• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2019 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 SKVX_DEFINED
9 #define SKVX_DEFINED
10 
11 // skvx::Vec<N,T> are SIMD vectors of N T's, a v1.5 successor to SkNx<N,T>.
12 //
13 // This time we're leaning a bit less on platform-specific intrinsics and a bit
14 // more on Clang/GCC vector extensions, but still keeping the option open to
15 // drop in platform-specific intrinsics, actually more easily than before.
16 //
17 // We've also fixed a few of the caveats that used to make SkNx awkward to work
18 // with across translation units.  skvx::Vec<N,T> always has N*sizeof(T) size
19 // and alignof(T) alignment and is safe to use across translation units freely.
20 
21 #include "SkTypes.h"         // SK_CPU_SSE_LEVEL*, etc.
22 #include <algorithm>         // std::min, std::max
23 #include <cmath>             // std::ceil, std::floor, std::trunc, std::round, std::sqrt, etc.
24 #include <cstdint>           // intXX_t
25 #include <cstring>           // memcpy()
26 #include <initializer_list>  // std::initializer_list
27 
28 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE1
29     #include <immintrin.h>
30 #elif defined(SK_ARM_HAS_NEON)
31     #include <arm_neon.h>
32 #endif
33 
34 
35 namespace skvx {
36 
37 // All Vec have the same simple memory layout, the same as `T vec[N]`.
38 // This gives Vec a consistent ABI, letting them pass between files compiled with
39 // different instruction sets (e.g. SSE2 and AVX2) without fear of ODR violation.
40 template <int N, typename T>
41 struct Vec {
42     static_assert((N & (N-1)) == 0, "N must be a power of 2.");
43 
44     Vec<N/2,T> lo, hi;
45 
46     // Methods belong here in the class declaration of Vec only if:
47     //   - they must be here, like constructors or operator[];
48     //   - they'll definitely never want a specialized implementation.
49     // Other operations on Vec should be defined outside the type.
50 
51     Vec() = default;
VecVec52     Vec(T x) : lo(x), hi(x) {}
53 
VecVec54     Vec(std::initializer_list<T> xs) {
55         T vals[N] = {0};
56         memcpy(vals, xs.begin(), std::min(xs.size(), (size_t)N)*sizeof(T));
57 
58         lo = Vec<N/2,T>::Load(vals +   0);
59         hi = Vec<N/2,T>::Load(vals + N/2);
60     }
61 
62     T  operator[](int i) const { return i < N/2 ? lo[i] : hi[i-N/2]; }
63     T& operator[](int i)       { return i < N/2 ? lo[i] : hi[i-N/2]; }
64 
LoadVec65     static Vec Load(const void* ptr) {
66         Vec v;
67         memcpy(&v, ptr, sizeof(Vec));
68         return v;
69     }
storeVec70     void store(void* ptr) const {
71         memcpy(ptr, this, sizeof(Vec));
72     }
73 };
74 
75 template <typename T>
76 struct Vec<1,T> {
77     T val;
78 
79     Vec() = default;
80     Vec(T x) : val(x) {}
81 
82     Vec(std::initializer_list<T> xs) : val(xs.size() ? *xs.begin() : 0) {}
83 
84     T  operator[](int) const { return val; }
85     T& operator[](int)       { return val; }
86 
87     static Vec Load(const void* ptr) {
88         Vec v;
89         memcpy(&v, ptr, sizeof(Vec));
90         return v;
91     }
92     void store(void* ptr) const {
93         memcpy(ptr, this, sizeof(Vec));
94     }
95 };
96 
97 #if defined(__GNUC__) && !defined(__clang__) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE1
98     // GCC warns about ABI changes when returning >= 32 byte vectors when -mavx is not enabled.
99     // This only happens for types like VExt whose ABI we don't care about, not for Vec itself.
100     #pragma GCC diagnostic ignored "-Wpsabi"
101 #endif
102 
103 // Helps tamp down on the repetitive boilerplate.
