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 #include "SkTypes.h"
12 
13 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
14     #include <smmintrin.h>
15 #elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
16     #include <tmmintrin.h>
17 #else
18     #include <emmintrin.h>
19 #endif
20 
21 // This file may assume <= SSE2, but must check SK_CPU_SSE_LEVEL for anything more recent.
22 // If you do, make sure this is in a static inline function... anywhere else risks violating ODR.
23 
24 namespace {  // NOLINT(google-build-namespaces)
25 
26 // Emulate _mm_floor_ps() with SSE2:
27 //   - roundtrip through integers via truncation
28 //   - subtract 1 if that's too big (possible for negative values).
29 // This restricts the domain of our inputs to a maximum somehwere around 2^31.
30 // Seems plenty big.
emulate_mm_floor_ps(__m128 v)31 AI static __m128 emulate_mm_floor_ps(__m128 v) {
32     __m128 roundtrip = _mm_cvtepi32_ps(_mm_cvttps_epi32(v));
33     __m128 too_big = _mm_cmpgt_ps(roundtrip, v);
34     return _mm_sub_ps(roundtrip, _mm_and_ps(too_big, _mm_set1_ps(1.0f)));
35 }
36 
37 template <>
38 class SkNx<2, float> {
39 public:
SkNx(const __m128 & vec)40     AI SkNx(const __m128& vec) : fVec(vec) {}
41 
SkNx()42     AI SkNx() {}
SkNx(float val)43     AI SkNx(float val) : fVec(_mm_set1_ps(val)) {}
Load(const void * ptr)44     AI static SkNx Load(const void* ptr) {
45         return _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*)ptr));
46     }
SkNx(float a,float b)47     AI SkNx(float a, float b) : fVec(_mm_setr_ps(a,b,0,0)) {}
48 
store(void * ptr)49     AI void store(void* ptr) const { _mm_storel_pi((__m64*)ptr, fVec); }
50 
Load2(const void * ptr,SkNx * x,SkNx * y)51     AI static void Load2(const void* ptr, SkNx* x, SkNx* y) {
52         const float* m = (const float*)ptr;
53         *x = SkNx{m[0], m[2]};
54         *y = SkNx{m[1], m[3]};
55     }
56 
Store2(void * dst,const SkNx & a,const SkNx & b)57     AI static void Store2(void* dst, const SkNx& a, const SkNx& b) {
58         auto vals = _mm_unpacklo_ps(a.fVec, b.fVec);
59         _mm_storeu_ps((float*)dst, vals);
60     }
61 
Store3(void * dst,const SkNx & a,const SkNx & b,const SkNx & c)62     AI static void Store3(void* dst, const SkNx& a, const SkNx& b, const SkNx& c) {
63         auto lo = _mm_setr_ps(a[0], b[0], c[0], a[1]),
64              hi = _mm_setr_ps(b[1], c[1],    0,    0);
65         _mm_storeu_ps((float*)dst, lo);
66         _mm_storel_pi(((__m64*)dst) + 2, hi);
67     }
68 
Store4(void * dst,const SkNx & a,const SkNx & b,const SkNx & c,const SkNx & d)69     AI static void Store4(void* dst, const SkNx& a, const SkNx& b, const SkNx& c, const SkNx& d) {
70         auto lo = _mm_setr_ps(a[0], b[0], c[0], d[0]),
71              hi = _mm_setr_ps(a[1], b[1], c[1], d[1]);
72         _mm_storeu_ps((float*)dst, lo);
73         _mm_storeu_ps(((float*)dst) + 4, hi);
74     }
75 
76     AI SkNx operator - () const { return _mm_xor_ps(_mm_set1_ps(-0.0f), fVec); }
77 
78     AI SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
79     AI SkNx operator - (const SkNx& o) const { return _mm_sub_ps(fVec, o.fVec); }
80     AI SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); }
81     AI SkNx operator / (const SkNx& o) const { return _mm_div_ps(fVec, o.fVec); }
82 
83     AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_ps (fVec, o.fVec); }
84     AI SkNx operator != (const SkNx& o) const { return _mm_cmpneq_ps(fVec, o.fVec); }
85     AI SkNx operator  < (const SkNx& o) const { return _mm_cmplt_ps (fVec, o.fVec); }
86     AI SkNx operator  > (const SkNx& o) const { return _mm_cmpgt_ps (fVec, o.fVec); }
87     AI SkNx operator <= (const SkNx& o) const { return _mm_cmple_ps (fVec, o.fVec); }
88     AI SkNx operator >= (const SkNx& o) const { return _mm_cmpge_ps (fVec, o.fVec); }
89 
Min(const SkNx & l,const SkNx & r)90     AI static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_ps(l.fVec, r.fVec); }
Max(const SkNx & l,const SkNx & r)91     AI static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r.fVec); }
92 
abs()93     AI SkNx abs() const { return _mm_andnot_ps(_mm_set1_ps(-0.0f), fVec); }
floor()94     AI SkNx floor() const {
95     #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
96         return _mm_floor_ps(fVec);
97     #else
98         return emulate_mm_floor_ps(fVec);
99     #endif
100     }
101 
sqrt()102     AI SkNx   sqrt() const { return _mm_sqrt_ps (fVec);  }
rsqrt()103     AI SkNx  rsqrt() const { return _mm_rsqrt_ps(fVec); }
invert()104     AI SkNx invert() const { return _mm_rcp_ps(fVec); }
105 
106     AI float operator[](int k) const {
107         SkASSERT(0 <= k && k < 2);
108         union { __m128 v; float fs[4]; } pun = {fVec};
109         return pun.fs[k&1];
110     }
111 
allTrue()112     AI bool allTrue() const { return 0xff == (_mm_movemask_epi8(_mm_castps_si128(fVec)) & 0xff); }
anyTrue()113     AI bool anyTrue() const { return 0x00 != (_mm_movemask_epi8(_mm_castps_si128(fVec)) & 0xff); }
114 
thenElse(const SkNx & t,const SkNx & e)115     AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
