Mercurial > libavcodec.hg
annotate dct-test.c @ 12512:58bd8cb86f5d libavcodec
Remove unused variable.
author | rbultje |
---|---|
date | Fri, 24 Sep 2010 15:31:46 +0000 |
parents | 9fef0a8ddd63 |
children |
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 /** |
11644
7dd2a45249a9
Remove explicit filename from Doxygen @file commands.
diego
parents:
11414
diff
changeset
|
23 * @file |
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 |
12475
9fef0a8ddd63
Move mm_support() from libavcodec to libavutil, make it a public
stefano
parents:
12456
diff
changeset
|
35 #include "libavutil/cpu.h" |
7130
601509a430f7
Replace redundant MAX macro declaration by proper use of FFMAX.
diego
parents:
7125
diff
changeset
|
36 #include "libavutil/common.h" |
9199
ea0e5e9a520f
Replace random() usage in test programs by av_lfg_*().
diego
parents:
9189
diff
changeset
|
37 #include "libavutil/lfg.h" |
0 | 38 |
633 | 39 #include "simple_idct.h" |
8223 | 40 #include "aandcttab.h" |
1557 | 41 #include "faandct.h" |
6407 | 42 #include "faanidct.h" |
8430 | 43 #include "x86/idct_xvid.h" |
11408 | 44 #include "dctref.h" |
33 | 45 |
2872 | 46 #undef printf |
47 | |
8250 | 48 void ff_mmx_idct(DCTELEM *data); |
49 void ff_mmxext_idct(DCTELEM *data); | |
33 | 50 |
8250 | 51 void odivx_idct_c(short *block); |
633 | 52 |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
53 // BFIN |
8250 | 54 void ff_bfin_idct(DCTELEM *block); |
55 void ff_bfin_fdct(DCTELEM *block); | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
56 |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
57 // ALTIVEC |
8250 | 58 void fdct_altivec(DCTELEM *block); |
59 //void idct_altivec(DCTELEM *block);?? no routine | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
60 |
8351 | 61 // ARM |
11414 | 62 void ff_j_rev_dct_arm(DCTELEM *data); |
63 void ff_simple_idct_arm(DCTELEM *data); | |
64 void ff_simple_idct_armv5te(DCTELEM *data); | |
8351 | 65 void ff_simple_idct_armv6(DCTELEM *data); |
66 void ff_simple_idct_neon(DCTELEM *data); | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
67 |
8622 | 68 void ff_simple_idct_axp(DCTELEM *data); |
69 | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
70 struct algo { |
7125
65e7e714c770
Mark constant structure member as const to avoid some warnings.
diego
parents:
6602
diff
changeset
|
71 const char *name; |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
72 enum { FDCT, IDCT } is_idct; |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
73 void (* func) (DCTELEM *block); |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
74 void (* ref) (DCTELEM *block); |
8350 | 75 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
|
76 int mm_support; |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
77 }; |
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 #ifndef FAAN_POSTSCALE |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
80 #define FAAN_SCALE SCALE_PERM |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
81 #else |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
82 #define FAAN_SCALE NO_PERM |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
83 #endif |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
84 |
7155 | 85 static int cpu_flags; |
86 | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
87 struct algo algos[] = { |
9293 | 88 {"REF-DBL", 0, ff_ref_fdct, ff_ref_fdct, NO_PERM}, |
89 {"FAAN", 0, ff_faandct, ff_ref_fdct, FAAN_SCALE}, | |
90 {"FAANI", 1, ff_faanidct, ff_ref_idct, NO_PERM}, | |
91 {"IJG-AAN-INT", 0, fdct_ifast, ff_ref_fdct, SCALE_PERM}, | |
92 {"IJG-LLM-INT", 0, ff_jpeg_fdct_islow, ff_ref_fdct, NO_PERM}, | |
