Vector Optimized Library of Kernels  2.2
Architecture-tuned implementations of math kernels
volk_32f_s32f_stddev_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 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_s32f_stddev_32f_a_H
69 #define INCLUDED_volk_32f_s32f_stddev_32f_a_H
70 
71 #include <inttypes.h>
72 #include <math.h>
73 #include <stdio.h>
74 #include <volk/volk_common.h>
75 
76 #ifdef LV_HAVE_SSE4_1
77 #include <smmintrin.h>
78 
79 static inline void volk_32f_s32f_stddev_32f_a_sse4_1(float* stddev,
80  const float* inputBuffer,
81  const float mean,
82  unsigned int num_points)
83 {
84  float returnValue = 0;
85  if (num_points > 0) {
86  unsigned int number = 0;
87  const unsigned int sixteenthPoints = num_points / 16;
88 
89  const float* aPtr = inputBuffer;
90 
91  __VOLK_ATTR_ALIGNED(16) float squareBuffer[4];
92 
93  __m128 squareAccumulator = _mm_setzero_ps();
94  __m128 aVal1, aVal2, aVal3, aVal4;
95  __m128 cVal1, cVal2, cVal3, cVal4;
96  for (; number < sixteenthPoints; number++) {
97  aVal1 = _mm_load_ps(aPtr);
98  aPtr += 4;
99  cVal1 = _mm_dp_ps(aVal1, aVal1, 0xF1);
100 
101  aVal2 = _mm_load_ps(aPtr);
102  aPtr += 4;
103  cVal2 = _mm_dp_ps(aVal2, aVal2, 0xF2);
104 
105  aVal3 = _mm_load_ps(aPtr);
106  aPtr += 4;
107  cVal3 = _mm_dp_ps(aVal3, aVal3, 0xF4);
108 
109  aVal4 = _mm_load_ps(aPtr);
110  aPtr += 4;
111  cVal4 = _mm_dp_ps(aVal4, aVal4, 0xF8);
112 
113  cVal1 = _mm_or_ps(cVal1, cVal2);
114  cVal3 = _mm_or_ps(cVal3, cVal4);
115  cVal1 = _mm_or_ps(cVal1, cVal3);
116 
117  squareAccumulator =
118  _mm_add_ps(squareAccumulator, cVal1); // squareAccumulator += x^2
119  }
120  _mm_store_ps(squareBuffer,
121  squareAccumulator); // Store the results back into the C container
122  returnValue = squareBuffer[0];
123  returnValue += squareBuffer[1];
124  returnValue += squareBuffer[2];
125  returnValue += squareBuffer[3];
126 
127  number = sixteenthPoints * 16;
128  for (; number < num_points; number++) {
129  returnValue += (*aPtr) * (*aPtr);
130  aPtr++;
131  }
132  returnValue /= num_points;
133  returnValue -= (mean * mean);
134  returnValue = sqrtf(returnValue);
135  }
136  *stddev = returnValue;
137 }
138 
139 #endif /* LV_HAVE_SSE4_1 */
140 
141 #ifdef LV_HAVE_SSE
142 #include <xmmintrin.h>
143 
144 static inline void volk_32f_s32f_stddev_32f_a_sse(float* stddev,
145  const float* inputBuffer,
146  const float mean,
147  unsigned int num_points)
148 {
149  float returnValue = 0;
150  if (num_points > 0) {
151  unsigned int number = 0;
152  const unsigned int quarterPoints = num_points / 4;
153 
154  const float* aPtr = inputBuffer;
155 
156  __VOLK_ATTR_ALIGNED(16) float squareBuffer[4];
157 
158  __m128 squareAccumulator = _mm_setzero_ps();
159  __m128 aVal = _mm_setzero_ps();
160  for (; number < quarterPoints; number++) {
161  aVal = _mm_load_ps(aPtr); // aVal = x
162  aVal = _mm_mul_ps(aVal, aVal); // squareAccumulator += x^2
163  squareAccumulator = _mm_add_ps(squareAccumulator, aVal);
164  aPtr += 4;
165  }
166  _mm_store_ps(squareBuffer,