104 #define SINT template <int N, typename T> static inline
105 #define SIT  template <       typename T> static inline
106 #define SI                                static inline
107 
108 template <typename D, typename S>
109 SI D bit_pun(S s) {
110     static_assert(sizeof(D) == sizeof(S), "");
111     D d;
112     memcpy(&d, &s, sizeof(D));
113     return d;
114 }
115 
116 // Translate from a value type T to its corresponding Mask, the result of a comparison.
117 template <typename T> struct Mask { using type = T; };
118 template <> struct Mask<float > { using type = int32_t; };
119 template <> struct Mask<double> { using type = int64_t; };
120 template <typename T> using M = typename Mask<T>::type;
121 
122 // Join two Vec<N,T> into one Vec<2N,T>.
123 SINT Vec<2*N,T> join(Vec<N,T> lo, Vec<N,T> hi) {
124     Vec<2*N,T> v;
125     v.lo = lo;
126     v.hi = hi;
127     return v;
128 }
129 
130 // We have two default strategies for implementing most operations:
131 //    1) lean on Clang/GCC vector extensions when available;
132 //    2) recurse to scalar portable implementations when not.
133 // At the end we can drop in platform-specific implementations that override either default.
134 
135 #if !defined(SKNX_NO_SIMD) && (defined(__clang__) || defined(__GNUC__))
136 
137     // VExt<N,T> types have the same size as Vec<N,T> and support most operations directly.
138     // N.B. VExt<N,T> alignment is N*alignof(T), stricter than Vec<N,T>'s alignof(T).
139     #if defined(__clang__)
140         template <int N, typename T>
141         using VExt = T __attribute__((ext_vector_type(N)));
142 
143     #elif defined(__GNUC__)
144         template <int N, typename T>
145         struct VExtHelper {
146             typedef T __attribute__((vector_size(N*sizeof(T)))) type;
147         };
148 
149         template <int N, typename T>
150         using VExt = typename VExtHelper<N,T>::type;
151 
152         // For some reason some (new!) versions of GCC cannot seem to deduce N in the generic
153         // to_vec<N,T>() below for N=4 and T=float.  This workaround seems to help...
154         SI Vec<4,float> to_vec(VExt<4,float> v) { return bit_pun<Vec<4,float>>(v); }
155     #endif
156 
157     SINT VExt<N,T> to_vext(Vec<N,T> v) { return bit_pun<VExt<N,T>>(v); }
158     SINT Vec <N,T> to_vec(VExt<N,T> v) { return bit_pun<Vec <N,T>>(v); }
159 
160     SINT Vec<N,T> operator+(Vec<N,T> x, Vec<N,T> y) { return to_vec(to_vext(x) + to_vext(y)); }
161     SINT Vec<N,T> operator-(Vec<N,T> x, Vec<N,T> y) { return to_vec(to_vext(x) - to_vext(y)); }
162     SINT Vec<N,T> operator*(Vec<N,T> x, Vec<N,T> y) { return to_vec(to_vext(x) * to_vext(y)); }
163     SINT Vec<N,T> operator/(Vec<N,T> x, Vec<N,T> y) { return to_vec(to_vext(x) / to_vext(y)); }
164 
165     SINT Vec<N,T> operator^(Vec<N,T> x, Vec<N,T> y) { return to_vec(to_vext(x) ^ to_vext(y)); }
166     SINT Vec<N,T> operator&(Vec<N,T> x, Vec<N,T> y) { return to_vec(to_vext(x) & to_vext(y)); }
167     SINT Vec<N,T> operator|(Vec<N,T> x, Vec<N,T> y) { return to_vec(to_vext(x) | to_vext(y)); }
