OLD | NEW |
1 /* | 1 /* |
2 * Copyright 2014 Google Inc. | 2 * Copyright 2014 Google Inc. |
3 * | 3 * |
4 * Use of this source code is governed by a BSD-style license that can be | 4 * Use of this source code is governed by a BSD-style license that can be |
5 * found in the LICENSE file. | 5 * found in the LICENSE file. |
6 */ | 6 */ |
7 | 7 |
8 #include "SkTextureCompressor.h" | 8 #include "SkTextureCompressor.h" |
9 | 9 |
10 #include "SkBitmap.h" | 10 #include "SkBitmap.h" |
(...skipping 262 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
273 // This is really just for correctness, in all of my tests we | 273 // This is really just for correctness, in all of my tests we |
274 // never take this step. We don't lose too much perf here because | 274 // never take this step. We don't lose too much perf here because |
275 // most of the processing in this function is worth it for the | 275 // most of the processing in this function is worth it for the |
276 // 1 == nUniquePixels optimization. | 276 // 1 == nUniquePixels optimization. |
277 return compress_latc_block_bb(pixels); | 277 return compress_latc_block_bb(pixels); |
278 } else { | 278 } else { |
279 return compress_latc_block_bb_ignore_extremal(pixels); | 279 return compress_latc_block_bb_ignore_extremal(pixels); |
280 } | 280 } |
281 } | 281 } |
282 | 282 |
283 static bool compress_a8_to_latc(uint8_t* dst, const uint8_t* src, | 283 static inline bool compress_a8_to_latc(uint8_t* dst, const uint8_t* src, |
284 int width, int height, int rowBytes) { | 284 int width, int height, int rowBytes) { |
285 return compress_4x4_a8_to_64bit(dst, src, width, height, rowBytes, compress_
latc_block); | 285 return compress_4x4_a8_to_64bit(dst, src, width, height, rowBytes, compress_
latc_block); |
286 } | 286 } |
287 | 287 |
288 //////////////////////////////////////////////////////////////////////////////// | 288 //////////////////////////////////////////////////////////////////////////////// |
289 // | 289 // |
290 // R11 EAC Compressor | 290 // R11 EAC Compressor |
291 // | 291 // |
292 //////////////////////////////////////////////////////////////////////////////// | 292 //////////////////////////////////////////////////////////////////////////////// |
293 | 293 |
| 294 // #define COMPRESS_R11_EAC_SLOW 1 |
| 295 // #define COMPRESS_R11_EAC_FAST 1 |
| 296 #define COMPRESS_R11_EAC_FASTEST 1 |
| 297 |
294 // Blocks compressed into R11 EAC are represented as follows: | 298 // Blocks compressed into R11 EAC are represented as follows: |
295 // 0000000000000000000000000000000000000000000000000000000000000000 | 299 // 0000000000000000000000000000000000000000000000000000000000000000 |
296 // |base_cw|mod|mul| ----------------- indices ------------------- | 300 // |base_cw|mod|mul| ----------------- indices ------------------- |
297 // | 301 // |
298 // To reconstruct the value of a given pixel, we use the formula: | 302 // To reconstruct the value of a given pixel, we use the formula: |
299 // clamp[0, 2047](base_cw * 8 + 4 + mod_val*mul*8) | 303 // clamp[0, 2047](base_cw * 8 + 4 + mod_val*mul*8) |
300 // | 304 // |
301 // mod_val is chosen from a palette of values based on the index of the | 305 // mod_val is chosen from a palette of values based on the index of the |
302 // given pixel. The palette is chosen by the value stored in mod. | 306 // given pixel. The palette is chosen by the value stored in mod. |
303 // This formula returns a value between 0 and 2047, which is converted | 307 // This formula returns a value between 0 and 2047, which is converted |
(...skipping 16 matching lines...) Expand all Loading... |
