From 6cf4371eb4adbe9df164f724ffad46713e2dbdf5 Mon Sep 17 00:00:00 2001
From: chacha21 <pierre@chachatelier.fr>
Date: Fri, 3 Mar 2017 18:24:12 +0100
Subject: [PATCH 1/3] make cuda::absdiff support multi-channel scalars

I took the subScalar.cu code and changed the inner operation
---
 modules/cudaarithm/src/cuda/absdiff_scalar.cu | 109 ++++++++++++++----
 1 file changed, 88 insertions(+), 21 deletions(-)

diff --git a/modules/cudaarithm/src/cuda/absdiff_scalar.cu b/modules/cudaarithm/src/cuda/absdiff_scalar.cu
index 3ffd0661b7..225298b291 100644
--- a/modules/cudaarithm/src/cuda/absdiff_scalar.cu
+++ b/modules/cudaarithm/src/cuda/absdiff_scalar.cu
@@ -56,14 +56,14 @@ void absDiffScalar(const GpuMat& src, cv::Scalar val, bool, GpuMat& dst, const G
 
 namespace
 {
-    template <typename T, typename S> struct AbsDiffScalarOp : unary_function<T, T>
+    template <typename SrcType, typename ScalarType, typename DstType> struct AbsDiffScalarOp : unary_function<SrcType, DstType>
     {
-        S val;
+        ScalarType val;
 
-        __device__ __forceinline__ T operator ()(T a) const
+        __device__ __forceinline__ DstType operator ()(SrcType a) const
         {
-            abs_func<S> f;
-            return saturate_cast<T>(f(a - val));
+            abs_func<ScalarType> f;
+            return saturate_cast<DstType>(f(saturate_cast<ScalarType>(a) - val));
         }
     };
 
@@ -77,34 +77,101 @@ namespace
         };
     };
 
-    template <typename SrcType, typename ScalarDepth>
-    void absDiffScalarImpl(const GpuMat& src, double value, GpuMat& dst, Stream& stream)
+    template <typename SrcType, typename ScalarDepth, typename DstType>
+    void absDiffScalarImpl(const GpuMat& src, cv::Scalar value, GpuMat& dst, Stream& stream)
     {
-        AbsDiffScalarOp<SrcType, ScalarDepth> op;
-        op.val = static_cast<ScalarDepth>(value);
+        typedef typename MakeVec<ScalarDepth, VecTraits<SrcType>::cn>::type ScalarType;
+
+        cv::Scalar_<ScalarDepth> value_ = value;
+
+        AbsDiffScalarOp<SrcType, ScalarType, DstType> op;
+        op.val = VecTraits<ScalarType>::make(value_.val);
         gridTransformUnary_< TransformPolicy<ScalarDepth> >(globPtr<SrcType>(src), globPtr<SrcType>(dst), op, stream);
     }
 }
 