168 
169     SINT Vec<N,T> operator!(Vec<N,T> x) { return to_vec(!to_vext(x)); }
170     SINT Vec<N,T> operator-(Vec<N,T> x) { return to_vec(-to_vext(x)); }
171     SINT Vec<N,T> operator~(Vec<N,T> x) { return to_vec(~to_vext(x)); }
172 
173     SINT Vec<N,T> operator<<(Vec<N,T> x, int bits) { return to_vec(to_vext(x) << bits); }
174     SINT Vec<N,T> operator>>(Vec<N,T> x, int bits) { return to_vec(to_vext(x) >> bits); }
175 
176     SINT Vec<N,M<T>> operator==(Vec<N,T> x, Vec<N,T> y) { return bit_pun<Vec<N,M<T>>>(to_vext(x) == to_vext(y)); }
177     SINT Vec<N,M<T>> operator!=(Vec<N,T> x, Vec<N,T> y) { return bit_pun<Vec<N,M<T>>>(to_vext(x) != to_vext(y)); }
178     SINT Vec<N,M<T>> operator<=(Vec<N,T> x, Vec<N,T> y) { return bit_pun<Vec<N,M<T>>>(to_vext(x) <= to_vext(y)); }
179     SINT Vec<N,M<T>> operator>=(Vec<N,T> x, Vec<N,T> y) { return bit_pun<Vec<N,M<T>>>(to_vext(x) >= to_vext(y)); }
180     SINT Vec<N,M<T>> operator< (Vec<N,T> x, Vec<N,T> y) { return bit_pun<Vec<N,M<T>>>(to_vext(x) <  to_vext(y)); }
181     SINT Vec<N,M<T>> operator> (Vec<N,T> x, Vec<N,T> y) { return bit_pun<Vec<N,M<T>>>(to_vext(x) >  to_vext(y)); }
182 
183 #else
184 
185     // Either SKNX_NO_SIMD is defined, or Clang/GCC vector extensions are not available.
186     // We'll implement things portably, in a way that should be easily autovectorizable.
187 
188     // N == 1 scalar implementations.
189     SIT Vec<1,T> operator+(Vec<1,T> x, Vec<1,T> y) { return x.val + y.val; }
190     SIT Vec<1,T> operator-(Vec<1,T> x, Vec<1,T> y) { return x.val - y.val; }
191     SIT Vec<1,T> operator*(Vec<1,T> x, Vec<1,T> y) { return x.val * y.val; }
192     SIT Vec<1,T> operator/(Vec<1,T> x, Vec<1,T> y) { return x.val / y.val; }
193 
194     SIT Vec<1,T> operator^(Vec<1,T> x, Vec<1,T> y) { return x.val ^ y.val; }
195     SIT Vec<1,T> operator&(Vec<1,T> x, Vec<1,T> y) { return x.val & y.val; }
196     SIT Vec<1,T> operator|(Vec<1,T> x, Vec<1,T> y) { return x.val | y.val; }
197 
198     SIT Vec<1,T> operator!(Vec<1,T> x) { return !x.val; }
199     SIT Vec<1,T> operator-(Vec<1,T> x) { return -x.val; }
200     SIT Vec<1,T> operator~(Vec<1,T> x) { return ~x.val; }
201 
202     SIT Vec<1,T> operator<<(Vec<1,T> x, int bits) { return x.val << bits; }
203     SIT Vec<1,T> operator>>(Vec<1,T> x, int bits) { return x.val >> bits; }
204 
205     SIT Vec<1,M<T>> operator==(Vec<1,T> x, Vec<1,T> y) { return x.val == y.val ? ~0 : 0; }
206     SIT Vec<1,M<T>> operator!=(Vec<1,T> x, Vec<1,T> y) { return x.val != y.val ? ~0 : 0; }
207     SIT Vec<1,M<T>> operator<=(Vec<1,T> x, Vec<1,T> y) { return x.val <= y.val ? ~0 : 0; }
208     SIT Vec<1,M<T>> operator>=(Vec<1,T> x, Vec<1,T> y) { return x.val >= y.val ? ~0 : 0; }
209     SIT Vec<1,M<T>> operator< (Vec<1,T> x, Vec<1,T> y) { return x.val <  y.val ? ~0 : 0; }
210     SIT Vec<1,M<T>> operator> (Vec<1,T> x, Vec<1,T> y) { return x.val >  y.val ? ~0 : 0; }
211 
212     // All default N != 1 implementations just recurse on lo and hi halves.
