dct-test.c File Reference

#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <sys/time.h>
#include <unistd.h>
#include <math.h>
#include "libavutil/common.h"
#include "simple_idct.h"
#include "faandct.h"
#include "faanidct.h"
#include "i386/idct_xvid.h"

Go to the source code of this file.

Data Structures
struct	algo
Defines
#define	FAAN_SCALE   NO_PERM
#define	AANSCALE_BITS   12
#define	NB_ITS   20000
#define	NB_ITS_SPEED   50000
Functions
void *	fast_memcpy (void *a, const void *b, size_t c)
void	fdct (DCTELEM *block)
void	idct (DCTELEM *block)
void	init_fdct ()
void	ff_mmx_idct (DCTELEM *data)
void	ff_mmxext_idct (DCTELEM *data)
void	odivx_idct_c (short *block)
void	ff_bfin_idct (DCTELEM *block)
void	ff_bfin_fdct (DCTELEM *block)
void	fdct_altivec (DCTELEM *block)
int64_t	gettime (void)
void	idct_mmx_init (void)
static DCTELEM block[64]	__attribute__ ((aligned(16)))
static DCTELEM block1[64]	__attribute__ ((aligned(8)))
static void	mmx_emms (void)
void	dct_error (const char *name, int is_idct, void(*fdct_func)(DCTELEM *block), void(*fdct_ref)(DCTELEM *block), int form, int test)
void	idct248_ref (uint8_t *dest, int linesize, int16_t *block)
void	idct248_error (const char *name, void(*idct248_put)(uint8_t *dest, int line_size, int16_t *block))
void	help (void)
int	main (int argc, char **argv)
Variables
static int	cpu_flags
algo	algos []
static const unsigned short	aanscales [64]
uint8_t	cropTbl [256+2 *MAX_NEG_CROP]
static short	idct_mmx_perm [64]
static short	idct_simple_mmx_perm [64]
static const uint8_t	idct_sse2_row_perm [8] = {0, 4, 1, 5, 2, 6, 3, 7}

---

Detailed Description

DCT test. (c) 2001 Fabrice Bellard. Started from sample code by Juan J. Sierralta P.

Definition in file dct-test.c.

---

Define Documentation

#define AANSCALE_BITS   12

Definition at line 121 of file dct-test.c.

Referenced by dct_error().

#define FAAN_SCALE   NO_PERM

Definition at line 78 of file dct-test.c.

#define NB_ITS   20000

Definition at line 143 of file dct-test.c.

Referenced by dct_error(), idct248_error(), and test_motion().

#define NB_ITS_SPEED   50000

Definition at line 144 of file dct-test.c.

Referenced by dct_error(), and idct248_error().

---

Function Documentation

static DCTELEM block1 [64] __attribute__

(

(aligned(8))

)

[static]

Definition at line 44 of file yuv2rgb.c.

                                                              {
{  1,   3,   1,   3,   1,   3,   1,   3, },
{  2,   0,   2,   0,   2,   0,   2,   0, },
};

static DCTELEM block [64] __attribute__

(

(aligned(16))

)

[static]

void dct_error	(	const char *	name,
		int	is_idct,
		void(*)(DCTELEM *block)	fdct_func,
		void(*)(DCTELEM *block)	fdct_ref,
		int	form,
		int	test
	)

Definition at line 184 of file dct-test.c.

References AANSCALE_BITS, aanscales, fdct(), FFABS, FFMAX, gettime(), idct_mmx_perm, idct_simple_mmx_perm, idct_sse2_row_perm, mmx_emms(), NB_ITS, NB_ITS_SPEED, printf, and random.

Referenced by main().

{
    int it, i, scale;
    int err_inf, v;
    int64_t err2, ti, ti1, it1;
    int64_t sysErr[64], sysErrMax=0;
    int maxout=0;
    int blockSumErrMax=0, blockSumErr;

    srandom(0);

    err_inf = 0;
    err2 = 0;
    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] = (random() % 512) -256;
            if (is_idct){
                fdct(block1);

                for(i=0;i<64;i++)
                    block1[i]>>=3;
            }
        break;
        case 1:{
            int num= (random()%10)+1;
            for(i=0;i<num;i++)
                block1[random()%64] = (random() % 512) -256;
        }break;
        case 2:
            block1[0]= (random()%4096)-2048;
            block1[63]= (block1[0]&1)^1;
        break;
        }

#if 0 // simulate mismatch control
{ int sum=0;
        for(i=0;i<64;i++)
           sum+=block1[i];

        if((sum&1)==0) block1[63]^=1;
}
#endif

        for(i=0; i<64; i++)
            block_org[i]= block1[i];

