| OLD | NEW |
| 1 /* | 1 /* |
| 2 * Copyright 2009 The Android Open Source Project | 2 * Copyright 2009 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 "SkColorPriv.h" | 10 #include "SkColorPriv.h" |
| 11 #include "SkPaint.h" | 11 #include "SkPaint.h" |
| 12 #include "SkUtils.h" | 12 #include "SkUtils.h" |
| 13 | 13 |
| 14 void S32_opaque_D32_filter_DX_SSE2(const SkBitmapProcState& s, | 14 void S32_opaque_D32_filter_DX_SSE2(const SkBitmapProcState& s, |
| 15 const uint32_t* xy, | 15 const uint32_t* xy, |
| 16 int count, uint32_t* colors) { | 16 int count, uint32_t* colors) { |
| 17 SkASSERT(count > 0 && colors != NULL); | 17 SkASSERT(count > 0 && colors != NULL); |
| 18 SkASSERT(s.fFilterLevel != kNone_SkFilterQuality); | 18 SkASSERT(s.fFilterLevel != kNone_SkFilterQuality); |
| 19 SkASSERT(kN32_SkColorType == s.fBitmap->colorType()); | 19 SkASSERT(kN32_SkColorType == s.fPixmap.colorType()); |
| 20 SkASSERT(s.fAlphaScale == 256); | 20 SkASSERT(s.fAlphaScale == 256); |
| 21 | 21 |
| 22 const char* srcAddr = static_cast<const char*>(s.fBitmap->getPixels()); | 22 const char* srcAddr = static_cast<const char*>(s.fPixmap.addr()); |
| 23 size_t rb = s.fBitmap->rowBytes(); | 23 size_t rb = s.fPixmap.rowBytes(); |
| 24 uint32_t XY = *xy++; | 24 uint32_t XY = *xy++; |
| 25 unsigned y0 = XY >> 14; | 25 unsigned y0 = XY >> 14; |
| 26 const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4)
* rb); | 26 const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4)
* rb); |
| 27 const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3
FFF) * rb); | 27 const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3
FFF) * rb); |
| 28 unsigned subY = y0 & 0xF; | 28 unsigned subY = y0 & 0xF; |
| 29 | 29 |
| 30 // ( 0, 0, 0, 0, 0, 0, 0, 16) | 30 // ( 0, 0, 0, 0, 0, 0, 0, 16) |
| 31 __m128i sixteen = _mm_cvtsi32_si128(16); | 31 __m128i sixteen = _mm_cvtsi32_si128(16); |
| 32 | 32 |
| 33 // ( 0, 0, 0, 0, 16, 16, 16, 16) | 33 // ( 0, 0, 0, 0, 16, 16, 16, 16) |
| (...skipping 81 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 115 // Extract low int and store. | 115 // Extract low int and store. |
| 116 *colors++ = _mm_cvtsi128_si32(sum); | 116 *colors++ = _mm_cvtsi128_si32(sum); |
| 117 } while (--count > 0); | 117 } while (--count > 0); |
| 118 } | 118 } |
| 119 | 119 |
| 120 void S32_alpha_D32_filter_DX_SSE2(const SkBitmapProcState& s, | 120 void S32_alpha_D32_filter_DX_SSE2(const SkBitmapProcState& s, |
| 121 const uint32_t* xy, | 121 const uint32_t* xy, |
| 122 int count, uint32_t* colors) { | 122 int count, uint32_t* colors) { |
| 123 SkASSERT(count > 0 && colors != NULL); | 123 SkASSERT(count > 0 && colors != NULL); |
| 124 SkASSERT(s.fFilterLevel != kNone_SkFilterQuality); | 124 SkASSERT(s.fFilterLevel != kNone_SkFilterQuality); |
| 125 SkASSERT(kN32_SkColorType == s.fBitmap->colorType()); | 125 SkASSERT(kN32_SkColorType == s.fPixmap.colorType()); |
| 126 SkASSERT(s.fAlphaScale < 256); | 126 SkASSERT(s.fAlphaScale < 256); |
