xvid_bench.c

mpeg4 video codec mpeg4 video codec

		TEST_MB(interpolate8x8_halfpel_v, 1);
		printf("%s -           round1 %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, iCrc,
			   (iCrc!=0x42202efe)?"| ERROR": "" );


		TEST_MB(interpolate8x8_halfpel_hv, 0);
		printf("%s - interp-hv-round0 %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, iCrc,
			   (iCrc!=0xd198d387)?"| ERROR": "" );

		TEST_MB(interpolate8x8_halfpel_hv, 1);
		printf("%s -           round1 %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, iCrc,
			   (iCrc!=0x9ecfd921)?"| ERROR": "" );


		/* this is a new function, as of 06.06.2002 */
#if 0
		TEST_MB2(interpolate8x8_avrg);
		printf("%s - interpolate8x8_c %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, iCrc,
			   (iCrc!=8107)?"| ERROR": "" );
#endif

    /* New functions for field prediction by CK 1.10.2005 */
#pragma NEW8X4
		TEST_MB(interpolate8x4_halfpel_h, 0);
		printf("%s - interpfield-h -round0 %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, iCrc,
			   (iCrc!=0x9538d6df)?"| ERROR": "" );

		TEST_MB(interpolate8x4_halfpel_h, 1);
		printf("%s -                round1 %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, iCrc,
			   (iCrc!=0xde5f1db4)?"| ERROR": "" );


		TEST_MB(interpolate8x4_halfpel_v, 0);
		printf("%s - interpfield- v-round0 %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, iCrc,
			   (iCrc!=0xea5a69ef)?"| ERROR": "" );

		TEST_MB(interpolate8x4_halfpel_v, 1);
		printf("%s -                round1 %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, iCrc,
			   (iCrc!=0x4f10ec0f)?"| ERROR": "" );


		TEST_MB(interpolate8x4_halfpel_hv, 0);
		printf("%s - interpfield-hv-round0 %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, iCrc,
			   (iCrc!=0xf97ee367)?"| ERROR": "" );

		TEST_MB(interpolate8x4_halfpel_hv, 1);
		printf("%s -                round1 %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, iCrc,
			   (iCrc!=0xb6a9f581)?"| ERROR": "" );
/* End of 8x4 functions */

		printf( " --- \n" );
	}
}

/*********************************************************************
 * test transfer
 *********************************************************************/

#define INIT_TRANSFER \
for(i=0; i<8*32; ++i) {             \
Src8[i] = i; Src16[i] = i;        \
Dst8[i] = 0; Dst16[i] = 0;        \
Ref1[i] = i^0x27;                 \
Ref2[i] = i^0x51;                 \
}

#define TEST_TRANSFER_BEGIN(DST)              \
INIT_TRANSFER                             \
overhead = -gettime_usec();               \
for(tst=0; tst<nb_tests; ++tst) {         \
  for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a;\
}                                         \
overhead += gettime_usec();               \
t = gettime_usec();                       \
emms();                                   \
for(tst=0; tst<nb_tests; ++tst) {         \
  for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a;


#define TEST_TRANSFER_END(DST)                \
}                                         \
emms();                                   \
t = (gettime_usec()-t -overhead) / nb_tests;\
byte_swap((uint8_t*)(DST), 8*32*sizeof((DST)[0]), sizeof((DST)[0]));  \
s = calc_crc((uint8_t*)(DST), 8*32*sizeof((DST)[0]), CRC32_INITIAL)

#define TEST_TRANSFER(FUNC, DST, SRC)         \
TEST_TRANSFER_BEGIN(DST);                 \
  (FUNC)((DST), (SRC), 32);               \
TEST_TRANSFER_END(DST)


#define TEST_TRANSFER2_BEGIN(DST, SRC)        \
INIT_TRANSFER                             \
overhead = -gettime_usec();               \
for(tst=0; tst<nb_tests; ++tst) {         \
  for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a;\
  for(i=0; i<8*32; ++i) (SRC)[i] = i^0x3e;\
}                                         \
overhead += gettime_usec();               \
t = gettime_usec();                       \
emms();                                   \
for(tst=0; tst<nb_tests; ++tst) {         \
  for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a;\
  for(i=0; i<8*32; ++i) (SRC)[i] = i^0x3e;

#define TEST_TRANSFER2_END(DST)               \
}                                         \
emms();                                   \
t = (gettime_usec()-t -overhead) / nb_tests;\
byte_swap((uint8_t*)(DST), 8*32*sizeof((DST)[0]), sizeof((DST)[0]));  \
s = calc_crc((uint8_t*)(DST), 8*32*sizeof((DST)[0]), CRC32_INITIAL)