        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 {
            for(i=0; i<64; i++)
                block[i]= block1[i];
        }
#if 0 // simulate mismatch control for tested IDCT but not the ref
{ int sum=0;
        for(i=0;i<64;i++)
           sum+=block[i];

        if((sum&1)==0) block[63]^=1;
}
#endif

        fdct_func(block);
        mmx_emms();

        if (form == SCALE_PERM) {
            for(i=0; i<64; i++) {
                scale = 8*(1 << (AANSCALE_BITS + 11)) / aanscales[i];
                block[i] = (block[i] * scale /*+ (1<<(AANSCALE_BITS-1))*/) >> AANSCALE_BITS;
            }
        }

        fdct_ref(block1);

        blockSumErr=0;
        for(i=0;i<64;i++) {
            v = abs(block[i] - block1[i]);
            if (v > err_inf)
                err_inf = v;
            err2 += v * v;
            sysErr[i] += block[i] - block1[i];
            blockSumErr += v;
            if( abs(block[i])>maxout) maxout=abs(block[i]);
        }
        if(blockSumErrMax < blockSumErr) blockSumErrMax= blockSumErr;
#if 0 // print different matrix pairs
        if(blockSumErr){
            printf("\n");
            for(i=0; i<64; i++){
                if((i&7)==0) printf("\n");
                printf("%4d ", block_org[i]);
            }
            for(i=0; i<64; i++){
                if((i&7)==0) printf("\n");
                printf("%4d ", block[i] - block1[i]);
            }
        }
#endif
    }
    for(i=0; i<64; i++) sysErrMax= FFMAX(sysErrMax, FFABS(sysErr[i]));

#if 1 // dump systematic errors
    for(i=0; i<64; i++){
        if(i%8==0) printf("\n");
        printf("%5d ", (int)sysErr[i]);
    }
    printf("\n");
#endif

    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, maxout, blockSumErrMax);
#if 1 //Speed test
    /* speed test */
    for(i=0;i<64;i++)
        block1[i] = 0;
    switch(test){
    case 0:
        for(i=0;i<64;i++)
            block1[i] = (random() % 512) -256;
        if (is_idct){
            fdct(block1);

            for(i=0;i<64;i++)
                block1[i]>>=3;
        }
    break;
    case 1:{
    case 2:
        block1[0] = (random() % 512) -256;
        block1[1] = (random() % 512) -256;
        block1[2] = (random() % 512) -256;
        block1[3] = (random() % 512) -256;
    }break;
    }

    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];
    }

    ti = gettime();
    it1 = 0;
    do {
        for(it=0;it<NB_ITS_SPEED;it++) {
            for(i=0; i<64; i++)
                block[i]= block1[i];
//            memcpy(block, block1, sizeof(DCTELEM) * 64);
// do not memcpy especially not fastmemcpy because it does movntq !!!
            fdct_func(block);
        }
        it1 += NB_ITS_SPEED;
        ti1 = gettime() - ti;
    } while (ti1 < 1000000);
    mmx_emms();

    printf("%s %s: %0.1f kdct/s\n",
           is_idct ? "IDCT" : "DCT",
           name, (double)it1 * 1000.0 / (double)ti1);
#endif
}

void* fast_memcpy	(	void *	a,
		const void *	b,
		size_t	c
	)

Definition at line 45 of file dct-test.c.

{return memcpy(a,b,c);};

void fdct

(

DCTELEM *

block

)

Definition at line 63 of file fdctref.c.

References c.

Referenced by dct_error().

{
        register int i, j;
        double s;
        double tmp[64];

        for(i = 0; i < 8; i++)
            for(j = 0; j < 8; j++)
            {
                    s = 0.0;

/*
 *                     for(k = 0; k < 8; k++)
 *                         s += c[j][k] * block[8 * i + k];
 */
                s += c[j][0] * block[8 * i + 0];
                s += c[j][1] * block[8 * i + 1];
                s += c[j][2] * block[8 * i + 2];
                s += c[j][3] * block[8 * i + 3];
                s += c[j][4] * block[8 * i + 4];
                s += c[j][5] * block[8 * i + 5];
                s += c[j][6] * block[8 * i + 6];
                s += c[j][7] * block[8 * i + 7];

                    tmp[8 * i + j] = s;
            }

        for(j = 0; j < 8; j++)
            for(i = 0; i < 8; i++)
            {
                    s = 0.0;

