OLD | NEW |
| (Empty) |
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 | |
8 namespace { // See SkPMFloat.h | |
9 | |
10 // For SkPMFloat(SkPMColor), we widen our 8 bit components (fix8) to 8-bit compo
nents in 16 bits | |
11 // (fix8_16), then widen those to 8-bit-in-32-bits (fix8_32), and finally conver
t those to floats. | |
12 | |
13 // round() and roundClamp() do the opposite, working from floats to 8-bit-in-32-
bit, | |
14 // to 8-bit-in-16-bit, back down to 8-bit components. | |
15 // _mm_packus_epi16() gives us clamping for free while narrowing. | |
16 | |
17 inline SkPMFloat::SkPMFloat(SkPMColor c) { | |
18 SkPMColorAssert(c); | |
19 __m128i fix8 = _mm_set_epi32(0,0,0,c), | |
20 fix8_16 = _mm_unpacklo_epi8 (fix8, _mm_setzero_si128()), | |
21 fix8_32 = _mm_unpacklo_epi16(fix8_16, _mm_setzero_si128()); | |
22 fVec = _mm_cvtepi32_ps(fix8_32); | |
23 SkASSERT(this->isValid()); | |
24 } | |
25 | |
26 inline SkPMColor SkPMFloat::round() const { | |
27 return this->roundClamp(); // Haven't beaten this yet. | |
28 } | |
29 | |
30 inline SkPMColor SkPMFloat::roundClamp() const { | |
31 // We don't use _mm_cvtps_epi32, because we want precise control over how 0.
5 rounds (up). | |
32 __m128i fix8_32 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), fVec)), | |
33 fix8_16 = _mm_packus_epi16(fix8_32, fix8_32), | |
34 fix8 = _mm_packus_epi16(fix8_16, fix8_16); | |
35 SkPMColor c = _mm_cvtsi128_si32(fix8); | |
36 SkPMColorAssert(c); | |
37 return c; | |
38 } | |
39 | |
40 inline SkPMColor SkPMFloat::trunc() const { | |
41 // Basically, same as roundClamp(), but no rounding. | |
42 __m128i fix8_32 = _mm_cvttps_epi32(fVec), | |
43 fix8_16 = _mm_packus_epi16(fix8_32, fix8_32), | |
44 fix8 = _mm_packus_epi16(fix8_16, fix8_16); | |
45 SkPMColor c = _mm_cvtsi128_si32(fix8); | |
46 SkPMColorAssert(c); | |
47 return c; | |
48 } | |
49 | |
50 inline void SkPMFloat::From4PMColors(const SkPMColor colors[4], | |
51 SkPMFloat* a, SkPMFloat* b, SkPMFloat* c, S
kPMFloat* d) { | |
52 // Haven't beaten this yet. | |
53 *a = FromPMColor(colors[0]); | |
54 *b = FromPMColor(colors[1]); | |
55 *c = FromPMColor(colors[2]); | |
56 *d = FromPMColor(colors[3]); | |
57 } | |
58 | |
59 inline void SkPMFloat::RoundTo4PMColors( | |
60 const SkPMFloat& a, const SkPMFloat& b, const SkPMFloat&c, const SkPMFlo
at& d, | |
61 SkPMColor colors[4]) { | |
62 // Haven't beaten this yet. | |
63 RoundClampTo4PMColors(a,b,c,d, colors); | |
64 } | |
65 | |
66 inline void SkPMFloat::RoundClampTo4PMColors( | |
67 const SkPMFloat& a, const SkPMFloat& b, const SkPMFloat&c, const SkPMFlo
at& d, | |
68 SkPMColor colors[4]) { | |
69 // Same as _SSSE3.h's. We use 3 _mm_packus_epi16() where the naive loop use
s 8. | |
70 // We don't use _mm_cvtps_epi32, because we want precise control over how 0.
5 rounds (up). | |
71 __m128i c0 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), a.fVec)), | |
72 c1 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), b.fVec)), | |
73 c2 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), c.fVec)), | |
74 c3 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), d.fVec)); | |
75 __m128i c3210 = _mm_packus_epi16(_mm_packus_epi16(c0, c1), | |
76 _mm_packus_epi16(c2, c3)); | |
77 _mm_storeu_si128((__m128i*)colors, c3210); | |
78 SkPMColorAssert(colors[0]); | |
79 SkPMColorAssert(colors[1]); | |
80 SkPMColorAssert(colors[2]); | |
81 SkPMColorAssert(colors[3]); | |
82 } | |
83 | |
84 } // namespace | |
OLD | NEW |