address C1_MacroAssembler::call_c_with_frame_resize(address dest, int frame_resize) {
if (frame_resize) { resize_frame(-frame_resize, R0); }
#if defined(ABI_ELFv2)
address return_pc = call_c(dest, relocInfo::runtime_call_type);
#else
address return_pc = call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, dest), relocInfo::runtime_call_type);
#endif
if (frame_resize) { resize_frame(frame_resize, R0); }
return return_pc;
}
static int check_cabac( int cpu_ref, int cpu_new )
{
int ret = 0, ok, used_asm = 1;
if( cpu_ref || run_cabac_c == run_cabac_asm)
return 0;
set_func_name( "cabac_encode_decision" );
memcpy( buf4, buf3, 0x1000 );
call_c( run_cabac_c, buf3 );
call_a( run_cabac_asm, buf4 );
ok = !memcmp( buf3, buf4, 0x1000 );
report( "cabac :" );
return ret;
}
static int check_mc( int cpu_ref, int cpu_new )
{
x264_mc_functions_t mc_c;
x264_mc_functions_t mc_ref;
x264_mc_functions_t mc_a;
x264_pixel_function_t pixel;
uint8_t *src = &buf1[2*32+2];
uint8_t *src2[4] = { &buf1[3*64+2], &buf1[5*64+2],
&buf1[7*64+2], &buf1[9*64+2] };
uint8_t *dst1 = buf3;
uint8_t *dst2 = buf4;
int dx, dy, i, j, k, w;
int ret = 0, ok, used_asm;
x264_mc_init( 0, &mc_c );
x264_mc_init( cpu_ref, &mc_ref );
x264_mc_init( cpu_new, &mc_a );
x264_pixel_init( 0, &pixel );
#define MC_TEST_LUMA( w, h ) \
if( mc_a.mc_luma != mc_ref.mc_luma && !(w&(w-1)) && h<=16 ) \
{ \
set_func_name( "mc_luma_%dx%d", w, h );\
used_asm = 1; \
memset(buf3, 0xCD, 1024); \
memset(buf4, 0xCD, 1024); \
call_c( mc_c.mc_luma, dst1, 32, src2, 64, dx, dy, w, h ); \
call_a( mc_a.mc_luma, dst2, 32, src2, 64, dx, dy, w, h ); \
if( memcmp( buf3, buf4, 1024 ) ) \
{ \
fprintf( stderr, "mc_luma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
ok = 0; \
} \
} \
if( mc_a.get_ref != mc_ref.get_ref ) \
{ \
uint8_t *ref = dst2; \
int ref_stride = 32; \
set_func_name( "get_ref_%dx%d", w, h );\
used_asm = 1; \
memset(buf3, 0xCD, 1024); \
memset(buf4, 0xCD, 1024); \
call_c( mc_c.mc_luma, dst1, 32, src2, 64, dx, dy, w, h ); \
ref = (uint8_t*) call_a( mc_a.get_ref, ref, &ref_stride, src2, 64, dx, dy, w, h ); \
for( i=0; i<h; i++ ) \
if( memcmp( dst1+i*32, ref+i*ref_stride, w ) ) \
{ \
fprintf( stderr, "get_ref[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
ok = 0; \
break; \
} \
}
#define MC_TEST_CHROMA( w, h ) \
if( mc_a.mc_chroma != mc_ref.mc_chroma ) \
{ \
set_func_name( "mc_chroma_%dx%d", w, h );\
used_asm = 1; \
memset(buf3, 0xCD, 1024); \
memset(buf4, 0xCD, 1024); \
call_c( mc_c.mc_chroma, dst1, 16, src, 32, dx, dy, w, h ); \
call_a( mc_a.mc_chroma, dst2, 16, src, 32, dx, dy, w, h ); \
