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
16namespace 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(a)) {}
44 __forceinline vint(int a, int b, int c, int d) : v(_mm_set_epi32(d, c, b, a)) {}
45
46 __forceinline explicit vint(__m128 a) : v(_mm_cvtps_epi32(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((__m128)a)) {}
51#endif
52
53 __forceinline vint(long long a, long long b) : v(_mm_set_epi64x(b,a)) {}
54
55 ////////////////////////////////////////////////////////////////////////////////
56 /// Constants
57 ////////////////////////////////////////////////////////////////////////////////
58
59 __forceinline vint(ZeroTy) : v(_mm_setzero_si128()) {}
60 __forceinline vint(OneTy) : v(_mm_set_epi32(1, 1, 1, 1)) {}
61 __forceinline vint(PosInfTy) : v(_mm_set_epi32(pos_inf, pos_inf, pos_inf, pos_inf)) {}
62 __forceinline vint(NegInfTy) : v(_mm_set_epi32(neg_inf, neg_inf, neg_inf, neg_inf)) {}
63 __forceinline vint(StepTy) : v(_mm_set_epi32(3, 2, 1, 0)) {}
64 __forceinline vint(ReverseStepTy) : v(_mm_set_epi32(0, 1, 2, 3)) {}
65
66 __forceinline vint(TrueTy) { v = _mm_cmpeq_epi32(v,v); }
67 __forceinline vint(UndefinedTy) : v(_mm_castps_si128(_mm_undefined_ps())) {}
68
69
70 ////////////////////////////////////////////////////////////////////////////////
71 /// Loads and Stores
72 ////////////////////////////////////////////////////////////////////////////////
73
74 static __forceinline vint4 load (const void* a) { return _mm_load_si128((__m128i*)a); }
75 static __forceinline vint4 loadu(const void* a) { return _mm_loadu_si128((__m128i*)a); }
76
77 static __forceinline void store (void* ptr, const vint4& v) { _mm_store_si128((__m128i*)ptr,v); }
78 static __forceinline void storeu(void* ptr, const vint4& v) { _mm_storeu_si128((__m128i*)ptr,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(_mm_load_si128 ((__m128i*)a),mask); }
102 static __forceinline vint4 loadu(const vbool4& mask, const void* a) { return _mm_and_si128(_mm_loadu_si128((__m128i*)a),mask); }
103
104 static __forceinline void store (const vboolf4& mask, void* ptr, const vint4& i) { store (ptr,select(mask,i,load (ptr))); }
105 static __forceinline void storeu(const vboolf4& mask, void* ptr, const vint4& i) { storeu(ptr,select(mask,i,loadu(ptr))); }
106#endif
107
108
109#if defined(__aarch64__)
110 static __forceinline vint4 load(const unsigned char* ptr) {
111 return _mm_load4epu8_epi32(((__m128i*)ptr));
112 }
113 static __forceinline vint4 loadu(const unsigned char* ptr) {
114 return _mm_load4epu8_epi32(((__m128i*)ptr));
115 }
116#elif defined(__SSE4_1__)
117 static __forceinline vint4 load(const unsigned char* ptr) {
118 return _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr));
119 }
120
121 static __forceinline vint4 loadu(const unsigned char* ptr) {
122 return _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr));
123 }
124#else
125
126 static __forceinline vint4 load(const unsigned char* ptr) {
127 return vint4(ptr[0],ptr[1],ptr[2],ptr[3]);
128 }
129
130 static __forceinline vint4 loadu(const unsigned char* ptr) {
131 return vint4(ptr[0],ptr[1],ptr[2],ptr[3]);
132 }
133
134#endif
135
136 static __forceinline vint4 load(const unsigned short* ptr) {
137#if defined(__aarch64__)
138 return __m128i(vmovl_u16(vld1_u16(ptr)));
139#elif defined (__SSE4_1__)
140 return _mm_cvtepu16_epi32(_mm_loadu_si128((__m128i*)ptr));
141#else
142 return vint4(ptr[0],ptr[1],ptr[2],ptr[3]);
143#endif
144 }
145
146 static __forceinline void store(unsigned char* ptr, const vint4& v) {
