Mercurial > libavcodec.hg
annotate dct-test.c @ 5109:8d4f1823d98d libavcodec
Fix linking when GPL code has been disabled.
author | diego |
---|---|
date | Thu, 07 Jun 2007 15:54:49 +0000 |
parents | 6ea82ac05f60 |
children | 8fe26e4fd918 |
rev | line source |
---|---|
3699
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
1 /* |
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
2 * (c) 2001 Fabrice Bellard |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
3 * 2007 Marc Hoffman <marc.hoffman@analog.com> |
3699
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
4 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3699
diff
changeset
|
5 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3699
diff
changeset
|
6 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3699
diff
changeset
|
7 * FFmpeg is free software; you can redistribute it and/or |
3699
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
8 * modify it under the terms of the GNU Lesser General Public |
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
9 * License as published by the Free Software Foundation; either |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3699
diff
changeset
|
10 * version 2.1 of the License, or (at your option) any later version. |
3699
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
11 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3699
diff
changeset
|
12 * FFmpeg is distributed in the hope that it will be useful, |
3699
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
15 * Lesser General Public License for more details. |
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
16 * |
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
17 * You should have received a copy of the GNU Lesser General Public |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3699
diff
changeset
|
18 * License along with FFmpeg; if not, write to the Free Software |
3699
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
20 */ |
c537a97eec66
Add official LGPL license headers to the files that were missing them.
diego
parents:
3398
diff
changeset
|
21 |
1106 | 22 /** |
23 * @file dct-test.c | |
2967 | 24 * DCT test. (c) 2001 Fabrice Bellard. |
1106 | 25 * Started from sample code by Juan J. Sierralta P. |
26 */ | |
27 | |
0 | 28 #include <stdlib.h> |
29 #include <stdio.h> | |
30 #include <string.h> | |
31 #include <sys/time.h> | |
32 #include <unistd.h> | |
33 | |
34 #include "dsputil.h" | |
35 | |
633 | 36 #include "simple_idct.h" |
1557 | 37 #include "faandct.h" |
33 | 38 |
980 | 39 #ifndef MAX |
40 #define MAX(a, b) (((a) > (b)) ? (a) : (b)) | |
41 #endif | |
42 | |
2872 | 43 #undef printf |
44 | |
45 void *fast_memcpy(void *a, const void *b, size_t c){return memcpy(a,b,c);}; | |
46 | |
33 | 47 /* reference fdct/idct */ |
0 | 48 extern void fdct(DCTELEM *block); |
33 | 49 extern void idct(DCTELEM *block); |
2872 | 50 extern void ff_idct_xvid_mmx(DCTELEM *block); |
51 extern void ff_idct_xvid_mmx2(DCTELEM *block); | |
0 | 52 extern void init_fdct(); |
53 | |
33 | 54 extern void ff_mmx_idct(DCTELEM *data); |
55 extern void ff_mmxext_idct(DCTELEM *data); | |
56 | |
633 | 57 extern void odivx_idct_c (short *block); |
58 | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
59 // BFIN |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
60 extern void ff_bfin_idct (DCTELEM *block) ; |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
61 extern void ff_bfin_fdct (DCTELEM *block) ; |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
62 |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
63 // ALTIVEC |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
64 extern void fdct_altivec (DCTELEM *block); |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
65 //extern void idct_altivec (DCTELEM *block);?? no routine |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
66 |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
67 |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
68 struct algo { |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
69 char *name; |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
70 enum { FDCT, IDCT } is_idct; |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