#define TEST_TRANSFER2(FUNC, DST, SRC, R1)    \
TEST_TRANSFER2_BEGIN(DST,SRC);            \
  (FUNC)((DST), (SRC), (R1), 32);         \
TEST_TRANSFER2_END(DST)

#define TEST_TRANSFER3(FUNC, DST, SRC, R1, R2)\
TEST_TRANSFER_BEGIN(DST);                 \
  (FUNC)((DST), (SRC), (R1), (R2), 32);   \
TEST_TRANSFER_END(DST)

void test_transfer()
{
	const int nb_tests = 4000*speed_ref;
	int i;
	CPU *cpu;
//	uint8_t  Src8[8*32], Dst8[8*32], Ref1[8*32], Ref2[8*32];
//	int16_t Src16[8*32], Dst16[8*32];
  DECLARE_ALIGNED_MATRIX(Src8, 8, 32, uint8_t, CACHE_LINE);
  DECLARE_ALIGNED_MATRIX(Dst8, 8, 32, uint8_t, CACHE_LINE);
  DECLARE_ALIGNED_MATRIX(Ref1, 8, 32, uint8_t, CACHE_LINE);
  DECLARE_ALIGNED_MATRIX(Ref2, 8, 32, uint8_t, CACHE_LINE);
  DECLARE_ALIGNED_MATRIX(Src16, 8, 32, uint16_t, CACHE_LINE);
  DECLARE_ALIGNED_MATRIX(Dst16, 8, 32, uint16_t, CACHE_LINE);

	printf( "\n ===  test transfer ===\n" );

	for(cpu = cpu_list; cpu->name!=0; ++cpu)
	{
		double t, overhead;
		int tst, s;

		if (!init_cpu(cpu))
			continue;

		TEST_TRANSFER(transfer_8to16copy, Dst16, Src8);
		printf("%s - 8to16     %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, s,
			   (s!=0x115814bb)?"| ERROR": "");

		TEST_TRANSFER(transfer_16to8copy, Dst8, Src16);
		printf( "%s - 16to8     %.3f usec       crc32=0x%08x %s\n",
				cpu->name, t, s,
				(s!=0xee7ccbb4)?"| ERROR": "");

    /* New functions for field prediction by CK 1.10.2005 */
#pragma NEW8X4
		TEST_TRANSFER(transfer8x4_copy, Dst8, Src8);
		printf("%s - 8to4      %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, s,
			   (s!=0xbb9c3db5)?"| ERROR": "");
/* End of new functions */

		TEST_TRANSFER(transfer8x8_copy, Dst8, Src8);
		printf("%s - 8to8      %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, s,
			   (s!=0xd37b3295)?"| ERROR": "");

		TEST_TRANSFER(transfer_16to8add, Dst8, Src16);
		printf("%s - 16to8add  %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, s,
			   (s!=0xdd817bf4)?"| ERROR": "" );

		TEST_TRANSFER2(transfer_8to16sub, Dst16, Src8, Ref1);
		{
			int s1, s2;
			s1 = calc_crc((uint8_t*)Dst16, 8*32*sizeof(Dst16[0]), CRC32_INITIAL);
			s2 = calc_crc((uint8_t*)Src8, 8*32*sizeof(Src8[0]), CRC32_INITIAL);
			printf("%s - 8to16sub  %.3f usec       crc32(1)=0x%08x crc32(2)=0x%08x %s %s\n",
				   cpu->name, t, s1, s2,
				   (s1!=0xa1e07163)?"| ERROR1": "",
				   (s2!=0xd86c5d23)?"| ERROR2": "" );
		}

		TEST_TRANSFER3(transfer_8to16sub2, Dst16, Src8, Ref1, Ref2);
		printf("%s - 8to16sub2 %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, s,
			   (s!=0x99b6c4c7)?"| ERROR": "" );

		printf( " --- \n" );
	}
}

/*********************************************************************
 * test quantization
 *********************************************************************/

#define TEST_QUANT(FUNC, DST, SRC)              \
t = gettime_usec();                         \
for(s=CRC32_INITIAL,qm=1; qm<=255; ++qm) {              \
  for(i=0; i<8*8; ++i) Quant[i] = qm;       \
  set_inter_matrix( mpeg_quant_matrices, Quant );                \
  emms();                                   \
  for(q=1; q<=max_Q; ++q) {                 \
	for(tst=0; tst<nb_tests; ++tst)         \
	  (FUNC)((DST), (SRC), q, mpeg_quant_matrices);              \
	byte_swap((uint8_t*)(DST), 64*sizeof((DST)[0]), sizeof((DST)[0]));  \
	s = calc_crc((uint8_t*)(DST), 64*sizeof((DST)[0]), s); \
  }                                         \
  emms();                                   \
}                                           \
t = (gettime_usec()-t-overhead)/nb_tests/qm