147#if defined(__aarch64__)
148 int32x4_t x = v;
149 uint16x4_t y = vqmovn_u32(uint32x4_t(x));
150 uint8x8_t z = vqmovn_u16(vcombine_u16(y, y));
151 vst1_lane_u32((uint32_t *)ptr,uint32x2_t(z), 0);
152#elif defined(__SSE4_1__)
153 __m128i x = v;
154 x = _mm_packus_epi32(x, x);
155 x = _mm_packus_epi16(x, x);
156 *(int*)ptr = _mm_cvtsi128_si32(x);
157#else
158 for (size_t i=0;i<4;i++)
159 ptr[i] = (unsigned char)v[i];
160#endif
161 }
162
163 static __forceinline void store(unsigned short* ptr, const vint4& v) {
164#if defined(__aarch64__)
165 uint32x4_t x = uint32x4_t(v.v);
166 uint16x4_t y = vqmovn_u32(x);
167 vst1_u16(ptr, y);
168#else
169 for (size_t i=0;i<4;i++)
170 ptr[i] = (unsigned short)v[i];
171#endif
172 }
173
174 static __forceinline vint4 load_nt(void* ptr) {
175#if defined(__aarch64__) || defined(__SSE4_1__)
176 return _mm_stream_load_si128((__m128i*)ptr);
177#else
178 return _mm_load_si128((__m128i*)ptr);
179#endif
180 }
181
182 static __forceinline void store_nt(void* ptr, const vint4& v) {
183#if !defined(__aarch64__) && defined(__SSE4_1__)
184 _mm_stream_ps((float*)ptr, _mm_castsi128_ps(v));
185#else
186 _mm_store_si128((__m128i*)ptr,v);
187#endif
188 }
189
190 template<int scale = 4>
191 static __forceinline vint4 gather(const int* ptr, const vint4& index) {
192#if defined(__AVX2__) && !defined(__aarch64__)
193 return _mm_i32gather_epi32(ptr, index, scale);
194#else
195 return vint4(
196 *(int*)(((char*)ptr)+scale*index[0]),
197 *(int*)(((char*)ptr)+scale*index[1]),
198 *(int*)(((char*)ptr)+scale*index[2]),
199 *(int*)(((char*)ptr)+scale*index[3]));
200#endif
201 }
202
203 template<int scale = 4>
204 static __forceinline vint4 gather(const vboolf4& mask, const int* ptr, const vint4& index) {
205 vint4 r = zero;
206#if defined(__AVX512VL__)
207 return _mm_mmask_i32gather_epi32(r, mask, index, ptr, scale);
208#elif defined(__AVX2__) && !defined(__aarch64__)
209 return _mm_mask_i32gather_epi32(r, ptr, index, mask, scale);
210#else
211 if (likely(mask[0])) r[0] = *(int*)(((char*)ptr)+scale*index[0]);
212 if (likely(mask[1])) r[1] = *(int*)(((char*)ptr)+scale*index[1]);
213 if (likely(mask[2])) r[2] = *(int*)(((char*)ptr)+scale*index[2]);
214 if (likely(mask[3])) r[3] = *(int*)(((char*)ptr)+scale*index[3]);
215 return r;
216#endif
217 }
218
219 template<int scale = 4>
220 static __forceinline void scatter(void* ptr, const vint4& index, const vint4& v)
221 {
222#if defined(__AVX512VL__)
223 _mm_i32scatter_epi32((int*)ptr, index, v, scale);
224#else
225 *(int*)(((char*)ptr)+scale*index[0]) = v[0];
226 *(int*)(((char*)ptr)+scale*index[1]) = v[1];
227 *(int*)(((char*)ptr)+scale*index[2]) = v[2];
228 *(int*)(((char*)ptr)+scale*index[3]) = v[3];
229#endif
230 }
231
232 template<int scale = 4>
233 static __forceinline void scatter(const vboolf4& mask, void* ptr, const vint4& index, const vint4& v)
234 {
235#if defined(__AVX512VL__)
236 _mm_mask_i32scatter_epi32((int*)ptr, mask, index, v, scale);
237#else
238 if (likely(mask[0])) *(int*)(((char*)ptr)+scale*index[0]) = v[0];
239 if (likely(mask[1])) *(int*)(((char*)ptr)+scale*index[1]) = v[1];
240 if (likely(mask[2])) *(int*)(((char*)ptr)+scale*index[2]) = v[2];
241 if (likely(mask[3])) *(int*)(((char*)ptr)+scale*index[3]) = v[3];
242#endif
243 }
244
245#if defined(__x86_64__) || defined(__aarch64__)
246 static __forceinline vint4 broadcast64(long long a) { return _mm_set1_epi64x(a); }
247#endif
248
249 ////////////////////////////////////////////////////////////////////////////////
250 /// Array Access
