Mercurial > libavcodec.hg
annotate dct-test.c @ 11408:4b6b3ffbaee3 libavcodec
Move dctref prototypes to a header file
author | mru |
---|---|
date | Mon, 08 Mar 2010 18:43:33 +0000 |
parents | 34a65026fa06 |
children | 9b1db41afb30 |
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 /** |
8718
e9d9d946f213
Use full internal pathname in doxygen @file directives.
diego
parents:
8693
diff
changeset
|
23 * @file libavcodec/dct-test.c |
8724
2c5662c41129
cosmetics: Remove period after copyright statement non-sentence.
diego
parents:
8718
diff
changeset
|
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> | |
5118
3b190bc34546
Add some #includes to allow compilation without HAVE_AV_CONFIG_H.
diego
parents:
5110
diff
changeset
|
33 #include <math.h> |
0 | 34 |
7130
601509a430f7
Replace redundant MAX macro declaration by proper use of FFMAX.
diego
parents:
7125
diff
changeset
|
35 #include "libavutil/common.h" |
9199
ea0e5e9a520f
Replace random() usage in test programs by av_lfg_*().
diego
parents:
9189
diff
changeset
|
36 #include "libavutil/lfg.h" |
0 | 37 |
633 | 38 #include "simple_idct.h" |
8223 | 39 #include "aandcttab.h" |
1557 | 40 #include "faandct.h" |
6407 | 41 #include "faanidct.h" |
8430 | 42 #include "x86/idct_xvid.h" |
11408 | 43 #include "dctref.h" |
33 | 44 |
2872 | 45 #undef printf |
46 | |
47 void *fast_memcpy(void *a, const void *b, size_t c){return memcpy(a,b,c);}; | |
48 | |
8250 | 49 void ff_mmx_idct(DCTELEM *data); |
50 void ff_mmxext_idct(DCTELEM *data); | |
33 | 51 |
8250 | 52 void odivx_idct_c(short *block); |
633 | 53 |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
54 // BFIN |
8250 | 55 void ff_bfin_idct(DCTELEM *block); |
56 void ff_bfin_fdct(DCTELEM *block); | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
57 |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
58 // ALTIVEC |
8250 | 59 void fdct_altivec(DCTELEM *block); |
60 //void idct_altivec(DCTELEM *block);?? no routine | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
61 |
8351 | 62 // ARM |
63 void j_rev_dct_ARM(DCTELEM *data); | |
64 void simple_idct_ARM(DCTELEM *data); | |
65 void simple_idct_armv5te(DCTELEM *data); | |
66 void ff_simple_idct_armv6(DCTELEM *data); | |
67 void ff_simple_idct_neon(DCTELEM *data); | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
68 |
8622 | 69 void ff_simple_idct_axp(DCTELEM *data); |
70 | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
71 struct algo { |
7125
65e7e714c770
Mark constant structure member as const to avoid some warnings.
diego
parents:
6602
diff
changeset
|
72 const char *name; |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
73 enum { FDCT, IDCT } is_idct; |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
74 void (* func) (DCTELEM *block); |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
75 void (* ref) (DCTELEM *block); |
8350 | 76 enum formattag { NO_PERM,MMX_PERM, MMX_SIMPLE_PERM, SCALE_PERM, SSE2_PERM, PARTTRANS_PERM } format; |
6543
948d9453432b
check mm_flags for each DCT and skips the ones that can't be run
gpoirier
parents:
6542
diff
changeset
|
77 int mm_support; |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
78 }; |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
79 |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
80 #ifndef FAAN_POSTSCALE |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
81 #define FAAN_SCALE SCALE_PERM |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
82 #else |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
83 #define FAAN_SCALE NO_PERM |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
84 #endif |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
85 |
7155 | 86 static int cpu_flags; |
87 | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
88 struct algo algos[] = { |
9293 | 89 {"REF-DBL", 0, ff_ref_fdct, ff_ref_fdct, NO_PERM}, |
90 {"FAAN", 0, ff_faandct, ff_ref_fdct, FAAN_SCALE}, | |
91 {"FAANI", 1, ff_faanidct, ff_ref_idct, NO_PERM}, | |
92 {"IJG-AAN-INT", 0, fdct_ifast, ff_ref_fdct, SCALE_PERM}, | |
93 {"IJG-LLM-INT", 0, ff_jpeg_fdct_islow, ff_ref_fdct, NO_PERM}, | |
94 {"REF-DBL", 1, ff_ref_idct, ff_ref_idct, NO_PERM}, | |
95 {"INT", 1, j_rev_dct, ff_ref_idct, MMX_PERM}, | |
96 {"SIMPLE-C", 1, ff_simple_idct, ff_ref_idct, NO_PERM}, | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
97 |
8590 | 98 #if HAVE_MMX |
9293 | 99 {"MMX", 0, ff_fdct_mmx, ff_ref_fdct, NO_PERM, FF_MM_MMX}, |
8590 | 100 #if HAVE_MMX2 |
9406
8d51b340393e
Replace deprecated FF_MM_MMXEXT flag with FF_MM_MMX2.