| 127 | 127 |
| 128 const char* srcAddr = static_cast<const char*>(s.fBitmap->getPixels()); | 128 const char* srcAddr = static_cast<const char*>(s.fPixmap.addr()); |
| 129 size_t rb = s.fBitmap->rowBytes(); | 129 size_t rb = s.fPixmap.rowBytes(); |
| 130 uint32_t XY = *xy++; | 130 uint32_t XY = *xy++; |
| 131 unsigned y0 = XY >> 14; | 131 unsigned y0 = XY >> 14; |
| 132 const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4)
* rb); | 132 const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4)
* rb); |
| 133 const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3
FFF) * rb); | 133 const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3
FFF) * rb); |
| 134 unsigned subY = y0 & 0xF; | 134 unsigned subY = y0 & 0xF; |
| 135 | 135 |
| 136 // ( 0, 0, 0, 0, 0, 0, 0, 16) | 136 // ( 0, 0, 0, 0, 0, 0, 0, 16) |
| 137 __m128i sixteen = _mm_cvtsi32_si128(16); | 137 __m128i sixteen = _mm_cvtsi32_si128(16); |
| 138 | 138 |
| 139 // ( 0, 0, 0, 0, 16, 16, 16, 16) | 139 // ( 0, 0, 0, 0, 16, 16, 16, 16) |
| (...skipping 102 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 242 | 242 |
| 243 /* SSE version of ClampX_ClampY_filter_scale() | 243 /* SSE version of ClampX_ClampY_filter_scale() |
| 244 * portable version is in core/SkBitmapProcState_matrix.h | 244 * portable version is in core/SkBitmapProcState_matrix.h |
| 245 */ | 245 */ |
| 246 void ClampX_ClampY_filter_scale_SSE2(const SkBitmapProcState& s, uint32_t xy[], | 246 void ClampX_ClampY_filter_scale_SSE2(const SkBitmapProcState& s, uint32_t xy[], |
| 247 int count, int x, int y) { | 247 int count, int x, int y) { |
| 248 SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask | | 248 SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask | |
| 249 SkMatrix::kScale_Mask)) == 0); | 249 SkMatrix::kScale_Mask)) == 0); |
| 250 SkASSERT(s.fInvKy == 0); | 250 SkASSERT(s.fInvKy == 0); |
| 251 | 251 |
| 252 const unsigned maxX = s.fBitmap->width() - 1; | 252 const unsigned maxX = s.fPixmap.width() - 1; |
| 253 const SkFixed one = s.fFilterOneX; | 253 const SkFixed one = s.fFilterOneX; |
| 254 const SkFixed dx = s.fInvSx; | 254 const SkFixed dx = s.fInvSx; |
| 255 SkFixed fx; | 255 SkFixed fx; |
| 256 | 256 |
| 257 SkPoint pt; | 257 SkPoint pt; |
| 258 s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf, | 258 s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf, |
| 259 SkIntToScalar(y) + SK_ScalarHalf, &pt); | 259 SkIntToScalar(y) + SK_ScalarHalf, &pt); |
| 260 const SkFixed fy = SkScalarToFixed(pt.fY) - (s.fFilterOneY >> 1); | 260 const SkFixed fy = SkScalarToFixed(pt.fY) - (s.fFilterOneY >> 1); |
| 261 const unsigned maxY = s.fBitmap->height() - 1; | 261 const unsigned maxY = s.fPixmap.height() - 1; |
| 262 // compute our two Y values up front | 262 // compute our two Y values up front |
| 263 *xy++ = ClampX_ClampY_pack_filter(fy, maxY, s.fFilterOneY); | 263 *xy++ = ClampX_ClampY_pack_filter(fy, maxY, s.fFilterOneY); |
