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 16 bits
9// (fix8_16), then widen those to 8-bit-in-32-bits (fix8_32), and finally convert those to floats.
10
11// round() and roundClamp() do the opposite, working from floats to 8-bit-in-32-bit,
12// to 8-bit-in-16-bit, back down to 8-bit components.
13// _mm_packus_epi16() gives us clamping for free while narrowing.
14
15inline SkPMFloat::SkPMFloat(SkPMColor c) {
16    SkPMColorAssert(c);
17    __m128i fix8    = _mm_set_epi32(0,0,0,c),
18            fix8_16 = _mm_unpacklo_epi8 (fix8,    _mm_setzero_si128()),
19            fix8_32 = _mm_unpacklo_epi16(fix8_16, _mm_setzero_si128());
20    fVec = _mm_cvtepi32_ps(fix8_32);
21    SkASSERT(this->isValid());
22}
23
24inline SkPMColor SkPMFloat::round() const {
25    return this->roundClamp();  // Haven't beaten this yet.
26}
27
28inline SkPMColor SkPMFloat::roundClamp() const {
29    // We don't use _mm_cvtps_epi32, because we want precise control over how 0.5 rounds (up).
30    __m128i fix8_32 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), fVec)),
31            fix8_16 = _mm_packus_epi16(fix8_32, fix8_32),
32            fix8    = _mm_packus_epi16(fix8_16, fix8_16);
33    SkPMColor c = _mm_cvtsi128_si32(fix8);
34    SkPMColorAssert(c);
35    return c;
36}
37
38inline SkPMColor SkPMFloat::trunc() const {
39    // Basically, same as roundClamp(), but no rounding.
40    __m128i fix8_32 = _mm_cvttps_epi32(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.
61    RoundClampTo4PMColors(a,b,c,d, colors);
62}
63
64inline void SkPMFloat::RoundClampTo4PMColors(
65        const SkPMFloat& a, const SkPMFloat& b, const SkPMFloat&c, const SkPMFloat& d,
66        SkPMColor colors[4]) {
67    // Same as _SSSE3.h's.  We use 3 _mm_packus_epi16() where the naive loop uses 8.
68    // We don't use _mm_cvtps_epi32, because we want precise control over how 0.5 rounds (up).
69    __m128i c0 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), a.fVec)),
70            c1 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), b.fVec)),
71            c2 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), c.fVec)),
72            c3 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), d.fVec));
73    __m128i c3210 = _mm_packus_epi16(_mm_packus_epi16(c0, c1),
74                                     _mm_packus_epi16(c2, c3));
75    _mm_storeu_si128((__m128i*)colors, c3210);
76    SkPMColorAssert(colors[0]);
77    SkPMColorAssert(colors[1]);
78    SkPMColorAssert(colors[2]);
79    SkPMColorAssert(colors[3]);
80}
81