OLD | NEW |
1 /* | 1 /* |
2 * Copyright 2012 The Android Open Source Project | 2 * Copyright 2012 The Android Open Source Project |
3 * | 3 * |
4 * Use of this source code is governed by a BSD-style license that can be | 4 * Use of this source code is governed by a BSD-style license that can be |
5 * found in the LICENSE file. | 5 * found in the LICENSE file. |
6 */ | 6 */ |
7 | 7 |
8 #include <emmintrin.h> | 8 #include <emmintrin.h> |
9 #include "SkBitmapProcState_opts_SSE2.h" | 9 #include "SkBitmapProcState_opts_SSE2.h" |
10 #include "SkBlitRow_opts_SSE2.h" | 10 #include "SkBlitRow_opts_SSE2.h" |
11 #include "SkColorPriv.h" | 11 #include "SkColorPriv.h" |
12 #include "SkColor_opts_SSE2.h" | 12 #include "SkColor_opts_SSE2.h" |
13 #include "SkDither.h" | 13 #include "SkDither.h" |
14 #include "SkMSAN.h" | 14 #include "SkMSAN.h" |
15 #include "SkUtils.h" | 15 #include "SkUtils.h" |
16 | 16 |
17 /* SSE2 version of S32_Blend_BlitRow32() | 17 /* SSE2 version of S32_Blend_BlitRow32() |
18 * portable version is in core/SkBlitRow_D32.cpp | 18 * portable version is in core/SkBlitRow_D32.cpp |
19 */ | 19 */ |
20 void S32_Blend_BlitRow32_SSE2(SkPMColor* SK_RESTRICT dst, | 20 void S32_Blend_BlitRow32_SSE2(SkPMColor* SK_RESTRICT dst, |
21 const SkPMColor* SK_RESTRICT src, | 21 const SkPMColor* SK_RESTRICT src, |
22 int count, U8CPU alpha) { | 22 int count, U8CPU alpha) { |
23 SkASSERT(alpha <= 255); | 23 SkASSERT(alpha <= 255); |
24 if (count <= 0) { | 24 if (count <= 0) { |
25 return; | 25 return; |
26 } | 26 } |
27 | 27 |
28 uint32_t src_scale = SkAlpha255To256(alpha); | 28 uint32_t src_scale = SkAlpha255To256(alpha); |
29 uint32_t dst_scale = 256 - src_scale; | |
30 | 29 |
31 if (count >= 4) { | 30 if (count >= 4) { |
32 SkASSERT(((size_t)dst & 0x03) == 0); | 31 SkASSERT(((size_t)dst & 0x03) == 0); |
33 while (((size_t)dst & 0x0F) != 0) { | 32 while (((size_t)dst & 0x0F) != 0) { |
34 *dst = SkAlphaMulQ(*src, src_scale) + SkAlphaMulQ(*dst, dst_scale); | 33 *dst = SkPMLerp(*src, *dst, src_scale); |
35 src++; | 34 src++; |
36 dst++; | 35 dst++; |
37 count--; | 36 count--; |
38 } | 37 } |
39 | 38 |
40 const __m128i *s = reinterpret_cast<const __m128i*>(src); | 39 const __m128i *s = reinterpret_cast<const __m128i*>(src); |
41 __m128i *d = reinterpret_cast<__m128i*>(dst); | 40 __m128i *d = reinterpret_cast<__m128i*>(dst); |
42 | 41 |
43 while (count >= 4) { | 42 while (count >= 4) { |
44 // Load 4 pixels each of src and dest. | 43 // Load 4 pixels each of src and dest. |
45 __m128i src_pixel = _mm_loadu_si128(s); | 44 __m128i src_pixel = _mm_loadu_si128(s); |
46 __m128i dst_pixel = _mm_load_si128(d); | 45 __m128i dst_pixel = _mm_load_si128(d); |
47 | 46 |
48 src_pixel = SkAlphaMulQ_SSE2(src_pixel, src_scale); | 47 __m128i result = SkPMLerp_SSE2(src_pixel, dst_pixel, src_scale); |
49 dst_pixel = SkAlphaMulQ_SSE2(dst_pixel, dst_scale); | |
50 | |
51 // Add result | |
52 __m128i result = _mm_add_epi8(src_pixel, dst_pixel); | |
53 _mm_store_si128(d, result); | 48 _mm_store_si128(d, result); |
54 s++; | 49 s++; |
55 d++; | 50 d++; |
56 count -= 4; | 51 count -= 4; |
57 } | 52 } |
58 src = reinterpret_cast<const SkPMColor*>(s); | 53 src = reinterpret_cast<const SkPMColor*>(s); |
59 dst = reinterpret_cast<SkPMColor*>(d); | 54 dst = reinterpret_cast<SkPMColor*>(d); |
60 } | 55 } |
61 | 56 |
62 while (count > 0) { | 57 while (count > 0) { |
63 *dst = SkAlphaMulQ(*src, src_scale) + SkAlphaMulQ(*dst, dst_scale); | 58 *dst = SkPMLerp(*src, *dst, src_scale); |
64 src++; | 59 src++; |
65 dst++; | 60 dst++; |
66 count--; | 61 count--; |
67 } | 62 } |
68 } | 63 } |
69 | 64 |
70 void S32A_Blend_BlitRow32_SSE2(SkPMColor* SK_RESTRICT dst, | 65 void S32A_Blend_BlitRow32_SSE2(SkPMColor* SK_RESTRICT dst, |
71 const SkPMColor* SK_RESTRICT src, | 66 const SkPMColor* SK_RESTRICT src, |
72 int count, U8CPU alpha) { | 67 int count, U8CPU alpha) { |
73 SkASSERT(alpha <= 255); | 68 SkASSERT(alpha <= 255); |
(...skipping 912 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
986 uint32_t dst_expanded = SkExpand_rgb_16(*dst); | 981 uint32_t dst_expanded = SkExpand_rgb_16(*dst); |
987 dst_expanded = dst_expanded * (SkAlpha255To256(255 - a) >> 3); | 982 dst_expanded = dst_expanded * (SkAlpha255To256(255 - a) >> 3); |
988 // now src and dst expanded are in g:11 r:10 x:1 b:10 | 983 // now src and dst expanded are in g:11 r:10 x:1 b:10 |
989 *dst = SkCompact_rgb_16((src_expanded + dst_expanded) >> 5); | 984 *dst = SkCompact_rgb_16((src_expanded + dst_expanded) >> 5); |
990 } | 985 } |
991 dst += 1; | 986 dst += 1; |
992 DITHER_INC_X(x); | 987 DITHER_INC_X(x); |
993 } while (--count != 0); | 988 } while (--count != 0); |
994 } | 989 } |
995 } | 990 } |
OLD | NEW |