From a48b3854266a8b5c310d55960854816fcc45f536 Mon Sep 17 00:00:00 2001 From: homm Date: Fri, 23 Sep 2016 14:52:54 +0300 Subject: [PATCH] SIMD AlphaComposite. move declarations to beginning of the blocks --- src/libImaging/AlphaComposite.c | 80 +++++++++++++++++++-------------- 1 file changed, 46 insertions(+), 34 deletions(-) diff --git a/src/libImaging/AlphaComposite.c b/src/libImaging/AlphaComposite.c index 52fce4e47..0a1d7bf11 100644 --- a/src/libImaging/AlphaComposite.c +++ b/src/libImaging/AlphaComposite.c @@ -84,40 +84,46 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) { _mm256_srli_epi16(_mm256_add_epi16(src, _mm256_srli_epi16(src, 8)), 8) for (; x < xsize - 7; x += 8) { - __m256i mm_dst = _mm256_loadu_si256((__m256i *) &dst[x]); - __m256i mm_dst_lo = _mm256_unpacklo_epi8(mm_dst, vmm_zero); - __m256i mm_dst_hi = _mm256_unpackhi_epi8(mm_dst, vmm_zero); - __m256i mm_src = _mm256_loadu_si256((__m256i *) &src[x]); - __m256i mm_src_lo = _mm256_unpacklo_epi8(mm_src, vmm_zero); - __m256i mm_src_hi = _mm256_unpackhi_epi8(mm_src, vmm_zero); + __m256i mm_dst, mm_dst_lo, mm_dst_hi; + __m256i mm_src, mm_src_lo, mm_src_hi; + __m256i mm_dst_a, mm_src_a, mm_out_a, mm_blend; + __m256i mm_coef1, mm_coef2, mm_out_lo, mm_out_hi; - __m256i mm_dst_a = _mm256_srli_epi32(mm_dst, 24); - __m256i mm_src_a = _mm256_srli_epi32(mm_src, 24); + mm_dst = _mm256_loadu_si256((__m256i *) &dst[x]); + mm_dst_lo = _mm256_unpacklo_epi8(mm_dst, vmm_zero); + mm_dst_hi = _mm256_unpackhi_epi8(mm_dst, vmm_zero); + mm_src = _mm256_loadu_si256((__m256i *) &src[x]); + mm_src_lo = _mm256_unpacklo_epi8(mm_src, vmm_zero); + mm_src_hi = _mm256_unpackhi_epi8(mm_src, vmm_zero); + + mm_dst_a = _mm256_srli_epi32(mm_dst, 24); + mm_src_a = _mm256_srli_epi32(mm_src, 24); // Compute coefficients // blend = dst->a * (255 - src->a); 16 bits - __m256i mm_blend = _mm256_mullo_epi16(mm_dst_a, _mm256_sub_epi32(vmm_max_alpha, mm_src_a)); + mm_blend = _mm256_mullo_epi16(mm_dst_a, _mm256_sub_epi32(vmm_max_alpha, mm_src_a)); // outa = src->a * 255 + dst->a * (255 - src->a); 16 bits - __m256i mm_outa = _mm256_add_epi32(_mm256_mullo_epi16(mm_src_a, vmm_max_alpha), mm_blend); - __m256i mm_coef1 = _mm256_mullo_epi32(mm_src_a, vmm_max_alpha2); + mm_out_a = _mm256_add_epi32(_mm256_mullo_epi16(mm_src_a, vmm_max_alpha), mm_blend); + mm_coef1 = _mm256_mullo_epi32(mm_src_a, vmm_max_alpha2); // 8 bits - mm_coef1 = _mm256_cvtps_epi32(_mm256_div_ps(_mm256_cvtepi32_ps(mm_coef1), _mm256_cvtepi32_ps(mm_outa))); + mm_coef1 = _mm256_cvtps_epi32(_mm256_div_ps(_mm256_cvtepi32_ps(mm_coef1), + _mm256_cvtepi32_ps(mm_out_a))); // 8 bits - __m256i mm_coef2 = _mm256_sub_epi32(vmm_max_alpha, mm_coef1); + mm_coef2 = _mm256_sub_epi32(vmm_max_alpha, mm_coef1); - __m256i mm_out_lo = _mm256_add_epi16( + mm_out_lo = _mm256_add_epi16( _mm256_mullo_epi16(mm_src_lo, _mm256_shuffle_epi8(mm_coef1, vmm_get_lo)), _mm256_mullo_epi16(mm_dst_lo, _mm256_shuffle_epi8(mm_coef2, vmm_get_lo))); mm_out_lo = _mm256_or_si256(mm_out_lo, _mm256_slli_epi64( - _mm256_unpacklo_epi32(mm_outa, vmm_zero), 48)); + _mm256_unpacklo_epi32(mm_out_a, vmm_zero), 48)); mm_out_lo = _mm256_add_epi16(mm_out_lo, vmm_half); mm_out_lo = MM_SHIFTDIV255_epi16(mm_out_lo); - __m256i mm_out_hi = _mm256_add_epi16( + mm_out_hi = _mm256_add_epi16( _mm256_mullo_epi16(mm_src_hi, _mm256_shuffle_epi8(mm_coef1, vmm_get_hi)), _mm256_mullo_epi16(mm_dst_hi, _mm256_shuffle_epi8(mm_coef2, vmm_get_hi))); mm_out_hi = _mm256_or_si256(mm_out_hi, _mm256_slli_epi64( - _mm256_unpackhi_epi32(mm_outa, vmm_zero), 