251 ////////////////////////////////////////////////////////////////////////////////
252
253 __forceinline const int& operator [](size_t index) const { assert(index < 4); return i[index]; }
254 __forceinline int& operator [](size_t index) { assert(index < 4); return i[index]; }
255
256 friend __forceinline vint4 select(const vboolf4& m, const vint4& t, const vint4& f) {
257#if defined(__AVX512VL__)
258 return _mm_mask_blend_epi32(m, (__m128i)f, (__m128i)t);
259#elif defined(__aarch64__)
260 return _mm_castps_si128(_mm_blendv_ps((__m128)f.v,(__m128) t.v, (__m128)m.v));
261#elif defined(__SSE4_1__)
262 return _mm_castps_si128(_mm_blendv_ps(_mm_castsi128_ps(f), _mm_castsi128_ps(t), m));
263#else
264 return _mm_or_si128(_mm_and_si128(m, t), _mm_andnot_si128(m, f));
265#endif
266 }
267 };
268
269 ////////////////////////////////////////////////////////////////////////////////
270 /// Unary Operators
271 ////////////////////////////////////////////////////////////////////////////////
272
273#if defined(__AVX512VL__)
274 __forceinline vboolf4 asBool(const vint4& a) { return _mm_movepi32_mask(a); }
275#else
276 __forceinline vboolf4 asBool(const vint4& a) { return _mm_castsi128_ps(a); }
277#endif
278
279 __forceinline vint4 operator +(const vint4& a) { return a; }
280 __forceinline vint4 operator -(const vint4& a) { return _mm_sub_epi32(_mm_setzero_si128(), a); }
281#if defined(__aarch64__)
282 __forceinline vint4 abs(const vint4& a) { return vabsq_s32(a.v); }
283#elif defined(__SSSE3__)
284 __forceinline vint4 abs(const vint4& a) { return _mm_abs_epi32(a); }
285#endif
286
287 ////////////////////////////////////////////////////////////////////////////////
288 /// Binary Operators
289 ////////////////////////////////////////////////////////////////////////////////
290
291 __forceinline vint4 operator +(const vint4& a, const vint4& b) { return _mm_add_epi32(a, 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, const vint4& b) { return _mm_sub_epi32(a, b); }
296 __forceinline vint4 operator -(const vint4& a, int b) { return a - vint4(b); }
297 __forceinline vint4 operator -(int a, const vint4& b) { return vint4(a) - b; }
298
299#if (defined(__aarch64__)) || defined(__SSE4_1__)
300 __forceinline vint4 operator *(const vint4& a, const vint4& b) { return _mm_mullo_epi32(a, b); }
301#else
302 __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]); }
303#endif
304 __forceinline vint4 operator *(const vint4& a, int b) { return a * vint4(b); }
305 __forceinline vint4 operator *(int a, const vint4& b) { return vint4(a) * b; }
306
307 __forceinline vint4 operator &(const vint4& a, const vint4& b) { return _mm_and_si128(a, b); }
308 __forceinline vint4 operator &(const vint4& a, int b) { return a & vint4(b); }
309 __forceinline vint4 operator &(int a, const vint4& b) { return vint4(a) & b; }
310
311 __forceinline vint4 operator |(const vint4& a, const vint4& b) { return _mm_or_si128(a, b); }
312 __forceinline vint4 operator |(const vint4& a, int b) { return a | vint4(b); }
313 __forceinline vint4 operator |(int a, const vint4& b) { return vint4(a) | b; }
314
315 __forceinline vint4 operator ^(const vint4& a, const vint4& b) { return _mm_xor_si128(a, b); }
316 __forceinline vint4 operator ^(const vint4& a, int b) { return a ^ vint4(b); }
317 __forceinline vint4 operator ^(int a, const vint4& b) { return vint4(a) ^ b; }
318
319 __forceinline vint4 operator <<(const vint4& a, const int n) { return _mm_slli_epi32(a, n); }
320 __forceinline vint4 operator >>(const vint4& a, const int n) { return _mm_srai_epi32(a, n); }
