Commit | Line | Data |
---|---|---|
04d7f601 DB |
1 | /* |
2 | * (c) 2001 Fabrice Bellard | |
3ac35bdb | 3 | * 2007 Marc Hoffman <marc.hoffman@analog.com> |
04d7f601 | 4 | * |
b78e7197 DB |
5 | * This file is part of FFmpeg. |
6 | * | |
7 | * FFmpeg is free software; you can redistribute it and/or | |
04d7f601 DB |
8 | * modify it under the terms of the GNU Lesser General Public |
9 | * License as published by the Free Software Foundation; either | |
b78e7197 | 10 | * version 2.1 of the License, or (at your option) any later version. |
04d7f601 | 11 | * |
b78e7197 | 12 | * FFmpeg is distributed in the hope that it will be useful, |
04d7f601 DB |
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 | |
b78e7197 | 18 | * License along with FFmpeg; if not, write to the Free Software |
04d7f601 DB |
19 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
20 | */ | |
21 | ||
983e3246 | 22 | /** |
bad5537e | 23 | * @file libavcodec/dct-test.c |
94f694a4 | 24 | * DCT test (c) 2001 Fabrice Bellard |
983e3246 MN |
25 | * Started from sample code by Juan J. Sierralta P. |
26 | */ | |
27 | ||
de6d9b64 FB |
28 | #include <stdlib.h> |
29 | #include <stdio.h> | |
30 | #include <string.h> | |
31 | #include <sys/time.h> | |
32 | #include <unistd.h> | |
12807c8d | 33 | #include <math.h> |
de6d9b64 | 34 | |
ae32e509 | 35 | #include "libavutil/common.h" |
294eaa26 | 36 | #include "libavutil/lfg.h" |
de6d9b64 | 37 | |
86748dbc | 38 | #include "simple_idct.h" |
10ac3618 | 39 | #include "aandcttab.h" |
65e4c8c9 | 40 | #include "faandct.h" |
6f08c541 | 41 | #include "faanidct.h" |
a6493a8f | 42 | #include "x86/idct_xvid.h" |
9e1586fc | 43 | |
434df899 MN |
44 | #undef printf |
45 | ||
46 | void *fast_memcpy(void *a, const void *b, size_t c){return memcpy(a,b,c);}; | |
47 | ||
9e1586fc | 48 | /* reference fdct/idct */ |
0de74546 DY |
49 | void ff_ref_fdct(DCTELEM *block); |
50 | void ff_ref_idct(DCTELEM *block); | |
51 | void ff_ref_dct_init(void); | |
de6d9b64 | 52 | |
9686df2b DB |
53 | void ff_mmx_idct(DCTELEM *data); |
54 | void ff_mmxext_idct(DCTELEM *data); | |
9e1586fc | 55 | |
9686df2b | 56 | void odivx_idct_c(short *block); |
86748dbc | 57 | |
3ac35bdb | 58 | // BFIN |
9686df2b DB |
59 | void ff_bfin_idct(DCTELEM *block); |
60 | void ff_bfin_fdct(DCTELEM *block); | |
3ac35bdb MH |
61 | |
62 | // ALTIVEC | |
9686df2b DB |
63 | void fdct_altivec(DCTELEM *block); |
64 | //void idct_altivec(DCTELEM *block);?? no routine | |
3ac35bdb | 65 | |
479044ce MR |
66 | // ARM |
67 | void j_rev_dct_ARM(DCTELEM *data); | |
68 | void simple_idct_ARM(DCTELEM *data); | |
69 | void simple_idct_armv5te(DCTELEM *data); | |
70 | void ff_simple_idct_armv6(DCTELEM *data); | |
71 | void ff_simple_idct_neon(DCTELEM *data); | |
3ac35bdb | 72 | |
2a839eeb MR |
