OLD | NEW |
1 /* | 1 /* |
2 * Copyright 2012 The Android Open Source Project | 2 * Copyright 2012 The Android Open Source Project |
3 * | 3 * |
4 * Use of this source code is governed by a BSD-style license that can be | 4 * Use of this source code is governed by a BSD-style license that can be |
5 * found in the LICENSE file. | 5 * found in the LICENSE file. |
6 */ | 6 */ |
7 | 7 |
8 #include "SkBlitRow_opts_arm_neon.h" | 8 #include "SkBlitRow_opts_arm_neon.h" |
9 | 9 |
10 #include "SkBlitMask.h" | 10 #include "SkBlitMask.h" |
(...skipping 447 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
458 SkPMColorAssert(c); | 458 SkPMColorAssert(c); |
459 if (c) { | 459 if (c) { |
460 *dst = SkSrcOver32To16(c, *dst); | 460 *dst = SkSrcOver32To16(c, *dst); |
461 } | 461 } |
462 dst += 1; | 462 dst += 1; |
463 } while (--count != 0); | 463 } while (--count != 0); |
464 } | 464 } |
465 } | 465 } |
466 #endif // #ifdef SK_CPU_ARM32 | 466 #endif // #ifdef SK_CPU_ARM32 |
467 | 467 |
| 468 static uint32_t pmcolor_to_expand16(SkPMColor c) { |
| 469 unsigned r = SkGetPackedR32(c); |
| 470 unsigned g = SkGetPackedG32(c); |
| 471 unsigned b = SkGetPackedB32(c); |
| 472 return (g << 24) | (r << 13) | (b << 2); |
| 473 } |
| 474 |
| 475 void Color32A_D565_neon(uint16_t dst[], SkPMColor src, int count, int x, int y)
{ |
| 476 uint32_t src_expand; |
| 477 unsigned scale; |
| 478 uint16x8_t vmask_blue; |
| 479 |
| 480 if (count <= 0) return; |
| 481 SkASSERT(((size_t)dst & 0x01) == 0); |
| 482 |
| 483 /* |
| 484 * This preamble code is in order to make dst aligned to 8 bytes |
| 485 * in the next mutiple bytes read & write access. |
| 486 */ |
| 487 src_expand = pmcolor_to_expand16(src); |
| 488 scale = SkAlpha255To256(0xFF - SkGetPackedA32(src)) >> 3; |
| 489 |
| 490 #define DST_ALIGN 8 |
| 491 |
| 492 /* |
| 493 * preamble_size is in byte, meantime, this blend32_16_row_neon updates 2 by
tes at a time. |
| 494 */ |
| 495 int preamble_size = (DST_ALIGN - (size_t)dst) & (DST_ALIGN - 1); |
| 496 |
| 497 for (int i = 0; i < preamble_size; i+=2, dst++) { |
| 498 uint32_t dst_expand = SkExpand_rgb_16(*dst) * scale; |
| 499 *dst = SkCompact_rgb_16((src_expand + dst_expand) >> 5); |
| 500 if (--count == 0) |
| 501 break; |
| 502 } |
| 503 |
| 504 int count16 = 0; |
| 505 count16 = count >> 4; |
| 506 vmask_blue = vmovq_n_u16(SK_B16_MASK); |
| 507 |
| 508 if (count16) { |
| 509 uint16x8_t wide_sr; |
| 510 uint16x8_t wide_sg; |
| 511 uint16x8_t wide_sb; |
| 512 uint16x8_t wide_256_sa; |
| 513 |
| 514 unsigned sr = SkGetPackedR32(src); |
| 515 unsigned sg = SkGetPackedG32(src); |
| 516 unsigned sb = SkGetPackedB32(src); |
| 517 unsigned sa = SkGetPackedA32(src); |
| 518 |
| 519 // Operation: dst_rgb = src_rgb + ((256 - src_a) >> 3) x dst_rgb |
| 520 // sr: 8-bit based, dr: 5-bit based, with dr x ((256-sa)>>3), 5-bit left
shifted, |
| 521 //thus, for sr, do 2-bit left shift to match MSB : (8 + 2 = 5 + 5) |
| 522 wide_sr = vshlq_n_u16(vmovl_u8(vdup_n_u8(sr)), 2); // widen and src_red
shift |
| 523 |
| 524 // sg: 8-bit based, dg: 6-bit based, with dg x ((256-sa)>>3), 5-bit left
shifted, |
| 525 //thus, for sg, do 3-bit left shift to match MSB : (8 + 3 = 6 + 5) |
| 526 wide_sg = vshlq_n_u16(vmovl_u8(vdup_n_u8(sg)), 3); // widen and src_grn
shift |
| 527 |
| 528 // sb: 8-bit based, db: 5-bit based, with db x ((256-sa)>>3), 5-bit left
