SkNx_sse.h revision fd5a26080d4647cc913f5f2b2dc72cb35abac3ab 
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_sse_DEFINED
9#define SkNx_sse_DEFINED
10
11// This file may assume <= SSE2, but must check SK_CPU_SSE_LEVEL for anything more recent.
12// If you do, make sure this is in a static inline function... anywhere else risks violating ODR.
13
14#define SKNX_IS_FAST
15
16// SSE 4.1 has _mm_floor_ps to floor 4 floats.  We emulate it:
17//   - roundtrip through integers via truncation
18//   - subtract 1 if that's too big (possible for negative values).
19// This restricts the domain of our inputs to a maximum somehwere around 2^31.  Seems plenty big.
20static inline __m128 sse2_mm_floor_ps(__m128 v) {
21    __m128 roundtrip = _mm_cvtepi32_ps(_mm_cvttps_epi32(v));
22    __m128 too_big = _mm_cmpgt_ps(roundtrip, v);
23    return _mm_sub_ps(roundtrip, _mm_and_ps(too_big, _mm_set1_ps(1.0f)));
24}
25
26template <>
27class SkNx<2, float> {
28public:
29    SkNx(const __m128& vec) : fVec(vec) {}
30
31    SkNx() {}
32    SkNx(float val) : fVec(_mm_set1_ps(val)) {}
33    static SkNx Load(const void* ptr) {
34        return _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*)ptr));
35    }
36    SkNx(float a, float b) : fVec(_mm_setr_ps(a,b,0,0)) {}
37
38    void store(void* ptr) const { _mm_storel_pi((__m64*)ptr, fVec); }
39
40    SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
41    SkNx operator - (const SkNx& o) const { return _mm_sub_ps(fVec, o.fVec); }
42    SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); }
43    SkNx operator / (const SkNx& o) const { return _mm_div_ps(fVec, o.fVec); }
44
45    SkNx operator == (const SkNx& o) const { return _mm_cmpeq_ps (fVec, o.fVec); }
46    SkNx operator != (const SkNx& o) const { return _mm_cmpneq_ps(fVec, o.fVec); }
47    SkNx operator  < (const SkNx& o) const { return _mm_cmplt_ps (fVec, o.fVec); }
48    SkNx operator  > (const SkNx& o) const { return _mm_cmpgt_ps (fVec, o.fVec); }
49    SkNx operator <= (const SkNx& o) const { return _mm_cmple_ps (fVec, o.fVec); }
50    SkNx operator >= (const SkNx& o) const { return _mm_cmpge_ps (fVec, o.fVec); }
51
52    static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_ps(l.fVec, r.fVec); }
53    static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r.fVec); }
54
55    SkNx  sqrt () const { return _mm_sqrt_ps (fVec);  }
56    SkNx rsqrt0() const { return _mm_rsqrt_ps(fVec); }
57    SkNx rsqrt1() const { return this->rsqrt0(); }
58    SkNx rsqrt2() const { return this->rsqrt1(); }
59
60    SkNx       invert() const { return SkNx(1) / *this; }
61    SkNx approxInvert() const { return _mm_rcp_ps(fVec); }
62
63    float operator[](int k) const {
64        SkASSERT(0 <= k && k < 2);
65        union { __m128 v; float fs[4]; } pun = {fVec};
66        return pun.fs[k&1];
67    }
68
69    bool allTrue() const { return 0xff == (_mm_movemask_epi8(_mm_castps_si128(fVec)) & 0xff); }
70    bool anyTrue() const { return 0x00 != (_mm_movemask_epi8(_mm_castps_si128(fVec)) & 0xff); }
71
72    __m128 fVec;
73};
74
75template <>
76class SkNx<4, float> {
77public:
78    SkNx(const __m128& vec) : fVec(vec) {}
79
80    SkNx() {}
81    SkNx(float val)           : fVec( _mm_set1_ps(val) ) {}
82    static SkNx Load(const void* ptr) { return _mm_loadu_ps((const float*)ptr); }
83
84    SkNx(float a, float b, float c, float d) : fVec(_mm_setr_ps(a,b,c,d)) {}
85
86    void store(void* ptr) const { _mm_storeu_ps((float*)ptr, fVec); }
87
88    SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
89    SkNx operator - (const SkNx& o) const { return _mm_sub_ps(fVec, o.fVec); }
90    SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); }
91    SkNx operator / (const SkNx& o) const { return _mm_div_ps(fVec, o.fVec); }
92
93    SkNx operator == (const SkNx& o) const { return _mm_cmpeq_ps (fVec, o.fVec); }
