SkBitmapProcState.cpp revision 0c2da0c99f3553d406e937fc9026f31cfc77507f
1/*
2 * Copyright 2011 Google Inc.
3 *
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#include "SkBitmapCache.h"
9#include "SkBitmapProcState.h"
10#include "SkColorPriv.h"
11#include "SkFilterProc.h"
12#include "SkPaint.h"
13#include "SkShader.h"   // for tilemodes
14#include "SkUtilsArm.h"
15#include "SkBitmapScaler.h"
16#include "SkMipMap.h"
17#include "SkPixelRef.h"
18#include "SkImageEncoder.h"
19#include "SkResourceCache.h"
20
21#if !SK_ARM_NEON_IS_NONE
22// These are defined in src/opts/SkBitmapProcState_arm_neon.cpp
23extern const SkBitmapProcState::SampleProc16 gSkBitmapProcStateSample16_neon[];
24extern const SkBitmapProcState::SampleProc32 gSkBitmapProcStateSample32_neon[];
25extern void  S16_D16_filter_DX_neon(const SkBitmapProcState&, const uint32_t*, int, uint16_t*);
26extern void  Clamp_S16_D16_filter_DX_shaderproc_neon(const SkBitmapProcState&, int, int, uint16_t*, int);
27extern void  Repeat_S16_D16_filter_DX_shaderproc_neon(const SkBitmapProcState&, int, int, uint16_t*, int);
28extern void  SI8_opaque_D32_filter_DX_neon(const SkBitmapProcState&, const uint32_t*, int, SkPMColor*);
29extern void  SI8_opaque_D32_filter_DX_shaderproc_neon(const SkBitmapProcState&, int, int, uint32_t*, int);
30extern void  Clamp_SI8_opaque_D32_filter_DX_shaderproc_neon(const SkBitmapProcState&, int, int, uint32_t*, int);
31#endif
32
33extern void Clamp_S32_opaque_D32_nofilter_DX_shaderproc(const SkBitmapProcState&, int, int, uint32_t*, int);
34
35#define   NAME_WRAP(x)  x
36#include "SkBitmapProcState_filter.h"
37#include "SkBitmapProcState_procs.h"
38
39///////////////////////////////////////////////////////////////////////////////
40
41// true iff the matrix contains, at most, scale and translate elements
42static bool matrix_only_scale_translate(const SkMatrix& m) {
43    return m.getType() <= (SkMatrix::kScale_Mask | SkMatrix::kTranslate_Mask);
44}
45
46/**
47 *  For the purposes of drawing bitmaps, if a matrix is "almost" translate
48 *  go ahead and treat it as if it were, so that subsequent code can go fast.
49 */
50static bool just_trans_clamp(const SkMatrix& matrix, const SkBitmap& bitmap) {
51    SkASSERT(matrix_only_scale_translate(matrix));
52
53    if (matrix.getType() & SkMatrix::kScale_Mask) {
54        SkRect src, dst;
55        bitmap.getBounds(&src);
56
57        // Can't call mapRect(), since that will fix up inverted rectangles,
58        // e.g. when scale is negative, and we don't want to return true for
59        // those.
60        matrix.mapPoints(SkTCast<SkPoint*>(&dst),
61                         SkTCast<const SkPoint*>(&src),
62                         2);
63
64        // Now round all 4 edges to device space, and then compare the device
65        // width/height to the original. Note: we must map all 4 and subtract
66        // rather than map the "width" and compare, since we care about the
67        // phase (in pixel space) that any translate in the matrix might impart.
68        SkIRect idst;
69        dst.round(&idst);
70        return idst.width() == bitmap.width() && idst.height() == bitmap.height();
71    }
72    // if we got here, we're either kTranslate_Mask or identity
73    return true;
74}
75
76static bool just_trans_general(const SkMatrix& matrix) {
77    SkASSERT(matrix_only_scale_translate(matrix));
78
79    if (matrix.getType() & SkMatrix::kScale_Mask) {
80        const SkScalar tol = SK_Scalar1 / 32768;
81
82        if (!SkScalarNearlyZero(matrix[SkMatrix::kMScaleX] - SK_Scalar1, tol)) {
83            return false;
84        }
85        if (!SkScalarNearlyZero(matrix[SkMatrix::kMScaleY] - SK_Scalar1, tol)) {
86            return false;
87        }
88    }
89    // if we got here, treat us as either kTranslate_Mask or identity
90    return true;
91}
92
93///////////////////////////////////////////////////////////////////////////////
94
95static bool valid_for_filtering(unsigned dimension) {
96    // for filtering, width and height must fit in 14bits, since we use steal
97    // 2 bits from each to store our 4bit subpixel data
98    return (dimension & ~0x3FFF) == 0;
99}
100
101static SkScalar effective_matrix_scale(const SkMatrix& mat) {
102    SkScalar dx = SkVector::Length(mat.getScaleX(), mat.getSkewY());
103    SkScalar dy = SkVector::Length(mat.getSkewX(), mat.getScaleY());
104#ifdef SK_SUPPORT_LEGACY_MIPMAP_EFFECTIVE_SCALE
105    return SkMaxScalar(dx, dy);
106#else
107    return SkScalarSqrt(dx * dy);
108#endif
109}
110
111// Check to see that the size of the bitmap that would be produced by
112// scaling by the given inverted matrix is less than the maximum allowed.
