Chiebot-Mirror
/
opencv
mirror of https://github.com/opencv/opencv.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity

Merge pull request #13524 from terfendail:filter_unroll

Browse Source
pull/13451/head
Alexander Alekhin 6 years ago
parent ffc9aa050d 7beb24553a
commit d9089741bc
2 changed files with 483 additions and 92 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 541
      modules/imgproc/src/filter.cpp
  2. 34
      modules/imgproc/test/test_filter.cpp

541
modules/imgproc/src/filter.cpp
Unescape View File

@ -587,6 +587,7 @@ struct RowVec_8u32s
i += v_uint32::nlanes; i += v_uint32::nlanes;
} }
} }
vx_cleanup();
return i; return i;
} }
@ -1083,6 +1084,7 @@ struct SymmRowSmallVec_8u32s
} }
} }
vx_cleanup();
return i; return i;
} }
@ -1106,6 +1108,8 @@ struct SymmColumnVec_32s8u
int operator()(const uchar** _src, uchar* dst, int width) const int operator()(const uchar** _src, uchar* dst, int width) const
{ {
int _ksize = kernel.rows + kernel.cols - 1; int _ksize = kernel.rows + kernel.cols - 1;
if( _ksize == 1 )
return 0;
int ksize2 = _ksize/2; int ksize2 = _ksize/2;
const float* ky = kernel.ptr<float>() + ksize2; const float* ky = kernel.ptr<float>() + ksize2;
int i = 0, k; int i = 0, k;
@ -1115,9 +1119,8 @@ struct SymmColumnVec_32s8u
v_float32 d4 = vx_setall_f32(delta); v_float32 d4 = vx_setall_f32(delta);
if( symmetrical ) if( symmetrical )
{ {
if (_ksize == 1)
return 0;
v_float32 f0 = vx_setall_f32(ky[0]); v_float32 f0 = vx_setall_f32(ky[0]);
v_float32 f1 = vx_setall_f32(ky[1]);
for( ; i <= width - v_uint8::nlanes; i += v_uint8::nlanes ) for( ; i <= width - v_uint8::nlanes; i += v_uint8::nlanes )
{ {
const int* S = src[0] + i; const int* S = src[0] + i;
@ -1125,11 +1128,17 @@ struct SymmColumnVec_32s8u
v_float32 s1 = v_muladd(v_cvt_f32(vx_load(S + v_int32::nlanes)), f0, d4); v_float32 s1 = v_muladd(v_cvt_f32(vx_load(S + v_int32::nlanes)), f0, d4);
v_float32 s2 = v_muladd(v_cvt_f32(vx_load(S + 2*v_int32::nlanes)), f0, d4); v_float32 s2 = v_muladd(v_cvt_f32(vx_load(S + 2*v_int32::nlanes)), f0, d4);
v_float32 s3 = v_muladd(v_cvt_f32(vx_load(S + 3*v_int32::nlanes)), f0, d4); v_float32 s3 = v_muladd(v_cvt_f32(vx_load(S + 3*v_int32::nlanes)), f0, d4);
for( k = 1; k <= ksize2; k++ ) const int* S0 = src[1] + i;
const int* S1 = src[-1] + i;
s0 = v_muladd(v_cvt_f32(vx_load(S0) + vx_load(S1)), f1, s0);
s1 = v_muladd(v_cvt_f32(vx_load(S0 + v_int32::nlanes) + vx_load(S1 + v_int32::nlanes)), f1, s1);
s2 = v_muladd(v_cvt_f32(vx_load(S0 + 2 * v_int32::nlanes) + vx_load(S1 + 2 * v_int32::nlanes)), f1, s2);
s3 = v_muladd(v_cvt_f32(vx_load(S0 + 3 * v_int32::nlanes) + vx_load(S1 + 3 * v_int32::nlanes)), f1, s3);
for( k = 2; k <= ksize2; k++ )
{ {
v_float32 f = vx_setall_f32(ky[k]); v_float32 f = vx_setall_f32(ky[k]);
const int* S0 = src[k] + i; S0 = src[k] + i;
const int* S1 = src[-k] + i; S1 = src[-k] + i;
s0 = v_muladd(v_cvt_f32(vx_load(S0) + vx_load(S1)), f, s0); s0 = v_muladd(v_cvt_f32(vx_load(S0) + vx_load(S1)), f, s0);
s1 = v_muladd(v_cvt_f32(vx_load(S0 + v_int32::nlanes) + vx_load(S1 + v_int32::nlanes)), f, s1); s1 = v_muladd(v_cvt_f32(vx_load(S0 + v_int32::nlanes) + vx_load(S1 + v_int32::nlanes)), f, s1);
s2 = v_muladd(v_cvt_f32(vx_load(S0 + 2*v_int32::nlanes) + vx_load(S1 + 2*v_int32::nlanes)), f, s2); s2 = v_muladd(v_cvt_f32(vx_load(S0 + 2*v_int32::nlanes) + vx_load(S1 + 2*v_int32::nlanes)), f, s2);
@ -1142,11 +1151,15 @@ struct SymmColumnVec_32s8u
const int* S = src[0] + i; const int* S = src[0] + i;
v_float32 s0 = v_muladd(v_cvt_f32(vx_load(S)), f0, d4); v_float32 s0 = v_muladd(v_cvt_f32(vx_load(S)), f0, d4);
v_float32 s1 = v_muladd(v_cvt_f32(vx_load(S + v_int32::nlanes)), f0, d4); v_float32 s1 = v_muladd(v_cvt_f32(vx_load(S + v_int32::nlanes)), f0, d4);
for( k = 1; k <= ksize2; k++ ) const int* S0 = src[1] + i;
const int* S1 = src[-1] + i;
s0 = v_muladd(v_cvt_f32(vx_load(S0) + vx_load(S1)), f1, s0);
s1 = v_muladd(v_cvt_f32(vx_load(S0 + v_int32::nlanes) + vx_load(S1 + v_int32::nlanes)), f1, s1);
for( k = 2; k <= ksize2; k++ )
{ {
v_float32 f = vx_setall_f32(ky[k]); v_float32 f = vx_setall_f32(ky[k]);
const int* S0 = src[k] + i; S0 = src[k] + i;
const int* S1 = src[-k] + i; S1 = src[-k] + i;
s0 = v_muladd(v_cvt_f32(vx_load(S0) + vx_load(S1)), f, s0); s0 = v_muladd(v_cvt_f32(vx_load(S0) + vx_load(S1)), f, s0);
s1 = v_muladd(v_cvt_f32(vx_load(S0 + v_int32::nlanes) + vx_load(S1 + v_int32::nlanes)), f, s1); s1 = v_muladd(v_cvt_f32(vx_load(S0 + v_int32::nlanes) + vx_load(S1 + v_int32::nlanes)), f, s1);
} }
@ -1160,7 +1173,8 @@ struct SymmColumnVec_32s8u
#endif #endif
{ {
v_float32x4 s0 = v_muladd(v_cvt_f32(v_load(src[0] + i)), v_setall_f32(ky[0]), v_setall_f32(delta)); v_float32x4 s0 = v_muladd(v_cvt_f32(v_load(src[0] + i)), v_setall_f32(ky[0]), v_setall_f32(delta));
for( k = 1; k <= ksize2; k++ ) s0 = v_muladd(v_cvt_f32(v_load(src[1] + i) + v_load(src[-1] + i)), v_setall_f32(ky[1]), s0);
for( k = 2; k <= ksize2; k++ )
s0 = v_muladd(v_cvt_f32(v_load(src[k] + i) + v_load(src[-k] + i)), v_setall_f32(ky[k]), s0); s0 = v_muladd(v_cvt_f32(v_load(src[k] + i) + v_load(src[-k] + i)), v_setall_f32(ky[k]), s0);
v_int32x4 s32 = v_round(s0); v_int32x4 s32 = v_round(s0);
v_int16x8 s16 = v_pack(s32, s32); v_int16x8 s16 = v_pack(s32, s32);
@ -1170,17 +1184,20 @@ struct SymmColumnVec_32s8u
} }
else else
{ {
v_float32 f1 = vx_setall_f32(ky[1]);
for( ; i <= width - v_uint8::nlanes; i += v_uint8::nlanes ) for( ; i <= width - v_uint8::nlanes; i += v_uint8::nlanes )
{ {
v_float32 s0 = d4; const int* S0 = src[1] + i;
v_float32 s1 = d4; const int* S1 = src[-1] + i;
v_float32 s2 = d4; v_float32 s0 = v_muladd(v_cvt_f32(vx_load(S0) - vx_load(S1)), f1, d4);
v_float32 s3 = d4; v_float32 s1 = v_muladd(v_cvt_f32(vx_load(S0 + v_int32::nlanes) - vx_load(S1 + v_int32::nlanes)), f1, d4);
for ( k = 1; k <= ksize2; k++ ) v_float32 s2 = v_muladd(v_cvt_f32(vx_load(S0 + 2 * v_int32::nlanes) - vx_load(S1 + 2 * v_int32::nlanes)), f1, d4);
v_float32 s3 = v_muladd(v_cvt_f32(vx_load(S0 + 3 * v_int32::nlanes) - vx_load(S1 + 3 * v_int32::nlanes)), f1, d4);
for ( k = 2; k <= ksize2; k++ )
{ {
v_float32 f = vx_setall_f32(ky[k]); v_float32 f = vx_setall_f32(ky[k]);
const int* S0 = src[k] + i; S0 = src[k] + i;
const int* S1 = src[-k] + i; S1 = src[-k] + i;
s0 = v_muladd(v_cvt_f32(vx_load(S0) - vx_load(S1)), f, s0); s0 = v_muladd(v_cvt_f32(vx_load(S0) - vx_load(S1)), f, s0);
