annotate dct-test.c @ 1998:5bc1a9ad6c33 libavcodec

mmx dct optimization
author michael
date Fri, 30 Apr 2004 19:53:54 +0000
parents 1396e33a9cb6
children 062b2d5d1eba
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
1106
1e39f273ecd6 per file doxy
michaelni
parents: 1064
diff changeset
1 /**
1e39f273ecd6 per file doxy
michaelni
parents: 1064
diff changeset
2 * @file dct-test.c
1e39f273ecd6 per file doxy
michaelni
parents: 1064
diff changeset
3 * DCT test. (c) 2001 Fabrice Bellard.
1e39f273ecd6 per file doxy
michaelni
parents: 1064
diff changeset
4 * Started from sample code by Juan J. Sierralta P.
1e39f273ecd6 per file doxy
michaelni
parents: 1064
diff changeset
5 */
1e39f273ecd6 per file doxy
michaelni
parents: 1064
diff changeset
6
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
7 #include <stdlib.h>
986e461dc072 Initial revision
glantau
parents:
diff changeset
8 #include <stdio.h>
986e461dc072 Initial revision
glantau
parents:
diff changeset
9 #include <string.h>
986e461dc072 Initial revision
glantau
parents:
diff changeset
10 #include <sys/time.h>
986e461dc072 Initial revision
glantau
parents:
diff changeset
11 #include <unistd.h>
986e461dc072 Initial revision
glantau
parents:
diff changeset
12
986e461dc072 Initial revision
glantau
parents:
diff changeset
13 #include "dsputil.h"
986e461dc072 Initial revision
glantau
parents:
diff changeset
14
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
15 #include "i386/mmx.h"
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
16 #include "simple_idct.h"
1557
5d53c03186a1 floating point AAN DCT
michael
parents: 1106
diff changeset
17 #include "faandct.h"
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
18
980
0c32f81b42b2 suppressed getopt.h
bellard
parents: 722
diff changeset
19 #ifndef MAX
0c32f81b42b2 suppressed getopt.h
bellard
parents: 722
diff changeset
20 #define MAX(a, b) (((a) > (b)) ? (a) : (b))
0c32f81b42b2 suppressed getopt.h
bellard
parents: 722
diff changeset
21 #endif
0c32f81b42b2 suppressed getopt.h
bellard
parents: 722
diff changeset
22
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
23 /* reference fdct/idct */
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
24 extern void fdct(DCTELEM *block);
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
25 extern void idct(DCTELEM *block);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
26 extern void init_fdct();
986e461dc072 Initial revision
glantau
parents:
diff changeset
27
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
28 extern void j_rev_dct(DCTELEM *data);
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
29 extern void ff_mmx_idct(DCTELEM *data);
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
30 extern void ff_mmxext_idct(DCTELEM *data);
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
31
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
32 extern void odivx_idct_c (short *block);
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
33
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
34 #define AANSCALE_BITS 12
986e461dc072 Initial revision
glantau
parents:
diff changeset
35 static const unsigned short aanscales[64] = {
986e461dc072 Initial revision
glantau
parents:
diff changeset
36 /* precomputed values scaled up by 14 bits */
986e461dc072 Initial revision
glantau
parents:
diff changeset
37 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
986e461dc072 Initial revision
glantau
parents:
diff changeset
38 22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270,
986e461dc072 Initial revision
glantau
parents:
diff changeset
39 21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906,
986e461dc072 Initial revision
glantau
parents:
diff changeset
40 19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315,
986e461dc072 Initial revision
glantau
parents:
diff changeset
41 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
986e461dc072 Initial revision
glantau
parents:
diff changeset
42 12873, 17855, 16819, 15137, 12873, 10114, 6967, 3552,
986e461dc072 Initial revision
glantau
parents:
diff changeset
43 8867, 12299, 11585, 10426, 8867, 6967, 4799, 2446,
986e461dc072 Initial revision
glantau
parents:
diff changeset
44 4520, 6270, 5906, 5315, 4520, 3552, 2446, 1247
986e461dc072 Initial revision
glantau
parents:
diff changeset
45 };
986e461dc072 Initial revision
glantau
parents:
diff changeset
46
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 980
diff changeset
47 uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
48
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 980
diff changeset
49 int64_t gettime(void)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
50 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
51 struct timeval tv;
986e461dc072 Initial revision
glantau
parents:
diff changeset
52 gettimeofday(&tv,NULL);
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 980
diff changeset
53 return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
54 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
55
986e461dc072 Initial revision
glantau
parents:
diff changeset
56 #define NB_ITS 20000
986e461dc072 Initial revision
glantau
parents:
diff changeset
57 #define NB_ITS_SPEED 50000
986e461dc072 Initial revision
glantau
parents:
diff changeset
58
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
59 static short idct_mmx_perm[64];
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
60
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
61 static short idct_simple_mmx_perm[64]={
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
62 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
63 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
64 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
65 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
66 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
67 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
68 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
69 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
70 };
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
71
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
72 void idct_mmx_init(void)
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
73 {
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
74 int i;
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
75
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
76 /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
77 for (i = 0; i < 64; i++) {
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
78 idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