| 264 // now initialize fx | 264 // now initialize fx |
| 265 fx = SkScalarToFixed(pt.fX) - (one >> 1); | 265 fx = SkScalarToFixed(pt.fX) - (one >> 1); |
| 266 | 266 |
| 267 // test if we don't need to apply the tile proc | 267 // test if we don't need to apply the tile proc |
| 268 if (dx > 0 && (unsigned)(fx >> 16) <= maxX && | 268 if (dx > 0 && (unsigned)(fx >> 16) <= maxX && |
| 269 (unsigned)((fx + dx * (count - 1)) >> 16) < maxX) { | 269 (unsigned)((fx + dx * (count - 1)) >> 16) < maxX) { |
| 270 if (count >= 4) { | 270 if (count >= 4) { |
| 271 // SSE version of decal_filter_scale | 271 // SSE version of decal_filter_scale |
| (...skipping 94 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 366 | 366 |
| 367 /* SSE version of ClampX_ClampY_nofilter_scale() | 367 /* SSE version of ClampX_ClampY_nofilter_scale() |
| 368 * portable version is in core/SkBitmapProcState_matrix.h | 368 * portable version is in core/SkBitmapProcState_matrix.h |
| 369 */ | 369 */ |
| 370 void ClampX_ClampY_nofilter_scale_SSE2(const SkBitmapProcState& s, | 370 void ClampX_ClampY_nofilter_scale_SSE2(const SkBitmapProcState& s, |
| 371 uint32_t xy[], int count, int x, int y) { | 371 uint32_t xy[], int count, int x, int y) { |
| 372 SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask | | 372 SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask | |
| 373 SkMatrix::kScale_Mask)) == 0); | 373 SkMatrix::kScale_Mask)) == 0); |
| 374 | 374 |
| 375 // we store y, x, x, x, x, x | 375 // we store y, x, x, x, x, x |
| 376 const unsigned maxX = s.fBitmap->width() - 1; | 376 const unsigned maxX = s.fPixmap.width() - 1; |
| 377 SkFixed fx; | 377 SkFixed fx; |
| 378 SkPoint pt; | 378 SkPoint pt; |
| 379 s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf, | 379 s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf, |
| 380 SkIntToScalar(y) + SK_ScalarHalf, &pt); | 380 SkIntToScalar(y) + SK_ScalarHalf, &pt); |
| 381 fx = SkScalarToFixed(pt.fY); | 381 fx = SkScalarToFixed(pt.fY); |
| 382 const unsigned maxY = s.fBitmap->height() - 1; | 382 const unsigned maxY = s.fPixmap.height() - 1; |
| 383 *xy++ = SkClampMax(fx >> 16, maxY); | 383 *xy++ = SkClampMax(fx >> 16, maxY); |
| 384 fx = SkScalarToFixed(pt.fX); | 384 fx = SkScalarToFixed(pt.fX); |
| 385 | 385 |
| 386 if (0 == maxX) { | 386 if (0 == maxX) { |
| 387 // all of the following X values must be 0 | 387 // all of the following X values must be 0 |
| 388 memset(xy, 0, count * sizeof(uint16_t)); | 388 memset(xy, 0, count * sizeof(uint16_t)); |
| 389 return; | 389 return; |
| 390 } | 390 } |
| 391 | 391 |
| 392 const SkFixed dx = s.fInvSx; | 392 const SkFixed dx = s.fInvSx; |
| (...skipping 100 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 493 s.fInvProc(s.fInvMatrix, | 493 s.fInvProc(s.fInvMatrix, |
| 494 SkIntToScalar(x) + SK_ScalarHalf, | 494 SkIntToScalar(x) + SK_ScalarHalf, |
| 495 SkIntToScalar(y) + SK_ScalarHalf, &srcPt); | 495 SkIntToScalar(y) + SK_ScalarHalf, &srcPt); |
