) ;r + mov rax, arg(3) ;rp + + pxor xmm0, xmm0 + pxor xmm15,xmm15 ;sum_s + pxor xmm14,xmm14 ;sum_r + pxor xmm13,xmm13 ;sum_sq_s + pxor xmm12,xmm12 ;sum_sq_r + pxor xmm11,xmm11 ;sum_sxr + + mov rdx, 8 ;row counter +.NextRow: + + ;grab source and reference pixels + movq xmm3, [rsi] + movq xmm4, [rdi] + punpcklbw xmm3, xmm0 ; low_s + punpcklbw xmm4, xmm0 ; low_r + + TABULATE_SSIM + + add rsi, rcx ; next s row + add rdi, rax ; next r row + + dec rdx ; counter + jnz .NextRow + + SUM_ACROSS_W xmm15 + SUM_ACROSS_W xmm14 + SUM_ACROSS_Q xmm13 + SUM_ACROSS_Q xmm12 + SUM_ACROSS_Q xmm11 + + mov rdi,arg(4) + movd [rdi], xmm15; + mov rdi,arg(5) + movd [rdi], xmm14; + mov rdi,arg(6) + movd [rdi], xmm13; + mov rdi,arg(7) + movd [rdi], xmm12; + mov rdi,arg(8) + movd [rdi], xmm11; + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/av1/encoder/x86/av1_txfm1d_sse4.h b/third_party/aom/av1/encoder/x86/av1_txfm1d_sse4.h new file mode 100644 index 0000000000..6df2a8bdbb --- /dev/null +++ b/third_party/aom/av1/encoder/x86/av1_txfm1d_sse4.h @@ -0,0 +1,142 @@ +/* + * Copyright (c) 2018, 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_ENCODER_X86_AV1_TXFM1D_SSE4_H_ +#define AOM_AV1_ENCODER_X86_AV1_TXFM1D_SSE4_H_ + +#include <smmintrin.h> +#include "av1/common/av1_txfm.h" +#include "av1/common/x86/av1_txfm_sse4.h" + +#ifdef __cplusplus +extern "C" { +#endif + +void av1_fdct4_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); +void av1_fdct8_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); +void av1_fdct16_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); +void av1_fdct32_new_sse4_1(const __m128i *input, __m128i *output, + int8_t cos_bit); +void av1_fdct64_new_sse4_1(const __m128i *input, __m128i *output, + int8_t cos_bit, const int instride, + const int outstride); + +void av1_fadst4_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); +void av1_fadst8_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); +void av1_fadst16_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); + +void av1_idct4_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); +void av1_idct8_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); +void av1_idct16_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); +void av1_idct32_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); +void av1_idct64_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); + +void av1_iadst4_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); +void av1_iadst8_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); +void av1_iadst16_new_sse4_1(const __m128i *input, __m128i *output, + const int8_t cos_bit, const int8_t *stage_range); +static INLINE void transpose_32_4x4(int stride, const __m128i *input, + __m128i *output) { + __m128i temp0 = _mm_unpacklo_epi32(input[0 * stride], input[2 * stride]); + __m128i temp1 = _mm_unpackhi_epi32(input[0 * stride], input[2 * stride]); + __m128i temp2 = _mm_unpacklo_epi32(input[1 * stride], input[3 * stride]); + __m128i temp3 = _mm_unpackhi_epi32(input[1 * stride], input[3 * stride]); + + output[0 * stride] = _mm_unpacklo_epi32(temp0, temp2); + output[1 * stride] = _mm_unpackhi_epi32(temp0, temp2); + output[2 * stride] = _mm_unpacklo_epi32(temp1, temp3); + output[3 * stride] = _mm_unpackhi_epi32(temp1, temp3); +} + +// the entire input block can be represent by a grid of 4x4 blocks +// each 4x4 blocks can be represent by 4 vertical __m128i +// we first transpose each 4x4 block internally +// then transpose the grid +static INLINE void transpose_32(int txfm_size, const __m128i *input, + __m128i *output) { + const int num_per_128 = 4; + const int row_size = txfm_size; + const int col_size = txfm_size / num_per_128; + int r, c; + + // transpose each 4x4 block internally + for (r = 0; r < row_size; r += 4) { + for (c = 0; c < col_size; c++) { + transpose_32_4x4(col_size, &input[r * col_size + c], + &output[c * 4 * col_size + r / 4]); + } + } +} + +// out0 = in0*w0 + in1*w1 +// out1 = -in1*w0 + in0*w1 +#define btf_32_sse4_1_type0(w0, w1, in0, in1, out0, out1, bit) \ + do { \ + const __m128i ww0 = _mm_set1_epi32(w0); \ + const __m128i ww1 = _mm_set1_epi32(w1); \ + const __m128i in0_w0 = _mm_mullo_epi32(in0, ww0); \ + const __m128i in1_w1 = _mm_mullo_epi32(in1, ww1); \ + out0 = _mm_add_epi32(in0_w0, in1_w1); \ + out0 = av1_round_shift_32_sse4_1(out0, bit); \ + const __m128i in0_w1 = _mm_mullo_epi32(in0, ww1); \ + const __m128i in1_w0 = _mm_mullo_epi32(in1, ww0); \ + out1 = _mm_sub_epi32(in0_w1, in1_w0); \ + out1 = av1_round_shift_32_sse4_1(out1, bit); \ + } while (0) + +// out0 = in0*w0 + in1*w1 +// out1 = in1*w0 - in0*w1 +#define btf_32_sse4_1_type1(w0, w1, in0, in1, out0, out1, bit) \ + do { \ + btf_32_sse4_1_type0(w1, w0, in1, in0, out0, out1, bit); \ + } while (0) + +// out0 = in0*w0 + in1*w1 +// out1 = -in1*w0 + in0*w1 +#define btf_32_type0_sse4_1_new(ww0, ww1, in0, in1, out0, out1, r, bit) \ + do { \ + const __m128i in0_w0 = _mm_mullo_epi32(in0, ww0); \ + const __m128i in1_w1 = _mm_mullo_epi32(in1, ww1); \ + out0 = _mm_add_epi32(in0_w0, in1_w1); \ + out0 = _mm_add_epi32(out0, r); \ + out0 = _mm_srai_epi32(out0, bit); \ + const __m128i in0_w1 = _mm_mullo_epi32(in0, ww1); \ + const __m128i in1_w0 = _mm_mullo_epi32(in1, ww0); \ + out1 = _mm_sub_epi32(in0_w1, in1_w0); \ + out1 = _mm_add_epi32(out1, r); \ + out1 = _mm_srai_epi32(out1, bit); \ + } while (0) + +// out0 = in0*w0 + in1*w1 +// out1 = in1*w0 - in0*w1 +#define btf_32_type1_sse4_1_new(ww0, ww1, in0, in1, out0, out1, r, bit) \ + do { \ + btf_32_type0_sse4_1_new(ww1, ww0, in1, in0, out0, out1, r, bit); \ + } while (0) + +#ifdef __cplusplus +} +#endif + +#endif // AOM_AV1_ENCODER_X86_AV1_TXFM1D_SSE4_H_ diff --git a/third_party/aom/av1/encoder/x86/corner_match_sse4.c b/third_party/aom/av1/encoder/x86/corner_match_sse4.c new file mode 100644 index 0000000000..93f37b71d3 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/corner_match_sse4.c @@ -0,0 +1,103 @@ +/* + * Copyright (c) 2018, 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 <stdlib.h> +#include <memory.h> +#include <math.h> +#include <assert.h> + +#include <smmintrin.h> + +#include "config/av1_rtcd.h" + +#include "aom_ports/mem.h" +#include "av1/encoder/corner_match.h" + +DECLARE_ALIGNED(16, static const uint8_t, byte_mask[16]) = { + 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0 +}; +#if MATCH_SZ != 13 +#error "Need to change byte_mask in corner_match_sse4.c if MATCH_SZ != 13" +#endif + +/* Compute corr(im1, im2) * MATCH_SZ * stddev(im1), where the + correlation/standard deviation are taken over MATCH_SZ by MATCH_SZ windows + of each image, centered at (x1, y1) and (x2, y2) respectively. +*/ +double compute_cross_correlation_sse4_1(unsigned char *im1, int stride1, int x1, + int y1, unsigned char *im2, int stride2, + int x2, int y2) { + int i; + // 2 16-bit partial sums in lanes 0, 4 (== 2 32-bit partial sums in lanes 0, + // 2) + __m128i sum1_vec = _mm_setzero_si128(); + __m128i sum2_vec = _mm_setzero_si128(); + // 4 32-bit partial sums of squares + __m128i sumsq2_vec = _mm_setzero_si128(); + __m128i cross_vec = _mm_setzero_si128(); + + const __m128i mask = _mm_load_si128((__m128i *)byte_mask); + const __m128i zero = _mm_setzero_si128(); + + im1 += (y1 - MATCH_SZ_BY2) * stride1 + (x1 - MATCH_SZ_BY2); + im2 += (y2 - MATCH_SZ_BY2) * stride2 + (x2 - MATCH_SZ_BY2); + + for (i = 0; i < MATCH_SZ; ++i) { + const __m128i v1 = + _mm_and_si128(_mm_loadu_si128((__m128i *)&im1[i * stride1]), mask); + const __m128i v2 = + _mm_and_si128(_mm_loadu_si128((__m128i *)&im2[i * stride2]), mask); + + // Using the 'sad' intrinsic here is a bit faster than adding + // v1_l + v1_r and v2_l + v2_r, plus it avoids the need for a 16->32 bit + // conversion step later, for a net speedup of ~10% + sum1_vec = _mm_add_epi16(sum1_vec, _mm_sad_epu8(v1, zero)); + sum2_vec = _mm_add_epi16(sum2_vec, _mm_sad_epu8(v2, zero)); + + const __m128i v1_l = _mm_cvtepu8_epi16(v1); + const __m128i v1_r = _mm_cvtepu8_epi16(_mm_srli_si128(v1, 8)); + const __m128i v2_l = _mm_cvtepu8_epi16(v2); + const __m128i v2_r = _mm_cvtepu8_epi16(_mm_srli_si128(v2, 8)); + + sumsq2_vec = _mm_add_epi32( + sumsq2_vec, + _mm_add_epi32(_mm_madd_epi16(v2_l, v2_l), _mm_madd_epi16(v2_r, v2_r))); + cross_vec = _mm_add_epi32( + cross_vec, + _mm_add_epi32(_mm_madd_epi16(v1_l, v2_l), _mm_madd_epi16(v1_r, v2_r))); + } + + // Now we can treat the four registers (sum1_vec, sum2_vec, sumsq2_vec, + // cross_vec) + // as holding 4 32-bit elements each, which we want to sum horizontally. + // We do this by transposing and then summing vertically. + __m128i tmp_0 = _mm_unpacklo_epi32(sum1_vec, sum2_vec); + __m128i tmp_1 = _mm_unpackhi_epi32(sum1_vec, sum2_vec); + __m128i tmp_2 = _mm_unpacklo_epi32(sumsq2_vec, cross_vec); + __m128i tmp_3 = _mm_unpackhi_epi32(sumsq2_vec, cross_vec); + + __m128i tmp_4 = _mm_unpacklo_epi64(tmp_0, tmp_2); + __m128i tmp_5 = _mm_unpackhi_epi64(tmp_0, tmp_2); + __m128i tmp_6 = _mm_unpacklo_epi64(tmp_1, tmp_3); + __m128i tmp_7 = _mm_unpackhi_epi64(tmp_1, tmp_3); + + __m128i res = + _mm_add_epi32(_mm_add_epi32(tmp_4, tmp_5), _mm_add_epi32(tmp_6, tmp_7)); + + int sum1 = _mm_extract_epi32(res, 0); + int sum2 = _mm_extract_epi32(res, 1); + int sumsq2 = _mm_extract_epi32(res, 2); + int cross = _mm_extract_epi32(res, 3); + + int var2 = sumsq2 * MATCH_SZ_SQ - sum2 * sum2; + int cov = cross * MATCH_SZ_SQ - sum1 * sum2; + return cov / sqrt((double)var2); +} diff --git a/third_party/aom/av1/encoder/x86/dct_sse2.asm b/third_party/aom/av1/encoder/x86/dct_sse2.asm new file mode 100644 index 0000000000..b185548184 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/dct_sse2.asm @@ -0,0 +1,82 @@ +; +; Copyright (c) 2016, 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. +; + +%define private_prefix av1 + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +%macro TRANSFORM_COLS 0 + paddw m0, m1 + movq m4, m0 + psubw m3, m2 + psubw m4, m3 + psraw m4, 1 + movq m5, m4 + psubw m5, m1 ;b1 + psubw m4, m2 ;c1 + psubw m0, m4 + paddw m3, m5 + ; m0 a0 + SWAP 1, 4 ; m1 c1 + SWAP 2, 3 ; m2 d1 + SWAP 3, 5 ; m3 b1 +%endmacro + +%macro TRANSPOSE_4X4 0 + ; 00 01 02 03 + ; 10 11 12 13 + ; 20 21 22 23 + ; 30 31 32 33 + punpcklwd m0, m1 ; 00 10 01 11 02 12 03 13 + punpcklwd m2, m3 ; 20 30 21 31 22 32 23 33 + mova m1, m0 + punpckldq m0, m2 ; 00 10 20 30 01 11 21 31 + punpckhdq m1, m2 ; 02 12 22 32 03 13 23 33 +%endmacro + +INIT_XMM sse2 +cglobal fwht4x4, 3, 4, 8, input, output, stride + lea r3q, [inputq + strideq*4] + movq m0, [inputq] ;a1 + movq m1, [inputq + strideq*2] ;b1 + movq m2, [r3q] ;c1 + movq m3, [r3q + strideq*2] ;d1 + + TRANSFORM_COLS + TRANSPOSE_4X4 + SWAP 1, 2 + psrldq m1, m0, 8 + psrldq m3, m2, 8 + TRANSFORM_COLS + TRANSPOSE_4X4 + + psllw m0, 2 + psllw m1, 2 + + ; sign extension + mova m2, m0 + mova m3, m1 + punpcklwd m0, m0 + punpcklwd m1, m1 + punpckhwd m2, m2 + punpckhwd m3, m3 + psrad m0, 16 + psrad m1, 16 + psrad m2, 16 + psrad m3, 16 + mova [outputq], m0 + mova [outputq + 16], m2 + mova [outputq + 32], m1 + mova [outputq + 48], m3 + + RET diff --git a/third_party/aom/av1/encoder/x86/encodetxb_avx2.c b/third_party/aom/av1/encoder/x86/encodetxb_avx2.c new file mode 100644 index 0000000000..7642f57d18 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/encodetxb_avx2.c @@ -0,0 +1,130 @@ +/* + * Copyright (c) 2018, 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 <emmintrin.h> // SSE2 +#include <smmintrin.h> /* SSE4.1 */ +#include <immintrin.h> /* AVX2 */ + +#include "aom/aom_integer.h" +#include "aom_dsp/x86/mem_sse2.h" +#include "av1/common/onyxc_int.h" +#include "av1/common/txb_common.h" +#include "aom_dsp/x86/synonyms.h" +#include "aom_dsp/x86/synonyms_avx2.h" + +void av1_txb_init_levels_avx2(const tran_low_t *const coeff, const int width, + const int height, uint8_t *const levels) { + const int stride = width + TX_PAD_HOR; + const __m256i y_zeros = _mm256_setzero_si256(); + + const int32_t pre_len = sizeof(*levels) * TX_PAD_TOP * stride; + uint8_t *pre_buf = levels - TX_PAD_TOP * stride; + uint8_t *pre_buf_end = pre_buf + pre_len; + do { + yy_storeu_256(pre_buf, y_zeros); + pre_buf += 32; + } while (pre_buf < pre_buf_end); + + const int32_t bottom_len = sizeof(*levels) * (TX_PAD_BOTTOM * stride); + uint8_t *bottom_buf_end = levels + (height + TX_PAD_BOTTOM) * stride; + uint8_t *bottom_buf = bottom_buf_end - ((bottom_len + 31) & (~31)); + + do { + yy_storeu_256(bottom_buf, y_zeros); + bottom_buf += 32; + } while (bottom_buf < bottom_buf_end); + + int i = 0; + uint8_t *ls = levels; + const tran_low_t *cf = coeff; + if (width == 4) { + do { + const __m256i c0 = yy_loadu_256(cf); + const __m256i c1 = yy_loadu_256(cf + 8); + const __m256i abs01 = _mm256_abs_epi16(_mm256_packs_epi32(c0, c1)); + const __m256i abs01_8 = _mm256_packs_epi16(abs01, y_zeros); + const __m256i res_ = _mm256_shuffle_epi32(abs01_8, 0xd8); + const __m256i res = _mm256_permute4x64_epi64(res_, 0xd8); + yy_storeu_256(ls, res); + ls += 32; + cf += 16; + i += 4; + } while (i < height); + } else if (width == 8) { + do { + const __m256i coeffA = yy_loadu_256(cf); + const __m256i coeffB = yy_loadu_256(cf + 8); + const __m256i coeffC = yy_loadu_256(cf + 16); + const __m256i coeffD = yy_loadu_256(cf + 24); + const __m256i coeffAB = _mm256_packs_epi32(coeffA, coeffB); + const __m256i coeffCD = _mm256_packs_epi32(coeffC, coeffD); + const __m256i absAB = _mm256_abs_epi16(coeffAB); + const __m256i absCD = _mm256_abs_epi16(coeffCD); + const __m256i absABCD = _mm256_packs_epi16(absAB, absCD); + const __m256i res_ = _mm256_permute4x64_epi64(absABCD, 0xd8); + const __m256i res = _mm256_shuffle_epi32(res_, 0xd8); + const __m128i res0 = _mm256_castsi256_si128(res); + const __m128i res1 = _mm256_extracti128_si256(res, 1); + xx_storel_64(ls, res0); + *(int32_t *)(ls + width) = 0; + xx_storel_64(ls + stride, _mm_srli_si128(res0, 8)); + *(int32_t *)(ls + width + stride) = 0; + xx_storel_64(ls + stride * 2, res1); + *(int32_t *)(ls + width + stride * 2) = 0; + xx_storel_64(ls + stride * 3, _mm_srli_si128(res1, 8)); + *(int32_t *)(ls + width + stride * 3) = 0; + cf += 32; + ls += stride << 2; + i += 4; + } while (i < height); + } else if (width == 16) { + do { + const __m256i coeffA = yy_loadu_256(cf); + const __m256i coeffB = yy_loadu_256(cf + 8); + const __m256i coeffC = yy_loadu_256(cf + 16); + const __m256i coeffD = yy_loadu_256(cf + 24); + const __m256i coeffAB = _mm256_packs_epi32(coeffA, coeffB); + const __m256i coeffCD = _mm256_packs_epi32(coeffC, coeffD); + const __m256i absAB = _mm256_abs_epi16(coeffAB); + const __m256i absCD = _mm256_abs_epi16(coeffCD); + const __m256i absABCD = _mm256_packs_epi16(absAB, absCD); + const __m256i res_ = _mm256_permute4x64_epi64(absABCD, 0xd8); + const __m256i res = _mm256_shuffle_epi32(res_, 0xd8); + xx_storeu_128(ls, _mm256_castsi256_si128(res)); + xx_storeu_128(ls + stride, _mm256_extracti128_si256(res, 1)); + cf += 32; + *(int32_t *)(ls + width) = 0; + *(int32_t *)(ls + stride + width) = 0; + ls += stride << 1; + i += 2; + } while (i < height); + } else { + do { + const __m256i coeffA = yy_loadu_256(cf); + const __m256i coeffB = yy_loadu_256(cf + 8); + const __m256i coeffC = yy_loadu_256(cf + 16); + const __m256i coeffD = yy_loadu_256(cf + 24); + const __m256i coeffAB = _mm256_packs_epi32(coeffA, coeffB); + const __m256i coeffCD = _mm256_packs_epi32(coeffC, coeffD); + const __m256i absAB = _mm256_abs_epi16(coeffAB); + const __m256i absCD = _mm256_abs_epi16(coeffCD); + const __m256i absABCD = _mm256_packs_epi16(absAB, absCD); + const __m256i res_ = _mm256_permute4x64_epi64(absABCD, 0xd8); + const __m256i res = _mm256_shuffle_epi32(res_, 0xd8); + yy_storeu_256(ls, res); + cf += 32; + *(int32_t *)(ls + width) = 0; + ls += stride; + i += 1; + } while (i < height); + } +} diff --git a/third_party/aom/av1/encoder/x86/encodetxb_sse2.c b/third_party/aom/av1/encoder/x86/encodetxb_sse2.c new file mode 100644 index 0000000000..dedb4d02f6 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/encodetxb_sse2.c @@ -0,0 +1,505 @@ +/* + * Copyright (c) 2017, 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 <emmintrin.h> // SSE2 + +#include "aom/aom_integer.h" +#include "aom_dsp/x86/mem_sse2.h" +#include "av1/common/onyxc_int.h" +#include "av1/common/txb_common.h" + +static INLINE void load_levels_4x4x5_sse2(const uint8_t *const src, + const int stride, + const ptrdiff_t *const offsets, + __m128i *const level) { + level[0] = load_8bit_4x4_to_1_reg_sse2(src + 1, stride); + level[1] = load_8bit_4x4_to_1_reg_sse2(src + stride, stride); + level[2] = load_8bit_4x4_to_1_reg_sse2(src + offsets[0], stride); + level[3] = load_8bit_4x4_to_1_reg_sse2(src + offsets[1], stride); + level[4] = load_8bit_4x4_to_1_reg_sse2(src + offsets[2], stride); +} + +static INLINE void load_levels_8x2x5_sse2(const uint8_t *const src, + const int stride, + const ptrdiff_t *const offsets, + __m128i *const level) { + level[0] = load_8bit_8x2_to_1_reg_sse2(src + 1, stride); + level[1] = load_8bit_8x2_to_1_reg_sse2(src + stride, stride); + level[2] = load_8bit_8x2_to_1_reg_sse2(src + offsets[0], stride); + level[3] = load_8bit_8x2_to_1_reg_sse2(src + offsets[1], stride); + level[4] = load_8bit_8x2_to_1_reg_sse2(src + offsets[2], stride); +} + +static INLINE void load_levels_16x1x5_sse2(const uint8_t *const src, + const int stride, + const ptrdiff_t *const offsets, + __m128i *const level) { + level[0] = _mm_loadu_si128((__m128i *)(src + 1)); + level[1] = _mm_loadu_si128((__m128i *)(src + stride)); + level[2] = _mm_loadu_si128((__m128i *)(src + offsets[0])); + level[3] = _mm_loadu_si128((__m128i *)(src + offsets[1])); + level[4] = _mm_loadu_si128((__m128i *)(src + offsets[2])); +} + +static INLINE __m128i get_coeff_contexts_kernel_sse2(__m128i *const level) { + const __m128i const_3 = _mm_set1_epi8(3); + const __m128i const_4 = _mm_set1_epi8(4); + __m128i count; + + count = _mm_min_epu8(level[0], const_3); + level[1] = _mm_min_epu8(level[1], const_3); + level[2] = _mm_min_epu8(level[2], const_3); + level[3] = _mm_min_epu8(level[3], const_3); + level[4] = _mm_min_epu8(level[4], const_3); + count = _mm_add_epi8(count, level[1]); + count = _mm_add_epi8(count, level[2]); + count = _mm_add_epi8(count, level[3]); + count = _mm_add_epi8(count, level[4]); + count = _mm_avg_epu8(count, _mm_setzero_si128()); + count = _mm_min_epu8(count, const_4); + return count; +} + +static INLINE void get_4_nz_map_contexts_2d(const uint8_t *levels, + const int height, + const ptrdiff_t *const offsets, + int8_t *const coeff_contexts) { + const int stride = 4 + TX_PAD_HOR; + const __m128i pos_to_offset_large = _mm_set1_epi8(21); + __m128i pos_to_offset = + (height == 4) + ? _mm_setr_epi8(0, 1, 6, 6, 1, 6, 6, 21, 6, 6, 21, 21, 6, 21, 21, 21) + : _mm_setr_epi8(0, 11, 11, 11, 11, 11, 11, 11, 6, 6, 21, 21, 6, 21, + 21, 21); + __m128i count; + __m128i level[5]; + int8_t *cc = coeff_contexts; + int row = height; + + assert(!(height % 4)); + + do { + load_levels_4x4x5_sse2(levels, stride, offsets, level); + count = get_coeff_contexts_kernel_sse2(level); + count = _mm_add_epi8(count, pos_to_offset); + _mm_store_si128((__m128i *)cc, count); + pos_to_offset = pos_to_offset_large; + levels += 4 * stride; + cc += 16; + row -= 4; + } while (row); + + coeff_contexts[0] = 0; +} + +static INLINE void get_4_nz_map_contexts_hor(const uint8_t *levels, + const int height, + const ptrdiff_t *const offsets, + int8_t *coeff_contexts) { + const int stride = 4 + TX_PAD_HOR; + const __m128i pos_to_offset = + _mm_setr_epi8(SIG_COEF_CONTEXTS_2D + 0, SIG_COEF_CONTEXTS_2D + 5, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 0, SIG_COEF_CONTEXTS_2D + 5, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 0, SIG_COEF_CONTEXTS_2D + 5, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 0, SIG_COEF_CONTEXTS_2D + 5, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10); + __m128i count; + __m128i level[5]; + int row = height; + + assert(!