* Started from sample code by Juan J. Sierralta P.
*/
+#include "config.h"
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
-#include <sys/time.h>
+#if HAVE_UNISTD_H
#include <unistd.h>
+#endif
#include <math.h>
#include "libavutil/cpu.h"
#include "libavutil/common.h"
#include "libavutil/lfg.h"
+#include "libavutil/time.h"
+#include "dct.h"
+#include "idctdsp.h"
#include "simple_idct.h"
+#include "xvididct.h"
#include "aandcttab.h"
#include "faandct.h"
#include "faanidct.h"
-#include "x86/idct_xvid.h"
#include "dctref.h"
-#undef printf
-
-void ff_mmx_idct(DCTELEM *data);
-void ff_mmxext_idct(DCTELEM *data);
-
-void odivx_idct_c(short *block);
-
-// BFIN
-void ff_bfin_idct(DCTELEM *block);
-void ff_bfin_fdct(DCTELEM *block);
-
-// ALTIVEC
-void fdct_altivec(DCTELEM *block);
-//void idct_altivec(DCTELEM *block);?? no routine
-
-// ARM
-void ff_j_rev_dct_arm(DCTELEM *data);
-void ff_simple_idct_arm(DCTELEM *data);
-void ff_simple_idct_armv5te(DCTELEM *data);
-void ff_simple_idct_armv6(DCTELEM *data);
-void ff_simple_idct_neon(DCTELEM *data);
-
-void ff_simple_idct_axp(DCTELEM *data);
-
struct algo {
const char *name;
- enum { FDCT, IDCT } is_idct;
- void (*func)(DCTELEM *block);
- void (*ref) (DCTELEM *block);
- enum formattag { NO_PERM, MMX_PERM, MMX_SIMPLE_PERM, SCALE_PERM,
- SSE2_PERM, PARTTRANS_PERM } format;
- int mm_support;
+ void (*func)(int16_t *block);
+ enum idct_permutation_type perm_type;
+ int cpu_flag;
+ int nonspec;
};
-#ifndef FAAN_POSTSCALE
-#define FAAN_SCALE SCALE_PERM
-#else
-#define FAAN_SCALE NO_PERM
-#endif
-
-static int cpu_flags;
-
-struct algo algos[] = {
- {"REF-DBL", 0, ff_ref_fdct, ff_ref_fdct, NO_PERM},
- {"FAAN", 0, ff_faandct, ff_ref_fdct, FAAN_SCALE},
- {"FAANI", 1, ff_faanidct, ff_ref_idct, NO_PERM},
- {"IJG-AAN-INT", 0, fdct_ifast, ff_ref_fdct, SCALE_PERM},
- {"IJG-LLM-INT", 0, ff_jpeg_fdct_islow, ff_ref_fdct, NO_PERM},
- {"REF-DBL", 1, ff_ref_idct, ff_ref_idct, NO_PERM},
- {"INT", 1, j_rev_dct, ff_ref_idct, MMX_PERM},
- {"SIMPLE-C", 1, ff_simple_idct, ff_ref_idct, NO_PERM},
-
-#if HAVE_MMX
- {"MMX", 0, ff_fdct_mmx, ff_ref_fdct, NO_PERM, AV_CPU_FLAG_MMX},
-#if HAVE_MMX2
- {"MMX2", 0, ff_fdct_mmx2, ff_ref_fdct, NO_PERM, AV_CPU_FLAG_MMX2},
- {"SSE2", 0, ff_fdct_sse2, ff_ref_fdct, NO_PERM, AV_CPU_FLAG_SSE2},
-#endif
-
-#if CONFIG_GPL
- {"LIBMPEG2-MMX", 1, ff_mmx_idct, ff_ref_idct, MMX_PERM, AV_CPU_FLAG_MMX},
- {"LIBMPEG2-MMX2", 1, ff_mmxext_idct, ff_ref_idct, MMX_PERM, AV_CPU_FLAG_MMX2},
-#endif
- {"SIMPLE-MMX", 1, ff_simple_idct_mmx, ff_ref_idct, MMX_SIMPLE_PERM, AV_CPU_FLAG_MMX},
- {"XVID-MMX", 1, ff_idct_xvid_mmx, ff_ref_idct, NO_PERM, AV_CPU_FLAG_MMX},
- {"XVID-MMX2", 1, ff_idct_xvid_mmx2, ff_ref_idct, NO_PERM, AV_CPU_FLAG_MMX2},
- {"XVID-SSE2", 1, ff_idct_xvid_sse2, ff_ref_idct, SSE2_PERM, AV_CPU_FLAG_SSE2},
-#endif
-
-#if HAVE_ALTIVEC