320 {-2, -6, -8, -10, 1, 5, 7, 9}, | 324 {-2, -6, -8, -10, 1, 5, 7, 9}, |
321 {-2, -5, -8, -10, 1, 4, 7, 9}, | 325 {-2, -5, -8, -10, 1, 4, 7, 9}, |
322 {-2, -4, -8, -10, 1, 3, 7, 9}, | 326 {-2, -4, -8, -10, 1, 3, 7, 9}, |
323 {-2, -5, -7, -10, 1, 4, 6, 9}, | 327 {-2, -5, -7, -10, 1, 4, 6, 9}, |
324 {-3, -4, -7, -10, 2, 3, 6, 9}, | 328 {-3, -4, -7, -10, 2, 3, 6, 9}, |
325 {-1, -2, -3, -10, 0, 1, 2, 9}, | 329 {-1, -2, -3, -10, 0, 1, 2, 9}, |
326 {-4, -6, -8, -9, 3, 5, 7, 8}, | 330 {-4, -6, -8, -9, 3, 5, 7, 8}, |
327 {-3, -5, -7, -9, 2, 4, 6, 8} | 331 {-3, -5, -7, -9, 2, 4, 6, 8} |
328 }; | 332 }; |
329 | 333 |
| 334 #if COMPRESS_R11_EAC_SLOW |
330 // Pack the base codeword, palette, and multiplier into the 64 bits necessary | 335 // Pack the base codeword, palette, and multiplier into the 64 bits necessary |
331 // to decode it. | 336 // to decode it. |
332 static uint64_t pack_r11eac_block(uint16_t base_cw, uint16_t palette, uint16_t m
ultiplier, | 337 static uint64_t pack_r11eac_block(uint16_t base_cw, uint16_t palette, uint16_t m
ultiplier, |
333 uint64_t indices) { | 338 uint64_t indices) { |
334 SkASSERT(palette < 16); | 339 SkASSERT(palette < 16); |
335 SkASSERT(multiplier < 16); | 340 SkASSERT(multiplier < 16); |
336 SkASSERT(indices < (static_cast<uint64_t>(1) << 48)); | 341 SkASSERT(indices < (static_cast<uint64_t>(1) << 48)); |
337 | 342 |
338 const uint64_t b = static_cast<uint64_t>(base_cw) << 56; | 343 const uint64_t b = static_cast<uint64_t>(base_cw) << 56; |
339 const uint64_t m = static_cast<uint64_t>(multiplier) << 52; | 344 const uint64_t m = static_cast<uint64_t>(multiplier) << 52; |
340 const uint64_t p = static_cast<uint64_t>(palette) << 48; | 345 const uint64_t p = static_cast<uint64_t>(palette) << 48; |
341 return SkEndian_SwapBE64(b | m | p | indices); | 346 return SkEndian_SwapBE64(b | m | p | indices); |
342 } | 347 } |
343 | 348 |
344 // Given a base codeword, a modifier, and a multiplier, compute the proper | 349 // Given a base codeword, a modifier, and a multiplier, compute the proper |
345 // pixel value in the range [0, 2047]. | 350 // pixel value in the range [0, 2047]. |
346 static uint16_t compute_r11eac_pixel(int base_cw, int modifier, int multiplier)
{ | 351 static uint16_t compute_r11eac_pixel(int base_cw, int modifier, int multiplier)
{ |
347 int ret = (base_cw * 8 + 4) + (modifier * multiplier * 8); | 352 int ret = (base_cw * 8 + 4) + (modifier * multiplier * 8); |
348 return (ret > 2047)? 2047 : ((ret < 0)? 0 : ret); | 353 return (ret > 2047)? 2047 : ((ret < 0)? 0 : ret); |
349 } | 354 } |
350 | 355 |
351 // Compress a block into R11 EAC format. | 356 // Compress a block into R11 EAC format. |
352 // The compression works as follows: | 357 // The compression works as follows: |
353 // 1. Find the center of the span of the block's values. Use this as the base co
deword. | 358 // 1. Find the center of the span of the block's values. Use this as the base co
deword. |
354 // 2. Choose a multiplier based roughly on the size of the span of block values | 359 // 2. Choose a multiplier based roughly on the size of the span of block values |
355 // 3. Iterate through each palette and choose the one with the most accurate | 360 // 3. Iterate through each palette and choose the one with the most accurate |
356 // modifiers. | 361 // modifiers. |
357 static uint64_t compress_heterogeneous_r11eac_block(const uint8_t block[16]) { | 362 static inline uint64_t compress_heterogeneous_r11eac_block(const uint8_t block[1
6]) { |
358 // Find the center of the data... | 363 // Find the center of the data... |