73 | void ff_simple_idct_axp(DCTELEM *data); |
74 | ||
3ac35bdb | 75 | struct algo { |
f5b67781 | 76 | const char *name; |
3ac35bdb MH |
77 | enum { FDCT, IDCT } is_idct; |
78 | void (* func) (DCTELEM *block); | |
79 | void (* ref) (DCTELEM *block); | |
875f3125 | 80 | enum formattag { NO_PERM,MMX_PERM, MMX_SIMPLE_PERM, SCALE_PERM, SSE2_PERM, PARTTRANS_PERM } format; |
3794b928 | 81 | int mm_support; |
3ac35bdb MH |
82 | }; |
83 | ||
84 | #ifndef FAAN_POSTSCALE | |
85 | #define FAAN_SCALE SCALE_PERM | |
86 | #else | |
87 | #define FAAN_SCALE NO_PERM | |
88 | #endif | |
89 | ||
aadd27cd MN |
90 | static int cpu_flags; |
91 | ||
3ac35bdb | 92 | struct algo algos[] = { |
0de74546 DY |
93 | {"REF-DBL", 0, ff_ref_fdct, ff_ref_fdct, NO_PERM}, |
94 | {"FAAN", 0, ff_faandct, ff_ref_fdct, FAAN_SCALE}, | |
95 | {"FAANI", 1, ff_faanidct, ff_ref_idct, NO_PERM}, | |
96 | {"IJG-AAN-INT", 0, fdct_ifast, ff_ref_fdct, SCALE_PERM}, | |
97 | {"IJG-LLM-INT", 0, ff_jpeg_fdct_islow, ff_ref_fdct, NO_PERM}, | |
98 | {"REF-DBL", 1, ff_ref_idct, ff_ref_idct, NO_PERM}, | |
99 | {"INT", 1, j_rev_dct, ff_ref_idct, MMX_PERM}, | |
100 | {"SIMPLE-C", 1, ff_simple_idct, ff_ref_idct, NO_PERM}, | |
3ac35bdb | 101 | |
b250f9c6 | 102 | #if HAVE_MMX |
0de74546 | 103 | {"MMX", 0, ff_fdct_mmx, ff_ref_fdct, NO_PERM, FF_MM_MMX}, |
b250f9c6 | 104 | #if HAVE_MMX2 |
0de74546 DY |
105 | {"MMX2", 0, ff_fdct_mmx2, ff_ref_fdct, NO_PERM, FF_MM_MMXEXT}, |
106 | {"SSE2", 0, ff_fdct_sse2, ff_ref_fdct, NO_PERM, FF_MM_SSE2}, | |
94254fc0 | 107 | #endif |
3ac35bdb | 108 | |
b250f9c6 | 109 | #if CONFIG_GPL |
0de74546 DY |
110 | {"LIBMPEG2-MMX", 1, ff_mmx_idct, ff_ref_idct, MMX_PERM, FF_MM_MMX}, |
111 | {"LIBMPEG2-MMXEXT", 1, ff_mmxext_idct, ff_ref_idct, MMX_PERM, FF_MM_MMXEXT}, | |
b9702de5 | 112 | #endif |
0de74546 DY |
113 | {"SIMPLE-MMX", 1, ff_simple_idct_mmx, ff_ref_idct, MMX_SIMPLE_PERM, FF_MM_MMX}, |
114 | {"XVID-MMX", 1, ff_idct_xvid_mmx, ff_ref_idct, NO_PERM, FF_MM_MMX}, | |
115 | {"XVID-MMX2", 1, ff_idct_xvid_mmx2, ff_ref_idct, NO_PERM, FF_MM_MMXEXT}, | |
116 | {"XVID-SSE2", 1, ff_idct_xvid_sse2, ff_ref_idct, SSE2_PERM, FF_MM_SSE2}, | |
3ac35bdb MH |
117 | #endif |
118 | ||
b250f9c6 | 119 | #if HAVE_ALTIVEC |
0de74546 | 120 | {"altivecfdct", 0, fdct_altivec, ff_ref_fdct, NO_PERM, FF_MM_ALTIVEC}, |
3ac35bdb MH |
121 | #endif |
122 | ||
b250f9c6 | 123 | #if ARCH_BFIN |
0de74546 DY |
124 | {"BFINfdct", 0, ff_bfin_fdct, ff_ref_fdct, NO_PERM}, |
125 | {"BFINidct", 1, ff_bfin_idct, ff_ref_idct, NO_PERM}, | |
3ac35bdb MH |
126 | #endif |
127 | ||
b250f9c6 | 128 | #if ARCH_ARM |
0de74546 DY |
129 | {"SIMPLE-ARM", 1, simple_idct_ARM, ff_ref_idct, NO_PERM }, |
130 | {"INT-ARM", 1, j_rev_dct_ARM, ff_ref_idct, MMX_PERM }, | |
b250f9c6 | 131 | #if HAVE_ARMV5TE |
