SkBitmapProcState.cpp revision 0dc546c37c7dff3885188054d191cf852d899e32
1
2/*
3 * Copyright 2011 Google Inc.
4 *
5 * Use of this source code is governed by a BSD-style license that can be
6 * found in the LICENSE file.
7 */
8#include "SkBitmapProcState.h"
9#include "SkColorPriv.h"
10#include "SkFilterProc.h"
11#include "SkPaint.h"
12#include "SkShader.h"   // for tilemodes
13#include "SkUtilsArm.h"
14#include "SkBitmapScaler.h"
15#include "SkMipMap.h"
16#include "SkScaledImageCache.h"
17
18#if !SK_ARM_NEON_IS_NONE
19// These are defined in src/opts/SkBitmapProcState_arm_neon.cpp
20extern const SkBitmapProcState::SampleProc16 gSkBitmapProcStateSample16_neon[];
21extern const SkBitmapProcState::SampleProc32 gSkBitmapProcStateSample32_neon[];
22extern void  S16_D16_filter_DX_neon(const SkBitmapProcState&, const uint32_t*, int, uint16_t*);
23extern void  Clamp_S16_D16_filter_DX_shaderproc_neon(const SkBitmapProcState&, int, int, uint16_t*, int);
24extern void  Repeat_S16_D16_filter_DX_shaderproc_neon(const SkBitmapProcState&, int, int, uint16_t*, int);
25extern void  SI8_opaque_D32_filter_DX_neon(const SkBitmapProcState&, const uint32_t*, int, SkPMColor*);
26extern void  SI8_opaque_D32_filter_DX_shaderproc_neon(const SkBitmapProcState&, int, int, uint32_t*, int);
27extern void  Clamp_SI8_opaque_D32_filter_DX_shaderproc_neon(const SkBitmapProcState&, int, int, uint32_t*, int);
28#endif
29
30#define   NAME_WRAP(x)  x
31#include "SkBitmapProcState_filter.h"
32#include "SkBitmapProcState_procs.h"
33
34///////////////////////////////////////////////////////////////////////////////
35
36// true iff the matrix contains, at most, scale and translate elements
37static bool matrix_only_scale_translate(const SkMatrix& m) {
38    return m.getType() <= (SkMatrix::kScale_Mask | SkMatrix::kTranslate_Mask);
39}
40
41/**
42 *  For the purposes of drawing bitmaps, if a matrix is "almost" translate
43 *  go ahead and treat it as if it were, so that subsequent code can go fast.
44 */
45static bool just_trans_clamp(const SkMatrix& matrix, const SkBitmap& bitmap) {
46    SkASSERT(matrix_only_scale_translate(matrix));
47
48    if (matrix.getType() & SkMatrix::kScale_Mask) {
49        SkRect src, dst;
50        bitmap.getBounds(&src);
51
52        // Can't call mapRect(), since that will fix up inverted rectangles,
53        // e.g. when scale is negative, and we don't want to return true for
54        // those.
55        matrix.mapPoints(SkTCast<SkPoint*>(&dst),
56                         SkTCast<const SkPoint*>(&src),
57                         2);
58
59        // Now round all 4 edges to device space, and then compare the device
60        // width/height to the original. Note: we must map all 4 and subtract
61        // rather than map the "width" and compare, since we care about the
62        // phase (in pixel space) that any translate in the matrix might impart.
63        SkIRect idst;
64        dst.round(&idst);
65        return idst.width() == bitmap.width() && idst.height() == bitmap.height();
66    }
67    // if we got here, we're either kTranslate_Mask or identity
68    return true;
69}
70
71static bool just_trans_general(const SkMatrix& matrix) {
72    SkASSERT(matrix_only_scale_translate(matrix));
73
74    if (matrix.getType() & SkMatrix::kScale_Mask) {
75        const SkScalar tol = SK_Scalar1 / 32768;
76
77        if (!SkScalarNearlyZero(matrix[SkMatrix::kMScaleX] - SK_Scalar1, tol)) {
78            return false;
79        }
80        if (!SkScalarNearlyZero(matrix[SkMatrix::kMScaleY] - SK_Scalar1, tol)) {
81            return false;
82        }
83    }
84    // if we got here, treat us as either kTranslate_Mask or identity
85    return true;
86}
87
88///////////////////////////////////////////////////////////////////////////////
89
90static bool valid_for_filtering(unsigned dimension) {
91    // for filtering, width and height must fit in 14bits, since we use steal
92    // 2 bits from each to store our 4bit subpixel data
93    return (dimension & ~0x3FFF) == 0;
94}
95
96static SkScalar effective_matrix_scale_sqrd(const SkMatrix& mat) {
97    SkPoint v1, v2;
98
99    v1.fX = mat.getScaleX();
100    v1.fY = mat.getSkewY();
101
102    v2.fX = mat.getSkewX();
103    v2.fY = mat.getScaleY();
104
105    return SkMaxScalar(v1.lengthSqd(), v2.lengthSqd());
106}
107
108// TODO -- we may want to pass the clip into this function so we only scale
109// the portion of the image that we're going to need.  This will complicate
110// the interface to the cache, but might be well worth it.
