Vector Optimized Library of Kernels  3.3.0
Architecture-tuned implementations of math kernels
volk_32f_x2_add_32f.h
Go to the documentation of this file.
1 /* -*- c++ -*- */
2 /*
3  * Copyright 2012, 2014 Free Software Foundation, Inc.
4  *
5  * This file is part of VOLK
6  *
7  * SPDX-License-Identifier: LGPL-3.0-or-later
8  */
9 
61 #ifndef INCLUDED_volk_32f_x2_add_32f_u_H
62 #define INCLUDED_volk_32f_x2_add_32f_u_H
63 
64 #include <inttypes.h>
65 #include <stdio.h>
66 
67 #ifdef LV_HAVE_AVX512F
68 #include <immintrin.h>
69 
70 static inline void volk_32f_x2_add_32f_u_avx512f(float* cVector,
71  const float* aVector,
72  const float* bVector,
73  unsigned int num_points)
74 {
75  unsigned int number = 0;
76  const unsigned int sixteenthPoints = num_points / 16;
77 
78  float* cPtr = cVector;
79  const float* aPtr = aVector;
80  const float* bPtr = bVector;
81 
82  __m512 aVal, bVal, cVal;
83  for (; number < sixteenthPoints; number++) {
84 
85  aVal = _mm512_loadu_ps(aPtr);
86  bVal = _mm512_loadu_ps(bPtr);
87 
88  cVal = _mm512_add_ps(aVal, bVal);
89 
90  _mm512_storeu_ps(cPtr, cVal); // Store the results back into the C container
91 
92  aPtr += 16;
93  bPtr += 16;
94  cPtr += 16;
95  }
96 
97  number = sixteenthPoints * 16;
98 
99  for (; number < num_points; number++) {
100  *cPtr++ = (*aPtr++) + (*bPtr++);
101  }
102 }
103 
104 #endif /* LV_HAVE_AVX512F */
105 
106 
107 #ifdef LV_HAVE_AVX
108 #include <immintrin.h>
109 
110 static inline void volk_32f_x2_add_32f_u_avx(float* cVector,
111  const float* aVector,
112  const float* bVector,
113  unsigned int num_points)
114 {
115  unsigned int number = 0;
116  const unsigned int eighthPoints = num_points / 8;
117  float* cPtr = cVector;
118  const float* aPtr = aVector;
119  const float* bPtr = bVector;
120  __m256 aVal, bVal, cVal;
121  for (; number < eighthPoints; number++) {
122 
123  aVal = _mm256_loadu_ps(aPtr);
124  bVal = _mm256_loadu_ps(bPtr);
125 
126  cVal = _mm256_add_ps(aVal, bVal);
127 
128  _mm256_storeu_ps(cPtr, cVal); // Store the results back into the C container
129 
130  aPtr += 8;
131  bPtr += 8;
132  cPtr += 8;
133  }
134 
135  number = eighthPoints * 8;
136 
137  for (; number < num_points; number++) {
138  *cPtr++ = (*aPtr++) + (*bPtr++);
139  }
140 }
141 #endif /* LV_HAVE_AVX */
142 
143 
144 #ifdef LV_HAVE_SSE
145 #include <xmmintrin.h>
146 
147 static inline void volk_32f_x2_add_32f_u_sse(float* cVector,
148  const float* aVector,
149  const float* bVector,
150  unsigned int num_points)
151 {
152  unsigned int number = 0;
153  const unsigned int quarterPoints = num_points / 4;
154 
155  float* cPtr = cVector;
156  const float* aPtr = aVector;
157  const float* bPtr = bVector;
158 
159  __m128 aVal, bVal, cVal;
160  for (; number < quarterPoints; number++) {
161 
162  aVal = _mm_loadu_ps(aPtr);
163  bVal = _mm_loadu_ps(bPtr);
164 
165  cVal = _mm_add_ps(aVal, bVal);
166 
167  _mm_storeu_ps(cPtr, cVal); // Store the results back into the C container
168 
169  aPtr += 4;
170  bPtr += 4;
171  cPtr += 4;
172  }
173 
174  number = quarterPoints * 4;
175  for (; number < num_points; number++) {
176  *cPtr++ = (*aPtr++) + (*bPtr++);
177  }
178 }
179 #endif /* LV_HAVE_SSE */
180 
181 
182 #ifdef LV_HAVE_GENERIC
183 
184 static inline void volk_32f_x2_add_32f_generic(float* cVector,
185  const float* aVector,
186  const float* bVector,
187  unsigned int num_points)
188 {
189  float* cPtr = cVector;
190  const float* aPtr = aVector;
191  const float* bPtr = bVector;
192  unsigned int number = 0;
193 
194  for (number = 0; number < num_points; number++) {
195  *cPtr++ = (*aPtr++) + (*bPtr++);
196  }
197 }
198 #endif /* LV_HAVE_GENERIC */
199 
200 
201 #endif /* INCLUDED_volk_32f_x2_add_32f_u_H */
202 #ifndef INCLUDED_volk_32f_x2_add_32f_a_H
203 #define INCLUDED_volk_32f_x2_add_32f_a_H
204 
205 #include <inttypes.h>
