2006-09-10 17:02:42 +03:00
|
|
|
/*
|
|
|
|
* (c) 2001 Fabrice Bellard
|
2007-03-31 20:24:09 +03:00
|
|
|
* 2007 Marc Hoffman <marc.hoffman@analog.com>
|
2006-09-10 17:02:42 +03:00
|
|
|
*
|
2006-10-07 18:30:46 +03:00
|
|
|
* This file is part of FFmpeg.
|
|
|
|
*
|
|
|
|
* FFmpeg is free software; you can redistribute it and/or
|
2006-09-10 17:02:42 +03:00
|
|
|
* modify it under the terms of the GNU Lesser General Public
|
|
|
|
* License as published by the Free Software Foundation; either
|
2006-10-07 18:30:46 +03:00
|
|
|
* version 2.1 of the License, or (at your option) any later version.
|
2006-09-10 17:02:42 +03:00
|
|
|
*
|
2006-10-07 18:30:46 +03:00
|
|
|
* FFmpeg is distributed in the hope that it will be useful,
|
2006-09-10 17:02:42 +03:00
|
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
|
* Lesser General Public License for more details.
|
|
|
|
*
|
|
|
|
* You should have received a copy of the GNU Lesser General Public
|
2006-10-07 18:30:46 +03:00
|
|
|
* License along with FFmpeg; if not, write to the Free Software
|
2006-09-10 17:02:42 +03:00
|
|
|
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
|
|
|
*/
|
|
|
|
|
2003-03-06 13:32:04 +02:00
|
|
|
/**
|
2010-04-20 17:45:34 +03:00
|
|
|
* @file
|
2009-02-01 17:52:01 +02:00
|
|
|
* DCT test (c) 2001 Fabrice Bellard
|
2003-03-06 13:32:04 +02:00
|
|
|
* Started from sample code by Juan J. Sierralta P.
|
|
|
|
*/
|
|
|
|
|
2012-07-04 06:14:09 +03:00
|
|
|
#include "config.h"
|
2001-07-22 17:18:56 +03:00
|
|
|
#include <stdlib.h>
|
|
|
|
#include <stdio.h>
|
|
|
|
#include <string.h>
|
2012-07-04 06:14:09 +03:00
|
|
|
#if HAVE_UNISTD_H
|
2001-07-22 17:18:56 +03:00
|
|
|
#include <unistd.h>
|
2012-07-04 06:14:09 +03:00
|
|
|
#endif
|
2007-06-09 20:52:16 +03:00
|
|
|
#include <math.h>
|
2001-07-22 17:18:56 +03:00
|
|
|
|
2010-09-08 18:07:14 +03:00
|
|
|
#include "libavutil/cpu.h"
|
2008-06-24 23:15:31 +03:00
|
|
|
#include "libavutil/common.h"
|
2009-03-20 13:48:27 +02:00
|
|
|
#include "libavutil/lfg.h"
|
2012-06-20 03:34:38 +03:00
|
|
|
#include "libavutil/time.h"
|
2001-07-22 17:18:56 +03:00
|
|
|
|
2002-09-01 12:33:05 +03:00
|
|
|
#include "simple_idct.h"
|
2008-11-28 02:32:24 +02:00
|
|
|
#include "aandcttab.h"
|
2003-10-22 21:26:38 +03:00
|
|
|
#include "faandct.h"
|
2008-02-26 16:59:11 +02:00
|
|
|
#include "faanidct.h"
|
2008-12-22 11:12:42 +02:00
|
|
|
#include "x86/idct_xvid.h"
|
2010-03-08 20:43:33 +02:00
|
|
|
#include "dctref.h"
|
2001-08-08 01:42:45 +03:00
|
|
|
|
2005-09-11 02:00:13 +03:00
|
|
|
#undef printf
|
|
|
|
|
2008-12-03 17:23:30 +02:00
|
|
|
void ff_mmx_idct(DCTELEM *data);
|
|
|
|
void ff_mmxext_idct(DCTELEM *data);
|
2001-08-08 01:42:45 +03:00
|
|
|
|
2007-03-31 20:24:09 +03:00
|
|
|
// BFIN
|
2008-12-03 17:23:30 +02:00
|
|
|
void ff_bfin_idct(DCTELEM *block);
|
|
|
|
void ff_bfin_fdct(DCTELEM *block);
|
2007-03-31 20:24:09 +03:00
|
|
|
|
|
|
|
// ALTIVEC
|
2012-02-16 08:15:40 +03:00
|
|
|
void ff_fdct_altivec(DCTELEM *block);
|
2012-02-18 00:12:25 +03:00
|
|
|
//void ff_idct_altivec(DCTELEM *block);?? no routine
|
2007-03-31 20:24:09 +03:00
|
|
|
|
2008-12-16 22:00:07 +02:00
|
|
|
// ARM
|
2010-03-08 21:00:05 +02:00
|
|
|
