1 #include "precompiled.h"
2 //
3 // Copyright (c) 2002-2012 The ANGLE Project Authors. All rights reserved.
4 // Use of this source code is governed by a BSD-style license that can be
5 // found in the LICENSE file.
6 //
7
8 // ImageSSE2.cpp: Implements SSE2-based functions of rx::Image class. It's
9 // in a separated file for GCC, which can enable SSE usage only per-file,
10 // not for code blocks that use SSE2 explicitly.
11
12 #include "libGLESv2/Texture.h"
13 #include "libGLESv2/renderer/Image.h"
14
15 namespace rx
16 {
17
loadRGBAUByteDataToBGRASSE2(GLsizei width,GLsizei height,int inputPitch,const void * input,size_t outputPitch,void * output)18 void Image::loadRGBAUByteDataToBGRASSE2(GLsizei width, GLsizei height,
19 int inputPitch, const void *input, size_t outputPitch, void *output)
20 {
21 const unsigned int *source = NULL;
22 unsigned int *dest = NULL;
23 __m128i brMask = _mm_set1_epi32(0x00ff00ff);
24
25 for (int y = 0; y < height; y++)
26 {
27 source = reinterpret_cast<const unsigned int*>(static_cast<const unsigned char*>(input) + y * inputPitch);
28 dest = reinterpret_cast<unsigned int*>(static_cast<unsigned char*>(output) + y * outputPitch);
29 int x = 0;
30
31 // Make output writes aligned
32 for (x = 0; ((reinterpret_cast<intptr_t>(&dest[x]) & 15) != 0) && x < width; x++)
33 {
34 unsigned int rgba = source[x];
35 dest[x] = (_rotl(rgba, 16) & 0x00ff00ff) | (rgba & 0xff00ff00);
36 }
37
38 for (; x + 3 < width; x += 4)
39 {
40 __m128i sourceData = _mm_loadu_si128(reinterpret_cast<const __m128i*>(&source[x]));
41 // Mask out g and a, which don't change
42 __m128i gaComponents = _mm_andnot_si128(brMask, sourceData);
43 // Mask out b and r
44 __m128i brComponents = _mm_and_si128(sourceData, brMask);
45 // Swap b and r
46 __m128i brSwapped = _mm_shufflehi_epi16(_mm_shufflelo_epi16(brComponents, _MM_SHUFFLE(2, 3, 0, 1)), _MM_SHUFFLE(2, 3, 0, 1));
47 __m128i result = _mm_or_si128(gaComponents, brSwapped);
48 _mm_store_si128(reinterpret_cast<__m128i*>(&dest[x]), result);
49 }
50
51 // Perform leftover writes
52 for (; x < width; x++)
53 {
54 unsigned int rgba = source[x];
55 dest[x] = (_rotl(rgba, 16) & 0x00ff00ff) | (rgba & 0xff00ff00);
56 }
57 }
58 }
59
loadAlphaDataToBGRASSE2(GLsizei width,GLsizei height,int inputPitch,const void * input,size_t outputPitch,void * output)60 void Image::loadAlphaDataToBGRASSE2(GLsizei width, GLsizei height,
61 int inputPitch, const void *input, size_t outputPitch, void *output)
62 {
63 const unsigned char *source = NULL;
64 unsigned int *dest = NULL;
65 __m128i zeroWide = _mm_setzero_si128();
66
67 for (int y = 0; y < height; y++)
68 {
69 source = static_cast<const unsigned char*>(input) + y * inputPitch;
70 dest = reinterpret_cast<unsigned int*>(static_cast<unsigned char*>(output) + y * outputPitch);
71
72 int x;
73 // Make output writes aligned
74 for (x = 0; ((reinterpret_cast<intptr_t>(&dest[x]) & 0xF) != 0 && x < width); x++)
75 {
76 dest[x] = static_cast<unsigned int>(source[x]) << 24;
77 }
78
79 for (; x + 7 < width; x += 8)
80 {
81 __m128i sourceData = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(&source[x]));
82 // Interleave each byte to 16bit, make the lower byte to zero
83 sourceData = _mm_unpacklo_epi8(zeroWide, sourceData);
84 // Interleave each 16bit to 32bit, make the lower 16bit to zero
85 __m128i lo = _mm_unpacklo_epi16(zeroWide, sourceData);
86 __m128i hi = _mm_unpackhi_epi16(zeroWide, sourceData);
87
88 _mm_store_si128(reinterpret_cast<__m128i*>(&dest[x]), lo);
89 _mm_store_si128(reinterpret_cast<__m128i*>(&dest[x + 4]), hi);
90 }
91
92 // Handle the remainder
93 for (; x < width; x++)
94 {
95 dest[x] = static_cast<unsigned int>(source[x]) << 24;
96 }
97 }
98 }
99
100 }
101