more fix of mismatch

modules/ocl/src/opencl/match_template.cl

@@ -45,22 +45,28 @@
 
 #pragma OPENCL EXTENSION cl_amd_printf : enable
 
-#if defined (__ATI__)
-#pragma OPENCL EXTENSION cl_amd_fp64:enable
+#if defined (DOUBLE_SUPPORT)
 
-#elif defined (__NVIDIA__)
+#ifdef cl_khr_fp64
 #pragma OPENCL EXTENSION cl_khr_fp64:enable
+#elif defined (cl_amd_fp64)
+#pragma OPENCL EXTENSION cl_amd_fp64:enable
 #endif
 
-#if !defined(USE_SQR_INTEGRAL) && (defined (__ATI__) || defined (__NVIDIA__))
 #define TYPE_IMAGE_SQSUM double
 #else
-#define TYPE_IMAGE_SQSUM ulong
+#define TYPE_IMAGE_SQSUM float
+#endif
+
+#ifndef CN4
+#define CN4 1
+#else
+#define CN4 4
 #endif
 
 //////////////////////////////////////////////////
 // utilities
-#define SQSUMS_PTR(ox, oy) mad24(gidy + oy, img_sqsums_step, gidx + img_sqsums_offset + ox)
+#define SQSUMS_PTR(ox, oy) mad24(gidy + oy, img_sqsums_step, (gidx + img_sqsums_offset + ox) * CN4)
 #define SUMS_PTR(ox, oy) mad24(gidy + oy, img_sums_step, gidx + img_sums_offset + ox)
 // normAcc* are accurate normalization routines which make GPU matchTemplate
 // consistent with CPU one
@@ -95,7 +101,7 @@ float normAcc_SQDIFF(float num, float denum)
 __kernel
 void normalizeKernel_C1_D0
 (
-    __global const TYPE_IMAGE_SQSUM * img_sqsums,
+    __global const float * img_sqsums,
     __global float * res,
     ulong tpl_sqsum,
     int res_rows,
@@ -161,7 +167,7 @@ void matchTemplate_Prepared_SQDIFF_C1_D0
 __kernel
 void matchTemplate_Prepared_SQDIFF_NORMED_C1_D0
 (
-    __global const TYPE_IMAGE_SQSUM * img_sqsums,
+    __global const float * img_sqsums,
     __global float * res,
     ulong tpl_sqsum,
     int res_rows,
@@ -702,7 +708,7 @@ void matchTemplate_Prepared_CCOFF_NORMED_C1_D0
     __global const uint * img_sums,
     int img_sums_offset,
     int img_sums_step,
-    __global const TYPE_IMAGE_SQSUM * img_sqsums,
+    __global const float * img_sqsums,
     int img_sqsums_offset,
     int img_sqsums_step,
     float tpl_sum,
@@ -754,10 +760,10 @@ void matchTemplate_Prepared_CCOFF_NORMED_C4_D0
     __global const uint * img_sums_c3,
     int img_sums_offset,
     int img_sums_step,
-    __global const TYPE_IMAGE_SQSUM * img_sqsums_c0,
-    __global const TYPE_IMAGE_SQSUM * img_sqsums_c1,
-    __global const TYPE_IMAGE_SQSUM * img_sqsums_c2,
-    __global const TYPE_IMAGE_SQSUM * img_sqsums_c3,
+    __global const float * img_sqsums_c0,
+    __global const float * img_sqsums_c1,
+    __global const float * img_sqsums_c2,
+    __global const float * img_sqsums_c3,
     int img_sqsums_offset,
     int img_sqsums_step,
     float tpl_sum_c0,
@@ -821,3 +827,32 @@ void matchTemplate_Prepared_CCOFF_NORMED_C4_D0
         res[res_idx] = normAcc(num, denum);
     }
 }
+
+//////////////////////////////////////////////////////////////////////
+// extractFirstChannel
+__kernel
+void extractFirstChannel
+(
+    const __global float4* img,
+    __global float* res,
+    int rows,
+    int cols,
+    int img_offset,
+    int res_offset,
+    int img_step,
+    int res_step
+)
+{
+    img_step   /= sizeof(float4);
+    res_step   /= sizeof(float);
+    img_offset /= sizeof(float4);
+    res_offset /= sizeof(float);
+    img += img_offset;
+    res += res_offset;
+    int gidx = get_global_id(0);
+    int gidy = get_global_id(1);
+    if(gidx < cols && gidy < rows)
+    {
+        res[gidx + gidy * res_step] = img[gidx + gidy * img_step].x;
+    }
+}

modules/ocl/test/test_match_template.cpp

@@ -75,7 +75,7 @@ PARAM_TEST_CASE(MatchTemplate8U, cv::Size, TemplateSize, Channels, TemplateMetho
     }
 };
 
-TEST_P(MatchTemplate8U, DISABLED_Accuracy)
+TEST_P(MatchTemplate8U, Accuracy)
 {
 
     std::cout << "Method: " << TEMPLATE_METHOD_NAMES[method] << std::endl;
@@ -138,18 +138,18 @@ TEST_P(MatchTemplate32F, Accuracy)
     EXPECT_MAT_NEAR(dst_gold, mat_dst, templ_size.area() * 1e-1, sss);
 }
 
-INSTANTIATE_TEST_CASE_P(GPU_ImgProc, MatchTemplate8U,
+INSTANTIATE_TEST_CASE_P(OCL_ImgProc, MatchTemplate8U,
                         testing::Combine(
                             MTEMP_SIZES,
-                            testing::Values(TemplateSize(cv::Size(5, 5)), TemplateSize(cv::Size(16, 16))/*, TemplateSize(cv::Size(30, 30))*/),
+                            testing::Values(TemplateSize(cv::Size(5, 5)), TemplateSize(cv::Size(16, 16)), TemplateSize(cv::Size(30, 30))),
                             testing::Values(Channels(1), Channels(3), Channels(4)),
                             ALL_TEMPLATE_METHODS
                         )
                        );
 
-INSTANTIATE_TEST_CASE_P(GPU_ImgProc, MatchTemplate32F, testing::Combine(
+INSTANTIATE_TEST_CASE_P(OCL_ImgProc, MatchTemplate32F, testing::Combine(
                             MTEMP_SIZES,
-                            testing::Values(TemplateSize(cv::Size(5, 5)), TemplateSize(cv::Size(16, 16))/*, TemplateSize(cv::Size(30, 30))*/),
+                            testing::Values(TemplateSize(cv::Size(5, 5)), TemplateSize(cv::Size(16, 16)), TemplateSize(cv::Size(30, 30))),
                             testing::Values(Channels(1), Channels(3), Channels(4)),
                             testing::Values(TemplateMethod(cv::TM_SQDIFF), TemplateMethod(cv::TM_CCORR))));
 #endif