116     #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
117         return _mm_blendv_ps(e.fVec, t.fVec, fVec);
118     #else
119         return _mm_or_ps(_mm_and_ps   (fVec, t.fVec),
120                          _mm_andnot_ps(fVec, e.fVec));
121     #endif
122     }
123 
124     __m128 fVec;
125 };
126 
127 template <>
128 class SkNx<4, float> {
129 public:
SkNx(const __m128 & vec)130     AI SkNx(const __m128& vec) : fVec(vec) {}
131 
SkNx()132     AI SkNx() {}
SkNx(float val)133     AI SkNx(float val)           : fVec( _mm_set1_ps(val) ) {}
SkNx(float a,float b,float c,float d)134     AI SkNx(float a, float b, float c, float d) : fVec(_mm_setr_ps(a,b,c,d)) {}
135 
Load(const void * ptr)136     AI static SkNx Load(const void* ptr) { return _mm_loadu_ps((const float*)ptr); }
store(void * ptr)137     AI void store(void* ptr) const { _mm_storeu_ps((float*)ptr, fVec); }
138 
Load2(const void * ptr,SkNx * x,SkNx * y)139     AI static void Load2(const void* ptr, SkNx* x, SkNx* y) {
140         SkNx lo = SkNx::Load((const float*)ptr+0),
141              hi = SkNx::Load((const float*)ptr+4);
142         *x = SkNx{lo[0], lo[2], hi[0], hi[2]};
143         *y = SkNx{lo[1], lo[3], hi[1], hi[3]};
144     }
145 
Load4(const void * ptr,SkNx * r,SkNx * g,SkNx * b,SkNx * a)146     AI static void Load4(const void* ptr, SkNx* r, SkNx* g, SkNx* b, SkNx* a) {
147         __m128 v0 = _mm_loadu_ps(((float*)ptr) +  0),
148                v1 = _mm_loadu_ps(((float*)ptr) +  4),
149                v2 = _mm_loadu_ps(((float*)ptr) +  8),
150                v3 = _mm_loadu_ps(((float*)ptr) + 12);
151         _MM_TRANSPOSE4_PS(v0, v1, v2, v3);
152         *r = v0;
153         *g = v1;
154         *b = v2;
155         *a = v3;
156     }
Store4(void * dst,const SkNx & r,const SkNx & g,const SkNx & b,const SkNx & a)157     AI static void Store4(void* dst, const SkNx& r, const SkNx& g, const SkNx& b, const SkNx& a) {
158         __m128 v0 = r.fVec,
159                v1 = g.fVec,
160                v2 = b.fVec,
161                v3 = a.fVec;
162         _MM_TRANSPOSE4_PS(v0, v1, v2, v3);
163         _mm_storeu_ps(((float*) dst) +  0, v0);
164         _mm_storeu_ps(((float*) dst) +  4, v1);
165         _mm_storeu_ps(((float*) dst) +  8, v2);
166         _mm_storeu_ps(((float*) dst) + 12, v3);
167     }
168 
169     AI SkNx operator - () const { return _mm_xor_ps(_mm_set1_ps(-0.0f), fVec); }
170 
171     AI SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
172     AI SkNx operator - (const SkNx& o) const { return _mm_sub_ps(fVec, o.fVec); }
173     AI SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); }
174     AI SkNx operator / (const SkNx& o) const { return _mm_div_ps(fVec, o.fVec); }
175 
176     AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_ps (fVec, o.fVec); }
177     AI SkNx operator != (const SkNx& o) const { return _mm_cmpneq_ps(fVec, o.fVec); }
178     AI SkNx operator  < (const SkNx& o) const { return _mm_cmplt_ps (fVec, o.fVec); }
179     AI SkNx operator  > (const SkNx& o) const { return _mm_cmpgt_ps (fVec, o.fVec); }
180     AI SkNx operator <= (const SkNx& o) const { return _mm_cmple_ps (fVec, o.fVec); }
181     AI SkNx operator >= (const SkNx& o) const { return _mm_cmpge_ps (fVec, o.fVec); }
182 
Min(const SkNx & l,const SkNx & r)183     AI static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_ps(l.fVec, r.fVec); }
Max(const SkNx & l,const SkNx & r)184     AI static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r.fVec); }
185 
abs()186     AI SkNx abs() const { return _mm_andnot_ps(_mm_set1_ps(-0.0f), fVec); }
floor()187     AI SkNx floor() const {
188     #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
189         return _mm_floor_ps(fVec);
190     #else
191         return emulate_mm_floor_ps(fVec);
192     #endif
193     }
194 
sqrt()195     AI SkNx   sqrt() const { return _mm_sqrt_ps (fVec);  }
rsqrt()196     AI SkNx  rsqrt() const { return _mm_rsqrt_ps(fVec); }
invert()197     AI SkNx invert() const { return _mm_rcp_ps(fVec); }
198 
199     AI float operator[](int k) const {
200         SkASSERT(0 <= k && k < 4);
201         union { __m128 v; float fs[4]; } pun = {fVec};
202         return pun.fs[k&3];
203     }
204 
min()205     AI float min() const {
206         SkNx min = Min(*this, _mm_shuffle_ps(fVec, fVec, _MM_SHUFFLE(2,3,0,1)));
207         min = Min(min, _mm_shuffle_ps(min.fVec, min.fVec, _MM_SHUFFLE(0,1,2,3)));
208         return min[0];
209     }
210 
max()211     AI float max() const {
212         SkNx max = Max(*this, _mm_shuffle_ps(fVec, fVec, _MM_SHUFFLE(2,3,0,1)));
213         max = Max(max, _mm_shuffle_ps(max.fVec, max.fVec, _MM_SHUFFLE(0,1,2,3)));
214         return max[0];
215     }
216 
allTrue()217     AI bool allTrue() const { return 0xffff == _mm_movemask_epi8(_mm_castps_si128(fVec)); }
anyTrue()218     AI bool anyTrue() const { return 0x0000 != _mm_movemask_epi8(_mm_castps_si128(fVec)); }
219 
thenElse(const SkNx & t,const SkNx & e)220     AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
221     #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