0de74546 | 132 | {"SIMPLE-ARMV5TE", 1, simple_idct_armv5te, ff_ref_idct, NO_PERM }, |
479044ce | 133 | #endif |
b250f9c6 | 134 | #if HAVE_ARMV6 |
0de74546 | 135 | {"SIMPLE-ARMV6", 1, ff_simple_idct_armv6, ff_ref_idct, MMX_PERM }, |
479044ce | 136 | #endif |
b250f9c6 | 137 | #if HAVE_NEON |
0de74546 | 138 | {"SIMPLE-NEON", 1, ff_simple_idct_neon, ff_ref_idct, PARTTRANS_PERM }, |
479044ce | 139 | #endif |
a2fc0f6a | 140 | #endif /* ARCH_ARM */ |
479044ce | 141 | |
2a839eeb | 142 | #if ARCH_ALPHA |
0de74546 | 143 | {"SIMPLE-ALPHA", 1, ff_simple_idct_axp, ff_ref_idct, NO_PERM }, |
2a839eeb MR |
144 | #endif |
145 | ||
3ac35bdb MH |
146 | { 0 } |
147 | }; | |
148 | ||
de6d9b64 | 149 | #define AANSCALE_BITS 12 |
de6d9b64 | 150 | |
486497e0 | 151 | uint8_t cropTbl[256 + 2 * MAX_NEG_CROP]; |
86748dbc | 152 | |
504ffed1 | 153 | static int64_t gettime(void) |
de6d9b64 FB |
154 | { |
155 | struct timeval tv; | |
156 | gettimeofday(&tv,NULL); | |
0c1a9eda | 157 | return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec; |
de6d9b64 FB |
158 | } |
159 | ||
160 | #define NB_ITS 20000 | |
161 | #define NB_ITS_SPEED 50000 | |
162 | ||
9e1586fc FB |
163 | static short idct_mmx_perm[64]; |
164 | ||
86748dbc | 165 | static short idct_simple_mmx_perm[64]={ |
bb270c08 DB |
166 | 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D, |
167 | 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D, | |
168 | 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D, | |
169 | 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F, | |
170 | 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F, | |
171 | 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D, | |
172 | 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F, | |
173 | 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F, | |
86748dbc MN |
174 | }; |
175 | ||
ad246860 AS |
176 | static const uint8_t idct_sse2_row_perm[8] = {0, 4, 1, 5, 2, 6, 3, 7}; |
177 | ||
504ffed1 | 178 | static void idct_mmx_init(void) |
9e1586fc FB |
179 | { |
180 | int i; | |
181 | ||
182 | /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */ | |
183 | for (i = 0; i < 64; i++) { | |
bb270c08 DB |
184 | idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2); |
185 | // idct_simple_mmx_perm[i] = simple_block_permute_op(i); | |
9e1586fc FB |
186 | } |
187 | } | |
188 | ||
55727a83 | 189 | static DCTELEM block[64] __attribute__ ((aligned (16))); |
9e1586fc | 190 | static DCTELEM block1[64] __attribute__ ((aligned (8))); |
86748dbc | 191 | static DCTELEM block_org[64] __attribute__ ((aligned (8))); |
9e1586fc | 192 | |
aadd27cd MN |
193 | static inline void mmx_emms(void) |
194 | { | |
b250f9c6 | 195 | #if HAVE_MMX |
82d1605f | 196 | if (cpu_flags & FF_MM_MMX) |
be449fca | 197 | __asm__ volatile ("emms\n\t"); |