s1 = v_muladd(v_cvt_f32(vx_load(S0 + v_int32::nlanes) - vx_load(S1 + v_int32::nlanes)), f, s1); s1 = v_muladd(v_cvt_f32(vx_load(S0 + v_int32::nlanes) - vx_load(S1 + v_int32::nlanes)), f, s1);
s2 = v_muladd(v_cvt_f32(vx_load(S0 + 2*v_int32::nlanes) - vx_load(S1 + 2*v_int32::nlanes)), f, s2); s2 = v_muladd(v_cvt_f32(vx_load(S0 + 2*v_int32::nlanes) - vx_load(S1 + 2*v_int32::nlanes)), f, s2);
@ -1190,13 +1207,15 @@ struct SymmColumnVec_32s8u
} }
if( i <= width - v_uint16::nlanes ) if( i <= width - v_uint16::nlanes )
{ {
v_float32 s0 = d4; const int* S0 = src[1] + i;
v_float32 s1 = d4; const int* S1 = src[-1] + i;
for ( k = 1; k <= ksize2; k++ ) v_float32 s0 = v_muladd(v_cvt_f32(vx_load(S0) - vx_load(S1)), f1, d4);
v_float32 s1 = v_muladd(v_cvt_f32(vx_load(S0 + v_int32::nlanes) - vx_load(S1 + v_int32::nlanes)), f1, d4);
for ( k = 2; k <= ksize2; k++ )
{ {
v_float32 f = vx_setall_f32(ky[k]); v_float32 f = vx_setall_f32(ky[k]);
const int* S0 = src[k] + i; S0 = src[k] + i;
const int* S1 = src[-k] + i; S1 = src[-k] + i;
s0 = v_muladd(v_cvt_f32(vx_load(S0) - vx_load(S1)), f, s0); s0 = v_muladd(v_cvt_f32(vx_load(S0) - vx_load(S1)), f, s0);
s1 = v_muladd(v_cvt_f32(vx_load(S0 + v_int32::nlanes) - vx_load(S1 + v_int32::nlanes)), f, s1); s1 = v_muladd(v_cvt_f32(vx_load(S0 + v_int32::nlanes) - vx_load(S1 + v_int32::nlanes)), f, s1);
} }
@ -1209,8 +1228,8 @@ struct SymmColumnVec_32s8u
if( i <= width - v_int32x4::nlanes ) if( i <= width - v_int32x4::nlanes )
#endif #endif
{ {
v_float32x4 s0 = v_setall_f32(delta); v_float32x4 s0 = v_muladd(v_cvt_f32(v_load(src[1] + i) - v_load(src[-1] + i)), v_setall_f32(ky[1]), v_setall_f32(delta));
for (k = 1; k <= ksize2; k++) for (k = 2; k <= ksize2; k++)
s0 = v_muladd(v_cvt_f32(v_load(src[k] + i) - v_load(src[-k] + i)), v_setall_f32(ky[k]), s0); s0 = v_muladd(v_cvt_f32(v_load(src[k] + i) - v_load(src[-k] + i)), v_setall_f32(ky[k]), s0);
v_int32x4 s32 = v_round(s0); v_int32x4 s32 = v_round(s0);
v_int16x8 s16 = v_pack(s32, s32); v_int16x8 s16 = v_pack(s32, s32);
@ -1219,6 +1238,7 @@ struct SymmColumnVec_32s8u
} }
} }
vx_cleanup();
return i; return i;
} }
@ -1250,57 +1270,104 @@ struct SymmColumnSmallVec_32s16s
short* dst = (short*)_dst; short* dst = (short*)_dst;
v_float32 df4 = vx_setall_f32(delta); v_float32 df4 = vx_setall_f32(delta);
v_int32 d4 = v_round(df4); int d = cvRound(delta);
v_int16 d8 = vx_setall_s16((short)d);
if( symmetrical ) if( symmetrical )
{ {
if( ky[0] == 2 && ky[1] == 1 ) if( ky[0] == 2 && ky[1] == 1 )
{ {
for( ; i <= width - v_int16::nlanes; i += v_int16::nlanes ) for( ; i <= width - 2*v_int16::nlanes; i += 2*v_int16::nlanes )
{
v_int32 s0 = vx_load(S1 + i);
v_int32 s1 = vx_load(S1 + i + v_int32::nlanes);
v_int32 s2 = vx_load(S1 + i + 2*v_int32::nlanes);
v_int32 s3 = vx_load(S1 + i + 3*v_int32::nlanes);
v_store(dst + i, v_pack(vx_load(S0 + i) + vx_load(S2 + i) + (s0 + s0), vx_load(S0 + i + v_int32::nlanes) + vx_load(S2 + i + v_int32::nlanes) + (s1 + s1)) + d8);
v_store(dst + i + v_int16::nlanes, v_pack(vx_load(S0 + i + 2*v_int32::nlanes) + vx_load(S2 + i + 2*v_int32::nlanes) + (s2 + s2),
vx_load(S0 + i + 3*v_int32::nlanes) + vx_load(S2 + i + 3*v_int32::nlanes) + (s3 + s3)) + d8);
}
if( i <= width - v_int16::nlanes )
{ {
v_int32 sl = vx_load(S1 + i); v_int32 sl = vx_load(S1 + i);
v_int32 sh = vx_load(S1 + i + v_int32::nlanes); v_int32 sh = vx_load(S1 + i + v_int32::nlanes);
v_store(dst + i, v_pack(vx_load(S0 + i) + vx_load(S2 + i) + d4 + (sl + sl), vx_load(S0 + i + v_int32::nlanes) + vx_load(S2 + i + v_int32::nlanes) + d4 + (sh + sh))); v_store(dst + i, v_pack(vx_load(S0 + i) + vx_load(S2 + i) + (sl + sl), vx_load(S0 + i + v_int32::nlanes) + vx_load(S2 + i + v_int32::nlanes) + (sh + sh)) + d8);
i += v_int16::nlanes;
} }
if( i <= width - v_int32::nlanes ) if( i <= width - v_int32::nlanes )
{ {
v_int32 s = vx_load(S1 + i); v_int32 s = vx_load(S1 + i);
v_pack_store(dst + i, vx_load(S0 + i) + vx_load(S2 + i) + d4 + (s + s)); v_pack_store(dst + i, vx_load(S0 + i) + vx_load(S2 + i) + vx_setall_s32(d) + (s + s));
i += v_int32::nlanes; i += v_int32::nlanes;
} }
} }
else if( ky[0] == -2 && ky[1] == 1 ) else if( ky[0] == -2 && ky[1] == 1 )
{ {
for( ; i <= width - v_int16::nlanes; i += v_int16::nlanes ) for( ; i <= width - 2*v_int16::nlanes; i += 2*v_int16::nlanes )
{
v_int32 s0 = vx_load(S1 + i);
v_int32 s1 = vx_load(S1 + i + v_int32::nlanes);
v_int32 s2 = vx_load(S1 + i + 2*v_int32::nlanes);
v_int32 s3 = vx_load(S1 + i + 3*v_int32::nlanes);
v_store(dst + i, v_pack(vx_load(S0 + i) + vx_load(S2 + i) - (s0 + s0),
vx_load(S0 + i + v_int32::nlanes) + vx_load(S2 + i + v_int32::nlanes) - (s1 + s1)) + d8);
v_store(dst + i + v_int16::nlanes, v_pack(vx_load(S0 + i + 2*v_int32::nlanes) + vx_load(S2 + i + 2*v_int32::nlanes) - (s2 + s2),
vx_load(S0 + i + 3*v_int32::nlanes) + vx_load(S2 + i + 3*v_int32::nlanes) - (s3 + s3)) + d8);
}
if( i <= width - v_int16::nlanes )
{ {
v_int32 sl = vx_load(S1 + i); v_int32 sl = vx_load(S1 + i);
v_int32 sh = vx_load(S1 + i + v_int32::nlanes); v_int32 sh = vx_load(S1 + i + v_int32::nlanes);
v_store(dst + i, v_pack(vx_load(S0 + i) + vx_load(S2 + i) + d4 - (sl + sl), vx_load(S0 + i + v_int32::nlanes) + vx_load(S2 + i + v_int32::nlanes) + d4 - (sh + sh))); v_store(dst + i, v_pack(vx_load(S0 + i) + vx_load(S2 + i) - (sl + sl), vx_load(S0 + i + v_int32::nlanes) + vx_load(S2 + i + v_int32::nlanes) - (sh + sh)) + d8);
i += v_int16::nlanes;
} }
if( i <= width - v_int32::nlanes ) if( i <= width - v_int32::nlanes )
{ {
v_int32 s = vx_load(S1 + i); v_int32 s = vx_load(S1 + i);
v_pack_store(dst + i, vx_load(S0 + i) + vx_load(S2 + i) + d4 - (s + s)); v_pack_store(dst + i, vx_load(S0 + i) + vx_load(S2 + i) + vx_setall_s32(d) - (s + s));
i += v_int32::nlanes; i += v_int32::nlanes;
} }
} }
#if CV_NEON
else if( ky[0] == (float)((int)ky[0]) && ky[1] == (float)((int)ky[1]) ) else if( ky[0] == (float)((int)ky[0]) && ky[1] == (float)((int)ky[1]) )
{ {
v_int32 k0 = vx_setall_s32((int)ky[0]), k1 = vx_setall_s32((int)ky[1]); v_int32 k0 = vx_setall_s32((int)ky[0]), k1 = vx_setall_s32((int)ky[1]);
for( ; i <= width - v_int16::nlanes; i += v_int16::nlanes ) v_int32 d4 = vx_setall_s32(d);
for( ; i <= width - 2*v_int16::nlanes; i += 2*v_int16::nlanes )
{