93 {"REF-DBL", 1, ff_ref_idct, ff_ref_idct, NO_PERM}, | |
94 {"INT", 1, j_rev_dct, ff_ref_idct, MMX_PERM}, | |
95 {"SIMPLE-C", 1, ff_simple_idct, ff_ref_idct, NO_PERM}, | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
96 |
8590 | 97 #if HAVE_MMX |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
11644
diff
changeset
|
98 {"MMX", 0, ff_fdct_mmx, ff_ref_fdct, NO_PERM, AV_CPU_FLAG_MMX}, |
8590 | 99 #if HAVE_MMX2 |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
11644
diff
changeset
|
100 {"MMX2", 0, ff_fdct_mmx2, ff_ref_fdct, NO_PERM, AV_CPU_FLAG_MMX2}, |
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
11644
diff
changeset
|
101 {"SSE2", 0, ff_fdct_sse2, ff_ref_fdct, NO_PERM, AV_CPU_FLAG_SSE2}, |
5110 | 102 #endif |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
103 |
8590 | 104 #if CONFIG_GPL |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
11644
diff
changeset
|
105 {"LIBMPEG2-MMX", 1, ff_mmx_idct, ff_ref_idct, MMX_PERM, AV_CPU_FLAG_MMX}, |
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
11644
diff
changeset
|
106 {"LIBMPEG2-MMX2", 1, ff_mmxext_idct, ff_ref_idct, MMX_PERM, AV_CPU_FLAG_MMX2}, |
5109 | 107 #endif |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
11644
diff
changeset
|
108 {"SIMPLE-MMX", 1, ff_simple_idct_mmx, ff_ref_idct, MMX_SIMPLE_PERM, AV_CPU_FLAG_MMX}, |
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
11644
diff
changeset
|
109 {"XVID-MMX", 1, ff_idct_xvid_mmx, ff_ref_idct, NO_PERM, AV_CPU_FLAG_MMX}, |
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
11644
diff
changeset
|
110 {"XVID-MMX2", 1, ff_idct_xvid_mmx2, ff_ref_idct, NO_PERM, AV_CPU_FLAG_MMX2}, |
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
11644
diff
changeset
|
111 {"XVID-SSE2", 1, ff_idct_xvid_sse2, ff_ref_idct, SSE2_PERM, AV_CPU_FLAG_SSE2}, |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
112 #endif |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
113 |
8590 | 114 #if HAVE_ALTIVEC |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
11644
diff
changeset
|
115 {"altivecfdct", 0, fdct_altivec, ff_ref_fdct, NO_PERM, AV_CPU_FLAG_ALTIVEC}, |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
116 #endif |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
117 |
8590 | 118 #if ARCH_BFIN |
9293 | 119 {"BFINfdct", 0, ff_bfin_fdct, ff_ref_fdct, NO_PERM}, |
120 {"BFINidct", 1, ff_bfin_idct, ff_ref_idct, NO_PERM}, | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
121 #endif |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
122 |
8590 | 123 #if ARCH_ARM |
11414 | 124 {"SIMPLE-ARM", 1, ff_simple_idct_arm, ff_ref_idct, NO_PERM }, |
125 {"INT-ARM", 1, ff_j_rev_dct_arm, ff_ref_idct, MMX_PERM }, | |
8590 | 126 #if HAVE_ARMV5TE |
11414 | 127 {"SIMPLE-ARMV5TE", 1, ff_simple_idct_armv5te, ff_ref_idct, NO_PERM }, |
8351 | 128 #endif |
8590 | 129 #if HAVE_ARMV6 |
9293 | 130 {"SIMPLE-ARMV6", 1, ff_simple_idct_armv6, ff_ref_idct, MMX_PERM }, |
8351 | 131 #endif |
8590 | 132 #if HAVE_NEON |
9293 | 133 {"SIMPLE-NEON", 1, ff_simple_idct_neon, ff_ref_idct, PARTTRANS_PERM }, |
8351 | 134 #endif |
8359 | 135 #endif /* ARCH_ARM */ |
8351 | 136 |
8622 | 137 #if ARCH_ALPHA |
9293 | 138 {"SIMPLE-ALPHA", 1, ff_simple_idct_axp, ff_ref_idct, NO_PERM }, |
8622 | 139 #endif |
140 | |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
141 { 0 } |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
142 }; |
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
143 |
0 | 144 #define AANSCALE_BITS 12 |
145 | |
4197 | 146 uint8_t cropTbl[256 + 2 * MAX_NEG_CROP]; |
633 | 147 |
9295
b225f51903af
Mark non-exported functions in test and example programs as static.