aadd27cd MN |
198 | #endif |
199 | } | |
200 | ||
504ffed1 | 201 | static void dct_error(const char *name, int is_idct, |
9e1586fc | 202 | void (*fdct_func)(DCTELEM *block), |
3ac35bdb | 203 | void (*fdct_ref)(DCTELEM *block), int form, int test) |
de6d9b64 FB |
204 | { |
205 | int it, i, scale; | |
de6d9b64 | 206 | int err_inf, v; |
0c1a9eda ZK |
207 | int64_t err2, ti, ti1, it1; |
208 | int64_t sysErr[64], sysErrMax=0; | |
86748dbc | 209 | int maxout=0; |
86748dbc | 210 | int blockSumErrMax=0, blockSumErr; |
294eaa26 | 211 | AVLFG prn; |
de6d9b64 | 212 | |
294eaa26 | 213 | av_lfg_init(&prn, 1); |
de6d9b64 FB |
214 | |
215 | err_inf = 0; | |
216 | err2 = 0; | |
86748dbc | 217 | for(i=0; i<64; i++) sysErr[i]=0; |
de6d9b64 | 218 | for(it=0;it<NB_ITS;it++) { |
86748dbc MN |
219 | for(i=0;i<64;i++) |
220 | block1[i] = 0; | |
221 | switch(test){ | |
115329f1 | 222 | case 0: |
86748dbc | 223 | for(i=0;i<64;i++) |
294eaa26 | 224 | block1[i] = (av_lfg_get(&prn) % 512) -256; |
ad324c93 | 225 | if (is_idct){ |
0de74546 | 226 | ff_ref_fdct(block1); |
ad324c93 MN |
227 | |
228 | for(i=0;i<64;i++) | |
229 | block1[i]>>=3; | |
230 | } | |
86748dbc MN |
231 | break; |
232 | case 1:{ | |
294eaa26 | 233 | int num = av_lfg_get(&prn) % 10 + 1; |
86748dbc | 234 | for(i=0;i<num;i++) |
294eaa26 | 235 | block1[av_lfg_get(&prn) % 64] = av_lfg_get(&prn) % 512 -256; |
86748dbc MN |
236 | }break; |
237 | case 2: | |
294eaa26 | 238 | block1[0] = av_lfg_get(&prn) % 4096 - 2048; |
86748dbc MN |
239 | block1[63]= (block1[0]&1)^1; |
240 | break; | |
241 | } | |
9e1586fc | 242 | |
86748dbc MN |
243 | #if 0 // simulate mismatch control |
244 | { int sum=0; | |
245 | for(i=0;i<64;i++) | |
246 | sum+=block1[i]; | |
247 | ||
115329f1 | 248 | if((sum&1)==0) block1[63]^=1; |
86748dbc MN |
249 | } |
250 | #endif | |
251 | ||
252 | for(i=0; i<64; i++) | |
253 | block_org[i]= block1[i]; | |
9e1586fc | 254 | |
3ac35bdb | 255 | if (form == MMX_PERM) { |
86748dbc | 256 | for(i=0;i<64;i++) |
9e1586fc | 257 | block[idct_mmx_perm[i]] = block1[i]; |
3ac35bdb | 258 | } else if (form == MMX_SIMPLE_PERM) { |
86748dbc MN |
259 | for(i=0;i<64;i++) |
260 | block[idct_simple_mmx_perm[i]] = block1[i]; | |
261 | ||
ad246860 AS |
262 | } else if (form == SSE2_PERM) { |
263 | for(i=0; i<64; i++) | |
264 | block[(i&0x38) | idct_sse2_row_perm[i&7]] = block1[i]; | |
875f3125 MR |
265 | } else if (form == PARTTRANS_PERM) { |
266 | for(i=0; i<64; i++) | |
267 | block[(i&0x24) | ((i&3)<<3) | ((i>>3)&3)] = block1[i]; | |
bb270c08 | 268 | } else { |
86748dbc MN |
269 | for(i=0; i<64; i++) |
270 | block[i]= block1[i]; | |
9e1586fc | 271 | } |
86748dbc MN |
272 | #if 0 // simulate mismatch control for tested IDCT but not the ref |
273 | { int sum=0; | |
274 | for(i=0;i<64;i++) | |
