ocl: dynamic load of OpenCL libraries

Support functions from OpenCL 1.1 / OpenCL 1.2

CMakeLists.txt

@@ -824,7 +824,14 @@ endif()
 
 if(HAVE_OPENCL)
   status("")
-  status("  OpenCL")
+  status("  OpenCL:")
+  set(__opencl_ver "invalid")
+  if(HAVE_OPENCL12)
+    set(__opencl_ver "1.2")
+  elseif(HAVE_OPENCL11)
+    set(__opencl_ver "1.1")
+  endif()
+  status("    Version:"       ${__opencl_ver})
   if(OPENCL_INCLUDE_DIR)
     status("    Include path:"       ${OPENCL_INCLUDE_DIRS})
   endif()

cmake/OpenCVDetectOpenCL.cmake

@@ -6,7 +6,7 @@ if(APPLE)
 else(APPLE)
   #find_package(OpenCL QUIET)
 
-  if (NOT OPENCL_FOUND)
+  if(NOT OPENCL_FOUND)
     find_path(OPENCL_ROOT_DIR
               NAMES OpenCL/cl.h CL/cl.h include/CL/cl.h include/nvidia-current/CL/cl.h
               PATHS ENV OCLROOT ENV AMDAPPSDKROOT ENV CUDA_PATH ENV INTELOCLSDKROOT
@@ -20,18 +20,7 @@ else(APPLE)
               DOC "OpenCL include directory"
               NO_DEFAULT_PATH)
 
-    if (X86_64)
+    set(OPENCL_LIBRARY "OPENCL_DYNAMIC_LOAD")
-      set(OPENCL_POSSIBLE_LIB_SUFFIXES lib/Win64 lib/x86_64 lib/x64)
-    elseif (X86)
-      set(OPENCL_POSSIBLE_LIB_SUFFIXES lib/Win32 lib/x86)
-    endif()
-
-    find_library(OPENCL_LIBRARY
-              NAMES OpenCL
-              HINTS ${OPENCL_ROOT_DIR}
-              PATH_SUFFIXES ${OPENCL_POSSIBLE_LIB_SUFFIXES}
-              DOC "OpenCL library"
-              NO_DEFAULT_PATH)
 
     mark_as_advanced(OPENCL_INCLUDE_DIR OPENCL_LIBRARY)
     include(FindPackageHandleStandardArgs)
@@ -40,20 +29,30 @@ else(APPLE)
 endif(APPLE)
 
 if(OPENCL_FOUND)
-  set(HAVE_OPENCL 1)
+  try_compile(HAVE_OPENCL11
-  set(OPENCL_INCLUDE_DIRS ${OPENCL_INCLUDE_DIR})
+    "${OpenCV_BINARY_DIR}"
-  set(OPENCL_LIBRARIES    ${OPENCL_LIBRARY})
+    "${OpenCV_SOURCE_DIR}/cmake/checks/opencl11.cpp"
-
+    CMAKE_FLAGS "-DINCLUDE_DIRECTORIES:STRING=${OPENCL_INCLUDE_DIR}"
-  if(WIN32 AND X86_64)
+    )
-    set(CLAMD_POSSIBLE_LIB_SUFFIXES lib64/import)
+  if(NOT HAVE_OPENCL11)
-  elseif(WIN32)
+    message(STATUS "OpenCL 1.1 not found, ignore OpenCL SDK")
-    set(CLAMD_POSSIBLE_LIB_SUFFIXES lib32/import)
+    return()
+  endif()
+  try_compile(HAVE_OPENCL12
+    "${OpenCV_BINARY_DIR}"
+    "${OpenCV_SOURCE_DIR}/cmake/checks/opencl12.cpp"
+    CMAKE_FLAGS "-DINCLUDE_DIRECTORIES:STRING=${OPENCL_INCLUDE_DIR}"
+    )
+  if(NOT HAVE_OPENCL12)
+    message(STATUS "OpenCL 1.2 not found, will use OpenCL 1.1")
   endif()
 
-  if(X86_64 AND UNIX)
+  set(HAVE_OPENCL 1)
-    set(CLAMD_POSSIBLE_LIB_SUFFIXES lib64)
+  set(OPENCL_INCLUDE_DIRS ${OPENCL_INCLUDE_DIR})
-  elseif(X86 AND UNIX)
+  if(OPENCL_LIBRARY MATCHES "OPENCL_DYNAMIC_LOAD")
-    set(CLAMD_POSSIBLE_LIB_SUFFIXES lib32)
+    unset(OPENCL_LIBRARIES)
+  else()
+    set(OPENCL_LIBRARIES "${OPENCL_LIBRARY}")
   endif()
 
   if(WITH_OPENCLAMDFFT)
@@ -70,16 +69,9 @@ if(OPENCL_FOUND)
               PATH_SUFFIXES include
               DOC "clAmdFft include directory")
 
