FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
float_dsp.c
Go to the documentation of this file.
1 /*
2  * Copyright 2005 Balatoni Denes
3  * Copyright 2006 Loren Merritt
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg 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 GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21 
22 #include "config.h"
23 #include "attributes.h"
24 #include "float_dsp.h"
25 
26 static void vector_fmul_c(float *dst, const float *src0, const float *src1,
27  int len)
28 {
29  int i;
30  for (i = 0; i < len; i++)
31  dst[i] = src0[i] * src1[i];
32 }
33 
34 static void vector_fmac_scalar_c(float *dst, const float *src, float mul,
35  int len)
36 {
37  int i;
38  for (i = 0; i < len; i++)
39  dst[i] += src[i] * mul;
40 }
41 
42 static void vector_fmul_scalar_c(float *dst, const float *src, float mul,
43  int len)
44 {
45  int i;
46  for (i = 0; i < len; i++)
47  dst[i] = src[i] * mul;
48 }
49 
50 static void vector_dmul_scalar_c(double *dst, const double *src, double mul,
51  int len)
52 {
53  int i;
54  for (i = 0; i < len; i++)
55  dst[i] = src[i] * mul;
56 }
57 
58 static void vector_fmul_window_c(float *dst, const float *src0,
59  const float *src1, const float *win, int len)
60 {
61  int i, j;
62 
63  dst += len;
64  win += len;
65  src0 += len;
66 
67  for (i = -len, j = len - 1; i < 0; i++, j--) {
68  float s0 = src0[i];
69  float s1 = src1[j];
70  float wi = win[i];
71  float wj = win[j];
72  dst[i] = s0 * wj - s1 * wi;
73  dst[j] = s0 * wi + s1 * wj;
74  }
75 }
76 
77 static void vector_fmul_add_c(float *dst, const float *src0, const float *src1,
78  const float *src2, int len){
79  int i;
80 
81  for (i = 0; i < len; i++)
82  dst[i] = src0[i] * src1[i] + src2[i];
83 }
84 
85 static void vector_fmul_reverse_c(float *dst, const float *src0,
86  const float *src1, int len)
87 {
88  int i;
89 
90  src1 += len-1;
91  for (i = 0; i < len; i++)
92  dst[i] = src0[i] * src1[-i];
93 }
94 
95 static void butterflies_float_c(float *av_restrict v1, float *av_restrict v2,
96  int len)
97 {
98  int i;
99 
100  for (i = 0; i < len; i++) {
101  float t = v1[i] - v2[i];
102  v1[i] += v2[i];
103  v2[i] = t;
104  }
105 }
106 
107 float avpriv_scalarproduct_float_c(const float *v1, const float *v2, int len)
108 {
109  float p = 0.0;
110  int i;
111 
112  for (i = 0; i < len; i++)
113  p += v1[i] * v2[i];
114 
115  return p;
116 }
117 
119 {
120  fdsp->vector_fmul = vector_fmul_c;
129 
130  if (ARCH_AARCH64)
132  if (ARCH_ARM)
133  ff_float_dsp_init_arm(fdsp);
134  if (ARCH_PPC)
135  ff_float_dsp_init_ppc(fdsp, bit_exact);
136  if (ARCH_X86)
137  ff_float_dsp_init_x86(fdsp);
138  if (ARCH_MIPS)
140 }
141 
142 #ifdef TEST
143 
144 #include <float.h>
145 #include <math.h>
146 #include <stdint.h>
147 #include <stdlib.h>
148 #include <string.h>
149 #if HAVE_UNISTD_H
150 #include <unistd.h> /* for getopt */
151 #endif
152 #if !HAVE_GETOPT