94    SkNx operator != (const SkNx& o) const { return _mm_cmpneq_ps(fVec, o.fVec); }
95    SkNx operator  < (const SkNx& o) const { return _mm_cmplt_ps (fVec, o.fVec); }
96    SkNx operator  > (const SkNx& o) const { return _mm_cmpgt_ps (fVec, o.fVec); }
97    SkNx operator <= (const SkNx& o) const { return _mm_cmple_ps (fVec, o.fVec); }
98    SkNx operator >= (const SkNx& o) const { return _mm_cmpge_ps (fVec, o.fVec); }
99
100    static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_ps(l.fVec, r.fVec); }
101    static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r.fVec); }
102
103    SkNx abs() const { return _mm_andnot_ps(_mm_set1_ps(-0.0f), fVec); }
104    SkNx floor() const { return sse2_mm_floor_ps(fVec); }
105
106    SkNx  sqrt () const { return _mm_sqrt_ps (fVec);  }
107    SkNx rsqrt0() const { return _mm_rsqrt_ps(fVec); }
108    SkNx rsqrt1() const { return this->rsqrt0(); }
109    SkNx rsqrt2() const { return this->rsqrt1(); }
110
111    SkNx       invert() const { return SkNx(1) / *this; }
112    SkNx approxInvert() const { return _mm_rcp_ps(fVec); }
113
114    float operator[](int k) const {
115        SkASSERT(0 <= k && k < 4);
116        union { __m128 v; float fs[4]; } pun = {fVec};
117        return pun.fs[k&3];
118    }
119
120    bool allTrue() const { return 0xffff == _mm_movemask_epi8(_mm_castps_si128(fVec)); }
121    bool anyTrue() const { return 0x0000 != _mm_movemask_epi8(_mm_castps_si128(fVec)); }
122
123    SkNx thenElse(const SkNx& t, const SkNx& e) const {
124        return _mm_or_ps(_mm_and_ps   (fVec, t.fVec),
125                         _mm_andnot_ps(fVec, e.fVec));
126    }
127
128    __m128 fVec;
129};
130
131template <>
132class SkNx<4, int> {
133public:
134    SkNx(const __m128i& vec) : fVec(vec) {}
135
136    SkNx() {}
137    SkNx(int val) : fVec(_mm_set1_epi32(val)) {}
138    static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
139    SkNx(int a, int b, int c, int d) : fVec(_mm_setr_epi32(a,b,c,d)) {}
140
141    void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
142
143    SkNx operator + (const SkNx& o) const { return _mm_add_epi32(fVec, o.fVec); }
144    SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); }
145    SkNx operator * (const SkNx& o) const {
146        __m128i mul20 = _mm_mul_epu32(fVec, o.fVec),
147                mul31 = _mm_mul_epu32(_mm_srli_si128(fVec, 4), _mm_srli_si128(o.fVec, 4));
148        return _mm_unpacklo_epi32(_mm_shuffle_epi32(mul20, _MM_SHUFFLE(0,0,2,0)),
149                                  _mm_shuffle_epi32(mul31, _MM_SHUFFLE(0,0,2,0)));
150    }
151
152    SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); }
153    SkNx operator >> (int bits) const { return _mm_srai_epi32(fVec, bits); }
154
155    int operator[](int k) const {
156        SkASSERT(0 <= k && k < 4);
157        union { __m128i v; int is[4]; } pun = {fVec};
158        return pun.is[k&3];
159    }
160
161    __m128i fVec;
162};
163
164template <>
165class SkNx<4, uint16_t> {
166public:
167    SkNx(const __m128i& vec) : fVec(vec) {}
168
169    SkNx() {}
170    SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
171    static SkNx Load(const void* ptr) { return _mm_loadl_epi64((const __m128i*)ptr); }
172    SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d) : fVec(_mm_setr_epi16(a,b,c,d,0,0,0,0)) {}
173
174    void store(void* ptr) const { _mm_storel_epi64((__m128i*)ptr, fVec); }
175
176    SkNx operator + (const SkNx& o) const { return _mm_add_epi16(fVec, o.fVec); }
177    SkNx operator - (const SkNx& o) const { return _mm_sub_epi16(fVec, o.fVec); }
178    SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec); }
179
180    SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); }
181    SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); }
182