diego
parents:
9293
diff
changeset
|
148 static int64_t gettime(void) |
0 | 149 { |
150 struct timeval tv; | |
151 gettimeofday(&tv,NULL); | |
1064 | 152 return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec; |
0 | 153 } |
154 | |
155 #define NB_ITS 20000 | |
156 #define NB_ITS_SPEED 50000 | |
157 | |
33 | 158 static short idct_mmx_perm[64]; |
159 | |
633 | 160 static short idct_simple_mmx_perm[64]={ |
2979 | 161 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D, |
162 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D, | |
163 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D, | |
164 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F, | |
165 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F, | |
166 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D, | |
167 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F, | |
168 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F, | |
633 | 169 }; |
170 | |
6602 | 171 static const uint8_t idct_sse2_row_perm[8] = {0, 4, 1, 5, 2, 6, 3, 7}; |
172 | |
9295
b225f51903af
Mark non-exported functions in test and example programs as static.
diego
parents:
9293
diff
changeset
|
173 static void idct_mmx_init(void) |
33 | 174 { |
175 int i; | |
176 | |
177 /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */ | |
178 for (i = 0; i < 64; i++) { | |
2979 | 179 idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2); |
180 // idct_simple_mmx_perm[i] = simple_block_permute_op(i); | |
33 | 181 } |
182 } | |
183 | |
10961
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
9793
diff
changeset
|
184 DECLARE_ALIGNED(16, static DCTELEM, block)[64]; |
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
9793
diff
changeset
|
185 DECLARE_ALIGNED(8, static DCTELEM, block1)[64]; |
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
9793
diff
changeset
|
186 DECLARE_ALIGNED(8, static DCTELEM, block_org)[64]; |
33 | 187 |
7155 | 188 static inline void mmx_emms(void) |
189 { | |
8590 | 190 #if HAVE_MMX |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
11644
diff
changeset
|
191 if (cpu_flags & AV_CPU_FLAG_MMX) |
8031 | 192 __asm__ volatile ("emms\n\t"); |
7155 | 193 #endif |
194 } | |
195 | |
9295
b225f51903af
Mark non-exported functions in test and example programs as static.
diego
parents:
9293
diff
changeset
|
196 static void dct_error(const char *name, int is_idct, |
33 | 197 void (*fdct_func)(DCTELEM *block), |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
198 void (*fdct_ref)(DCTELEM *block), int form, int test) |
0 | 199 { |
200 int it, i, scale; | |
201 int err_inf, v; | |
1064 | 202 int64_t err2, ti, ti1, it1; |
203 int64_t sysErr[64], sysErrMax=0; | |
633 | 204 int maxout=0; |
205 int blockSumErrMax=0, blockSumErr; | |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
206 AVLFG prng; |
0 | 207 |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
208 av_lfg_init(&prng, 1); |
0 | 209 |
210 err_inf = 0; | |
211 err2 = 0; | |
633 | 212 for(i=0; i<64; i++) sysErr[i]=0; |
0 | 213 for(it=0;it<NB_ITS;it++) { |
633 | 214 for(i=0;i<64;i++) |
215 block1[i] = 0; | |
216 switch(test){ | |
2967 | 217 case 0: |
633 | 218 for(i=0;i<64;i++) |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