stefano
parents:
9388
diff
changeset
|
101 {"MMX2", 0, ff_fdct_mmx2, ff_ref_fdct, NO_PERM, FF_MM_MMX2}, |
9293 | 102 {"SSE2", 0, ff_fdct_sse2, ff_ref_fdct, NO_PERM, FF_MM_SSE2}, |
5110 | 103 #endif |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
104 |
8590 | 105 #if CONFIG_GPL |
9293 | 106 {"LIBMPEG2-MMX", 1, ff_mmx_idct, ff_ref_idct, MMX_PERM, FF_MM_MMX}, |
9406
8d51b340393e
Replace deprecated FF_MM_MMXEXT flag with FF_MM_MMX2.
stefano
parents:
9388
diff
changeset
|
107 {"LIBMPEG2-MMX2", 1, ff_mmxext_idct, ff_ref_idct, MMX_PERM, FF_MM_MMX2}, |
5109 | 108 #endif |
9293 | 109 {"SIMPLE-MMX", 1, ff_simple_idct_mmx, ff_ref_idct, MMX_SIMPLE_PERM, FF_MM_MMX}, |
110 {"XVID-MMX", 1, ff_idct_xvid_mmx, ff_ref_idct, NO_PERM, FF_MM_MMX}, | |
9406
8d51b340393e
Replace deprecated FF_MM_MMXEXT flag with FF_MM_MMX2.
stefano
parents:
9388
diff
changeset
|
111 {"XVID-MMX2", 1, ff_idct_xvid_mmx2, ff_ref_idct, NO_PERM, FF_MM_MMX2}, |
9293 | 112 {"XVID-SSE2", 1, ff_idct_xvid_sse2, ff_ref_idct, SSE2_PERM, FF_MM_SSE2}, |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
113 #endif |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
114 |
8590 | 115 #if HAVE_ALTIVEC |
9293 | 116 {"altivecfdct", 0, fdct_altivec, ff_ref_fdct, NO_PERM, FF_MM_ALTIVEC}, |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
117 #endif |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
118 |
8590 | 119 #if ARCH_BFIN |
9293 | 120 {"BFINfdct", 0, ff_bfin_fdct, ff_ref_fdct, NO_PERM}, |
121 {"BFINidct", 1, ff_bfin_idct, ff_ref_idct, NO_PERM}, | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
122 #endif |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
123 |
8590 | 124 #if ARCH_ARM |
9293 | 125 {"SIMPLE-ARM", 1, simple_idct_ARM, ff_ref_idct, NO_PERM }, |
126 {"INT-ARM", 1, j_rev_dct_ARM, ff_ref_idct, MMX_PERM }, | |
8590 | 127 #if HAVE_ARMV5TE |
9293 | 128 {"SIMPLE-ARMV5TE", 1, simple_idct_armv5te, ff_ref_idct, NO_PERM }, |
8351 | 129 #endif |
8590 | 130 #if HAVE_ARMV6 |
9293 | 131 {"SIMPLE-ARMV6", 1, ff_simple_idct_armv6, ff_ref_idct, MMX_PERM }, |
8351 | 132 #endif |
8590 | 133 #if HAVE_NEON |
9293 | 134 {"SIMPLE-NEON", 1, ff_simple_idct_neon, ff_ref_idct, PARTTRANS_PERM }, |
8351 | 135 #endif |
8359 | 136 #endif /* ARCH_ARM */ |
8351 | 137 |
8622 | 138 #if ARCH_ALPHA |
9293 | 139 {"SIMPLE-ALPHA", 1, ff_simple_idct_axp, ff_ref_idct, NO_PERM }, |
8622 | 140 #endif |
141 | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
142 { 0 } |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
143 }; |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
144 |
0 | 145 #define AANSCALE_BITS 12 |
146 | |
4197 | 147 uint8_t cropTbl[256 + 2 * MAX_NEG_CROP]; |
633 | 148 |
9295
b225f51903af
Mark non-exported functions in test and example programs as static.