71 void (* func) (DCTELEM *block); |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
72 void (* ref) (DCTELEM *block); |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
73 enum formattag { NO_PERM,MMX_PERM, MMX_SIMPLE_PERM, SCALE_PERM } format; |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
74 }; |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
75 |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
76 #ifndef FAAN_POSTSCALE |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
77 #define FAAN_SCALE SCALE_PERM |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
78 #else |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
79 #define FAAN_SCALE NO_PERM |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
80 #endif |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
81 |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
82 #define DCT_ERROR(name,is_idct,func,ref,form) {name,is_idct,func,ref,form} |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
83 |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
84 |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
85 struct algo algos[] = { |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
86 DCT_ERROR( "REF-DBL", 0, fdct, fdct, NO_PERM), |
5107 | 87 DCT_ERROR("FAAN", 0, ff_faandct, fdct, FAAN_SCALE), |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
88 DCT_ERROR("IJG-AAN-INT", 0, fdct_ifast, fdct, SCALE_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
89 DCT_ERROR("IJG-LLM-INT", 0, ff_jpeg_fdct_islow, fdct, NO_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
90 DCT_ERROR("REF-DBL", 1, idct, idct, NO_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
91 DCT_ERROR("INT", 1, j_rev_dct, idct, MMX_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
92 DCT_ERROR("SIMPLE-C", 1, simple_idct, idct, NO_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
93 |
5108 | 94 #ifdef HAVE_MMX |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
95 DCT_ERROR("MMX", 0, ff_fdct_mmx, fdct, NO_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
96 DCT_ERROR("MMX2", 0, ff_fdct_mmx2, fdct, NO_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
97 |
5109 | 98 #ifdef CONFIG_GPL |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
99 DCT_ERROR("LIBMPEG2-MMX", 1, ff_mmx_idct, idct, MMX_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
100 DCT_ERROR("LIBMPEG2-MMXEXT", 1, ff_mmxext_idct, idct, MMX_PERM), |
5109 | 101 #endif |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
102 DCT_ERROR("SIMPLE-MMX", 1, ff_simple_idct_mmx, idct, MMX_SIMPLE_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
103 DCT_ERROR("XVID-MMX", 1, ff_idct_xvid_mmx, idct, NO_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
104 DCT_ERROR("XVID-MMX2", 1, ff_idct_xvid_mmx2, idct, NO_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
105 #endif |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
106 |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
107 #ifdef HAVE_ALTIVEC |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
108 DCT_ERROR("altivecfdct", 0, fdct_altivec, fdct, NO_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
109 #endif |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
110 |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
111 #ifdef ARCH_BFIN |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
112 DCT_ERROR("BFINfdct", 0, ff_bfin_fdct, fdct, NO_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
113 DCT_ERROR("BFINidct", 1, ff_bfin_idct, idct, NO_PERM), |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
114 #endif |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
115 |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
116 { 0 } |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
117 }; |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
118 |
0 | 119 #define AANSCALE_BITS 12 |
120 static const unsigned short aanscales[64] = { | |
121 /* precomputed values scaled up by 14 bits */ | |
122 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520, | |
123 22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270, | |