-    find_library(CLAMDFFT_LIBRARY
+    if(CLAMDFFT_INCLUDE_DIR)
-              NAMES clAmdFft.Runtime
-              HINTS ${CLAMDFFT_ROOT_DIR}
-              PATH_SUFFIXES ${CLAMD_POSSIBLE_LIB_SUFFIXES}
-              DOC "clAmdFft library")
-
-    if(CLAMDFFT_LIBRARY AND CLAMDFFT_INCLUDE_DIR)
       set(HAVE_CLAMDFFT 1)
       list(APPEND OPENCL_INCLUDE_DIRS "${CLAMDFFT_INCLUDE_DIR}")
-      list(APPEND OPENCL_LIBRARIES    "${CLAMDFFT_LIBRARY}")
     endif()
   endif()
 
@@ -97,16 +89,9 @@ if(OPENCL_FOUND)
               PATH_SUFFIXES include
               DOC "clAmdFft include directory")
 
-    find_library(CLAMDBLAS_LIBRARY
+    if(CLAMDBLAS_INCLUDE_DIR)
-              NAMES clAmdBlas
-              HINTS ${CLAMDBLAS_ROOT_DIR}
-              PATH_SUFFIXES ${CLAMD_POSSIBLE_LIB_SUFFIXES}
-              DOC "clAmdBlas library")
-
-    if(CLAMDBLAS_LIBRARY AND CLAMDBLAS_INCLUDE_DIR)
       set(HAVE_CLAMDBLAS 1)
       list(APPEND OPENCL_INCLUDE_DIRS "${CLAMDBLAS_INCLUDE_DIR}")
-      list(APPEND OPENCL_LIBRARIES    "${CLAMDBLAS_LIBRARY}")
     endif()
   endif()
 endif()

cmake/checks/opencl11.cpp

@@ -0,0 +1,14 @@
+#if defined __APPLE__
+#include <OpenCL/cl.h>
+#else
+#include <CL/cl.h>
+#endif
+
+int main(int argc, char** argv)
+{
+#ifdef CL_VERSION_1_1
+#else
+#error OpenCL 1.1 not found
+#endif
+    return 0;
+}

cmake/checks/opencl12.cpp

@@ -0,0 +1,14 @@
+#if defined __APPLE__
+#include <OpenCL/cl.h>
+#else
+#include <CL/cl.h>
+#endif
+
+int main(int argc, char** argv)
+{
+#ifdef CL_VERSION_1_2
+#else
+#error OpenCL 1.2 not found
+#endif
+    return 0;
+}

cmake/templates/cvconfig.h.cmake

@@ -108,6 +108,8 @@
 
 /* OpenCL Support */
 #cmakedefine HAVE_OPENCL
+#cmakedefine HAVE_OPENCL11
+#cmakedefine HAVE_OPENCL12
 
 /* OpenEXR codec */
 #cmakedefine HAVE_OPENEXR

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

@@ -0,0 +1,19 @@
+//
+// AUTOGENERATED, DO NOT EDIT
+//
+#ifndef __OPENCV_OCL_CL_RUNTIME_HPP__
+#define __OPENCV_OCL_CL_RUNTIME_HPP__
+
+#ifdef HAVE_OPENCL
+
+#if defined(HAVE_OPENCL12)
+#include "cl_runtime_opencl12.hpp"
+#elif defined(HAVE_OPENCL11)
+#include "cl_runtime_opencl11.hpp"
+#else
+#error Invalid OpenCL configuration
+#endif
+
+#endif
+
+#endif // __OPENCV_OCL_CL_RUNTIME_HPP__

modules/ocl/include/opencv2/ocl/private/util.hpp

@@ -48,11 +48,7 @@
 
 #include "opencv2/ocl/ocl.hpp"
 