(height % 4)); + + do { + load_levels_4x4x5_sse2(levels, stride, offsets, level); + count = get_coeff_contexts_kernel_sse2(level); + count = _mm_add_epi8(count, pos_to_offset); + _mm_store_si128((__m128i *)coeff_contexts, count); + levels += 4 * stride; + coeff_contexts += 16; + row -= 4; + } while (row); +} + +static INLINE void get_4_nz_map_contexts_ver(const uint8_t *levels, + const int height, + const ptrdiff_t *const offsets, + int8_t *coeff_contexts) { + const int stride = 4 + TX_PAD_HOR; + const __m128i pos_to_offset_large = _mm_set1_epi8(SIG_COEF_CONTEXTS_2D + 10); + __m128i pos_to_offset = + _mm_setr_epi8(SIG_COEF_CONTEXTS_2D + 0, SIG_COEF_CONTEXTS_2D + 0, + SIG_COEF_CONTEXTS_2D + 0, SIG_COEF_CONTEXTS_2D + 0, + SIG_COEF_CONTEXTS_2D + 5, SIG_COEF_CONTEXTS_2D + 5, + SIG_COEF_CONTEXTS_2D + 5, SIG_COEF_CONTEXTS_2D + 5, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10); + __m128i count; + __m128i level[5]; + int row = height; + + assert(!(height % 4)); + + do { + load_levels_4x4x5_sse2(levels, stride, offsets, level); + count = get_coeff_contexts_kernel_sse2(level); + count = _mm_add_epi8(count, pos_to_offset); + _mm_store_si128((__m128i *)coeff_contexts, count); + pos_to_offset = pos_to_offset_large; + levels += 4 * stride; + coeff_contexts += 16; + row -= 4; + } while (row); +} + +static INLINE void get_8_coeff_contexts_2d(const uint8_t *levels, + const int height, + const ptrdiff_t *const offsets, + int8_t *coeff_contexts) { + const int stride = 8 + TX_PAD_HOR; + int8_t *cc = coeff_contexts; + int row = height; + __m128i count; + __m128i level[5]; + __m128i pos_to_offset[3]; + + assert(!(height % 2)); + + if (height == 8) { + pos_to_offset[0] = + _mm_setr_epi8(0, 1, 6, 6, 21, 21, 21, 21, 1, 6, 6, 21, 21, 21, 21, 21); + pos_to_offset[1] = _mm_setr_epi8(6, 6, 21, 21, 21, 21, 21, 21, 6, 21, 21, + 21, 21, 21, 21, 21); + } else if (height < 8) { + pos_to_offset[0] = _mm_setr_epi8(0, 16, 6, 6, 21, 21, 21, 21, 16, 16, 6, 21, + 21, 21, 21, 21); + pos_to_offset[1] = _mm_setr_epi8(16, 16, 21, 21, 21, 21, 21, 21, 16, 16, 21, + 21, 21, 21, 21, 21); + } else { + pos_to_offset[0] = _mm_setr_epi8(0, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, + 11, 11, 11, 11, 11); + pos_to_offset[1] = _mm_setr_epi8(6, 6, 21, 21, 21, 21, 21, 21, 6, 21, 21, + 21, 21, 21, 21, 21); + } + pos_to_offset[2] = _mm_set1_epi8(21); + + do { + load_levels_8x2x5_sse2(levels, stride, offsets, level); + count = get_coeff_contexts_kernel_sse2(level); + count = _mm_add_epi8(count, pos_to_offset[0]); + _mm_store_si128((__m128i *)cc, count); + pos_to_offset[0] = pos_to_offset[1]; + pos_to_offset[1] = pos_to_offset[2]; + levels += 2 * stride; + cc += 16; + row -= 2; + } while (row); + + coeff_contexts[0] = 0; +} + +static INLINE void get_8_coeff_contexts_hor(const uint8_t *levels, + const int height, + const ptrdiff_t *const offsets, + int8_t *coeff_contexts) { + const int stride = 8 + TX_PAD_HOR; + const __m128i pos_to_offset = + _mm_setr_epi8(SIG_COEF_CONTEXTS_2D + 0, SIG_COEF_CONTEXTS_2D + 5, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 0, SIG_COEF_CONTEXTS_2D + 5, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10); + int row = height; + __m128i count; + __m128i level[5]; + + assert(!(height % 2)); + + do { + load_levels_8x2x5_sse2(levels, stride, offsets, level); + count = get_coeff_contexts_kernel_sse2(level); + count = _mm_add_epi8(count, pos_to_offset); + _mm_store_si128((__m128i *)coeff_contexts, count); + levels += 2 * stride; + coeff_contexts += 16; + row -= 2; + } while (row); +} + +static INLINE void get_8_coeff_contexts_ver(const uint8_t *levels, + const int height, + const ptrdiff_t *const offsets, + int8_t *coeff_contexts) { + const int stride = 8 + TX_PAD_HOR; + const __m128i pos_to_offset_large = _mm_set1_epi8(SIG_COEF_CONTEXTS_2D + 10); + __m128i pos_to_offset = + _mm_setr_epi8(SIG_COEF_CONTEXTS_2D + 0, SIG_COEF_CONTEXTS_2D + 0, + SIG_COEF_CONTEXTS_2D + 0, SIG_COEF_CONTEXTS_2D + 0, + SIG_COEF_CONTEXTS_2D + 0, SIG_COEF_CONTEXTS_2D + 0, + SIG_COEF_CONTEXTS_2D + 0, SIG_COEF_CONTEXTS_2D + 0, + SIG_COEF_CONTEXTS_2D + 5, SIG_COEF_CONTEXTS_2D + 5, + SIG_COEF_CONTEXTS_2D + 5, SIG_COEF_CONTEXTS_2D + 5, + SIG_COEF_CONTEXTS_2D + 5, SIG_COEF_CONTEXTS_2D + 5, + SIG_COEF_CONTEXTS_2D + 5, SIG_COEF_CONTEXTS_2D + 5); + int row = height; + __m128i count; + __m128i level[5]; + + assert(!(height % 2)); + + do { + load_levels_8x2x5_sse2(levels, stride, offsets, level); + count = get_coeff_contexts_kernel_sse2(level); + count = _mm_add_epi8(count, pos_to_offset); + _mm_store_si128((__m128i *)coeff_contexts, count); + pos_to_offset = pos_to_offset_large; + levels += 2 * stride; + coeff_contexts += 16; + row -= 2; + } while (row); +} + +static INLINE void get_16n_coeff_contexts_2d(const uint8_t *levels, + const int real_width, + const int real_height, + const int width, const int height, + const ptrdiff_t *const offsets, + int8_t *coeff_contexts) { + const int stride = width + TX_PAD_HOR; + int8_t *cc = coeff_contexts; + int row = height; + __m128i pos_to_offset[5]; + __m128i pos_to_offset_large[3]; + __m128i count; + __m128i level[5]; + + assert(!(width % 16)); + + pos_to_offset_large[2] = _mm_set1_epi8(21); + if (real_width == real_height) { + pos_to_offset[0] = _mm_setr_epi8(0, 1, 6, 6, 21, 21, 21, 21, 21, 21, 21, 21, + 21, 21, 21, 21); + pos_to_offset[1] = _mm_setr_epi8(1, 6, 6, 21, 21, 21, 21, 21, 21, 21, 21, + 21, 21, 21, 21, 21); + pos_to_offset[2] = _mm_setr_epi8(6, 6, 21, 21, 21, 21, 21, 21, 21, 21, 21, + 21, 21, 21, 21, 21); + pos_to_offset[3] = _mm_setr_epi8(6, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, + 21, 21, 21, 21, 21); + pos_to_offset[4] = pos_to_offset_large[0] = pos_to_offset_large[1] = + pos_to_offset_large[2]; + } else if (real_width > real_height) { + pos_to_offset[0] = _mm_setr_epi8(0, 16, 6, 6, 21, 21, 21, 21, 21, 21, 21, + 21, 21, 21, 21, 21); + pos_to_offset[1] = _mm_setr_epi8(16, 16, 6, 21, 21, 21, 21, 21, 21, 21, 21, + 21, 21, 21, 21, 21); + pos_to_offset[2] = pos_to_offset[3] = pos_to_offset[4] = _mm_setr_epi8( + 16, 16, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21); + pos_to_offset_large[0] = pos_to_offset_large[1] = pos_to_offset_large[2]; + } else { // real_width < real_height + pos_to_offset[0] = pos_to_offset[1] = _mm_setr_epi8( + 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11); + pos_to_offset[2] = _mm_setr_epi8(6, 6, 21, 21, 21, 21, 21, 21, 21, 21, 21, + 21, 21, 21, 21, 21); + pos_to_offset[3] = _mm_setr_epi8(6, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, + 21, 21, 21, 21, 21); + pos_to_offset[4] = pos_to_offset_large[2]; + pos_to_offset_large[0] = pos_to_offset_large[1] = _mm_set1_epi8(11); + } + + do { + int w = width; + + do { + load_levels_16x1x5_sse2(levels, stride, offsets, level); + count = get_coeff_contexts_kernel_sse2(level); + count = _mm_add_epi8(count, pos_to_offset[0]); + _mm_store_si128((__m128i *)cc, count); + levels += 16; + cc += 16; + w -= 16; + pos_to_offset[0] = pos_to_offset_large[0]; + } while (w); + + pos_to_offset[0] = pos_to_offset[1]; + pos_to_offset[1] = pos_to_offset[2]; + pos_to_offset[2] = pos_to_offset[3]; + pos_to_offset[3] = pos_to_offset[4]; + pos_to_offset_large[0] = pos_to_offset_large[1]; + pos_to_offset_large[1] = pos_to_offset_large[2]; + levels += TX_PAD_HOR; + } while (--row); + + coeff_contexts[0] = 0; +} + +static INLINE void get_16n_coeff_contexts_hor(const uint8_t *levels, + const int width, const int height, + const ptrdiff_t *const offsets, + int8_t *coeff_contexts) { + const int stride = width + TX_PAD_HOR; + const __m128i pos_to_offset_large = + _mm_setr_epi8(SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10); + __m128i count; + __m128i level[5]; + int row = height; + + assert(!(width % 16)); + + do { + __m128i pos_to_offset = + _mm_setr_epi8(SIG_COEF_CONTEXTS_2D + 0, SIG_COEF_CONTEXTS_2D + 5, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10, + SIG_COEF_CONTEXTS_2D + 10, SIG_COEF_CONTEXTS_2D + 10); + int w = width; + + do { + load_levels_16x1x5_sse2(levels, stride, offsets, level); + count = get_coeff_contexts_kernel_sse2(level); + count = _mm_add_epi8(count, pos_to_offset); + _mm_store_si128((__m128i *)coeff_contexts, count); + pos_to_offset = pos_to_offset_large; + levels += 16; + coeff_contexts += 16; + w -= 16; + } while (w); + + levels += TX_PAD_HOR; + } while (--row); +} + +static INLINE void get_16n_coeff_contexts_ver(const uint8_t *levels, + const int width, const int height, + const ptrdiff_t *const offsets, + int8_t *coeff_contexts) { + const int stride = width + TX_PAD_HOR; + __m128i pos_to_offset[3]; + __m128i count; + __m128i level[5]; + int row = height; + + assert(!(width % 16)); + + pos_to_offset[0] = _mm_set1_epi8(SIG_COEF_CONTEXTS_2D + 0); + pos_to_offset[1] = _mm_set1_epi8(SIG_COEF_CONTEXTS_2D + 5); + pos_to_offset[2] = _mm_set1_epi8(SIG_COEF_CONTEXTS_2D + 10); + + do { + int w = width; + + do { + load_levels_16x1x5_sse2(levels, stride, offsets, level); + count = get_coeff_contexts_kernel_sse2(level); + count = _mm_add_epi8(count, pos_to_offset[0]); + _mm_store_si128((__m128i *)coeff_contexts, count); + levels += 16; + coeff_contexts += 16; + w -= 16; + } while (w); + + pos_to_offset[0] = pos_to_offset[1]; + pos_to_offset[1] = pos_to_offset[2]; + levels += TX_PAD_HOR; + } while (--row); +} + +// Note: levels[] must be in the range [0, 127], inclusive. +void av1_get_nz_map_contexts_sse2(const uint8_t *const levels, + const int16_t *const scan, const uint16_t eob, + const TX_SIZE tx_size, + const TX_CLASS tx_class, + int8_t *const coeff_contexts) { + const int last_idx = eob - 1; + if (!last_idx) { + coeff_contexts[0] = 0; + return; + } + + const int real_width = tx_size_wide[tx_size]; + const int real_height = tx_size_high[tx_size]; + const int width = get_txb_wide(tx_size); + const int height = get_txb_high(tx_size); + const int stride = width + TX_PAD_HOR; + ptrdiff_t offsets[3]; + + /* coeff_contexts must be 16 byte aligned. */ + assert(!((intptr_t)coeff_contexts & 0xf)); + + if (tx_class == TX_CLASS_2D) { + offsets[0] = 0 * stride + 2; + offsets[1] = 1 * stride + 1; + offsets[2] = 2 * stride + 0; + + if (width == 4) { + get_4_nz_map_contexts_2d(levels, height, offsets, coeff_contexts); + } else if (width == 8) { + get_8_coeff_contexts_2d(levels, height, offsets, coeff_contexts); + } else if (width == 16) { + get_16n_coeff_contexts_2d(levels, real_width, real_height, width, height, + offsets, coeff_contexts); + } else { + get_16n_coeff_contexts_2d(levels, real_width, real_height, width, height, + offsets, coeff_contexts); + } + } else if (tx_class == TX_CLASS_HORIZ) { + offsets[0] = 2; + offsets[1] = 3; + offsets[2] = 4; + if (width == 4) { + get_4_nz_map_contexts_hor(levels, height, offsets, coeff_contexts); + } else if (width == 8) { + get_8_coeff_contexts_hor(levels, height, offsets, coeff_contexts); + } else { + get_16n_coeff_contexts_hor(levels, width, height, offsets, + coeff_contexts); + } + } else { // TX_CLASS_VERT + offsets[0] = 2 * stride; + offsets[1] = 3 * stride; + offsets[2] = 4 * stride; + if (width == 4) { + get_4_nz_map_contexts_ver(levels, height, offsets, coeff_contexts); + } else if (width == 8) { + get_8_coeff_contexts_ver(levels, height, offsets, coeff_contexts); + } else { + get_16n_coeff_contexts_ver(levels, width, height, offsets, + coeff_contexts); + } + } + + const int bwl = get_txb_bwl(tx_size); + const int pos = scan[last_idx]; + if (last_idx <= (height << bwl) / 8) + coeff_contexts[pos] = 1; + else if (last_idx <= (height << bwl) / 4) + coeff_contexts[pos] = 2; + else + coeff_contexts[pos] = 3; +} diff --git a/third_party/aom/av1/encoder/x86/encodetxb_sse4.c b/third_party/aom/av1/encoder/x86/encodetxb_sse4.c new file mode 100644 index 0000000000..5e0687cd38 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/encodetxb_sse4.c @@ -0,0 +1,92 @@ +/* + * Copyright (c) 2017, 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 <emmintrin.h> // SSE2 +#include <smmintrin.h> /* SSE4.1 */ + +#include "aom/aom_integer.h" +#include "av1/common/onyxc_int.h" +#include "av1/common/txb_common.h" +#include "aom_dsp/x86/synonyms.h" + +void av1_txb_init_levels_sse4_1(const tran_low_t *const coeff, const int width, + const int height, uint8_t *const levels) { + const int stride = width + TX_PAD_HOR; + const __m128i zeros = _mm_setzero_si128(); + + const int32_t pre_len = sizeof(*levels) * TX_PAD_TOP * stride; + uint8_t *pre_buf = levels - TX_PAD_TOP * stride; + uint8_t *pre_buf_end = pre_buf + pre_len; + do { + _mm_storeu_si128((__m128i *)(pre_buf), zeros); + pre_buf += 16; + } while (pre_buf < pre_buf_end); + + const int32_t bottom_len = sizeof(*levels) * (TX_PAD_BOTTOM * stride); + uint8_t *bottom_buf = levels + stride * height; + uint8_t *bottom_buf_end = bottom_buf + bottom_len; + do { + _mm_storeu_si128((__m128i *)(bottom_buf), zeros); + bottom_buf += 16; + } while (bottom_buf < bottom_buf_end); + + int i = 0; + uint8_t *ls = levels; + const tran_low_t *cf = coeff; + if (width == 4) { + do { + const __m128i coeffA = xx_loadu_128(cf); + const __m128i coeffB = xx_loadu_128(cf + 4); + const __m128i coeffAB = _mm_packs_epi32(coeffA, coeffB); + const __m128i absAB = _mm_abs_epi16(coeffAB); + const __m128i absAB8 = _mm_packs_epi16(absAB, zeros); + const __m128i lsAB = _mm_unpacklo_epi32(absAB8, zeros); + xx_storeu_128(ls, lsAB); + ls += (stride << 1); + cf += (width << 1); + i += 2; + } while (i < height); + } else if (width == 8) { + do { + const __m128i coeffA = xx_loadu_128(cf); + const __m128i coeffB = xx_loadu_128(cf + 4); + const __m128i coeffAB = _mm_packs_epi32(coeffA, coeffB); + const __m128i absAB = _mm_abs_epi16(coeffAB); + const __m128i absAB8 = _mm_packs_epi16(absAB, zeros); + xx_storeu_128(ls, absAB8); + ls += stride; + cf += width; + i += 1; + } while (i < height); + } else { + do { + int j = 0; + do { + const __m128i coeffA = xx_loadu_128(cf); + const __m128i coeffB = xx_loadu_128(cf + 4); + const __m128i coeffC = xx_loadu_128(cf + 8); + const __m128i coeffD = xx_loadu_128(cf + 12); + const __m128i coeffAB = _mm_packs_epi32(coeffA, coeffB); + const __m128i coeffCD = _mm_packs_epi32(coeffC, coeffD); + const __m128i absAB = _mm_abs_epi16(coeffAB); + const __m128i absCD = _mm_abs_epi16(coeffCD); + const __m128i absABCD = _mm_packs_epi16(absAB, absCD); + xx_storeu_128(ls + j, absABCD); + j += 16; + cf += 16; + } while (j < width); + *(int32_t *)(ls + width) = 0; + ls += stride; + i += 1; + } while (i < height); + } +} diff --git a/third_party/aom/av1/encoder/x86/error_intrin_avx2.c b/third_party/aom/av1/encoder/x86/error_intrin_avx2.c new file mode 100644 index 0000000000..7d4f695854 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/error_intrin_avx2.c @@ -0,0 +1,88 @@ +/* + * Copyright (c) 2016, 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 <immintrin.h> // AVX2 + +#include "config/av1_rtcd.h" + +#include "aom/aom_integer.h" + +static INLINE void read_coeff(const tran_low_t *coeff, intptr_t offset, + __m256i *c) { + const tran_low_t *addr = coeff + offset; + + if (sizeof(tran_low_t) == 4) { + const __m256i x0 = _mm256_loadu_si256((const __m256i *)addr); + const __m256i x1 = _mm256_loadu_si256((const __m256i *)addr + 1); + const __m256i y = _mm256_packs_epi32(x0, x1); + *c = _mm256_permute4x64_epi64(y, 0xD8); + } else { + *c = _mm256_loadu_si256((const __m256i *)addr); + } +} + +int64_t av1_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, + intptr_t block_size, int64_t *ssz) { + __m256i sse_reg, ssz_reg, coeff_reg, dqcoeff_reg; + __m256i exp_dqcoeff_lo, exp_dqcoeff_hi, exp_coeff_lo, exp_coeff_hi; + __m256i sse_reg_64hi, ssz_reg_64hi; + __m128i sse_reg128, ssz_reg128; + int64_t sse; + int i; + const __m256i zero_reg = _mm256_setzero_si256(); + + // init sse and ssz registerd to zero + sse_reg = _mm256_setzero_si256(); + ssz_reg = _mm256_setzero_si256(); + + for (i = 0; i < block_size; i += 16) { + // load 32 bytes from coeff and dqcoeff + read_coeff(coeff, i, &coeff_reg); + read_coeff(dqcoeff, i, &dqcoeff_reg); + // dqcoeff - coeff + dqcoeff_reg = _mm256_sub_epi16(dqcoeff_reg, coeff_reg); + // madd (dqcoeff - coeff) + dqcoeff_reg = _mm256_madd_epi16(dqcoeff_reg, dqcoeff_reg); + // madd coeff + coeff_reg = _mm256_madd_epi16(coeff_reg, coeff_reg); + // expand each double word of madd (dqcoeff - coeff) to quad word + exp_dqcoeff_lo = _mm256_unpacklo_epi32(dqcoeff_reg, zero_reg); + exp_dqcoeff_hi = _mm256_unpackhi_epi32(dqcoeff_reg, zero_reg); + // expand each double word of madd (coeff) to quad word + exp_coeff_lo = _mm256_unpacklo_epi32(coeff_reg, zero_reg); + exp_coeff_hi = _mm256_unpackhi_epi32(coeff_reg, zero_reg); + // add each quad word of madd (dqcoeff - coeff) and madd (coeff) + sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_lo); + ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_lo); + sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_hi); + ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_hi); + } + // save the higher 64 bit of each 128 bit lane + sse_reg_64hi = _mm256_srli_si256(sse_reg, 8); + ssz_reg_64hi = _mm256_srli_si256(ssz_reg, 8); + // add the higher 64 bit to the low 64 bit + sse_reg = _mm256_add_epi64(sse_reg, sse_reg_64hi); + ssz_reg = _mm256_add_epi64(ssz_reg, ssz_reg_64hi); + + // add each 64 bit from each of the 128 bit lane of the 256 bit + sse_reg128 = _mm_add_epi64(_mm256_castsi256_si128(sse_reg), + _mm256_extractf128_si256(sse_reg, 1)); + + ssz_reg128 = _mm_add_epi64(_mm256_castsi256_si128(ssz_reg), + _mm256_extractf128_si256(ssz_reg, 1)); + + // store the results + _mm_storel_epi64((__m128i *)(&sse), sse_reg128); + + _mm_storel_epi64((__m128i *)(ssz), ssz_reg128); + _mm256_zeroupper(); + return sse; +} diff --git a/third_party/aom/av1/encoder/x86/error_sse2.asm b/third_party/aom/av1/encoder/x86/error_sse2.asm new file mode 100644 index 0000000000..72e9e22b18 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/error_sse2.asm @@ -0,0 +1,79 @@ +; +; Copyright (c) 2016, 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. +; + +; + +%define private_prefix av1 + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +; int64_t av1_block_error(int16_t *coeff, int16_t *dqcoeff, intptr_t block_size, +; int64_t *ssz) + +INIT_XMM sse2 +cglobal block_error, 3, 3, 8, uqc, dqc, size, ssz + pxor m4, m4 ; sse accumulator + pxor m6, m6 ; ssz accumulator + pxor m5, m5 ; dedicated zero register + lea uqcq, [uqcq+sizeq*2] + lea dqcq, [dqcq+sizeq*2] + neg sizeq +.loop: + mova m2, [uqcq+sizeq*2] + mova m0, [dqcq+sizeq*2] + mova m3, [uqcq+sizeq*2+mmsize] + mova m1, [dqcq+sizeq*2+mmsize] + psubw m0, m2 + psubw m1, m3 + ; individual errors are max. 15bit+sign, so squares are 30bit, and + ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit) + pmaddwd m0, m0 + pmaddwd m1, m1 + pmaddwd m2, m2 + pmaddwd m3, m3 + ; accumulate in 64bit + punpckldq m7, m0, m5 + punpckhdq m0, m5 + paddq m4, m7 + punpckldq m7, m1, m5 + paddq m4, m0 + punpckhdq m1, m5 + paddq m4, m7 + punpckldq m7, m2, m5 + paddq m4, m1 + punpckhdq m2, m5 + paddq m6, m7 + punpckldq m7, m3, m5 + paddq m6, m2 + punpckhdq m3, m5 + paddq m6, m7 + paddq m6, m3 + add sizeq, mmsize + jl .loop + + ; accumulate horizontally and store in return value + movhlps m5, m4 + movhlps m7, m6 + paddq m4, m5 + paddq m6, m7 +%if ARCH_X86_64 + movq rax, m4 + movq [sszq], m6 +%else + mov eax, sszm + pshufd m5, m4, 0x1 + movq [eax], m6 + movd eax, m4 + movd edx, m5 +%endif + RET diff --git a/third_party/aom/av1/encoder/x86/hash_sse42.c b/third_party/aom/av1/encoder/x86/hash_sse42.c new file mode 100644 index 0000000000..65fa463117 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/hash_sse42.c @@ -0,0 +1,51 @@ +/* + * Copyright (c) 2018, 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 <stdint.h> +#include <smmintrin.h> + +// Byte-boundary alignment issues +#define ALIGN_SIZE 8 +#define ALIGN_MASK (ALIGN_SIZE - 1) + +#define CALC_CRC(op, crc, type, buf, len) \ + while ((len) >= sizeof(type)) { \ + (crc) = op((crc), *(type *)(buf)); \ + (len) -= sizeof(type); \ + buf += sizeof(type); \ + } + +/** + * Calculates 32-bit CRC for the input buffer + * polynomial is 0x11EDC6F41 + * @return A 32-bit unsigned integer representing the CRC + */ +uint32_t av1_get_crc32c_value_sse4_2(void *crc_calculator, uint8_t *p, + size_t len) { + (void)crc_calculator; + const uint8_t *buf = p; + uint32_t crc = 0xFFFFFFFF; + + // Align the input to the word boundary + for (; (len > 0) && ((intptr_t)buf & ALIGN_MASK); len--, buf++) { + crc = _mm_crc32_u8(crc, *buf); + } + +#ifdef __x86_64__ + uint64_t crc64 = crc; + CALC_CRC(_mm_crc32_u64, crc64, uint64_t, buf, len); + crc = (uint32_t)crc64; +#endif + CALC_CRC(_mm_crc32_u32, crc, uint32_t, buf, len); + CALC_CRC(_mm_crc32_u16, crc, uint16_t, buf, len); + CALC_CRC(_mm_crc32_u8, crc, uint8_t, buf, len); + return (crc ^= 0xFFFFFFFF); +} diff --git a/third_party/aom/av1/encoder/x86/highbd_block_error_intrin_sse2.c b/third_party/aom/av1/encoder/x86/highbd_block_error_intrin_sse2.c new file mode 100644 index 0000000000..777304ace7 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/highbd_block_error_intrin_sse2.c @@ -0,0 +1,72 @@ +/* + * Copyright (c) 2016, 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 <emmintrin.h> +#include <stdio.h> + +#include "av1/common/common.h" + +int64_t av1_highbd_block_error_sse2(tran_low_t *coeff, tran_low_t *dqcoeff, + intptr_t block_size, int64_t *ssz, + int bps) { + int i, j, test; + uint32_t temp[4]; + __m128i max, min, cmp0, cmp1, cmp2, cmp3; + int64_t error = 0, sqcoeff = 0; + const int shift = 2 * (bps - 8); + const int rounding = shift > 0 ? 