167  squareAccumulator); // Store the results back into the C container
168  returnValue = squareBuffer[0];
169  returnValue += squareBuffer[1];
170  returnValue += squareBuffer[2];
171  returnValue += squareBuffer[3];
172 
173  number = quarterPoints * 4;
174  for (; number < num_points; number++) {
175  returnValue += (*aPtr) * (*aPtr);
176  aPtr++;
177  }
178  returnValue /= num_points;
179  returnValue -= (mean * mean);
180  returnValue = sqrtf(returnValue);
181  }
182  *stddev = returnValue;
183 }
184 #endif /* LV_HAVE_SSE */
185 
186 
187 #ifdef LV_HAVE_AVX
188 #include <immintrin.h>
189 
190 static inline void volk_32f_s32f_stddev_32f_a_avx(float* stddev,
191  const float* inputBuffer,
192  const float mean,
193  unsigned int num_points)
194 {
195  float stdDev = 0;
196  if (num_points > 0) {
197  unsigned int number = 0;
198  const unsigned int thirtySecondthPoints = num_points / 32;
199 
200  const float* aPtr = inputBuffer;
201  __VOLK_ATTR_ALIGNED(32) float squareBuffer[8];
202 
203  __m256 squareAccumulator = _mm256_setzero_ps();
204  __m256 aVal1, aVal2, aVal3, aVal4;
205  __m256 cVal1, cVal2, cVal3, cVal4;
206  for (; number < thirtySecondthPoints; number++) {
207  aVal1 = _mm256_load_ps(aPtr);
208  aPtr += 8;
209  cVal1 = _mm256_dp_ps(aVal1, aVal1, 0xF1);
210 
211  aVal2 = _mm256_load_ps(aPtr);
212  aPtr += 8;
213  cVal2 = _mm256_dp_ps(aVal2, aVal2, 0xF2);
214 
215  aVal3 = _mm256_load_ps(aPtr);
216  aPtr += 8;
217  cVal3 = _mm256_dp_ps(aVal3, aVal3, 0xF4);
218 
219  aVal4 = _mm256_load_ps(aPtr);
220  aPtr += 8;
221  cVal4 = _mm256_dp_ps(aVal4, aVal4, 0xF8);
222 
223  cVal1 = _mm256_or_ps(cVal1, cVal2);
224  cVal3 = _mm256_or_ps(cVal3, cVal4);
225  cVal1 = _mm256_or_ps(cVal1, cVal3);
226 
227  squareAccumulator =
228  _mm256_add_ps(squareAccumulator, cVal1); // squareAccumulator += x^2
229  }
230  _mm256_store_ps(squareBuffer,
231  squareAccumulator); // Store the results back into the C container
232  stdDev = squareBuffer[0];
233  stdDev += squareBuffer[1];
234  stdDev += squareBuffer[2];
235  stdDev += squareBuffer[3];
236  stdDev += squareBuffer[4];
237  stdDev += squareBuffer[5];
238  stdDev += squareBuffer[6];
239  stdDev += squareBuffer[7];
240 
241  number = thirtySecondthPoints * 32;
242  for (; number < num_points; number++) {
243  stdDev += (*aPtr) * (*aPtr);
244  aPtr++;
245  }
246  stdDev /= num_points;
247  stdDev -= (mean * mean);
248  stdDev = sqrtf(stdDev);
249  }
250  *stddev = stdDev;
251 }
252 #endif /* LV_HAVE_AVX */
253 
254 
255 #ifdef LV_HAVE_GENERIC
256 
257 static inline void volk_32f_s32f_stddev_32f_generic(float* stddev,
258  const float* inputBuffer,
259  const float mean,
260  unsigned int num_points)
261 {
262  float returnValue = 0;
263  if (num_points > 0) {
264  const float* aPtr = inputBuffer;
265  unsigned int number = 0;
266 
267  for (number = 0; number < num_points; number++) {
268  returnValue += (*aPtr) * (*aPtr);
269  aPtr++;
270  }
271 
272  returnValue /= num_points;
273  returnValue -= (mean * mean);
274  returnValue = sqrtf(returnValue);
275  }
276  *stddev = returnValue;
277 }
278 