 void absDiffScalar(const GpuMat& src, cv::Scalar val, bool, GpuMat& dst, const GpuMat&, double, Stream& stream, int)
 {
-    typedef void (*func_t)(const GpuMat& src, double val, GpuMat& dst, Stream& stream);
-    static const func_t funcs[] =
+    typedef void (*func_t)(const GpuMat& src, cv::Scalar val, GpuMat& dst, Stream& stream);
+    static const func_t funcs[7][7][4] =
     {
-        absDiffScalarImpl<uchar, float>,
-        absDiffScalarImpl<schar, float>,
-        absDiffScalarImpl<ushort, float>,
-        absDiffScalarImpl<short, float>,
-        absDiffScalarImpl<int, float>,
-        absDiffScalarImpl<float, float>,
-        absDiffScalarImpl<double, double>
+        {
+            {absDiffScalarImpl<uchar, float, uchar>, absDiffScalarImpl<uchar2, float, uchar2>, absDiffScalarImpl<uchar3, float, uchar3>, absDiffScalarImpl<uchar4, float, uchar4>},
+            {absDiffScalarImpl<uchar, float, schar>, absDiffScalarImpl<uchar2, float, char2>, absDiffScalarImpl<uchar3, float, char3>, absDiffScalarImpl<uchar4, float, char4>},
+            {absDiffScalarImpl<uchar, float, ushort>, absDiffScalarImpl<uchar2, float, ushort2>, absDiffScalarImpl<uchar3, float, ushort3>, absDiffScalarImpl<uchar4, float, ushort4>},
+            {absDiffScalarImpl<uchar, float, short>, absDiffScalarImpl<uchar2, float, short2>, absDiffScalarImpl<uchar3, float, short3>, absDiffScalarImpl<uchar4, float, short4>},
+            {absDiffScalarImpl<uchar, float, int>, absDiffScalarImpl<uchar2, float, int2>, absDiffScalarImpl<uchar3, float, int3>, absDiffScalarImpl<uchar4, float, int4>},
+            {absDiffScalarImpl<uchar, float, float>, absDiffScalarImpl<uchar2, float, float2>, absDiffScalarImpl<uchar3, float, float3>, absDiffScalarImpl<uchar4, float, float4>},
+            {absDiffScalarImpl<uchar, double, double>, absDiffScalarImpl<uchar2, double, double2>, absDiffScalarImpl<uchar3, double, double3>, absDiffScalarImpl<uchar4, double, double4>}
+        },
+        {
+            {absDiffScalarImpl<schar, float, uchar>, absDiffScalarImpl<char2, float, uchar2>, absDiffScalarImpl<char3, float, uchar3>, absDiffScalarImpl<char4, float, uchar4>},
+            {absDiffScalarImpl<schar, float, schar>, absDiffScalarImpl<char2, float, char2>, absDiffScalarImpl<char3, float, char3>, absDiffScalarImpl<char4, float, char4>},
+            {absDiffScalarImpl<schar, float, ushort>, absDiffScalarImpl<char2, float, ushort2>, absDiffScalarImpl<char3, float, ushort3>, absDiffScalarImpl<char4, float, ushort4>},
+            {absDiffScalarImpl<schar, float, short>, absDiffScalarImpl<char2, float, short2>, absDiffScalarImpl<char3, float, short3>, absDiffScalarImpl<char4, float, short4>},
+            {absDiffScalarImpl<schar, float, int>, absDiffScalarImpl<char2, float, int2>, absDiffScalarImpl<char3, float, int3>, absDiffScalarImpl<char4, float, int4>},
+            {absDiffScalarImpl<schar, float, float>, absDiffScalarImpl<char2, float, float2>, absDiffScalarImpl<char3, float, float3>, absDiffScalarImpl<char4, float, float4>},
+            {absDiffScalarImpl<schar, double, double>, absDiffScalarImpl<char2, double, double2>, absDiffScalarImpl<char3, double, double3>, absDiffScalarImpl<char4, double, double4>}
+        },
+        {
+            {0 /*absDiffScalarImpl<ushort, float, uchar>*/, 0 /*absDiffScalarImpl<ushort2, float, uchar2>*/, 0 /*absDiffScalarImpl<ushort3, float, uchar3>*/, 0 /*absDiffScalarImpl<ushort4, float, uchar4>*/},
+            {0 /*absDiffScalarImpl<ushort, float, schar>*/, 0 /*absDiffScalarImpl<ushort2, float, char2>*/, 0 /*absDiffScalarImpl<ushort3, float, char3>*/, 0 /*absDiffScalarImpl<ushort4, float, char4>*/},
+            {absDiffScalarImpl<ushort, float, ushort>, absDiffScalarImpl<ushort2, float, ushort2>, absDiffScalarImpl<ushort3, float, ushort3>, absDiffScalarImpl<ushort4, float, ushort4>},
+            {absDiffScalarImpl<ushort, float, short>, absDiffScalarImpl<ushort2, float, short2>, absDiffScalarImpl<ushort3, float, short3>, absDiffScalarImpl<ushort4, float, short4>},
+            {absDiffScalarImpl<ushort, float, int>, absDiffScalarImpl<ushort2, float, int2>, absDiffScalarImpl<ushort3, float, int3>, absDiffScalarImpl<ushort4, float, int4>},
+            {absDiffScalarImpl<ushort, float, float>, absDiffScalarImpl<ushort2, float, float2>, absDiffScalarImpl<ushort3, float, float3>, absDiffScalarImpl<ushort4, float, float4>},
+            {absDiffScalarImpl<ushort, double, double>, absDiffScalarImpl<ushort2, double, double2>, absDiffScalarImpl<ushort3, double, double3>, absDiffScalarImpl<ushort4, double, double4>}
+        },
+        {
+            {0 /*absDiffScalarImpl<short, float, uchar>*/, 0 /*absDiffScalarImpl<short2, float, uchar2>*/, 0 /*absDiffScalarImpl<short3, float, uchar3>*/, 0 /*absDiffScalarImpl<short4, float, uchar4>*/},
+            {0 /*absDiffScalarImpl<short, float, schar>*/, 0 /*absDiffScalarImpl<short2, float, char2>*/, 0 /*absDiffScalarImpl<short3, float, char3>*/, 0 /*absDiffScalarImpl<short4, float, char4>*/},