113static inline bool cache_size_okay(const SkBitmap& bm, const SkMatrix& invMat) {
114    size_t maximumAllocation = SkResourceCache::GetEffectiveSingleAllocationByteLimit();
115    if (0 == maximumAllocation) {
116        return true;
117    }
118    // float matrixScaleFactor = 1.0 / (invMat.scaleX * invMat.scaleY);
119    // return ((origBitmapSize * matrixScaleFactor) < maximumAllocationSize);
120    // Skip the division step:
121    return bm.info().getSafeSize(bm.info().minRowBytes())
122        < (maximumAllocation * invMat.getScaleX() * invMat.getScaleY());
123}
124
125/*
126 *  Extract the "best" scale factors from a matrix.
127 */
128static bool extract_scale(const SkMatrix& matrix, SkVector* scale) {
129    SkASSERT(!matrix.hasPerspective());
130    SkScalar sx = SkPoint::Length(matrix[SkMatrix::kMScaleX], matrix[SkMatrix::kMSkewY]);
131    SkScalar sy = SkPoint::Length(matrix[SkMatrix::kMSkewX],  matrix[SkMatrix::kMScaleY]);
132    if (!SkScalarIsFinite(sx) || !SkScalarIsFinite(sy) ||
133        SkScalarNearlyZero(sx) || SkScalarNearlyZero(sy))
134    {
135        return false;
136    }
137    scale->set(sx, sy);
138    return true;
139}
140
141/*
142 *  High quality is implemented by performing up-right scale-only filtering and then
143 *  using bilerp for any remaining transformations.
144 */
145void SkBitmapProcState::processHQRequest() {
146    SkASSERT(SkPaint::kHigh_FilterLevel == fFilterLevel);
147
148    // Our default return state is to downgrade the request to Medium, w/ or w/o setting fBitmap
149    // to a valid bitmap. If we succeed, we will set this to Low instead.
150    fFilterLevel = SkPaint::kMedium_FilterLevel;
151
152    if (kN32_SkColorType != fOrigBitmap.colorType() || !cache_size_okay(fOrigBitmap, fInvMatrix) ||
153        fInvMatrix.hasPerspective())
154    {
155        return; // can't handle the reqeust
156    }
157
158    SkScalar invScaleX = fInvMatrix.getScaleX();
159    SkScalar invScaleY = fInvMatrix.getScaleY();
160    if (fInvMatrix.getType() & SkMatrix::kAffine_Mask) {
161        SkVector scale;
162        if (!extract_scale(fInvMatrix, &scale)) {
163            return; // can't find suitable scale factors
164        }
165        invScaleX = scale.x();
166        invScaleY = scale.y();
167    }
168    if (SkScalarNearlyEqual(invScaleX, 1) && SkScalarNearlyEqual(invScaleY, 1)) {
169        return; // no need for HQ
170    }
171
172    SkScalar trueDestWidth  = fOrigBitmap.width() / invScaleX;
173    SkScalar trueDestHeight = fOrigBitmap.height() / invScaleY;
174    SkScalar roundedDestWidth = SkScalarRoundToScalar(trueDestWidth);
175    SkScalar roundedDestHeight = SkScalarRoundToScalar(trueDestHeight);
176
177    if (!SkBitmapCache::Find(fOrigBitmap, roundedDestWidth, roundedDestHeight, &fScaledBitmap)) {
178        if (!SkBitmapScaler::Resize(&fScaledBitmap,
179                                    fOrigBitmap,
180                                    SkBitmapScaler::RESIZE_BEST,
181                                    roundedDestWidth,
182                                    roundedDestHeight,
183                                    SkResourceCache::GetAllocator())) {
184            return; // we failed to create fScaledBitmap
185        }
186
187        SkASSERT(fScaledBitmap.getPixels());
188        fScaledBitmap.setImmutable();
189        SkBitmapCache::Add(fOrigBitmap, roundedDestWidth, roundedDestHeight, fScaledBitmap);
190    }
191
192    SkASSERT(fScaledBitmap.getPixels());
193    fBitmap = &fScaledBitmap;
194
195    fInvMatrix.postScale(roundedDestWidth / fOrigBitmap.width(),
196                         roundedDestHeight / fOrigBitmap.height());
197    fFilterLevel = SkPaint::kLow_FilterLevel;
198}
199
200/*
201 *  Modulo internal errors, this should always succeed *if* the matrix is downscaling
202 *  (in this case, we have the inverse, so it succeeds if fInvMatrix is upscaling)
203 */
204void SkBitmapProcState::processMediumRequest() {
205    SkASSERT(SkPaint::kMedium_FilterLevel == fFilterLevel);
206
207    // Our default return state is to downgrade the request to Low, w/ or w/o setting fBitmap
208    // to a valid bitmap.
209    fFilterLevel = SkPaint::kLow_FilterLevel;
210
211    SkScalar invScale = effective_matrix_scale(fInvMatrix);
212
213    if (invScale > SK_Scalar1) {
214        fCurrMip.reset(SkMipMapCache::FindAndRef(fOrigBitmap));
215        if (NULL == fCurrMip.get()) {
216            fCurrMip.reset(SkMipMapCache::AddAndRef(fOrigBitmap));
217            if (NULL == fCurrMip.get()) {
218                return;
219            }
220        }
221        // diagnostic for a crasher...