/* mc_chroma width=2 may write garbage to the right of dst. ignore that. */\
for( j=0; j<h; j++ ) \
for( i=w; i<4; i++ ) \
dst2[i+j*16] = dst1[i+j*16]; \
if( memcmp( buf3, buf4, 1024 ) ) \
{ \
fprintf( stderr, "mc_chroma[mv(%d,%d) %2dx%-2d] [FAILED]\n", dx, dy, w, h ); \
ok = 0; \
} \
}
ok = 1; used_asm = 0;
for( dy = -8; dy < 8; dy++ )
for( dx = -128; dx < 128; dx++ )
{
if( rand()&15 ) continue; // running all of them is too slow
MC_TEST_LUMA( 20, 18 );
MC_TEST_LUMA( 16, 16 );
MC_TEST_LUMA( 16, 8 );
MC_TEST_LUMA( 12, 10 );
MC_TEST_LUMA( 8, 16 );
MC_TEST_LUMA( 8, 8 );
MC_TEST_LUMA( 8, 4 );
MC_TEST_LUMA( 4, 8 );
MC_TEST_LUMA( 4, 4 );
}
report( "mc luma :" );
ok = 1; used_asm = 0;
for( dy = -1; dy < 9; dy++ )
for( dx = -1; dx < 9; dx++ )
{
MC_TEST_CHROMA( 8, 8 );
MC_TEST_CHROMA( 8, 4 );
MC_TEST_CHROMA( 4, 8 );
MC_TEST_CHROMA( 4, 4 );
MC_TEST_CHROMA( 4, 2 );
MC_TEST_CHROMA( 2, 4 );
MC_TEST_CHROMA( 2, 2 );
}
report( "mc chroma :" );
#undef MC_TEST_LUMA
#undef MC_TEST_CHROMA
#define MC_TEST_AVG( name, ... ) \
for( i = 0, ok = 1, used_asm = 0; i < 10; i++ ) \
{ \
memcpy( buf3, buf1, 1024 ); \
memcpy( buf4, buf1, 1024 ); \
if( mc_a.name[i] != mc_ref.name[i] ) \
{ \
set_func_name( "%s_%s", #name, pixel_names[i] );\
used_asm = 1; \
call_c1( mc_c.name[i], buf3, 32, buf2, 16, ##__VA_ARGS__ ); \
call_a1( mc_a.name[i], buf4, 32, buf2, 16, ##__VA_ARGS__ ); \
if( memcmp( buf3, buf4, 1024 ) ) \
{ \
ok = 0; \
fprintf( stderr, #name "[%d]: [FAILED]\n", i ); \
} \
call_c2( mc_c.name[i], buf3, 32, buf2, 16, ##__VA_ARGS__ ); \
call_a2( mc_a.name[i], buf4, 32, buf2, 16, ##__VA_ARGS__ ); \
} \
}
MC_TEST_AVG( avg );
report( "mc avg :" );
ok = 1; used_asm = 0;
for( w = -64; w <= 128 && ok; w++ )
MC_TEST_AVG( avg_weight, w );
report( "mc wpredb :" );
if( mc_a.hpel_filter != mc_ref.hpel_filter )
{
uint8_t *src = buf1+8+2*64;
uint8_t *dstc[3] = { buf3+8, buf3+8+16*64, buf3+8+32*64 };
uint8_t *dsta[3] = { buf4+8, buf4+8+16*64, buf4+8+32*64 };
set_func_name( "hpel_filter" );
ok = 1; used_asm = 1;
memset( buf3, 0, 4096 );
memset( buf4, 0, 4096 );
call_c( mc_c.hpel_filter, dstc[0], dstc[1], dstc[2], src, 64, 48, 10 );
call_a( mc_a.hpel_filter, dsta[0], dsta[1], dsta[2], src, 64, 48, 10 );
for( i=0; i<3; i++ )
for( j=0; j<10; j++ )
//FIXME ideally the first pixels would match too, but they aren't actually used
if( memcmp( dstc[i]+j*64+2, dsta[i]+j*64+2, 43 ) )
{
ok = 0;
fprintf( stderr, "hpel filter differs at plane %c line %d\n", "hvc"[i], j );
for( k=0; k<48; k++ )
printf("%02x%s", dstc[i][j*64+k], (k+1)&3 ? "" : " ");
printf("\n");
for( k=0; k<48; k++ )
printf("%02x%s", dsta[i][j*64+k], (k+1)&3 ? "" : " ");
printf("\n");
break;