/*
 *                       for(k = 0; k < 8; k++)
 *                    s += c[i][k] * tmp[8 * k + j];
 */
                s += c[i][0] * tmp[8 * 0 + j];
                s += c[i][1] * tmp[8 * 1 + j];
                s += c[i][2] * tmp[8 * 2 + j];
                s += c[i][3] * tmp[8 * 3 + j];
                s += c[i][4] * tmp[8 * 4 + j];
                s += c[i][5] * tmp[8 * 5 + j];
                s += c[i][6] * tmp[8 * 6 + j];
                s += c[i][7] * tmp[8 * 7 + j];
                s*=8.0;

                    block[8 * i + j] = (short)floor(s + 0.499999);
/*
 * reason for adding 0.499999 instead of 0.5:
 * s is quite often x.5 (at least for i and/or j = 0 or 4)
 * and setting the rounding threshold exactly to 0.5 leads to an
 * extremely high arithmetic implementation dependency of the result;
 * s being between x.5 and x.500001 (which is now incorrectly rounded
 * downwards instead of upwards) is assumed to occur less often
 * (if at all)
 */
      }
}

void fdct_altivec

(

DCTELEM *

block

)

Referenced by dsputil_init_ppc().

void ff_bfin_fdct

(

DCTELEM *

block

)

void ff_bfin_idct

(

DCTELEM *

block

)

void ff_mmx_idct

(

DCTELEM *

data

)

Referenced by dsputil_init_mmx().

void ff_mmxext_idct

(

DCTELEM *

data

)

Referenced by dsputil_init_mmx().

int64_t gettime

(

void

)

Definition at line 136 of file dct-test.c.

References NULL.

Referenced by dct_error(), idct248_error(), main(), and test_motion().

{
    struct timeval tv;
    gettimeofday(&tv,NULL);
    return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec;
}

void help

(

void

)

Definition at line 525 of file dct-test.c.

References printf.

Referenced by main(), and printHelp().

{
    printf("dct-test [-i] [<test-number>]\n"
           "test-number 0 -> test with random matrixes\n"
           "            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");
}

void idct

(

DCTELEM *

block

)

Definition at line 125 of file fdctref.c.

References c.

{
  int i, j, k, v;
  double partial_product;
  double tmp[64];

  for (i=0; i<8; i++)
    for (j=0; j<8; j++)
    {
      partial_product = 0.0;

      for (k=0; k<8; k++)
        partial_product+= c[k][j]*block[8*i+k];

      tmp[8*i+j] = partial_product;
    }

  /* Transpose operation is integrated into address mapping by switching
     loop order of i and j */

  for (j=0; j<8; j++)
    for (i=0; i<8; i++)
    {
      partial_product = 0.0;

      for (k=0; k<8; k++)
        partial_product+= c[k][i]*tmp[8*k+j];

      v = (int) floor(partial_product+0.5);
      block[8*i+j] = v;
    }
}

void idct248_error	(	const char *	name,
		void(*)(uint8_t *dest, int line_size, int16_t *block)	idct248_put
	)

Definition at line 449 of file dct-test.c.

References gettime(), idct248_ref(), mmx_emms(), NB_ITS, NB_ITS_SPEED, printf, and random.

Referenced by main().

{
    int it, i, it1, ti, ti1, err_max, v;

    srandom(0);

    /* just one test to see if code is correct (precision is less
       important here) */
    err_max = 0;
    for(it=0;it<NB_ITS;it++) {

        /* XXX: use forward transform to generate values */
        for(i=0;i<64;i++)
            block1[i] = (random() % 256) - 128;
        block1[0] += 1024;

        for(i=0; i<64; i++)
            block[i]= block1[i];
        idct248_ref(img_dest1, 8, block);

        for(i=0; i<64; i++)
            block[i]= block1[i];
        idct248_put(img_dest, 8, block);

        for(i=0;i<64;i++) {
            v = abs((int)img_dest[i] - (int)img_dest1[i]);
            if (v == 255)
                printf("%d %d\n", img_dest[i], img_dest1[i]);
            if (v > err_max)
                err_max = v;
        }
#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");
        }

        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
    }
    printf("%s %s: err_inf=%d\n",
           1 ? "IDCT248" : "DCT248",
           name, err_max);

    ti = gettime();
    it1 = 0;
    do {
        for(it=0;it<NB_ITS_SPEED;it++) {
            for(i=0; i<64; i++)
                block[i]= block1[i];
//            memcpy(block, block1, sizeof(DCTELEM) * 64);
// do not memcpy especially not fastmemcpy because it does movntq !!!
            idct248_put(img_dest, 8, block);
        }
        it1 += NB_ITS_SPEED;
        ti1 = gettime() - ti;
    } while (ti1 < 1000000);
    mmx_emms();

    printf("%s %s: %0.1f kdct/s\n",
           1 ? "IDCT248" : "DCT248",
           name, (double)it1 * 1000.0 / (double)ti1);
}

void idct248_ref	(	uint8_t *	dest,
		int	linesize,
		int16_t *	block
	)

Definition at line 369 of file dct-test.c.