275 | sum+=block[i]; | |
276 | ||
115329f1 | 277 | if((sum&1)==0) block[63]^=1; |
86748dbc MN |
278 | } |
279 | #endif | |
9e1586fc | 280 | |
de6d9b64 | 281 | fdct_func(block); |
aadd27cd | 282 | mmx_emms(); |
9e1586fc | 283 | |
3ac35bdb | 284 | if (form == SCALE_PERM) { |
de6d9b64 | 285 | for(i=0; i<64; i++) { |
10ac3618 | 286 | scale = 8*(1 << (AANSCALE_BITS + 11)) / ff_aanscales[i]; |
86748dbc MN |
287 | block[i] = (block[i] * scale /*+ (1<<(AANSCALE_BITS-1))*/) >> AANSCALE_BITS; |
288 | } | |
289 | } | |
290 | ||
9e1586fc | 291 | fdct_ref(block1); |
de6d9b64 | 292 | |
86748dbc | 293 | blockSumErr=0; |
de6d9b64 FB |
294 | for(i=0;i<64;i++) { |
295 | v = abs(block[i] - block1[i]); | |
296 | if (v > err_inf) | |
297 | err_inf = v; | |
298 | err2 += v * v; | |
bb270c08 DB |
299 | sysErr[i] += block[i] - block1[i]; |
300 | blockSumErr += v; | |
301 | if( abs(block[i])>maxout) maxout=abs(block[i]); | |
de6d9b64 | 302 | } |
86748dbc MN |
303 | if(blockSumErrMax < blockSumErr) blockSumErrMax= blockSumErr; |
304 | #if 0 // print different matrix pairs | |
305 | if(blockSumErr){ | |
306 | printf("\n"); | |
307 | for(i=0; i<64; i++){ | |
308 | if((i&7)==0) printf("\n"); | |
309 | printf("%4d ", block_org[i]); | |
310 | } | |
311 | for(i=0; i<64; i++){ | |
312 | if((i&7)==0) printf("\n"); | |
313 | printf("%4d ", block[i] - block1[i]); | |
314 | } | |
315 | } | |
316 | #endif | |
317 | } | |
ae32e509 | 318 | for(i=0; i<64; i++) sysErrMax= FFMAX(sysErrMax, FFABS(sysErr[i])); |
115329f1 | 319 | |
86748dbc MN |
320 | #if 1 // dump systematic errors |
321 | for(i=0; i<64; i++){ | |
bb270c08 | 322 | if(i%8==0) printf("\n"); |
2029e934 | 323 | printf("%7d ", (int)sysErr[i]); |
de6d9b64 | 324 | } |
86748dbc MN |
325 | printf("\n"); |
326 | #endif | |
115329f1 | 327 | |
86748dbc | 328 | printf("%s %s: err_inf=%d err2=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n", |
9e1586fc | 329 | is_idct ? "IDCT" : "DCT", |
86748dbc MN |
330 | name, err_inf, (double)err2 / NB_ITS / 64.0, (double)sysErrMax / NB_ITS, maxout, blockSumErrMax); |
331 | #if 1 //Speed test | |
de6d9b64 | 332 | /* speed test */ |
86748dbc MN |
333 | for(i=0;i<64;i++) |
334 | block1[i] = 0; | |
335 | switch(test){ | |
115329f1 | 336 | case 0: |
86748dbc | 337 | for(i=0;i<64;i++) |
294eaa26 | 338 | block1[i] = av_lfg_get(&prn) % 512 -256; |
ad324c93 | 339 | if (is_idct){ |
0de74546 | 340 | ff_ref_fdct(block1); |
ad324c93 MN |
341 | |
342 | for(i=0;i<64;i++) | |
343 | block1[i]>>=3; | |
344 | } | |
86748dbc MN |
345 | break; |
346 | case 1:{ | |
347 | case 2: | |
294eaa26 DB |
348 | block1[0] = av_lfg_get(&prn) % 512 -256; |
349 | block1[1] = av_lfg_get(&prn) % 512 -256; | |
350 | block1[2] = av_lfg_get(&prn) % 512 -256; | |
351 | block1[3] = av_lfg_get(&prn) % 512 -256; | |
