diff options
Diffstat (limited to 'vp9/encoder/x86')
-rw-r--r-- | vp9/encoder/x86/vp9_dct_sse2.c | 36 | ||||
-rw-r--r-- | vp9/encoder/x86/vp9_error_intrin_avx2.c | 72 | ||||
-rw-r--r-- | vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm | 49 |
3 files changed, 117 insertions, 40 deletions
diff --git a/vp9/encoder/x86/vp9_dct_sse2.c b/vp9/encoder/x86/vp9_dct_sse2.c index 686582238..1f58d872e 100644 --- a/vp9/encoder/x86/vp9_dct_sse2.c +++ b/vp9/encoder/x86/vp9_dct_sse2.c @@ -1187,7 +1187,7 @@ void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride) { 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_m24_m08 = 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); @@ -1513,8 +1513,8 @@ void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride) { 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_m24_m08); - const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m24_m08); + 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); @@ -1535,8 +1535,8 @@ void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride) { 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_m08_p24); - const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m08_p24); + 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); @@ -1554,10 +1554,10 @@ void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride) { { step1_0 = _mm_add_epi16(step3_0, step2_1); step1_1 = _mm_sub_epi16(step3_0, step2_1); - step1_2 = _mm_sub_epi16(step3_3, step2_2); - step1_3 = _mm_add_epi16(step3_3, step2_2); - step1_4 = _mm_add_epi16(step3_4, step2_5); - step1_5 = _mm_sub_epi16(step3_4, step2_5); + 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); } @@ -1848,7 +1848,7 @@ void fdct16_8col(__m128i *in) { const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); const __m128i k__cospi_m16_p16 = 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_m24_m08 = 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); @@ -2052,10 +2052,10 @@ void fdct16_8col(__m128i *in) { v[0] = _mm_madd_epi16(u[0], k__cospi_m08_p24); v[1] = _mm_madd_epi16(u[1], k__cospi_m08_p24); - v[2] = _mm_madd_epi16(u[2], k__cospi_m24_m08); - v[3] = _mm_madd_epi16(u[3], k__cospi_m24_m08); - v[4] = _mm_madd_epi16(u[2], k__cospi_m08_p24); - v[5] = _mm_madd_epi16(u[3], k__cospi_m08_p24); + v[2] = _mm_madd_epi16(u[2], k__cospi_p24_p08); + v[3] = _mm_madd_epi16(u[3], k__cospi_p24_p08); + v[4] = _mm_madd_epi16(u[2], k__cospi_p08_m24); + v[5] = _mm_madd_epi16(u[3], k__cospi_p08_m24); v[6] = _mm_madd_epi16(u[0], k__cospi_p24_p08); v[7] = _mm_madd_epi16(u[1], k__cospi_p24_p08); @@ -2085,10 +2085,10 @@ void fdct16_8col(__m128i *in) { // stage 5 s[0] = _mm_add_epi16(p[0], t[1]); s[1] = _mm_sub_epi16(p[0], t[1]); - s[2] = _mm_sub_epi16(p[3], t[2]); - s[3] = _mm_add_epi16(p[3], t[2]); - s[4] = _mm_add_epi16(p[4], t[5]); - s[5] = _mm_sub_epi16(p[4], t[5]); + s[2] = _mm_add_epi16(p[3], t[2]); + s[3] = _mm_sub_epi16(p[3], t[2]); + s[4] = _mm_sub_epi16(p[4], t[5]); + s[5] = _mm_add_epi16(p[4], t[5]); s[6] = _mm_sub_epi16(p[7], t[6]); s[7] = _mm_add_epi16(p[7], t[6]); diff --git a/vp9/encoder/x86/vp9_error_intrin_avx2.c b/vp9/encoder/x86/vp9_error_intrin_avx2.c new file mode 100644 index 000000000..c67490fad --- /dev/null +++ b/vp9/encoder/x86/vp9_error_intrin_avx2.c @@ -0,0 +1,72 @@ +/* + * Copyright (c) 2014 The WebM project authors. All Rights Reserved. + * + * Usee 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 <immintrin.h> // AVX2 +#include "vpx/vpx_integer.h" + + +int64_t vp9_block_error_avx2(const int16_t *coeff, + const int16_t *dqcoeff, + intptr_t block_size, + int64_t *ssz) { + __m256i sse_reg, ssz_reg, coeff_reg, dqcoeff_reg; + __m256i exp_dqcoeff_lo, exp_dqcoeff_hi, exp_coeff_lo, exp_coeff_hi; + __m256i sse_reg_64hi, ssz_reg_64hi; + __m128i sse_reg128, ssz_reg128; + int64_t sse; + int i; + const __m256i zero_reg = _mm256_set1_epi16(0); + + // init sse and ssz registerd to zero + sse_reg = _mm256_set1_epi16(0); + ssz_reg = _mm256_set1_epi16(0); + + for (i = 0 ; i < block_size ; i+= 16) { + // load 32 bytes from coeff and dqcoeff + coeff_reg = _mm256_loadu_si256((const __m256i *)(coeff + i)); + dqcoeff_reg = _mm256_loadu_si256((const __m256i *)(dqcoeff + i)); + // dqcoeff - coeff + dqcoeff_reg = _mm256_sub_epi16(dqcoeff_reg, coeff_reg); + // madd (dqcoeff - coeff) + dqcoeff_reg = _mm256_madd_epi16(dqcoeff_reg, dqcoeff_reg); + // madd coeff + coeff_reg = _mm256_madd_epi16(coeff_reg, coeff_reg); + // expand each double word of madd (dqcoeff - coeff) to quad word + exp_dqcoeff_lo = _mm256_unpacklo_epi32(dqcoeff_reg, zero_reg); + exp_dqcoeff_hi = _mm256_unpackhi_epi32(dqcoeff_reg, zero_reg); + // expand each double word of madd (coeff) to quad word + exp_coeff_lo = _mm256_unpacklo_epi32(coeff_reg, zero_reg); + exp_coeff_hi = _mm256_unpackhi_epi32(coeff_reg, zero_reg); + // add each quad word of madd (dqcoeff - coeff) and madd (coeff) + sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_lo); + ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_lo); + sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_hi); + ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_hi); + } + // save the higher 64 bit of each 128 bit lane + sse_reg_64hi = _mm256_srli_si256(sse_reg, 8); + ssz_reg_64hi = _mm256_srli_si256(ssz_reg, 8); + // add the higher 64 bit to the low 64 bit + sse_reg = _mm256_add_epi64(sse_reg, sse_reg_64hi); + ssz_reg = _mm256_add_epi64(ssz_reg, ssz_reg_64hi); + + // add each 64 bit from each of the 128 bit lane of the 256 bit + sse_reg128 = _mm_add_epi64(_mm256_castsi256_si128(sse_reg), + _mm256_extractf128_si256(sse_reg, 1)); + + ssz_reg128 = _mm_add_epi64(_mm256_castsi256_si128(ssz_reg), + _mm256_extractf128_si256(ssz_reg, 1)); + + // store the results + _mm_storel_epi64((__m128i*)(&sse), sse_reg128); + + _mm_storel_epi64((__m128i*)(ssz), ssz_reg128); + return sse; +} diff --git a/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm b/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm index d2d13b383..673e0b3a6 100644 --- a/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm +++ b/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm @@ -15,41 +15,45 @@ ; (unsigned char *frame1, | 0 ; unsigned int stride, | 1 ; unsigned char *frame2, | 2 -; unsigned int block_size, | 3 -; int strength, | 4 -; int filter_weight, | 5 -; unsigned int *accumulator, | 6 -; unsigned short *count) | 7 +; unsigned int block_width, | 3 +; unsigned int block_height, | 4 +; int strength, | 5 +; int filter_weight, | 6 +; unsigned int *accumulator, | 7 +; unsigned short *count) | 8 global sym(vp9_temporal_filter_apply_sse2) PRIVATE sym(vp9_temporal_filter_apply_sse2): push rbp mov rbp, rsp - SHADOW_ARGS_TO_STACK 8 + SHADOW_ARGS_TO_STACK 9 SAVE_XMM 7 GET_GOT rbx push rsi push rdi ALIGN_STACK 16, rax - %define block_size 0 - %define strength 16 - %define filter_weight 32 - %define rounding_bit 48 - %define rbp_backup 64 - %define stack_size 80 + %define block_width 0 + %define block_height 16 + %define strength 32 + %define filter_weight 48 + %define rounding_bit 64 + %define rbp_backup 80 + %define stack_size 96 sub rsp, stack_size mov [rsp + rbp_backup], rbp ; end prolog mov rdx, arg(3) - mov [rsp + block_size], rdx - movd xmm6, arg(4) + mov [rsp + block_width], rdx + mov rdx, arg(4) + mov [rsp + block_height], rdx + movd xmm6, arg(5) movdqa [rsp + strength], xmm6 ; where strength is used, all 16 bytes are read ; calculate the rounding bit outside the loop ; 0x8000 >> (16 - strength) mov rdx, 16 - sub rdx, arg(4) ; 16 - strength + sub rdx, arg(5) ; 16 - strength movq xmm4, rdx ; can't use rdx w/ shift movdqa xmm5, [GLOBAL(_const_top_bit)] psrlw xmm5, xmm4 @@ -57,11 +61,11 @@ sym(vp9_temporal_filter_apply_sse2): mov rsi, arg(0) ; src/frame1 mov rdx, arg(2) ; predictor frame - mov rdi, arg(6) ; accumulator - mov rax, arg(7) ; count + mov rdi, arg(7) ; accumulator + mov rax, arg(8) ; count ; dup the filter weight and store for later - movd xmm0, arg(5) ; filter_weight + movd xmm0, arg(6) ; filter_weight pshuflw xmm0, xmm0, 0 punpcklwd xmm0, xmm0 movdqa [rsp + filter_weight], xmm0 @@ -69,10 +73,11 @@ sym(vp9_temporal_filter_apply_sse2): mov rbp, arg(1) ; stride pxor xmm7, xmm7 ; zero for extraction - lea rcx, [rdx + 16*16*1] - cmp dword ptr [rsp + block_size], 8 + mov rcx, [rsp + block_width] + imul rcx, [rsp + block_height] + add rcx, rdx + cmp dword ptr [rsp + block_width], 8 jne .temporal_filter_apply_load_16 - lea rcx, [rdx + 8*8*1] .temporal_filter_apply_load_8: movq xmm0, [rsi] ; first row @@ -178,7 +183,7 @@ sym(vp9_temporal_filter_apply_sse2): cmp rdx, rcx je .temporal_filter_apply_epilog pxor xmm7, xmm7 ; zero for extraction - cmp dword ptr [rsp + block_size], 16 + cmp dword ptr [rsp + block_width], 16 je .temporal_filter_apply_load_16 jmp .temporal_filter_apply_load_8 |