124 21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906, | |
125 19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315, | |
126 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520, | |
127 12873, 17855, 16819, 15137, 12873, 10114, 6967, 3552, | |
128 8867, 12299, 11585, 10426, 8867, 6967, 4799, 2446, | |
129 4520, 6270, 5906, 5315, 4520, 3552, 2446, 1247 | |
130 }; | |
131 | |
4197 | 132 uint8_t cropTbl[256 + 2 * MAX_NEG_CROP]; |
633 | 133 |
1064 | 134 int64_t gettime(void) |
0 | 135 { |
136 struct timeval tv; | |
137 gettimeofday(&tv,NULL); | |
1064 | 138 return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec; |
0 | 139 } |
140 | |
141 #define NB_ITS 20000 | |
142 #define NB_ITS_SPEED 50000 | |
143 | |
33 | 144 static short idct_mmx_perm[64]; |
145 | |
633 | 146 static short idct_simple_mmx_perm[64]={ |
2979 | 147 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D, |
148 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D, | |
149 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D, | |
150 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F, | |
151 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F, | |
152 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D, | |
153 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F, | |
154 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F, | |
633 | 155 }; |
156 | |
33 | 157 void idct_mmx_init(void) |
158 { | |
159 int i; | |
160 | |
161 /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */ | |
162 for (i = 0; i < 64; i++) { | |
2979 | 163 idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2); |
164 // idct_simple_mmx_perm[i] = simple_block_permute_op(i); | |
33 | 165 } |
166 } | |
167 | |
168 static DCTELEM block[64] __attribute__ ((aligned (8))); | |
169 static DCTELEM block1[64] __attribute__ ((aligned (8))); | |
633 | 170 static DCTELEM block_org[64] __attribute__ ((aligned (8))); |
33 | 171 |
172 void dct_error(const char *name, int is_idct, | |
173 void (*fdct_func)(DCTELEM *block), | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
174 void (*fdct_ref)(DCTELEM *block), int form, int test) |
0 | 175 { |
176 int it, i, scale; | |
177 int err_inf, v; | |
1064 | 178 int64_t err2, ti, ti1, it1; |
179 int64_t sysErr[64], sysErrMax=0; | |
633 | 180 int maxout=0; |
181 int blockSumErrMax=0, blockSumErr; | |
0 | 182 |
183 srandom(0); | |
184 | |
185 err_inf = 0; | |
186 err2 = 0; | |
633 | 187 for(i=0; i<64; i++) sysErr[i]=0; |
0 | 188 for(it=0;it<NB_ITS;it++) { |
633 | 189 for(i=0;i<64;i++) |
190 block1[i] = 0; | |
191 switch(test){ | |
2967 | 192 case 0: |
633 | 193 for(i=0;i<64;i++) |
194 block1[i] = (random() % 512) -256; | |
635 | 195 if (is_idct){ |
633 | 196 fdct(block1); |
635 | 197 |
198 for(i=0;i<64;i++) | |
199 block1[i]>>=3; | |
200 } | |
633 | 201 break; |
202 case 1:{ | |
203 int num= (random()%10)+1; | |
204 for(i=0;i<num;i++) | |
205 block1[random()%64] = (random() % 512) -256; | |
206 }break; | |
207 case 2: | |
208 block1[0]= (random()%4096)-2048; | |
209 block1[63]= (block1[0]&1)^1; | |
210 break; | |
211 } | |
33 | 212 |
633 | 213 #if 0 // simulate mismatch control |
214 { int sum=0; | |
215 for(i=0;i<64;i++) | |
216 sum+=block1[i]; | |
217 | |
2967 | 218 if((sum&1)==0) block1[63]^=1; |
633 | 219 } |
220 #endif | |
221 | |
222 for(i=0; i<64; i++) | |
223 block_org[i]= block1[i]; | |
33 | 224 |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
225 if (form == MMX_PERM) { |
633 | 226 for(i=0;i<64;i++) |
33 | 227 block[idct_mmx_perm[i]] = block1[i]; |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
228 } else if (form == MMX_SIMPLE_PERM) { |
633 | 229 for(i=0;i<64;i++) |
230 block[idct_simple_mmx_perm[i]] = block1[i]; | |
231 | |
2979 | 232 } else { |
633 | 233 for(i=0; i<64; i++) |
234 block[i]= block1[i]; | |
33 | 235 } |
633 | 236 #if 0 // simulate mismatch control for tested IDCT but not the ref |
237 { int sum=0; | |
238 for(i=0;i<64;i++) | |
239 sum+=block[i]; | |
240 | |
2967 | 241 if((sum&1)==0) block[63]^=1; |
633 | 242 } |
243 #endif | |
33 | 244 |
0 | 245 fdct_func(block); |
4752
d49f8b3e8c4a
Replace emms() with emms_c(), patch by Marc Hoffman, mmh pleasantst com.