-#if defined __APPLE__
+#include "opencv2/ocl/cl_runtime/cl_runtime.hpp"
-#include <OpenCL/opencl.h>
-#else
-#include <CL/opencl.h>
-#endif
 
 namespace cv
 {

modules/ocl/src/cl_runtime/cl_runtime.cpp

@@ -0,0 +1,105 @@
+#include "precomp.hpp"
+
+#if defined(HAVE_OPENCL) && (!defined(__APPLE__) || defined(IOS))
+
+#if defined(BUILD_SHARED_LIBS) && (defined WIN32 || defined _WIN32 || defined WINCE)
+#define CL_RUNTIME_EXPORT __declspec(dllexport)
+#else
+#define CL_RUNTIME_EXPORT
+#endif
+
+
+#include "opencv2/ocl/cl_runtime/cl_runtime.hpp"
+
+#if defined(__APPLE__)
+    #include <dlfcn.h>
+
+    static void* AppleCLGetProcAddress(const char* name)
+    {
+        static void * image = NULL;
+        if (!image)
+        {
+            image = dlopen("/System/Library/Frameworks/OpenCL.framework/Versions/Current/OpenCL", RTLD_LAZY | RTLD_GLOBAL);
+            if (!image)
+                return NULL;
+        }
+
+        return dlsym(image, name);
+    }
+    #define CV_CL_GET_PROC_ADDRESS(name) AppleCLGetProcAddress(name)
+#endif // __APPLE__
+
+#if defined(_WIN32)
+    static void* WinGetProcAddress(const char* name)
+    {
+        static HMODULE opencl_module = NULL;
+        if (!opencl_module)
+        {
+            opencl_module = GetModuleHandleA("OpenCL.dll");
+            if (!opencl_module)
+            {
+                const char* name = "OpenCL.dll";
+                const char* envOpenCLBinary = getenv("OPENCV_OPENCL_BINARY");
+                if (envOpenCLBinary)
+                    name = envOpenCLBinary;
+                opencl_module = LoadLibraryA(name);
+                if (!opencl_module)
+                    return NULL;
+            }
+        }
+        return (void*)GetProcAddress(opencl_module, name);
+    }
+    #define CV_CL_GET_PROC_ADDRESS(name) WinGetProcAddress(name)
+#endif // _WIN32
+
+#if defined(linux)
+    #include <dlfcn.h>
+    #include <stdio.h>
+
+    static void* GetProcAddress (const char* name)
+    {
+        static void* h = NULL;
+        if (!h)
+        {
+            const char* name = "libOpenCL.so";
+            const char* envOpenCLBinary = getenv("OPENCV_OPENCL_BINARY");
+            if (envOpenCLBinary)
+                name = envOpenCLBinary;
+            h = dlopen(name, RTLD_LAZY | RTLD_GLOBAL);
+            if (!h)
+                return NULL;
+        }
+
+        return dlsym(h, name);
+    }
+    #define CV_CL_GET_PROC_ADDRESS(name) GetProcAddress(name)
+#endif
+
+#ifndef CV_CL_GET_PROC_ADDRESS
+#define CV_CL_GET_PROC_ADDRESS(name) NULL
+#endif
+
+static void* opencl_check_fn(int ID)
+{
+    extern const char* opencl_fn_names[];
+    void* func = CV_CL_GET_PROC_ADDRESS(opencl_fn_names[ID]);
+    if (!func)
+    {
+        std::ostringstream msg;
+        msg << "OpenCL function is not available: [" << opencl_fn_names[ID] << "]";
+        CV_Error(CV_StsBadFunc, msg.str());
+    }
+    extern void* opencl_fn_ptrs[];
+    *(void**)(opencl_fn_ptrs[ID]) = func;
+    return func;
+}
+
+#if defined(HAVE_OPENCL12)
+#include "cl_runtime_opencl12_impl.hpp"
+#elif defined(HAVE_OPENCL11)
+#include "cl_runtime_opencl11_impl.hpp"
+#else
+#error Invalid OpenCL configuration
+#endif
+
+#endif