321
322 __forceinline vint4 sll (const vint4& a, int b) { return _mm_slli_epi32(a, b); }
323 __forceinline vint4 sra (const vint4& a, int b) { return _mm_srai_epi32(a, b); }
324 __forceinline vint4 srl (const vint4& a, int b) { return _mm_srli_epi32(a, b); }
325
326 ////////////////////////////////////////////////////////////////////////////////
327 /// Assignment Operators
328 ////////////////////////////////////////////////////////////////////////////////
329
330 __forceinline vint4& operator +=(vint4& a, const vint4& b) { return a = a + b; }
331 __forceinline vint4& operator +=(vint4& a, int b) { return a = a + b; }
332
333 __forceinline vint4& operator -=(vint4& a, const vint4& b) { return a = a - b; }
334 __forceinline vint4& operator -=(vint4& a, int b) { return a = a - b; }
335
336#if (defined(__aarch64__)) || defined(__SSE4_1__)
337 __forceinline vint4& operator *=(vint4& a, const vint4& b) { return a = a * b; }
338 __forceinline vint4& operator *=(vint4& a, int b) { return a = a * b; }
339#endif
340
341 __forceinline vint4& operator &=(vint4& a, const vint4& b) { return a = a & b; }
342 __forceinline vint4& operator &=(vint4& a, int b) { return a = a & b; }
343
344 __forceinline vint4& operator |=(vint4& a, const vint4& b) { return a = a | b; }
345 __forceinline vint4& operator |=(vint4& a, int b) { return a = a | b; }
346
347 __forceinline vint4& operator <<=(vint4& a, int b) { return a = a << b; }
348 __forceinline vint4& operator >>=(vint4& a, int b) { return a = a >> b; }
349
350 ////////////////////////////////////////////////////////////////////////////////
351 /// Comparison Operators + Select
352 ////////////////////////////////////////////////////////////////////////////////
353
354#if defined(__AVX512VL__)
355 __forceinline vboolf4 operator ==(const vint4& a, const vint4& b) { return _mm_cmp_epi32_mask(a,b,_MM_CMPINT_EQ); }
356 __forceinline vboolf4 operator !=(const vint4& a, const vint4& b) { return _mm_cmp_epi32_mask(a,b,_MM_CMPINT_NE); }
357 __forceinline vboolf4 operator < (const vint4& a, const vint4& b) { return _mm_cmp_epi32_mask(a,b,_MM_CMPINT_LT); }
358 __forceinline vboolf4 operator >=(const vint4& a, const vint4& b) { return _mm_cmp_epi32_mask(a,b,_MM_CMPINT_GE); }
359 __forceinline vboolf4 operator > (const vint4& a, const vint4& b) { return _mm_cmp_epi32_mask(a,b,_MM_CMPINT_GT); }
360 __forceinline vboolf4 operator <=(const vint4& a, const vint4& b) { return _mm_cmp_epi32_mask(a,b,_MM_CMPINT_LE); }
361#else
362 __forceinline vboolf4 operator ==(const vint4& a, const vint4& b) { return _mm_castsi128_ps(_mm_cmpeq_epi32(a, b)); }
363 __forceinline vboolf4 operator !=(const vint4& a, const vint4& b) { return !(a == b); }
364 __forceinline vboolf4 operator < (const vint4& a, const vint4& b) { return _mm_castsi128_ps(_mm_cmplt_epi32(a, b)); }
365 __forceinline vboolf4 operator >=(const vint4& a, const vint4& b) { return !(a < b); }
366 __forceinline vboolf4 operator > (const vint4& a, const vint4& b) { return _mm_castsi128_ps(_mm_cmpgt_epi32(a, b)); }
367 __forceinline vboolf4 operator <=(const vint4& a, const vint4& b) { return !(a > b); }
368#endif
369
370 __forceinline vboolf4 operator ==(const vint4& a, int b) { return a == vint4(b); }
371 __forceinline vboolf4 operator ==(int a, const vint4& b) { return vint4(a) == b; }
372
373 __forceinline vboolf4 operator !=(const vint4& a, int b) { return a != vint4(b); }
374 __forceinline vboolf4 operator !=(int a, const vint4& b) { return vint4(a) != b; }