79 // idct_simple_mmx_perm[i] = simple_block_permute_op(i);
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
80 }
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
81 }
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
82
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
83 static DCTELEM block[64] __attribute__ ((aligned (8)));
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
84 static DCTELEM block1[64] __attribute__ ((aligned (8)));
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
85 static DCTELEM block_org[64] __attribute__ ((aligned (8)));
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
86
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
87 void dct_error(const char *name, int is_idct,
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
88 void (*fdct_func)(DCTELEM *block),
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
89 void (*fdct_ref)(DCTELEM *block), int test)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
90 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
91 int it, i, scale;
986e461dc072 Initial revision
glantau
parents:
diff changeset
92 int err_inf, v;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 980
diff changeset
93 int64_t err2, ti, ti1, it1;
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 980
diff changeset
94 int64_t sysErr[64], sysErrMax=0;
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
95 int maxout=0;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
96 int blockSumErrMax=0, blockSumErr;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
97
986e461dc072 Initial revision
glantau
parents:
diff changeset
98 srandom(0);
986e461dc072 Initial revision
glantau
parents:
diff changeset
99
986e461dc072 Initial revision
glantau
parents:
diff changeset
100 err_inf = 0;
986e461dc072 Initial revision
glantau
parents:
diff changeset
101 err2 = 0;
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
102 for(i=0; i<64; i++) sysErr[i]=0;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
103 for(it=0;it<NB_ITS;it++) {
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
104 for(i=0;i<64;i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
105 block1[i] = 0;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
106 switch(test){
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
107 case 0:
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
108 for(i=0;i<64;i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
109 block1[i] = (random() % 512) -256;
635
3e0f62e5eed6 dct cleanup
michaelni
parents: 633
diff changeset
110 if (is_idct){
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
111 fdct(block1);
635
3e0f62e5eed6 dct cleanup
michaelni
parents: 633
diff changeset
112
3e0f62e5eed6 dct cleanup
michaelni
parents: 633
diff changeset
113 for(i=0;i<64;i++)
3e0f62e5eed6 dct cleanup
michaelni
parents: 633
diff changeset
114 block1[i]>>=3;
3e0f62e5eed6 dct cleanup
michaelni
parents: 633
diff changeset
115 }
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
116 break;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
117 case 1:{
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
118 int num= (random()%10)+1;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
119 for(i=0;i<num;i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
120 block1[random()%64] = (random() % 512) -256;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
121 }break;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
122 case 2:
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
123 block1[0]= (random()%4096)-2048;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
124 block1[63]= (block1[0]&1)^1;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
125 break;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
126 }
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
127
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
128 #if 0 // simulate mismatch control
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
129 { int sum=0;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
130 for(i=0;i<64;i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
131 sum+=block1[i];
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
132
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
133 if((sum&1)==0) block1[63]^=1;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
134 }
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
135 #endif
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
136
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
137 for(i=0; i<64; i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
138 block_org[i]= block1[i];
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
139
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
140 if (fdct_func == ff_mmx_idct ||
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
141 fdct_func == j_rev_dct || fdct_func == ff_mmxext_idct) {
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
142 for(i=0;i<64;i++)
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
143 block[idct_mmx_perm[i]] = block1[i];
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
144 } else if(fdct_func == ff_simple_idct_mmx ) {
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
145 for(i=0;i<64;i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
146 block[idct_simple_mmx_perm[i]] = block1[i];
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
147
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
148 } else {
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
149 for(i=0; i<64; i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
150 block[i]= block1[i];
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
151 }
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
152 #if 0 // simulate mismatch control for tested IDCT but not the ref
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
153 { int sum=0;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
154 for(i=0;i<64;i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
155 sum+=block[i];
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
156
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
157 if((sum&1)==0) block[63]^=1;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
158 }
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
159 #endif
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
160
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
161 fdct_func(block);
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
162 emms(); /* for ff_mmx_idct */
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
163
1562
bf452704100f optionally merge postscale into quantization table for the float aan dct