+            {absDiffScalarImpl<short, float, ushort>, absDiffScalarImpl<short2, float, ushort2>, absDiffScalarImpl<short3, float, ushort3>, absDiffScalarImpl<short4, float, ushort4>},
+            {absDiffScalarImpl<short, float, short>, absDiffScalarImpl<short2, float, short2>, absDiffScalarImpl<short3, float, short3>, absDiffScalarImpl<short4, float, short4>},
+            {absDiffScalarImpl<short, float, int>, absDiffScalarImpl<short2, float, int2>, absDiffScalarImpl<short3, float, int3>, absDiffScalarImpl<short4, float, int4>},
+            {absDiffScalarImpl<short, float, float>, absDiffScalarImpl<short2, float, float2>, absDiffScalarImpl<short3, float, float3>, absDiffScalarImpl<short4, float, float4>},
+            {absDiffScalarImpl<short, double, double>, absDiffScalarImpl<short2, double, double2>, absDiffScalarImpl<short3, double, double3>, absDiffScalarImpl<short4, double, double4>}
+        },
+        {
+            {0 /*absDiffScalarImpl<int, float, uchar>*/, 0 /*absDiffScalarImpl<int2, float, uchar2>*/, 0 /*absDiffScalarImpl<int3, float, uchar3>*/, 0 /*absDiffScalarImpl<int4, float, uchar4>*/},
+            {0 /*absDiffScalarImpl<int, float, schar>*/, 0 /*absDiffScalarImpl<int2, float, char2>*/, 0 /*absDiffScalarImpl<int3, float, char3>*/, 0 /*absDiffScalarImpl<int4, float, char4>*/},
+            {0 /*absDiffScalarImpl<int, float, ushort>*/, 0 /*absDiffScalarImpl<int2, float, ushort2>*/, 0 /*absDiffScalarImpl<int3, float, ushort3>*/, 0 /*absDiffScalarImpl<int4, float, ushort4>*/},
+            {0 /*absDiffScalarImpl<int, float, short>*/, 0 /*absDiffScalarImpl<int2, float, short2>*/, 0 /*absDiffScalarImpl<int3, float, short3>*/, 0 /*absDiffScalarImpl<int4, float, short4>*/},
+            {absDiffScalarImpl<int, float, int>, absDiffScalarImpl<int2, float, int2>, absDiffScalarImpl<int3, float, int3>, absDiffScalarImpl<int4, float, int4>},
+            {absDiffScalarImpl<int, float, float>, absDiffScalarImpl<int2, float, float2>, absDiffScalarImpl<int3, float, float3>, absDiffScalarImpl<int4, float, float4>},
+            {absDiffScalarImpl<int, double, double>, absDiffScalarImpl<int2, double, double2>, absDiffScalarImpl<int3, double, double3>, absDiffScalarImpl<int4, double, double4>}
+        },
+        {
+            {0 /*absDiffScalarImpl<float, float, uchar>*/, 0 /*absDiffScalarImpl<float2, float, uchar2>*/, 0 /*absDiffScalarImpl<float3, float, uchar3>*/, 0 /*absDiffScalarImpl<float4, float, uchar4>*/},
+            {0 /*absDiffScalarImpl<float, float, schar>*/, 0 /*absDiffScalarImpl<float2, float, char2>*/, 0 /*absDiffScalarImpl<float3, float, char3>*/, 0 /*absDiffScalarImpl<float4, float, char4>*/},
+            {0 /*absDiffScalarImpl<float, float, ushort>*/, 0 /*absDiffScalarImpl<float2, float, ushort2>*/, 0 /*absDiffScalarImpl<float3, float, ushort3>*/, 0 /*absDiffScalarImpl<float4, float, ushort4>*/},
+            {0 /*absDiffScalarImpl<float, float, short>*/, 0 /*absDiffScalarImpl<float2, float, short2>*/, 0 /*absDiffScalarImpl<float3, float, short3>*/, 0 /*absDiffScalarImpl<float4, float, short4>*/},
+            {0 /*absDiffScalarImpl<float, float, int>*/, 0 /*absDiffScalarImpl<float2, float, int2>*/, 0 /*absDiffScalarImpl<float3, float, int3>*/, 0 /*absDiffScalarImpl<float4, float, int4>*/},
+            {absDiffScalarImpl<float, float, float>, absDiffScalarImpl<float2, float, float2>, absDiffScalarImpl<float3, float, float3>, absDiffScalarImpl<float4, float, float4>},
+            {absDiffScalarImpl<float, double, double>, absDiffScalarImpl<float2, double, double2>, absDiffScalarImpl<float3, double, double3>, absDiffScalarImpl<float4, double, double4>}
+        },
+        {
+            {0 /*absDiffScalarImpl<double, double, uchar>*/, 0 /*absDiffScalarImpl<double2, double, uchar2>*/, 0 /*absDiffScalarImpl<double3, double, uchar3>*/, 0 /*absDiffScalarImpl<double4, double, uchar4>*/},
+            {0 /*absDiffScalarImpl<double, double, schar>*/, 0 /*absDiffScalarImpl<double2, double, char2>*/, 0 /*absDiffScalarImpl<double3, double, char3>*/, 0 /*absDiffScalarImpl<double4, double, char4>*/},
+            {0 /*absDiffScalarImpl<double, double, ushort>*/, 0 /*absDiffScalarImpl<double2, double, ushort2>*/, 0 /*absDiffScalarImpl<double3, double, ushort3>*/, 0 /*absDiffScalarImpl<double4, double, ushort4>*/},
+            {0 /*absDiffScalarImpl<double, double, short>*/, 0 /*absDiffScalarImpl<double2, double, short2>*/, 0 /*absDiffScalarImpl<double3, double, short3>*/, 0 /*absDiffScalarImpl<double4, double, short4>*/},
+            {0 /*absDiffScalarImpl<double, double, int>*/, 0 /*absDiffScalarImpl<double2, double, int2>*/, 0 /*absDiffScalarImpl<double3, double, int3>*/, 0 /*absDiffScalarImpl<double4, double, int4>*/},
+            {0 /*absDiffScalarImpl<double, double, float>*/, 0 /*absDiffScalarImpl<double2, double, float2>*/, 0 /*absDiffScalarImpl<double3, double, float3>*/, 0 /*absDiffScalarImpl<double4, double, float4>*/},
+            {absDiffScalarImpl<double, double, double>, absDiffScalarImpl<double2, double, double2>, absDiffScalarImpl<double3, double, double3>, absDiffScalarImpl<double4, double, double4>}
+        }
     };
 