void ff_j_rev_dct_arm(DCTELEM *data);
|
|
|
|
void ff_simple_idct_arm(DCTELEM *data);
|
|
|
|
void ff_simple_idct_armv5te(DCTELEM *data);
|
2008-12-16 22:00:07 +02:00
|
|
|
void ff_simple_idct_armv6(DCTELEM *data);
|
|
|
|
void ff_simple_idct_neon(DCTELEM *data);
|
2007-03-31 20:24:09 +03:00
|
|
|
|
2009-01-18 07:52:20 +02:00
|
|
|
void ff_simple_idct_axp(DCTELEM *data);
|
|
|
|
|
2007-03-31 20:24:09 +03:00
|
|
|
struct algo {
|
2011-07-04 00:59:03 +03:00
|
|
|
const char *name;
|
|
|
|
void (*func)(DCTELEM *block);
|
|
|
|
enum formattag { NO_PERM, MMX_PERM, MMX_SIMPLE_PERM, SCALE_PERM,
|
2011-10-12 23:01:44 +03:00
|
|
|
SSE2_PERM, PARTTRANS_PERM, TRANSPOSE_PERM } format;
|
2011-07-04 00:59:03 +03:00
|
|
|
int mm_support;
|
2011-07-04 02:53:39 +03:00
|
|
|
int nonspec;
|
2007-03-31 20:24:09 +03:00
|
|
|
};
|
|
|
|
|
2008-06-28 13:42:31 +03:00
|
|
|
static int cpu_flags;
|
|
|
|
|
2011-07-04 01:03:11 +03:00
|
|
|
static const struct algo fdct_tab[] = {
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "REF-DBL", ff_ref_fdct, NO_PERM },
|
2012-03-27 17:42:23 +03:00
|
|
|
{ "FAAN", ff_faandct, NO_PERM },
|
2012-02-15 13:10:21 +03:00
|
|
|
{ "IJG-AAN-INT", ff_fdct_ifast, SCALE_PERM },
|
2011-07-20 22:01:56 +03:00
|
|
|
{ "IJG-LLM-INT", ff_jpeg_fdct_islow_8, NO_PERM },
|
2007-03-31 20:24:09 +03:00
|
|
|
|
2012-08-01 06:30:29 +03:00
|
|
|
#if HAVE_MMX && HAVE_INLINE_ASM
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "MMX", ff_fdct_mmx, NO_PERM, AV_CPU_FLAG_MMX },
|
2012-07-08 19:42:12 +03:00
|
|
|
{ "MMXEXT", ff_fdct_mmx2, NO_PERM, AV_CPU_FLAG_MMXEXT },
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "SSE2", ff_fdct_sse2, NO_PERM, AV_CPU_FLAG_SSE2 },
|
2007-06-07 18:59:15 +03:00
|
|
|
#endif
|
2007-03-31 20:24:09 +03:00
|
|
|
|
2011-07-04 01:03:11 +03:00
|
|
|
#if HAVE_ALTIVEC
|
2012-02-16 08:15:40 +03:00
|
|
|
{ "altivecfdct", ff_fdct_altivec, NO_PERM, AV_CPU_FLAG_ALTIVEC },
|
2007-06-07 18:54:49 +03:00
|
|
|
#endif
|
2011-07-04 01:03:11 +03:00
|
|
|
|
|
|
|
#if ARCH_BFIN
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "BFINfdct", ff_bfin_fdct, NO_PERM },
|
2007-03-31 20:24:09 +03:00
|
|
|
#endif
|
|
|
|
|
2011-07-04 01:03:11 +03:00
|
|
|
{ 0 }
|
|
|
|
};
|
|
|
|
|
2012-06-10 04:15:38 +03:00
|
|
|
#if ARCH_X86_64 && HAVE_MMX && HAVE_YASM
|
2011-10-12 23:02:54 +03:00
|
|
|
void ff_prores_idct_put_10_sse2(uint16_t *dst, int linesize,
|
|
|
|
DCTELEM *block, int16_t *qmat);
|
|
|
|
|
2012-04-02 22:26:08 +03:00
|
|
|
static void ff_prores_idct_put_10_sse2_wrap(DCTELEM *dst){
|
2012-06-10 21:45:55 +03:00
|
|
|
DECLARE_ALIGNED(16, static int16_t, qmat)[64];
|
|
|
|
DECLARE_ALIGNED(16, static int16_t, tmp)[64];
|
|
|
|
int i;
|
2011-10-12 23:02:54 +03:00
|
|
|
|
|
|
|
for(i=0; i<64; i++){
|
|
|
|
qmat[i]=4;
|
|
|
|
tmp[i]= dst[i];
|
|
|
|
}
|
|
|
|
ff_prores_idct_put_10_sse2(dst, 16, tmp, qmat);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2011-07-04 01:03:11 +03:00
|
|
|
static const struct algo idct_tab[] = {
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "FAANI", ff_faanidct, NO_PERM },
|
|
|
|
{ "REF-DBL", ff_ref_idct, NO_PERM },
|
2012-02-15 13:29:24 +03:00
|
|
|
{ "INT", ff_j_rev_dct, MMX_PERM },
|
2011-07-20 18:05:05 +03:00
