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 #include "mem.h"
26 
27 static void vector_fmul_c(float *dst, const float *src0, const float *src1,
28  int len)
29 {
30  int i;
31  for (i = 0; i < len; i++)
32  dst[i] = src0[i] * src1[i];
33 }
34 
35 static void vector_fmac_scalar_c(float *dst, const float *src, float mul,
36  int len)
37 {
38  int i;
39  for (i = 0; i < len; i++)
40  dst[i] += src[i] * mul;
41 }
42 
43 static void vector_fmul_scalar_c(float *dst, const float *src, float mul,
44  int len)
45 {
46  int i;
47  for (i = 0; i < len; i++)
48  dst[i] = src[i] * mul;
49 }
50 
51 static void vector_dmul_scalar_c(double *dst, const double *src, double mul,
52  int len)
53 {
54  int i;
55  for (i = 0; i < len; i++)
56  dst[i] = src[i] * mul;
57 }
58 
59 static void vector_fmul_window_c(float *dst, const float *src0,
60  const float *src1, const float *win, int len)
61 {
62  int i, j;
63 
64  dst += len;
65  win += len;
66  src0 += len;
67 
68  for (i = -len, j = len - 1; i < 0; i++, j--) {
69  float s0 = src0[i];
70  float s1 = src1[j];
71  float wi = win[i];
72  float wj = win[j];
73  dst[i] = s0 * wj - s1 * wi;
74  dst[j] = s0 * wi + s1 * wj;
75  }
76 }
77 
78 static void vector_fmul_add_c(float *dst, const float *src0, const float *src1,
79  const float *src2, int len){
80  int i;
81 
82  for (i = 0; i < len; i++)
83  dst[i] = src0[i] * src1[i] + src2[i];
84 }
85 
86 static void vector_fmul_reverse_c(float *dst, const float *src0,
87  const float *src1, int len)
88 {
89  int i;
90 
91  src1 += len-1;
92  for (i = 0; i < len; i++)
93  dst[i] = src0[i] * src1[-i];
94 }
95 
96 static void butterflies_float_c(float *av_restrict v1, float *av_restrict v2,
97  int len)
98 {
99  int i;
100 
101  for (i = 0; i < len; i++) {
102  float t = v1[i] - v2[i];
103  v1[i] += v2[i];
104  v2[i] = t;
105  }
106 }
107 
108 float avpriv_scalarproduct_float_c(const float *v1, const float *v2, int len)
109 {
110  float p = 0.0;
111  int i;
112 
113  for (i = 0; i < len; i++)
114  p += v1[i] * v2[i];
115 
116  return p;
117 }
118 
120 {
122  if (!fdsp)
123  return NULL;
124 
125  fdsp->vector_fmul = vector_fmul_c;
134 
135  if (ARCH_AARCH64)
137  if (ARCH_ARM)
138  ff_float_dsp_init_arm(fdsp);
139  if (ARCH_PPC)
140  ff_float_dsp_init_ppc(fdsp, bit_exact);
141  if (ARCH_X86)
142  ff_float_dsp_init_x86(fdsp);
143  if (ARCH_MIPS)
145  return fdsp;
146 }
147 
148 
149 #ifdef TEST
150 
151 #include <float.h>
152 #include <math.h>
153 #include <stdint.h>
154 #include <stdlib.h>
155 #include <string.h>
156 #if HAVE_UNISTD_H
157 #include <unistd.h> /* for getopt */
158 #endif
159 #if !HAVE_GETOPT
160 #include "compat/getopt.c"
161 #endif
162 
163 #include "common.h"
164 #include "cpu.h"
165 #include "internal.h"
166 #include "lfg.h"
167 #include "log.h"
168 #include "random_seed.h"
169 
170 #define LEN 240
171 
172 static void fill_float_array(AVLFG *lfg, float *a, int len)
173 {
174  int i;
175  double bmg[2], stddev = 10.0, mean = 0.0;
176 
177  for (i = 0; i < len; i += 2) {
178  av_bmg_get(lfg, bmg);
179  a[i] = bmg[0] * stddev + mean;
180  a[i + 1] = bmg[1] * stddev + mean;
181  }
182 }
183 static int compare_floats(const float *a, const float *b, int len,
