1 | // Copyright 2009-2021 Intel Corporation |
2 | // SPDX-License-Identifier: Apache-2.0 |
3 | |
4 | #pragma once |
5 | |
6 | #include "../math/math.h" |
7 | |
8 | #define vboolf vboolf_impl |
9 | #define vboold vboold_impl |
10 | #define vint vint_impl |
11 | #define vuint vuint_impl |
12 | #define vllong vllong_impl |
13 | #define vfloat vfloat_impl |
14 | #define vdouble vdouble_impl |
15 | |
16 | namespace embree |
17 | { |
18 | /* 4-wide SSE integer type */ |
19 | template<> |
20 | struct vint<4> |
21 | { |
22 | ALIGNED_STRUCT_(16); |
23 | |
24 | typedef vboolf4 Bool; |
25 | typedef vint4 Int; |
26 | typedef vfloat4 Float; |
27 | |
28 | enum { size = 4 }; // number of SIMD elements |
29 | union { __m128i v; int i[4]; }; // data |
30 | |
31 | //////////////////////////////////////////////////////////////////////////////// |
32 | /// Constructors, Assignment & Cast Operators |
33 | //////////////////////////////////////////////////////////////////////////////// |
34 | |
35 | __forceinline vint() {} |
36 | __forceinline vint(const vint4& a) { v = a.v; } |
37 | __forceinline vint4& operator =(const vint4& a) { v = a.v; return *this; } |
38 | |
39 | __forceinline vint(__m128i a) : v(a) {} |
40 | __forceinline operator const __m128i&() const { return v; } |
41 | __forceinline operator __m128i&() { return v; } |
42 | |
43 | __forceinline vint(int a) : v(_mm_set1_epi32(i: a)) {} |
44 | __forceinline vint(int a, int b, int c, int d) : v(_mm_set_epi32(i3: d, i2: c, i1: b, i0: a)) {} |
45 | |
46 | __forceinline explicit vint(__m128 a) : v(_mm_cvtps_epi32(a: a)) {} |
47 | #if defined(__AVX512VL__) |
48 | __forceinline explicit vint(const vboolf4& a) : v(_mm_movm_epi32(a)) {} |
49 | #else |
50 | __forceinline explicit vint(const vboolf4& a) : v(_mm_castps_si128(a: (__m128)a)) {} |
51 | #endif |
52 | |
53 | __forceinline vint(long long a, long long b) : v(_mm_set_epi64x(q1: b,q0: a)) {} |
54 | |
55 | //////////////////////////////////////////////////////////////////////////////// |
56 | /// Constants |
57 | //////////////////////////////////////////////////////////////////////////////// |
58 | |
59 | __forceinline vint(ZeroTy) : v(_mm_setzero_si128()) {} |
60 | __forceinline vint(OneTy) : v(_mm_set_epi32(i3: 1, i2: 1, i1: 1, i0: 1)) {} |
61 | __forceinline vint(PosInfTy) : v(_mm_set_epi32(i3: pos_inf, i2: pos_inf, i1: pos_inf, i0: pos_inf)) {} |
62 | __forceinline vint(NegInfTy) : v(_mm_set_epi32(i3: neg_inf, i2: neg_inf, i1: neg_inf, i0: neg_inf)) {} |
63 | __forceinline vint(StepTy) : v(_mm_set_epi32(i3: 3, i2: 2, i1: 1, i0: 0)) {} |
64 | __forceinline vint(ReverseStepTy) : v(_mm_set_epi32(i3: 0, i2: 1, i1: 2, i0: 3)) {} |
65 | |
66 | __forceinline vint(TrueTy) { v = _mm_cmpeq_epi32(a: v,b: v); } |
67 | __forceinline vint(UndefinedTy) : v(_mm_castps_si128(a: _mm_undefined_ps())) {} |
68 | |
69 | |
70 | //////////////////////////////////////////////////////////////////////////////// |
71 | /// Loads and Stores |
72 | //////////////////////////////////////////////////////////////////////////////// |
73 | |
74 | static __forceinline vint4 load (const void* a) { return _mm_load_si128(p: (__m128i*)a); } |
75 | static __forceinline vint4 loadu(const void* a) { return _mm_loadu_si128(p: (__m128i*)a); } |
76 | |
77 | static __forceinline void store (void* ptr, const vint4& v) { _mm_store_si128(p: (__m128i*)ptr,b: v); } |
78 | static __forceinline void storeu(void* ptr, const vint4& v) { _mm_storeu_si128(p: (__m128i*)ptr,b: v); } |
79 | |
80 | #if defined(__AVX512VL__) |
81 | |
82 | static __forceinline vint4 compact(const vboolf4& mask, vint4 &v) { |
83 | return _mm_mask_compress_epi32(v, mask, v); |
84 | } |
85 | static __forceinline vint4 compact(const vboolf4& mask, vint4 &a, const vint4& b) { |
86 | return _mm_mask_compress_epi32(a, mask, b); |
87 | } |
88 | |
