| 1 | /* |
| 2 | * Copyright 2006 The Android Open Source Project |
| 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 "src/core/SkScanPriv.h" |
| 9 | |
| 10 | #include "include/core/SkMatrix.h" |
| 11 | #include "include/core/SkPath.h" |
| 12 | #include "include/core/SkRegion.h" |
| 13 | #include "include/private/SkTo.h" |
| 14 | #include "src/core/SkAntiRun.h" |
| 15 | #include "src/core/SkBlitter.h" |
| 16 | #include "src/core/SkPathPriv.h" |
| 17 | |
| 18 | #define SHIFT SK_SUPERSAMPLE_SHIFT |
| 19 | #define SCALE (1 << SHIFT) |
| 20 | #define MASK (SCALE - 1) |
| 21 | |
| 22 | /** @file |
| 23 | We have two techniques for capturing the output of the supersampler: |
| 24 | - SUPERMASK, which records a large mask-bitmap |
| 25 | this is often faster for small, complex objects |
| 26 | - RLE, which records a rle-encoded scanline |
| 27 | this is often faster for large objects with big spans |
| 28 | |
| 29 | These blitters use two coordinate systems: |
| 30 | - destination coordinates, scale equal to the output - often |
| 31 | abbreviated with 'i' or 'I' in variable names |
| 32 | - supersampled coordinates, scale equal to the output * SCALE |
| 33 | */ |
| 34 | |
| 35 | //#define FORCE_SUPERMASK |
| 36 | //#define FORCE_RLE |
| 37 | |
| 38 | /////////////////////////////////////////////////////////////////////////////// |
| 39 | |
| 40 | /// Base class for a single-pass supersampled blitter. |
| 41 | class BaseSuperBlitter : public SkBlitter { |
| 42 | public: |
| 43 | BaseSuperBlitter(SkBlitter* realBlitter, const SkIRect& ir, |
| 44 | const SkIRect& clipBounds, bool isInverse); |
| 45 | |
| 46 | /// Must be explicitly defined on subclasses. |
| 47 | virtual void blitAntiH(int x, int y, const SkAlpha antialias[], |
| 48 | const int16_t runs[]) override { |
| 49 | SkDEBUGFAIL("How did I get here?" ); |
| 50 | } |
| 51 | /// May not be called on BaseSuperBlitter because it blits out of order. |
| 52 | void blitV(int x, int y, int height, SkAlpha alpha) override { |
| 53 | SkDEBUGFAIL("How did I get here?" ); |
| 54 | } |
| 55 | |
| 56 | protected: |
| 57 | SkBlitter* fRealBlitter; |
| 58 | /// Current y coordinate, in destination coordinates. |
| 59 | int fCurrIY; |
| 60 | /// Widest row of region to be blitted, in destination coordinates. |
| 61 | int fWidth; |
| 62 | /// Leftmost x coordinate in any row, in destination coordinates. |
| 63 | int fLeft; |
| 64 | /// Leftmost x coordinate in any row, in supersampled coordinates. |
| 65 | int fSuperLeft; |
| 66 | |
| 67 | SkDEBUGCODE(int fCurrX;) |
| 68 | /// Current y coordinate in supersampled coordinates. |
| 69 | int fCurrY; |
| 70 | /// Initial y coordinate (top of bounds). |
| 71 | int fTop; |
| 72 | |
| 73 | SkIRect fSectBounds; |
| 74 | }; |
| 75 | |
| 76 | BaseSuperBlitter::BaseSuperBlitter(SkBlitter* realBlit, const SkIRect& ir, |
| 77 | const SkIRect& clipBounds, bool isInverse) { |
| 78 | fRealBlitter = realBlit; |
| 79 | |
| 80 | SkIRect sectBounds; |
| 81 | if (isInverse) { |
| 82 | // We use the clip bounds instead of the ir, since we may be asked to |
| 83 | //draw outside of the rect when we're a inverse filltype |
| 84 | sectBounds = clipBounds; |
| 85 | } else { |
| 86 | if (!sectBounds.intersect(ir, clipBounds)) { |
| 87 | sectBounds.setEmpty(); |
| 88 | } |
| 89 | } |
| 90 | |
| 91 | const int left = sectBounds.left(); |
| 92 | const int right = sectBounds.right(); |
| 93 | |
| 94 | fLeft = left; |
| 95 | fSuperLeft = SkLeftShift(left, SHIFT); |
| 96 | fWidth = right - left; |
| 97 | fTop = sectBounds.top(); |
| 98 | fCurrIY = fTop - 1; |
| 99 | fCurrY = SkLeftShift(fTop, SHIFT) - 1; |
| 100 | |
| 101 | SkDEBUGCODE(fCurrX = -1;) |
| 102 | } |
| 103 | |
| 104 | /// Run-length-encoded supersampling antialiased blitter. |
| 105 | class SuperBlitter : public BaseSuperBlitter { |
| 106 | public: |
| 107 | SuperBlitter(SkBlitter* realBlitter, const SkIRect& ir, const SkIRect& clipBounds, |