}
report( "hpel filter :" );
}
if( mc_a.frame_init_lowres_core != mc_ref.frame_init_lowres_core )
{
uint8_t *dstc[4] = { buf3, buf3+1024, buf3+2048, buf3+3072 };
uint8_t *dsta[4] = { buf4, buf4+1024, buf4+2048, buf3+3072 };
set_func_name( "lowres_init" );
for( w=40; w<=48; w+=8 )
if( mc_a.frame_init_lowres_core != mc_ref.frame_init_lowres_core )
{
int stride = (w+8)&~15;
used_asm = 1;
call_c( mc_c.frame_init_lowres_core, buf1, dstc[0], dstc[1], dstc[2], dstc[3], w*2, stride, w, 16 );
call_a( mc_a.frame_init_lowres_core, buf1, dsta[0], dsta[1], dsta[2], dsta[3], w*2, stride, w, 16 );
for( i=0; i<16; i++)
{
for( j=0; j<4; j++)
if( memcmp( dstc[j]+i*stride, dsta[j]+i*stride, w ) )
{
ok = 0;
fprintf( stderr, "frame_init_lowres differs at plane %d line %d\n", j, i );
for( k=0; k<w; k++ )
printf( "%d ", dstc[j][k+i*stride] );
printf("\n");
for( k=0; k<w; k++ )
printf( "%d ", dsta[j][k+i*stride] );
printf("\n");
break;
}
}
}
report( "lowres init :" );
}
return ret;
}
static int check_dct( int cpu_ref, int cpu_new )
{
x264_dct_function_t dct_c;
x264_dct_function_t dct_ref;
x264_dct_function_t dct_asm;
x264_quant_function_t qf;
int ret = 0, ok, used_asm, i, interlace;
DECLARE_ALIGNED_16( int16_t dct1[16][4][4] );
DECLARE_ALIGNED_16( int16_t dct2[16][4][4] );
DECLARE_ALIGNED_16( int16_t dct4[16][4][4] );
DECLARE_ALIGNED_16( int16_t dct8[4][8][8] );
x264_t h_buf;
x264_t *h = &h_buf;
x264_dct_init( 0, &dct_c );
x264_dct_init( cpu_ref, &dct_ref);
x264_dct_init( cpu_new, &dct_asm );
memset( h, 0, sizeof(*h) );
h->pps = h->pps_array;
x264_param_default( &h->param );
h->param.analyse.i_luma_deadzone[0] = 0;
h->param.analyse.i_luma_deadzone[1] = 0;
h->param.analyse.b_transform_8x8 = 1;
for( i=0; i<6; i++ )
h->pps->scaling_list[i] = x264_cqm_flat16;
x264_cqm_init( h );
x264_quant_init( h, 0, &qf );
#define TEST_DCT( name, t1, t2, size ) \
if( dct_asm.name != dct_ref.name ) \
{ \
set_func_name( #name );\
used_asm = 1; \
call_c( dct_c.name, t1, buf1, buf2 ); \
call_a( dct_asm.name, t2, buf1, buf2 ); \
if( memcmp( t1, t2, size ) ) \
{ \
ok = 0; \
fprintf( stderr, #name " [FAILED]\n" ); \
} \
}
ok = 1; used_asm = 0;
TEST_DCT( sub4x4_dct, dct1[0], dct2[0], 16*2 );
TEST_DCT( sub8x8_dct, dct1, dct2, 16*2*4 );
TEST_DCT( sub16x16_dct, dct1, dct2, 16*2*16 );
report( "sub_dct4 :" );
ok = 1; used_asm = 0;
TEST_DCT( sub8x8_dct8, (void*)dct1[0], (void*)dct2[0], 64*2 );
TEST_DCT( sub16x16_dct8, (void*)dct1, (void*)dct2, 64*2*4 );
report( "sub_dct8 :" );
#undef TEST_DCT
// fdct and idct are denormalized by different factors, so quant/dequant
// is needed to force the coefs into the right range.
