| OLD | NEW |
| 1 /* | 1 /* |
| 2 * Copyright 2011 Google Inc. | 2 * Copyright 2011 Google Inc. |
| 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 "SkBlitRect_opts_SSE2.h" | 9 #include "SkBlitRect_opts_SSE2.h" |
| 9 #include "SkBlitRow.h" | 10 #include "SkBlitRow.h" |
| 10 #include "SkColorPriv.h" | 11 #include "SkColorPriv.h" |
| 11 | 12 |
| 12 #include <emmintrin.h> | 13 /* Simple blitting of opaque rectangles less than 31 pixels wide: |
| 13 | 14 * inlines and merges sections of Color32_SSE2 and sk_memset32_SSE2. |
| 14 /** Simple blitting of opaque rectangles less than 31 pixels wide: | 15 */ |
| 15 inlines and merges sections of Color32_SSE2 and sk_memset32_SSE2. | |
| 16 */ | |
| 17 static void BlitRect32_OpaqueNarrow_SSE2(SkPMColor* SK_RESTRICT destination, | 16 static void BlitRect32_OpaqueNarrow_SSE2(SkPMColor* SK_RESTRICT destination, |
| 18 int width, int height, | 17 int width, int height, |
| 19 size_t rowBytes, uint32_t color) { | 18 size_t rowBytes, uint32_t color) { |
| 20 SkASSERT(255 == SkGetPackedA32(color)); | 19 SkASSERT(255 == SkGetPackedA32(color)); |
| 21 SkASSERT(width > 0); | 20 SkASSERT(width > 0); |
| 22 SkASSERT(width < 31); | 21 SkASSERT(width < 31); |
| 23 | 22 |
| 24 while (--height >= 0) { | 23 while (--height >= 0) { |
| 25 SkPMColor* dst = destination; | 24 SkPMColor* dst = destination; |
| 26 int count = width; | 25 int count = width; |
| 27 | 26 |
| 28 while (count > 4) { | 27 while (count > 4) { |
| 29 *dst++ = color; | 28 *dst++ = color; |
| 30 *dst++ = color; | 29 *dst++ = color; |
| 31 *dst++ = color; | 30 *dst++ = color; |
| 32 *dst++ = color; | 31 *dst++ = color; |
| 33 count -= 4; | 32 count -= 4; |
| 34 } | 33 } |
| 35 | 34 |
| 36 while (count > 0) { | 35 while (count > 0) { |
| 37 *dst++ = color; | 36 *dst++ = color; |
| 38 --count; | 37 --count; |
| 39 } | 38 } |
| 40 | 39 |
| 41 destination = (uint32_t*)((char*)destination + rowBytes); | 40 destination = (uint32_t*)((char*)destination + rowBytes); |
| 42 } | 41 } |
| 43 } | 42 } |
| 44 | 43 |
| 45 /** | 44 /* |
| 46 Fast blitting of opaque rectangles at least 31 pixels wide: | 45 * Fast blitting of opaque rectangles at least 31 pixels wide: |
| 47 inlines and merges sections of Color32_SSE2 and sk_memset32_SSE2. | 46 * inlines and merges sections of Color32_SSE2 and sk_memset32_SSE2. |
| 48 A 31 pixel rectangle is guaranteed to have at least one | 47 * A 31 pixel rectangle is guaranteed to have at least one |
| 49 16-pixel aligned span that can take advantage of mm_store. | 48 * 16-pixel aligned span that can take advantage of mm_store. |
| 50 */ | 49 */ |
| 51 static void BlitRect32_OpaqueWide_SSE2(SkPMColor* SK_RESTRICT destination, | 50 static void BlitRect32_OpaqueWide_SSE2(SkPMColor* SK_RESTRICT destination, |
| 52 int width, int height, | 51 int width, int height, |
| 53 size_t rowBytes, uint32_t color) { | 52 size_t rowBytes, uint32_t color) { |
| 54 SkASSERT(255 == SkGetPackedA32(color)); | 53 SkASSERT(255 == SkGetPackedA32(color)); |
| 55 SkASSERT(width >= 31); | 54 SkASSERT(width >= 31); |
| 56 | 55 |
| 57 __m128i color_wide = _mm_set1_epi32(color); | 56 __m128i color_wide = _mm_set1_epi32(color); |
| 58 while (--height >= 0) { | 57 while (--height >= 0) { |
| 59 // Prefetching one row ahead to L1 cache can equal hardware | 58 // Prefetching one row ahead to L1 cache can equal hardware |
| 60 // performance for large/tall rects, but never *beats* | 59 // performance for large/tall rects, but never *beats* |
| (...skipping 63 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 124 BlitRect32_OpaqueNarrow_SSE2(destination, width, height, | 123 BlitRect32_OpaqueNarrow_SSE2(destination, width, height, |
| 125 rowBytes, color); | 124 rowBytes, color); |
| 126 } else { | 125 } else { |
| 127 BlitRect32_OpaqueWide_SSE2(destination, width, height, | 126 BlitRect32_OpaqueWide_SSE2(destination, width, height, |
| 128 rowBytes, color); | 127 rowBytes, color); |
| 129 } | 128 } |
| 130 } else { | 129 } else { |
| 131 SkBlitRow::ColorRect32(destination, width, height, rowBytes, color); | 130 SkBlitRow::ColorRect32(destination, width, height, rowBytes, color); |
| 132 } | 131 } |
| 133 } | 132 } |
| OLD | NEW |