diego
parents:
9293
diff
changeset
|
149 static int64_t gettime(void) |
0 | 150 { |
151 struct timeval tv; | |
152 gettimeofday(&tv,NULL); | |
1064 | 153 return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec; |
0 | 154 } |
155 | |
156 #define NB_ITS 20000 | |
157 #define NB_ITS_SPEED 50000 | |
158 | |
33 | 159 static short idct_mmx_perm[64]; |
160 | |
633 | 161 static short idct_simple_mmx_perm[64]={ |
2979 | 162 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D, |
163 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D, | |
164 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D, | |
165 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F, | |
166 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F, | |
167 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D, | |
168 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F, | |
169 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F, | |
633 | 170 }; |
171 | |
6602 | 172 static const uint8_t idct_sse2_row_perm[8] = {0, 4, 1, 5, 2, 6, 3, 7}; |
173 | |
9295
b225f51903af
Mark non-exported functions in test and example programs as static.
diego
parents:
9293
diff
changeset
|
174 static void idct_mmx_init(void) |
33 | 175 { |
176 int i; | |
177 | |
178 /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */ | |
179 for (i = 0; i < 64; i++) { | |
2979 | 180 idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2); |
181 // idct_simple_mmx_perm[i] = simple_block_permute_op(i); | |
33 | 182 } |
183 } | |
184 | |
10961
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
9793
diff
changeset
|
185 DECLARE_ALIGNED(16, static DCTELEM, block)[64]; |
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
9793
diff
changeset
|
186 DECLARE_ALIGNED(8, static DCTELEM, block1)[64]; |
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
9793
diff
changeset
|
187 DECLARE_ALIGNED(8, static DCTELEM, block_org)[64]; |
33 | 188 |
7155 | 189 static inline void mmx_emms(void) |
190 { | |
8590 | 191 #if HAVE_MMX |
8104
0d108ec85620
Remove duplicated MM_* macros for CPU capabilities from dsputil.h.
rathann
parents:
8031
diff
changeset
|
192 if (cpu_flags & FF_MM_MMX) |
8031 | 193 __asm__ volatile ("emms\n\t"); |
7155 | 194 #endif |
195 } | |
196 | |
9295
b225f51903af
Mark non-exported functions in test and example programs as static.
diego
parents:
9293
diff
changeset
|
197 static void dct_error(const char *name, int is_idct, |
33 | 198 void (*fdct_func)(DCTELEM *block), |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
199 void (*fdct_ref)(DCTELEM *block), int form, int test) |
0 | 200 { |
201 int it, i, scale; | |
202 int err_inf, v; | |
1064 | 203 int64_t err2, ti, ti1, it1; |
204 int64_t sysErr[64], sysErrMax=0; | |
633 | 205 int maxout=0; |
206 int blockSumErrMax=0, blockSumErr; | |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
207 AVLFG prng; |
0 | 208 |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
209 av_lfg_init(&prng, 1); |
0 | 210 |
211 err_inf = 0; | |
212 err2 = 0; | |
633 | 213 for(i=0; i<64; i++) sysErr[i]=0; |
0 | 214 for(it=0;it<NB_ITS;it++) { |
633 | 215 for(i=0;i<64;i++) |
216 block1[i] = 0; | |
217 switch(test){ | |
2967 | 218 case 0: |