213     SINT Vec<N,T> operator+(Vec<N,T> x, Vec<N,T> y) { return join(x.lo + y.lo, x.hi + y.hi); }
214     SINT Vec<N,T> operator-(Vec<N,T> x, Vec<N,T> y) { return join(x.lo - y.lo, x.hi - y.hi); }
215     SINT Vec<N,T> operator*(Vec<N,T> x, Vec<N,T> y) { return join(x.lo * y.lo, x.hi * y.hi); }
216     SINT Vec<N,T> operator/(Vec<N,T> x, Vec<N,T> y) { return join(x.lo / y.lo, x.hi / y.hi); }
217 
218     SINT Vec<N,T> operator^(Vec<N,T> x, Vec<N,T> y) { return join(x.lo ^ y.lo, x.hi ^ y.hi); }
219     SINT Vec<N,T> operator&(Vec<N,T> x, Vec<N,T> y) { return join(x.lo & y.lo, x.hi & y.hi); }
220     SINT Vec<N,T> operator|(Vec<N,T> x, Vec<N,T> y) { return join(x.lo | y.lo, x.hi | y.hi); }
221 
222     SINT Vec<N,T> operator!(Vec<N,T> x) { return join(!x.lo, !x.hi); }
223     SINT Vec<N,T> operator-(Vec<N,T> x) { return join(-x.lo, -x.hi); }
224     SINT Vec<N,T> operator~(Vec<N,T> x) { return join(~x.lo, ~x.hi); }
225 
226     SINT Vec<N,T> operator<<(Vec<N,T> x, int bits) { return join(x.lo << bits, x.hi << bits); }
227     SINT Vec<N,T> operator>>(Vec<N,T> x, int bits) { return join(x.lo >> bits, x.hi >> bits); }
228 
229     SINT Vec<N,M<T>> operator==(Vec<N,T> x, Vec<N,T> y) { return join(x.lo == y.lo, x.hi == y.hi); }
230     SINT Vec<N,M<T>> operator!=(Vec<N,T> x, Vec<N,T> y) { return join(x.lo != y.lo, x.hi != y.hi); }
231     SINT Vec<N,M<T>> operator<=(Vec<N,T> x, Vec<N,T> y) { return join(x.lo <= y.lo, x.hi <= y.hi); }
232     SINT Vec<N,M<T>> operator>=(Vec<N,T> x, Vec<N,T> y) { return join(x.lo >= y.lo, x.hi >= y.hi); }
233     SINT Vec<N,M<T>> operator< (Vec<N,T> x, Vec<N,T> y) { return join(x.lo <  y.lo, x.hi <  y.hi); }
234     SINT Vec<N,M<T>> operator> (Vec<N,T> x, Vec<N,T> y) { return join(x.lo >  y.lo, x.hi >  y.hi); }
235 #endif
236 
237 // Some operations we want are not expressible with Clang/GCC vector
238 // extensions, so we implement them using the recursive approach.
239 
240 // N == 1 scalar implementations.