222         return _mm_blendv_ps(e.fVec, t.fVec, fVec);
223     #else
224         return _mm_or_ps(_mm_and_ps   (fVec, t.fVec),
225                          _mm_andnot_ps(fVec, e.fVec));
226     #endif
227     }
228 
229     __m128 fVec;
230 };
231 
mullo32(__m128i a,__m128i b)232 AI static __m128i mullo32(__m128i a, __m128i b) {
233 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
234     return _mm_mullo_epi32(a, b);
235 #else
236     __m128i mul20 = _mm_mul_epu32(a, b),
237             mul31 = _mm_mul_epu32(_mm_srli_si128(a, 4), _mm_srli_si128(b, 4));
238     return _mm_unpacklo_epi32(_mm_shuffle_epi32(mul20, _MM_SHUFFLE(0,0,2,0)),
239                               _mm_shuffle_epi32(mul31, _MM_SHUFFLE(0,0,2,0)));
240 #endif
241 }
242 
243 template <>
244 class SkNx<4, int32_t> {
245 public:
SkNx(const __m128i & vec)246     AI SkNx(const __m128i& vec) : fVec(vec) {}
247 
SkNx()248     AI SkNx() {}
SkNx(int32_t val)249     AI SkNx(int32_t val) : fVec(_mm_set1_epi32(val)) {}
Load(const void * ptr)250     AI static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
SkNx(int32_t a,int32_t b,int32_t c,int32_t d)251     AI SkNx(int32_t a, int32_t b, int32_t c, int32_t d) : fVec(_mm_setr_epi32(a,b,c,d)) {}
252 
store(void * ptr)253     AI void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
254 
255     AI SkNx operator + (const SkNx& o) const { return _mm_add_epi32(fVec, o.fVec); }
256     AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); }
257     AI SkNx operator * (const SkNx& o) const { return mullo32(fVec, o.fVec);       }
258 
259     AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
260     AI SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec);  }
261     AI SkNx operator ^ (const SkNx& o) const { return _mm_xor_si128(fVec, o.fVec); }
262 
263     AI SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); }
264     AI SkNx operator >> (int bits) const { return _mm_srai_epi32(fVec, bits); }
265 
266     AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_epi32 (fVec, o.fVec); }
267     AI SkNx operator  < (const SkNx& o) const { return _mm_cmplt_epi32 (fVec, o.fVec); }
268     AI SkNx operator  > (const SkNx& o) const { return _mm_cmpgt_epi32 (fVec, o.fVec); }
269 
270     AI int32_t operator[](int k) const {
271         SkASSERT(0 <= k && k < 4);
272         union { __m128i v; int32_t is[4]; } pun = {fVec};
273         return pun.is[k&3];
274     }
275 
thenElse(const SkNx & t,const SkNx & e)276     AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
277     #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
278         return _mm_blendv_epi8(e.fVec, t.fVec, fVec);
279     #else
280         return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
281                             _mm_andnot_si128(fVec, e.fVec));
282     #endif
283     }
284 
abs()285     AI SkNx abs() const {
286 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
287         return _mm_abs_epi32(fVec);
288 #else
289         SkNx mask = (*this) >> 31;
290         return (mask ^ (*this)) - mask;
291 #endif
292     }
293 
Min(const SkNx & x,const SkNx & y)294     AI static SkNx Min(const SkNx& x, const SkNx& y) {
295 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
296         return _mm_min_epi32(x.fVec, y.fVec);
297 #else
298         return (x < y).thenElse(x, y);
299 #endif
300     }
301 
Max(const SkNx & x,const SkNx & y)302     AI static SkNx Max(const SkNx& x, const SkNx& y) {
303 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
304         return _mm_max_epi32(x.fVec, y.fVec);
305 #else
306         return (x > y).thenElse(x, y);
307 #endif
308     }
309 
310     __m128i fVec;
311 };
312 
313 template <>
314 class SkNx<2, uint32_t> {
315 public:
SkNx(const __m128i & vec)316     AI SkNx(const __m128i& vec) : fVec(vec) {}
317 
SkNx()318     AI SkNx() {}
SkNx(uint32_t val)319     AI SkNx(uint32_t val) : fVec(_mm_set1_epi32(val)) {}
Load(const void * ptr)320     AI static SkNx Load(const void* ptr) { return _mm_loadl_epi64((const __m128i*)ptr); }
SkNx(uint32_t a,uint32_t b)321     AI SkNx(uint32_t a, uint32_t b) : fVec(_mm_setr_epi32(a,b,0,0)) {}
322 
store(void * ptr)323     AI void store(void* ptr) const { _mm_storel_epi64((__m128i*)ptr, fVec); }
324 
325     AI SkNx operator + (const SkNx& o) const { return _mm_add_epi32(fVec, o.fVec); }
326     AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); }
327     AI SkNx operator * (const SkNx& o) const { return mullo32(fVec, o.fVec);       }
328 
329     AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
330     AI SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec);  }
331     AI SkNx operator ^ (const SkNx& o) const { return _mm_xor_si128(fVec, o.fVec); }
332 
333     AI SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); }
334     AI SkNx operator >> (int bits) const { return _mm_srli_epi32(fVec, bits); }
335 
336     AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_epi32 (fVec, o.fVec); }
337     AI SkNx operator != (const SkNx& o) const { return (*this == o) ^ 0xffffffff; }
338     // operator < and > take a little extra fiddling to make work for unsigned ints.
