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 | 8 #include <emmintrin.h> |
| 9 #include "SkBitmapProcState_opts_SSE2.h" |
9 #include "SkBlitRow_opts_SSE2.h" | 10 #include "SkBlitRow_opts_SSE2.h" |
10 #include "SkBitmapProcState_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 "SkUtils.h" | 14 #include "SkUtils.h" |
15 | 15 |
16 #include <emmintrin.h> | |
17 | |
18 /* SSE2 version of S32_Blend_BlitRow32() | 16 /* SSE2 version of S32_Blend_BlitRow32() |
19 * portable version is in core/SkBlitRow_D32.cpp | 17 * portable version is in core/SkBlitRow_D32.cpp |
20 */ | 18 */ |
21 void S32_Blend_BlitRow32_SSE2(SkPMColor* SK_RESTRICT dst, | 19 void S32_Blend_BlitRow32_SSE2(SkPMColor* SK_RESTRICT dst, |
22 const SkPMColor* SK_RESTRICT src, | 20 const SkPMColor* SK_RESTRICT src, |
23 int count, U8CPU alpha) { | 21 int count, U8CPU alpha) { |
24 SkASSERT(alpha <= 255); | 22 SkASSERT(alpha <= 255); |
25 if (count <= 0) { | 23 if (count <= 0) { |
26 return; | 24 return; |
27 } | 25 } |
(...skipping 144 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
172 // Combine back into RGBA. | 170 // Combine back into RGBA. |
173 dst_pixel = _mm_or_si128(dst_rb, dst_ag); | 171 dst_pixel = _mm_or_si128(dst_rb, dst_ag); |
174 | 172 |
175 // Add result | 173 // Add result |
176 __m128i result = _mm_add_epi8(src_pixel, dst_pixel); | 174 __m128i result = _mm_add_epi8(src_pixel, dst_pixel); |
177 _mm_store_si128(d, result); | 175 _mm_store_si128(d, result); |
178 s++; | 176 s++; |
179 d++; | 177 d++; |
180 count -= 4; | 178 count -= 4; |
181 } | 179 } |
182 #else | 180 #else |
183 __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); | 181 __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); |
184 __m128i c_256 = _mm_set1_epi16(0x0100); // 8 copies of 256 (16-bit) | 182 __m128i c_256 = _mm_set1_epi16(0x0100); // 8 copies of 256 (16-bit) |
185 while (count >= 4) { | 183 while (count >= 4) { |
186 // Load 4 pixels | 184 // Load 4 pixels |
187 __m128i src_pixel = _mm_loadu_si128(s); | 185 __m128i src_pixel = _mm_loadu_si128(s); |
188 __m128i dst_pixel = _mm_load_si128(d); | 186 __m128i dst_pixel = _mm_load_si128(d); |
189 | 187 |
190 __m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel); | 188 __m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel); |
191 __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8); | 189 __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8); |
192 | 190 |
(...skipping 142 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
335 dst++; | 333 dst++; |
336 count--; | 334 count--; |
337 } | 335 } |
338 } | 336 } |
339 | 337 |
340 /* SSE2 version of Color32() | 338 /* SSE2 version of Color32() |
341 * portable version is in core/SkBlitRow_D32.cpp | 339 * portable version is in core/SkBlitRow_D32.cpp |
342 */ | 340 */ |
343 void Color32_SSE2(SkPMColor dst[], const SkPMColor src[], int count, | 341 void Color32_SSE2(SkPMColor dst[], const SkPMColor src[], int count, |
344 SkPMColor color) { | 342 SkPMColor color) { |
345 | |
346 if (count <= 0) { | 343 if (count <= 0) { |
347 return; | 344 return; |
348 } | 345 } |