111
112void SkBitmapProcState::possiblyScaleImage() {
113
114    if (fFilterLevel <= SkPaint::kLow_FilterLevel) {
115        // none or low (bilerp) does not need to look any further
116        return;
117    }
118
119    // see if our platform has any specialized convolution code.
120
121
122    // Set up a pointer to a local (instead of storing the structure in the
123    // proc state) to avoid introducing a header dependency; this makes
124    // recompiles a lot less painful.
125
126    SkConvolutionProcs simd;
127    fConvolutionProcs = &simd;
128
129    fConvolutionProcs->fExtraHorizontalReads = 0;
130    fConvolutionProcs->fConvolveVertically = NULL;
131    fConvolutionProcs->fConvolve4RowsHorizontally = NULL;
132    fConvolutionProcs->fConvolveHorizontally = NULL;
133    fConvolutionProcs->fApplySIMDPadding = NULL;
134
135    this->platformConvolutionProcs();
136
137    // STEP 1: Highest quality direct scale?
138
139    // Check to see if the transformation matrix is simple, and if we're
140    // doing high quality scaling.  If so, do the bitmap scale here and
141    // remove the scaling component from the matrix.
142
143    if (SkPaint::kHigh_FilterLevel == fFilterLevel &&
144        fInvMatrix.getType() <= (SkMatrix::kScale_Mask | SkMatrix::kTranslate_Mask) &&
145        fOrigBitmap.config() == SkBitmap::kARGB_8888_Config) {
146
147        SkScalar invScaleX = fInvMatrix.getScaleX();
148        SkScalar invScaleY = fInvMatrix.getScaleY();
149
150        SkASSERT(NULL == fScaledCacheID);
151        fScaledCacheID = SkScaledImageCache::FindAndLock(fOrigBitmap,
152                                                         invScaleX, invScaleY,
153                                                         &fScaledBitmap);
154        if (NULL == fScaledCacheID) {
155            int dest_width  = SkScalarCeilToInt(fOrigBitmap.width() / invScaleX);
156            int dest_height = SkScalarCeilToInt(fOrigBitmap.height() / invScaleY);
157
158            // All the criteria are met; let's make a new bitmap.
159
160            if (!SkBitmapScaler::Resize(&fScaledBitmap,
161                                        fOrigBitmap,
162                                        SkBitmapScaler::RESIZE_BEST,
163                                        dest_width,
164                                        dest_height,
165                                        fConvolutionProcs)) {
166                // we failed to create fScaledBitmap, so just return and let
167                // the scanline proc handle it.
168                return;
169
170            }
171            fScaledCacheID = SkScaledImageCache::AddAndLock(fOrigBitmap,
172                                                            invScaleX,
173                                                            invScaleY,
174                                                            fScaledBitmap);
175        }
176        fScaledBitmap.lockPixels();
177
178        fBitmap = &fScaledBitmap;
179
180        // set the inv matrix type to translate-only;
181
182        fInvMatrix.setTranslate( 1/fInvMatrix.getScaleX() * fInvMatrix.getTranslateX(),
183                                 1/fInvMatrix.getScaleY() * fInvMatrix.getTranslateY() );
184
185        // no need for any further filtering; we just did it!
186
187        fFilterLevel = SkPaint::kNone_FilterLevel;
188
189        return;
190    }
191
192    /*
193     *  If we get here, the caller has requested either Med or High filter-level
194     *
195     *  If High, then our special-case for scale-only did not take, and so we
196     *  have to make a choice:
197     *      1. fall back on mipmaps + bilerp
198     *      2. fall back on scanline bicubic filter
199     *  For now, we compute the "scale" value from the matrix, and have a
200     *  threshold to decide when bicubic is better, and when mips are better.
201     *  No doubt a fancier decision tree could be used uere.
202     *
203     *  If Medium, then we just try to build a mipmap and select a level,
204     *  setting the filter-level to kLow to signal that we just need bilerp
205     *  to process the selected level.
206     */
207
208    SkScalar scaleSqd = effective_matrix_scale_sqrd(fInvMatrix);
209
210    if (SkPaint::kHigh_FilterLevel == fFilterLevel) {
211        // Set the limit at 0.25 for the CTM... if the CTM is scaling smaller
212        // than this, then the mipmaps quality may be greater (certainly faster)
213        // so we only keep High quality if the scale is greater than this.
214        //
215        // Since we're dealing with the inverse, we compare against its inverse.