-    const int depth = src.depth();
+    const int sdepth = src.depth();
+    const int ddepth = dst.depth();
+    const int cn = src.channels();
 
-    CV_DbgAssert( depth <= CV_64F );
+    CV_DbgAssert( sdepth <= CV_64F && ddepth <= CV_64F && cn <= 4 );
 
-    funcs[depth](src, val[0], dst, stream);
+    const func_t func = funcs[sdepth][ddepth][cn - 1];
+
+    if (!func)
+        CV_Error(cv::Error::StsUnsupportedFormat, "Unsupported combination of source and destination types");
+
+    func(src, val, dst, stream);
 }
 
 #endif

From 2375e587056b2f44962c1afce4e1aa5f41ba6857 Mon Sep 17 00:00:00 2001
From: chacha21 <pierre@chachatelier.fr>
Date: Mon, 6 Mar 2017 15:08:59 +0100
Subject: [PATCH 2/3] assert SrcDepth == DstDepth

simplify code by requesting srcDepth == dstDepth
---
 modules/cudaarithm/src/cuda/absdiff_scalar.cu | 73 ++++---------------
 1 file changed, 15 insertions(+), 58 deletions(-)

diff --git a/modules/cudaarithm/src/cuda/absdiff_scalar.cu b/modules/cudaarithm/src/cuda/absdiff_scalar.cu
index 225298b291..be3046b3cb 100644
--- a/modules/cudaarithm/src/cuda/absdiff_scalar.cu
+++ b/modules/cudaarithm/src/cuda/absdiff_scalar.cu
@@ -56,14 +56,14 @@ void absDiffScalar(const GpuMat& src, cv::Scalar val, bool, GpuMat& dst, const G
 
 namespace
 {
-    template <typename SrcType, typename ScalarType, typename DstType> struct AbsDiffScalarOp : unary_function<SrcType, DstType>
+    template <typename SrcType, typename ScalarType> struct AbsDiffScalarOp : unary_function<SrcType, SrcType>
     {
         ScalarType val;
 
-        __device__ __forceinline__ DstType operator ()(SrcType a) const
+        __device__ __forceinline__ SrcType operator ()(SrcType a) const
         {
             abs_func<ScalarType> f;
-            return saturate_cast<DstType>(f(saturate_cast<ScalarType>(a) - val));
+            return saturate_cast<SrcType>(f(saturate_cast<ScalarType>(a) - val));
         }
     };
 
@@ -77,14 +77,14 @@ namespace
         };
     };
 
-    template <typename SrcType, typename ScalarDepth, typename DstType>
+    template <typename SrcType, typename ScalarDepth>
     void absDiffScalarImpl(const GpuMat& src, cv::Scalar value, GpuMat& dst, Stream& stream)
     {
         typedef typename MakeVec<ScalarDepth, VecTraits<SrcType>::cn>::type ScalarType;
 
         cv::Scalar_<ScalarDepth> value_ = value;
 