v_store(dst + i, v_pack(v_muladd(vx_load(S0 + i) + vx_load(S2 + i), k1, v_muladd(vx_load(S1 + i), k0, d4)), v_store(dst + i, v_pack(v_muladd(vx_load(S0 + i) + vx_load(S2 + i), k1, v_muladd(vx_load(S1 + i), k0, d4)),
v_muladd(vx_load(S0 + i + v_int32::nlanes) + vx_load(S2 + i + v_int32::nlanes), k1, v_muladd(vx_load(S1 + i + v_int32::nlanes), k0, d4)))); v_muladd(vx_load(S0 + i + v_int32::nlanes) + vx_load(S2 + i + v_int32::nlanes), k1, v_muladd(vx_load(S1 + i + v_int32::nlanes), k0, d4))));
v_store(dst + i + v_int16::nlanes, v_pack(v_muladd(vx_load(S0 + i + 2*v_int32::nlanes) + vx_load(S2 + i + 2*v_int32::nlanes), k1, v_muladd(vx_load(S1 + i + 2*v_int32::nlanes), k0, d4)),
v_muladd(vx_load(S0 + i + 3*v_int32::nlanes) + vx_load(S2 + i + 3*v_int32::nlanes), k1, v_muladd(vx_load(S1 + i + 3*v_int32::nlanes), k0, d4))));
}
if( i <= width - v_int16::nlanes )
{
v_store(dst + i, v_pack(v_muladd(vx_load(S0 + i) + vx_load(S2 + i), k1, v_muladd(vx_load(S1 + i), k0, d4)),
v_muladd(vx_load(S0 + i + v_int32::nlanes) + vx_load(S2 + i + v_int32::nlanes), k1, v_muladd(vx_load(S1 + i + v_int32::nlanes), k0, d4))));
i += v_int16::nlanes;
}
if( i <= width - v_int32::nlanes ) if( i <= width - v_int32::nlanes )
{ {
v_pack_store(dst + i, v_muladd(vx_load(S0 + i) + vx_load(S2 + i), k1, v_muladd(vx_load(S1 + i), k0, d4))); v_pack_store(dst + i, v_muladd(vx_load(S0 + i) + vx_load(S2 + i), k1, v_muladd(vx_load(S1 + i), k0, d4)));
i += v_int32::nlanes; i += v_int32::nlanes;
} }
} }
#endif
else else
{ {
v_float32 k0 = vx_setall_f32(ky[0]), k1 = vx_setall_f32(ky[1]); v_float32 k0 = vx_setall_f32(ky[0]), k1 = vx_setall_f32(ky[1]);
for( ; i <= width - v_int16::nlanes; i += v_int16::nlanes ) for( ; i <= width - 2*v_int16::nlanes; i += 2*v_int16::nlanes )
{
v_store(dst + i, v_pack(v_round(v_muladd(v_cvt_f32(vx_load(S0 + i) + vx_load(S2 + i)), k1, v_muladd(v_cvt_f32(vx_load(S1 + i)), k0, df4))),
v_round(v_muladd(v_cvt_f32(vx_load(S0 + i + v_int32::nlanes) + vx_load(S2 + i + v_int32::nlanes)), k1, v_muladd(v_cvt_f32(vx_load(S1 + i + v_int32::nlanes)), k0, df4)))));
v_store(dst + i + v_int16::nlanes, v_pack(v_round(v_muladd(v_cvt_f32(vx_load(S0 + i + 2*v_int32::nlanes) + vx_load(S2 + i + 2*v_int32::nlanes)), k1, v_muladd(v_cvt_f32(vx_load(S1 + i + 2*v_int32::nlanes)), k0, df4))),
v_round(v_muladd(v_cvt_f32(vx_load(S0 + i + 3*v_int32::nlanes) + vx_load(S2 + i + 3*v_int32::nlanes)), k1, v_muladd(v_cvt_f32(vx_load(S1 + i + 3*v_int32::nlanes)), k0, df4)))));
}
if( i <= width - v_int16::nlanes )
{
v_store(dst + i, v_pack(v_round(v_muladd(v_cvt_f32(vx_load(S0 + i) + vx_load(S2 + i)), k1, v_muladd(v_cvt_f32(vx_load(S1 + i)), k0, df4))), v_store(dst + i, v_pack(v_round(v_muladd(v_cvt_f32(vx_load(S0 + i) + vx_load(S2 + i)), k1, v_muladd(v_cvt_f32(vx_load(S1 + i)), k0, df4))),
v_round(v_muladd(v_cvt_f32(vx_load(S0 + i + v_int32::nlanes) + vx_load(S2 + i + v_int32::nlanes)), k1, v_muladd(v_cvt_f32(vx_load(S1 + i + v_int32::nlanes)), k0, df4))))); v_round(v_muladd(v_cvt_f32(vx_load(S0 + i + v_int32::nlanes) + vx_load(S2 + i + v_int32::nlanes)), k1, v_muladd(v_cvt_f32(vx_load(S1 + i + v_int32::nlanes)), k0, df4)))));
i += v_int16::nlanes;
}
if( i <= width - v_int32::nlanes ) if( i <= width - v_int32::nlanes )
{ {
v_pack_store(dst + i, v_round(v_muladd(v_cvt_f32(vx_load(S0 + i) + vx_load(S2 + i)), k1, v_muladd(v_cvt_f32(vx_load(S1 + i)), k0, df4)))); v_pack_store(dst + i, v_round(v_muladd(v_cvt_f32(vx_load(S0 + i) + vx_load(S2 + i)), k1, v_muladd(v_cvt_f32(vx_load(S1 + i)), k0, df4))));
@ -1314,20 +1381,38 @@ struct SymmColumnSmallVec_32s16s
{ {
if( ky[1] < 0 ) if( ky[1] < 0 )
std::swap(S0, S2); std::swap(S0, S2);
for( ; i <= width - v_int16::nlanes; i += v_int16::nlanes ) for( ; i <= width - 2*v_int16::nlanes; i += 2*v_int16::nlanes )
v_store(dst + i, v_pack(vx_load(S2 + i) - vx_load(S0 + i) + d4, vx_load(S2 + i + v_int32::nlanes) - vx_load(S0 + i + v_int32::nlanes) + d4)); {
v_store(dst + i, v_pack(vx_load(S2 + i) - vx_load(S0 + i), vx_load(S2 + i + v_int32::nlanes) - vx_load(S0 + i + v_int32::nlanes)) + d8);
v_store(dst + i + v_int16::nlanes, v_pack(vx_load(S2 + i + 2*v_int32::nlanes) - vx_load(S0 + i + 2*v_int32::nlanes), vx_load(S2 + i + 3*v_int32::nlanes) - vx_load(S0 + i + 3*v_int32::nlanes)) + d8);
}
if( i <= width - v_int16::nlanes )
{
v_store(dst + i, v_pack(vx_load(S2 + i) - vx_load(S0 + i), vx_load(S2 + i + v_int32::nlanes) - vx_load(S0 + i + v_int32::nlanes)) + d8);
i += v_int16::nlanes;
}
if( i <= width - v_int32::nlanes ) if( i <= width - v_int32::nlanes )
{ {
v_pack_store(dst + i, vx_load(S2 + i) - vx_load(S0 + i) + d4); v_pack_store(dst + i, vx_load(S2 + i) - vx_load(S0 + i) + vx_setall_s32(d));
i += v_int32::nlanes; i += v_int32::nlanes;
} }
} }
else else
{ {
v_float32 k1 = vx_setall_f32(ky[1]); v_float32 k1 = vx_setall_f32(ky[1]);
for( ; i <= width - v_int16::nlanes; i += v_int16::nlanes ) for( ; i <= width - 2*v_int16::nlanes; i += 2*v_int16::nlanes )
{
v_store(dst + i, v_pack(v_round(v_muladd(v_cvt_f32(vx_load(S2 + i) - vx_load(S0 + i)), k1, df4)),
v_round(v_muladd(v_cvt_f32(vx_load(S2 + i + v_int32::nlanes) - vx_load(S0 + i + v_int32::nlanes)), k1, df4))));
v_store(dst + i + v_int16::nlanes, v_pack(v_round(v_muladd(v_cvt_f32(vx_load(S2 + i + 2*v_int32::nlanes) - vx_load(S0 + i + 2*v_int32::nlanes)), k1, df4)),
v_round(v_muladd(v_cvt_f32(vx_load(S2 + i + 3*v_int32::nlanes) - vx_load(S0 + i + 3*v_int32::nlanes)), k1, df4))));
}
if( i <= width - v_int16::nlanes )
{
v_store(dst + i, v_pack(v_round(v_muladd(v_cvt_f32(vx_load(S2 + i) - vx_load(S0 + i)), k1, df4)), v_store(dst + i, v_pack(v_round(v_muladd(v_cvt_f32(vx_load(S2 + i) - vx_load(S0 + i)), k1, df4)),
v_round(v_muladd(v_cvt_f32(vx_load(S2 + i + v_int32::nlanes) - vx_load(S0 + i + v_int32::nlanes)), k1, df4)))); v_round(v_muladd(v_cvt_f32(vx_load(S2 + i + v_int32::nlanes) - vx_load(S0 + i + v_int32::nlanes)), k1, df4))));
i += v_int16::nlanes;
}
if( i <= width - v_int32::nlanes ) if( i <= width - v_int32::nlanes )
{ {
v_pack_store(dst + i, v_round(v_muladd(v_cvt_f32(vx_load(S2 + i) - vx_load(S0 + i)), k1, df4))); v_pack_store(dst + i, v_round(v_muladd(v_cvt_f32(vx_load(S2 + i) - vx_load(S0 + i)), k1, df4)));
@ -1336,6 +1421,7 @@ struct SymmColumnSmallVec_32s16s
} }
} }
vx_cleanup();
return i; return i;
} }
@ -1362,19 +1448,43 @@ struct RowVec_16s32f
const float* _kx = kernel.ptr<float>(); const float* _kx = kernel.ptr<float>();
width *= cn; width *= cn;
