Sk4px_SSE2.h 2.8 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576
  1. /*
  2. * Copyright 2015 Google Inc.
  3. *
  4. * Use of this source code is governed by a BSD-style license that can be
  5. * found in the LICENSE file.
  6. */
  7. namespace { // NOLINT(google-build-namespaces)
  8. inline Sk4px::Wide Sk4px::widen() const {
  9. return Sk16h(_mm_unpacklo_epi8(this->fVec, _mm_setzero_si128()),
  10. _mm_unpackhi_epi8(this->fVec, _mm_setzero_si128()));
  11. }
  12. inline Sk4px::Wide Sk4px::mulWiden(const Sk16b& other) const {
  13. return this->widen() * Sk4px(other).widen();
  14. }
  15. inline Sk4px Sk4px::Wide::addNarrowHi(const Sk16h& other) const {
  16. Sk4px::Wide r = (*this + other) >> 8;
  17. return Sk4px(_mm_packus_epi16(r.fLo.fVec, r.fHi.fVec));
  18. }
  19. inline Sk4px Sk4px::Wide::div255() const {
  20. // (x + 127) / 255 == ((x+128) * 257)>>16,
  21. // and _mm_mulhi_epu16 makes the (_ * 257)>>16 part very convenient.
  22. const __m128i _128 = _mm_set1_epi16(128),
  23. _257 = _mm_set1_epi16(257);
  24. return Sk4px(_mm_packus_epi16(_mm_mulhi_epu16(_mm_add_epi16(fLo.fVec, _128), _257),
  25. _mm_mulhi_epu16(_mm_add_epi16(fHi.fVec, _128), _257)));
  26. }
  27. // Load4Alphas and Load2Alphas use possibly-unaligned loads (SkAlpha[] -> uint16_t or uint32_t).
  28. // These are safe on x86, often with no speed penalty.
  29. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
  30. inline Sk4px Sk4px::alphas() const {
  31. static_assert(SK_A32_SHIFT == 24, "Intel's always little-endian.");
  32. __m128i splat = _mm_set_epi8(15,15,15,15, 11,11,11,11, 7,7,7,7, 3,3,3,3);
  33. return Sk16b(_mm_shuffle_epi8(this->fVec, splat));
  34. }
  35. inline Sk4px Sk4px::Load4Alphas(const SkAlpha a[4]) {
  36. uint32_t as;
  37. memcpy(&as, a, 4);
  38. __m128i splat = _mm_set_epi8(3,3,3,3, 2,2,2,2, 1,1,1,1, 0,0,0,0);
  39. return Sk16b(_mm_shuffle_epi8(_mm_cvtsi32_si128(as), splat));
  40. }
  41. #else
  42. inline Sk4px Sk4px::alphas() const {
  43. static_assert(SK_A32_SHIFT == 24, "Intel's always little-endian.");
  44. // We exploit that A >= rgb for any premul pixel.
  45. __m128i as = fVec; // 3xxx 2xxx 1xxx 0xxx
  46. as = _mm_max_epu8(as, _mm_srli_epi32(as, 8)); // 33xx 22xx 11xx 00xx
  47. as = _mm_max_epu8(as, _mm_srli_epi32(as, 16)); // 3333 2222 1111 0000
  48. return Sk16b(as);
  49. }
  50. inline Sk4px Sk4px::Load4Alphas(const SkAlpha a[4]) {
  51. __m128i as;
  52. memcpy(&as, a, 4); // ____ ____ ____ 3210
  53. as = _mm_unpacklo_epi8 (as, as); // ____ ____ 3322 1100
  54. as = _mm_unpacklo_epi16(as, as); // 3333 2222 1111 0000
  55. return Sk16b(as);
  56. }
  57. #endif
  58. inline Sk4px Sk4px::Load2Alphas(const SkAlpha a[2]) {
  59. uint16_t alphas;
  60. memcpy(&alphas, a, 2);
  61. uint32_t alphas_and_two_zeros = alphas; // Aa -> Aa00
  62. return Load4Alphas((const SkAlpha*)&alphas_and_two_zeros);
  63. }
  64. } // namespace