|
|
|
{ "SIMPLE-C", ff_simple_idct_8, NO_PERM },
|
2011-07-04 01:03:11 +03:00
|
|
|
|
2012-07-23 00:14:20 +03:00
|
|
|
#if HAVE_MMX && HAVE_INLINE_ASM
|
2009-01-14 01:44:16 +02:00
|
|
|
#if CONFIG_GPL
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "LIBMPEG2-MMX", ff_mmx_idct, MMX_PERM, AV_CPU_FLAG_MMX, 1 },
|
|
|
|
{ "LIBMPEG2-MMX2", ff_mmxext_idct, MMX_PERM, AV_CPU_FLAG_MMX2, 1 },
|
2007-06-07 18:54:49 +03:00
|
|
|
#endif
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "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 },
|
2012-07-08 19:42:12 +03:00
|
|
|
{ "XVID-MMXEXT", ff_idct_xvid_mmx2, NO_PERM, AV_CPU_FLAG_MMXEXT, 1 },
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "XVID-SSE2", ff_idct_xvid_sse2, SSE2_PERM, AV_CPU_FLAG_SSE2, 1 },
|
2012-03-31 01:03:44 +03:00
|
|
|
#if ARCH_X86_64 && HAVE_YASM
|
2011-10-12 23:02:54 +03:00
|
|
|
{ "PR-SSE2", ff_prores_idct_put_10_sse2_wrap, TRANSPOSE_PERM, AV_CPU_FLAG_SSE2, 1 },
|
2007-03-31 20:24:09 +03:00
|
|
|
#endif
|
2011-11-01 17:47:15 +03:00
|
|
|
#endif
|
2007-03-31 20:24:09 +03:00
|
|
|
|
2009-01-14 01:44:16 +02:00
|
|
|
#if ARCH_BFIN
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "BFINidct", ff_bfin_idct, NO_PERM },
|
2007-03-31 20:24:09 +03:00
|
|
|
#endif
|
|
|
|
|
2009-01-14 01:44:16 +02:00
|
|
|
#if ARCH_ARM
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "SIMPLE-ARM", ff_simple_idct_arm, NO_PERM },
|
|
|
|
{ "INT-ARM", ff_j_rev_dct_arm, MMX_PERM },
|
2011-07-04 01:03:11 +03:00
|
|
|
#endif
|
2009-01-14 01:44:16 +02:00
|
|
|
#if HAVE_ARMV5TE
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "SIMPLE-ARMV5TE", ff_simple_idct_armv5te,NO_PERM },
|
2008-12-16 22:00:07 +02:00
|
|
|
#endif
|
2009-01-14 01:44:16 +02:00
|
|
|
#if HAVE_ARMV6
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "SIMPLE-ARMV6", ff_simple_idct_armv6, MMX_PERM },
|
2008-12-16 22:00:07 +02:00
|
|
|
#endif
|
2009-01-14 01:44:16 +02:00
|
|
|
#if HAVE_NEON
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "SIMPLE-NEON", ff_simple_idct_neon, PARTTRANS_PERM },
|
2008-12-16 22:00:07 +02:00
|
|
|
#endif
|
|
|
|
|
2009-01-18 07:52:20 +02:00
|
|
|
#if ARCH_ALPHA
|
2011-07-04 13:54:22 +03:00
|
|
|
{ "SIMPLE-ALPHA", ff_simple_idct_axp, NO_PERM },
|
2009-01-18 07:52:20 +02:00
|
|
|
#endif
|
|
|
|
|
2011-07-04 00:59:03 +03:00
|
|
|
{ 0 }
|
2007-03-31 20:24:09 +03:00
|
|
|
};
|
|
|
|
|
2001-07-22 17:18:56 +03:00
|
|
|
#define AANSCALE_BITS 12
|
|
|
|
|
|
|
|
#define NB_ITS 20000
|
|
|
|
#define NB_ITS_SPEED 50000
|
|
|
|
|
2001-08-08 01:42:45 +03:00
|
|
|
static short idct_mmx_perm[64];
|
|
|
|
|
2011-07-04 00:59:03 +03:00
|
|
|
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,
|
2002-09-01 12:33:05 +03:00
|
|
|
};
|
|
|
|
|
2011-07-04 00:59:03 +03:00
|
|
|
static const uint8_t idct_sse2_row_perm[8] = { 0, 4, 1, 5, 2, 6, 3, 7 };
|
2008-04-16 04:54:16 +03:00
|
|
|
|
2009-03-31 12:32:59 +03:00
|
|
|
static void idct_mmx_init(void)
|
2001-08-08 01:42:45 +03:00
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
|
|
|
/* the mmx/mmxext idct uses a reordered input, so we patch scan tables */
|
|
|
|
for (i = 0; i < 64; i++) {
|
2005-12-22 03:10:11 +02:00
|
|
|
idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
|
2001-08-08 01:42:45 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2010-01-22 05:25:11 +02:00
|
|
|
DECLARE_ALIGNED(16, static DCTELEM, block)[64];
|
2011-07-04 00:59:03 +03:00
|
|
|
DECLARE_ALIGNED(8, static DCTELEM, block1)[64];
|
2001-08-08 01:42:45 +03:00
|
|
|
|
2011-09-21 22:25:43 +03:00
|
|
|
static void init_block(DCTELEM block[64], int test, int is_idct, AVLFG *prng, int vals)
|
2011-09-21 17:23:34 +03:00
|
|
|
{
|
|
|
|
int i, j;
|
|
|
|
|
|
|
|
memset(block, 0, 64 * sizeof(*block));
|
|
|
|
|
|
|
|
switch (test) {
|
|
|
|
case 0:
|
|
|
|
for (i = 0; i < 64; i++)
|
2011-09-21 22:25:43 +03:00
|
|
|
block[i] = (av_lfg_get(prng) % (2*vals)) -vals;
|
2011-09-21 17:23:34 +03:00
|
|
|
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++)
|
2011-09-21 22:25:43 +03:00
|
|
|
block[av_lfg_get(prng) % 64] = av_lfg_get(prng) % (2*vals) -vals;
|
2011-09-21 17:23:34 +03:00
|
|
|
break;
|
|
|
|
case 2:
|
2011-09-21 22:25:43 +03:00
|
|
|
block[ 0] = av_lfg_get(prng) % (16*vals) - (8*vals);
|
2011-09-21 17:23:34 +03:00
|
|
|
block[63] = (block[0] & 1) ^ 1;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void permute(DCTELEM dst[64], const DCTELEM src[64], int perm)
|
|
|
|
{
|
|
|
|
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) {
|
|
|
|
for (i = 0; i < 64; i++)
|
|
|
|
dst[(i & 0x38) | idct_sse2_row_perm[i & 7]] = src[i];
|
|
|
|
} else if (perm == PARTTRANS_PERM) {
|
|
|
|
for (i = 0; i < 64; i++)
|
|
|
|
dst[(i & 0x24) | ((i & 3) << 3) | ((i >> 3) & 3)] = src[i];
|
2011-10-12 23:01:44 +03:00
|
|
|
} else if (perm == TRANSPOSE_PERM) {
|
|
|
|
for (i = 0; i < 64; i++)
|
|
|
|
dst[(i>>3) | ((i<<3)&0x38)] = src[i];
|
2011-09-21 17:23:34 +03:00
|
|
|
} else {
|
|
|
|
for (i = 0; i < 64; i++)
|
|
|
|
dst[i] = src[i];
|
|
|
|
}
|
|
|
|
}
|
2011-07-05 02:46:03 +03:00
|
|
|
|
|
|
|
static int dct_error(const struct algo *dct, int test, int is_idct, int speed, const int bits)
|
2001-07-22 17:18:56 +03:00
|
|
|
{
|
2011-07-04 13:54:22 +03:00
|
|
|
void (*ref)(DCTELEM *block) = is_idct ? ff_ref_idct : ff_ref_fdct;
|
2001-07-22 17:18:56 +03:00
|
|
|
int it, i, scale;
|
|
|
|
int err_inf, v;
|
2011-07-04 02:53:39 +03:00
|
|
|
int64_t err2, ti, ti1, it1, err_sum = 0;
|
2011-07-04 00:59:03 +03:00
|
|
|
int64_t sysErr[64], sysErrMax = 0;
|
|
|
|
int maxout = 0;
|
|
|
|
int blockSumErrMax = 0, blockSumErr;
|
2009-04-10 20:23:38 +03:00
|
|
|
AVLFG prng;
|
2011-05-01 18:50:34 +03:00
|
|
|
const int vals=1<<bits;
|
2011-07-04 02:53:39 +03:00
|
|
|
double omse, ome;
|
|
|
|
int spec_err;
|
2001-07-22 17:18:56 +03:00
|
|
|
|
2009-04-10 20:23:38 +03:00
|
|
|
av_lfg_init(&prng, 1);
|
2001-07-22 17:18:56 +03:00
|
|
|
|
|
|
|
err_inf = 0;
|
|
|
|
err2 = 0;
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 64; i++)
|
|
|
|
sysErr[i] = 0;
|
|
|
|
for (it = 0; it < NB_ITS; it++) {
|
2011-09-21 22:25:43 +03:00
|
|
|
init_block(block1, test, is_idct, &prng, vals);
|
2011-09-21 17:23:34 +03:00
|
|
|
permute(block, block1, dct->format);
|
2001-08-08 01:42:45 +03:00
|
|
|
|
2011-07-04 00:48:53 +03:00
|
|
|
dct->func(block);
|
2012-06-20 15:50:45 +03:00
|
|
|
emms_c();
|
2001-08-08 01:42:45 +03:00
|
|
|
|