241 SIT Vec<1,T> if_then_else(Vec<1,M<T>> cond, Vec<1,T> t, Vec<1,T> e) {
242     auto t_bits = bit_pun<M<T>>(t),
243          e_bits = bit_pun<M<T>>(e);
244     return bit_pun<T>( (cond.val & t_bits) | (~cond.val & e_bits) );
245 }
246 
247 SIT bool any(Vec<1,T> x) { return x.val != 0; }
248 SIT bool all(Vec<1,T> x) { return x.val != 0; }
249 
250 SIT T min(Vec<1,T> x) { return x.val; }
251 SIT T max(Vec<1,T> x) { return x.val; }
252 
253 SIT Vec<1,T> min(Vec<1,T> x, Vec<1,T> y) { return std::min(x.val, y.val); }
254 SIT Vec<1,T> max(Vec<1,T> x, Vec<1,T> y) { return std::max(x.val, y.val); }
255 
256 SIT Vec<1,T>  ceil(Vec<1,T> x) { return std:: ceil(x.val); }
257 SIT Vec<1,T> floor(Vec<1,T> x) { return std::floor(x.val); }
258 SIT Vec<1,T> trunc(Vec<1,T> x) { return std::trunc(x.val); }
259 SIT Vec<1,T> round(Vec<1,T> x) { return std::round(x.val); }
260 SIT Vec<1,T>  sqrt(Vec<1,T> x) { return std:: sqrt(x.val); }
261 SIT Vec<1,T>   abs(Vec<1,T> x) { return std::  abs(x.val); }
262 
263 SIT Vec<1,T>   rcp(Vec<1,T> x) { return 1 / x.val; }
264 SIT Vec<1,T> rsqrt(Vec<1,T> x) { return rcp(sqrt(x)); }
265 SIT Vec<1,T>   mad(Vec<1,T> f,
266                    Vec<1,T> m,
267                    Vec<1,T> a) { return f*m+a; }
268 
269 // All default N != 1 implementations just recurse on lo and hi halves.
270 SINT Vec<N,T> if_then_else(Vec<N,M<T>> cond, Vec<N,T> t, Vec<N,T> e) {
271     return join(if_then_else(cond.lo, t.lo, e.lo),
272                 if_then_else(cond.hi, t.hi, e.hi));
273 }
274 
275 SINT bool any(Vec<N,T> x) { return any(x.lo) || any(x.hi); }
276 SINT bool all(Vec<N,T> x) { return all(x.lo) && all(x.hi); }
277 
278 SINT T min(Vec<N,T> x) { return std::min(min(x.lo), min(x.hi)); }
279 SINT T max(Vec<N,T> x) { return std::max(max(x.lo), max(x.hi)); }
280 
281 SINT Vec<N,T> min(Vec<N,T> x, Vec<N,T> y) { return join(min(x.lo, y.lo), min(x.hi, y.hi)); }
282 SINT Vec<N,T> max(Vec<N,T> x, Vec<N,T> y) { return join(max(x.lo, y.lo), max(x.hi, y.hi)); }
283 
284 SINT Vec<N,T>  ceil(Vec<N,T> x) { return join( ceil(x.lo),  ceil(x.hi)); }
285 SINT Vec<N,T> floor(Vec<N,T> x) { return join(floor(x.lo), floor(x.hi)); }
286 SINT Vec<N,T> trunc(Vec<N,T> x) { return join(trunc(x.lo), trunc(x.hi)); }
287 SINT Vec<N,T> round(Vec<N,T> x) { return join(round(x.lo), round(x.hi)); }
288 SINT Vec<N,T>  sqrt(Vec<N,T> x) { return join( sqrt(x.lo),  sqrt(x.hi)); }
289 SINT Vec<N,T>   abs(Vec<N,T> x) { return join(  abs(x.lo),   abs(x.hi)); }
290 
291 SINT Vec<N,T>   rcp(Vec<N,T> x) { return join(  rcp(x.lo),   rcp(x.hi)); }
292 SINT Vec<N,T> rsqrt(Vec<N,T> x) { return join(rsqrt(x.lo), rsqrt(x.hi)); }
293 SINT Vec<N,T>   mad(Vec<N,T> f,
294                     Vec<N,T> m,
295                     Vec<N,T> a) { return join(mad(f.lo, m.lo, a.lo), mad(f.hi, m.hi, a.hi)); }
296 
297 
298 // Scalar/vector operations just splat the scalar to a vector...
