static int32_t scalarproduct_int16_altivec(int16_t * v1, int16_t * v2, int order, const int shift)
{
int i;
LOAD_ZERO;
register vec_s16_t vec1, *pv;
register vec_s32_t res = vec_splat_s32(0), t;
register vec_u32_t shifts;
DECLARE_ALIGNED_16(int32_t, ires);
shifts = zero_u32v;
if(shift & 0x10) shifts = vec_add(shifts, vec_sl(vec_splat_u32(0x08), vec_splat_u32(0x1)));
if(shift & 0x08) shifts = vec_add(shifts, vec_splat_u32(0x08));
if(shift & 0x04) shifts = vec_add(shifts, vec_splat_u32(0x04));
if(shift & 0x02) shifts = vec_add(shifts, vec_splat_u32(0x02));
if(shift & 0x01) shifts = vec_add(shifts, vec_splat_u32(0x01));
for(i = 0; i < order; i += 8){
pv = (vec_s16_t*)v1;
vec1 = vec_perm(pv[0], pv[1], vec_lvsl(0, v1));
t = vec_msum(vec1, vec_ld(0, v2), zero_s32v);
t = vec_sr(t, shifts);
res = vec_sums(t, res);
v1 += 8;
v2 += 8;
}
res = vec_splat(res, 3);
vec_ste(res, 0, &ires);
return ires;
}
void x264_me_search_ref( x264_t *h, x264_me_t *m, int16_t (*mvc)[2], int i_mvc, int *p_halfpel_thresh )
{
const int bw = x264_pixel_size[m->i_pixel].w;
const int bh = x264_pixel_size[m->i_pixel].h;
const int i_pixel = m->i_pixel;
int i_me_range = h->param.analyse.i_me_range;
int bmx, bmy, bcost;
int bpred_mx = 0, bpred_my = 0, bpred_cost = COST_MAX;
int omx, omy, pmx, pmy;
uint8_t *p_fref = m->p_fref[0];
DECLARE_ALIGNED_16( uint8_t pix[16*16] );
int i, j;
int dir;
int costs[6];
int mv_x_min = h->mb.mv_min_fpel[0];
int mv_y_min = h->mb.mv_min_fpel[1];
int mv_x_max = h->mb.mv_max_fpel[0];
int mv_y_max = h->mb.mv_max_fpel[1];
#define CHECK_MVRANGE(mx,my) ( mx >= mv_x_min && mx <= mv_x_max && my >= mv_y_min && my <= mv_y_max )
const int16_t *p_cost_mvx = m->p_cost_mv - m->mvp[0];
const int16_t *p_cost_mvy = m->p_cost_mv - m->mvp[1];
bmx = x264_clip3( m->mvp[0], mv_x_min*4, mv_x_max*4 );
bmy = x264_clip3( m->mvp[1], mv_y_min*4, mv_y_max*4 );
pmx = ( bmx + 2 ) >> 2;
pmy = ( bmy + 2 ) >> 2;
bcost = COST_MAX;
/* try extra predictors if provided */
if( h->mb.i_subpel_refine >= 3 )
{
uint32_t bmv = pack16to32_mask(bmx,bmy);
COST_MV_HPEL( bmx, bmy );
for( i = 0; i < i_mvc; i++ )
{
if( *(uint32_t*)mvc[i] && (bmv - *(uint32_t*)mvc[i]) )
{
int mx = x264_clip3( mvc[i][0], mv_x_min*4, mv_x_max*4 );
int my = x264_clip3( mvc[i][1], mv_y_min*4, mv_y_max*4 );
COST_MV_HPEL( mx, my );
}
}
bmx = ( bpred_mx + 2 ) >> 2;
bmy = ( bpred_my + 2 ) >> 2;
COST_MV( bmx, bmy );
}
static int check_quant( int cpu_ref, int cpu_new )
{
x264_quant_function_t qf_c;
x264_quant_function_t qf_ref;
x264_quant_function_t qf_a;
DECLARE_ALIGNED_16( int16_t dct1[64] );
DECLARE_ALIGNED_16( int16_t dct2[64] );
DECLARE_ALIGNED_16( uint8_t cqm_buf[64] );
int ret = 0, ok, used_asm;
int oks[2] = {1,1}, used_asms[2] = {0,0};
int i, i_cqm, qp;
x264_t h_buf;
x264_t *h = &h_buf;
memset( h, 0, sizeof(*h) );
h->pps = h->pps_array;
x264_param_default( &h->param );
h->param.rc.i_qp_min = 26;
h->param.analyse.b_transform_8x8 = 1;
for( i_cqm = 0; i_cqm < 4; i_cqm++ )
{
if( i_cqm == 0 )
{
for( i = 0; i < 6; i++ )
h->pps->scaling_list[i] = x264_cqm_flat16;
h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_FLAT;
}
else if( i_cqm == 1 )
{
for( i = 0; i < 6; i++ )
h->pps->scaling_list[i] = x264_cqm_jvt[i];
h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_JVT;
}
else
{