183    uint16_t operator[](int k) const {
184        SkASSERT(0 <= k && k < 4);
185        union { __m128i v; uint16_t us[8]; } pun = {fVec};
186        return pun.us[k&3];
187    }
188
189    __m128i fVec;
190};
191
192template <>
193class SkNx<8, uint16_t> {
194public:
195    SkNx(const __m128i& vec) : fVec(vec) {}
196
197    SkNx() {}
198    SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
199    static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
200    SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d,
201         uint16_t e, uint16_t f, uint16_t g, uint16_t h) : fVec(_mm_setr_epi16(a,b,c,d,e,f,g,h)) {}
202
203    void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
204
205    SkNx operator + (const SkNx& o) const { return _mm_add_epi16(fVec, o.fVec); }
206    SkNx operator - (const SkNx& o) const { return _mm_sub_epi16(fVec, o.fVec); }
207    SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec); }
208
209    SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); }
210    SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); }
211
212    static SkNx Min(const SkNx& a, const SkNx& b) {
213        // No unsigned _mm_min_epu16, so we'll shift into a space where we can use the
214        // signed version, _mm_min_epi16, then shift back.
215        const uint16_t top = 0x8000; // Keep this separate from _mm_set1_epi16 or MSVC will whine.
216        const __m128i top_8x = _mm_set1_epi16(top);
217        return _mm_add_epi8(top_8x, _mm_min_epi16(_mm_sub_epi8(a.fVec, top_8x),
218                                                  _mm_sub_epi8(b.fVec, top_8x)));
219    }
220
221    SkNx thenElse(const SkNx& t, const SkNx& e) const {
222        return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
223                            _mm_andnot_si128(fVec, e.fVec));
224    }
225
226    uint16_t operator[](int k) const {
227        SkASSERT(0 <= k && k < 8);
228        union { __m128i v; uint16_t us[8]; } pun = {fVec};
229        return pun.us[k&7];
230    }
231
232    __m128i fVec;
233};
234
235template <>
236class SkNx<4, uint8_t> {
237public:
238    SkNx() {}
239    SkNx(const __m128i& vec) : fVec(vec) {}
240    SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d)
241        : fVec(_mm_setr_epi8(a,b,c,d, 0,0,0,0, 0,0,0,0, 0,0,0,0)) {}
242
243
244    static SkNx Load(const void* ptr) { return _mm_cvtsi32_si128(*(const int*)ptr); }
245    void store(void* ptr) const { *(int*)ptr = _mm_cvtsi128_si32(fVec); }
246
247    uint8_t operator[](int k) const {
248        SkASSERT(0 <= k && k < 4);
249        union { __m128i v; uint8_t us[16]; } pun = {fVec};
250        return pun.us[k&3];
251    }
252
253    // TODO as needed
254
255    __m128i fVec;
256};
257
258template <>
259class SkNx<16, uint8_t> {
260public:
261    SkNx(const __m128i& vec) : fVec(vec) {}
262
263    SkNx() {}
264    SkNx(uint8_t val) : fVec(_mm_set1_epi8(val)) {}
265    static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
266    SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d,
267         uint8_t e, uint8_t f, uint8_t g, uint8_t h,
268         uint8_t i, uint8_t j, uint8_t k, uint8_t l,
269         uint8_t m, uint8_t n, uint8_t o, uint8_t p)
270        : fVec(_mm_setr_epi8(a,b,c,d, e,f,g,h, i,j,k,l, m,n,o,p)) {}
271
272    void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
273
274    SkNx saturatedAdd(const SkNx& o) const { return _mm_adds_epu8(fVec, o.fVec); }
275
276    SkNx operator + (const SkNx& o) const { return _mm_add_epi8(fVec, o.fVec); }
277    SkNx operator - (const SkNx& o) const { return _mm_sub_epi8(fVec, o.fVec); }
278
279    static SkNx Min(const SkNx& a, const SkNx& b) { return _mm_min_epu8(a.fVec, b.fVec); }
280    SkNx operator < (const SkNx& o) const {
281        // There's no unsigned _mm_cmplt_epu8, so we flip the sign bits then use a signed compare.