184  float max_diff)
185 {
186  int i;
187  for (i = 0; i < len; i++) {
188  if (fabsf(a[i] - b[i]) > max_diff) {
189  av_log(NULL, AV_LOG_ERROR, "%d: %- .12f - %- .12f = % .12g\n",
190  i, a[i], b[i], a[i] - b[i]);
191  return -1;
192  }
193  }
194  return 0;
195 }
196 
197 static void fill_double_array(AVLFG *lfg, double *a, int len)
198 {
199  int i;
200  double bmg[2], stddev = 10.0, mean = 0.0;
201 
202  for (i = 0; i < len; i += 2) {
203  av_bmg_get(lfg, bmg);
204  a[i] = bmg[0] * stddev + mean;
205  a[i + 1] = bmg[1] * stddev + mean;
206  }
207 }
208 
209 static int compare_doubles(const double *a, const double *b, int len,
210  double max_diff)
211 {
212  int i;
213 
214  for (i = 0; i < len; i++) {
215  if (fabs(a[i] - b[i]) > max_diff) {
216  av_log(NULL, AV_LOG_ERROR, "%d: %- .12f - %- .12f = % .12g\n",
217  i, a[i], b[i], a[i] - b[i]);
218  return -1;
219  }
220  }
221  return 0;
222 }
223 
224 static int test_vector_fmul(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
225  const float *v1, const float *v2)
226 {
227  LOCAL_ALIGNED(32, float, cdst, [LEN]);
228  LOCAL_ALIGNED(32, float, odst, [LEN]);
229  int ret;
230 
231  cdsp->vector_fmul(cdst, v1, v2, LEN);
232  fdsp->vector_fmul(odst, v1, v2, LEN);
233 
234  if (ret = compare_floats(cdst, odst, LEN, FLT_EPSILON))
235  av_log(NULL, AV_LOG_ERROR, "vector_fmul failed\n");
236 
237  return ret;
238 }
239 
240 #define ARBITRARY_FMAC_SCALAR_CONST 0.005
241 static int test_vector_fmac_scalar(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
242  const float *v1, const float *src0, float scale)
243 {
244  LOCAL_ALIGNED(32, float, cdst, [LEN]);
245  LOCAL_ALIGNED(32, float, odst, [LEN]);
246  int ret;
247 
248  memcpy(cdst, v1, LEN * sizeof(*v1));
249  memcpy(odst, v1, LEN * sizeof(*v1));
250 
251  cdsp->vector_fmac_scalar(cdst, src0, scale, LEN);
252  fdsp->vector_fmac_scalar(odst, src0, scale, LEN);
253 
254  if (ret = compare_floats(cdst, odst, LEN, ARBITRARY_FMAC_SCALAR_CONST))
255  av_log(NULL, AV_LOG_ERROR, "vector_fmac_scalar failed\n");
256 
257  return ret;
258 }
259 
260 static int test_vector_fmul_scalar(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
261  const float *v1, float scale)
262 {
263  LOCAL_ALIGNED(32, float, cdst, [LEN]);
264  LOCAL_ALIGNED(32, float, odst, [LEN]);
265  int ret;
266 
267  cdsp->vector_fmul_scalar(cdst, v1, scale, LEN);
268  fdsp->vector_fmul_scalar(odst, v1, scale, LEN);
269 
270  if (ret = compare_floats(cdst, odst, LEN, FLT_EPSILON))
271  av_log(NULL, AV_LOG_ERROR, "vector_fmul_scalar failed\n");
272 
273  return ret;
274 }
275 
276 static int test_vector_dmul_scalar(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
277  const double *v1, double scale)
278 {
279  LOCAL_ALIGNED(32, double, cdst, [LEN]);
280  LOCAL_ALIGNED(32, double, odst, [LEN]);
281  int ret;
282 
283  cdsp->vector_dmul_scalar(cdst, v1, scale, LEN);
284  fdsp->vector_dmul_scalar(odst, v1, scale, LEN);
285 
286  if (ret = compare_doubles(cdst, odst, LEN, DBL_EPSILON))
287  av_log(NULL, AV_LOG_ERROR, "vector_dmul_scalar failed\n");
288 
289  return ret;
290 }
291 
292 #define ARBITRARY_FMUL_WINDOW_CONST 0.008