222        if (NULL == fCurrMip->data()) {
223            sk_throw();
224        }
225
226        SkScalar levelScale = SkScalarInvert(invScale);
227        SkMipMap::Level level;
228        if (fCurrMip->extractLevel(levelScale, &level)) {
229            SkScalar invScaleFixup = level.fScale;
230            fInvMatrix.postScale(invScaleFixup, invScaleFixup);
231
232            const SkImageInfo info = fOrigBitmap.info().makeWH(level.fWidth, level.fHeight);
233            // todo: if we could wrap the fCurrMip in a pixelref, then we could just install
234            //       that here, and not need to explicitly track it ourselves.
235            fScaledBitmap.installPixels(info, level.fPixels, level.fRowBytes);
236            fBitmap = &fScaledBitmap;
237        } else {
238            // failed to extract, so release the mipmap
239            fCurrMip.reset(NULL);
240        }
241    }
242}
243
244static bool get_locked_pixels(const SkBitmap& src, int pow2, SkBitmap* dst) {
245    SkPixelRef* pr = src.pixelRef();
246    if (pr && pr->decodeInto(pow2, dst)) {
247        return true;
248    }
249
250    /*
251     *  If decodeInto() fails, it is possibe that we have an old subclass that
252     *  does not, or cannot, implement that. In that case we fall back to the
253     *  older protocol of having the pixelRef handle the caching for us.
254     */
255    *dst = src;
256    dst->lockPixels();
257    return SkToBool(dst->getPixels());
258}
259
260bool SkBitmapProcState::lockBaseBitmap() {
261    SkPixelRef* pr = fOrigBitmap.pixelRef();
262
263    if (pr->isLocked() || !pr->implementsDecodeInto()) {
264        // fast-case, no need to look in our cache
265        fScaledBitmap = fOrigBitmap;
266        fScaledBitmap.lockPixels();
267        if (NULL == fScaledBitmap.getPixels()) {
268            return false;
269        }
270    } else {
271        if (!SkBitmapCache::Find(fOrigBitmap, 1, 1, &fScaledBitmap)) {
272            if (!get_locked_pixels(fOrigBitmap, 0, &fScaledBitmap)) {
273                return false;
274            }
275
276            // TODO: if fScaled comes back at a different width/height than fOrig,
277            // we need to update the matrix we are using to sample from this guy.
278
279            SkBitmapCache::Add(fOrigBitmap, 1, 1, fScaledBitmap);
280        }
281    }
282    fBitmap = &fScaledBitmap;
283    return true;
284}
285
286static bool valid_for_drawing(const SkBitmap& bm) {
287    if (0 == bm.width() || 0 == bm.height()) {
288        return false;   // nothing to draw
289    }
290    if (NULL == bm.pixelRef()) {
291        return false;   // no pixels to read
292    }
293    if (bm.getTexture()) {
294        // we can handle texture (ugh) since lockPixels will perform a read-back
295        return true;
296    }
297    if (kIndex_8_SkColorType == bm.colorType()) {
298        SkAutoLockPixels alp(bm); // but we need to call it before getColorTable() is safe.
299        if (!bm.getColorTable()) {
300            return false;
301        }
302    }
303    return true;
304}
305
306/*
307 *  Analyze filter-quality and matrix, and decide how to implement that.
308 *
309 *  In general, we cascade down the request level [ High ... None ]
310 *  - for a given level, if we can fulfill it, fine, else
311 *    - else we downgrade to the next lower level and try again.
312 *  We can always fulfill requests for Low and None
313 *  - sometimes we will "ignore" Low and give None, but this is likely a legacy perf hack
314 *    and may be removed.
315 */
316bool SkBitmapProcState::chooseProcs(const SkMatrix& inv, const SkPaint& paint) {
317    if (!valid_for_drawing(fOrigBitmap)) {
318        return false;
319    }
320
321    fBitmap = NULL;
322    fInvMatrix = inv;
323    fFilterLevel = paint.getFilterLevel();
324
325    if (SkPaint::kHigh_FilterLevel == fFilterLevel) {
326        this->processHQRequest();
327    }
328    SkASSERT(fFilterLevel < SkPaint::kHigh_FilterLevel);
329
330    if (SkPaint::kMedium_FilterLevel == fFilterLevel) {
331        this->processMediumRequest();
332    }
333    SkASSERT(fFilterLevel < SkPaint::kMedium_FilterLevel);
334
335    if (NULL == fBitmap) {
336        if (!this->lockBaseBitmap()) {
337            return false;
338        }
339    }
340    SkASSERT(fBitmap);
341
342    bool trivialMatrix = (fInvMatrix.getType() & ~SkMatrix::kTranslate_Mask) == 0;
343    bool clampClamp = SkShader::kClamp_TileMode == fTileModeX &&
344                      SkShader::kClamp_TileMode == fTileModeY;
345
346    // Most of the scanline procs deal with "unit" texture coordinates, as this
347    // makes it easy to perform tiling modes (repeat = (x & 0xFFFF)). To generate
348    // those, we divide the matrix by its dimensions here.
349    //
350    // We don't do this if we're either trivial (can ignore the matrix) or clamping
351    // in both X and Y since clamping to width,height is just as easy as to 0xFFFF.
352
353    if (!(clampClamp || trivialMatrix)) {
354        fInvMatrix.postIDiv(fBitmap->width(), fBitmap->height());
355    }
356
357    // Now that all possible changes to the matrix have taken place, check
358    // to see if we're really close to a no-scale matrix.  If so, explicitly
359    // set it to be so.  Subsequent code may inspect this matrix to choose
360    // a faster path in this case.
361
362    // This code will only execute if the matrix has some scale component;
363    // if it's already pure translate then we won't do this inversion.