375
376 __forceinline vboolf4 operator < (const vint4& a, int b) { return a < vint4(b); }
377 __forceinline vboolf4 operator < (int a, const vint4& b) { return vint4(a) < b; }
378
379 __forceinline vboolf4 operator >=(const vint4& a, int b) { return a >= vint4(b); }
380 __forceinline vboolf4 operator >=(int a, const vint4& b) { return vint4(a) >= b; }
381
382 __forceinline vboolf4 operator > (const vint4& a, int b) { return a > vint4(b); }
383 __forceinline vboolf4 operator > (int a, const vint4& b) { return vint4(a) > b; }
384
385 __forceinline vboolf4 operator <=(const vint4& a, int b) { return a <= vint4(b); }
386 __forceinline vboolf4 operator <=(int a, const vint4& b) { return vint4(a) <= b; }
387
388 __forceinline vboolf4 eq(const vint4& a, const vint4& b) { return a == b; }
389 __forceinline vboolf4 ne(const vint4& a, const vint4& b) { return a != b; }
390 __forceinline vboolf4 lt(const vint4& a, const vint4& b) { return a < b; }
391 __forceinline vboolf4 ge(const vint4& a, const vint4& b) { return a >= b; }
392 __forceinline vboolf4 gt(const vint4& a, const vint4& b) { return a > b; }
393 __forceinline vboolf4 le(const vint4& a, const vint4& b) { return a <= b; }
394
395#if defined(__AVX512VL__)
396 __forceinline vboolf4 eq(const vboolf4& mask, const vint4& a, const vint4& b) { return _mm_mask_cmp_epi32_mask(mask, a, b, _MM_CMPINT_EQ); }
397 __forceinline vboolf4 ne(const vboolf4& mask, const vint4& a, const vint4& b) { return _mm_mask_cmp_epi32_mask(mask, a, b, _MM_CMPINT_NE); }
398 __forceinline vboolf4 lt(const vboolf4& mask, const vint4& a, const vint4& b) { return _mm_mask_cmp_epi32_mask(mask, a, b, _MM_CMPINT_LT); }
399 __forceinline vboolf4 ge(const vboolf4& mask, const vint4& a, const vint4& b) { return _mm_mask_cmp_epi32_mask(mask, a, b, _MM_CMPINT_GE); }
400 __forceinline vboolf4 gt(const vboolf4& mask, const vint4& a, const vint4& b) { return _mm_mask_cmp_epi32_mask(mask, a, b, _MM_CMPINT_GT); }
401 __forceinline vboolf4 le(const vboolf4& mask, const vint4& a, const vint4& b) { return _mm_mask_cmp_epi32_mask(mask, a, b, _MM_CMPINT_LE); }
402#else
403 __forceinline vboolf4 eq(const vboolf4& mask, const vint4& a, const vint4& b) { return mask & (a == b); }
404 __forceinline vboolf4 ne(const vboolf4& mask, const vint4& a, const vint4& b) { return mask & (a != b); }
405 __forceinline vboolf4 lt(const vboolf4& mask, const vint4& a, const vint4& b) { return mask & (a < b); }
406 __forceinline vboolf4 ge(const vboolf4& mask, const vint4& a, const vint4& b) { return mask & (a >= b); }
407 __forceinline vboolf4 gt(const vboolf4& mask, const vint4& a, const vint4& b) { return mask & (a > b); }
408 __forceinline vboolf4 le(const vboolf4& mask, const vint4& a, const vint4& b) { return mask & (a <= b); }
409#endif
410
411 template<int mask>
412 __forceinline vint4 select(const vint4& t, const vint4& f) {
413#if defined(__SSE4_1__)
414 return _mm_castps_si128(_mm_blend_ps(_mm_castsi128_ps(f), _mm_castsi128_ps(t), mask));
415#else
416 return select(vboolf4(mask), t, f);
417#endif
418 }
419
420#if defined(__aarch64__) || defined(__SSE4_1__)
421 __forceinline vint4 min(const vint4& a, const vint4& b) { return _mm_min_epi32(a, b); }
422 __forceinline vint4 max(const vint4& a, const vint4& b) { return _mm_max_epi32(a, b); }
423
424 __forceinline vint4 umin(const vint4& a, const vint4& b) { return _mm_min_epu32(a, b); }
425 __forceinline vint4 umax(const vint4& a, const vint4& b) { return _mm_max_epu32(a, b); }
426
427#else