339 
340     AI uint32_t operator[](int k) const {
341         SkASSERT(0 <= k && k < 2);
342         union { __m128i v; uint32_t us[4]; } pun = {fVec};
343         return pun.us[k&1];
344     }
345 
thenElse(const SkNx & t,const SkNx & e)346     AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
347 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
348         return _mm_blendv_epi8(e.fVec, t.fVec, fVec);
349 #else
350         return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
351                             _mm_andnot_si128(fVec, e.fVec));
352 #endif
353     }
354 
allTrue()355     AI bool allTrue() const { return 0xff == (_mm_movemask_epi8(fVec) & 0xff); }
356 
357     __m128i fVec;
358 };
359 
360 template <>
361 class SkNx<4, uint32_t> {
362 public:
SkNx(const __m128i & vec)363     AI SkNx(const __m128i& vec) : fVec(vec) {}
364 
SkNx()365     AI SkNx() {}
SkNx(uint32_t val)366     AI SkNx(uint32_t val) : fVec(_mm_set1_epi32(val)) {}
Load(const void * ptr)367     AI static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
SkNx(uint32_t a,uint32_t b,uint32_t c,uint32_t d)368     AI SkNx(uint32_t a, uint32_t b, uint32_t c, uint32_t d) : fVec(_mm_setr_epi32(a,b,c,d)) {}
369 
store(void * ptr)370     AI void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
371 
372     AI SkNx operator + (const SkNx& o) const { return _mm_add_epi32(fVec, o.fVec); }
373     AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); }
374     AI SkNx operator * (const SkNx& o) const { return mullo32(fVec, o.fVec);       }
375 
376     AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
377     AI SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec);  }
378     AI SkNx operator ^ (const SkNx& o) const { return _mm_xor_si128(fVec, o.fVec); }
379 
380     AI SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); }
381     AI SkNx operator >> (int bits) const { return _mm_srli_epi32(fVec, bits); }
382 
383     AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_epi32 (fVec, o.fVec); }
384     AI SkNx operator != (const SkNx& o) const { return (*this == o) ^ 0xffffffff; }
385 
386     // operator < and > take a little extra fiddling to make work for unsigned ints.
387 
388     AI uint32_t operator[](int k) const {
389         SkASSERT(0 <= k && k < 4);
390         union { __m128i v; uint32_t us[4]; } pun = {fVec};
391         return pun.us[k&3];
392     }
393 
thenElse(const SkNx & t,const SkNx & e)394     AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
395     #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
396         return _mm_blendv_epi8(e.fVec, t.fVec, fVec);
397     #else
398         return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
399                             _mm_andnot_si128(fVec, e.fVec));
400     #endif
401     }
402 
mulHi(SkNx m)403     AI SkNx mulHi(SkNx m) const {
404         SkNx v20{_mm_mul_epu32(m.fVec, fVec)};
405         SkNx v31{_mm_mul_epu32(_mm_srli_si128(m.fVec, 4), _mm_srli_si128(fVec, 4))};
406 
407         return SkNx{v20[1], v31[1], v20[3], v31[3]};
408     }
409 
410     __m128i fVec;
411 };
412 
413 template <>
414 class SkNx<4, uint16_t> {
415 public:
SkNx(const __m128i & vec)416     AI SkNx(const __m128i& vec) : fVec(vec) {}
417 
SkNx()418     AI SkNx() {}
SkNx(uint16_t val)419     AI SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
SkNx(uint16_t a,uint16_t b,uint16_t c,uint16_t d)420     AI SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d)
421         : fVec(_mm_setr_epi16(a,b,c,d,0,0,0,0)) {}
422 
Load(const void * ptr)423     AI static SkNx Load(const void* ptr) { return _mm_loadl_epi64((const __m128i*)ptr); }
store(void * ptr)424     AI void store(void* ptr) const { _mm_storel_epi64((__m128i*)ptr, fVec); }
425 
Load4(const void * ptr,SkNx * r,SkNx * g,SkNx * b,SkNx * a)426     AI static void Load4(const void* ptr, SkNx* r, SkNx* g, SkNx* b, SkNx* a) {
427         __m128i lo = _mm_loadu_si128(((__m128i*)ptr) + 0),
428                 hi = _mm_loadu_si128(((__m128i*)ptr) + 1);
429         __m128i even = _mm_unpacklo_epi16(lo, hi),   // r0 r2 g0 g2 b0 b2 a0 a2
430                  odd = _mm_unpackhi_epi16(lo, hi);   // r1 r3 ...