michael
parents: 1557
diff changeset
164 if (fdct_func == fdct_ifast
bf452704100f optionally merge postscale into quantization table for the float aan dct
michael
parents: 1557
diff changeset
165 #ifndef FAAN_POSTSCALE
bf452704100f optionally merge postscale into quantization table for the float aan dct
michael
parents: 1557
diff changeset
166 || fdct_func == ff_faandct
bf452704100f optionally merge postscale into quantization table for the float aan dct
michael
parents: 1557
diff changeset
167 #endif
bf452704100f optionally merge postscale into quantization table for the float aan dct
michael
parents: 1557
diff changeset
168 ) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
169 for(i=0; i<64; i++) {
635
3e0f62e5eed6 dct cleanup
michaelni
parents: 633
diff changeset
170 scale = 8*(1 << (AANSCALE_BITS + 11)) / aanscales[i];
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
171 block[i] = (block[i] * scale /*+ (1<<(AANSCALE_BITS-1))*/) >> AANSCALE_BITS;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
172 }
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
173 }
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
174
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
175 fdct_ref(block1);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
176
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
177 blockSumErr=0;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
178 for(i=0;i<64;i++) {
986e461dc072 Initial revision
glantau
parents:
diff changeset
179 v = abs(block[i] - block1[i]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
180 if (v > err_inf)
986e461dc072 Initial revision
glantau
parents:
diff changeset
181 err_inf = v;
986e461dc072 Initial revision
glantau
parents:
diff changeset
182 err2 += v * v;
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
183 sysErr[i] += block[i] - block1[i];
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
184 blockSumErr += v;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
185 if( abs(block[i])>maxout) maxout=abs(block[i]);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
186 }
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
187 if(blockSumErrMax < blockSumErr) blockSumErrMax= blockSumErr;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
188 #if 0 // print different matrix pairs
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
189 if(blockSumErr){
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
190 printf("\n");
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
191 for(i=0; i<64; i++){
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
192 if((i&7)==0) printf("\n");
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
193 printf("%4d ", block_org[i]);
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
194 }
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
195 for(i=0; i<64; i++){
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
196 if((i&7)==0) printf("\n");
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
197 printf("%4d ", block[i] - block1[i]);
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
198 }
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
199 }
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
200 #endif
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
201 }
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
202 for(i=0; i<64; i++) sysErrMax= MAX(sysErrMax, ABS(sysErr[i]));
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
203
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
204 #if 1 // dump systematic errors
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
205 for(i=0; i<64; i++){
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
206 if(i%8==0) printf("\n");
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
207 printf("%5d ", (int)sysErr[i]);
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
208 }
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
209 printf("\n");
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
210 #endif
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
211
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
212 printf("%s %s: err_inf=%d err2=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
213 is_idct ? "IDCT" : "DCT",
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
214 name, err_inf, (double)err2 / NB_ITS / 64.0, (double)sysErrMax / NB_ITS, maxout, blockSumErrMax);
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
215 #if 1 //Speed test
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
216 /* speed test */
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
217 for(i=0;i<64;i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
218 block1[i] = 0;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
219 switch(test){
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
220 case 0:
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
221 for(i=0;i<64;i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
222 block1[i] = (random() % 512) -256;
635
3e0f62e5eed6 dct cleanup
michaelni
parents: 633
diff changeset
223 if (is_idct){
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
224 fdct(block1);
635
3e0f62e5eed6 dct cleanup
michaelni
parents: 633
diff changeset
225
3e0f62e5eed6 dct cleanup
michaelni
parents: 633
diff changeset
226 for(i=0;i<64;i++)
3e0f62e5eed6 dct cleanup
michaelni
parents: 633
diff changeset
227 block1[i]>>=3;
3e0f62e5eed6 dct cleanup
michaelni
parents: 633
diff changeset
228 }
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
229 break;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
230 case 1:{
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
231 case 2:
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
232 block1[0] = (random() % 512) -256;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
233 block1[1] = (random() % 512) -256;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
234 block1[2] = (random() % 512) -256;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
235 block1[3] = (random() % 512) -256;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
236 }break;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
237 }
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
238
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
239 if (fdct_func == ff_mmx_idct ||
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
240 fdct_func == j_rev_dct || fdct_func == ff_mmxext_idct) {
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
241 for(i=0;i<64;i++)
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
242 block[idct_mmx_perm[i]] = block1[i];
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
