Vector Optimized Library of Kernels  2.5.0
Architecture-tuned implementations of math kernels
volk_32f_tanh_32f.h
Go to the documentation of this file.
1 /* -*- c++ -*- */
2 /*
3  * Copyright 2014 Free Software Foundation, Inc.
4  *
5  * This file is part of GNU Radio
6  *
7  * GNU Radio is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License as published by
9  * the Free Software Foundation; either version 3, or (at your option)
10  * any later version.
11  *
12  * GNU Radio is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15  * GNU General Public License for more details.
16  *
17  * You should have received a copy of the GNU General Public License
18  * along with GNU Radio; see the file COPYING. If not, write to
19  * the Free Software Foundation, Inc., 51 Franklin Street,
20  * Boston, MA 02110-1301, USA.
21  */
22 
68 #ifndef INCLUDED_volk_32f_tanh_32f_a_H
69 #define INCLUDED_volk_32f_tanh_32f_a_H
70 
71 #include <inttypes.h>
72 #include <math.h>
73 #include <stdio.h>
74 #include <string.h>
75 
76 
77 #ifdef LV_HAVE_GENERIC
78 
79 static inline void
80 volk_32f_tanh_32f_generic(float* cVector, const float* aVector, unsigned int num_points)
81 {
82  unsigned int number = 0;
83  float* cPtr = cVector;
84  const float* aPtr = aVector;
85  for (; number < num_points; number++) {
86  *cPtr++ = tanhf(*aPtr++);
87  }
88 }
89 
90 #endif /* LV_HAVE_GENERIC */
91 
92 
93 #ifdef LV_HAVE_GENERIC
94 
95 static inline void
96 volk_32f_tanh_32f_series(float* cVector, const float* aVector, unsigned int num_points)
97 {
98  float* cPtr = cVector;
99  const float* aPtr = aVector;
100  for (unsigned int number = 0; number < num_points; number++) {
101  if (*aPtr > 4.97)
102  *cPtr++ = 1;
103  else if (*aPtr <= -4.97)
104  *cPtr++ = -1;
105  else {
106  float x2 = (*aPtr) * (*aPtr);
107  float a = (*aPtr) * (135135.0f + x2 * (17325.0f + x2 * (378.0f + x2)));
108  float b = 135135.0f + x2 * (62370.0f + x2 * (3150.0f + x2 * 28.0f));
109  *cPtr++ = a / b;
110  aPtr++;
111  }
112  }
113 }
114 
115 #endif /* LV_HAVE_GENERIC */
116 
117 
118 #ifdef LV_HAVE_SSE
119 #include <xmmintrin.h>
120 
121 static inline void
122 volk_32f_tanh_32f_a_sse(float* cVector, const float* aVector, unsigned int num_points)
123 {
124  unsigned int number = 0;
125  const unsigned int quarterPoints = num_points / 4;
126 
127  float* cPtr = cVector;
128  const float* aPtr = aVector;
129 
130  __m128 aVal, cVal, x2, a, b;
131  __m128 const1, const2, const3, const4, const5, const6;
132  const1 = _mm_set_ps1(135135.0f);
133  const2 = _mm_set_ps1(17325.0f);
134  const3 = _mm_set_ps1(378.0f);
135  const4 = _mm_set_ps1(62370.0f);
136  const5 = _mm_set_ps1(3150.0f);
137  const6 = _mm_set_ps1(28.0f);
138  for (; number < quarterPoints; number++) {
139 
140  aVal = _mm_load_ps(aPtr);
141  x2 = _mm_mul_ps(aVal, aVal);
142  a = _mm_mul_ps(
143  aVal,
144  _mm_add_ps(
145  const1,
