diff options
Diffstat (limited to 'vp9/encoder')
| -rw-r--r-- | vp9/encoder/vp9_encodeframe.c | 4 | ||||
| -rw-r--r-- | vp9/encoder/x86/vp9_dct32x32_sse2.c | 891 | ||||
| -rw-r--r-- | vp9/encoder/x86/vp9_dct_impl_sse2.c | 1011 | ||||
| -rw-r--r-- | vp9/encoder/x86/vp9_dct_mmx.asm | 31 | ||||
| -rw-r--r-- | vp9/encoder/x86/vp9_dct_sse2.c | 1357 | ||||
| -rw-r--r-- | vp9/encoder/x86/vp9_dct_sse2.h | 373 |
6 files changed, 2331 insertions, 1336 deletions
diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c index a5adcbb7a..d5122d0bc 100644 --- a/vp9/encoder/vp9_encodeframe.c +++ b/vp9/encoder/vp9_encodeframe.c @@ -3535,9 +3535,9 @@ static void encode_frame_internal(VP9_COMP *cpi) { #if CONFIG_VP9_HIGHBITDEPTH if (cm->use_highbitdepth) - x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4; - else x->fwd_txm4x4 = xd->lossless ? vp9_highbd_fwht4x4 : vp9_highbd_fdct4x4; + else + x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4; x->highbd_itxm_add = xd->lossless ? vp9_highbd_iwht4x4_add : vp9_highbd_idct4x4_add; #else diff --git a/vp9/encoder/x86/vp9_dct32x32_sse2.c b/vp9/encoder/x86/vp9_dct32x32_sse2.c index dc36cc471..7ec126e4b 100644 --- a/vp9/encoder/x86/vp9_dct32x32_sse2.c +++ b/vp9/encoder/x86/vp9_dct32x32_sse2.c @@ -10,30 +10,50 @@ #include <emmintrin.h> // SSE2 #include "vp9/common/vp9_idct.h" // for cospi constants +#include "vp9/encoder/x86/vp9_dct_sse2.h" +#include "vp9/encoder/vp9_dct.h" #include "vpx_ports/mem.h" -#define pair_set_epi32(a, b) \ - _mm_set_epi32((int)(b), (int)(a), (int)(b), (int)(a)) - +#if DCT_HIGH_BIT_DEPTH +#define ADD_EPI16 _mm_adds_epi16 +#define SUB_EPI16 _mm_subs_epi16 #if FDCT32x32_HIGH_PRECISION -static INLINE __m128i k_madd_epi32(__m128i a, __m128i b) { - __m128i buf0, buf1; - buf0 = _mm_mul_epu32(a, b); - a = _mm_srli_epi64(a, 32); - b = _mm_srli_epi64(b, 32); - buf1 = _mm_mul_epu32(a, b); - return _mm_add_epi64(buf0, buf1); +void vp9_fdct32x32_rows_c(const int16_t *intermediate, tran_low_t *out) { + int i, j; + for (i = 0; i < 32; ++i) { + tran_high_t temp_in[32], temp_out[32]; + for (j = 0; j < 32; ++j) + temp_in[j] = intermediate[j * 32 + i]; + vp9_fdct32(temp_in, temp_out, 0); + for (j = 0; j < 32; ++j) + out[j + i * 32] = (temp_out[j] + 1 + (temp_out[j] < 0)) >> 2; + } } - -static INLINE __m128i k_packs_epi64(__m128i a, __m128i b) { - __m128i buf0 = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0)); - __m128i buf1 = _mm_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0)); - return _mm_unpacklo_epi64(buf0, buf1); + #define HIGH_FDCT32x32_2D_C vp9_highbd_fdct32x32_c + #define HIGH_FDCT32x32_2D_ROWS_C vp9_fdct32x32_rows_c +#else +void vp9_fdct32x32_rd_rows_c(const int16_t *intermediate, tran_low_t *out) { + int i, j; + for (i = 0; i < 32; ++i) { + tran_high_t temp_in[32], temp_out[32]; + for (j = 0; j < 32; ++j) + temp_in[j] = intermediate[j * 32 + i]; + vp9_fdct32(temp_in, temp_out, 1); + for (j = 0; j < 32; ++j) + out[j + i * 32] = temp_out[j]; + } } -#endif + #define HIGH_FDCT32x32_2D_C vp9_highbd_fdct32x32_rd_c + #define HIGH_FDCT32x32_2D_ROWS_C vp9_fdct32x32_rd_rows_c +#endif // FDCT32x32_HIGH_PRECISION +#else +#define ADD_EPI16 _mm_add_epi16 +#define SUB_EPI16 _mm_sub_epi16 +#endif // DCT_HIGH_BIT_DEPTH + void FDCT32x32_2D(const int16_t *input, - int16_t *output_org, int stride) { + tran_low_t *output_org, int stride) { // Calculate pre-multiplied strides const int str1 = stride; const int str2 = 2 * stride; @@ -84,6 +104,9 @@ void FDCT32x32_2D(const int16_t *input, const __m128i kOne = _mm_set1_epi16(1); // Do the two transform/transpose passes int pass; +#if DCT_HIGH_BIT_DEPTH + int overflow; +#endif for (pass = 0; pass < 2; ++pass) { // We process eight columns (transposed rows in second pass) at a time. int column_start; @@ -237,14 +260,22 @@ void FDCT32x32_2D(const int16_t *input, __m128i in29 = _mm_loadu_si128((const __m128i *)(in + 29 * 32)); __m128i in30 = _mm_loadu_si128((const __m128i *)(in + 30 * 32)); __m128i in31 = _mm_loadu_si128((const __m128i *)(in + 31 * 32)); - step1[ 0] = _mm_add_epi16(in00, in31); - step1[ 1] = _mm_add_epi16(in01, in30); - step1[ 2] = _mm_add_epi16(in02, in29); - step1[ 3] = _mm_add_epi16(in03, in28); - step1[28] = _mm_sub_epi16(in03, in28); - step1[29] = _mm_sub_epi16(in02, in29); - step1[30] = _mm_sub_epi16(in01, in30); - step1[31] = _mm_sub_epi16(in00, in31); + step1[0] = ADD_EPI16(in00, in31); + step1[1] = ADD_EPI16(in01, in30); + step1[2] = ADD_EPI16(in02, in29); + step1[3] = ADD_EPI16(in03, in28); + step1[28] = SUB_EPI16(in03, in28); + step1[29] = SUB_EPI16(in02, in29); + step1[30] = SUB_EPI16(in01, in30); + step1[31] = SUB_EPI16(in00, in31); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(step1[0], step1[1], step1[2], + step1[3], step1[28], step1[29], step1[30], step1[31]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { __m128i in04 = _mm_loadu_si128((const __m128i *)(in + 4 * 32)); @@ -255,14 +286,23 @@ void FDCT32x32_2D(const int16_t *input, __m128i in25 = _mm_loadu_si128((const __m128i *)(in + 25 * 32)); __m128i in26 = _mm_loadu_si128((const __m128i *)(in + 26 * 32)); __m128i in27 = _mm_loadu_si128((const __m128i *)(in + 27 * 32)); - step1[ 4] = _mm_add_epi16(in04, in27); - step1[ 5] = _mm_add_epi16(in05, in26); - step1[ 6] = _mm_add_epi16(in06, in25); - step1[ 7] = _mm_add_epi16(in07, in24); - step1[24] = _mm_sub_epi16(in07, in24); - step1[25] = _mm_sub_epi16(in06, in25); - step1[26] = _mm_sub_epi16(in05, in26); - step1[27] = _mm_sub_epi16(in04, in27); + step1[4] = ADD_EPI16(in04, in27); + step1[5] = ADD_EPI16(in05, in26); + step1[6] = ADD_EPI16(in06, in25); + step1[7] = ADD_EPI16(in07, in24); + step1[24] = SUB_EPI16(in07, in24); + step1[25] = SUB_EPI16(in06, in25); + step1[26] = SUB_EPI16(in05, in26); + step1[27] = SUB_EPI16(in04, in27); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(step1[4], step1[5], step1[6], + step1[7], step1[24], step1[25], + step1[26], step1[27]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { __m128i in08 = _mm_loadu_si128((const __m128i *)(in + 8 * 32)); @@ -273,14 +313,23 @@ void FDCT32x32_2D(const int16_t *input, __m128i in21 = _mm_loadu_si128((const __m128i *)(in + 21 * 32)); __m128i in22 = _mm_loadu_si128((const __m128i *)(in + 22 * 32)); __m128i in23 = _mm_loadu_si128((const __m128i *)(in + 23 * 32)); - step1[ 8] = _mm_add_epi16(in08, in23); - step1[ 9] = _mm_add_epi16(in09, in22); - step1[10] = _mm_add_epi16(in10, in21); - step1[11] = _mm_add_epi16(in11, in20); - step1[20] = _mm_sub_epi16(in11, in20); - step1[21] = _mm_sub_epi16(in10, in21); - step1[22] = _mm_sub_epi16(in09, in22); - step1[23] = _mm_sub_epi16(in08, in23); + step1[8] = ADD_EPI16(in08, in23); + step1[9] = ADD_EPI16(in09, in22); + step1[10] = ADD_EPI16(in10, in21); + step1[11] = ADD_EPI16(in11, in20); + step1[20] = SUB_EPI16(in11, in20); + step1[21] = SUB_EPI16(in10, in21); + step1[22] = SUB_EPI16(in09, in22); + step1[23] = SUB_EPI16(in08, in23); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(step1[8], step1[9], step1[10], + step1[11], step1[20], step1[21], + step1[22], step1[23]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { __m128i in12 = _mm_loadu_si128((const __m128i *)(in + 12 * 32)); @@ -291,34 +340,57 @@ void FDCT32x32_2D(const int16_t *input, __m128i in17 = _mm_loadu_si128((const __m128i *)(in + 17 * 32)); __m128i in18 = _mm_loadu_si128((const __m128i *)(in + 18 * 32)); __m128i in19 = _mm_loadu_si128((const __m128i *)(in + 19 * 32)); - step1[12] = _mm_add_epi16(in12, in19); - step1[13] = _mm_add_epi16(in13, in18); - step1[14] = _mm_add_epi16(in14, in17); - step1[15] = _mm_add_epi16(in15, in16); - step1[16] = _mm_sub_epi16(in15, in16); - step1[17] = _mm_sub_epi16(in14, in17); - step1[18] = _mm_sub_epi16(in13, in18); - step1[19] = _mm_sub_epi16(in12, in19); + step1[12] = ADD_EPI16(in12, in19); + step1[13] = ADD_EPI16(in13, in18); + step1[14] = ADD_EPI16(in14, in17); + step1[15] = ADD_EPI16(in15, in16); + step1[16] = SUB_EPI16(in15, in16); + step1[17] = SUB_EPI16(in14, in17); + step1[18] = SUB_EPI16(in13, in18); + step1[19] = SUB_EPI16(in12, in19); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(step1[12], step1[13], step1[14], + step1[15], step1[16], step1[17], + step1[18], step1[19]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } } // Stage 2 { - step2[ 0] = _mm_add_epi16(step1[0], step1[15]); - step2[ 1] = _mm_add_epi16(step1[1], step1[14]); - step2[ 2] = _mm_add_epi16(step1[2], step1[13]); - step2[ 3] = _mm_add_epi16(step1[3], step1[12]); - step2[ 4] = _mm_add_epi16(step1[4], step1[11]); - step2[ 5] = _mm_add_epi16(step1[5], step1[10]); - step2[ 6] = _mm_add_epi16(step1[6], step1[ 9]); - step2[ 7] = _mm_add_epi16(step1[7], step1[ 8]); - step2[ 8] = _mm_sub_epi16(step1[7], step1[ 8]); - step2[ 9] = _mm_sub_epi16(step1[6], step1[ 9]); - step2[10] = _mm_sub_epi16(step1[5], step1[10]); - step2[11] = _mm_sub_epi16(step1[4], step1[11]); - step2[12] = _mm_sub_epi16(step1[3], step1[12]); - step2[13] = _mm_sub_epi16(step1[2], step1[13]); - step2[14] = _mm_sub_epi16(step1[1], step1[14]); - step2[15] = _mm_sub_epi16(step1[0], step1[15]); + step2[0] = ADD_EPI16(step1[0], step1[15]); + step2[1] = ADD_EPI16(step1[1], step1[14]); + step2[2] = ADD_EPI16(step1[2], step1[13]); + step2[3] = ADD_EPI16(step1[3], step1[12]); + step2[4] = ADD_EPI16(step1[4], step1[11]); + step2[5] = ADD_EPI16(step1[5], step1[10]); + step2[6] = ADD_EPI16(step1[6], step1[ 9]); + step2[7] = ADD_EPI16(step1[7], step1[ 8]); + step2[8] = SUB_EPI16(step1[7], step1[ 8]); + step2[9] = SUB_EPI16(step1[6], step1[ 9]); + step2[10] = SUB_EPI16(step1[5], step1[10]); + step2[11] = SUB_EPI16(step1[4], step1[11]); + step2[12] = SUB_EPI16(step1[3], step1[12]); + step2[13] = SUB_EPI16(step1[2], step1[13]); + step2[14] = SUB_EPI16(step1[1], step1[14]); + step2[15] = SUB_EPI16(step1[0], step1[15]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x16( + step2[0], step2[1], step2[2], step2[3], + step2[4], step2[5], step2[6], step2[7], + step2[8], step2[9], step2[10], step2[11], + step2[12], step2[13], step2[14], step2[15]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { const __m128i s2_20_0 = _mm_unpacklo_epi16(step1[27], step1[20]); @@ -387,6 +459,18 @@ void FDCT32x32_2D(const int16_t *input, step2[25] = _mm_packs_epi32(s2_25_6, s2_25_7); step2[26] = _mm_packs_epi32(s2_26_6, s2_26_7); step2[27] = _mm_packs_epi32(s2_27_6, s2_27_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(step2[20], step2[21], step2[22], + step2[23], step2[24], step2[25], + step2[26], step2[27]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } #if !FDCT32x32_HIGH_PRECISION @@ -426,49 +510,63 @@ void FDCT32x32_2D(const int16_t *input, __m128i s3_30_0 = _mm_cmplt_epi16(step1[30], kZero); __m128i s3_31_0 = _mm_cmplt_epi16(step1[31], kZero); - step2[ 0] = _mm_sub_epi16(step2[ 0], s3_00_0); - step2[ 1] = _mm_sub_epi16(step2[ 1], s3_01_0); - step2[ 2] = _mm_sub_epi16(step2[ 2], s3_02_0); - step2[ 3] = _mm_sub_epi16(step2[ 3], s3_03_0); - step2[ 4] = _mm_sub_epi16(step2[ 4], s3_04_0); - step2[ 5] = _mm_sub_epi16(step2[ 5], s3_05_0); - step2[ 6] = _mm_sub_epi16(step2[ 6], s3_06_0); - step2[ 7] = _mm_sub_epi16(step2[ 7], s3_07_0); - step2[ 8] = _mm_sub_epi16(step2[ 8], s2_08_0); - step2[ 9] = _mm_sub_epi16(step2[ 9], s2_09_0); - step2[10] = _mm_sub_epi16(step2[10], s3_10_0); - step2[11] = _mm_sub_epi16(step2[11], s3_11_0); - step2[12] = _mm_sub_epi16(step2[12], s3_12_0); - step2[13] = _mm_sub_epi16(step2[13], s3_13_0); - step2[14] = _mm_sub_epi16(step2[14], s2_14_0); - step2[15] = _mm_sub_epi16(step2[15], s2_15_0); - step1[16] = _mm_sub_epi16(step1[16], s3_16_0); - step1[17] = _mm_sub_epi16(step1[17], s3_17_0); - step1[18] = _mm_sub_epi16(step1[18], s3_18_0); - step1[19] = _mm_sub_epi16(step1[19], s3_19_0); - step2[20] = _mm_sub_epi16(step2[20], s3_20_0); - step2[21] = _mm_sub_epi16(step2[21], s3_21_0); - step2[22] = _mm_sub_epi16(step2[22], s3_22_0); - step2[23] = _mm_sub_epi16(step2[23], s3_23_0); - step2[24] = _mm_sub_epi16(step2[24], s3_24_0); - step2[25] = _mm_sub_epi16(step2[25], s3_25_0); - step2[26] = _mm_sub_epi16(step2[26], s3_26_0); - step2[27] = _mm_sub_epi16(step2[27], s3_27_0); - step1[28] = _mm_sub_epi16(step1[28], s3_28_0); - step1[29] = _mm_sub_epi16(step1[29], s3_29_0); - step1[30] = _mm_sub_epi16(step1[30], s3_30_0); - step1[31] = _mm_sub_epi16(step1[31], s3_31_0); - - step2[ 0] = _mm_add_epi16(step2[ 0], kOne); - step2[ 1] = _mm_add_epi16(step2[ 1], kOne); - step2[ 2] = _mm_add_epi16(step2[ 2], kOne); - step2[ 3] = _mm_add_epi16(step2[ 3], kOne); - step2[ 4] = _mm_add_epi16(step2[ 4], kOne); - step2[ 5] = _mm_add_epi16(step2[ 5], kOne); - step2[ 6] = _mm_add_epi16(step2[ 6], kOne); - step2[ 7] = _mm_add_epi16(step2[ 7], kOne); - step2[ 8] = _mm_add_epi16(step2[ 8], kOne); - step2[ 9] = _mm_add_epi16(step2[ 9], kOne); + step2[0] = SUB_EPI16(step2[ 0], s3_00_0); + step2[1] = SUB_EPI16(step2[ 1], s3_01_0); + step2[2] = SUB_EPI16(step2[ 2], s3_02_0); + step2[3] = SUB_EPI16(step2[ 3], s3_03_0); + step2[4] = SUB_EPI16(step2[ 4], s3_04_0); + step2[5] = SUB_EPI16(step2[ 5], s3_05_0); + step2[6] = SUB_EPI16(step2[ 6], s3_06_0); + step2[7] = SUB_EPI16(step2[ 7], s3_07_0); + step2[8] = SUB_EPI16(step2[ 8], s2_08_0); + step2[9] = SUB_EPI16(step2[ 9], s2_09_0); + step2[10] = SUB_EPI16(step2[10], s3_10_0); + step2[11] = SUB_EPI16(step2[11], s3_11_0); + step2[12] = SUB_EPI16(step2[12], s3_12_0); + step2[13] = SUB_EPI16(step2[13], s3_13_0); + step2[14] = SUB_EPI16(step2[14], s2_14_0); + step2[15] = SUB_EPI16(step2[15], s2_15_0); + step1[16] = SUB_EPI16(step1[16], s3_16_0); + step1[17] = SUB_EPI16(step1[17], s3_17_0); + step1[18] = SUB_EPI16(step1[18], s3_18_0); + step1[19] = SUB_EPI16(step1[19], s3_19_0); + step2[20] = SUB_EPI16(step2[20], s3_20_0); + step2[21] = SUB_EPI16(step2[21], s3_21_0); + step2[22] = SUB_EPI16(step2[22], s3_22_0); + step2[23] = SUB_EPI16(step2[23], s3_23_0); + step2[24] = SUB_EPI16(step2[24], s3_24_0); + step2[25] = SUB_EPI16(step2[25], s3_25_0); + step2[26] = SUB_EPI16(step2[26], s3_26_0); + step2[27] = SUB_EPI16(step2[27], s3_27_0); + step1[28] = SUB_EPI16(step1[28], s3_28_0); + step1[29] = SUB_EPI16(step1[29], s3_29_0); + step1[30] = SUB_EPI16(step1[30], s3_30_0); + step1[31] = SUB_EPI16(step1[31], s3_31_0); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x32( + step2[0], step2[1], step2[2], step2[3], + step2[4], step2[5], step2[6], step2[7], + step2[8], step2[9], step2[10], step2[11], + step2[12], step2[13], step2[14], step2[15], + step1[16], step1[17], step1[18], step1[19], + step2[20], step2[21], step2[22], step2[23], + step2[24], step2[25], step2[26], step2[27], + step1[28], step1[29], step1[30], step1[31]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + step2[0] = _mm_add_epi16(step2[ 0], kOne); + step2[1] = _mm_add_epi16(step2[ 1], kOne); + step2[2] = _mm_add_epi16(step2[ 2], kOne); + step2[3] = _mm_add_epi16(step2[ 3], kOne); + step2[4] = _mm_add_epi16(step2[ 4], kOne); + step2[5] = _mm_add_epi16(step2[ 5], kOne); + step2[6] = _mm_add_epi16(step2[ 6], kOne); + step2[7] = _mm_add_epi16(step2[ 7], kOne); + step2[8] = _mm_add_epi16(step2[ 8], kOne); + step2[9] = _mm_add_epi16(step2[ 9], kOne); step2[10] = _mm_add_epi16(step2[10], kOne); step2[11] = _mm_add_epi16(step2[11], kOne); step2[12] = _mm_add_epi16(step2[12], kOne); @@ -492,16 +590,16 @@ void FDCT32x32_2D(const int16_t *input, step1[30] = _mm_add_epi16(step1[30], kOne); step1[31] = _mm_add_epi16(step1[31], kOne); - step2[ 0] = _mm_srai_epi16(step2[ 0], 2); - step2[ 1] = _mm_srai_epi16(step2[ 1], 2); - step2[ 2] = _mm_srai_epi16(step2[ 2], 2); - step2[ 3] = _mm_srai_epi16(step2[ 3], 2); - step2[ 4] = _mm_srai_epi16(step2[ 4], 2); - step2[ 5] = _mm_srai_epi16(step2[ 5], 2); - step2[ 6] = _mm_srai_epi16(step2[ 6], 2); - step2[ 7] = _mm_srai_epi16(step2[ 7], 2); - step2[ 8] = _mm_srai_epi16(step2[ 8], 2); - step2[ 9] = _mm_srai_epi16(step2[ 9], 2); + step2[0] = _mm_srai_epi16(step2[ 0], 2); + step2[1] = _mm_srai_epi16(step2[ 1], 2); + step2[2] = _mm_srai_epi16(step2[ 2], 2); + step2[3] = _mm_srai_epi16(step2[ 3], 2); + step2[4] = _mm_srai_epi16(step2[ 4], 2); + step2[5] = _mm_srai_epi16(step2[ 5], 2); + step2[6] = _mm_srai_epi16(step2[ 6], 2); + step2[7] = _mm_srai_epi16(step2[ 7], 2); + step2[8] = _mm_srai_epi16(step2[ 8], 2); + step2[9] = _mm_srai_epi16(step2[ 9], 2); step2[10] = _mm_srai_epi16(step2[10], 2); step2[11] = _mm_srai_epi16(step2[11], 2); step2[12] = _mm_srai_epi16(step2[12], 2); @@ -525,21 +623,33 @@ void FDCT32x32_2D(const int16_t *input, step1[30] = _mm_srai_epi16(step1[30], 2); step1[31] = _mm_srai_epi16(step1[31], 2); } -#endif +#endif // !FDCT32x32_HIGH_PRECISION #if FDCT32x32_HIGH_PRECISION if (pass == 0) { #endif // Stage 3 { - step3[0] = _mm_add_epi16(step2[(8 - 1)], step2[0]); - step3[1] = _mm_add_epi16(step2[(8 - 2)], step2[1]); - step3[2] = _mm_add_epi16(step2[(8 - 3)], step2[2]); - step3[3] = _mm_add_epi16(step2[(8 - 4)], step2[3]); - step3[4] = _mm_sub_epi16(step2[(8 - 5)], step2[4]); - step3[5] = _mm_sub_epi16(step2[(8 - 6)], step2[5]); - step3[6] = _mm_sub_epi16(step2[(8 - 7)], step2[6]); - step3[7] = _mm_sub_epi16(step2[(8 - 8)], step2[7]); + step3[0] = ADD_EPI16(step2[(8 - 1)], step2[0]); + step3[1] = ADD_EPI16(step2[(8 - 2)], step2[1]); + step3[2] = ADD_EPI16(step2[(8 - 3)], step2[2]); + step3[3] = ADD_EPI16(step2[(8 - 4)], step2[3]); + step3[4] = SUB_EPI16(step2[(8 - 5)], step2[4]); + step3[5] = SUB_EPI16(step2[(8 - 6)], step2[5]); + step3[6] = SUB_EPI16(step2[(8 - 7)], step2[6]); + step3[7] = SUB_EPI16(step2[(8 - 8)], step2[7]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(step3[0], step3[1], step3[2], + step3[3], step3[4], step3[5], + step3[6], step3[7]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]); @@ -576,40 +686,79 @@ void FDCT32x32_2D(const int16_t *input, step3[11] = _mm_packs_epi32(s3_11_6, s3_11_7); step3[12] = _mm_packs_epi32(s3_12_6, s3_12_7); step3[13] = _mm_packs_epi32(s3_13_6, s3_13_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(step3[10], step3[11], + step3[12], step3[13]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { - step3[16] = _mm_add_epi16(step2[23], step1[16]); - step3[17] = _mm_add_epi16(step2[22], step1[17]); - step3[18] = _mm_add_epi16(step2[21], step1[18]); - step3[19] = _mm_add_epi16(step2[20], step1[19]); - step3[20] = _mm_sub_epi16(step1[19], step2[20]); - step3[21] = _mm_sub_epi16(step1[18], step2[21]); - step3[22] = _mm_sub_epi16(step1[17], step2[22]); - step3[23] = _mm_sub_epi16(step1[16], step2[23]); - step3[24] = _mm_sub_epi16(step1[31], step2[24]); - step3[25] = _mm_sub_epi16(step1[30], step2[25]); - step3[26] = _mm_sub_epi16(step1[29], step2[26]); - step3[27] = _mm_sub_epi16(step1[28], step2[27]); - step3[28] = _mm_add_epi16(step2[27], step1[28]); - step3[29] = _mm_add_epi16(step2[26], step1[29]); - step3[30] = _mm_add_epi16(step2[25], step1[30]); - step3[31] = _mm_add_epi16(step2[24], step1[31]); + step3[16] = ADD_EPI16(step2[23], step1[16]); + step3[17] = ADD_EPI16(step2[22], step1[17]); + step3[18] = ADD_EPI16(step2[21], step1[18]); + step3[19] = ADD_EPI16(step2[20], step1[19]); + step3[20] = SUB_EPI16(step1[19], step2[20]); + step3[21] = SUB_EPI16(step1[18], step2[21]); + step3[22] = SUB_EPI16(step1[17], step2[22]); + step3[23] = SUB_EPI16(step1[16], step2[23]); + step3[24] = SUB_EPI16(step1[31], step2[24]); + step3[25] = SUB_EPI16(step1[30], step2[25]); + step3[26] = SUB_EPI16(step1[29], step2[26]); + step3[27] = SUB_EPI16(step1[28], step2[27]); + step3[28] = ADD_EPI16(step2[27], step1[28]); + step3[29] = ADD_EPI16(step2[26], step1[29]); + step3[30] = ADD_EPI16(step2[25], step1[30]); + step3[31] = ADD_EPI16(step2[24], step1[31]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x16( + step3[16], step3[17], step3[18], step3[19], + step3[20], step3[21], step3[22], step3[23], + step3[24], step3[25], step3[26], step3[27], + step3[28], step3[29], step3[30], step3[31]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } // Stage 4 { - step1[ 0] = _mm_add_epi16(step3[ 3], step3[ 0]); - step1[ 1] = _mm_add_epi16(step3[ 2], step3[ 1]); - step1[ 2] = _mm_sub_epi16(step3[ 1], step3[ 2]); - step1[ 3] = _mm_sub_epi16(step3[ 0], step3[ 3]); - step1[ 8] = _mm_add_epi16(step3[11], step2[ 8]); - step1[ 9] = _mm_add_epi16(step3[10], step2[ 9]); - step1[10] = _mm_sub_epi16(step2[ 9], step3[10]); - step1[11] = _mm_sub_epi16(step2[ 8], step3[11]); - step1[12] = _mm_sub_epi16(step2[15], step3[12]); - step1[13] = _mm_sub_epi16(step2[14], step3[13]); - step1[14] = _mm_add_epi16(step3[13], step2[14]); - step1[15] = _mm_add_epi16(step3[12], step2[15]); + step1[0] = ADD_EPI16(step3[ 3], step3[ 0]); + step1[1] = ADD_EPI16(step3[ 2], step3[ 1]); + step1[2] = SUB_EPI16(step3[ 1], step3[ 2]); + step1[3] = SUB_EPI16(step3[ 0], step3[ 3]); + step1[8] = ADD_EPI16(step3[11], step2[ 8]); + step1[9] = ADD_EPI16(step3[10], step2[ 9]); + step1[10] = SUB_EPI16(step2[ 9], step3[10]); + step1[11] = SUB_EPI16(step2[ 8], step3[11]); + step1[12] = SUB_EPI16(step2[15], step3[12]); + step1[13] = SUB_EPI16(step2[14], step3[13]); + step1[14] = ADD_EPI16(step3[13], step2[14]); + step1[15] = ADD_EPI16(step3[12], step2[15]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x16( + step1[0], step1[1], step1[2], step1[3], + step1[4], step1[5], step1[6], step1[7], + step1[8], step1[9], step1[10], step1[11], + step1[12], step1[13], step1[14], step1[15]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { const __m128i s1_05_0 = _mm_unpacklo_epi16(step3[6], step3[5]); @@ -630,6 +779,16 @@ void FDCT32x32_2D(const int16_t *input, // Combine step1[5] = _mm_packs_epi32(s1_05_6, s1_05_7); step1[6] = _mm_packs_epi32(s1_06_6, s1_06_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x2(step1[5], step1[6]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { const __m128i s1_18_0 = _mm_unpacklo_epi16(step3[18], step3[29]); @@ -698,13 +857,36 @@ void FDCT32x32_2D(const int16_t *input, step1[27] = _mm_packs_epi32(s1_27_6, s1_27_7); step1[28] = _mm_packs_epi32(s1_28_6, s1_28_7); step1[29] = _mm_packs_epi32(s1_29_6, s1_29_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(step1[18], step1[19], step1[20], + step1[21], step1[26], step1[27], + step1[28], step1[29]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } // Stage 5 { - step2[4] = _mm_add_epi16(step1[5], step3[4]); - step2[5] = _mm_sub_epi16(step3[4], step1[5]); - step2[6] = _mm_sub_epi16(step3[7], step1[6]); - step2[7] = _mm_add_epi16(step1[6], step3[7]); + step2[4] = ADD_EPI16(step1[5], step3[4]); + step2[5] = SUB_EPI16(step3[4], step1[5]); + step2[6] = SUB_EPI16(step3[7], step1[6]); + step2[7] = ADD_EPI16(step1[6], step3[7]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(step2[4], step2[5], + step2[6], step2[7]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { const __m128i out_00_0 = _mm_unpacklo_epi16(step1[0], step1[1]); @@ -741,6 +923,16 @@ void FDCT32x32_2D(const int16_t *input, out[16] = _mm_packs_epi32(out_16_6, out_16_7); out[ 8] = _mm_packs_epi32(out_08_6, out_08_7); out[24] = _mm_packs_epi32(out_24_6, out_24_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(out[0], out[16], out[8], out[24]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { const __m128i s2_09_0 = _mm_unpacklo_epi16(step1[ 9], step1[14]); @@ -777,24 +969,49 @@ void FDCT32x32_2D(const int16_t *input, step2[10] = _mm_packs_epi32(s2_10_6, s2_10_7); step2[13] = _mm_packs_epi32(s2_13_6, s2_13_7); step2[14] = _mm_packs_epi32(s2_14_6, s2_14_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(step2[9], step2[10], + step2[13], step2[14]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { - step2[16] = _mm_add_epi16(step1[19], step3[16]); - step2[17] = _mm_add_epi16(step1[18], step3[17]); - step2[18] = _mm_sub_epi16(step3[17], step1[18]); - step2[19] = _mm_sub_epi16(step3[16], step1[19]); - step2[20] = _mm_sub_epi16(step3[23], step1[20]); - step2[21] = _mm_sub_epi16(step3[22], step1[21]); - step2[22] = _mm_add_epi16(step1[21], step3[22]); - step2[23] = _mm_add_epi16(step1[20], step3[23]); - step2[24] = _mm_add_epi16(step1[27], step3[24]); - step2[25] = _mm_add_epi16(step1[26], step3[25]); - step2[26] = _mm_sub_epi16(step3[25], step1[26]); - step2[27] = _mm_sub_epi16(step3[24], step1[27]); - step2[28] = _mm_sub_epi16(step3[31], step1[28]); - step2[29] = _mm_sub_epi16(step3[30], step1[29]); - step2[30] = _mm_add_epi16(step1[29], step3[30]); - step2[31] = _mm_add_epi16(step1[28], step3[31]); + step2[16] = ADD_EPI16(step1[19], step3[16]); + step2[17] = ADD_EPI16(step1[18], step3[17]); + step2[18] = SUB_EPI16(step3[17], step1[18]); + step2[19] = SUB_EPI16(step3[16], step1[19]); + step2[20] = SUB_EPI16(step3[23], step1[20]); + step2[21] = SUB_EPI16(step3[22], step1[21]); + step2[22] = ADD_EPI16(step1[21], step3[22]); + step2[23] = ADD_EPI16(step1[20], step3[23]); + step2[24] = ADD_EPI16(step1[27], step3[24]); + step2[25] = ADD_EPI16(step1[26], step3[25]); + step2[26] = SUB_EPI16(step3[25], step1[26]); + step2[27] = SUB_EPI16(step3[24], step1[27]); + step2[28] = SUB_EPI16(step3[31], step1[28]); + step2[29] = SUB_EPI16(step3[30], step1[29]); + step2[30] = ADD_EPI16(step1[29], step3[30]); + step2[31] = ADD_EPI16(step1[28], step3[31]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x16( + step2[16], step2[17], step2[18], step2[19], + step2[20], step2[21], step2[22], step2[23], + step2[24], step2[25], step2[26], step2[27], + step2[28], step2[29], step2[30], step2[31]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } // Stage 6 { @@ -832,20 +1049,42 @@ void FDCT32x32_2D(const int16_t *input, const __m128i out_28_6 = _mm_srai_epi32(out_28_4, DCT_CONST_BITS); const __m128i out_28_7 = _mm_srai_epi32(out_28_5, DCT_CONST_BITS); // Combine - out[ 4] = _mm_packs_epi32(out_04_6, out_04_7); + out[4] = _mm_packs_epi32(out_04_6, out_04_7); out[20] = _mm_packs_epi32(out_20_6, out_20_7); out[12] = _mm_packs_epi32(out_12_6, out_12_7); out[28] = _mm_packs_epi32(out_28_6, out_28_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(out[4], out[20], out[12], out[28]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { - step3[ 8] = _mm_add_epi16(step2[ 9], step1[ 8]); - step3[ 9] = _mm_sub_epi16(step1[ 8], step2[ 9]); - step3[10] = _mm_sub_epi16(step1[11], step2[10]); - step3[11] = _mm_add_epi16(step2[10], step1[11]); - step3[12] = _mm_add_epi16(step2[13], step1[12]); - step3[13] = _mm_sub_epi16(step1[12], step2[13]); - step3[14] = _mm_sub_epi16(step1[15], step2[14]); - step3[15] = _mm_add_epi16(step2[14], step1[15]); + step3[8] = ADD_EPI16(step2[ 9], step1[ 8]); + step3[9] = SUB_EPI16(step1[ 8], step2[ 9]); + step3[10] = SUB_EPI16(step1[11], step2[10]); + step3[11] = ADD_EPI16(step2[10], step1[11]); + step3[12] = ADD_EPI16(step2[13], step1[12]); + step3[13] = SUB_EPI16(step1[12], step2[13]); + step3[14] = SUB_EPI16(step1[15], step2[14]); + step3[15] = ADD_EPI16(step2[14], step1[15]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(step3[8], step3[9], step3[10], + step3[11], step3[12], step3[13], + step3[14], step3[15]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { const __m128i s3_17_0 = _mm_unpacklo_epi16(step2[17], step2[30]); @@ -915,6 +1154,18 @@ void FDCT32x32_2D(const int16_t *input, step3[26] = _mm_packs_epi32(s3_26_6, s3_26_7); step3[29] = _mm_packs_epi32(s3_29_6, s3_29_7); step3[30] = _mm_packs_epi32(s3_30_6, s3_30_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(step3[17], step3[18], step3[21], + step3[22], step3[25], step3[26], + step3[29], step3[30]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } // Stage 7 { @@ -984,24 +1235,49 @@ void FDCT32x32_2D(const int16_t *input, out[22] = _mm_packs_epi32(out_22_6, out_22_7); out[14] = _mm_packs_epi32(out_14_6, out_14_7); out[30] = _mm_packs_epi32(out_30_6, out_30_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(out[2], out[18], out[10], out[26], + out[6], out[22], out[14], out[30]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { - step1[16] = _mm_add_epi16(step3[17], step2[16]); - step1[17] = _mm_sub_epi16(step2[16], step3[17]); - step1[18] = _mm_sub_epi16(step2[19], step3[18]); - step1[19] = _mm_add_epi16(step3[18], step2[19]); - step1[20] = _mm_add_epi16(step3[21], step2[20]); - step1[21] = _mm_sub_epi16(step2[20], step3[21]); - step1[22] = _mm_sub_epi16(step2[23], step3[22]); - step1[23] = _mm_add_epi16(step3[22], step2[23]); - step1[24] = _mm_add_epi16(step3[25], step2[24]); - step1[25] = _mm_sub_epi16(step2[24], step3[25]); - step1[26] = _mm_sub_epi16(step2[27], step3[26]); - step1[27] = _mm_add_epi16(step3[26], step2[27]); - step1[28] = _mm_add_epi16(step3[29], step2[28]); - step1[29] = _mm_sub_epi16(step2[28], step3[29]); - step1[30] = _mm_sub_epi16(step2[31], step3[30]); - step1[31] = _mm_add_epi16(step3[30], step2[31]); + step1[16] = ADD_EPI16(step3[17], step2[16]); + step1[17] = SUB_EPI16(step2[16], step3[17]); + step1[18] = SUB_EPI16(step2[19], step3[18]); + step1[19] = ADD_EPI16(step3[18], step2[19]); + step1[20] = ADD_EPI16(step3[21], step2[20]); + step1[21] = SUB_EPI16(step2[20], step3[21]); + step1[22] = SUB_EPI16(step2[23], step3[22]); + step1[23] = ADD_EPI16(step3[22], step2[23]); + step1[24] = ADD_EPI16(step3[25], step2[24]); + step1[25] = SUB_EPI16(step2[24], step3[25]); + step1[26] = SUB_EPI16(step2[27], step3[26]); + step1[27] = ADD_EPI16(step3[26], step2[27]); + step1[28] = ADD_EPI16(step3[29], step2[28]); + step1[29] = SUB_EPI16(step2[28], step3[29]); + step1[30] = SUB_EPI16(step2[31], step3[30]); + step1[31] = ADD_EPI16(step3[30], step2[31]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x16( + step1[16], step1[17], step1[18], step1[19], + step1[20], step1[21], step1[22], step1[23], + step1[24], step1[25], step1[26], step1[27], + step1[28], step1[29], step1[30], step1[31]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } // Final stage --- outputs indices are bit-reversed. { @@ -1071,6 +1347,17 @@ void FDCT32x32_2D(const int16_t *input, out[23] = _mm_packs_epi32(out_23_6, out_23_7); out[15] = _mm_packs_epi32(out_15_6, out_15_7); out[31] = _mm_packs_epi32(out_31_6, out_31_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(out[1], out[17], out[9], out[25], + out[7], out[23], out[15], out[31]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { const __m128i out_05_0 = _mm_unpacklo_epi16(step1[20], step1[27]); @@ -1139,6 +1426,17 @@ void FDCT32x32_2D(const int16_t *input, out[19] = _mm_packs_epi32(out_19_6, out_19_7); out[11] = _mm_packs_epi32(out_11_6, out_11_7); out[27] = _mm_packs_epi32(out_27_6, out_27_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(out[5], out[21], out[13], out[29], + out[3], out[19], out[11], out[27]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } #if FDCT32x32_HIGH_PRECISION } else { @@ -1390,15 +1688,22 @@ void FDCT32x32_2D(const int16_t *input, // TODO(jingning): manually inline k_madd_epi32_ to further hide // instruction latency. - v[ 0] = k_madd_epi32(u[0], k32_p16_m16); - v[ 1] = k_madd_epi32(u[1], k32_p16_m16); - v[ 2] = k_madd_epi32(u[2], k32_p16_m16); - v[ 3] = k_madd_epi32(u[3], k32_p16_m16); - v[ 4] = k_madd_epi32(u[0], k32_p16_p16); - v[ 5] = k_madd_epi32(u[1], k32_p16_p16); - v[ 6] = k_madd_epi32(u[2], k32_p16_p16); - v[ 7] = k_madd_epi32(u[3], k32_p16_p16); - + v[0] = k_madd_epi32(u[0], k32_p16_m16); + v[1] = k_madd_epi32(u[1], k32_p16_m16); + v[2] = k_madd_epi32(u[2], k32_p16_m16); + v[3] = k_madd_epi32(u[3], k32_p16_m16); + v[4] = k_madd_epi32(u[0], k32_p16_p16); + v[5] = k_madd_epi32(u[1], k32_p16_p16); + v[6] = k_madd_epi32(u[2], k32_p16_p16); + v[7] = k_madd_epi32(u[3], k32_p16_p16); +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_8(v[0], v[1], v[2], v[3], v[4], v[5], + v[6], v[7], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH u[0] = k_packs_epi64(v[0], v[1]); u[1] = k_packs_epi64(v[2], v[3]); u[2] = k_packs_epi64(v[4], v[5]); @@ -1469,6 +1774,17 @@ void FDCT32x32_2D(const int16_t *input, v[30] = k_madd_epi32(u[ 2], k32_p24_p08); v[31] = k_madd_epi32(u[ 3], k32_p24_p08); +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_32( + v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7], + v[8], v[9], v[10], v[11], v[12], v[13], v[14], v[15], + v[16], v[17], v[18], v[19], v[20], v[21], v[22], v[23], + v[24], v[25], v[26], v[27], v[28], v[29], v[30], v[31], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH u[ 0] = k_packs_epi64(v[ 0], v[ 1]); u[ 1] = k_packs_epi64(v[ 2], v[ 3]); u[ 2] = k_packs_epi64(v[ 4], v[ 5]); @@ -1565,6 +1881,15 @@ void FDCT32x32_2D(const int16_t *input, v[14] = k_madd_epi32(u[6], k32_m08_p24); v[15] = k_madd_epi32(u[7], k32_m08_p24); +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_16( + v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7], + v[8], v[9], v[10], v[11], v[12], v[13], v[14], v[15], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH u[0] = k_packs_epi64(v[0], v[1]); u[1] = k_packs_epi64(v[2], v[3]); u[2] = k_packs_epi64(v[4], v[5]); @@ -1633,6 +1958,13 @@ void FDCT32x32_2D(const int16_t *input, out[16] = _mm_packs_epi32(u[2], u[3]); out[ 8] = _mm_packs_epi32(u[4], u[5]); out[24] = _mm_packs_epi32(u[6], u[7]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(out[0], out[16], out[8], out[24]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64); @@ -1665,6 +1997,15 @@ void FDCT32x32_2D(const int16_t *input, v[14] = k_madd_epi32(u[2], k32_p24_p08); v[15] = k_madd_epi32(u[3], k32_p24_p08); +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_16( + v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7], + v[8], v[9], v[10], v[11], v[12], v[13], v[14], v[15], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH u[0] = k_packs_epi64(v[0], v[1]); u[1] = k_packs_epi64(v[2], v[3]); u[2] = k_packs_epi64(v[4], v[5]); @@ -1767,6 +2108,15 @@ void FDCT32x32_2D(const int16_t *input, v[14] = k_madd_epi32(u[14], k32_m04_p28); v[15] = k_madd_epi32(u[15], k32_m04_p28); +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_16( + v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7], + v[8], v[9], v[10], v[11], v[12], v[13], v[14], v[15], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH u[0] = k_packs_epi64(v[0], v[1]); u[1] = k_packs_epi64(v[2], v[3]); u[2] = k_packs_epi64(v[4], v[5]); @@ -1834,6 +2184,13 @@ void FDCT32x32_2D(const int16_t *input, out[20] = _mm_packs_epi32(u[2], u[3]); out[12] = _mm_packs_epi32(u[4], u[5]); out[28] = _mm_packs_epi32(u[6], u[7]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(out[4], out[20], out[12], out[28]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { lstep3[16] = _mm_add_epi32(lstep2[18], lstep1[16]); @@ -1912,6 +2269,17 @@ void FDCT32x32_2D(const int16_t *input, v[30] = k_madd_epi32(u[ 2], k32_p28_p04); v[31] = k_madd_epi32(u[ 3], k32_p28_p04); +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_32( + v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7], + v[8], v[9], v[10], v[11], v[12], v[13], v[14], v[15], + v[16], v[17], v[18], v[19], v[20], v[21], v[22], v[23], + v[24], v[25], v[26], v[27], v[28], v[29], v[30], v[31], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH u[ 0] = k_packs_epi64(v[ 0], v[ 1]); u[ 1] = k_packs_epi64(v[ 2], v[ 3]); u[ 2] = k_packs_epi64(v[ 4], v[ 5]); @@ -2024,6 +2392,17 @@ void FDCT32x32_2D(const int16_t *input, v[30] = k_madd_epi32(u[ 2], k32_m02_p30); v[31] = k_madd_epi32(u[ 3], k32_m02_p30); +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_32( + v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7], + v[8], v[9], v[10], v[11], v[12], v[13], v[14], v[15], + v[16], v[17], v[18], v[19], v[20], v[21], v[22], v[23], + v[24], v[25], v[26], v[27], v[28], v[29], v[30], v[31], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH u[ 0] = k_packs_epi64(v[ 0], v[ 1]); u[ 1] = k_packs_epi64(v[ 2], v[ 3]); u[ 2] = k_packs_epi64(v[ 4], v[ 5]); @@ -2151,6 +2530,14 @@ void FDCT32x32_2D(const int16_t *input, out[22] = _mm_packs_epi32(u[10], u[11]); out[14] = _mm_packs_epi32(u[12], u[13]); out[30] = _mm_packs_epi32(u[14], u[15]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(out[2], out[18], out[10], out[26], + out[6], out[22], out[14], out[30]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { lstep1[32] = _mm_add_epi32(lstep3[34], lstep2[32]); @@ -2247,6 +2634,17 @@ void FDCT32x32_2D(const int16_t *input, v[30] = k_madd_epi32(u[ 2], k32_m01_p31); v[31] = k_madd_epi32(u[ 3], k32_m01_p31); +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_32( + v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7], + v[8], v[9], v[10], v[11], v[12], v[13], v[14], v[15], + v[16], v[17], v[18], v[19], v[20], v[21], v[22], v[23], + v[24], v[25], v[26], v[27], v[28], v[29], v[30], v[31], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH u[ 0] = k_packs_epi64(v[ 0], v[ 1]); u[ 1] = k_packs_epi64(v[ 2], v[ 3]); u[ 2] = k_packs_epi64(v[ 4], v[ 5]); @@ -2374,6 +2772,14 @@ void FDCT32x32_2D(const int16_t *input, out[23] = _mm_packs_epi32(u[10], u[11]); out[15] = _mm_packs_epi32(u[12], u[13]); out[31] = _mm_packs_epi32(u[14], u[15]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(out[1], out[17], out[9], out[25], + out[7], out[23], out[15], out[31]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } { const __m128i k32_p27_p05 = pair_set_epi32(cospi_27_64, cospi_5_64); @@ -2435,6 +2841,17 @@ void FDCT32x32_2D(const int16_t *input, v[30] = k_madd_epi32(u[ 2], k32_m05_p27); v[31] = k_madd_epi32(u[ 