219 block1[i] = (av_lfg_get(&prng) % 512) -256; |
635 | 220 if (is_idct){ |
9293 | 221 ff_ref_fdct(block1); |
635 | 222 |
223 for(i=0;i<64;i++) | |
224 block1[i]>>=3; | |
225 } | |
633 | 226 break; |
227 case 1:{ | |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
228 int num = av_lfg_get(&prng) % 10 + 1; |
633 | 229 for(i=0;i<num;i++) |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
230 block1[av_lfg_get(&prng) % 64] = av_lfg_get(&prng) % 512 -256; |
633 | 231 }break; |
232 case 2: | |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
233 block1[0] = av_lfg_get(&prng) % 4096 - 2048; |
633 | 234 block1[63]= (block1[0]&1)^1; |
235 break; | |
236 } | |
33 | 237 |
633 | 238 #if 0 // simulate mismatch control |
239 { int sum=0; | |
240 for(i=0;i<64;i++) | |
241 sum+=block1[i]; | |
242 | |
2967 | 243 if((sum&1)==0) block1[63]^=1; |
633 | 244 } |
245 #endif | |
246 | |
247 for(i=0; i<64; i++) | |
248 block_org[i]= block1[i]; | |
33 | 249 |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
250 if (form == MMX_PERM) { |
633 | 251 for(i=0;i<64;i++) |
33 | 252 block[idct_mmx_perm[i]] = block1[i]; |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
253 } else if (form == MMX_SIMPLE_PERM) { |
633 | 254 for(i=0;i<64;i++) |
255 block[idct_simple_mmx_perm[i]] = block1[i]; | |
256 | |
6602 | 257 } else if (form == SSE2_PERM) { |
258 for(i=0; i<64; i++) | |
259 block[(i&0x38) | idct_sse2_row_perm[i&7]] = block1[i]; | |
8350 | 260 } else if (form == PARTTRANS_PERM) { |
261 for(i=0; i<64; i++) | |
262 block[(i&0x24) | ((i&3)<<3) | ((i>>3)&3)] = block1[i]; | |
2979 | 263 } else { |
633 | 264 for(i=0; i<64; i++) |
265 block[i]= block1[i]; | |
33 | 266 } |
633 | 267 #if 0 // simulate mismatch control for tested IDCT but not the ref |
268 { int sum=0; | |
269 for(i=0;i<64;i++) | |
270 sum+=block[i]; | |
271 | |
2967 | 272 if((sum&1)==0) block[63]^=1; |
633 | 273 } |
274 #endif | |
33 | 275 |
0 | 276 fdct_func(block); |
7155 | 277 mmx_emms(); |
33 | 278 |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
279 if (form == SCALE_PERM) { |
0 | 280 for(i=0; i<64; i++) { |
8223 | 281 scale = 8*(1 << (AANSCALE_BITS + 11)) / ff_aanscales[i]; |
633 | 282 block[i] = (block[i] * scale /*+ (1<<(AANSCALE_BITS-1))*/) >> AANSCALE_BITS; |
283 } | |
284 } | |
285 | |
33 | 286 fdct_ref(block1); |
0 | 287 |
633 | 288 blockSumErr=0; |
0 | 289 for(i=0;i<64;i++) { |
290 v = abs(block[i] - block1[i]); | |
291 if (v > err_inf) | |
292 err_inf = v; | |
293 err2 += v * v; | |
2979 | 294 sysErr[i] += block[i] - block1[i]; |
295 blockSumErr += v; | |
296 if( abs(block[i])>maxout) maxout=abs(block[i]); | |
0 | 297 } |
633 | 298 if(blockSumErrMax < blockSumErr) blockSumErrMax= blockSumErr; |
299 #if 0 // print different matrix pairs | |
300 if(blockSumErr){ | |
301 printf("\n"); | |
302 for(i=0; i<64; i++){ | |
303 if((i&7)==0) printf("\n"); | |
304 printf("%4d ", block_org[i]); | |
305 } | |
306 for(i=0; i<64; i++){ | |
307 if((i&7)==0) printf("\n"); | |
308 printf("%4d ", block[i] - block1[i]); | |
309 } | |
310 } | |
311 #endif | |
0 | 312 } |
7130
601509a430f7
Replace redundant MAX macro declaration by proper use of FFMAX.