349 | 346 |
350 if (0 == color) { | 347 if (0 == color) { |
351 if (src != dst) { | 348 if (src != dst) { |
352 memcpy(dst, src, count * sizeof(SkPMColor)); | 349 memcpy(dst, src, count * sizeof(SkPMColor)); |
353 } | 350 } |
354 return; | 351 return; |
355 } | 352 } |
(...skipping 43 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
399 __m128i result = _mm_add_epi8(color_wide, src_pixel); | 396 __m128i result = _mm_add_epi8(color_wide, src_pixel); |
400 | 397 |
401 // Store result. | 398 // Store result. |
402 _mm_store_si128(d, result); | 399 _mm_store_si128(d, result); |
403 s++; | 400 s++; |
404 d++; | 401 d++; |
405 count -= 4; | 402 count -= 4; |
406 } | 403 } |
407 src = reinterpret_cast<const SkPMColor*>(s); | 404 src = reinterpret_cast<const SkPMColor*>(s); |
408 dst = reinterpret_cast<SkPMColor*>(d); | 405 dst = reinterpret_cast<SkPMColor*>(d); |
409 } | 406 } |
410 | 407 |
411 while (count > 0) { | 408 while (count > 0) { |
412 *dst = color + SkAlphaMulQ(*src, scale); | 409 *dst = color + SkAlphaMulQ(*src, scale); |
413 src += 1; | 410 src += 1; |
414 dst += 1; | 411 dst += 1; |
415 count--; | 412 count--; |
416 } | 413 } |
417 } | 414 } |
418 } | 415 } |
419 | 416 |
(...skipping 77 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
497 // Add two pixels into result. | 494 // Add two pixels into result. |
498 __m128i result = _mm_add_epi8(tmp_src_pixel, dst_pixel); | 495 __m128i result = _mm_add_epi8(tmp_src_pixel, dst_pixel); |
499 _mm_store_si128(d, result); | 496 _mm_store_si128(d, result); |
500 // load the next 4 pixel | 497 // load the next 4 pixel |
501 mask = mask + 4; | 498 mask = mask + 4; |
502 d++; | 499 d++; |
503 count -= 4; | 500 count -= 4; |
504 } | 501 } |
505 dst = reinterpret_cast<SkPMColor *>(d); | 502 dst = reinterpret_cast<SkPMColor *>(d); |
506 } | 503 } |
507 while(count > 0) { | 504 while (count > 0) { |
508 *dst= SkBlendARGB32(color, *dst, *mask); | 505 *dst= SkBlendARGB32(color, *dst, *mask); |
509 dst += 1; | 506 dst += 1; |
510 mask++; | 507 mask++; |
511 count --; | 508 count --; |
512 } | 509 } |
513 dst = (SkPMColor *)((char*)dst + dstOffset); | 510 dst = (SkPMColor *)((char*)dst + dstOffset); |
514 mask += maskOffset; | 511 mask += maskOffset; |
515 } while (--height != 0); | 512 } while (--height != 0); |
516 } | 513 } |
517 | 514 |
(...skipping 837 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
1355 uint32_t dst_expanded = SkExpand_rgb_16(*dst); | 1352 uint32_t dst_expanded = SkExpand_rgb_16(*dst); |
1356 dst_expanded = dst_expanded * (SkAlpha255To256(255 - a) >> 3); | 1353 dst_expanded = dst_expanded * (SkAlpha255To256(255 - a) >> 3); |
1357 // now src and dst expanded are in g:11 r:10 x:1 b:10 | 1354 // now src and dst expanded are in g:11 r:10 x:1 b:10 |
1358 *dst = SkCompact_rgb_16((src_expanded + dst_expanded) >> 5); | 1355 *dst = SkCompact_rgb_16((src_expanded + dst_expanded) >> 5); |
1359 } | 1356 } |
1360 dst += 1; | 1357 dst += 1; |
1361 DITHER_INC_X(x); | 1358 DITHER_INC_X(x); |
1362 } while (--count != 0); | 1359 } while (--count != 0); |
1363 } | 1360 } |
1364 } | 1361 } |
OLD | NEW |