428 __forceinline vint4 min(const vint4& a, const vint4& b) { return select(a < b,a,b); }
429 __forceinline vint4 max(const vint4& a, const vint4& b) { return select(a < b,b,a); }
430#endif
431
432 __forceinline vint4 min(const vint4& a, int b) { return min(a,vint4(b)); }
433 __forceinline vint4 min(int a, const vint4& b) { return min(vint4(a),b); }
434 __forceinline vint4 max(const vint4& a, int b) { return max(a,vint4(b)); }
435 __forceinline vint4 max(int a, const vint4& b) { return max(vint4(a),b); }
436
437 ////////////////////////////////////////////////////////////////////////////////
438 // Movement/Shifting/Shuffling Functions
439 ////////////////////////////////////////////////////////////////////////////////
440
441 __forceinline vint4 unpacklo(const vint4& a, const vint4& b) { return _mm_castps_si128(_mm_unpacklo_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b))); }
442 __forceinline vint4 unpackhi(const vint4& a, const vint4& b) { return _mm_castps_si128(_mm_unpackhi_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b))); }
443
444#if defined(__aarch64__)
445 template<int i0, int i1, int i2, int i3>
446 __forceinline vint4 shuffle(const vint4& v) {
447 return vreinterpretq_s32_u8(vqtbl1q_u8( (uint8x16_t)v.v, _MN_SHUFFLE(i0, i1, i2, i3)));
448 }
449 template<int i0, int i1, int i2, int i3>
450 __forceinline vint4 shuffle(const vint4& a, const vint4& b) {
451 return vreinterpretq_s32_u8(vqtbl2q_u8( (uint8x16x2_t){(uint8x16_t)a.v, (uint8x16_t)b.v}, _MF_SHUFFLE(i0, i1, i2, i3)));
452 }
453#else
454 template<int i0, int i1, int i2, int i3>
455 __forceinline vint4 shuffle(const vint4& v) {
456 return _mm_shuffle_epi32(v, _MM_SHUFFLE(i3, i2, i1, i0));
457 }
458
459 template<int i0, int i1, int i2, int i3>
460 __forceinline vint4 shuffle(const vint4& a, const vint4& b) {
461 return _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), _MM_SHUFFLE(i3, i2, i1, i0)));
462 }
463#endif
464#if defined(__SSE3__)
465 template<> __forceinline vint4 shuffle<0, 0, 2, 2>(const vint4& v) { return _mm_castps_si128(_mm_moveldup_ps(_mm_castsi128_ps(v))); }
466 template<> __forceinline vint4 shuffle<1, 1, 3, 3>(const vint4& v) { return _mm_castps_si128(_mm_movehdup_ps(_mm_castsi128_ps(v))); }
467 template<> __forceinline vint4 shuffle<0, 1, 0, 1>(const vint4& v) { return _mm_castpd_si128(_mm_movedup_pd (_mm_castsi128_pd(v))); }
468#endif
469
470 template<int i>
471 __forceinline vint4 shuffle(const vint4& v) {
472 return shuffle<i,i,i,i>(v);
473 }
474
475#if defined(__SSE4_1__) && !defined(__aarch64__)
476 template<int src> __forceinline int extract(const vint4& b) { return _mm_extract_epi32(b, src); }
477 template<int dst> __forceinline vint4 insert(const vint4& a, const int b) { return _mm_insert_epi32(a, b, dst); }
478#else
479 template<int src> __forceinline int extract(const vint4& b) { return b[src&3]; }
480 template<int dst> __forceinline vint4 insert(const vint4& a, int b) { vint4 c = a; c[dst&3] = b; return c; }
481#endif
482
483 template<> __forceinline int extract<0>(const vint4& b) { return _mm_cvtsi128_si32(b); }
484
485 __forceinline int toScalar(const vint4& v) { return _mm_cvtsi128_si32(v); }
486
487#if defined(__aarch64__)
488 __forceinline size_t toSizeT(const vint4& v) {
489 uint64x2_t x = uint64x2_t(v.v);
490 return x[0];
491 }
492#else
493__forceinline size_t toSizeT(const vint4& v) {
494#if defined(__WIN32__) && !defined(__X86_64__) // win32 workaround
495 return toScalar(v);
496#elif defined(__ARM_NEON)
497 // FIXME(LTE): Do we need a swap(i.e. use lane 1)?