359 uint16_t bmin = block[0]; | 364 uint16_t bmin = block[0]; |
360 uint16_t bmax = block[0]; | 365 uint16_t bmax = block[0]; |
361 for (int i = 1; i < 16; ++i) { | 366 for (int i = 1; i < 16; ++i) { |
362 bmin = SkTMin<uint16_t>(bmin, block[i]); | 367 bmin = SkTMin<uint16_t>(bmin, block[i]); |
363 bmax = SkTMax<uint16_t>(bmax, block[i]); | 368 bmax = SkTMax<uint16_t>(bmax, block[i]); |
364 } | 369 } |
365 | 370 |
366 uint16_t center = (bmax + bmin) >> 1; | 371 uint16_t center = (bmax + bmin) >> 1; |
367 SkASSERT(center <= 255); | 372 SkASSERT(center <= 255); |
368 | 373 |
369 // Based on the min and max, we can guesstimate a proper multiplier | 374 // Based on the min and max, we can guesstimate a proper multiplier |
370 // This is kind of a magic choice to start with. | 375 // This is kind of a magic choice to start with. |
371 uint16_t multiplier = (bmax - center) / 10; | 376 uint16_t multiplier = (bmax - center) / 10; |
372 | 377 |
373 // Now convert the block to 11 bits and transpose it to match | 378 // Now convert the block to 11 bits and transpose it to match |
374 // the proper layout | 379 // the proper layout |
375 uint16_t cblock[16]; | 380 uint16_t cblock[16]; |
376 for (int i = 0; i < 4; ++i) { | 381 for (int i = 0; i < 4; ++i) { |
377 for (int j = 0; j < 4; ++j) { | 382 for (int j = 0; j < 4; ++j) { |
378 int srcIdx = i*4+j; | 383 int srcIdx = i*4+j; |
379 int dstIdx = j*4+i; | 384 int dstIdx = j*4+i; |
380 cblock[dstIdx] = (block[srcIdx] << 3) | (block[srcIdx] >> 5); | 385 cblock[dstIdx] = (block[srcIdx] << 3) | (block[srcIdx] >> 5); |
381 } | 386 } |
382 } | 387 } |
383 | 388 |
384 // Finally, choose the proper palette and indices | 389 // Finally, choose the proper palette and indices |
385 uint32_t bestError = static_cast<uint32_t>(-1); | 390 uint32_t bestError = 0xFFFFFFFF; |
386 uint64_t bestIndices = 0; | 391 uint64_t bestIndices = 0; |
387 uint16_t bestPalette = 0; | 392 uint16_t bestPalette = 0; |
388 for (uint16_t paletteIdx = 0; paletteIdx < kNumR11EACPalettes; ++paletteIdx)
{ | 393 for (uint16_t paletteIdx = 0; paletteIdx < kNumR11EACPalettes; ++paletteIdx)
{ |
389 const int *palette = kR11EACModifierPalettes[paletteIdx]; | 394 const int *palette = kR11EACModifierPalettes[paletteIdx]; |
390 | 395 |
391 // Iterate through each pixel to find the best palette index | 396 // Iterate through each pixel to find the best palette index |
392 // and update the indices with the choice. Also store the error | 397 // and update the indices with the choice. Also store the error |
393 // for this palette to be compared against the best error... | 398 // for this palette to be compared against the best error... |
394 uint32_t error = 0; | 399 uint32_t error = 0; |
395 uint64_t indices = 0; | 400 uint64_t indices = 0; |
(...skipping 29 matching lines...) Expand all Loading... |
425 if (error < bestError) { | 430 if (error < bestError) { |
426 bestPalette = paletteIdx; | 431 bestPalette = paletteIdx; |
427 bestIndices = indices; | 432 bestIndices = indices; |
428 bestError = error; | 433 bestError = error; |
429 } | 434 } |
430 } | 435 } |
431 | 436 |
432 // Finally, pack everything together... | 437 // Finally, pack everything together... |
433 return pack_r11eac_block(center, bestPalette, multiplier, bestIndices); | 438 return pack_r11eac_block(center, bestPalette, multiplier, bestIndices); |
434 } | 439 } |
| 440 #endif // COMPRESS_R11_EAC_SLOW |
435 | 441 |
| 442 #if COMPRESS_R11_EAC_FAST |
| 443 // This function takes into account that most blocks that we compress have a gra
dation from |
| 444 // fully opaque to fully transparent. The compression scheme works by selecting