364
365    if (matrix_only_scale_translate(fInvMatrix)) {
366        SkMatrix forward;
367        if (fInvMatrix.invert(&forward)) {
368            if (clampClamp ? just_trans_clamp(forward, *fBitmap)
369                           : just_trans_general(forward)) {
370                SkScalar tx = -SkScalarRoundToScalar(forward.getTranslateX());
371                SkScalar ty = -SkScalarRoundToScalar(forward.getTranslateY());
372                fInvMatrix.setTranslate(tx, ty);
373            }
374        }
375    }
376
377    fInvProc        = fInvMatrix.getMapXYProc();
378    fInvType        = fInvMatrix.getType();
379    fInvSx          = SkScalarToFixed(fInvMatrix.getScaleX());
380    fInvSxFractionalInt = SkScalarToFractionalInt(fInvMatrix.getScaleX());
381    fInvKy          = SkScalarToFixed(fInvMatrix.getSkewY());
382    fInvKyFractionalInt = SkScalarToFractionalInt(fInvMatrix.getSkewY());
383
384    fAlphaScale = SkAlpha255To256(paint.getAlpha());
385
386    fShaderProc32 = NULL;
387    fShaderProc16 = NULL;
388    fSampleProc32 = NULL;
389    fSampleProc16 = NULL;
390
391    // recompute the triviality of the matrix here because we may have
392    // changed it!
393
394    trivialMatrix = (fInvMatrix.getType() & ~SkMatrix::kTranslate_Mask) == 0;
395
396    if (SkPaint::kLow_FilterLevel == fFilterLevel) {
397        // Only try bilerp if the matrix is "interesting" and
398        // the image has a suitable size.
399
400        if (fInvType <= SkMatrix::kTranslate_Mask ||
401            !valid_for_filtering(fBitmap->width() | fBitmap->height()))
402        {
403            fFilterLevel = SkPaint::kNone_FilterLevel;
404        }
405    }
406
407    return this->chooseScanlineProcs(trivialMatrix, clampClamp, paint);
408}
409
410bool SkBitmapProcState::chooseScanlineProcs(bool trivialMatrix, bool clampClamp,
411                                            const SkPaint& paint) {
412    fMatrixProc = this->chooseMatrixProc(trivialMatrix);
413    // TODO(dominikg): SkASSERT(fMatrixProc) instead? chooseMatrixProc never returns NULL.
414    if (NULL == fMatrixProc) {
415        return false;
416    }
417
418    ///////////////////////////////////////////////////////////////////////
419
420    const SkAlphaType at = fBitmap->alphaType();
421
422    // No need to do this if we're doing HQ sampling; if filter quality is
423    // still set to HQ by the time we get here, then we must have installed
424    // the shader procs above and can skip all this.
425
426    if (fFilterLevel < SkPaint::kHigh_FilterLevel) {
427
428        int index = 0;
429        if (fAlphaScale < 256) {  // note: this distinction is not used for D16
430            index |= 1;
431        }
432        if (fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)) {
433            index |= 2;
434        }
435        if (fFilterLevel > SkPaint::kNone_FilterLevel) {
436            index |= 4;
437        }
438        // bits 3,4,5 encoding the source bitmap format
439        switch (fBitmap->colorType()) {
440            case kN32_SkColorType:
441                if (kPremul_SkAlphaType != at && kOpaque_SkAlphaType != at) {
442                    return false;
443                }
444                index |= 0;
445                break;
446            case kRGB_565_SkColorType:
447                index |= 8;
448                break;
449            case kIndex_8_SkColorType:
450                if (kPremul_SkAlphaType != at && kOpaque_SkAlphaType != at) {
451                    return false;
452                }
453                index |= 16;
454                break;
455            case kARGB_4444_SkColorType:
456                if (kPremul_SkAlphaType != at && kOpaque_SkAlphaType != at) {
457                    return false;
458                }
459                index |= 24;
460                break;
461            case kAlpha_8_SkColorType:
462                index |= 32;
463                fPaintPMColor = SkPreMultiplyColor(paint.getColor());
464                break;
465            default:
466                // TODO(dominikg): Should we ever get here? SkASSERT(false) instead?