206 #include <stdio.h>
207 
208 #ifdef LV_HAVE_AVX512F
209 #include <immintrin.h>
210 
211 static inline void volk_32f_x2_add_32f_a_avx512f(float* cVector,
212  const float* aVector,
213  const float* bVector,
214  unsigned int num_points)
215 {
216  unsigned int number = 0;
217  const unsigned int sixteenthPoints = num_points / 16;
218 
219  float* cPtr = cVector;
220  const float* aPtr = aVector;
221  const float* bPtr = bVector;
222 
223  __m512 aVal, bVal, cVal;
224  for (; number < sixteenthPoints; number++) {
225 
226  aVal = _mm512_load_ps(aPtr);
227  bVal = _mm512_load_ps(bPtr);
228 
229  cVal = _mm512_add_ps(aVal, bVal);
230 
231  _mm512_store_ps(cPtr, cVal); // Store the results back into the C container
232 
233  aPtr += 16;
234  bPtr += 16;
235  cPtr += 16;
236  }
237 
238  number = sixteenthPoints * 16;
239 
240  for (; number < num_points; number++) {
241  *cPtr++ = (*aPtr++) + (*bPtr++);
242  }
243 }
244 
245 #endif /* LV_HAVE_AVX512F */
246 
247 
248 #ifdef LV_HAVE_AVX
249 #include <immintrin.h>
250 
251 static inline void volk_32f_x2_add_32f_a_avx(float* cVector,
252  const float* aVector,
253  const float* bVector,
254  unsigned int num_points)
255 {
256  unsigned int number = 0;
257  const unsigned int eighthPoints = num_points / 8;
258 
259  float* cPtr = cVector;
260  const float* aPtr = aVector;
261  const float* bPtr = bVector;
262 
263  __m256 aVal, bVal, cVal;
264  for (; number < eighthPoints; number++) {
265 
266  aVal = _mm256_load_ps(aPtr);
267  bVal = _mm256_load_ps(bPtr);
268 
269  cVal = _mm256_add_ps(aVal, bVal);
270 
271  _mm256_store_ps(cPtr, cVal); // Store the results back into the C container
272 
273  aPtr += 8;
274  bPtr += 8;
275  cPtr += 8;
276  }
277 
278  number = eighthPoints * 8;
279  for (; number < num_points; number++) {
280  *cPtr++ = (*aPtr++) + (*bPtr++);
281  }
282 }
283 #endif /* LV_HAVE_AVX */
284 
285 #ifdef LV_HAVE_SSE
286 #include <xmmintrin.h>
287 
288 static inline void volk_32f_x2_add_32f_a_sse(float* cVector,
289  const float* aVector,
290  const float* bVector,
291  unsigned int num_points)
292 {
293  unsigned int number = 0;
294  const unsigned int quarterPoints = num_points / 4;
295 
296  float* cPtr = cVector;
297  const float* aPtr = aVector;
298  const float* bPtr = bVector;
299 
300  __m128 aVal, bVal, cVal;
301  for (; number < quarterPoints; number++) {
302  aVal = _mm_load_ps(aPtr);
303  bVal = _mm_load_ps(bPtr);
304 
305  cVal = _mm_add_ps(aVal, bVal);
306 
307  _mm_store_ps(cPtr, cVal); // Store the results back into the C container
308 
309  aPtr += 4;
310  bPtr += 4;
311  cPtr += 4;
312  }
313 
314  number = quarterPoints * 4;
315  for (; number < num_points; number++) {
316  *cPtr++ = (*aPtr++) + (*bPtr++);
317  }
318 }
319 #endif /* LV_HAVE_SSE */
320 
321 
322 #ifdef LV_HAVE_NEON
323 #include <arm_neon.h>
324 
325 static inline void volk_32f_x2_add_32f_u_neon(float* cVector,
326  const float* aVector,
327  const float* bVector,
328  unsigned int num_points)
329 {
330  unsigned int number = 0;
331  const unsigned int quarterPoints = num_points / 4;
332 
333  float* cPtr = cVector;
334  const float* aPtr = aVector;
335  const float* bPtr = bVector;
336  float32x4_t aVal, bVal, cVal;
337  for (number = 0; number < quarterPoints; number++) {
338  // Load in to NEON registers
339  aVal = vld1q_f32(aPtr);
340  bVal = vld1q_f32(bPtr);
341  __VOLK_PREFETCH(aPtr + 4);
342  __VOLK_PREFETCH(bPtr + 4);
343 
344  // vector add
345  cVal = vaddq_f32(aVal, bVal);
346  // Store the results back into the C container
347  vst1q_f32(cPtr, cVal);
348 
349  aPtr += 4; // q uses quadwords, 4 floats per vadd
350  bPtr += 4;
351  cPtr += 4;
352  }
353 
354  number = quarterPoints * 4; // should be = num_points
355  for (; number < num_points; number++) {
356  *cPtr++ = (*aPtr++) + (*bPtr++);
357  }
358 }
359 