1 << (shift - 1) : 0; + + for (i = 0; i < block_size; i += 8) { + // Load the data into xmm registers + __m128i mm_coeff = _mm_load_si128((__m128i *)(coeff + i)); + __m128i mm_coeff2 = _mm_load_si128((__m128i *)(coeff + i + 4)); + __m128i mm_dqcoeff = _mm_load_si128((__m128i *)(dqcoeff + i)); + __m128i mm_dqcoeff2 = _mm_load_si128((__m128i *)(dqcoeff + i + 4)); + // Check if any values require more than 15 bit + max = _mm_set1_epi32(0x3fff); + min = _mm_set1_epi32(0xffffc000); + cmp0 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff, max), + _mm_cmplt_epi32(mm_coeff, min)); + cmp1 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff2, max), + _mm_cmplt_epi32(mm_coeff2, min)); + cmp2 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff, max), + _mm_cmplt_epi32(mm_dqcoeff, min)); + cmp3 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff2, max), + _mm_cmplt_epi32(mm_dqcoeff2, min)); + test = _mm_movemask_epi8( + _mm_or_si128(_mm_or_si128(cmp0, cmp1), _mm_or_si128(cmp2, cmp3))); + + if (!test) { + __m128i mm_diff, error_sse2, sqcoeff_sse2; + mm_coeff = _mm_packs_epi32(mm_coeff, mm_coeff2); + mm_dqcoeff = _mm_packs_epi32(mm_dqcoeff, mm_dqcoeff2); + mm_diff = _mm_sub_epi16(mm_coeff, mm_dqcoeff); + error_sse2 = _mm_madd_epi16(mm_diff, mm_diff); + sqcoeff_sse2 = _mm_madd_epi16(mm_coeff, mm_coeff); + _mm_storeu_si128((__m128i *)temp, error_sse2); + error = error + temp[0] + temp[1] + temp[2] + temp[3]; + _mm_storeu_si128((__m128i *)temp, sqcoeff_sse2); + sqcoeff += temp[0] + temp[1] + temp[2] + temp[3]; + } else { + for (j = 0; j < 8; j++) { + const int64_t diff = coeff[i + j] - dqcoeff[i + j]; + error += diff * diff; + sqcoeff += (int64_t)coeff[i + j] * (int64_t)coeff[i + j]; + } + } + } + assert(error >= 0 && sqcoeff >= 0); + error = (error + rounding) >> shift; + sqcoeff = (sqcoeff + rounding) >> shift; + + *ssz = sqcoeff; + return error; +} diff --git a/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c b/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c new file mode 100644 index 0000000000..535485ae8e --- /dev/null +++ b/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c @@ -0,0 +1,1783 @@ +/* + * Copyright (c) 2016, 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 <smmintrin.h> /* SSE4.1 */ + +#include "config/aom_config.h" +#include "config/av1_rtcd.h" + +#include "av1/common/av1_txfm.h" +#include "av1/common/x86/highbd_txfm_utility_sse4.h" +#include "av1/encoder/av1_fwd_txfm1d_cfg.h" +#include "av1/encoder/x86/av1_txfm1d_sse4.h" +#include "aom_dsp/txfm_common.h" +#include "aom_dsp/x86/txfm_common_sse2.h" +#include "aom_ports/mem.h" + +static INLINE void load_buffer_4x4(const int16_t *input, __m128i *in, + int stride, int flipud, int fliplr, + int shift) { + if (!flipud) { + in[0] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride)); + in[1] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride)); + in[2] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride)); + in[3] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride)); + } else { + in[0] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride)); + in[1] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride)); + in[2] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride)); + in[3] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride)); + } + + if (fliplr) { + in[0] = _mm_shufflelo_epi16(in[0], 0x1b); + in[1] = _mm_shufflelo_epi16(in[1], 0x1b); + in[2] = _mm_shufflelo_epi16(in[2], 0x1b); + in[3] = _mm_shufflelo_epi16(in[3], 0x1b); + } + + in[0] = _mm_cvtepi16_epi32(in[0]); + in[1] = _mm_cvtepi16_epi32(in[1]); + in[2] = _mm_cvtepi16_epi32(in[2]); + in[3] = _mm_cvtepi16_epi32(in[3]); + + in[0] = _mm_slli_epi32(in[0], shift); + in[1] = _mm_slli_epi32(in[1], shift); + in[2] = _mm_slli_epi32(in[2], shift); + in[3] = _mm_slli_epi32(in[3], shift); +} + +// We only use stage-2 bit; +// shift[0] is used in load_buffer_4x4() +// shift[1] is used in txfm_func_col() +// shift[2] is used in txfm_func_row() +static void fdct4x4_sse4_1(__m128i *in, int bit) { + const int32_t *cospi = cospi_arr(bit); + const __m128i cospi32 = _mm_set1_epi32(cospi[32]); + const __m128i cospi48 = _mm_set1_epi32(cospi[48]); + const __m128i cospi16 = _mm_set1_epi32(cospi[16]); + const __m128i rnding = _mm_set1_epi32(1 << (bit - 1)); + __m128i s0, s1, s2, s3; + __m128i u0, u1, u2, u3; + __m128i v0, v1, v2, v3; + + s0 = _mm_add_epi32(in[0], in[3]); + s1 = _mm_add_epi32(in[1], in[2]); + s2 = _mm_sub_epi32(in[1], in[2]); + s3 = _mm_sub_epi32(in[0], in[3]); + + // btf_32_sse4_1_type0(cospi32, cospi32, s[01], u[02], bit); + u0 = _mm_mullo_epi32(s0, cospi32); + u1 = _mm_mullo_epi32(s1, cospi32); + u2 = _mm_add_epi32(u0, u1); + v0 = _mm_sub_epi32(u0, u1); + + u3 = _mm_add_epi32(u2, rnding); + v1 = _mm_add_epi32(v0, rnding); + + u0 = _mm_srai_epi32(u3, bit); + u2 = _mm_srai_epi32(v1, bit); + + // btf_32_sse4_1_type1(cospi48, cospi16, s[23], u[13], bit); + v0 = _mm_mullo_epi32(s2, cospi48); + v1 = _mm_mullo_epi32(s3, cospi16); + v2 = _mm_add_epi32(v0, v1); + + v3 = _mm_add_epi32(v2, rnding); + u1 = _mm_srai_epi32(v3, bit); + + v0 = _mm_mullo_epi32(s2, cospi16); + v1 = _mm_mullo_epi32(s3, cospi48); + v2 = _mm_sub_epi32(v1, v0); + + v3 = _mm_add_epi32(v2, rnding); + u3 = _mm_srai_epi32(v3, bit); + + // Note: shift[1] and shift[2] are zeros + + // Transpose 4x4 32-bit + v0 = _mm_unpacklo_epi32(u0, u1); + v1 = _mm_unpackhi_epi32(u0, u1); + v2 = _mm_unpacklo_epi32(u2, u3); + v3 = _mm_unpackhi_epi32(u2, u3); + + in[0] = _mm_unpacklo_epi64(v0, v2); + in[1] = _mm_unpackhi_epi64(v0, v2); + in[2] = _mm_unpacklo_epi64(v1, v3); + in[3] = _mm_unpackhi_epi64(v1, v3); +} + +static INLINE void write_buffer_4x4(__m128i *res, int32_t *output) { + _mm_store_si128((__m128i *)(output + 0 * 4), res[0]); + _mm_store_si128((__m128i *)(output + 1 * 4), res[1]); + _mm_store_si128((__m128i *)(output + 2 * 4), res[2]); + _mm_store_si128((__m128i *)(output + 3 * 4), res[3]); +} + +static void fadst4x4_sse4_1(__m128i *in, int bit) { + const int32_t *sinpi = sinpi_arr(bit); + const __m128i rnding = _mm_set1_epi32(1 << (bit - 1)); + const __m128i sinpi1 = _mm_set1_epi32((int)sinpi[1]); + const __m128i sinpi2 = _mm_set1_epi32((int)sinpi[2]); + const __m128i sinpi3 = _mm_set1_epi32((int)sinpi[3]); + const __m128i sinpi4 = _mm_set1_epi32((int)sinpi[4]); + __m128i t; + __m128i s0, s1, s2, s3, s4, s5, s6, s7; + __m128i x0, x1, x2, x3; + __m128i u0, u1, u2, u3; + __m128i v0, v1, v2, v3; + + s0 = _mm_mullo_epi32(in[0], sinpi1); + s1 = _mm_mullo_epi32(in[0], sinpi4); + s2 = _mm_mullo_epi32(in[1], sinpi2); + s3 = _mm_mullo_epi32(in[1], sinpi1); + s4 = _mm_mullo_epi32(in[2], sinpi3); + s5 = _mm_mullo_epi32(in[3], sinpi4); + s6 = _mm_mullo_epi32(in[3], sinpi2); + t = _mm_add_epi32(in[0], in[1]); + s7 = _mm_sub_epi32(t, in[3]); + + t = _mm_add_epi32(s0, s2); + x0 = _mm_add_epi32(t, s5); + x1 = _mm_mullo_epi32(s7, sinpi3); + t = _mm_sub_epi32(s1, s3); + x2 = _mm_add_epi32(t, s6); + x3 = s4; + + s0 = _mm_add_epi32(x0, x3); + s1 = x1; + s2 = _mm_sub_epi32(x2, x3); + t = _mm_sub_epi32(x2, x0); + s3 = _mm_add_epi32(t, x3); + + u0 = _mm_add_epi32(s0, rnding); + u0 = _mm_srai_epi32(u0, bit); + + u1 = _mm_add_epi32(s1, rnding); + u1 = _mm_srai_epi32(u1, bit); + + u2 = _mm_add_epi32(s2, rnding); + u2 = _mm_srai_epi32(u2, bit); + + u3 = _mm_add_epi32(s3, rnding); + u3 = _mm_srai_epi32(u3, bit); + + v0 = _mm_unpacklo_epi32(u0, u1); + v1 = _mm_unpackhi_epi32(u0, u1); + v2 = _mm_unpacklo_epi32(u2, u3); + v3 = _mm_unpackhi_epi32(u2, u3); + + in[0] = _mm_unpacklo_epi64(v0, v2); + in[1] = _mm_unpackhi_epi64(v0, v2); + in[2] = _mm_unpacklo_epi64(v1, v3); + in[3] = _mm_unpackhi_epi64(v1, v3); +} + +void av1_fwd_txfm2d_4x4_sse4_1(const int16_t *input, int32_t *coeff, + int input_stride, TX_TYPE tx_type, int bd) { + __m128i in[4]; + const int8_t *shift = fwd_txfm_shift_ls[TX_4X4]; + const int txw_idx = get_txw_idx(TX_4X4); + const int txh_idx = get_txh_idx(TX_4X4); + + switch (tx_type) { + case DCT_DCT: + load_buffer_4x4(input, in, input_stride, 0, 0, shift[0]); + fdct4x4_sse4_1(in, fwd_cos_bit_col[txw_idx][txh_idx]); + fdct4x4_sse4_1(in, fwd_cos_bit_row[txw_idx][txh_idx]); + write_buffer_4x4(in, coeff); + break; + case ADST_DCT: + load_buffer_4x4(input, in, input_stride, 0, 0, shift[0]); + fadst4x4_sse4_1(in, fwd_cos_bit_col[txw_idx][txh_idx]); + fdct4x4_sse4_1(in, fwd_cos_bit_row[txw_idx][txh_idx]); + write_buffer_4x4(in, coeff); + break; + case DCT_ADST: + load_buffer_4x4(input, in, input_stride, 0, 0, shift[0]); + fdct4x4_sse4_1(in, fwd_cos_bit_col[txw_idx][txh_idx]); + fadst4x4_sse4_1(in, fwd_cos_bit_row[txw_idx][txh_idx]); + write_buffer_4x4(in, coeff); + break; + case ADST_ADST: + load_buffer_4x4(input, in, input_stride, 0, 0, shift[0]); + fadst4x4_sse4_1(in, fwd_cos_bit_col[txw_idx][txh_idx]); + fadst4x4_sse4_1(in, fwd_cos_bit_row[txw_idx][txh_idx]); + write_buffer_4x4(in, coeff); + break; + case FLIPADST_DCT: + load_buffer_4x4(input, in, input_stride, 1, 0, shift[0]); + fadst4x4_sse4_1(in, fwd_cos_bit_col[txw_idx][txh_idx]); + fdct4x4_sse4_1(in, fwd_cos_bit_row[txw_idx][txh_idx]); + write_buffer_4x4(in, coeff); + break; + case DCT_FLIPADST: + load_buffer_4x4(input, in, input_stride, 0, 1, shift[0]); + fdct4x4_sse4_1(in, fwd_cos_bit_col[txw_idx][txh_idx]); + fadst4x4_sse4_1(in, fwd_cos_bit_row[txw_idx][txh_idx]); + write_buffer_4x4(in, coeff); + break; + case FLIPADST_FLIPADST: + load_buffer_4x4(input, in, input_stride, 1, 1, shift[0]); + fadst4x4_sse4_1(in, fwd_cos_bit_col[txw_idx][txh_idx]); + fadst4x4_sse4_1(in, fwd_cos_bit_row[txw_idx][txh_idx]); + write_buffer_4x4(in, coeff); + break; + case ADST_FLIPADST: + load_buffer_4x4(input, in, input_stride, 0, 1, shift[0]); + fadst4x4_sse4_1(in, fwd_cos_bit_col[txw_idx][txh_idx]); + fadst4x4_sse4_1(in, fwd_cos_bit_row[txw_idx][txh_idx]); + write_buffer_4x4(in, coeff); + break; + case FLIPADST_ADST: + load_buffer_4x4(input, in, input_stride, 1, 0, shift[0]); + fadst4x4_sse4_1(in, fwd_cos_bit_col[txw_idx][txh_idx]); + fadst4x4_sse4_1(in, fwd_cos_bit_row[txw_idx][txh_idx]); + write_buffer_4x4(in, coeff); + break; + default: assert(0); + } + (void)bd; +} + +static INLINE void load_buffer_8x8(const int16_t *input, __m128i *in, + int stride, int flipud, int fliplr, + int shift) { + __m128i u; + if (!flipud) { + in[0] = _mm_load_si128((const __m128i *)(input + 0 * stride)); + in[1] = _mm_load_si128((const __m128i *)(input + 1 * stride)); + in[2] = _mm_load_si128((const __m128i *)(input + 2 * stride)); + in[3] = _mm_load_si128((const __m128i *)(input + 3 * stride)); + in[4] = _mm_load_si128((const __m128i *)(input + 4 * stride)); + in[5] = _mm_load_si128((const __m128i *)(input + 5 * stride)); + in[6] = _mm_load_si128((const __m128i *)(input + 6 * stride)); + in[7] = _mm_load_si128((const __m128i *)(input + 7 * stride)); + } else { + in[0] = _mm_load_si128((const __m128i *)(input + 7 * stride)); + in[1] = _mm_load_si128((const __m128i *)(input + 6 * stride)); + in[2] = _mm_load_si128((const __m128i *)(input + 5 * stride)); + in[3] = _mm_load_si128((const __m128i *)(input + 4 * stride)); + in[4] = _mm_load_si128((const __m128i *)(input + 3 * stride)); + in[5] = _mm_load_si128((const __m128i *)(input + 2 * stride)); + in[6] = _mm_load_si128((const __m128i *)(input + 1 * stride)); + in[7] = _mm_load_si128((const __m128i *)(input + 0 * stride)); + } + + if (fliplr) { + in[0] = mm_reverse_epi16(in[0]); + in[1] = mm_reverse_epi16(in[1]); + in[2] = mm_reverse_epi16(in[2]); + in[3] = mm_reverse_epi16(in[3]); + in[4] = mm_reverse_epi16(in[4]); + in[5] = mm_reverse_epi16(in[5]); + in[6] = mm_reverse_epi16(in[6]); + in[7] = mm_reverse_epi16(in[7]); + } + + u = _mm_unpackhi_epi64(in[4], in[4]); + in[8] = _mm_cvtepi16_epi32(in[4]); + in[9] = _mm_cvtepi16_epi32(u); + + u = _mm_unpackhi_epi64(in[5], in[5]); + in[10] = _mm_cvtepi16_epi32(in[5]); + in[11] = _mm_cvtepi16_epi32(u); + + u = _mm_unpackhi_epi64(in[6], in[6]); + in[12] = _mm_cvtepi16_epi32(in[6]); + in[13] = _mm_cvtepi16_epi32(u); + + u = _mm_unpackhi_epi64(in[7], in[7]); + in[14] = _mm_cvtepi16_epi32(in[7]); + in[15] = _mm_cvtepi16_epi32(u); + + u = _mm_unpackhi_epi64(in[3], in[3]); + in[6] = _mm_cvtepi16_epi32(in[3]); + in[7] = _mm_cvtepi16_epi32(u); + + u = _mm_unpackhi_epi64(in[2], in[2]); + in[4] = _mm_cvtepi16_epi32(in[2]); + in[5] = _mm_cvtepi16_epi32(u); + + u = _mm_unpackhi_epi64(in[1], in[1]); + in[2] = _mm_cvtepi16_epi32(in[1]); + in[3] = _mm_cvtepi16_epi32(u); + + u = _mm_unpackhi_epi64(in[0], in[0]); + in[0] = _mm_cvtepi16_epi32(in[0]); + in[1] = _mm_cvtepi16_epi32(u); + + in[0] = _mm_slli_epi32(in[0], shift); + in[1] = _mm_slli_epi32(in[1], shift); + in[2] = _mm_slli_epi32(in[2], shift); + in[3] = _mm_slli_epi32(in[3], shift); + in[4] = _mm_slli_epi32(in[4], shift); + in[5] = _mm_slli_epi32(in[5], shift); + in[6] = _mm_slli_epi32(in[6], shift); + in[7] = _mm_slli_epi32(in[7], shift); + + in[8] = _mm_slli_epi32(in[8], shift); + in[9] = _mm_slli_epi32(in[9], shift); + in[10] = _mm_slli_epi32(in[10], shift); + in[11] = _mm_slli_epi32(in[11], shift); + in[12] = _mm_slli_epi32(in[12], shift); + in[13] = _mm_slli_epi32(in[13], shift); + in[14] = _mm_slli_epi32(in[14], shift); + in[15] = _mm_slli_epi32(in[15], shift); +} + +static INLINE void col_txfm_8x8_rounding(__m128i *in, int shift) { + const __m128i rounding = _mm_set1_epi32(1 << (shift - 1)); + + in[0] = _mm_add_epi32(in[0], rounding); + in[1] = _mm_add_epi32(in[1], rounding); + in[2] = _mm_add_epi32(in[2], rounding); + in[3] = _mm_add_epi32(in[3], rounding); + in[4] = _mm_add_epi32(in[4], rounding); + in[5] = _mm_add_epi32(in[5], rounding); + in[6] = _mm_add_epi32(in[6], rounding); + in[7] = _mm_add_epi32(in[7], rounding); + in[8] = _mm_add_epi32(in[8], rounding); + in[9] = _mm_add_epi32(in[9], rounding); + in[10] = _mm_add_epi32(in[10], rounding); + in[11] = _mm_add_epi32(in[11], rounding); + in[12] = _mm_add_epi32(in[12], rounding); + in[13] = _mm_add_epi32(in[13], rounding); + in[14] = _mm_add_epi32(in[14], rounding); + in[15] = _mm_add_epi32(in[15], rounding); + + in[0] = _mm_srai_epi32(in[0], shift); + in[1] = _mm_srai_epi32(in[1], shift); + in[2] = _mm_srai_epi32(in[2], shift); + in[3] = _mm_srai_epi32(in[3], shift); + in[4] = _mm_srai_epi32(in[4], shift); + in[5] = _mm_srai_epi32(in[5], shift); + in[6] = _mm_srai_epi32(in[6], shift); + in[7] = _mm_srai_epi32(in[7], shift); + in[8] = _mm_srai_epi32(in[8], shift); + in[9] = _mm_srai_epi32(in[9], shift); + in[10] = _mm_srai_epi32(in[10], shift); + in[11] = _mm_srai_epi32(in[11], shift); + in[12] = _mm_srai_epi32(in[12], shift); + in[13] = _mm_srai_epi32(in[13], shift); + in[14] = _mm_srai_epi32(in[14], shift); + in[15] = _mm_srai_epi32(in[15], shift); +} + +static INLINE void write_buffer_8x8(const __m128i *res, int32_t *output) { + _mm_store_si128((__m128i *)(output + 0 * 4), res[0]); + _mm_store_si128((__m128i *)(output + 1 * 4), res[1]); + _mm_store_si128((__m128i *)(output + 2 * 4), res[2]); + _mm_store_si128((__m128i *)(output + 3 * 4), res[3]); + + _mm_store_si128((__m128i *)(output + 4 * 4), res[4]); + _mm_store_si128((__m128i *)(output + 5 * 4), res[5]); + _mm_store_si128((__m128i *)(output + 6 * 4), res[6]); + _mm_store_si128((__m128i *)(output + 7 * 4), res[7]); + + _mm_store_si128((__m128i *)(output + 8 * 4), res[8]); + _mm_store_si128((__m128i *)(output + 9 * 4), res[9]); + _mm_store_si128((__m128i *)(output + 10 * 4), res[10]); + _mm_store_si128((__m128i *)(output + 11 * 4), res[11]); + + _mm_store_si128((__m128i *)(output + 12 * 4), res[12]); + _mm_store_si128((__m128i *)(output + 13 * 4), res[13]); + _mm_store_si128((__m128i *)(output + 14 * 4), res[14]); + _mm_store_si128((__m128i *)(output + 15 * 4), res[15]); +} + +static INLINE void write_buffer_16x8(const __m128i *res, int32_t *output, + const int stride) { + _mm_storeu_si128((__m128i *)(output), res[0]); + _mm_storeu_si128((__m128i *)(output + 4), res[1]); + _mm_storeu_si128((__m128i *)(output + stride), res[2]); + _mm_storeu_si128((__m128i *)(output + stride + 4), res[3]); + + _mm_storeu_si128((__m128i *)(output + (stride * 2)), res[4]); + _mm_storeu_si128((__m128i *)(output + (stride * 2) + 4), res[5]); + _mm_storeu_si128((__m128i *)(output + (stride * 3)), res[6]); + _mm_storeu_si128((__m128i *)(output + (stride * 3) + 4), res[7]); + + _mm_storeu_si128((__m128i *)(output + (stride * 4)), res[8]); + _mm_storeu_si128((__m128i *)(output + (stride * 4) + 4), res[9]); + _mm_storeu_si128((__m128i *)(output + (stride * 5)), res[10]); + _mm_storeu_si128((__m128i *)(output + (stride * 5) + 4), res[11]); + + _mm_storeu_si128((__m128i *)(output + (stride * 6)), res[12]); + _mm_storeu_si128((__m128i *)(output + (stride * 6) + 4), res[13]); + _mm_storeu_si128((__m128i *)(output + (stride * 7)), res[14]); + _mm_storeu_si128((__m128i *)(output + (stride * 7) + 4), res[15]); +} + +static void fdct8x8_sse4_1(__m128i *in, __m128i *out, int bit, + const int col_num) { + (void)(col_num); + const int32_t *cospi = cospi_arr(bit); + const __m128i cospi32 = _mm_set1_epi32(cospi[32]); + const __m128i cospim32 = _mm_set1_epi32(-cospi[32]); + const __m128i cospi48 = _mm_set1_epi32(cospi[48]); + const __m128i cospi16 = _mm_set1_epi32(cospi[16]); + const __m128i cospi56 = _mm_set1_epi32(cospi[56]); + const __m128i cospi8 = _mm_set1_epi32(cospi[8]); + const __m128i cospi24 = _mm_set1_epi32(cospi[24]); + const __m128i cospi40 = _mm_set1_epi32(cospi[40]); + const __m128i rnding = _mm_set1_epi32(1 << (bit - 1)); + __m128i u[8], v[8]; + + // Even 8 points 0, 2, ..., 14 + // stage 0 + // stage 1 + u[0] = _mm_add_epi32(in[0], in[14]); + v[7] = _mm_sub_epi32(in[0], in[14]); // v[7] + u[1] = _mm_add_epi32(in[2], in[12]); + u[6] = _mm_sub_epi32(in[2], in[12]); + u[2] = _mm_add_epi32(in[4], in[10]); + u[5] = _mm_sub_epi32(in[4], in[10]); + u[3] = _mm_add_epi32(in[6], in[8]); + v[4] = _mm_sub_epi32(in[6], in[8]); // v[4] + + // stage 2 + v[0] = _mm_add_epi32(u[0], u[3]); + v[3] = _mm_sub_epi32(u[0], u[3]); + v[1] = _mm_add_epi32(u[1], u[2]); + v[2] = _mm_sub_epi32(u[1], u[2]); + + v[5] = _mm_mullo_epi32(u[5], cospim32); + v[6] = _mm_mullo_epi32(u[6], cospi32); + v[5] = _mm_add_epi32(v[5], v[6]); + v[5] = _mm_add_epi32(v[5], rnding); + v[5] = _mm_srai_epi32(v[5], bit); + + u[0] = _mm_mullo_epi32(u[5], cospi32); + v[6] = _mm_mullo_epi32(u[6], cospim32); + v[6] = _mm_sub_epi32(u[0], v[6]); + v[6] = _mm_add_epi32(v[6], rnding); + v[6] = _mm_srai_epi32(v[6], bit); + + // stage 3 + // type 0 + v[0] = _mm_mullo_epi32(v[0], cospi32); + v[1] = _mm_mullo_epi32(v[1], cospi32); + u[0] = _mm_add_epi32(v[0], v[1]); + u[0] = _mm_add_epi32(u[0], rnding); + u[0] = _mm_srai_epi32(u[0], bit); + + u[1] = _mm_sub_epi32(v[0], v[1]); + u[1] = _mm_add_epi32(u[1], rnding); + u[1] = _mm_srai_epi32(u[1], bit); + + // type 1 + v[0] = _mm_mullo_epi32(v[2], cospi48); + v[1] = _mm_mullo_epi32(v[3], cospi16); + u[2] = _mm_add_epi32(v[0], v[1]); + u[2] = _mm_add_epi32(u[2], rnding); + u[2] = _mm_srai_epi32(u[2], bit); + + v[0] = _mm_mullo_epi32(v[2], cospi16); + v[1] = _mm_mullo_epi32(v[3], cospi48); + u[3] = _mm_sub_epi32(v[1], v[0]); + u[3] = _mm_add_epi32(u[3], rnding); + u[3] = _mm_srai_epi32(u[3], bit); + + u[4] = _mm_add_epi32(v[4], v[5]); + u[5] = _mm_sub_epi32(v[4], v[5]); + u[6] = _mm_sub_epi32(v[7], v[6]); + u[7] = _mm_add_epi32(v[7], v[6]); + + // stage 4 + // stage 5 + v[0] = _mm_mullo_epi32(u[4], cospi56); + v[1] = _mm_mullo_epi32(u[7], cospi8); + v[0] = _mm_add_epi32(v[0], v[1]); + v[0] = _mm_add_epi32(v[0], rnding); + out[2] = _mm_srai_epi32(v[0], bit); // buf0[4] + + v[0] = _mm_mullo_epi32(u[4], cospi8); + v[1] = _mm_mullo_epi32(u[7], cospi56); + v[0] = _mm_sub_epi32(v[1], v[0]); + v[0] = _mm_add_epi32(v[0], rnding); + out[14] = _mm_srai_epi32(v[0], bit); // buf0[7] + + v[0] = _mm_mullo_epi32(u[5], cospi24); + v[1] = _mm_mullo_epi32(u[6], cospi40); + v[0] = _mm_add_epi32(v[0], v[1]); + v[0] = _mm_add_epi32(v[0], rnding); + out[10] = _mm_srai_epi32(v[0], bit); // buf0[5] + + v[0] = _mm_mullo_epi32(u[5], cospi40); + v[1] = _mm_mullo_epi32(u[6], cospi24); + v[0] = _mm_sub_epi32(v[1], v[0]); + v[0] = _mm_add_epi32(v[0], rnding); + out[6] = _mm_srai_epi32(v[0], bit); // buf0[6] + + out[0] = u[0]; // buf0[0] + out[8] = u[1]; // buf0[1] + out[4] = u[2]; // buf0[2] + out[12] = u[3]; // buf0[3] + + // Odd 8 points: 1, 3, ..., 15 + // stage 0 + // stage 1 + u[0] = _mm_add_epi32(in[1], in[15]); + v[7] = _mm_sub_epi32(in[1], in[15]); // v[7] + u[1] = _mm_add_epi32(in[3], in[13]); + u[6] = _mm_sub_epi32(in[3], in[13]); + u[2] = _mm_add_epi32(in[5], in[11]); + u[5] = _mm_sub_epi32(in[5], in[11]); + u[3] = _mm_add_epi32(in[7], in[9]); + v[4] = _mm_sub_epi32(in[7], in[9]); // v[4] + + // stage 2 + v[0] = _mm_add_epi32(u[0], u[3]); + v[3] = _mm_sub_epi32(u[0], u[3]); + v[1] = _mm_add_epi32(u[1], u[2]); + v[2] = _mm_sub_epi32(u[1], u[2]); + + v[5] = _mm_mullo_epi32(u[5], cospim32); + v[6] = _mm_mullo_epi32(u[6], cospi32); + v[5] = _mm_add_epi32(v[5], v[6]); + v[5] = _mm_add_epi32(v[5], rnding); + v[5] = _mm_srai_epi32(v[5], bit); + + u[0] = _mm_mullo_epi32(u[5], cospi32); + v[6] = _mm_mullo_epi32(u[6], cospim32); + v[6] = _mm_sub_epi32(u[0], v[6]); + v[6] = _mm_add_epi32(v[6], rnding); + v[6] = _mm_srai_epi32(v[6], bit); + + // stage 3 + // type 0 + v[0] = _mm_mullo_epi32(v[0], cospi32); + v[1] = _mm_mullo_epi32(v[1], cospi32); + u[0] = _mm_add_epi32(v[0], v[1]); + u[0] = _mm_add_epi32(u[0], rnding); + u[0] = _mm_srai_epi32(u[0], bit); + + u[1] = _mm_sub_epi32(v[0], v[1]); + u[1] = _mm_add_epi32(u[1], rnding); + u[1] = _mm_srai_epi32(u[1], bit); + + // type 1 + v[0] = _mm_mullo_epi32(v[2], cospi48); + v[1] = _mm_mullo_epi32(v[3], cospi16); + u[2] = _mm_add_epi32(v[0], v[1]); + u[2] = _mm_add_epi32(u[2], rnding); + u[2] = _mm_srai_epi32(u[2], bit); + + v[0] = _mm_mullo_epi32(v[2], cospi16); + v[1] = _mm_mullo_epi32(v[3], cospi48); + u[3] = _mm_sub_epi32(v[1], v[0]); + u[3] = _mm_add_epi32(u[3], rnding); + u[3] = _mm_srai_epi32(u[3], bit); + + u[4] = _mm_add_epi32(v[4], v[5]); + u[5] = _mm_sub_epi32(v[4], v[5]); + u[6] = _mm_sub_epi32(v[7], v[6]); + u[7] = _mm_add_epi32(v[7], v[6]); + + // stage 4 + // stage 5 + v[0] = _mm_mullo_epi32(u[4], cospi56); + v[1] = _mm_mullo_epi32(u[7], cospi8); + v[0] = _mm_add_epi32(v[0], v[1]); + v[0] = _mm_add_epi32(v[0], rnding); + out[3] = _mm_srai_epi32(v[0], bit); // buf0[4] + + v[0] = _mm_mullo_epi32(u[4], cospi8); + v[1] = _mm_mullo_epi32(u[7], cospi56); + v[0] = _mm_sub_epi32(v[1], v[0]); + v[0] = _mm_add_epi32(v[0], rnding); + out[15] = _mm_srai_epi32(v[0], bit); // buf0[7] + + v[0] = _mm_mullo_epi32(u[5], cospi24); + v[1] = _mm_mullo_epi32(u[6], cospi40); + v[0] = _mm_add_epi32(v[0], v[1]); + v[0] = _mm_add_epi32(v[0], rnding); + out[11] = _mm_srai_epi32(v[0], bit); // buf0[5] + + v[0] = _mm_mullo_epi32(u[5], cospi40); + v[1] = _mm_mullo_epi32(u[6], cospi24); + v[0] = _mm_sub_epi32(v[1], v[0]); + v[0] = _mm_add_epi32(v[0], rnding); + out[7] = _mm_srai_epi32(v[0], bit); // buf0[6] + + out[1] = u[0]; // buf0[0] + out[9] = u[1]; // buf0[1] + out[5] = u[2]; // buf0[2] + out[13] = u[3]; // buf0[3] +} + +static void fadst8x8_sse4_1(__m128i *in, __m128i *out, int bit, + const int col_num) { + (void)(col_num); + const int32_t *cospi = cospi_arr(bit); + const __m128i cospi32 = _mm_set1_epi32(cospi[32]); + const __m128i cospi16 = _mm_set1_epi32(cospi[16]); + const __m128i cospim16 = _mm_set1_epi32(-cospi[16]); + const __m128i cospi48 = _mm_set1_epi32(cospi[48]); + const __m128i cospim48 = _mm_set1_epi32(-cospi[48]); + const __m128i cospi4 = _mm_set1_epi32(cospi[4]); + const __m128i cospim4 = _mm_set1_epi32(-cospi[4]); + const __m128i cospi60 = _mm_set1_epi32(cospi[60]); + const __m128i cospi20 = _mm_set1_epi32(cospi[20]); + const __m128i cospim20 = _mm_set1_epi32(-cospi[20]); + const __m128i cospi44 = _mm_set1_epi32(cospi[44]); + const __m128i cospi28 = _mm_set1_epi32(cospi[28]); + const __m128i cospi36 = _mm_set1_epi32(cospi[36]); + const __m128i cospim36 = _mm_set1_epi32(-cospi[36]); + const __m128i cospi52 = _mm_set1_epi32(cospi[52]); + const __m128i cospim52 = _mm_set1_epi32(-cospi[52]); + const __m128i cospi12 = _mm_set1_epi32(cospi[12]); + const __m128i rnding = _mm_set1_epi32(1 << (bit - 1)); + const __m128i zero = _mm_setzero_si128(); + __m128i u0, u1, u2, u3, u4, u5, u6, u7; + __m128i v0, v1, v2, v3, v4, v5, v6, v7; + __m128i x, y; + int col; + + // Note: + // Even column: 0, 2, ..., 14 + // Odd column: 1, 3, ..., 15 + // one even column plus one odd column constructs one row (8 coeffs) + // total we have 8 rows (8x8). + for (col = 0; col < 2; ++col) { + // stage 0 + // stage 1 + u0 = in[2 * 0 + col]; + u1 = _mm_sub_epi32(zero, in[2 * 7 + col]); + u2 = _mm_sub_epi32(zero, in[2 * 3 + col]); + u3 = in[2 * 4 + col]; + u4 = _mm_sub_epi32(zero, in[2 * 1 + col]); + u5 = in[2 * 6 + col]; + u6 = in[2 * 2 + col]; + u7 = _mm_sub_epi32(zero, in[2 * 5 + col]); + + // stage 2 + v0 = u0; + v1 = u1; + + x = _mm_mullo_epi32(u2, cospi32); + y = _mm_mullo_epi32(u3, cospi32); + v2 = _mm_add_epi32(x, y); + v2 = _mm_add_epi32(v2, rnding); + v2 = _mm_srai_epi32(v2, bit); + + v3 = _mm_sub_epi32(x, y); + v3 = _mm_add_epi32(v3, rnding); + v3 = _mm_srai_epi32(v3, bit); + + v4 = u4; + v5 = u5; + + x = _mm_mullo_epi32(u6, cospi32); + y = _mm_mullo_epi32(u7, cospi32); + v6 = _mm_add_epi32(x, y); + v6 = _mm_add_epi32(v6, rnding); + v6 = _mm_srai_epi32(v6, bit); + + v7 = _mm_sub_epi32(x, y); + v7 = _mm_add_epi32(v7, rnding); + v7 = _mm_srai_epi32(v7, bit); + + // stage 3 + u0 = _mm_add_epi32(v0, v2); + u1 = _mm_add_epi32(v1, v3); + u2 = _mm_sub_epi32(v0, v2); + u3 = _mm_sub_epi32(v1, v3); + u4 = _mm_add_epi32(v4, v6); + u5 = _mm_add_epi32(v5, v7); + u6 = _mm_sub_epi32(v4, v6); + u7 = _mm_sub_epi32(v5, v7); + + // stage 4 + v0 = u0; + v1 = u1; + v2 = u2; + v3 = u3; + + x = _mm_mullo_epi32(u4, cospi16); + y = _mm_mullo_epi32(u5, cospi48); + v4 = _mm_add_epi32(x, y); + v4 = _mm_add_epi32(v4, rnding); + v4 = _mm_srai_epi32(v4, bit); + + x = _mm_mullo_epi32(u4, cospi48); + y = _mm_mullo_epi32(u5, cospim16); + v5 = _mm_add_epi32(x, y); + v5 = _mm_add_epi32(v5, rnding); + v5 = _mm_srai_epi32(v5, bit); + + x = _mm_mullo_epi32(u6, cospim48); + y = _mm_mullo_epi32(u7, cospi16); + v6 = _mm_add_epi32(x, y); + v6 = _mm_add_epi32(v6, rnding); + v6 = _mm_srai_epi32(v6, bit); + + x = _mm_mullo_epi32(u6, cospi16); + y = _mm_mullo_epi32(u7, cospi48); + v7 = _mm_add_epi32(x, y); + v7 = _mm_add_epi32(v7, rnding); + v7 = _mm_srai_epi32(v7, bit); + + // stage 5 + u0 = _mm_add_epi32(v0, v4); + u1 = _mm_add_epi32(v1, v5); + u2 = _mm_add_epi32(v2, v6); + u3 = _mm_add_epi32(v3, v7); + u4 = _mm_sub_epi32(v0, v4); + u5 = _mm_sub_epi32(v1, v5); + u6 = _mm_sub_epi32(v2, v6); + u7 = _mm_sub_epi32(v3, v7); + + // stage 6 + x = _mm_mullo_epi32(u0, cospi4); + y = _mm_mullo_epi32(u1, cospi60); + v0 = _mm_add_epi32(x, y); + v0 = _mm_add_epi32(v0, rnding); + v0 = _mm_srai_epi32(v0, bit); + + x = _mm_mullo_epi32(u0, cospi60); + y = _mm_mullo_epi32(u1, cospim4); + v1 = _mm_add_epi32(x, y); + v1 = _mm_add_epi32(v1, rnding); + v1 = _mm_srai_epi32(v1, bit); + + x = _mm_mullo_epi32(u2, cospi20); + y = _mm_mullo_epi32(u3, cospi44); + v2 = _mm_add_epi32(x, y); + v2 = _mm_add_epi32(v2, rnding); + v2 = _mm_srai_epi32(v2, bit); + + x = _mm_mullo_epi32(u2, cospi44); + y = _mm_mullo_epi32(u3, cospim20); + v3 = _mm_add_epi32(x, y); + v3 = _mm_add_epi32(v3, rnding); + v3 = _mm_srai_epi32(v3, bit); + + x = _mm_mullo_epi32(u4, cospi36); + y = _mm_mullo_epi32(u5, cospi28); + v4 = _mm_add_epi32(x, y); + v4 = _mm_add_epi32(v4, rnding); + v4 = _mm_srai_epi32(v4, bit); + + x = _mm_mullo_epi32(u4, cospi28); + y = _mm_mullo_epi32(u5, cospim36); + v5 = _mm_add_epi32(x, y); + v5 = _mm_add_epi32(v5, rnding); + v5 = _mm_srai_epi32(v5, bit); + + x = _mm_mullo_epi32(u6, cospi52); + y = _mm_mullo_epi32(u7, cospi12); + v6 = _mm_add_epi32(x, y); + v6 = _mm_add_epi32(v6, rnding); + v6 = _mm_srai_epi32(v6, bit); + + x = _mm_mullo_epi32(u6, cospi12); + y = _mm_mullo_epi32(u7, cospim52); + v7 = _mm_add_epi32(x, y); + v7 = _mm_add_epi32(v7, rnding); + v7 = _mm_srai_epi32(v7, bit); + + // stage 7 + out[2 * 0 + col] = v1; + out[2 * 1 + col] = v6; + out[2 * 2 + col] = v3; + out[2 * 3 + col] = v4; + out[2 * 4 + col] = v5; + out[2 * 5 + col] = v2; + out[2 * 6 + col] = v7; + out[2 * 7 + col] = v0; + } +} + +void av1_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *coeff, int stride, + TX_TYPE tx_type, int bd) { + __m128i in[16], out[16]; + const int8_t *shift = fwd_txfm_shift_ls[TX_8X8]; + const int txw_idx = get_txw_idx(TX_8X8); + const int txh_idx = get_txh_idx(TX_8X8); + + switch (tx_type) { + case DCT_DCT: + load_buffer_8x8(input, in, stride, 0, 0, shift[0]); + fdct8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0); + col_txfm_8x8_rounding(out, -shift[1]); + transpose_8x8(out, in); + fdct8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0); + transpose_8x8(out, in); + write_buffer_8x8(in, coeff); + break; + case ADST_DCT: + load_buffer_8x8(input, in, stride, 0, 0, shift[0]); + fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0); + col_txfm_8x8_rounding(out, -shift[1]); + transpose_8x8(out, in); + fdct8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0); + transpose_8x8(out, in); + write_buffer_8x8(in, coeff); + break; + case DCT_ADST: + load_buffer_8x8(input, in, stride, 0, 0, shift[0]); + fdct8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0); + col_txfm_8x8_rounding(out, -shift[1]); + transpose_8x8(out, in); + fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0); + transpose_8x8(out, in); + write_buffer_8x8(in, coeff); + break; + case ADST_ADST: + load_buffer_8x8(input, in, stride, 0, 0, shift[0]); + fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0); + col_txfm_8x8_rounding(out, -shift[1]); + transpose_8x8(out, in); + fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0); + transpose_8x8(out, in); + write_buffer_8x8(in, coeff); + break; + case FLIPADST_DCT: + load_buffer_8x8(input, in, stride, 1, 0, shift[0]); + fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0); + col_txfm_8x8_rounding(out, -shift[1]); + transpose_8x8(out, in); + fdct8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0); + transpose_8x8(out, in); + write_buffer_8x8(in, coeff); + break; + case DCT_FLIPADST: + load_buffer_8x8(input, in, stride, 0, 1, shift[0]); + fdct8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0); + col_txfm_8x8_rounding(out, -shift[1]); + transpose_8x8(out, in); + fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0); + transpose_8x8(out, in); + write_buffer_8x8(in, coeff); + break; + case FLIPADST_FLIPADST: + load_buffer_8x8(input, in, stride, 1, 1, shift[0]); + fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0); + col_txfm_8x8_rounding(out, -shift[1]); + transpose_8x8(out, in); + fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0); + transpose_8x8(out, in); + write_buffer_8x8(in, coeff); + break; + case ADST_FLIPADST: + load_buffer_8x8(input, in, stride, 0, 1, shift[0]); + fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0); + col_txfm_8x8_rounding(out, -shift[1]); + transpose_8x8(out, in); + fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0); + transpose_8x8(out, in); + write_buffer_8x8(in, coeff); + break; + case FLIPADST_ADST: + load_buffer_8x8(input, in, stride, 1, 0, shift[0]); + fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0); + col_txfm_8x8_rounding(out, -shift[1]); + transpose_8x8(out, in); + fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0); + transpose_8x8(out, in); + write_buffer_8x8(in, coeff); + break; + default: assert(0); + } + (void)bd; +} + +// Hybrid Transform 16x16 + +static INLINE void convert_8x8_to_16x16(const __m128i *in, __m128i *out) { + int row_index = 0; + int dst_index = 0; + int src_index = 0; + + // row 0, 1, .., 7 + do { + out[dst_index] = in[src_index]; + out[dst_index + 1] = in[src_index + 1]; + out[dst_index + 2] = in[src_index + 16]; + out[dst_index + 3] = in[src_index + 17]; + dst_index += 4; + src_index += 2; + row_index += 1; + } while (row_index < 8); + + // row 8, 9, ..., 15 + src_index += 16; + do { + out[dst_index] = in[src_index]; + out[dst_index + 1] = in[src_index + 1]; + out[dst_index + 2] = in[src_index + 16]; + out[dst_index + 3] = in[src_index + 17]; + dst_index += 4; + src_index += 2; + row_index += 1; + } while (row_index < 16); +} + +static INLINE void load_buffer_16x16(const int16_t *input, __m128i *out, + int stride, int flipud, int fliplr, + int shift) { + __m128i in[64]; + // Load 4 8x8 blocks + const int16_t *topL = input; + const int16_t *topR = input + 8; + const int16_t *botL = input + 8 * stride; + const int16_t *botR = input + 8 * stride + 8; + + const int16_t *tmp; + + if (flipud) { + // Swap left columns + tmp = topL; + topL = botL; + botL = tmp; + // Swap right columns + tmp = topR; + topR = botR; + botR = tmp; + } + + if (fliplr) { + // Swap top rows + tmp = topL; + topL = topR; + topR = tmp; + // Swap bottom rows + tmp = botL; + botL = botR; + botR = tmp; + } + + // load first 8 columns + load_buffer_8x8(topL, &in[0], stride, flipud, fliplr, shift); + load_buffer_8x8(botL, &in[32], stride, flipud, fliplr, shift); + + // load second 8 columns + load_buffer_8x8(topR, &in[16], stride, flipud, fliplr, shift); + load_buffer_8x8(botR, &in[48], stride, flipud, fliplr, shift); + + convert_8x8_to_16x16(in, out); +} + +static INLINE void load_buffer_8x16(const int16_t *input, __m128i *out, + int stride, int flipud, int fliplr, + int shift) { + const int16_t *topL = input; + const int16_t *botL = input + 8 * stride; + + const int16_t *tmp; + + if (flipud) { + tmp = topL; + topL = botL; + botL = tmp; + } + + load_buffer_8x8(topL, out, stride, flipud, fliplr, shift); + load_buffer_8x8(botL, out + 16, stride, flipud, fliplr, shift); +} + +static void fdct16x16_sse4_1(__m128i *in, __m128i *out, int bit, + const int col_num) { + const int32_t *cospi = cospi_arr(bit); + const __m128i cospi32 = _mm_set1_epi32(cospi[32]); + const __m128i cospim32 = _mm_set1_epi32(-cospi[32]); + const __m128i cospi48 = _mm_set1_epi32(cospi[48]); + const __m128i cospi16 = _mm_set1_epi32(cospi[16]); + const __m128i cospim48 = _mm_set1_epi32(-cospi[48]); + const __m128i cospim16 = _mm_set1_epi32(-cospi[16]); + const __m128i cospi56 = _mm_set1_epi32(cospi[56]); + const __m128i cospi8 = _mm_set1_epi32(cospi[8]); + const __m128i cospi24 = _mm_set1_epi32(cospi[24]); + const __m128i cospi40 = _mm_set1_epi32(cospi[40]); + const __m128i cospi60 = _mm_set1_epi32(cospi[60]); + const __m128i cospi4 = _mm_set1_epi32(cospi[4]); + const __m128i cospi28 = _mm_set1_epi32(cospi[28]); + const __m128i cospi36 = _mm_set1_epi32(cospi[36]); + const __m128i cospi44 = _mm_set1_epi32(cospi[44]); + const __m128i cospi20 = _mm_set1_epi32(cospi[20]); + const __m128i cospi12 = _mm_set1_epi32(cospi[12]); + const __m128i cospi52 = _mm_set1_epi32(cospi[52]); + const __m128i rnding = _mm_set1_epi32(1 << (bit - 1)); + __m128i u[16], v[16], x; + int col; + + // Calculate the column 0, 1, 2, 3 + for (col = 0; col < col_num; ++col) { + // stage 0 + // stage 1 + u[0] = _mm_add_epi32(in[0 * col_num + col], in[15 * col_num + col]); + u[15] = _mm_sub_epi32(in[0 * col_num + col], in[15 * col_num + col]); + u[1] = _mm_add_epi32(in[1 * col_num + col], in[14 * col_num + col]); + u[14] = _mm_sub_epi32(in[1 * col_num + col], in[14 * col_num + col]); + u[2] = _mm_add_epi32(in[2 * col_num + col], in[13 * col_num + col]); + u[13] = _mm_sub_epi32(in[2 * col_num + col], in[13 * col_num + col]); + u[3] = _mm_add_epi32(in[3 * col_num + col], in[12 * col_num + col]); + u[12] = _mm_sub_epi32(in[3 * col_num + col], in[12 * col_num + col]); + u[4] = _mm_add_epi32(in[4 * col_num + col], in[11 * col_num + col]); + u[11] = _mm_sub_epi32(in[4 * col_num + col], in[11 * col_num + col]); + u[5] = _mm_add_epi32(in[5 * col_num + col], in[10 * col_num + col]); + u[10] = _mm_sub_epi32(in[5 * col_num + col], in[10 * col_num + col]); + u[6] = _mm_add_epi32(in[6 * col_num + col], in[9 * col_num + col]); + u[9] = _mm_sub_epi32(in[6 * col_num + col], in[9 * col_num + col]); + u[7] = _mm_add_epi32(in[7 * col_num + col], in[8 * col_num + col]); + u[8] = _mm_sub_epi32(in[7 * col_num + col], in[8 * col_num + col]); + + // stage 