89 | static __forceinline vint4 load (const vboolf4& mask, const void* ptr) { return _mm_mask_load_epi32 (_mm_setzero_si128(),mask,ptr); } |
90 | static __forceinline vint4 loadu(const vboolf4& mask, const void* ptr) { return _mm_mask_loadu_epi32(_mm_setzero_si128(),mask,ptr); } |
91 | |
92 | static __forceinline void store (const vboolf4& mask, void* ptr, const vint4& v) { _mm_mask_store_epi32 (ptr,mask,v); } |
93 | static __forceinline void storeu(const vboolf4& mask, void* ptr, const vint4& v) { _mm_mask_storeu_epi32(ptr,mask,v); } |
94 | #elif defined(__AVX__) |
95 | static __forceinline vint4 load (const vbool4& mask, const void* a) { return _mm_castps_si128(_mm_maskload_ps((float*)a,mask)); } |
96 | static __forceinline vint4 loadu(const vbool4& mask, const void* a) { return _mm_castps_si128(_mm_maskload_ps((float*)a,mask)); } |
97 | |
98 | static __forceinline void store (const vboolf4& mask, void* ptr, const vint4& i) { _mm_maskstore_ps((float*)ptr,(__m128i)mask,_mm_castsi128_ps(i)); } |
99 | static __forceinline void storeu(const vboolf4& mask, void* ptr, const vint4& i) { _mm_maskstore_ps((float*)ptr,(__m128i)mask,_mm_castsi128_ps(i)); } |
100 | #else |
101 | static __forceinline vint4 load (const vbool4& mask, const void* a) { return _mm_and_si128(a: _mm_load_si128 (p: (__m128i*)a),b: mask); } |
102 | static __forceinline vint4 loadu(const vbool4& mask, const void* a) { return _mm_and_si128(a: _mm_loadu_si128(p: (__m128i*)a),b: mask); } |
103 | |
104 | static __forceinline void store (const vboolf4& mask, void* ptr, const vint4& i) { store (ptr,v: select(m: mask,t: i,f: load (a: ptr))); } |
105 | static __forceinline void storeu(const vboolf4& mask, void* ptr, const vint4& i) { storeu(ptr,v: select(m: mask,t: i,f: loadu(a: ptr))); } |
106 | #endif |
107 | |
108 | |
109 | #if defined(__SSE4_1__) |
110 | static __forceinline vint4 load(const unsigned char* ptr) { |
111 | return _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr)); |
112 | } |
113 | |
114 | static __forceinline vint4 loadu(const unsigned char* ptr) { |
115 | return _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr)); |
116 | } |
117 | #else |
118 | |
119 | static __forceinline vint4 load(const unsigned char* ptr) { |
120 | return vint4(ptr[0],ptr[1],ptr[2],ptr[3]); |
121 | } |
122 | |
123 | static __forceinline vint4 loadu(const unsigned char* ptr) { |
124 | return vint4(ptr[0],ptr[1],ptr[2],ptr[3]); |
125 | } |
126 | |
127 | #endif |
128 | |
129 | static __forceinline vint4 load(const unsigned short* ptr) { |
130 | #if defined (__SSE4_1__) |
131 | return _mm_cvtepu16_epi32(_mm_loadu_si128((__m128i*)ptr)); |
132 | #else |
133 | return vint4(ptr[0],ptr[1],ptr[2],ptr[3]); |
134 | #endif |
135 | } |
136 | |
137 | static __forceinline void store(unsigned char* ptr, const vint4& v) { |
138 | #if defined(__SSE4_1__) |
139 | __m128i x = v; |
140 | x = _mm_packus_epi32(x, x); |
141 | x = _mm_packus_epi16(x, x); |
142 | *(int*)ptr = _mm_cvtsi128_si32(x); |
143 | #else |
144 | for (size_t i=0;i<4;i++) |
145 | ptr[i] = (unsigned char)v[i]; |
146 | #endif |
147 | } |
148 | |
149 | static __forceinline void store(unsigned short* ptr, const vint4& v) { |
150 | for (size_t i=0;i<4;i++) |
151 | ptr[i] = (unsigned short)v[i]; |
152 | } |
153 | |
154 | static __forceinline vint4 load_nt(void* ptr) { |
155 | #if defined(__SSE4_1__) |
156 | return _mm_stream_load_si128((__m128i*)ptr); |
157 | #else |
158 | return _mm_load_si128(p: (__m128i*)ptr); |
159 | #endif |
160 | } |
161 | |
162 | static __forceinline void store_nt(void* ptr, const vint4& v) { |
163 | #if defined(__SSE4_1__) |
164 | _mm_stream_ps((float*)ptr, _mm_castsi128_ps(v)); |
165 | #else |
166 | _mm_store_si128(p: (__m128i*)ptr,b: v); |
167 | #endif |
168 | } |
169 | |
170 | template<int scale = 4> |
