Final refactoring, fixes

modules/core/perf/opencl/perf_arithm.cpp

@@ -292,7 +292,7 @@ OCL_PERF_TEST_P(MagnitudeFixture, Magnitude, ::testing::Combine(
 typedef Size_MatType TransposeFixture;
 
 OCL_PERF_TEST_P(TransposeFixture, Transpose, ::testing::Combine(
-                OCL_TEST_SIZES, Values(CV_8UC1, CV_32FC1, CV_8UC2, CV_32FC2, CV_8UC4, CV_32FC4)))
+                OCL_TEST_SIZES, OCL_TEST_TYPES_134))
 {
     const Size_MatType_t params = GetParam();
     const Size srcSize = get<0>(params);

modules/core/perf/opencl/perf_dxt.cpp

@@ -54,40 +54,21 @@ namespace ocl {
 
 ///////////// dft ////////////////////////
 
-enum OCL_FFT_TYPE
-{
-    R2R = 0, // real to real (CCS)
-    C2R = 1, // complex to real
-    R2C = 2, // real to complex
-    C2C = 3  // complex to complex
-};
-
-typedef tuple<OCL_FFT_TYPE, Size, int> DftParams;
+typedef tuple<Size, int> DftParams;
 typedef TestBaseWithParam<DftParams> DftFixture;
 
-OCL_PERF_TEST_P(DftFixture, Dft, ::testing::Combine(Values(C2C, R2R, C2R, R2C),
-                                                Values(OCL_SIZE_1, OCL_SIZE_2, OCL_SIZE_3, Size(1024, 1024), Size(512, 512), Size(2048, 2048)),
-                                                Values((int) 0, (int)DFT_ROWS, (int)DFT_SCALE/*, (int)DFT_INVERSE,
-                                                       (int)DFT_INVERSE | DFT_SCALE, (int)DFT_ROWS | DFT_INVERSE*/)))
+OCL_PERF_TEST_P(DftFixture, Dft, ::testing::Combine(Values(OCL_SIZE_1, OCL_SIZE_2, OCL_SIZE_3),
+                                                Values((int)DFT_ROWS, (int)DFT_SCALE, (int)DFT_INVERSE,
+                                                       (int)DFT_INVERSE | DFT_SCALE, (int)DFT_ROWS | DFT_INVERSE)))
 {
     const DftParams params = GetParam();
-    const int dft_type = get<0>(params);
-    const Size srcSize = get<1>(params);
-    int flags = get<2>(params);
-    
-    int in_cn, out_cn;
-    switch (dft_type)
-    {
-    case R2R: flags |= cv::DFT_REAL_OUTPUT; in_cn = 1; out_cn = 1; break;
-    case C2R: flags |= cv::DFT_REAL_OUTPUT; in_cn = 2; out_cn = 2; break;
-    case R2C: flags |= cv::DFT_COMPLEX_OUTPUT; in_cn = 1; out_cn = 2; break;
-    case C2C: flags |= cv::DFT_COMPLEX_OUTPUT; in_cn = 2; out_cn = 2; break;
-    }
-
-    UMat src(srcSize, CV_MAKE_TYPE(CV_32F, in_cn)), dst(srcSize, CV_MAKE_TYPE(CV_32F, out_cn));
+    const Size srcSize = get<0>(params);
+    const int flags = get<1>(params);
+
+    UMat src(srcSize, CV_32FC2), dst(srcSize, CV_32FC2);
     declare.in(src, WARMUP_RNG).out(dst);
 
-    OCL_TEST_CYCLE() cv::dft(src, dst, flags);
+    OCL_TEST_CYCLE() cv::dft(src, dst, flags | DFT_COMPLEX_OUTPUT);
 
     SANITY_CHECK(dst, 1e-3);
 }

modules/core/src/ocl.cpp

@@ -3002,8 +3002,7 @@ bool Kernel::run(int dims, size_t _globalsize[], size_t _localsize[],
                                            sync ? 0 : &p->e);
     if( sync || retval != CL_SUCCESS )
     {
-        int a = clFinish(qq);
-        CV_OclDbgAssert(a == CL_SUCCESS);
+        CV_OclDbgAssert(clFinish(qq) == CL_SUCCESS);
         p->cleanupUMats();
     }
     else
@@ -3899,9 +3898,8 @@ public:
         if( (accessFlags & ACCESS_READ) != 0 && u->hostCopyObsolete() )
         {
             AlignedDataPtr<false, true> alignedPtr(u->data, u->size, CV_OPENCL_DATA_PTR_ALIGNMENT);
-            int a = clEnqueueReadBuffer(q, (cl_mem)u->handle, CL_TRUE, 0,
-                                           u->size, alignedPtr.getAlignedPtr(), 0, 0, 0);
-            CV_Assert( a == CL_SUCCESS );
+            CV_Assert( clEnqueueReadBuffer(q, (cl_mem)u->handle, CL_TRUE, 0,
+                                           u->size, alignedPtr.getAlignedPtr(), 0, 0, 0) == CL_SUCCESS );
             u->markHostCopyObsolete(false);
         }
     }

modules/core/src/opencl/fft.cl

@@ -571,10 +571,15 @@ __kernel void fft_multi_radix_rows(__global const uchar* src_ptr, int src_step,
     }
     else
     {
+        // fill with zero other rows
+#ifdef COMPLEX_OUTPUT
         __global float2* dst = (__global float2*)(dst_ptr + mad24(y, dst_step, dst_offset));
+#else
+        __global float* dst = (__global float*)(dst_ptr + mad24(y, dst_step, dst_offset));
+#endif
         #pragma unroll
         for (int i=x; i<dst_cols; i+=block_size)
-            dst[i] = (float2) 0.f;
+            dst[i] = 0.f;
     }
 }
 