86748dbc MN |
352 | }break; |
353 | } | |
de6d9b64 | 354 | |
3ac35bdb | 355 | if (form == MMX_PERM) { |
86748dbc | 356 | for(i=0;i<64;i++) |
9e1586fc | 357 | block[idct_mmx_perm[i]] = block1[i]; |
3ac35bdb | 358 | } else if(form == MMX_SIMPLE_PERM) { |
86748dbc MN |
359 | for(i=0;i<64;i++) |
360 | block[idct_simple_mmx_perm[i]] = block1[i]; | |
361 | } else { | |
362 | for(i=0; i<64; i++) | |
363 | block[i]= block1[i]; | |
9e1586fc FB |
364 | } |
365 | ||
de6d9b64 FB |
366 | ti = gettime(); |
367 | it1 = 0; | |
368 | do { | |
369 | for(it=0;it<NB_ITS_SPEED;it++) { | |
86748dbc MN |
370 | for(i=0; i<64; i++) |
371 | block[i]= block1[i]; | |
372 | // memcpy(block, block1, sizeof(DCTELEM) * 64); | |
755bfeab | 373 | // do not memcpy especially not fastmemcpy because it does movntq !!! |
de6d9b64 FB |
374 | fdct_func(block); |
375 | } | |
376 | it1 += NB_ITS_SPEED; | |
377 | ti1 = gettime() - ti; | |
378 | } while (ti1 < 1000000); | |
aadd27cd | 379 | mmx_emms(); |
de6d9b64 | 380 | |
86748dbc | 381 | printf("%s %s: %0.1f kdct/s\n", |
9e1586fc | 382 | is_idct ? "IDCT" : "DCT", |
de6d9b64 | 383 | name, (double)it1 * 1000.0 / (double)ti1); |
86748dbc | 384 | #endif |
de6d9b64 FB |
385 | } |
386 | ||
0c1a9eda ZK |
387 | static uint8_t img_dest[64] __attribute__ ((aligned (8))); |
388 | static uint8_t img_dest1[64] __attribute__ ((aligned (8))); | |
a46a3ce4 | 389 | |
504ffed1 | 390 | static void idct248_ref(uint8_t *dest, int linesize, int16_t *block) |
a46a3ce4 FB |
391 | { |
392 | static int init; | |
393 | static double c8[8][8]; | |
394 | static double c4[4][4]; | |
395 | double block1[64], block2[64], block3[64]; | |
396 | double s, sum, v; | |
397 | int i, j, k; | |
398 | ||
399 | if (!init) { | |
400 | init = 1; | |
401 | ||
402 | for(i=0;i<8;i++) { | |
403 | sum = 0; | |
404 | for(j=0;j<8;j++) { | |
405 | s = (i==0) ? sqrt(1.0/8.0) : sqrt(1.0/4.0); | |
406 | c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0); | |
407 | sum += c8[i][j] * c8[i][j]; | |
408 | } | |
409 | } | |
115329f1 | 410 | |
a46a3ce4 FB |
411 | for(i=0;i<4;i++) { |
412 | sum = 0; | |
413 | for(j=0;j<4;j++) { | |
414 | s = (i==0) ? sqrt(1.0/4.0) : sqrt(1.0/2.0); | |
415 | c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0); | |
416 | sum += c4[i][j] * c4[i][j]; | |
417 | } | |
418 | } | |
419 | } | |
420 | ||
421 | /* butterfly */ | |
652f0197 | 422 | s = 0.5 * sqrt(2.0); |
a46a3ce4 FB |
423 | for(i=0;i<4;i++) { |
424 | for(j=0;j<8;j++) { | |
652f0197 FB |
425 | block1[8*(2*i)+j] = (block[8*(2*i)+j] + block[8*(2*i+1)+j]) * s; |
426 | block1[8*(2*i+1)+j] = (block[8*(2*i)+j] - block[8*(2*i+1)+j]) * s; | |
a46a3ce4 FB |
427 | } |
428 | } | |
429 | ||
430 | /* idct8 on lines */ | |
431 | for(i=0;i<8;i++) { | |
