potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 07:43:03 +04:00
Code Packages Releases Wiki Activity

core: follow IEEE 754 rules for floating-point division

Browse Source
This commit is contained in:
Alexander Alekhin
2018-10-14 00:37:10 +00:00
parent 09cb329d73
commit fd832bb57d
5 changed files with 63 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/core/src/arithm_core.hpp
+10 -4
View File
@@ -516,7 +516,10 @@ div_i( const T* src1, size_t step1, const T* src2, size_t step2,
for( ; i < width; i++ )
{
T num = src1[i], denom = src2[i];
dst[i] = denom != 0 ? saturate_cast<T>(num*scale_f/denom) : (T)0;
T v = 0;
if (denom != 0)
v = saturate_cast<T>(num*scale_f/denom);
dst[i] = v;
}
}
}
@@ -538,7 +541,7 @@ div_f( const T* src1, size_t step1, const T* src2, size_t step2,
for( ; i < width; i++ )
{
T num = src1[i], denom = src2[i];
dst[i] = denom != 0 ? saturate_cast<T>(num*scale_f/denom) : (T)0;
dst[i] = saturate_cast<T>(num*scale_f/denom);
}
}
}
@@ -559,7 +562,10 @@ recip_i( const T* src2, size_t step2,
for( ; i < width; i++ )
{
T denom = src2[i];
dst[i] = denom != 0 ? saturate_cast<T>(scale_f/denom) : (T)0;
T v = 0;
if (denom != 0)
v = saturate_cast<T>(scale_f/denom);
dst[i] = v;
}
}
}
@@ -580,7 +586,7 @@ recip_f( const T* src2, size_t step2,
for( ; i < width; i++ )
{
T denom = src2[i];
dst[i] = denom != 0 ? saturate_cast<T>(scale_f/denom) : (T)0;
dst[i] = saturate_cast<T>(scale_f/denom);
}
}
}
modules/core/src/arithm_simd.hpp
-16
View File
@@ -1433,7 +1433,6 @@ struct Div_SIMD<float>
return x;
v_float32x4 v_scale = v_setall_f32((float)scale);
v_float32x4 v_zero = v_setzero_f32();
for ( ; x <= width - 8; x += 8)
{
@@ -1445,9 +1444,6 @@ struct Div_SIMD<float>
v_float32x4 res0 = f0 * v_scale / f2;
v_float32x4 res1 = f1 * v_scale / f3;
res0 = v_select(f2 == v_zero, v_zero, res0);
res1 = v_select(f3 == v_zero, v_zero, res1);
v_store(dst + x, res0);
v_store(dst + x + 4, res1);
}
@@ -1675,7 +1671,6 @@ struct Recip_SIMD<float>
return x;
v_float32x4 v_scale = v_setall_f32((float)scale);
v_float32x4 v_zero = v_setzero_f32();
for ( ; x <= width - 8; x += 8)
{
@@ -1685,9 +1680,6 @@ struct Recip_SIMD<float>
v_float32x4 res0 = v_scale / f0;
v_float32x4 res1 = v_scale / f1;
res0 = v_select(f0 == v_zero, v_zero, res0);
res1 = v_select(f1 == v_zero, v_zero, res1);
v_store(dst + x, res0);
v_store(dst + x + 4, res1);
}
@@ -1712,7 +1704,6 @@ struct Div_SIMD<double>
return x;
v_float64x2 v_scale = v_setall_f64(scale);
v_float64x2 v_zero = v_setzero_f64();
for ( ; x <= width - 4; x += 4)
{
@@ -1724,9 +1715,6 @@ struct Div_SIMD<double>
v_float64x2 res0 = f0 * v_scale / f2;
v_float64x2 res1 = f1 * v_scale / f3;
res0 = v_select(f2 == v_zero, v_zero, res0);
res1 = v_select(f3 == v_zero, v_zero, res1);
v_store(dst + x, res0);
v_store(dst + x + 2, res1);
}
@@ -1749,7 +1737,6 @@ struct Recip_SIMD<double>
return x;
v_float64x2 v_scale = v_setall_f64(scale);
v_float64x2 v_zero = v_setzero_f64();
for ( ; x <= width - 4; x += 4)
{
@@ -1759,9 +1746,6 @@ struct Recip_SIMD<double>
v_float64x2 res0 = v_scale / f0;
v_float64x2 res1 = v_scale / f1;
res0 = v_select(f0 == v_zero, v_zero, res0);
res1 = v_select(f1 == v_zero, v_zero, res1);
v_store(dst + x, res0);
v_store(dst + x + 2, res1);
}
modules/core/src/opencl/arithm.cl
+24
View File
@@ -227,9 +227,15 @@
#define PROCESS_ELEM storedst(convertToDT(srcelem1 * scale * srcelem2))
#elif defined OP_DIV
#ifdef CV_DST_TYPE_IS_INTEGER
#define PROCESS_ELEM \
workT e2 = srcelem2, zero = (workT)(0); \
storedst(convertToDT(e2 != zero ? srcelem1 / e2 : zero))
#else
#define PROCESS_ELEM \
workT e2 = srcelem2; \
storedst(convertToDT(srcelem1 / e2))
#endif
#elif defined OP_DIV_SCALE
#undef EXTRA_PARAMS
@@ -240,9 +246,15 @@
#else
#define EXTRA_PARAMS , scaleT scale
#endif
#ifdef CV_DST_TYPE_IS_INTEGER
#define PROCESS_ELEM \
workT e2 = srcelem2, zero = (workT)(0); \
storedst(convertToDT(e2 == zero ? zero : (srcelem1 * (workT)(scale) / e2)))
#else
#define PROCESS_ELEM \
workT e2 = srcelem2; \
storedst(convertToDT(srcelem1 * (workT)(scale) / e2))
#endif
#elif defined OP_RDIV_SCALE
#undef EXTRA_PARAMS
@@ -253,16 +265,28 @@
#else
#define EXTRA_PARAMS , scaleT scale
#endif
#ifdef CV_DST_TYPE_IS_INTEGER
#define PROCESS_ELEM \
workT e1 = srcelem1, zero = (workT)(0); \
storedst(convertToDT(e1 == zero ? zero : (srcelem2 * (workT)(scale) / e1)))
#else
#define PROCESS_ELEM \
workT e1 = srcelem1; \
storedst(convertToDT(srcelem2 * (workT)(scale) / e1))
#endif
#elif defined OP_RECIP_SCALE
#undef EXTRA_PARAMS
#define EXTRA_PARAMS , scaleT scale
#ifdef CV_DST_TYPE_IS_INTEGER
#define PROCESS_ELEM \
workT e1 = srcelem1, zero = (workT)(0); \
storedst(convertToDT(e1 != zero ? scale / e1 : zero))
#else
#define PROCESS_ELEM \
workT e1 = srcelem1; \
storedst(convertToDT(scale / e1))
#endif
#elif defined OP_ADDW
#undef EXTRA_PARAMS