146  _mm_mul_ps(x2,
147  _mm_add_ps(const2, _mm_mul_ps(x2, _mm_add_ps(const3, x2))))));
148  b = _mm_add_ps(
149  const1,
150  _mm_mul_ps(
151  x2,
152  _mm_add_ps(const4,
153  _mm_mul_ps(x2, _mm_add_ps(const5, _mm_mul_ps(x2, const6))))));
154 
155  cVal = _mm_div_ps(a, b);
156 
157  _mm_store_ps(cPtr, cVal); // Store the results back into the C container
158 
159  aPtr += 4;
160  cPtr += 4;
161  }
162 
163  number = quarterPoints * 4;
164  volk_32f_tanh_32f_series(cPtr, aPtr, num_points - number);
165 }
166 #endif /* LV_HAVE_SSE */
167 
168 
169 #ifdef LV_HAVE_AVX
170 #include <immintrin.h>
171 
172 static inline void
173 volk_32f_tanh_32f_a_avx(float* cVector, const float* aVector, unsigned int num_points)
174 {
175  unsigned int number = 0;
176  const unsigned int eighthPoints = num_points / 8;
177 
178  float* cPtr = cVector;
179  const float* aPtr = aVector;
180 
181  __m256 aVal, cVal, x2, a, b;
182  __m256 const1, const2, const3, const4, const5, const6;
183  const1 = _mm256_set1_ps(135135.0f);
184  const2 = _mm256_set1_ps(17325.0f);
185  const3 = _mm256_set1_ps(378.0f);
186  const4 = _mm256_set1_ps(62370.0f);
187  const5 = _mm256_set1_ps(3150.0f);
188  const6 = _mm256_set1_ps(28.0f);
189  for (; number < eighthPoints; number++) {
190 
191  aVal = _mm256_load_ps(aPtr);
192  x2 = _mm256_mul_ps(aVal, aVal);
193  a = _mm256_mul_ps(
194  aVal,
195  _mm256_add_ps(
196  const1,
197  _mm256_mul_ps(
198  x2,
199  _mm256_add_ps(const2,
200  _mm256_mul_ps(x2, _mm256_add_ps(const3, x2))))));
201  b = _mm256_add_ps(
202  const1,
203  _mm256_mul_ps(
204  x2,
205  _mm256_add_ps(
206  const4,
207  _mm256_mul_ps(x2,
208  _mm256_add_ps(const5, _mm256_mul_ps(x2, const6))))));
209 
210  cVal = _mm256_div_ps(a, b);
211 
212  _mm256_store_ps(cPtr, cVal); // Store the results back into the C container
213 
214  aPtr += 8;
215  cPtr += 8;
216  }
217 
218  number = eighthPoints * 8;
219  volk_32f_tanh_32f_series(cPtr, aPtr, num_points - number);
220 }
221 #endif /* LV_HAVE_AVX */
222 
223 #if LV_HAVE_AVX && LV_HAVE_FMA
224 #include <immintrin.h>
225 
226 static inline void
227 volk_32f_tanh_32f_a_avx_fma(float* cVector, const float* aVector, unsigned int num_points)
228 {
229  unsigned int number = 0;
230  const unsigned int eighthPoints = num_points / 8;
231 
232  float* cPtr = cVector;
233  const float* aPtr = aVector;
234 
235  __m256 aVal, cVal, x2, a, b;
236  __m256 const1, const2, const3, const4, const5, const6;
237  const1 = _mm256_set1_ps(135135.0f);
238  const2 = _mm256_set1_ps(17325.0f);
239  const3 = _mm256_set1_ps(378.0f);
240  const4 = _mm256_set1_ps(62370.0f);
241  const5 = _mm256_set1_ps(3150.0f);
242  const6 = _mm256_set1_ps(28.0f);
243  for (; number < eighthPoints; number++) {
244 
245  aVal = _mm256_load_ps(aPtr);
246  x2 = _mm256_mul_ps(aVal, aVal);
247  a = _mm256_mul_ps(
248  aVal,
249  _mm256_fmadd_ps(
