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#include <immintrin.h>
13
14template <>
15class SkNb<2, 4> {
16public:
17    SkNb(const __m128i& vec) : fVec(vec) {}
18
19    SkNb() {}
20    bool allTrue() const { return 0xff == (_mm_movemask_epi8(fVec) & 0xff); }
21    bool anyTrue() const { return 0x00 != (_mm_movemask_epi8(fVec) & 0xff); }
22
23    __m128i fVec;
24};
25
26template <>
27class SkNb<4, 4> {
28public:
29    SkNb(const __m128i& vec) : fVec(vec) {}
30
31    SkNb() {}
32    bool allTrue() const { return 0xffff == _mm_movemask_epi8(fVec); }
33    bool anyTrue() const { return 0x0000 != _mm_movemask_epi8(fVec); }
34
35    __m128i fVec;
36};
37
38template <>
39class SkNb<2, 8> {
40public:
41    SkNb(const __m128i& vec) : fVec(vec) {}
42
43    SkNb() {}
44    bool allTrue() const { return 0xffff == _mm_movemask_epi8(fVec); }
45    bool anyTrue() const { return 0x0000 != _mm_movemask_epi8(fVec); }
46
47    __m128i fVec;
48};
49
50
51template <>
52class SkNf<2, float> {
53    typedef SkNb<2, 4> Nb;
54public:
55    SkNf(const __m128& vec) : fVec(vec) {}
56
57    SkNf() {}
58    explicit SkNf(float val) : fVec(_mm_set1_ps(val)) {}
59    static SkNf Load(const float vals[2]) {
60        return _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*)vals));
61    }
62    SkNf(float a, float b) : fVec(_mm_setr_ps(a,b,0,0)) {}
63
64    void store(float vals[2]) const { _mm_storel_pi((__m64*)vals, fVec); }
65
66    SkNf operator + (const SkNf& o) const { return _mm_add_ps(fVec, o.fVec); }
67    SkNf operator - (const SkNf& o) const { return _mm_sub_ps(fVec, o.fVec); }
68    SkNf operator * (const SkNf& o) const { return _mm_mul_ps(fVec, o.fVec); }
69    SkNf operator / (const SkNf& o) const { return _mm_div_ps(fVec, o.fVec); }
70
71    Nb operator == (const SkNf& o) const { return _mm_castps_si128(_mm_cmpeq_ps (fVec, o.fVec)); }
72    Nb operator != (const SkNf& o) const { return _mm_castps_si128(_mm_cmpneq_ps(fVec, o.fVec)); }
73    Nb operator  < (const SkNf& o) const { return _mm_castps_si128(_mm_cmplt_ps (fVec, o.fVec)); }
74    Nb operator  > (const SkNf& o) const { return _mm_castps_si128(_mm_cmpgt_ps (fVec, o.fVec)); }
75    Nb operator <= (const SkNf& o) const { return _mm_castps_si128(_mm_cmple_ps (fVec, o.fVec)); }
76    Nb operator >= (const SkNf& o) const { return _mm_castps_si128(_mm_cmpge_ps (fVec, o.fVec)); }
77
78    static SkNf Min(const SkNf& l, const SkNf& r) { return _mm_min_ps(l.fVec, r.fVec); }
79    static SkNf Max(const SkNf& l, const SkNf& r) { return _mm_max_ps(l.fVec, r.fVec); }
80
81    SkNf  sqrt() const { return _mm_sqrt_ps (fVec);  }
82    SkNf rsqrt0() const { return _mm_rsqrt_ps(fVec); }
83    SkNf rsqrt1() const { return this->rsqrt0(); }
84    SkNf rsqrt2() const { return this->rsqrt1(); }
85
86    SkNf       invert() const { return SkNf(1) / *this; }
87    SkNf approxInvert() const { return _mm_rcp_ps(fVec); }
88
89    template <int k> float kth() const {
90        SkASSERT(0 <= k && k < 2);
91        union { __m128 v; float fs[4]; } pun = {fVec};
92        return pun.fs[k&1];
93    }
94
95    __m128 fVec;
96};
97
98template <>
99class SkNf<2, double> {
100    typedef SkNb<2, 8> Nb;
101public:
102    SkNf(const __m128d& vec) : fVec(vec) {}
103
104    SkNf() {}
105    explicit SkNf(double val)           : fVec( _mm_set1_pd(val) ) {}
106    static SkNf Load(const double vals[2]) { return _mm_loadu_pd(vals); }