467                return false;
468        }
469
470#if !SK_ARM_NEON_IS_ALWAYS
471        static const SampleProc32 gSkBitmapProcStateSample32[] = {
472            S32_opaque_D32_nofilter_DXDY,
473            S32_alpha_D32_nofilter_DXDY,
474            S32_opaque_D32_nofilter_DX,
475            S32_alpha_D32_nofilter_DX,
476            S32_opaque_D32_filter_DXDY,
477            S32_alpha_D32_filter_DXDY,
478            S32_opaque_D32_filter_DX,
479            S32_alpha_D32_filter_DX,
480
481            S16_opaque_D32_nofilter_DXDY,
482            S16_alpha_D32_nofilter_DXDY,
483            S16_opaque_D32_nofilter_DX,
484            S16_alpha_D32_nofilter_DX,
485            S16_opaque_D32_filter_DXDY,
486            S16_alpha_D32_filter_DXDY,
487            S16_opaque_D32_filter_DX,
488            S16_alpha_D32_filter_DX,
489
490            SI8_opaque_D32_nofilter_DXDY,
491            SI8_alpha_D32_nofilter_DXDY,
492            SI8_opaque_D32_nofilter_DX,
493            SI8_alpha_D32_nofilter_DX,
494            SI8_opaque_D32_filter_DXDY,
495            SI8_alpha_D32_filter_DXDY,
496            SI8_opaque_D32_filter_DX,
497            SI8_alpha_D32_filter_DX,
498
499            S4444_opaque_D32_nofilter_DXDY,
500            S4444_alpha_D32_nofilter_DXDY,
501            S4444_opaque_D32_nofilter_DX,
502            S4444_alpha_D32_nofilter_DX,
503            S4444_opaque_D32_filter_DXDY,
504            S4444_alpha_D32_filter_DXDY,
505            S4444_opaque_D32_filter_DX,
506            S4444_alpha_D32_filter_DX,
507
508            // A8 treats alpha/opaque the same (equally efficient)
509            SA8_alpha_D32_nofilter_DXDY,
510            SA8_alpha_D32_nofilter_DXDY,
511            SA8_alpha_D32_nofilter_DX,
512            SA8_alpha_D32_nofilter_DX,
513            SA8_alpha_D32_filter_DXDY,
514            SA8_alpha_D32_filter_DXDY,
515            SA8_alpha_D32_filter_DX,
516            SA8_alpha_D32_filter_DX
517        };
518
519        static const SampleProc16 gSkBitmapProcStateSample16[] = {
520            S32_D16_nofilter_DXDY,
521            S32_D16_nofilter_DX,
522            S32_D16_filter_DXDY,
523            S32_D16_filter_DX,
524
525            S16_D16_nofilter_DXDY,
526            S16_D16_nofilter_DX,
527            S16_D16_filter_DXDY,
528            S16_D16_filter_DX,
529
530            SI8_D16_nofilter_DXDY,
531            SI8_D16_nofilter_DX,
532            SI8_D16_filter_DXDY,
533            SI8_D16_filter_DX,
534
535            // Don't support 4444 -> 565
536            NULL, NULL, NULL, NULL,
537            // Don't support A8 -> 565
538            NULL, NULL, NULL, NULL
539        };
540#endif
541
542        fSampleProc32 = SK_ARM_NEON_WRAP(gSkBitmapProcStateSample32)[index];
543        index >>= 1;    // shift away any opaque/alpha distinction
544        fSampleProc16 = SK_ARM_NEON_WRAP(gSkBitmapProcStateSample16)[index];
545
546        // our special-case shaderprocs
547        if (SK_ARM_NEON_WRAP(S16_D16_filter_DX) == fSampleProc16) {
548            if (clampClamp) {
549                fShaderProc16 = SK_ARM_NEON_WRAP(Clamp_S16_D16_filter_DX_shaderproc);
550            } else if (SkShader::kRepeat_TileMode == fTileModeX &&
551                       SkShader::kRepeat_TileMode == fTileModeY) {
552                fShaderProc16 = SK_ARM_NEON_WRAP(Repeat_S16_D16_filter_DX_shaderproc);
553            }
554        } else if (SK_ARM_NEON_WRAP(SI8_opaque_D32_filter_DX) == fSampleProc32 && clampClamp) {
555            fShaderProc32 = SK_ARM_NEON_WRAP(Clamp_SI8_opaque_D32_filter_DX_shaderproc);
556        } else if (S32_opaque_D32_nofilter_DX == fSampleProc32 && clampClamp) {
557            fShaderProc32 = Clamp_S32_opaque_D32_nofilter_DX_shaderproc;
558        }
559
560        if (NULL == fShaderProc32) {
561            fShaderProc32 = this->chooseShaderProc32();
562        }
563    }
564
565    // see if our platform has any accelerated overrides
566    this->platformProcs();
567
568    return true;
569}
570
571static void Clamp_S32_D32_nofilter_trans_shaderproc(const SkBitmapProcState& s,
572                                                    int x, int y,
573                                                    SkPMColor* SK_RESTRICT colors,
574                                                    int count) {
575    SkASSERT(((s.fInvType & ~SkMatrix::kTranslate_Mask)) == 0);
576    SkASSERT(s.fInvKy == 0);
577    SkASSERT(count > 0 && colors != NULL);
578    SkASSERT(SkPaint::kNone_FilterLevel == s.fFilterLevel);
579
580    const int maxX = s.fBitmap->width() - 1;
581    const int maxY = s.fBitmap->height() - 1;
582    int ix = s.fFilterOneX + x;
583    int iy = SkClampMax(s.fFilterOneY + y, maxY);
584#ifdef SK_DEBUG
585    {
586        SkPoint pt;
587        s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf,
588                   SkIntToScalar(y) + SK_ScalarHalf, &pt);
589        int iy2 = SkClampMax(SkScalarFloorToInt(pt.fY), maxY);
590        int ix2 = SkScalarFloorToInt(pt.fX);
591
592        SkASSERT(iy == iy2);
593        SkASSERT(ix == ix2);
594    }
595#endif
596    const SkPMColor* row = s.fBitmap->getAddr32(0, iy);
597
598    // clamp to the left
599    if (ix < 0) {