if( i_cqm == 2 )
for( i = 0; i < 64; i++ )
cqm_buf[i] = 10 + rand() % 246;
else
for( i = 0; i < 64; i++ )
cqm_buf[i] = 1;
for( i = 0; i < 6; i++ )
h->pps->scaling_list[i] = cqm_buf;
h->param.i_cqm_preset = h->pps->i_cqm_preset = X264_CQM_CUSTOM;
}
x264_cqm_init( h );
x264_quant_init( h, 0, &qf_c );
x264_quant_init( h, cpu_ref, &qf_ref );
x264_quant_init( h, cpu_new, &qf_a );
#define INIT_QUANT8() \
{ \
static const int scale1d[8] = {32,31,24,31,32,31,24,31}; \
int x, y; \
for( y = 0; y < 8; y++ ) \
for( x = 0; x < 8; x++ ) \
{ \
unsigned int scale = (255*scale1d[y]*scale1d[x])/16; \
dct1[y*8+x] = dct2[y*8+x] = (rand()%(2*scale+1))-scale; \
} \
}
#define INIT_QUANT4() \
{ \
static const int scale1d[4] = {4,6,4,6}; \
int x, y; \
for( y = 0; y < 4; y++ ) \
for( x = 0; x < 4; x++ ) \
{ \
unsigned int scale = 255*scale1d[y]*scale1d[x]; \
dct1[y*4+x] = dct2[y*4+x] = (rand()%(2*scale+1))-scale; \
} \
}
#define TEST_QUANT_DC( name, cqm ) \
if( qf_a.name != qf_ref.name ) \
{ \
set_func_name( #name ); \
used_asms[0] = 1; \
for( qp = 51; qp > 0; qp-- ) \
{ \
for( i = 0; i < 16; i++ ) \
dct1[i] = dct2[i] = (rand() & 0x1fff) - 0xfff; \
call_c1( qf_c.name, (void*)dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
call_a1( qf_a.name, (void*)dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
if( memcmp( dct1, dct2, 16*2 ) ) \
{ \
oks[0] = 0; \
fprintf( stderr, #name "(cqm=%d): [FAILED]\n", i_cqm ); \
break; \
} \
call_c2( qf_c.name, (void*)dct1, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
call_a2( qf_a.name, (void*)dct2, h->quant4_mf[CQM_4IY][qp][0], h->quant4_bias[CQM_4IY][qp][0] ); \
} \
}
#define TEST_QUANT( qname, block, w ) \
if( qf_a.qname != qf_ref.qname ) \
{ \
set_func_name( #qname ); \
used_asms[0] = 1; \
for( qp = 51; qp > 0; qp-- ) \
{ \
INIT_QUANT##w() \
call_c1( qf_c.qname, (void*)dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
call_a1( qf_a.qname, (void*)dct2, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
if( memcmp( dct1, dct2, w*w*2 ) ) \
{ \
oks[0] = 0; \
fprintf( stderr, #qname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
break; \
} \
call_c2( qf_c.qname, (void*)dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
call_a2( qf_a.qname, (void*)dct2, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
} \
}
TEST_QUANT( quant_8x8, CQM_8IY, 8 );
TEST_QUANT( quant_8x8, CQM_8PY, 8 );
TEST_QUANT( quant_4x4, CQM_4IY, 4 );
TEST_QUANT( quant_4x4, CQM_4PY, 4 );
TEST_QUANT_DC( quant_4x4_dc, **h->quant4_mf[CQM_4IY] );
TEST_QUANT_DC( quant_2x2_dc, **h->quant4_mf[CQM_4IC] );
#define TEST_DEQUANT( qname, dqname, block, w ) \
if( qf_a.dqname != qf_ref.dqname ) \
{ \
set_func_name( "%s_%s", #dqname, i_cqm?"cqm":"flat" ); \
used_asms[1] = 1; \
for( qp = 51; qp > 0; qp-- ) \
{ \
INIT_QUANT##w() \
call_c( qf_c.qname, (void*)dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
memcpy( dct2, dct1, w*w*2 ); \
call_c1( qf_c.dqname, (void*)dct1, h->dequant##w##_mf[block], qp ); \
call_a1( qf_a.dqname, (void*)dct2, h->dequant##w##_mf[block], qp ); \
if( memcmp( dct1, dct2, w*w*2 ) ) \
{ \
oks[1] = 0; \
fprintf( stderr, #dqname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
break; \
} \
call_c2( qf_c.dqname, (void*)dct1, h->dequant##w##_mf[block], qp ); \