@@ -667,12 +672,9 @@ __kernel void ifft_multi_radix_rows(__global const uchar* src_ptr, int src_step,
             smem[x+i*block_size].y = -src[i*block_size].y;
         }
 #else
-    __global const float2* src;
+
     #if !defined(REAL_INPUT) && defined(NO_CONJUGATE)
-        src = (__global const float2*)(src_ptr + mad24(y, src_step, mad24(2, (int)sizeof(float), src_offset)));
-    #else
-        src = (__global const float2*)(src_ptr + mad24(y, src_step, mad24(1, (int)sizeof(float), src_offset)));
-    #endif
+        __global const float2* src = (__global const float2*)(src_ptr + mad24(y, src_step, mad24(2, (int)sizeof(float), src_offset)));
 
         #pragma unroll
         for (int i=x; i<(LOCAL_SIZE-1)/2; i+=block_size)
@@ -681,6 +683,20 @@ __kernel void ifft_multi_radix_rows(__global const uchar* src_ptr, int src_step,
             smem[i+1].y = -src[i].y;
             smem[LOCAL_SIZE-i-1] = src[i];
         }
+    #else
+
+        #pragma unroll
+        for (int i=x; i<(LOCAL_SIZE-1)/2; i+=block_size)
+        {
+            float2 src = vload2(0, (__global const float*)(src_ptr + mad24(y, src_step, mad24(2*i+1, (int)sizeof(float), src_offset))));
+
+            smem[i+1].x = src.x;
+            smem[i+1].y = -src.y;
+            smem[LOCAL_SIZE-i-1] = src;
+        }
+
+    #endif
+
         if (x==0)
         {
             smem[0].x = *(__global const float*)(src_ptr + mad24(y, src_step, src_offset));
@@ -688,7 +704,11 @@ __kernel void ifft_multi_radix_rows(__global const uchar* src_ptr, int src_step,
 
             if(LOCAL_SIZE % 2 ==0)
             {
+                #if !defined(REAL_INPUT) && defined(NO_CONJUGATE)
                 smem[LOCAL_SIZE/2].x = src[LOCAL_SIZE/2-1].x;
+                #else
+                smem[LOCAL_SIZE/2].x = *(__global const float*)(src_ptr + mad24(y, src_step, mad24(LOCAL_SIZE-1, (int)sizeof(float), src_offset)));
+                #endif
                 smem[LOCAL_SIZE/2].y = 0.f;
             }
         }
@@ -718,10 +738,15 @@ __kernel void ifft_multi_radix_rows(__global const uchar* src_ptr, int src_step,
     }
     else
     {
-        __global float2* dst = (__global float*)(dst_ptr + mad24(y, dst_step, mad24(x, (int)(sizeof(float)*2), dst_offset)));
+        // fill with zero other rows
+#ifdef COMPLEX_OUTPUT
+        __global float2* dst = (__global float2*)(dst_ptr + mad24(y, dst_step, dst_offset));
+#else
+        __global float* dst = (__global float*)(dst_ptr + mad24(y, dst_step, dst_offset));
+#endif
         #pragma unroll
-        for (int i=0; i<kercn; i++)
-            dst[i*block_size] = (float2) 0.f;
+        for (int i=x; i<dst_cols; i+=block_size)
+            dst[i] = 0.f;
     }
 }
 
@@ -781,7 +806,7 @@ __kernel void ifft_multi_radix_cols(__global const uchar* src_ptr, int src_step,
             #pragma unroll
             for (int i=0; i<kercn; i++)
             {
-                float2 temp = *((__global const float2*)(src + i*block_size*src_step));
+                float2 temp = vload2(0, (__global const float*)(src + i*block_size*src_step));
                 smem[y+i*block_size].x = temp.x;
                 smem[y+i*block_size].y = -temp.y;
             }

modules/core/test/ocl/test_dft.cpp

@@ -48,26 +48,17 @@
 
 #ifdef HAVE_OPENCL
 
-enum OCL_FFT_TYPE
-{
-    R2R = 0, 
-    C2R = 1, 
-    R2C = 2, 
-    C2C = 3
-};
-
 namespace cvtest {
 namespace ocl {
 
 ////////////////////////////////////////////////////////////////////////////
 // Dft
 