modules/ocl/src/cl_runtime/generator/common.py

@@ -0,0 +1,199 @@
+import sys, os, re
+
+#
+# Parser helpers
+#
+
+def remove_comments(s):
+    def replacer(match):
+        s = match.group(0)
+        if s.startswith('/'):
+            return ""
+        else:
+            return s
+    pattern = re.compile(
+        r'//.*?$|/\*.*?\*/|\'(?:\\.|[^\\\'])*\'|"(?:\\.|[^\\"])*"',
+        re.DOTALL | re.MULTILINE
+    )
+    return re.sub(pattern, replacer, s)
+
+
+def getTokens(s):
+    return re.findall(r'[a-z_A-Z0-9_]+|[^[a-z_A-Z0-9_ \n\r\t]', s)
+
+
+def getParameter(pos, tokens):
+    deep = 0
+    p = []
+    while True:
+        if pos >= len(tokens):
+            break
+        if (tokens[pos] == ')' or tokens[pos] == ',') and deep == 0:
+            if tokens[pos] == ')':
+                pos = len(tokens)
+            else:
+                pos += 1
+            break
+        if tokens[pos] == '(':
+            deep += 1
+        if tokens[pos] == ')':
+            deep -= 1
+        p.append(tokens[pos])
+        pos += 1
+    return (' '.join(p), pos)
+
+
+def getParameters(i, tokens):
+    assert tokens[i] == '('
+    i += 1
+
+    params = []
+    while True:
+        if i >= len(tokens) or tokens[i] == ')':
+            break
+
+        (param, i) = getParameter(i, tokens)
+        if len(param) > 0:
+            params.append(param)
+        else:
+            assert False
+            break
+
+    if len(params) > 0 and params[0] == 'void':
+        del params[0]
+
+    return params
+
+def postProcessParameters(fns):
+    for fn in fns:
+        fn['params_full'] = list(fn['params'])
+        for i in range(len(fn['params'])):
+            p = fn['params'][i]
+            if p.find('(') != -1:
+                p = re.sub(r'\* *([a-zA-Z0-9_]*) ?\)', '*)', p, 1)
+                fn['params'][i] = p
+                continue
+            parts = re.findall(r'[a-z_A-Z0-9]+|\*', p)
+            if len(parts) > 1:
+                if parts[-1].find('*') == -1:
+                    del parts[-1]
+            fn['params'][i] = ' '.join(parts)
+
+#
+# Generator helpers
+#
+
+def outputToString(f):
+    def wrapped(*args, **kwargs):
+        from cStringIO import StringIO
+        old_stdout = sys.stdout
+        sys.stdout = str_stdout = StringIO()
+        res = f(*args, **kwargs)
+        assert res is None
+        sys.stdout = old_stdout
+        result = str_stdout.getvalue()
+        result = re.sub(r'([^\n ]) [ ]+', r'\1 ', result)  # don't remove spaces at start of line
+        result = re.sub(r' ,', ',', result)
+        result = re.sub(r' \*', '*', result)
+        result = re.sub(r'\( ', '(', result)
+        result = re.sub(r' \)', ')', result)
+        return result
+    return wrapped
+
+@outputToString
+def generateEnums(fns, prefix='OPENCL_FN'):
+    print '// generated by %s' % os.path.basename(sys.argv[0])
+    print 'enum %s_ID {' % prefix
+    first = True
+    for fn in fns:
+        print '    %s_%s%s,' % (prefix, fn['name'], ' = 0' if first else '')
+        first = False
+    print '};'
+
+@outputToString
+def generateNames(fns, prefix='opencl_fn'):
+    print '// generated by %s' % os.path.basename(sys.argv[0])
+    print 'const char* %s_names[] = {' % prefix
+    for fn in fns:
+        print '    "%s",' % (fn['name'])
+    print '};'
+
+@outputToString
+def generatePtrs(fns, prefix='opencl_fn'):
+    print '// generated by %s' % os.path.basename(sys.argv[0])
+    print 'void* %s_ptrs[] = {' % prefix
+    for fn in fns:
+        print '    &%s,' % (fn['name'])
+    print '};'
+
+@outputToString
+def generateRemapOrigin(fns):
+    print '// generated by %s' % os.path.basename(sys.argv[0])
+    for fn in fns:
+        print '#define %s %s_' % (fn['name'], fn['name'])
+
+@outputToString
+def generateRemapDynamic(fns):
+    print '// generated by %s' % os.path.basename(sys.argv[0])
+    for fn in fns:
+        print '#undef %s' % (fn['name'])
+        print '#define %s %s_pfn' % (fn['name'], fn['name'])
+
+@outputToString
+def generateParamsCfg(fns):
+    for fn in fns:
+        print '%s %d' % (fn['name'], len(fn['params']))
+
+@outputToString
+def generateFnDeclaration(fns):
+    print '// generated by %s' % os.path.basename(sys.argv[0])
+    for fn in fns:
+        print 'extern CL_RUNTIME_EXPORT %s %s (%s *%s)(%s);' % (' '.join(fn['modifiers']), ' '.join(fn['ret']), ' '.join(fn['calling']),
+                                  fn['name'], ', '.join(fn['params'] if not fn.has_key('params_full') else fn['params_full']))
+
+@outputToString
+def generateFnDefinition(fns, lprefix='opencl_fn', uprefix='OPENCL_FN'):
+    print '// generated by %s' % os.path.basename(sys.argv[0])
+    for fn in fns:
+        print '%s%s (%s *%s)(%s) = %s%d<%s_%s, %s%s>::switch_fn;' % \
+            ((' '.join(fn['modifiers'] + ' ') if len(fn['modifiers']) > 0 else ''),
+             ' '.join(fn['ret']), ' '.join(fn['calling']), fn['name'], ', '.join(fn['params']), \
+             lprefix, len(fn['params']), uprefix, fn['name'], ' '.join(fn['ret']), ('' if len(fn['params']) == 0 else ', ' + ', '.join(fn['params'])))
+
+@outputToString
+def generateTemplates(sz, lprefix, switch_name, calling_convention=''):
+    print '// generated by %s' % os.path.basename(sys.argv[0])
+    for sz in range(sz):
+        template_params = ['int ID', 'typename _R']
+        types = []
+        types_with_params = []
+        params = []
+        for i in range(1, sz + 1):
+            template_params.append('typename _T%d' % i)
+            types.append('_T%d' % i)
+            types_with_params.append('_T%d p%d' % (i, i))
+            params.append('p%d' % i)
+        print 'template <%s>' % ', '.join(template_params)
+        print 'struct %s%d' % (lprefix, sz)
+        print '{'
+        print '    typedef _R (%s *FN)(%s);' % (calling_convention, ', '.join(types))
+        print '    static _R %s switch_fn(%s)' % (calling_convention, ', '.join(types_with_params))
+        print '    { return ((FN)%s(ID))(%s); }' % (switch_name, ', '.join(params))
+        print '};'
+        print ''
+
+
+def ProcessTemplate(inputFile, ctx, noteLine='//\n// AUTOGENERATED, DO NOT EDIT\n//'):
+    f = open(inputFile, "r")
+    if noteLine:
+        print noteLine
+    for line in f:
+        if line.startswith('@'):
+            assert line[-1] == '\n'
+            line = line[:-1]  # remove '\n'
+            assert line[-1] == '@'
+            name = line[1:-1]
+            assert ctx.has_key(name), name
+            line = ctx[name]
+        print line,
+    f.close()