107    SkNf(double a, double b) : fVec(_mm_setr_pd(a,b)) {}
108
109    void store(double vals[2]) const { _mm_storeu_pd(vals, fVec); }
110
111    SkNf operator + (const SkNf& o) const { return _mm_add_pd(fVec, o.fVec); }
112    SkNf operator - (const SkNf& o) const { return _mm_sub_pd(fVec, o.fVec); }
113    SkNf operator * (const SkNf& o) const { return _mm_mul_pd(fVec, o.fVec); }
114    SkNf operator / (const SkNf& o) const { return _mm_div_pd(fVec, o.fVec); }
115
116    Nb operator == (const SkNf& o) const { return _mm_castpd_si128(_mm_cmpeq_pd (fVec, o.fVec)); }
117    Nb operator != (const SkNf& o) const { return _mm_castpd_si128(_mm_cmpneq_pd(fVec, o.fVec)); }
118    Nb operator  < (const SkNf& o) const { return _mm_castpd_si128(_mm_cmplt_pd (fVec, o.fVec)); }
119    Nb operator  > (const SkNf& o) const { return _mm_castpd_si128(_mm_cmpgt_pd (fVec, o.fVec)); }
120    Nb operator <= (const SkNf& o) const { return _mm_castpd_si128(_mm_cmple_pd (fVec, o.fVec)); }
121    Nb operator >= (const SkNf& o) const { return _mm_castpd_si128(_mm_cmpge_pd (fVec, o.fVec)); }
122
123    static SkNf Min(const SkNf& l, const SkNf& r) { return _mm_min_pd(l.fVec, r.fVec); }
124    static SkNf Max(const SkNf& l, const SkNf& r) { return _mm_max_pd(l.fVec, r.fVec); }
125
126    SkNf  sqrt() const { return _mm_sqrt_pd(fVec);  }
127    SkNf rsqrt0() const { return _mm_cvtps_pd(_mm_rsqrt_ps(_mm_cvtpd_ps(fVec))); }
128    SkNf rsqrt1() const { return this->rsqrt0(); }
129    SkNf rsqrt2() const { return this->rsqrt1(); }
130
131    SkNf       invert() const { return SkNf(1) / *this; }
132    SkNf approxInvert() const { return _mm_cvtps_pd(_mm_rcp_ps(_mm_cvtpd_ps(fVec))); }
133
134    template <int k> double kth() const {
135        SkASSERT(0 <= k && k < 2);
136        union { __m128d v; double ds[2]; } pun = {fVec};
137        return pun.ds[k&1];
138    }
139
140    __m128d fVec;
141};
142
143template <>
144class SkNi<4, int> {
145public:
146    SkNi(const __m128i& vec) : fVec(vec) {}
147
148    SkNi() {}
149    explicit SkNi(int val) : fVec(_mm_set1_epi32(val)) {}
150    static SkNi Load(const int vals[4]) { return _mm_loadu_si128((const __m128i*)vals); }
151    SkNi(int a, int b, int c, int d) : fVec(_mm_setr_epi32(a,b,c,d)) {}
152
153    void store(int vals[4]) const { _mm_storeu_si128((__m128i*)vals, fVec); }
154
155    SkNi operator + (const SkNi& o) const { return _mm_add_epi32(fVec, o.fVec); }
156    SkNi operator - (const SkNi& o) const { return _mm_sub_epi32(fVec, o.fVec); }
157    SkNi operator * (const SkNi& o) const {
158        __m128i mul20 = _mm_mul_epu32(fVec, o.fVec),
159                mul31 = _mm_mul_epu32(_mm_srli_si128(fVec, 4), _mm_srli_si128(o.fVec, 4));
160        return _mm_unpacklo_epi32(_mm_shuffle_epi32(mul20, _MM_SHUFFLE(0,0,2,0)),
161                                  _mm_shuffle_epi32(mul31, _MM_SHUFFLE(0,0,2,0)));
162    }
163
164    SkNi operator << (int bits) const { return _mm_slli_epi32(fVec, bits); }
165    SkNi operator >> (int bits) const { return _mm_srai_epi32(fVec, bits); }
166
167    template <int k> int kth() const {
168        SkASSERT(0 <= k && k < 4);
169        switch (k) {
170            case 0: return _mm_cvtsi128_si32(fVec);
171            case 1: return _mm_cvtsi128_si32(_mm_srli_si128(fVec,  4));
172            case 2: return _mm_cvtsi128_si32(_mm_srli_si128(fVec,  8));
173            case 3: return _mm_cvtsi128_si32(_mm_srli_si128(fVec, 12));