171 | static __forceinline vint4 gather(const int* ptr, const vint4& index) { |
172 | #if defined(__AVX2__) |
173 | return _mm_i32gather_epi32(ptr, index, scale); |
174 | #else |
175 | return vint4( |
176 | *(int*)(((char*)ptr)+scale*index[0]), |
177 | *(int*)(((char*)ptr)+scale*index[1]), |
178 | *(int*)(((char*)ptr)+scale*index[2]), |
179 | *(int*)(((char*)ptr)+scale*index[3])); |
180 | #endif |
181 | } |
182 | |
183 | template<int scale = 4> |
184 | static __forceinline vint4 gather(const vboolf4& mask, const int* ptr, const vint4& index) { |
185 | vint4 r = zero; |
186 | #if defined(__AVX512VL__) |
187 | return _mm_mmask_i32gather_epi32(r, mask, index, ptr, scale); |
188 | #elif defined(__AVX2__) |
189 | return _mm_mask_i32gather_epi32(r, ptr, index, mask, scale); |
190 | #else |
191 | if (likely(mask[0])) r[0] = *(int*)(((char*)ptr)+scale*index[0]); |
192 | if (likely(mask[1])) r[1] = *(int*)(((char*)ptr)+scale*index[1]); |
193 | if (likely(mask[2])) r[2] = *(int*)(((char*)ptr)+scale*index[2]); |
194 | if (likely(mask[3])) r[3] = *(int*)(((char*)ptr)+scale*index[3]); |
195 | return r; |
196 | #endif |
197 | } |
198 | |
199 | template<int scale = 4> |
200 | static __forceinline void scatter(void* ptr, const vint4& index, const vint4& v) |
201 | { |
202 | #if defined(__AVX512VL__) |
203 | _mm_i32scatter_epi32((int*)ptr, index, v, scale); |
204 | #else |
205 | *(int*)(((char*)ptr)+scale*index[0]) = v[0]; |
206 | *(int*)(((char*)ptr)+scale*index[1]) = v[1]; |
207 | *(int*)(((char*)ptr)+scale*index[2]) = v[2]; |
208 | *(int*)(((char*)ptr)+scale*index[3]) = v[3]; |
209 | #endif |
210 | } |
211 | |
212 | template<int scale = 4> |
213 | static __forceinline void scatter(const vboolf4& mask, void* ptr, const vint4& index, const vint4& v) |
214 | { |
215 | #if defined(__AVX512VL__) |
216 | _mm_mask_i32scatter_epi32((int*)ptr, mask, index, v, scale); |
217 | #else |
218 | if (likely(mask[0])) *(int*)(((char*)ptr)+scale*index[0]) = v[0]; |
219 | if (likely(mask[1])) *(int*)(((char*)ptr)+scale*index[1]) = v[1]; |
220 | if (likely(mask[2])) *(int*)(((char*)ptr)+scale*index[2]) = v[2]; |
221 | if (likely(mask[3])) *(int*)(((char*)ptr)+scale*index[3]) = v[3]; |
222 | #endif |
223 | } |
224 | |
225 | #if defined(__x86_64__) |
226 | static __forceinline vint4 broadcast64(long long a) { return _mm_set1_epi64x(q: a); } |
227 | #endif |
228 | |
229 | //////////////////////////////////////////////////////////////////////////////// |
230 | /// Array Access |
231 | //////////////////////////////////////////////////////////////////////////////// |
232 | |
233 | __forceinline const int& operator [](size_t index) const { assert(index < 4); return i[index]; } |
234 | __forceinline int& operator [](size_t index) { assert(index < 4); return i[index]; } |
235 | |
236 | friend __forceinline vint4 select(const vboolf4& m, const vint4& t, const vint4& f) { |
237 | #if defined(__AVX512VL__) |
238 | return _mm_mask_blend_epi32(m, (__m128i)f, (__m128i)t); |
239 | #elif defined(__SSE4_1__) |
240 | return _mm_castps_si128(_mm_blendv_ps(_mm_castsi128_ps(f), _mm_castsi128_ps(t), m)); |
241 | #else |
242 | return _mm_or_si128(a: _mm_and_si128(a: m, b: t), b: _mm_andnot_si128(a: m, b: f)); |
243 | #endif |
244 | } |
245 | }; |
246 | |
247 | //////////////////////////////////////////////////////////////////////////////// |
248 | /// Unary Operators |
249 | //////////////////////////////////////////////////////////////////////////////// |
250 | |
251 | #if defined(__AVX512VL__) |
252 | __forceinline vboolf4 asBool(const vint4& a) { return _mm_movepi32_mask(a); } |
253 | #else |
254 | __forceinline vboolf4 asBool(const vint4& a) { return _mm_castsi128_ps(a: a); } |
255 | #endif |
256 | |
257 | __forceinline vint4 operator +(const vint4& a) { return a; } |
258 | __forceinline vint4 operator -(const vint4& a) { return _mm_sub_epi32(a: _mm_setzero_si128(), b: a); } |