48)); + _mm256_unpackhi_epi32(mm_out_a, vmm_zero), 48)); mm_out_hi = _mm256_add_epi16(mm_out_hi, vmm_half); mm_out_hi = MM_SHIFTDIV255_epi16(mm_out_hi); @@ -133,48 +139,54 @@ ImagingAlphaComposite(Imaging imDst, Imaging imSrc) { _mm_srli_epi16(_mm_add_epi16(src, _mm_srli_epi16(src, 8)), 8) for (; x < xsize - 3; x += 4) { + __m128i mm_dst, mm_dst_lo, mm_dst_hi; + __m128i mm_src, mm_src_hi, mm_src_lo; + __m128i mm_dst_a, mm_src_a, mm_out_a, mm_blend; + __m128i mm_coef1, mm_coef2, mm_out_lo, mm_out_hi; + // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 - __m128i mm_dst = _mm_loadu_si128((__m128i *) &dst[x]); + mm_dst = _mm_loadu_si128((__m128i *) &dst[x]); // [16] a1 b1 g1 r1 a0 b0 g0 r0 - __m128i mm_dst_lo = _mm_unpacklo_epi8(mm_dst, mm_zero); + mm_dst_lo = _mm_unpacklo_epi8(mm_dst, mm_zero); // [16] a3 b3 g3 r3 a2 b2 g2 r2 - __m128i mm_dst_hi = _mm_unpackhi_epi8(mm_dst, mm_zero); + mm_dst_hi = _mm_unpackhi_epi8(mm_dst, mm_zero); // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0 - __m128i mm_src = _mm_loadu_si128((__m128i *) &src[x]); - __m128i mm_src_lo = _mm_unpacklo_epi8(mm_src, mm_zero); - __m128i mm_src_hi = _mm_unpackhi_epi8(mm_src, mm_zero); + mm_src = _mm_loadu_si128((__m128i *) &src[x]); + mm_src_lo = _mm_unpacklo_epi8(mm_src, mm_zero); + mm_src_hi = _mm_unpackhi_epi8(mm_src, mm_zero); // [32] a3 a2 a1 a0 - __m128i mm_dst_a = _mm_srli_epi32(mm_dst, 24); - __m128i mm_src_a = _mm_srli_epi32(mm_src, 24); + mm_dst_a = _mm_srli_epi32(mm_dst, 24); + mm_src_a = _mm_srli_epi32(mm_src, 24); // Compute coefficients // blend = dst->a * (255 - src->a) // [16] xx b3 xx b2 xx b1 xx b0 - __m128i mm_blend = _mm_mullo_epi16(mm_dst_a, _mm_sub_epi32(mm_max_alpha, mm_src_a)); + mm_blend = _mm_mullo_epi16(mm_dst_a, _mm_sub_epi32(mm_max_alpha, mm_src_a)); // outa = src->a * 255 + blend // [16] xx a3 xx a2 xx a1 xx a0 - __m128i mm_outa = _mm_add_epi32(_mm_mullo_epi16(mm_src_a, mm_max_alpha), mm_blend); + mm_out_a = _mm_add_epi32(_mm_mullo_epi16(mm_src_a, mm_max_alpha), mm_blend); // coef1 = src->a * 255 * 255 / outa - __m128i mm_coef1 = _mm_mullo_epi32(mm_src_a, mm_max_alpha2); + mm_coef1 = _mm_mullo_epi32(mm_src_a, mm_max_alpha2); // [8] xx xx xx c3 xx xx xx c2 xx xx xx c1 xx xx xx c0 - mm_coef1 = _mm_cvtps_epi32(_mm_div_ps(_mm_cvtepi32_ps(mm_coef1), _mm_cvtepi32_ps(mm_outa))); + mm_coef1 = _mm_cvtps_epi32(_mm_div_ps(_mm_cvtepi32_ps(mm_coef1), + _mm_cvtepi32_ps(mm_out_a))); // [8] xx xx xx c3 xx xx xx c2 xx xx xx c1 xx xx xx c0 - __m128i mm_coef2 = _mm_sub_epi32(mm_max_alpha, mm_coef1); + mm_coef2 = _mm_sub_epi32(mm_max_alpha, mm_coef1); - __m128i mm_out_lo = _mm_add_epi16( + mm_out_lo = _mm_add_epi16( _mm_mullo_epi16(mm_src_lo, _mm_shuffle_epi8(mm_coef1, mm_get_lo)), _mm_mullo_epi16(mm_dst_lo, _mm_shuffle_epi8(mm_coef2, mm_get_lo))); mm_out_lo = _mm_or_si128(mm_out_lo, _mm_slli_epi64( - _mm_unpacklo_epi32(mm_outa, mm_zero), 48)); + _mm_unpacklo_epi32(mm_out_a, mm_zero), 48)); mm_out_lo = _mm_add_epi16(mm_out_lo, mm_half); mm_out_lo = MM_SHIFTDIV255_epi16(mm_out_lo); - __m128i mm_out_hi = _mm_add_epi16( + mm_out_hi = _mm_add_epi16( _mm_mullo_epi16(mm_src_hi, _mm_shuffle_epi8(mm_coef1, mm_get_hi)), _mm_mullo_epi16(mm_dst_hi, _mm_shuffle_epi8(mm_coef2, mm_get_hi))); mm_out_hi = _mm_or_si128(mm_out_hi, _mm_slli_epi64( - _mm_unpackhi_epi32(mm_outa, mm_zero), 48)); + _mm_unpackhi_epi32(mm_out_a, mm_zero), 48)); mm_out_hi = _mm_add_epi16(mm_out_hi, mm_half); mm_out_hi = MM_SHIFTDIV255_epi16(mm_out_hi);