OLD | NEW |
1 inline SkPMFloat& SkPMFloat::operator=(const SkPMFloat& that) { | 1 inline SkPMFloat& SkPMFloat::operator=(const SkPMFloat& that) { |
2 fColors = that.fColors; | 2 fColors = that.fColors; |
3 return *this; | 3 return *this; |
4 } | 4 } |
5 | 5 |
6 // For SkPMFloat(SkPMColor), we widen our 8 bit components (fix8) to 8-bit compo
nents in 16 bits | 6 // For SkPMFloat(SkPMColor), we widen our 8 bit components (fix8) to 8-bit compo
nents in 16 bits |
7 // (fix8_16), then widen those to 8-bit-in-32-bits (fix8_32), and finally conver
t those to floats. | 7 // (fix8_16), then widen those to 8-bit-in-32-bits (fix8_32), and finally conver
t those to floats. |
8 | 8 |
9 // get() and clamped() do the opposite, working from floats to 8-bit-in-32-bit, | 9 // get() and clamped() do the opposite, working from floats to 8-bit-in-32-bit, |
10 // to 8-bit-in-16-bit, back down to 8-bit components. | 10 // to 8-bit-in-16-bit, back down to 8-bit components. |
11 // _mm_packus_epi16() gives us clamping for free while narrowing. | 11 // _mm_packus_epi16() gives us clamping for free while narrowing. |
12 | 12 |
13 inline SkPMFloat::SkPMFloat(SkPMColor c) { | 13 inline SkPMFloat::SkPMFloat(SkPMColor c) { |
14 SkPMColorAssert(c); | 14 SkPMColorAssert(c); |
15 __m128i fix8 = _mm_set_epi32(0,0,0,c), | 15 __m128i fix8 = _mm_set_epi32(0,0,0,c), |
16 fix8_16 = _mm_unpacklo_epi8 (fix8, _mm_setzero_si128()), | 16 fix8_16 = _mm_unpacklo_epi8 (fix8, _mm_setzero_si128()), |
17 fix8_32 = _mm_unpacklo_epi16(fix8_16, _mm_setzero_si128()); | 17 fix8_32 = _mm_unpacklo_epi16(fix8_16, _mm_setzero_si128()); |
18 fColors = _mm_cvtepi32_ps(fix8_32); | 18 fColors = _mm_cvtepi32_ps(fix8_32); |
19 SkASSERT(this->isValid()); | 19 SkASSERT(this->isValid()); |
20 } | 20 } |
21 | 21 |
22 inline SkPMColor SkPMFloat::get() const { | 22 inline SkPMColor SkPMFloat::get() const { |
23 SkASSERT(this->isValid()); | 23 SkASSERT(this->isValid()); |
24 return this->clamped(); // Haven't beaten this yet. | 24 return this->clamped(); // Haven't beaten this yet. |
25 } | 25 } |
26 | 26 |
27 inline SkPMColor SkPMFloat::clamped() const { | 27 inline SkPMColor SkPMFloat::clamped() const { |
28 __m128i fix8_32 = _mm_cvtps_epi32(fColors), // _mm_cvtps_epi32 rounds for u
s! | 28 // We don't use _mm_cvtps_epi32, because we want precise control over how 0.
5 rounds (up). |
| 29 __m128i fix8_32 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), fColors)), |
29 fix8_16 = _mm_packus_epi16(fix8_32, fix8_32), | 30 fix8_16 = _mm_packus_epi16(fix8_32, fix8_32), |
30 fix8 = _mm_packus_epi16(fix8_16, fix8_16); | 31 fix8 = _mm_packus_epi16(fix8_16, fix8_16); |
31 SkPMColor c = _mm_cvtsi128_si32(fix8); | 32 SkPMColor c = _mm_cvtsi128_si32(fix8); |
32 SkPMColorAssert(c); | 33 SkPMColorAssert(c); |
33 return c; | 34 return c; |
34 } | 35 } |
35 | 36 |
36 inline void SkPMFloat::From4PMColors(SkPMFloat floats[4], const SkPMColor colors
[4]) { | 37 inline void SkPMFloat::From4PMColors(SkPMFloat floats[4], const SkPMColor colors
[4]) { |
37 // Haven't beaten this yet. | 38 // Haven't beaten this yet. |
38 for (int i = 0; i < 4; i++) { floats[i] = FromPMColor(colors[i]); } | 39 for (int i = 0; i < 4; i++) { floats[i] = FromPMColor(colors[i]); } |
39 } | 40 } |
40 | 41 |
41 inline void SkPMFloat::To4PMColors(SkPMColor colors[4], const SkPMFloat floats[4
]) { | 42 inline void SkPMFloat::To4PMColors(SkPMColor colors[4], const SkPMFloat floats[4
]) { |
42 SkASSERT(floats[0].isValid() && floats[1].isValid() | 43 SkASSERT(floats[0].isValid() && floats[1].isValid() |
43 && floats[2].isValid() && floats[3].isValid()); | 44 && floats[2].isValid() && floats[3].isValid()); |
44 // Haven't beaten this yet. | 45 // Haven't beaten this yet. |
45 ClampTo4PMColors(colors, floats); | 46 ClampTo4PMColors(colors, floats); |
46 } | 47 } |
47 | 48 |
48 inline void SkPMFloat::ClampTo4PMColors(SkPMColor colors[4], const SkPMFloat flo
ats[4]) { | 49 inline void SkPMFloat::ClampTo4PMColors(SkPMColor colors[4], const SkPMFloat flo
ats[4]) { |
49 // Same as _SSSE3.h's. We use 3 _mm_packus_epi16() where the naive loop use
s 8. | 50 // Same as _SSSE3.h's. We use 3 _mm_packus_epi16() where the naive loop use
s 8. |
50 __m128i c0 = _mm_cvtps_epi32(floats[0].fColors), // _mm_cvtps_epi32 rounds
for us! | 51 // We don't use _mm_cvtps_epi32, because we want precise control over how 0.
5 rounds (up). |
51 c1 = _mm_cvtps_epi32(floats[1].fColors), | 52 __m128i c0 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), floats[0].fColor
s)), |
52 c2 = _mm_cvtps_epi32(floats[2].fColors), | 53 c1 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), floats[1].fColor
s)), |
53 c3 = _mm_cvtps_epi32(floats[3].fColors); | 54 c2 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), floats[2].fColor
s)), |
| 55 c3 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), floats[3].fColor
s)); |
54 __m128i c3210 = _mm_packus_epi16(_mm_packus_epi16(c0, c1), | 56 __m128i c3210 = _mm_packus_epi16(_mm_packus_epi16(c0, c1), |
55 _mm_packus_epi16(c2, c3)); | 57 _mm_packus_epi16(c2, c3)); |
56 _mm_storeu_si128((__m128i*)colors, c3210); | 58 _mm_storeu_si128((__m128i*)colors, c3210); |
57 SkPMColorAssert(colors[0]); | 59 SkPMColorAssert(colors[0]); |
58 SkPMColorAssert(colors[1]); | 60 SkPMColorAssert(colors[1]); |
59 SkPMColorAssert(colors[2]); | 61 SkPMColorAssert(colors[2]); |
60 SkPMColorAssert(colors[3]); | 62 SkPMColorAssert(colors[3]); |
61 } | 63 } |
OLD | NEW |