Merge pull request #705 from bitwangyaoyao:2.4_oclFix

cmake/OpenCVDetectOpenCL.cmake

@@ -43,7 +43,7 @@ if(OPENCL_FOUND)
   set(OPENCL_LIBRARIES    ${OPENCL_LIBRARY})
 
   if (X86_64)
-    set(CLAMD_POSSIBLE_LIB_SUFFIXES lib32/import)
+    set(CLAMD_POSSIBLE_LIB_SUFFIXES lib64/import)
   elseif (X86)
     set(CLAMD_POSSIBLE_LIB_SUFFIXES lib32/import)
   endif()

cmake/cl2cpp.cmake

@@ -18,6 +18,7 @@ foreach(cl ${cl_list})
   string(REPLACE "\t" "  " lines "${lines}")
 
   string(REGEX REPLACE "/\\*([^*]/|\\*[^/]|[^*/])*\\*/" ""   lines "${lines}") # multiline comments
+  string(REGEX REPLACE "/\\*([^\n])*\\*/"               ""   lines "${lines}") # single-line comments
   string(REGEX REPLACE "[ ]*//[^\n]*\n"                 "\n" lines "${lines}") # single-line comments
   string(REGEX REPLACE "\n[ ]*(\n[ ]*)*"                "\n" lines "${lines}") # empty lines & leading whitespace
   string(REGEX REPLACE "^\n"                            ""   lines "${lines}") # leading new line

modules/ocl/src/haar.cpp

@@ -953,8 +953,8 @@ CvSeq *cv::ocl::OclCascadeClassifier::oclHaarDetectObjects( oclMat &gimg, CvMemS
         //int flag = 0;
 
         oclMat gimg1(gimg.rows, gimg.cols, CV_8UC1);
-        oclMat gsum(totalheight, gimg.cols + 1, CV_32SC1);
-        oclMat gsqsum(totalheight, gimg.cols + 1, CV_32FC1);
+        oclMat gsum(totalheight + 4, gimg.cols + 1, CV_32SC1);
+        oclMat gsqsum(totalheight + 4, gimg.cols + 1, CV_32FC1);
 
         //cl_mem cascadebuffer;
         cl_mem stagebuffer;

modules/ocl/src/moments.cpp

@@ -277,16 +277,7 @@ static void ocl_cvMoments( const void* array, CvMoments* mom, int binary )
         blocky = size.height/TILE_SIZE;
     else
         blocky = size.height/TILE_SIZE + 1;
-    cv::ocl::oclMat dst_m00(blocky, blockx, CV_64FC1);
-    cv::ocl::oclMat dst_m10(blocky, blockx, CV_64FC1);
-    cv::ocl::oclMat dst_m01(blocky, blockx, CV_64FC1);
-    cv::ocl::oclMat dst_m20(blocky, blockx, CV_64FC1);
-    cv::ocl::oclMat dst_m11(blocky, blockx, CV_64FC1);
-    cv::ocl::oclMat dst_m02(blocky, blockx, CV_64FC1);
-    cv::ocl::oclMat dst_m30(blocky, blockx, CV_64FC1);
-    cv::ocl::oclMat dst_m21(blocky, blockx, CV_64FC1);
-    cv::ocl::oclMat dst_m12(blocky, blockx, CV_64FC1);
-    cv::ocl::oclMat dst_m03(blocky, blockx, CV_64FC1);
+    cv::ocl::oclMat dst_m(blocky * 10, blockx, CV_64FC1);
     cl_mem sum = openCLCreateBuffer(src.clCxt,CL_MEM_READ_WRITE,10*sizeof(double));
     int tile_width  = std::min(size.width,TILE_SIZE);
     int tile_height = std::min(size.height,TILE_SIZE);
@@ -299,25 +290,17 @@ static void ocl_cvMoments( const void* array, CvMoments* mom, int binary )
     args.push_back( make_pair( sizeof(cl_int) , (void *)&src.step ));
     args.push_back( make_pair( sizeof(cl_int) , (void *)&tileSize.width ));
     args.push_back( make_pair( sizeof(cl_int) , (void *)&tileSize.height ));
-    args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m00.data ));
-    args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m10.data ));
-    args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m01.data ));
-    args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m20.data ));
-    args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m11.data ));
-    args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m02.data ));
-    args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m30.data ));
-    args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m21.data ));
-    args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m12.data ));
-    args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m03.data ));
-    args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_m00.cols ));
-    args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_m00.step ));
+    args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m.data ));
+    args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_m.cols ));
+    args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_m.step ));
+    args.push_back( make_pair( sizeof(cl_int) , (void *)&blocky ));
     args.push_back( make_pair( sizeof(cl_int) , (void *)&type ));
     args.push_back( make_pair( sizeof(cl_int) , (void *)&depth ));
     args.push_back( make_pair( sizeof(cl_int) , (void *)&cn ));
     args.push_back( make_pair( sizeof(cl_int) , (void *)&coi ));
     args.push_back( make_pair( sizeof(cl_int) , (void *)&binary ));
     args.push_back( make_pair( sizeof(cl_int) , (void *)&TILE_SIZE ));
-    openCLExecuteKernel(dst_m00.clCxt, &moments, "CvMoments", globalThreads, localThreads, args, -1, depth);
+    openCLExecuteKernel(dst_m.clCxt, &moments, "CvMoments", globalThreads, localThreads, args, -1, depth);
 