-PARAM_TEST_CASE(Dft, cv::Size, OCL_FFT_TYPE, bool, bool, bool, bool)
+PARAM_TEST_CASE(Dft, cv::Size, MatDepth, bool, bool, bool, bool)
 {
     cv::Size dft_size;
-    int	dft_flags, depth, cn, dft_type;
-    bool hint;
-    bool is1d;
+    int	dft_flags, depth;
+    bool inplace;
 
     TEST_DECLARE_INPUT_PARAMETER(src);
     TEST_DECLARE_OUTPUT_PARAMETER(dst);
@@ -75,60 +66,34 @@ PARAM_TEST_CASE(Dft, cv::Size, OCL_FFT_TYPE, bool, bool, bool, bool)
     virtual void SetUp()
     {
         dft_size = GET_PARAM(0);
-        dft_type = GET_PARAM(1);
-        depth = CV_32F;
+        depth = GET_PARAM(1);
+        inplace = GET_PARAM(2);
 
         dft_flags = 0;
-        switch (dft_type)
-        {
-        case R2R: dft_flags |= cv::DFT_REAL_OUTPUT; cn = 1; break;
-        case C2R: dft_flags |= cv::DFT_REAL_OUTPUT; cn = 2; break;
-        case R2C: dft_flags |= cv::DFT_COMPLEX_OUTPUT; cn = 1; break;
-        case C2C: dft_flags |= cv::DFT_COMPLEX_OUTPUT; cn = 2; break;
-        }
-
-        if (GET_PARAM(2))
-            dft_flags |= cv::DFT_INVERSE;
         if (GET_PARAM(3))
             dft_flags |= cv::DFT_ROWS;
         if (GET_PARAM(4))
             dft_flags |= cv::DFT_SCALE;
-        hint = GET_PARAM(5);
-        is1d = (dft_flags & DFT_ROWS) != 0 || dft_size.height == 1;
+        if (GET_PARAM(5))
+            dft_flags |= cv::DFT_INVERSE;
     }
 
-    void generateTestData()
+    void generateTestData(int cn = 2)
     {
         src = randomMat(dft_size, CV_MAKE_TYPE(depth, cn), 0.0, 100.0);
         usrc = src.getUMat(ACCESS_READ);
+
+        if (inplace)
+            dst = src, udst = usrc;
     }
 };
 
-OCL_TEST_P(Dft, Mat)
+OCL_TEST_P(Dft, C2C)
 {
     generateTestData();
 
-    int nonzero_rows = hint ? src.cols - randomInt(1, src.rows-1) : 0;
-    OCL_OFF(cv::dft(src, dst, dft_flags, nonzero_rows));
-    OCL_ON(cv::dft(usrc, udst, dft_flags, nonzero_rows));
-
-    if (dft_type == R2C && is1d && (dft_flags & cv::DFT_INVERSE) == 0)
-    {
-        dst = dst(cv::Range(0, dst.rows), cv::Range(0, dst.cols/2 + 1));
-        udst = udst(cv::Range(0, udst.rows), cv::Range(0, udst.cols/2 + 1));
-    }
-    
-    //Mat gpu = udst.getMat(ACCESS_READ);
-    //std::cout << dst << std::endl;
-    //std::cout << gpu << std::endl;
-
-    //int cn = udst.channels();
-    //
-    //Mat dst1ch = dst.reshape(1);
-    //Mat gpu1ch = gpu.reshape(1);
-    //Mat df;
-    //absdiff(dst1ch, gpu1ch, df);
-    //std::cout << Mat_<int>(df) << std::endl;
+    OCL_OFF(cv::dft(src, dst, dft_flags | cv::DFT_COMPLEX_OUTPUT));
+    OCL_ON(cv::dft(usrc, udst, dft_flags | cv::DFT_COMPLEX_OUTPUT));
 
     double eps = src.size().area() * 1e-4;
     EXPECT_MAT_NEAR(dst, udst, eps);
@@ -185,13 +150,13 @@ OCL_TEST_P(MulSpectrums, Mat)
 
 OCL_INSTANTIATE_TEST_CASE_P(OCL_ImgProc, MulSpectrums, testing::Combine(Bool(), Bool()));
 
-OCL_INSTANTIATE_TEST_CASE_P(Core, Dft, Combine(Values(cv::Size(10, 10), cv::Size(36, 36), cv::Size(512, 1), cv::Size(1280, 768)),
-                                               Values((OCL_FFT_TYPE) R2C, (OCL_FFT_TYPE) C2C, (OCL_FFT_TYPE) R2R, (OCL_FFT_TYPE) C2R),
-                                               Bool(), // DFT_INVERSE
+OCL_INSTANTIATE_TEST_CASE_P(Core, Dft, Combine(Values(cv::Size(2, 3), cv::Size(5, 4), cv::Size(25, 20),
+                                                       cv::Size(512, 1), cv::Size(1024, 768)),
+                                               Values(CV_32F, CV_64F),
+                                               Bool(), // inplace
                                                Bool(), // DFT_ROWS
                                                Bool(), // DFT_SCALE
-                                               Bool()  // hint
-                                               )
+                                               Bool()) // DFT_INVERSE
                             );
 
 } } // namespace cvtest::ocl