216        const SkScalar bicubicLimit = SkFloatToScalar(4.0f);
217        const SkScalar bicubicLimitSqd = bicubicLimit * bicubicLimit;
218        if (scaleSqd < bicubicLimitSqd) {  // use bicubic scanline
219            return;
220        }
221
222        // else set the filter-level to Medium, since we're scaling down and
223        // want to reqeust mipmaps
224        fFilterLevel = SkPaint::kMedium_FilterLevel;
225    }
226
227    SkASSERT(SkPaint::kMedium_FilterLevel == fFilterLevel);
228
229    /**
230     *  Medium quality means use a mipmap for down-scaling, and just bilper
231     *  for upscaling. Since we're examining the inverse matrix, we look for
232     *  a scale > 1 to indicate down scaling by the CTM.
233     */
234    if (scaleSqd > SK_Scalar1) {
235        const SkMipMap* mip = NULL;
236
237        SkASSERT(NULL == fScaledCacheID);
238        fScaledCacheID = SkScaledImageCache::FindAndLockMip(fOrigBitmap, &mip);
239        if (!fScaledCacheID) {
240            SkASSERT(NULL == mip);
241            mip = SkMipMap::Build(fOrigBitmap);
242            if (mip) {
243                fScaledCacheID = SkScaledImageCache::AddAndLockMip(fOrigBitmap,
244                                                                   mip);
245                mip->unref();   // the cache took a ref
246                SkASSERT(fScaledCacheID);
247            }
248        } else {
249            SkASSERT(mip);
250        }
251
252        if (mip) {
253            SkScalar levelScale = SkScalarInvert(SkScalarSqrt(scaleSqd));
254            SkMipMap::Level level;
255            if (mip->extractLevel(levelScale, &level)) {
256                SkScalar invScaleFixup = level.fScale;
257                fInvMatrix.postScale(invScaleFixup, invScaleFixup);
258
259                fScaledBitmap.setConfig(fOrigBitmap.config(),
260                                        level.fWidth, level.fHeight,
261                                        level.fRowBytes);
262                fScaledBitmap.setPixels(level.fPixels);
263                fBitmap = &fScaledBitmap;
264            }
265        }
266    }
267
268    /*
269     *  At this point, we may or may not have built a mipmap. Regardless, we
270     *  now fall back on Low so will bilerp whatever fBitmap now points at.
271     */
272    fFilterLevel = SkPaint::kLow_FilterLevel;
273}
274
275void SkBitmapProcState::endContext() {
276    SkDELETE(fBitmapFilter);
277    fBitmapFilter = NULL;
278    fScaledBitmap.reset();
279
280    if (fScaledCacheID) {
281        SkScaledImageCache::Unlock(fScaledCacheID);
282        fScaledCacheID = NULL;
283    }
284}
285
286SkBitmapProcState::~SkBitmapProcState() {
287    if (fScaledCacheID) {
288        SkScaledImageCache::Unlock(fScaledCacheID);
289    }
290    SkDELETE(fBitmapFilter);
291}
292
293bool SkBitmapProcState::chooseProcs(const SkMatrix& inv, const SkPaint& paint) {
294    if (fOrigBitmap.width() == 0 || fOrigBitmap.height() == 0) {
295        return false;
296    }
297
298    bool trivialMatrix = (inv.getType() & ~SkMatrix::kTranslate_Mask) == 0;
299    bool clampClamp = SkShader::kClamp_TileMode == fTileModeX &&
300                       SkShader::kClamp_TileMode == fTileModeY;
301
302    fInvMatrix = inv;
303    if (!(clampClamp || trivialMatrix)) {
304        fInvMatrix.postIDiv(fOrigBitmap.width(), fOrigBitmap.height());
305    }
306
307    fBitmap = &fOrigBitmap;
308
309    // initialize our filter quality to the one requested by the caller.
310    // We may downgrade it later if we determine that we either don't need
311    // or can't provide as high a quality filtering as the user requested.
312
313    fFilterLevel = paint.getFilterLevel();
314
315#ifndef SK_IGNORE_IMAGE_PRESCALE
316    // possiblyScaleImage will look to see if it can rescale the image as a
317    // preprocess; either by scaling up to the target size, or by selecting
318    // a nearby mipmap level.  If it does, it will adjust the working
319    // matrix as well as the working bitmap.  It may also adjust the filter
320    // quality to avoid re-filtering an already perfectly scaled image.
321
322    this->possiblyScaleImage();
323#endif
324
325    // Now that all possible changes to the matrix have taken place, check
326    // to see if we're really close to a no-scale matrix.  If so, explicitly
327    // set it to be so.  Subsequent code may inspect this matrix to choose
328    // a faster path in this case.
329
330    // This code will only execute if the matrix has some scale component;
331    // if it's already pure translate then we won't do this inversion.