-        AbsDiffScalarOp<SrcType, ScalarType, DstType> op;
+        AbsDiffScalarOp<SrcType, ScalarType> op;
         op.val = VecTraits<ScalarType>::make(value_.val);
         gridTransformUnary_< TransformPolicy<ScalarDepth> >(globPtr<SrcType>(src), globPtr<SrcType>(dst), op, stream);
     }
@@ -93,70 +93,28 @@ namespace
 void absDiffScalar(const GpuMat& src, cv::Scalar val, bool, GpuMat& dst, const GpuMat&, double, Stream& stream, int)
 {
     typedef void (*func_t)(const GpuMat& src, cv::Scalar val, GpuMat& dst, Stream& stream);
-    static const func_t funcs[7][7][4] =
+    static const func_t funcs[7][4] =
     {
         {
-            {absDiffScalarImpl<uchar, float, uchar>, absDiffScalarImpl<uchar2, float, uchar2>, absDiffScalarImpl<uchar3, float, uchar3>, absDiffScalarImpl<uchar4, float, uchar4>},
-            {absDiffScalarImpl<uchar, float, schar>, absDiffScalarImpl<uchar2, float, char2>, absDiffScalarImpl<uchar3, float, char3>, absDiffScalarImpl<uchar4, float, char4>},
-            {absDiffScalarImpl<uchar, float, ushort>, absDiffScalarImpl<uchar2, float, ushort2>, absDiffScalarImpl<uchar3, float, ushort3>, absDiffScalarImpl<uchar4, float, ushort4>},
-            {absDiffScalarImpl<uchar, float, short>, absDiffScalarImpl<uchar2, float, short2>, absDiffScalarImpl<uchar3, float, short3>, absDiffScalarImpl<uchar4, float, short4>},
-            {absDiffScalarImpl<uchar, float, int>, absDiffScalarImpl<uchar2, float, int2>, absDiffScalarImpl<uchar3, float, int3>, absDiffScalarImpl<uchar4, float, int4>},
-            {absDiffScalarImpl<uchar, float, float>, absDiffScalarImpl<uchar2, float, float2>, absDiffScalarImpl<uchar3, float, float3>, absDiffScalarImpl<uchar4, float, float4>},
-            {absDiffScalarImpl<uchar, double, double>, absDiffScalarImpl<uchar2, double, double2>, absDiffScalarImpl<uchar3, double, double3>, absDiffScalarImpl<uchar4, double, double4>}
+            absDiffScalarImpl<uchar, double>, absDiffScalarImpl<uchar2, double>, absDiffScalarImpl<uchar3, double>, absDiffScalarImpl<uchar4, double>
         },
         {
-            {absDiffScalarImpl<schar, float, uchar>, absDiffScalarImpl<char2, float, uchar2>, absDiffScalarImpl<char3, float, uchar3>, absDiffScalarImpl<char4, float, uchar4>},
-            {absDiffScalarImpl<schar, float, schar>, absDiffScalarImpl<char2, float, char2>, absDiffScalarImpl<char3, float, char3>, absDiffScalarImpl<char4, float, char4>},
-            {absDiffScalarImpl<schar, float, ushort>, absDiffScalarImpl<char2, float, ushort2>, absDiffScalarImpl<char3, float, ushort3>, absDiffScalarImpl<char4, float, ushort4>},
-            {absDiffScalarImpl<schar, float, short>, absDiffScalarImpl<char2, float, short2>, absDiffScalarImpl<char3, float, short3>, absDiffScalarImpl<char4, float, short4>},
-            {absDiffScalarImpl<schar, float, int>, absDiffScalarImpl<char2, float, int2>, absDiffScalarImpl<char3, float, int3>, absDiffScalarImpl<char4, float, int4>},
-            {absDiffScalarImpl<schar, float, float>, absDiffScalarImpl<char2, float, float2>, absDiffScalarImpl<char3, float, float3>, absDiffScalarImpl<char4, float, float4>},
-            {absDiffScalarImpl<schar, double, double>, absDiffScalarImpl<char2, double, double2>, absDiffScalarImpl<char3, double, double3>, absDiffScalarImpl<char4, double, double4>}
+            absDiffScalarImpl<schar, double>, absDiffScalarImpl<char2, double>, absDiffScalarImpl<char3, double>, absDiffScalarImpl<char4, double>
         },
         {
-            {0 /*absDiffScalarImpl<ushort, float, uchar>*/, 0 /*absDiffScalarImpl<ushort2, float, uchar2>*/, 0 /*absDiffScalarImpl<ushort3, float, uchar3>*/, 0 /*absDiffScalarImpl<ushort4, float, uchar4>*/},