dct_c.sub16x16_dct( dct4, buf1, buf2 );
dct_c.sub16x16_dct8( dct8, buf1, buf2 );
for( i=0; i<16; i++ )
{
qf.quant_4x4( dct4[i], h->quant4_mf[CQM_4IY][20], h->quant4_bias[CQM_4IY][20] );
qf.dequant_4x4( dct4[i], h->dequant4_mf[CQM_4IY], 20 );
}
for( i=0; i<4; i++ )
{
qf.quant_8x8( dct8[i], h->quant8_mf[CQM_8IY][20], h->quant8_bias[CQM_8IY][20] );
qf.dequant_8x8( dct8[i], h->dequant8_mf[CQM_8IY], 20 );
}
#define TEST_IDCT( name, src ) \
if( dct_asm.name != dct_ref.name ) \
{ \
set_func_name( #name );\
used_asm = 1; \
memcpy( buf3, buf1, 32*32 ); \
memcpy( buf4, buf1, 32*32 ); \
memcpy( dct1, src, 512 ); \
memcpy( dct2, src, 512 ); \
call_c1( dct_c.name, buf3, (void*)dct1 ); \
call_a1( dct_asm.name, buf4, (void*)dct2 ); \
if( memcmp( buf3, buf4, 32*32 ) ) \
{ \
ok = 0; \
fprintf( stderr, #name " [FAILED]\n" ); \
} \
call_c2( dct_c.name, buf3, (void*)dct1 ); \
call_a2( dct_asm.name, buf4, (void*)dct2 ); \
}
ok = 1; used_asm = 0;
TEST_IDCT( add4x4_idct, dct4 );
TEST_IDCT( add8x8_idct, dct4 );
TEST_IDCT( add16x16_idct, dct4 );
report( "add_idct4 :" );
ok = 1; used_asm = 0;
TEST_IDCT( add8x8_idct8, dct8 );
TEST_IDCT( add16x16_idct8, dct8 );
report( "add_idct8 :" );
#undef TEST_IDCT
ok = 1; used_asm = 0;
if( dct_asm.dct4x4dc != dct_ref.dct4x4dc )
{
DECLARE_ALIGNED_16( int16_t dct1[4][4] ) = {{-12, 42, 23, 67},{2, 90, 89,56},{67,43,-76,91},{56,-78,-54,1}};
DECLARE_ALIGNED_16( int16_t dct2[4][4] ) = {{-12, 42, 23, 67},{2, 90, 89,56},{67,43,-76,91},{56,-78,-54,1}};
set_func_name( "dct4x4dc" );
used_asm = 1;
call_c1( dct_c.dct4x4dc, dct1 );
call_a1( dct_asm.dct4x4dc, dct2 );
if( memcmp( dct1, dct2, 32 ) )
{
ok = 0;
fprintf( stderr, " - dct4x4dc : [FAILED]\n" );
}
call_c2( dct_c.dct4x4dc, dct1 );
call_a2( dct_asm.dct4x4dc, dct2 );
}
if( dct_asm.idct4x4dc != dct_ref.idct4x4dc )
{
DECLARE_ALIGNED_16( int16_t dct1[4][4] ) = {{-12, 42, 23, 67},{2, 90, 89,56},{67,43,-76,91},{56,-78,-54,1}};
DECLARE_ALIGNED_16( int16_t dct2[4][4] ) = {{-12, 42, 23, 67},{2, 90, 89,56},{67,43,-76,91},{56,-78,-54,1}};
set_func_name( "idct4x4dc" );
used_asm = 1;
call_c1( dct_c.idct4x4dc, dct1 );
call_a1( dct_asm.idct4x4dc, dct2 );
if( memcmp( dct1, dct2, 32 ) )
{
ok = 0;
fprintf( stderr, " - idct4x4dc : [FAILED]\n" );
}
call_c2( dct_c.idct4x4dc, dct1 );
call_a2( dct_asm.idct4x4dc, dct2 );
}
report( "(i)dct4x4dc :" );
ok = 1; used_asm = 0;
if( dct_asm.dct2x2dc != dct_ref.dct2x2dc )
{
DECLARE_ALIGNED_16( int16_t dct1[2][2] ) = {{-12, 42},{2, 90}};
DECLARE_ALIGNED_16( int16_t dct2[2][2] ) = {{-12, 42},{2, 90}};
set_func_name( "dct2x2dc" );
used_asm = 1;
call_c( dct_c.dct2x2dc, dct1 );
call_a( dct_asm.dct2x2dc, dct2 );
if( memcmp( dct1, dct2, 4*2 ) )
{
ok = 0;
fprintf( stderr, " - dct2x2dc : [FAILED]\n" );
}
}
if( dct_asm.idct2x2dc != dct_ref.idct2x2dc )
{
DECLARE_ALIGNED_16( int16_t dct1[2][2] ) = {{-12, 42},{2, 90}};
DECLARE_ALIGNED_16( int16_t dct2[2][2] ) = {{-12, 42},{2, 90}};
set_func_name( "idct2x2dc" );
used_asm = 1;