332
333    if (matrix_only_scale_translate(fInvMatrix)) {
334        SkMatrix forward;
335        if (fInvMatrix.invert(&forward)) {
336            if (clampClamp ? just_trans_clamp(forward, *fBitmap)
337                            : just_trans_general(forward)) {
338                SkScalar tx = -SkScalarRoundToScalar(forward.getTranslateX());
339                SkScalar ty = -SkScalarRoundToScalar(forward.getTranslateY());
340                fInvMatrix.setTranslate(tx, ty);
341
342            }
343        }
344    }
345
346    fInvProc        = fInvMatrix.getMapXYProc();
347    fInvType        = fInvMatrix.getType();
348    fInvSx          = SkScalarToFixed(fInvMatrix.getScaleX());
349    fInvSxFractionalInt = SkScalarToFractionalInt(fInvMatrix.getScaleX());
350    fInvKy          = SkScalarToFixed(fInvMatrix.getSkewY());
351    fInvKyFractionalInt = SkScalarToFractionalInt(fInvMatrix.getSkewY());
352
353    fAlphaScale = SkAlpha255To256(paint.getAlpha());
354
355    fShaderProc32 = NULL;
356    fShaderProc16 = NULL;
357    fSampleProc32 = NULL;
358    fSampleProc16 = NULL;
359
360    // recompute the triviality of the matrix here because we may have
361    // changed it!
362
363    trivialMatrix = (fInvMatrix.getType() & ~SkMatrix::kTranslate_Mask) == 0;
364
365    if (SkPaint::kHigh_FilterLevel == fFilterLevel) {
366        // If this is still set, that means we wanted HQ sampling
367        // but couldn't do it as a preprocess.  Let's try to install
368        // the scanline version of the HQ sampler.  If that process fails,
369        // downgrade to bilerp.
370
371        // NOTE: Might need to be careful here in the future when we want
372        // to have the platform proc have a shot at this; it's possible that
373        // the chooseBitmapFilterProc will fail to install a shader but a
374        // platform-specific one might succeed, so it might be premature here
375        // to fall back to bilerp.  This needs thought.
376
377        SkASSERT(fInvType > SkMatrix::kTranslate_Mask);
378
379        if (!this->setBitmapFilterProcs()) {
380            fFilterLevel = SkPaint::kLow_FilterLevel;
381        }
382    }
383
384    if (SkPaint::kLow_FilterLevel == fFilterLevel) {
385        // Only try bilerp if the matrix is "interesting" and
386        // the image has a suitable size.
387
388        if (fInvType <= SkMatrix::kTranslate_Mask ||
389                !valid_for_filtering(fBitmap->width() | fBitmap->height())) {
390            fFilterLevel = SkPaint::kNone_FilterLevel;
391        }
392    }
393
394    // At this point, we know exactly what kind of sampling the per-scanline
395    // shader will perform.
396
397    fMatrixProc = this->chooseMatrixProc(trivialMatrix);
398    if (NULL == fMatrixProc) {
399        return false;
400    }
401
402    ///////////////////////////////////////////////////////////////////////
403
404    // No need to do this if we're doing HQ sampling; if filter quality is
405    // still set to HQ by the time we get here, then we must have installed
406    // the shader procs above and can skip all this.
407
408    if (fFilterLevel < SkPaint::kHigh_FilterLevel) {
409
410        int index = 0;
411        if (fAlphaScale < 256) {  // note: this distinction is not used for D16
412            index |= 1;
413        }
414        if (fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)) {
415            index |= 2;
416        }
417        if (fFilterLevel > SkPaint::kNone_FilterLevel) {
418            index |= 4;
419        }
420        // bits 3,4,5 encoding the source bitmap format
421        switch (fBitmap->config()) {
422            case SkBitmap::kARGB_8888_Config:
423                index |= 0;
424                break;
425            case SkBitmap::kRGB_565_Config:
426                index |= 8;
427                break;
428            case SkBitmap::kIndex8_Config:
429                index |= 16;
430                break;
431            case SkBitmap::kARGB_4444_Config:
432                index |= 24;
433                break;
434            case SkBitmap::kA8_Config:
435                index |= 32;
436                fPaintPMColor = SkPreMultiplyColor(paint.getColor());
437                break;
438            default:
439                return false;
440        }
441
442    #if !SK_ARM_NEON_IS_ALWAYS
443        static const SampleProc32 gSkBitmapProcStateSample32[] = {
444            S32_opaque_D32_nofilter_DXDY,