431         __m128i rg = _mm_unpacklo_epi16(even, odd),  // r0 r1 r2 r3 g0 g1 g2 g3
432                 ba = _mm_unpackhi_epi16(even, odd);  // b0 b1 ...   a0 a1 ...
433         *r = rg;
434         *g = _mm_srli_si128(rg, 8);
435         *b = ba;
436         *a = _mm_srli_si128(ba, 8);
437     }
Load3(const void * ptr,SkNx * r,SkNx * g,SkNx * b)438     AI static void Load3(const void* ptr, SkNx* r, SkNx* g, SkNx* b) {
439         // The idea here is to get 4 vectors that are R G B _ _ _ _ _.
440         // The second load is at a funny location to make sure we don't read past
441         // the bounds of memory.  This is fine, we just need to shift it a little bit.
442         const uint8_t* ptr8 = (const uint8_t*) ptr;
443         __m128i rgb0 = _mm_loadu_si128((const __m128i*) (ptr8 + 0));
444         __m128i rgb1 = _mm_srli_si128(rgb0, 3*2);
445         __m128i rgb2 = _mm_srli_si128(_mm_loadu_si128((const __m128i*) (ptr8 + 4*2)), 2*2);
446         __m128i rgb3 = _mm_srli_si128(rgb2, 3*2);
447 
448         __m128i rrggbb01 = _mm_unpacklo_epi16(rgb0, rgb1);
449         __m128i rrggbb23 = _mm_unpacklo_epi16(rgb2, rgb3);
450         *r = _mm_unpacklo_epi32(rrggbb01, rrggbb23);
451         *g = _mm_srli_si128(r->fVec, 4*2);
452         *b = _mm_unpackhi_epi32(rrggbb01, rrggbb23);
453     }
Store4(void * dst,const SkNx & r,const SkNx & g,const SkNx & b,const SkNx & a)454     AI static void Store4(void* dst, const SkNx& r, const SkNx& g, const SkNx& b, const SkNx& a) {
455         __m128i rg = _mm_unpacklo_epi16(r.fVec, g.fVec);
456         __m128i ba = _mm_unpacklo_epi16(b.fVec, a.fVec);
457         __m128i lo = _mm_unpacklo_epi32(rg, ba);
458         __m128i hi = _mm_unpackhi_epi32(rg, ba);
459         _mm_storeu_si128(((__m128i*) dst) + 0, lo);
460         _mm_storeu_si128(((__m128i*) dst) + 1, hi);
461     }
462 
463     AI SkNx operator + (const SkNx& o) const { return _mm_add_epi16(fVec, o.fVec); }
464     AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi16(fVec, o.fVec); }
465     AI SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec); }
466     AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
467     AI SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); }
468 
469     AI SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); }
470     AI SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); }
471 
472     AI uint16_t operator[](int k) const {
473         SkASSERT(0 <= k && k < 4);
474         union { __m128i v; uint16_t us[8]; } pun = {fVec};
475         return pun.us[k&3];
476     }
477 
478     __m128i fVec;
479 };
480 
481 template <>
482 class SkNx<8, uint16_t> {
483 public:
SkNx(const __m128i & vec)484     AI SkNx(const __m128i& vec) : fVec(vec) {}
485 
SkNx()486     AI SkNx() {}
SkNx(uint16_t val)487     AI SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
SkNx(uint16_t a,uint16_t b,uint16_t c,uint16_t d,uint16_t e,uint16_t f,uint16_t g,uint16_t h)488     AI SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d,
489             uint16_t e, uint16_t f, uint16_t g, uint16_t h)
490         : fVec(_mm_setr_epi16(a,b,c,d,e,f,g,h)) {}
491 
Load(const void * ptr)492     AI static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
store(void * ptr)493     AI void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
494 
Load4(const void * ptr,SkNx * r,SkNx * g,SkNx * b,SkNx * a)495     AI static void Load4(const void* ptr, SkNx* r, SkNx* g, SkNx* b, SkNx* a) {
496         __m128i _01 = _mm_loadu_si128(((__m128i*)ptr) + 0),
497                 _23 = _mm_loadu_si128(((__m128i*)ptr) + 1),
498                 _45 = _mm_loadu_si128(((__m128i*)ptr) + 2),
499                 _67 = _mm_loadu_si128(((__m128i*)ptr) + 3);
500 
501         __m128i _02 = _mm_unpacklo_epi16(_01, _23),  // r0 r2 g0 g2 b0 b2 a0 a2