diego
parents:
7125
diff
changeset
|
313 for(i=0; i<64; i++) sysErrMax= FFMAX(sysErrMax, FFABS(sysErr[i])); |
2967 | 314 |
633 | 315 #if 1 // dump systematic errors |
316 for(i=0; i<64; i++){ | |
2979 | 317 if(i%8==0) printf("\n"); |
9189 | 318 printf("%7d ", (int)sysErr[i]); |
633 | 319 } |
320 printf("\n"); | |
321 #endif | |
2967 | 322 |
633 | 323 printf("%s %s: err_inf=%d err2=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n", |
33 | 324 is_idct ? "IDCT" : "DCT", |
633 | 325 name, err_inf, (double)err2 / NB_ITS / 64.0, (double)sysErrMax / NB_ITS, maxout, blockSumErrMax); |
326 #if 1 //Speed test | |
0 | 327 /* speed test */ |
633 | 328 for(i=0;i<64;i++) |
329 block1[i] = 0; | |
330 switch(test){ | |
2967 | 331 case 0: |
633 | 332 for(i=0;i<64;i++) |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
333 block1[i] = av_lfg_get(&prng) % 512 -256; |
635 | 334 if (is_idct){ |
9293 | 335 ff_ref_fdct(block1); |
635 | 336 |
337 for(i=0;i<64;i++) | |
338 block1[i]>>=3; | |
339 } | |
633 | 340 break; |
341 case 1:{ | |
342 case 2: | |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
343 block1[0] = av_lfg_get(&prng) % 512 -256; |
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
344 block1[1] = av_lfg_get(&prng) % 512 -256; |
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
345 block1[2] = av_lfg_get(&prng) % 512 -256; |
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
346 block1[3] = av_lfg_get(&prng) % 512 -256; |
633 | 347 }break; |
348 } | |
0 | 349 |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
350 if (form == MMX_PERM) { |
633 | 351 for(i=0;i<64;i++) |
33 | 352 block[idct_mmx_perm[i]] = block1[i]; |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
353 } else if(form == MMX_SIMPLE_PERM) { |
633 | 354 for(i=0;i<64;i++) |
355 block[idct_simple_mmx_perm[i]] = block1[i]; | |
356 } else { | |
357 for(i=0; i<64; i++) | |
358 block[i]= block1[i]; | |
33 | 359 } |
360 | |
0 | 361 ti = gettime(); |
362 it1 = 0; | |
363 do { | |
364 for(it=0;it<NB_ITS_SPEED;it++) { | |
633 | 365 for(i=0; i<64; i++) |
366 block[i]= block1[i]; | |
367 // memcpy(block, block1, sizeof(DCTELEM) * 64); | |
5127 | 368 // do not memcpy especially not fastmemcpy because it does movntq !!! |
0 | 369 fdct_func(block); |
370 } | |
371 it1 += NB_ITS_SPEED; | |
372 ti1 = gettime() - ti; | |
373 } while (ti1 < 1000000); | |
7155 | 374 mmx_emms(); |
0 | 375 |
633 | 376 printf("%s %s: %0.1f kdct/s\n", |
33 | 377 is_idct ? "IDCT" : "DCT", |
0 | 378 name, (double)it1 * 1000.0 / (double)ti1); |
633 | 379 #endif |
0 | 380 } |
381 | |
10961
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
9793
diff
changeset
|
382 DECLARE_ALIGNED(8, static uint8_t, img_dest)[64]; |
34a65026fa06
Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents:
9793
diff
changeset
|
383 DECLARE_ALIGNED(8, static uint8_t, img_dest1)[64]; |
720 | 384 |
9295
b225f51903af
Mark non-exported functions in test and example programs as static.