the |
| 445 // palette and multiplier that has the tightest fit to the 0-255 range. This is
encoded |
| 446 // as the block header (0x8490). The indices are then selected by considering th
e top |
| 447 // three bits of each alpha value. For alpha masks, this reduces the dynamic ran
ge from |
| 448 // 17 to 8, but the quality is still acceptable. |
| 449 // |
| 450 // There are a few caveats that need to be taken care of... |
| 451 // |
| 452 // 1. The block is read in as scanlines, so the indices are stored as: |
| 453 // 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 |
| 454 // However, the decomrpession routine reads them in column-major order, so th
ey |
| 455 // need to be packed as: |
| 456 // 0 4 8 12 1 5 9 13 2 6 10 14 3 7 11 15 |
| 457 // So when reading, they must be transposed. |
| 458 // |
| 459 // 2. We cannot use the top three bits as an index directly, since the R11 EAC p
alettes |
| 460 // above store the modulation values first decreasing and then increasing: |
| 461 // e.g. {-3, -6, -9, -15, 2, 5, 8, 14} |
| 462 // Hence, we need to convert the indices with the following mapping: |
| 463 // From: 0 1 2 3 4 5 6 7 |
| 464 // To: 3 2 1 0 4 5 6 7 |
| 465 static inline uint64_t compress_heterogeneous_r11eac_block(const uint8_t block[1
6]) { |
| 466 uint64_t retVal = static_cast<uint64_t>(0x8490) << 48; |
| 467 for(int i = 0; i < 4; ++i) { |
| 468 for(int j = 0; j < 4; ++j) { |
| 469 const int shift = 45-3*(j*4+i); |
| 470 SkASSERT(shift <= 45); |
| 471 const uint64_t idx = block[i*4+j] >> 5; |
| 472 SkASSERT(idx < 8); |
| 473 |
| 474 // !SPEED! This is slightly faster than having an if-statement. |
| 475 switch(idx) { |
| 476 case 0: |
| 477 case 1: |
| 478 case 2: |
| 479 case 3: |
| 480 retVal |= (3-idx) << shift; |
| 481 break; |
| 482 default: |
| 483 retVal |= idx << shift; |
| 484 break; |
| 485 } |
| 486 } |
| 487 } |
| 488 |
| 489 return SkEndian_SwapBE64(retVal); |
| 490 } |
| 491 #endif // COMPRESS_R11_EAC_FAST |
| 492 |
| 493 #if (COMPRESS_R11_EAC_SLOW) || (COMPRESS_R11_EAC_FAST) |
436 static uint64_t compress_r11eac_block(const uint8_t block[16]) { | 494 static uint64_t compress_r11eac_block(const uint8_t block[16]) { |
437 // Are all blocks a solid color? | 495 // Are all blocks a solid color? |
438 bool solid = true; | 496 bool solid = true; |
439 for (int i = 1; i < 16; ++i) { | 497 for (int i = 1; i < 16; ++i) { |
440 if (block[i] != block[0]) { | 498 if (block[i] != block[0]) { |
441 solid = false; | 499 solid = false; |
442 break; | 500 break; |
443 } | 501 } |
444 } | 502 } |
445 | 503 |
446 // Fully transparent? We know the encoding... | 504 if (solid) { |
447 if (solid && 0 == block[0]) { | 505 switch(block[0]) { |
448 // (0x0060 << 48) produces the following: | 506 // Fully transparent? We know the encoding... |
449 // basw_cw: 0 | 507 case 0: |
450 // mod: 6, palette: {-4, -7, -8, -11, 3, 6, 7, 10} | 508 // (0x0020 << 48) produces the following: |
451 // mod_val: -3 | 509 // basw_cw: 0 |
452 // | 510 // mod: 0, palette: {-3, -6, -9, -15, 2, 5, 8, 14} |
453 // this gives the following formula: | 511 // multiplier: 2 |
454 // clamp[0, 2047](0*8+4+(-4)) = 0 | 512 // mod_val: -3 |
455 return SkEndian_SwapBE64(static_cast<uint64_t>(0x0060) << 48); | 513 // |
456 | 514 // this gives the following formula: |
457 // Fully opaque? We know this encoding too... | 515 // clamp[0, 2047](0*8+4+(-3)*2*8) = 0 |
458 } else if (solid && 255 == block[0]) { | 516 // |
459 // -1 produces the following: | 517 // Furthermore, it is impervious to endianness: |
460 // basw_cw: 255 | 518 // 0x0020000000002000ULL |