-            {0 /*absDiffScalarImpl<ushort, float, schar>*/, 0 /*absDiffScalarImpl<ushort2, float, char2>*/, 0 /*absDiffScalarImpl<ushort3, float, char3>*/, 0 /*absDiffScalarImpl<ushort4, float, char4>*/},
-            {absDiffScalarImpl<ushort, float, ushort>, absDiffScalarImpl<ushort2, float, ushort2>, absDiffScalarImpl<ushort3, float, ushort3>, absDiffScalarImpl<ushort4, float, ushort4>},
-            {absDiffScalarImpl<ushort, float, short>, absDiffScalarImpl<ushort2, float, short2>, absDiffScalarImpl<ushort3, float, short3>, absDiffScalarImpl<ushort4, float, short4>},
-            {absDiffScalarImpl<ushort, float, int>, absDiffScalarImpl<ushort2, float, int2>, absDiffScalarImpl<ushort3, float, int3>, absDiffScalarImpl<ushort4, float, int4>},
-            {absDiffScalarImpl<ushort, float, float>, absDiffScalarImpl<ushort2, float, float2>, absDiffScalarImpl<ushort3, float, float3>, absDiffScalarImpl<ushort4, float, float4>},
-            {absDiffScalarImpl<ushort, double, double>, absDiffScalarImpl<ushort2, double, double2>, absDiffScalarImpl<ushort3, double, double3>, absDiffScalarImpl<ushort4, double, double4>}
+            absDiffScalarImpl<ushort, double>, absDiffScalarImpl<ushort2, double>, absDiffScalarImpl<ushort3, double>, absDiffScalarImpl<ushort4, double>
         },
         {
-            {0 /*absDiffScalarImpl<short, float, uchar>*/, 0 /*absDiffScalarImpl<short2, float, uchar2>*/, 0 /*absDiffScalarImpl<short3, float, uchar3>*/, 0 /*absDiffScalarImpl<short4, float, uchar4>*/},
-            {0 /*absDiffScalarImpl<short, float, schar>*/, 0 /*absDiffScalarImpl<short2, float, char2>*/, 0 /*absDiffScalarImpl<short3, float, char3>*/, 0 /*absDiffScalarImpl<short4, float, char4>*/},
-            {absDiffScalarImpl<short, float, ushort>, absDiffScalarImpl<short2, float, ushort2>, absDiffScalarImpl<short3, float, ushort3>, absDiffScalarImpl<short4, float, ushort4>},
-            {absDiffScalarImpl<short, float, short>, absDiffScalarImpl<short2, float, short2>, absDiffScalarImpl<short3, float, short3>, absDiffScalarImpl<short4, float, short4>},
-            {absDiffScalarImpl<short, float, int>, absDiffScalarImpl<short2, float, int2>, absDiffScalarImpl<short3, float, int3>, absDiffScalarImpl<short4, float, int4>},
-            {absDiffScalarImpl<short, float, float>, absDiffScalarImpl<short2, float, float2>, absDiffScalarImpl<short3, float, float3>, absDiffScalarImpl<short4, float, float4>},
-            {absDiffScalarImpl<short, double, double>, absDiffScalarImpl<short2, double, double2>, absDiffScalarImpl<short3, double, double3>, absDiffScalarImpl<short4, double, double4>}
+            absDiffScalarImpl<short, double>, absDiffScalarImpl<short2, double>, absDiffScalarImpl<short3, double>, absDiffScalarImpl<short4, double>
         },
         {
-            {0 /*absDiffScalarImpl<int, float, uchar>*/, 0 /*absDiffScalarImpl<int2, float, uchar2>*/, 0 /*absDiffScalarImpl<int3, float, uchar3>*/, 0 /*absDiffScalarImpl<int4, float, uchar4>*/},
-            {0 /*absDiffScalarImpl<int, float, schar>*/, 0 /*absDiffScalarImpl<int2, float, char2>*/, 0 /*absDiffScalarImpl<int3, float, char3>*/, 0 /*absDiffScalarImpl<int4, float, char4>*/},
-            {0 /*absDiffScalarImpl<int, float, ushort>*/, 0 /*absDiffScalarImpl<int2, float, ushort2>*/, 0 /*absDiffScalarImpl<int3, float, ushort3>*/, 0 /*absDiffScalarImpl<int4, float, ushort4>*/},
-            {0 /*absDiffScalarImpl<int, float, short>*/, 0 /*absDiffScalarImpl<int2, float, short2>*/, 0 /*absDiffScalarImpl<int3, float, short3>*/, 0 /*absDiffScalarImpl<int4, float, short4>*/},
-            {absDiffScalarImpl<int, float, int>, absDiffScalarImpl<int2, float, int2>, absDiffScalarImpl<int3, float, int3>, absDiffScalarImpl<int4, float, int4>},
