1 // 2 // Copyright (c) 2002-2014 The ANGLE Project Authors. All rights reserved. 3 // Use of this source code is governed by a BSD-style license that can be 4 // found in the LICENSE file. 5 // 6 7 // loadimageSSE2.cpp: Defines image loading functions. It's 8 // in a separated file for GCC, which can enable SSE usage only per-file, 9 // not for code blocks that use SSE2 explicitly. 10 11 #include "libGLESv2/renderer/loadimage.h" 12 13 namespace rx 14 { 15 16 void LoadA8ToBGRA8_SSE2(size_t width, size_t height, size_t depth, 17 const uint8_t *input, size_t inputRowPitch, size_t inputDepthPitch, 18 uint8_t *output, size_t outputRowPitch, size_t outputDepthPitch) 19 { 20 __m128i zeroWide = _mm_setzero_si128(); 21 22 for (size_t z = 0; z < depth; z++) 23 { 24 for (size_t y = 0; y < height; y++) 25 { 26 const uint8_t *source = OffsetDataPointer<uint8_t>(input, y, z, inputRowPitch, inputDepthPitch); 27 uint32_t *dest = OffsetDataPointer<uint32_t>(output, y, z, outputRowPitch, outputDepthPitch); 28 29 size_t x = 0; 30 31 // Make output writes aligned 32 for (; ((reinterpret_cast<intptr_t>(&dest[x]) & 0xF) != 0 && x < width); x++) 33 { 34 dest[x] = static_cast<uint32_t>(source[x]) << 24; 35 } 36 37 for (; x + 7 < width; x += 8) 38 { 39 __m128i sourceData = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(&source[x])); 40 // Interleave each byte to 16bit, make the lower byte to zero 41 sourceData = _mm_unpacklo_epi8(zeroWide, sourceData); 42 // Interleave each 16bit to 32bit, make the lower 16bit to zero 43 __m128i lo = _mm_unpacklo_epi16(zeroWide, sourceData); 44 __m128i hi = _mm_unpackhi_epi16(zeroWide, sourceData); 45 46 _mm_store_si128(reinterpret_cast<__m128i*>(&dest[x]), lo); 47 _mm_store_si128(reinterpret_cast<__m128i*>(&dest[x + 4]), hi); 48 } 49 50 // Handle the remainder 51 for (; x < width; x++) 52 { 53 dest[x] = static_cast<uint32_t>(source[x]) << 24; 54 } 55 } 56 } 57 } 58 59 void LoadRGBA8ToBGRA8_SSE2(size_t width, size_t height, size_t depth, 60 const uint8_t *input, size_t inputRowPitch, size_t inputDepthPitch, 61 uint8_t *output, size_t outputRowPitch, size_t outputDepthPitch) 62 { 63 __m128i brMask = _mm_set1_epi32(0x00ff00ff); 64 65 for (size_t z = 0; z < depth; z++) 66 { 67 for (size_t y = 0; y < height; y++) 68 { 69 const uint32_t *source = OffsetDataPointer<uint32_t>(input, y, z, inputRowPitch, inputDepthPitch); 70 uint32_t *dest = OffsetDataPointer<uint32_t>(output, y, z, outputRowPitch, outputDepthPitch); 71 72 size_t x = 0; 73 74 // Make output writes aligned 75 for (; ((reinterpret_cast<intptr_t>(&dest[x]) & 15) != 0) && x < width; x++) 76 { 77 uint32_t rgba = source[x]; 78 dest[x] = (_rotl(rgba, 16) & 0x00ff00ff) | (rgba & 0xff00ff00); 79 } 80 81 for (; x + 3 < width; x += 4) 82 { 83 __m128i sourceData = _mm_loadu_si128(reinterpret_cast<const __m128i*>(&source[x])); 84 // Mask out g and a, which don't change 85 __m128i gaComponents = _mm_andnot_si128(brMask, sourceData); 86 // Mask out b and r 87 __m128i brComponents = _mm_and_si128(sourceData, brMask); 88 // Swap b and r 89 __m128i brSwapped = _mm_shufflehi_epi16(_mm_shufflelo_epi16(brComponents, _MM_SHUFFLE(2, 3, 0, 1)), _MM_SHUFFLE(2, 3, 0, 1)); 90 __m128i result = _mm_or_si128(gaComponents, brSwapped); 91 _mm_store_si128(reinterpret_cast<__m128i*>(&dest[x]), result); 92 } 93 94 // Perform leftover writes 95 for (; x < width; x++) 96 { 97 uint32_t rgba = source[x]; 98 dest[x] = (_rotl(rgba, 16) & 0x00ff00ff) | (rgba & 0xff00ff00); 99 } 100 } 101 } 102 } 103 104 } 105