| 108 | bool isInverse); |
| 109 | |
| 110 | ~SuperBlitter() override { |
| 111 | this->flush(); |
| 112 | } |
| 113 | |
| 114 | /// Once fRuns contains a complete supersampled row, flush() blits |
| 115 | /// it out through the wrapped blitter. |
| 116 | void flush(); |
| 117 | |
| 118 | /// Blits a row of pixels, with location and width specified |
| 119 | /// in supersampled coordinates. |
| 120 | void blitH(int x, int y, int width) override; |
| 121 | /// Blits a rectangle of pixels, with location and size specified |
| 122 | /// in supersampled coordinates. |
| 123 | void blitRect(int x, int y, int width, int height) override; |
| 124 | |
| 125 | private: |
| 126 | // The next three variables are used to track a circular buffer that |
| 127 | // contains the values used in SkAlphaRuns. These variables should only |
| 128 | // ever be updated in advanceRuns(), and fRuns should always point to |
| 129 | // a valid SkAlphaRuns... |
| 130 | int fRunsToBuffer; |
| 131 | void* fRunsBuffer; |
| 132 | int fCurrentRun; |
| 133 | SkAlphaRuns fRuns; |
| 134 | |
| 135 | // extra one to store the zero at the end |
| 136 | int getRunsSz() const { return (fWidth + 1 + (fWidth + 2)/2) * sizeof(int16_t); } |
| 137 | |
| 138 | // This function updates the fRuns variable to point to the next buffer space |
| 139 | // with adequate storage for a SkAlphaRuns. It mostly just advances fCurrentRun |
| 140 | // and resets fRuns to point to an empty scanline. |
| 141 | void advanceRuns() { |
| 142 | const size_t kRunsSz = this->getRunsSz(); |
| 143 | fCurrentRun = (fCurrentRun + 1) % fRunsToBuffer; |
| 144 | fRuns.fRuns = reinterpret_cast<int16_t*>( |
| 145 | reinterpret_cast<uint8_t*>(fRunsBuffer) + fCurrentRun * kRunsSz); |
| 146 | fRuns.fAlpha = reinterpret_cast<SkAlpha*>(fRuns.fRuns + fWidth + 1); |
| 147 | fRuns.reset(fWidth); |
| 148 | } |
| 149 | |
| 150 | int fOffsetX; |
| 151 | }; |
| 152 | |
| 153 | SuperBlitter::SuperBlitter(SkBlitter* realBlitter, const SkIRect& ir, const SkIRect& clipBounds, |
| 154 | bool isInverse) |
| 155 | : BaseSuperBlitter(realBlitter, ir, clipBounds, isInverse) |
| 156 | { |
| 157 | fRunsToBuffer = realBlitter->requestRowsPreserved(); |
| 158 | fRunsBuffer = realBlitter->allocBlitMemory(fRunsToBuffer * this->getRunsSz()); |
| 159 | fCurrentRun = -1; |
| 160 | |
| 161 | this->advanceRuns(); |
| 162 | |
| 163 | fOffsetX = 0; |
| 164 | } |
| 165 | |
| 166 | void SuperBlitter::flush() { |
| 167 | if (fCurrIY >= fTop) { |
| 168 | |
| 169 | SkASSERT(fCurrentRun < fRunsToBuffer); |
| 170 | if (!fRuns.empty()) { |
| 171 | // SkDEBUGCODE(fRuns.dump();) |
| 172 | fRealBlitter->blitAntiH(fLeft, fCurrIY, fRuns.fAlpha, fRuns.fRuns); |
| 173 | this->advanceRuns(); |
| 174 | fOffsetX = 0; |
| 175 | } |
| 176 | |
| 177 | fCurrIY = fTop - 1; |
| 178 | SkDEBUGCODE(fCurrX = -1;) |
| 179 | } |
| 180 | } |
| 181 | |
| 182 | /** coverage_to_partial_alpha() is being used by SkAlphaRuns, which |
| 183 | *accumulates* SCALE pixels worth of "alpha" in [0,(256/SCALE)] |
| 184 | to produce a final value in [0, 255] and handles clamping 256->255 |
| 185 | itself, with the same (alpha - (alpha >> 8)) correction as |
| 186 | coverage_to_exact_alpha(). |
| 187 | */ |
| 188 | static inline int coverage_to_partial_alpha(int aa) { |
| 189 | aa <<= 8 - 2*SHIFT; |
| 190 | return aa; |
| 191 | } |
| 192 | |
| 193 | /** coverage_to_exact_alpha() is being used by our blitter, which wants |
| 194 | a final value in [0, 255]. |
| 195 | */ |
| 196 | static inline int coverage_to_exact_alpha(int aa) { |
| 197 | int alpha = (256 >> SHIFT) * aa; |
| 198 | // clamp 256->255 |
| 199 | return alpha - (alpha >> 8); |
| 200 | } |
| 201 | |
| 202 | void SuperBlitter::blitH(int x, int y, int width) { |
| 203 | SkASSERT(width > 0); |
| 204 | |
| 205 | int iy = y >> SHIFT; |
| 206 | SkASSERT(iy >= fCurrIY); |
| 207 | |