-            {absDiffScalarImpl<int, float, float>, absDiffScalarImpl<int2, float, float2>, absDiffScalarImpl<int3, float, float3>, absDiffScalarImpl<int4, float, float4>},
-            {absDiffScalarImpl<int, double, double>, absDiffScalarImpl<int2, double, double2>, absDiffScalarImpl<int3, double, double3>, absDiffScalarImpl<int4, double, double4>}
+            absDiffScalarImpl<int, double>, absDiffScalarImpl<int2, double>, absDiffScalarImpl<int3, double>, absDiffScalarImpl<int4, double>
         },
         {
-            {0 /*absDiffScalarImpl<float, float, uchar>*/, 0 /*absDiffScalarImpl<float2, float, uchar2>*/, 0 /*absDiffScalarImpl<float3, float, uchar3>*/, 0 /*absDiffScalarImpl<float4, float, uchar4>*/},
-            {0 /*absDiffScalarImpl<float, float, schar>*/, 0 /*absDiffScalarImpl<float2, float, char2>*/, 0 /*absDiffScalarImpl<float3, float, char3>*/, 0 /*absDiffScalarImpl<float4, float, char4>*/},
-            {0 /*absDiffScalarImpl<float, float, ushort>*/, 0 /*absDiffScalarImpl<float2, float, ushort2>*/, 0 /*absDiffScalarImpl<float3, float, ushort3>*/, 0 /*absDiffScalarImpl<float4, float, ushort4>*/},
-            {0 /*absDiffScalarImpl<float, float, short>*/, 0 /*absDiffScalarImpl<float2, float, short2>*/, 0 /*absDiffScalarImpl<float3, float, short3>*/, 0 /*absDiffScalarImpl<float4, float, short4>*/},
-            {0 /*absDiffScalarImpl<float, float, int>*/, 0 /*absDiffScalarImpl<float2, float, int2>*/, 0 /*absDiffScalarImpl<float3, float, int3>*/, 0 /*absDiffScalarImpl<float4, float, int4>*/},
-            {absDiffScalarImpl<float, float, float>, absDiffScalarImpl<float2, float, float2>, absDiffScalarImpl<float3, float, float3>, absDiffScalarImpl<float4, float, float4>},
-            {absDiffScalarImpl<float, double, double>, absDiffScalarImpl<float2, double, double2>, absDiffScalarImpl<float3, double, double3>, absDiffScalarImpl<float4, double, double4>}
+          absDiffScalarImpl<float, double>, absDiffScalarImpl<float2, double>, absDiffScalarImpl<float3, double>, absDiffScalarImpl<float4, double>
         },
         {
-            {0 /*absDiffScalarImpl<double, double, uchar>*/, 0 /*absDiffScalarImpl<double2, double, uchar2>*/, 0 /*absDiffScalarImpl<double3, double, uchar3>*/, 0 /*absDiffScalarImpl<double4, double, uchar4>*/},
-            {0 /*absDiffScalarImpl<double, double, schar>*/, 0 /*absDiffScalarImpl<double2, double, char2>*/, 0 /*absDiffScalarImpl<double3, double, char3>*/, 0 /*absDiffScalarImpl<double4, double, char4>*/},
-            {0 /*absDiffScalarImpl<double, double, ushort>*/, 0 /*absDiffScalarImpl<double2, double, ushort2>*/, 0 /*absDiffScalarImpl<double3, double, ushort3>*/, 0 /*absDiffScalarImpl<double4, double, ushort4>*/},
-            {0 /*absDiffScalarImpl<double, double, short>*/, 0 /*absDiffScalarImpl<double2, double, short2>*/, 0 /*absDiffScalarImpl<double3, double, short3>*/, 0 /*absDiffScalarImpl<double4, double, short4>*/},
-            {0 /*absDiffScalarImpl<double, double, int>*/, 0 /*absDiffScalarImpl<double2, double, int2>*/, 0 /*absDiffScalarImpl<double3, double, int3>*/, 0 /*absDiffScalarImpl<double4, double, int4>*/},
-            {0 /*absDiffScalarImpl<double, double, float>*/, 0 /*absDiffScalarImpl<double2, double, float2>*/, 0 /*absDiffScalarImpl<double3, double, float3>*/, 0 /*absDiffScalarImpl<double4, double, float4>*/},
-            {absDiffScalarImpl<double, double, double>, absDiffScalarImpl<double2, double, double2>, absDiffScalarImpl<double3, double, double3>, absDiffScalarImpl<double4, double, double4>}
+          absDiffScalarImpl<double, double>, absDiffScalarImpl<double2, double>, absDiffScalarImpl<double3, double>, absDiffScalarImpl<double4, double>
         }
     };
 