259 | #if defined(__SSSE3__) |
260 | __forceinline vint4 abs(const vint4& a) { return _mm_abs_epi32(a); } |
261 | #endif |
262 | |
263 | //////////////////////////////////////////////////////////////////////////////// |
264 | /// Binary Operators |
265 | //////////////////////////////////////////////////////////////////////////////// |
266 | |
267 | __forceinline vint4 operator +(const vint4& a, const vint4& b) { return _mm_add_epi32(a: a, b: b); } |
268 | __forceinline vint4 operator +(const vint4& a, int b) { return a + vint4(b); } |
269 | __forceinline vint4 operator +(int a, const vint4& b) { return vint4(a) + b; } |
270 | |
271 | __forceinline vint4 operator -(const vint4& a, const vint4& b) { return _mm_sub_epi32(a: a, b: b); } |
272 | __forceinline vint4 operator -(const vint4& a, int b) { return a - vint4(b); } |
273 | __forceinline vint4 operator -(int a, const vint4& b) { return vint4(a) - b; } |
274 | |
275 | #if defined(__SSE4_1__) |
276 | __forceinline vint4 operator *(const vint4& a, const vint4& b) { return _mm_mullo_epi32(a, b); } |
277 | #else |
278 | __forceinline vint4 operator *(const vint4& a, const vint4& b) { return vint4(a[0]*b[0],a[1]*b[1],a[2]*b[2],a[3]*b[3]); } |
279 | #endif |
280 | __forceinline vint4 operator *(const vint4& a, int b) { return a * vint4(b); } |
281 | __forceinline vint4 operator *(int a, const vint4& b) { return vint4(a) * b; } |
282 | |
283 | __forceinline vint4 operator &(const vint4& a, const vint4& b) { return _mm_and_si128(a: a, b: b); } |
284 | __forceinline vint4 operator &(const vint4& a, int b) { return a & vint4(b); } |
285 | __forceinline vint4 operator &(int a, const vint4& b) { return vint4(a) & b; } |
286 | |
287 | __forceinline vint4 operator |(const vint4& a, const vint4& b) { return _mm_or_si128(a: a, b: b); } |
288 | __forceinline vint4 operator |(const vint4& a, int b) { return a | vint4(b); } |
289 | __forceinline vint4 operator |(int a, const vint4& b) { return vint4(a) | b; } |
290 | |
291 | __forceinline vint4 operator ^(const vint4& a, const vint4& b) { return _mm_xor_si128(a: a, b: b); } |
292 | __forceinline vint4 operator ^(const vint4& a, int b) { return a ^ vint4(b); } |
293 | __forceinline vint4 operator ^(int a, const vint4& b) { return vint4(a) ^ b; } |
294 | |
295 | __forceinline vint4 operator <<(const vint4& a, int n) { return _mm_slli_epi32(a: a, count: n); } |
296 | __forceinline vint4 operator >>(const vint4& a, int n) { return _mm_srai_epi32(a: a, count: n); } |
297 | |
298 | __forceinline vint4 sll (const vint4& a, int b) { return _mm_slli_epi32(a: a, count: b); } |
299 | __forceinline vint4 sra (const vint4& a, int b) { return _mm_srai_epi32(a: a, count: b); } |
300 | __forceinline vint4 srl (const vint4& a, int b) { return _mm_srli_epi32(a: a, count: b); } |
301 | |
302 | //////////////////////////////////////////////////////////////////////////////// |
303 | /// Assignment Operators |
304 | //////////////////////////////////////////////////////////////////////////////// |
305 | |
306 | __forceinline vint4& operator +=(vint4& a, const vint4& b) { return a = a + b; } |
307 | __forceinline vint4& operator +=(vint4& a, int b) { return a = a + b; } |
308 | |
309 | __forceinline vint4& operator -=(vint4& a, const vint4& b) { return a = a - b; } |
310 | __forceinline vint4& operator -=(vint4& a, int b) { return a = a - b; } |
311 | |
312 | #if defined(__SSE4_1__) |
313 | __forceinline vint4& operator *=(vint4& a, const vint4& b) { return a = a * b; } |
314 | __forceinline vint4& operator *=(vint4& a, int b) { return a = a * b; } |
315 | #endif |
316 | |
317 | __forceinline vint4& operator &=(vint4& a, const vint4& b) { return a = a & b; } |
318 | __forceinline vint4& operator &=(vint4& a, int b) { return a = a & b; } |
319 | |