153 #include "compat/getopt.c"
154 #endif
155 
156 #include "common.h"
157 #include "cpu.h"
158 #include "internal.h"
159 #include "lfg.h"
160 #include "log.h"
161 #include "mem.h"
162 #include "random_seed.h"
163 
164 #define LEN 240
165 
166 static void fill_float_array(AVLFG *lfg, float *a, int len)
167 {
168  int i;
169  double bmg[2], stddev = 10.0, mean = 0.0;
170 
171  for (i = 0; i < len; i += 2) {
172  av_bmg_get(lfg, bmg);
173  a[i] = bmg[0] * stddev + mean;
174  a[i + 1] = bmg[1] * stddev + mean;
175  }
176 }
177 static int compare_floats(const float *a, const float *b, int len,
178  float max_diff)
179 {
180  int i;
181  for (i = 0; i < len; i++) {
182  if (fabsf(a[i] - b[i]) > max_diff) {
183  av_log(NULL, AV_LOG_ERROR, "%d: %- .12f - %- .12f = % .12g\n",
184  i, a[i], b[i], a[i] - b[i]);
185  return -1;
186  }
187  }
188  return 0;
189 }
190 
191 static void fill_double_array(AVLFG *lfg, double *a, int len)
192 {
193  int i;
194  double bmg[2], stddev = 10.0, mean = 0.0;
195 
196  for (i = 0; i < len; i += 2) {
197  av_bmg_get(lfg, bmg);
198  a[i] = bmg[0] * stddev + mean;
199  a[i + 1] = bmg[1] * stddev + mean;
200  }
201 }
202 
203 static int compare_doubles(const double *a, const double *b, int len,
204  double max_diff)
205 {
206  int i;
207 
208  for (i = 0; i < len; i++) {
209  if (fabs(a[i] - b[i]) > max_diff) {
210  av_log(NULL, AV_LOG_ERROR, "%d: %- .12f - %- .12f = % .12g\n",
211  i, a[i], b[i], a[i] - b[i]);
212  return -1;
213  }
214  }
215  return 0;
216 }
217 
218 static int test_vector_fmul(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
219  const float *v1, const float *v2)
220 {
221  LOCAL_ALIGNED(32, float, cdst, [LEN]);
222  LOCAL_ALIGNED(32, float, odst, [LEN]);
223  int ret;
224 
225  cdsp->vector_fmul(cdst, v1, v2, LEN);
226  fdsp->vector_fmul(odst, v1, v2, LEN);
227 
228  if (ret = compare_floats(cdst, odst, LEN, FLT_EPSILON))
229  av_log(NULL, AV_LOG_ERROR, "vector_fmul failed\n");
230 
231  return ret;
232 }
233 
234 #define ARBITRARY_FMAC_SCALAR_CONST 0.005
235 static int test_vector_fmac_scalar(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
236  const float *v1, const float *src0, float scale)
237 {
238  LOCAL_ALIGNED(32, float, cdst, [LEN]);
239  LOCAL_ALIGNED(32, float, odst, [LEN]);
240  int ret;
241 
242  memcpy(cdst, v1, LEN * sizeof(*v1));
243  memcpy(odst, v1, LEN * sizeof(*v1));
244 
245  cdsp->vector_fmac_scalar(cdst, src0, scale, LEN);
246  fdsp->vector_fmac_scalar(odst, src0, scale, LEN);
247 
248  if (ret = compare_floats(cdst, odst, LEN, ARBITRARY_FMAC_SCALAR_CONST))
249  av_log(NULL, AV_LOG_ERROR, "vector_fmac_scalar failed\n");
250 
251  return ret;
252 }
253 
254 static int test_vector_fmul_scalar(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
255  const float *v1, float scale)
256 {
257  LOCAL_ALIGNED(32, float, cdst, [LEN]);
258  LOCAL_ALIGNED(32, float, odst, [LEN]);
259  int ret;
260 