600        int n = SkMin32(-ix, count);
601        sk_memset32(colors, row[0], n);
602        count -= n;
603        if (0 == count) {
604            return;
605        }
606        colors += n;
607        SkASSERT(-ix == n);
608        ix = 0;
609    }
610    // copy the middle
611    if (ix <= maxX) {
612        int n = SkMin32(maxX - ix + 1, count);
613        memcpy(colors, row + ix, n * sizeof(SkPMColor));
614        count -= n;
615        if (0 == count) {
616            return;
617        }
618        colors += n;
619    }
620    SkASSERT(count > 0);
621    // clamp to the right
622    sk_memset32(colors, row[maxX], count);
623}
624
625static inline int sk_int_mod(int x, int n) {
626    SkASSERT(n > 0);
627    if ((unsigned)x >= (unsigned)n) {
628        if (x < 0) {
629            x = n + ~(~x % n);
630        } else {
631            x = x % n;
632        }
633    }
634    return x;
635}
636
637static inline int sk_int_mirror(int x, int n) {
638    x = sk_int_mod(x, 2 * n);
639    if (x >= n) {
640        x = n + ~(x - n);
641    }
642    return x;
643}
644
645static void Repeat_S32_D32_nofilter_trans_shaderproc(const SkBitmapProcState& s,
646                                                     int x, int y,
647                                                     SkPMColor* SK_RESTRICT colors,
648                                                     int count) {
649    SkASSERT(((s.fInvType & ~SkMatrix::kTranslate_Mask)) == 0);
650    SkASSERT(s.fInvKy == 0);
651    SkASSERT(count > 0 && colors != NULL);
652    SkASSERT(SkPaint::kNone_FilterLevel == s.fFilterLevel);
653
654    const int stopX = s.fBitmap->width();
655    const int stopY = s.fBitmap->height();
656    int ix = s.fFilterOneX + x;
657    int iy = sk_int_mod(s.fFilterOneY + y, stopY);
658#ifdef SK_DEBUG
659    {
660        SkPoint pt;
661        s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf,
662                   SkIntToScalar(y) + SK_ScalarHalf, &pt);
663        int iy2 = sk_int_mod(SkScalarFloorToInt(pt.fY), stopY);
664        int ix2 = SkScalarFloorToInt(pt.fX);
665
666        SkASSERT(iy == iy2);
667        SkASSERT(ix == ix2);
668    }
669#endif
670    const SkPMColor* row = s.fBitmap->getAddr32(0, iy);
671
672    ix = sk_int_mod(ix, stopX);
673    for (;;) {
674        int n = SkMin32(stopX - ix, count);
675        memcpy(colors, row + ix, n * sizeof(SkPMColor));
676        count -= n;
677        if (0 == count) {
678            return;
679        }
680        colors += n;
681        ix = 0;
682    }
683}
684
685static void S32_D32_constX_shaderproc(const SkBitmapProcState& s,
686                                      int x, int y,
687                                      SkPMColor* SK_RESTRICT colors,
688                                      int count) {
689    SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)) == 0);
690    SkASSERT(s.fInvKy == 0);
691    SkASSERT(count > 0 && colors != NULL);
692    SkASSERT(1 == s.fBitmap->width());
693
694    int iY0;
695    int iY1   SK_INIT_TO_AVOID_WARNING;
696    int iSubY SK_INIT_TO_AVOID_WARNING;
697
698    if (SkPaint::kNone_FilterLevel != s.fFilterLevel) {
699        SkBitmapProcState::MatrixProc mproc = s.getMatrixProc();
700        uint32_t xy[2];
701
702        mproc(s, xy, 1, x, y);
703
704        iY0 = xy[0] >> 18;
705        iY1 = xy[0] & 0x3FFF;
706        iSubY = (xy[0] >> 14) & 0xF;
707    } else {
708        int yTemp;
709
710        if (s.fInvType > SkMatrix::kTranslate_Mask) {
711            SkPoint pt;
712            s.fInvProc(s.fInvMatrix,
713                       SkIntToScalar(x) + SK_ScalarHalf,
714                       SkIntToScalar(y) + SK_ScalarHalf,
715                       &pt);
716            // When the matrix has a scale component the setup code in
717            // chooseProcs multiples the inverse matrix by the inverse of the
718            // bitmap's width and height. Since this method is going to do
719            // its own tiling and sampling we need to undo that here.
720            if (SkShader::kClamp_TileMode != s.fTileModeX ||
721                SkShader::kClamp_TileMode != s.fTileModeY) {
722                yTemp = SkScalarFloorToInt(pt.fY * s.fBitmap->height());
723            } else {
724                yTemp = SkScalarFloorToInt(pt.fY);
725            }
726        } else {
727            yTemp = s.fFilterOneY + y;
728        }
729
730        const int stopY = s.fBitmap->height();
731        switch (s.fTileModeY) {
732            case SkShader::kClamp_TileMode:
733                iY0 = SkClampMax(yTemp, stopY-1);
734                break;
735            case SkShader::kRepeat_TileMode:
736                iY0 = sk_int_mod(yTemp, stopY);
737                break;
738            case SkShader::kMirror_TileMode:
739            default:
740                iY0 = sk_int_mirror(yTemp, stopY);
741                break;
742        }
743
744#ifdef SK_DEBUG
745        {
746            SkPoint pt;
747            s.fInvProc(s.fInvMatrix,
748                       SkIntToScalar(x) + SK_ScalarHalf,
749                       SkIntToScalar(y) + SK_ScalarHalf,
750                       &pt);
751            if (s.fInvType > SkMatrix::kTranslate_Mask &&