modules/ocl/src/cl_runtime/generator/generate.sh

@@ -0,0 +1,7 @@
+#!/bin/bash -e
+echo "Generate files for CL runtime..."
+cat sources/opencl11/cl.h | python parser_cl.py cl_runtime_opencl11
+cat sources/opencl12/cl.h | python parser_cl.py cl_runtime_opencl12
+cat sources/clAmdBlas.h | python parser_clamdblas.py
+cat sources/clAmdFft.h | python parser_clamdfft.py
+echo "Generate files for CL runtime... Done"

modules/ocl/src/cl_runtime/generator/parser_cl.py

@@ -0,0 +1,112 @@
+#!/bin/python
+# usage:
+#     cat opencl11/cl.h | $0 cl_runtime_opencl11
+#     cat opencl12/cl.h | $0 cl_runtime_opencl12
+import sys, re;
+
+from common import remove_comments, getTokens, getParameters, postProcessParameters
+
+try:
+    if len(sys.argv) > 1:
+        outfile = open('../../../include/opencv2/ocl/cl_runtime/' + sys.argv[1] + '.hpp', "w")
+        outfile_impl = open('../' + sys.argv[1] + '_impl.hpp', "w")
+        if len(sys.argv) > 2:
+            f = open(sys.argv[2], "r")
+        else:
+            f = sys.stdin
+    else:
+        sys.exit("ERROR. Specify output file")
+except:
+    sys.exit("ERROR. Can't open input/output file, check parameters")
+
+fns = []
+
+while True:
+    line = f.readline()
+    if len(line) == 0:
+        break
+    assert isinstance(line, str)
+    parts = line.split();
+    if line.startswith('extern') and line.find('CL_API_CALL') != -1:
+        # read block of lines
+        while True:
+            nl = f.readline()
+            nl = nl.strip()
+            nl = re.sub(r'\n', r'', nl)
+            if len(nl) == 0:
+                break;
+            line += ' ' + nl
+
+        line = remove_comments(line)
+
+        parts = getTokens(line)
+
+        fn = {}
+        modifiers = []
+        ret = []
+        calling = []
+        i = 1
+        while (i < len(parts)):
+            if parts[i].startswith('CL_'):
+                modifiers.append(parts[i])
+            else:
+                break
+            i += 1
+        while (i < len(parts)):
+            if not parts[i].startswith('CL_'):
+                ret.append(parts[i])
+            else:
+                break
+            i += 1
+        while (i < len(parts)):
+            calling.append(parts[i])
+            i += 1
+            if parts[i - 1] == 'CL_API_CALL':
+                break
+
+        fn['modifiers'] = []  # modifiers
+        fn['ret'] = ret
+        fn['calling'] = []  # calling
+
+        # print 'modifiers='+' '.join(modifiers)
+        # print 'ret='+' '.join(type)
+        # print 'calling='+' '.join(calling)
+
+        name = parts[i]; i += 1;
+        fn['name'] = name
+        print 'name=' + name
+
+        params = getParameters(i, parts)
+
+        fn['params'] = params
+        # print 'params="'+','.join(params)+'"'
+
+        fns.append(fn)
+
+f.close()
+
+print 'Found %d functions' % len(fns)
+
+postProcessParameters(fns)
+
+from pprint import pprint
+pprint(fns)
+
+from common import *
+
+ctx = {}
+ctx['CL_REMAP_ORIGIN'] = generateRemapOrigin(fns)
+ctx['CL_REMAP_DYNAMIC'] = generateRemapDynamic(fns)
+ctx['CL_FN_DECLARATIONS'] = generateFnDeclaration(fns)
+
+sys.stdout = outfile
+ProcessTemplate('template/cl_runtime_opencl.hpp.in', ctx)
+
+ctx['CL_FN_ENUMS'] = generateEnums(fns)
+ctx['CL_FN_NAMES'] = generateNames(fns)
+ctx['CL_FN_DEFINITIONS'] = generateFnDefinition(fns)
+ctx['CL_FN_PTRS'] = generatePtrs(fns)
+ctx['CL_FN_SWITCH'] = generateTemplates(15, 'opencl_fn', 'opencl_check_fn')
+
+sys.stdout = outfile_impl
+ProcessTemplate('template/cl_runtime_impl_opencl.hpp.in', ctx)

