Add dft and gemm to ocl module, using AMD's clAmdFft and clAmdBlas li…

…braries

CMakeLists.txt

@@ -140,6 +140,9 @@ OCV_OPTION(WITH_XIMEA          "Include XIMEA cameras support"               OFF
 OCV_OPTION(WITH_XINE           "Include Xine support (GPL)"                  OFF  IF (UNIX AND NOT APPLE AND NOT ANDROID) )
 OCV_OPTION(WITH_CLP            "Include Clp support (EPL)"                   OFF)
 OCV_OPTION(WITH_OPENCL         "Include OpenCL Runtime support"              OFF  IF (NOT ANDROID AND NOT IOS) )
+OCV_OPTION(WITH_OPENCLAMDFFT   "Include AMD OpenCL FFT library support"      OFF  IF (NOT ANDROID AND NOT IOS) )
+OCV_OPTION(WITH_OPENCLAMDBLAS  "Include AMD OpenCL BLAS library support"     OFF  IF (NOT ANDROID AND NOT IOS) )
+
 
 # OpenCV build components
 # ===================================================
@@ -396,6 +399,12 @@ if(WITH_OPENCL)
   if(OPENCL_FOUND)
     set(HAVE_OPENCL 1)
   endif()
+  if(WITH_OPENCLAMDFFT)
+    set(HAVE_CLAMDFFT 1)
+  endif()
+  if(WITH_OPENCLAMDBLAS)
+    set(HAVE_CLAMDBLAS 1)
+  endif()
 endif()
 
 # ----------------------------------------------------------------------------

cmake/OpenCVDetectOpenCL.cmake

@@ -2,8 +2,19 @@ if(APPLE)
     set(OPENCL_FOUND YES)
     set(OPENCL_LIBRARIES "-framework OpenCL")
 else()
-    find_package(OpenCL QUIET)
-
+    #find_package(OpenCL QUIET)
+	if(WITH_OPENCLAMDFFT)
+            find_path(CLAMDFFT_INCLUDE_DIR
+                NAMES clAmdFft.h)
+            find_library(CLAMDFFT_LIBRARIES
+                NAMES clAmdFft.Runtime)
+	endif()
+	if(WITH_OPENCLAMDBLAS)
+            find_path(CLAMDBLAS_INCLUDE_DIR
+                NAMES clAmdBlas.h)
+            find_library(CLAMDBLAS_LIBRARIES
+                NAMES clAmdBlas)
+	endif()
     # Try AMD/ATI Stream SDK
     if (NOT OPENCL_FOUND)
         set(ENV_AMDSTREAMSDKROOT $ENV{AMDAPPSDKROOT})

cmake/templates/cvconfig.h.cmake

@@ -175,6 +175,12 @@
 /* OpenCL Support */
 #cmakedefine HAVE_OPENCL
 
+/* AMD's OpenCL Fast Fourier Transform Library*/ 
+#cmakedefine HAVE_CLAMDFFT
+
+/* AMD's Basic Linear Algebra Subprograms Library*/
+#cmakedefine HAVE_CLAMDBLAS
+
 /* NVidia Cuda Fast Fourier Transform (FFT) API*/
 #cmakedefine HAVE_CUFFT
 

modules/ocl/CMakeLists.txt

@@ -29,6 +29,14 @@ if (HAVE_OPENCL)
   if(OPENCL_INCLUDE_DIR)
       ocv_include_directories(${OPENCL_INCLUDE_DIR})
   endif()
+  if (HAVE_CLAMDFFT)
+	set(ocl_link_libs ${ocl_link_libs} ${CLAMDFFT_LIBRARIES})
+	ocv_include_directories(${CLAMDFFT_INCLUDE_DIR})
+  endif()
+  if (HAVE_CLAMDBLAS)
+	set(ocl_link_libs ${ocl_link_libs} ${CLAMDBLAS_LIBRARIES})
+	ocv_include_directories(${CLAMDBLAS_INCLUDE_DIR})
+  endif()
 endif()
 
 ocv_set_module_sources(

modules/ocl/include/opencv2/ocl/ocl.hpp

@@ -894,7 +894,35 @@ namespace cv
 		// Supports TM_SQDIFF, TM_CCORR for type 32FC1 and 32FC4
 		CV_EXPORTS void matchTemplate(const oclMat& image, const oclMat& templ, oclMat& result, int method, MatchTemplateBuf& buf);
 
-
+#ifdef HAVE_CLAMDFFT
+            ///////////////////////////////////////// clAmdFft related /////////////////////////////////////////
+            // the two functions must be called before/after run any fft library functions.
+            CV_EXPORTS void fft_setup();    // this will be implicitly invoked
+            CV_EXPORTS void fft_teardown(); // you need to teardown fft library manually
+
+		    /////////////////////////////////////// DFT /////////////////////////////////////////////////////
+		    //! Performs a forward or inverse discrete Fourier transform (1D or 2D) of floating point matrix.
+		    //! Param dft_size is the size of DFT transform.
+		    //!
+		    //! For complex-to-real transform it is assumed that the source matrix is packed in CLFFT's format.
+		    // support src type of CV32FC1, CV32FC2
+		    // support flags: DFT_INVERSE, DFT_REAL_OUTPUT, DFT_COMPLEX_OUTPUT, DFT_ROWS
+		    // dft_size is the size of original input, which is used for transformation from complex to real.
+		    // dft_size must be powers of 2, 3 and 5
+		    // real to complex dft requires at least v1.8 clAmdFft
+		    // real to complex dft output is not the same with cpu version
+		    // real to complex and complex to real does not support DFT_ROWS
+		    CV_EXPORTS void dft(const oclMat& src, oclMat& dst, Size dft_size = Size(0, 0), int flags = 0);
+#endif // HAVE_CLAMDFFT
+
+#ifdef HAVE_CLAMDBLAS
+		//! implements generalized matrix product algorithm GEMM from BLAS
+		// The functionality requires clAmdBlas library
+		// only support type CV_32FC1
+		// flag GEMM_3_T is not supported
+		CV_EXPORTS void gemm(const oclMat& src1, const oclMat& src2, double alpha,
+		const oclMat& src3, double beta, oclMat& dst, int flags = 0);
+#endif
 
     }
 }