502                 _13 = _mm_unpackhi_epi16(_01, _23),  // r1 r3 g1 g3 b1 b3 a1 a3
503                 _46 = _mm_unpacklo_epi16(_45, _67),
504                 _57 = _mm_unpackhi_epi16(_45, _67);
505 
506         __m128i rg0123 = _mm_unpacklo_epi16(_02, _13),  // r0 r1 r2 r3 g0 g1 g2 g3
507                 ba0123 = _mm_unpackhi_epi16(_02, _13),  // b0 b1 b2 b3 a0 a1 a2 a3
508                 rg4567 = _mm_unpacklo_epi16(_46, _57),
509                 ba4567 = _mm_unpackhi_epi16(_46, _57);
510 
511         *r = _mm_unpacklo_epi64(rg0123, rg4567);
512         *g = _mm_unpackhi_epi64(rg0123, rg4567);
513         *b = _mm_unpacklo_epi64(ba0123, ba4567);
514         *a = _mm_unpackhi_epi64(ba0123, ba4567);
515     }
Load3(const void * ptr,SkNx * r,SkNx * g,SkNx * b)516     AI static void Load3(const void* ptr, SkNx* r, SkNx* g, SkNx* b) {
517         const uint8_t* ptr8 = (const uint8_t*) ptr;
518         __m128i rgb0 = _mm_loadu_si128((const __m128i*) (ptr8 +  0*2));
519         __m128i rgb1 = _mm_srli_si128(rgb0, 3*2);
520         __m128i rgb2 = _mm_loadu_si128((const __m128i*) (ptr8 +  6*2));
521         __m128i rgb3 = _mm_srli_si128(rgb2, 3*2);
522         __m128i rgb4 = _mm_loadu_si128((const __m128i*) (ptr8 + 12*2));
523         __m128i rgb5 = _mm_srli_si128(rgb4, 3*2);
524         __m128i rgb6 = _mm_srli_si128(_mm_loadu_si128((const __m128i*) (ptr8 + 16*2)), 2*2);
525         __m128i rgb7 = _mm_srli_si128(rgb6, 3*2);
526 
527         __m128i rgb01 = _mm_unpacklo_epi16(rgb0, rgb1);
528         __m128i rgb23 = _mm_unpacklo_epi16(rgb2, rgb3);
529         __m128i rgb45 = _mm_unpacklo_epi16(rgb4, rgb5);
530         __m128i rgb67 = _mm_unpacklo_epi16(rgb6, rgb7);
531 
532         __m128i rg03 = _mm_unpacklo_epi32(rgb01, rgb23);
533         __m128i bx03 = _mm_unpackhi_epi32(rgb01, rgb23);
534         __m128i rg47 = _mm_unpacklo_epi32(rgb45, rgb67);
535         __m128i bx47 = _mm_unpackhi_epi32(rgb45, rgb67);
536 
537         *r = _mm_unpacklo_epi64(rg03, rg47);
538         *g = _mm_unpackhi_epi64(rg03, rg47);
539         *b = _mm_unpacklo_epi64(bx03, bx47);
540     }
Store4(void * ptr,const SkNx & r,const SkNx & g,const SkNx & b,const SkNx & a)541     AI static void Store4(void* ptr, const SkNx& r, const SkNx& g, const SkNx& b, const SkNx& a) {
542         __m128i rg0123 = _mm_unpacklo_epi16(r.fVec, g.fVec),  // r0 g0 r1 g1 r2 g2 r3 g3
543                 rg4567 = _mm_unpackhi_epi16(r.fVec, g.fVec),  // r4 g4 r5 g5 r6 g6 r7 g7
544                 ba0123 = _mm_unpacklo_epi16(b.fVec, a.fVec),
545                 ba4567 = _mm_unpackhi_epi16(b.fVec, a.fVec);
546 
547         _mm_storeu_si128((__m128i*)ptr + 0, _mm_unpacklo_epi32(rg0123, ba0123));
548         _mm_storeu_si128((__m128i*)ptr + 1, _mm_unpackhi_epi32(rg0123, ba0123));
549         _mm_storeu_si128((__m128i*)ptr + 2, _mm_unpacklo_epi32(rg4567, ba4567));
550         _mm_storeu_si128((__m128i*)ptr + 3, _mm_unpackhi_epi32(rg4567, ba4567));
551     }
552 
553     AI SkNx operator + (const SkNx& o) const { return _mm_add_epi16(fVec, o.fVec); }
554     AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi16(fVec, o.fVec); }
555     AI SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec); }
556     AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
557     AI SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); }
558 
559     AI SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); }
560     AI SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); }
561 
Min(const SkNx & a,const SkNx & b)562     AI static SkNx Min(const SkNx& a, const SkNx& b) {
563         // No unsigned _mm_min_epu16, so we'll shift into a space where we can use the
564         // signed version, _mm_min_epi16, then shift back.
565         const uint16_t top = 0x8000; // Keep this separate from _mm_set1_epi16 or MSVC will whine.