call_c( dct_c.idct2x2dc, dct1 );
call_a( dct_asm.idct2x2dc, dct2 );
if( memcmp( dct1, dct2, 4*2 ) )
{
ok = 0;
fprintf( stderr, " - idct2x2dc : [FAILED]\n" );
}
}
report( "(i)dct2x2dc :" );
x264_zigzag_function_t zigzag_c;
x264_zigzag_function_t zigzag_ref;
x264_zigzag_function_t zigzag_asm;
DECLARE_ALIGNED_16( int16_t level1[64] );
DECLARE_ALIGNED_16( int16_t level2[64] );
#define TEST_ZIGZAG_SCAN( name, t1, t2, dct, size ) \
if( zigzag_asm.name != zigzag_ref.name ) \
{ \
set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" );\
used_asm = 1; \
call_c( zigzag_c.name, t1, dct ); \
call_a( zigzag_asm.name, t2, dct ); \
if( memcmp( t1, t2, size*sizeof(int16_t) ) ) \
{ \
ok = 0; \
fprintf( stderr, #name " [FAILED]\n" ); \
} \
}
#define TEST_ZIGZAG_SUB( name, t1, t2, size ) \
if( zigzag_asm.name != zigzag_ref.name ) \
{ \
set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" );\
used_asm = 1; \
memcpy( buf3, buf1, 16*FDEC_STRIDE ); \
memcpy( buf4, buf1, 16*FDEC_STRIDE ); \
call_c1( zigzag_c.name, t1, buf2, buf3 ); \
call_a1( zigzag_asm.name, t2, buf2, buf4 ); \
if( memcmp( t1, t2, size*sizeof(int16_t) )|| memcmp( buf3, buf4, 16*FDEC_STRIDE ) ) \
{ \
ok = 0; \
fprintf( stderr, #name " [FAILED]\n" ); \
} \
call_c2( zigzag_c.name, t1, buf2, buf3 ); \
call_a2( zigzag_asm.name, t2, buf2, buf4 ); \
}
interlace = 0;
x264_zigzag_init( 0, &zigzag_c, 0 );
x264_zigzag_init( cpu_ref, &zigzag_ref, 0 );
x264_zigzag_init( cpu_new, &zigzag_asm, 0 );
ok = 1; used_asm = 0;
TEST_ZIGZAG_SCAN( scan_8x8, level1, level2, (void*)dct1, 64 );
TEST_ZIGZAG_SCAN( scan_4x4, level1, level2, dct1[0], 16 );
TEST_ZIGZAG_SUB( sub_4x4, level1, level2, 16 );
report( "zigzag_frame :" );
interlace = 1;
x264_zigzag_init( 0, &zigzag_c, 1 );
x264_zigzag_init( cpu_ref, &zigzag_ref, 1 );
x264_zigzag_init( cpu_new, &zigzag_asm, 1 );
ok = 1; used_asm = 0;
TEST_ZIGZAG_SCAN( scan_8x8, level1, level2, (void*)dct1, 64 );
TEST_ZIGZAG_SCAN( scan_4x4, level1, level2, dct1[0], 16 );
TEST_ZIGZAG_SUB( sub_4x4, level1, level2, 16 );
report( "zigzag_field :" );
#undef TEST_ZIGZAG_SCAN
#undef TEST_ZIGZAG_SUB
return ret;
}
static int check_pixel( int cpu_ref, int cpu_new )
{
x264_pixel_function_t pixel_c;
x264_pixel_function_t pixel_ref;
x264_pixel_function_t pixel_asm;
x264_predict_t predict_16x16[4+3];
x264_predict_t predict_8x8c[4+3];
x264_predict_t predict_4x4[9+3];
x264_predict8x8_t predict_8x8[9+3];
DECLARE_ALIGNED_16( uint8_t edge[33] );
uint16_t cost_mv[32];
int ret = 0, ok, used_asm;
int i, j;
x264_pixel_init( 0, &pixel_c );
x264_pixel_init( cpu_ref, &pixel_ref );
x264_pixel_init( cpu_new, &pixel_asm );
x264_predict_16x16_init( 0, predict_16x16 );
x264_predict_8x8c_init( 0, predict_8x8c );
x264_predict_8x8_init( 0, predict_8x8 );
x264_predict_4x4_init( 0, predict_4x4 );
x264_predict_8x8_filter( buf2+40, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
#define TEST_PIXEL( name, align ) \
for( i = 0, ok = 1, used_asm = 0; i < 7; i++ ) \