633 | 219 for(i=0;i<64;i++) |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
220 block1[i] = (av_lfg_get(&prng) % 512) -256; |
635 | 221 if (is_idct){ |
9293 | 222 ff_ref_fdct(block1); |
635 | 223 |
224 for(i=0;i<64;i++) | |
225 block1[i]>>=3; | |
226 } | |
633 | 227 break; |
228 case 1:{ | |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
229 int num = av_lfg_get(&prng) % 10 + 1; |
633 | 230 for(i=0;i<num;i++) |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
231 block1[av_lfg_get(&prng) % 64] = av_lfg_get(&prng) % 512 -256; |
633 | 232 }break; |
233 case 2: | |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
234 block1[0] = av_lfg_get(&prng) % 4096 - 2048; |
633 | 235 block1[63]= (block1[0]&1)^1; |
236 break; | |
237 } | |
33 | 238 |
633 | 239 #if 0 // simulate mismatch control |
240 { int sum=0; | |
241 for(i=0;i<64;i++) | |
242 sum+=block1[i]; | |
243 | |
2967 | 244 if((sum&1)==0) block1[63]^=1; |
633 | 245 } |
246 #endif | |
247 | |
248 for(i=0; i<64; i++) | |
249 block_org[i]= block1[i]; | |
33 | 250 |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
251 if (form == MMX_PERM) { |
633 | 252 for(i=0;i<64;i++) |
33 | 253 block[idct_mmx_perm[i]] = block1[i]; |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
254 } else if (form == MMX_SIMPLE_PERM) { |
633 | 255 for(i=0;i<64;i++) |
256 block[idct_simple_mmx_perm[i]] = block1[i]; | |
257 | |
6602 | 258 } else if (form == SSE2_PERM) { |
259 for(i=0; i<64; i++) | |
260 block[(i&0x38) | idct_sse2_row_perm[i&7]] = block1[i]; | |
8350 | 261 } else if (form == PARTTRANS_PERM) { |
262 for(i=0; i<64; i++) | |
263 block[(i&0x24) | ((i&3)<<3) | ((i>>3)&3)] = block1[i]; | |
2979 | 264 } else { |
633 | 265 for(i=0; i<64; i++) |
266 block[i]= block1[i]; | |
33 | 267 } |
633 | 268 #if 0 // simulate mismatch control for tested IDCT but not the ref |
269 { int sum=0; | |
270 for(i=0;i<64;i++) | |
271 sum+=block[i]; | |
272 | |
2967 | 273 if((sum&1)==0) block[63]^=1; |
633 | 274 } |
275 #endif | |
33 | 276 |
0 | 277 fdct_func(block); |
7155 | 278 mmx_emms(); |
33 | 279 |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
280 if (form == SCALE_PERM) { |
0 | 281 for(i=0; i<64; i++) { |
8223 | 282 scale = 8*(1 << (AANSCALE_BITS + 11)) / ff_aanscales[i]; |
633 | 283 block[i] = (block[i] * scale /*+ (1<<(AANSCALE_BITS-1))*/) >> AANSCALE_BITS; |
284 } | |
285 } | |
286 | |
33 | 287 fdct_ref(block1); |
0 | 288 |
633 | 289 blockSumErr=0; |
0 | 290 for(i=0;i<64;i++) { |
291 v = abs(block[i] - block1[i]); | |
292 if (v > err_inf) | |
293 err_inf = v; | |
294 err2 += v * v; | |
2979 | 295 sysErr[i] += block[i] - block1[i]; |
296 blockSumErr += v; | |
297 if( abs(block[i])>maxout) maxout=abs(block[i]); | |
0 | 298 } |
633 | 299 if(blockSumErrMax < blockSumErr) blockSumErrMax= blockSumErr; |
300 #if 0 // print different matrix pairs | |
301 if(blockSumErr){ | |
302 printf("\n"); | |
303 for(i=0; i<64; i++){ | |
304 if((i&7)==0) printf("\n"); | |
305 printf("%4d ", block_org[i]); | |
306 } | |
307 for(i=0; i<64; i++){ | |
308 if((i&7)==0) printf("\n"); | |
309 printf("%4d ", block[i] - block1[i]); | |
310 } | |
311 } | |
312 #endif | |
0 | 313 } |
7130
601509a430f7
Replace redundant MAX macro declaration by proper use of FFMAX.