| 208 | x -= fSuperLeft; |
| 209 | // hack, until I figure out why my cubics (I think) go beyond the bounds |
| 210 | if (x < 0) { |
| 211 | width += x; |
| 212 | x = 0; |
| 213 | } |
| 214 | |
| 215 | #ifdef SK_DEBUG |
| 216 | SkASSERT(y != fCurrY || x >= fCurrX); |
| 217 | #endif |
| 218 | SkASSERT(y >= fCurrY); |
| 219 | if (fCurrY != y) { |
| 220 | fOffsetX = 0; |
| 221 | fCurrY = y; |
| 222 | } |
| 223 | |
| 224 | if (iy != fCurrIY) { // new scanline |
| 225 | this->flush(); |
| 226 | fCurrIY = iy; |
| 227 | } |
| 228 | |
| 229 | int start = x; |
| 230 | int stop = x + width; |
| 231 | |
| 232 | SkASSERT(start >= 0 && stop > start); |
| 233 | // integer-pixel-aligned ends of blit, rounded out |
| 234 | int fb = start & MASK; |
| 235 | int fe = stop & MASK; |
| 236 | int n = (stop >> SHIFT) - (start >> SHIFT) - 1; |
| 237 | |
| 238 | if (n < 0) { |
| 239 | fb = fe - fb; |
| 240 | n = 0; |
| 241 | fe = 0; |
| 242 | } else { |
| 243 | if (fb == 0) { |
| 244 | n += 1; |
| 245 | } else { |
| 246 | fb = SCALE - fb; |
| 247 | } |
| 248 | } |
| 249 | |
| 250 | fOffsetX = fRuns.add(x >> SHIFT, coverage_to_partial_alpha(fb), |
| 251 | n, coverage_to_partial_alpha(fe), |
| 252 | (1 << (8 - SHIFT)) - (((y & MASK) + 1) >> SHIFT), |
| 253 | fOffsetX); |
| 254 | |
| 255 | #ifdef SK_DEBUG |
| 256 | fRuns.assertValid(y & MASK, (1 << (8 - SHIFT))); |
| 257 | fCurrX = x + width; |
| 258 | #endif |
| 259 | } |
| 260 | |
| 261 | #if 0 // UNUSED |
| 262 | static void set_left_rite_runs(SkAlphaRuns& runs, int ileft, U8CPU leftA, |
| 263 | int n, U8CPU riteA) { |
| 264 | SkASSERT(leftA <= 0xFF); |
| 265 | SkASSERT(riteA <= 0xFF); |
| 266 | |
| 267 | int16_t* run = runs.fRuns; |
| 268 | uint8_t* aa = runs.fAlpha; |
| 269 | |
| 270 | if (ileft > 0) { |
| 271 | run[0] = ileft; |
| 272 | aa[0] = 0; |
| 273 | run += ileft; |
| 274 | aa += ileft; |
| 275 | } |
| 276 | |
| 277 | SkASSERT(leftA < 0xFF); |
| 278 | if (leftA > 0) { |
| 279 | *run++ = 1; |
| 280 | *aa++ = leftA; |
| 281 | } |
| 282 | |
| 283 | if (n > 0) { |
| 284 | run[0] = n; |
| 285 | aa[0] = 0xFF; |
| 286 | run += n; |
| 287 | aa += n; |
| 288 | } |
| 289 | |
| 290 | SkASSERT(riteA < 0xFF); |
| 291 | if (riteA > 0) { |
| 292 | *run++ = 1; |
| 293 | *aa++ = riteA; |
| 294 | } |
| 295 | run[0] = 0; |
| 296 | } |
| 297 | #endif |
| 298 | |
| 299 | void SuperBlitter::blitRect(int x, int y, int width, int height) { |
| 300 | SkASSERT(width > 0); |
| 301 | SkASSERT(height > 0); |
| 302 | |
| 303 | // blit leading rows |
| 304 | while ((y & MASK)) { |
| 305 | this->blitH(x, y++, width); |
| 306 | if (--height <= 0) { |
| 307 | return; |
| 308 | } |
| 309 | } |
| 310 | SkASSERT(height > 0); |
| 311 | |
| 312 | // Since this is a rect, instead of blitting supersampled rows one at a |
| 313 | // time and then resolving to the destination canvas, we can blit |
| 314 | // directly to the destintion canvas one row per SCALE supersampled rows. |
| 315 | int start_y = y >> SHIFT; |
| 316 | int stop_y = (y + height) >> SHIFT; |
| 317 | int count = stop_y - start_y; |
| 318 | if (count > 0) { |
| 319 | y += count << SHIFT; |
| 320 | height -= count << SHIFT; |
| 321 | |
| 322 | // save original X for our tail blitH() loop at the bottom |
| 323 | int origX = x; |
| 324 | |
| 325 | x -= fSuperLeft; |
| 326 | // hack, until I figure out why my cubics (I think) go beyond the bounds |
| 327 | if (x < 0) { |
| 328 | width += x; |
| 329 | x = 0; |
| 330 | } |
| 331 | |
| 332 | // There is always a left column, a middle, and a right column. |
| 333 | // ileft is the destination x of the first pixel of the entire rect. |
| 334 | // xleft is (SCALE - # of covered supersampled pixels) in that |
| 335 | // destination pixel. |
| 336 | int ileft = x >> SHIFT; |
| 337 | int xleft = x & MASK; |
| 338 | // irite is the destination x of the last pixel of the OPAQUE section. |
| 339 | // xrite is the number of supersampled pixels extending beyond irite; |