752                (SkShader::kClamp_TileMode != s.fTileModeX ||
753                 SkShader::kClamp_TileMode != s.fTileModeY)) {
754                pt.fY *= s.fBitmap->height();
755            }
756            int iY2;
757
758            switch (s.fTileModeY) {
759            case SkShader::kClamp_TileMode:
760                iY2 = SkClampMax(SkScalarFloorToInt(pt.fY), stopY-1);
761                break;
762            case SkShader::kRepeat_TileMode:
763                iY2 = sk_int_mod(SkScalarFloorToInt(pt.fY), stopY);
764                break;
765            case SkShader::kMirror_TileMode:
766            default:
767                iY2 = sk_int_mirror(SkScalarFloorToInt(pt.fY), stopY);
768                break;
769            }
770
771            SkASSERT(iY0 == iY2);
772        }
773#endif
774    }
775
776    const SkPMColor* row0 = s.fBitmap->getAddr32(0, iY0);
777    SkPMColor color;
778
779    if (SkPaint::kNone_FilterLevel != s.fFilterLevel) {
780        const SkPMColor* row1 = s.fBitmap->getAddr32(0, iY1);
781
782        if (s.fAlphaScale < 256) {
783            Filter_32_alpha(iSubY, *row0, *row1, &color, s.fAlphaScale);
784        } else {
785            Filter_32_opaque(iSubY, *row0, *row1, &color);
786        }
787    } else {
788        if (s.fAlphaScale < 256) {
789            color = SkAlphaMulQ(*row0, s.fAlphaScale);
790        } else {
791            color = *row0;
792        }
793    }
794
795    sk_memset32(colors, color, count);
796}
797
798static void DoNothing_shaderproc(const SkBitmapProcState&, int x, int y,
799                                 SkPMColor* SK_RESTRICT colors, int count) {
800    // if we get called, the matrix is too tricky, so we just draw nothing
801    sk_memset32(colors, 0, count);
802}
803
804bool SkBitmapProcState::setupForTranslate() {
805    SkPoint pt;
806    fInvProc(fInvMatrix, SK_ScalarHalf, SK_ScalarHalf, &pt);
807
808    /*
809     *  if the translate is larger than our ints, we can get random results, or
810     *  worse, we might get 0x80000000, which wreaks havoc on us, since we can't
811     *  negate it.
812     */
813    const SkScalar too_big = SkIntToScalar(1 << 30);
814    if (SkScalarAbs(pt.fX) > too_big || SkScalarAbs(pt.fY) > too_big) {
815        return false;
816    }
817
818    // Since we know we're not filtered, we re-purpose these fields allow
819    // us to go from device -> src coordinates w/ just an integer add,
820    // rather than running through the inverse-matrix
821    fFilterOneX = SkScalarFloorToInt(pt.fX);
822    fFilterOneY = SkScalarFloorToInt(pt.fY);
823    return true;
824}
825
826SkBitmapProcState::ShaderProc32 SkBitmapProcState::chooseShaderProc32() {
827
828    if (kN32_SkColorType != fBitmap->colorType()) {
829        return NULL;
830    }
831
832    static const unsigned kMask = SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask;
833
834    if (1 == fBitmap->width() && 0 == (fInvType & ~kMask)) {
835        if (SkPaint::kNone_FilterLevel == fFilterLevel &&
836            fInvType <= SkMatrix::kTranslate_Mask &&
837            !this->setupForTranslate()) {
838            return DoNothing_shaderproc;
839        }
840        return S32_D32_constX_shaderproc;
841    }
842
843    if (fAlphaScale < 256) {
844        return NULL;
845    }
846    if (fInvType > SkMatrix::kTranslate_Mask) {
847        return NULL;
848    }
849    if (SkPaint::kNone_FilterLevel != fFilterLevel) {
850        return NULL;
851    }
852
853    SkShader::TileMode tx = (SkShader::TileMode)fTileModeX;
854    SkShader::TileMode ty = (SkShader::TileMode)fTileModeY;
855
856    if (SkShader::kClamp_TileMode == tx && SkShader::kClamp_TileMode == ty) {
857        if (this->setupForTranslate()) {
858            return Clamp_S32_D32_nofilter_trans_shaderproc;
859        }
860        return DoNothing_shaderproc;
861    }
862    if (SkShader::kRepeat_TileMode == tx && SkShader::kRepeat_TileMode == ty) {
863        if (this->setupForTranslate()) {
864            return Repeat_S32_D32_nofilter_trans_shaderproc;
865        }
866        return DoNothing_shaderproc;
867    }
868    return NULL;
869}
870
871///////////////////////////////////////////////////////////////////////////////
872
873#ifdef SK_DEBUG
874
875static void check_scale_nofilter(uint32_t bitmapXY[], int count,
876                                 unsigned mx, unsigned my) {
877    unsigned y = *bitmapXY++;
878    SkASSERT(y < my);
879
880    const uint16_t* xptr = reinterpret_cast<const uint16_t*>(bitmapXY);
881    for (int i = 0; i < count; ++i) {
882        SkASSERT(xptr[i] < mx);
883    }
884}
885
886static void check_scale_filter(uint32_t bitmapXY[], int count,
887                                 unsigned mx, unsigned my) {
888    uint32_t YY = *bitmapXY++;
889    unsigned y0 = YY >> 18;
890    unsigned y1 = YY & 0x3FFF;
891    SkASSERT(y0 < my);
892    SkASSERT(y1 < my);
893
894    for (int i = 0; i < count; ++i) {
895        uint32_t XX = bitmapXY[i];
896        unsigned x0 = XX >> 18;
897        unsigned x1 = XX & 0x3FFF;
898        SkASSERT(x0 < mx);
899        SkASSERT(x1 < mx);
900    }
901}
902
903static void check_affine_nofilter(uint32_t bitmapXY[], int count,