diego
parents:
7125
diff
changeset
|
314 for(i=0; i<64; i++) sysErrMax= FFMAX(sysErrMax, FFABS(sysErr[i])); |
2967 | 315 |
633 | 316 #if 1 // dump systematic errors |
317 for(i=0; i<64; i++){ | |
2979 | 318 if(i%8==0) printf("\n"); |
9189 | 319 printf("%7d ", (int)sysErr[i]); |
633 | 320 } |
321 printf("\n"); | |
322 #endif | |
2967 | 323 |
633 | 324 printf("%s %s: err_inf=%d err2=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n", |
33 | 325 is_idct ? "IDCT" : "DCT", |
633 | 326 name, err_inf, (double)err2 / NB_ITS / 64.0, (double)sysErrMax / NB_ITS, maxout, blockSumErrMax); |
327 #if 1 //Speed test | |
0 | 328 /* speed test */ |
633 | 329 for(i=0;i<64;i++) |
330 block1[i] = 0; | |
331 switch(test){ | |
2967 | 332 case 0: |
633 | 333 for(i=0;i<64;i++) |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
334 block1[i] = av_lfg_get(&prng) % 512 -256; |
635 | 335 if (is_idct){ |
9293 | 336 ff_ref_fdct(block1); |
635 | 337 |
338 for(i=0;i<64;i++) | |
339 block1[i]>>=3; | |
340 } | |
633 | 341 break; |
342 case 1:{ | |
343 case 2: | |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
344 block1[0] = av_lfg_get(&prng) % 512 -256; |
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
345 block1[1] = av_lfg_get(&prng) % 512 -256; |
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
346 block1[2] = av_lfg_get(&prng) % 512 -256; |
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
347 block1[3] = av_lfg_get(&prng) % 512 -256; |
633 | 348 }break; |
349 } | |
0 | 350 |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
351 if (form == MMX_PERM) { |
633 | 352 for(i=0;i<64;i++) |
33 | 353 block[idct_mmx_perm[i]] = block1[i]; |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
354 } else if(form == MMX_SIMPLE_PERM) { |
633 | 355 for(i=0;i<64;i++) |
356 block[idct_simple_mmx_perm[i]] = block1[i]; | |
357 } else { | |
358 for(i=0; i<64; i++) | |
359 block[i]= block1[i]; | |
33 | 360 } |
361 | |
0 | 362 ti = gettime(); |
363 it1 = 0; | |
364 do { | |
365 for(it=0;it<NB_ITS_SPEED;it++) { | |
633 | 366 for(i=0; i<64; i++) |
367 block[i]= block1[i]; | |
368 // memcpy(block, block1, sizeof(DCTELEM) * 64); | |
5127 | 369 // do not memcpy especially not fastmemcpy because it does movntq !!! |
0 | 370 fdct_func(block); |
371 } | |
372 it1 += NB_ITS_SPEED; | |
373 ti1 = gettime() - ti; | |
374 } while (ti1 < 1000000); | |
7155 | 375 mmx_emms(); |
0 | 376 |
633 | 377 printf("%s %s: %0.1f kdct/s\n", |
33 | 378 is_idct ? "IDCT" : "DCT", |
0 | 379 name, (double)it1 * 1000.0 / (double)ti1); |
633 | 380 #endif |
0 | 381 } |
382 | |
10961
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
9793
diff
changeset
|
383 DECLARE_ALIGNED(8, static uint8_t, img_dest)[64]; |
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
9793
diff
changeset
|
384 DECLARE_ALIGNED(8, static uint8_t, img_dest1)[64]; |
720 | 385 |
9295
b225f51903af
Mark non-exported functions in test and example programs as static.