293 static int test_vector_fmul_window(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
294  const float *v1, const float *v2, const float *v3)
295 {
296  LOCAL_ALIGNED(32, float, cdst, [LEN]);
297  LOCAL_ALIGNED(32, float, odst, [LEN]);
298  int ret;
299 
300  cdsp->vector_fmul_window(cdst, v1, v2, v3, LEN / 2);
301  fdsp->vector_fmul_window(odst, v1, v2, v3, LEN / 2);
302 
303  if (ret = compare_floats(cdst, odst, LEN, ARBITRARY_FMUL_WINDOW_CONST))
304  av_log(NULL, AV_LOG_ERROR, "vector_fmul_window failed\n");
305 
306  return ret;
307 }
308 
309 #define ARBITRARY_FMUL_ADD_CONST 0.005
310 static int test_vector_fmul_add(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
311  const float *v1, const float *v2, const float *v3)
312 {
313  LOCAL_ALIGNED(32, float, cdst, [LEN]);
314  LOCAL_ALIGNED(32, float, odst, [LEN]);
315  int ret;
316 
317  cdsp->vector_fmul_add(cdst, v1, v2, v3, LEN);
318  fdsp->vector_fmul_add(odst, v1, v2, v3, LEN);
319 
320  if (ret = compare_floats(cdst, odst, LEN, ARBITRARY_FMUL_ADD_CONST))
321  av_log(NULL, AV_LOG_ERROR, "vector_fmul_add failed\n");
322 
323  return ret;
324 }
325 
326 static int test_vector_fmul_reverse(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
327  const float *v1, const float *v2)
328 {
329  LOCAL_ALIGNED(32, float, cdst, [LEN]);
330  LOCAL_ALIGNED(32, float, odst, [LEN]);
331  int ret;
332 
333  cdsp->vector_fmul_reverse(cdst, v1, v2, LEN);
334  fdsp->vector_fmul_reverse(odst, v1, v2, LEN);
335 
336  if (ret = compare_floats(cdst, odst, LEN, FLT_EPSILON))
337  av_log(NULL, AV_LOG_ERROR, "vector_fmul_reverse failed\n");
338 
339  return ret;
340 }
341 
342 static int test_butterflies_float(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
343  const float *v1, const float *v2)
344 {
345  LOCAL_ALIGNED(32, float, cv1, [LEN]);
346  LOCAL_ALIGNED(32, float, cv2, [LEN]);
347  LOCAL_ALIGNED(32, float, ov1, [LEN]);
348  LOCAL_ALIGNED(32, float, ov2, [LEN]);
349  int ret;
350 
351  memcpy(cv1, v1, LEN * sizeof(*v1));
352  memcpy(cv2, v2, LEN * sizeof(*v2));
353  memcpy(ov1, v1, LEN * sizeof(*v1));
354  memcpy(ov2, v2, LEN * sizeof(*v2));
355 
356  cdsp->butterflies_float(cv1, cv2, LEN);
357  fdsp->butterflies_float(ov1, ov2, LEN);
358 
359  if ((ret = compare_floats(cv1, ov1, LEN, FLT_EPSILON)) ||
360  (ret = compare_floats(cv2, ov2, LEN, FLT_EPSILON)))
361  av_log(NULL, AV_LOG_ERROR, "butterflies_float failed\n");
362 
363  return ret;
364 }
365 
366 #define ARBITRARY_SCALARPRODUCT_CONST 0.2
367 static int test_scalarproduct_float(AVFloatDSPContext *fdsp, AVFloatDSPContext *cdsp,
368  const float *v1, const float *v2)
369 {
370  float cprod, oprod;
371  int ret;
372 
373  cprod = cdsp->scalarproduct_float(v1, v2, LEN);
374  oprod = fdsp->scalarproduct_float(v1, v2, LEN);
375 
376  if (ret = compare_floats(&cprod, &oprod, 1, ARBITRARY_SCALARPRODUCT_CONST))
377  av_log(NULL, AV_LOG_ERROR, "scalarproduct_float failed\n");
378 
379  return ret;
380 }
381 
382 int main(int argc, char **argv)
383 {
384  int ret = 0, seeded = 0;
385  uint32_t seed;
386  AVFloatDSPContext *fdsp, *cdsp;
387  AVLFG lfg;
388 
389  LOCAL_ALIGNED(32, float, src0, [LEN]);
390  LOCAL_ALIGNED(32, float, src1, [LEN]);