445            S32_alpha_D32_nofilter_DXDY,
446            S32_opaque_D32_nofilter_DX,
447            S32_alpha_D32_nofilter_DX,
448            S32_opaque_D32_filter_DXDY,
449            S32_alpha_D32_filter_DXDY,
450            S32_opaque_D32_filter_DX,
451            S32_alpha_D32_filter_DX,
452
453            S16_opaque_D32_nofilter_DXDY,
454            S16_alpha_D32_nofilter_DXDY,
455            S16_opaque_D32_nofilter_DX,
456            S16_alpha_D32_nofilter_DX,
457            S16_opaque_D32_filter_DXDY,
458            S16_alpha_D32_filter_DXDY,
459            S16_opaque_D32_filter_DX,
460            S16_alpha_D32_filter_DX,
461
462            SI8_opaque_D32_nofilter_DXDY,
463            SI8_alpha_D32_nofilter_DXDY,
464            SI8_opaque_D32_nofilter_DX,
465            SI8_alpha_D32_nofilter_DX,
466            SI8_opaque_D32_filter_DXDY,
467            SI8_alpha_D32_filter_DXDY,
468            SI8_opaque_D32_filter_DX,
469            SI8_alpha_D32_filter_DX,
470
471            S4444_opaque_D32_nofilter_DXDY,
472            S4444_alpha_D32_nofilter_DXDY,
473            S4444_opaque_D32_nofilter_DX,
474            S4444_alpha_D32_nofilter_DX,
475            S4444_opaque_D32_filter_DXDY,
476            S4444_alpha_D32_filter_DXDY,
477            S4444_opaque_D32_filter_DX,
478            S4444_alpha_D32_filter_DX,
479
480            // A8 treats alpha/opaque the same (equally efficient)
481            SA8_alpha_D32_nofilter_DXDY,
482            SA8_alpha_D32_nofilter_DXDY,
483            SA8_alpha_D32_nofilter_DX,
484            SA8_alpha_D32_nofilter_DX,
485            SA8_alpha_D32_filter_DXDY,
486            SA8_alpha_D32_filter_DXDY,
487            SA8_alpha_D32_filter_DX,
488            SA8_alpha_D32_filter_DX
489        };
490
491        static const SampleProc16 gSkBitmapProcStateSample16[] = {
492            S32_D16_nofilter_DXDY,
493            S32_D16_nofilter_DX,
494            S32_D16_filter_DXDY,
495            S32_D16_filter_DX,
496
497            S16_D16_nofilter_DXDY,
498            S16_D16_nofilter_DX,
499            S16_D16_filter_DXDY,
500            S16_D16_filter_DX,
501
502            SI8_D16_nofilter_DXDY,
503            SI8_D16_nofilter_DX,
504            SI8_D16_filter_DXDY,
505            SI8_D16_filter_DX,
506
507            // Don't support 4444 -> 565
508            NULL, NULL, NULL, NULL,
509            // Don't support A8 -> 565
510            NULL, NULL, NULL, NULL
511        };
512    #endif
513
514        fSampleProc32 = SK_ARM_NEON_WRAP(gSkBitmapProcStateSample32)[index];
515        index >>= 1;    // shift away any opaque/alpha distinction
516        fSampleProc16 = SK_ARM_NEON_WRAP(gSkBitmapProcStateSample16)[index];
517
518        // our special-case shaderprocs
519        if (SK_ARM_NEON_WRAP(S16_D16_filter_DX) == fSampleProc16) {
520            if (clampClamp) {
521                fShaderProc16 = SK_ARM_NEON_WRAP(Clamp_S16_D16_filter_DX_shaderproc);
522            } else if (SkShader::kRepeat_TileMode == fTileModeX &&
523                       SkShader::kRepeat_TileMode == fTileModeY) {
524                fShaderProc16 = SK_ARM_NEON_WRAP(Repeat_S16_D16_filter_DX_shaderproc);
525            }
526        } else if (SK_ARM_NEON_WRAP(SI8_opaque_D32_filter_DX) == fSampleProc32 && clampClamp) {
527            fShaderProc32 = SK_ARM_NEON_WRAP(Clamp_SI8_opaque_D32_filter_DX_shaderproc);
528        }
529
530        if (NULL == fShaderProc32) {
531            fShaderProc32 = this->chooseShaderProc32();
532        }
533    }
534
535    // see if our platform has any accelerated overrides
536    this->platformProcs();
537
538    return true;
539}
540
541static void Clamp_S32_D32_nofilter_trans_shaderproc(const SkBitmapProcState& s,
542                                                    int x, int y,
543                                                    SkPMColor* SK_RESTRICT colors,
544                                                    int count) {
545    SkASSERT(((s.fInvType & ~SkMatrix::kTranslate_Mask)) == 0);
546    SkASSERT(s.fInvKy == 0);
547    SkASSERT(count > 0 && colors != NULL);
548    SkASSERT(SkPaint::kNone_FilterLevel == s.fFilterLevel);
549
550    const int maxX = s.fBitmap->width() - 1;
551    const int maxY = s.fBitmap->height() - 1;
552    int ix = s.fFilterOneX + x;
553    int iy = SkClampMax(s.fFilterOneY + y, maxY);
554#ifdef SK_DEBUG
555    {
556        SkPoint pt;