2 + v[0] = _mm_add_epi32(u[0], u[7]); + v[7] = _mm_sub_epi32(u[0], u[7]); + v[1] = _mm_add_epi32(u[1], u[6]); + v[6] = _mm_sub_epi32(u[1], u[6]); + v[2] = _mm_add_epi32(u[2], u[5]); + v[5] = _mm_sub_epi32(u[2], u[5]); + v[3] = _mm_add_epi32(u[3], u[4]); + v[4] = _mm_sub_epi32(u[3], u[4]); + v[8] = u[8]; + v[9] = u[9]; + + v[10] = _mm_mullo_epi32(u[10], cospim32); + x = _mm_mullo_epi32(u[13], cospi32); + v[10] = _mm_add_epi32(v[10], x); + v[10] = _mm_add_epi32(v[10], rnding); + v[10] = _mm_srai_epi32(v[10], bit); + + v[13] = _mm_mullo_epi32(u[10], cospi32); + x = _mm_mullo_epi32(u[13], cospim32); + v[13] = _mm_sub_epi32(v[13], x); + v[13] = _mm_add_epi32(v[13], rnding); + v[13] = _mm_srai_epi32(v[13], bit); + + v[11] = _mm_mullo_epi32(u[11], cospim32); + x = _mm_mullo_epi32(u[12], cospi32); + v[11] = _mm_add_epi32(v[11], x); + v[11] = _mm_add_epi32(v[11], rnding); + v[11] = _mm_srai_epi32(v[11], bit); + + v[12] = _mm_mullo_epi32(u[11], cospi32); + x = _mm_mullo_epi32(u[12], cospim32); + v[12] = _mm_sub_epi32(v[12], x); + v[12] = _mm_add_epi32(v[12], rnding); + v[12] = _mm_srai_epi32(v[12], bit); + v[14] = u[14]; + v[15] = u[15]; + + // stage 3 + u[0] = _mm_add_epi32(v[0], v[3]); + u[3] = _mm_sub_epi32(v[0], v[3]); + u[1] = _mm_add_epi32(v[1], v[2]); + u[2] = _mm_sub_epi32(v[1], v[2]); + u[4] = v[4]; + + u[5] = _mm_mullo_epi32(v[5], cospim32); + x = _mm_mullo_epi32(v[6], cospi32); + u[5] = _mm_add_epi32(u[5], x); + u[5] = _mm_add_epi32(u[5], rnding); + u[5] = _mm_srai_epi32(u[5], bit); + + u[6] = _mm_mullo_epi32(v[5], cospi32); + x = _mm_mullo_epi32(v[6], cospim32); + u[6] = _mm_sub_epi32(u[6], x); + u[6] = _mm_add_epi32(u[6], rnding); + u[6] = _mm_srai_epi32(u[6], bit); + + u[7] = v[7]; + u[8] = _mm_add_epi32(v[8], v[11]); + u[11] = _mm_sub_epi32(v[8], v[11]); + u[9] = _mm_add_epi32(v[9], v[10]); + u[10] = _mm_sub_epi32(v[9], v[10]); + u[12] = _mm_sub_epi32(v[15], v[12]); + u[15] = _mm_add_epi32(v[15], v[12]); + u[13] = _mm_sub_epi32(v[14], v[13]); + u[14] = _mm_add_epi32(v[14], v[13]); + + // stage 4 + u[0] = _mm_mullo_epi32(u[0], cospi32); + u[1] = _mm_mullo_epi32(u[1], cospi32); + v[0] = _mm_add_epi32(u[0], u[1]); + v[0] = _mm_add_epi32(v[0], rnding); + v[0] = _mm_srai_epi32(v[0], bit); + + v[1] = _mm_sub_epi32(u[0], u[1]); + v[1] = _mm_add_epi32(v[1], rnding); + v[1] = _mm_srai_epi32(v[1], bit); + + v[2] = _mm_mullo_epi32(u[2], cospi48); + x = _mm_mullo_epi32(u[3], cospi16); + v[2] = _mm_add_epi32(v[2], x); + v[2] = _mm_add_epi32(v[2], rnding); + v[2] = _mm_srai_epi32(v[2], bit); + + v[3] = _mm_mullo_epi32(u[2], cospi16); + x = _mm_mullo_epi32(u[3], cospi48); + v[3] = _mm_sub_epi32(x, v[3]); + v[3] = _mm_add_epi32(v[3], rnding); + v[3] = _mm_srai_epi32(v[3], bit); + + v[4] = _mm_add_epi32(u[4], u[5]); + v[5] = _mm_sub_epi32(u[4], u[5]); + v[6] = _mm_sub_epi32(u[7], u[6]); + v[7] = _mm_add_epi32(u[7], u[6]); + v[8] = u[8]; + + v[9] = _mm_mullo_epi32(u[9], cospim16); + x = _mm_mullo_epi32(u[14], cospi48); + v[9] = _mm_add_epi32(v[9], x); + v[9] = _mm_add_epi32(v[9], rnding); + v[9] = _mm_srai_epi32(v[9], bit); + + v[14] = _mm_mullo_epi32(u[9], cospi48); + x = _mm_mullo_epi32(u[14], cospim16); + v[14] = _mm_sub_epi32(v[14], x); + v[14] = _mm_add_epi32(v[14], rnding); + v[14] = _mm_srai_epi32(v[14], bit); + + v[10] = _mm_mullo_epi32(u[10], cospim48); + x = _mm_mullo_epi32(u[13], cospim16); + v[10] = _mm_add_epi32(v[10], x); + v[10] = _mm_add_epi32(v[10], rnding); + v[10] = _mm_srai_epi32(v[10], bit); + + v[13] = _mm_mullo_epi32(u[10], cospim16); + x = _mm_mullo_epi32(u[13], cospim48); + v[13] = _mm_sub_epi32(v[13], x); + v[13] = _mm_add_epi32(v[13], rnding); + v[13] = _mm_srai_epi32(v[13], bit); + + v[11] = u[11]; + v[12] = u[12]; + v[15] = u[15]; + + // stage 5 + u[0] = v[0]; + u[1] = v[1]; + u[2] = v[2]; + u[3] = v[3]; + + u[4] = _mm_mullo_epi32(v[4], cospi56); + x = _mm_mullo_epi32(v[7], cospi8); + u[4] = _mm_add_epi32(u[4], x); + u[4] = _mm_add_epi32(u[4], rnding); + u[4] = _mm_srai_epi32(u[4], bit); + + u[7] = _mm_mullo_epi32(v[4], cospi8); + x = _mm_mullo_epi32(v[7], cospi56); + u[7] = _mm_sub_epi32(x, u[7]); + u[7] = _mm_add_epi32(u[7], rnding); + u[7] = _mm_srai_epi32(u[7], bit); + + u[5] = _mm_mullo_epi32(v[5], cospi24); + x = _mm_mullo_epi32(v[6], cospi40); + u[5] = _mm_add_epi32(u[5], x); + u[5] = _mm_add_epi32(u[5], rnding); + u[5] = _mm_srai_epi32(u[5], bit); + + u[6] = _mm_mullo_epi32(v[5], cospi40); + x = _mm_mullo_epi32(v[6], cospi24); + u[6] = _mm_sub_epi32(x, u[6]); + u[6] = _mm_add_epi32(u[6], rnding); + u[6] = _mm_srai_epi32(u[6], bit); + + u[8] = _mm_add_epi32(v[8], v[9]); + u[9] = _mm_sub_epi32(v[8], v[9]); + u[10] = _mm_sub_epi32(v[11], v[10]); + u[11] = _mm_add_epi32(v[11], v[10]); + u[12] = _mm_add_epi32(v[12], v[13]); + u[13] = _mm_sub_epi32(v[12], v[13]); + u[14] = _mm_sub_epi32(v[15], v[14]); + u[15] = _mm_add_epi32(v[15], v[14]); + + // stage 6 + v[0] = u[0]; + v[1] = u[1]; + v[2] = u[2]; + v[3] = u[3]; + v[4] = u[4]; + v[5] = u[5]; + v[6] = u[6]; + v[7] = u[7]; + + v[8] = _mm_mullo_epi32(u[8], cospi60); + x = _mm_mullo_epi32(u[15], cospi4); + v[8] = _mm_add_epi32(v[8], x); + v[8] = _mm_add_epi32(v[8], rnding); + v[8] = _mm_srai_epi32(v[8], bit); + + v[15] = _mm_mullo_epi32(u[8], cospi4); + x = _mm_mullo_epi32(u[15], cospi60); + v[15] = _mm_sub_epi32(x, v[15]); + v[15] = _mm_add_epi32(v[15], rnding); + v[15] = _mm_srai_epi32(v[15], bit); + + v[9] = _mm_mullo_epi32(u[9], cospi28); + x = _mm_mullo_epi32(u[14], cospi36); + v[9] = _mm_add_epi32(v[9], x); + v[9] = _mm_add_epi32(v[9], rnding); + v[9] = _mm_srai_epi32(v[9], bit); + + v[14] = _mm_mullo_epi32(u[9], cospi36); + x = _mm_mullo_epi32(u[14], cospi28); + v[14] = _mm_sub_epi32(x, v[14]); + v[14] = _mm_add_epi32(v[14], rnding); + v[14] = _mm_srai_epi32(v[14], bit); + + v[10] = _mm_mullo_epi32(u[10], cospi44); + x = _mm_mullo_epi32(u[13], cospi20); + v[10] = _mm_add_epi32(v[10], x); + v[10] = _mm_add_epi32(v[10], rnding); + v[10] = _mm_srai_epi32(v[10], bit); + + v[13] = _mm_mullo_epi32(u[10], cospi20); + x = _mm_mullo_epi32(u[13], cospi44); + v[13] = _mm_sub_epi32(x, v[13]); + v[13] = _mm_add_epi32(v[13], rnding); + v[13] = _mm_srai_epi32(v[13], bit); + + v[11] = _mm_mullo_epi32(u[11], cospi12); + x = _mm_mullo_epi32(u[12], cospi52); + v[11] = _mm_add_epi32(v[11], x); + v[11] = _mm_add_epi32(v[11], rnding); + v[11] = _mm_srai_epi32(v[11], bit); + + v[12] = _mm_mullo_epi32(u[11], cospi52); + x = _mm_mullo_epi32(u[12], cospi12); + v[12] = _mm_sub_epi32(x, v[12]); + v[12] = _mm_add_epi32(v[12], rnding); + v[12] = _mm_srai_epi32(v[12], bit); + + out[0 * col_num + col] = v[0]; + out[1 * col_num + col] = v[8]; + out[2 * col_num + col] = v[4]; + out[3 * col_num + col] = v[12]; + out[4 * col_num + col] = v[2]; + out[5 * col_num + col] = v[10]; + out[6 * col_num + col] = v[6]; + out[7 * col_num + col] = v[14]; + out[8 * col_num + col] = v[1]; + out[9 * col_num + col] = v[9]; + out[10 * col_num + col] = v[5]; + out[11 * col_num + col] = v[13]; + out[12 * col_num + col] = v[3]; + out[13 * col_num + col] = v[11]; + out[14 * col_num + col] = v[7]; + out[15 * col_num + col] = v[15]; + } +} + +static void fadst16x16_sse4_1(__m128i *in, __m128i *out, int bit, + const int num_cols) { + const int32_t *cospi = cospi_arr(bit); + const __m128i cospi32 = _mm_set1_epi32(cospi[32]); + const __m128i cospi48 = _mm_set1_epi32(cospi[48]); + const __m128i cospi16 = _mm_set1_epi32(cospi[16]); + const __m128i cospim16 = _mm_set1_epi32(-cospi[16]); + const __m128i cospim48 = _mm_set1_epi32(-cospi[48]); + const __m128i cospi8 = _mm_set1_epi32(cospi[8]); + const __m128i cospi56 = _mm_set1_epi32(cospi[56]); + const __m128i cospim56 = _mm_set1_epi32(-cospi[56]); + const __m128i cospim8 = _mm_set1_epi32(-cospi[8]); + const __m128i cospi24 = _mm_set1_epi32(cospi[24]); + const __m128i cospim24 = _mm_set1_epi32(-cospi[24]); + const __m128i cospim40 = _mm_set1_epi32(-cospi[40]); + const __m128i cospi40 = _mm_set1_epi32(cospi[40]); + const __m128i cospi2 = _mm_set1_epi32(cospi[2]); + const __m128i cospi62 = _mm_set1_epi32(cospi[62]); + const __m128i cospim2 = _mm_set1_epi32(-cospi[2]); + const __m128i cospi10 = _mm_set1_epi32(cospi[10]); + const __m128i cospi54 = _mm_set1_epi32(cospi[54]); + const __m128i cospim10 = _mm_set1_epi32(-cospi[10]); + const __m128i cospi18 = _mm_set1_epi32(cospi[18]); + const __m128i cospi46 = _mm_set1_epi32(cospi[46]); + const __m128i cospim18 = _mm_set1_epi32(-cospi[18]); + const __m128i cospi26 = _mm_set1_epi32(cospi[26]); + const __m128i cospi38 = _mm_set1_epi32(cospi[38]); + const __m128i cospim26 = _mm_set1_epi32(-cospi[26]); + const __m128i cospi34 = _mm_set1_epi32(cospi[34]); + const __m128i cospi30 = _mm_set1_epi32(cospi[30]); + const __m128i cospim34 = _mm_set1_epi32(-cospi[34]); + const __m128i cospi42 = _mm_set1_epi32(cospi[42]); + const __m128i cospi22 = _mm_set1_epi32(cospi[22]); + const __m128i cospim42 = _mm_set1_epi32(-cospi[42]); + const __m128i cospi50 = _mm_set1_epi32(cospi[50]); + const __m128i cospi14 = _mm_set1_epi32(cospi[14]); + const __m128i cospim50 = _mm_set1_epi32(-cospi[50]); + const __m128i cospi58 = _mm_set1_epi32(cospi[58]); + const __m128i cospi6 = _mm_set1_epi32(cospi[6]); + const __m128i cospim58 = _mm_set1_epi32(-cospi[58]); + const __m128i rnding = _mm_set1_epi32(1 << (bit - 1)); + const __m128i zero = _mm_setzero_si128(); + + __m128i u[16], v[16], x, y; + int col; + + for (col = 0; col < num_cols; ++col) { + // stage 0 + // stage 1 + u[0] = in[0 * num_cols + col]; + u[1] = _mm_sub_epi32(zero, in[15 * num_cols + col]); + u[2] = _mm_sub_epi32(zero, in[7 * num_cols + col]); + u[3] = in[8 * num_cols + col]; + u[4] = _mm_sub_epi32(zero, in[3 * num_cols + col]); + u[5] = in[12 * num_cols + col]; + u[6] = in[4 * num_cols + col]; + u[7] = _mm_sub_epi32(zero, in[11 * num_cols + col]); + u[8] = _mm_sub_epi32(zero, in[1 * num_cols + col]); + u[9] = in[14 * num_cols + col]; + u[10] = in[6 * num_cols + col]; + u[11] = _mm_sub_epi32(zero, in[9 * num_cols + col]); + u[12] = in[2 * num_cols + col]; + u[13] = _mm_sub_epi32(zero, in[13 * num_cols + col]); + u[14] = _mm_sub_epi32(zero, in[5 * num_cols + col]); + u[15] = in[10 * num_cols + col]; + + // stage 2 + v[0] = u[0]; + v[1] = u[1]; + + x = _mm_mullo_epi32(u[2], cospi32); + y = _mm_mullo_epi32(u[3], cospi32); + v[2] = _mm_add_epi32(x, y); + v[2] = _mm_add_epi32(v[2], rnding); + v[2] = _mm_srai_epi32(v[2], bit); + + v[3] = _mm_sub_epi32(x, y); + v[3] = _mm_add_epi32(v[3], rnding); + v[3] = _mm_srai_epi32(v[3], bit); + + v[4] = u[4]; + v[5] = u[5]; + + x = _mm_mullo_epi32(u[6], cospi32); + y = _mm_mullo_epi32(u[7], cospi32); + v[6] = _mm_add_epi32(x, y); + v[6] = _mm_add_epi32(v[6], rnding); + v[6] = _mm_srai_epi32(v[6], bit); + + v[7] = _mm_sub_epi32(x, y); + v[7] = _mm_add_epi32(v[7], rnding); + v[7] = _mm_srai_epi32(v[7], bit); + + v[8] = u[8]; + v[9] = u[9]; + + x = _mm_mullo_epi32(u[10], cospi32); + y = _mm_mullo_epi32(u[11], cospi32); + v[10] = _mm_add_epi32(x, y); + v[10] = _mm_add_epi32(v[10], rnding); + v[10] = _mm_srai_epi32(v[10], bit); + + v[11] = _mm_sub_epi32(x, y); + v[11] = _mm_add_epi32(v[11], rnding); + v[11] = _mm_srai_epi32(v[11], bit); + + v[12] = u[12]; + v[13] = u[13]; + + x = _mm_mullo_epi32(u[14], cospi32); + y = _mm_mullo_epi32(u[15], cospi32); + v[14] = _mm_add_epi32(x, y); + v[14] = _mm_add_epi32(v[14], rnding); + v[14] = _mm_srai_epi32(v[14], bit); + + v[15] = _mm_sub_epi32(x, y); + v[15] = _mm_add_epi32(v[15], rnding); + v[15] = _mm_srai_epi32(v[15], bit); + + // stage 3 + u[0] = _mm_add_epi32(v[0], v[2]); + u[1] = _mm_add_epi32(v[1], v[3]); + u[2] = _mm_sub_epi32(v[0], v[2]); + u[3] = _mm_sub_epi32(v[1], v[3]); + u[4] = _mm_add_epi32(v[4], v[6]); + u[5] = _mm_add_epi32(v[5], v[7]); + u[6] = _mm_sub_epi32(v[4], v[6]); + u[7] = _mm_sub_epi32(v[5], v[7]); + u[8] = _mm_add_epi32(v[8], v[10]); + u[9] = _mm_add_epi32(v[9], v[11]); + u[10] = _mm_sub_epi32(v[8], v[10]); + u[11] = _mm_sub_epi32(v[9], v[11]); + u[12] = _mm_add_epi32(v[12], v[14]); + u[13] = _mm_add_epi32(v[13], v[15]); + u[14] = _mm_sub_epi32(v[12], v[14]); + u[15] = _mm_sub_epi32(v[13], v[15]); + + // stage 4 + v[0] = u[0]; + v[1] = u[1]; + v[2] = u[2]; + v[3] = u[3]; + v[4] = half_btf_sse4_1(&cospi16, &u[4], &cospi48, &u[5], &rnding, bit); + v[5] = half_btf_sse4_1(&cospi48, &u[4], &cospim16, &u[5], &rnding, bit); + v[6] = half_btf_sse4_1(&cospim48, &u[6], &cospi16, &u[7], &rnding, bit); + v[7] = half_btf_sse4_1(&cospi16, &u[6], &cospi48, &u[7], &rnding, bit); + v[8] = u[8]; + v[9] = u[9]; + v[10] = u[10]; + v[11] = u[11]; + v[12] = half_btf_sse4_1(&cospi16, &u[12], &cospi48, &u[13], &rnding, bit); + v[13] = half_btf_sse4_1(&cospi48, &u[12], &cospim16, &u[13], &rnding, bit); + v[14] = half_btf_sse4_1(&cospim48, &u[14], &cospi16, &u[15], &rnding, bit); + v[15] = half_btf_sse4_1(&cospi16, &u[14], &cospi48, &u[15], &rnding, bit); + + // stage 5 + u[0] = _mm_add_epi32(v[0], v[4]); + u[1] = _mm_add_epi32(v[1], v[5]); + u[2] = _mm_add_epi32(v[2], v[6]); + u[3] = _mm_add_epi32(v[3], v[7]); + u[4] = _mm_sub_epi32(v[0], v[4]); + u[5] = _mm_sub_epi32(v[1], v[5]); + u[6] = _mm_sub_epi32(v[2], v[6]); + u[7] = _mm_sub_epi32(v[3], v[7]); + u[8] = _mm_add_epi32(v[8], v[12]); + u[9] = _mm_add_epi32(v[9], v[13]); + u[10] = _mm_add_epi32(v[10], v[14]); + u[11] = _mm_add_epi32(v[11], v[15]); + u[12] = _mm_sub_epi32(v[8], v[12]); + u[13] = _mm_sub_epi32(v[9], v[13]); + u[14] = _mm_sub_epi32(v[10], v[14]); + u[15] = _mm_sub_epi32(v[11], v[15]); + + // stage 6 + v[0] = u[0]; + v[1] = u[1]; + v[2] = u[2]; + v[3] = u[3]; + v[4] = u[4]; + v[5] = u[5]; + v[6] = u[6]; + v[7] = u[7]; + v[8] = half_btf_sse4_1(&cospi8, &u[8], &cospi56, &u[9], &rnding, bit); + v[9] = half_btf_sse4_1(&cospi56, &u[8], &cospim8, &u[9], &rnding, bit); + v[10] = half_btf_sse4_1(&cospi40, &u[10], &cospi24, &u[11], &rnding, bit); + v[11] = half_btf_sse4_1(&cospi24, &u[10], &cospim40, &u[11], &rnding, bit); + v[12] = half_btf_sse4_1(&cospim56, &u[12], &cospi8, &u[13], &rnding, bit); + v[13] = half_btf_sse4_1(&cospi8, &u[12], &cospi56, &u[13], &rnding, bit); + v[14] = half_btf_sse4_1(&cospim24, &u[14], &cospi40, &u[15], &rnding, bit); + v[15] = half_btf_sse4_1(&cospi40, &u[14], &cospi24, &u[15], &rnding, bit); + + // stage 7 + u[0] = _mm_add_epi32(v[0], v[8]); + u[1] = _mm_add_epi32(v[1], v[9]); + u[2] = _mm_add_epi32(v[2], v[10]); + u[3] = _mm_add_epi32(v[3], v[11]); + u[4] = _mm_add_epi32(v[4], v[12]); + u[5] = _mm_add_epi32(v[5], v[13]); + u[6] = _mm_add_epi32(v[6], v[14]); + u[7] = _mm_add_epi32(v[7], v[15]); + u[8] = _mm_sub_epi32(v[0], v[8]); + u[9] = _mm_sub_epi32(v[1], v[9]); + u[10] = _mm_sub_epi32(v[2], v[10]); + u[11] = _mm_sub_epi32(v[3], v[11]); + u[12] = _mm_sub_epi32(v[4], v[12]); + u[13] = _mm_sub_epi32(v[5], v[13]); + u[14] = _mm_sub_epi32(v[6], v[14]); + u[15] = _mm_sub_epi32(v[7], v[15]); + + // stage 8 + v[0] = half_btf_sse4_1(&cospi2, &u[0], &cospi62, &u[1], &rnding, bit); + v[1] = half_btf_sse4_1(&cospi62, &u[0], &cospim2, &u[1], &rnding, bit); + v[2] = half_btf_sse4_1(&cospi10, &u[2], &cospi54, &u[3], &rnding, bit); + v[3] = half_btf_sse4_1(&cospi54, &u[2], &cospim10, &u[3], &rnding, bit); + v[4] = half_btf_sse4_1(&cospi18, &u[4], &cospi46, &u[5], &rnding, bit); + v[5] = half_btf_sse4_1(&cospi46, &u[4], &cospim18, &u[5], &rnding, bit); + v[6] = half_btf_sse4_1(&cospi26, &u[6], &cospi38, &u[7], &rnding, bit); + v[7] = half_btf_sse4_1(&cospi38, &u[6], &cospim26, &u[7], &rnding, bit); + v[8] = half_btf_sse4_1(&cospi34, &u[8], &cospi30, &u[9], &rnding, bit); + v[9] = half_btf_sse4_1(&cospi30, &u[8], &cospim34, &u[9], &rnding, bit); + v[10] = half_btf_sse4_1(&cospi42, &u[10], &cospi22, &u[11], &rnding, bit); + v[11] = half_btf_sse4_1(&cospi22, &u[10], &cospim42, &u[11], &rnding, bit); + v[12] = half_btf_sse4_1(&cospi50, &u[12], &cospi14, &u[13], &rnding, bit); + v[13] = half_btf_sse4_1(&cospi14, &u[12], &cospim50, &u[13], &rnding, bit); + v[14] = half_btf_sse4_1(&cospi58, &u[14], &cospi6, &u[15], &rnding, bit); + v[15] = half_btf_sse4_1(&cospi6, &u[14], &cospim58, &u[15], &rnding, bit); + + // stage 9 + out[0 * num_cols + col] = v[1]; + out[1 * num_cols + col] = v[14]; + out[2 * num_cols + col] = v[3]; + out[3 * num_cols + col] = v[12]; + out[4 * num_cols + col] = v[5]; + out[5 * num_cols + col] = v[10]; + out[6 * num_cols + col] = v[7]; + out[7 * num_cols + col] = v[8]; + out[8 * num_cols + col] = v[9]; + out[9 * num_cols + col] = v[6]; + out[10 * num_cols + col] = v[11]; + out[11 * num_cols + col] = v[4]; + out[12 * num_cols + col] = v[13]; + out[13 * num_cols + col] = v[2]; + out[14 * num_cols + col] = v[15]; + out[15 * num_cols + col] = v[0]; + } +} + +static void col_txfm_16x16_rounding(__m128i *in, int shift) { + // Note: + // We split 16x16 rounding into 4 sections of 8x8 rounding, + // instead of 4 columns + col_txfm_8x8_rounding(&in[0], shift); + col_txfm_8x8_rounding(&in[16], shift); + col_txfm_8x8_rounding(&in[32], shift); + col_txfm_8x8_rounding(&in[48], shift); +} + +static void col_txfm_8x16_rounding(__m128i *in, int shift) { + col_txfm_8x8_rounding(&in[0], shift); + col_txfm_8x8_rounding(&in[16], shift); +} + +static void write_buffer_16x16(const __m128i *in, int32_t *output) { + const int size_8x8 = 16 * 4; + write_buffer_8x8(&in[0], output); + output += size_8x8; + write_buffer_8x8(&in[16], output); + output += size_8x8; + write_buffer_8x8(&in[32], output); + output += size_8x8; + write_buffer_8x8(&in[48], output); +} + +void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *coeff, + int stride, TX_TYPE tx_type, int bd) { + __m128i in[64], out[64]; + const int8_t *shift = fwd_txfm_shift_ls[TX_16X16]; + const int txw_idx = get_txw_idx(TX_16X16); + const int txh_idx = get_txh_idx(TX_16X16); + const int col_num = 4; + switch (tx_type) { + case DCT_DCT: + load_buffer_16x16(input, in, stride, 0, 0, shift[0]); + fdct16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num); + col_txfm_16x16_rounding(out, -shift[1]); + transpose_16x16(out, in); + fdct16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num); + transpose_16x16(out, in); + write_buffer_16x16(in, coeff); + break; + case ADST_DCT: + load_buffer_16x16(input, in, stride, 0, 0, shift[0]); + fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num); + col_txfm_16x16_rounding(out, -shift[1]); + transpose_16x16(out, in); + fdct16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num); + transpose_16x16(out, in); + write_buffer_16x16(in, coeff); + break; + case DCT_ADST: + load_buffer_16x16(input, in, stride, 0, 0, shift[0]); + fdct16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num); + col_txfm_16x16_rounding(out, -shift[1]); + transpose_16x16(out, in); + fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num); + transpose_16x16(out, in); + write_buffer_16x16(in, coeff); + break; + case ADST_ADST: + load_buffer_16x16(input, in, stride, 0, 0, shift[0]); + fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num); + col_txfm_16x16_rounding(out, -shift[1]); + transpose_16x16(out, in); + fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num); + transpose_16x16(out, in); + write_buffer_16x16(in, coeff); + break; + case FLIPADST_DCT: + load_buffer_16x16(input, in, stride, 1, 0, shift[0]); + fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num); + col_txfm_16x16_rounding(out, -shift[1]); + transpose_16x16(out, in); + fdct16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num); + transpose_16x16(out, in); + write_buffer_16x16(in, coeff); + break; + case