391  LOCAL_ALIGNED(32, float, src2, [LEN]);
392  LOCAL_ALIGNED(32, double, dbl_src0, [LEN]);
393  LOCAL_ALIGNED(32, double, dbl_src1, [LEN]);
394 
395  for (;;) {
396  int arg = getopt(argc, argv, "s:c:");
397  if (arg == -1)
398  break;
399  switch (arg) {
400  case 's':
401  seed = strtoul(optarg, NULL, 10);
402  seeded = 1;
403  break;
404  case 'c':
405  {
406  int cpuflags = av_get_cpu_flags();
407 
408  if (av_parse_cpu_caps(&cpuflags, optarg) < 0)
409  return 1;
410 
411  av_force_cpu_flags(cpuflags);
412  break;
413  }
414  }
415  }
416  if (!seeded)
417  seed = av_get_random_seed();
418 
419  av_log(NULL, AV_LOG_INFO, "float_dsp-test: %s %u\n", seeded ? "seed" : "random seed", seed);
420 
421  fdsp = avpriv_float_dsp_alloc(1);
423  cdsp = avpriv_float_dsp_alloc(1);
424 
425  if (!fdsp || !cdsp) {
426  ret = 1;
427  goto end;
428  }
429 
430  av_lfg_init(&lfg, seed);
431 
432  fill_float_array(&lfg, src0, LEN);
433  fill_float_array(&lfg, src1, LEN);
434  fill_float_array(&lfg, src2, LEN);
435 
436  fill_double_array(&lfg, dbl_src0, LEN);
437  fill_double_array(&lfg, dbl_src1, LEN);
438 
439  if (test_vector_fmul(fdsp, cdsp, src0, src1))
440  ret -= 1 << 0;
441  if (test_vector_fmac_scalar(fdsp, cdsp, src2, src0, src1[0]))
442  ret -= 1 << 1;
443  if (test_vector_fmul_scalar(fdsp, cdsp, src0, src1[0]))
444  ret -= 1 << 2;
445  if (test_vector_fmul_window(fdsp, cdsp, src0, src1, src2))
446  ret -= 1 << 3;
447  if (test_vector_fmul_add(fdsp, cdsp, src0, src1, src2))
448  ret -= 1 << 4;
449  if (test_vector_fmul_reverse(fdsp, cdsp, src0, src1))
450  ret -= 1 << 5;
451  if (test_butterflies_float(fdsp, cdsp, src0, src1))
452  ret -= 1 << 6;
453  if (test_scalarproduct_float(fdsp, cdsp, src0, src1))
454  ret -= 1 << 7;
455  if (test_vector_dmul_scalar(fdsp, cdsp, dbl_src0, dbl_src1[0]))
456  ret -= 1 << 8;
457 
458 end:
459  av_freep(&fdsp);
460  av_freep(&cdsp);
461  return ret;
462 }
463 
464 #endif /* TEST */
Definition: lfg.h:25
#define NULL
Definition: coverity.c:32
void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp)
static void vector_fmul_scalar_c(float *dst, const float *src, float mul, int len)
Definition: float_dsp.c:43
memory handling functions
const char * b
Definition: vf_curves.c:109
float(* scalarproduct_float)(const float *v1, const float *v2, int len)
Calculate the scalar product of two vectors of floats.
Definition: float_dsp.h:159
void(* vector_fmul_reverse)(float *dst, const float *src0, const float *src1, int len)
Calculate the entry wise product of two vectors of floats, and store the result in a vector of floats...
Definition: float_dsp.h:138
void ff_float_dsp_init_mips(AVFloatDSPContext *fdsp)
Macro definitions for various function/variable attributes.
void(* vector_fmac_scalar)(float *dst, const float *src, float mul, int len)
Multiply a vector of floats by a scalar float and add to destination vector.
Definition: float_dsp.h:54
float avpriv_scalarproduct_float_c(const float *v1, const float *v2, int len)
Return the scalar product of two vectors.
Definition: float_dsp.c:108
void(* vector_fmul_window)(float *dst, const float *src0, const float *src1, const float *win, int len)
Overlap/add with window function.
