OLD | NEW |
1 /* | 1 /* |
2 * Copyright 2014 The Android Open Source Project | 2 * Copyright 2014 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 #ifndef SkColor_opts_SSE2_DEFINED | 8 #ifndef SkColor_opts_SSE2_DEFINED |
9 #define SkColor_opts_SSE2_DEFINED | 9 #define SkColor_opts_SSE2_DEFINED |
10 | 10 |
(...skipping 24 matching lines...) Expand all Loading... |
35 static inline __m128i SkAlphaMulAlpha_SSE2(const __m128i& a, | 35 static inline __m128i SkAlphaMulAlpha_SSE2(const __m128i& a, |
36 const __m128i& b) { | 36 const __m128i& b) { |
37 __m128i prod = _mm_mullo_epi16(a, b); | 37 __m128i prod = _mm_mullo_epi16(a, b); |
38 prod = _mm_add_epi32(prod, _mm_set1_epi32(128)); | 38 prod = _mm_add_epi32(prod, _mm_set1_epi32(128)); |
39 prod = _mm_add_epi32(prod, _mm_srli_epi32(prod, 8)); | 39 prod = _mm_add_epi32(prod, _mm_srli_epi32(prod, 8)); |
40 prod = _mm_srli_epi32(prod, 8); | 40 prod = _mm_srli_epi32(prod, 8); |
41 | 41 |
42 return prod; | 42 return prod; |
43 } | 43 } |
44 | 44 |
| 45 static const __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); |
| 46 static const __m128i ag_mask = _mm_set1_epi32(0xFF00FF00); |
| 47 |
45 // Portable version SkAlphaMulQ is in SkColorPriv.h. | 48 // Portable version SkAlphaMulQ is in SkColorPriv.h. |
46 static inline __m128i SkAlphaMulQ_SSE2(const __m128i& c, const __m128i& scale) { | 49 static inline __m128i SkAlphaMulQ_SSE2(const __m128i& c, const __m128i& scale) { |
47 __m128i mask = _mm_set1_epi32(0xFF00FF); | |
48 __m128i s = _mm_or_si128(_mm_slli_epi32(scale, 16), scale); | 50 __m128i s = _mm_or_si128(_mm_slli_epi32(scale, 16), scale); |
49 | 51 |
50 // uint32_t rb = ((c & mask) * scale) >> 8 | 52 // uint32_t rb = ((c & mask) * scale) >> 8 |
51 __m128i rb = _mm_and_si128(mask, c); | 53 __m128i rb = _mm_and_si128(rb_mask, c); |
52 rb = _mm_mullo_epi16(rb, s); | 54 rb = _mm_mullo_epi16(rb, s); |
53 rb = _mm_srli_epi16(rb, 8); | 55 rb = _mm_srli_epi16(rb, 8); |
54 | 56 |
55 // uint32_t ag = ((c >> 8) & mask) * scale | 57 // uint32_t ag = ((c >> 8) & mask) * scale |
56 __m128i ag = _mm_srli_epi16(c, 8); | 58 __m128i ag = _mm_srli_epi16(c, 8); |
57 ASSERT_EQ(ag, _mm_and_si128(mask, ag)); // ag = _mm_srli_epi16(c, 8) did th
is for us. | 59 ASSERT_EQ(ag, _mm_and_si128(rb_mask, ag)); // ag = _mm_srli_epi16(c, 8) did
this for us. |
58 ag = _mm_mullo_epi16(ag, s); | 60 ag = _mm_mullo_epi16(ag, s); |
59 | 61 |
60 // (rb & mask) | (ag & ~mask) | 62 // (rb & mask) | (ag & ~mask) |
61 ASSERT_EQ(rb, _mm_and_si128(mask, rb)); // rb = _mm_srli_epi16(rb, 8) did t
his for us. | 63 ASSERT_EQ(rb, _mm_and_si128(rb_mask, rb)); // rb = _mm_srli_epi16(rb, 8) di
d this for us. |
62 ag = _mm_andnot_si128(mask, ag); | 64 ag = _mm_and_si128(ag_mask, ag); |
63 return _mm_or_si128(rb, ag); | 65 return _mm_or_si128(rb, ag); |
64 } | 66 } |
65 | 67 |
| 68 // Fast path for SkAlphaMulQ_SSE2 with a constant scale factor. |
| 69 static inline __m128i SkAlphaMulQ_SSE2(const __m128i& c, const unsigned scale) { |
| 70 __m128i s = _mm_set1_epi16(scale << 8); // Move scale factor to upper byte o
f word. |
| 71 |
| 72 // With mulhi, red and blue values are already in the right place and |
| 73 // don't need to be divided by 256. |
| 74 __m128i rb = _mm_and_si128(rb_mask, c); |
| 75 rb = _mm_mulhi_epu16(rb, s); |
| 76 |
| 77 __m128i ag = _mm_and_si128(ag_mask, c); |
| 78 ag = _mm_mulhi_epu16(ag, s); // Alpha and green values are in the higher
byte of each word. |
| 79 ag = _mm_and_si128(ag_mask, ag); |
| 80 |
| 81 return _mm_or_si128(rb, ag); |
| 82 } |
| 83 |
66 static inline __m128i SkGetPackedA32_SSE2(const __m128i& src) { | 84 static inline __m128i SkGetPackedA32_SSE2(const __m128i& src) { |
67 __m128i a = _mm_slli_epi32(src, (24 - SK_A32_SHIFT)); | 85 __m128i a = _mm_slli_epi32(src, (24 - SK_A32_SHIFT)); |
68 return _mm_srli_epi32(a, 24); | 86 return _mm_srli_epi32(a, 24); |
69 } | 87 } |
70 | 88 |
71 static inline __m128i SkGetPackedR32_SSE2(const __m128i& src) { | 89 static inline __m128i SkGetPackedR32_SSE2(const __m128i& src) { |
72 __m128i r = _mm_slli_epi32(src, (24 - SK_R32_SHIFT)); | 90 __m128i r = _mm_slli_epi32(src, (24 - SK_R32_SHIFT)); |
73 return _mm_srli_epi32(r, 24); | 91 return _mm_srli_epi32(r, 24); |
74 } | 92 } |
75 | 93 |
(...skipping 104 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
180 __m128i b = _mm_packs_epi32(b1, b2); | 198 __m128i b = _mm_packs_epi32(b1, b2); |
181 | 199 |
182 // Store 8 16-bit colors in dst. | 200 // Store 8 16-bit colors in dst. |
183 __m128i d_pixel = SkPackRGB16_SSE2(r, g, b); | 201 __m128i d_pixel = SkPackRGB16_SSE2(r, g, b); |
184 | 202 |
185 return d_pixel; | 203 return d_pixel; |
186 } | 204 } |
187 | 205 |
188 #undef ASSERT_EQ | 206 #undef ASSERT_EQ |
189 #endif // SkColor_opts_SSE2_DEFINED | 207 #endif // SkColor_opts_SSE2_DEFINED |
OLD | NEW |