From 7beb24553a16dd5af4e91d0cb067758797c953ca Mon Sep 17 00:00:00 2001 From: Vitaly Tuzov Date: Tue, 25 Dec 2018 14:40:48 +0300 Subject: [PATCH] Speedup filter2d by loop unrolling Added filter2d tests for 16S --- modules/imgproc/src/filter.cpp | 541 ++++++++++++++++++++++----- modules/imgproc/test/test_filter.cpp | 34 +- 2 files changed, 483 insertions(+), 92 deletions(-) diff --git a/modules/imgproc/src/filter.cpp b/modules/imgproc/src/filter.cpp index 9107d0019b..3d865408eb 100644 --- a/modules/imgproc/src/filter.cpp +++ b/modules/imgproc/src/filter.cpp @@ -587,6 +587,7 @@ struct RowVec_8u32s i += v_uint32::nlanes; } } + vx_cleanup(); return i; } @@ -1083,6 +1084,7 @@ struct SymmRowSmallVec_8u32s } } + vx_cleanup(); return i; } @@ -1106,6 +1108,8 @@ struct SymmColumnVec_32s8u int operator()(const uchar** _src, uchar* dst, int width) const { int _ksize = kernel.rows + kernel.cols - 1; + if( _ksize == 1 ) + return 0; int ksize2 = _ksize/2; const float* ky = kernel.ptr() + ksize2; int i = 0, k; @@ -1115,9 +1119,8 @@ struct SymmColumnVec_32s8u v_float32 d4 = vx_setall_f32(delta); if( symmetrical ) { - if (_ksize == 1) - return 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 ) { 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 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); - 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]); - const int* S0 = src[k] + i; - const int* S1 = src[-k] + i; + S0 = src[k] + i; + S1 = src[-k] + i; 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); 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; 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); - 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]); - const int* S0 = src[k] + i; - const int* S1 = src[-k] + i; + S0 = src[k] + i; + S1 = src[-k] + i; 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); } @@ -1160,7 +1173,8 @@ struct SymmColumnVec_32s8u #endif { 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); v_int32x4 s32 = v_round(s0); v_int16x8 s16 = v_pack(s32, s32); @@ -1170,17 +1184,20 @@ struct SymmColumnVec_32s8u } else { + v_float32 f1 = vx_setall_f32(ky[1]); for( ; i <= width - v_uint8::nlanes; i += v_uint8::nlanes ) { - v_float32 s0 = d4; - v_float32 s1 = d4; - v_float32 s2 = d4; - v_float32 s3 = d4; - for ( k = 1; k <= ksize2; k++ ) + const int* S0 = src[1] + i; + const int* S1 = src[-1] + i; + 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); + 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]); - const int* S0 = src[k] + i; - const int* S1 = src[-k] + i; + S0 = src[k] + i; + S1 = src[-k] + i; 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); 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 ) { - v_float32 s0 = d4; - v_float32 s1 = d4; - for ( k = 1; k <= ksize2; k++ ) + const int* S0 = src[1] + i; + const int* S1 = src[-1] + i; + 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]); - const int* S0 = src[k] + i; - const int* S1 = src[-k] + i; + S0 = src[k] + i; + S1 = src[-k] + i; 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); } @@ -1209,8 +1228,8 @@ struct SymmColumnVec_32s8u if( i <= width - v_int32x4::nlanes ) #endif { - v_float32x4 s0 = v_setall_f32(delta); - for (k = 1; k <= ksize2; k++) + 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 = 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); v_int32x4 s32 = v_round(s0); v_int16x8 s16 = v_pack(s32, s32); @@ -1219,6 +1238,7 @@ struct SymmColumnVec_32s8u } } + vx_cleanup(); return i; } @@ -1250,57 +1270,104 @@ struct SymmColumnSmallVec_32s16s short* dst = (short*)_dst; 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( 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 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 ) { 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; } } 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 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 ) { 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; } } +#if CV_NEON 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]); - 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_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 ) { 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; } } +#endif else { 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_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 ) { 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 ) std::swap(S0, S2); - for( ; i <= width - v_int16::nlanes; i += 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)); + 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), 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 ) { - 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; } } else { 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_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 ) { 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; } @@ -1362,19 +1448,43 @@ struct RowVec_16s32f const float* _kx = kernel.ptr(); 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; v_float32 s0 = vx_setzero_f32(); v_float32 s1 = vx_setzero_f32(); for( k = 0; k < _ksize; k++, src += cn ) { + v_float32 f = vx_setall_f32(_kx[k]); v_int16 x = vx_load(src); - s0 = v_muladd(v_cvt_f32(v_expand_low(x)), vx_setall_f32(_kx[k]), s0); - s1 = v_muladd(v_cvt_f32(v_expand_high(x)), vx_setall_f32(_kx[k]), s1); + s0 = v_muladd(v_cvt_f32(v_expand_low(x)), f, s0); + s1 = v_muladd(v_cvt_f32(v_expand_high(x)), f, s1); } v_store(dst + i, s0); v_store(dst + i + v_float32::nlanes, s1); + i += v_int16::nlanes; } if( i <= width - v_float32::nlanes ) { @@ -1385,6 +1495,7 @@ struct RowVec_16s32f v_store(dst + i, s0); i += v_float32::nlanes; } + vx_cleanup(); return i; } @@ -1406,6 +1517,8 @@ struct SymmColumnVec_32f16s int operator()(const uchar** _src, uchar* _dst, int width) const { int _ksize = kernel.rows + kernel.cols - 1; + if( _ksize == 1 ) + return 0; int ksize2 = _ksize / 2; const float* ky = kernel.ptr() + ksize2; int i = 0, k; @@ -1416,25 +1529,49 @@ struct SymmColumnVec_32f16s v_float32 d4 = vx_setall_f32(delta); if( symmetrical ) { - if (_ksize == 1) - return 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 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]); - 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_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, 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))); + i += v_int16::nlanes; } if( i <= width - v_float32::nlanes ) { 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); v_pack_store(dst + i, v_round(s0)); i += v_float32::nlanes; @@ -1442,28 +1579,48 @@ struct SymmColumnVec_32f16s } 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 s1 = d4; - for( k = 1; k <= ksize2; k++ ) + 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); + 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]); - 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_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, 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))); + i += v_int16::nlanes; } if( i <= width - v_float32::nlanes ) { - v_float32 s0 = d4; - for( k = 1; k <= ksize2; k++ ) + 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); v_pack_store(dst + i, v_round(s0)); i += v_float32::nlanes; } } + vx_cleanup(); return i; } @@ -1505,6 +1662,7 @@ struct RowVec_32f } #endif int _ksize = kernel.rows + kernel.cols - 1; + CV_DbgAssert(_ksize > 0); const float* src0 = (const float*)_src; float* dst = (float*)_dst; const float* _kx = kernel.ptr(); @@ -1516,14 +1674,55 @@ struct RowVec_32f if (haveAVX2) return RowVec_32f_AVX(src0, _kx, dst, width, cn, _ksize); #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; - v_float32 s0 = vx_setzero_f32(); - for( k = 0; k < _ksize; k++, src += cn ) + v_float32 s0 = vx_load(src) * k0; + 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); v_store(dst + i, s0); + i += v_float32::nlanes; } + vx_cleanup(); return i; } @@ -1584,6 +1783,8 @@ struct SymmRowSmallVec_32f int operator()(const uchar* _src, uchar* _dst, int width, int cn) const { int i = 0, _ksize = kernel.rows + kernel.cols - 1; + if( _ksize == 1 ) + return 0; float* dst = (float*)_dst; const float* src = (const float*)_src + (_ksize/2)*cn; bool symmetrical = (symmetryType & KERNEL_SYMMETRICAL) != 0; @@ -1592,15 +1793,28 @@ struct SymmRowSmallVec_32f if( symmetrical ) { - if( _ksize == 1 ) - return 0; if( _ksize == 3 ) { if( fabs(kx[0]) == 2 && kx[1] == 1 ) { +#if CV_FMA3 || CV_AVX2 v_float32 k0 = vx_setall_f32(kx[0]); 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))); +#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 { @@ -1613,9 +1827,17 @@ struct SymmRowSmallVec_32f { if( kx[0] == -2 && kx[1] == 0 && kx[2] == 1 ) { +#if CV_FMA3 || CV_AVX2 v_float32 k0 = vx_setall_f32(-2); 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))); +#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 { @@ -1647,6 +1869,7 @@ struct SymmRowSmallVec_32f } } + vx_cleanup(); return i; } @@ -1688,12 +1911,47 @@ struct SymmColumnVec_32f return SymmColumnVec_32f_Symm_AVX(src, ky, dst, delta, width, ksize2); #endif 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++ ) s0 = v_muladd(vx_load(src[k] + i) + vx_load(src[-k] + i), vx_setall_f32(ky[k]), s0); v_store(dst + i, s0); + i += v_float32::nlanes; } } else @@ -1702,16 +1960,53 @@ struct SymmColumnVec_32f if (haveAVX2) return SymmColumnVec_32f_Unsymm_AVX(src, ky, dst, delta, width, ksize2); #endif + CV_DbgAssert(ksize2 > 0); 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; - for( k = 1; k <= ksize2; k++ ) + 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); + 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); v_store(dst + i, s0); + i += v_float32::nlanes; } } + vx_cleanup(); return i; } @@ -1748,9 +2043,24 @@ struct SymmColumnSmallVec_32f { if( fabs(ky[0]) == 2 && ky[1] == 1 ) { +#if CV_FMA3 || CV_AVX2 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)); +#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 { @@ -1776,6 +2086,7 @@ struct SymmColumnSmallVec_32f } } + vx_cleanup(); return i; } @@ -1804,19 +2115,27 @@ struct FilterVec_8u int operator()(const uchar** src, uchar* dst, int width) const { + CV_DbgAssert(_nz > 0); const float* kf = (const float*)&coeffs[0]; int i = 0, k, nz = _nz; 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 ) { - v_float32 s0 = d4, s1 = d4, s2 = d4, s3 = d4; - for( k = 0; k < nz; k++ ) + v_uint16 xl, xh; + 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_uint16 xl, xh; v_expand(vx_load(src[k] + i), xl, xh); - v_uint32 x0, x1, x2, x3; v_expand(xl, x0, x1); v_expand(xh, x2, x3); 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 ) { - v_float32 s0 = d4, s1 = d4; - for( k = 0; k < nz; k++ ) + v_uint32 x0, x1; + 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_uint32 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); 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 ) #endif { - v_float32x4 s0 = v_setall_f32(delta); - for( k = 0; k < nz; k++ ) + 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 = 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); v_int32x4 s32 = v_round(s0); v_int16x8 s16 = v_pack(s32, s32); @@ -1855,6 +2176,7 @@ struct FilterVec_8u i += v_int32x4::nlanes; } + vx_cleanup(); return i; } @@ -1879,18 +2201,24 @@ struct FilterVec_8u16s int operator()(const uchar** src, uchar* _dst, int width) const { + CV_DbgAssert(_nz > 0); const float* kf = (const float*)&coeffs[0]; short* dst = (short*)_dst; int i = 0, k, nz = _nz; 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 ) { - v_float32 s0 = d4, s1 = d4, s2 = d4, s3 = d4; - for( k = 0; k < nz; k++ ) + v_uint16 xl, xh; + 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_uint16 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); 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 ) { - v_float32 s0 = d4, s1 = d4; - for( k = 0; k < nz; k++ ) + v_uint16 x = vx_load_expand(src[0] + i); + 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_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); 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 ) { - v_float32 s0 = d4; - for( k = 0; k < nz; k++ ) + v_float32 s0 = v_muladd(v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(src[0] + i))), f0, d4); + 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); v_pack_store(dst + i, v_round(s0)); i += v_int32::nlanes; } + vx_cleanup(); return i; } @@ -1950,14 +2281,50 @@ struct FilterVec_32f int i = 0, k, nz = _nz; 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; - for( k = 0; k < nz; k++ ) + 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); + 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); v_store(dst + i, s0); + i += v_float32::nlanes; } + vx_cleanup(); return i; } diff --git a/modules/imgproc/test/test_filter.cpp b/modules/imgproc/test/test_filter.cpp index 963de092e2..46a96fc958 100644 --- a/modules/imgproc/test/test_filter.cpp +++ b/modules/imgproc/test/test_filter.cpp @@ -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 ); 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]); - 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); } @@ -457,10 +457,11 @@ void CV_DerivBaseTest::get_test_array_types_and_sizes( int test_case_idx, { RNG& rng = ts->get_rng(); CV_FilterBaseTest::get_test_array_types_and_sizes( test_case_idx, sizes, types ); - int depth = cvtest::randInt(rng) % 2; - depth = depth == 0 ? CV_8U : CV_32F; + int depth = cvtest::randInt(rng) % 4; + depth = depth == 0 ? CV_8U : depth == 1 ? CV_16U : depth == 2 ? CV_16S : CV_32F; 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; 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)); } + +TEST(Imgproc_Sobel, s16_regression_13506) +{ + Mat src = (Mat_(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_(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 -- GitLab