diego
parents:
9293
diff
changeset
|
386 static void idct248_ref(uint8_t *dest, int linesize, int16_t *block) |
720 | 387 { |
388 static int init; | |
389 static double c8[8][8]; | |
390 static double c4[4][4]; | |
391 double block1[64], block2[64], block3[64]; | |
392 double s, sum, v; | |
393 int i, j, k; | |
394 | |
395 if (!init) { | |
396 init = 1; | |
397 | |
398 for(i=0;i<8;i++) { | |
399 sum = 0; | |
400 for(j=0;j<8;j++) { | |
401 s = (i==0) ? sqrt(1.0/8.0) : sqrt(1.0/4.0); | |
402 c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0); | |
403 sum += c8[i][j] * c8[i][j]; | |
404 } | |
405 } | |
2967 | 406 |
720 | 407 for(i=0;i<4;i++) { |
408 sum = 0; | |
409 for(j=0;j<4;j++) { | |
410 s = (i==0) ? sqrt(1.0/4.0) : sqrt(1.0/2.0); | |
411 c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0); | |
412 sum += c4[i][j] * c4[i][j]; | |
413 } | |
414 } | |
415 } | |
416 | |
417 /* butterfly */ | |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
418 s = 0.5 * sqrt(2.0); |
720 | 419 for(i=0;i<4;i++) { |
420 for(j=0;j<8;j++) { | |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
421 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
|
422 block1[8*(2*i+1)+j] = (block[8*(2*i)+j] - block[8*(2*i+1)+j]) * s; |
720 | 423 } |
424 } | |
425 | |
426 /* idct8 on lines */ | |
427 for(i=0;i<8;i++) { | |
428 for(j=0;j<8;j++) { | |
429 sum = 0; | |
430 for(k=0;k<8;k++) | |
431 sum += c8[k][j] * block1[8*i+k]; | |
432 block2[8*i+j] = sum; | |
433 } | |
434 } | |
435 | |
436 /* idct4 */ | |
437 for(i=0;i<8;i++) { | |
438 for(j=0;j<4;j++) { | |
439 /* top */ | |
440 sum = 0; | |
441 for(k=0;k<4;k++) | |
442 sum += c4[k][j] * block2[8*(2*k)+i]; | |
443 block3[8*(2*j)+i] = sum; | |
444 | |
445 /* bottom */ | |
446 sum = 0; | |
447 for(k=0;k<4;k++) | |
448 sum += c4[k][j] * block2[8*(2*k+1)+i]; | |
449 block3[8*(2*j+1)+i] = sum; | |
450 } | |
451 } | |
452 | |
453 /* clamp and store the result */ | |
454 for(i=0;i<8;i++) { | |
455 for(j=0;j<8;j++) { | |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
456 v = block3[8*i+j]; |
720 | 457 if (v < 0) |
458 v = 0; | |
459 else if (v > 255) | |
460 v = 255; | |
461 dest[i * linesize + j] = (int)rint(v); | |
462 } | |
463 } | |
464 } | |
465 | |
9295
b225f51903af
Mark non-exported functions in test and example programs as static.
diego
parents:
9293
diff
changeset
|
466 static void idct248_error(const char *name, |
1064 | 467 void (*idct248_put)(uint8_t *dest, int line_size, int16_t *block)) |
720 | 468 { |
469 int it, i, it1, ti, ti1, err_max, v; | |
470 | |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
471 AVLFG prng; |
9199
ea0e5e9a520f
Replace random() usage in test programs by av_lfg_*().
diego
parents:
9189
diff
changeset
|
472 |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
473 av_lfg_init(&prng, 1); |
2967 | 474 |
720 | 475 /* just one test to see if code is correct (precision is less |
476 important here) */ | |
477 err_max = 0; | |
478 for(it=0;it<NB_ITS;it++) { | |
2967 | 479 |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
480 /* XXX: use forward transform to generate values */ |
720 | 481 for(i=0;i<64;i++) |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
482 block1[i] = av_lfg_get(&prng) % 256 - 128; |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
483 block1[0] += 1024; |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
484 |
720 | 485 for(i=0; i<64; i++) |
486 block[i]= block1[i]; | |
487 idct248_ref(img_dest1, 8, block); | |
2967 | 488 |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
489 for(i=0; i<64; i++) |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
490 block[i]= block1[i]; |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
491 idct248_put(img_dest, 8, block); |