for( ; i <= width - v_int16::nlanes; i += v_int16::nlanes ) for( ; i <= width - 2*v_int16::nlanes; i += 2*v_int16::nlanes )
{
const short* src = (const short*)_src + i;
v_float32 s0 = vx_setzero_f32();
v_float32 s1 = vx_setzero_f32();
v_float32 s2 = vx_setzero_f32();
v_float32 s3 = vx_setzero_f32();
for( k = 0; k < _ksize; k++, src += cn )
{
v_float32 f = vx_setall_f32(_kx[k]);
v_int16 xl = vx_load(src);
v_int16 xh = vx_load(src + v_int16::nlanes);
s0 = v_muladd(v_cvt_f32(v_expand_low(xl)), f, s0);
s1 = v_muladd(v_cvt_f32(v_expand_high(xl)), f, s1);
s2 = v_muladd(v_cvt_f32(v_expand_low(xh)), f, s2);
s3 = v_muladd(v_cvt_f32(v_expand_high(xh)), f, s3);
}
v_store(dst + i, s0);
v_store(dst + i + v_float32::nlanes, s1);
v_store(dst + i + 2*v_float32::nlanes, s2);
v_store(dst + i + 3*v_float32::nlanes, s3);
}
if( i <= width - v_int16::nlanes )
{ {
const short* src = (const short*)_src + i; const short* src = (const short*)_src + i;
v_float32 s0 = vx_setzero_f32(); v_float32 s0 = vx_setzero_f32();
v_float32 s1 = vx_setzero_f32(); v_float32 s1 = vx_setzero_f32();
for( k = 0; k < _ksize; k++, src += cn ) for( k = 0; k < _ksize; k++, src += cn )
{ {
v_float32 f = vx_setall_f32(_kx[k]);
v_int16 x = vx_load(src); v_int16 x = vx_load(src);
s0 = v_muladd(v_cvt_f32(v_expand_low(x)), vx_setall_f32(_kx[k]), s0); s0 = v_muladd(v_cvt_f32(v_expand_low(x)), f, s0);
s1 = v_muladd(v_cvt_f32(v_expand_high(x)), vx_setall_f32(_kx[k]), s1); s1 = v_muladd(v_cvt_f32(v_expand_high(x)), f, s1);
} }
v_store(dst + i, s0); v_store(dst + i, s0);
v_store(dst + i + v_float32::nlanes, s1); v_store(dst + i + v_float32::nlanes, s1);
i += v_int16::nlanes;
} }
if( i <= width - v_float32::nlanes ) if( i <= width - v_float32::nlanes )
{ {
@ -1385,6 +1495,7 @@ struct RowVec_16s32f
v_store(dst + i, s0); v_store(dst + i, s0);
i += v_float32::nlanes; i += v_float32::nlanes;
} }
vx_cleanup();
return i; return i;
} }
@ -1406,6 +1517,8 @@ struct SymmColumnVec_32f16s
int operator()(const uchar** _src, uchar* _dst, int width) const int operator()(const uchar** _src, uchar* _dst, int width) const
{ {
int _ksize = kernel.rows + kernel.cols - 1; int _ksize = kernel.rows + kernel.cols - 1;
if( _ksize == 1 )
return 0;
int ksize2 = _ksize / 2; int ksize2 = _ksize / 2;
const float* ky = kernel.ptr<float>() + ksize2; const float* ky = kernel.ptr<float>() + ksize2;
int i = 0, k; int i = 0, k;
@ -1416,25 +1529,49 @@ struct SymmColumnVec_32f16s
v_float32 d4 = vx_setall_f32(delta); v_float32 d4 = vx_setall_f32(delta);
if( symmetrical ) if( symmetrical )
{ {
if (_ksize == 1)
return 0;
v_float32 k0 = vx_setall_f32(ky[0]); v_float32 k0 = vx_setall_f32(ky[0]);
for( ; i <= width - v_int16::nlanes; i += v_int16::nlanes ) v_float32 k1 = vx_setall_f32(ky[1]);
for( ; i <= width - 2*v_int16::nlanes; i += 2*v_int16::nlanes )
{ {
v_float32 s0 = v_muladd(vx_load(src[0] + i), k0, d4); v_float32 s0 = v_muladd(vx_load(src[0] + i), k0, d4);
v_float32 s1 = v_muladd(vx_load(src[0] + i + v_float32::nlanes), k0, d4); v_float32 s1 = v_muladd(vx_load(src[0] + i + v_float32::nlanes), k0, d4);
for( k = 1; k <= ksize2; k++ ) v_float32 s2 = v_muladd(vx_load(src[0] + i + 2*v_float32::nlanes), k0, d4);
v_float32 s3 = v_muladd(vx_load(src[0] + i + 3*v_float32::nlanes), k0, d4);
s0 = v_muladd(vx_load(src[1] + i) + vx_load(src[-1] + i), k1, s0);
s1 = v_muladd(vx_load(src[1] + i + v_float32::nlanes) + vx_load(src[-1] + i + v_float32::nlanes), k1, s1);
s2 = v_muladd(vx_load(src[1] + i + 2*v_float32::nlanes) + vx_load(src[-1] + i + 2*v_float32::nlanes), k1, s2);
s3 = v_muladd(vx_load(src[1] + i + 3*v_float32::nlanes) + vx_load(src[-1] + i + 3*v_float32::nlanes), k1, s3);
for( k = 2; k <= ksize2; k++ )
{ {
v_float32 k1 = vx_setall_f32(ky[k]); v_float32 k2 = vx_setall_f32(ky[k]);
s0 = v_muladd(vx_load(src[k] + i) + vx_load(src[-k] + i), k1, s0); s0 = v_muladd(vx_load(src[k] + i) + vx_load(src[-k] + i), k2, s0);
s1 = v_muladd(vx_load(src[k] + i + v_float32::nlanes) + vx_load(src[-k] + i + v_float32::nlanes), k1, s1); s1 = v_muladd(vx_load(src[k] + i + v_float32::nlanes) + vx_load(src[-k] + i + v_float32::nlanes), k2, s1);
s2 = v_muladd(vx_load(src[k] + i + 2*v_float32::nlanes) + vx_load(src[-k] + i + 2*v_float32::nlanes), k2, s2);
s3 = v_muladd(vx_load(src[k] + i + 3*v_float32::nlanes) + vx_load(src[-k] + i + 3*v_float32::nlanes), k2, s3);
}
v_store(dst + i, v_pack(v_round(s0), v_round(s1)));
v_store(dst + i + v_int16::nlanes, v_pack(v_round(s2), v_round(s3)));
}
if( i <= width - v_int16::nlanes )
{
v_float32 s0 = v_muladd(vx_load(src[0] + i), k0, d4);
v_float32 s1 = v_muladd(vx_load(src[0] + i + v_float32::nlanes), k0, d4);
s0 = v_muladd(vx_load(src[1] + i) + vx_load(src[-1] + i), k1, s0);
s1 = v_muladd(vx_load(src[1] + i + v_float32::nlanes) + vx_load(src[-1] + i + v_float32::nlanes), k1, s1);
for( k = 2; k <= ksize2; k++ )
{
v_float32 k2 = vx_setall_f32(ky[k]);
s0 = v_muladd(vx_load(src[k] + i) + vx_load(src[-k] + i), k2, s0);
s1 = v_muladd(vx_load(src[k] + i + v_float32::nlanes) + vx_load(src[-k] + i + v_float32::nlanes), k2, s1);
} }
v_store(dst + i, v_pack(v_round(s0), v_round(s1))); v_store(dst + i, v_pack(v_round(s0), v_round(s1)));
i += v_int16::nlanes;
} }
if( i <= width - v_float32::nlanes ) if( i <= width - v_float32::nlanes )
{ {
v_float32 s0 = v_muladd(vx_load(src[0] + i), k0, d4); v_float32 s0 = v_muladd(vx_load(src[0] + i), k0, d4);
for( k = 1; k <= ksize2; k++ ) s0 = v_muladd(vx_load(src[1] + i) + vx_load(src[-1] + i), k1, s0);
for( k = 2; k <= ksize2; k++ )
s0 = v_muladd(vx_load(src[k] + i) + vx_load(src[-k] + i), vx_setall_f32(ky[k]), s0); s0 = v_muladd(vx_load(src[k] + i) + vx_load(src[-k] + i), vx_setall_f32(ky[k]), s0);
v_pack_store(dst + i, v_round(s0)); v_pack_store(dst + i, v_round(s0));
i += v_float32::nlanes; i += v_float32::nlanes;
@ -1442,28 +1579,48 @@ struct SymmColumnVec_32f16s
} }
else else
{ {
for( ; i <= width - v_int16::nlanes; i += v_int16::nlanes ) v_float32 k1 = vx_setall_f32(ky[1]);
for( ; i <= width - 2*v_int16::nlanes; i += 2*v_int16::nlanes )
{ {
v_float32 s0 = d4; v_float32 s0 = v_muladd(vx_load(src[1] + i) - vx_load(src[-1] + i), k1, d4);
v_float32 s1 = d4; v_float32 s1 = v_muladd(vx_load(src[1] + i + v_float32::nlanes) - vx_load(src[-1] + i + v_float32::nlanes), k1, d4);
