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
8namespace { // See Sk4px.h
9
10inline Sk4px Sk4px::DupPMColor(SkPMColor px) { return Sk16b(_mm_set1_epi32(px)); }
11
12inline Sk4px Sk4px::Load4(const SkPMColor px[4]) {
13    return Sk16b(_mm_loadu_si128((const __m128i*)px));
14}
15inline Sk4px Sk4px::Load2(const SkPMColor px[2]) {
16    return Sk16b(_mm_loadl_epi64((const __m128i*)px));
17}
18inline Sk4px Sk4px::Load1(const SkPMColor px[1]) { return Sk16b(_mm_cvtsi32_si128(*px)); }
19
20inline void Sk4px::store4(SkPMColor px[4]) const { _mm_storeu_si128((__m128i*)px, this->fVec); }
21inline void Sk4px::store2(SkPMColor px[2]) const { _mm_storel_epi64((__m128i*)px, this->fVec); }
22inline void Sk4px::store1(SkPMColor px[1]) const { *px = _mm_cvtsi128_si32(this->fVec); }
23
24inline Sk4px::Wide Sk4px::widenLo() const {
25    return Sk16h(_mm_unpacklo_epi8(this->fVec, _mm_setzero_si128()),
26                 _mm_unpackhi_epi8(this->fVec, _mm_setzero_si128()));
27}
28
29inline Sk4px::Wide Sk4px::widenHi() const {
30    return Sk16h(_mm_unpacklo_epi8(_mm_setzero_si128(), this->fVec),
31                 _mm_unpackhi_epi8(_mm_setzero_si128(), this->fVec));
32}
33
34inline Sk4px::Wide Sk4px::widenLoHi() const {
35    return Sk16h(_mm_unpacklo_epi8(this->fVec, this->fVec),
36                 _mm_unpackhi_epi8(this->fVec, this->fVec));
37}
38
39inline Sk4px::Wide Sk4px::mulWiden(const Sk16b& other) const {
40    return this->widenLo() * Sk4px(other).widenLo();
41}
42
43inline Sk4px Sk4px::Wide::addNarrowHi(const Sk16h& other) const {
44    Sk4px::Wide r = (*this + other) >> 8;
45    return Sk4px(_mm_packus_epi16(r.fLo.fVec, r.fHi.fVec));
46}
47
48inline Sk4px Sk4px::Wide::div255() const {
49    // (x + 127) / 255 == ((x+128) * 257)>>16,
50    // and _mm_mulhi_epu16 makes the (_ * 257)>>16 part very convenient.
51    const __m128i _128 = _mm_set1_epi16(128),
52                  _257 = _mm_set1_epi16(257);
53    return Sk4px(_mm_packus_epi16(_mm_mulhi_epu16(_mm_add_epi16(fLo.fVec, _128), _257),
54                                  _mm_mulhi_epu16(_mm_add_epi16(fHi.fVec, _128), _257)));
55}
56
57// Load4Alphas and Load2Alphas use possibly-unaligned loads (SkAlpha[] -> uint16_t or uint32_t).
58// These are safe on x86, often with no speed penalty.
59
60#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
61    inline Sk4px Sk4px::alphas() const {
62        static_assert(SK_A32_SHIFT == 24, "Intel's always little-endian.");
63        __m128i splat = _mm_set_epi8(15,15,15,15, 11,11,11,11, 7,7,7,7, 3,3,3,3);
64        return Sk16b(_mm_shuffle_epi8(this->fVec, splat));
65    }
66
67    inline Sk4px Sk4px::Load4Alphas(const SkAlpha a[4]) {
68        uint32_t as = *(const uint32_t*)a;
69        __m128i splat = _mm_set_epi8(3,3,3,3, 2,2,2,2, 1,1,1,1, 0,0,0,0);
70        return Sk16b(_mm_shuffle_epi8(_mm_cvtsi32_si128(as), splat));
71    }
72#else
73    inline Sk4px Sk4px::alphas() const {
74        static_assert(SK_A32_SHIFT == 24, "Intel's always little-endian.");
75        // We exploit that A >= rgb for any premul pixel.
76        __m128i as = fVec;                             // 3xxx 2xxx 1xxx 0xxx
77        as = _mm_max_epu8(as, _mm_srli_epi32(as,  8)); // 33xx 22xx 11xx 00xx
78        as = _mm_max_epu8(as, _mm_srli_epi32(as, 16)); // 3333 2222 1111 0000
79        return Sk16b(as);
80    }
81
82    inline Sk4px Sk4px::Load4Alphas(const SkAlpha a[4]) {
83        __m128i as = _mm_cvtsi32_si128(*(const uint32_t*)a);  // ____ ____ ____ 3210
84        as = _mm_unpacklo_epi8 (as, as);                      // ____ ____ 3322 1100
85        as = _mm_unpacklo_epi16(as, as);                      // 3333 2222 1111 0000
86        return Sk16b(as);
87    }
88#endif
89
90inline Sk4px Sk4px::Load2Alphas(const SkAlpha a[2]) {
91    uint32_t as = *(const uint16_t*)a;   // Aa -> Aa00
92    return Load4Alphas((const SkAlpha*)&as);
93}
94
95inline Sk4px Sk4px::zeroColors() const {
96    return Sk16b(_mm_and_si128(_mm_set1_epi32(0xFF << SK_A32_SHIFT), this->fVec));
97}
98
99inline Sk4px Sk4px::zeroAlphas() const {
100    // andnot(a,b) == ~a & b
101    return Sk16b(_mm_andnot_si128(_mm_set1_epi32(0xFF << SK_A32_SHIFT), this->fVec));
102}
103
104}  // namespace
105