498 return vgetq_lane_u64(*(reinterpret_cast<const uint64x2_t *>(&v)), 0);
499#else
500 return _mm_cvtsi128_si64(v);
501#endif
502 }
503#endif
504
505#if defined(__AVX512VL__)
506
507 __forceinline vint4 permute(const vint4 &a, const vint4 &index) {
508 return _mm_castps_si128(_mm_permutevar_ps(_mm_castsi128_ps(a),index));
509 }
510
511 template<int i>
512 __forceinline vint4 align_shift_right(const vint4& a, const vint4& b) {
513 return _mm_alignr_epi32(a, b, i);
514 }
515#endif
516
517 ////////////////////////////////////////////////////////////////////////////////
518 /// Reductions
519 ////////////////////////////////////////////////////////////////////////////////
520
521#if defined(__aarch64__) || defined(__SSE4_1__)
522
523#if defined(__aarch64__)
524 __forceinline vint4 vreduce_min(const vint4& v) { int h = vminvq_s32(v); return vdupq_n_s32(h); }
525 __forceinline vint4 vreduce_max(const vint4& v) { int h = vmaxvq_s32(v); return vdupq_n_s32(h); }
526 __forceinline vint4 vreduce_add(const vint4& v) { int h = vaddvq_s32(v); return vdupq_n_s32(h); }
527
528 __forceinline int reduce_min(const vint4& v) { return vminvq_s32(v); }
529 __forceinline int reduce_max(const vint4& v) { return vmaxvq_s32(v); }
530 __forceinline int reduce_add(const vint4& v) { return vaddvq_s32(v); }
531#else
532 __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); }
533 __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); }
534 __forceinline vint4 vreduce_add(const vint4& v) { vint4 h = shuffle<1,0,3,2>(v) + v ; return shuffle<2,3,0,1>(h) + h ; }
535
536 __forceinline int reduce_min(const vint4& v) { return toScalar(vreduce_min(v)); }
537 __forceinline int reduce_max(const vint4& v) { return toScalar(vreduce_max(v)); }
538 __forceinline int reduce_add(const vint4& v) { return toScalar(vreduce_add(v)); }
539#endif
540
541 __forceinline size_t select_min(const vint4& v) { return bsf(movemask(v == vreduce_min(v))); }
542 __forceinline size_t select_max(const vint4& v) { return bsf(movemask(v == vreduce_max(v))); }
543
544 __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)))); }
545 __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)))); }
546
547#else
548
549 __forceinline int reduce_min(const vint4& v) { return min(v[0],v[1],v[2],v[3]); }
550 __forceinline int reduce_max(const vint4& v) { return max(v[0],v[1],v[2],v[3]); }
551 __forceinline int reduce_add(const vint4& v) { return v[0]+v[1]+v[2]+v[3]; }
552
553#endif
554
555 ////////////////////////////////////////////////////////////////////////////////
556 /// Sorting networks
557 ////////////////////////////////////////////////////////////////////////////////
558
559#if (defined(__aarch64__)) || defined(__SSE4_1__)
560
561 __forceinline vint4 usort_ascending(const vint4& v)
562 {
563 const vint4 a0 = v;
564 const vint4 b0 = shuffle<1,0,3,2>(a0);
565 const vint4 c0 = umin(a0,b0);