250  x2, _mm256_fmadd_ps(x2, _mm256_add_ps(const3, x2), const2), const1));
251  b = _mm256_fmadd_ps(
252  x2, _mm256_fmadd_ps(x2, _mm256_fmadd_ps(x2, const6, const5), const4), const1);
253 
254  cVal = _mm256_div_ps(a, b);
255 
256  _mm256_store_ps(cPtr, cVal); // Store the results back into the C container
257 
258  aPtr += 8;
259  cPtr += 8;
260  }
261 
262  number = eighthPoints * 8;
263  volk_32f_tanh_32f_series(cPtr, aPtr, num_points - number);
264 }
265 #endif /* LV_HAVE_AVX && LV_HAVE_FMA */
266 
267 #endif /* INCLUDED_volk_32f_tanh_32f_a_H */
268 
269 
270 #ifndef INCLUDED_volk_32f_tanh_32f_u_H
271 #define INCLUDED_volk_32f_tanh_32f_u_H
272 
273 #include <inttypes.h>
274 #include <math.h>
275 #include <stdio.h>
276 #include <string.h>
277 
278 
279 #ifdef LV_HAVE_SSE
280 #include <xmmintrin.h>
281 
282 static inline void
283 volk_32f_tanh_32f_u_sse(float* cVector, const float* aVector, unsigned int num_points)
284 {
285  unsigned int number = 0;
286  const unsigned int quarterPoints = num_points / 4;
287 
288  float* cPtr = cVector;
289  const float* aPtr = aVector;
290 
291  __m128 aVal, cVal, x2, a, b;
292  __m128 const1, const2, const3, const4, const5, const6;
293  const1 = _mm_set_ps1(135135.0f);
294  const2 = _mm_set_ps1(17325.0f);
295  const3 = _mm_set_ps1(378.0f);
296  const4 = _mm_set_ps1(62370.0f);
297  const5 = _mm_set_ps1(3150.0f);
298  const6 = _mm_set_ps1(28.0f);
299  for (; number < quarterPoints; number++) {
300 
301  aVal = _mm_loadu_ps(aPtr);
302  x2 = _mm_mul_ps(aVal, aVal);
303  a = _mm_mul_ps(
304  aVal,
305  _mm_add_ps(
306  const1,
307  _mm_mul_ps(x2,
308  _mm_add_ps(const2, _mm_mul_ps(x2, _mm_add_ps(const3, x2))))));
309  b = _mm_add_ps(
310  const1,
311  _mm_mul_ps(
312  x2,
313  _mm_add_ps(const4,
314  _mm_mul_ps(x2, _mm_add_ps(const5, _mm_mul_ps(x2, const6))))));
315 
316  cVal = _mm_div_ps(a, b);
317 
318  _mm_storeu_ps(cPtr, cVal); // Store the results back into the C container
319 
320  aPtr += 4;
321  cPtr += 4;
322  }
323 
324  number = quarterPoints * 4;
325  volk_32f_tanh_32f_series(cPtr, aPtr, num_points - number);
326 }
327 #endif /* LV_HAVE_SSE */
328 
329 
330 #ifdef LV_HAVE_AVX
331 #include <immintrin.h>
332 
333 static inline void
334 volk_32f_tanh_32f_u_avx(float* cVector, const float* aVector, unsigned int num_points)
335 {
336  unsigned int number = 0;
337  const unsigned int eighthPoints = num_points / 8;
338 
339  float* cPtr = cVector;
340  const float* aPtr = aVector;
341 
342  __m256 aVal, cVal, x2, a, b;
343  __m256 const1, const2, const3, const4, const5, const6;
344  const1 = _mm256_set1_ps(135135.0f);
345  const2 = _mm256_set1_ps(17325.0f);
346  const3 = _mm256_set1_ps(378.0f);
347  const4 = _mm256_set1_ps(62370.0f);
348  const5 = _mm256_set1_ps(3150.0f);
349  const6 = _mm256_set1_ps(28.0f);
350  for (; number < eighthPoints; number++) {
351 
352  aVal = _mm256_loadu_ps(aPtr);
353  x2 = _mm256_mul_ps(aVal, aVal);