904                                 unsigned mx, unsigned my) {
905    for (int i = 0; i < count; ++i) {
906        uint32_t XY = bitmapXY[i];
907        unsigned x = XY & 0xFFFF;
908        unsigned y = XY >> 16;
909        SkASSERT(x < mx);
910        SkASSERT(y < my);
911    }
912}
913
914static void check_affine_filter(uint32_t bitmapXY[], int count,
915                                 unsigned mx, unsigned my) {
916    for (int i = 0; i < count; ++i) {
917        uint32_t YY = *bitmapXY++;
918        unsigned y0 = YY >> 18;
919        unsigned y1 = YY & 0x3FFF;
920        SkASSERT(y0 < my);
921        SkASSERT(y1 < my);
922
923        uint32_t XX = *bitmapXY++;
924        unsigned x0 = XX >> 18;
925        unsigned x1 = XX & 0x3FFF;
926        SkASSERT(x0 < mx);
927        SkASSERT(x1 < mx);
928    }
929}
930
931void SkBitmapProcState::DebugMatrixProc(const SkBitmapProcState& state,
932                                        uint32_t bitmapXY[], int count,
933                                        int x, int y) {
934    SkASSERT(bitmapXY);
935    SkASSERT(count > 0);
936
937    state.fMatrixProc(state, bitmapXY, count, x, y);
938
939    void (*proc)(uint32_t bitmapXY[], int count, unsigned mx, unsigned my);
940
941    // There are four formats possible:
942    //  scale -vs- affine
943    //  filter -vs- nofilter
944    if (state.fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)) {
945        proc = state.fFilterLevel != SkPaint::kNone_FilterLevel ? check_scale_filter : check_scale_nofilter;
946    } else {
947        proc = state.fFilterLevel != SkPaint::kNone_FilterLevel ? check_affine_filter : check_affine_nofilter;
948    }
949    proc(bitmapXY, count, state.fBitmap->width(), state.fBitmap->height());
950}
951
952SkBitmapProcState::MatrixProc SkBitmapProcState::getMatrixProc() const {
953    return DebugMatrixProc;
954}
955
956#endif
957
958///////////////////////////////////////////////////////////////////////////////
959/*
960    The storage requirements for the different matrix procs are as follows,
961    where each X or Y is 2 bytes, and N is the number of pixels/elements:
962
963    scale/translate     nofilter      Y(4bytes) + N * X
964    affine/perspective  nofilter      N * (X Y)
965    scale/translate     filter        Y Y + N * (X X)
966    affine/perspective  filter        N * (Y Y X X)
967 */
968int SkBitmapProcState::maxCountForBufferSize(size_t bufferSize) const {
969    int32_t size = static_cast<int32_t>(bufferSize);
970
971    size &= ~3; // only care about 4-byte aligned chunks
972    if (fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)) {
973        size -= 4;   // the shared Y (or YY) coordinate
974        if (size < 0) {
975            size = 0;
976        }
977        size >>= 1;
978    } else {
979        size >>= 2;
980    }
981
982    if (fFilterLevel != SkPaint::kNone_FilterLevel) {
983        size >>= 1;
984    }
985
986    return size;
987}
988
989///////////////////////
990
991void  Clamp_S32_opaque_D32_nofilter_DX_shaderproc(const SkBitmapProcState& s, int x, int y,
992                                                  SkPMColor* SK_RESTRICT dst, int count) {
993    SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask |
994                             SkMatrix::kScale_Mask)) == 0);
995
996    const unsigned maxX = s.fBitmap->width() - 1;
997    SkFractionalInt fx;
998    int dstY;
999    {
1000        SkPoint pt;
1001        s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf, SkIntToScalar(y) + SK_ScalarHalf,
1002                   &pt);
1003        fx = SkScalarToFractionalInt(pt.fY);
1004        const unsigned maxY = s.fBitmap->height() - 1;
1005        dstY = SkClampMax(SkFractionalIntToInt(fx), maxY);
1006        fx = SkScalarToFractionalInt(pt.fX);
1007    }
1008
1009    const SkPMColor* SK_RESTRICT src = s.fBitmap->getAddr32(0, dstY);
1010    const SkFractionalInt dx = s.fInvSxFractionalInt;
1011
1012    // Check if we're safely inside [0...maxX] so no need to clamp each computed index.
1013    //
1014    if ((uint64_t)SkFractionalIntToInt(fx) <= maxX &&
1015        (uint64_t)SkFractionalIntToInt(fx + dx * (count - 1)) <= maxX)
1016    {
1017        int count4 = count >> 2;
1018        for (int i = 0; i < count4; ++i) {
1019            SkPMColor src0 = src[SkFractionalIntToInt(fx)]; fx += dx;
1020            SkPMColor src1 = src[SkFractionalIntToInt(fx)]; fx += dx;
1021            SkPMColor src2 = src[SkFractionalIntToInt(fx)]; fx += dx;
1022            SkPMColor src3 = src[SkFractionalIntToInt(fx)]; fx += dx;
1023            dst[0] = src0;
1024            dst[1] = src1;
1025            dst[2] = src2;
1026            dst[3] = src3;
1027            dst += 4;
1028        }
1029        for (int i = (count4 << 2); i < count; ++i) {
1030            unsigned index = SkFractionalIntToInt(fx);
1031            SkASSERT(index <= maxX);
1032            *dst++ = src[index];
1033            fx += dx;
1034        }
1035    } else {
1036        for (int i = 0; i < count; ++i) {
1037            dst[i] = src[SkClampMax(SkFractionalIntToInt(fx), maxX)];
1038            fx += dx;
1039        }
1040    }
1041}
1042
1043