@@ -164,10 +122,9 @@ void absDiffScalar(const GpuMat& src, cv::Scalar val, bool, GpuMat& dst, const G
     const int ddepth = dst.depth();
     const int cn = src.channels();
 
-    CV_DbgAssert( sdepth <= CV_64F && ddepth <= CV_64F && cn <= 4 );
-
-    const func_t func = funcs[sdepth][ddepth][cn - 1];
+    CV_DbgAssert( sdepth <= CV_64F && ddepth <= CV_64F && cn <= 4 && src.type() == dst.type());
 
+    const func_t func = funcs[sdepth][cn - 1];
     if (!func)
         CV_Error(cv::Error::StsUnsupportedFormat, "Unsupported combination of source and destination types");
 

From bfd800342bfb1900bc2b1cdb0e94fbbff88311f2 Mon Sep 17 00:00:00 2001
From: chacha21 <pierre@chachatelier.fr>
Date: Tue, 7 Mar 2017 09:49:49 +0100
Subject: [PATCH 3/3] use "double" from scalar only for CV_64F input

Restore "DstType" in AbsDiffScalarOp template arguments, even if
eventually it will be equal to SrcType

use "double" from scalar only for CV_64F input
---
 modules/cudaarithm/src/cuda/absdiff_scalar.cu | 20 +++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/modules/cudaarithm/src/cuda/absdiff_scalar.cu b/modules/cudaarithm/src/cuda/absdiff_scalar.cu
index be3046b3cb..01770a5f23 100644
--- a/modules/cudaarithm/src/cuda/absdiff_scalar.cu
+++ b/modules/cudaarithm/src/cuda/absdiff_scalar.cu
@@ -56,14 +56,14 @@ void absDiffScalar(const GpuMat& src, cv::Scalar val, bool, GpuMat& dst, const G
 
 namespace
 {
-    template <typename SrcType, typename ScalarType> struct AbsDiffScalarOp : unary_function<SrcType, SrcType>
+    template <typename SrcType, typename ScalarType, typename DstType> struct AbsDiffScalarOp : unary_function<SrcType, DstType>
     {
         ScalarType val;
 
-        __device__ __forceinline__ SrcType operator ()(SrcType a) const
+        __device__ __forceinline__ DstType operator ()(SrcType a) const
         {
             abs_func<ScalarType> f;
-            return saturate_cast<SrcType>(f(saturate_cast<ScalarType>(a) - val));
+            return saturate_cast<DstType>(f(saturate_cast<ScalarType>(a) - val));
         }
     };
 
@@ -84,7 +84,7 @@ namespace
 
         cv::Scalar_<ScalarDepth> value_ = value;
 
-        AbsDiffScalarOp<SrcType, ScalarType> op;
+        AbsDiffScalarOp<SrcType, ScalarType, SrcType> op;
         op.val = VecTraits<ScalarType>::make(value_.val);
         gridTransformUnary_< TransformPolicy<ScalarDepth> >(globPtr<SrcType>(src), globPtr<SrcType>(dst), op, stream);
     }
@@ -96,22 +96,22 @@ void absDiffScalar(const GpuMat& src, cv::Scalar val, bool, GpuMat& dst, const G
     static const func_t funcs[7][4] =
     {
         {
-            absDiffScalarImpl<uchar, double>, absDiffScalarImpl<uchar2, double>, absDiffScalarImpl<uchar3, double>, absDiffScalarImpl<uchar4, double>
+            absDiffScalarImpl<uchar, float>, absDiffScalarImpl<uchar2, float>, absDiffScalarImpl<uchar3, float>, absDiffScalarImpl<uchar4, float>
         },
         {
-            absDiffScalarImpl<schar, double>, absDiffScalarImpl<char2, double>, absDiffScalarImpl<char3, double>, absDiffScalarImpl<char4, double>
+            absDiffScalarImpl<schar, float>, absDiffScalarImpl<char2, float>, absDiffScalarImpl<char3, float>, absDiffScalarImpl<char4, float>
         },
         {
-            absDiffScalarImpl<ushort, double>, absDiffScalarImpl<ushort2, double>, absDiffScalarImpl<ushort3, double>, absDiffScalarImpl<ushort4, double>
+            absDiffScalarImpl<ushort, float>, absDiffScalarImpl<ushort2, float>, absDiffScalarImpl<ushort3, float>, absDiffScalarImpl<ushort4, float>
         },
         {
-            absDiffScalarImpl<short, double>, absDiffScalarImpl<short2, double>, absDiffScalarImpl<short3, double>, absDiffScalarImpl<short4, double>
+            absDiffScalarImpl<short, float>, absDiffScalarImpl<short2, float>, absDiffScalarImpl<short3, float>, absDiffScalarImpl<short4, float>
         },
         {
-            absDiffScalarImpl<int, double>, absDiffScalarImpl<int2, double>, absDiffScalarImpl<int3, double>, absDiffScalarImpl<int4, double>
+            absDiffScalarImpl<int, float>, absDiffScalarImpl<int2, float>, absDiffScalarImpl<int3, float>, absDiffScalarImpl<int4, float>
         },
         {
-          absDiffScalarImpl<float, double>, absDiffScalarImpl<float2, double>, absDiffScalarImpl<float3, double>, absDiffScalarImpl<float4, double>
+          absDiffScalarImpl<float, float>, absDiffScalarImpl<float2, float>, absDiffScalarImpl<float3, float>, absDiffScalarImpl<float4, float>
         },
         {
           absDiffScalarImpl<double, double>, absDiffScalarImpl<double2, double>, absDiffScalarImpl<double3, double>, absDiffScalarImpl<double4, double>