| 496 | 496 |
| 497 SkFixed oneX = s.fFilterOneX; | 497 SkFixed oneX = s.fFilterOneX; |
| 498 SkFixed oneY = s.fFilterOneY; | 498 SkFixed oneY = s.fFilterOneY; |
| 499 SkFixed fx = SkScalarToFixed(srcPt.fX) - (oneX >> 1); | 499 SkFixed fx = SkScalarToFixed(srcPt.fX) - (oneX >> 1); |
| 500 SkFixed fy = SkScalarToFixed(srcPt.fY) - (oneY >> 1); | 500 SkFixed fy = SkScalarToFixed(srcPt.fY) - (oneY >> 1); |
| 501 SkFixed dx = s.fInvSx; | 501 SkFixed dx = s.fInvSx; |
| 502 SkFixed dy = s.fInvKy; | 502 SkFixed dy = s.fInvKy; |
| 503 unsigned maxX = s.fBitmap->width() - 1; | 503 unsigned maxX = s.fPixmap.width() - 1; |
| 504 unsigned maxY = s.fBitmap->height() - 1; | 504 unsigned maxY = s.fPixmap.height() - 1; |
| 505 | 505 |
| 506 if (count >= 2 && (maxX <= 0xFFFF)) { | 506 if (count >= 2 && (maxX <= 0xFFFF)) { |
| 507 SkFixed dx2 = dx + dx; | 507 SkFixed dx2 = dx + dx; |
| 508 SkFixed dy2 = dy + dy; | 508 SkFixed dy2 = dy + dy; |
| 509 | 509 |
| 510 __m128i wide_f = _mm_set_epi32(fx + dx, fy + dy, fx, fy); | 510 __m128i wide_f = _mm_set_epi32(fx + dx, fy + dy, fx, fy); |
| 511 __m128i wide_d2 = _mm_set_epi32(dx2, dy2, dx2, dy2); | 511 __m128i wide_d2 = _mm_set_epi32(dx2, dy2, dx2, dy2); |
| 512 __m128i wide_one = _mm_set_epi32(oneX, oneY, oneX, oneY); | 512 __m128i wide_one = _mm_set_epi32(oneX, oneY, oneX, oneY); |
| 513 __m128i wide_max = _mm_set_epi32(maxX, maxY, maxX, maxY); | 513 __m128i wide_max = _mm_set_epi32(maxX, maxY, maxX, maxY); |
| 514 __m128i wide_mask = _mm_set1_epi32(0xF); | 514 __m128i wide_mask = _mm_set1_epi32(0xF); |
| (...skipping 52 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 567 | 567 |
| 568 SkPoint srcPt; | 568 SkPoint srcPt; |
| 569 s.fInvProc(s.fInvMatrix, | 569 s.fInvProc(s.fInvMatrix, |
| 570 SkIntToScalar(x) + SK_ScalarHalf, | 570 SkIntToScalar(x) + SK_ScalarHalf, |
| 571 SkIntToScalar(y) + SK_ScalarHalf, &srcPt); | 571 SkIntToScalar(y) + SK_ScalarHalf, &srcPt); |
| 572 | 572 |
| 573 SkFixed fx = SkScalarToFixed(srcPt.fX); | 573 SkFixed fx = SkScalarToFixed(srcPt.fX); |
| 574 SkFixed fy = SkScalarToFixed(srcPt.fY); | 574 SkFixed fy = SkScalarToFixed(srcPt.fY); |
| 575 SkFixed dx = s.fInvSx; | 575 SkFixed dx = s.fInvSx; |
| 576 SkFixed dy = s.fInvKy; | 576 SkFixed dy = s.fInvKy; |
| 577 int maxX = s.fBitmap->width() - 1; | 577 int maxX = s.fPixmap.width() - 1; |
| 578 int maxY = s.fBitmap->height() - 1; | 578 int maxY = s.fPixmap.height() - 1; |
| 579 | 579 |
| 580 if (count >= 4 && (maxX <= 0xFFFF)) { | 580 if (count >= 4 && (maxX <= 0xFFFF)) { |
| 581 while (((size_t)xy & 0x0F) != 0) { | 581 while (((size_t)xy & 0x0F) != 0) { |
| 582 *xy++ = (SkClampMax(fy >> 16, maxY) << 16) | | 582 *xy++ = (SkClampMax(fy >> 16, maxY) << 16) | |
| 583 SkClampMax(fx >> 16, maxX); | 583 SkClampMax(fx >> 16, maxX); |
| 584 fx += dx; | 584 fx += dx; |
| 585 fy += dy; | 585 fy += dy; |
| 586 count--; | 586 count--; |
| 587 } | 587 } |
| 588 | 588 |