566 const vint4 d0 = umax(a0,b0);
567 const vint4 a1 = select<0x5 /* 0b0101 */>(c0,d0);
568 const vint4 b1 = shuffle<2,3,0,1>(a1);
569 const vint4 c1 = umin(a1,b1);
570 const vint4 d1 = umax(a1,b1);
571 const vint4 a2 = select<0x3 /* 0b0011 */>(c1,d1);
572 const vint4 b2 = shuffle<0,2,1,3>(a2);
573 const vint4 c2 = umin(a2,b2);
574 const vint4 d2 = umax(a2,b2);
575 const vint4 a3 = select<0x2 /* 0b0010 */>(c2,d2);
576 return a3;
577 }
578
579 __forceinline vint4 usort_descending(const vint4& v)
580 {
581 const vint4 a0 = v;
582 const vint4 b0 = shuffle<1,0,3,2>(a0);
583 const vint4 c0 = umax(a0,b0);
584 const vint4 d0 = umin(a0,b0);
585 const vint4 a1 = select<0x5 /* 0b0101 */>(c0,d0);
586 const vint4 b1 = shuffle<2,3,0,1>(a1);
587 const vint4 c1 = umax(a1,b1);
588 const vint4 d1 = umin(a1,b1);
589 const vint4 a2 = select<0x3 /* 0b0011 */>(c1,d1);
590 const vint4 b2 = shuffle<0,2,1,3>(a2);
591 const vint4 c2 = umax(a2,b2);
592 const vint4 d2 = umin(a2,b2);
593 const vint4 a3 = select<0x2 /* 0b0010 */>(c2,d2);
594 return a3;
595 }
596
597#else
598
599 __forceinline vint4 usort_ascending(const vint4& v)
600 {
601 const vint4 a0 = v-vint4(0x80000000);
602 const vint4 b0 = shuffle<1,0,3,2>(a0);
603 const vint4 c0 = min(a0,b0);
604 const vint4 d0 = max(a0,b0);
605 const vint4 a1 = select<0x5 /* 0b0101 */>(c0,d0);
606 const vint4 b1 = shuffle<2,3,0,1>(a1);
607 const vint4 c1 = min(a1,b1);
608 const vint4 d1 = max(a1,b1);
609 const vint4 a2 = select<0x3 /* 0b0011 */>(c1,d1);
610 const vint4 b2 = shuffle<0,2,1,3>(a2);
611 const vint4 c2 = min(a2,b2);
612 const vint4 d2 = max(a2,b2);
613 const vint4 a3 = select<0x2 /* 0b0010 */>(c2,d2);
614 return a3+vint4(0x80000000);
615 }
616
617 __forceinline vint4 usort_descending(const vint4& v)
618 {
619 const vint4 a0 = v-vint4(0x80000000);
620 const vint4 b0 = shuffle<1,0,3,2>(a0);
621 const vint4 c0 = max(a0,b0);
622 const vint4 d0 = min(a0,b0);
623 const vint4 a1 = select<0x5 /* 0b0101 */>(c0,d0);
624 const vint4 b1 = shuffle<2,3,0,1>(a1);
625 const vint4 c1 = max(a1,b1);
626 const vint4 d1 = min(a1,b1);
627 const vint4 a2 = select<0x3 /* 0b0011 */>(c1,d1);
628 const vint4 b2 = shuffle<0,2,1,3>(a2);
629 const vint4 c2 = max(a2,b2);
630 const vint4 d2 = min(a2,b2);
631 const vint4 a3 = select<0x2 /* 0b0010 */>(c2,d2);
632 return a3+vint4(0x80000000);
633 }
634
635#endif
636
637 ////////////////////////////////////////////////////////////////////////////////
638 /// Output Operators
639 ////////////////////////////////////////////////////////////////////////////////
640
641 __forceinline embree_ostream operator <<(embree_ostream cout, const vint4& a) {
642 return cout << "<" << a[0] << ", " << a[1] << ", " << a[2] << ", " << a[3] << ">";
643 }
644}
645
646#undef vboolf
647#undef vboold
648#undef vint
649#undef vuint
650#undef vllong
651#undef vfloat
652#undef vdouble
653