DCT_FLIPADST: + load_buffer_16x16(input, in, stride, 0, 1, shift[0]); + fdct16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num); + col_txfm_16x16_rounding(out, -shift[1]); + transpose_16x16(out, in); + fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num); + transpose_16x16(out, in); + write_buffer_16x16(in, coeff); + break; + case FLIPADST_FLIPADST: + load_buffer_16x16(input, in, stride, 1, 1, shift[0]); + fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num); + col_txfm_16x16_rounding(out, -shift[1]); + transpose_16x16(out, in); + fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num); + transpose_16x16(out, in); + write_buffer_16x16(in, coeff); + break; + case ADST_FLIPADST: + load_buffer_16x16(input, in, stride, 0, 1, shift[0]); + fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num); + col_txfm_16x16_rounding(out, -shift[1]); + transpose_16x16(out, in); + fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num); + transpose_16x16(out, in); + write_buffer_16x16(in, coeff); + break; + case FLIPADST_ADST: + load_buffer_16x16(input, in, stride, 1, 0, shift[0]); + fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num); + col_txfm_16x16_rounding(out, -shift[1]); + transpose_16x16(out, in); + fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num); + transpose_16x16(out, in); + write_buffer_16x16(in, coeff); + break; + default: assert(0); + } + (void)bd; +} + +static INLINE void flip_buf_sse4_1(__m128i *in, __m128i *out, int size) { + for (int i = 0; i < size; i += 2) in[30 - i] = out[i]; + for (int i = 1; i < size; i += 2) in[size - i] = out[i]; +} + +static const fwd_transform_1d_sse4_1 col_highbd_txfm8x8_arr[TX_TYPES] = { + fdct8x8_sse4_1, // DCT_DCT + fadst8x8_sse4_1, // ADST_DCT + fdct8x8_sse4_1, // DCT_ADST + fadst8x8_sse4_1, // ADST_ADST + fadst8x8_sse4_1, // FLIPADST_DCT + fdct8x8_sse4_1, // DCT_FLIPADST + fadst8x8_sse4_1, // FLIPADST_FLIPADST + fadst8x8_sse4_1, // ADST_FLIPADST + fadst8x8_sse4_1, // FLIPADST_ADST + NULL, // IDTX + NULL, // V_DCT + NULL, // H_DCT + NULL, // V_ADST + NULL, // H_ADST + NULL, // V_FLIPADST + NULL // H_FLIPADST +}; + +static const fwd_transform_1d_sse4_1 row_highbd_txfm8x16_arr[TX_TYPES] = { + fdct16x16_sse4_1, // DCT_DCT + fdct16x16_sse4_1, // ADST_DCT + fadst16x16_sse4_1, // DCT_ADST + fadst16x16_sse4_1, // ADST_ADST + fdct16x16_sse4_1, // FLIPADST_DCT + fadst16x16_sse4_1, // DCT_FLIPADST + fadst16x16_sse4_1, // FLIPADST_FLIPADST + fadst16x16_sse4_1, // ADST_FLIPADST + fadst16x16_sse4_1, // FLIPADST_ADST + NULL, // IDTX + NULL, // V_DCT + NULL, // H_DCT + NULL, // V_ADST + NULL, // H_ADST + NULL, // V_FLIPADST + NULL // H_FLIPADST +}; + +static const fwd_transform_1d_sse4_1 col_highbd_txfm8x16_arr[TX_TYPES] = { + fdct16x16_sse4_1, // DCT_DCT + fadst16x16_sse4_1, // ADST_DCT + fdct16x16_sse4_1, // DCT_ADST + fadst16x16_sse4_1, // ADST_ADST + fadst16x16_sse4_1, // FLIPADST_DCT + fdct16x16_sse4_1, // DCT_FLIPADST + fadst16x16_sse4_1, // FLIPADST_FLIPADST + fadst16x16_sse4_1, // ADST_FLIPADST + fadst16x16_sse4_1, // FLIPADST_ADST + NULL, // IDTX + NULL, // V_DCT + NULL, // H_DCT + NULL, // V_ADST + NULL, // H_ADST + NULL, // V_FLIPADST + NULL // H_FLIPADST +}; +static const fwd_transform_1d_sse4_1 row_highbd_txfm8x8_arr[TX_TYPES] = { + fdct8x8_sse4_1, // DCT_DCT + fdct8x8_sse4_1, // ADST_DCT + fadst8x8_sse4_1, // DCT_ADST + fadst8x8_sse4_1, // ADST_ADST + fdct8x8_sse4_1, // FLIPADST_DCT + fadst8x8_sse4_1, // DCT_FLIPADST + fadst8x8_sse4_1, // FLIPADST_FLIPADST + fadst8x8_sse4_1, // ADST_FLIPADST + fadst8x8_sse4_1, // FLIPADST_ADST + NULL, // IDTX + NULL, // V_DCT + NULL, // H_DCT + NULL, // V_ADST + NULL, // H_ADST + NULL, // V_FLIPADST + NULL // H_FLIPADST +}; + +void av1_fwd_txfm2d_16x8_sse4_1(const int16_t *input, int32_t *coeff, + int stride, TX_TYPE tx_type, int bd) { + __m128i in[32], out[32]; + const int8_t *shift = fwd_txfm_shift_ls[TX_16X8]; + const int txw_idx = get_txw_idx(TX_16X8); + const int txh_idx = get_txh_idx(TX_16X8); + const fwd_transform_1d_sse4_1 col_txfm = col_highbd_txfm8x8_arr[tx_type]; + const fwd_transform_1d_sse4_1 row_txfm = row_highbd_txfm8x16_arr[tx_type]; + int bit = fwd_cos_bit_col[txw_idx][txh_idx]; + int ud_flip, lr_flip; + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + + for (int i = 0; i < 2; i++) { + load_buffer_8x8(input + i * 8, in, stride, ud_flip, 0, shift[0]); + col_txfm(in, in, bit, 0); + col_txfm_8x8_rounding(in, -shift[1]); + transpose_8x8(in, out + i * 16); + } + + if (lr_flip) { + flip_buf_sse4_1(in, out, 32); + row_txfm(in, out, bit, 2); + } else { + row_txfm(out, out, bit, 2); + } + + for (int i = 0; i < 2; i++) { + transpose_8x8(out + i * 16, in); + av1_round_shift_rect_array_32_sse4_1(in, in, 16, -shift[2], NewSqrt2); + write_buffer_16x8(in, coeff + i * 8, 16); + } + + (void)bd; +} + +void av1_fwd_txfm2d_8x16_sse4_1(const int16_t *input, int32_t *coeff, + int stride, TX_TYPE tx_type, int bd) { + __m128i in[32], out[32]; + const int8_t *shift = fwd_txfm_shift_ls[TX_8X16]; + const int txw_idx = get_txw_idx(TX_8X16); + const int txh_idx = get_txh_idx(TX_8X16); + const fwd_transform_1d_sse4_1 col_txfm = col_highbd_txfm8x16_arr[tx_type]; + const fwd_transform_1d_sse4_1 row_txfm = row_highbd_txfm8x8_arr[tx_type]; + int bit = fwd_cos_bit_col[txw_idx][txh_idx]; + int ud_flip, lr_flip; + get_flip_cfg(tx_type, &ud_flip, &lr_flip); + + load_buffer_8x16(input, in, stride, ud_flip, lr_flip, shift[0]); + col_txfm(in, in, bit, 2); + col_txfm_8x16_rounding(in, -shift[1]); + transpose_8x8(in, out); + transpose_8x8(in + 16, out + 16); + + for (int i = 0; i < 2; i++) { + row_txfm(out + i * 16, out, bit, 0); + transpose_8x8(out, in); + av1_round_shift_rect_array_32_sse4_1(in, in, 16, -shift[2], NewSqrt2); + write_buffer_8x8(in, coeff + i * 64); + } + + (void)bd; +} diff --git a/third_party/aom/av1/encoder/x86/pickrst_avx2.c b/third_party/aom/av1/encoder/x86/pickrst_avx2.c new file mode 100644 index 0000000000..06aaaa7eee --- /dev/null +++ b/third_party/aom/av1/encoder/x86/pickrst_avx2.c @@ -0,0 +1,403 @@ +/* + * Copyright (c) 2018, 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 <immintrin.h> // AVX2 +#include "aom_dsp/x86/synonyms.h" +#include "aom_dsp/x86/synonyms_avx2.h" +#include "aom_dsp/x86/transpose_sse2.h" + +#include "config/av1_rtcd.h" +#include "av1/common/restoration.h" +#include "av1/encoder/pickrst.h" + +static INLINE void acc_stat_avx2(int32_t *dst, const uint8_t *src, + const __m128i *shuffle, const __m256i *kl) { + const __m128i s = _mm_shuffle_epi8(xx_loadu_128(src), *shuffle); + const __m256i d0 = _mm256_madd_epi16(*kl, _mm256_cvtepu8_epi16(s)); + const __m256i dst0 = yy_loadu_256(dst); + const __m256i r0 = _mm256_add_epi32(dst0, d0); + yy_storeu_256(dst, r0); +} + +static INLINE void acc_stat_win7_one_line_avx2( + const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, + int dgd_stride, const __m128i *shuffle, int32_t *sumX, + int32_t sumY[WIENER_WIN][WIENER_WIN], int32_t M_int[WIENER_WIN][WIENER_WIN], + int32_t H_int[WIENER_WIN2][WIENER_WIN * 8]) { + int j, k, l; + const int wiener_win = WIENER_WIN; + for (j = h_start; j < h_end; j += 2) { + const uint8_t X1 = src[j]; + const uint8_t X2 = src[j + 1]; + *sumX += X1 + X2; + const uint8_t *dgd_ij = dgd + j; + for (k = 0; k < wiener_win; k++) { + const uint8_t *dgd_ijk = dgd_ij + k * dgd_stride; + for (l = 0; l < wiener_win; l++) { + int32_t *H_ = &H_int[(l * wiener_win + k)][0]; + const uint8_t D1 = dgd_ijk[l]; + const uint8_t D2 = dgd_ijk[l + 1]; + sumY[k][l] += D1 + D2; + M_int[k][l] += D1 * X1 + D2 * X2; + + const __m256i kl = + _mm256_cvtepu8_epi16(_mm_set1_epi16(*((uint16_t *)(dgd_ijk + l)))); + acc_stat_avx2(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle, &kl); + acc_stat_avx2(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle, &kl); + acc_stat_avx2(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle, &kl); + acc_stat_avx2(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle, &kl); + acc_stat_avx2(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle, &kl); + acc_stat_avx2(H_ + 5 * 8, dgd_ij + 5 * dgd_stride, shuffle, &kl); + acc_stat_avx2(H_ + 6 * 8, dgd_ij + 6 * dgd_stride, shuffle, &kl); + } + } + } +} + +static INLINE void compute_stats_win7_opt_avx2( + const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, int v_start, + int v_end, int dgd_stride, int src_stride, double *M, double *H) { + int i, j, k, l, m, n; + const int wiener_win = WIENER_WIN; + const int pixel_count = (h_end - h_start) * (v_end - v_start); + const int wiener_win2 = wiener_win * wiener_win; + const int wiener_halfwin = (wiener_win >> 1); + const double avg = + find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride); + + int32_t M_int32[WIENER_WIN][WIENER_WIN] = { { 0 } }; + int64_t M_int64[WIENER_WIN][WIENER_WIN] = { { 0 } }; + int32_t H_int32[WIENER_WIN2][WIENER_WIN * 8] = { { 0 } }; + int64_t H_int64[WIENER_WIN2][WIENER_WIN * 8] = { { 0 } }; + int32_t sumY[WIENER_WIN][WIENER_WIN] = { { 0 } }; + int32_t sumX = 0; + const uint8_t *dgd_win = dgd - wiener_halfwin * dgd_stride - wiener_halfwin; + + const __m128i shuffle = xx_loadu_128(g_shuffle_stats_data); + for (j = v_start; j < v_end; j += 64) { + const int vert_end = AOMMIN(64, v_end - j) + j; + for (i = j; i < vert_end; i++) { + acc_stat_win7_one_line_avx2( + dgd_win + i * dgd_stride, src + i * src_stride, h_start, h_end, + dgd_stride, &shuffle, &sumX, sumY, M_int32, H_int32); + } + for (k = 0; k < wiener_win; ++k) { + for (l = 0; l < wiener_win; ++l) { + M_int64[k][l] += M_int32[k][l]; + M_int32[k][l] = 0; + } + } + for (k = 0; k < WIENER_WIN2; ++k) { + for (l = 0; l < WIENER_WIN * 8; ++l) { + H_int64[k][l] += H_int32[k][l]; + H_int32[k][l] = 0; + } + } + } + + const double avg_square_sum = avg * avg * pixel_count; + for (k = 0; k < wiener_win; k++) { + for (l = 0; l < wiener_win; l++) { + const int32_t idx0 = l * wiener_win + k; + M[idx0] = M_int64[k][l] + avg_square_sum - avg * (sumX + sumY[k][l]); + double *H_ = H + idx0 * wiener_win2; + int64_t *H_int_ = &H_int64[idx0][0]; + for (m = 0; m < wiener_win; m++) { + for (n = 0; n < wiener_win; n++) { + H_[m * wiener_win + n] = H_int_[n * 8 + m] + avg_square_sum - + avg * (sumY[k][l] + sumY[n][m]); + } + } + } + } +} + +static INLINE void acc_stat_win5_one_line_avx2( + const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, + int dgd_stride, const __m128i *shuffle, int32_t *sumX, + int32_t sumY[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA], + int32_t M_int[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA], + int32_t H_int[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8]) { + int j, k, l; + const int wiener_win = WIENER_WIN_CHROMA; + for (j = h_start; j < h_end; j += 2) { + const uint8_t X1 = src[j]; + const uint8_t X2 = src[j + 1]; + *sumX += X1 + X2; + const uint8_t *dgd_ij = dgd + j; + for (k = 0; k < wiener_win; k++) { + const uint8_t *dgd_ijk = dgd_ij + k * dgd_stride; + for (l = 0; l < wiener_win; l++) { + int32_t *H_ = &H_int[(l * wiener_win + k)][0]; + const uint8_t D1 = dgd_ijk[l]; + const uint8_t D2 = dgd_ijk[l + 1]; + sumY[k][l] += D1 + D2; + M_int[k][l] += D1 * X1 + D2 * X2; + + const __m256i kl = + _mm256_cvtepu8_epi16(_mm_set1_epi16(*((uint16_t *)(dgd_ijk + l)))); + acc_stat_avx2(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle, &kl); + acc_stat_avx2(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle, &kl); + acc_stat_avx2(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle, &kl); + acc_stat_avx2(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle, &kl); + acc_stat_avx2(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle, &kl); + } + } + } +} + +static INLINE void compute_stats_win5_opt_avx2( + const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, int v_start, + int v_end, int dgd_stride, int src_stride, double *M, double *H) { + int i, j, k, l, m, n; + const int wiener_win = WIENER_WIN_CHROMA; + const int pixel_count = (h_end - h_start) * (v_end - v_start); + const int wiener_win2 = wiener_win * wiener_win; + const int wiener_halfwin = (wiener_win >> 1); + const double avg = + find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride); + + int32_t M_int32[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } }; + int64_t M_int64[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } }; + int32_t H_int32[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8] = { { 0 } }; + int64_t H_int64[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8] = { { 0 } }; + int32_t sumY[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } }; + int32_t sumX = 0; + const uint8_t *dgd_win = dgd - wiener_halfwin * dgd_stride - wiener_halfwin; + + const __m128i shuffle = xx_loadu_128(g_shuffle_stats_data); + for (j = v_start; j < v_end; j += 64) { + const int vert_end = AOMMIN(64, v_end - j) + j; + for (i = j; i < vert_end; i++) { + acc_stat_win5_one_line_avx2( + dgd_win + i * dgd_stride, src + i * src_stride, h_start, h_end, + dgd_stride, &shuffle, &sumX, sumY, M_int32, H_int32); + } + for (k = 0; k < wiener_win; ++k) { + for (l = 0; l < wiener_win; ++l) { + M_int64[k][l] += M_int32[k][l]; + M_int32[k][l] = 0; + } + } + for (k = 0; k < WIENER_WIN2_CHROMA; ++k) { + for (l = 0; l < WIENER_WIN_CHROMA * 8; ++l) { + H_int64[k][l] += H_int32[k][l]; + H_int32[k][l] = 0; + } + } + } + + const double avg_square_sum = avg * avg * pixel_count; + for (k = 0; k < wiener_win; k++) { + for (l = 0; l < wiener_win; l++) { + const int32_t idx0 = l * wiener_win + k; + M[idx0] = M_int64[k][l] + avg_square_sum - avg * (sumX + sumY[k][l]); + double *H_ = H + idx0 * wiener_win2; + int64_t *H_int_ = &H_int64[idx0][0]; + for (m = 0; m < wiener_win; m++) { + for (n = 0; n < wiener_win; n++) { + H_[m * wiener_win + n] = H_int_[n * 8 + m] + avg_square_sum - + avg * (sumY[k][l] + sumY[n][m]); + } + } + } + } +} + +void av1_compute_stats_avx2(int wiener_win, const uint8_t *dgd, + const uint8_t *src, int h_start, int h_end, + int v_start, int v_end, int dgd_stride, + int src_stride, double *M, double *H) { + if (wiener_win == WIENER_WIN) { + compute_stats_win7_opt_avx2(dgd, src, h_start, h_end, v_start, v_end, + dgd_stride, src_stride, M, H); + } else if (wiener_win == WIENER_WIN_CHROMA) { + compute_stats_win5_opt_avx2(dgd, src, h_start, h_end, v_start, v_end, + dgd_stride, src_stride, M, H); + } else { + av1_compute_stats_c(wiener_win, dgd, src, h_start, h_end, v_start, v_end, + dgd_stride, src_stride, M, H); + } +} + +static INLINE __m256i pair_set_epi16(uint16_t a, uint16_t b) { + return _mm256_set1_epi32( + (int32_t)(((uint16_t)(a)) | (((uint32_t)(b)) << 16))); +} + +int64_t av1_lowbd_pixel_proj_error_avx2( + const uint8_t *src8, int width, int height, int src_stride, + const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, + int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params) { + int i, j, k; + const int32_t shift = SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS; + const __m256i rounding = _mm256_set1_epi32(1 << (shift - 1)); + __m256i sum64 = _mm256_setzero_si256(); + const uint8_t *src = src8; + const uint8_t *dat = dat8; + int64_t err = 0; + if (params->r[0] > 0 && params->r[1] > 0) { + __m256i xq_coeff = pair_set_epi16(xq[0], xq[1]); + for (i = 0; i < height; ++i) { + __m256i sum32 = _mm256_setzero_si256(); + for (j = 0; j <= width - 16; j += 16) { + const __m256i d0 = _mm256_cvtepu8_epi16(xx_loadu_128(dat + j)); + const __m256i s0 = _mm256_cvtepu8_epi16(xx_loadu_128(src + j)); + const __m256i flt0_16b = _mm256_permute4x64_epi64( + _mm256_packs_epi32(yy_loadu_256(flt0 + j), + yy_loadu_256(flt0 + j + 8)), + 0xd8); + const __m256i flt1_16b = _mm256_permute4x64_epi64( + _mm256_packs_epi32(yy_loadu_256(flt1 + j), + yy_loadu_256(flt1 + j + 8)), + 0xd8); + const __m256i u0 = _mm256_slli_epi16(d0, SGRPROJ_RST_BITS); + const __m256i flt0_0_sub_u = _mm256_sub_epi16(flt0_16b, u0); + const __m256i flt1_0_sub_u = _mm256_sub_epi16(flt1_16b, u0); + const __m256i v0 = _mm256_madd_epi16( + xq_coeff, _mm256_unpacklo_epi16(flt0_0_sub_u, flt1_0_sub_u)); + const __m256i v1 = _mm256_madd_epi16( + xq_coeff, _mm256_unpackhi_epi16(flt0_0_sub_u, flt1_0_sub_u)); + const __m256i vr0 = + _mm256_srai_epi32(_mm256_add_epi32(v0, rounding), shift); + const __m256i vr1 = + _mm256_srai_epi32(_mm256_add_epi32(v1, rounding), shift); + const __m256i e0 = _mm256_sub_epi16( + _mm256_add_epi16(_mm256_packs_epi32(vr0, vr1), d0), s0); + const __m256i err0 = _mm256_madd_epi16(e0, e0); + sum32 = _mm256_add_epi32(sum32, err0); + } + for (k = j; k < width; ++k) { + const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS); + int32_t v = xq[0] * (flt0[k] - u) + xq[1] * (flt1[k] - u); + const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k]; + err += e * e; + } + dat += dat_stride; + src += src_stride; + flt0 += flt0_stride; + flt1 += flt1_stride; + const __m256i sum64_0 = + _mm256_cvtepi32_epi64(_mm256_castsi256_si128(sum32)); + const __m256i sum64_1 = + _mm256_cvtepi32_epi64(_mm256_extracti128_si256(sum32, 1)); + sum64 = _mm256_add_epi64(sum64, sum64_0); + sum64 = _mm256_add_epi64(sum64, sum64_1); + } + } else if (params->r[0] > 0) { + __m256i xq_coeff = + pair_set_epi16(xq[0], (-xq[0] * (1 << SGRPROJ_RST_BITS))); + for (i = 0; i < height; ++i) { + __m256i sum32 = _mm256_setzero_si256(); + for (j = 0; j <= width - 16; j += 16) { + const __m256i d0 = _mm256_cvtepu8_epi16(xx_loadu_128(dat + j)); + const __m256i s0 = _mm256_cvtepu8_epi16(xx_loadu_128(src + j)); + const __m256i flt0_16b = _mm256_permute4x64_epi64( + _mm256_packs_epi32(yy_loadu_256(flt0 + j), + yy_loadu_256(flt0 + j + 8)), + 0xd8); + const __m256i v0 = + _mm256_madd_epi16(xq_coeff, _mm256_unpacklo_epi16(flt0_16b, d0)); + const __m256i v1 = + _mm256_madd_epi16(xq_coeff, _mm256_unpackhi_epi16(flt0_16b, d0)); + const __m256i vr0 = + _mm256_srai_epi32(_mm256_add_epi32(v0, rounding), shift); + const __m256i vr1 = + _mm256_srai_epi32(_mm256_add_epi32(v1, rounding), shift); + const __m256i e0 = _mm256_sub_epi16( + _mm256_add_epi16(_mm256_packs_epi32(vr0, vr1), d0), s0); + const __m256i err0 = _mm256_madd_epi16(e0, e0); + sum32 = _mm256_add_epi32(sum32, err0); + } + for (k = j; k < width; ++k) { + const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS); + int32_t v = xq[0] * (flt0[k] - u); + const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k]; + err += e * e; + } + dat += dat_stride; + src += src_stride; + flt0 += flt0_stride; + const __m256i sum64_0 = + _mm256_cvtepi32_epi64(_mm256_castsi256_si128(sum32)); + const __m256i sum64_1 = + _mm256_cvtepi32_epi64(_mm256_extracti128_si256(sum32, 1)); + sum64 = _mm256_add_epi64(sum64, sum64_0); + sum64 = _mm256_add_epi64(sum64, sum64_1); + } + } else if (params->r[1] > 0) { + __m256i xq_coeff = pair_set_epi16(xq[1], -(xq[1] << SGRPROJ_RST_BITS)); + for (i = 0; i < height; ++i) { + __m256i sum32 = _mm256_setzero_si256(); + for (j = 0; j <= width - 16; j += 16) { + const __m256i d0 = _mm256_cvtepu8_epi16(xx_loadu_128(dat + j)); + const __m256i s0 = _mm256_cvtepu8_epi16(xx_loadu_128(src + j)); + const __m256i flt1_16b = _mm256_permute4x64_epi64( + _mm256_packs_epi32(yy_loadu_256(flt1 + j), + yy_loadu_256(flt1 + j + 8)), + 0xd8); + const __m256i v0 = + _mm256_madd_epi16(xq_coeff, _mm256_unpacklo_epi16(flt1_16b, d0)); + const __m256i v1 = + _mm256_madd_epi16(xq_coeff, _mm256_unpackhi_epi16(flt1_16b, d0)); + const __m256i vr0 = + _mm256_srai_epi32(_mm256_add_epi32(v0, rounding), shift); + const __m256i vr1 = + _mm256_srai_epi32(_mm256_add_epi32(v1, rounding), shift); + const __m256i e0 = _mm256_sub_epi16( + _mm256_add_epi16(_mm256_packs_epi32(vr0, vr1), d0), s0); + const __m256i err0 = _mm256_madd_epi16(e0, e0); + sum32 = _mm256_add_epi32(sum32, err0); + } + for (k = j; k < width; ++k) { + const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS); + int32_t v = xq[1] * (flt1[k] - u); + const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k]; + err += e * e; + } + dat += dat_stride; + src += src_stride; + flt1 += flt1_stride; + const __m256i sum64_0 = + _mm256_cvtepi32_epi64(_mm256_castsi256_si128(sum32)); + const __m256i sum64_1 = + _mm256_cvtepi32_epi64(_mm256_extracti128_si256(sum32, 1)); + sum64 = _mm256_add_epi64(sum64, sum64_0); + sum64 = _mm256_add_epi64(sum64, sum64_1); + } + } else { + __m256i sum32 = _mm256_setzero_si256(); + for (i = 0; i < height; ++i) { + for (j = 0; j <= width - 16; j += 16) { + const __m256i d0 = _mm256_cvtepu8_epi16(xx_loadu_128(dat + j)); + const __m256i s0 = _mm256_cvtepu8_epi16(xx_loadu_128(src + j)); + const __m256i diff0 = _mm256_sub_epi16(d0, s0); + const __m256i err0 = _mm256_madd_epi16(diff0, diff0); + sum32 = _mm256_add_epi32(sum32, err0); + } + for (k = j; k < width; ++k) { + const int32_t e = (int32_t)(dat[k]) - src[k]; + err += e * e; + } + dat += dat_stride; + src += src_stride; + } + const __m256i sum64_0 = + _mm256_cvtepi32_epi64(_mm256_castsi256_si128(sum32)); + const __m256i sum64_1 = + _mm256_cvtepi32_epi64(_mm256_extracti128_si256(sum32, 1)); + sum64 = _mm256_add_epi64(sum64_0, sum64_1); + } + int64_t sum[4]; + yy_storeu_256(sum, sum64); + err += sum[0] + sum[1] + sum[2] + sum[3]; + return err; +} diff --git a/third_party/aom/av1/encoder/x86/pickrst_sse4.c b/third_party/aom/av1/encoder/x86/pickrst_sse4.c new file mode 100644 index 0000000000..04e4d1afc4 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/pickrst_sse4.c @@ -0,0 +1,389 @@ +/* + * Copyright (c) 2018, 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 <emmintrin.h> +#include "aom_dsp/x86/synonyms.h" + +#include "config/av1_rtcd.h" +#include "av1/common/restoration.h" +#include "av1/encoder/pickrst.h" + +static INLINE void acc_stat_sse41(int32_t *dst, const uint8_t *src, + const __m128i *shuffle, const __m128i *kl) { + const __m128i s = _mm_shuffle_epi8(xx_loadu_128(src), *shuffle); + const __m128i d0 = _mm_madd_epi16(*kl, _mm_cvtepu8_epi16(s)); + const __m128i d1 = + _mm_madd_epi16(*kl, _mm_cvtepu8_epi16(_mm_srli_si128(s, 8))); + const __m128i dst0 = xx_loadu_128(dst); + const __m128i dst1 = xx_loadu_128(dst + 4); + const __m128i r0 = _mm_add_epi32(dst0, d0); + const __m128i r1 = _mm_add_epi32(dst1, d1); + xx_storeu_128(dst, r0); + xx_storeu_128(dst + 4, r1); +} + +static INLINE void acc_stat_win7_one_line_sse4_1( + const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, + int dgd_stride, const __m128i *shuffle, int32_t *sumX, + int32_t sumY[WIENER_WIN][WIENER_WIN], int32_t M_int[WIENER_WIN][WIENER_WIN], + int32_t H_int[WIENER_WIN2][WIENER_WIN * 8]) { + const int wiener_win = 7; + int j, k, l; + for (j = h_start; j < h_end; j += 2) { + const uint8_t *dgd_ij = dgd + j; + const uint8_t X1 = src[j]; + const uint8_t X2 = src[j + 1]; + *sumX += X1 + X2; + for (k = 0; k < wiener_win; k++) { + const uint8_t *dgd_ijk = dgd_ij + k * dgd_stride; + for (l = 0; l < wiener_win; l++) { + int32_t *H_ = &H_int[(l * wiener_win + k)][0]; + const uint8_t D1 = dgd_ijk[l]; + const uint8_t D2 = dgd_ijk[l + 1]; + sumY[k][l] += D1 + D2; + M_int[k][l] += D1 * X1 + D2 * X2; + + const __m128i kl = + _mm_cvtepu8_epi16(_mm_set1_epi16(*((uint16_t *)(dgd_ijk + l)))); + acc_stat_sse41(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle, &kl); + acc_stat_sse41(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle, &kl); + acc_stat_sse41(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle, &kl); + acc_stat_sse41(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle, &kl); + acc_stat_sse41(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle, &kl); + acc_stat_sse41(H_ + 5 * 8, dgd_ij + 5 * dgd_stride, shuffle, &kl); + acc_stat_sse41(H_ + 6 * 8, dgd_ij + 6 * dgd_stride, shuffle, &kl); + } + } + } +} + +static INLINE void compute_stats_win7_opt_sse4_1( + const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, int v_start, + int v_end, int dgd_stride, int src_stride, double *M, double *H) { + int i, j, k, l, m, n; + const int wiener_win = WIENER_WIN; + const int pixel_count = (h_end - h_start) * (v_end - v_start); + const int wiener_win2 = wiener_win * wiener_win; + const int wiener_halfwin = (wiener_win >> 1); + const double avg = + find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride); + + int32_t M_int32[WIENER_WIN][WIENER_WIN] = { { 0 } }; + int64_t M_int64[WIENER_WIN][WIENER_WIN] = { { 0 } }; + int32_t H_int32[WIENER_WIN2][WIENER_WIN * 8] = { { 0 } }; + int64_t H_int64[WIENER_WIN2][WIENER_WIN * 8] = { { 0 } }; + int32_t sumY[WIENER_WIN][WIENER_WIN] = { { 0 } }; + int32_t sumX = 0; + const uint8_t *dgd_win = dgd - wiener_halfwin * dgd_stride - wiener_halfwin; + + const __m128i shuffle = xx_loadu_128(g_shuffle_stats_data); + for (j = v_start; j < v_end; j += 64) { + const int vert_end = AOMMIN(64, v_end - j) + j; + for (i = j; i < vert_end; i++) { + acc_stat_win7_one_line_sse4_1( + dgd_win + i * dgd_stride, src + i * src_stride, h_start, h_end, + dgd_stride, &shuffle, &sumX, sumY, M_int32, H_int32); + } + for (k = 0; k < wiener_win; ++k) { + for (l = 0; l < wiener_win; ++l) { + M_int64[k][l] += M_int32[k][l]; + M_int32[k][l] = 0; + } + } + for (k = 0; k < WIENER_WIN2; ++k) { + for (l = 0; l < WIENER_WIN * 8; ++l) { + H_int64[k][l] += H_int32[k][l]; + H_int32[k][l] = 0; + } + } + } + + const double avg_square_sum = avg * avg * pixel_count; + for (k = 0; k < wiener_win; k++) { + for (l = 0; l < wiener_win; l++) { + const int32_t idx0 = l * wiener_win + k; + M[idx0] = M_int64[k][l] + avg_square_sum - avg * (sumX + sumY[k][l]); + double *H_ = H + idx0 * wiener_win2; + int64_t *H_int_ = &H_int64[idx0][0]; + for (m = 0; m < wiener_win; m++) { + for (n = 0; n < wiener_win; n++) { + H_[m * wiener_win + n] = H_int_[n * 8 + m] + avg_square_sum - + avg * (sumY[k][l] + sumY[n][m]); + } + } + } + } +} + +static INLINE void acc_stat_win5_one_line_sse4_1( + const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, + int dgd_stride, const __m128i *shuffle, int32_t *sumX, + int32_t sumY[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA], + int32_t M_int[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA], + int32_t H_int[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8]) { + const int wiener_win = WIENER_WIN_CHROMA; + int j, k, l; + for (j = h_start; j < h_end; j += 2) { + const uint8_t *dgd_ij = dgd + j; + const uint8_t X1 = src[j]; + const uint8_t X2 = src[j + 1]; + *sumX += X1 + X2; + for (k = 0; k < wiener_win; k++) { + const uint8_t *dgd_ijk = dgd_ij + k * dgd_stride; + for (l = 0; l < wiener_win; l++) { + int32_t *H_ = &H_int[(l * wiener_win + k)][0]; + const uint8_t D1 = dgd_ijk[l]; + const uint8_t D2 = dgd_ijk[l + 1]; + sumY[k][l] += D1 + D2; + M_int[k][l] += D1 * X1 + D2 * X2; + + const __m128i kl = + _mm_cvtepu8_epi16(_mm_set1_epi16(*((uint16_t *)(dgd_ijk + l)))); + acc_stat_sse41(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle, &kl); + acc_stat_sse41(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle, &kl); + acc_stat_sse41(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle, &kl); + acc_stat_sse41(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle, &kl); + acc_stat_sse41(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle, &kl); + } + } + } +} + +static INLINE void compute_stats_win5_opt_sse4_1( + const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, int v_start, + int v_end, int dgd_stride, int src_stride, double *M, double *H) { + int i, j, k, l, m, n; + const int wiener_win = WIENER_WIN_CHROMA; + const int pixel_count = (h_end - h_start) * (v_end - v_start); + const int wiener_win2 = wiener_win * wiener_win; + const int wiener_halfwin = (wiener_win >> 1); + const double avg = + find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride); + + int32_t M_int32[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } }; + int64_t M_int64[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } }; + int32_t H_int32[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8] = { { 0 } }; + int64_t H_int64[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8] = { { 0 } }; + int32_t sumY[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } }; + int32_t sumX = 0; + const uint8_t *dgd_win = dgd - wiener_halfwin * dgd_stride - wiener_halfwin; + + const __m128i shuffle = xx_loadu_128(g_shuffle_stats_data); + for (j = v_start; j < v_end; j += 64) { + const int vert_end = AOMMIN(64, v_end - j) + j; + for (i = j; i < vert_end; i++) { + acc_stat_win5_one_line_sse4_1( + dgd_win + i * dgd_stride, src + i * src_stride, h_start, h_end, + dgd_stride, &shuffle, &sumX, sumY, M_int32, H_int32); + } + for (k = 0; k < wiener_win; ++k) { + for (l = 0; l < wiener_win; ++l) { + M_int64[k][l] += M_int32[k][l]; + M_int32[k][l] = 0; + } + } + for (k = 0; k < WIENER_WIN_CHROMA * WIENER_WIN_CHROMA; ++k) { + for (l = 0; l < WIENER_WIN_CHROMA * 8; ++l) { + H_int64[k][l] += H_int32[k][l]; + H_int32[k][l] = 0; + } + } + } + + const double avg_square_sum = avg * avg * pixel_count; + for (k = 0; k < wiener_win; k++) { + for (l = 0; l < wiener_win; l++) { + const int32_t idx0 = l * wiener_win + k; + M[idx0] = M_int64[k][l] + avg_square_sum - avg * (sumX + sumY[k][l]); + double *H_ = H + idx0 * wiener_win2; + int64_t *H_int_ = &H_int64[idx0][0]; + for (m = 0; m < wiener_win; m++) { + for (n = 0; n < wiener_win; n++) { + H_[m * wiener_win + n] = H_int_[n * 8 + m] + avg_square_sum - + avg * (sumY[k][l] + sumY[n][m]); + } + } + } + } +} +void av1_compute_stats_sse4_1(int wiener_win, const uint8_t *dgd, + const uint8_t *src, int h_start, int h_end, + int v_start, int v_end, int dgd_stride, + int src_stride, double *M, double *H) { + if (wiener_win == WIENER_WIN) { + compute_stats_win7_opt_sse4_1(dgd, src, h_start, h_end, v_start, v_end, + dgd_stride, src_stride, M, H); + } else if (wiener_win == WIENER_WIN_CHROMA) { + compute_stats_win5_opt_sse4_1(dgd, src, h_start, h_end, v_start, v_end, + dgd_stride, src_stride, M, H); + } else { + av1_compute_stats_c(wiener_win, dgd, src, h_start, h_end, v_start, v_end, + dgd_stride, src_stride, M, H); + } +} + +static INLINE __m128i pair_set_epi16(uint16_t a, uint16_t b) { + return _mm_set1_epi32((int32_t)(((uint16_t)(a)) | (((uint32_t)(b)) << 16))); +} + +int64_t av1_lowbd_pixel_proj_error_sse4_1( + const uint8_t *src8, int width, int height, int src_stride, + const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, + int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params) { + int i, j, k; + const int32_t shift = SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS; + const __m128i rounding = _mm_set1_epi32(1 << (shift - 1)); + __m128i sum64 = _mm_setzero_si128(); + const uint8_t *src = src8; + const uint8_t *dat = dat8; + int64_t err = 0; + if (params->r[0] > 0 && params->r[1] > 0) { + __m128i xq_coeff = pair_set_epi16(xq[0], xq[1]); + for (i = 0; i < height; ++i) { + __m128i sum32 = _mm_setzero_si128(); + for (j = 0; j < width - 8; j += 8) { + const __m128i d0 = _mm_cvtepu8_epi16(xx_loadl_64(dat + j)); + const __m128i s0 = _mm_cvtepu8_epi16(xx_loadl_64(src + j)); + const __m128i flt0_16b = + _mm_packs_epi32(xx_loadu_128(flt0 + j), xx_loadu_128(flt0 + j + 4)); + const __m128i flt1_16b = + _mm_packs_epi32(xx_loadu_128(flt1 + j), xx_loadu_128(flt1 + j + 4)); + const __m128i u0 = _mm_slli_epi16(d0, SGRPROJ_RST_BITS); + const __m128i flt0_0_sub_u = _mm_sub_epi16(flt0_16b, u0); + const __m128i flt1_0_sub_u = _mm_sub_epi16(flt1_16b, u0); + const __m128i v0 = _mm_madd_epi16( + xq_coeff, _mm_unpacklo_epi16(flt0_0_sub_u, flt1_0_sub_u)); + const __m128i v1 = _mm_madd_epi16( + xq_coeff, _mm_unpackhi_epi16(flt0_0_sub_u, flt1_0_sub_u)); + const __m128i vr0 = _mm_srai_epi32(_mm_add_epi32(v0, rounding), shift); + const __m128i vr1 = _mm_srai_epi32(_mm_add_epi32(v1, rounding), shift); + const __m128i e0 = + _mm_sub_epi16(_mm_add_epi16(_mm_packs_epi32(vr0, vr1), d0), s0); + const __m128i err0 = _mm_madd_epi16(e0, e0); + sum32 = _mm_add_epi32(sum32, err0); + } + for (k = j; k < width; ++k) { + const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS); + int32_t v = xq[0] * (flt0[k] - u) + xq[1] * (flt1[k] - u); + const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k]; + err += e * e; + } + dat += dat_stride; + src += src_stride; + flt0 += flt0_stride; + flt1 += flt1_stride; + const __m128i sum64_0 = _mm_cvtepi32_epi64(sum32); + const __m128i sum64_1 = _mm_cvtepi32_epi64(_mm_srli_si128(sum32, 8)); + sum64 = _mm_add_epi64(sum64, sum64_0); + sum64 = _mm_add_epi64(sum64, sum64_1); + } + } else if (params->r[0] > 0) { + __m128i xq_coeff = pair_set_epi16(xq[0], -(xq[0] << SGRPROJ_RST_BITS)); + for (i = 0; i < height; ++i) { + __m128i sum32 = _mm_setzero_si128(); + for (j = 0; j < width - 8; j += 8) { + const __m128i d0 = _mm_cvtepu8_epi16(xx_loadl_64(dat + j)); + const __m128i s0 = _mm_cvtepu8_epi16(xx_loadl_64(src + j)); + const __m128i flt0_16b = + _mm_packs_epi32(xx_loadu_128(flt0 + j), xx_loadu_128(flt0 + j + 4)); + const __m128i v0 = + _mm_madd_epi16(xq_coeff, _mm_unpacklo_epi16(flt0_16b, d0)); + const __m128i v1 = + _mm_madd_epi16(xq_coeff, _mm_unpackhi_epi16(flt0_16b, d0)); + const __m128i vr0 = _mm_srai_epi32(_mm_add_epi32(v0, rounding), shift); + const __m128i vr1 = _mm_srai_epi32(_mm_add_epi32(v1, rounding), shift); + const __m128i e0 = + _mm_sub_epi16(_mm_add_epi16(_mm_packs_epi32(vr0, vr1), d0), s0); + const __m128i err0 = _mm_madd_epi16(e0, e0); + sum32 = _mm_add_epi32(sum32, err0); + } + for (k = j; k < width; ++k) { + const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS); + int32_t v = xq[0] * (flt0[k] - u); + const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k]; + err += e * e; + } + dat += dat_stride; + src += src_stride; + flt0 += flt0_stride; + const __m128i sum64_0 = _mm_cvtepi32_epi64(sum32); + const __m128i sum64_1 = _mm_cvtepi32_epi64(_mm_srli_si128(sum32, 8)); + sum64 = _mm_add_epi64(sum64, sum64_0); + sum64 = _mm_add_epi64(sum64, sum64_1); + } + } else if (params->r[1] > 0) { + __m128i xq_coeff = pair_set_epi16(xq[1], -(xq[1] << SGRPROJ_RST_BITS)); + for (i = 0; i < height; ++i) { + __m128i sum32 = _mm_setzero_si128(); + for (j = 0; j < width - 8; j += 8) { + const __m128i d0 = _mm_cvtepu8_epi16(xx_loadl_64(dat + j)); + const __m128i s0 = _mm_cvtepu8_epi16(xx_loadl_64(src + j)); + const __m128i flt1_16b = + _mm_packs_epi32(xx_loadu_128(flt1 + j), xx_loadu_128(flt1 + j + 4)); + const __m128i v0 = + _mm_madd_epi16(xq_coeff, _mm_unpacklo_epi16(flt1_16b, d0)); + const __m128i v1 = + _mm_madd_epi16(xq_coeff, _mm_unpackhi_epi16(flt1_16b, d0)); + const __m128i vr0 = _mm_srai_epi32(_mm_add_epi32(v0, rounding), shift); + const __m128i vr1 = _mm_srai_epi32(_mm_add_epi32(v1, rounding), shift); + const __m128i e0 = + _mm_sub_epi16(_mm_add_epi16(_mm_packs_epi32(vr0, vr1), d0), s0); + const __m128i err0 = _mm_madd_epi16(e0, e0); + sum32 = _mm_add_epi32(sum32, err0); + } + for (k = j; k < width; ++k) { + const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS); + int32_t v = xq[1] * (flt1[k] - u); + const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k]; + err += e * e; + } + dat += dat_stride; + src += src_stride; + flt1 += flt1_stride; + const __m128i sum64_0 = _mm_cvtepi32_epi64(sum32); + const __m128i sum64_1 = _mm_cvtepi32_epi64(_mm_srli_si128(sum32, 8)); + sum64 = _mm_add_epi64(sum64, sum64_0); + sum64 = _mm_add_epi64(sum64, sum64_1); + } + } else { + __m128i sum32 = _mm_setzero_si128(); + for (i = 0; i < height; ++i) { + for (j = 0; j < width - 16; j += 16) { + const __m128i d = xx_loadu_128(dat + j); + const __m128i s = xx_loadu_128(src + j); + const __m128i d0 = _mm_cvtepu8_epi16(d); + const __m128i d1 = _mm_cvtepu8_epi16(_mm_srli_si128(d, 8)); + const __m128i s0 = _mm_cvtepu8_epi16(s); + const __m128i s1 = _mm_cvtepu8_epi16(_mm_srli_si128(s, 8)); + const __m128i diff0 = _mm_sub_epi16(d0, s0); + const __m128i diff1 = _mm_sub_epi16(d1, s1); + const __m128i err0 = _mm_madd_epi16(diff0, diff0); + const __m128i err1 = _mm_madd_epi16(diff1, diff1); + sum32 = _mm_add_epi32(sum32, err0); + sum32 = _mm_add_epi32(sum32, err1); + } + for (k = j; k < width; ++k) { + const int32_t e = (int32_t)(dat[k]) - src[k]; + err += e * e; + } + dat += dat_stride; + src += src_stride; + } + const __m128i sum64_0 = _mm_cvtepi32_epi64(sum32); + const __m128i sum64_1 = _mm_cvtepi32_epi64(_mm_srli_si128(sum32, 8)); + sum64 = _mm_add_epi64(sum64_0, sum64_1); + } + int64_t sum[2]; + xx_storeu_128(sum, sum64); + err += sum[0] + sum[1]; + return err; +} diff --git a/third_party/aom/av1/encoder/x86/temporal_filter_apply_sse2.asm b/third_party/aom/av1/encoder/x86/temporal_filter_apply_sse2.asm new file mode 100644 index 0000000000..30983d1c10 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/temporal_filter_apply_sse2.asm @@ -0,0 +1,217 @@ +; +; Copyright (c) 2016, 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 "aom_ports/x86_abi_support.asm" + +SECTION .text + +; void av1_temporal_filter_apply_sse2 | arg +; (unsigned char *frame1, | 0 +; unsigned int stride, | 1 +; unsigned char *frame2, | 2 +; unsigned int block_width, | 3 +; unsigned int block_height, | 4 +; int strength, | 5 +; int filter_weight, | 6 +; unsigned int *accumulator, | 7 +; unsigned short *count) | 8 +global sym(av1_temporal_filter_apply_sse2) PRIVATE +sym(av1_temporal_filter_apply_sse2): + + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 9 + SAVE_XMM 7 + GET_GOT rbx + push rsi + push rdi + ALIGN_STACK 16, rax + %define block_width 0 + %define block_height 16 + %define strength 32 + %define filter_weight 48 + %define rounding_bit 64 + %define rbp_backup 80 + %define stack_size 96 + sub rsp, stack_size + mov [rsp + rbp_backup], rbp + ; end prolog + + mov edx, arg(3) + mov [rsp + block_width], rdx + mov edx, arg(4) + mov [rsp + block_height], rdx + movd xmm6, arg(5) + movdqa [rsp + strength], xmm6 ; where strength is used, all 16 bytes are read + + ; calculate the rounding bit outside the loop + ; 0x8000 >> (16 - strength) + mov rdx, 16 + sub rdx, arg(5) ; 16 - strength + movq xmm4, rdx ; can't use rdx w/ shift + movdqa xmm5, [GLOBAL(_const_top_bit)] + psrlw xmm5, xmm4 + movdqa [rsp + rounding_bit], xmm5 + + mov rsi, arg(0) ; src/frame1 + mov rdx, arg(2) ; predictor frame + mov rdi, arg(7) ; accumulator + mov rax, arg(8) ; count + + ; dup the filter weight and store for later + movd xmm0, arg(6) ; filter_weight + pshuflw xmm0, xmm0, 0 + punpcklwd xmm0, xmm0 + movdqa [rsp + filter_weight], xmm0 + + mov rbp, arg(1) ; stride + pxor xmm7, xmm7 ; zero for extraction + + mov rcx, [rsp + block_width] + imul rcx, [rsp + block_height] + add rcx, rdx + cmp dword ptr [rsp + block_width], 8 + jne .temporal_filter_apply_load_16 + +.temporal_filter_apply_load_8: + movq xmm0, [rsi] ; first row + lea rsi, [rsi + rbp] ; += stride + punpcklbw xmm0, xmm7 ; src[ 0- 7] + movq xmm1, [rsi] ; second row + lea rsi, [rsi + rbp] ; += stride + punpcklbw xmm1, xmm7 ; src[ 8-15] + jmp .temporal_filter_apply_load_finished + +.temporal_filter_apply_load_16: + movdqa