299 SINT Vec<N,T>    operator+ (T x, Vec<N,T> y) { return Vec<N,T>(x) +  y; }
300 SINT Vec<N,T>    operator- (T x, Vec<N,T> y) { return Vec<N,T>(x) -  y; }
301 SINT Vec<N,T>    operator* (T x, Vec<N,T> y) { return Vec<N,T>(x) *  y; }
302 SINT Vec<N,T>    operator/ (T x, Vec<N,T> y) { return Vec<N,T>(x) /  y; }
303 SINT Vec<N,T>    operator^ (T x, Vec<N,T> y) { return Vec<N,T>(x) ^  y; }
304 SINT Vec<N,T>    operator& (T x, Vec<N,T> y) { return Vec<N,T>(x) &  y; }
305 SINT Vec<N,T>    operator| (T x, Vec<N,T> y) { return Vec<N,T>(x) |  y; }
306 SINT Vec<N,M<T>> operator==(T x, Vec<N,T> y) { return Vec<N,T>(x) == y; }
307 SINT Vec<N,M<T>> operator!=(T x, Vec<N,T> y) { return Vec<N,T>(x) != y; }
308 SINT Vec<N,M<T>> operator<=(T x, Vec<N,T> y) { return Vec<N,T>(x) <= y; }
309 SINT Vec<N,M<T>> operator>=(T x, Vec<N,T> y) { return Vec<N,T>(x) >= y; }
310 SINT Vec<N,M<T>> operator< (T x, Vec<N,T> y) { return Vec<N,T>(x) <  y; }
311 SINT Vec<N,M<T>> operator> (T x, Vec<N,T> y) { return Vec<N,T>(x) >  y; }
312 SINT Vec<N,T>           min(T x, Vec<N,T> y) { return min(Vec<N,T>(x), y); }
313 SINT Vec<N,T>           max(T x, Vec<N,T> y) { return max(Vec<N,T>(x), y); }
314 
315 // ... and same deal for vector/scalar operations.
316 SINT Vec<N,T>    operator+ (Vec<N,T> x, T y) { return x +  Vec<N,T>(y); }
317 SINT Vec<N,T>    operator- (Vec<N,T> x, T y) { return x -  Vec<N,T>(y); }
318 SINT Vec<N,T>    operator* (Vec<N,T> x, T y) { return x *  Vec<N,T>(y); }
319 SINT Vec<N,T>    operator/ (Vec<N,T> x, T y) { return x /  Vec<N,T>(y); }
320 SINT Vec<N,T>    operator^ (Vec<N,T> x, T y) { return x ^  Vec<N,T>(y); }
321 SINT Vec<N,T>    operator& (Vec<N,T> x, T y) { return x &  Vec<N,T>(y); }
322 SINT Vec<N,T>    operator| (Vec<N,T> x, T y) { return x |  Vec<N,T>(y); }
323 SINT Vec<N,M<T>> operator==(Vec<N,T> x, T y) { return x == Vec<N,T>(y); }
324 SINT Vec<N,M<T>> operator!=(Vec<N,T> x, T y) { return x != Vec<N,T>(y); }
325 SINT Vec<N,M<T>> operator<=(Vec<N,T> x, T y) { return x <= Vec<N,T>(y); }
326 SINT Vec<N,M<T>> operator>=(Vec<N,T> x, T y) { return x >= Vec<N,T>(y); }
327 SINT Vec<N,M<T>> operator< (Vec<N,T> x, T y) { return x <  Vec<N,T>(y); }
328 SINT Vec<N,M<T>> operator> (Vec<N,T> x, T y) { return x >  Vec<N,T>(y); }
329 SINT Vec<N,T>           min(Vec<N,T> x, T y) { return min(x, Vec<N,T>(y)); }
330 SINT Vec<N,T>           max(Vec<N,T> x, T y) { return max(x, Vec<N,T>(y)); }
331 
332 // All vector/scalar combinations for mad() with at least one vector.