557        s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf,
558                   SkIntToScalar(y) + SK_ScalarHalf, &pt);
559        int iy2 = SkClampMax(SkScalarFloorToInt(pt.fY), maxY);
560        int ix2 = SkScalarFloorToInt(pt.fX);
561
562        SkASSERT(iy == iy2);
563        SkASSERT(ix == ix2);
564    }
565#endif
566    const SkPMColor* row = s.fBitmap->getAddr32(0, iy);
567
568    // clamp to the left
569    if (ix < 0) {
570        int n = SkMin32(-ix, count);
571        sk_memset32(colors, row[0], n);
572        count -= n;
573        if (0 == count) {
574            return;
575        }
576        colors += n;
577        SkASSERT(-ix == n);
578        ix = 0;
579    }
580    // copy the middle
581    if (ix <= maxX) {
582        int n = SkMin32(maxX - ix + 1, count);
583        memcpy(colors, row + ix, n * sizeof(SkPMColor));
584        count -= n;
585        if (0 == count) {
586            return;
587        }
588        colors += n;
589    }
590    SkASSERT(count > 0);
591    // clamp to the right
592    sk_memset32(colors, row[maxX], count);
593}
594
595static inline int sk_int_mod(int x, int n) {
596    SkASSERT(n > 0);
597    if ((unsigned)x >= (unsigned)n) {
598        if (x < 0) {
599            x = n + ~(~x % n);
600        } else {
601            x = x % n;
602        }
603    }
604    return x;
605}
606
607static inline int sk_int_mirror(int x, int n) {
608    x = sk_int_mod(x, 2 * n);
609    if (x >= n) {
610        x = n + ~(x - n);
611    }
612    return x;
613}
614
615static void Repeat_S32_D32_nofilter_trans_shaderproc(const SkBitmapProcState& s,
616                                                     int x, int y,
617                                                     SkPMColor* SK_RESTRICT colors,
618                                                     int count) {
619    SkASSERT(((s.fInvType & ~SkMatrix::kTranslate_Mask)) == 0);
620    SkASSERT(s.fInvKy == 0);
621    SkASSERT(count > 0 && colors != NULL);
622    SkASSERT(SkPaint::kNone_FilterLevel == s.fFilterLevel);
623
624    const int stopX = s.fBitmap->width();
625    const int stopY = s.fBitmap->height();
626    int ix = s.fFilterOneX + x;
627    int iy = sk_int_mod(s.fFilterOneY + y, stopY);
628#ifdef SK_DEBUG
629    {
630        SkPoint pt;
631        s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf,
632                   SkIntToScalar(y) + SK_ScalarHalf, &pt);
633        int iy2 = sk_int_mod(SkScalarFloorToInt(pt.fY), stopY);
634        int ix2 = SkScalarFloorToInt(pt.fX);
635
636        SkASSERT(iy == iy2);
637        SkASSERT(ix == ix2);
638    }
639#endif
640    const SkPMColor* row = s.fBitmap->getAddr32(0, iy);
641
642    ix = sk_int_mod(ix, stopX);
643    for (;;) {
644        int n = SkMin32(stopX - ix, count);
645        memcpy(colors, row + ix, n * sizeof(SkPMColor));
646        count -= n;
647        if (0 == count) {
648            return;
649        }
650        colors += n;
651        ix = 0;
652    }
653}
654
655static void S32_D32_constX_shaderproc(const SkBitmapProcState& s,
656                                      int x, int y,
657                                      SkPMColor* SK_RESTRICT colors,
658                                      int count) {
659    SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)) == 0);
660    SkASSERT(s.fInvKy == 0);
661    SkASSERT(count > 0 && colors != NULL);
662    SkASSERT(1 == s.fBitmap->width());
663
664    int iY0;
665    int iY1   SK_INIT_TO_AVOID_WARNING;
666    int iSubY SK_INIT_TO_AVOID_WARNING;
667
668    if (SkPaint::kNone_FilterLevel != s.fFilterLevel) {
669        SkBitmapProcState::MatrixProc mproc = s.getMatrixProc();
670        uint32_t xy[2];
671
672        mproc(s, xy, 1, x, y);
673
674        iY0 = xy[0] >> 18;
675        iY1 = xy[0] & 0x3FFF;
676        iSubY = (xy[0] >> 14) & 0xF;
677    } else {
678        int yTemp;
679
680        if (s.fInvType > SkMatrix::kTranslate_Mask) {
681            SkPoint pt;
682            s.fInvProc(s.fInvMatrix,
683                       SkIntToScalar(x) + SK_ScalarHalf,
684                       SkIntToScalar(y) + SK_ScalarHalf,
685                       &pt);
686            // When the matrix has a scale component the setup code in
687            // chooseProcs multiples the inverse matrix by the inverse of the
688            // bitmap's width and height. Since this method is going to do
689            // its own tiling and sampling we need to undo that here.