- {"altivecfdct", 0, fdct_altivec, ff_ref_fdct, NO_PERM, AV_CPU_FLAG_ALTIVEC},
-#endif
+static const struct algo fdct_tab[] = {
+ { "REF-DBL", ff_ref_fdct, FF_IDCT_PERM_NONE },
+ { "IJG-AAN-INT", ff_fdct_ifast, FF_IDCT_PERM_NONE },
+ { "IJG-LLM-INT", ff_jpeg_fdct_islow_8, FF_IDCT_PERM_NONE },
+#if CONFIG_FAANDCT
+ { "FAAN", ff_faandct, FF_IDCT_PERM_NONE },
+#endif /* CONFIG_FAANDCT */
+};
-#if ARCH_BFIN
- {"BFINfdct", 0, ff_bfin_fdct, ff_ref_fdct, NO_PERM},
- {"BFINidct", 1, ff_bfin_idct, ff_ref_idct, NO_PERM},
-#endif
+static const struct algo idct_tab[] = {
+ { "REF-DBL", ff_ref_idct, FF_IDCT_PERM_NONE },
+ { "INT", ff_j_rev_dct, FF_IDCT_PERM_LIBMPEG2 },
+ { "SIMPLE-C", ff_simple_idct_8, FF_IDCT_PERM_NONE },
+#if CONFIG_FAANIDCT
+ { "FAANI", ff_faanidct, FF_IDCT_PERM_NONE },
+#endif /* CONFIG_FAANIDCT */
+#if CONFIG_MPEG4_DECODER
+ { "XVID", ff_xvid_idct, FF_IDCT_PERM_NONE, 0, 1 },
+#endif /* CONFIG_MPEG4_DECODER */
+};
#if ARCH_ARM
- {"SIMPLE-ARM", 1, ff_simple_idct_arm, ff_ref_idct, NO_PERM },
- {"INT-ARM", 1, ff_j_rev_dct_arm, ff_ref_idct, MMX_PERM },
-#if HAVE_ARMV5TE
- {"SIMPLE-ARMV5TE", 1, ff_simple_idct_armv5te, ff_ref_idct, NO_PERM },
-#endif
-#if HAVE_ARMV6
- {"SIMPLE-ARMV6", 1, ff_simple_idct_armv6, ff_ref_idct, MMX_PERM },
-#endif
-#if HAVE_NEON
- {"SIMPLE-NEON", 1, ff_simple_idct_neon, ff_ref_idct, PARTTRANS_PERM },
-#endif
-#endif /* ARCH_ARM */
-
-#if ARCH_ALPHA
- {"SIMPLE-ALPHA", 1, ff_simple_idct_axp, ff_ref_idct, NO_PERM },
+#include "arm/dct-test.c"
+#elif ARCH_PPC
+#include "ppc/dct-test.c"
+#elif ARCH_X86
+#include "x86/dct-test.c"
+#else
+static const struct algo fdct_tab_arch[] = { 0 };
+static const struct algo idct_tab_arch[] = { 0 };
#endif
- { 0 }
-};
-
#define AANSCALE_BITS 12
-uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
-
-static int64_t gettime(void)
-{
- struct timeval tv;
- gettimeofday(&tv, NULL);
- return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec;
-}
-
#define NB_ITS 20000
#define NB_ITS_SPEED 50000
-static short idct_mmx_perm[64];
-
-static short idct_simple_mmx_perm[64] = {
- 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
- 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
- 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
- 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
- 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
- 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
- 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
- 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
-};
-
-static const uint8_t idct_sse2_row_perm[8] = { 0, 4, 1, 5, 2, 6, 3, 7 };
+DECLARE_ALIGNED(16, static int16_t, block)[64];
+DECLARE_ALIGNED(8, static int16_t, block1)[64];
-static void idct_mmx_init(void)
+static void init_block(int16_t block[64], int test, int is_idct, AVLFG *prng)
{
- int i;
+ int i, j;
- /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */
- for (i = 0; i < 64; i++) {
- idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
+ memset(block, 0, 64 * sizeof(*block));
+
+ switch (test) {
+ case 0:
+ for (i = 0; i < 64; i++)
+ block[i] = (av_lfg_get(prng) % 512) - 256;
+ if (is_idct) {
+ ff_ref_fdct(block);
+ for (i = 0; i < 64; i++)
+ block[i] >>= 3;
+ }
+ break;
+ case 1:
+ j = av_lfg_get(prng) % 10 + 1;
+ for (i = 0; i < j; i++)
+ block[av_lfg_get(prng) % 64] = av_lfg_get(prng) % 512 - 256;
+ break;
+ case 2:
+ block[ 0] = av_lfg_get(prng) % 4096 - 2048;
+ block[63] = (block[0] & 1) ^ 1;
+ break;
}
}
-DECLARE_ALIGNED(16, static DCTELEM, block)[64];
-DECLARE_ALIGNED(8, static DCTELEM, block1)[64];
-DECLARE_ALIGNED(8, static DCTELEM, block_org)[64];
-
-static inline void mmx_emms(void)
+static void permute(int16_t dst[64], const int16_t src[64],
+ enum idct_permutation_type perm_type)
{
-#if HAVE_MMX
- if (cpu_flags & AV_CPU_FLAG_MMX)
- __asm__ volatile ("emms\n\t");
+ int i;
+
+#if ARCH_X86
+ if (permute_x86(dst, src, perm_type))
+ return;
#endif
+
+ switch (perm_type) {
+ case FF_IDCT_PERM_LIBMPEG2:
+ for (i = 0; i < 64; i++)
+ dst[(i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2)] = src[i];
+ break;
+ case FF_IDCT_PERM_PARTTRANS:
+ for (i = 0; i < 64; i++)
+ dst[(i & 0x24) | ((i & 3) << 3) | ((i >> 3) & 3)] = src[i];
+ break;
+ default:
+ for (i = 0; i < 64; i++)
+ dst[i] = src[i];
+ break;
+ }
}
-static void dct_error(const char *name, int is_idct,
- void (*fdct_func)(DCTELEM *block),
- void (*fdct_ref)(DCTELEM *block), int form,
- int test)
+static int dct_error(const struct algo *dct, int test, int is_idct, int speed)
{
+ void (*ref)(int16_t *block) = is_idct ? ff_ref_idct : ff_ref_fdct;
int it, i, scale;
int err_inf, v;
- int64_t err2, ti, ti1, it1;
+ int64_t err2, ti, ti1, it1, err_sum = 0;
int64_t sysErr[64], sysErrMax = 0;
int maxout = 0;
int blockSumErrMax = 0, blockSumErr;
AVLFG prng;
+ double omse, ome;
+ int spec_err;
av_lfg_init(&prng, 1);
for (i = 0; i < 64; i++)
sysErr[i] = 0;
for (it = 0; it < NB_ITS; it++) {
- for (i = 0; i < 64; i++)
- block1[i] = 0;
- switch (test) {
- case 0:
- for (i = 0; i < 64; i++)
- block1[i] = (av_lfg_get(&prng) % 512) - 256;
- if (is_idct) {
- ff_ref_fdct(block1);
- for (i = 0; i < 64; i++)
- block1[i] >>= 3;
- }
- break;
- case 1: {
- int num = av_lfg_get(&prng) % 10 + 1;
- for (i = 0; i < num; i++)
- block1[av_lfg_get(&prng) % 64] =
- av_lfg_get(&prng) % 512 - 256;
- }
- break;
- case 2:
- block1[0] = av_lfg_get(&prng) % 4096 - 2048;
- block1[63] = (block1[0] & 1) ^ 1;
- break;
- }
+ init_block(block1, test, is_idct, &prng);
+ permute(block, block1, dct->perm_type);
- for (i = 0; i < 64; i++)
- block_org[i] = block1[i];
+ dct->func(block);
+ emms_c();
- if (form == MMX_PERM) {
- for (i = 0; i < 64; i++)
- block[idct_mmx_perm[i]] = block1[i];
- } else if (form == MMX_SIMPLE_PERM) {