461 // mod: 15, palette: {-3, -5, -7, -9, 2, 4, 6, 8} | 519 // Will produce one pixel with index 2, which gives: |
462 // mod_val: 8 | 520 // clamp[0, 2047](0*8+4+(-9)*2*8) = 0 |
463 // | 521 return 0x0020000000002000ULL; |
464 // this gives the following formula: | 522 |
465 // clamp[0, 2047](255*8+4+8*8*8) = clamp[0, 2047](2556) = 2047 | 523 // Fully opaque? We know this encoding too... |
466 return static_cast<uint64_t>(-1); | 524 case 255: |
467 } | 525 |
468 | 526 // -1 produces the following: |
469 #if 0 | 527 // basw_cw: 255 |
470 else if (solid) { | 528 // mod: 15, palette: {-3, -5, -7, -9, 2, 4, 6, 8} |
471 // !TODO! krajcevski: | 529 // mod_val: 8 |
472 // This will probably never happen, since we're using this format | 530 // |
473 // primarily for compressing alpha maps. Usually the only | 531 // this gives the following formula: |
474 // non-fullly opaque or fully transparent blocks are not a solid | 532 // clamp[0, 2047](255*8+4+8*8*8) = clamp[0, 2047](2556) = 2047 |
475 // intermediate color. If we notice that they are, then we can | 533 return 0xFFFFFFFFFFFFFFFFULL; |
476 // add another optimization... | 534 |
477 } | 535 default: |
| 536 // !TODO! krajcevski: |
| 537 // This will probably never happen, since we're using this forma
t |
| 538 // primarily for compressing alpha maps. Usually the only |
| 539 // non-fullly opaque or fully transparent blocks are not a solid |
| 540 // intermediate color. If we notice that they are, then we can |
| 541 // add another optimization... |
| 542 break; |
| 543 } |
| 544 } |
| 545 |
| 546 return compress_heterogeneous_r11eac_block(block); |
| 547 } |
| 548 #endif // (COMPRESS_R11_EAC_SLOW) || (COMPRESS_R11_EAC_FAST) |
| 549 |
| 550 #if COMPRESS_R11_EAC_FASTEST |
| 551 static inline uint64_t interleave6(uint64_t topRows, uint64_t bottomRows) { |
| 552 // If our 3-bit block indices are laid out as: |
| 553 // a b c d |
| 554 // e f g h |
| 555 // i j k l |
| 556 // m n o p |
| 557 // |
| 558 // This function expects topRows and bottomRows to contain the first two row
s |
| 559 // of indices interleaved in the least significant bits of a and b. In other
words... |
| 560 // |
| 561 // If the architecture is big endian, then topRows and bottomRows will conta
in the following: |
| 562 // Bits 31-0: |
| 563 // a: 00 a e 00 b f 00 c g 00 d h |
| 564 // b: 00 i m 00 j n 00 k o 00 l p |
| 565 // |
| 566 // If the architecture is little endian, then topRows and bottomRows will co
ntain |
| 567 // the following: |
| 568 // Bits 31-0: |
| 569 // a: 00 d h 00 c g 00 b f 00 a e |
| 570 // b: 00 l p 00 k o 00 j n 00 i m |
| 571 // |
| 572 // This function returns a 48-bit packing of the form: |
| 573 // a e i m b f j n c g k o d h l p |
| 574 // |
| 575 // !SPEED! this function might be even faster if certain SIMD intrinsics are |
| 576 // used.. |
| 577 |
| 578 // For both architectures, we can figure out a packing of the bits by |
| 579 // using a shuffle and a few shift-rotates... |
| 580 uint64_t x = (static_cast<uint64_t>(topRows) << 32) | static_cast<uint64_t>(
bottomRows); |
| 581 |
| 582 // x: 00 a e 00 b f 00 c g 00 d h 00 i m 00 j n 00 k o 00 l p |
| 583 |
| 584 uint64_t t = (x ^ (x >> 10)) & 0x3FC0003FC00000ULL; |
| 585 x = x ^ t ^ (t << 10); |
| 586 |
| 587 // x: b f 00 00 00 a e c g i m 00 00 00 d h j n 00 k o 00 l p |
| 588 |
| 589 x |= ((x << 52) & (0x3FULL << 52)); |
| 590 x = (x | ((x << 20) & (0x3FULL << 28))) >> 16; |
| 591 |
| 592 #if defined (SK_CPU_BENDIAN) |
| 593 // x: 00 00 00 00 00 00 00 00 b f l p a e c g i m k o d h j n |
| 594 |
| 595 t = (x ^ (x >> 6)) & 0xFC0000ULL; |
| 596 x = x ^ t ^ (t << 6); |