360 #endif /* LV_HAVE_NEON */
361 
362 #ifdef LV_HAVE_NEONV7
363 extern void volk_32f_x2_add_32f_a_neonasm(float* cVector,
364  const float* aVector,
365  const float* bVector,
366  unsigned int num_points);
367 #endif /* LV_HAVE_NEONV7 */
368 
369 #ifdef LV_HAVE_NEONV7
370 extern void volk_32f_x2_add_32f_a_neonpipeline(float* cVector,
371  const float* aVector,
372  const float* bVector,
373  unsigned int num_points);
374 #endif /* LV_HAVE_NEONV7 */
375 
376 
377 #ifdef LV_HAVE_NEONV8
378 #include <arm_neon.h>
379 
380 static inline void volk_32f_x2_add_32f_neonv8(float* cVector,
381  const float* aVector,
382  const float* bVector,
383  unsigned int num_points)
384 {
385  unsigned int n = num_points;
386  float* c = cVector;
387  const float* a = aVector;
388  const float* b = bVector;
389 
390  /* Process 8 floats per iteration (2x unroll) */
391  while (n >= 8) {
392  float32x4_t a0 = vld1q_f32(a);
393  float32x4_t a1 = vld1q_f32(a + 4);
394  float32x4_t b0 = vld1q_f32(b);
395  float32x4_t b1 = vld1q_f32(b + 4);
396  __VOLK_PREFETCH(a + 16);
397  __VOLK_PREFETCH(b + 16);
398 
399  vst1q_f32(c, vaddq_f32(a0, b0));
400  vst1q_f32(c + 4, vaddq_f32(a1, b1));
401 
402  a += 8;
403  b += 8;
404  c += 8;
405  n -= 8;
406  }
407 
408  /* Process remaining 4 floats */
409  if (n >= 4) {
410  vst1q_f32(c, vaddq_f32(vld1q_f32(a), vld1q_f32(b)));
411  a += 4;
412  b += 4;
413  c += 4;
414  n -= 4;
415  }
416 
417  /* Scalar tail */
418  while (n > 0) {
419  *c++ = *a++ + *b++;
420  n--;
421  }
422 }
423 
424 #endif /* LV_HAVE_NEONV8 */
425 
426 
427 #ifdef LV_HAVE_ORC
428 
429 extern void volk_32f_x2_add_32f_a_orc_impl(float* cVector,
430  const float* aVector,
431  const float* bVector,
432  int num_points);
433 
434 static inline void volk_32f_x2_add_32f_u_orc(float* cVector,
435  const float* aVector,
436  const float* bVector,
437  unsigned int num_points)
438 {
439  volk_32f_x2_add_32f_a_orc_impl(cVector, aVector, bVector, num_points);
440 }
441 
442 #endif /* LV_HAVE_ORC */
443 
444 #ifdef LV_HAVE_RVV
445 #include <riscv_vector.h>
446 
447 static inline void volk_32f_x2_add_32f_rvv(float* cVector,
448  const float* aVector,
449  const float* bVector,
450  unsigned int num_points)
451 {
452  size_t n = num_points;
453  for (size_t vl; n > 0; n -= vl, aVector += vl, bVector += vl, cVector += vl) {
454  vl = __riscv_vsetvl_e32m8(n);
455  vfloat32m8_t va = __riscv_vle32_v_f32m8(aVector, vl);
456  vfloat32m8_t vb = __riscv_vle32_v_f32m8(bVector, vl);
457  __riscv_vse32(cVector, __riscv_vfadd(va, vb, vl), vl);
458  }
459 }
460 #endif /*LV_HAVE_RVV*/
461 
462 #endif /* INCLUDED_volk_32f_x2_add_32f_a_H */
volk_32f_x2_add_32f_a_sse
static void volk_32f_x2_add_32f_a_sse(float *cVector, const float *aVector, const float *bVector, unsigned int num_points)
Definition: volk_32f_x2_add_32f.h:288
__VOLK_PREFETCH
#define __VOLK_PREFETCH(addr)
Definition: volk_common.h:68
volk_32f_x2_add_32f_u_neon
static void volk_32f_x2_add_32f_u_neon(float *cVector, const float *aVector, const float *bVector, unsigned int num_points)
Definition: volk_32f_x2_add_32f.h:325
volk_32f_x2_add_32f_generic
static void volk_32f_x2_add_32f_generic(float *cVector, const float *aVector, const float *bVector, unsigned int num_points)
Definition: volk_32f_x2_add_32f.h:184
volk_32f_x2_add_32f_u_avx
static void volk_32f_x2_add_32f_u_avx(float *cVector, const float *aVector, const float *bVector, unsigned int num_points)
Definition: volk_32f_x2_add_32f.h:110
volk_32f_x2_add_32f_u_sse
static void volk_32f_x2_add_32f_u_sse(float *cVector, const float *aVector, const float *bVector, unsigned int num_points)
Definition: volk_32f_x2_add_32f.h:147
volk_32f_x2_add_32f_a_avx
static void volk_32f_x2_add_32f_a_avx(float *cVector, const float *aVector, const float *bVector, unsigned int num_points)
Definition: volk_32f_x2_add_32f.h:251