243 } else if(fdct_func == ff_simple_idct_mmx ) {
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
244 for(i=0;i<64;i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
245 block[idct_simple_mmx_perm[i]] = block1[i];
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
246 } else {
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
247 for(i=0; i<64; i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
248 block[i]= block1[i];
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
249 }
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
250
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
251 ti = gettime();
986e461dc072 Initial revision
glantau
parents:
diff changeset
252 it1 = 0;
986e461dc072 Initial revision
glantau
parents:
diff changeset
253 do {
986e461dc072 Initial revision
glantau
parents:
diff changeset
254 for(it=0;it<NB_ITS_SPEED;it++) {
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
255 for(i=0; i<64; i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
256 block[i]= block1[i];
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
257 // memcpy(block, block1, sizeof(DCTELEM) * 64);
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
258 // dont memcpy especially not fastmemcpy because it does movntq !!!
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
259 fdct_func(block);
986e461dc072 Initial revision
glantau
parents:
diff changeset
260 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
261 it1 += NB_ITS_SPEED;
986e461dc072 Initial revision
glantau
parents:
diff changeset
262 ti1 = gettime() - ti;
986e461dc072 Initial revision
glantau
parents:
diff changeset
263 } while (ti1 < 1000000);
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
264 emms();
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
265
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
266 printf("%s %s: %0.1f kdct/s\n",
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
267 is_idct ? "IDCT" : "DCT",
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
268 name, (double)it1 * 1000.0 / (double)ti1);
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
269 #endif
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
270 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
271
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 980
diff changeset
272 static uint8_t img_dest[64] __attribute__ ((aligned (8)));
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 980
diff changeset
273 static uint8_t img_dest1[64] __attribute__ ((aligned (8)));
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
274
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 980
diff changeset
275 void idct248_ref(uint8_t *dest, int linesize, int16_t *block)
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
276 {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
277 static int init;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
278 static double c8[8][8];
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
279 static double c4[4][4];
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
280 double block1[64], block2[64], block3[64];
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
281 double s, sum, v;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
282 int i, j, k;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
283
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
284 if (!init) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
285 init = 1;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
286
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
287 for(i=0;i<8;i++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
288 sum = 0;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
289 for(j=0;j<8;j++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
290 s = (i==0) ? sqrt(1.0/8.0) : sqrt(1.0/4.0);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
291 c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
292 sum += c8[i][j] * c8[i][j];
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
293 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
294 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
295
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
296 for(i=0;i<4;i++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
297 sum = 0;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
298 for(j=0;j<4;j++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
299 s = (i==0) ? sqrt(1.0/4.0) : sqrt(1.0/2.0);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
300 c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
301 sum += c4[i][j] * c4[i][j];
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
302 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
303 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
304 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
305
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
306 /* butterfly */
722
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
307 s = 0.5 * sqrt(2.0);
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
308 for(i=0;i<4;i++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
309 for(j=0;j<8;j++) {
722
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
310 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
311 block1[8*(2*i+1)+j] = (block[8*(2*i)+j] - block[8*(2*i+1)+j]) * s;
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
312 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
313 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
314
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
315 /* idct8 on lines */
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
316 for(i=0;i<8;i++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
317 for(j=0;j<8;j++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
318 sum = 0;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
319 for(k=0;k<8;k++)
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
320 sum += c8[k][j] * block1[8*i+k];
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
321 block2[8*i+j] = sum;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
322 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
323 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
324
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
325 /* idct4 */
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
326 for(i=0;i<8;i++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
327 for(j=0;j<4;j++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
328 /* top */
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
329 sum = 0;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