#define TEST_QUANT2(FUNC, DST, SRC)             \
t = gettime_usec();                         \
for(s=CRC32_INITIAL,qm=1; qm<=255; ++qm) {              \
  for(i=0; i<8*8; ++i) Quant[i] = qm;       \
  set_intra_matrix( mpeg_quant_matrices, Quant );                \
  emms();                                   \
  for(q=1; q<=max_Q; ++q) {                 \
	for(tst=0; tst<nb_tests; ++tst)         \
	  (FUNC)((DST), (SRC), q, q, mpeg_quant_matrices);           \
	byte_swap((uint8_t*)(DST), 64*sizeof((DST)[0]), sizeof((DST)[0]));  \
	s = calc_crc((uint8_t*)(DST), 64*sizeof((DST)[0]), s); \
  }                                         \
  emms();                                   \
}                                           \
t = (gettime_usec()-t-overhead)/nb_tests/qm

#define TEST_INTRA(REFFUNC, NEWFUNC, RANGE)              \
{ int i,q,s;\
	DECLARE_ALIGNED_MATRIX(Src, 8, 8, int16_t, 16); \
  DECLARE_ALIGNED_MATRIX(Dst, 8, 8, int16_t, 16); \
  DECLARE_ALIGNED_MATRIX(Dst2,8, 8, int16_t, 16); \
  for(q=1;q<=max_Q;q++)          \
    for(s=-RANGE;s<RANGE;s++) { \
      for(i=0;i<64;i++) Src[i]=s; \
      (REFFUNC)((Dst),(Src),q,q,mpeg_quant_matrices);   \
      (NEWFUNC)((Dst2),(Src),q,q,mpeg_quant_matrices);  \
      for(i=0;i<64;i++)     \
        if(Dst[i]!=Dst2[i]) printf("ERROR : " #NEWFUNC " i%d quant:%d input:%d C_result:%d ASM_result:%d\n",i,q,s,Dst[i],Dst2[i]);  \
    }      \
}

#define TEST_INTER(REFFUNC, NEWFUNC, RANGE)              \
{ int i,q,s;  \
	DECLARE_ALIGNED_MATRIX(Src, 8, 8, int16_t, 16); \
  DECLARE_ALIGNED_MATRIX(Dst, 8, 8, int16_t, 16); \
  DECLARE_ALIGNED_MATRIX(Dst2,8, 8, int16_t, 16); \
  for(q=1;q<=max_Q;q++)  \
    for(s=-RANGE;s<RANGE;s++) {   \
      for(i=0;i<64;i++) Src[i]=s; \
      (REFFUNC)((Dst),(Src),q,mpeg_quant_matrices);  \
      (NEWFUNC)((Dst2),(Src),q,mpeg_quant_matrices); \
      emms();           \
      for(i=0;i<64;i++) \
        if(Dst[i]!=Dst2[i]) printf("ERROR : " #NEWFUNC " i%d quant:%d input:%d C_result:%d ASM_result:%d\n",i,q,s,Dst[i],Dst2[i]); \
    } \
}

void test_quant()
{
	const int nb_tests = 1*speed_ref;
	const int max_Q = 31;
	DECLARE_ALIGNED_MATRIX(mpeg_quant_matrices, 8, 64, uint16_t, 16);

	int i, qm;
	CPU *cpu;
	DECLARE_ALIGNED_MATRIX(Src, 8, 8, int16_t, 16);
	DECLARE_ALIGNED_MATRIX(Dst, 8, 8, int16_t, 16);
	DECLARE_ALIGNED_MATRIX(Dst2,8, 8, int16_t, 16);
	uint8_t Quant[8*8];

	printf( "\n =====  test quant =====\n" );

/* we deliberately enfringe the norm's specified range [-127,127], */
/* to test the robustness of the iquant module */
	for(i=0; i<64; ++i) {
		Src[i] = 1 + (i-32) * (i&6);
		Dst[i] = 0;
	}

	for(cpu = cpu_list; cpu->name!=0; ++cpu)
	{
		double t, overhead;
		int tst, q;
		uint32_t s;

		if (!init_cpu(cpu))
			continue;