| 340 | // xrite/SCALE should give us alpha. |
| 341 | int irite = (x + width) >> SHIFT; |
| 342 | int xrite = (x + width) & MASK; |
| 343 | if (!xrite) { |
| 344 | xrite = SCALE; |
| 345 | irite--; |
| 346 | } |
| 347 | |
| 348 | // Need to call flush() to clean up pending draws before we |
| 349 | // even consider blitV(), since otherwise it can look nonmonotonic. |
| 350 | SkASSERT(start_y > fCurrIY); |
| 351 | this->flush(); |
| 352 | |
| 353 | int n = irite - ileft - 1; |
| 354 | if (n < 0) { |
| 355 | // If n < 0, we'll only have a single partially-transparent column |
| 356 | // of pixels to render. |
| 357 | xleft = xrite - xleft; |
| 358 | SkASSERT(xleft <= SCALE); |
| 359 | SkASSERT(xleft > 0); |
| 360 | fRealBlitter->blitV(ileft + fLeft, start_y, count, |
| 361 | coverage_to_exact_alpha(xleft)); |
| 362 | } else { |
| 363 | // With n = 0, we have two possibly-transparent columns of pixels |
| 364 | // to render; with n > 0, we have opaque columns between them. |
| 365 | |
| 366 | xleft = SCALE - xleft; |
| 367 | |
| 368 | // Using coverage_to_exact_alpha is not consistent with blitH() |
| 369 | const int coverageL = coverage_to_exact_alpha(xleft); |
| 370 | const int coverageR = coverage_to_exact_alpha(xrite); |
| 371 | |
| 372 | SkASSERT(coverageL > 0 || n > 0 || coverageR > 0); |
| 373 | SkASSERT((coverageL != 0) + n + (coverageR != 0) <= fWidth); |
| 374 | |
| 375 | fRealBlitter->blitAntiRect(ileft + fLeft, start_y, n, count, |
| 376 | coverageL, coverageR); |
| 377 | } |
| 378 | |
| 379 | // preamble for our next call to blitH() |
| 380 | fCurrIY = stop_y - 1; |
| 381 | fOffsetX = 0; |
| 382 | fCurrY = y - 1; |
| 383 | fRuns.reset(fWidth); |
| 384 | x = origX; |
| 385 | } |
| 386 | |
| 387 | // catch any remaining few rows |
| 388 | SkASSERT(height <= MASK); |
| 389 | while (--height >= 0) { |
| 390 | this->blitH(x, y++, width); |
| 391 | } |
| 392 | } |
| 393 | |
| 394 | /////////////////////////////////////////////////////////////////////////////// |
| 395 | |
| 396 | /// Masked supersampling antialiased blitter. |
| 397 | class MaskSuperBlitter : public BaseSuperBlitter { |
| 398 | public: |
| 399 | MaskSuperBlitter(SkBlitter* realBlitter, const SkIRect& ir, const SkIRect&, bool isInverse); |
| 400 | ~MaskSuperBlitter() override { |
| 401 | fRealBlitter->blitMask(fMask, fClipRect); |
| 402 | } |
| 403 | |
| 404 | void blitH(int x, int y, int width) override; |
| 405 | |
| 406 | static bool CanHandleRect(const SkIRect& bounds) { |
| 407 | #ifdef FORCE_RLE |
| 408 | return false; |
| 409 | #endif |
| 410 | int width = bounds.width(); |
| 411 | int64_t rb = SkAlign4(width); |
| 412 | // use 64bits to detect overflow |
| 413 | int64_t storage = rb * bounds.height(); |
| 414 | |
| 415 | return (width <= MaskSuperBlitter::kMAX_WIDTH) && |
| 416 | (storage <= MaskSuperBlitter::kMAX_STORAGE); |
| 417 | } |
| 418 | |
| 419 | private: |
| 420 | enum { |
| 421 | #ifdef FORCE_SUPERMASK |
| 422 | kMAX_WIDTH = 2048, |
| 423 | kMAX_STORAGE = 1024 * 1024 * 2 |
| 424 | #else |
| 425 | kMAX_WIDTH = 32, // so we don't try to do very wide things, where the RLE blitter would be faster |
| 426 | kMAX_STORAGE = 1024 |
| 427 | #endif |
| 428 | }; |
| 429 | |
| 430 | SkMask fMask; |
| 431 | SkIRect fClipRect; |
| 432 | // we add 1 because add_aa_span can write (unchanged) 1 extra byte at the end, rather than |
| 433 | // perform a test to see if stopAlpha != 0 |
| 434 | uint32_t fStorage[(kMAX_STORAGE >> 2) + 1]; |
| 435 | }; |
| 436 | |
| 437 | MaskSuperBlitter::MaskSuperBlitter(SkBlitter* realBlitter, const SkIRect& ir, |
| 438 | const SkIRect& clipBounds, bool isInverse) |
| 439 | : BaseSuperBlitter(realBlitter, ir, clipBounds, isInverse) |
| 440 | { |
| 441 | SkASSERT(CanHandleRect(ir)); |
| 442 | SkASSERT(!isInverse); |
| 443 | |
| 444 | fMask.fImage = (uint8_t*)fStorage; |
| 445 | fMask.fBounds = ir; |