2011-07-04 00:48:53 +03:00
|
|
|
if (dct->format == SCALE_PERM) {
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 64; i++) {
|
|
|
|
scale = 8 * (1 << (AANSCALE_BITS + 11)) / ff_aanscales[i];
|
|
|
|
block[i] = (block[i] * scale) >> AANSCALE_BITS;
|
2002-09-01 12:33:05 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-07-04 13:54:22 +03:00
|
|
|
ref(block1);
|
2001-07-22 17:18:56 +03:00
|
|
|
|
2011-07-04 00:59:03 +03:00
|
|
|
blockSumErr = 0;
|
|
|
|
for (i = 0; i < 64; i++) {
|
2011-07-04 02:53:39 +03:00
|
|
|
int err = block[i] - block1[i];
|
|
|
|
err_sum += err;
|
|
|
|
v = abs(err);
|
2001-07-22 17:18:56 +03:00
|
|
|
if (v > err_inf)
|
|
|
|
err_inf = v;
|
|
|
|
err2 += v * v;
|
2005-12-22 03:10:11 +02:00
|
|
|
sysErr[i] += block[i] - block1[i];
|
|
|
|
blockSumErr += v;
|
2011-07-04 00:59:03 +03:00
|
|
|
if (abs(block[i]) > maxout)
|
|
|
|
maxout = abs(block[i]);
|
2001-07-22 17:18:56 +03:00
|
|
|
}
|
2011-07-04 00:59:03 +03:00
|
|
|
if (blockSumErrMax < blockSumErr)
|
|
|
|
blockSumErrMax = blockSumErr;
|
2002-09-01 12:33:05 +03:00
|
|
|
}
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 64; i++)
|
|
|
|
sysErrMax = FFMAX(sysErrMax, FFABS(sysErr[i]));
|
2005-12-17 20:14:38 +02:00
|
|
|
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 64; i++) {
|
|
|
|
if (i % 8 == 0)
|
|
|
|
printf("\n");
|
|
|
|
printf("%7d ", (int) sysErr[i]);
|
2001-07-22 17:18:56 +03:00
|
|
|
}
|
2002-09-01 12:33:05 +03:00
|
|
|
printf("\n");
|
2005-12-17 20:14:38 +02:00
|
|
|
|
2011-07-04 02:53:39 +03:00
|
|
|
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);
|
|
|
|
|
2011-10-08 20:57:30 +03:00
|
|
|
printf("%s %s: max_err=%d omse=%0.8f ome=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
|
2011-07-04 01:05:21 +03:00
|
|
|
is_idct ? "IDCT" : "DCT", dct->name, err_inf,
|
2011-07-04 02:53:39 +03:00
|
|
|
omse, ome, (double) sysErrMax / NB_ITS,
|
2011-07-04 00:59:03 +03:00
|
|
|
maxout, blockSumErrMax);
|
2011-04-26 14:52:12 +03:00
|
|
|
|
2011-07-04 02:53:39 +03:00
|
|
|
if (spec_err && !dct->nonspec)
|
|
|
|
return 1;
|
|
|
|
|
2011-07-04 01:19:13 +03:00
|
|
|
if (!speed)
|
2011-07-04 02:53:39 +03:00
|
|
|
return 0;
|
2011-04-26 14:52:12 +03:00
|
|
|
|
2001-07-22 17:18:56 +03:00
|
|
|
/* speed test */
|
2011-07-04 00:59:03 +03:00
|
|
|
|
2011-09-21 22:25:43 +03:00
|
|
|
init_block(block, test, is_idct, &prng, vals);
|
2011-09-21 17:23:34 +03:00
|
|
|
permute(block1, block, dct->format);
|
2001-08-08 01:42:45 +03:00
|
|
|
|
2012-06-20 03:34:38 +03:00
|
|
|
ti = av_gettime();
|
2001-07-22 17:18:56 +03:00
|
|
|
it1 = 0;
|
|
|
|
do {
|
2011-07-04 00:59:03 +03:00
|
|
|
for (it = 0; it < NB_ITS_SPEED; it++) {
|
2011-09-21 17:23:34 +03:00
|
|
|
memcpy(block, block1, sizeof(block));
|
2011-07-04 00:48:53 +03:00
|
|
|
dct->func(block);
|
2001-07-22 17:18:56 +03:00
|
|
|
}
|
2012-06-20 21:47:46 +03:00
|
|
|
emms_c();
|
2001-07-22 17:18:56 +03:00
|
|
|
it1 += NB_ITS_SPEED;
|
2012-06-20 03:34:38 +03:00
|
|
|
ti1 = av_gettime() - ti;
|
2001-07-22 17:18:56 +03:00
|
|
|
} while (ti1 < 1000000);
|
|
|
|
|
2011-07-04 01:05:21 +03:00
|
|
|
printf("%s %s: %0.1f kdct/s\n", is_idct ? "IDCT" : "DCT", dct->name,
|
2011-07-04 00:59:03 +03:00
|
|