		// exhaustive tests to compare against the (ref) C-version
		TEST_INTRA(quant_h263_intra_c,   quant_h263_intra,    2048);
		TEST_INTRA(dequant_h263_intra_c, dequant_h263_intra , 512 );
		TEST_INTER(quant_h263_inter_c,   quant_h263_inter ,   2048);
		TEST_INTER(dequant_h263_inter_c, dequant_h263_inter , 512 );

		overhead = -gettime_usec();
		for(s=0,qm=1; qm<=255; ++qm) {
			for(i=0; i<8*8; ++i) Quant[i] = qm;
			set_inter_matrix(mpeg_quant_matrices, Quant );
			for(q=1; q<=max_Q; ++q)
				for(i=0; i<64; ++i) s+=Dst[i]^i^qm;
		}
		overhead += gettime_usec();

		TEST_QUANT2(quant_mpeg_intra, Dst, Src);
		printf("%s -   quant_mpeg_intra %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, s,
			   (s!=0xfd6a21a4)? "| ERROR": "");

		TEST_QUANT(quant_mpeg_inter, Dst, Src);
		printf("%s -   quant_mpeg_inter %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, s,
			   (s!=0xf6de7757)?"| ERROR": "");

		TEST_QUANT2(dequant_mpeg_intra, Dst, Src);
		printf("%s - dequant_mpeg_intra %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, s,
			   (s!=0x2def7bc7)?"| ERROR": "");

		TEST_QUANT(dequant_mpeg_inter, Dst, Src);
		printf("%s - dequant_mpeg_inter %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, s,
			   (s!=0xd878c722)?"| ERROR": "");

		TEST_QUANT2(quant_h263_intra, Dst, Src);
		printf("%s -   quant_h263_intra %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, s,
			   (s!=0x2eba9d43)?"| ERROR": "");

		TEST_QUANT(quant_h263_inter, Dst, Src);
		printf("%s -   quant_h263_inter %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, s,
			   (s!=0xbd315a7e)?"| ERROR": "");

		TEST_QUANT2(dequant_h263_intra, Dst, Src);
		printf("%s - dequant_h263_intra %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, s,
			   (s!=0x9841212a)?"| ERROR": "");

		TEST_QUANT(dequant_h263_inter, Dst, Src);
		printf("%s - dequant_h263_inter %.3f usec       crc32=0x%08x %s\n",
			   cpu->name, t, s,
			   (s!=0xe7df8fba)?"| ERROR": "");

		printf( " --- \n" );
	}
}

/*********************************************************************
 * test distortion operators
 *********************************************************************/

static void ieee_reseed(long s);
static long ieee_rand(int Min, int Max);

#define TEST_SSE(FUNCTION, SRC1, SRC2, STRIDE) \
  do { \
    t = gettime_usec(); \
    tst = nb_tests; \
    while((tst--)>0) sse = (FUNCTION)((SRC1), (SRC2), (STRIDE)); \
    emms(); \
    t = (gettime_usec() - t)/(double)nb_tests;	\
  } while(0)


void test_sse()
{
	const int nb_tests = 100000*speed_ref;
	int i;
	CPU *cpu;
	DECLARE_ALIGNED_MATRIX(Src1, 8, 8, int16_t, 16);
	DECLARE_ALIGNED_MATRIX(Src2, 8, 8, int16_t, 16);
	DECLARE_ALIGNED_MATRIX(Src3, 8, 8, int16_t, 16);
	DECLARE_ALIGNED_MATRIX(Src4, 8, 8, int16_t, 16);

	printf( "\n =====  test sse =====\n" );

	ieee_reseed(1);
	for(i=0; i<64; ++i) {
		Src1[i] = ieee_rand(-2048, 2047);
		Src2[i] = ieee_rand(-2048, 2047);
		Src3[i] = ieee_rand(-2048, 2047);
		Src4[i] = ieee_rand(-2048, 2047);
	}

	for(cpu = cpu_list; cpu->name!=0; ++cpu)
	{
		double t;
		int tst, sse;

		if (!init_cpu(cpu))
			continue;