| 597 |
| 598 // x: 00 00 00 00 00 00 00 00 b f l p a e i m c g k o d h j n |
| 599 |
| 600 t = (x ^ (x >> 36)) & 0x3FULL; |
| 601 x = x ^ t ^ (t << 36); |
| 602 |
| 603 // x: 00 00 00 00 00 00 00 00 b f j n a e i m c g k o d h l p |
| 604 |
| 605 t = (x ^ (x >> 12)) & 0xFFF000000ULL; |
| 606 x = x ^ t ^ (t << 12); |
| 607 |
| 608 // x: 00 00 00 00 00 00 00 00 a e i m b f j n c g k o d h l p |
| 609 return x; |
| 610 #else |
| 611 // If our CPU is little endian, then the above logic will |
| 612 // produce the following indices: |
| 613 // x: 00 00 00 00 00 00 00 00 c g i m d h b f l p j n a e k o |
| 614 |
| 615 t = (x ^ (x >> 6)) & 0xFC0000ULL; |
| 616 x = x ^ t ^ (t << 6); |
| 617 |
| 618 // x: 00 00 00 00 00 00 00 00 c g i m d h l p b f j n a e k o |
| 619 |
| 620 t = (x ^ (x >> 36)) & 0xFC0ULL; |
| 621 x = x ^ t ^ (t << 36); |
| 622 |
| 623 // x: 00 00 00 00 00 00 00 00 a e i m d h l p b f j n c g k o |
| 624 |
| 625 x = (x & (0xFFFULL << 36)) | ((x & 0xFFFFFFULL) << 12) | ((x >> 24) & 0xFFFU
LL); |
| 626 |
| 627 // x: 00 00 00 00 00 00 00 00 a e i m b f j n c g k o d h l p |
| 628 |
| 629 return x; |
478 #endif | 630 #endif |
479 | 631 } |
480 return compress_heterogeneous_r11eac_block(block); | 632 |
481 } | 633 // This function converts an integer containing four bytes of alpha |
482 | 634 // values into an integer containing four bytes of indices into R11 EAC. |
483 static bool compress_a8_to_r11eac(uint8_t* dst, const uint8_t* src, | 635 // Note, there needs to be a mapping of indices: |
484 int width, int height, int rowBytes) { | 636 // 0 1 2 3 4 5 6 7 |
| 637 // 3 2 1 0 4 5 6 7 |
| 638 // |
| 639 // To compute this, we first negate each byte, and then add three, which |
| 640 // gives the mapping |
| 641 // 3 2 1 0 -1 -2 -3 -4 |
| 642 // |
| 643 // Then we mask out the negative values, take their absolute value, and |
| 644 // add three. |
| 645 // |
| 646 // Most of the voodoo in this function comes from Hacker's Delight, section 2-18 |
| 647 static inline uint32_t convert_indices(uint32_t x) { |
| 648 // Take the top three bits... |
| 649 x = (x & 0xE0E0E0E0) >> 5; |
| 650 |
| 651 // Negate... |
| 652 x = ~((0x80808080 - x) ^ 0x7F7F7F7F); |
| 653 |
| 654 // Add three |
| 655 const uint32_t s = (x & 0x7F7F7F7F) + 0x03030303; |
| 656 x = ((x ^ 0x03030303) & 0x80808080) ^ s; |
| 657 |
| 658 // Absolute value |
| 659 const uint32_t a = x & 0x80808080; |
| 660 const uint32_t b = a >> 7; |
| 661 |
| 662 // Aside: mask negatives (m is three if the byte was negative) |
| 663 const uint32_t m = (a >> 6) | b; |
| 664 |
| 665 // .. continue absolute value |
| 666 x = (x ^ ((a - b) | a)) + b; |
| 667 |
| 668 // Add three |
| 669 return x + m; |
| 670 } |
| 671 |
| 672 // This function follows the same basic procedure as compress_heterogeneous_r11e
ac_block |
| 673 // above when COMPRESS_R11_EAC_FAST is defined, but it avoids a few loads/stores
and |
| 674 // tries to optimize where it can using SIMD. |
| 675 static uint64_t compress_r11eac_block_fast(const uint8_t* src, int rowBytes) { |
| 676 // Store each row of alpha values in an integer |
| 677 const uint32_t alphaRow1 = *(reinterpret_cast<const uint32_t*>(src)); |
| 678 const uint32_t alphaRow2 = *(reinterpret_cast<const uint32_t*>(src + rowByte
s)); |
| 679 const uint32_t alphaRow3 = *(reinterpret_cast<const uint32_t*>(src + 2*rowBy
tes)); |
| 680 const uint32_t alphaRow4 = *(reinterpret_cast<const uint32_t*>(src + 3*rowBy
tes)); |
| 681 |
| 682 // Check for solid blocks. The explanations for these values |
| 683 // can be found in the comments of compress_r11eac_block above |