modules/ocl/src/cl_runtime/generator/parser_clamdblas.py

@@ -0,0 +1,107 @@
+#!/bin/python
+# usage:
+#     cat clAmdBlas.h | $0
+import sys, re;
+
+from common import remove_comments, getTokens, getParameters, postProcessParameters
+
+try:
+    if len(sys.argv) > 1:
+        f = open(sys.argv[1], "r")
+    else:
+        f = sys.stdin
+except:
+    sys.exit("ERROR. Can't open input file")
+
+fns = []
+
+while True:
+    line = f.readline()
+    if len(line) == 0:
+        break
+    assert isinstance(line, str)
+    line = line.strip()
+    parts = line.split();
+    if (line.startswith('clAmd') or line.startswith('cl_') or line == 'void') and len(line.split()) == 1 and line.find('(') == -1:
+        fn = {}
+        modifiers = []
+        ret = []
+        calling = []
+        i = 0
+        while (i < len(parts)):
+            if parts[i].startswith('CL_'):
+                modifiers.append(parts[i])
+            else:
+                break
+            i += 1
+        while (i < len(parts)):
+            if not parts[i].startswith('CL_'):
+                ret.append(parts[i])
+            else:
+                break
+            i += 1
+        while (i < len(parts)):
+            calling.append(parts[i])
+            i += 1
+        fn['modifiers'] = []  # modifiers
+        fn['ret'] = ret
+        fn['calling'] = calling
+
+        # print 'modifiers='+' '.join(modifiers)
+        # print 'ret='+' '.join(type)
+        # print 'calling='+' '.join(calling)
+
+        # read block of lines
+        line = f.readline()
+        while True:
+            nl = f.readline()
+            nl = nl.strip()
+            nl = re.sub(r'\n', r'', nl)
+            if len(nl) == 0:
+                break;
+            line += ' ' + nl
+
+        line = remove_comments(line)
+
+        parts = getTokens(line)
+
+        i = 0;
+
+        name = parts[i]; i += 1;
+        fn['name'] = name
+        print 'name=' + name
+
+        params = getParameters(i, parts)
+
+        fn['params'] = params
+        # print 'params="'+','.join(params)+'"'
+
+        fns.append(fn)
+
+f.close()
+
+print 'Found %d functions' % len(fns)
+
+postProcessParameters(fns)
+
+from pprint import pprint
+pprint(fns)
+
+from common import *
+
+ctx = {}
+ctx['CLAMDBLAS_REMAP_ORIGIN'] = generateRemapOrigin(fns)
+ctx['CLAMDBLAS_REMAP_DYNAMIC'] = generateRemapDynamic(fns)
+ctx['CLAMDBLAS_FN_DECLARATIONS'] = generateFnDeclaration(fns)
+
+sys.stdout = open('../../../include/opencv2/ocl/cl_runtime/clamdblas_runtime.hpp', 'w')
+ProcessTemplate('template/clamdblas_runtime.hpp.in', ctx)
+
+ctx['CL_FN_ENUMS'] = generateEnums(fns, 'OPENCLAMDBLAS_FN')
+ctx['CL_FN_NAMES'] = generateNames(fns, 'openclamdblas_fn')
+ctx['CL_FN_DEFINITIONS'] = generateFnDefinition(fns, 'openclamdblas_fn', 'OPENCLAMDBLAS_FN')
+ctx['CL_FN_PTRS'] = generatePtrs(fns, 'openclamdblas_fn')
+ctx['CL_FN_SWITCH'] = generateTemplates(23, 'openclamdblas_fn', 'openclamdblas_check_fn', '')
+
+sys.stdout = open('../clamdblas_runtime.cpp', 'w')
+ProcessTemplate('template/clamdblas_runtime.cpp.in', ctx)