566         const __m128i top_8x = _mm_set1_epi16(top);
567         return _mm_add_epi8(top_8x, _mm_min_epi16(_mm_sub_epi8(a.fVec, top_8x),
568                                                   _mm_sub_epi8(b.fVec, top_8x)));
569     }
570 
mulHi(const SkNx & m)571     AI SkNx mulHi(const SkNx& m) const {
572         return _mm_mulhi_epu16(fVec, m.fVec);
573     }
574 
thenElse(const SkNx & t,const SkNx & e)575     AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
576         return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
577                             _mm_andnot_si128(fVec, e.fVec));
578     }
579 
580     AI uint16_t operator[](int k) const {
581         SkASSERT(0 <= k && k < 8);
582         union { __m128i v; uint16_t us[8]; } pun = {fVec};
583         return pun.us[k&7];
584     }
585 
586     __m128i fVec;
587 };
588 
589 template <>
590 class SkNx<4, uint8_t> {
591 public:
SkNx()592     AI SkNx() {}
SkNx(const __m128i & vec)593     AI SkNx(const __m128i& vec) : fVec(vec) {}
SkNx(uint8_t a,uint8_t b,uint8_t c,uint8_t d)594     AI SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d)
595         : fVec(_mm_setr_epi8(a,b,c,d, 0,0,0,0, 0,0,0,0, 0,0,0,0)) {}
596 
Load(const void * ptr)597     AI static SkNx Load(const void* ptr) { return _mm_cvtsi32_si128(*(const int*)ptr); }
store(void * ptr)598     AI void store(void* ptr) const { *(int*)ptr = _mm_cvtsi128_si32(fVec); }
599 
600     AI uint8_t operator[](int k) const {
601         SkASSERT(0 <= k && k < 4);
602         union { __m128i v; uint8_t us[16]; } pun = {fVec};
603         return pun.us[k&3];
604     }
605 
606     // TODO as needed
607 
608     __m128i fVec;
609 };
610 
611 template <>
612 class SkNx<8, uint8_t> {
613 public:
SkNx(const __m128i & vec)614     AI SkNx(const __m128i& vec) : fVec(vec) {}
615 
SkNx()616     AI SkNx() {}
SkNx(uint8_t val)617     AI SkNx(uint8_t val) : fVec(_mm_set1_epi8(val)) {}
Load(const void * ptr)618     AI static SkNx Load(const void* ptr) { return _mm_loadl_epi64((const __m128i*)ptr); }
SkNx(uint8_t a,uint8_t b,uint8_t c,uint8_t d,uint8_t e,uint8_t f,uint8_t g,uint8_t h)619     AI SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d,
620             uint8_t e, uint8_t f, uint8_t g, uint8_t h)
621             : fVec(_mm_setr_epi8(a,b,c,d, e,f,g,h, 0,0,0,0, 0,0,0,0)) {}
622 
store(void * ptr)623     AI void store(void* ptr) const {_mm_storel_epi64((__m128i*)ptr, fVec);}
624 
saturatedAdd(const SkNx & o)625     AI SkNx saturatedAdd(const SkNx& o) const { return _mm_adds_epu8(fVec, o.fVec); }
626 
627     AI SkNx operator + (const SkNx& o) const { return _mm_add_epi8(fVec, o.fVec); }
628     AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi8(fVec, o.fVec); }
629 
Min(const SkNx & a,const SkNx & b)630     AI static SkNx Min(const SkNx& a, const SkNx& b) { return _mm_min_epu8(a.fVec, b.fVec); }
631     AI SkNx operator < (const SkNx& o) const {
632         // There's no unsigned _mm_cmplt_epu8, so we flip the sign bits then use a signed compare.
633         auto flip = _mm_set1_epi8(char(0x80));
634         return _mm_cmplt_epi8(_mm_xor_si128(flip, fVec), _mm_xor_si128(flip, o.fVec));
635     }
636 
637     AI uint8_t operator[](int k) const {
638         SkASSERT(0 <= k && k < 16);
639         union { __m128i v; uint8_t us[16]; } pun = {fVec};
640         return pun.us[k&15];
641     }
642 
thenElse(const SkNx & t,const SkNx & e)643     AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
644         return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
645                             _mm_andnot_si128(fVec, e.fVec));
646     }
647 
648     __m128i fVec;
649 };
650 
651 template <>
652 class SkNx<16, uint8_t> {
653 public:
SkNx(const __m128i & vec)654     AI SkNx(const __m128i& vec) : fVec(vec) {}
655 
SkNx()656     AI SkNx() {}
SkNx(uint8_t val)657     AI SkNx(uint8_t val) : fVec(_mm_set1_epi8(val)) {}
Load(const void * ptr)658     AI static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
SkNx(uint8_t a,uint8_t b,uint8_t c,uint8_t d,uint8_t e,uint8_t f,uint8_t g,uint8_t h,uint8_t i,uint8_t j,uint8_t k,uint8_t l,uint8_t m,uint8_t n,uint8_t o,uint8_t p)659     AI SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d,
660             uint8_t e, uint8_t f, uint8_t g, uint8_t h,
661             uint8_t i, uint8_t j, uint8_t k, uint8_t l,
662             uint8_t m, uint8_t n, uint8_t o, uint8_t p)
663         : fVec(_mm_setr_epi8(a,b,c,d, e,f,g,h, i,j,k,l, m,n,o,p)) {}
664 
store(void * ptr)665     AI void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
666 
saturatedAdd(const SkNx & o)667     AI SkNx saturatedAdd(const SkNx& o) const { return _mm_adds_epu8(fVec, o.fVec); }
668 
669     AI SkNx operator + (const SkNx& o) const { return _mm_add_epi8(fVec, o.fVec); }
670     AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi8(fVec, o.fVec); }
671     AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
672 
Min(const SkNx & a,const SkNx & b)673     AI static SkNx Min(const SkNx& a, const SkNx& b) { return _mm_min_epu8(a.fVec, b.fVec); }
674     AI SkNx operator < (const SkNx& o) const {
675         // There's no unsigned _mm_cmplt_epu8, so we flip the sign bits then use a signed compare.
676         auto flip = _mm_set1_epi8(char(0x80));
677         return _mm_cmplt_epi8(_mm_xor_si128(flip, fVec), _mm_xor_si128(flip, o.fVec));
678     }
679 
680     AI uint8_t operator[](int k) const {
681         SkASSERT(0 <= k && k < 16);
682         union { __m128i v; uint8_t us[16]; } pun = {fVec};
683         return pun.us[k&15];
684     }
685 
thenElse(const SkNx & t,const SkNx & e)686     AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
687         return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
688                             _mm_andnot_si128(fVec, e.fVec));
689     }
690 
691     __m128i fVec;
692 };
693 
694 template<> AI /*static*/ Sk4f SkNx_cast<float, int32_t>(const Sk4i& src) {
695     return _mm_cvtepi32_ps(src.fVec);
696 }
697 
698 template<> AI /*static*/ Sk4f SkNx_cast<float, uint32_t>(const Sk4u& src) {
699     return SkNx_cast<float>(Sk4i::Load(&src));
700 }
701 
702 template <> AI /*static*/ Sk4i SkNx_cast<int32_t, float>(const Sk4f& src) {
703     return _mm_cvttps_epi32(src.fVec);
704 }
705 
706 template<> AI /*static*/ Sk4h SkNx_cast<uint16_t, int32_t>(const Sk4i& src) {
707 #if 0 && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
708     // TODO: This seems to be causing code generation problems.   Investigate?