{ \
int res_c, res_asm; \
if( pixel_asm.name[i] != pixel_ref.name[i] ) \
{ \
set_func_name( "%s_%s", #name, pixel_names[i] ); \
for( j=0; j<64; j++ ) \
{ \
used_asm = 1; \
res_c = call_c( pixel_c.name[i], buf1, 16, buf2+j*!align, 64 ); \
res_asm = call_a( pixel_asm.name[i], buf1, 16, buf2+j*!align, 64 ); \
if( res_c != res_asm ) \
{ \
ok = 0; \
fprintf( stderr, #name "[%d]: %d != %d [FAILED]\n", i, res_c, res_asm ); \
break; \
} \
} \
} \
} \
report( "pixel " #name " :" );
TEST_PIXEL( sad, 0 );
TEST_PIXEL( ssd, 1 );
TEST_PIXEL( satd, 0 );
TEST_PIXEL( sa8d, 0 );
#define TEST_PIXEL_X( N ) \
for( i = 0, ok = 1, used_asm = 0; i < 7; i++ ) \
{ \
int res_c[4]={0}, res_asm[4]={0}; \
if( pixel_asm.sad_x##N[i] && pixel_asm.sad_x##N[i] != pixel_ref.sad_x##N[i] ) \
{ \
set_func_name( "sad_x%d_%s", N, pixel_names[i] ); \
for( j=0; j<64; j++) \
{ \
uint8_t *pix2 = buf2+j; \
used_asm = 1; \
res_c[0] = pixel_c.sad[i]( buf1, 16, pix2, 64 ); \
res_c[1] = pixel_c.sad[i]( buf1, 16, pix2+6, 64 ); \
res_c[2] = pixel_c.sad[i]( buf1, 16, pix2+1, 64 ); \
if(N==4) \
{ \
res_c[3] = pixel_c.sad[i]( buf1, 16, pix2+10, 64 ); \
call_a( pixel_asm.sad_x4[i], buf1, pix2, pix2+6, pix2+1, pix2+10, 64, res_asm ); \
} \
else \
call_a( pixel_asm.sad_x3[i], buf1, pix2, pix2+6, pix2+1, 64, res_asm ); \
if( memcmp(res_c, res_asm, sizeof(res_c)) ) \
{ \
ok = 0; \
fprintf( stderr, "sad_x"#N"[%d]: %d,%d,%d,%d != %d,%d,%d,%d [FAILED]\n", \
i, res_c[0], res_c[1], res_c[2], res_c[3], \
res_asm[0], res_asm[1], res_asm[2], res_asm[3] ); \
} \
if(N==4) \
call_c2( pixel_c.sad_x4[i], buf1, pix2, pix2+6, pix2+1, pix2+10, 64, res_asm ); \
else \
call_c2( pixel_c.sad_x3[i], buf1, pix2, pix2+6, pix2+1, 64, res_asm ); \
} \
} \
} \
report( "pixel sad_x"#N" :" );
TEST_PIXEL_X(3);
TEST_PIXEL_X(4);
#define TEST_PIXEL_VAR( i ) \
if( pixel_asm.var[i] != pixel_ref.var[i] ) \
{ \
uint32_t res_c, res_asm; \
uint32_t sad_c, sad_asm; \
set_func_name( "%s_%s", "var", pixel_names[i] ); \
used_asm = 1; \
res_c = call_c( pixel_c.var[i], buf1, 16, &sad_c ); \
res_asm = call_a( pixel_asm.var[i], buf1, 16, &sad_asm ); \
if( (res_c != res_asm) || (sad_c != sad_asm) ) \
{ \
ok = 0; \
fprintf( stderr, "var[%d]: %d,%d != %d,%d [FAILED]\n", i, res_c, sad_c, res_asm, sad_asm ); \
} \
}
ok = 1; used_asm = 0;
TEST_PIXEL_VAR( PIXEL_16x16 );
TEST_PIXEL_VAR( PIXEL_8x8 );
report( "pixel var :" );
#define TEST_INTRA_SATD( name, pred, satd, i8x8, ... ) \
if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
{ \
int res_c[3], res_asm[3]; \
set_func_name( #name );\
used_asm = 1; \
memcpy( buf3, buf2, 1024 ); \
for( i=0; i<3; i++ ) \
{ \
pred[i]( buf3+40, ##__VA_ARGS__ ); \
res_c[i] = pixel_c.satd( buf1+40, 16, buf3+40, 32 ); \
} \
call_a( pixel_asm.name, buf1+40, i8x8 ? edge : buf3+40, res_asm ); \
if( memcmp(res_c, res_asm, sizeof(res_c)) ) \
{ \
ok = 0; \
fprintf( stderr, #name": %d,%d,%d != %d,%d,%d [FAILED]\n", \
res_c[0], res_c[1], res_c[2], \