432 | for(j=0;j<8;j++) { | |
433 | sum = 0; | |
434 | for(k=0;k<8;k++) | |
435 | sum += c8[k][j] * block1[8*i+k]; | |
436 | block2[8*i+j] = sum; | |
437 | } | |
438 | } | |
439 | ||
440 | /* idct4 */ | |
441 | for(i=0;i<8;i++) { | |
442 | for(j=0;j<4;j++) { | |
443 | /* top */ | |
444 | sum = 0; | |
445 | for(k=0;k<4;k++) | |
446 | sum += c4[k][j] * block2[8*(2*k)+i]; | |
447 | block3[8*(2*j)+i] = sum; | |
448 | ||
449 | /* bottom */ | |
450 | sum = 0; | |
451 | for(k=0;k<4;k++) | |
452 | sum += c4[k][j] * block2[8*(2*k+1)+i]; | |
453 | block3[8*(2*j+1)+i] = sum; | |
454 | } | |
455 | } | |
456 | ||
457 | /* clamp and store the result */ | |
458 | for(i=0;i<8;i++) { | |
459 | for(j=0;j<8;j++) { | |
652f0197 | 460 | v = block3[8*i+j]; |
a46a3ce4 FB |
461 | if (v < 0) |
462 | v = 0; | |
463 | else if (v > 255) | |
464 | v = 255; | |
465 | dest[i * linesize + j] = (int)rint(v); | |
466 | } | |
467 | } | |
468 | } | |
469 | ||
504ffed1 | 470 | static void idct248_error(const char *name, |
0c1a9eda | 471 | void (*idct248_put)(uint8_t *dest, int line_size, int16_t *block)) |
a46a3ce4 FB |
472 | { |
473 | int it, i, it1, ti, ti1, err_max, v; | |
474 | ||
294eaa26 DB |
475 | AVLFG prn; |
476 | ||
477 | av_lfg_init(&prn, 1); | |
115329f1 | 478 | |
a46a3ce4 FB |
479 | /* just one test to see if code is correct (precision is less |
480 | important here) */ | |
481 | err_max = 0; | |
482 | for(it=0;it<NB_ITS;it++) { | |
115329f1 | 483 | |
652f0197 FB |
484 | /* XXX: use forward transform to generate values */ |
485 | for(i=0;i<64;i++) | |
294eaa26 | 486 | block1[i] = av_lfg_get(&prn) % 256 - 128; |
652f0197 FB |
487 | block1[0] += 1024; |
488 | ||
a46a3ce4 FB |
489 | for(i=0; i<64; i++) |
490 | block[i]= block1[i]; | |
491 | idct248_ref(img_dest1, 8, block); | |
115329f1 | 492 | |
652f0197 FB |
493 | for(i=0; i<64; i++) |
494 | block[i]= block1[i]; | |
495 | idct248_put(img_dest, 8, block); | |
115329f1 | 496 | |
652f0197 FB |
497 | for(i=0;i<64;i++) { |
498 | v = abs((int)img_dest[i] - (int)img_dest1[i]); | |
499 | if (v == 255) | |
500 | printf("%d %d\n", img_dest[i], img_dest1[i]); | |
501 | if (v > err_max) | |
502 | err_max = v; | |
503 | } | |
a46a3ce4 FB |
504 | #if 0 |
505 | printf("ref=\n"); | |
506 | for(i=0;i<8;i++) { | |
507 | int j; | |
508 | for(j=0;j<8;j++) { | |
509 | printf(" %3d", img_dest1[i*8+j]); | |
510 | } | |
511 | printf("\n"); | |
512 | } | |
115329f1 | 513 | |
a46a3ce4 FB |
514 | printf("out=\n"); |
515 | for(i=0;i<8;i++) { | |
516 | int j; | |
517 | for(j=0;j<8;j++) { | |
518 | printf(" %3d", img_dest[i*8+j]); | |
519 | } | |
520 | printf("\n"); | |
521 | } | |
522 | #endif | |
a46a3ce4 FB |
523 | } |
524 | printf("%s %s: err_inf=%d\n", | |
525 | 1 ? "IDCT248" : "DCT248", | |
526 | name, err_max); | |
527 | ||