261  cdsp->vector_fmul_scalar(cdst, v1, scale, LEN);
262  fdsp->vector_fmul_scalar(odst, v1, scale, LEN);
263 
264  if (ret = compare_floats(cdst, odst, LEN, FLT_EPSILON))
265  av_log(NULL, AV_LOG_ERROR, "vector_fmul_scalar failed\n");
266 
267  return ret;
268 }
269 
270 static int test_vector_dmul_scalar(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
271  const double *v1, double scale)
272 {
273  LOCAL_ALIGNED(32, double, cdst, [LEN]);
274  LOCAL_ALIGNED(32, double, odst, [LEN]);
275  int ret;
276 
277  cdsp->vector_dmul_scalar(cdst, v1, scale, LEN);
278  fdsp->vector_dmul_scalar(odst, v1, scale, LEN);
279 
280  if (ret = compare_doubles(cdst, odst, LEN, DBL_EPSILON))
281  av_log(NULL, AV_LOG_ERROR, "vector_dmul_scalar failed\n");
282 
283  return ret;
284 }
285 
286 #define ARBITRARY_FMUL_WINDOW_CONST 0.008
287 static int test_vector_fmul_window(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
288  const float *v1, const float *v2, const float *v3)
289 {
290  LOCAL_ALIGNED(32, float, cdst, [LEN]);
291  LOCAL_ALIGNED(32, float, odst, [LEN]);
292  int ret;
293 
294  cdsp->vector_fmul_window(cdst, v1, v2, v3, LEN / 2);
295  fdsp->vector_fmul_window(odst, v1, v2, v3, LEN / 2);
296 
297  if (ret = compare_floats(cdst, odst, LEN, ARBITRARY_FMUL_WINDOW_CONST))
298  av_log(NULL, AV_LOG_ERROR, "vector_fmul_window failed\n");
299 
300  return ret;
301 }
302 
303 #define ARBITRARY_FMUL_ADD_CONST 0.005
304 static int test_vector_fmul_add(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
305  const float *v1, const float *v2, const float *v3)
306 {
307  LOCAL_ALIGNED(32, float, cdst, [LEN]);
308  LOCAL_ALIGNED(32, float, odst, [LEN]);
309  int ret;
310 
311  cdsp->vector_fmul_add(cdst, v1, v2, v3, LEN);
312  fdsp->vector_fmul_add(odst, v1, v2, v3, LEN);
313 
314  if (ret = compare_floats(cdst, odst, LEN, ARBITRARY_FMUL_ADD_CONST))
315  av_log(NULL, AV_LOG_ERROR, "vector_fmul_add failed\n");
316 
317  return ret;
318 }
319 
320 static int test_vector_fmul_reverse(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
321  const float *v1, const float *v2)
322 {
323  LOCAL_ALIGNED(32, float, cdst, [LEN]);
324  LOCAL_ALIGNED(32, float, odst, [LEN]);
325  int ret;
326 
327  cdsp->vector_fmul_reverse(cdst, v1, v2, LEN);
328  fdsp->vector_fmul_reverse(odst, v1, v2, LEN);
329 
330  if (ret = compare_floats(cdst, odst, LEN, FLT_EPSILON))
331  av_log(NULL, AV_LOG_ERROR, "vector_fmul_reverse failed\n");
332 
333  return ret;
334 }
335 
336 static int test_butterflies_float(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
337  const float *v1, const float *v2)
338 {
339  LOCAL_ALIGNED(32, float, cv1, [LEN]);
340  LOCAL_ALIGNED(32, float, cv2, [LEN]);
341  LOCAL_ALIGNED(32, float, ov1, [LEN]);
342  LOCAL_ALIGNED(32, float, ov2, [LEN]);
343  int ret;
344 
345  memcpy(cv1, v1, LEN * sizeof(*v1));
346  memcpy(cv2, v2, LEN * sizeof(*v2));
347  memcpy(ov1, v1, LEN * sizeof(*v1));
348  memcpy(ov2, v2, LEN * sizeof(*v2));