References c4, and M_PI.

Referenced by idct248_error().

{
    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;

        for(i=0;i<8;i++) {
            sum = 0;
            for(j=0;j<8;j++) {
                s = (i==0) ? sqrt(1.0/8.0) : sqrt(1.0/4.0);
                c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0);
                sum += c8[i][j] * c8[i][j];
            }
        }

        for(i=0;i<4;i++) {
            sum = 0;
            for(j=0;j<4;j++) {
                s = (i==0) ? sqrt(1.0/4.0) : sqrt(1.0/2.0);
                c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0);
                sum += c4[i][j] * c4[i][j];
            }
        }
    }

    /* butterfly */
    s = 0.5 * sqrt(2.0);
    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;
        }
    }

    /* idct8 on lines */
    for(i=0;i<8;i++) {
        for(j=0;j<8;j++) {
            sum = 0;
            for(k=0;k<8;k++)
                sum += c8[k][j] * block1[8*i+k];
            block2[8*i+j] = sum;
        }
    }

    /* idct4 */
    for(i=0;i<8;i++) {
        for(j=0;j<4;j++) {
            /* top */
            sum = 0;
            for(k=0;k<4;k++)
                sum += c4[k][j] * block2[8*(2*k)+i];
            block3[8*(2*j)+i] = sum;

            /* bottom */
            sum = 0;
            for(k=0;k<4;k++)
                sum += c4[k][j] * block2[8*(2*k+1)+i];
            block3[8*(2*j+1)+i] = sum;
        }
    }

    /* clamp and store the result */
    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);
        }
    }
}

void idct_mmx_init

(

void

)

Definition at line 161 of file dct-test.c.

References idct_mmx_perm.

Referenced by main().

{
    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);
//        idct_simple_mmx_perm[i] = simple_block_permute_op(i);
    }
}

void init_fdct

(

)

Definition at line 49 of file fdctref.c.

References c, and PI.

Referenced by main().

{
  int i, j;
  double s;

  for (i=0; i<8; i++)
  {
    s = (i==0) ? sqrt(0.125) : 0.5;

    for (j=0; j<8; j++)
      c[i][j] = s * cos((PI/8.0)*i*(j+0.5));
  }
}

int main	(	int	argc,
		char **	argv
	)

Definition at line 535 of file dct-test.c.

References algos, c, cpu_flags, cropTbl, dct_error(), ff_simple_idct248_put(), getopt(), help(), idct248_error(), idct_mmx_init(), init_fdct(), MAX_NEG_CROP, mm_support, algo::name, optind, printf, and test().

{
    int test_idct = 0, test_248_dct = 0;
    int c,i;
    int test=1;
    cpu_flags = mm_support();

    init_fdct();
    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");
        if (c == -1)
            break;
        switch(c) {
        case 'i':
            test_idct = 1;
            break;
        case '4':
            test_248_dct = 1;
            break;
        default :
        case 'h':
            help();
            return 0;
        }
    }

    if(optind <argc) test= atoi(argv[optind]);

    printf("ffmpeg DCT/IDCT test\n");

    if (test_248_dct) {
        idct248_error("SIMPLE-C", ff_simple_idct248_put);
    } 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);
        }
    }
    return 0;
}

static void mmx_emms

(

void

)

[inline, static]

Definition at line 176 of file dct-test.c.

References cpu_flags.

Referenced by dct_error(), and idct248_error().

{
#ifdef HAVE_MMX
    if (cpu_flags & MM_MMX)
        asm volatile ("emms\n\t");
#endif
}

void odivx_idct_c

(

short *

block

)

---

Variable Documentation

const unsigned short aanscales[64] [static]

Initial value:

 {
    
    16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
    22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
    21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
    19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
    16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
    12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
    8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
    4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
}

Definition at line 122 of file dct-test.c.

Referenced by dct_error().

struct algo algos[]

Definition at line 83 of file dct-test.c.

Referenced by main().

int cpu_flags [static]

Definition at line 81 of file dct-test.c.

Referenced by main(), and mmx_emms().

uint8_t cropTbl[256+2 *MAX_NEG_CROP]

Definition at line 134 of file dct-test.c.

Referenced by main().

short idct_mmx_perm[64] [static]

Definition at line 146 of file dct-test.c.

Referenced by dct_error(), and idct_mmx_init().

short idct_simple_mmx_perm[64] [static]

Initial value:

{
        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,
}

Definition at line 148 of file dct-test.c.

Referenced by dct_error().

const uint8_t idct_sse2_row_perm[8] = {0, 4, 1, 5, 2, 6, 3, 7} [static]

Definition at line 159 of file dct-test.c.

Referenced by dct_error(), and dsputil_init().

---