for( k = 1; k <= ksize2; k++ ) v_float32 s2 = v_muladd(vx_load(src[1] + i + 2*v_float32::nlanes) - vx_load(src[-1] + i + 2*v_float32::nlanes), k1, d4);
v_float32 s3 = v_muladd(vx_load(src[1] + i + 3*v_float32::nlanes) - vx_load(src[-1] + i + 3*v_float32::nlanes), k1, d4);
for( k = 2; k <= ksize2; k++ )
{ {
v_float32 k1 = vx_setall_f32(ky[k]); v_float32 k2 = vx_setall_f32(ky[k]);
s0 = v_muladd(vx_load(src[k] + i) - vx_load(src[-k] + i), k1, s0); s0 = v_muladd(vx_load(src[k] + i) - vx_load(src[-k] + i), k2, s0);
s1 = v_muladd(vx_load(src[k] + i + v_float32::nlanes) - vx_load(src[-k] + i + v_float32::nlanes), k1, s1); s1 = v_muladd(vx_load(src[k] + i + v_float32::nlanes) - vx_load(src[-k] + i + v_float32::nlanes), k2, s1);
s2 = v_muladd(vx_load(src[k] + i + 2*v_float32::nlanes) - vx_load(src[-k] + i + 2*v_float32::nlanes), k2, s2);
s3 = v_muladd(vx_load(src[k] + i + 3*v_float32::nlanes) - vx_load(src[-k] + i + 3*v_float32::nlanes), k2, s3);
}
v_store(dst + i, v_pack(v_round(s0), v_round(s1)));
v_store(dst + i + v_int16::nlanes, v_pack(v_round(s2), v_round(s3)));
}
if( i <= width - v_int16::nlanes )
{
v_float32 s0 = v_muladd(vx_load(src[1] + i) - vx_load(src[-1] + i), k1, d4);
v_float32 s1 = v_muladd(vx_load(src[1] + i + v_float32::nlanes) - vx_load(src[-1] + i + v_float32::nlanes), k1, d4);
for( k = 2; k <= ksize2; k++ )
{
v_float32 k2 = vx_setall_f32(ky[k]);
s0 = v_muladd(vx_load(src[k] + i) - vx_load(src[-k] + i), k2, s0);
s1 = v_muladd(vx_load(src[k] + i + v_float32::nlanes) - vx_load(src[-k] + i + v_float32::nlanes), k2, s1);
} }
v_store(dst + i, v_pack(v_round(s0), v_round(s1))); v_store(dst + i, v_pack(v_round(s0), v_round(s1)));
i += v_int16::nlanes;
} }
if( i <= width - v_float32::nlanes ) if( i <= width - v_float32::nlanes )
{ {
v_float32 s0 = d4; v_float32 s0 = v_muladd(vx_load(src[1] + i) - vx_load(src[-1] + i), k1, d4);
for( k = 1; k <= ksize2; k++ ) for( k = 2; k <= ksize2; k++ )
s0 = v_muladd(vx_load(src[k] + i) - vx_load(src[-k] + i), vx_setall_f32(ky[k]), s0); s0 = v_muladd(vx_load(src[k] + i) - vx_load(src[-k] + i), vx_setall_f32(ky[k]), s0);
v_pack_store(dst + i, v_round(s0)); v_pack_store(dst + i, v_round(s0));
i += v_float32::nlanes; i += v_float32::nlanes;
} }
} }
vx_cleanup();
return i; return i;
} }
@ -1505,6 +1662,7 @@ struct RowVec_32f
} }
#endif #endif
int _ksize = kernel.rows + kernel.cols - 1; int _ksize = kernel.rows + kernel.cols - 1;
CV_DbgAssert(_ksize > 0);
const float* src0 = (const float*)_src; const float* src0 = (const float*)_src;
float* dst = (float*)_dst; float* dst = (float*)_dst;
const float* _kx = kernel.ptr<float>(); const float* _kx = kernel.ptr<float>();
@ -1516,14 +1674,55 @@ struct RowVec_32f
if (haveAVX2) if (haveAVX2)
return RowVec_32f_AVX(src0, _kx, dst, width, cn, _ksize); return RowVec_32f_AVX(src0, _kx, dst, width, cn, _ksize);
#endif #endif
for( ; i <= width - v_float32::nlanes; i += v_float32::nlanes ) v_float32 k0 = vx_setall_f32(_kx[0]);
for( ; i <= width - 4*v_float32::nlanes; i += 4*v_float32::nlanes )
{ {
const float* src = src0 + i; const float* src = src0 + i;
v_float32 s0 = vx_setzero_f32(); v_float32 s0 = vx_load(src) * k0;
for( k = 0; k < _ksize; k++, src += cn ) v_float32 s1 = vx_load(src + v_float32::nlanes) * k0;
v_float32 s2 = vx_load(src + 2*v_float32::nlanes) * k0;
v_float32 s3 = vx_load(src + 3*v_float32::nlanes) * k0;
src += cn;
for( k = 1; k < _ksize; k++, src += cn )
{
v_float32 k1 = vx_setall_f32(_kx[k]);
s0 = v_muladd(vx_load(src), k1, s0);
s1 = v_muladd(vx_load(src + v_float32::nlanes), k1, s1);
s2 = v_muladd(vx_load(src + 2*v_float32::nlanes), k1, s2);
s3 = v_muladd(vx_load(src + 3*v_float32::nlanes), k1, s3);
}
v_store(dst + i, s0);
v_store(dst + i + v_float32::nlanes, s1);
v_store(dst + i + 2*v_float32::nlanes, s2);
v_store(dst + i + 3*v_float32::nlanes, s3);
}
if( i <= width - 2*v_float32::nlanes )
{
const float* src = src0 + i;
v_float32 s0 = vx_load(src) * k0;
v_float32 s1 = vx_load(src + v_float32::nlanes) * k0;
src += cn;
for( k = 1; k < _ksize; k++, src += cn )
{
v_float32 k1 = vx_setall_f32(_kx[k]);
s0 = v_muladd(vx_load(src), k1, s0);
s1 = v_muladd(vx_load(src + v_float32::nlanes), k1, s1);
}
v_store(dst + i, s0);
v_store(dst + i + v_float32::nlanes, s1);
i += 2*v_float32::nlanes;
}
if( i <= width - v_float32::nlanes )
{
const float* src = src0 + i;
v_float32 s0 = vx_load(src) * k0;
src += cn;
for( k = 1; k < _ksize; k++, src += cn )
s0 = v_muladd(vx_load(src), vx_setall_f32(_kx[k]), s0); s0 = v_muladd(vx_load(src), vx_setall_f32(_kx[k]), s0);
v_store(dst + i, s0); v_store(dst + i, s0);
i += v_float32::nlanes;
} }
vx_cleanup();
return i; return i;
} }
@ -1584,6 +1783,8 @@ struct SymmRowSmallVec_32f
int operator()(const uchar* _src, uchar* _dst, int width, int cn) const int operator()(const uchar* _src, uchar* _dst, int width, int cn) const
{ {
int i = 0, _ksize = kernel.rows + kernel.cols - 1; int i = 0, _ksize = kernel.rows + kernel.cols - 1;
if( _ksize == 1 )
return 0;
float* dst = (float*)_dst; float* dst = (float*)_dst;
const float* src = (const float*)_src + (_ksize/2)*cn; const float* src = (const float*)_src + (_ksize/2)*cn;
bool symmetrical = (symmetryType & KERNEL_SYMMETRICAL) != 0; bool symmetrical = (symmetryType & KERNEL_SYMMETRICAL) != 0;
@ -1592,15 +1793,28 @@ struct SymmRowSmallVec_32f
if( symmetrical ) if( symmetrical )
{ {
if( _ksize == 1 )
return 0;
if( _ksize == 3 ) if( _ksize == 3 )
{ {
if( fabs(kx[0]) == 2 && kx[1] == 1 ) if( fabs(kx[0]) == 2 && kx[1] == 1 )
{ {
#if CV_FMA3 || CV_AVX2
v_float32 k0 = vx_setall_f32(kx[0]); v_float32 k0 = vx_setall_f32(kx[0]);
for( ; i <= width - v_float32::nlanes; i += v_float32::nlanes, src += v_float32::nlanes ) for( ; i <= width - v_float32::nlanes; i += v_float32::nlanes, src += v_float32::nlanes )
v_store(dst + i, v_muladd(vx_load(src), k0, vx_load(src - cn) + vx_load(src + cn))); v_store(dst + i, v_muladd(vx_load(src), k0, vx_load(src - cn) + vx_load(src + cn)));
#else
if( kx[0] > 0 )
for( ; i <= width - v_float32::nlanes; i += v_float32::nlanes, src += v_float32::nlanes )
{
v_float32 x = vx_load(src);
v_store(dst + i, vx_load(src - cn) + vx_load(src + cn) + (x + x));
}
else
for( ; i <= width - v_float32::nlanes; i += v_float32::nlanes, src += v_float32::nlanes )
{
v_float32 x = vx_load(src);
v_store(dst + i, vx_load(src - cn) + vx_load(src + cn) - (x + x));
}
#endif
} }
else else
{ {
@ -1613,9 +1827,17 @@ struct SymmRowSmallVec_32f