528 | ti = gettime(); | |
529 | it1 = 0; | |
530 | do { | |
531 | for(it=0;it<NB_ITS_SPEED;it++) { | |
532 | for(i=0; i<64; i++) | |
533 | block[i]= block1[i]; | |
534 | // memcpy(block, block1, sizeof(DCTELEM) * 64); | |
755bfeab | 535 | // do not memcpy especially not fastmemcpy because it does movntq !!! |
a46a3ce4 FB |
536 | idct248_put(img_dest, 8, block); |
537 | } | |
538 | it1 += NB_ITS_SPEED; | |
539 | ti1 = gettime() - ti; | |
540 | } while (ti1 < 1000000); | |
aadd27cd | 541 | mmx_emms(); |
a46a3ce4 FB |
542 | |
543 | printf("%s %s: %0.1f kdct/s\n", | |
544 | 1 ? "IDCT248" : "DCT248", | |
545 | name, (double)it1 * 1000.0 / (double)ti1); | |
546 | } | |
547 | ||
504ffed1 | 548 | static void help(void) |
9e1586fc | 549 | { |
86748dbc MN |
550 | printf("dct-test [-i] [<test-number>]\n" |
551 | "test-number 0 -> test with random matrixes\n" | |
552 | " 1 -> test with random sparse matrixes\n" | |
553 | " 2 -> do 3. test from mpeg4 std\n" | |
a46a3ce4 FB |
554 | "-i test IDCT implementations\n" |
555 | "-4 test IDCT248 implementations\n"); | |
9e1586fc FB |
556 | } |
557 | ||
de6d9b64 FB |
558 | int main(int argc, char **argv) |
559 | { | |
a46a3ce4 | 560 | int test_idct = 0, test_248_dct = 0; |
86748dbc MN |
561 | int c,i; |
562 | int test=1; | |
aadd27cd | 563 | cpu_flags = mm_support(); |
9e1586fc | 564 | |
0de74546 | 565 | ff_ref_dct_init(); |
9e1586fc | 566 | idct_mmx_init(); |
f67a10cd | 567 | |
486497e0 | 568 | for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i; |
86748dbc | 569 | for(i=0;i<MAX_NEG_CROP;i++) { |
486497e0 MR |
570 | cropTbl[i] = 0; |
571 | cropTbl[i + MAX_NEG_CROP + 256] = 255; | |
86748dbc | 572 | } |
115329f1 | 573 | |
9e1586fc | 574 | for(;;) { |
a46a3ce4 | 575 | c = getopt(argc, argv, "ih4"); |
9e1586fc FB |
576 | if (c == -1) |
577 | break; | |
578 | switch(c) { | |
579 | case 'i': | |
580 | test_idct = 1; | |
581 | break; | |
a46a3ce4 FB |
582 | case '4': |
583 | test_248_dct = 1; | |
584 | break; | |
86748dbc | 585 | default : |
9e1586fc FB |
586 | case 'h': |
587 | help(); | |
c6bdc908 | 588 | return 0; |
9e1586fc FB |
589 | } |
590 | } | |
115329f1 | 591 | |
86748dbc | 592 | if(optind <argc) test= atoi(argv[optind]); |
115329f1 | 593 | |
9e1586fc FB |
594 | printf("ffmpeg DCT/IDCT test\n"); |
595 | ||
a46a3ce4 | 596 | if (test_248_dct) { |
59e6f60a | 597 | idct248_error("SIMPLE-C", ff_simple_idct248_put); |
9e1586fc | 598 | } else { |
3ac35bdb | 599 | for (i=0;algos[i].name;i++) |
dafe8824 | 600 | if (algos[i].is_idct == test_idct && !(~cpu_flags & algos[i].mm_support)) { |
3ac35bdb | 601 | dct_error (algos[i].name, algos[i].is_idct, algos[i].func, algos[i].ref, algos[i].format, test); |
a46a3ce4 | 602 | } |
9e1586fc | 603 | } |
de6d9b64 FB |
604 | return 0; |
605 | } |