|
(double) it1 * 1000.0 / (double) ti1);
|
2011-07-04 02:53:39 +03:00
|
|
|
|
|
|
|
return 0;
|
2001-07-22 17:18:56 +03:00
|
|
|
}
|
|
|
|
|
2010-01-22 05:25:11 +02:00
|
|
|
DECLARE_ALIGNED(8, static uint8_t, img_dest)[64];
|
|
|
|
DECLARE_ALIGNED(8, static uint8_t, img_dest1)[64];
|
2002-10-03 22:49:23 +03:00
|
|
|
|
2009-03-31 12:32:59 +03:00
|
|
|
static void idct248_ref(uint8_t *dest, int linesize, int16_t *block)
|
2002-10-03 22:49:23 +03:00
|
|
|
{
|
|
|
|
static int init;
|
|
|
|
static double c8[8][8];
|
|
|
|
static double c4[4][4];
|
|
|
|
double block1[64], block2[64], block3[64];
|
|
|
|
double s, sum, v;
|
|
|
|
int i, j, k;
|
|
|
|
|
|
|
|
if (!init) {
|
|
|
|
init = 1;
|
|
|
|
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 8; i++) {
|
2002-10-03 22:49:23 +03:00
|
|
|
sum = 0;
|
2011-07-04 00:59:03 +03:00
|
|
|
for (j = 0; j < 8; j++) {
|
|
|
|
s = (i == 0) ? sqrt(1.0 / 8.0) : sqrt(1.0 / 4.0);
|
2002-10-03 22:49:23 +03:00
|
|
|
c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0);
|
|
|
|
sum += c8[i][j] * c8[i][j];
|
|
|
|
}
|
|
|
|
}
|
2005-12-17 20:14:38 +02:00
|
|
|
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 4; i++) {
|
2002-10-03 22:49:23 +03:00
|
|
|
sum = 0;
|
2011-07-04 00:59:03 +03:00
|
|
|
for (j = 0; j < 4; j++) {
|
|
|
|
s = (i == 0) ? sqrt(1.0 / 4.0) : sqrt(1.0 / 2.0);
|
2002-10-03 22:49:23 +03:00
|
|
|
c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0);
|
|
|
|
sum += c4[i][j] * c4[i][j];
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* butterfly */
|
2002-10-04 00:03:59 +03:00
|
|
|
s = 0.5 * sqrt(2.0);
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 4; i++) {
|
|
|
|
for (j = 0; j < 8; j++) {
|
|
|
|
block1[8 * (2 * i) + j] =
|
|
|
|
(block[8 * (2 * i) + j] + block[8 * (2 * i + 1) + j]) * s;
|
|
|
|
block1[8 * (2 * i + 1) + j] =
|
|
|
|
(block[8 * (2 * i) + j] - block[8 * (2 * i + 1) + j]) * s;
|
2002-10-03 22:49:23 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* idct8 on lines */
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 8; i++) {
|
|
|
|
for (j = 0; j < 8; j++) {
|
2002-10-03 22:49:23 +03:00
|
|
|
sum = 0;
|
2011-07-04 00:59:03 +03:00
|
|
|
for (k = 0; k < 8; k++)
|
|
|
|
sum += c8[k][j] * block1[8 * i + k];
|
|
|
|
block2[8 * i + j] = sum;
|
2002-10-03 22:49:23 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* idct4 */
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 8; i++) {
|
|
|
|
for (j = 0; j < 4; j++) {
|
2002-10-03 22:49:23 +03:00
|
|
|
/* top */
|
|
|
|
sum = 0;
|
2011-07-04 00:59:03 +03:00
|
|
|
for (k = 0; k < 4; k++)
|
|
|
|
sum += c4[k][j] * block2[8 * (2 * k) + i];
|
|
|
|
block3[8 * (2 * j) + i] = sum;
|
2002-10-03 22:49:23 +03:00
|
|
|
|
|
|
|
/* bottom */
|
|
|
|
sum = 0;
|
2011-07-04 00:59:03 +03:00
|
|
|
for (k = 0; k < 4; k++)
|
|
|
|
sum += c4[k][j] * block2[8 * (2 * k + 1) + i];
|
|
|
|
block3[8 * (2 * j + 1) + i] = sum;
|
2002-10-03 22:49:23 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* clamp and store the result */
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 8; i++) {
|
|
|
|
for (j = 0; j < 8; j++) {
|
|
|
|
v = block3[8 * i + j];
|
|
|
|
if (v < 0) v = 0;
|
|
|
|
else if (v > 255) v = 255;