shifted, |
| 529 //thus, for sb, do 2-bit left shift to match MSB : (8 + 2 = 5 + 5) |
| 530 wide_sb = vshlq_n_u16(vmovl_u8(vdup_n_u8(sb)), 2); // widen and src blu
shift |
| 531 |
| 532 wide_256_sa = |
| 533 vshrq_n_u16(vsubw_u8(vdupq_n_u16(256), vdup_n_u8(sa)), 3); // (256 -
sa) >> 3 |
| 534 |
| 535 while (count16-- > 0) { |
| 536 uint16x8_t vdst1, vdst1_r, vdst1_g, vdst1_b; |
| 537 uint16x8_t vdst2, vdst2_r, vdst2_g, vdst2_b; |
| 538 vdst1 = vld1q_u16(dst); |
| 539 dst += 8; |
| 540 vdst2 = vld1q_u16(dst); |
| 541 dst -= 8; //to store dst again. |
| 542 |
| 543 vdst1_g = vshlq_n_u16(vdst1, SK_R16_BITS); // shift
green to top of lanes |
| 544 vdst1_b = vdst1 & vmask_blue; // extrac
t blue |
| 545 vdst1_r = vshrq_n_u16(vdst1, SK_R16_SHIFT); // extrac
t red |
| 546 vdst1_g = vshrq_n_u16(vdst1_g, SK_R16_BITS + SK_B16_BITS); // extrac
t green |
| 547 |
| 548 vdst2_g = vshlq_n_u16(vdst2, SK_R16_BITS); // shift
green to top of lanes |
| 549 vdst2_b = vdst2 & vmask_blue; // extrac
t blue |
| 550 vdst2_r = vshrq_n_u16(vdst2, SK_R16_SHIFT); // extrac
t red |
| 551 vdst2_g = vshrq_n_u16(vdst2_g, SK_R16_BITS + SK_B16_BITS); // extrac
t green |
| 552 |
| 553 vdst1_r = vmlaq_u16(wide_sr, wide_256_sa, vdst1_r); // sr + (
256-sa) x dr1 |
| 554 vdst1_g = vmlaq_u16(wide_sg, wide_256_sa, vdst1_g); // sg + (
256-sa) x dg1 |
| 555 vdst1_b = vmlaq_u16(wide_sb, wide_256_sa, vdst1_b); // sb + (
256-sa) x db1 |
| 556 |
| 557 vdst2_r = vmlaq_u16(wide_sr, wide_256_sa, vdst2_r); // sr + (
256-sa) x dr2 |
| 558 vdst2_g = vmlaq_u16(wide_sg, wide_256_sa, vdst2_g); // sg + (
256-sa) x dg2 |
| 559 vdst2_b = vmlaq_u16(wide_sb, wide_256_sa, vdst2_b); // sb + (
256-sa) x db2 |
| 560 |
| 561 vdst1_r = vshrq_n_u16(vdst1_r, 5); // 5-bit
right shift for 5-bit red |
| 562 vdst1_g = vshrq_n_u16(vdst1_g, 5); // 5-bit
right shift for 6-bit green |
| 563 vdst1_b = vshrq_n_u16(vdst1_b, 5); // 5-bit
right shift for 5-bit blue |
| 564 |
| 565 vdst1 = vsliq_n_u16(vdst1_b, vdst1_g, SK_G16_SHIFT); // insert
green into blue |
| 566 vdst1 = vsliq_n_u16(vdst1, vdst1_r, SK_R16_SHIFT); // insert
red into green/blue |
| 567 |
| 568 vdst2_r = vshrq_n_u16(vdst2_r, 5); // 5-bit
right shift for 5-bit red |
| 569 vdst2_g = vshrq_n_u16(vdst2_g, 5); // 5-bit
right shift for 6-bit green |
| 570 vdst2_b = vshrq_n_u16(vdst2_b, 5); // 5-bit
right shift for 5-bit blue |
| 571 |
| 572 vdst2 = vsliq_n_u16(vdst2_b, vdst2_g, SK_G16_SHIFT); // insert
green into blue |
| 573 vdst2 = vsliq_n_u16(vdst2, vdst2_r, SK_R16_SHIFT); // insert
red into green/blue |
| 574 |
| 575 vst1q_u16(dst, vdst1); |
| 576 dst += 8; |
| 577 vst1q_u16(dst, vdst2); |
| 578 dst += 8; |
| 579 } |
| 580 } |
| 581 |
| 582 count &= 0xF; |
| 583 if (count > 0) { |
| 584 do { |
| 585 uint32_t dst_expand = SkExpand_rgb_16(*dst) * scale; |
| 586 *dst = SkCompact_rgb_16((src_expand + dst_expand) >> 5); |
| 587 dst += 1; |
| 588 } while (--count != 0); |
| 589 } |
| 590 } |
| 591 |
468 static inline uint16x8_t SkDiv255Round_neon8(uint16x8_t prod) { | 592 static inline uint16x8_t SkDiv255Round_neon8(uint16x8_t prod) { |
469 prod += vdupq_n_u16(128); | 593 prod += vdupq_n_u16(128); |
470 prod += vshrq_n_u16(prod, 8); | 594 prod += vshrq_n_u16(prod, 8); |
471 return vshrq_n_u16(prod, 8); | 595 return vshrq_n_u16(prod, 8); |
472 } | 596 } |
473 | 597 |