diego
parents:
4197
diff
changeset
|
246 emms_c(); /* for ff_mmx_idct */ |
33 | 247 |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
248 if (form == SCALE_PERM) { |
0 | 249 for(i=0; i<64; i++) { |
635 | 250 scale = 8*(1 << (AANSCALE_BITS + 11)) / aanscales[i]; |
633 | 251 block[i] = (block[i] * scale /*+ (1<<(AANSCALE_BITS-1))*/) >> AANSCALE_BITS; |
252 } | |
253 } | |
254 | |
33 | 255 fdct_ref(block1); |
0 | 256 |
633 | 257 blockSumErr=0; |
0 | 258 for(i=0;i<64;i++) { |
259 v = abs(block[i] - block1[i]); | |
260 if (v > err_inf) | |
261 err_inf = v; | |
262 err2 += v * v; | |
2979 | 263 sysErr[i] += block[i] - block1[i]; |
264 blockSumErr += v; | |
265 if( abs(block[i])>maxout) maxout=abs(block[i]); | |
0 | 266 } |
633 | 267 if(blockSumErrMax < blockSumErr) blockSumErrMax= blockSumErr; |
268 #if 0 // print different matrix pairs | |
269 if(blockSumErr){ | |
270 printf("\n"); | |
271 for(i=0; i<64; i++){ | |
272 if((i&7)==0) printf("\n"); | |
273 printf("%4d ", block_org[i]); | |
274 } | |
275 for(i=0; i<64; i++){ | |
276 if((i&7)==0) printf("\n"); | |
277 printf("%4d ", block[i] - block1[i]); | |
278 } | |
279 } | |
280 #endif | |
0 | 281 } |
4001 | 282 for(i=0; i<64; i++) sysErrMax= MAX(sysErrMax, FFABS(sysErr[i])); |
2967 | 283 |
633 | 284 #if 1 // dump systematic errors |
285 for(i=0; i<64; i++){ | |
2979 | 286 if(i%8==0) printf("\n"); |
633 | 287 printf("%5d ", (int)sysErr[i]); |
288 } | |
289 printf("\n"); | |
290 #endif | |
2967 | 291 |
633 | 292 printf("%s %s: err_inf=%d err2=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n", |
33 | 293 is_idct ? "IDCT" : "DCT", |
633 | 294 name, err_inf, (double)err2 / NB_ITS / 64.0, (double)sysErrMax / NB_ITS, maxout, blockSumErrMax); |
295 #if 1 //Speed test | |
0 | 296 /* speed test */ |
633 | 297 for(i=0;i<64;i++) |
298 block1[i] = 0; | |
299 switch(test){ | |
2967 | 300 case 0: |
633 | 301 for(i=0;i<64;i++) |
302 block1[i] = (random() % 512) -256; | |
635 | 303 if (is_idct){ |
633 | 304 fdct(block1); |
635 | 305 |
306 for(i=0;i<64;i++) | |
307 block1[i]>>=3; | |
308 } | |
633 | 309 break; |
310 case 1:{ | |
311 case 2: | |
312 block1[0] = (random() % 512) -256; | |
313 block1[1] = (random() % 512) -256; | |
314 block1[2] = (random() % 512) -256; | |
315 block1[3] = (random() % 512) -256; | |
316 }break; | |
317 } | |
0 | 318 |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
319 if (form == MMX_PERM) { |
633 | 320 for(i=0;i<64;i++) |
33 | 321 block[idct_mmx_perm[i]] = block1[i]; |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
322 } else if(form == MMX_SIMPLE_PERM) { |
633 | 323 for(i=0;i<64;i++) |
324 block[idct_simple_mmx_perm[i]] = block1[i]; | |
325 } else { | |
326 for(i=0; i<64; i++) | |
327 block[i]= block1[i]; | |
33 | 328 } |
329 | |
0 | 330 ti = gettime(); |
331 it1 = 0; | |
332 do { | |
333 for(it=0;it<NB_ITS_SPEED;it++) { | |
633 | 334 for(i=0; i<64; i++) |
335 block[i]= block1[i]; | |
336 // memcpy(block, block1, sizeof(DCTELEM) * 64); | |
337 // dont memcpy especially not fastmemcpy because it does movntq !!! | |
0 | 338 fdct_func(block); |
339 } | |
340 it1 += NB_ITS_SPEED; | |
341 ti1 = gettime() - ti; | |
342 } while (ti1 < 1000000); | |
4752
d49f8b3e8c4a
Replace emms() with emms_c(), patch by Marc Hoffman, mmh pleasantst com.