320 | __forceinline vint4& operator |=(vint4& a, const vint4& b) { return a = a | b; } |
321 | __forceinline vint4& operator |=(vint4& a, int b) { return a = a | b; } |
322 | |
323 | __forceinline vint4& operator <<=(vint4& a, int b) { return a = a << b; } |
324 | __forceinline vint4& operator >>=(vint4& a, int b) { return a = a >> b; } |
325 | |
326 | //////////////////////////////////////////////////////////////////////////////// |
327 | /// Comparison Operators + Select |
328 | //////////////////////////////////////////////////////////////////////////////// |
329 | |
330 | #if defined(__AVX512VL__) |
331 | __forceinline vboolf4 operator ==(const vint4& a, const vint4& b) { return _mm_cmp_epi32_mask(a,b,_MM_CMPINT_EQ); } |
332 | __forceinline vboolf4 operator !=(const vint4& a, const vint4& b) { return _mm_cmp_epi32_mask(a,b,_MM_CMPINT_NE); } |
333 | __forceinline vboolf4 operator < (const vint4& a, const vint4& b) { return _mm_cmp_epi32_mask(a,b,_MM_CMPINT_LT); } |
334 | __forceinline vboolf4 operator >=(const vint4& a, const vint4& b) { return _mm_cmp_epi32_mask(a,b,_MM_CMPINT_GE); } |
335 | __forceinline vboolf4 operator > (const vint4& a, const vint4& b) { return _mm_cmp_epi32_mask(a,b,_MM_CMPINT_GT); } |
336 | __forceinline vboolf4 operator <=(const vint4& a, const vint4& b) { return _mm_cmp_epi32_mask(a,b,_MM_CMPINT_LE); } |
337 | #else |
338 | __forceinline vboolf4 operator ==(const vint4& a, const vint4& b) { return _mm_castsi128_ps(a: _mm_cmpeq_epi32(a: a, b: b)); } |
339 | __forceinline vboolf4 operator !=(const vint4& a, const vint4& b) { return !(a == b); } |
340 | __forceinline vboolf4 operator < (const vint4& a, const vint4& b) { return _mm_castsi128_ps(a: _mm_cmplt_epi32(a: a, b: b)); } |
341 | __forceinline vboolf4 operator >=(const vint4& a, const vint4& b) { return !(a < b); } |
342 | __forceinline vboolf4 operator > (const vint4& a, const vint4& b) { return _mm_castsi128_ps(a: _mm_cmpgt_epi32(a: a, b: b)); } |
343 | __forceinline vboolf4 operator <=(const vint4& a, const vint4& b) { return !(a > b); } |
344 | #endif |
345 | |
346 | __forceinline vboolf4 operator ==(const vint4& a, int b) { return a == vint4(b); } |
347 | __forceinline vboolf4 operator ==(int a, const vint4& b) { return vint4(a) == b; } |
348 | |
349 | __forceinline vboolf4 operator !=(const vint4& a, int b) { return a != vint4(b); } |
350 | __forceinline vboolf4 operator !=(int a, const vint4& b) { return vint4(a) != b; } |
351 | |
352 | __forceinline vboolf4 operator < (const vint4& a, int b) { return a < vint4(b); } |
353 | __forceinline vboolf4 operator < (int a, const vint4& b) { return vint4(a) < b; } |
354 | |
355 | __forceinline vboolf4 operator >=(const vint4& a, int b) { return a >= vint4(b); } |
356 | __forceinline vboolf4 operator >=(int a, const vint4& b) { return vint4(a) >= b; } |
357 | |
358 | __forceinline vboolf4 operator > (const vint4& a, int b) { return a > vint4(b); } |
359 | __forceinline vboolf4 operator > (int a, const vint4& b) { return vint4(a) > b; } |
360 | |
361 | __forceinline vboolf4 operator <=(const vint4& a, int b) { return a <= vint4(b); } |
362 | __forceinline vboolf4 operator <=(int a, const vint4& b) { return vint4(a) <= b; } |
363 | |
364 | __forceinline vboolf4 eq(const vint4& a, const vint4& b) { return a == b; } |
365 | __forceinline vboolf4 ne(const vint4& a, const vint4& b) { return a != b; } |
366 | __forceinline vboolf4 lt(const vint4& a, const vint4& b) { return a < b; } |
367 | __forceinline vboolf4 ge(const vint4& a, const vint4& b) { return a >= b; } |
368 | __forceinline vboolf4 gt(const vint4& a, const vint4& b) { return a > b; } |
369 | __forceinline vboolf4 le(const vint4& a, const vint4& b) { return a <= b; } |
370 | |
371 | #if defined(__AVX512VL__) |
372 | __forceinline vboolf4 eq(const vboolf4& mask, const vint4& a, const vint4& b) { return _mm_mask_cmp_epi32_mask(mask, a, b, _MM_CMPINT_EQ); } |