modules/ocl/src/cl_runtime/generator/parser_clamdfft.py

@@ -0,0 +1,104 @@
+#!/bin/python
+# usage:
+#     cat clAmdFft.h | $0
+import sys, re;
+
+from common import remove_comments, getTokens, getParameters, postProcessParameters
+
+
+try:
+    if len(sys.argv) > 1:
+        f = open(sys.argv[1], "r")
+    else:
+        f = sys.stdin
+except:
+    sys.exit("ERROR. Can't open input file")
+
+fns = []
+
+while True:
+    line = f.readline()
+    if len(line) == 0:
+        break
+    assert isinstance(line, str)
+    line = line.strip()
+    if line.startswith('CLAMDFFTAPI'):
+        line = re.sub(r'\n', r'', line)
+        while True:
+            nl = f.readline()
+            nl = nl.strip()
+            nl = re.sub(r'\n', r'', nl)
+            if len(nl) == 0:
+                break;
+            line += ' ' + nl
+
+        line = remove_comments(line)
+
+        parts = getTokens(line)
+
+        fn = {}
+        modifiers = []
+        ret = []
+        calling = []
+
+        i = 0
+        while True:
+            if parts[i] == "CLAMDFFTAPI":
+                modifiers.append(parts[i])
+            else:
+                break
+            i += 1
+        while (i < len(parts)):
+            if not parts[i] == '(':
+                ret.append(parts[i])
+            else:
+                del ret[-1]
+                i -= 1
+                break
+            i += 1
+
+        fn['modifiers'] = []  # modifiers
+        fn['ret'] = ret
+        fn['calling'] = calling
+
+        name = parts[i]; i += 1;
+        fn['name'] = name
+        print 'name=' + name
+
+        params = getParameters(i, parts)
+
+        if len(params) > 0 and params[0] == 'void':
+            del params[0]
+
+        fn['params'] = params
+        # print 'params="'+','.join(params)+'"'
+
+        fns.append(fn)
+
+f.close()
+
+print 'Found %d functions' % len(fns)
+
+postProcessParameters(fns)
+
+from pprint import pprint
+pprint(fns)
+
+from common import *
+
+ctx = {}
+ctx['CLAMDFFT_REMAP_ORIGIN'] = generateRemapOrigin(fns)
+ctx['CLAMDFFT_REMAP_DYNAMIC'] = generateRemapDynamic(fns)
+ctx['CLAMDFFT_FN_DECLARATIONS'] = generateFnDeclaration(fns)
+
+sys.stdout = open('../../../include/opencv2/ocl/cl_runtime/clamdfft_runtime.hpp', 'w')
+ProcessTemplate('template/clamdfft_runtime.hpp.in', ctx)
+
+ctx['CL_FN_ENUMS'] = generateEnums(fns, 'OPENCLAMDFFT_FN')
+ctx['CL_FN_NAMES'] = generateNames(fns, 'openclamdfft_fn')
+ctx['CL_FN_DEFINITIONS'] = generateFnDefinition(fns, 'openclamdfft_fn', 'OPENCLAMDFFT_FN')
+ctx['CL_FN_PTRS'] = generatePtrs(fns, 'openclamdfft_fn')
+ctx['CL_FN_SWITCH'] = generateTemplates(23, 'openclamdfft_fn', 'openclamdfft_check_fn', '')
+
+sys.stdout = open('../clamdfft_runtime.cpp', 'w')
+ProcessTemplate('template/clamdfft_runtime.cpp.in', ctx)