| 446 | fMask.fRowBytes = ir.width(); |
| 447 | fMask.fFormat = SkMask::kA8_Format; |
| 448 | |
| 449 | fClipRect = ir; |
| 450 | if (!fClipRect.intersect(clipBounds)) { |
| 451 | SkASSERT(0); |
| 452 | fClipRect.setEmpty(); |
| 453 | } |
| 454 | |
| 455 | // For valgrind, write 1 extra byte at the end so we don't read |
| 456 | // uninitialized memory. See comment in add_aa_span and fStorage[]. |
| 457 | memset(fStorage, 0, fMask.fBounds.height() * fMask.fRowBytes + 1); |
| 458 | } |
| 459 | |
| 460 | static void add_aa_span(uint8_t* alpha, U8CPU startAlpha) { |
| 461 | /* I should be able to just add alpha[x] + startAlpha. |
| 462 | However, if the trailing edge of the previous span and the leading |
| 463 | edge of the current span round to the same super-sampled x value, |
| 464 | I might overflow to 256 with this add, hence the funny subtract. |
| 465 | */ |
| 466 | unsigned tmp = *alpha + startAlpha; |
| 467 | SkASSERT(tmp <= 256); |
| 468 | *alpha = SkToU8(tmp - (tmp >> 8)); |
| 469 | } |
| 470 | |
| 471 | static inline uint32_t quadplicate_byte(U8CPU value) { |
| 472 | uint32_t pair = (value << 8) | value; |
| 473 | return (pair << 16) | pair; |
| 474 | } |
| 475 | |
| 476 | // Perform this tricky subtract, to avoid overflowing to 256. Our caller should |
| 477 | // only ever call us with at most enough to hit 256 (never larger), so it is |
| 478 | // enough to just subtract the high-bit. Actually clamping with a branch would |
| 479 | // be slower (e.g. if (tmp > 255) tmp = 255;) |
| 480 | // |
| 481 | static inline void saturated_add(uint8_t* ptr, U8CPU add) { |
| 482 | unsigned tmp = *ptr + add; |
| 483 | SkASSERT(tmp <= 256); |
| 484 | *ptr = SkToU8(tmp - (tmp >> 8)); |
| 485 | } |
| 486 | |
| 487 | // minimum count before we want to setup an inner loop, adding 4-at-a-time |
| 488 | #define MIN_COUNT_FOR_QUAD_LOOP 16 |
| 489 | |
| 490 | static void add_aa_span(uint8_t* alpha, U8CPU startAlpha, int middleCount, |
| 491 | U8CPU stopAlpha, U8CPU maxValue) { |
| 492 | SkASSERT(middleCount >= 0); |
| 493 | |
| 494 | saturated_add(alpha, startAlpha); |
| 495 | alpha += 1; |
| 496 | |
| 497 | if (middleCount >= MIN_COUNT_FOR_QUAD_LOOP) { |
| 498 | // loop until we're quad-byte aligned |
| 499 | while (reinterpret_cast<intptr_t>(alpha) & 0x3) { |
| 500 | alpha[0] = SkToU8(alpha[0] + maxValue); |
| 501 | alpha += 1; |
| 502 | middleCount -= 1; |
| 503 | } |
| 504 | |
| 505 | int bigCount = middleCount >> 2; |
| 506 | uint32_t* qptr = reinterpret_cast<uint32_t*>(alpha); |
| 507 | uint32_t qval = quadplicate_byte(maxValue); |
| 508 | do { |
| 509 | *qptr++ += qval; |
| 510 | } while (--bigCount > 0); |
| 511 | |
| 512 | middleCount &= 3; |
| 513 | alpha = reinterpret_cast<uint8_t*> (qptr); |
| 514 | // fall through to the following while-loop |
| 515 | } |
| 516 | |
| 517 | while (--middleCount >= 0) { |
| 518 | alpha[0] = SkToU8(alpha[0] + maxValue); |
| 519 | alpha += 1; |
| 520 | } |
| 521 | |
| 522 | // potentially this can be off the end of our "legal" alpha values, but that |
| 523 | // only happens if stopAlpha is also 0. Rather than test for stopAlpha != 0 |
| 524 | // every time (slow), we just do it, and ensure that we've allocated extra space |
| 525 | // (see the + 1 comment in fStorage[] |
| 526 | saturated_add(alpha, stopAlpha); |
| 527 | } |
| 528 | |
| 529 | void MaskSuperBlitter::blitH(int x, int y, int width) { |
| 530 | int iy = (y >> SHIFT); |
| 531 | |
| 532 | SkASSERT(iy >= fMask.fBounds.fTop && iy < fMask.fBounds.fBottom); |
| 533 | iy -= fMask.fBounds.fTop; // make it relative to 0 |
| 534 | |
| 535 | // This should never happen, but it does. Until the true cause is |
| 536 | // discovered, let's skip this span instead of crashing. |
| 537 | // See http://crbug.com/17569. |
| 538 | if (iy < 0) { |
| 539 | return; |
| 540 | } |
| 541 | |
| 542 | #ifdef SK_DEBUG |
| 543 | { |
| 544 | int ix = x >> SHIFT; |