373 | __forceinline vboolf4 ne(const vboolf4& mask, const vint4& a, const vint4& b) { return _mm_mask_cmp_epi32_mask(mask, a, b, _MM_CMPINT_NE); } |
374 | __forceinline vboolf4 lt(const vboolf4& mask, const vint4& a, const vint4& b) { return _mm_mask_cmp_epi32_mask(mask, a, b, _MM_CMPINT_LT); } |
375 | __forceinline vboolf4 ge(const vboolf4& mask, const vint4& a, const vint4& b) { return _mm_mask_cmp_epi32_mask(mask, a, b, _MM_CMPINT_GE); } |
376 | __forceinline vboolf4 gt(const vboolf4& mask, const vint4& a, const vint4& b) { return _mm_mask_cmp_epi32_mask(mask, a, b, _MM_CMPINT_GT); } |
377 | __forceinline vboolf4 le(const vboolf4& mask, const vint4& a, const vint4& b) { return _mm_mask_cmp_epi32_mask(mask, a, b, _MM_CMPINT_LE); } |
378 | #else |
379 | __forceinline vboolf4 eq(const vboolf4& mask, const vint4& a, const vint4& b) { return mask & (a == b); } |
380 | __forceinline vboolf4 ne(const vboolf4& mask, const vint4& a, const vint4& b) { return mask & (a != b); } |
381 | __forceinline vboolf4 lt(const vboolf4& mask, const vint4& a, const vint4& b) { return mask & (a < b); } |
382 | __forceinline vboolf4 ge(const vboolf4& mask, const vint4& a, const vint4& b) { return mask & (a >= b); } |
383 | __forceinline vboolf4 gt(const vboolf4& mask, const vint4& a, const vint4& b) { return mask & (a > b); } |
384 | __forceinline vboolf4 le(const vboolf4& mask, const vint4& a, const vint4& b) { return mask & (a <= b); } |
385 | #endif |
386 | |
387 | template<int mask> |
388 | __forceinline vint4 select(const vint4& t, const vint4& f) { |
389 | #if defined(__SSE4_1__) |
390 | return _mm_castps_si128(_mm_blend_ps(_mm_castsi128_ps(f), _mm_castsi128_ps(t), mask)); |
391 | #else |
392 | return select(m: vboolf4(mask), t, f); |
393 | #endif |
394 | } |
395 | |
396 | #if defined(__SSE4_1__) |
397 | __forceinline vint4 min(const vint4& a, const vint4& b) { return _mm_min_epi32(a, b); } |
398 | __forceinline vint4 max(const vint4& a, const vint4& b) { return _mm_max_epi32(a, b); } |
399 | |
400 | __forceinline vint4 umin(const vint4& a, const vint4& b) { return _mm_min_epu32(a, b); } |
401 | __forceinline vint4 umax(const vint4& a, const vint4& b) { return _mm_max_epu32(a, b); } |
402 | |
403 | #else |
404 | __forceinline vint4 min(const vint4& a, const vint4& b) { return select(m: a < b,t: a,f: b); } |
405 | __forceinline vint4 max(const vint4& a, const vint4& b) { return select(m: a < b,t: b,f: a); } |
406 | #endif |
407 | |
408 | __forceinline vint4 min(const vint4& a, int b) { return min(a,b: vint4(b)); } |
409 | __forceinline vint4 min(int a, const vint4& b) { return min(a: vint4(a),b); } |
410 | __forceinline vint4 max(const vint4& a, int b) { return max(a,b: vint4(b)); } |
411 | __forceinline vint4 max(int a, const vint4& b) { return max(a: vint4(a),b); } |
412 | |
413 | //////////////////////////////////////////////////////////////////////////////// |
414 | // Movement/Shifting/Shuffling Functions |
415 | //////////////////////////////////////////////////////////////////////////////// |
416 | |
417 | __forceinline vint4 unpacklo(const vint4& a, const vint4& b) { return _mm_castps_si128(a: _mm_unpacklo_ps(a: _mm_castsi128_ps(a: a), b: _mm_castsi128_ps(a: b))); } |
418 | __forceinline vint4 unpackhi(const vint4& a, const vint4& b) { return _mm_castps_si128(a: _mm_unpackhi_ps(a: _mm_castsi128_ps(a: a), b: _mm_castsi128_ps(a: b))); } |
419 | |
420 | template<int i0, int i1, int i2, int i3> |
421 | __forceinline vint4 shuffle(const vint4& v) { |
422 | return _mm_shuffle_epi32(v, _MM_SHUFFLE(i3, i2, i1, i0)); |
423 | } |
424 | |
425 | template<int i0, int i1, int i2, int i3> |
426 | __forceinline vint4 shuffle(const vint4& a, const vint4& b) { |