{ {
if( kx[0] == -2 && kx[1] == 0 && kx[2] == 1 ) if( kx[0] == -2 && kx[1] == 0 && kx[2] == 1 )
{ {
#if CV_FMA3 || CV_AVX2
v_float32 k0 = vx_setall_f32(-2); v_float32 k0 = vx_setall_f32(-2);
for( ; i <= width - v_float32::nlanes; i += v_float32::nlanes, src += v_float32::nlanes ) for( ; i <= width - v_float32::nlanes; i += v_float32::nlanes, src += v_float32::nlanes )
v_store(dst + i, v_muladd(vx_load(src), k0, vx_load(src - 2*cn) + vx_load(src + 2*cn))); v_store(dst + i, v_muladd(vx_load(src), k0, vx_load(src - 2*cn) + vx_load(src + 2*cn)));
#else
for( ; i <= width - v_float32::nlanes; i += v_float32::nlanes, src += v_float32::nlanes )
{
v_float32 x = vx_load(src);
v_store(dst + i, vx_load(src - 2*cn) + vx_load(src + 2*cn) - (x + x));
}
#endif
} }
else else
{ {
@ -1647,6 +1869,7 @@ struct SymmRowSmallVec_32f
} }
} }
vx_cleanup();
return i; return i;
} }
@ -1688,12 +1911,47 @@ struct SymmColumnVec_32f
return SymmColumnVec_32f_Symm_AVX(src, ky, dst, delta, width, ksize2); return SymmColumnVec_32f_Symm_AVX(src, ky, dst, delta, width, ksize2);
#endif #endif
const v_float32 d4 = vx_setall_f32(delta); const v_float32 d4 = vx_setall_f32(delta);
for ( ; i <= width - v_float32::nlanes; i += v_float32::nlanes ) const v_float32 k0 = vx_setall_f32(ky[0]);
for( ; i <= width - 4*v_float32::nlanes; i += 4*v_float32::nlanes )
{
v_float32 s0 = v_muladd(vx_load(src[0] + i), k0, d4);
v_float32 s1 = v_muladd(vx_load(src[0] + i + v_float32::nlanes), k0, d4);
v_float32 s2 = v_muladd(vx_load(src[0] + i + 2*v_float32::nlanes), k0, d4);
v_float32 s3 = v_muladd(vx_load(src[0] + i + 3*v_float32::nlanes), k0, d4);
for( k = 1; k <= ksize2; k++ )
{
v_float32 k1 = vx_setall_f32(ky[k]);
s0 = v_muladd(vx_load(src[k] + i) + vx_load(src[-k] + i), k1, s0);
s1 = v_muladd(vx_load(src[k] + i + v_float32::nlanes) + vx_load(src[-k] + i + v_float32::nlanes), k1, s1);
s2 = v_muladd(vx_load(src[k] + i + 2*v_float32::nlanes) + vx_load(src[-k] + i + 2*v_float32::nlanes), k1, s2);
s3 = v_muladd(vx_load(src[k] + i + 3*v_float32::nlanes) + vx_load(src[-k] + i + 3*v_float32::nlanes), k1, s3);
}
v_store(dst + i, s0);
v_store(dst + i + v_float32::nlanes, s1);
v_store(dst + i + 2*v_float32::nlanes, s2);
v_store(dst + i + 3*v_float32::nlanes, s3);
}
if( i <= width - 2*v_float32::nlanes )
{
v_float32 s0 = v_muladd(vx_load(src[0] + i), k0, d4);
v_float32 s1 = v_muladd(vx_load(src[0] + i + v_float32::nlanes), k0, d4);
for( k = 1; k <= ksize2; k++ )
{
v_float32 k1 = vx_setall_f32(ky[k]);
s0 = v_muladd(vx_load(src[k] + i) + vx_load(src[-k] + i), k1, s0);
s1 = v_muladd(vx_load(src[k] + i + v_float32::nlanes) + vx_load(src[-k] + i + v_float32::nlanes), k1, s1);
}
v_store(dst + i, s0);
v_store(dst + i + v_float32::nlanes, s1);
i += 2*v_float32::nlanes;
}
if( i <= width - v_float32::nlanes )
{ {
v_float32 s0 = v_muladd(vx_load(src[0] + i), vx_setall_f32(ky[0]), d4); v_float32 s0 = v_muladd(vx_load(src[0] + i), k0, d4);
for( k = 1; k <= ksize2; k++ ) for( k = 1; k <= ksize2; k++ )
s0 = v_muladd(vx_load(src[k] + i) + vx_load(src[-k] + i), vx_setall_f32(ky[k]), s0); s0 = v_muladd(vx_load(src[k] + i) + vx_load(src[-k] + i), vx_setall_f32(ky[k]), s0);
v_store(dst + i, s0); v_store(dst + i, s0);
i += v_float32::nlanes;
} }
} }
else else
@ -1702,16 +1960,53 @@ struct SymmColumnVec_32f
if (haveAVX2) if (haveAVX2)
return SymmColumnVec_32f_Unsymm_AVX(src, ky, dst, delta, width, ksize2); return SymmColumnVec_32f_Unsymm_AVX(src, ky, dst, delta, width, ksize2);
#endif #endif
CV_DbgAssert(ksize2 > 0);
const v_float32 d4 = vx_setall_f32(delta); const v_float32 d4 = vx_setall_f32(delta);
for ( ; i <= width - v_float32::nlanes; i += v_float32::nlanes ) const v_float32 k1 = vx_setall_f32(ky[1]);
for( ; i <= width - 4*v_float32::nlanes; i += 4*v_float32::nlanes )
{ {
v_float32 s0 = d4; v_float32 s0 = v_muladd(vx_load(src[1] + i) - vx_load(src[-1] + i), k1, d4);
for( k = 1; k <= ksize2; k++ ) v_float32 s1 = v_muladd(vx_load(src[1] + i + v_float32::nlanes) - vx_load(src[-1] + i + v_float32::nlanes), k1, d4);
v_float32 s2 = v_muladd(vx_load(src[1] + i + 2*v_float32::nlanes) - vx_load(src[-1] + i + 2*v_float32::nlanes), k1, d4);
v_float32 s3 = v_muladd(vx_load(src[1] + i + 3*v_float32::nlanes) - vx_load(src[-1] + i + 3*v_float32::nlanes), k1, d4);
for( k = 2; k <= ksize2; k++ )
{
v_float32 k2 = vx_setall_f32(ky[k]);
s0 = v_muladd(vx_load(src[k] + i) - vx_load(src[-k] + i), k2, s0);
s1 = v_muladd(vx_load(src[k] + i + v_float32::nlanes) - vx_load(src[-k] + i + v_float32::nlanes), k2, s1);
s2 = v_muladd(vx_load(src[k] + i + 2*v_float32::nlanes) - vx_load(src[-k] + i + 2*v_float32::nlanes), k2, s2);
s3 = v_muladd(vx_load(src[k] + i + 3*v_float32::nlanes) - vx_load(src[-k] + i + 3*v_float32::nlanes), k2, s3);
}
v_store(dst + i, s0);
v_store(dst + i + v_float32::nlanes, s1);
v_store(dst + i + 2*v_float32::nlanes, s2);
v_store(dst + i + 3*v_float32::nlanes, s3);
}
if( i <= width - 2*v_float32::nlanes )
{
v_float32 s0 = v_muladd(vx_load(src[1] + i) - vx_load(src[-1] + i), k1, d4);
v_float32 s1 = v_muladd(vx_load(src[1] + i + v_float32::nlanes) - vx_load(src[-1] + i + v_float32::nlanes), k1, d4);
for( k = 2; k <= ksize2; k++ )
{
v_float32 k2 = vx_setall_f32(ky[k]);
s0 = v_muladd(vx_load(src[k] + i) - vx_load(src[-k] + i), k2, s0);
s1 = v_muladd(vx_load(src[k] + i + v_float32::nlanes) - vx_load(src[-k] + i + v_float32::nlanes), k2, s1);
}
v_store(dst + i, s0);
v_store(dst + i + v_float32::nlanes, s1);
i += 2*v_float32::nlanes;
}
if( i <= width - v_float32::nlanes )
{
v_float32 s0 = v_muladd(vx_load(src[1] + i) - vx_load(src[-1] + i), k1, d4);
for( k = 2; k <= ksize2; k++ )
s0 = v_muladd(vx_load(src[k] + i) - vx_load(src[-k] + i), vx_setall_f32(ky[k]), s0); s0 = v_muladd(vx_load(src[k] + i) - vx_load(src[-k] + i), vx_setall_f32(ky[k]), s0);
v_store(dst + i, s0); v_store(dst + i, s0);
i += v_float32::nlanes;
} }
} }
vx_cleanup();
return i; return i;
} }
@ -1748,9 +2043,24 @@ struct SymmColumnSmallVec_32f
{ {
if( fabs(ky[0]) == 2 && ky[1] == 1 ) if( fabs(ky[0]) == 2 && ky[1] == 1 )
{ {
#if CV_FMA3 || CV_AVX2
v_float32 k0 = vx_setall_f32(ky[0]); v_float32 k0 = vx_setall_f32(ky[0]);
for ( ; i <= width - v_float32::nlanes; i += v_float32::nlanes ) for( ; i <= width - v_float32::nlanes; i += v_float32::nlanes )
