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// For SkPMFloat(SkPMColor), we widen our 8 bit components (fix8) to 8-bit components in 32 bits
9// (fix8_32), then convert those to floats.
10
11// round() does the opposite, working from floats to 8-bit-in-32-bits, then back to packed 8 bit.
12
13// roundClamp() is the same as _SSE2: floats to 8-in-32, to 8-in-16, to packed 8 bit, with
14// _mm_packus_epi16() both clamping and narrowing.
15
16inline SkPMFloat::SkPMFloat(SkPMColor c) {
17    SkPMColorAssert(c);
18    const int _ = 255;  // _ means to zero that byte.
19    __m128i fix8    = _mm_set_epi32(0,0,0,c),
20            fix8_32 = _mm_shuffle_epi8(fix8, _mm_set_epi8(_,_,_,3, _,_,_,2, _,_,_,1, _,_,_,0));
21    fVec = _mm_cvtepi32_ps(fix8_32);
22    SkASSERT(this->isValid());
23}
24
25inline SkPMColor SkPMFloat::trunc() const {
26    const int _ = 255;  // _ means to zero that byte.
27    __m128i fix8_32 = _mm_cvttps_epi32(fVec),
28            fix8    = _mm_shuffle_epi8(fix8_32, _mm_set_epi8(_,_,_,_, _,_,_,_, _,_,_,_, 12,8,4,0));
29    SkPMColor c = _mm_cvtsi128_si32(fix8);
30    SkPMColorAssert(c);
31    return c;
32}
33
34inline SkPMColor SkPMFloat::round() const {
35    return SkPMFloat(Sk4f(0.5f) + *this).trunc();
36}
37
38inline SkPMColor SkPMFloat::roundClamp() const {
39    // We don't use _mm_cvtps_epi32, because we want precise control over how 0.5 rounds (up).
40    __m128i fix8_32 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), fVec)),
41            fix8_16 = _mm_packus_epi16(fix8_32, fix8_32),
42            fix8    = _mm_packus_epi16(fix8_16, fix8_16);
43    SkPMColor c = _mm_cvtsi128_si32(fix8);
44    SkPMColorAssert(c);
45    return c;
46}
47
48inline void SkPMFloat::From4PMColors(const SkPMColor colors[4],
49                                     SkPMFloat* a, SkPMFloat* b, SkPMFloat* c, SkPMFloat* d) {
50    // Haven't beaten this yet.
51    *a = FromPMColor(colors[0]);
52    *b = FromPMColor(colors[1]);
53    *c = FromPMColor(colors[2]);
54    *d = FromPMColor(colors[3]);
55}
56
57inline void SkPMFloat::RoundTo4PMColors(
58        const SkPMFloat& a, const SkPMFloat& b, const SkPMFloat&c, const SkPMFloat& d,
59        SkPMColor colors[4]) {
60    // Haven't beaten this yet.  Still faster than RoundClampTo4PMColors?
61    colors[0] = a.round();
62    colors[1] = b.round();
63    colors[2] = c.round();
64    colors[3] = d.round();
65}
66
67inline void SkPMFloat::RoundClampTo4PMColors(
68        const SkPMFloat& a, const SkPMFloat& b, const SkPMFloat&c, const SkPMFloat& d,
69        SkPMColor colors[4]) {
70    // Same as _SSE2.h's.  We use 3 _mm_packus_epi16() where the naive loop uses 8.
71    // We don't use _mm_cvtps_epi32, because we want precise control over how 0.5 rounds (up).
72    __m128i c0 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), a.fVec)),
73            c1 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), b.fVec)),
74            c2 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), c.fVec)),
75            c3 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), d.fVec));
76    __m128i c3210 = _mm_packus_epi16(_mm_packus_epi16(c0, c1),
77                                     _mm_packus_epi16(c2, c3));
78    _mm_storeu_si128((__m128i*)colors, c3210);
79    SkPMColorAssert(colors[0]);
80    SkPMColorAssert(colors[1]);
81    SkPMColorAssert(colors[2]);
82    SkPMColorAssert(colors[3]);
83}
84