427 | return _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), _MM_SHUFFLE(i3, i2, i1, i0))); |
428 | } |
429 | |
430 | #if defined(__SSE3__) |
431 | template<> __forceinline vint4 shuffle<0, 0, 2, 2>(const vint4& v) { return _mm_castps_si128(_mm_moveldup_ps(_mm_castsi128_ps(v))); } |
432 | template<> __forceinline vint4 shuffle<1, 1, 3, 3>(const vint4& v) { return _mm_castps_si128(_mm_movehdup_ps(_mm_castsi128_ps(v))); } |
433 | template<> __forceinline vint4 shuffle<0, 1, 0, 1>(const vint4& v) { return _mm_castpd_si128(_mm_movedup_pd (_mm_castsi128_pd(v))); } |
434 | #endif |
435 | |
436 | template<int i> |
437 | __forceinline vint4 shuffle(const vint4& v) { |
438 | return shuffle<i,i,i,i>(v); |
439 | } |
440 | |
441 | #if defined(__SSE4_1__) |
442 | template<int src> __forceinline int extract(const vint4& b) { return _mm_extract_epi32(b, src); } |
443 | template<int dst> __forceinline vint4 insert(const vint4& a, const int b) { return _mm_insert_epi32(a, b, dst); } |
444 | #else |
445 | template<int src> __forceinline int (const vint4& b) { return b[src&3]; } |
446 | template<int dst> __forceinline vint4 insert(const vint4& a, int b) { vint4 c = a; c[dst&3] = b; return c; } |
447 | #endif |
448 | |
449 | |
450 | template<> __forceinline int <0>(const vint4& b) { return _mm_cvtsi128_si32(a: b); } |
451 | |
452 | __forceinline int toScalar(const vint4& v) { return _mm_cvtsi128_si32(a: v); } |
453 | |
454 | __forceinline size_t toSizeT(const vint4& v) { |
455 | #if defined(__WIN32__) && !defined(__X86_64__) // win32 workaround |
456 | return toScalar(v); |
457 | #else |
458 | return _mm_cvtsi128_si64(a: v); |
459 | #endif |
460 | } |
461 | |
462 | #if defined(__AVX512VL__) |
463 | |
464 | __forceinline vint4 permute(const vint4 &a, const vint4 &index) { |
465 | return _mm_castps_si128(_mm_permutevar_ps(_mm_castsi128_ps(a),index)); |
466 | } |
467 | |
468 | template<int i> |
469 | __forceinline vint4 align_shift_right(const vint4& a, const vint4& b) { |
470 | return _mm_alignr_epi32(a, b, i); |
471 | } |
472 | #endif |
473 | |
474 | //////////////////////////////////////////////////////////////////////////////// |
475 | /// Reductions |
476 | //////////////////////////////////////////////////////////////////////////////// |
477 | |
478 | #if defined(__SSE4_1__) |
479 | __forceinline vint4 vreduce_min(const vint4& v) { vint4 h = min(shuffle<1,0,3,2>(v),v); return min(shuffle<2,3,0,1>(h),h); } |
480 | __forceinline vint4 vreduce_max(const vint4& v) { vint4 h = max(shuffle<1,0,3,2>(v),v); return max(shuffle<2,3,0,1>(h),h); } |
481 | __forceinline vint4 vreduce_add(const vint4& v) { vint4 h = shuffle<1,0,3,2>(v) + v ; return shuffle<2,3,0,1>(h) + h ; } |
482 | |
483 | __forceinline int reduce_min(const vint4& v) { return toScalar(vreduce_min(v)); } |
484 | __forceinline int reduce_max(const vint4& v) { return toScalar(vreduce_max(v)); } |
485 | __forceinline int reduce_add(const vint4& v) { return toScalar(vreduce_add(v)); } |
486 | |
487 | __forceinline size_t select_min(const vint4& v) { return bsf(movemask(v == vreduce_min(v))); } |
488 | __forceinline size_t select_max(const vint4& v) { return bsf(movemask(v == vreduce_max(v))); } |
489 | |
490 | __forceinline size_t select_min(const vboolf4& valid, const vint4& v) { const vint4 a = select(valid,v,vint4(pos_inf)); return bsf(movemask(valid & (a == vreduce_min(a)))); } |
491 | __forceinline size_t select_max(const vboolf4& valid, const vint4& v) { const vint4 a = select(valid,v,vint4(neg_inf)); return bsf(movemask(valid & (a == vreduce_max(a)))); } |
492 | |
493 | #else |
494 | |
495 | __forceinline int reduce_min(const vint4& v) { return min(a: v[0],b: v[1],c: v[2],d: v[3]); } |
496 | __forceinline int reduce_max(const vint4& v) { return max(a: v[0],b: v[1],c: v[2],d: v[3]); } |
497 | __forceinline int reduce_add(const vint4& v) { return v[0]+v[1]+v[2]+v[3]; } |