| (...skipping 46 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 635 | 635 |
| 636 /* SSE version of S32_D16_filter_DX_SSE2 | 636 /* SSE version of S32_D16_filter_DX_SSE2 |
| 637 * Definition is in section of "D16 functions for SRC == 8888" in SkBitmapProcS
tate.cpp | 637 * Definition is in section of "D16 functions for SRC == 8888" in SkBitmapProcS
tate.cpp |
| 638 * It combines S32_opaque_D32_filter_DX_SSE2 and SkPixel32ToPixel16 | 638 * It combines S32_opaque_D32_filter_DX_SSE2 and SkPixel32ToPixel16 |
| 639 */ | 639 */ |
| 640 void S32_D16_filter_DX_SSE2(const SkBitmapProcState& s, | 640 void S32_D16_filter_DX_SSE2(const SkBitmapProcState& s, |
| 641 const uint32_t* xy, | 641 const uint32_t* xy, |
| 642 int count, uint16_t* colors) { | 642 int count, uint16_t* colors) { |
| 643 SkASSERT(count > 0 && colors != NULL); | 643 SkASSERT(count > 0 && colors != NULL); |
| 644 SkASSERT(s.fFilterLevel != kNone_SkFilterQuality); | 644 SkASSERT(s.fFilterLevel != kNone_SkFilterQuality); |
| 645 SkASSERT(kN32_SkColorType == s.fBitmap->colorType()); | 645 SkASSERT(kN32_SkColorType == s.fPixmap.colorType()); |
| 646 SkASSERT(s.fBitmap->isOpaque()); | 646 SkASSERT(s.fPixmap.isOpaque()); |
| 647 | 647 |
| 648 SkPMColor dstColor; | 648 SkPMColor dstColor; |
| 649 const char* srcAddr = static_cast<const char*>(s.fBitmap->getPixels()); | 649 const char* srcAddr = static_cast<const char*>(s.fPixmap.addr()); |
| 650 size_t rb = s.fBitmap->rowBytes(); | 650 size_t rb = s.fPixmap.rowBytes(); |
| 651 uint32_t XY = *xy++; | 651 uint32_t XY = *xy++; |
| 652 unsigned y0 = XY >> 14; | 652 unsigned y0 = XY >> 14; |
| 653 const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4)
* rb); | 653 const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4)
* rb); |
| 654 const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3
FFF) * rb); | 654 const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3
FFF) * rb); |
| 655 unsigned subY = y0 & 0xF; | 655 unsigned subY = y0 & 0xF; |
| 656 | 656 |
| 657 // ( 0, 0, 0, 0, 0, 0, 0, 16) | 657 // ( 0, 0, 0, 0, 0, 0, 0, 16) |
| 658 __m128i sixteen = _mm_cvtsi32_si128(16); | 658 __m128i sixteen = _mm_cvtsi32_si128(16); |
| 659 | 659 |
| 660 // ( 0, 0, 0, 0, 16, 16, 16, 16) | 660 // ( 0, 0, 0, 0, 16, 16, 16, 16) |
| (...skipping 78 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 739 | 739 |
| 740 // Pack lower 4 16 bit values of sum into lower 4 bytes. | 740 // Pack lower 4 16 bit values of sum into lower 4 bytes. |
| 741 sum = _mm_packus_epi16(sum, zero); | 741 sum = _mm_packus_epi16(sum, zero); |
| 742 | 742 |
| 743 // Extract low int and store. | 743 // Extract low int and store. |
| 744 dstColor = _mm_cvtsi128_si32(sum); | 744 dstColor = _mm_cvtsi128_si32(sum); |
| 745 | 745 |
| 746 *colors++ = SkPixel32ToPixel16(dstColor); | 746 *colors++ = SkPixel32ToPixel16(dstColor); |
| 747 } while (--count > 0); | 747 } while (--count > 0); |
| 748 } | 748 } |
| OLD | NEW |