#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);
-
-// BFIN
-void ff_bfin_idct(DCTELEM *block);
-void ff_bfin_fdct(DCTELEM *block);
-
-// ALTIVEC
-void ff_fdct_altivec(DCTELEM *block);
-//void ff_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;
- void (*func)(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;
};
-static int cpu_flags;
-
static const struct algo fdct_tab[] = {
- { "REF-DBL", ff_ref_fdct, NO_PERM },
- { "FAAN", ff_faandct, NO_PERM },
- { "IJG-AAN-INT", ff_fdct_ifast, SCALE_PERM },
- { "IJG-LLM-INT", ff_jpeg_fdct_islow_8, NO_PERM },
-
-#if HAVE_MMX
- { "MMX", ff_fdct_mmx, NO_PERM, AV_CPU_FLAG_MMX },
- { "MMX2", ff_fdct_mmx2, NO_PERM, AV_CPU_FLAG_MMX2 },
- { "SSE2", ff_fdct_sse2, NO_PERM, AV_CPU_FLAG_SSE2 },
-#endif
-
-#if HAVE_ALTIVEC
- { "altivecfdct", ff_fdct_altivec, NO_PERM, AV_CPU_FLAG_ALTIVEC },
-#endif
-
-#if ARCH_BFIN
- { "BFINfdct", ff_bfin_fdct, NO_PERM },
-#endif
-
- { 0 }
+ { "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 */
};
static const struct algo idct_tab[] = {
- { "FAANI", ff_faanidct, NO_PERM },
- { "REF-DBL", ff_ref_idct, NO_PERM },
- { "INT", ff_j_rev_dct, MMX_PERM },
- { "SIMPLE-C", ff_simple_idct_8, NO_PERM },
-
-#if HAVE_MMX
-#if CONFIG_GPL
- { "LIBMPEG2-MMX", ff_mmx_idct, MMX_PERM, AV_CPU_FLAG_MMX, 1 },
- { "LIBMPEG2-MMX2", ff_mmxext_idct, MMX_PERM, AV_CPU_FLAG_MMX2, 1 },
-#endif
- { "SIMPLE-MMX", ff_simple_idct_mmx, MMX_SIMPLE_PERM, AV_CPU_FLAG_MMX },
- { "XVID-MMX", ff_idct_xvid_mmx, NO_PERM, AV_CPU_FLAG_MMX, 1 },
- { "XVID-MMX2", ff_idct_xvid_mmx2, NO_PERM, AV_CPU_FLAG_MMX2, 1 },
- { "XVID-SSE2", ff_idct_xvid_sse2, SSE2_PERM, AV_CPU_FLAG_SSE2, 1 },
-#endif
-
-#if ARCH_BFIN
- { "BFINidct", ff_bfin_idct, NO_PERM },
-#endif
+ { "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", ff_simple_idct_arm, NO_PERM },
- { "INT-ARM", ff_j_rev_dct_arm, MMX_PERM },
-#endif
-#if HAVE_ARMV5TE
- { "SIMPLE-ARMV5TE", ff_simple_idct_armv5te,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
-#if HAVE_ARMV6
- { "SIMPLE-ARMV6", ff_simple_idct_armv6, MMX_PERM },
-#endif
-#if HAVE_NEON
- { "SIMPLE-NEON", ff_simple_idct_neon, PARTTRANS_PERM },
-#endif
-
-#if ARCH_ALPHA
- { "SIMPLE-ALPHA", ff_simple_idct_axp, NO_PERM },
-#endif
-
- { 0 }
-};
#define AANSCALE_BITS 12
#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 };
-
-static void idct_mmx_init(void)
-{
- int i;
-
- /* 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);
- }
-}
+DECLARE_ALIGNED(16, static int16_t, block)[64];
+DECLARE_ALIGNED(8, static int16_t, block1)[64];
-DECLARE_ALIGNED(16, static DCTELEM, block)[64];
-DECLARE_ALIGNED(8, static DCTELEM, block1)[64];
-
-static void init_block(DCTELEM block[64], int test, int is_idct, AVLFG *prng)
+static void init_block(int16_t block[64], int test, int is_idct, AVLFG *prng)
{
int i, j;
}
}
-static void permute(DCTELEM dst[64], const DCTELEM src[64], int perm)
+static void permute(int16_t dst[64], const int16_t src[64],
+ enum idct_permutation_type perm_type)
{
int i;
- if (perm == MMX_PERM) {
- for (i = 0; i < 64; i++)
- dst[idct_mmx_perm[i]] = src[i];
- } else if (perm == MMX_SIMPLE_PERM) {
- for (i = 0; i < 64; i++)
- dst[idct_simple_mmx_perm[i]] = src[i];
- } else if (perm == SSE2_PERM) {
+#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) | idct_sse2_row_perm[i & 7]] = src[i];
- } else if (perm == PARTTRANS_PERM) {
+ 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];
- } else {
+ break;
+ default:
for (i = 0; i < 64; i++)
dst[i] = src[i];
+ break;
}
}
static int dct_error(const struct algo *dct, int test, int is_idct, int speed)
{
- void (*ref)(DCTELEM *block) = is_idct ? ff_ref_idct : ff_ref_fdct;
+ 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, err_sum = 0;
sysErr[i] = 0;
for (it = 0; it < NB_ITS; it++) {
init_block(block1, test, is_idct, &prng);
- permute(block, block1, dct->format);
+ permute(block, block1, dct->perm_type);
dct->func(block);
emms_c();
- if (dct->format == 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;
/* speed test */
init_block(block, test, is_idct, &prng);
- permute(block1, block, dct->format);
+ permute(block1, block, dct->perm_type);
- ti = av_gettime();
+ ti = av_gettime_relative();
it1 = 0;
do {
for (it = 0; it < NB_ITS_SPEED; it++) {
dct->func(block);
}
it1 += NB_ITS_SPEED;
- ti1 = av_gettime() - ti;
+ ti1 = av_gettime_relative() - ti;
} while (ti1 < 1000000);
emms_c();
if (!speed)
return;
- ti = av_gettime();
+ 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 = av_gettime() - ti;
+ ti1 = av_gettime_relative() - ti;
} while (ti1 < 1000000);
emms_c();
int speed = 0;
int err = 0;
- cpu_flags = av_get_cpu_flags();
-
ff_ref_dct_init();
- idct_mmx_init();
for (;;) {
c = getopt(argc, argv, "ih4t");
if (test_248_dct) {
idct248_error("SIMPLE-C", ff_simple_idct248_put, speed);
} else {
- const struct algo *algos = test_idct ? idct_tab : fdct_tab;
- for (i = 0; algos[i].name; i++)
- if (!(~cpu_flags & algos[i].mm_support)) {
- err |= dct_error(&algos[i], test, test_idct, speed);
- }
+ 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 err;
+ if (err)
+ printf("Error: %d.\n", err);
+
+ return !!err;
}