709     return _mm_packus_epi32(src.fVec);
710 #elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
711     // With SSSE3, we can just shuffle the low 2 bytes from each lane right into place.
712     const int _ = ~0;
713     return _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_,_,_));
714 #else
715     // With SSE2, we have to sign extend our input, making _mm_packs_epi32 do the pack we want.
716     __m128i x = _mm_srai_epi32(_mm_slli_epi32(src.fVec, 16), 16);
717     return _mm_packs_epi32(x,x);
718 #endif
719 }
720 
721 template<> AI /*static*/ Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) {
722     return SkNx_cast<uint16_t>(SkNx_cast<int32_t>(src));
723 }
724 
725 template<> AI /*static*/ Sk4b SkNx_cast<uint8_t, float>(const Sk4f& src) {
726     auto _32 = _mm_cvttps_epi32(src.fVec);
727 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
728     const int _ = ~0;
729     return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,4,8,12, _,_,_,_, _,_,_,_, _,_,_,_));
730 #else
731     auto _16 = _mm_packus_epi16(_32, _32);
732     return     _mm_packus_epi16(_16, _16);
733 #endif
734 }
735 
736 template<> AI /*static*/ Sk4u SkNx_cast<uint32_t, uint8_t>(const Sk4b& src) {
737 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
738     const int _ = ~0;
739     return _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,_,_,_, 1,_,_,_, 2,_,_,_, 3,_,_,_));
740 #else
741     auto _16 = _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128());
742     return _mm_unpacklo_epi16(_16, _mm_setzero_si128());
743 #endif
744 }
745 
746 template<> AI /*static*/ Sk4i SkNx_cast<int32_t, uint8_t>(const Sk4b& src) {
747     return SkNx_cast<uint32_t>(src).fVec;
748 }
749 
750 template<> AI /*static*/ Sk4f SkNx_cast<float, uint8_t>(const Sk4b& src) {
751     return _mm_cvtepi32_ps(SkNx_cast<int32_t>(src).fVec);
752 }
753 
754 template<> AI /*static*/ Sk4f SkNx_cast<float, uint16_t>(const Sk4h& src) {
755     auto _32 = _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128());
756     return _mm_cvtepi32_ps(_32);
757 }
758 
759 template<> AI /*static*/ Sk8b SkNx_cast<uint8_t, int32_t>(const Sk8i& src) {
760     Sk4i lo, hi;
761     SkNx_split(src, &lo, &hi);
762 
763     auto t = _mm_packs_epi32(lo.fVec, hi.fVec);
764     return _mm_packus_epi16(t, t);
765 }
766 
767 template<> AI /*static*/ Sk16b SkNx_cast<uint8_t, float>(const Sk16f& src) {
768     Sk8f ab, cd;
769     SkNx_split(src, &ab, &cd);
770 
771     Sk4f a,b,c,d;
772     SkNx_split(ab, &a, &b);
773     SkNx_split(cd, &c, &d);
774 
775     return _mm_packus_epi16(_mm_packus_epi16(_mm_cvttps_epi32(a.fVec),
776                                              _mm_cvttps_epi32(b.fVec)),
777                             _mm_packus_epi16(_mm_cvttps_epi32(c.fVec),
778                                              _mm_cvttps_epi32(d.fVec)));
779 }
780 
781 template<> AI /*static*/ Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src) {
782     return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128());
783 }
784 
785 template<> AI /*static*/ Sk8h SkNx_cast<uint16_t, uint8_t>(const Sk8b& src) {
786     return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128());
787 }
788 
789 template<> AI /*static*/ Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) {
790     return _mm_packus_epi16(src.fVec, src.fVec);
791 }
792 
793 template<> AI /*static*/ Sk8b SkNx_cast<uint8_t, uint16_t>(const Sk8h& src) {
794     return _mm_packus_epi16(src.fVec, src.fVec);
795 }
796 
797 template<> AI /*static*/ Sk4i SkNx_cast<int32_t, uint16_t>(const Sk4h& src) {
798     return _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128());
799 }
800 
801 
802 template<> AI /*static*/ Sk4b SkNx_cast<uint8_t, int32_t>(const Sk4i& src) {
803     return _mm_packus_epi16(_mm_packus_epi16(src.fVec, src.fVec), src.fVec);
804 }
805 
806 template<> AI /*static*/ Sk4b SkNx_cast<uint8_t, uint32_t>(const Sk4u& src) {
807     return _mm_packus_epi16(_mm_packus_epi16(src.fVec, src.fVec), src.fVec);
808 }
809 
810 template<> AI /*static*/ Sk4i SkNx_cast<int32_t, uint32_t>(const Sk4u& src) {
811     return src.fVec;
812 }
813 
Sk4f_round(const Sk4f & x)814 AI static Sk4i Sk4f_round(const Sk4f& x) {
815     return _mm_cvtps_epi32(x.fVec);
816 }
817 
818 }  // namespace
819 
820 #endif//SkNx_sse_DEFINED
821