3], k32_m05_p27); +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_32( + v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7], + v[8], v[9], v[10], v[11], v[12], v[13], v[14], v[15], + v[16], v[17], v[18], v[19], v[20], v[21], v[22], v[23], + v[24], v[25], v[26], v[27], v[28], v[29], v[30], v[31], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH u[ 0] = k_packs_epi64(v[ 0], v[ 1]); u[ 1] = k_packs_epi64(v[ 2], v[ 3]); u[ 2] = k_packs_epi64(v[ 4], v[ 5]); @@ -2562,18 +2979,22 @@ void FDCT32x32_2D(const int16_t *input, out[19] = _mm_packs_epi32(u[10], u[11]); out[11] = _mm_packs_epi32(u[12], u[13]); out[27] = _mm_packs_epi32(u[14], u[15]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(out[5], out[21], out[13], out[29], + out[3], out[19], out[11], out[27]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH } } -#endif +#endif // FDCT32x32_HIGH_PRECISION // Transpose the results, do it as four 8x8 transposes. { int transpose_block; - int16_t *output; - if (0 == pass) { - output = &intermediate[column_start * 32]; - } else { - output = &output_org[column_start * 32]; - } + int16_t *output0 = &intermediate[column_start * 32]; + tran_low_t *output1 = &output_org[column_start * 32]; for (transpose_block = 0; transpose_block < 4; ++transpose_block) { __m128i *this_out = &out[8 * transpose_block]; // 00 01 02 03 04 05 06 07 @@ -2674,18 +3095,36 @@ void FDCT32x32_2D(const int16_t *input, } // Note: even though all these stores are aligned, using the aligned // intrinsic make the code slightly slower. - _mm_storeu_si128((__m128i *)(output + 0 * 32), tr2_0); - _mm_storeu_si128((__m128i *)(output + 1 * 32), tr2_1); - _mm_storeu_si128((__m128i *)(output + 2 * 32), tr2_2); - _mm_storeu_si128((__m128i *)(output + 3 * 32), tr2_3); - _mm_storeu_si128((__m128i *)(output + 4 * 32), tr2_4); - _mm_storeu_si128((__m128i *)(output + 5 * 32), tr2_5); - _mm_storeu_si128((__m128i *)(output + 6 * 32), tr2_6); - _mm_storeu_si128((__m128i *)(output + 7 * 32), tr2_7); - // Process next 8x8 - output += 8; + if (pass == 0) { + _mm_storeu_si128((__m128i *)(output0 + 0 * 32), tr2_0); + _mm_storeu_si128((__m128i *)(output0 + 1 * 32), tr2_1); + _mm_storeu_si128((__m128i *)(output0 + 2 * 32), tr2_2); + _mm_storeu_si128((__m128i *)(output0 + 3 * 32), tr2_3); + _mm_storeu_si128((__m128i *)(output0 + 4 * 32), tr2_4); + _mm_storeu_si128((__m128i *)(output0 + 5 * 32), tr2_5); + _mm_storeu_si128((__m128i *)(output0 + 6 * 32), tr2_6); + _mm_storeu_si128((__m128i *)(output0 + 7 * 32), tr2_7); + // Process next 8x8 + output0 += 8; + } else { + storeu_output(tr2_0, (output1 + 0 * 32)); + storeu_output(tr2_1, (output1 + 1 * 32)); + storeu_output(tr2_2, (output1 + 2 * 32)); + storeu_output(tr2_3, (output1 + 3 * 32)); + storeu_output(tr2_4, (output1 + 4 * 32)); + storeu_output(tr2_5, (output1 + 5 * 32)); + storeu_output(tr2_6, (output1 + 6 * 32)); + storeu_output(tr2_7, (output1 + 7 * 32)); + // Process next 8x8 + output1 += 8; + } } } } } } // NOLINT + +#undef ADD_EPI16 +#undef SUB_EPI16 +#undef HIGH_FDCT32x32_2D_C +#undef HIGH_FDCT32x32_2D_ROWS_C diff --git a/vp9/encoder/x86/vp9_dct_impl_sse2.c b/vp9/encoder/x86/vp9_dct_impl_sse2.c new file mode 100644 index 000000000..3fdde83da --- /dev/null +++ b/vp9/encoder/x86/vp9_dct_impl_sse2.c @@ -0,0 +1,1011 @@ +/* + * Copyright (c) 2014 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include <emmintrin.h> // SSE2 +#include "vp9/common/vp9_idct.h" // for cospi constants +#include "vp9/encoder/vp9_dct.h" +#include "vp9/encoder/x86/vp9_dct_sse2.h" +#include "vpx_ports/mem.h" + +#if DCT_HIGH_BIT_DEPTH +#define ADD_EPI16 _mm_adds_epi16 +#define SUB_EPI16 _mm_subs_epi16 + +#else +#define ADD_EPI16 _mm_add_epi16 +#define SUB_EPI16 _mm_sub_epi16 +#endif + +void FDCT4x4_2D(const int16_t *input, tran_low_t *output, int stride) { + // This 2D transform implements 4 vertical 1D transforms followed + // by 4 horizontal 1D transforms. The multiplies and adds are as given + // by Chen, Smith and Fralick ('77). The commands for moving the data + // around have been minimized by hand. + // For the purposes of the comments, the 16 inputs are referred to at i0 + // through iF (in raster order), intermediate variables are a0, b0, c0 + // through f, and correspond to the in-place computations mapped to input + // locations. The outputs, o0 through oF are labeled according to the + // output locations. + + // Constants + // These are the coefficients used for the multiplies. + // In the comments, pN means cos(N pi /64) and mN is -cos(N pi /64), + // where cospi_N_64 = cos(N pi /64) + const __m128i k__cospi_A = _mm_setr_epi16(cospi_16_64, cospi_16_64, + cospi_16_64, cospi_16_64, + cospi_16_64, -cospi_16_64, + cospi_16_64, -cospi_16_64); + const __m128i k__cospi_B = _mm_setr_epi16(cospi_16_64, -cospi_16_64, + cospi_16_64, -cospi_16_64, + cospi_16_64, cospi_16_64, + cospi_16_64, cospi_16_64); + const __m128i k__cospi_C = _mm_setr_epi16(cospi_8_64, cospi_24_64, + cospi_8_64, cospi_24_64, + cospi_24_64, -cospi_8_64, + cospi_24_64, -cospi_8_64); + const __m128i k__cospi_D = _mm_setr_epi16(cospi_24_64, -cospi_8_64, + cospi_24_64, -cospi_8_64, + cospi_8_64, cospi_24_64, + cospi_8_64, cospi_24_64); + const __m128i k__cospi_E = _mm_setr_epi16(cospi_16_64, cospi_16_64, + cospi_16_64, cospi_16_64, + cospi_16_64, cospi_16_64, + cospi_16_64, cospi_16_64); + const __m128i k__cospi_F = _mm_setr_epi16(cospi_16_64, -cospi_16_64, + cospi_16_64, -cospi_16_64, + cospi_16_64, -cospi_16_64, + cospi_16_64, -cospi_16_64); + const __m128i k__cospi_G = _mm_setr_epi16(cospi_8_64, cospi_24_64, + cospi_8_64, cospi_24_64, + -cospi_8_64, -cospi_24_64, + -cospi_8_64, -cospi_24_64); + const __m128i k__cospi_H = _mm_setr_epi16(cospi_24_64, -cospi_8_64, + cospi_24_64, -cospi_8_64, + -cospi_24_64, cospi_8_64, + -cospi_24_64, cospi_8_64); + + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); + // This second rounding constant saves doing some extra adds at the end + const __m128i k__DCT_CONST_ROUNDING2 = _mm_set1_epi32(DCT_CONST_ROUNDING + +(DCT_CONST_ROUNDING << 1)); + const int DCT_CONST_BITS2 = DCT_CONST_BITS+2; + const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1); + const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0); + __m128i in0, in1; +#if DCT_HIGH_BIT_DEPTH + __m128i cmp0, cmp1; + int test, overflow; +#endif + + // Load inputs. + in0 = _mm_loadl_epi64((const __m128i *)(input + 0 * stride)); + in1 = _mm_loadl_epi64((const __m128i *)(input + 1 * stride)); + in1 = _mm_unpacklo_epi64(in1, _mm_loadl_epi64((const __m128i *) + (input + 2 * stride))); + in0 = _mm_unpacklo_epi64(in0, _mm_loadl_epi64((const __m128i *) + (input + 3 * stride))); + // in0 = [i0 i1 i2 i3 iC iD iE iF] + // in1 = [i4 i5 i6 i7 i8 i9 iA iB] +#if DCT_HIGH_BIT_DEPTH + // Check inputs small enough to use optimised code + cmp0 = _mm_xor_si128(_mm_cmpgt_epi16(in0, _mm_set1_epi16(0x3ff)), + _mm_cmplt_epi16(in0, _mm_set1_epi16(0xfc00))); + cmp1 = _mm_xor_si128(_mm_cmpgt_epi16(in1, _mm_set1_epi16(0x3ff)), + _mm_cmplt_epi16(in1, _mm_set1_epi16(0xfc00))); + test = _mm_movemask_epi8(_mm_or_si128(cmp0, cmp1)); + if (test) { + vp9_highbd_fdct4x4_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + + // multiply by 16 to give some extra precision + in0 = _mm_slli_epi16(in0, 4); + in1 = _mm_slli_epi16(in1, 4); + // if (i == 0 && input[0]) input[0] += 1; + // add 1 to the upper left pixel if it is non-zero, which helps reduce + // the round-trip error + { + // The mask will only contain whether the first value is zero, all + // other comparison will fail as something shifted by 4 (above << 4) + // can never be equal to one. To increment in the non-zero case, we + // add the mask and one for the first element: + // - if zero, mask = -1, v = v - 1 + 1 = v + // - if non-zero, mask = 0, v = v + 0 + 1 = v + 1 + __m128i mask = _mm_cmpeq_epi16(in0, k__nonzero_bias_a); + in0 = _mm_add_epi16(in0, mask); + in0 = _mm_add_epi16(in0, k__nonzero_bias_b); + } + // There are 4 total stages, alternating between an add/subtract stage + // followed by an multiply-and-add stage. + { + // Stage 1: Add/subtract + + // in0 = [i0 i1 i2 i3 iC iD iE iF] + // in1 = [i4 i5 i6 i7 i8 i9 iA iB] + const __m128i r0 = _mm_unpacklo_epi16(in0, in1); + const __m128i r1 = _mm_unpackhi_epi16(in0, in1); + // r0 = [i0 i4 i1 i5 i2 i6 i3 i7] + // r1 = [iC i8 iD i9 iE iA iF iB] + const __m128i r2 = _mm_shuffle_epi32(r0, 0xB4); + const __m128i r3 = _mm_shuffle_epi32(r1, 0xB4); + // r2 = [i0 i4 i1 i5 i3 i7 i2 i6] + // r3 = [iC i8 iD i9 iF iB iE iA] + + const __m128i t0 = _mm_add_epi16(r2, r3); + const __m128i t1 = _mm_sub_epi16(r2, r3); + // t0 = [a0 a4 a1 a5 a3 a7 a2 a6] + // t1 = [aC a8 aD a9 aF aB aE aA] + + // Stage 2: multiply by constants (which gets us into 32 bits). + // The constants needed here are: + // k__cospi_A = [p16 p16 p16 p16 p16 m16 p16 m16] + // k__cospi_B = [p16 m16 p16 m16 p16 p16 p16 p16] + // k__cospi_C = [p08 p24 p08 p24 p24 m08 p24 m08] + // k__cospi_D = [p24 m08 p24 m08 p08 p24 p08 p24] + const __m128i u0 = _mm_madd_epi16(t0, k__cospi_A); + const __m128i u2 = _mm_madd_epi16(t0, k__cospi_B); + const __m128i u1 = _mm_madd_epi16(t1, k__cospi_C); + const __m128i u3 = _mm_madd_epi16(t1, k__cospi_D); + // Then add and right-shift to get back to 16-bit range + const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); + const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); + const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); + const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); + const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); + const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); + const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); + const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); + // w0 = [b0 b1 b7 b6] + // w1 = [b8 b9 bF bE] + // w2 = [b4 b5 b3 b2] + // w3 = [bC bD bB bA] + const __m128i x0 = _mm_packs_epi32(w0, w1); + const __m128i x1 = _mm_packs_epi32(w2, w3); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x2(x0, x1); + if (overflow) { + vp9_highbd_fdct4x4_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // x0 = [b0 b1 b7 b6 b8 b9 bF bE] + // x1 = [b4 b5 b3 b2 bC bD bB bA] + in0 = _mm_shuffle_epi32(x0, 0xD8); + in1 = _mm_shuffle_epi32(x1, 0x8D); + // in0 = [b0 b1 b8 b9 b7 b6 bF bE] + // in1 = [b3 b2 bB bA b4 b5 bC bD] + } + { + // vertical DCTs finished. Now we do the horizontal DCTs. + // Stage 3: Add/subtract + + const __m128i t0 = ADD_EPI16(in0, in1); + const __m128i t1 = SUB_EPI16(in0, in1); + // t0 = [c0 c1 c8 c9 c4 c5 cC cD] + // t1 = [c3 c2 cB cA -c7 -c6 -cF -cE] +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x2(t0, t1); + if (overflow) { + vp9_highbd_fdct4x4_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + + // Stage 4: multiply by constants (which gets us into 32 bits). + { + // The constants needed here are: + // k__cospi_E = [p16 p16 p16 p16 p16 p16 p16 p16] + // k__cospi_F = [p16 m16 p16 m16 p16 m16 p16 m16] + // k__cospi_G = [p08 p24 p08 p24 m08 m24 m08 m24] + // k__cospi_H = [p24 m08 p24 m08 m24 p08 m24 p08] + const __m128i u0 = _mm_madd_epi16(t0, k__cospi_E); + const __m128i u1 = _mm_madd_epi16(t0, k__cospi_F); + const __m128i u2 = _mm_madd_epi16(t1, k__cospi_G); + const __m128i u3 = _mm_madd_epi16(t1, k__cospi_H); + // Then add and right-shift to get back to 16-bit range + // but this combines the final right-shift as well to save operations + // This unusual rounding operations is to maintain bit-accurate + // compatibility with the c version of this function which has two + // rounding steps in a row. + const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING2); + const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING2); + const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING2); + const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING2); + const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS2); + const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS2); + const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS2); + const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS2); + // w0 = [o0 o4 o8 oC] + // w1 = [o2 o6 oA oE] + // w2 = [o1 o5 o9 oD] + // w3 = [o3 o7 oB oF] + // remember the o's are numbered according to the correct output location + const __m128i x0 = _mm_packs_epi32(w0, w1); + const __m128i x1 = _mm_packs_epi32(w2, w3); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x2(x0, x1); + if (overflow) { + vp9_highbd_fdct4x4_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + { + // x0 = [o0 o4 o8 oC o2 o6 oA oE] + // x1 = [o1 o5 o9 oD o3 o7 oB oF] + const __m128i y0 = _mm_unpacklo_epi16(x0, x1); + const __m128i y1 = _mm_unpackhi_epi16(x0, x1); + // y0 = [o0 o1 o4 o5 o8 o9 oC oD] + // y1 = [o2 o3 o6 o7 oA oB oE oF] + in0 = _mm_unpacklo_epi32(y0, y1); + // in0 = [o0 o1 o2 o3 o4 o5 o6 o7] + in1 = _mm_unpackhi_epi32(y0, y1); + // in1 = [o8 o9 oA oB oC oD oE oF] + } + } + } + // Post-condition (v + 1) >> 2 is now incorporated into previous + // add and right-shift commands. Only 2 store instructions needed + // because we are using the fact that 1/3 are stored just after 0/2. + storeu_output(in0, output + 0 * 4); + storeu_output(in1, output + 2 * 4); +} + + +void FDCT8x8_2D(const int16_t *input, tran_low_t *output, int stride) { + int pass; + // Constants + // When we use them, in one case, they are all the same. In all others + // it's a pair of them that we need to repeat four times. This is done + // by constructing the 32 bit constant corresponding to that pair. + const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); + const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); + const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); + const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); +#if DCT_HIGH_BIT_DEPTH + int overflow; +#endif + // Load input + __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride)); + __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride)); + __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride)); + __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride)); + __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride)); + __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride)); + __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride)); + __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride)); + // Pre-condition input (shift by two) + in0 = _mm_slli_epi16(in0, 2); + in1 = _mm_slli_epi16(in1, 2); + in2 = _mm_slli_epi16(in2, 2); + in3 = _mm_slli_epi16(in3, 2); + in4 = _mm_slli_epi16(in4, 2); + in5 = _mm_slli_epi16(in5, 2); + in6 = _mm_slli_epi16(in6, 2); + in7 = _mm_slli_epi16(in7, 2); + + // We do two passes, first the columns, then the rows. The results of the + // first pass are transposed so that the same column code can be reused. The + // results of the second pass are also transposed so that the rows (processed + // as columns) are put back in row positions. + for (pass = 0; pass < 2; pass++) { + // To store results of each pass before the transpose. + __m128i res0, res1, res2, res3, res4, res5, res6, res7; + // Add/subtract + const __m128i q0 = ADD_EPI16(in0, in7); + const __m128i q1 = ADD_EPI16(in1, in6); + const __m128i q2 = ADD_EPI16(in2, in5); + const __m128i q3 = ADD_EPI16(in3, in4); + const __m128i q4 = SUB_EPI16(in3, in4); + const __m128i q5 = SUB_EPI16(in2, in5); + const __m128i q6 = SUB_EPI16(in1, in6); + const __m128i q7 = SUB_EPI16(in0, in7); +#if DCT_HIGH_BIT_DEPTH + if (pass == 1) { + overflow = check_epi16_overflow_x8(q0, q1, q2, q3, q4, q5, q6, q7); + if (overflow) { + vp9_highbd_fdct8x8_c(input, output, stride); + return; + } + } +#endif // DCT_HIGH_BIT_DEPTH + // Work on first four results + { + // Add/subtract + const __m128i r0 = ADD_EPI16(q0, q3); + const __m128i r1 = ADD_EPI16(q1, q2); + const __m128i r2 = SUB_EPI16(q1, q2); + const __m128i r3 = SUB_EPI16(q0, q3); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(r0, r1, r2, r3); + if (overflow) { + vp9_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // Interleave to do the multiply by constants which gets us into 32bits + { + const __m128i t0 = _mm_unpacklo_epi16(r0, r1); + const __m128i t1 = _mm_unpackhi_epi16(r0, r1); + const __m128i t2 = _mm_unpacklo_epi16(r2, r3); + const __m128i t3 = _mm_unpackhi_epi16(r2, r3); + const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); + const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); + const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); + const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16); + const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); + const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08); + const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); + const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24); + // dct_const_round_shift + const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); + const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); + const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); + const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); + const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); + const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); + const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); + const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); + const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); + const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); + const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); + const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); + const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); + const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); + const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); + const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); + // Combine + res0 = _mm_packs_epi32(w0, w1); + res4 = _mm_packs_epi32(w2, w3); + res2 = _mm_packs_epi32(w4, w5); + res6 = _mm_packs_epi32(w6, w7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(res0, res4, res2, res6); + if (overflow) { + vp9_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } + // Work on next four results + { + // Interleave to do the multiply by constants which gets us into 32bits + const __m128i d0 = _mm_unpacklo_epi16(q6, q5); + const __m128i d1 = _mm_unpackhi_epi16(q6, q5); + const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16); + const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16); + const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16); + const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16); + // dct_const_round_shift + const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING); + const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING); + const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING); + const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING); + const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS); + const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS); + const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS); + const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS); + // Combine + const __m128i r0 = _mm_packs_epi32(s0, s1); + const __m128i r1 = _mm_packs_epi32(s2, s3); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x2(r0, r1); + if (overflow) { + vp9_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + { + // Add/subtract + const __m128i x0 = ADD_EPI16(q4, r0); + const __m128i x1 = SUB_EPI16(q4, r0); + const __m128i x2 = SUB_EPI16(q7, r1); + const __m128i x3 = ADD_EPI16(q7, r1); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(x0, x1, x2, x3); + if (overflow) { + vp9_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // Interleave to do the multiply by constants which gets us into 32bits + { + const __m128i t0 = _mm_unpacklo_epi16(x0, x3); + const __m128i t1 = _mm_unpackhi_epi16(x0, x3); + const __m128i t2 = _mm_unpacklo_epi16(x1, x2); + const __m128i t3 = _mm_unpackhi_epi16(x1, x2); + const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04); + const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04); + const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28); + const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28); + const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20); + const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20); + const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12); + const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12); + // dct_const_round_shift + const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); + const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); + const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); + const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); + const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); + const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); + const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); + const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); + const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); + const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); + const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); + const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); + const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); + const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); + const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); + const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); + // Combine + res1 = _mm_packs_epi32(w0, w1); + res7 = _mm_packs_epi32(w2, w3); + res5 = _mm_packs_epi32(w4, w5); + res3 = _mm_packs_epi32(w6, w7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(res1, res7, res5, res3); + if (overflow) { + vp9_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } + } + // Transpose the 8x8. + { + // 00 01 02 03 04 05 06 07 + // 10 11 12 13 14 15 16 17 + // 20 21 22 23 24 25 26 27 + // 30 31 32 33 34 35 36 37 + // 40 41 42 43 44 45 46 47 + // 50 51 52 53 54 55 56 57 + // 60 61 62 63 64 65 66 67 + // 70 71 72 73 74 75 76 77 + const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1); + const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3); + const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1); + const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3); + const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5); + const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7); + const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5); + const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7); + // 00 10 01 11 02 12 03 13 + // 20 30 21 31 22 32 23 33 + // 04 14 05 15 06 16 07 17 + // 24 34 25 35 26 36 27 37 + // 40 50 41 51 42 52 43 53 + // 60 70 61 71 62 72 63 73 + // 54 54 55 55 56 56 57 57 + // 64 74 65 75 66 76 67 77 + const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); + const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); + const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); + const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); + const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); + const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); + const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); + const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); + // 00 10 20 30 01 11 21 31 + // 40 50 60 70 41 51 61 71 + // 02 12 22 32 03 13 23 33 + // 42 52 62 72 43 53 63 73 + // 04 14 24 34 05 15 21 36 + // 44 54 64 74 45 55 61 76 + // 06 16 26 36 07 17 27 37 + // 46 56 66 76 47 57 67 77 + in0 = _mm_unpacklo_epi64(tr1_0, tr1_4); + in1 = _mm_unpackhi_epi64(tr1_0, tr1_4); + in2 = _mm_unpacklo_epi64(tr1_2, tr1_6); + in3 = _mm_unpackhi_epi64(tr1_2, tr1_6); + in4 = _mm_unpacklo_epi64(tr1_1, tr1_5); + in5 = _mm_unpackhi_epi64(tr1_1, tr1_5); + in6 = _mm_unpacklo_epi64(tr1_3, tr1_7); + in7 = _mm_unpackhi_epi64(tr1_3, tr1_7); + // 00 10 20 30 40 50 60 70 + // 01 11 21 31 41 51 61 71 + // 02 12 22 32 42 52 62 72 + // 03 13 23 33 43 53 63 73 + // 04 14 24 34 44 54 64 74 + // 05 15 25 35 45 55 65 75 + // 06 16 26 36 46 56 66 76 + // 07 17 27 37 47 57 67 77 + } + } + // Post-condition output and store it + { + // Post-condition (division by two) + // division of two 16 bits signed numbers using shifts + // n / 2 = (n - (n >> 15)) >> 1 + const __m128i sign_in0 = _mm_srai_epi16(in0, 15); + const __m128i sign_in1 = _mm_srai_epi16(in1, 15); + const __m128i sign_in2 = _mm_srai_epi16(in2, 15); + const __m128i sign_in3 = _mm_srai_epi16(in3, 15); + const __m128i sign_in4 = _mm_srai_epi16(in4, 15); + const __m128i sign_in5 = _mm_srai_epi16(in5, 15); + const __m128i sign_in6 = _mm_srai_epi16(in6, 15); + const __m128i sign_in7 = _mm_srai_epi16(in7, 15); + in0 = _mm_sub_epi16(in0, sign_in0); + in1 = _mm_sub_epi16(in1, sign_in1); + in2 = _mm_sub_epi16(in2, sign_in2); + in3 = _mm_sub_epi16(in3, sign_in3); + in4 = _mm_sub_epi16(in4, sign_in4); + in5 = _mm_sub_epi16(in5, sign_in5); + in6 = _mm_sub_epi16(in6, sign_in6); + in7 = _mm_sub_epi16(in7, sign_in7); + in0 = _mm_srai_epi16(in0, 1); + in1 = _mm_srai_epi16(in1, 1); + in2 = _mm_srai_epi16(in2, 1); + in3 = _mm_srai_epi16(in3, 1); + in4 = _mm_srai_epi16(in4, 1); + in5 = _mm_srai_epi16(in5, 1); + in6 = _mm_srai_epi16(in6, 1); + in7 = _mm_srai_epi16(in7, 1); + // store results + store_output(in0, (output + 0 * 8)); + store_output(in1, (output + 1 * 8)); + store_output(in2, (output + 2 * 8)); + store_output(in3, (output + 3 * 8)); + store_output(in4, (output + 4 * 8)); + store_output(in5, (output + 5 * 8)); + store_output(in6, (output + 6 * 8)); + store_output(in7, (output + 7 * 8)); + } +} + +void FDCT16x16_2D(const int16_t *input, tran_low_t *output, int stride) { + // The 2D transform is done with two passes which are actually pretty + // similar. In the first one, we transform the columns and transpose + // the results. In the second one, we transform the rows. To achieve that, + // as the first pass results are transposed, we transpose the columns (that + // is the transposed rows) and transpose the results (so that it goes back + // in normal/row positions). + int pass; + // We need an intermediate buffer between passes. + DECLARE_ALIGNED_ARRAY(16, int16_t, intermediate, 256); + const int16_t *in = input; + int16_t *out0 = intermediate; + tran_low_t *out1 = output; + // Constants + // When we use them, in one case, they are all the same. In all others + // it's a pair of them that we need to repeat four times. This is done + // by constructing the 32 bit constant corresponding to that pair. + const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); + const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64); + const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); + const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); + const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64); + const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64); + const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64); + const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64); + const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64); + const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64); + const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64); + const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); + const __m128i kOne = _mm_set1_epi16(1); + // Do the two transform/transpose passes + for (pass = 0; pass < 2; ++pass) { + // We process eight columns (transposed rows in second pass) at a time. + int column_start; +#if DCT_HIGH_BIT_DEPTH + int overflow; +#endif + for (column_start = 0; column_start < 16; column_start += 8) { + __m128i in00, in01, in02, in03, in04, in05, in06, in07; + __m128i in08, in09, in10, in11, in12, in13, in14, in15; + __m128i input0, input1, input2, input3, input4, input5, input6, input7; + __m128i step1_0, step1_1, step1_2, step1_3; + __m128i step1_4, step1_5, step1_6, step1_7; + __m128i step2_1, step2_2, step2_3, step2_4, step2_5, step2_6; + __m128i step3_0, step3_1, step3_2, step3_3; + __m128i step3_4, step3_5, step3_6, step3_7; + __m128i res00, res01, res02, res03, res04, res05, res06, res07; + __m128i res08, res09, res10, res11, res12, res13, res14, res15; + // Load and pre-condition input. + if (0 == pass) { + in00 = _mm_load_si128((const __m128i *)(in + 0 * stride)); + in01 = _mm_load_si128((const __m128i *)(in + 1 * stride)); + in02 = _mm_load_si128((const __m128i *)(in + 2 * stride)); + in03 = _mm_load_si128((const __m128i *)(in + 3 * stride)); + in04 = _mm_load_si128((const __m128i *)(in + 4 * stride)); + in05 = _mm_load_si128((const __m128i *)(in + 5 * stride)); + in06 = _mm_load_si128((const __m128i *)(in + 6 * stride)); + in07 = _mm_load_si128((const __m128i *)(in + 7 * stride)); + in08 = _mm_load_si128((const __m128i *)(in + 8 * stride)); + in09 = _mm_load_si128((const __m128i *)(in + 9 * stride)); + in10 = _mm_load_si128((const __m128i *)(in + 10 * stride)); + in11 = _mm_load_si128((const __m128i *)(in + 11 * stride)); + in12 = _mm_load_si128((const __m128i *)(in + 12 * stride)); + in13 = _mm_load_si128((const __m128i *)(in + 13 * stride)); + in14 = _mm_load_si128((const __m128i *)(in + 14 * stride)); + in15 = _mm_load_si128((const __m128i *)(in + 15 * stride)); + // x = x << 2 + in00 = _mm_slli_epi16(in00, 2); + in01 = _mm_slli_epi16(in01, 2); + in02 = _mm_slli_epi16(in02, 2); + in03 = _mm_slli_epi16(in03, 2); + in04 = _mm_slli_epi16(in04, 2); + in05 = _mm_slli_epi16(in05, 2); + in06 = _mm_slli_epi16(in06, 2); + in07 = _mm_slli_epi16(in07, 2); + in08 = _mm_slli_epi16(in08, 2); + in09 = _mm_slli_epi16(in09, 2); + in10 = _mm_slli_epi16(in10, 2); + in11 = _mm_slli_epi16(in11, 2); + in12 = _mm_slli_epi16(in12, 2); + in13 = _mm_slli_epi16(in13, 2); + in14 = _mm_slli_epi16(in14, 2); + in15 = _mm_slli_epi16(in15, 2); + } else { + in00 = _mm_load_si128((const __m128i *)(in + 0 * 16)); + in01 = _mm_load_si128((const __m128i *)(in + 1 * 16)); + in02 = _mm_load_si128((const __m128i *)(in + 2 * 16)); + in03 = _mm_load_si128((const __m128i *)(in + 3 * 16)); + in04 = _mm_load_si128((const __m128i *)(in + 4 * 16)); + in05 = _mm_load_si128((const __m128i *)(in + 5 * 16)); + in06 = _mm_load_si128((const __m128i *)(in + 6 * 16)); + in07 = _mm_load_si128((const __m128i *)(in + 7 * 16)); + in08 = _mm_load_si128((const __m128i *)(in + 8 * 16)); + in09 = _mm_load_si128((const __m128i *)(in + 9 * 16)); + in10 = _mm_load_si128((const __m128i *)(in + 10 * 16)); + in11 = _mm_load_si128((const __m128i *)(in + 11 * 16)); + in12 = _mm_load_si128((const __m128i *)(in + 12 * 16)); + in13 = _mm_load_si128((const __m128i *)(in + 13 * 16)); + in14 = _mm_load_si128((const __m128i *)(in + 14 * 16)); + in15 = _mm_load_si128((const __m128i *)(in + 15 * 16)); + // x = (x + 1) >> 2 + in00 = _mm_add_epi16(in00, kOne); + in01 = _mm_add_epi16(in01, kOne); + in02 = _mm_add_epi16(in02, kOne); + in03 = _mm_add_epi16(in03, kOne); + in04 = _mm_add_epi16(in04, kOne); + in05 = _mm_add_epi16(in05, kOne); + in06 = _mm_add_epi16(in06, kOne); + in07 = _mm_add_epi16(in07, kOne); + in08 = _mm_add_epi16(in08, kOne); + in09 = _mm_add_epi16(in09, kOne); + in10 = _mm_add_epi16(in10, kOne); + in11 = _mm_add_epi16(in11, kOne); + in12 = _mm_add_epi16(in12, kOne); + in13 = _mm_add_epi16(in13, kOne); + in14 = _mm_add_epi16(in14, kOne); + in15 = _mm_add_epi16(in15, kOne); + in00 = _mm_srai_epi16(in00, 2); + in01 = _mm_srai_epi16(in01, 2); + in02 = _mm_srai_epi16(in02, 2); + in03 = _mm_srai_epi16(in03, 2); + in04 = _mm_srai_epi16(in04, 2); + in05 = _mm_srai_epi16(in05, 2); + in06 = _mm_srai_epi16(in06, 2); + in07 = _mm_srai_epi16(in07, 2); + in08 = _mm_srai_epi16(in08, 2); + in09 = _mm_srai_epi16(in09, 2); + in10 = _mm_srai_epi16(in10, 2); + in11 = _mm_srai_epi16(in11, 2); + in12 = _mm_srai_epi16(in12, 2); + in13 = _mm_srai_epi16(in13, 2); + in14 = _mm_srai_epi16(in14, 2); + in15 = _mm_srai_epi16(in15, 2); + } + in += 8; + // Calculate input for the first 8 results. + { + input0 = ADD_EPI16(in00, in15); + input1 = ADD_EPI16(in01, in14); + input2 = ADD_EPI16(in02, in13); + input3 = ADD_EPI16(in03, in12); + input4 = ADD_EPI16(in04, in11); + input5 = ADD_EPI16(in05, in10); + input6 = ADD_EPI16(in06, in09); + input7 = ADD_EPI16(in07, in08); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(input0, input1, input2, input3, + input4, input5, input6, input7); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // Calculate input for the next 8 results. + { + step1_0 = SUB_EPI16(in07, in08); + step1_1 = SUB_EPI16(in06, in09); + step1_2 = SUB_EPI16(in05, in10); + step1_3 = SUB_EPI16(in04, in11); + step1_4 = SUB_EPI16(in03, in12); + step1_5 = SUB_EPI16(in02, in13); + step1_6 = SUB_EPI16(in01, in14); + step1_7 = SUB_EPI16(in00, in15); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(step1_0, step1_1, step1_2, step1_3, + step1_4, step1_5, step1_6, step1_7); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // Work on the first eight values; fdct8(input, even_results); + { + // Add/subtract + const __m128i q0 = ADD_EPI16(input0, input7); + const __m128i q1 = ADD_EPI16(input1, input6); + const __m128i q2 = ADD_EPI16(input2, input5); + const __m128i q3 = ADD_EPI16(input3, input4); + const __m128i q4 = SUB_EPI16(input3, input4); + const __m128i q5 = SUB_EPI16(input2, input5); + const __m128i q6 = SUB_EPI16(input1, input6); + const __m128i q7 = SUB_EPI16(input0, input7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(q0, q1, q2, q3, q4, q5, q6, q7); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // Work on first four results + { + // Add/subtract + const __m128i r0 = ADD_EPI16(q0, q3); + const __m128i r1 = ADD_EPI16(q1, q2); + const __m128i r2 = SUB_EPI16(q1, q2); + const __m128i r3 = SUB_EPI16(q0, q3); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(r0, r1, r2, r3); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // Interleave to do the multiply by constants which gets us + // into 32 bits. + { + const __m128i t0 = _mm_unpacklo_epi16(r0, r1); + const __m128i t1 = _mm_unpackhi_epi16(r0, r1); + const __m128i t2 = _mm_unpacklo_epi16(r2, r3); + const __m128i t3 = _mm_unpackhi_epi16(r2, r3); + res00 = mult_round_shift(t0, t1, k__cospi_p16_p16, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res08 = mult_round_shift(t0, t1, k__cospi_p16_m16, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res04 = mult_round_shift(t2, t3, k__cospi_p24_p08, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res12 = mult_round_shift(t2, t3, k__cospi_m08_p24, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(res00, res08, res04, res12); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } + // Work on next four results + { + // Interleave to do the multiply by constants which gets us + // into 32 bits. + const __m128i d0 = _mm_unpacklo_epi16(q6, q5); + const __m128i d1 = _mm_unpackhi_epi16(q6, q5); + const __m128i r0 = mult_round_shift(d0, d1, k__cospi_p16_m16, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + const __m128i r1 = mult_round_shift(d0, d1, k__cospi_p16_p16, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x2(r0, r1); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + { + // Add/subtract + const __m128i x0 = ADD_EPI16(q4, r0); + const __m128i x1 = SUB_EPI16(q4, r0); + const __m128i x2 = SUB_EPI16(q7, r1); + const __m128i x3 = ADD_EPI16(q7, r1); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(x0, x1, x2, x3); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // Interleave to do the multiply by constants which gets us + // into 32 bits. + { + const __m128i t0 = _mm_unpacklo_epi16(x0, x3); + const __m128i t1 = _mm_unpackhi_epi16(x0, x3); + const __m128i t2 = _mm_unpacklo_epi16(x1, x2); + const __m128i t3 = _mm_unpackhi_epi16(x1, x2); + res02 = mult_round_shift(t0, t1, k__cospi_p28_p04, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res14 = mult_round_shift(t0, t1, k__cospi_m04_p28, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res10 = mult_round_shift(t2, t3, k__cospi_p12_p20, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res06 = mult_round_shift(t2, t3, k__cospi_m20_p12, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(res02, res14, res10, res06); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } + } + } + // Work on the next eight values; step1 -> odd_results + { + // step 2 + { + const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2); + const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2); + const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3); + const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3); + step2_2 = mult_round_shift(t0, t1, k__cospi_p16_m16, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2_3 = mult_round_shift(t2, t3, k__cospi_p16_m16, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2_5 = mult_round_shift(t0, t1, k__cospi_p16_p16, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2_4 = mult_round_shift(t2, t3, k__cospi_p16_p16, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(step2_2, step2_3, step2_5, + step2_4); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // step 3 + { + step3_0 = ADD_EPI16(step1_0, step2_3); + step3_1 = ADD_EPI16(step1_1, step2_2); + step3_2 = SUB_EPI16(step1_1, step2_2); + step3_3 = SUB_EPI16(step1_0, step2_3); + step3_4 = SUB_EPI16(step1_7, step2_4); + step3_5 = SUB_EPI16(step1_6, step2_5); + step3_6 = ADD_EPI16(step1_6, step2_5); + step3_7 = ADD_EPI16(step1_7, step2_4); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(step3_0, step3_1, step3_2, step3_3, + step3_4, step3_5, step3_6, step3_7); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // step 4 + { + const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6); + const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6); + const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5); + const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5); + step2_1 = mult_round_shift(t0, t1, k__cospi_m08_p24, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2_2 = mult_round_shift(t2, t3, k__cospi_p24_p08, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2_6 = mult_round_shift(t0, t1, k__cospi_p24_p08, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2_5 = mult_round_shift(t2, t3, k__cospi_p08_m24, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(step2_1, step2_2, step2_6, + step2_5); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // step 5 + { + step1_0 = ADD_EPI16(step3_0, step2_1); + step1_1 = SUB_EPI16(step3_0, step2_1); + step1_2 = ADD_EPI16(step3_3, step2_2); + step1_3 = SUB_EPI16(step3_3, step2_2); + step1_4 = SUB_EPI16(step3_4, step2_5); + step1_5 = ADD_EPI16(step3_4, step2_5); + step1_6 = SUB_EPI16(step3_7, step2_6); + step1_7 = ADD_EPI16(step3_7, step2_6); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(step1_0, step1_1, step1_2, step1_3, + step1_4, step1_5, step1_6, step1_7); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // step 6 + { + const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7); + const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7); + const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6); + const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6); + res01 = mult_round_shift(t0, t1, k__cospi_p30_p02, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res09 = mult_round_shift(t2, t3, k__cospi_p14_p18, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res15 = mult_round_shift(t0, t1, k__cospi_m02_p30, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res07 = mult_round_shift(t2, t3, k__cospi_m18_p14, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(res01, res09, res15, res07); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5); + const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5); + const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4); + const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4); + res05 = mult_round_shift(t0, t1, k__cospi_p22_p10, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res13 = mult_round_shift(t2, t3, k__cospi_p06_p26, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res11 = mult_round_shift(t0, t1, k__cospi_m10_p22, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res03 = mult_round_shift(t2, t3, k__cospi_m26_p06, + k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(res05, res13, res11, res03); + if (overflow) { + vp9_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } + // Transpose the results, do it as two 8x8 transposes. + transpose_and_output8x8(res00, res01, res02, res03, + res04, res05, res06, res07, + pass, out0, out1); + transpose_and_output8x8(res08, res09, res10, res11, + res12, res13, res14, res15, + pass, out0 + 8, out1 + 8); + if (pass == 0) { + out0 += 8*16; + } else { + out1 += 8*16; + } + } + // Setup in/out for next pass. + in = intermediate; + } +} + +#undef ADD_EPI16 +#undef SUB_EPI16 diff --git a/vp9/encoder/x86/vp9_dct_mmx.asm b/vp9/encoder/x86/vp9_dct_mmx.asm index f71181c5e..b41fbc8b3 100644 --- a/vp9/encoder/x86/vp9_dct_mmx.asm +++ b/vp9/encoder/x86/vp9_dct_mmx.asm @@ -62,9 +62,40 @@ cglobal fwht4x4, 3, 4, 8, input, output, stride psllw m2, 2 psllw m3, 2 +%if CONFIG_VP9_HIGHBITDEPTH + pxor m4, m4 + pxor m5, m5 + pcmpgtw m4, m0 + pcmpgtw m5, m1 + movq m6, m0 + movq m7, m1 + punpcklwd m0, m4 + punpcklwd m1, m5 + punpckhwd m6, m4 + punpckhwd m7, m5 + movq [outputq], m0 + movq [outputq + 8], m6 + movq [outputq + 16], m1 + movq [outputq + 24], m7 + pxor m4, m4 + pxor m5, m5 + pcmpgtw m4, m2 + pcmpgtw m5, m3 + movq m6, m2 + movq m7, m3 + punpcklwd m2, m4 + punpcklwd m3, m5 + punpckhwd m6, m4 + punpckhwd m7, m5 + movq [outputq + 32], m2 + movq [outputq + 40], m6 + movq [outputq + 48], m3 + movq [outputq + 56], m7 +%else movq [outputq], m0 movq [outputq + 8], m1 movq [outputq + 16], m2 movq [outputq + 24], m3 +%endif RET diff --git a/vp9/encoder/x86/vp9_dct_sse2.c b/vp9/encoder/x86/vp9_dct_sse2.c index f2b6a6fa4..81da34306 100644 --- a/vp9/encoder/x86/vp9_dct_sse2.c +++ b/vp9/encoder/x86/vp9_dct_sse2.c @@ -10,11 +10,11 @@ #include <emmintrin.h> // SSE2 #include "vp9/common/vp9_idct.h" // for cospi constants +#include "vp9/encoder/vp9_dct.h" +#include "vp9/encoder/x86/vp9_dct_sse2.h" #include "vpx_ports/mem.h" -#include "vp9/common/x86/vp9_idct_intrin_sse2.h" - -void vp9_fdct4x4_1_sse2(const int16_t *input, int16_t *output, int stride) { +void vp9_fdct4x4_1_sse2(const int16_t *input, tran_low_t *output, int stride) { __m128i in0, in1; __m128i tmp; const __m128i zero = _mm_setzero_si128(); @@ -40,209 +40,9 @@ void vp9_fdct4x4_1_sse2(const int16_t *input, int16_t *output, int stride) { in1 = _mm_add_epi32(tmp, in0); in0 = _mm_slli_epi32(in1, 1); - _mm_store_si128((__m128i *)(output), in0); + store_output(in0, output); } -void vp9_fdct4x4_sse2(const int16_t *input, int16_t *output, int stride) { - // This 2D transform implements 4 vertical 1D transforms followed - // by 4 horizontal 1D transforms. The multiplies and adds are as given - // by Chen, Smith and Fralick ('77). The commands for moving the data - // around have been minimized by hand. - // For the purposes of the comments, the 16 inputs are referred to at i0 - // through iF (in raster order), intermediate variables are a0, b0, c0 - // through f, and correspond to the in-place computations mapped to input - // locations. The outputs, o0 through oF are labeled according to the - // output locations. - - // Constants - // These are the coefficients used for the multiplies. - // In the comments, pN means cos(N pi /64) and mN is -cos(N pi /64), - // where cospi_N_64 = cos(N pi /64) - const __m128i k__cospi_A = _mm_setr_epi16((int16_t)cospi_16_64, (int16_t)cospi_16_64, - (int16_t)cospi_16_64, (int16_t)cospi_16_64, - (int16_t)cospi_16_64, (int16_t)-cospi_16_64, - (int16_t)cospi_16_64, (int16_t)-cospi_16_64); - const __m128i k__cospi_B = _mm_setr_epi16((int16_t)cospi_16_64, (int16_t)-cospi_16_64, - (int16_t)cospi_16_64, (int16_t)-cospi_16_64, - (int16_t)cospi_16_64, (int16_t)cospi_16_64, - (int16_t)cospi_16_64, (int16_t)cospi_16_64); - const __m128i k__cospi_C = _mm_setr_epi16((int16_t)cospi_8_64, (int16_t)cospi_24_64, - (int16_t)cospi_8_64, (int16_t)cospi_24_64, - (int16_t)cospi_24_64, (int16_t)-cospi_8_64, - (int16_t)cospi_24_64, (int16_t)-cospi_8_64); - const __m128i k__cospi_D = _mm_setr_epi16((int16_t)cospi_24_64, (int16_t)-cospi_8_64, - (int16_t)cospi_24_64, (int16_t)-cospi_8_64, - (int16_t)cospi_8_64, (int16_t)cospi_24_64, - (int16_t)cospi_8_64, (int16_t)cospi_24_64); - const __m128i k__cospi_E = _mm_setr_epi16((int16_t)cospi_16_64, (int16_t)cospi_16_64, - (int16_t)cospi_16_64, (int16_t)cospi_16_64, - (int16_t)cospi_16_64, (int16_t)cospi_16_64, - (int16_t)cospi_16_64, (int16_t)cospi_16_64); - const __m128i k__cospi_F = _mm_setr_epi16((int16_t)cospi_16_64, (int16_t)-cospi_16_64, - (int16_t)cospi_16_64, (int16_t)-cospi_16_64, - (int16_t)cospi_16_64, (int16_t)-cospi_16_64, - (int16_t)cospi_16_64, (int16_t)-cospi_16_64); - const __m128i k__cospi_G = _mm_setr_epi16((int16_t)cospi_8_64, (int16_t)cospi_24_64, - (int16_t)cospi_8_64, (int16_t)cospi_24_64, - (int16_t)-cospi_8_64, (int16_t)-cospi_24_64, - (int16_t)-cospi_8_64, (int16_t)-cospi_24_64); - const __m128i k__cospi_H = _mm_setr_epi16((int16_t)cospi_24_64, (int16_t)-cospi_8_64, - (int16_t)cospi_24_64, (int16_t)-cospi_8_64, - (int16_t)-cospi_24_64, (int16_t)cospi_8_64, - (int16_t)-cospi_24_64, (int16_t)cospi_8_64); - - const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); - // This second rounding constant saves doing some extra adds at the end - const __m128i k__DCT_CONST_ROUNDING2 = _mm_set1_epi32(DCT_CONST_ROUNDING - +(DCT_CONST_ROUNDING << 1)); - const int DCT_CONST_BITS2 = DCT_CONST_BITS+2; - const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1); - const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0); - __m128i in0, in1; - - // Load inputs. - { - in0 = _mm_loadl_epi64((const __m128i *)(input + 0 * stride)); - in1 = _mm_loadl_epi64((const __m128i *)(input + 1 * stride)); - in1 = _mm_unpacklo_epi64(in1, _mm_loadl_epi64((const __m128i *) - (input + 2 * stride))); - in0 = _mm_unpacklo_epi64(in0, _mm_loadl_epi64((const __m128i *) - (input + 3 * stride))); - // in0 = [i0 i1 i2 i3 iC iD iE iF] - // in1 = [i4 i5 i6 i7 i8 i9 iA iB] - - - // multiply by 16 to give some extra precision - in0 = _mm_slli_epi16(in0, 4); - in1 = _mm_slli_epi16(in1, 4); - // if (i == 0 && input[0]) input[0] += 1; - // add 1 to the upper left pixel if it is non-zero, which helps reduce - // the round-trip error - { - // The mask will only contain whether the first value is zero, all - // other comparison will fail as something shifted by 4 (above << 4) - // can never be equal to one. To increment in the non-zero case, we - // add the mask and one for the first element: - // - if zero, mask = -1, v = v - 1 + 1 = v - // - if non-zero, mask = 0, v = v + 0 + 1 = v + 1 - __m128i mask = _mm_cmpeq_epi16(in0, k__nonzero_bias_a); - in0 = _mm_add_epi16(in0, mask); - in0 = _mm_add_epi16(in0, k__nonzero_bias_b); - } - } - // There are 4 total stages, alternating between an add/subtract stage - // followed by an multiply-and-add stage. - { - // Stage 1: Add/subtract - - // in0 = [i0 i1 i2 i3 iC iD iE iF] - // in1 = [i4 i5 i6 i7 i8 i9 iA iB] - const __m128i r0 = _mm_unpacklo_epi16(in0, in1); - const __m128i r1 = _mm_unpackhi_epi16(in0, in1); - // r0 = [i0 i4 i1 i5 i2 i6 i3 i7] - // r1 = [iC i8 iD i9 iE iA iF iB] - const __m128i r2 = _mm_shuffle_epi32(r0, 0xB4); - const __m128i r3 = _mm_shuffle_epi32(r1, 0xB4); - // r2 = [i0 i4 i1 i5 i3 i7 i2 i6] - // r3 = [iC i8 iD i9 iF iB iE iA] - - const __m128i t0 = _mm_add_epi16(r2, r3); - const __m128i t1 = _mm_sub_epi16(r2, r3); - // t0 = [a0 a4 a1 a5 a3 a7 a2 a6] - // t1 = [aC a8 aD a9 aF aB aE aA] - - // Stage 2: multiply by constants (which gets us into 32 bits). - // The constants needed here are: - // k__cospi_A = [p16 p16 p16 p16 p16 m16 p16 m16] - // k__cospi_B = [p16 m16 p16 m16 p16 p16 p16 p16] - // k__cospi_C = [p08 p24 p08 p24 p24 m08 p24 m08] - // k__cospi_D = [p24 m08 p24 m08 p08 p24 p08 p24] - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_A); - const __m128i u2 = _mm_madd_epi16(t0, k__cospi_B); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_C); - const __m128i u3 = _mm_madd_epi16(t1, k__cospi_D); - // Then add and right-shift to get back to 16-bit range - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - // w0 = [b0 b1 b7 b6] - // w1 = [b8 b9 bF bE] - // w2 = [b4 b5 b3 b2] - // w3 = [bC bD bB bA] - const __m128i x0 = _mm_packs_epi32(w0, w1); - const __m128i x1 = _mm_packs_epi32(w2, w3); - // x0 = [b0 b1 b7 b6 b8 b9 bF bE] - // x1 = [b4 b5 b3 b2 bC bD bB bA] - in0 = _mm_shuffle_epi32(x0, 0xD8); - in1 = _mm_shuffle_epi32(x1, 0x8D); - // in0 = [b0 b1 b8 b9 b7 b6 bF bE] - // in1 = [b3 b2 bB bA b4 b5 bC bD] - } - { - // vertical DCTs finished. Now we do the horizontal DCTs. - // Stage 3: Add/subtract - - const __m128i t0 = _mm_add_epi16(in0, in1); - const __m128i t1 = _mm_sub_epi16(in0, in1); - // t0 = [c0 c1 c8 c9 c4 c5 cC cD] - // t1 = [c3 c2 cB cA -c7 -c6 -cF -cE] - - // Stage 4: multiply by constants (which gets us into 32 bits). - // The constants needed here are: - // k__cospi_E = [p16 p16 p16 p16 p16 p16 p16 p16] - // k__cospi_F = [p16 m16 p16 m16 p16 m16 p16 m16] - // k__cospi_G = [p08 p24 p08 p24 m08 m24 m08 m24] - // k__cospi_H = [p24 m08 p24 m08 m24 p08 m24 p08] - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_E); - const __m128i u1 = _mm_madd_epi16(t0, k__cospi_F); - const __m128i u2 = _mm_madd_epi16(t1, k__cospi_G); - const __m128i u3 = _mm_madd_epi16(t1, k__cospi_H); - // Then add and right-shift to get back to 16-bit range - // but this combines the final right-shift as well to save operations - // This unusual rounding operations is to maintain bit-accurate - // compatibility with the c version of this function which has two - // rounding steps in a row. - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING2); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING2); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING2); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING2); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS2); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS2); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS2); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS2); - // w0 = [o0 o4 o8 oC] - // w1 = [o2 o6 oA oE] - // w2 = [o1 o5 o9 oD] - // w3 = [o3 o7 oB oF] - // remember the o's are numbered according to the correct output location - const __m128i x0 = _mm_packs_epi32(w0, w1); - const __m128i x1 = _mm_packs_epi32(w2, w3); - // x0 = [o0 o4 o8 oC o2 o6 oA oE] - // x1 = [o1 o5 o9 oD o3 o7 oB oF] - const __m128i y0 = _mm_unpacklo_epi16(x0, x1); - const __m128i y1 = _mm_unpackhi_epi16(x0, x1); - // y0 = [o0 o1 o4 o5 o8 o9 oC oD] - // y1 = [o2 o3 o6 o7 oA oB oE oF] - in0 = _mm_unpacklo_epi32(y0, y1); - // in0 = [o0 o1 o2 o3 o4 o5 o6 o7] - in1 = _mm_unpackhi_epi32(y0, y1); - // in1 = [o8 o9 oA oB oC oD oE oF] - } - // Post-condition (v + 1) >> 2 is now incorporated into previous - // add and right-shift commands. Only 2 store instructions needed - // because we are using the fact that 1/3 are stored just after 0/2. - { - _mm_storeu_si128((__m128i *)(output + 0 * 4), in0); - _mm_storeu_si128((__m128i *)(output + 2 * 4), in1); - } -} - - static INLINE void load_buffer_4x4(const int16_t *input, __m128i *in, int stride) { const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1); @@ -264,7 +64,7 @@ static INLINE void load_buffer_4x4(const int16_t *input, __m128i *in, in[0] = _mm_add_epi16(in[0], k__nonzero_bias_b); } -static INLINE void write_buffer_4x4(int16_t *output, __m128i *res) { +static INLINE void write_buffer_4x4(tran_low_t *output, __m128i *res) { const __m128i kOne = _mm_set1_epi16(1); __m128i in01 = _mm_unpacklo_epi64(res[0], res[1]); __m128i in23 = _mm_unpacklo_epi64(res[2], res[3]); @@ -272,8 +72,8 @@ static INLINE void write_buffer_4x4(int16_t *output, __m128i *res) { __m128i out23 = _mm_add_epi16(in23, kOne); out01 = _mm_srai_epi16(out01, 2); out23 = _mm_srai_epi16(out23, 2); - _mm_store_si128((__m128i *)(output + 0 * 8), out01); - _mm_store_si128((__m128i *)(output + 1 * 8), out23); + store_output(out01, (output + 0 * 8)); + store_output(out23, (output + 1 * 8)); } static INLINE void transpose_4x4(__m128i *res) { @@ -376,7 +176,7 @@ void fadst4_sse2(__m128i *in) { transpose_4x4(in); } -void vp9_fht4x4_sse2(const int16_t *input, int16_t *output, +void vp9_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type) { __m128i in[4]; @@ -408,7 +208,7 @@ void vp9_fht4x4_sse2(const int16_t *input, int16_t *output, } } -void vp9_fdct8x8_1_sse2(const int16_t *input, int16_t *output, int stride) { +void vp9_fdct8x8_1_sse2(const int16_t *input, tran_low_t *output, int stride) { __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride)); __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride)); __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride)); @@ -445,263 +245,7 @@ void vp9_fdct8x8_1_sse2(const int16_t *input, int16_t *output, int stride) { in0 = _mm_srli_si128(sum, 8); in1 = _mm_add_epi32(sum, in0); - _mm_store_si128((__m128i *)(output), in1); -} - -void vp9_fdct8x8_sse2(const int16_t *input, int16_t *output, int stride) { - int pass; - // Constants - // When we use them, in one case, they are all the same. In all others - // it's a pair of them that we need to repeat four times. This is done - // by constructing the 32 bit constant corresponding to that pair. - const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64); - const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); - const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); - const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); - const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); - const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); - const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); - const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); - const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); - // Load input - __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride)); - __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride)); - __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride)); - __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride)); - __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride)); - __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride)); - __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride)); - __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride)); - // Pre-condition input (shift by two) - in0 = _mm_slli_epi16(in0, 2); - in1 = _mm_slli_epi16(in1, 2); - in2 = _mm_slli_epi16(in2, 2); - in3 = _mm_slli_epi16(in3, 2); - in4 = _mm_slli_epi16(in4, 2); - in5 = _mm_slli_epi16(in5, 2); - in6 = _mm_slli_epi16(in6, 2); - in7 = _mm_slli_epi16(in7, 2); - - // We do two passes, first the columns, then the rows. The results of the - // first pass are transposed so that the same column code can be reused. The - // results of the second pass are also transposed so that the rows (processed - // as columns) are put back in row positions. - for (pass = 0; pass < 2; pass++) { - // To store results of each pass before the transpose. - __m128i res0, res1, res2, res3, res4, res5, res6, res7; - // Add/subtract - const __m128i q0 = _mm_add_epi16(in0, in7); - const __m128i q1 = _mm_add_epi16(in1, in6); - const __m128i q2 = _mm_add_epi16(in2, in5); - const __m128i q3 = _mm_add_epi16(in3, in4); - const __m128i q4 = _mm_sub_epi16(in3, in4); - const __m128i q5 = _mm_sub_epi16(in2, in5); - const __m128i q6 = _mm_sub_epi16(in1, in6); - const __m128i q7 = _mm_sub_epi16(in0, in7); - // Work on first four results - { - // Add/subtract - const __m128i r0 = _mm_add_epi16(q0, q3); - const __m128i r1 = _mm_add_epi16(q1, q2); - const __m128i r2 = _mm_sub_epi16(q1, q2); - const __m128i r3 = _mm_sub_epi16(q0, q3); - // Interleave to do the multiply by constants which gets us into 32bits - const __m128i t0 = _mm_unpacklo_epi16(r0, r1); - const __m128i t1 = _mm_unpackhi_epi16(r0, r1); - const __m128i t2 = _mm_unpacklo_epi16(r2, r3); - const __m128i t3 = _mm_unpackhi_epi16(r2, r3); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); - const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); - const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16); - const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); - const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08); - const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); - const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); - const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); - const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); - const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); - const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); - const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); - const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); - // Combine - res0 = _mm_packs_epi32(w0, w1); - res4 = _mm_packs_epi32(w2, w3); - res2 = _mm_packs_epi32(w4, w5); - res6 = _mm_packs_epi32(w6, w7); - } - // Work on next four results - { - // Interleave to do the multiply by constants which gets us into 32bits - const __m128i d0 = _mm_unpacklo_epi16(q6, q5); - const __m128i d1 = _mm_unpackhi_epi16(q6, q5); - const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16); - const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16); - const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16); - const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16); - // dct_const_round_shift - const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING); - const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING); - const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING); - const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING); - const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS); - const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS); - const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS); - const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS); - // Combine - const __m128i r0 = _mm_packs_epi32(s0, s1); - const __m128i r1 = _mm_packs_epi32(s2, s3); - // Add/subtract - const __m128i x0 = _mm_add_epi16(q4, r0); - const __m128i x1 = _mm_sub_epi16(q4, r0); - const __m128i x2 = _mm_sub_epi16(q7, r1); - const __m128i x3 = _mm_add_epi16(q7, r1); - // Interleave to do the multiply by constants which gets us into 32bits - const __m128i t0 = _mm_unpacklo_epi16(x0, x3); - const __m128i t1 = _mm_unpackhi_epi16(x0, x3); - const __m128i t2 = _mm_unpacklo_epi16(x1, x2); - const __m128i t3 = _mm_unpackhi_epi16(x1, x2); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04); - const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28); - const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28); - const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20); - const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20); - const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12); - const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); - const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); - const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); - const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); - const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); - const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); - const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); - // Combine - res1 = _mm_packs_epi32(w0, w1); - res7 = _mm_packs_epi32(w2, w3); - res5 = _mm_packs_epi32(w4, w5); - res3 = _mm_packs_epi32(w6, w7); - } - // Transpose the 8x8. - { - // 00 01 02 03 04 05 06 07 - // 10 11 12 13 14 15 16 17 - // 20 21 22 23 24 25 26 27 - // 30 31 32 33 34 35 36 37 - // 40 41 42 43 44 45 46 47 - // 50 51 52 53 54 55 56 57 - // 60 61 62 63 64 65 66 67 - // 70 71 72 73 74 75 76 77 - const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1); - const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3); - const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1); - const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3); - const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5); - const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7); - const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5); - const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7); - // 00 10 01 11 02 12 03 13 - // 20 30 21 31 22 32 23 33 - // 04 14 05 15 06 16 07 17 - // 24 34 25 35 26 36 27 37 - // 40 50 41 51 42 52 43 53 - // 60 70 61 71 62 72 63 73 - // 54 54 55 55 56 56 57 57 - // 64 74 65 75 66 76 67 77 - const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); - const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); - const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); - const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); - const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); - const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); - const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); - const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); - // 00 10 20 30 01 11 21 31 - // 40 50 60 70 41 51 61 71 - // 02 12 22 32 03 13 23 33 - // 42 52 62 72 43 53 63 73 - // 04 14 24 34 05 15 21 36 - // 44 54 64 74 45 55 61 76 - // 06 16 26 36 07 17 27 37 - // 46 56 66 76 47 57 67 77 - in0 = _mm_unpacklo_epi64(tr1_0, tr1_4); - in1 = _mm_unpackhi_epi64(tr1_0, tr1_4); - in2 = _mm_unpacklo_epi64(tr1_2, tr1_6); - in3 = _mm_unpackhi_epi64(tr1_2, tr1_6); - in4 = _mm_unpacklo_epi64(tr1_1, tr1_5); - in5 = _mm_unpackhi_epi64(tr1_1, tr1_5); - in6 = _mm_unpacklo_epi64(tr1_3, tr1_7); - in7 = _mm_unpackhi_epi64(tr1_3, tr1_7); - // 00 10 20 30 40 50 60 70 - // 01 11 21 31 41 51 61 71 - // 02 12 22 32 42 52 62 72 - // 03 13 23 33 43 53 63 73 - // 04 14 24 34 44 54 64 74 - // 05 15 25 35 45 55 65 75 - // 06 16 26 36 46 56 66 76 - // 07 17 27 37 47 57 67 77 - } - } - // Post-condition output and store it - { - // Post-condition (division by two) - // division of two 16 bits signed numbers using shifts - // n / 2 = (n - (n >> 15)) >> 1 - const __m128i sign_in0 = _mm_srai_epi16(in0, 15); - const __m128i sign_in1 = _mm_srai_epi16(in1, 15); - const __m128i sign_in2 = _mm_srai_epi16(in2, 15); - const __m128i sign_in3 = _mm_srai_epi16(in3, 15); - const __m128i sign_in4 = _mm_srai_epi16(in4, 15); - const __m128i sign_in5 = _mm_srai_epi16(in5, 15); - const __m128i sign_in6 = _mm_srai_epi16(in6, 15); - const __m128i sign_in7 = _mm_srai_epi16(in7, 15); - in0 = _mm_sub_epi16(in0, sign_in0); - in1 = _mm_sub_epi16(in1, sign_in1); - in2 = _mm_sub_epi16(in2, sign_in2); - in3 = _mm_sub_epi16(in3, sign_in3); - in4 = _mm_sub_epi16(in4, sign_in4); - in5 = _mm_sub_epi16(in5, sign_in5); - in6 = _mm_sub_epi16(in6, sign_in6); - in7 = _mm_sub_epi16(in7, sign_in7); - in0 = _mm_srai_epi16(in0, 1); - in1 = _mm_srai_epi16(in1, 1); - in2 = _mm_srai_epi16(in2, 1); - in3 = _mm_srai_epi16(in3, 1); - in4 = _mm_srai_epi16(in4, 1); - in5 = _mm_srai_epi16(in5, 1); - in6 = _mm_srai_epi16(in6, 1); - in7 = _mm_srai_epi16(in7, 1); - // store results - _mm_store_si128((__m128i *)(output + 0 * 8), in0); - _mm_store_si128((__m128i *)(output + 1 * 8), in1); - _mm_store_si128((__m128i *)(output + 2 * 8), in2); - _mm_store_si128((__m128i *)(output + 3 * 8), in3); - _mm_store_si128((__m128i *)(output + 4 * 8), in4); - _mm_store_si128((__m128i *)(output + 5 * 8), in5); - _mm_store_si128((__m128i *)(output + 6 * 8), in6); - _mm_store_si128((__m128i *)(output + 7 * 8), in7); - } + store_output(in1, output); } void vp9_fdct8x8_quant_sse2(const int16_t *input, int stride, @@ -1213,15 +757,68 @@ static INLINE void right_shift_8x8(__m128i *res, int const bit) { } // write 8x8 array -static INLINE void write_buffer_8x8(int16_t *output, __m128i *res, int stride) { - _mm_store_si128((__m128i *)(output + 0 * stride), res[0]); - _mm_store_si128((__m128i *)(output + 1 * stride), res[1]); - _mm_store_si128((__m128i *)(output + 2 * stride), res[2]); - _mm_store_si128((__m128i *)(output + 3 * stride), res[3]); - _mm_store_si128((__m128i *)(output + 4 * stride), res[4]); - _mm_store_si128((__m128i *)(output + 5 * stride), res[5]); - _mm_store_si128((__m128i *)(output + 6 * stride), res[6]); - _mm_store_si128((__m128i *)(output + 7 * stride), res[7]); +static INLINE void write_buffer_8x8(tran_low_t *output, __m128i *res, + int stride) { + store_output(res[0], (output + 0 * stride)); + store_output(res[1], (output + 1 * stride)); + store_output(res[2], (output + 2 * stride)); + store_output(res[3], (output + 3 * stride)); + store_output(res[4], (output + 4 * stride)); + store_output(res[5], (output + 5 * stride)); + store_output(res[6], (output + 6 * stride)); + store_output(res[7], (output + 7 * stride)); +} + +// perform in-place transpose +static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) { + const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]); + const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]); + const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]); + const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]); + const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]); + const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]); + const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]); + const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]); + // 00 10 01 11 02 12 03 13 + // 20 30 21 31 22 32 23 33 + // 04 14 05 15 06 16 07 17 + // 24 34 25 35 26 36 27 37 + // 40 50 41 51 42 52 43 53 + // 60 70 61 71 62 72 63 73 + // 44 54 45 55 46 56 47 57 + // 64 74 65 75 66 76 67 77 + const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); + const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5); + const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); + const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5); + const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3); + const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); + const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3); + const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); + // 00 10 20 30 01 11 21 31 + // 40 50 60 70 41 51 61 71 + // 02 12 22 32 03 13 23 33 + // 42 52 62 72 43 53 63 73 + // 04 14 24 34 05 15 25 35 + // 44 54 64 74 45 55 65 75 + // 06 16 26 36 07 17 27 37 + // 46 56 66 76 47 57 67 77 + res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1); + res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1); + res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3); + res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3); + res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5); + res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5); + res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7); + res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7); + // 00 10 20 30 40 50 60 70 + // 01 11 21 31 41 51 61 71 + // 02 12 22 32 42 52 62 72 + // 03 13 23 33 43 53 63 73 + // 04 14 24 34 44 54 64 74 + // 05 15 25 35 45 55 65 75 + // 06 16 26 36 46 56 66 76 + // 07 17 27 37 47 57 67 77 } void fdct8_sse2(__m128i *in) { @@ -1594,7 +1191,7 @@ void fadst8_sse2(__m128i *in) { array_transpose_8x8(in, in); } -void vp9_fht8x8_sse2(const int16_t *input, int16_t *output, +void vp9_fht8x8_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type) { __m128i in[8]; @@ -1629,7 +1226,8 @@ void vp9_fht8x8_sse2(const int16_t *input, int16_t *output, } } -void vp9_fdct16x16_1_sse2(const int16_t *input, int16_t *output, int stride) { +void vp9_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output, + int stride) { __m128i in0, in1, in2, in3; __m128i u0, u1; __m128i sum = _mm_setzero_si128(); @@ -1694,632 +1292,7 @@ void vp9_fdct16x16_1_sse2(const int16_t *input, int16_t *output, int stride) { in1 = _mm_add_epi32(sum, in0); in1 = _mm_srai_epi32(in1, 1); - _mm_store_si128((__m128i *)(output), in1); -} - -void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride) { - // The 2D transform is done with two passes which are actually pretty - // similar. In the first one, we transform the columns and transpose - // the results. In the second one, we transform the rows. To achieve that, - // as the first pass results are transposed, we transpose the columns (that - // is the transposed rows) and transpose the results (so that it goes back - // in normal/row positions). - int pass; - // We need an intermediate buffer between passes. - DECLARE_ALIGNED_ARRAY(16, int16_t, intermediate, 256); - const int16_t *in = input; - int16_t *out = intermediate; - // Constants - // When we use them, in one case, they are all the same. In all others - // it's a pair of them that we need to repeat four times. This is done - // by constructing the 32 bit constant corresponding to that pair. - const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64); - const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); - const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); - const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64); - const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); - const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); - const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); - const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); - const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); - const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64); - const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64); - const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64); - const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64); - const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64); - const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64); - const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64); - const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64); - const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); - const __m128i kOne = _mm_set1_epi16(1); - // Do the two transform/transpose passes - for (pass = 0; pass < 2; ++pass) { - // We process eight columns (transposed rows in second pass) at a time. - int column_start; - for (column_start = 0; column_start < 16; column_start += 8) { - __m128i in00, in01, in02, in03, in04, in05, in06, in07; - __m128i in08, in09, in10, in11, in12, in13, in14, in15; - __m128i input0, input1, input2, input3, input4, input5, input6, input7; - __m128i step1_0, step1_1, step1_2, step1_3; - __m128i step1_4, step1_5, step1_6, step1_7; - __m128i step2_1, step2_2, step2_3, step2_4, step2_5, step2_6; - __m128i step3_0, step3_1, step3_2, step3_3; - __m128i step3_4, step3_5, step3_6, step3_7; - __m128i res00, res01, res02, res03, res04, res05, res06, res07; - __m128i res08, res09, res10, res11, res12, res13, res14, res15; - // Load and pre-condition input. - if (0 == pass) { - in00 = _mm_load_si128((const __m128i *)(in + 0 * stride)); - in01 = _mm_load_si128((const __m128i *)(in + 1 * stride)); - in02 = _mm_load_si128((const __m128i *)(in + 2 * stride)); - in03 = _mm_load_si128((const __m128i *)(in + 3 * stride)); - in04 = _mm_load_si128((const __m128i *)(in + 4 * stride)); - in05 = _mm_load_si128((const __m128i *)(in + 5 * stride)); - in06 = _mm_load_si128((const __m128i *)(in + 6 * stride)); - in07 = _mm_load_si128((const __m128i *)(in + 7 * stride)); - in08 = _mm_load_si128((const __m128i *)(in + 8 * stride)); - in09 = _mm_load_si128((const __m128i *)(in + 9 * stride)); - in10 = _mm_load_si128((const __m128i *)(in + 10 * stride)); - in11 = _mm_load_si128((const __m128i *)(in + 11 * stride)); - in12 = _mm_load_si128((const __m128i *)(in + 12 * stride)); - in13 = _mm_load_si128((const __m128i *)(in + 13 * stride)); - in14 = _mm_load_si128((const __m128i *)(in + 14 * stride)); - in15 = _mm_load_si128((const __m128i *)(in + 15 * stride)); - // x = x << 2 - in00 = _mm_slli_epi16(in00, 2); - in01 = _mm_slli_epi16(in01, 2); - in02 = _mm_slli_epi16(in02, 2); - in03 = _mm_slli_epi16(in03, 2); - in04 = _mm_slli_epi16(in04, 2); - in05 = _mm_slli_epi16(in05, 2); - in06 = _mm_slli_epi16(in06, 2); - in07 = _mm_slli_epi16(in07, 2); - in08 = _mm_slli_epi16(in08, 2); - in09 = _mm_slli_epi16(in09, 2); - in10 = _mm_slli_epi16(in10, 2); - in11 = _mm_slli_epi16(in11, 2); - in12 = _mm_slli_epi16(in12, 2); - in13 = _mm_slli_epi16(in13, 2); - in14 = _mm_slli_epi16(in14, 2); - in15 = _mm_slli_epi16(in15, 2); - } else { - in00 = _mm_load_si128((const __m128i *)(in + 0 * 16)); - in01 = _mm_load_si128((const __m128i *)(in + 1 * 16)); - in02 = _mm_load_si128((const __m128i *)(in + 2 * 16)); - in03 = _mm_load_si128((const __m128i *)(in + 3 * 16)); - in04 = _mm_load_si128((const __m128i *)(in + 4 * 16)); - in05 = _mm_load_si128((const __m128i *)(in + 5 * 16)); - in06 = _mm_load_si128((const __m128i *)(in + 6 * 16)); - in07 = _mm_load_si128((const __m128i *)(in + 7 * 16)); - in08 = _mm_load_si128((const __m128i *)(in + 8 * 16)); - in09 = _mm_load_si128((const __m128i *)(in + 9 * 16)); - in10 = _mm_load_si128((const __m128i *)(in + 10 * 16)); - in11 = _mm_load_si128((const __m128i *)(in + 11 * 16)); - in12 = _mm_load_si128((const __m128i *)(in + 12 * 16)); - in13 = _mm_load_si128((const __m128i *)(in + 13 * 16)); - in14 = _mm_load_si128((const __m128i *)(in + 14 * 16)); - in15 = _mm_load_si128((const __m128i *)(in + 15 * 16)); - // x = (x + 1) >> 2 - in00 = _mm_add_epi16(in00, kOne); - in01 = _mm_add_epi16(in01, kOne); - in02 = _mm_add_epi16(in02, kOne); - in03 = _mm_add_epi16(in03, kOne); - in04 = _mm_add_epi16(in04, kOne); - in05 = _mm_add_epi16(in05, kOne); - in06 = _mm_add_epi16(in06, kOne); - in07 = _mm_add_epi16(in07, kOne); - in08 = _mm_add_epi16(in08, kOne); - in09 = _mm_add_epi16(in09, kOne); - in10 = _mm_add_epi16(in10, kOne); - in11 = _mm_add_epi16(in11, kOne); - in12 = _mm_add_epi16(in12, kOne); - in13 = _mm_add_epi16(in13, kOne); - in14 = _mm_add_epi16(in14, kOne); - in15 = _mm_add_epi16(in15, kOne); - in00 = _mm_srai_epi16(in00, 2); - in01 = _mm_srai_epi16(in01, 2); - in02 = _mm_srai_epi16(in02, 2); - in03 = _mm_srai_epi16(in03, 2); - in04 = _mm_srai_epi16(in04, 2); - in05 = _mm_srai_epi16(in05, 2); - in06 = _mm_srai_epi16(in06, 2); - in07 = _mm_srai_epi16(in07, 2); - in08 = _mm_srai_epi16(in08, 2); - in09 = _mm_srai_epi16(in09, 2); - in10 = _mm_srai_epi16(in10, 2); - in11 = _mm_srai_epi16(in11, 2); - in12 = _mm_srai_epi16(in12, 2); - in13 = _mm_srai_epi16(in13, 2); - in14 = _mm_srai_epi16(in14, 2); - in15 = _mm_srai_epi16(in15, 2); - } - in += 8; - // Calculate input for the first 8 results. - { - input0 = _mm_add_epi16(in00, in15); - input1 = _mm_add_epi16(in01, in14); - input2 = _mm_add_epi16(in02, in13); - input3 = _mm_add_epi16(in03, in12); - input4 = _mm_add_epi16(in04, in11); - input5 = _mm_add_epi16(in05, in10); - input6 = _mm_add_epi16(in06, in09); - input7 = _mm_add_epi16(in07, in08); - } - // Calculate input for the next 8 results. - { - step1_0 = _mm_sub_epi16(in07, in08); - step1_1 = _mm_sub_epi16(in06, in09); - step1_2 = _mm_sub_epi16(in05, in10); - step1_3 = _mm_sub_epi16(in04, in11); - step1_4 = _mm_sub_epi16(in03, in12); - step1_5 = _mm_sub_epi16(in02, in13); - step1_6 = _mm_sub_epi16(in01, in14); - step1_7 = _mm_sub_epi16(in00, in15); - } - // Work on the first eight values; fdct8(input, even_results); - { - // Add/subtract - const __m128i q0 = _mm_add_epi16(input0, input7); - const __m128i q1 = _mm_add_epi16(input1, input6); - const __m128i q2 = _mm_add_epi16(input2, input5); - const __m128i q3 = _mm_add_epi16(input3, input4); - const __m128i q4 = _mm_sub_epi16(input3, input4); - const __m128i q5 = _mm_sub_epi16(input2, input5); - const __m128i q6 = _mm_sub_epi16(input1, input6); - const __m128i q7 = _mm_sub_epi16(input0, input7); - // Work on first four results - { - // Add/subtract - const __m128i r0 = _mm_add_epi16(q0, q3); - const __m128i r1 = _mm_add_epi16(q1, q2); - const __m128i r2 = _mm_sub_epi16(q1, q2); - const __m128i r3 = _mm_sub_epi16(q0, q3); - // Interleave to do the multiply by constants which gets us - // into 32 bits. - const __m128i t0 = _mm_unpacklo_epi16(r0, r1); - const __m128i t1 = _mm_unpackhi_epi16(r0, r1); - const __m128i t2 = _mm_unpacklo_epi16(r2, r3); - const __m128i t3 = _mm_unpackhi_epi16(r2, r3); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); - const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); - const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16); - const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); - const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08); - const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); - const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); - const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); - const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); - const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); - const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); - const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); - const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); - // Combine - res00 = _mm_packs_epi32(w0, w1); - res08 = _mm_packs_epi32(w2, w3); - res04 = _mm_packs_epi32(w4, w5); - res12 = _mm_packs_epi32(w6, w7); - } - // Work on next four results - { - // Interleave to do the multiply by constants which gets us - // into 32 bits. - const __m128i d0 = _mm_unpacklo_epi16(q6, q5); - const __m128i d1 = _mm_unpackhi_epi16(q6, q5); - const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16); - const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16); - const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16); - const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16); - // dct_const_round_shift - const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING); - const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING); - const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING); - const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING); - const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS); - const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS); - const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS); - const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS); - // Combine - const __m128i r0 = _mm_packs_epi32(s0, s1); - const __m128i r1 = _mm_packs_epi32(s2, s3); - // Add/subtract - const __m128i x0 = _mm_add_epi16(q4, r0); - const __m128i x1 = _mm_sub_epi16(q4, r0); - const __m128i x2 = _mm_sub_epi16(q7, r1); - const __m128i x3 = _mm_add_epi16(q7, r1); - // Interleave to do the multiply by constants which gets us - // into 32 bits. - const __m128i t0 = _mm_unpacklo_epi16(x0, x3); - const __m128i t1 = _mm_unpackhi_epi16(x0, x3); - const __m128i t2 = _mm_unpacklo_epi16(x1, x2); - const __m128i t3 = _mm_unpackhi_epi16(x1, x2); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04); - const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28); - const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28); - const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20); - const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20); - const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12); - const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); - const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); - const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); - const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); - const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); - const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); - const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); - // Combine - res02 = _mm_packs_epi32(w0, w1); - res14 = _mm_packs_epi32(w2, w3); - res10 = _mm_packs_epi32(w4, w5); - res06 = _mm_packs_epi32(w6, w7); - } - } - // Work on the next eight values; step1 -> odd_results - { - // step 2 - { - const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2); - const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2); - const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3); - const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_m16); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_m16); - const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p16_m16); - const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p16_m16); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - // Combine - step2_2 = _mm_packs_epi32(w0, w1); - step2_3 = _mm_packs_epi32(w2, w3); - } - { - const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2); - const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2); - const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3); - const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); - const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p16_p16); - const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p16_p16); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - // Combine - step2_5 = _mm_packs_epi32(w0, w1); - step2_4 = _mm_packs_epi32(w2, w3); - } - // step 3 - { - step3_0 = _mm_add_epi16(step1_0, step2_3); - step3_1 = _mm_add_epi16(step1_1, step2_2); - step3_2 = _mm_sub_epi16(step1_1, step2_2); - step3_3 = _mm_sub_epi16(step1_0, step2_3); - step3_4 = _mm_sub_epi16(step1_7, step2_4); - step3_5 = _mm_sub_epi16(step1_6, step2_5); - step3_6 = _mm_add_epi16(step1_6, step2_5); - step3_7 = _mm_add_epi16(step1_7, step2_4); - } - // step 4 - { - const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6); - const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6); - const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5); - const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m08_p24); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m08_p24); - const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p24_p08); - const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p24_p08); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - // Combine - step2_1 = _mm_packs_epi32(w0, w1); - step2_2 = _mm_packs_epi32(w2, w3); - } - { - const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6); - const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6); - const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5); - const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p24_p08); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p24_p08); - const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p08_m24); - const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p08_m24); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - // Combine - step2_6 = _mm_packs_epi32(w0, w1); - step2_5 = _mm_packs_epi32(w2, w3); - } - // step 5 - { - step1_0 = _mm_add_epi16(step3_0, step2_1); - step1_1 = _mm_sub_epi16(step3_0, step2_1); - step1_2 = _mm_add_epi16(step3_3, step2_2); - step1_3 = _mm_sub_epi16(step3_3, step2_2); - step1_4 = _mm_sub_epi16(step3_4, step2_5); - step1_5 = _mm_add_epi16(step3_4, step2_5); - step1_6 = _mm_sub_epi16(step3_7, step2_6); - step1_7 = _mm_add_epi16(step3_7, step2_6); - } - // step 6 - { - const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7); - const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7); - const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6); - const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p30_p02); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p30_p02); - const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p14_p18); - const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p14_p18); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - // Combine - res01 = _mm_packs_epi32(w0, w1); - res09 = _mm_packs_epi32(w2, w3); - } - { - const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5); - const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5); - const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4); - const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p22_p10); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p22_p10); - const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p06_p26); - const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p06_p26); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - // Combine - res05 = _mm_packs_epi32(w0, w1); - res13 = _mm_packs_epi32(w2, w3); - } - { - const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5); - const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5); - const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4); - const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m10_p22); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m10_p22); - const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m26_p06); - const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m26_p06); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - // Combine - res11 = _mm_packs_epi32(w0, w1); - res03 = _mm_packs_epi32(w2, w3); - } - { - const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7); - const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7); - const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6); - const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m02_p30); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m02_p30); - const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m18_p14); - const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m18_p14); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - // Combine - res15 = _mm_packs_epi32(w0, w1); - res07 = _mm_packs_epi32(w2, w3); - } - } - // Transpose the results, do it as two 8x8 transposes. - { - // 00 01 02 03 04 05 06 07 - // 10 11 12 13 14 15 16 17 - // 20 21 22 23 24 25 26 27 - // 30 31 32 33 34 35 36 37 - // 40 41 42 43 44 45 46 47 - // 50 51 52 53 54 55 56 57 - // 60 61 62 63 64 65 66 67 - // 70 71 72 73 74 75 76 77 - const __m128i tr0_0 = _mm_unpacklo_epi16(res00, res01); - const __m128i tr0_1 = _mm_unpacklo_epi16(res02, res03); - const __m128i tr0_2 = _mm_unpackhi_epi16(res00, res01); - const __m128i tr0_3 = _mm_unpackhi_epi16(res02, res03); - const __m128i tr0_4 = _mm_unpacklo_epi16(res04, res05); - const __m128i tr0_5 = _mm_unpacklo_epi16(res06, res07); - const __m128i tr0_6 = _mm_unpackhi_epi16(res04, res05); - const __m128i tr0_7 = _mm_unpackhi_epi16(res06, res07); - // 00 10 01 11 02 12 03 13 - // 20 30 21 31 22 32 23 33 - // 04 14 05 15 06 16 07 17 - // 24 34 25 35 26 36 27 37 - // 40 50 41 51 42 52 43 53 - // 60 70 61 71 62 72 63 73 - // 54 54 55 55 56 56 57 57 - // 64 74 65 75 66 76 67 77 - const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); - const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); - const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); - const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); - const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); - const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); - const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); - const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); - // 00 10 20 30 01 11 21 31 - // 40 50 60 70 41 51 61 71 - // 02 12 22 32 03 13 23 33 - // 42 52 62 72 43 53 63 73 - // 04 14 24 34 05 15 21 36 - // 44 54 64 74 45 55 61 76 - // 06 16 26 36 07 17 27 37 - // 46 56 66 76 47 57 67 77 - const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4); - const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4); - const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6); - const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6); - const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5); - const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5); - const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7); - const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7); - // 00 10 20 30 40 50 60 70 - // 01 11 21 31 41 51 61 71 - // 02 12 22 32 42 52 62 72 - // 03 13 23 33 43 53 63 73 - // 04 14 24 34 44 54 64 74 - // 05 15 25 35 45 55 65 75 - // 06 16 26 36 46 56 66 76 - // 07 17 27 37 47 57 67 77 - _mm_storeu_si128((__m128i *)(out + 0 * 16), tr2_0); - _mm_storeu_si128((__m128i *)(out + 1 * 16), tr2_1); - _mm_storeu_si128((__m128i *)(out + 2 * 16), tr2_2); - _mm_storeu_si128((__m128i *)(out + 3 * 16), tr2_3); - _mm_storeu_si128((__m128i *)(out + 4 * 16), tr2_4); - _mm_storeu_si128((__m128i *)(out + 5 * 16), tr2_5); - _mm_storeu_si128((__m128i *)(out + 6 * 16), tr2_6); - _mm_storeu_si128((__m128i *)(out + 7 * 16), tr2_7); - } - { - // 00 01 02 03 04 05 06 07 - // 10 11 12 13 14 15 16 17 - // 20 21 22 23 24 25 26 27 - // 30 31 32 33 34 35 36 37 - // 40 41 42 43 44 45 46 47 - // 50 51 52 53 54 55 56 57 - // 60 61 62 63 64 65 66 67 - // 70 71 72 73 74 75 76 77 - const __m128i tr0_0 = _mm_unpacklo_epi16(res08, res09); - const __m128i tr0_1 = _mm_unpacklo_epi16(res10, res11); - const __m128i tr0_2 = _mm_unpackhi_epi16(res08, res09); - const __m128i tr0_3 = _mm_unpackhi_epi16(res10, res11); - const __m128i tr0_4 = _mm_unpacklo_epi16(res12, res13); - const __m128i tr0_5 = _mm_unpacklo_epi16(res14, res15); - const __m128i tr0_6 = _mm_unpackhi_epi16(res12, res13); - const __m128i tr0_7 = _mm_unpackhi_epi16(res14, res15); - // 00 10 01 11 02 12 03 13 - // 20 30 21 31 22 32 23 33 - // 04 14 05 15 06 16 07 17 - // 24 34 25 35 26 36 27 37 - // 40 50 41 51 42 52 43 53 - // 60 70 61 71 62 72 63 73 - // 54 54 55 55 56 56 57 57 - // 64 74 65 75 66 76 67 77 - const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); - const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); - const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); - const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); - const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); - const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); - const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); - const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); - // 00 10 20 30 01 11 21 31 - // 40 50 60 70 41 51 61 71 - // 02 12 22 32 03 13 23 33 - // 42 52 62 72 43 53 63 73 - // 04 14 24 34 05 15 21 36 - // 44 54 64 74 45 55 61 76 - // 06 16 26 36 07 17 27 37 - // 46 56 66 76 47 57 67 77 - const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4); - const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4); - const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6); - const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6); - const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5); - const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5); - const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7); - const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7); - // 00 10 20 30 40 50 60 70 - // 01 11 21 31 41 51 61 71 - // 02 12 22 32 42 52 62 72 - // 03 13 23 33 43 53 63 73 - // 04 14 24 34 44 54 64 74 - // 05 15 25 35 45 55 65 75 - // 06 16 26 36 46 56 66 76 - // 07 17 27 37 47 57 67 77 - // Store results - _mm_store_si128((__m128i *)(out + 8 + 0 * 16), tr2_0); - _mm_store_si128((__m128i *)(out + 8 + 1 * 16), tr2_1); - _mm_store_si128((__m128i *)(out + 8 + 2 * 16), tr2_2); - _mm_store_si128((__m128i *)(out + 8 + 3 * 16), tr2_3); - _mm_store_si128((__m128i *)(out + 8 + 4 * 16), tr2_4); - _mm_store_si128((__m128i *)(out + 8 + 5 * 16), tr2_5); - _mm_store_si128((__m128i *)(out + 8 + 6 * 16), tr2_6); - _mm_store_si128((__m128i *)(out + 8 + 7 * 16), tr2_7); - } - out += 8*16; - } - // Setup in/out for next pass. - in = intermediate; - out = output; - } + store_output(in1, output); } static INLINE void load_buffer_16x16(const int16_t* input, __m128i *in0, @@ -2334,7 +1307,7 @@ static INLINE void load_buffer_16x16(const int16_t* input, __m128i *in0, load_buffer_8x8(input + 8 * stride, in1 + 8, stride); } -static INLINE void write_buffer_16x16(int16_t *output, __m128i *in0, +static INLINE void write_buffer_16x16(tran_low_t *output, __m128i *in0, __m128i *in1, int stride) { // write first 8 columns write_buffer_8x8(output, in0, stride); @@ -2345,6 +1318,23 @@ static INLINE void write_buffer_16x16(int16_t *output, __m128i *in0, write_buffer_8x8(output + 8 * stride, in1 + 8, stride); } +static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) { + __m128i tbuf[8]; + array_transpose_8x8(res0, res0); + array_transpose_8x8(res1, tbuf); + array_transpose_8x8(res0 + 8, res1); + array_transpose_8x8(res1 + 8, res1 + 8); + + res0[8] = tbuf[0]; + res0[9] = tbuf[1]; + res0[10] = tbuf[2]; + res0[11] = tbuf[3]; + res0[12] = tbuf[4]; + res0[13] = tbuf[5]; + res0[14] = tbuf[6]; + res0[15] = tbuf[7]; +} + static INLINE void right_shift_16x16(__m128i *res0, __m128i *res1) { // perform rounding operations right_shift_8x8(res0, 2); @@ -3157,7 +2147,7 @@ void fadst16_sse2(__m128i *in0, __m128i *in1) { array_transpose_16x16(in0, in1); } -void vp9_fht16x16_sse2(const int16_t *input, int16_t *output, +void vp9_fht16x16_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type) { __m128i in0[16], in1[16]; @@ -3192,7 +2182,8 @@ void vp9_fht16x16_sse2(const int16_t *input, int16_t *output, } } -void vp9_fdct32x32_1_sse2(const int16_t *input, int16_t *output, int stride) { +void vp9_fdct32x32_1_sse2(const int16_t *input, tran_low_t *output, + int stride) { __m128i in0, in1, in2, in3; __m128i u0, u1; __m128i sum = _mm_setzero_si128(); @@ -3260,17 +2251,167 @@ void vp9_fdct32x32_1_sse2(const int16_t *input, int16_t *output, int stride) { in1 = _mm_add_epi32(sum, in0); in1 = _mm_srai_epi32(in1, 3); - _mm_store_si128((__m128i *)(output), in1); + store_output(in1, output); } +#if CONFIG_VP9_HIGHBITDEPTH +/* These SSE2 versions of the FHT functions only actually use SSE2 in the + * DCT_DCT case in all other cases, they revert to C code which is identical + * to that used by the C versions of them. + */ + +void vp9_highbd_fht4x4_sse2(const int16_t *input, tran_low_t *output, + int stride, int tx_type) { + if (tx_type == DCT_DCT) { + vp9_highbd_fdct4x4_sse2(input, output, stride); + } else { + tran_low_t out[4 * 4]; + tran_low_t *outptr = &out[0]; + int i, j; + tran_low_t temp_in[4], temp_out[4]; + const transform_2d ht = FHT_4[tx_type]; + + // Columns + for (i = 0; i < 4; ++i) { + for (j = 0; j < 4; ++j) + temp_in[j] = input[j * stride + i] * 16; + if (i == 0 && temp_in[0]) + temp_in[0] += 1; + ht.cols(temp_in, temp_out); + for (j = 0; j < 4; ++j) + outptr[j * 4 + i] = temp_out[j]; + } + + // Rows + for (i = 0; i < 4; ++i) { + for (j = 0; j < 4; ++j) + temp_in[j] = out[j + i * 4]; + ht.rows(temp_in, temp_out); + for (j = 0; j < 4; ++j) + output[j + i * 4] = (temp_out[j] + 1) >> 2; + } + } +} + +void vp9_highbd_fht8x8_sse2(const int16_t *input, tran_low_t *output, + int stride, int tx_type) { + if (tx_type == DCT_DCT) { + vp9_highbd_fdct8x8_sse2(input, output, stride); + } else { + tran_low_t out[64]; + tran_low_t *outptr = &out[0]; + int i, j; + tran_low_t temp_in[8], temp_out[8]; + const transform_2d ht = FHT_8[tx_type]; + + // Columns + for (i = 0; i < 8; ++i) { + for (j = 0; j < 8; ++j) + temp_in[j] = input[j * stride + i] * 4; + ht.cols(temp_in, temp_out); + for (j = 0; j < 8; ++j) + outptr[j * 8 + i] = temp_out[j]; + } + + // Rows + for (i = 0; i < 8; ++i) { + for (j = 0; j < 8; ++j) + temp_in[j] = out[j + i * 8]; + ht.rows(temp_in, temp_out); + for (j = 0; j < 8; ++j) + output[j + i * 8] = (temp_out[j] + (temp_out[j] < 0)) >> 1; + } + } +} + +void vp9_highbd_fht16x16_sse2(int16_t *input, tran_low_t *output, + int stride, int tx_type) { + if (tx_type == DCT_DCT) { + vp9_highbd_fdct16x16_sse2(input, output, stride); + } else { + tran_low_t out[256]; + tran_low_t *outptr = &out[0]; + int i, j; + tran_low_t temp_in[16], temp_out[16]; + const transform_2d ht = FHT_16[tx_type]; + + // Columns + for (i = 0; i < 16; ++i) { + for (j = 0; j < 16; ++j) + temp_in[j] = input[j * stride + i] * 4; + ht.cols(temp_in, temp_out); + for (j = 0; j < 16; ++j) + outptr[j * 16 + i] = (temp_out[j] + 1 + (temp_out[j] < 0)) >> 2; + } + + // Rows + for (i = 0; i < 16; ++i) { + for (j = 0; j < 16; ++j) + temp_in[j] = out[j + i * 16]; + ht.rows(temp_in, temp_out); + for (j = 0; j < 16; ++j) + output[j + i * 16] = temp_out[j]; + } + } +} +#endif // CONFIG_VP9_HIGHBITDEPTH + +/* + * The DCTnxn functions are defined using the macros below. The main code for + * them is in separate files (vp9/encoder/x86/vp9_dct_impl_sse2.c & + * vp9/encoder/x86/vp9_dct32x32_sse2.c) which are used by both the 8 bit code + * and the high bit depth code. + */ + +#define DCT_HIGH_BIT_DEPTH 0 + +#define FDCT4x4_2D vp9_fdct4x4_sse2 +#define FDCT8x8_2D vp9_fdct8x8_sse2 +#define FDCT16x16_2D vp9_fdct16x16_sse2 +#include "vp9/encoder/x86/vp9_dct_impl_sse2.c" +#undef FDCT4x4_2D +#undef FDCT8x8_2D +#undef FDCT16x16_2D + #define FDCT32x32_2D vp9_fdct32x32_rd_sse2 #define FDCT32x32_HIGH_PRECISION 0 #include "vp9/encoder/x86/vp9_dct32x32_sse2.c" -#undef FDCT32x32_HIGH_PRECISION #undef FDCT32x32_2D +#undef FDCT32x32_HIGH_PRECISION #define FDCT32x32_2D vp9_fdct32x32_sse2 #define FDCT32x32_HIGH_PRECISION 1 #include "vp9/encoder/x86/vp9_dct32x32_sse2.c" // NOLINT +#undef FDCT32x32_2D +#undef FDCT32x32_HIGH_PRECISION + +#undef DCT_HIGH_BIT_DEPTH + + +#if CONFIG_VP9_HIGHBITDEPTH + +#define DCT_HIGH_BIT_DEPTH 1 + +#define FDCT4x4_2D vp9_highbd_fdct4x4_sse2 +#define FDCT8x8_2D vp9_highbd_fdct8x8_sse2 +#define FDCT16x16_2D vp9_highbd_fdct16x16_sse2 +#include "vp9/encoder/x86/vp9_dct_impl_sse2.c" // NOLINT +#undef FDCT4x4_2D +#undef FDCT8x8_2D +#undef FDCT16x16_2D + +#define FDCT32x32_2D vp9_highbd_fdct32x32_rd_sse2 +#define FDCT32x32_HIGH_PRECISION 0 +#include "vp9/encoder/x86/vp9_dct32x32_sse2.c" // NOLINT +#undef FDCT32x32_2D #undef FDCT32x32_HIGH_PRECISION + +#define FDCT32x32_2D vp9_highbd_fdct32x32_sse2 +#define FDCT32x32_HIGH_PRECISION 1 +#include "vp9/encoder/x86/vp9_dct32x32_sse2.c" // NOLINT #undef FDCT32x32_2D +#undef FDCT32x32_HIGH_PRECISION + +#undef DCT_HIGH_BIT_DEPTH + +#endif // CONFIG_VP9_HIGHBITDEPTH diff --git a/vp9/encoder/x86/vp9_dct_sse2.h b/vp9/encoder/x86/vp9_dct_sse2.h new file mode 100644 index 000000000..2d322103e --- /dev/null +++ b/vp9/encoder/x86/vp9_dct_sse2.h @@ -0,0 +1,373 @@ +/* + * Copyright (c) 2014 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef VP9_ENCODER_X86_VP9_DCT_SSE2_H_ +#define VP9_ENCODER_X86_VP9_DCT_SSE2_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#define pair_set_epi32(a, b) \ + _mm_set_epi32((int)(b), (int)(a), (int)(b), (int)(a)) + +void vp9_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride); +void vp9_fdct8x8_sse2(const int16_t *input, tran_low_t *output, int stride); +void vp9_fdct16x16_sse2(const int16_t *input, tran_low_t *output, int stride); +void vp9_highbd_fdct4x4_sse2(const int16_t *input, tran_low_t *output, + int stride); +void vp9_highbd_fdct8x8_sse2(const int16_t *input, tran_low_t *output, + int stride); +void vp9_highbd_fdct16x16_sse2(const int16_t *input, tran_low_t *output, + int stride); + +static INLINE __m128i k_madd_epi32(__m128i a, __m128i b) { + __m128i buf0, buf1; + buf0 = _mm_mul_epu32(a, b); + a = _mm_srli_epi64(a, 32); + b = _mm_srli_epi64(b, 32); + buf1 = _mm_mul_epu32(a, b); + return _mm_add_epi64(buf0, buf1); +} + +static INLINE __m128i k_packs_epi64(__m128i a, __m128i b) { + __m128i buf0 = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0)); + __m128i buf1 = _mm_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0)); + return _mm_unpacklo_epi64(buf0, buf1); +} + +static INLINE int check_epi16_overflow_x2(__m128i reg0, __m128i reg1) { + const __m128i max_overflow = _mm_set1_epi16(0x7fff); + const __m128i min_overflow = _mm_set1_epi16(0x8000); + __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(reg0, max_overflow), + _mm_cmpeq_epi16(reg0, min_overflow)); + __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(reg1, max_overflow), + _mm_cmpeq_epi16(reg1, min_overflow)); + cmp0 = _mm_or_si128(cmp0, cmp1); + return _mm_movemask_epi8(cmp0); +} + +static INLINE int check_epi16_overflow_x4(__m128i reg0, __m128i reg1, + __m128i reg2, __m128i reg3) { + const __m128i max_overflow = _mm_set1_epi16(0x7fff); + const __m128i min_overflow = _mm_set1_epi16(0x8000); + __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(reg0, max_overflow), + _mm_cmpeq_epi16(reg0, min_overflow)); + __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(reg1, max_overflow), + _mm_cmpeq_epi16(reg1, min_overflow)); + __m128i cmp2 = _mm_or_si128(_mm_cmpeq_epi16(reg2, max_overflow), + _mm_cmpeq_epi16(reg2, min_overflow)); + __m128i cmp3 = _mm_or_si128(_mm_cmpeq_epi16(reg3, max_overflow), + _mm_cmpeq_epi16(reg3, min_overflow)); + cmp0 = _mm_or_si128(_mm_or_si128(cmp0, cmp1), _mm_or_si128(cmp2, cmp3)); + return _mm_movemask_epi8(cmp0); +} + +static INLINE int check_epi16_overflow_x8(__m128i reg0, __m128i reg1, + __m128i reg2, __m128i reg3, + __m128i reg4, __m128i reg5, + __m128i reg6, __m128i reg7) { + int res0, res1; + res0 = check_epi16_overflow_x4(reg0, reg1, reg2, reg3); + res1 = check_epi16_overflow_x4(reg4, reg5, reg6, reg7); + return res0 + res1; +} + +static INLINE int check_epi16_overflow_x12(__m128i reg0, __m128i reg1, + __m128i reg2, __m128i reg3, __m128i reg4, + __m128i reg5, __m128i reg6, __m128i reg7, + __m128i reg8, __m128i reg9, __m128i reg10, + __m128i reg11) { + int res0, res1; + res0 = check_epi16_overflow_x4(reg0, reg1, reg2, reg3); + res1 = check_epi16_overflow_x4(reg4, reg5, reg6, reg7); + if (!res0) + res0 = check_epi16_overflow_x4(reg8, reg9, reg10, reg11); + return res0 + res1; +} + +static INLINE int check_epi16_overflow_x16(__m128i reg0, __m128i reg1, + __m128i reg2, __m128i reg3, __m128i reg4, + __m128i reg5, __m128i reg6, __m128i reg7, + __m128i reg8, __m128i reg9, __m128i reg10, + __m128i reg11, __m128i reg12, __m128i reg13, + __m128i reg14, __m128i reg15) { + int res0, res1; + res0 = check_epi16_overflow_x4(reg0, reg1, reg2, reg3); + res1 = check_epi16_overflow_x4(reg4, reg5, reg6, reg7); + if (!res0) { + res0 = check_epi16_overflow_x4(reg8, reg9, reg10, reg11); + if (!res1) + res1 = check_epi16_overflow_x4(reg12, reg13, reg14, reg15); + } + return res0 + res1; +} + +static INLINE int check_epi16_overflow_x32(__m128i reg0, __m128i reg1, + __m128i reg2, __m128i reg3, __m128i reg4, + __m128i reg5, __m128i reg6, __m128i reg7, + __m128i reg8, __m128i reg9, __m128i reg10, + __m128i reg11, __m128i reg12, __m128i reg13, + __m128i reg14, __m128i reg15, __m128i reg16, + __m128i reg17, __m128i reg18, __m128i reg19, + __m128i reg20, __m128i reg21, __m128i reg22, + __m128i reg23, __m128i reg24, __m128i reg25, + __m128i reg26, __m128i reg27, __m128i reg28, + __m128i reg29, __m128i reg30, __m128i reg31) { + int res0, res1; + res0 = check_epi16_overflow_x4(reg0, reg1, reg2, reg3); + res1 = check_epi16_overflow_x4(reg4, reg5, reg6, reg7); + if (!res0) { + res0 = check_epi16_overflow_x4(reg8, reg9, reg10, reg11); + if (!res1) { + res1 = check_epi16_overflow_x4(reg12, reg13, reg14, reg15); + if (!res0) { + res0 = check_epi16_overflow_x4(reg16, reg17, reg18, reg19); + if (!res1) { + res1 = check_epi16_overflow_x4(reg20, reg21, reg22, reg23); + if (!res0) { + res0 = check_epi16_overflow_x4(reg24, reg25, reg26, reg27); + if (!res1) + res1 = check_epi16_overflow_x4(reg28, reg29, reg30, reg31); + } + } + } + } + } + return res0 + res1; +} + +static INLINE int k_check_epi32_overflow_4(__m128i reg0, __m128i reg1, + __m128i reg2, __m128i reg3, const __m128i* zero) { + __m128i minus_one = _mm_set1_epi32(-1); + // Check for overflows + __m128i reg0_shifted = _mm_slli_epi64(reg0, 1); + __m128i reg1_shifted = _mm_slli_epi64(reg1, 1); + __m128i reg2_shifted = _mm_slli_epi64(reg2, 1); + __m128i reg3_shifted = _mm_slli_epi64(reg3, 1); + __m128i reg0_top_dwords = _mm_shuffle_epi32( + reg0_shifted, _MM_SHUFFLE(0, 0, 3, 1)); + __m128i reg1_top_dwords = _mm_shuffle_epi32( + reg1_shifted, _MM_SHUFFLE(0, 0, 3, 1)); + __m128i reg2_top_dwords = _mm_shuffle_epi32( + reg2_shifted, _MM_SHUFFLE(0, 0, 3, 1)); + __m128i reg3_top_dwords = _mm_shuffle_epi32( + reg3_shifted, _MM_SHUFFLE(0, 0, 3, 1)); + __m128i top_dwords_01 = _mm_unpacklo_epi64(reg0_top_dwords, reg1_top_dwords); + __m128i top_dwords_23 = _mm_unpacklo_epi64(reg2_top_dwords, reg3_top_dwords); + __m128i valid_positve_01 = _mm_cmpeq_epi32(top_dwords_01, *zero); + __m128i valid_positve_23 = _mm_cmpeq_epi32(top_dwords_23, *zero); + __m128i valid_negative_01 = _mm_cmpeq_epi32(top_dwords_01, minus_one); + __m128i valid_negative_23 = _mm_cmpeq_epi32(top_dwords_23, minus_one); + int overflow_01 = _mm_movemask_epi8( + _mm_cmpeq_epi32(valid_positve_01, valid_negative_01)); + int overflow_23 = _mm_movemask_epi8( + _mm_cmpeq_epi32(valid_positve_23, valid_negative_23)); + return (overflow_01 + overflow_23); +} + +static INLINE int k_check_epi32_overflow_8(__m128i reg0, __m128i reg1, + __m128i reg2, __m128i reg3, + __m128i reg4, __m128i reg5, + __m128i reg6, __m128i reg7, + const __m128i* zero) { + int overflow = k_check_epi32_overflow_4(reg0, reg1, reg2, reg3, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(reg4, reg5, reg6, reg7, zero); + } + return overflow; +} + +static INLINE int k_check_epi32_overflow_16( + __m128i reg0, __m128i reg1, __m128i reg2, __m128i reg3, + __m128i reg4, __m128i reg5, __m128i reg6, __m128i reg7, + __m128i reg8, __m128i reg9, __m128i reg10, __m128i reg11, + __m128i reg12, __m128i reg13, __m128i reg14, __m128i reg15, + const __m128i* zero) { + int overflow = k_check_epi32_overflow_4(reg0, reg1, reg2, reg3, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(reg4, reg5, reg6, reg7, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(reg8, reg9, reg10, reg11, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(reg12, reg13, reg14, reg15, zero); + } + } + } + return overflow; +} + +static INLINE int k_check_epi32_overflow_32( + __m128i reg0, __m128i reg1, __m128i reg2, __m128i reg3, + __m128i reg4, __m128i reg5, __m128i reg6, __m128i reg7, + __m128i reg8, __m128i reg9, __m128i reg10, __m128i reg11, + __m128i reg12, __m128i reg13, __m128i reg14, __m128i reg15, + __m128i reg16, __m128i reg17, __m128i reg18, __m128i reg19, + __m128i reg20, __m128i reg21, __m128i reg22, __m128i reg23, + __m128i reg24, __m128i reg25, __m128i reg26, __m128i reg27, + __m128i reg28, __m128i reg29, __m128i reg30, __m128i reg31, + const __m128i* zero) { + int overflow = k_check_epi32_overflow_4(reg0, reg1, reg2, reg3, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(reg4, reg5, reg6, reg7, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(reg8, reg9, reg10, reg11, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(reg12, reg13, reg14, reg15, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(reg16, reg17, reg18, reg19, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(reg20, reg21, + reg22, reg23, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(reg24, reg25, + reg26, reg27, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(reg28, reg29, + reg30, reg31, zero); + } + } + } + } + } + } + } + return overflow; +} + +static INLINE void store_output(const __m128i output, tran_low_t* dst_ptr) { +#if CONFIG_VP9_HIGHBITDEPTH + const __m128i zero = _mm_setzero_si128(); + const __m128i sign_bits = _mm_cmplt_epi16(output, zero); + __m128i out0 = _mm_unpacklo_epi16(output, sign_bits); + __m128i out1 = _mm_unpackhi_epi16(output, sign_bits); + _mm_store_si128((__m128i *)(dst_ptr), out0); + _mm_store_si128((__m128i *)(dst_ptr + 4), out1); +#else + _mm_store_si128((__m128i *)(dst_ptr), output); +#endif // CONFIG_VP9_HIGHBITDEPTH +} + +static INLINE void storeu_output(const __m128i output, tran_low_t* dst_ptr) { +#if CONFIG_VP9_HIGHBITDEPTH + const __m128i zero = _mm_setzero_si128(); + const __m128i sign_bits = _mm_cmplt_epi16(output, zero); + __m128i out0 = _mm_unpacklo_epi16(output, sign_bits); + __m128i out1 = _mm_unpackhi_epi16(output, sign_bits); + _mm_storeu_si128((__m128i *)(dst_ptr), out0); + _mm_storeu_si128((__m128i *)(dst_ptr + 4), out1); +#else + _mm_storeu_si128((__m128i *)(dst_ptr), output); +#endif // CONFIG_VP9_HIGHBITDEPTH +} + + +static INLINE __m128i mult_round_shift(const __m128i in0, const __m128i in1, + const __m128i multiplier, + const __m128i rounding, + const int shift) { + const __m128i u0 = _mm_madd_epi16(in0, multiplier); + const __m128i u1 = _mm_madd_epi16(in1, multiplier); + const __m128i v0 = _mm_add_epi32(u0, rounding); + const __m128i v1 = _mm_add_epi32(u1, rounding); + const __m128i w0 = _mm_srai_epi32(v0, shift); + const __m128i w1 = _mm_srai_epi32(v1, shift); + return _mm_packs_epi32(w0, w1); +} + +static INLINE void transpose_and_output8x8( + const __m128i in00, const __m128i in01, + const __m128i in02, const __m128i in03, + const __m128i in04, const __m128i in05, + const __m128i in06, const __m128i in07, + const int pass, int16_t* out0_ptr, + tran_low_t* out1_ptr) { + // 00 01 02 03 04 05 06 07 + // 10 11 12 13 14 15 16 17 + // 20 21 22 23 24 25 26 27 + // 30 31 32 33 34 35 36 37 + // 40 41 42 43 44 45 46 47 + // 50 51 52 53 54 55 56 57 + // 60 61 62 63 64 65 66 67 + // 70 71 72 73 74 75 76 77 + const __m128i tr0_0 = _mm_unpacklo_epi16(in00, in01); + const __m128i tr0_1 = _mm_unpacklo_epi16(in02, in03); + const __m128i tr0_2 = _mm_unpackhi_epi16(in00, in01); + const __m128i tr0_3 = _mm_unpackhi_epi16(in02, in03); + const __m128i tr0_4 = _mm_unpacklo_epi16(in04, in05); + const __m128i tr0_5 = _mm_unpacklo_epi16(in06, in07); + const __m128i tr0_6 = _mm_unpackhi_epi16(in04, in05); + const __m128i tr0_7 = _mm_unpackhi_epi16(in06, in07); + // 00 10 01 11 02 12 03 13 + // 20 30 21 31 22 32 23 33 + // 04 14 05 15 06 16 07 17 + // 24 34 25 35 26 36 27 37 + // 40 50 41 51 42 52 43 53 + // 60 70 61 71 62 72 63 73 + // 54 54 55 55 56 56 57 57 + // 64 74 65 75 66 76 67 77 + const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); + const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); + const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); + const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); + const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); + const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); + const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); + const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); + // 00 10 20 30 01 11 21 31 + // 40 50 60 70 41 51 61 71 + // 02 12 22 32 03 13 23 33 + // 42 52 62 72 43 53 63 73 + // 04 14 24 34 05 15 21 36 + // 44 54 64 74 45 55 61 76 + // 06 16 26 36 07 17 27 37 + // 46 56 66 76 47 57 67 77 + const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4); + const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4); + const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6); + const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6); + const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5); + const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5); + const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7); + const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7); + // 00 10 20 30 40 50 60 70 + // 01 11 21 31 41 51 61 71 + // 02 12 22 32 42 52 62 72 + // 03 13 23 33 43 53 63 73 + // 04 14 24 34 44 54 64 74 + // 05 15 25 35 45 55 65 75 + // 06 16 26 36 46 56 66 76 + // 07 17 27 37 47 57 67 77 + if (pass == 0) { + _mm_storeu_si128((__m128i*)(out0_ptr + 0 * 16), tr2_0); + _mm_storeu_si128((__m128i*)(out0_ptr + 1 * 16), tr2_1); + _mm_storeu_si128((__m128i*)(out0_ptr + 2 * 16), tr2_2); + _mm_storeu_si128((__m128i*)(out0_ptr + 3 * 16), tr2_3); + _mm_storeu_si128((__m128i*)(out0_ptr + 4 * 16), tr2_4); + _mm_storeu_si128((__m128i*)(out0_ptr + 5 * 16), tr2_5); + _mm_storeu_si128((__m128i*)(out0_ptr + 6 * 16), tr2_6); + _mm_storeu_si128((__m128i*)(out0_ptr + 7 * 16), tr2_7); + } else { + storeu_output(tr2_0, (out1_ptr + 0 * 16)); + storeu_output(tr2_1, (out1_ptr + 1 * 16)); + storeu_output(tr2_2, (out1_ptr + 2 * 16)); + storeu_output(tr2_3, (out1_ptr + 3 * 16)); + storeu_output(tr2_4, (out1_ptr + 4 * 16)); + storeu_output(tr2_5, (out1_ptr + 5 * 16)); + storeu_output(tr2_6, (out1_ptr + 6 * 16)); + storeu_output(tr2_7, (out1_ptr + 7 * 16)); + } +} + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // VP9_ENCODER_X86_VP9_DCT_SSE2_H_ |