349 
350  cdsp->butterflies_float(cv1, cv2, LEN);
351  fdsp->butterflies_float(ov1, ov2, LEN);
352 
353  if ((ret = compare_floats(cv1, ov1, LEN, FLT_EPSILON)) ||
354  (ret = compare_floats(cv2, ov2, LEN, FLT_EPSILON)))
355  av_log(NULL, AV_LOG_ERROR, "butterflies_float failed\n");
356 
357  return ret;
358 }
359 
360 #define ARBITRARY_SCALARPRODUCT_CONST 0.2
361 static int test_scalarproduct_float(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
362  const float *v1, const float *v2)
363 {
364  float cprod, oprod;
365  int ret;
366 
367  cprod = cdsp->scalarproduct_float(v1, v2, LEN);
368  oprod = fdsp->scalarproduct_float(v1, v2, LEN);
369 
370  if (ret = compare_floats(&cprod, &oprod, 1, ARBITRARY_SCALARPRODUCT_CONST))
371  av_log(NULL, AV_LOG_ERROR, "scalarproduct_float failed\n");
372 
373  return ret;
374 }
375 
376 int main(int argc, char **argv)
377 {
378  int ret = 0, seeded = 0;
379  uint32_t seed;
380  AVFloatDSPContext fdsp, cdsp;
381  AVLFG lfg;
382 
383  LOCAL_ALIGNED(32, float, src0, [LEN]);
384  LOCAL_ALIGNED(32, float, src1, [LEN]);
385  LOCAL_ALIGNED(32, float, src2, [LEN]);
386  LOCAL_ALIGNED(32, double, dbl_src0, [LEN]);
387  LOCAL_ALIGNED(32, double, dbl_src1, [LEN]);
388 
389  for (;;) {
390  int arg = getopt(argc, argv, "s:c:");
391  if (arg == -1)
392  break;
393  switch (arg) {
394  case 's':
395  seed = strtoul(optarg, NULL, 10);
396  seeded = 1;
397  break;
398  case 'c':
399  {
400  int cpuflags = av_get_cpu_flags();
401 
402  if (av_parse_cpu_caps(&cpuflags, optarg) < 0)
403  return 1;
404 
405  av_force_cpu_flags(cpuflags);
406  break;
407  }
408  }
409  }
410  if (!seeded)
411  seed = av_get_random_seed();
412 
413  av_log(NULL, AV_LOG_INFO, "float_dsp-test: %s %u\n", seeded ? "seed" : "random seed", seed);
414 
415  av_lfg_init(&lfg, seed);
416 
417  fill_float_array(&lfg, src0, LEN);
418  fill_float_array(&lfg, src1, LEN);
419  fill_float_array(&lfg, src2, LEN);
420 
421  fill_double_array(&lfg, dbl_src0, LEN);
422  fill_double_array(&lfg, dbl_src1, LEN);
423 
424  avpriv_float_dsp_init(&fdsp, 1);
426  avpriv_float_dsp_init(&cdsp, 1);
427 
428  if (test_vector_fmul(&fdsp, &cdsp, src0, src1))
429  ret -= 1 << 0;
430  if (test_vector_fmac_scalar(&fdsp, &cdsp, src2, src0, src1[0]))
431  ret -= 1 << 1;
432  if (test_vector_fmul_scalar(&fdsp, &cdsp, src0, src1[0]))
433  ret -= 1 << 2;
434  if (test_vector_fmul_window(&fdsp, &cdsp, src0, src1, src2))
435  ret -= 1 << 3;
436  if (test_vector_fmul_add(&fdsp, &cdsp, src0, src1, src2))
437  ret -= 1 << 4;
438  if (test_vector_fmul_reverse(&fdsp, &cdsp, src0, src1))
439  ret -= 1 << 5;
440  if (test_butterflies_float(&fdsp, &cdsp, src0, src1))
441  ret -= 1 << 6;
442  if (test_scalarproduct_float(&fdsp, &cdsp, src0, src1))
443  ret -= 1 << 7;
444  if (test_vector_dmul_scalar(&fdsp, &cdsp, dbl_src0, dbl_src1[0]))
445  ret -= 1 << 8;
446 
447  return ret;
448 }
449 
450 #endif /* TEST */