diego
parents:
9293
diff
changeset
|
385 static void idct248_ref(uint8_t *dest, int linesize, int16_t *block) |
720 | 386 { |
387 static int init; | |
388 static double c8[8][8]; | |
389 static double c4[4][4]; | |
390 double block1[64], block2[64], block3[64]; | |
391 double s, sum, v; | |
392 int i, j, k; | |
393 | |
394 if (!init) { | |
395 init = 1; | |
396 | |
397 for(i=0;i<8;i++) { | |
398 sum = 0; | |
399 for(j=0;j<8;j++) { | |
400 s = (i==0) ? sqrt(1.0/8.0) : sqrt(1.0/4.0); | |
401 c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0); | |
402 sum += c8[i][j] * c8[i][j]; | |
403 } | |
404 } | |
2967 | 405 |
720 | 406 for(i=0;i<4;i++) { |
407 sum = 0; | |
408 for(j=0;j<4;j++) { | |
409 s = (i==0) ? sqrt(1.0/4.0) : sqrt(1.0/2.0); | |
410 c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0); | |
411 sum += c4[i][j] * c4[i][j]; | |
412 } | |
413 } | |
414 } | |
415 | |
416 /* butterfly */ | |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
417 s = 0.5 * sqrt(2.0); |
720 | 418 for(i=0;i<4;i++) { |
419 for(j=0;j<8;j++) { | |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
420 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
|
421 block1[8*(2*i+1)+j] = (block[8*(2*i)+j] - block[8*(2*i+1)+j]) * s; |
720 | 422 } |
423 } | |
424 | |
425 /* idct8 on lines */ | |
426 for(i=0;i<8;i++) { | |
427 for(j=0;j<8;j++) { | |
428 sum = 0; | |
429 for(k=0;k<8;k++) | |
430 sum += c8[k][j] * block1[8*i+k]; | |
431 block2[8*i+j] = sum; | |
432 } | |
433 } | |
434 | |
435 /* idct4 */ | |
436 for(i=0;i<8;i++) { | |
437 for(j=0;j<4;j++) { | |
438 /* top */ | |
439 sum = 0; | |
440 for(k=0;k<4;k++) | |
441 sum += c4[k][j] * block2[8*(2*k)+i]; | |
442 block3[8*(2*j)+i] = sum; | |
443 | |
444 /* bottom */ | |
445 sum = 0; | |
446 for(k=0;k<4;k++) | |
447 sum += c4[k][j] * block2[8*(2*k+1)+i]; | |
448 block3[8*(2*j+1)+i] = sum; | |
449 } | |
450 } | |
451 | |
452 /* clamp and store the result */ | |
453 for(i=0;i<8;i++) { | |
454 for(j=0;j<8;j++) { | |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
455 v = block3[8*i+j]; |
720 | 456 if (v < 0) |
457 v = 0; | |
458 else if (v > 255) | |
459 v = 255; | |
460 dest[i * linesize + j] = (int)rint(v); | |
461 } | |
462 } | |
463 } | |
464 | |
9295
b225f51903af
Mark non-exported functions in test and example programs as static.
diego
parents:
9293
diff
changeset
|
465 static void idct248_error(const char *name, |
1064 | 466 void (*idct248_put)(uint8_t *dest, int line_size, int16_t *block)) |
720 | 467 { |
468 int it, i, it1, ti, ti1, err_max, v; | |
469 | |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
470 AVLFG prng; |
9199
ea0e5e9a520f
Replace random() usage in test programs by av_lfg_*().
diego
parents:
9189
diff
changeset
|
471 |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
472 av_lfg_init(&prng, 1); |
2967 | 473 |
720 | 474 /* just one test to see if code is correct (precision is less |
475 important here) */ | |
476 err_max = 0; | |
477 for(it=0;it<NB_ITS;it++) { | |
2967 | 478 |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
479 /* XXX: use forward transform to generate values */ |
720 | 480 for(i=0;i<64;i++) |
9388
2313bf51945b
cosmetics: Rename prn variable to prng (Pseudo Random Number Generator).