354  a = _mm256_mul_ps(
355  aVal,
356  _mm256_add_ps(
357  const1,
358  _mm256_mul_ps(
359  x2,
360  _mm256_add_ps(const2,
361  _mm256_mul_ps(x2, _mm256_add_ps(const3, x2))))));
362  b = _mm256_add_ps(
363  const1,
364  _mm256_mul_ps(
365  x2,
366  _mm256_add_ps(
367  const4,
368  _mm256_mul_ps(x2,
369  _mm256_add_ps(const5, _mm256_mul_ps(x2, const6))))));
370 
371  cVal = _mm256_div_ps(a, b);
372 
373  _mm256_storeu_ps(cPtr, cVal); // Store the results back into the C container
374 
375  aPtr += 8;
376  cPtr += 8;
377  }
378 
379  number = eighthPoints * 8;
380  volk_32f_tanh_32f_series(cPtr, aPtr, num_points - number);
381 }
382 #endif /* LV_HAVE_AVX */
383 
384 #if LV_HAVE_AVX && LV_HAVE_FMA
385 #include <immintrin.h>
386 
387 static inline void
388 volk_32f_tanh_32f_u_avx_fma(float* cVector, const float* aVector, unsigned int num_points)
389 {
390  unsigned int number = 0;
391  const unsigned int eighthPoints = num_points / 8;
392 
393  float* cPtr = cVector;
394  const float* aPtr = aVector;
395 
396  __m256 aVal, cVal, x2, a, b;
397  __m256 const1, const2, const3, const4, const5, const6;
398  const1 = _mm256_set1_ps(135135.0f);
399  const2 = _mm256_set1_ps(17325.0f);
400  const3 = _mm256_set1_ps(378.0f);
401  const4 = _mm256_set1_ps(62370.0f);
402  const5 = _mm256_set1_ps(3150.0f);
403  const6 = _mm256_set1_ps(28.0f);
404  for (; number < eighthPoints; number++) {
405 
406  aVal = _mm256_loadu_ps(aPtr);
407  x2 = _mm256_mul_ps(aVal, aVal);
408  a = _mm256_mul_ps(
409  aVal,
410  _mm256_fmadd_ps(
411  x2, _mm256_fmadd_ps(x2, _mm256_add_ps(const3, x2), const2), const1));
412  b = _mm256_fmadd_ps(
413  x2, _mm256_fmadd_ps(x2, _mm256_fmadd_ps(x2, const6, const5), const4), const1);
414 
415  cVal = _mm256_div_ps(a, b);
416 
417  _mm256_storeu_ps(cPtr, cVal); // Store the results back into the C container
418 
419  aPtr += 8;
420  cPtr += 8;
421  }
422 
423  number = eighthPoints * 8;
424  volk_32f_tanh_32f_series(cPtr, aPtr, num_points - number);
425 }
426 #endif /* LV_HAVE_AVX && LV_HAVE_FMA */
427 
428 #endif /* INCLUDED_volk_32f_tanh_32f_u_H */
static void volk_32f_tanh_32f_u_avx(float *cVector, const float *aVector, unsigned int num_points)
Definition: volk_32f_tanh_32f.h:334
static void volk_32f_tanh_32f_generic(float *cVector, const float *aVector, unsigned int num_points)
Definition: volk_32f_tanh_32f.h:80
static void volk_32f_tanh_32f_a_avx(float *cVector, const float *aVector, unsigned int num_points)
Definition: volk_32f_tanh_32f.h:173
static void volk_32f_tanh_32f_a_sse(float *cVector, const float *aVector, unsigned int num_points)
Definition: volk_32f_tanh_32f.h:122
static void volk_32f_tanh_32f_series(float *cVector, const float *aVector, unsigned int num_points)
Definition: volk_32f_tanh_32f.h:96
static void volk_32f_tanh_32f_u_sse(float *cVector, const float *aVector, unsigned int num_points)
Definition: volk_32f_tanh_32f.h:283