call_a2( qf_a.dqname, (void*)dct2, h->dequant##w##_mf[block], qp ); \
} \
}
TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8IY, 8 );
TEST_DEQUANT( quant_8x8, dequant_8x8, CQM_8PY, 8 );
TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4IY, 4 );
TEST_DEQUANT( quant_4x4, dequant_4x4, CQM_4PY, 4 );
x264_cqm_delete( h );
}
ok = oks[0]; used_asm = used_asms[0];
report( "quant :" );
ok = oks[1]; used_asm = used_asms[1];
report( "dequant :" );
if( qf_a.denoise_dct != qf_ref.denoise_dct )
{
int size;
for( size = 16; size <= 64; size += 48 )
{
set_func_name( "denoise_dct" );
used_asm = 1;
memcpy(dct1, buf1, size*2);
memcpy(dct2, buf1, size*2);
memcpy(buf3+256, buf3, 256);
call_c1( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (uint16_t*)buf2, size );
call_a1( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (uint16_t*)buf2, size );
if( memcmp( dct1, dct2, size*2 ) || memcmp( buf3+4, buf3+256+4, (size-1)*sizeof(uint32_t) ) )
ok = 0;
call_c2( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (uint16_t*)buf2, size );
call_a2( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (uint16_t*)buf2, size );
}
}
report( "denoise dct :" );
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;
}
static int check_intra( int cpu_ref, int cpu_new )
{
int ret = 0, ok = 1, used_asm = 0;
int i;
DECLARE_ALIGNED_16( uint8_t edge[33] );
struct
{
x264_predict_t predict_16x16[4+3];
x264_predict_t predict_8x8c[4+3];
x264_predict8x8_t predict_8x8[9+3];
x264_predict_t predict_4x4[9+3];
} ip_c, ip_ref, ip_a;
x264_predict_16x16_init( 0, ip_c.predict_16x16 );
x264_predict_8x8c_init( 0, ip_c.predict_8x8c );
x264_predict_8x8_init( 0, ip_c.predict_8x8 );
x264_predict_4x4_init( 0, ip_c.predict_4x4 );
x264_predict_16x16_init( cpu_ref, ip_ref.predict_16x16 );
x264_predict_8x8c_init( cpu_ref, ip_ref.predict_8x8c );
x264_predict_8x8_init( cpu_ref, ip_ref.predict_8x8 );
x264_predict_4x4_init( cpu_ref, ip_ref.predict_4x4 );
x264_predict_16x16_init( cpu_new, ip_a.predict_16x16 );
x264_predict_8x8c_init( cpu_new, ip_a.predict_8x8c );
x264_predict_8x8_init( cpu_new, ip_a.predict_8x8 );
x264_predict_4x4_init( cpu_new, ip_a.predict_4x4 );
x264_predict_8x8_filter( buf1+48, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
#define INTRA_TEST( name, dir, w, ... ) \
if( ip_a.name[dir] != ip_ref.name[dir] )\
{ \
set_func_name( "intra_%s_%s", #name, intra_##name##_names[dir] );\
used_asm = 1; \
memcpy( buf3, buf1, 32*20 );\
memcpy( buf4, buf1, 32*20 );\
call_c( ip_c.name[dir], buf3+48, ##__VA_ARGS__ );\
call_a( ip_a.name[dir], buf4+48, ##__VA_ARGS__ );\
if( memcmp( buf3, buf4, 32*20 ) )\
{\
fprintf( stderr, #name "[%d] : [FAILED]\n", dir );\
ok = 0;\
int j,k;\
for(k=-1; k<16; k++)\
printf("%2x ", edge[16+k]);\
printf("\n");\
for(j=0; j<w; j++){\
printf("%2x ", edge[14-j]);\
for(k=0; k<w; k++)\
printf("%2x ", buf4[48+k+j*32]);\
printf("\n");\
}\
printf("\n");\
for(j=0; j<w; j++){\
printf(" ");\
for(k=0; k<w; k++)\
printf("%2x ", buf3[48+k+j*32]);\
printf("\n");\
}\
}\
}
for( i = 0; i < 12; i++ )
INTRA_TEST( predict_4x4, i, 4 );
for( i = 0; i < 7; i++ )
INTRA_TEST( predict_8x8c, i, 8 );
for( i = 0; i < 7; i++ )
INTRA_TEST( predict_16x16, i, 16 );
for( i = 0; i < 12; i++ )
INTRA_TEST( predict_8x8, i, 8, edge );
report( "intra pred :" );
return ret;
}