diego
parents:
4197
diff
changeset
|
343 emms_c(); |
0 | 344 |
633 | 345 printf("%s %s: %0.1f kdct/s\n", |
33 | 346 is_idct ? "IDCT" : "DCT", |
0 | 347 name, (double)it1 * 1000.0 / (double)ti1); |
633 | 348 #endif |
0 | 349 } |
350 | |
1064 | 351 static uint8_t img_dest[64] __attribute__ ((aligned (8))); |
352 static uint8_t img_dest1[64] __attribute__ ((aligned (8))); | |
720 | 353 |
1064 | 354 void idct248_ref(uint8_t *dest, int linesize, int16_t *block) |
720 | 355 { |
356 static int init; | |
357 static double c8[8][8]; | |
358 static double c4[4][4]; | |
359 double block1[64], block2[64], block3[64]; | |
360 double s, sum, v; | |
361 int i, j, k; | |
362 | |
363 if (!init) { | |
364 init = 1; | |
365 | |
366 for(i=0;i<8;i++) { | |
367 sum = 0; | |
368 for(j=0;j<8;j++) { | |
369 s = (i==0) ? sqrt(1.0/8.0) : sqrt(1.0/4.0); | |
370 c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0); | |
371 sum += c8[i][j] * c8[i][j]; | |
372 } | |
373 } | |
2967 | 374 |
720 | 375 for(i=0;i<4;i++) { |
376 sum = 0; | |
377 for(j=0;j<4;j++) { | |
378 s = (i==0) ? sqrt(1.0/4.0) : sqrt(1.0/2.0); | |
379 c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0); | |
380 sum += c4[i][j] * c4[i][j]; | |
381 } | |
382 } | |
383 } | |
384 | |
385 /* butterfly */ | |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
386 s = 0.5 * sqrt(2.0); |
720 | 387 for(i=0;i<4;i++) { |
388 for(j=0;j<8;j++) { | |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
389 block1[8*(2*i)+j] = (block[8*(2*i)+j] + block[8*(2*i+1)+j]) * s; |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
390 block1[8*(2*i+1)+j] = (block[8*(2*i)+j] - block[8*(2*i+1)+j]) * s; |
720 | 391 } |
392 } | |
393 | |
394 /* idct8 on lines */ | |
395 for(i=0;i<8;i++) { | |
396 for(j=0;j<8;j++) { | |
397 sum = 0; | |
398 for(k=0;k<8;k++) | |
399 sum += c8[k][j] * block1[8*i+k]; | |
400 block2[8*i+j] = sum; | |
401 } | |
402 } | |
403 | |
404 /* idct4 */ | |
405 for(i=0;i<8;i++) { | |
406 for(j=0;j<4;j++) { | |
407 /* top */ | |
408 sum = 0; | |
409 for(k=0;k<4;k++) | |
410 sum += c4[k][j] * block2[8*(2*k)+i]; | |
411 block3[8*(2*j)+i] = sum; | |
412 | |
413 /* bottom */ | |
414 sum = 0; | |
415 for(k=0;k<4;k++) | |
416 sum += c4[k][j] * block2[8*(2*k+1)+i]; | |
417 block3[8*(2*j+1)+i] = sum; | |
418 } | |
419 } | |
420 | |
421 /* clamp and store the result */ | |
422 for(i=0;i<8;i++) { | |
423 for(j=0;j<8;j++) { | |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
424 v = block3[8*i+j]; |
720 | 425 if (v < 0) |
426 v = 0; | |
427 else if (v > 255) | |
428 v = 255; | |
429 dest[i * linesize + j] = (int)rint(v); | |
430 } | |
431 } | |
432 } | |
433 | |
2967 | 434 void idct248_error(const char *name, |
1064 | 435 void (*idct248_put)(uint8_t *dest, int line_size, int16_t *block)) |
720 | 436 { |
437 int it, i, it1, ti, ti1, err_max, v; | |
438 | |
439 srandom(0); | |
2967 | 440 |
720 | 441 /* just one test to see if code is correct (precision is less |
442 important here) */ | |
443 err_max = 0; | |
444 for(it=0;it<NB_ITS;it++) { | |
2967 | 445 |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
446 /* XXX: use forward transform to generate values */ |
720 | 447 for(i=0;i<64;i++) |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
448 block1[i] = (random() % 256) - 128; |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
449 block1[0] += 1024; |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
450 |
720 | 451 for(i=0; i<64; i++) |
452 block[i]= block1[i]; | |
453 idct248_ref(img_dest1, 8, block); | |
2967 | 454 |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
455 for(i=0; i<64; i++) |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