- for (i = 0; i < 64; i++)
- block[idct_simple_mmx_perm[i]] = block1[i];
- } else if (form == SSE2_PERM) {
- for (i = 0; i < 64; i++)
- block[(i & 0x38) | idct_sse2_row_perm[i & 7]] = block1[i];
- } else if (form == PARTTRANS_PERM) {
- for (i = 0; i < 64; i++)
- block[(i & 0x24) | ((i & 3) << 3) | ((i >> 3) & 3)] = block1[i];
- } else {
- for (i = 0; i < 64; i++)
- block[i] = block1[i];
- }
-
- fdct_func(block);
- mmx_emms();
-
- if (form == SCALE_PERM) {
+ if (!strcmp(dct->name, "IJG-AAN-INT")) {
for (i = 0; i < 64; i++) {
scale = 8 * (1 << (AANSCALE_BITS + 11)) / ff_aanscales[i];
block[i] = (block[i] * scale) >> AANSCALE_BITS;
}
}
- fdct_ref(block1);
+ ref(block1);
blockSumErr = 0;
for (i = 0; i < 64; i++) {
- v = abs(block[i] - block1[i]);
+ int err = block[i] - block1[i];
+ err_sum += err;
+ v = abs(err);
if (v > err_inf)
err_inf = v;
err2 += v * v;
}
printf("\n");
- printf("%s %s: err_inf=%d err2=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
- is_idct ? "IDCT" : "DCT", name, err_inf,
- (double) err2 / NB_ITS / 64.0, (double) sysErrMax / NB_ITS,
+ omse = (double) err2 / NB_ITS / 64;
+ ome = (double) err_sum / NB_ITS / 64;
+
+ spec_err = is_idct && (err_inf > 1 || omse > 0.02 || fabs(ome) > 0.0015);
+
+ printf("%s %s: ppe=%d omse=%0.8f ome=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
+ is_idct ? "IDCT" : "DCT", dct->name, err_inf,
+ omse, ome, (double) sysErrMax / NB_ITS,
maxout, blockSumErrMax);
- /* speed test */
- for (i = 0; i < 64; i++)
- block1[i] = 0;
+ if (spec_err && !dct->nonspec)
+ return 1;
- switch (test) {
- case 0:
- for (i = 0; i < 64; i++)
- block1[i] = av_lfg_get(&prng) % 512 - 256;
- if (is_idct) {
- ff_ref_fdct(block1);
- for (i = 0; i < 64; i++)
- block1[i] >>= 3;
- }
- break;
- case 1:
- case 2:
- block1[0] = av_lfg_get(&prng) % 512 - 256;
- block1[1] = av_lfg_get(&prng) % 512 - 256;
- block1[2] = av_lfg_get(&prng) % 512 - 256;
- block1[3] = av_lfg_get(&prng) % 512 - 256;
- break;
- }
+ if (!speed)
+ return 0;
- if (form == MMX_PERM) {
- for (i = 0; i < 64; i++)
- block[idct_mmx_perm[i]] = block1[i];
- } else if (form == MMX_SIMPLE_PERM) {
- for (i = 0; i < 64; i++)
- block[idct_simple_mmx_perm[i]] = block1[i];
- } else {
- for (i = 0; i < 64; i++)
- block[i] = block1[i];
- }
+ /* speed test */
+ init_block(block, test, is_idct, &prng);
+ permute(block1, block, dct->perm_type);
- ti = gettime();
+ ti = av_gettime_relative();
it1 = 0;
do {
for (it = 0; it < NB_ITS_SPEED; it++) {
- for (i = 0; i < 64; i++)
- block[i] = block1[i];
- fdct_func(block);
+ memcpy(block, block1, sizeof(block));
+ dct->func(block);
}
it1 += NB_ITS_SPEED;
- ti1 = gettime() - ti;
+ ti1 = av_gettime_relative() - ti;
} while (ti1 < 1000000);
- mmx_emms();
+ emms_c();
- printf("%s %s: %0.1f kdct/s\n", is_idct ? "IDCT" : "DCT", name,
+ printf("%s %s: %0.1f kdct/s\n", is_idct ? "IDCT" : "DCT", dct->name,