690            if (SkShader::kClamp_TileMode != s.fTileModeX ||
691                SkShader::kClamp_TileMode != s.fTileModeY) {
692                yTemp = SkScalarFloorToInt(pt.fY * s.fBitmap->height());
693            } else {
694                yTemp = SkScalarFloorToInt(pt.fY);
695            }
696        } else {
697            yTemp = s.fFilterOneY + y;
698        }
699
700        const int stopY = s.fBitmap->height();
701        switch (s.fTileModeY) {
702            case SkShader::kClamp_TileMode:
703                iY0 = SkClampMax(yTemp, stopY-1);
704                break;
705            case SkShader::kRepeat_TileMode:
706                iY0 = sk_int_mod(yTemp, stopY);
707                break;
708            case SkShader::kMirror_TileMode:
709            default:
710                iY0 = sk_int_mirror(yTemp, stopY);
711                break;
712        }
713
714#ifdef SK_DEBUG
715        {
716            SkPoint pt;
717            s.fInvProc(s.fInvMatrix,
718                       SkIntToScalar(x) + SK_ScalarHalf,
719                       SkIntToScalar(y) + SK_ScalarHalf,
720                       &pt);
721            if (s.fInvType > SkMatrix::kTranslate_Mask &&
722                (SkShader::kClamp_TileMode != s.fTileModeX ||
723                 SkShader::kClamp_TileMode != s.fTileModeY)) {
724                pt.fY *= s.fBitmap->height();
725            }
726            int iY2;
727
728            switch (s.fTileModeY) {
729            case SkShader::kClamp_TileMode:
730                iY2 = SkClampMax(SkScalarFloorToInt(pt.fY), stopY-1);
731                break;
732            case SkShader::kRepeat_TileMode:
733                iY2 = sk_int_mod(SkScalarFloorToInt(pt.fY), stopY);
734                break;
735            case SkShader::kMirror_TileMode:
736            default:
737                iY2 = sk_int_mirror(SkScalarFloorToInt(pt.fY), stopY);
738                break;
739            }
740
741            SkASSERT(iY0 == iY2);
742        }
743#endif
744    }
745
746    const SkPMColor* row0 = s.fBitmap->getAddr32(0, iY0);
747    SkPMColor color;
748
749    if (SkPaint::kNone_FilterLevel != s.fFilterLevel) {
750        const SkPMColor* row1 = s.fBitmap->getAddr32(0, iY1);
751
752        if (s.fAlphaScale < 256) {
753            Filter_32_alpha(iSubY, *row0, *row1, &color, s.fAlphaScale);
754        } else {
755            Filter_32_opaque(iSubY, *row0, *row1, &color);
756        }
757    } else {
758        if (s.fAlphaScale < 256) {
759            color = SkAlphaMulQ(*row0, s.fAlphaScale);
760        } else {
761            color = *row0;
762        }
763    }
764
765    sk_memset32(colors, color, count);
766}
767
768static void DoNothing_shaderproc(const SkBitmapProcState&, int x, int y,
769                                 SkPMColor* SK_RESTRICT colors, int count) {
770    // if we get called, the matrix is too tricky, so we just draw nothing
771    sk_memset32(colors, 0, count);
772}
773
774bool SkBitmapProcState::setupForTranslate() {
775    SkPoint pt;
776    fInvProc(fInvMatrix, SK_ScalarHalf, SK_ScalarHalf, &pt);
777
778    /*
779     *  if the translate is larger than our ints, we can get random results, or
780     *  worse, we might get 0x80000000, which wreaks havoc on us, since we can't
781     *  negate it.
782     */
783    const SkScalar too_big = SkIntToScalar(1 << 30);
784    if (SkScalarAbs(pt.fX) > too_big || SkScalarAbs(pt.fY) > too_big) {
785        return false;
786    }
787
788    // Since we know we're not filtered, we re-purpose these fields allow
789    // us to go from device -> src coordinates w/ just an integer add,
790    // rather than running through the inverse-matrix
791    fFilterOneX = SkScalarFloorToInt(pt.fX);
792    fFilterOneY = SkScalarFloorToInt(pt.fY);
793    return true;
794}
795
796SkBitmapProcState::ShaderProc32 SkBitmapProcState::chooseShaderProc32() {
797
798    if (SkBitmap::kARGB_8888_Config != fBitmap->config()) {
799        return NULL;
800    }
801
802    static const unsigned kMask = SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask;
803
804    if (1 == fBitmap->width() && 0 == (fInvType & ~kMask)) {
805        if (SkPaint::kNone_FilterLevel == fFilterLevel &&
806            fInvType <= SkMatrix::kTranslate_Mask &&
807            !this->setupForTranslate()) {
808            return DoNothing_shaderproc;
809        }
810        return S32_D32_constX_shaderproc;
811    }
812
813    if (fAlphaScale < 256) {
814        return NULL;
815    }
816    if (fInvType > SkMatrix::kTranslate_Mask) {
817        return NULL;
818    }
819    if (SkPaint::kNone_FilterLevel != fFilterLevel) {
820        return NULL;
821    }