330 for(k=0;k<4;k++)
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
331 sum += c4[k][j] * block2[8*(2*k)+i];
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
332 block3[8*(2*j)+i] = sum;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
333
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
334 /* bottom */
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
335 sum = 0;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
336 for(k=0;k<4;k++)
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
337 sum += c4[k][j] * block2[8*(2*k+1)+i];
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
338 block3[8*(2*j+1)+i] = sum;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
339 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
340 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
341
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
342 /* clamp and store the result */
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
343 for(i=0;i<8;i++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
344 for(j=0;j<8;j++) {
722
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
345 v = block3[8*i+j];
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
346 if (v < 0)
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
347 v = 0;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
348 else if (v > 255)
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
349 v = 255;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
350 dest[i * linesize + j] = (int)rint(v);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
351 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
352 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
353 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
354
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
355 void idct248_error(const char *name,
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 980
diff changeset
356 void (*idct248_put)(uint8_t *dest, int line_size, int16_t *block))
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
357 {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
358 int it, i, it1, ti, ti1, err_max, v;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
359
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
360 srandom(0);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
361
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
362 /* just one test to see if code is correct (precision is less
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
363 important here) */
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
364 err_max = 0;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
365 for(it=0;it<NB_ITS;it++) {
722
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
366
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
367 /* XXX: use forward transform to generate values */
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
368 for(i=0;i<64;i++)
722
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
369 block1[i] = (random() % 256) - 128;
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
370 block1[0] += 1024;
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
371
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
372 for(i=0; i<64; i++)
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
373 block[i]= block1[i];
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
374 idct248_ref(img_dest1, 8, block);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
375
722
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
376 for(i=0; i<64; i++)
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
377 block[i]= block1[i];
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
378 idct248_put(img_dest, 8, block);
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
379
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
380 for(i=0;i<64;i++) {
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
381 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
382 if (v == 255)
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
383 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
384 if (v > err_max)
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
385 err_max = v;
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 720
diff changeset
386 }
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
387 #if 0
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
388 printf("ref=\n");
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
389 for(i=0;i<8;i++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
390 int j;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
391 for(j=0;j<8;j++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
392 printf(" %3d", img_dest1[i*8+j]);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
393 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
394 printf("\n");
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
395 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
396
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
397 printf("out=\n");
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
398 for(i=0;i<8;i++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
399 int j;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
400 for(j=0;j<8;j++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
401 printf(" %3d", img_dest[i*8+j]);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
402 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
403 printf("\n");
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
404 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
405 #endif
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
406 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
407 printf("%s %s: err_inf=%d\n",
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
408 1 ? "IDCT248" : "DCT248",
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
409 name, err_max);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
410
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
411 ti = gettime();
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
412 it1 = 0;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
413 do {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
414 for(it=0;it<NB_ITS_SPEED;it++) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
415 for(i=0; i<64; i++)
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
416 block[i]= block1[i];
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
417 // memcpy(block, block1, sizeof(DCTELEM) * 64);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
418 // dont memcpy especially not fastmemcpy because it does movntq !!!