| 684 if (alphaRow1 == alphaRow2 && alphaRow1 == alphaRow3 && alphaRow1 == alphaRo
w4) { |
| 685 if (0 == alphaRow1) { |
| 686 // Fully transparent block |
| 687 return 0x0020000000002000ULL; |
| 688 } else if (0xFFFFFFFF == alphaRow1) { |
| 689 // Fully opaque block |
| 690 return 0xFFFFFFFFFFFFFFFFULL; |
| 691 } |
| 692 } |
| 693 |
| 694 // Convert each integer of alpha values into an integer of indices |
| 695 const uint32_t indexRow1 = convert_indices(alphaRow1); |
| 696 const uint32_t indexRow2 = convert_indices(alphaRow2); |
| 697 const uint32_t indexRow3 = convert_indices(alphaRow3); |
| 698 const uint32_t indexRow4 = convert_indices(alphaRow4); |
| 699 |
| 700 // Interleave the indices from the top two rows and bottom two rows |
| 701 // prior to passing them to interleave6. Since each index is at most |
| 702 // three bits, then each byte can hold two indices... The way that the |
| 703 // compression scheme expects the packing allows us to efficiently pack |
| 704 // the top two rows and bottom two rows. Interleaving each 6-bit sequence |
| 705 // and tightly packing it into a uint64_t is a little trickier, which is |
| 706 // taken care of in interleave6. |
| 707 const uint32_t r1r2 = (indexRow1 << 3) | indexRow2; |
| 708 const uint32_t r3r4 = (indexRow3 << 3) | indexRow4; |
| 709 const uint64_t indices = interleave6(r1r2, r3r4); |
| 710 |
| 711 // Return the packed incdices in the least significant bits with the magic h
eader |
| 712 return SkEndian_SwapBE64(0x8490000000000000ULL | indices); |
| 713 } |
| 714 |
| 715 static bool compress_a8_to_r11eac_fast(uint8_t* dst, const uint8_t* src, |
| 716 int width, int height, int rowBytes) { |
| 717 // Make sure that our data is well-formed enough to be considered for compre
ssion |
| 718 if (0 == width || 0 == height || (width % 4) != 0 || (height % 4) != 0) { |
| 719 return false; |
| 720 } |
| 721 |
| 722 const int blocksX = width >> 2; |
| 723 const int blocksY = height >> 2; |
| 724 |
| 725 uint64_t* encPtr = reinterpret_cast<uint64_t*>(dst); |
| 726 for (int y = 0; y < blocksY; ++y) { |
| 727 for (int x = 0; x < blocksX; ++x) { |
| 728 // Compress it |
| 729 *encPtr = compress_r11eac_block_fast(src + 4*x, rowBytes); |
| 730 ++encPtr; |
| 731 } |
| 732 src += 4 * rowBytes; |
| 733 } |
| 734 return true; |
| 735 } |
| 736 #endif // COMPRESS_R11_EAC_FASTEST |
| 737 |
| 738 static inline bool compress_a8_to_r11eac(uint8_t* dst, const uint8_t* src, |
| 739 int width, int height, int rowBytes) { |
| 740 #if (COMPRESS_R11_EAC_SLOW) || (COMPRESS_R11_EAC_FAST) |
485 return compress_4x4_a8_to_64bit(dst, src, width, height, rowBytes, compress_
r11eac_block); | 741 return compress_4x4_a8_to_64bit(dst, src, width, height, rowBytes, compress_
r11eac_block); |
| 742 #elif COMPRESS_R11_EAC_FASTEST |
| 743 return compress_a8_to_r11eac_fast(dst, src, width, height, rowBytes); |
| 744 #else |
| 745 #error "Must choose R11 EAC algorithm" |
| 746 #endif |
486 } | 747 } |
487 | 748 |
488 //////////////////////////////////////////////////////////////////////////////// | 749 //////////////////////////////////////////////////////////////////////////////// |
489 | 750 |
490 namespace SkTextureCompressor { | 751 namespace SkTextureCompressor { |
491 | 752 |
492 static size_t get_compressed_data_size(Format fmt, int width, int height) { | 753 static inline size_t get_compressed_data_size(Format fmt, int width, int height)
{ |
493 switch (fmt) { | 754 switch (fmt) { |
| 755 // These formats are 64 bits per 4x4 block. |
494 case kR11_EAC_Format: | 756 case kR11_EAC_Format: |
495 case kLATC_Format: | 757 case kLATC_Format: |
496 { | 758 { |