xmm0, [rsi] ; src (frame1) + lea rsi, [rsi + rbp] ; += stride + movdqa xmm1, xmm0 + punpcklbw xmm0, xmm7 ; src[ 0- 7] + punpckhbw xmm1, xmm7 ; src[ 8-15] + +.temporal_filter_apply_load_finished: + movdqa xmm2, [rdx] ; predictor (frame2) + movdqa xmm3, xmm2 + punpcklbw xmm2, xmm7 ; pred[ 0- 7] + punpckhbw xmm3, xmm7 ; pred[ 8-15] + + ; modifier = src_byte - pixel_value + psubw xmm0, xmm2 ; src - pred[ 0- 7] + psubw xmm1, xmm3 ; src - pred[ 8-15] + + ; modifier *= modifier + pmullw xmm0, xmm0 ; modifer[ 0- 7]^2 + pmullw xmm1, xmm1 ; modifer[ 8-15]^2 + + ; modifier *= 3 + pmullw xmm0, [GLOBAL(_const_3w)] + pmullw xmm1, [GLOBAL(_const_3w)] + + ; modifer += 0x8000 >> (16 - strength) + paddw xmm0, [rsp + rounding_bit] + paddw xmm1, [rsp + rounding_bit] + + ; modifier >>= strength + psrlw xmm0, [rsp + strength] + psrlw xmm1, [rsp + strength] + + ; modifier = 16 - modifier + ; saturation takes care of modifier > 16 + movdqa xmm3, [GLOBAL(_const_16w)] + movdqa xmm2, [GLOBAL(_const_16w)] + psubusw xmm3, xmm1 + psubusw xmm2, xmm0 + + ; modifier *= filter_weight + pmullw xmm2, [rsp + filter_weight] + pmullw xmm3, [rsp + filter_weight] + + ; count + movdqa xmm4, [rax] + movdqa xmm5, [rax+16] + ; += modifier + paddw xmm4, xmm2 + paddw xmm5, xmm3 + ; write back + movdqa [rax], xmm4 + movdqa [rax+16], xmm5 + lea rax, [rax + 16*2] ; count += 16*(sizeof(short)) + + ; load and extract the predictor up to shorts + pxor xmm7, xmm7 + movdqa xmm0, [rdx] + lea rdx, [rdx + 16*1] ; pred += 16*(sizeof(char)) + movdqa xmm1, xmm0 + punpcklbw xmm0, xmm7 ; pred[ 0- 7] + punpckhbw xmm1, xmm7 ; pred[ 8-15] + + ; modifier *= pixel_value + pmullw xmm0, xmm2 + pmullw xmm1, xmm3 + + ; expand to double words + movdqa xmm2, xmm0 + punpcklwd xmm0, xmm7 ; [ 0- 3] + punpckhwd xmm2, xmm7 ; [ 4- 7] + movdqa xmm3, xmm1 + punpcklwd xmm1, xmm7 ; [ 8-11] + punpckhwd xmm3, xmm7 ; [12-15] + + ; accumulator + movdqa xmm4, [rdi] + movdqa xmm5, [rdi+16] + movdqa xmm6, [rdi+32] + movdqa xmm7, [rdi+48] + ; += modifier + paddd xmm4, xmm0 + paddd xmm5, xmm2 + paddd xmm6, xmm1 + paddd xmm7, xmm3 + ; write back + movdqa [rdi], xmm4 + movdqa [rdi+16], xmm5 + movdqa [rdi+32], xmm6 + movdqa [rdi+48], xmm7 + lea rdi, [rdi + 16*4] ; accumulator += 16*(sizeof(int)) + + cmp rdx, rcx + je .temporal_filter_apply_epilog + pxor xmm7, xmm7 ; zero for extraction + cmp dword ptr [rsp + block_width], 16 + je .temporal_filter_apply_load_16 + jmp .temporal_filter_apply_load_8 + +.temporal_filter_apply_epilog: + ; begin epilog + mov rbp, [rsp + rbp_backup] + add rsp, stack_size + pop rsp + pop rdi + pop rsi + RESTORE_GOT + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +SECTION_RODATA +align 16 +_const_3w: + times 8 dw 3 +align 16 +_const_top_bit: + times 8 dw 1<<15 +align 16 +_const_16w: + times 8 dw 16 diff --git a/third_party/aom/av1/encoder/x86/wedge_utils_avx2.c b/third_party/aom/av1/encoder/x86/wedge_utils_avx2.c new file mode 100644 index 0000000000..2a792f14e6 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/wedge_utils_avx2.c @@ -0,0 +1,215 @@ +/* + * Copyright (c) 2018, 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 <immintrin.h> +#include <smmintrin.h> + +#include "aom_dsp/x86/synonyms.h" +#include "aom_dsp/x86/synonyms_avx2.h" +#include "aom/aom_integer.h" + +#include "av1/common/reconinter.h" + +#define MAX_MASK_VALUE (1 << WEDGE_WEIGHT_BITS) + +/** + * See av1_wedge_sse_from_residuals_c + */ +uint64_t av1_wedge_sse_from_residuals_avx2(const int16_t *r1, const int16_t *d, + const uint8_t *m, int N) { + int n = -N; + + uint64_t csse; + + const __m256i v_mask_max_w = _mm256_set1_epi16(MAX_MASK_VALUE); + const __m256i v_zext_q = yy_set1_64_from_32i(0xffffffff); + + __m256i v_acc0_q = _mm256_setzero_si256(); + + assert(N % 64 == 0); + + r1 += N; + d += N; + m += N; + + do { + const __m256i v_r0_w = _mm256_lddqu_si256((__m256i *)(r1 + n)); + const __m256i v_d0_w = _mm256_lddqu_si256((__m256i *)(d + n)); + const __m128i v_m01_b = _mm_lddqu_si128((__m128i *)(m + n)); + + const __m256i v_rd0l_w = _mm256_unpacklo_epi16(v_d0_w, v_r0_w); + const __m256i v_rd0h_w = _mm256_unpackhi_epi16(v_d0_w, v_r0_w); + const __m256i v_m0_w = _mm256_cvtepu8_epi16(v_m01_b); + + const __m256i v_m0l_w = _mm256_unpacklo_epi16(v_m0_w, v_mask_max_w); + const __m256i v_m0h_w = _mm256_unpackhi_epi16(v_m0_w, v_mask_max_w); + + const __m256i v_t0l_d = _mm256_madd_epi16(v_rd0l_w, v_m0l_w); + const __m256i v_t0h_d = _mm256_madd_epi16(v_rd0h_w, v_m0h_w); + + const __m256i v_t0_w = _mm256_packs_epi32(v_t0l_d, v_t0h_d); + + const __m256i v_sq0_d = _mm256_madd_epi16(v_t0_w, v_t0_w); + + const __m256i v_sum0_q = _mm256_add_epi64( + _mm256_and_si256(v_sq0_d, v_zext_q), _mm256_srli_epi64(v_sq0_d, 32)); + + v_acc0_q = _mm256_add_epi64(v_acc0_q, v_sum0_q); + + n += 16; + } while (n); + + v_acc0_q = _mm256_add_epi64(v_acc0_q, _mm256_srli_si256(v_acc0_q, 8)); + __m128i v_acc_q_0 = _mm256_castsi256_si128(v_acc0_q); + __m128i v_acc_q_1 = _mm256_extracti128_si256(v_acc0_q, 1); + v_acc_q_0 = _mm_add_epi64(v_acc_q_0, v_acc_q_1); +#if ARCH_X86_64 + csse = (uint64_t)_mm_extract_epi64(v_acc_q_0, 0); +#else + xx_storel_64(&csse, v_acc_q_0); +#endif + + return ROUND_POWER_OF_TWO(csse, 2 * WEDGE_WEIGHT_BITS); +} + +/** + * See av1_wedge_sign_from_residuals_c + */ +int av1_wedge_sign_from_residuals_avx2(const int16_t *ds, const uint8_t *m, + int N, int64_t limit) { + int64_t acc; + __m256i v_acc0_d = _mm256_setzero_si256(); + + // Input size limited to 8192 by the use of 32 bit accumulators and m + // being between [0, 64]. Overflow might happen at larger sizes, + // though it is practically impossible on real video input. + assert(N < 8192); + assert(N % 64 == 0); + + do { + const __m256i v_m01_b = _mm256_lddqu_si256((__m256i *)(m)); + const __m256i v_m23_b = _mm256_lddqu_si256((__m256i *)(m + 32)); + + const __m256i v_d0_w = _mm256_lddqu_si256((__m256i *)(ds)); + const __m256i v_d1_w = _mm256_lddqu_si256((__m256i *)(ds + 16)); + const __m256i v_d2_w = _mm256_lddqu_si256((__m256i *)(ds + 32)); + const __m256i v_d3_w = _mm256_lddqu_si256((__m256i *)(ds + 48)); + + const __m256i v_m0_w = + _mm256_cvtepu8_epi16(_mm256_castsi256_si128(v_m01_b)); + const __m256i v_m1_w = + _mm256_cvtepu8_epi16(_mm256_extracti128_si256(v_m01_b, 1)); + const __m256i v_m2_w = + _mm256_cvtepu8_epi16(_mm256_castsi256_si128(v_m23_b)); + const __m256i v_m3_w = + _mm256_cvtepu8_epi16(_mm256_extracti128_si256(v_m23_b, 1)); + + const __m256i v_p0_d = _mm256_madd_epi16(v_d0_w, v_m0_w); + const __m256i v_p1_d = _mm256_madd_epi16(v_d1_w, v_m1_w); + const __m256i v_p2_d = _mm256_madd_epi16(v_d2_w, v_m2_w); + const __m256i v_p3_d = _mm256_madd_epi16(v_d3_w, v_m3_w); + + const __m256i v_p01_d = _mm256_add_epi32(v_p0_d, v_p1_d); + const __m256i v_p23_d = _mm256_add_epi32(v_p2_d, v_p3_d); + + const __m256i v_p0123_d = _mm256_add_epi32(v_p01_d, v_p23_d); + + v_acc0_d = _mm256_add_epi32(v_acc0_d, v_p0123_d); + + ds += 64; + m += 64; + + N -= 64; + } while (N); + + __m256i v_sign_d = _mm256_srai_epi32(v_acc0_d, 31); + v_acc0_d = _mm256_add_epi64(_mm256_unpacklo_epi32(v_acc0_d, v_sign_d), + _mm256_unpackhi_epi32(v_acc0_d, v_sign_d)); + + __m256i v_acc_q = _mm256_add_epi64(v_acc0_d, _mm256_srli_si256(v_acc0_d, 8)); + + __m128i v_acc_q_0 = _mm256_castsi256_si128(v_acc_q); + __m128i v_acc_q_1 = _mm256_extracti128_si256(v_acc_q, 1); + v_acc_q_0 = _mm_add_epi64(v_acc_q_0, v_acc_q_1); + +#if ARCH_X86_64 + acc = (uint64_t)_mm_extract_epi64(v_acc_q_0, 0); +#else + xx_storel_64(&acc, v_acc_q_0); +#endif + + return acc > limit; +} + +/** + * av1_wedge_compute_delta_squares_c + */ +void av1_wedge_compute_delta_squares_avx2(int16_t *d, const int16_t *a, + const int16_t *b, int N) { + const __m256i v_neg_w = _mm256_set1_epi32(0xffff0001); + + assert(N % 64 == 0); + + do { + const __m256i v_a0_w = _mm256_lddqu_si256((__m256i *)(a)); + const __m256i v_b0_w = _mm256_lddqu_si256((__m256i *)(b)); + const __m256i v_a1_w = _mm256_lddqu_si256((__m256i *)(a + 16)); + const __m256i v_b1_w = _mm256_lddqu_si256((__m256i *)(b + 16)); + const __m256i v_a2_w = _mm256_lddqu_si256((__m256i *)(a + 32)); + const __m256i v_b2_w = _mm256_lddqu_si256((__m256i *)(b + 32)); + const __m256i v_a3_w = _mm256_lddqu_si256((__m256i *)(a + 48)); + const __m256i v_b3_w = _mm256_lddqu_si256((__m256i *)(b + 48)); + + const __m256i v_ab0l_w = _mm256_unpacklo_epi16(v_a0_w, v_b0_w); + const __m256i v_ab0h_w = _mm256_unpackhi_epi16(v_a0_w, v_b0_w); + const __m256i v_ab1l_w = _mm256_unpacklo_epi16(v_a1_w, v_b1_w); + const __m256i v_ab1h_w = _mm256_unpackhi_epi16(v_a1_w, v_b1_w); + const __m256i v_ab2l_w = _mm256_unpacklo_epi16(v_a2_w, v_b2_w); + const __m256i v_ab2h_w = _mm256_unpackhi_epi16(v_a2_w, v_b2_w); + const __m256i v_ab3l_w = _mm256_unpacklo_epi16(v_a3_w, v_b3_w); + const __m256i v_ab3h_w = _mm256_unpackhi_epi16(v_a3_w, v_b3_w); + + // Negate top word of pairs + const __m256i v_abl0n_w = _mm256_sign_epi16(v_ab0l_w, v_neg_w); + const __m256i v_abh0n_w = _mm256_sign_epi16(v_ab0h_w, v_neg_w); + const __m256i v_abl1n_w = _mm256_sign_epi16(v_ab1l_w, v_neg_w); + const __m256i v_abh1n_w = _mm256_sign_epi16(v_ab1h_w, v_neg_w); + const __m256i v_abl2n_w = _mm256_sign_epi16(v_ab2l_w, v_neg_w); + const __m256i v_abh2n_w = _mm256_sign_epi16(v_ab2h_w, v_neg_w); + const __m256i v_abl3n_w = _mm256_sign_epi16(v_ab3l_w, v_neg_w); + const __m256i v_abh3n_w = _mm256_sign_epi16(v_ab3h_w, v_neg_w); + + const __m256i v_r0l_w = _mm256_madd_epi16(v_ab0l_w, v_abl0n_w); + const __m256i v_r0h_w = _mm256_madd_epi16(v_ab0h_w, v_abh0n_w); + const __m256i v_r1l_w = _mm256_madd_epi16(v_ab1l_w, v_abl1n_w); + const __m256i v_r1h_w = _mm256_madd_epi16(v_ab1h_w, v_abh1n_w); + const __m256i v_r2l_w = _mm256_madd_epi16(v_ab2l_w, v_abl2n_w); + const __m256i v_r2h_w = _mm256_madd_epi16(v_ab2h_w, v_abh2n_w); + const __m256i v_r3l_w = _mm256_madd_epi16(v_ab3l_w, v_abl3n_w); + const __m256i v_r3h_w = _mm256_madd_epi16(v_ab3h_w, v_abh3n_w); + + const __m256i v_r0_w = _mm256_packs_epi32(v_r0l_w, v_r0h_w); + const __m256i v_r1_w = _mm256_packs_epi32(v_r1l_w, v_r1h_w); + const __m256i v_r2_w = _mm256_packs_epi32(v_r2l_w, v_r2h_w); + const __m256i v_r3_w = _mm256_packs_epi32(v_r3l_w, v_r3h_w); + + _mm256_store_si256((__m256i *)(d), v_r0_w); + _mm256_store_si256((__m256i *)(d + 16), v_r1_w); + _mm256_store_si256((__m256i *)(d + 32), v_r2_w); + _mm256_store_si256((__m256i *)(d + 48), v_r3_w); + + a += 64; + b += 64; + d += 64; + N -= 64; + } while (N); +} diff --git a/third_party/aom/av1/encoder/x86/wedge_utils_sse2.c b/third_party/aom/av1/encoder/x86/wedge_utils_sse2.c new file mode 100644 index 0000000000..4d2e99f258 --- /dev/null +++ b/third_party/aom/av1/encoder/x86/wedge_utils_sse2.c @@ -0,0 +1,254 @@ +/* + * Copyright (c) 2016, 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 <immintrin.h> + +#include "aom_dsp/x86/synonyms.h" + +#include "aom/aom_integer.h" + +#include "av1/common/reconinter.h" + +#define MAX_MASK_VALUE (1 << WEDGE_WEIGHT_BITS) + +/** + * See av1_wedge_sse_from_residuals_c + */ +uint64_t av1_wedge_sse_from_residuals_sse2(const int16_t *r1, const int16_t *d, + const uint8_t *m, int N) { + int n = -N; + int n8 = n + 8; + + uint64_t csse; + + const __m128i v_mask_max_w = _mm_set1_epi16(MAX_MASK_VALUE); + const __m128i v_zext_q = xx_set1_64_from_32i(0xffffffff); + + __m128i v_acc0_q = _mm_setzero_si128(); + + assert(N % 64 == 0); + + r1 += N; + d += N; + m += N; + + do { + const __m128i v_r0_w = xx_load_128(r1 + n); + const __m128i v_r1_w = xx_load_128(r1 + n8); + const __m128i v_d0_w = xx_load_128(d + n); + const __m128i v_d1_w = xx_load_128(d + n8); + const __m128i v_m01_b = xx_load_128(m + n); + + const __m128i v_rd0l_w = _mm_unpacklo_epi16(v_d0_w, v_r0_w); + const __m128i v_rd0h_w = _mm_unpackhi_epi16(v_d0_w, v_r0_w); + const __m128i v_rd1l_w = _mm_unpacklo_epi16(v_d1_w, v_r1_w); + const __m128i v_rd1h_w = _mm_unpackhi_epi16(v_d1_w, v_r1_w); + const __m128i v_m0_w = _mm_unpacklo_epi8(v_m01_b, _mm_setzero_si128()); + const __m128i v_m1_w = _mm_unpackhi_epi8(v_m01_b, _mm_setzero_si128()); + + const __m128i v_m0l_w = _mm_unpacklo_epi16(v_m0_w, v_mask_max_w); + const __m128i v_m0h_w = _mm_unpackhi_epi16(v_m0_w, v_mask_max_w); + const __m128i v_m1l_w = _mm_unpacklo_epi16(v_m1_w, v_mask_max_w); + const __m128i v_m1h_w = _mm_unpackhi_epi16(v_m1_w, v_mask_max_w); + + const __m128i v_t0l_d = _mm_madd_epi16(v_rd0l_w, v_m0l_w); + const __m128i v_t0h_d = _mm_madd_epi16(v_rd0h_w, v_m0h_w); + const __m128i v_t1l_d = _mm_madd_epi16(v_rd1l_w, v_m1l_w); + const __m128i v_t1h_d = _mm_madd_epi16(v_rd1h_w, v_m1h_w); + + const __m128i v_t0_w = _mm_packs_epi32(v_t0l_d, v_t0h_d); + const __m128i v_t1_w = _mm_packs_epi32(v_t1l_d, v_t1h_d); + + const __m128i v_sq0_d = _mm_madd_epi16(v_t0_w, v_t0_w); + const __m128i v_sq1_d = _mm_madd_epi16(v_t1_w, v_t1_w); + + const __m128i v_sum0_q = _mm_add_epi64(_mm_and_si128(v_sq0_d, v_zext_q), + _mm_srli_epi64(v_sq0_d, 32)); + const __m128i v_sum1_q = _mm_add_epi64(_mm_and_si128(v_sq1_d, v_zext_q), + _mm_srli_epi64(v_sq1_d, 32)); + + v_acc0_q = _mm_add_epi64(v_acc0_q, v_sum0_q); + v_acc0_q = _mm_add_epi64(v_acc0_q, v_sum1_q); + + n8 += 16; + n += 16; + } while (n); + + v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_srli_si128(v_acc0_q, 8)); + +#if ARCH_X86_64 + csse = (uint64_t)_mm_cvtsi128_si64(v_acc0_q); +#else + xx_storel_64(&csse, v_acc0_q); +#endif + + return ROUND_POWER_OF_TWO(csse, 2 * WEDGE_WEIGHT_BITS); +} + +/** + * See av1_wedge_sign_from_residuals_c + */ +int av1_wedge_sign_from_residuals_sse2(const int16_t *ds, const uint8_t *m, + int N, int64_t limit) { + int64_t acc; + + __m128i v_sign_d; + __m128i v_acc0_d = _mm_setzero_si128(); + __m128i v_acc1_d = _mm_setzero_si128(); + __m128i v_acc_q; + + // Input size limited to 8192 by the use of 32 bit accumulators and m + // being between [0, 64]. Overflow might happen at larger sizes, + // though it is practically impossible on real video input. + assert(N < 8192); + assert(N % 64 == 0); + + do { + const __m128i v_m01_b = xx_load_128(m); + const __m128i v_m23_b = xx_load_128(m + 16); + const __m128i v_m45_b = xx_load_128(m + 32); + const __m128i v_m67_b = xx_load_128(m + 48); + + const __m128i v_d0_w = xx_load_128(ds); + const __m128i v_d1_w = xx_load_128(ds + 8); + const __m128i v_d2_w = xx_load_128(ds + 16); + const __m128i v_d3_w = xx_load_128(ds + 24); + const __m128i v_d4_w = xx_load_128(ds + 32); + const __m128i v_d5_w = xx_load_128(ds + 40); + const __m128i v_d6_w = xx_load_128(ds + 48); + const __m128i v_d7_w = xx_load_128(ds + 56); + + const __m128i v_m0_w = _mm_unpacklo_epi8(v_m01_b, _mm_setzero_si128()); + const __m128i v_m1_w = _mm_unpackhi_epi8(v_m01_b, _mm_setzero_si128()); + const __m128i v_m2_w = _mm_unpacklo_epi8(v_m23_b, _mm_setzero_si128()); + const __m128i v_m3_w = _mm_unpackhi_epi8(v_m23_b, _mm_setzero_si128()); + const __m128i v_m4_w = _mm_unpacklo_epi8(v_m45_b, _mm_setzero_si128()); + const __m128i v_m5_w = _mm_unpackhi_epi8(v_m45_b, _mm_setzero_si128()); + const __m128i v_m6_w = _mm_unpacklo_epi8(v_m67_b, _mm_setzero_si128()); + const __m128i v_m7_w = _mm_unpackhi_epi8(v_m67_b, _mm_setzero_si128()); + + const __m128i v_p0_d = _mm_madd_epi16(v_d0_w, v_m0_w); + const __m128i v_p1_d = _mm_madd_epi16(v_d1_w, v_m1_w); + const __m128i v_p2_d = _mm_madd_epi16(v_d2_w, v_m2_w); + const __m128i v_p3_d = _mm_madd_epi16(v_d3_w, v_m3_w); + const __m128i v_p4_d = _mm_madd_epi16(v_d4_w, v_m4_w); + const __m128i v_p5_d = _mm_madd_epi16(v_d5_w, v_m5_w); + const __m128i v_p6_d = _mm_madd_epi16(v_d6_w, v_m6_w); + const __m128i v_p7_d = _mm_madd_epi16(v_d7_w, v_m7_w); + + const __m128i v_p01_d = _mm_add_epi32(v_p0_d, v_p1_d); + const __m128i v_p23_d = _mm_add_epi32(v_p2_d, v_p3_d); + const __m128i v_p45_d = _mm_add_epi32(v_p4_d, v_p5_d); + const __m128i v_p67_d = _mm_add_epi32(v_p6_d, v_p7_d); + + const __m128i v_p0123_d = _mm_add_epi32(v_p01_d, v_p23_d); + const __m128i v_p4567_d = _mm_add_epi32(v_p45_d, v_p67_d); + + v_acc0_d = _mm_add_epi32(v_acc0_d, v_p0123_d); + v_acc1_d = _mm_add_epi32(v_acc1_d, v_p4567_d); + + ds += 64; + m += 64; + + N -= 64; + } while (N); + + v_sign_d = _mm_cmplt_epi32(v_acc0_d, _mm_setzero_si128()); + v_acc0_d = _mm_add_epi64(_mm_unpacklo_epi32(v_acc0_d, v_sign_d), + _mm_unpackhi_epi32(v_acc0_d, v_sign_d)); + + v_sign_d = _mm_cmplt_epi32(v_acc1_d, _mm_setzero_si128()); + v_acc1_d = _mm_add_epi64(_mm_unpacklo_epi32(v_acc1_d, v_sign_d), + _mm_unpackhi_epi32(v_acc1_d, v_sign_d)); + + v_acc_q = _mm_add_epi64(v_acc0_d, v_acc1_d); + + v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_si128(v_acc_q, 8)); + +#if ARCH_X86_64 + acc = (uint64_t)_mm_cvtsi128_si64(v_acc_q); +#else + xx_storel_64(&acc, v_acc_q); +#endif + + return acc > limit; +} + +// Negate under mask +static INLINE __m128i negm_epi16(__m128i v_v_w, __m128i v_mask_w) { + return _mm_sub_epi16(_mm_xor_si128(v_v_w, v_mask_w), v_mask_w); +} + +/** + * av1_wedge_compute_delta_squares_c + */ +void av1_wedge_compute_delta_squares_sse2(int16_t *d, const int16_t *a, + const int16_t *b, int N) { + const __m128i v_neg_w = + _mm_set_epi16(0xffff, 0, 0xffff, 0, 0xffff, 0, 0xffff, 0); + + assert(N % 64 == 0); + + do { + const __m128i v_a0_w = xx_load_128(a); + const __m128i v_b0_w = xx_load_128(b); + const __m128i v_a1_w = xx_load_128(a + 8); + const __m128i v_b1_w = xx_load_128(b + 8); + const __m128i v_a2_w = xx_load_128(a + 16); + const __m128i v_b2_w = xx_load_128(b + 16); + const __m128i v_a3_w = xx_load_128(a + 24); + const __m128i v_b3_w = xx_load_128(b + 24); + + const __m128i v_ab0l_w = _mm_unpacklo_epi16(v_a0_w, v_b0_w); + const __m128i v_ab0h_w = _mm_unpackhi_epi16(v_a0_w, v_b0_w); + const __m128i v_ab1l_w = _mm_unpacklo_epi16(v_a1_w, v_b1_w); + const __m128i v_ab1h_w = _mm_unpackhi_epi16(v_a1_w, v_b1_w); + const __m128i v_ab2l_w = _mm_unpacklo_epi16(v_a2_w, v_b2_w); + const __m128i v_ab2h_w = _mm_unpackhi_epi16(v_a2_w, v_b2_w); + const __m128i v_ab3l_w = _mm_unpacklo_epi16(v_a3_w, v_b3_w); + const __m128i v_ab3h_w = _mm_unpackhi_epi16(v_a3_w, v_b3_w); + + // Negate top word of pairs + const __m128i v_abl0n_w = negm_epi16(v_ab0l_w, v_neg_w); + const __m128i v_abh0n_w = negm_epi16(v_ab0h_w, v_neg_w); + const __m128i v_abl1n_w = negm_epi16(v_ab1l_w, v_neg_w); + const __m128i v_abh1n_w = negm_epi16(v_ab1h_w, v_neg_w); + const __m128i v_abl2n_w = negm_epi16(v_ab2l_w, v_neg_w); + const __m128i v_abh2n_w = negm_epi16(v_ab2h_w, v_neg_w); + const __m128i v_abl3n_w = negm_epi16(v_ab3l_w, v_neg_w); + const __m128i v_abh3n_w = negm_epi16(v_ab3h_w, v_neg_w); + + const __m128i v_r0l_w = _mm_madd_epi16(v_ab0l_w, v_abl0n_w); + const __m128i v_r0h_w = _mm_madd_epi16(v_ab0h_w, v_abh0n_w); + const __m128i v_r1l_w = _mm_madd_epi16(v_ab1l_w, v_abl1n_w); + const __m128i v_r1h_w = _mm_madd_epi16(v_ab1h_w, v_abh1n_w); + const __m128i v_r2l_w = _mm_madd_epi16(v_ab2l_w, v_abl2n_w); + const __m128i v_r2h_w = _mm_madd_epi16(v_ab2h_w, v_abh2n_w); + const __m128i v_r3l_w = _mm_madd_epi16(v_ab3l_w, v_abl3n_w); + const __m128i v_r3h_w = _mm_madd_epi16(v_ab3h_w, v_abh3n_w); + + const __m128i v_r0_w = _mm_packs_epi32(v_r0l_w, v_r0h_w); + const __m128i v_r1_w = _mm_packs_epi32(v_r1l_w, v_r1h_w); + const __m128i v_r2_w = _mm_packs_epi32(v_r2l_w, v_r2h_w); + const __m128i v_r3_w = _mm_packs_epi32(v_r3l_w, v_r3h_w); + + xx_store_128(d, v_r0_w); + xx_store_128(d + 8, v_r1_w); + xx_store_128(d + 16, v_r2_w); + xx_store_128(d + 24, v_r3_w); + + a += 32; + b += 32; + d += 32; + N -= 32; + } while (N); +} |