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
419 idct248_put(img_dest, 8, block);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
420 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
421 it1 += NB_ITS_SPEED;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
422 ti1 = gettime() - ti;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
423 } while (ti1 < 1000000);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
424 emms();
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
425
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
426 printf("%s %s: %0.1f kdct/s\n",
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
427 1 ? "IDCT248" : "DCT248",
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
428 name, (double)it1 * 1000.0 / (double)ti1);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
429 }
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
430
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
431 void help(void)
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
432 {
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
433 printf("dct-test [-i] [<test-number>]\n"
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
434 "test-number 0 -> test with random matrixes\n"
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
435 " 1 -> test with random sparse matrixes\n"
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
436 " 2 -> do 3. test from mpeg4 std\n"
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
437 "-i test IDCT implementations\n"
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
438 "-4 test IDCT248 implementations\n");
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
439 exit(1);
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
440 }
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
441
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
442 int main(int argc, char **argv)
986e461dc072 Initial revision
glantau
parents:
diff changeset
443 {
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
444 int test_idct = 0, test_248_dct = 0;
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
445 int c,i;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
446 int test=1;
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
447
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
448 init_fdct();
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
449 idct_mmx_init();
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
450
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
451 for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
452 for(i=0;i<MAX_NEG_CROP;i++) {
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
453 cropTbl[i] = 0;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
454 cropTbl[i + MAX_NEG_CROP + 256] = 255;
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
455 }
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
456
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
457 for(;;) {
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
458 c = getopt(argc, argv, "ih4");
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
459 if (c == -1)
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
460 break;
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
461 switch(c) {
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
462 case 'i':
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
463 test_idct = 1;
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
464 break;
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
465 case '4':
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
466 test_248_dct = 1;
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
467 break;
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
468 default :
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
469 case 'h':
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
470 help();
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
471 break;
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
472 }
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
473 }
633
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
474
e7b72c1dfa1b dct-test update
michaelni
parents: 474
diff changeset
475 if(optind <argc) test= atoi(argv[optind]);
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
476
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
477 printf("ffmpeg DCT/IDCT test\n");
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
478
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
479 if (test_248_dct) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
480 idct248_error("SIMPLE-C", simple_idct248_put);
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
481 } else {
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
482 if (!test_idct) {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
483 dct_error("REF-DBL", 0, fdct, fdct, test); /* only to verify code ! */
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
484 dct_error("IJG-AAN-INT", 0, fdct_ifast, fdct, test);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
485 dct_error("IJG-LLM-INT", 0, ff_jpeg_fdct_islow, fdct, test);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
486 dct_error("MMX", 0, ff_fdct_mmx, fdct, test);
1574
michael
parents: 1562
diff changeset
487 dct_error("MMX2", 0, ff_fdct_mmx2, fdct, test);
1557
5d53c03186a1 floating point AAN DCT
michael
parents: 1106
diff changeset
488 dct_error("FAAN", 0, ff_faandct, fdct, test);
720
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
489 } else {
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
490 dct_error("REF-DBL", 1, idct, idct, test);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
491 dct_error("INT", 1, j_rev_dct, idct, test);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
492 dct_error("LIBMPEG2-MMX", 1, ff_mmx_idct, idct, test);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
493 dct_error("LIBMPEG2-MMXEXT", 1, ff_mmxext_idct, idct, test);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
494 dct_error("SIMPLE-C", 1, simple_idct, idct, test);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
495 dct_error("SIMPLE-MMX", 1, ff_simple_idct_mmx, idct, test);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
496 // dct_error("ODIVX-C", 1, odivx_idct_c, idct);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
497 //printf(" test against odivx idct\n");
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
498 // dct_error("REF", 1, idct, odivx_idct_c);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
499 // dct_error("INT", 1, j_rev_dct, odivx_idct_c);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
500 // dct_error("MMX", 1, ff_mmx_idct, odivx_idct_c);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
501 // dct_error("MMXEXT", 1, ff_mmxext_idct, odivx_idct_c);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
502 // dct_error("SIMPLE-C", 1, simple_idct, odivx_idct_c);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
503 // dct_error("SIMPLE-MMX", 1, ff_simple_idct_mmx, odivx_idct_c);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
504 // dct_error("ODIVX-C", 1, odivx_idct_c, odivx_idct_c);
409bdaa0b964 added IDCT248 testing
bellard
parents: 687
diff changeset
505 }
33
7cf705a32d1c updated dct-test to test IDCTs too
glantau
parents: 0
diff changeset
506 }
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
507 return 0;
986e461dc072 Initial revision
glantau
parents:
diff changeset
508 }