822
823    SkShader::TileMode tx = (SkShader::TileMode)fTileModeX;
824    SkShader::TileMode ty = (SkShader::TileMode)fTileModeY;
825
826    if (SkShader::kClamp_TileMode == tx && SkShader::kClamp_TileMode == ty) {
827        if (this->setupForTranslate()) {
828            return Clamp_S32_D32_nofilter_trans_shaderproc;
829        }
830        return DoNothing_shaderproc;
831    }
832    if (SkShader::kRepeat_TileMode == tx && SkShader::kRepeat_TileMode == ty) {
833        if (this->setupForTranslate()) {
834            return Repeat_S32_D32_nofilter_trans_shaderproc;
835        }
836        return DoNothing_shaderproc;
837    }
838    return NULL;
839}
840
841///////////////////////////////////////////////////////////////////////////////
842
843#ifdef SK_DEBUG
844
845static void check_scale_nofilter(uint32_t bitmapXY[], int count,
846                                 unsigned mx, unsigned my) {
847    unsigned y = *bitmapXY++;
848    SkASSERT(y < my);
849
850    const uint16_t* xptr = reinterpret_cast<const uint16_t*>(bitmapXY);
851    for (int i = 0; i < count; ++i) {
852        SkASSERT(xptr[i] < mx);
853    }
854}
855
856static void check_scale_filter(uint32_t bitmapXY[], int count,
857                                 unsigned mx, unsigned my) {
858    uint32_t YY = *bitmapXY++;
859    unsigned y0 = YY >> 18;
860    unsigned y1 = YY & 0x3FFF;
861    SkASSERT(y0 < my);
862    SkASSERT(y1 < my);
863
864    for (int i = 0; i < count; ++i) {
865        uint32_t XX = bitmapXY[i];
866        unsigned x0 = XX >> 18;
867        unsigned x1 = XX & 0x3FFF;
868        SkASSERT(x0 < mx);
869        SkASSERT(x1 < mx);
870    }
871}
872
873static void check_affine_nofilter(uint32_t bitmapXY[], int count,
874                                 unsigned mx, unsigned my) {
875    for (int i = 0; i < count; ++i) {
876        uint32_t XY = bitmapXY[i];
877        unsigned x = XY & 0xFFFF;
878        unsigned y = XY >> 16;
879        SkASSERT(x < mx);
880        SkASSERT(y < my);
881    }
882}
883
884static void check_affine_filter(uint32_t bitmapXY[], int count,
885                                 unsigned mx, unsigned my) {
886    for (int i = 0; i < count; ++i) {
887        uint32_t YY = *bitmapXY++;
888        unsigned y0 = YY >> 18;
889        unsigned y1 = YY & 0x3FFF;
890        SkASSERT(y0 < my);
891        SkASSERT(y1 < my);
892
893        uint32_t XX = *bitmapXY++;
894        unsigned x0 = XX >> 18;
895        unsigned x1 = XX & 0x3FFF;
896        SkASSERT(x0 < mx);
897        SkASSERT(x1 < mx);
898    }
899}
900
901void SkBitmapProcState::DebugMatrixProc(const SkBitmapProcState& state,
902                                        uint32_t bitmapXY[], int count,
903                                        int x, int y) {
904    SkASSERT(bitmapXY);
905    SkASSERT(count > 0);
906
907    state.fMatrixProc(state, bitmapXY, count, x, y);
908
909    void (*proc)(uint32_t bitmapXY[], int count, unsigned mx, unsigned my);
910
911    // There are four formats possible:
912    //  scale -vs- affine
913    //  filter -vs- nofilter
914    if (state.fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)) {
915        proc = state.fFilterLevel != SkPaint::kNone_FilterLevel ? check_scale_filter : check_scale_nofilter;
916    } else {
917        proc = state.fFilterLevel != SkPaint::kNone_FilterLevel ? check_affine_filter : check_affine_nofilter;
918    }
919    proc(bitmapXY, count, state.fBitmap->width(), state.fBitmap->height());
920}
921
922SkBitmapProcState::MatrixProc SkBitmapProcState::getMatrixProc() const {
923    return DebugMatrixProc;
924}
925
926#endif
927
928///////////////////////////////////////////////////////////////////////////////
929/*
930    The storage requirements for the different matrix procs are as follows,
931    where each X or Y is 2 bytes, and N is the number of pixels/elements:
932
933    scale/translate     nofilter      Y(4bytes) + N * X
934    affine/perspective  nofilter      N * (X Y)
935    scale/translate     filter        Y Y + N * (X X)
936    affine/perspective  filter        N * (Y Y X X)
937 */
938int SkBitmapProcState::maxCountForBufferSize(size_t bufferSize) const {
939    int32_t size = static_cast<int32_t>(bufferSize);
940
941    size &= ~3; // only care about 4-byte aligned chunks
942    if (fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)) {
943        size -= 4;   // the shared Y (or YY) coordinate
944        if (size < 0) {
945            size = 0;
946        }
947        size >>= 1;
948    } else {
949        size >>= 2;
950    }
951
952    if (fFilterLevel != SkPaint::kNone_FilterLevel) {
953        size >>= 1;
954    }
955
956    return size;
957}
958