333 SINT Vec<N,T> mad(T f, Vec<N,T> m, Vec<N,T> a) { return Vec<N,T>(f)*m + a; }
334 SINT Vec<N,T> mad(Vec<N,T> f, T m, Vec<N,T> a) { return f*Vec<N,T>(m) + a; }
335 SINT Vec<N,T> mad(Vec<N,T> f, Vec<N,T> m, T a) { return f*m + Vec<N,T>(a); }
336 SINT Vec<N,T> mad(Vec<N,T> f, T m, T a) { return f*Vec<N,T>(m) + Vec<N,T>(a); }
337 SINT Vec<N,T> mad(T f, Vec<N,T> m, T a) { return Vec<N,T>(f)*m + Vec<N,T>(a); }
338 SINT Vec<N,T> mad(T f, T m, Vec<N,T> a) { return Vec<N,T>(f)*Vec<N,T>(m) + a; }
339 
340 // The various op= operators, for vectors...
341 SINT Vec<N,T>& operator+=(Vec<N,T>& x, Vec<N,T> y) { return (x = x + y); }
342 SINT Vec<N,T>& operator-=(Vec<N,T>& x, Vec<N,T> y) { return (x = x - y); }
343 SINT Vec<N,T>& operator*=(Vec<N,T>& x, Vec<N,T> y) { return (x = x * y); }
344 SINT Vec<N,T>& operator/=(Vec<N,T>& x, Vec<N,T> y) { return (x = x / y); }
345 SINT Vec<N,T>& operator^=(Vec<N,T>& x, Vec<N,T> y) { return (x = x ^ y); }
346 SINT Vec<N,T>& operator&=(Vec<N,T>& x, Vec<N,T> y) { return (x = x & y); }
347 SINT Vec<N,T>& operator|=(Vec<N,T>& x, Vec<N,T> y) { return (x = x | y); }
348 
349 // ... for scalars...
350 SINT Vec<N,T>& operator+=(Vec<N,T>& x, T y) { return (x = x + Vec<N,T>(y)); }
351 SINT Vec<N,T>& operator-=(Vec<N,T>& x, T y) { return (x = x - Vec<N,T>(y)); }
352 SINT Vec<N,T>& operator*=(Vec<N,T>& x, T y) { return (x = x * Vec<N,T>(y)); }
353 SINT Vec<N,T>& operator/=(Vec<N,T>& x, T y) { return (x = x / Vec<N,T>(y)); }
354 SINT Vec<N,T>& operator^=(Vec<N,T>& x, T y) { return (x = x ^ Vec<N,T>(y)); }
355 SINT Vec<N,T>& operator&=(Vec<N,T>& x, T y) { return (x = x & Vec<N,T>(y)); }
356 SINT Vec<N,T>& operator|=(Vec<N,T>& x, T y) { return (x = x | Vec<N,T>(y)); }
357 
358 // ... and for shifts.
359 SINT Vec<N,T>& operator<<=(Vec<N,T>& x, int bits) { return (x = x << bits); }
360 SINT Vec<N,T>& operator>>=(Vec<N,T>& x, int bits) { return (x = x >> bits); }
361 
362 }  // namespace skvx
363 
364 // These next few routines take extra template arguments that prevent
365 // argument-dependent lookup.  They must live outside the skvx namespace,
366 // but since they operate only on skvx::Vec, that shouldn't be a big deal.
367 
368 // cast() Vec<N,S> to Vec<N,D>, as if applying a C-cast to each lane.