     size_t localThreadss[3]  = { 128, 1, 1};
     size_t globalThreadss[3] = { 128, 1, 1};
@@ -327,20 +310,12 @@ static void ocl_cvMoments( const void* array, CvMoments* mom, int binary )
     args_sum.push_back( make_pair( sizeof(cl_int) , (void *)&tile_width ));
     args_sum.push_back( make_pair( sizeof(cl_int) , (void *)&TILE_SIZE ));
     args_sum.push_back( make_pair( sizeof(cl_mem) , (void *)&sum ));
-    args_sum.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m00.data ));
-    args_sum.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m10.data ));
-    args_sum.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m01.data ));
-    args_sum.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m20.data ));
-    args_sum.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m11.data ));
-    args_sum.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m02.data ));
-    args_sum.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m30.data ));
-    args_sum.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m21.data ));
-    args_sum.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m12.data ));
-    args_sum.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m03.data ));
-    openCLExecuteKernel(dst_m00.clCxt, &moments, "dst_sum", globalThreadss, localThreadss, args_sum, -1, -1);
+    args_sum.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_m.data ));
+    args_sum.push_back( make_pair( sizeof(cl_int) , (void *)&dst_m.step ));
+    openCLExecuteKernel(dst_m.clCxt, &moments, "dst_sum", globalThreadss, localThreadss, args_sum, -1, -1);
     double* dstsum = new double[10];
     memset(dstsum,0,10*sizeof(double));
-    openCLReadBuffer(dst_m00.clCxt,sum,(void *)dstsum,10*sizeof(double));
+    openCLReadBuffer(dst_m.clCxt,sum,(void *)dstsum,10*sizeof(double));
     mom->m00 = dstsum[0];
     mom->m10 = dstsum[1];
     mom->m01 = dstsum[2];
@@ -351,6 +326,7 @@ static void ocl_cvMoments( const void* array, CvMoments* mom, int binary )
     mom->m21 = dstsum[7];
     mom->m12 = dstsum[8];
     mom->m03 = dstsum[9];
+    delete [] dstsum;
 
     icvCompleteMomentState( mom );
 }

modules/ocl/src/opencl/haarobjectdetect.cl

@@ -211,10 +211,14 @@ __kernel void __attribute__((reqd_work_group_size(8,8,1)))gpuRunHaarClassifierCa
                 int4 data = *(__global int4*)&sum[glb_off];
                 int lcl_off = mad24(lcl_y, readwidth, lcl_x<<2);
 
+#if OFF
                 lcldata[lcl_off] = data.x;
                 lcldata[lcl_off+1] = data.y;
                 lcldata[lcl_off+2] = data.z;
                 lcldata[lcl_off+3] = data.w;
+#else
+                vstore4(data, 0, &lcldata[lcl_off]);
+#endif
             }
 
             lcloutindex[lcl_id] = 0;
@@ -559,3 +563,7 @@ if(result)
 }
 }
 */
+
+
+
+

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/src/opencl/stereobm.cl

@@ -323,7 +323,7 @@ float sobel(__global unsigned char *input, int x, int y, int rows, int cols)
     float conv = 0;
     int y1 = y==0? 0 : y-1;
     int x1 = x==0? 0 : x-1;
-    if(x < cols && y < rows)
+    if(x < cols && y < rows && x > 0 && y > 0)
     {
         conv = (float)input[(y1)  * cols + (x1)] * (-1) + (float)input[(y1)  * cols + (x+1)] * (1) +
                (float)input[(y)   * cols + (x1)] * (-2) + (float)input[(y)   * cols + (x+1)] * (2) +

modules/ocl/test/test_brute_force_matcher.cpp

@@ -110,7 +110,7 @@ namespace
         }
     };
 
-    TEST_P(BruteForceMatcher, DISABLED_Match_Single)
+    TEST_P(BruteForceMatcher, Match_Single)
     {
         cv::ocl::BruteForceMatcher_OCL_base matcher(distType);
 
@@ -130,7 +130,7 @@ namespace
         ASSERT_EQ(0, badCount);
     }
 
-    TEST_P(BruteForceMatcher, DISABLED_KnnMatch_2_Single)
+    TEST_P(BruteForceMatcher, KnnMatch_2_Single)
     {
         const int knn = 2;
 

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