| 545 | SkASSERT(ix >= fMask.fBounds.fLeft && ix < fMask.fBounds.fRight); |
| 546 | } |
| 547 | #endif |
| 548 | |
| 549 | x -= SkLeftShift(fMask.fBounds.fLeft, SHIFT); |
| 550 | |
| 551 | // hack, until I figure out why my cubics (I think) go beyond the bounds |
| 552 | if (x < 0) { |
| 553 | width += x; |
| 554 | x = 0; |
| 555 | } |
| 556 | |
| 557 | uint8_t* row = fMask.fImage + iy * fMask.fRowBytes + (x >> SHIFT); |
| 558 | |
| 559 | int start = x; |
| 560 | int stop = x + width; |
| 561 | |
| 562 | SkASSERT(start >= 0 && stop > start); |
| 563 | int fb = start & MASK; |
| 564 | int fe = stop & MASK; |
| 565 | int n = (stop >> SHIFT) - (start >> SHIFT) - 1; |
| 566 | |
| 567 | |
| 568 | if (n < 0) { |
| 569 | SkASSERT(row >= fMask.fImage); |
| 570 | SkASSERT(row < fMask.fImage + kMAX_STORAGE + 1); |
| 571 | add_aa_span(row, coverage_to_partial_alpha(fe - fb)); |
| 572 | } else { |
| 573 | fb = SCALE - fb; |
| 574 | SkASSERT(row >= fMask.fImage); |
| 575 | SkASSERT(row + n + 1 < fMask.fImage + kMAX_STORAGE + 1); |
| 576 | add_aa_span(row, coverage_to_partial_alpha(fb), |
| 577 | n, coverage_to_partial_alpha(fe), |
| 578 | (1 << (8 - SHIFT)) - (((y & MASK) + 1) >> SHIFT)); |
| 579 | } |
| 580 | |
| 581 | #ifdef SK_DEBUG |
| 582 | fCurrX = x + width; |
| 583 | #endif |
| 584 | } |
| 585 | |
| 586 | /////////////////////////////////////////////////////////////////////////////// |
| 587 | |
| 588 | static SkIRect safeRoundOut(const SkRect& src) { |
| 589 | // roundOut will pin huge floats to max/min int |
| 590 | SkIRect dst = src.roundOut(); |
| 591 | |
| 592 | // intersect with a smaller huge rect, so the rect will not be considered empty for being |
| 593 | // too large. e.g. { -SK_MaxS32 ... SK_MaxS32 } is considered empty because its width |
| 594 | // exceeds signed 32bit. |
| 595 | const int32_t limit = SK_MaxS32 >> SK_SUPERSAMPLE_SHIFT; |
| 596 | (void)dst.intersect({ -limit, -limit, limit, limit}); |
| 597 | |
| 598 | return dst; |
| 599 | } |
| 600 | |
| 601 | constexpr int kSampleSize = 8; |
| 602 | #if !defined(SK_DISABLE_AAA) |
| 603 | constexpr SkScalar kComplexityThreshold = 0.25; |
| 604 | #endif |
| 605 | |
| 606 | static void compute_complexity(const SkPath& path, SkScalar& avgLength, SkScalar& complexity) { |
| 607 | int n = path.countPoints(); |
| 608 | if (n < kSampleSize || path.getBounds().isEmpty()) { |
| 609 | // set to invalid value to indicate that we failed to compute |
| 610 | avgLength = complexity = -1; |
| 611 | return; |
| 612 | } |
| 613 | |
| 614 | SkScalar sumLength = 0; |
| 615 | SkPoint lastPoint = path.getPoint(0); |
| 616 | for(int i = 1; i < kSampleSize; ++i) { |
| 617 | SkPoint point = path.getPoint(i); |
| 618 | sumLength += SkPoint::Distance(lastPoint, point); |
| 619 | lastPoint = point; |
| 620 | } |
| 621 | avgLength = sumLength / (kSampleSize - 1); |
| 622 | |
| 623 | auto sqr = [](SkScalar x) { return x*x; }; |
| 624 | |
| 625 | SkScalar diagonalSqr = sqr(path.getBounds().width()) + sqr(path.getBounds().height()); |
| 626 | |
| 627 | // If the path consists of random line segments, the number of intersections should be |
| 628 | // proportional to this. |
| 629 | SkScalar intersections = sk_ieee_float_divide(sqr(n) * sqr(avgLength), diagonalSqr); |
| 630 | |
| 631 | // The number of intersections per scanline should be proportional to this number. |
| 632 | complexity = sk_ieee_float_divide(intersections, path.getBounds().height()); |
| 633 | |
| 634 | if (sk_float_isnan(complexity)) { // it may be possible to have 0.0 / 0.0; inf is fine for us. |
| 635 | complexity = -1; |
| 636 | } |
| 637 | } |
| 638 | |
| 639 | static bool ShouldUseAAA(const SkPath& path, SkScalar avgLength, SkScalar complexity) { |
| 640 | #if defined(SK_DISABLE_AAA) |
| 641 | return false; |
| 642 | #else |
| 643 | if (gSkForceAnalyticAA) { |
| 644 | return true; |
| 645 | } |