456 block[i]= block1[i]; |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
457 idct248_put(img_dest, 8, block); |
2967 | 458 |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
459 for(i=0;i<64;i++) { |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
460 v = abs((int)img_dest[i] - (int)img_dest1[i]); |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
461 if (v == 255) |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
462 printf("%d %d\n", img_dest[i], img_dest1[i]); |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
463 if (v > err_max) |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
464 err_max = v; |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
465 } |
720 | 466 #if 0 |
467 printf("ref=\n"); | |
468 for(i=0;i<8;i++) { | |
469 int j; | |
470 for(j=0;j<8;j++) { | |
471 printf(" %3d", img_dest1[i*8+j]); | |
472 } | |
473 printf("\n"); | |
474 } | |
2967 | 475 |
720 | 476 printf("out=\n"); |
477 for(i=0;i<8;i++) { | |
478 int j; | |
479 for(j=0;j<8;j++) { | |
480 printf(" %3d", img_dest[i*8+j]); | |
481 } | |
482 printf("\n"); | |
483 } | |
484 #endif | |
485 } | |
486 printf("%s %s: err_inf=%d\n", | |
487 1 ? "IDCT248" : "DCT248", | |
488 name, err_max); | |
489 | |
490 ti = gettime(); | |
491 it1 = 0; | |
492 do { | |
493 for(it=0;it<NB_ITS_SPEED;it++) { | |
494 for(i=0; i<64; i++) | |
495 block[i]= block1[i]; | |
496 // memcpy(block, block1, sizeof(DCTELEM) * 64); | |
497 // dont memcpy especially not fastmemcpy because it does movntq !!! | |
498 idct248_put(img_dest, 8, block); | |
499 } | |
500 it1 += NB_ITS_SPEED; | |
501 ti1 = gettime() - ti; | |
502 } while (ti1 < 1000000); | |
4752
d49f8b3e8c4a
Replace emms() with emms_c(), patch by Marc Hoffman, mmh pleasantst com.
diego
parents:
4197
diff
changeset
|
503 emms_c(); |
720 | 504 |
505 printf("%s %s: %0.1f kdct/s\n", | |
506 1 ? "IDCT248" : "DCT248", | |
507 name, (double)it1 * 1000.0 / (double)ti1); | |
508 } | |
509 | |
33 | 510 void help(void) |
511 { | |
633 | 512 printf("dct-test [-i] [<test-number>]\n" |
513 "test-number 0 -> test with random matrixes\n" | |
514 " 1 -> test with random sparse matrixes\n" | |
515 " 2 -> do 3. test from mpeg4 std\n" | |
720 | 516 "-i test IDCT implementations\n" |
517 "-4 test IDCT248 implementations\n"); | |
33 | 518 } |
519 | |
0 | 520 int main(int argc, char **argv) |
521 { | |
720 | 522 int test_idct = 0, test_248_dct = 0; |
633 | 523 int c,i; |
524 int test=1; | |
33 | 525 |
0 | 526 init_fdct(); |
33 | 527 idct_mmx_init(); |
0 | 528 |
4197 | 529 for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i; |
633 | 530 for(i=0;i<MAX_NEG_CROP;i++) { |
4197 | 531 cropTbl[i] = 0; |
532 cropTbl[i + MAX_NEG_CROP + 256] = 255; | |
633 | 533 } |
2967 | 534 |
33 | 535 for(;;) { |
720 | 536 c = getopt(argc, argv, "ih4"); |
33 | 537 if (c == -1) |
538 break; | |
539 switch(c) { | |
540 case 'i': | |
541 test_idct = 1; | |
542 break; | |
720 | 543 case '4': |
544 test_248_dct = 1; | |
545 break; | |
633 | 546 default : |
33 | 547 case 'h': |
548 help(); | |
4754 | 549 return 0; |
33 | 550 } |
551 } | |
2967 | 552 |
633 | 553 if(optind <argc) test= atoi(argv[optind]); |
2967 | 554 |
33 | 555 printf("ffmpeg DCT/IDCT test\n"); |
556 | |
720 | 557 if (test_248_dct) { |
558 idct248_error("SIMPLE-C", simple_idct248_put); | |
33 | 559 } else { |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
560 for (i=0;algos[i].name;i++) |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
561 if (algos[i].is_idct == test_idct) { |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
562 dct_error (algos[i].name, algos[i].is_idct, algos[i].func, algos[i].ref, algos[i].format, test); |
720 | 563 } |
33 | 564 } |
0 | 565 return 0; |
566 } |