Definition: float_dsp.h:103
av_cold void ff_float_dsp_init_aarch64(AVFloatDSPContext *fdsp)
av_cold void ff_float_dsp_init_arm(AVFloatDSPContext *fdsp)
#define av_cold
Definition: attributes.h:82
static void butterflies_float_c(float *av_restrict v1, float *av_restrict v2, int len)
Definition: float_dsp.c:96
static av_cold int end(AVCodecContext *avctx)
Definition: avrndec.c:90
void av_bmg_get(AVLFG *lfg, double out[2])
Get the next two numbers generated by a Box-Muller Gaussian generator using the random numbers issued...
Definition: lfg.c:47
#define av_log(a,...)
#define AV_LOG_ERROR
Something went wrong and cannot losslessly be recovered.
Definition: log.h:176
void(* vector_fmul)(float *dst, const float *src0, const float *src1, int len)
Calculate the entry wise product of two vectors of floats and store the result in a vector of floats...
Definition: float_dsp.h:38
void(* butterflies_float)(float *av_restrict v1, float *av_restrict v2, int len)
Calculate the sum and difference of two vectors of floats.
Definition: float_dsp.h:148
#define s0
Definition: regdef.h:37
const char * arg
Definition: jacosubdec.c:66
static void vector_fmul_add_c(float *dst, const float *src0, const float *src1, const float *src2, int len)
Definition: float_dsp.c:78
void(* vector_dmul_scalar)(double *dst, const double *src, double mul, int len)
Multiply a vector of double by a scalar double.
Definition: float_dsp.h:84
int av_parse_cpu_caps(unsigned *flags, const char *s)
Parse CPU caps from a string and update the given AV_CPU_* flags based on that.
Definition: cpu.c:181
static void vector_dmul_scalar_c(double *dst, const double *src, double mul, int len)
Definition: float_dsp.c:51
void(* vector_fmul_scalar)(float *dst, const float *src, float mul, int len)
Multiply a vector of floats by a scalar float.
Definition: float_dsp.h:69
#define src
Definition: vp9dsp.c:530
#define src1
Definition: h264pred.c:139
#define AV_LOG_INFO
Standard information.
Definition: log.h:187
void ff_float_dsp_init_ppc(AVFloatDSPContext *fdsp, int strict)
static unsigned int seed
Definition: videogen.c:78
static int getopt(int argc, char *argv[], char *opts)
Definition: getopt.c:41
void(* vector_fmul_add)(float *dst, const float *src0, const float *src1, const float *src2, int len)
Calculate the entry wise product of two vectors of floats, add a third vector of floats and store the...
Definition: float_dsp.h:121
#define LEN
#define src0
Definition: h264pred.c:138
av_cold void av_lfg_init(AVLFG *c, unsigned int seed)
Definition: lfg.c:30
#define s1
Definition: regdef.h:38
av_cold AVFloatDSPContext * avpriv_float_dsp_alloc(int bit_exact)
Allocate a float DSP context.
Definition: float_dsp.c:119
int av_get_cpu_flags(void)
Return the flags which specify extensions supported by the CPU.
Definition: cpu.c:76
static void vector_fmul_window_c(float *dst, const float *src0, const float *src1, const float *win, int len)
Definition: float_dsp.c:59
common internal and external API header
#define LOCAL_ALIGNED(a, t, v,...)
Definition: internal.h:110
static void vector_fmul_c(float *dst, const float *src0, const float *src1, int len)
Definition: float_dsp.c:27
static char * optarg
Definition: getopt.c:39
static void vector_fmac_scalar_c(float *dst, const float *src, float mul, int len)
Definition: float_dsp.c:35
int len
#define av_freep(p)
uint32_t av_get_random_seed(void)
Get a seed to use in conjunction with random functions.
Definition: random_seed.c:114
void av_force_cpu_flags(int arg)
Disables cpu detection and forces the specified flags.
Definition: cpu.c:49
int main(int argc, char **argv)
Definition: main.c:22
void * av_mallocz(size_t size)
Allocate a block of size bytes with alignment suitable for all memory accesses (including vectors if ...
Definition: mem.c:252
static void vector_fmul_reverse_c(float *dst, const float *src0, const float *src1, int len)
Definition: float_dsp.c:86