474 void S32A_D565_Blend_neon(uint16_t* SK_RESTRICT dst, | 598 void S32A_D565_Blend_neon(uint16_t* SK_RESTRICT dst, |
475 const SkPMColor* SK_RESTRICT src, int count, | 599 const SkPMColor* SK_RESTRICT src, int count, |
476 U8CPU alpha, int /*x*/, int /*y*/) { | 600 U8CPU alpha, int /*x*/, int /*y*/) { |
477 SkASSERT(255 > alpha); | 601 SkASSERT(255 > alpha); |
(...skipping 1180 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
1658 // https://code.google.com/p/skia/issues/detail?id=2797 | 1782 // https://code.google.com/p/skia/issues/detail?id=2797 |
1659 #endif | 1783 #endif |
1660 | 1784 |
1661 // dither | 1785 // dither |
1662 S32_D565_Opaque_Dither_neon, | 1786 S32_D565_Opaque_Dither_neon, |
1663 S32_D565_Blend_Dither_neon, | 1787 S32_D565_Blend_Dither_neon, |
1664 S32A_D565_Opaque_Dither_neon, | 1788 S32A_D565_Opaque_Dither_neon, |
1665 NULL, // S32A_D565_Blend_Dither | 1789 NULL, // S32A_D565_Blend_Dither |
1666 }; | 1790 }; |
1667 | 1791 |
| 1792 const SkBlitRow::ColorProc16 sk_blitrow_platform_565_colorprocs_arm_neon[] = { |
| 1793 Color32A_D565_neon, // Color32_D565, |
| 1794 Color32A_D565_neon, // Color32A_D565, |
| 1795 Color32A_D565_neon, // Color32_D565_Dither, |
| 1796 Color32A_D565_neon, // Color32A_D565_Dither |
| 1797 }; |
| 1798 |
1668 const SkBlitRow::Proc32 sk_blitrow_platform_32_procs_arm_neon[] = { | 1799 const SkBlitRow::Proc32 sk_blitrow_platform_32_procs_arm_neon[] = { |
1669 NULL, // S32_Opaque, | 1800 NULL, // S32_Opaque, |
1670 S32_Blend_BlitRow32_neon, // S32_Blend, | 1801 S32_Blend_BlitRow32_neon, // S32_Blend, |
1671 /* | 1802 /* |
1672 * We have two choices for S32A_Opaque procs. The one reads the src alpha | 1803 * We have two choices for S32A_Opaque procs. The one reads the src alpha |
1673 * value and attempts to optimize accordingly. The optimization is | 1804 * value and attempts to optimize accordingly. The optimization is |
1674 * sensitive to the source content and is not a win in all cases. For | 1805 * sensitive to the source content and is not a win in all cases. For |
1675 * example, if there are a lot of transitions between the alpha states, | 1806 * example, if there are a lot of transitions between the alpha states, |
1676 * the performance will almost certainly be worse. However, for many | 1807 * the performance will almost certainly be worse. However, for many |
1677 * common cases the performance is equivalent or better than the standard | 1808 * common cases the performance is equivalent or better than the standard |
1678 * case where we do not inspect the src alpha. | 1809 * case where we do not inspect the src alpha. |
1679 */ | 1810 */ |
1680 #if SK_A32_SHIFT == 24 | 1811 #if SK_A32_SHIFT == 24 |
1681 // This proc assumes the alpha value occupies bits 24-32 of each SkPMColor | 1812 // This proc assumes the alpha value occupies bits 24-32 of each SkPMColor |
1682 S32A_Opaque_BlitRow32_neon_src_alpha, // S32A_Opaque, | 1813 S32A_Opaque_BlitRow32_neon_src_alpha, // S32A_Opaque, |
1683 #else | 1814 #else |
1684 S32A_Opaque_BlitRow32_neon, // S32A_Opaque, | 1815 S32A_Opaque_BlitRow32_neon, // S32A_Opaque, |
1685 #endif | 1816 #endif |
1686 #ifdef SK_CPU_ARM32 | 1817 #ifdef SK_CPU_ARM32 |
1687 S32A_Blend_BlitRow32_neon // S32A_Blend | 1818 S32A_Blend_BlitRow32_neon // S32A_Blend |
1688 #else | 1819 #else |
1689 NULL | 1820 NULL |
1690 #endif | 1821 #endif |
1691 }; | 1822 }; |
OLD | NEW |