diego
parents:
9295
diff
changeset
|
481 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
|
482 block1[0] += 1024; |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
483 |
720 | 484 for(i=0; i<64; i++) |
485 block[i]= block1[i]; | |
486 idct248_ref(img_dest1, 8, block); | |
2967 | 487 |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
488 for(i=0; i<64; i++) |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
489 block[i]= block1[i]; |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
490 idct248_put(img_dest, 8, block); |
2967 | 491 |
722
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
492 for(i=0;i<64;i++) { |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
493 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
|
494 if (v == 255) |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
495 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
|
496 if (v > err_max) |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
497 err_max = v; |
ff90043f4a2d
in fact IDCT248 needs to be normalized as I suspected
bellard
parents:
720
diff
changeset
|
498 } |
720 | 499 #if 0 |
500 printf("ref=\n"); | |
501 for(i=0;i<8;i++) { | |
502 int j; | |
503 for(j=0;j<8;j++) { | |
504 printf(" %3d", img_dest1[i*8+j]); | |
505 } | |
506 printf("\n"); | |
507 } | |
2967 | 508 |
720 | 509 printf("out=\n"); |
510 for(i=0;i<8;i++) { | |
511 int j; | |
512 for(j=0;j<8;j++) { | |
513 printf(" %3d", img_dest[i*8+j]); | |
514 } | |
515 printf("\n"); | |
516 } | |
517 #endif | |
518 } | |
519 printf("%s %s: err_inf=%d\n", | |
520 1 ? "IDCT248" : "DCT248", | |
521 name, err_max); | |
522 | |
523 ti = gettime(); | |
524 it1 = 0; | |
525 do { | |
526 for(it=0;it<NB_ITS_SPEED;it++) { | |
527 for(i=0; i<64; i++) | |
528 block[i]= block1[i]; | |
529 // memcpy(block, block1, sizeof(DCTELEM) * 64); | |
5127 | 530 // do not memcpy especially not fastmemcpy because it does movntq !!! |
720 | 531 idct248_put(img_dest, 8, block); |
532 } | |
533 it1 += NB_ITS_SPEED; | |
534 ti1 = gettime() - ti; | |
535 } while (ti1 < 1000000); | |
7155 | 536 mmx_emms(); |
720 | 537 |
538 printf("%s %s: %0.1f kdct/s\n", | |
539 1 ? "IDCT248" : "DCT248", | |
540 name, (double)it1 * 1000.0 / (double)ti1); | |
541 } | |
542 | |
9295
b225f51903af
Mark non-exported functions in test and example programs as static.
diego
parents:
9293
diff
changeset
|
543 static void help(void) |
33 | 544 { |
633 | 545 printf("dct-test [-i] [<test-number>]\n" |
546 "test-number 0 -> test with random matrixes\n" | |
547 " 1 -> test with random sparse matrixes\n" | |
548 " 2 -> do 3. test from mpeg4 std\n" | |
720 | 549 "-i test IDCT implementations\n" |
550 "-4 test IDCT248 implementations\n"); | |
33 | 551 } |
552 | |
0 | 553 int main(int argc, char **argv) |
554 { | |
720 | 555 int test_idct = 0, test_248_dct = 0; |
633 | 556 int c,i; |
557 int test=1; | |
12475
9fef0a8ddd63
Move mm_support() from libavcodec to libavutil, make it a public
stefano
parents:
12456
diff
changeset
|
558 cpu_flags = av_get_cpu_flags(); |
33 | 559 |
9293 | 560 ff_ref_dct_init(); |
33 | 561 idct_mmx_init(); |
7135
6bd6a2da306e
Define mm_flags/support to be 0 on architectures where they don't exist.
astrange
parents:
7130
diff
changeset
|
562 |
4197 | 563 for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i; |
633 | 564 for(i=0;i<MAX_NEG_CROP;i++) { |
4197 | 565 cropTbl[i] = 0; |
566 cropTbl[i + MAX_NEG_CROP + 256] = 255; | |
633 | 567 } |
2967 | 568 |
33 | 569 for(;;) { |
720 | 570 c = getopt(argc, argv, "ih4"); |
33 | 571 if (c == -1) |
572 break; | |
573 switch(c) { | |
574 case 'i': | |
575 test_idct = 1; | |
576 break; | |
720 | 577 case '4': |
578 test_248_dct = 1; | |
579 break; | |
633 | 580 default : |
33 | 581 case 'h': |
582 help(); | |
4754 | 583 return 0; |
33 | 584 } |
585 } | |
2967 | 586 |
633 | 587 if(optind <argc) test= atoi(argv[optind]); |
2967 | 588 |
33 | 589 printf("ffmpeg DCT/IDCT test\n"); |
590 | |
720 | 591 if (test_248_dct) { |
6001 | 592 idct248_error("SIMPLE-C", ff_simple_idct248_put); |
33 | 593 } else { |
4755
b7c27288e509
platform-independent restructuring and code simplification
diego
parents:
4754
diff
changeset
|
594 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
|
595 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
|
596 dct_error (algos[i].name, algos[i].is_idct, algos[i].func, algos[i].ref, algos[i].format, test); |
720 | 597 } |
33 | 598 } |
0 | 599 return 0; |
600 } |