(double) it1 * 1000.0 / (double) ti1);
+
+ return 0;
}
DECLARE_ALIGNED(8, static uint8_t, img_dest)[64];
static void idct248_error(const char *name,
void (*idct248_put)(uint8_t *dest, int line_size,
- int16_t *block))
+ int16_t *block),
+ int speed)
{
int it, i, it1, ti, ti1, err_max, v;
AVLFG prng;
}
printf("%s %s: err_inf=%d\n", 1 ? "IDCT248" : "DCT248", name, err_max);
- ti = gettime();
+ if (!speed)
+ return;
+
+ ti = av_gettime_relative();
it1 = 0;
do {
for (it = 0; it < NB_ITS_SPEED; it++) {
idct248_put(img_dest, 8, block);
}
it1 += NB_ITS_SPEED;
- ti1 = gettime() - ti;
+ ti1 = av_gettime_relative() - ti;
} while (ti1 < 1000000);
- mmx_emms();
+ emms_c();
printf("%s %s: %0.1f kdct/s\n", 1 ? "IDCT248" : "DCT248", name,
(double) it1 * 1000.0 / (double) ti1);
" 1 -> test with random sparse matrixes\n"
" 2 -> do 3. test from mpeg4 std\n"
"-i test IDCT implementations\n"
- "-4 test IDCT248 implementations\n");
+ "-4 test IDCT248 implementations\n"
+ "-t speed test\n");
}
+#if !HAVE_GETOPT
+#include "compat/getopt.c"
+#endif
+
int main(int argc, char **argv)
{
int test_idct = 0, test_248_dct = 0;
int c, i;
int test = 1;
-
- cpu_flags = av_get_cpu_flags();
+ int speed = 0;
+ int err = 0;
ff_ref_dct_init();
- idct_mmx_init();
-
- for (i = 0; i < 256; i++)
- cropTbl[i + MAX_NEG_CROP] = i;
- for (i = 0; i < MAX_NEG_CROP; i++) {
- cropTbl[i] = 0;
- cropTbl[i + MAX_NEG_CROP + 256] = 255;
- }
for (;;) {
- c = getopt(argc, argv, "ih4");
+ c = getopt(argc, argv, "ih4t");
if (c == -1)
break;
switch (c) {
case '4':
test_248_dct = 1;
break;
+ case 't':
+ speed = 1;
+ break;
default:
case 'h':
help();
if (optind < argc)
test = atoi(argv[optind]);
- printf("ffmpeg DCT/IDCT test\n");
+ printf("Libav DCT/IDCT test\n");
if (test_248_dct) {
- idct248_error("SIMPLE-C", ff_simple_idct248_put);
+ idct248_error("SIMPLE-C", ff_simple_idct248_put, speed);
} else {
- for (i = 0; algos[i].name; i++)
- if (algos[i].is_idct == test_idct &&
- !(~cpu_flags & algos[i].mm_support)) {
- dct_error(algos[i].name, algos[i].is_idct, algos[i].func,
- algos[i].ref, algos[i].format, test);
- }
+ const int cpu_flags = av_get_cpu_flags();
+ if (test_idct) {
+ for (i = 0; i < FF_ARRAY_ELEMS(idct_tab); i++)
+ err |= dct_error(&idct_tab[i], test, test_idct, speed);
+
+ for (i = 0; idct_tab_arch[i].name; i++)
+ if (!(~cpu_flags & idct_tab_arch[i].cpu_flag))
+ err |= dct_error(&idct_tab_arch[i], test, test_idct, speed);
+ }
+#if CONFIG_FDCTDSP
+ else {
+ for (i = 0; i < FF_ARRAY_ELEMS(fdct_tab); i++)
+ err |= dct_error(&fdct_tab[i], test, test_idct, speed);
+
+ for (i = 0; fdct_tab_arch[i].name; i++)
+ if (!(~cpu_flags & fdct_tab_arch[i].cpu_flag))
+ err |= dct_error(&fdct_tab_arch[i], test, test_idct, speed);
+ }
+#endif /* CONFIG_FDCTDSP */
}
- return 0;
+
+ if (err)
+ printf("Error: %d.\n", err);
+
+ return !!err;
}
projects / ffmpeg / blobdiff