174            default: SkASSERT(false); return 0;
175        }
176    }
177
178    __m128i fVec;
179};
180
181template <>
182class SkNf<4, float> {
183    typedef SkNb<4, 4> Nb;
184public:
185    SkNf(const __m128& vec) : fVec(vec) {}
186
187    SkNf() {}
188    explicit SkNf(float val)           : fVec( _mm_set1_ps(val) ) {}
189    static SkNf Load(const float vals[4]) { return _mm_loadu_ps(vals); }
190    SkNf(float a, float b, float c, float d) : fVec(_mm_setr_ps(a,b,c,d)) {}
191
192    void store(float vals[4]) const { _mm_storeu_ps(vals, fVec); }
193
194    SkNi<4, int> castTrunc() const { return _mm_cvttps_epi32(fVec); }
195
196    SkNf operator + (const SkNf& o) const { return _mm_add_ps(fVec, o.fVec); }
197    SkNf operator - (const SkNf& o) const { return _mm_sub_ps(fVec, o.fVec); }
198    SkNf operator * (const SkNf& o) const { return _mm_mul_ps(fVec, o.fVec); }
199    SkNf operator / (const SkNf& o) const { return _mm_div_ps(fVec, o.fVec); }
200
201    Nb operator == (const SkNf& o) const { return _mm_castps_si128(_mm_cmpeq_ps (fVec, o.fVec)); }
202    Nb operator != (const SkNf& o) const { return _mm_castps_si128(_mm_cmpneq_ps(fVec, o.fVec)); }
203    Nb operator  < (const SkNf& o) const { return _mm_castps_si128(_mm_cmplt_ps (fVec, o.fVec)); }
204    Nb operator  > (const SkNf& o) const { return _mm_castps_si128(_mm_cmpgt_ps (fVec, o.fVec)); }
205    Nb operator <= (const SkNf& o) const { return _mm_castps_si128(_mm_cmple_ps (fVec, o.fVec)); }
206    Nb operator >= (const SkNf& o) const { return _mm_castps_si128(_mm_cmpge_ps (fVec, o.fVec)); }
207
208    static SkNf Min(const SkNf& l, const SkNf& r) { return _mm_min_ps(l.fVec, r.fVec); }
209    static SkNf Max(const SkNf& l, const SkNf& r) { return _mm_max_ps(l.fVec, r.fVec); }
210
211    SkNf  sqrt() const { return _mm_sqrt_ps (fVec);  }
212    SkNf rsqrt0() const { return _mm_rsqrt_ps(fVec); }
213    SkNf rsqrt1() const { return this->rsqrt0(); }
214    SkNf rsqrt2() const { return this->rsqrt1(); }
215
216    SkNf       invert() const { return SkNf(1) / *this; }
217    SkNf approxInvert() const { return _mm_rcp_ps(fVec); }
218
219    template <int k> float kth() const {
220        SkASSERT(0 <= k && k < 4);
221        union { __m128 v; float fs[4]; } pun = {fVec};
222        return pun.fs[k&3];
223    }
224
225    __m128 fVec;
226};
227
228template <>
229class SkNi<4, uint16_t> {
230public:
231    SkNi(const __m128i& vec) : fVec(vec) {}
232
233    SkNi() {}
234    explicit SkNi(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
235    static SkNi Load(const uint16_t vals[4]) { return _mm_loadl_epi64((const __m128i*)vals); }
236    SkNi(uint16_t a, uint16_t b, uint16_t c, uint16_t d) : fVec(_mm_setr_epi16(a,b,c,d,0,0,0,0)) {}
237
238    void store(uint16_t vals[4]) const { _mm_storel_epi64((__m128i*)vals, fVec); }
239
240    SkNi operator + (const SkNi& o) const { return _mm_add_epi16(fVec, o.fVec); }
241    SkNi operator - (const SkNi& o) const { return _mm_sub_epi16(fVec, o.fVec); }
242    SkNi operator * (const SkNi& o) const { return _mm_mullo_epi16(fVec, o.fVec); }
243
244    SkNi operator << (int bits) const { return _mm_slli_epi16(fVec, bits); }
245    SkNi operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); }
246
247    template <int k> uint16_t kth() const {
248        SkASSERT(0 <= k && k < 4);
249        return _mm_extract_epi16(fVec, k);
250    }
251