498 | |
499 | #endif |
500 | |
501 | //////////////////////////////////////////////////////////////////////////////// |
502 | /// Sorting networks |
503 | //////////////////////////////////////////////////////////////////////////////// |
504 | |
505 | #if defined(__SSE4_1__) |
506 | |
507 | __forceinline vint4 usort_ascending(const vint4& v) |
508 | { |
509 | const vint4 a0 = v; |
510 | const vint4 b0 = shuffle<1,0,3,2>(a0); |
511 | const vint4 c0 = umin(a0,b0); |
512 | const vint4 d0 = umax(a0,b0); |
513 | const vint4 a1 = select<0x5 /* 0b0101 */>(c0,d0); |
514 | const vint4 b1 = shuffle<2,3,0,1>(a1); |
515 | const vint4 c1 = umin(a1,b1); |
516 | const vint4 d1 = umax(a1,b1); |
517 | const vint4 a2 = select<0x3 /* 0b0011 */>(c1,d1); |
518 | const vint4 b2 = shuffle<0,2,1,3>(a2); |
519 | const vint4 c2 = umin(a2,b2); |
520 | const vint4 d2 = umax(a2,b2); |
521 | const vint4 a3 = select<0x2 /* 0b0010 */>(c2,d2); |
522 | return a3; |
523 | } |
524 | |
525 | __forceinline vint4 usort_descending(const vint4& v) |
526 | { |
527 | const vint4 a0 = v; |
528 | const vint4 b0 = shuffle<1,0,3,2>(a0); |
529 | const vint4 c0 = umax(a0,b0); |
530 | const vint4 d0 = umin(a0,b0); |
531 | const vint4 a1 = select<0x5 /* 0b0101 */>(c0,d0); |
532 | const vint4 b1 = shuffle<2,3,0,1>(a1); |
533 | const vint4 c1 = umax(a1,b1); |
534 | const vint4 d1 = umin(a1,b1); |
535 | const vint4 a2 = select<0x3 /* 0b0011 */>(c1,d1); |
536 | const vint4 b2 = shuffle<0,2,1,3>(a2); |
537 | const vint4 c2 = umax(a2,b2); |
538 | const vint4 d2 = umin(a2,b2); |
539 | const vint4 a3 = select<0x2 /* 0b0010 */>(c2,d2); |
540 | return a3; |
541 | } |
542 | |
543 | #else |
544 | |
545 | __forceinline vint4 usort_ascending(const vint4& v) |
546 | { |
547 | const vint4 a0 = v-vint4(0x80000000); |
548 | const vint4 b0 = shuffle<1,0,3,2>(v: a0); |
549 | const vint4 c0 = min(a: a0,b: b0); |
550 | const vint4 d0 = max(a: a0,b: b0); |
551 | const vint4 a1 = select<0x5 /* 0b0101 */>(t: c0,f: d0); |
552 | const vint4 b1 = shuffle<2,3,0,1>(v: a1); |
553 | const vint4 c1 = min(a: a1,b: b1); |
554 | const vint4 d1 = max(a: a1,b: b1); |
555 | const vint4 a2 = select<0x3 /* 0b0011 */>(t: c1,f: d1); |
556 | const vint4 b2 = shuffle<0,2,1,3>(v: a2); |
557 | const vint4 c2 = min(a: a2,b: b2); |
558 | const vint4 d2 = max(a: a2,b: b2); |
559 | const vint4 a3 = select<0x2 /* 0b0010 */>(t: c2,f: d2); |
560 | return a3+vint4(0x80000000); |
561 | } |
562 | |
563 | __forceinline vint4 usort_descending(const vint4& v) |
564 | { |
565 | const vint4 a0 = v-vint4(0x80000000); |
566 | const vint4 b0 = shuffle<1,0,3,2>(v: a0); |
567 | const vint4 c0 = max(a: a0,b: b0); |
568 | const vint4 d0 = min(a: a0,b: b0); |
569 | const vint4 a1 = select<0x5 /* 0b0101 */>(t: c0,f: d0); |
570 | const vint4 b1 = shuffle<2,3,0,1>(v: a1); |
571 | const vint4 c1 = max(a: a1,b: b1); |
572 | const vint4 d1 = min(a: a1,b: b1); |
573 | const vint4 a2 = select<0x3 /* 0b0011 */>(t: c1,f: d1); |
574 | const vint4 b2 = shuffle<0,2,1,3>(v: a2); |
575 | const vint4 c2 = max(a: a2,b: b2); |
576 | const vint4 d2 = min(a: a2,b: b2); |
577 | const vint4 a3 = select<0x2 /* 0b0010 */>(t: c2,f: d2); |
578 | return a3+vint4(0x80000000); |
579 | } |
580 | |
581 | #endif |
582 | |
583 | //////////////////////////////////////////////////////////////////////////////// |
584 | /// Output Operators |
585 | //////////////////////////////////////////////////////////////////////////////// |
586 | |
587 | __forceinline embree_ostream operator <<(embree_ostream cout, const vint4& a) { |
588 | return cout << "<" << a[0] << ", " << a[1] << ", " << a[2] << ", " << a[3] << ">" ; |
589 | } |
590 | } |
591 | |
592 | #undef vboolf |
593 | #undef vboold |
594 | #undef vint |
595 | #undef vuint |
596 | #undef vllong |
597 | #undef vfloat |
598 | #undef vdouble |
599 | |