res_asm[0], res_asm[1], res_asm[2] ); \
} \
}
ok = 1; used_asm = 0;
TEST_INTRA_SATD( intra_satd_x3_16x16, predict_16x16, satd[PIXEL_16x16], 0 );
TEST_INTRA_SATD( intra_satd_x3_8x8c, predict_8x8c, satd[PIXEL_8x8], 0 );
TEST_INTRA_SATD( intra_satd_x3_4x4, predict_4x4, satd[PIXEL_4x4], 0 );
TEST_INTRA_SATD( intra_sa8d_x3_8x8, predict_8x8, sa8d[PIXEL_8x8], 1, edge );
report( "intra satd_x3 :" );
if( pixel_asm.ssim_4x4x2_core != pixel_ref.ssim_4x4x2_core ||
pixel_asm.ssim_end4 != pixel_ref.ssim_end4 )
{
float res_c, res_a;
int sums[5][4] = {{0}};
used_asm = ok = 1;
x264_emms();
res_c = x264_pixel_ssim_wxh( &pixel_c, buf1+2, 32, buf2+2, 32, 32, 28 );
res_a = x264_pixel_ssim_wxh( &pixel_asm, buf1+2, 32, buf2+2, 32, 32, 28 );
if( fabs(res_c - res_a) > 1e-6 )
{
ok = 0;
fprintf( stderr, "ssim: %.7f != %.7f [FAILED]\n", res_c, res_a );
}
set_func_name( "ssim_core" );
call_c2( pixel_c.ssim_4x4x2_core, buf1+2, 32, buf2+2, 32, sums );
call_a2( pixel_asm.ssim_4x4x2_core, buf1+2, 32, buf2+2, 32, sums );
set_func_name( "ssim_end" );
call_c2( pixel_c.ssim_end4, sums, sums, 4 );
call_a2( pixel_asm.ssim_end4, sums, sums, 4 );
report( "ssim :" );
}
ok = 1; used_asm = 0;
for( i=0; i<32; i++ )
cost_mv[i] = i*10;
for( i=0; i<100 && ok; i++ )
if( pixel_asm.ads[i&3] != pixel_ref.ads[i&3] )
{
DECLARE_ALIGNED_16( uint16_t sums[72] );
DECLARE_ALIGNED_16( int dc[4] );
int16_t mvs_a[32], mvs_c[32];
int mvn_a, mvn_c;
int thresh = rand() & 0x3fff;
set_func_name( "esa_ads" );
for( j=0; j<72; j++ )
sums[j] = rand() & 0x3fff;
for( j=0; j<4; j++ )
dc[j] = rand() & 0x3fff;
used_asm = 1;
mvn_c = call_c( pixel_c.ads[i&3], dc, sums, 32, cost_mv, mvs_c, 28, thresh );
mvn_a = call_a( pixel_asm.ads[i&3], dc, sums, 32, cost_mv, mvs_a, 28, thresh );
if( mvn_c != mvn_a || memcmp( mvs_c, mvs_a, mvn_c*sizeof(*mvs_c) ) )
{
ok = 0;
printf("c%d: ", i&3);
for(j=0; j<mvn_c; j++)
printf("%d ", mvs_c[j]);
printf("\na%d: ", i&3);
for(j=0; j<mvn_a; j++)
printf("%d ", mvs_a[j]);
printf("\n\n");
}
}
report( "esa ads:" );
return ret;
}
inline address MacroAssembler::call_stub(address function_entry) {
return call_c(function_entry);
}
int check_intra( int cpu_ref, int cpu_new ){
int ret = 0, ok = 1, used_asm = 0;
ALIGNED_ARRAY_32( pixel, edge,[36] );
ALIGNED_ARRAY_32( pixel, edge2,[36] );
ALIGNED_ARRAY_32( pixel, fdec,[FDEC_STRIDE*20] );
struct
{
vbench_predict_t predict_16x16[4+3];
vbench_predict_t predict_8x8c[4+3];
vbench_predict_t predict_8x16c[4+3];
vbench_predict8x8_t predict_8x8[9+3];
vbench_predict_t predict_4x4[9+3];
vbench_predict_8x8_filter_t predict_8x8_filter;
} ip_c, ip_ref, ip_a;
vbench_predict_16x16_init( 0, ip_c.predict_16x16 );
vbench_predict_8x8c_init( 0, ip_c.predict_8x8c );
vbench_predict_8x16c_init( 0, ip_c.predict_8x16c );
vbench_predict_8x8_init( 0, ip_c.predict_8x8, &ip_c.predict_8x8_filter );
vbench_predict_4x4_init( 0, ip_c.predict_4x4 );
vbench_predict_16x16_init( cpu_ref, ip_ref.predict_16x16 );