| 646 | if (!gSkUseAnalyticAA) { |
| 647 | return false; |
| 648 | } |
| 649 | if (path.isRect(nullptr)) { |
| 650 | return true; |
| 651 | } |
| 652 | |
| 653 | #ifdef SK_SUPPORT_LEGACY_AAA_CHOICE |
| 654 | const SkRect& bounds = path.getBounds(); |
| 655 | // When the path have so many points compared to the size of its |
| 656 | // bounds/resolution, it indicates that the path is not quite smooth in |
| 657 | // the current resolution: the expected number of turning points in |
| 658 | // every pixel row/column is significantly greater than zero. Hence |
| 659 | // Aanlytic AA is not likely to produce visible quality improvements, |
| 660 | // and Analytic AA might be slower than supersampling. |
| 661 | return path.countPoints() < std::max(bounds.width(), bounds.height()) / 2 - 10; |
| 662 | #else |
| 663 | if (path.countPoints() >= path.getBounds().height()) { |
| 664 | // SAA is faster than AAA in this case even if there are no |
| 665 | // intersections because AAA will have too many scan lines. See |
| 666 | // skbug.com/8272 |
| 667 | return false; |
| 668 | } |
| 669 | // We will use AAA if the number of verbs < kSampleSize and therefore complexity < 0 |
| 670 | return complexity < kComplexityThreshold; |
| 671 | #endif |
| 672 | #endif |
| 673 | } |
| 674 | |
| 675 | void SkScan::SAAFillPath(const SkPath& path, SkBlitter* blitter, const SkIRect& ir, |
| 676 | const SkIRect& clipBounds, bool forceRLE) { |
| 677 | bool containedInClip = clipBounds.contains(ir); |
| 678 | bool isInverse = path.isInverseFillType(); |
| 679 | |
| 680 | // MaskSuperBlitter can't handle drawing outside of ir, so we can't use it |
| 681 | // if we're an inverse filltype |
| 682 | if (!isInverse && MaskSuperBlitter::CanHandleRect(ir) && !forceRLE) { |
| 683 | MaskSuperBlitter superBlit(blitter, ir, clipBounds, isInverse); |
| 684 | SkASSERT(SkIntToScalar(ir.fTop) <= path.getBounds().fTop); |
| 685 | sk_fill_path(path, clipBounds, &superBlit, ir.fTop, ir.fBottom, SHIFT, containedInClip); |
| 686 | } else { |
| 687 | SuperBlitter superBlit(blitter, ir, clipBounds, isInverse); |
| 688 | sk_fill_path(path, clipBounds, &superBlit, ir.fTop, ir.fBottom, SHIFT, containedInClip); |
| 689 | } |
| 690 | } |
| 691 | |
| 692 | static int overflows_short_shift(int value, int shift) { |
| 693 | const int s = 16 + shift; |
| 694 | return (SkLeftShift(value, s) >> s) - value; |
| 695 | } |
| 696 | |
| 697 | /** |
| 698 | Would any of the coordinates of this rectangle not fit in a short, |
| 699 | when left-shifted by shift? |
| 700 | */ |
| 701 | static int rect_overflows_short_shift(SkIRect rect, int shift) { |
| 702 | SkASSERT(!overflows_short_shift(8191, shift)); |
| 703 | SkASSERT(overflows_short_shift(8192, shift)); |
| 704 | SkASSERT(!overflows_short_shift(32767, 0)); |
| 705 | SkASSERT(overflows_short_shift(32768, 0)); |
| 706 | |
| 707 | // Since we expect these to succeed, we bit-or together |
| 708 | // for a tiny extra bit of speed. |
| 709 | return overflows_short_shift(rect.fLeft, shift) | |
| 710 | overflows_short_shift(rect.fRight, shift) | |
| 711 | overflows_short_shift(rect.fTop, shift) | |
| 712 | overflows_short_shift(rect.fBottom, shift); |
| 713 | } |
| 714 | |
| 715 | void SkScan::AntiFillPath(const SkPath& path, const SkRegion& origClip, |
| 716 | SkBlitter* blitter, bool forceRLE) { |
| 717 | if (origClip.isEmpty()) { |
| 718 | return; |
| 719 | } |
| 720 | |
| 721 | const bool isInverse = path.isInverseFillType(); |
| 722 | SkIRect ir = safeRoundOut(path.getBounds()); |
| 723 | if (ir.isEmpty()) { |
| 724 | if (isInverse) { |
| 725 | blitter->blitRegion(origClip); |
| 726 | } |
| 727 | return; |
| 728 | } |
| 729 | |
| 730 | // If the intersection of the path bounds and the clip bounds |
| 731 | // will overflow 32767 when << by SHIFT, we can't supersample, |
| 732 | // so draw without antialiasing. |
| 733 | SkIRect clippedIR; |
