GpuMat::setTo optimization (call cudaMemcpy2D if possible)

diff --git a/modules/gpu/src/matrix_operations.cpp b/modules/gpu/src/matrix_operations.cpp
index e1c34dd..e4d5059 100644 (file)
--- a/modules/gpu/src/matrix_operations.cpp
+++ b/modules/gpu/src/matrix_operations.cpp
@@ -442,6 +442,22 @@ GpuMat& GpuMat::setTo(const Scalar& s, const GpuMat& mask)
 \r
     if (mask.empty())\r
     {\r
+        if (s[0] == 0.0 && s[1] == 0.0 && s[2] == 0.0 && s[3] == 0.0)\r
+        {\r
+            cudaSafeCall( cudaMemset2D(data, step, 0, cols * elemSize(), rows) );\r
+            return *this;\r
+        }\r
+        if (depth() == CV_8U)\r
+        {\r
+            int cn = channels();\r
+\r
+            if (cn == 1 || (cn == 2 && s[0] == s[1]) || (cn == 3 && s[0] == s[1] && s[0] == s[2]) || (cn == 4 && s[0] == s[1] && s[0] == s[2] && s[0] == s[3]))\r
+            {\r
+                int val = saturate_cast<uchar>(s[0]);\r
+                cudaSafeCall( cudaMemset2D(data, step, val, cols, rows) );\r
+                return *this;\r
+            }\r
+        }\r
         typedef void (*set_caller_t)(GpuMat& src, const Scalar& s);\r
         static const set_caller_t set_callers[8][4] =\r
         {\r