vbench_predict_8x8c_init( cpu_ref, ip_ref.predict_8x8c );
vbench_predict_8x16c_init( cpu_ref, ip_ref.predict_8x16c );
vbench_predict_8x8_init( cpu_ref, ip_ref.predict_8x8, &ip_ref.predict_8x8_filter );
vbench_predict_4x4_init( cpu_ref, ip_ref.predict_4x4 );
vbench_predict_16x16_init( cpu_new, ip_a.predict_16x16 );
vbench_predict_8x8c_init( cpu_new, ip_a.predict_8x8c );
vbench_predict_8x16c_init( cpu_new, ip_a.predict_8x16c );
vbench_predict_8x8_init( cpu_new, ip_a.predict_8x8, &ip_a.predict_8x8_filter );
vbench_predict_4x4_init( cpu_new, ip_a.predict_4x4 );
memcpy( fdec, pbuf1, 32*20 * sizeof(pixel) );\
ip_c.predict_8x8_filter( fdec+48, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
#define INTRA_TEST( name, dir, w, h, align, bench, ... )\
if( ip_a.name[dir] != ip_ref.name[dir] )\
{\
set_func_name( "intra_%s_%s", #name, intra_##name##_names[dir] );\
used_asm = 1;\
memcpy( pbuf3, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
memcpy( pbuf4, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
for( int a = 0; a < ( 64/sizeof(pixel) ); a += align )\
{\
call_c##bench( ip_c.name[dir], pbuf3+48+a, ##__VA_ARGS__ );\
call_a##bench( ip_a.name[dir], pbuf4+48+a, ##__VA_ARGS__ );\
if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*20 * sizeof(pixel) ) )\
{\
fprintf( stderr, #name "[%d] : [FAILED]\n", dir );\
ok = 0;\
if( ip_c.name == (void *)ip_c.predict_8x8 )\
{\
for( int k = -1; k < 16; k++ )\
printf( "%2x ", edge[16+k] );\
printf( "\n" );\
}\
for( int j = 0; j < h; j++ )\
{\
if( ip_c.name == (void *)ip_c.predict_8x8 )\
printf( "%2x ", edge[14-j] );\
for( int k = 0; k < w; k++ )\
printf( "%2x ", pbuf4[48+k+j*FDEC_STRIDE] );\
printf( "\n" );\
}\
printf( "\n" );\
for( int j = 0; j < h; j++ )\
{\
if( ip_c.name == (void *)ip_c.predict_8x8 )\
printf( " " );\
for( int k = 0; k < w; k++ )\
printf( "%2x ", pbuf3[48+k+j*FDEC_STRIDE] );\
printf( "\n" );\
}\
break;\
}\
}\
}
for( int i = 0; i < 12; i++ )
INTRA_TEST( predict_4x4, i, 4, 4, 4, );
for( int i = 0; i < 7; i++ )
INTRA_TEST( predict_8x8c, i, 8, 8, 16, );
for( int i = 0; i < 7; i++ )
INTRA_TEST( predict_8x16c, i, 8, 16, 16, );
for( int i = 0; i < 7; i++ )
INTRA_TEST( predict_16x16, i, 16, 16, 16, );
for( int i = 0; i < 12; i++ )
INTRA_TEST( predict_8x8, i, 8, 8, 8, , edge );
set_func_name("intra_predict_8x8_filter");
if( ip_a.predict_8x8_filter != ip_ref.predict_8x8_filter )
{
used_asm = 1;
for( int i = 0; i < 32; i++ )
{
if( !(i&7) || ((i&MB_TOPRIGHT) && !(i&MB_TOP)) )
continue;
int neighbor = (i&24)>>1;
memset( edge, 0, 36*sizeof(pixel) );
memset( edge2, 0, 36*sizeof(pixel) );
call_c( ip_c.predict_8x8_filter, pbuf1+48, edge, neighbor, i&7 );
call_a( ip_a.predict_8x8_filter, pbuf1+48, edge2, neighbor, i&7 );
if( !(neighbor&MB_TOPLEFT) )
edge[15] = edge2[15] = 0;
if( memcmp( edge+7, edge2+7, (i&MB_TOPRIGHT ? 26 : i&MB_TOP ? 17 : 8) * sizeof(pixel) ) )
{
fprintf( stderr, "predict_8x8_filter : [FAILED] %d %d\n", (i&24)>>1, i&7);
ok = 0;
}
}