| 734 | if (isInverse) { |
| 735 | // If the path is an inverse fill, it's going to fill the entire |
| 736 | // clip, and we care whether the entire clip exceeds our limits. |
| 737 | clippedIR = origClip.getBounds(); |
| 738 | } else { |
| 739 | if (!clippedIR.intersect(ir, origClip.getBounds())) { |
| 740 | return; |
| 741 | } |
| 742 | } |
| 743 | if (rect_overflows_short_shift(clippedIR, SHIFT)) { |
| 744 | SkScan::FillPath(path, origClip, blitter); |
| 745 | return; |
| 746 | } |
| 747 | |
| 748 | // Our antialiasing can't handle a clip larger than 32767, so we restrict |
| 749 | // the clip to that limit here. (the runs[] uses int16_t for its index). |
| 750 | // |
| 751 | // A more general solution (one that could also eliminate the need to |
| 752 | // disable aa based on ir bounds (see overflows_short_shift) would be |
| 753 | // to tile the clip/target... |
| 754 | SkRegion tmpClipStorage; |
| 755 | const SkRegion* clipRgn = &origClip; |
| 756 | { |
| 757 | static const int32_t kMaxClipCoord = 32767; |
| 758 | const SkIRect& bounds = origClip.getBounds(); |
| 759 | if (bounds.fRight > kMaxClipCoord || bounds.fBottom > kMaxClipCoord) { |
| 760 | SkIRect limit = { 0, 0, kMaxClipCoord, kMaxClipCoord }; |
| 761 | tmpClipStorage.op(origClip, limit, SkRegion::kIntersect_Op); |
| 762 | clipRgn = &tmpClipStorage; |
| 763 | } |
| 764 | } |
| 765 | // for here down, use clipRgn, not origClip |
| 766 | |
| 767 | SkScanClipper clipper(blitter, clipRgn, ir); |
| 768 | |
| 769 | if (clipper.getBlitter() == nullptr) { // clipped out |
| 770 | if (isInverse) { |
| 771 | blitter->blitRegion(*clipRgn); |
| 772 | } |
| 773 | return; |
| 774 | } |
| 775 | |
| 776 | SkASSERT(clipper.getClipRect() == nullptr || |
| 777 | *clipper.getClipRect() == clipRgn->getBounds()); |
| 778 | |
| 779 | // now use the (possibly wrapped) blitter |
| 780 | blitter = clipper.getBlitter(); |
| 781 | |
| 782 | if (isInverse) { |
| 783 | sk_blit_above(blitter, ir, *clipRgn); |
| 784 | } |
| 785 | |
| 786 | SkScalar avgLength, complexity; |
| 787 | compute_complexity(path, avgLength, complexity); |
| 788 | |
| 789 | if (ShouldUseAAA(path, avgLength, complexity)) { |
| 790 | // Do not use AAA if path is too complicated: |
| 791 | // there won't be any speedup or significant visual improvement. |
| 792 | SkScan::AAAFillPath(path, blitter, ir, clipRgn->getBounds(), forceRLE); |
| 793 | } else { |
| 794 | SkScan::SAAFillPath(path, blitter, ir, clipRgn->getBounds(), forceRLE); |
| 795 | } |
| 796 | |
| 797 | if (isInverse) { |
| 798 | sk_blit_below(blitter, ir, *clipRgn); |
| 799 | } |
| 800 | } |
| 801 | |
| 802 | /////////////////////////////////////////////////////////////////////////////// |
| 803 | |
| 804 | #include "src/core/SkRasterClip.h" |
| 805 | |
| 806 | void SkScan::FillPath(const SkPath& path, const SkRasterClip& clip, SkBlitter* blitter) { |
| 807 | if (clip.isEmpty() || !path.isFinite()) { |
| 808 | return; |
| 809 | } |
| 810 | |
| 811 | if (clip.isBW()) { |
| 812 | FillPath(path, clip.bwRgn(), blitter); |
| 813 | } else { |
| 814 | SkRegion tmp; |
| 815 | SkAAClipBlitter aaBlitter; |
| 816 | |
| 817 | tmp.setRect(clip.getBounds()); |
| 818 | aaBlitter.init(blitter, &clip.aaRgn()); |
| 819 | SkScan::FillPath(path, tmp, &aaBlitter); |
| 820 | } |
| 821 | } |
| 822 | |
| 823 | void SkScan::AntiFillPath(const SkPath& path, const SkRasterClip& clip, SkBlitter* blitter) { |
| 824 | if (clip.isEmpty() || !path.isFinite()) { |
| 825 | return; |
| 826 | } |
| 827 | |
| 828 | if (clip.isBW()) { |
| 829 | AntiFillPath(path, clip.bwRgn(), blitter, false); |
| 830 | } else { |
| 831 | SkRegion tmp; |
| 832 | SkAAClipBlitter aaBlitter; |
| 833 | |
| 834 | tmp.setRect(clip.getBounds()); |
| 835 | aaBlitter.init(blitter, &clip.aaRgn()); |
| 836 | AntiFillPath(path, tmp, &aaBlitter, true); // SkAAClipBlitter can blitMask, why forceRLE? |
| 837 | } |
| 838 | } |
| 839 | |