|
|
|
|
dest[i * linesize + j] = (int) rint(v);
|
2002-10-03 22:49:23 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2009-03-31 12:32:59 +03:00
|
|
|
static void idct248_error(const char *name,
|
2011-07-04 00:59:03 +03:00
|
|
|
void (*idct248_put)(uint8_t *dest, int line_size,
|
2011-07-04 01:19:13 +03:00
|
|
|
int16_t *block),
|
|
|
|
int speed)
|
2002-10-03 22:49:23 +03:00
|
|
|
{
|
|
|
|
int it, i, it1, ti, ti1, err_max, v;
|
2009-04-10 20:23:38 +03:00
|
|
|
AVLFG prng;
|
2009-03-20 13:48:27 +02:00
|
|
|
|
2009-04-10 20:23:38 +03:00
|
|
|
av_lfg_init(&prng, 1);
|
2005-12-17 20:14:38 +02:00
|
|
|
|
2002-10-03 22:49:23 +03:00
|
|
|
/* just one test to see if code is correct (precision is less
|
|
|
|
important here) */
|
|
|
|
err_max = 0;
|
2011-07-04 00:59:03 +03:00
|
|
|
for (it = 0; it < NB_ITS; it++) {
|
2002-10-04 00:03:59 +03:00
|
|
|
/* XXX: use forward transform to generate values */
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 64; i++)
|
2009-04-10 20:23:38 +03:00
|
|
|
block1[i] = av_lfg_get(&prng) % 256 - 128;
|
2002-10-04 00:03:59 +03:00
|
|
|
block1[0] += 1024;
|
|
|
|
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 64; i++)
|
|
|
|
block[i] = block1[i];
|
2002-10-03 22:49:23 +03:00
|
|
|
idct248_ref(img_dest1, 8, block);
|
2005-12-17 20:14:38 +02:00
|
|
|
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 64; i++)
|
|
|
|
block[i] = block1[i];
|
2002-10-04 00:03:59 +03:00
|
|
|
idct248_put(img_dest, 8, block);
|
2005-12-17 20:14:38 +02:00
|
|
|
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; i < 64; i++) {
|
|
|
|
v = abs((int) img_dest[i] - (int) img_dest1[i]);
|
2002-10-04 00:03:59 +03:00
|
|
|
if (v == 255)
|
|
|
|
printf("%d %d\n", img_dest[i], img_dest1[i]);
|
|
|
|
if (v > err_max)
|
|
|
|
err_max = v;
|
|
|
|
}
|
2002-10-03 22:49:23 +03:00
|
|
|
#if 0
|
|
|
|
printf("ref=\n");
|
|
|
|
for(i=0;i<8;i++) {
|
|
|
|
int j;
|
|
|
|
for(j=0;j<8;j++) {
|
|
|
|
printf(" %3d", img_dest1[i*8+j]);
|
|
|
|
}
|
|
|
|
printf("\n");
|
|
|
|
}
|
2005-12-17 20:14:38 +02:00
|
|
|
|
2002-10-03 22:49:23 +03:00
|
|
|
printf("out=\n");
|
|
|
|
for(i=0;i<8;i++) {
|
|
|
|
int j;
|
|
|
|
for(j=0;j<8;j++) {
|
|
|
|
printf(" %3d", img_dest[i*8+j]);
|
|
|
|
}
|
|
|
|
printf("\n");
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
}
|
2011-07-04 00:59:03 +03:00
|
|
|
printf("%s %s: err_inf=%d\n", 1 ? "IDCT248" : "DCT248", name, err_max);
|
2002-10-03 22:49:23 +03:00
|
|
|
|
2011-07-04 01:19:13 +03:00
|
|
|
if (!speed)
|
|
|
|
return;
|
2002-10-03 22:49:23 +03:00
|
|
|
|
2012-06-20 03:34:38 +03:00
|
|
|
ti = av_gettime();
|
2002-10-03 22:49:23 +03:00
|
|
|
it1 = 0;
|
|
|
|
do {
|
2011-07-04 00:59:03 +03:00
|
|
|
for (it = 0; it < NB_ITS_SPEED; it++) {
|
|
|
|
for (i = 0; i < 64; i++)
|
|
|
|
block[i] = block1[i];
|
2002-10-03 22:49:23 +03:00
|
|
|
idct248_put(img_dest, 8, block);
|
|
|
|
}
|
2012-06-20 21:47:46 +03:00
|
|
|
emms_c();
|
2002-10-03 22:49:23 +03:00
|
|
|
it1 += NB_ITS_SPEED;
|
2012-06-20 03:34:38 +03:00
|
|
|
ti1 = av_gettime() - ti;
|
2002-10-03 22:49:23 +03:00
|
|
|
} while (ti1 < 1000000);
|
|
|
|
|
2011-07-04 00:59:03 +03:00
|
|
|
printf("%s %s: %0.1f kdct/s\n", 1 ? "IDCT248" : "DCT248", name,