252    __m128i fVec;
253};
254
255template <>
256class SkNi<8, uint16_t> {
257public:
258    SkNi(const __m128i& vec) : fVec(vec) {}
259
260    SkNi() {}
261    explicit SkNi(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
262    static SkNi Load(const uint16_t vals[8]) { return _mm_loadu_si128((const __m128i*)vals); }
263    SkNi(uint16_t a, uint16_t b, uint16_t c, uint16_t d,
264         uint16_t e, uint16_t f, uint16_t g, uint16_t h) : fVec(_mm_setr_epi16(a,b,c,d,e,f,g,h)) {}
265
266    void store(uint16_t vals[8]) const { _mm_storeu_si128((__m128i*)vals, fVec); }
267
268    SkNi operator + (const SkNi& o) const { return _mm_add_epi16(fVec, o.fVec); }
269    SkNi operator - (const SkNi& o) const { return _mm_sub_epi16(fVec, o.fVec); }
270    SkNi operator * (const SkNi& o) const { return _mm_mullo_epi16(fVec, o.fVec); }
271
272    SkNi operator << (int bits) const { return _mm_slli_epi16(fVec, bits); }
273    SkNi operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); }
274
275    static SkNi Min(const SkNi& a, const SkNi& b) {
276        // No unsigned _mm_min_epu16, so we'll shift into a space where we can use the
277        // signed version, _mm_min_epi16, then shift back.
278        const uint16_t top = 0x8000; // Keep this separate from _mm_set1_epi16 or MSVC will whine.
279        const __m128i top_8x = _mm_set1_epi16(top);
280        return _mm_add_epi8(top_8x, _mm_min_epi16(_mm_sub_epi8(a.fVec, top_8x),
281                                                  _mm_sub_epi8(b.fVec, top_8x)));
282    }
283
284    template <int k> uint16_t kth() const {
285        SkASSERT(0 <= k && k < 8);
286        return _mm_extract_epi16(fVec, k);
287    }
288
289    __m128i fVec;
290};
291
292template <>
293class SkNi<16, uint8_t> {
294public:
295    SkNi(const __m128i& vec) : fVec(vec) {}
296
297    SkNi() {}
298    explicit SkNi(uint8_t val) : fVec(_mm_set1_epi8(val)) {}
299    static SkNi Load(const uint8_t vals[16]) { return _mm_loadu_si128((const __m128i*)vals); }
300    SkNi(uint8_t a, uint8_t b, uint8_t c, uint8_t d,
301         uint8_t e, uint8_t f, uint8_t g, uint8_t h,
302         uint8_t i, uint8_t j, uint8_t k, uint8_t l,
303         uint8_t m, uint8_t n, uint8_t o, uint8_t p)
304        : fVec(_mm_setr_epi8(a,b,c,d, e,f,g,h, i,j,k,l, m,n,o,p)) {}
305
306    void store(uint8_t vals[16]) const { _mm_storeu_si128((__m128i*)vals, fVec); }
307
308    SkNi saturatedAdd(const SkNi& o) const { return _mm_adds_epu8(fVec, o.fVec); }
309
310    SkNi operator + (const SkNi& o) const { return _mm_add_epi8(fVec, o.fVec); }
311    SkNi operator - (const SkNi& o) const { return _mm_sub_epi8(fVec, o.fVec); }
312
313    // SSE cannot multiply or shift vectors of uint8_t.
314    SkNi operator * (const SkNi& o) const { SkASSERT(false); return fVec; }
315    SkNi operator << (int bits) const { SkASSERT(false); return fVec; }
316    SkNi operator >> (int bits) const { SkASSERT(false); return fVec; }
317
318    static SkNi Min(const SkNi& a, const SkNi& b) { return _mm_min_epu8(a.fVec, b.fVec); }
319
320    template <int k> uint8_t kth() const {
321        SkASSERT(0 <= k && k < 16);
322        // SSE4.1 would just `return _mm_extract_epi8(fVec, k)`.  We have to read 16-bits instead.
323        int pair = _mm_extract_epi16(fVec, k/2);
324        return k % 2 == 0 ? pair : (pair >> 8);
325    }
326
327    __m128i fVec;
328};
329
330#endif//SkNx_sse_DEFINED
331