279 #endif /* LV_HAVE_GENERIC */
280 
281 
282 #endif /* INCLUDED_volk_32f_s32f_stddev_32f_a_H */
283 
284 #ifndef INCLUDED_volk_32f_s32f_stddev_32f_u_H
285 #define INCLUDED_volk_32f_s32f_stddev_32f_u_H
286 
287 #include <inttypes.h>
288 #include <math.h>
289 #include <stdio.h>
290 #include <volk/volk_common.h>
291 
292 #ifdef LV_HAVE_AVX
293 #include <immintrin.h>
294 
295 static inline void volk_32f_s32f_stddev_32f_u_avx(float* stddev,
296  const float* inputBuffer,
297  const float mean,
298  unsigned int num_points)
299 {
300  float stdDev = 0;
301  if (num_points > 0) {
302  unsigned int number = 0;
303  const unsigned int thirtySecondthPoints = num_points / 32;
304 
305  const float* aPtr = inputBuffer;
306  __VOLK_ATTR_ALIGNED(32) float squareBuffer[8];
307 
308  __m256 squareAccumulator = _mm256_setzero_ps();
309  __m256 aVal1, aVal2, aVal3, aVal4;
310  __m256 cVal1, cVal2, cVal3, cVal4;
311  for (; number < thirtySecondthPoints; number++) {
312  aVal1 = _mm256_loadu_ps(aPtr);
313  aPtr += 8;
314  cVal1 = _mm256_dp_ps(aVal1, aVal1, 0xF1);
315 
316  aVal2 = _mm256_loadu_ps(aPtr);
317  aPtr += 8;
318  cVal2 = _mm256_dp_ps(aVal2, aVal2, 0xF2);
319 
320  aVal3 = _mm256_loadu_ps(aPtr);
321  aPtr += 8;
322  cVal3 = _mm256_dp_ps(aVal3, aVal3, 0xF4);
323 
324  aVal4 = _mm256_loadu_ps(aPtr);
325  aPtr += 8;
326  cVal4 = _mm256_dp_ps(aVal4, aVal4, 0xF8);
327 
328  cVal1 = _mm256_or_ps(cVal1, cVal2);
329  cVal3 = _mm256_or_ps(cVal3, cVal4);
330  cVal1 = _mm256_or_ps(cVal1, cVal3);
331 
332  squareAccumulator =
333  _mm256_add_ps(squareAccumulator, cVal1); // squareAccumulator += x^2
334  }
335  _mm256_storeu_ps(
336  squareBuffer,
337  squareAccumulator); // Store the results back into the C container
338  stdDev = squareBuffer[0];
339  stdDev += squareBuffer[1];
340  stdDev += squareBuffer[2];
341  stdDev += squareBuffer[3];
342  stdDev += squareBuffer[4];
343  stdDev += squareBuffer[5];
344  stdDev += squareBuffer[6];
345  stdDev += squareBuffer[7];
346 
347  number = thirtySecondthPoints * 32;
348  for (; number < num_points; number++) {
349  stdDev += (*aPtr) * (*aPtr);
350  aPtr++;
351  }
352  stdDev /= num_points;
353  stdDev -= (mean * mean);
354  stdDev = sqrtf(stdDev);
355  }
356  *stddev = stdDev;
357 }
358 #endif /* LV_HAVE_AVX */
359 
360 #endif /* INCLUDED_volk_32f_s32f_stddev_32f_u_H */
static void volk_32f_s32f_stddev_32f_u_avx(float *stddev, const float *inputBuffer, const float mean, unsigned int num_points)
Definition: volk_32f_s32f_stddev_32f.h:295
static void volk_32f_s32f_stddev_32f_a_sse(float *stddev, const float *inputBuffer, const float mean, unsigned int num_points)
Definition: volk_32f_s32f_stddev_32f.h:144
static void volk_32f_s32f_stddev_32f_generic(float *stddev, const float *inputBuffer, const float mean, unsigned int num_points)
Definition: volk_32f_s32f_stddev_32f.h:257
#define __VOLK_ATTR_ALIGNED(x)
Definition: volk_common.h:56
static void volk_32f_s32f_stddev_32f_a_avx(float *stddev, const float *inputBuffer, const float mean, unsigned int num_points)
Definition: volk_32f_s32f_stddev_32f.h:190