v_store(dst + i, v_muladd(vx_load(S1 + i), k0, vx_load(S0 + i) + vx_load(S2 + i) + d4)); v_store(dst + i, v_muladd(vx_load(S1 + i), k0, vx_load(S0 + i) + vx_load(S2 + i) + d4));
#else
if(ky[0] > 0)
for( ; i <= width - v_float32::nlanes; i += v_float32::nlanes )
{
v_float32 x = vx_load(S1 + i);
v_store(dst + i, vx_load(S0 + i) + vx_load(S2 + i) + d4 + (x + x));
}
else
for( ; i <= width - v_float32::nlanes; i += v_float32::nlanes )
{
v_float32 x = vx_load(S1 + i);
v_store(dst + i, vx_load(S0 + i) + vx_load(S2 + i) + d4 - (x + x));
}
#endif
} }
else else
{ {
@ -1776,6 +2086,7 @@ struct SymmColumnSmallVec_32f
} }
} }
vx_cleanup();
return i; return i;
} }
@ -1804,19 +2115,27 @@ struct FilterVec_8u
int operator()(const uchar** src, uchar* dst, int width) const int operator()(const uchar** src, uchar* dst, int width) const
{ {
CV_DbgAssert(_nz > 0);
const float* kf = (const float*)&coeffs[0]; const float* kf = (const float*)&coeffs[0];
int i = 0, k, nz = _nz; int i = 0, k, nz = _nz;
v_float32 d4 = vx_setall_f32(delta); v_float32 d4 = vx_setall_f32(delta);
v_float32 f0 = vx_setall_f32(kf[0]);
for( ; i <= width - v_uint8::nlanes; i += v_uint8::nlanes ) for( ; i <= width - v_uint8::nlanes; i += v_uint8::nlanes )
{ {
v_float32 s0 = d4, s1 = d4, s2 = d4, s3 = d4; v_uint16 xl, xh;
for( k = 0; k < nz; k++ ) v_expand(vx_load(src[0] + i), xl, xh);
v_uint32 x0, x1, x2, x3;
v_expand(xl, x0, x1);
v_expand(xh, x2, x3);
v_float32 s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(x0)), f0, d4);
v_float32 s1 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(x1)), f0, d4);
v_float32 s2 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(x2)), f0, d4);
v_float32 s3 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(x3)), f0, d4);
for( k = 1; k < nz; k++ )
{ {
v_float32 f = vx_setall_f32(kf[k]); v_float32 f = vx_setall_f32(kf[k]);
v_uint16 xl, xh;
v_expand(vx_load(src[k] + i), xl, xh); v_expand(vx_load(src[k] + i), xl, xh);
v_uint32 x0, x1, x2, x3;
v_expand(xl, x0, x1); v_expand(xl, x0, x1);
v_expand(xh, x2, x3); v_expand(xh, x2, x3);
s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(x0)), f, s0); s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(x0)), f, s0);
@ -1828,11 +2147,13 @@ struct FilterVec_8u
} }
if( i <= width - v_uint16::nlanes ) if( i <= width - v_uint16::nlanes )
{ {
v_float32 s0 = d4, s1 = d4; v_uint32 x0, x1;
for( k = 0; k < nz; k++ ) v_expand(vx_load_expand(src[0] + i), x0, x1);
v_float32 s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(x0)), f0, d4);
v_float32 s1 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(x1)), f0, d4);
for( k = 1; k < nz; k++ )
{ {
v_float32 f = vx_setall_f32(kf[k]); v_float32 f = vx_setall_f32(kf[k]);
v_uint32 x0, x1;
v_expand(vx_load_expand(src[k] + i), x0, x1); v_expand(vx_load_expand(src[k] + i), x0, x1);
s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(x0)), f, s0); s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(x0)), f, s0);
s1 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(x1)), f, s1); s1 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(x1)), f, s1);
@ -1846,8 +2167,8 @@ struct FilterVec_8u
if( i <= width - v_int32x4::nlanes ) if( i <= width - v_int32x4::nlanes )
#endif #endif
{ {
v_float32x4 s0 = v_setall_f32(delta); v_float32x4 s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_load_expand_q(src[0] + i))), v_setall_f32(kf[0]), v_setall_f32(delta));
for( k = 0; k < nz; k++ ) for( k = 1; k < nz; k++ )
s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_load_expand_q(src[k] + i))), v_setall_f32(kf[k]), s0); s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_load_expand_q(src[k] + i))), v_setall_f32(kf[k]), s0);
v_int32x4 s32 = v_round(s0); v_int32x4 s32 = v_round(s0);
v_int16x8 s16 = v_pack(s32, s32); v_int16x8 s16 = v_pack(s32, s32);
@ -1855,6 +2176,7 @@ struct FilterVec_8u
i += v_int32x4::nlanes; i += v_int32x4::nlanes;
} }
vx_cleanup();
return i; return i;
} }
@ -1879,18 +2201,24 @@ struct FilterVec_8u16s
int operator()(const uchar** src, uchar* _dst, int width) const int operator()(const uchar** src, uchar* _dst, int width) const
{ {
CV_DbgAssert(_nz > 0);
const float* kf = (const float*)&coeffs[0]; const float* kf = (const float*)&coeffs[0];
short* dst = (short*)_dst; short* dst = (short*)_dst;
int i = 0, k, nz = _nz; int i = 0, k, nz = _nz;
v_float32 d4 = vx_setall_f32(delta); v_float32 d4 = vx_setall_f32(delta);
v_float32 f0 = vx_setall_f32(kf[0]);
for( ; i <= width - v_uint8::nlanes; i += v_uint8::nlanes ) for( ; i <= width - v_uint8::nlanes; i += v_uint8::nlanes )
{ {
v_float32 s0 = d4, s1 = d4, s2 = d4, s3 = d4; v_uint16 xl, xh;
for( k = 0; k < nz; k++ ) v_expand(vx_load(src[0] + i), xl, xh);
v_float32 s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_low(xl))), f0, d4);
v_float32 s1 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_high(xl))), f0, d4);
v_float32 s2 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_low(xh))), f0, d4);
v_float32 s3 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_high(xh))), f0, d4);
for( k = 1; k < nz; k++ )
{ {
v_float32 f = vx_setall_f32(kf[k]); v_float32 f = vx_setall_f32(kf[k]);
v_uint16 xl, xh;
v_expand(vx_load(src[k] + i), xl, xh); v_expand(vx_load(src[k] + i), xl, xh);
s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_low(xl))), f, s0); s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_low(xl))), f, s0);
s1 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_high(xl))), f, s1); s1 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_high(xl))), f, s1);
@ -1902,11 +2230,13 @@ struct FilterVec_8u16s
} }
if( i <= width - v_uint16::nlanes ) if( i <= width - v_uint16::nlanes )
{ {
v_float32 s0 = d4, s1 = d4; v_uint16 x = vx_load_expand(src[0] + i);
for( k = 0; k < nz; k++ ) v_float32 s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_low(x))), f0, d4);
v_float32 s1 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_high(x))), f0, d4);
for( k = 1; k < nz; k++ )
{ {
v_float32 f = vx_setall_f32(kf[k]); v_float32 f = vx_setall_f32(kf[k]);
v_uint16 x = vx_load_expand(src[k] + i); x = vx_load_expand(src[k] + i);
s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_low(x))), f, s0); s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_low(x))), f, s0);
s1 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_high(x))), f, s1); s1 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(v_expand_high(x))), f, s1);