369 template <typename D, typename S>
370 SI skvx::Vec<1,D> cast(skvx::Vec<1,S> src) { return (D)src.val; }
371 
372 template <typename D, int N, typename S>
373 SI skvx::Vec<N,D> cast(skvx::Vec<N,S> src) {
374 #if !defined(SKNX_NO_SIMD) && defined(__clang__)
375     return skvx::to_vec(__builtin_convertvector(skvx::to_vext(src), skvx::VExt<N,D>));
376 #else
377     return join(cast<D>(src.lo), cast<D>(src.hi));
378 #endif
379 }
380 
381 // Shuffle values from a vector pretty arbitrarily:
382 //    skvx::Vec<4,float> rgba = {R,G,B,A};
383 //    shuffle<2,1,0,3>        (rgba) ~> {B,G,R,A}
384 //    shuffle<2,1>            (rgba) ~> {B,G}
385 //    shuffle<2,1,2,1,2,1,2,1>(rgba) ~> {B,G,B,G,B,G,B,G}
386 //    shuffle<3,3,3,3>        (rgba) ~> {A,A,A,A}
387 // The only real restriction is that the output also be a legal N=power-of-two sknx::Vec.
388 template <int... Ix, int N, typename T>
389 SI skvx::Vec<sizeof...(Ix),T> shuffle(skvx::Vec<N,T> x) {
390     return { x[Ix]... };
391 }
392 
393 #if !defined(SKNX_NO_SIMD)
394 namespace skvx {
395     // Platform-specific specializations and overloads can now drop in here.
396 
397 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE1
398     SI Vec<4,float> sqrt(Vec<4,float> x) {
399         return bit_pun<Vec<4,float>>(_mm_sqrt_ps(bit_pun<__m128>(x)));
400     }
401     SI Vec<4,float> rsqrt(Vec<4,float> x) {
402         return bit_pun<Vec<4,float>>(_mm_rsqrt_ps(bit_pun<__m128>(x)));
403     }
404     SI Vec<4,float> rcp(Vec<4,float> x) {
405         return bit_pun<Vec<4,float>>(_mm_rcp_ps(bit_pun<__m128>(x)));
406     }
407 
408     SI Vec<2,float>  sqrt(Vec<2,float> x) { return shuffle<0,1>( sqrt(shuffle<0,1,0,1>(x))); }
409     SI Vec<2,float> rsqrt(Vec<2,float> x) { return shuffle<0,1>(rsqrt(shuffle<0,1,0,1>(x))); }
410     SI Vec<2,float>   rcp(Vec<2,float> x) { return shuffle<0,1>(  rcp(shuffle<0,1,0,1>(x))); }
411 #endif
412 
413 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
414     SI Vec<4,float> if_then_else(Vec<4,int> c, Vec<4,float> t, Vec<4,float> e) {
415         return bit_pun<Vec<4,float>>(_mm_blendv_ps(bit_pun<__m128>(e),
416                                                    bit_pun<__m128>(t),
417                                                    bit_pun<__m128>(c)));
418     }
419 #elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE1
420     SI Vec<4,float> if_then_else(Vec<4,int> c, Vec<4,float> t, Vec<4,float> e) {
421         return bit_pun<Vec<4,float>>(_mm_or_ps(_mm_and_ps   (bit_pun<__m128>(c),
422                                                              bit_pun<__m128>(t)),
423                                                _mm_andnot_ps(bit_pun<__m128>(c),
424                                                              bit_pun<__m128>(e))));
425     }
426 #elif defined(SK_ARM_HAS_NEON)
427     SI Vec<4,float> if_then_else(Vec<4,int> c, Vec<4,float> t, Vec<4,float> e) {
428         return bit_pun<Vec<4,float>>(vbslq_f32(bit_pun<uint32x4_t> (c),
429                                                bit_pun<float32x4_t>(t),
430                                                bit_pun<float32x4_t>(e)));
431     }
432 #endif
433 
434 }  // namespace skvx
435 #endif  // !defined(SKNX_NO_SIMD)
436 
437 #undef SINT
438 #undef SIT
439 #undef SI
440 
441 #endif//SKVX_DEFINED
442