282        auto flip = _mm_set1_epi8(char(0x80));
283        return _mm_cmplt_epi8(_mm_xor_si128(flip, fVec), _mm_xor_si128(flip, o.fVec));
284    }
285
286    uint8_t operator[](int k) const {
287        SkASSERT(0 <= k && k < 16);
288        union { __m128i v; uint8_t us[16]; } pun = {fVec};
289        return pun.us[k&15];
290    }
291
292    SkNx thenElse(const SkNx& t, const SkNx& e) const {
293        return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
294                            _mm_andnot_si128(fVec, e.fVec));
295    }
296
297    __m128i fVec;
298};
299
300template<> /*static*/ inline Sk4f SkNx_cast<float, int>(const Sk4i& src) {
301    return _mm_cvtepi32_ps(src.fVec);
302}
303
304template <> /*static*/ inline Sk4i SkNx_cast<int, float>(const Sk4f& src) {
305    return _mm_cvttps_epi32(src.fVec);
306}
307
308template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) {
309    auto _32 = _mm_cvttps_epi32(src.fVec);
310    // Ideally we'd use _mm_packus_epi32 here.  But that's SSE4.1+.
311#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
312    // With SSSE3, we can just shuffle the low 2 bytes from each lane right into place.
313    const int _ = ~0;
314    return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_,_,_));
315#else
316    // With SSE2, we have to emulate _mm_packus_epi32 with _mm_packs_epi32:
317    _32 = _mm_sub_epi32(_32, _mm_set1_epi32((int)0x00008000));
318    return _mm_add_epi16(_mm_packs_epi32(_32, _32), _mm_set1_epi16((short)0x8000));
319#endif
320}
321
322template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, float>(const Sk4f& src) {
323    auto _32 = _mm_cvttps_epi32(src.fVec);
324#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
325    const int _ = ~0;
326    return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,4,8,12, _,_,_,_, _,_,_,_, _,_,_,_));
327#else
328    auto _16 = _mm_packus_epi16(_32, _32);
329    return     _mm_packus_epi16(_16, _16);
330#endif
331}
332
333template<> /*static*/ inline Sk4f SkNx_cast<float, uint8_t>(const Sk4b& src) {
334#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
335    const int _ = ~0;
336    auto _32 = _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,_,_,_, 1,_,_,_, 2,_,_,_, 3,_,_,_));
337#else
338    auto _16 = _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()),
339         _32 = _mm_unpacklo_epi16(_16,     _mm_setzero_si128());
340#endif
341    return _mm_cvtepi32_ps(_32);
342}
343
344template<> /*static*/ inline Sk4f SkNx_cast<float, uint16_t>(const Sk4h& src) {
345    auto _32 = _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128());
346    return _mm_cvtepi32_ps(_32);
347}
348
349static inline void Sk4f_ToBytes(uint8_t bytes[16],
350                                const Sk4f& a, const Sk4f& b, const Sk4f& c, const Sk4f& d) {
351    _mm_storeu_si128((__m128i*)bytes,
352                     _mm_packus_epi16(_mm_packus_epi16(_mm_cvttps_epi32(a.fVec),
353                                                       _mm_cvttps_epi32(b.fVec)),
354                                      _mm_packus_epi16(_mm_cvttps_epi32(c.fVec),
355                                                       _mm_cvttps_epi32(d.fVec))));
356}
357
358template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src) {
359    return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128());
360}
361
362template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) {
363    return _mm_packus_epi16(src.fVec, src.fVec);
364}
365
366#endif//SkNx_sse_DEFINED
367