|
|
|
|
(double) it1 * 1000.0 / (double) ti1);
|
2002-10-03 22:49:23 +03:00
|
|
|
}
|
|
|
|
|
2009-03-31 12:32:59 +03:00
|
|
|
static void help(void)
|
2001-08-08 01:42:45 +03:00
|
|
|
{
|
2011-10-08 20:58:27 +03:00
|
|
|
printf("dct-test [-i] [<test-number>] [<bits>]\n"
|
2002-09-01 12:33:05 +03:00
|
|
|
"test-number 0 -> test with random matrixes\n"
|
|
|
|
" 1 -> test with random sparse matrixes\n"
|
|
|
|
" 2 -> do 3. test from mpeg4 std\n"
|
2011-10-08 20:58:27 +03:00
|
|
|
"bits Number of time domain bits to use, 8 is default\n"
|
2002-10-03 22:49:23 +03:00
|
|
|
"-i test IDCT implementations\n"
|
2011-07-04 01:19:13 +03:00
|
|
|
"-4 test IDCT248 implementations\n"
|
|
|
|
"-t speed test\n");
|
2001-08-08 01:42:45 +03:00
|
|
|
}
|
|
|
|
|
2012-07-04 06:14:09 +03:00
|
|
|
#if !HAVE_GETOPT
|
|
|
|
#include "compat/getopt.c"
|
|
|
|
#endif
|
|
|
|
|
2001-07-22 17:18:56 +03:00
|
|
|
int main(int argc, char **argv)
|
|
|
|
{
|
2002-10-03 22:49:23 +03:00
|
|
|
int test_idct = 0, test_248_dct = 0;
|
2011-07-04 00:59:03 +03:00
|
|
|
int c, i;
|
|
|
|
int test = 1;
|
2011-07-04 01:19:13 +03:00
|
|
|
int speed = 0;
|
2011-07-04 02:53:39 +03:00
|
|
|
int err = 0;
|
2011-05-01 18:50:34 +03:00
|
|
|
int bits=8;
|
2011-07-04 00:59:03 +03:00
|
|
|
|
2010-09-08 18:07:14 +03:00
|
|
|
cpu_flags = av_get_cpu_flags();
|
2001-08-08 01:42:45 +03:00
|
|
|
|
2009-03-31 11:55:22 +03:00
|
|
|
ff_ref_dct_init();
|
2001-08-08 01:42:45 +03:00
|
|
|
idct_mmx_init();
|
2008-06-25 02:51:17 +03:00
|
|
|
|
2011-07-04 00:59:03 +03:00
|
|
|
for (;;) {
|
2011-07-04 01:19:13 +03:00
|
|
|
c = getopt(argc, argv, "ih4t");
|
2001-08-08 01:42:45 +03:00
|
|
|
if (c == -1)
|
|
|
|
break;
|
2011-07-04 00:59:03 +03:00
|
|
|
switch (c) {
|
2001-08-08 01:42:45 +03:00
|
|
|
case 'i':
|
|
|
|
test_idct = 1;
|
|
|
|
break;
|
2002-10-03 22:49:23 +03:00
|
|
|
case '4':
|
|
|
|
test_248_dct = 1;
|
|
|
|
break;
|
2011-07-04 01:19:13 +03:00
|
|
|
case 't':
|
|
|
|
speed = 1;
|
|
|
|
break;
|
2011-07-04 00:59:03 +03:00
|
|
|
default:
|
2001-08-08 01:42:45 +03:00
|
|
|
case 'h':
|
|
|
|
help();
|
2007-03-31 20:14:59 +03:00
|
|
|
return 0;
|
2001-08-08 01:42:45 +03:00
|
|
|
}
|
|
|
|
}
|
2005-12-17 20:14:38 +02:00
|
|
|
|
2011-07-04 00:59:03 +03:00
|
|
|
if (optind < argc)
|
|
|
|
test = atoi(argv[optind]);
|
2011-05-01 18:50:34 +03:00
|
|
|
if(optind+1 < argc) bits= atoi(argv[optind+1]);
|
2005-12-17 20:14:38 +02:00
|
|
|
|
2001-08-08 01:42:45 +03:00
|
|
|
printf("ffmpeg DCT/IDCT test\n");
|
|
|
|
|
2002-10-03 22:49:23 +03:00
|
|
|
if (test_248_dct) {
|
2011-07-04 01:19:13 +03:00
|
|
|
idct248_error("SIMPLE-C", ff_simple_idct248_put, speed);
|
2001-08-08 01:42:45 +03:00
|
|
|
} else {
|
2011-07-04 01:03:11 +03:00
|
|
|
const struct algo *algos = test_idct ? idct_tab : fdct_tab;
|
2011-07-04 00:59:03 +03:00
|
|
|
for (i = 0; algos[i].name; i++)
|
2011-07-04 01:03:11 +03:00
|
|
|
if (!(~cpu_flags & algos[i].mm_support)) {
|
2011-07-05 02:46:03 +03:00
|
|
|
err |= dct_error(&algos[i], test, test_idct, speed, bits);
|
2011-07-04 00:59:03 +03:00
|
|
|
}
|
2001-08-08 01:42:45 +03:00
|
|
|
}
|
2011-07-04 02:53:39 +03:00
|
|
|
|
|
|
|
return err;
|
2001-07-22 17:18:56 +03:00
|
|
|
}
|