497 // The LATC format is 64 bits per 4x4 block. | |
498 static const int kLATCEncodedBlockSize = 8; | 759 static const int kLATCEncodedBlockSize = 8; |
499 | 760 |
500 int blocksX = width / kLATCBlockSize; | 761 const int blocksX = width / kLATCBlockSize; |
501 int blocksY = height / kLATCBlockSize; | 762 const int blocksY = height / kLATCBlockSize; |
502 | 763 |
503 return blocksX * blocksY * kLATCEncodedBlockSize; | 764 return blocksX * blocksY * kLATCEncodedBlockSize; |
504 } | 765 } |
505 | 766 |
506 default: | 767 default: |
507 SkFAIL("Unknown compressed format!"); | 768 SkFAIL("Unknown compressed format!"); |
508 return 0; | 769 return 0; |
509 } | 770 } |
510 } | 771 } |
511 | 772 |
512 typedef bool (*CompressBitmapProc)(uint8_t* dst, const uint8_t* src, | 773 typedef bool (*CompressBitmapProc)(uint8_t* dst, const uint8_t* src, |
513 int width, int height, int rowBytes); | 774 int width, int height, int rowBytes); |
514 | 775 |
515 bool CompressBufferToFormat(uint8_t* dst, const uint8_t* src, SkColorType srcCol
orType, | 776 bool CompressBufferToFormat(uint8_t* dst, const uint8_t* src, SkColorType srcCol
orType, |
516 int width, int height, int rowBytes, Format format)
{ | 777 int width, int height, int rowBytes, Format format)
{ |
517 | 778 |
518 CompressBitmapProc kProcMap[kFormatCnt][kLastEnum_SkColorType + 1]; | 779 CompressBitmapProc kProcMap[kFormatCnt][kLastEnum_SkColorType + 1]; |
519 memset(kProcMap, 0, sizeof(kProcMap)); | 780 memset(kProcMap, 0, sizeof(kProcMap)); |
520 | 781 |
521 kProcMap[kLATC_Format][kAlpha_8_SkColorType] = compress_a8_to_latc; | 782 kProcMap[kLATC_Format][kAlpha_8_SkColorType] = compress_a8_to_latc; |
522 kProcMap[kR11_EAC_Format][kAlpha_8_SkColorType] = compress_a8_to_r11eac; | 783 kProcMap[kR11_EAC_Format][kAlpha_8_SkColorType] = compress_a8_to_r11eac; |
523 | 784 |
524 CompressBitmapProc proc = kProcMap[format][srcColorType]; | 785 CompressBitmapProc proc = kProcMap[format][srcColorType]; |
525 if (NULL != proc) { | 786 if (NULL != proc) { |
526 return proc(dst, src, width, height, rowBytes); | 787 return proc(dst, src, width, height, rowBytes); |
527 } | 788 } |
528 | 789 |
529 return false; | 790 return false; |
530 } | 791 } |
531 | 792 |
532 SkData *CompressBitmapToFormat(const SkBitmap &bitmap, Format format) { | 793 SkData *CompressBitmapToFormat(const SkBitmap &bitmap, Format format) { |
533 SkAutoLockPixels alp(bitmap); | 794 SkAutoLockPixels alp(bitmap); |
534 | 795 |
535 int compressedDataSize = get_compressed_data_size(format, bitmap.width(), bi
tmap.height()); | 796 int compressedDataSize = get_compressed_data_size(format, bitmap.width(), bi
tmap.height()); |
536 const uint8_t* src = reinterpret_cast<const uint8_t*>(bitmap.getPixels()); | 797 const uint8_t* src = reinterpret_cast<const uint8_t*>(bitmap.getPixels()); |
537 uint8_t* dst = reinterpret_cast<uint8_t*>(sk_malloc_throw(compressedDataSize
)); | 798 uint8_t* dst = reinterpret_cast<uint8_t*>(sk_malloc_throw(compressedDataSize
)); |
538 if (CompressBufferToFormat(dst, src, bitmap.colorType(), bitmap.width(), bit
map.height(), | 799 if (CompressBufferToFormat(dst, src, bitmap.colorType(), bitmap.width(), bit
map.height(), |
539 bitmap.rowBytes(), format)) { | 800 bitmap.rowBytes(), format)) { |
540 return SkData::NewFromMalloc(dst, compressedDataSize); | 801 return SkData::NewFromMalloc(dst, compressedDataSize); |
541 } | 802 } |
542 | 803 |
543 sk_free(dst); | 804 sk_free(dst); |
544 return NULL; | 805 return NULL; |
545 } | 806 } |
546 | 807 |
547 } // namespace SkTextureCompressor | 808 } // namespace SkTextureCompressor |
OLD | NEW |