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 #ifndef SkNx_avx_DEFINED | |
9 #define SkNx_avx_DEFINED | |
10 | |
11 // This file may assume <= AVX, but must check SK_CPU_SSE_LEVEL for anything mor e recent. | |
12 | |
13 // All the SSE specializations are still good ideas. We'll just add Sk8f. | |
14 #include "SkNx_sse.h" | |
15 | |
16 namespace { // See SkNx.h | |
17 | |
18 template <> | |
19 class SkNf<8> { | |
20 public: | |
21 SkNf(const __m256& vec) : fVec(vec) {} | |
22 | |
23 SkNf() {} | |
24 SkNf(float val) : fVec(_mm256_set1_ps(val)) {} | |
25 static SkNf Load(const float vals[8]) { return _mm256_loadu_ps(vals); } | |
26 | |
27 static SkNf FromBytes(const uint8_t bytes[8]) { | |
28 __m128i fix8 = _mm_loadl_epi64((const __m128i*)bytes), | |
29 fix16 = _mm_unpacklo_epi8 (fix8 , _mm_setzero_si128()), | |
30 lo32 = _mm_unpacklo_epi16(fix16, _mm_setzero_si128()), | |
31 hi32 = _mm_unpackhi_epi16(fix16, _mm_setzero_si128()); | |
32 __m256i fix32 = _mm256_insertf128_si256(_mm256_castsi128_si256(lo32), hi 32, 1); | |
msarett
2015/11/09 23:25:06
Seems annoying to do all the unpacking in 128-bit
mtklein
2015/11/10 00:22:05
Right. Even then, the AVX and AVX2 packing, unpac
msarett
2015/11/10 14:54:16
Good point, forgot about that.
| |
33 return _mm256_cvtepi32_ps(fix32); | |
34 } | |
35 | |
36 SkNf(float a, float b, float c, float d, | |
37 float e, float f, float g, float h) : fVec(_mm256_setr_ps(a,b,c,d,e,f,g ,h)) {} | |
38 | |
39 void store(float vals[8]) const { _mm256_storeu_ps(vals, fVec); } | |
40 void toBytes(uint8_t bytes[8]) const { | |
41 __m256i fix32 = _mm256_cvttps_epi32(fVec); | |
42 __m128i lo32 = _mm256_extractf128_si256(fix32, 0), | |
43 hi32 = _mm256_extractf128_si256(fix32, 1), | |
44 fix16 = _mm_packus_epi32(lo32, hi32), | |
45 fix8 = _mm_packus_epi16(fix16, fix16); | |
46 _mm_storel_epi64((__m128i*)bytes, fix8); | |
47 } | |
48 | |
49 SkNf operator + (const SkNf& o) const { return _mm256_add_ps(fVec, o.fVec); } | |
50 SkNf operator - (const SkNf& o) const { return _mm256_sub_ps(fVec, o.fVec); } | |
51 SkNf operator * (const SkNf& o) const { return _mm256_mul_ps(fVec, o.fVec); } | |
52 SkNf operator / (const SkNf& o) const { return _mm256_div_ps(fVec, o.fVec); } | |
53 | |
54 SkNf operator == (const SkNf& o) const { return _mm256_cmp_ps(fVec, o.fVec, _CMP_EQ_OQ); } | |
55 SkNf operator != (const SkNf& o) const { return _mm256_cmp_ps(fVec, o.fVec, _CMP_NEQ_OQ); } | |
56 SkNf operator < (const SkNf& o) const { return _mm256_cmp_ps(fVec, o.fVec, _CMP_LT_OQ); } | |
57 SkNf operator > (const SkNf& o) const { return _mm256_cmp_ps(fVec, o.fVec, _CMP_GT_OQ); } | |
58 SkNf operator <= (const SkNf& o) const { return _mm256_cmp_ps(fVec, o.fVec, _CMP_LE_OQ); } | |
59 SkNf operator >= (const SkNf& o) const { return _mm256_cmp_ps(fVec, o.fVec, _CMP_GE_OQ); } | |
60 | |
61 static SkNf Min(const SkNf& l, const SkNf& r) { return _mm256_min_ps(l.fVec, r.fVec); } | |
62 static SkNf Max(const SkNf& l, const SkNf& r) { return _mm256_max_ps(l.fVec, r.fVec); } | |
63 | |
64 SkNf sqrt() const { return _mm256_sqrt_ps (fVec); } | |
65 SkNf rsqrt0() const { return _mm256_rsqrt_ps(fVec); } | |
66 SkNf rsqrt1() const { return this->rsqrt0(); } | |
67 SkNf rsqrt2() const { return this->rsqrt1(); } | |
68 | |
69 SkNf invert() const { return SkNf(1) / *this; } | |
70 SkNf approxInvert() const { return _mm256_rcp_ps(fVec); } | |
71 | |
72 template <int k> float kth() const { | |
73 SkASSERT(0 <= k && k < 8); | |
74 union { __m256 v; float fs[8]; } pun = {fVec}; | |
75 return pun.fs[k&7]; | |
76 } | |
77 | |
78 bool allTrue() const { return 0xff == _mm256_movemask_ps(fVec); } | |
79 bool anyTrue() const { return 0x00 != _mm256_movemask_ps(fVec); } | |
80 | |
81 SkNf thenElse(const SkNf& t, const SkNf& e) const { | |
82 return _mm256_blendv_ps(e.fVec, t.fVec, fVec); | |
83 } | |
84 | |
85 __m256 fVec; | |
86 }; | |
87 | |
88 } // namespace | |
89 | |
90 #endif//SkNx_avx_DEFINED | |
OLD | NEW |