2967 | 492 |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
493 for(i=0;i<64;i++) { |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
494 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
|
495 if (v == 255) |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
496 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
|
497 if (v > err_max) |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
498 err_max = v; |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
499 } |
720 | 500 #if 0 |
501 printf("ref=\n"); | |
502 for(i=0;i<8;i++) { | |
503 int j; | |
504 for(j=0;j<8;j++) { | |
505 printf(" %3d", img_dest1[i*8+j]); | |
506 } | |
507 printf("\n"); | |
508 } | |
2967 | 509 |
720 | 510 printf("out=\n"); |
511 for(i=0;i<8;i++) { | |
512 int j; | |
513 for(j=0;j<8;j++) { | |
514 printf(" %3d", img_dest[i*8+j]); | |
515 } | |
516 printf("\n"); | |
517 } | |
518 #endif | |
519 } | |
520 printf("%s %s: err_inf=%d\n", | |
521 1 ? "IDCT248" : "DCT248", | |
522 name, err_max); | |
523 | |
524 ti = gettime(); | |
525 it1 = 0; | |
526 do { | |
527 for(it=0;it<NB_ITS_SPEED;it++) { | |
528 for(i=0; i<64; i++) | |
529 block[i]= block1[i]; | |
530 // memcpy(block, block1, sizeof(DCTELEM) * 64); | |
5127 | 531 // do not memcpy especially not fastmemcpy because it does movntq !!! |
720 | 532 idct248_put(img_dest, 8, block); |
533 } | |
534 it1 += NB_ITS_SPEED; | |
535 ti1 = gettime() - ti; | |
536 } while (ti1 < 1000000); | |
7155 | 537 mmx_emms(); |
720 | 538 |
539 printf("%s %s: %0.1f kdct/s\n", | |
540 1 ? "IDCT248" : "DCT248", | |
541 name, (double)it1 * 1000.0 / (double)ti1); | |
542 } | |
543 | |
9295
b225f51903af
Mark non-exported functions in test and example programs as static.
diego
parents:
9293
diff
changeset
|
544 static void help(void) |
33 | 545 { |
633 | 546 printf("dct-test [-i] [<test-number>]\n" |
547 "test-number 0 -> test with random matrixes\n" | |
548 " 1 -> test with random sparse matrixes\n" | |
549 " 2 -> do 3. test from mpeg4 std\n" | |
720 | 550 "-i test IDCT implementations\n" |
551 "-4 test IDCT248 implementations\n"); | |
33 | 552 } |
553 | |
0 | 554 int main(int argc, char **argv) |
555 { | |
720 | 556 int test_idct = 0, test_248_dct = 0; |
633 | 557 int c,i; |
558 int test=1; | |
7155 | 559 cpu_flags = mm_support(); |
33 | 560 |
9293 | 561 ff_ref_dct_init(); |
33 | 562 idct_mmx_init(); |
7135
6bd6a2da306e
Define mm_flags/support to be 0 on architectures where they don't exist.
astrange
parents:
7130
diff
changeset
|
563 |
4197 | 564 for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i; |
633 | 565 for(i=0;i<MAX_NEG_CROP;i++) { |
4197 | 566 cropTbl[i] = 0; |
567 cropTbl[i + MAX_NEG_CROP + 256] = 255; | |
633 | 568 } |
2967 | 569 |
33 | 570 for(;;) { |
720 | 571 c = getopt(argc, argv, "ih4"); |
33 | 572 if (c == -1) |
573 break; | |
574 switch(c) { | |
575 case 'i': | |
576 test_idct = 1; | |
577 break; | |
720 | 578 case '4': |
579 test_248_dct = 1; | |
580 break; | |
633 | 581 default : |
33 | 582 case 'h': |
583 help(); | |
4754 | 584 return 0; |
33 | 585 } |
586 } | |
2967 | 587 |
633 | 588 if(optind <argc) test= atoi(argv[optind]); |
2967 | 589 |
33 | 590 printf("ffmpeg DCT/IDCT test\n"); |
591 | |
720 | 592 if (test_248_dct) { |
6001 | 593 idct248_error("SIMPLE-C", ff_simple_idct248_put); |
33 | 594 } else { |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
595 for (i=0;algos[i].name;i++) |
7140
654f063dc099
Use a local variable instead of mm_flags in dct-test.
astrange
parents:
7135
diff
changeset
|
596 if (algos[i].is_idct == test_idct && !(~cpu_flags & algos[i].mm_support)) { |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
597 dct_error (algos[i].name, algos[i].is_idct, algos[i].func, algos[i].ref, algos[i].format, test); |
720 | 598 } |
33 | 599 } |
0 | 600 return 0; |
601 } |