} }
@ -1915,13 +2245,14 @@ struct FilterVec_8u16s
} }
if( i <= width - v_int32::nlanes ) if( i <= width - v_int32::nlanes )
{ {
v_float32 s0 = d4; v_float32 s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(src[0] + i))), f0, d4);
for( k = 0; k < nz; k++ ) for( k = 1; k < nz; k++ )
s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(src[k] + i))), vx_setall_f32(kf[k]), s0); s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(src[k] + i))), vx_setall_f32(kf[k]), s0);
v_pack_store(dst + i, v_round(s0)); v_pack_store(dst + i, v_round(s0));
i += v_int32::nlanes; i += v_int32::nlanes;
} }
vx_cleanup();
return i; return i;
} }
@ -1950,14 +2281,50 @@ struct FilterVec_32f
int i = 0, k, nz = _nz; int i = 0, k, nz = _nz;
v_float32 d4 = vx_setall_f32(delta); v_float32 d4 = vx_setall_f32(delta);
for( ; i <= width - v_float32::nlanes; i += v_float32::nlanes ) v_float32 f0 = vx_setall_f32(kf[0]);
for( ; i <= width - 4*v_float32::nlanes; i += 4*v_float32::nlanes )
{ {
v_float32 s0 = d4; v_float32 s0 = v_muladd(vx_load(src[0] + i), f0, d4);
for( k = 0; k < nz; k++ ) v_float32 s1 = v_muladd(vx_load(src[0] + i + v_float32::nlanes), f0, d4);
v_float32 s2 = v_muladd(vx_load(src[0] + i + 2*v_float32::nlanes), f0, d4);
v_float32 s3 = v_muladd(vx_load(src[0] + i + 3*v_float32::nlanes), f0, d4);
for( k = 1; k < nz; k++ )
{
v_float32 f1 = vx_setall_f32(kf[k]);
s0 = v_muladd(vx_load(src[k] + i), f1, s0);
s1 = v_muladd(vx_load(src[k] + i + v_float32::nlanes), f1, s1);
s2 = v_muladd(vx_load(src[k] + i + 2*v_float32::nlanes), f1, s2);
s3 = v_muladd(vx_load(src[k] + i + 3*v_float32::nlanes), f1, s3);
}
v_store(dst + i, s0);
v_store(dst + i + v_float32::nlanes, s1);
v_store(dst + i + 2*v_float32::nlanes, s2);
v_store(dst + i + 3*v_float32::nlanes, s3);
}
if( i <= width - 2*v_float32::nlanes )
{
v_float32 s0 = v_muladd(vx_load(src[0] + i), f0, d4);
v_float32 s1 = v_muladd(vx_load(src[0] + i + v_float32::nlanes), f0, d4);
for( k = 1; k < nz; k++ )
{
v_float32 f1 = vx_setall_f32(kf[k]);
s0 = v_muladd(vx_load(src[k] + i), f1, s0);
s1 = v_muladd(vx_load(src[k] + i + v_float32::nlanes), f1, s1);
}
v_store(dst + i, s0);
v_store(dst + i + v_float32::nlanes, s1);
i += 2*v_float32::nlanes;
}
if( i <= width - v_float32::nlanes )
{
v_float32 s0 = v_muladd(vx_load(src[0] + i), f0, d4);
for( k = 1; k < nz; k++ )
s0 = v_muladd(vx_load(src[k] + i), vx_setall_f32(kf[k]), s0); s0 = v_muladd(vx_load(src[k] + i), vx_setall_f32(kf[k]), s0);
v_store(dst + i, s0); v_store(dst + i, s0);
i += v_float32::nlanes;
} }
vx_cleanup();
return i; return i;
} }

34
modules/imgproc/test/test_filter.cpp
Unescape View File

@ -403,9 +403,9 @@ void CV_FilterTest::get_test_array_types_and_sizes( int test_case_idx,
{ {
CV_FilterBaseTest::get_test_array_types_and_sizes( test_case_idx, sizes, types ); CV_FilterBaseTest::get_test_array_types_and_sizes( test_case_idx, sizes, types );
RNG& rng = ts->get_rng(); RNG& rng = ts->get_rng();
int depth = cvtest::randInt(rng)%3; int depth = cvtest::randInt(rng)%4;
int cn = CV_MAT_CN(types[INPUT][0]); int cn = CV_MAT_CN(types[INPUT][0]);
depth = depth == 0 ? CV_8U : depth == 1 ? CV_16U : CV_32F; depth = depth == 0 ? CV_8U : depth == 1 ? CV_16U : depth == 2 ? CV_16S : CV_32F;
types[INPUT][0] = types[OUTPUT][0] = types[REF_OUTPUT][0] = CV_MAKETYPE(depth, cn); types[INPUT][0] = types[OUTPUT][0] = types[REF_OUTPUT][0] = CV_MAKETYPE(depth, cn);
} }
@ -457,10 +457,11 @@ void CV_DerivBaseTest::get_test_array_types_and_sizes( int test_case_idx,
{ {
RNG& rng = ts->get_rng(); RNG& rng = ts->get_rng();
CV_FilterBaseTest::get_test_array_types_and_sizes( test_case_idx, sizes, types ); CV_FilterBaseTest::get_test_array_types_and_sizes( test_case_idx, sizes, types );
int depth = cvtest::randInt(rng) % 2; int depth = cvtest::randInt(rng) % 4;
depth = depth == 0 ? CV_8U : CV_32F; depth = depth == 0 ? CV_8U : depth == 1 ? CV_16U : depth == 2 ? CV_16S : CV_32F;
types[INPUT][0] = CV_MAKETYPE(depth,1); types[INPUT][0] = CV_MAKETYPE(depth,1);
types[OUTPUT][0] = types[REF_OUTPUT][0] = CV_MAKETYPE(depth==CV_8U?CV_16S:CV_32F,1); int sameDepth = cvtest::randInt(rng) % 2;
types[OUTPUT][0] = types[REF_OUTPUT][0] = sameDepth ? depth : CV_MAKETYPE(depth==CV_8U?CV_16S:CV_32F,1);
_aperture_size = (cvtest::randInt(rng)%5)*2 - 1; _aperture_size = (cvtest::randInt(rng)%5)*2 - 1;
sizes[INPUT][1] = aperture_size = cvSize(_aperture_size, _aperture_size); sizes[INPUT][1] = aperture_size = cvSize(_aperture_size, _aperture_size);
} }
@ -2211,4 +2212,27 @@ TEST(Imgproc_MedianBlur, hires_regression_13409)
ASSERT_EQ(0.0, cvtest::norm(dst_hires(Rect(516, 516, 1016, 1016)), dst_ref(Rect(4, 4, 1016, 1016)), NORM_INF)); ASSERT_EQ(0.0, cvtest::norm(dst_hires(Rect(516, 516, 1016, 1016)), dst_ref(Rect(4, 4, 1016, 1016)), NORM_INF));
} }
TEST(Imgproc_Sobel, s16_regression_13506)
{
Mat src = (Mat_<short>(8, 16) << 127, 138, 130, 102, 118, 97, 76, 84, 124, 90, 146, 63, 130, 87, 212, 85,
164, 3, 51, 124, 151, 89, 154, 117, 36, 88, 116, 117, 180, 112, 147, 124,
63, 50, 115, 103, 83, 148, 106, 79, 213, 106, 135, 53, 79, 106, 122, 112,
218, 107, 81, 126, 78, 138, 85, 142, 151, 108, 104, 158, 155, 81, 112, 178,
184, 96, 187, 148, 150, 112, 138, 162, 222, 146, 128, 49, 124, 46, 165, 104,
119, 164, 77, 144, 186, 98, 106, 148, 155, 157, 160, 151, 156, 149, 43, 122,
106, 155, 120, 132, 159, 115, 126, 188, 44, 79, 164, 201, 153, 97, 139, 133,
133, 98, 111, 165, 66, 106, 131, 85, 176, 156, 67, 108, 142, 91, 74, 137);
Mat ref = (Mat_<short>(8, 16) << 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-1020, -796, -489, -469, -247, 317, 760, 1429, 1983, 1384, 254, -459, -899, -1197, -1172, -1058,
2552, 2340, 1617, 591, 9, 96, 722, 1985, 2746, 1916, 676, 9, -635, -1115, -779, -380,
3546, 3349, 2838, 2206, 1388, 669, 938, 1880, 2252, 1785, 1083, 606, 180, -298, -464, -418,
816, 966, 1255, 1652, 1619, 924, 535, 288, 5, 601, 1581, 1870, 1520, 625, -627, -1260,
-782, -610, -395, -267, -122, -42, -317, -1378, -2293, -1451, 596, 1870, 1679, 763, -69, -394,
-882, -681, -463, -818, -1167, -732, -463, -1042, -1604, -1592, -1047, -334, -104, -117, 229, 512,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
Mat dst;
Sobel(src, dst, CV_16S, 0, 1, 5);
ASSERT_EQ(0.0, cvtest::norm(dst, ref, NORM_INF));
}
}} // namespace }} // namespace

Write Preview
Loading…
Cancel
Save