		/* 16 bit element blocks */
		TEST_SSE(sse8_16bit, Src1, Src2, 16);
		printf("%s -   sse8_16bit#1 %.3f usec       sse=%d %s\n",
			   cpu->name, t, sse, (sse!=182013834)?"| ERROR": "");
		TEST_SSE(sse8_16bit, Src1, Src3, 16);
		printf("%s -   sse8_16bit#2 %.3f usec       sse=%d %s\n",
			   cpu->name, t, sse, (sse!=142545203)?"| ERROR": "");
		TEST_SSE(sse8_16bit, Src1, Src4, 16);
		printf("%s -   sse8_16bit#3 %.3f usec       sse=%d %s\n",
			   cpu->name, t, sse, (sse!=146340935)?"| ERROR": "");
		TEST_SSE(sse8_16bit, Src2, Src3, 16);
		printf("%s -   sse8_16bit#4 %.3f usec       sse=%d %s\n",
			   cpu->name, t, sse, (sse!=130136661)?"| ERROR": "");
		TEST_SSE(sse8_16bit, Src2, Src4, 16);
		printf("%s -   sse8_16bit#5 %.3f usec       sse=%d %s\n",
			   cpu->name, t, sse, (sse!=136870353)?"| ERROR": "");
		TEST_SSE(sse8_16bit, Src3, Src4, 16);
		printf("%s -   sse8_16bit#6 %.3f usec       sse=%d %s\n",
			   cpu->name, t, sse, (sse!=164107772)?"| ERROR": "");

		/* 8 bit element blocks */
		TEST_SSE(sse8_8bit, (int8_t*)Src1, (int8_t*)Src2, 8);
		printf("%s -    sse8_8bit#1 %.3f usec       sse=%d %s\n",
			   cpu->name, t, sse, (sse!=1356423)?"| ERROR": "");
		TEST_SSE(sse8_8bit, (int8_t*)Src1, (int8_t*)Src3, 8);
		printf("%s -    sse8_8bit#2 %.3f usec       sse=%d %s\n",
			   cpu->name, t, sse, (sse!=1173074)?"| ERROR": "");
		TEST_SSE(sse8_8bit, (int8_t*)Src1, (int8_t*)Src4, 8);
		printf("%s -    sse8_8bit#3 %.3f usec       sse=%d %s\n",
			   cpu->name, t, sse, (sse!=1092357)?"| ERROR": "");
		TEST_SSE(sse8_8bit, (int8_t*)Src2, (int8_t*)Src3, 8);
		printf("%s -    sse8_8bit#4 %.3f usec       sse=%d %s\n",
			   cpu->name, t, sse, (sse!=1360239)?"| ERROR": "");
		TEST_SSE(sse8_8bit, (int8_t*)Src2, (int8_t*)Src4, 8);
		printf("%s -    sse8_8bit#5 %.3f usec       sse=%d %s\n",
			   cpu->name, t, sse, (sse!=1208414)?"| ERROR": "");
		TEST_SSE(sse8_8bit, (int8_t*)Src3, (int8_t*)Src4, 8);
		printf("%s -    sse8_8bit#6 %.3f usec       sse=%d %s\n",
			   cpu->name, t, sse, (sse!=1099285)?"| ERROR": "");

		printf(" ---\n");
	}
}

/*********************************************************************
 * test non-zero AC counting
 *********************************************************************/

#define TEST_CBP(FUNC, SRC, NB)           \
t = gettime_usec();                       \
emms();                                   \
for(tst=0; tst<NB; ++tst) {         \
  cbp = (FUNC)((SRC));                    \
}                                         \
emms();                                   \
t = (gettime_usec()-t ) / nb_tests;

void test_cbp()
{
	const int nb_tests = 10000*speed_ref;
	int i, n, m;
	CPU *cpu;
	DECLARE_ALIGNED_MATRIX(Src1, 6, 64, int16_t, 16);
	DECLARE_ALIGNED_MATRIX(Src2, 6, 64, int16_t, 16);
	DECLARE_ALIGNED_MATRIX(Src3, 6, 64, int16_t, 16);
	DECLARE_ALIGNED_MATRIX(Src4, 6, 64, int16_t, 16);
  DECLARE_ALIGNED_MATRIX(Src5, 6, 64, int16_t, 16);

	printf( "\n =====  test cbp =====\n" );

	for(i=0; i<6*64; ++i) {
		Src1[i] = (i*i*3/8192)&(i/64)&1;  /* 'random' */
		Src2[i] = (i<3*64);               /* half-full */
		Src3[i] = ((i+32)>3*64);
		Src4[i] = (i==(3*64+2) || i==(5*64+9));
    Src5[i] = ieee_rand(0,1) ? -1 : 1;  /* +/- test */
	}

	for(cpu = cpu_list; cpu->name!=0; ++cpu)
	{
		double t;
		int tst, cbp;

		if (!init_cpu(cpu))
			continue;

		TEST_CBP(calc_cbp, Src1, nb_tests);
		printf("%s -   calc_cbp#1 %.3f usec       cbp=0x%02x %s\n",