modules/ocl/src/cl_runtime/generator/template/cl_runtime_impl_opencl.hpp.in

@@ -0,0 +1,10 @@
+@CL_FN_ENUMS@
+@CL_FN_NAMES@
+
+namespace {
+@CL_FN_SWITCH@
+}
+
+@CL_FN_DEFINITIONS@
+
+@CL_FN_PTRS@

modules/ocl/src/cl_runtime/generator/template/cl_runtime_opencl.hpp.in

@@ -0,0 +1,34 @@
+#ifndef __OPENCV_OCL_CL_RUNTIME_OPENCL_HPP__
+#define __OPENCV_OCL_CL_RUNTIME_OPENCL_HPP__
+
+#ifdef HAVE_OPENCL
+
+#if defined __APPLE__ && !defined(IOS)
+#include <OpenCL/cl.h>
+#else
+
+@CL_REMAP_ORIGIN@
+
+#if defined __APPLE__
+#include <OpenCL/cl.h>
+#else
+#include <CL/cl.h>
+#endif
+
+@CL_REMAP_DYNAMIC@
+
+#ifndef CL_RUNTIME_EXPORT
+#if (defined(BUILD_SHARED_LIBS) || defined(OPENCV_OCL_SHARED)) && (defined WIN32 || defined _WIN32 || defined WINCE)
+#define CL_RUNTIME_EXPORT __declspec(dllimport)
+#else
+#define CL_RUNTIME_EXPORT
+#endif
+#endif
+
+@CL_FN_DECLARATIONS@
+
+#endif
+
+#endif
+
+#endif // __OPENCV_OCL_CL_RUNTIME_OPENCL_HPP__

modules/ocl/src/cl_runtime/generator/template/clamdblas_runtime.cpp.in

@@ -0,0 +1,75 @@
+#include "precomp.hpp"
+
+#ifdef HAVE_CLAMDBLAS
+
+#include "opencv2/ocl/cl_runtime/cl_runtime.hpp"
+#include "opencv2/ocl/cl_runtime/clamdblas_runtime.hpp"
+
+#if defined(_WIN32)
+    static void* WinGetProcAddress(const char* name)
+    {
+        static HMODULE opencl_module = NULL;
+        if (!opencl_module)
+        {
+            opencl_module = GetModuleHandleA("clAmdBlas.dll");
+            if (!opencl_module)
+            {
+                opencl_module = LoadLibraryA("clAmdBlas.dll");
+                if (!opencl_module)
+                    return NULL;
+            }
+        }
+        return (void*)GetProcAddress(opencl_module, name);
+    }
+    #define CV_CL_GET_PROC_ADDRESS(name) WinGetProcAddress(name)
+#endif // _WIN32
+
+#if defined(linux)
+    #include <dlfcn.h>
+    #include <stdio.h>
+
+    static void* GetProcAddress (const char* name)
+    {
+        static void* h = NULL;
+        if (!h)
+        {
+            h = dlopen("libclAmdBlas.so", RTLD_LAZY | RTLD_GLOBAL);
+            if (!h)
+                return NULL;
+        }
+
+        return dlsym(h, name);
+    }
+    #define CV_CL_GET_PROC_ADDRESS(name) GetProcAddress(name)
+#endif
+
+#ifndef CV_CL_GET_PROC_ADDRESS
+#define CV_CL_GET_PROC_ADDRESS(name) NULL
+#endif
+
+@CL_FN_ENUMS@
+@CL_FN_NAMES@
+
+static void* openclamdblas_check_fn(int ID)
+{
+    void* func = CV_CL_GET_PROC_ADDRESS(openclamdblas_fn_names[ID]);
+    if (!func)
+    {
+        std::ostringstream msg;
+        msg << "OpenCL AMD BLAS function is not available: [" << openclamdblas_fn_names[ID] << "]";
+        CV_Error(CV_StsBadFunc, msg.str());
+    }
+    extern void* openclamdblas_fn_ptrs[];
+    *(void**)(openclamdblas_fn_ptrs[ID]) = func;
+    return func;
+}
+
+namespace {
+@CL_FN_SWITCH@
+}
+
+@CL_FN_DEFINITIONS@
+
+@CL_FN_PTRS@
+
+#endif

modules/ocl/src/cl_runtime/generator/template/clamdblas_runtime.hpp.in

@@ -0,0 +1,25 @@
+#ifndef __OPENCV_OCL_CLAMDBLAS_RUNTIME_HPP__
+#define __OPENCV_OCL_CLAMDBLAS_RUNTIME_HPP__
+
+#ifdef HAVE_CLAMDBLAS
+
+@CLAMDBLAS_REMAP_ORIGIN@
+
+#include <clAmdBlas.h>
+
+@CLAMDBLAS_REMAP_DYNAMIC@
+
+#ifndef CL_RUNTIME_EXPORT
+#if (defined(BUILD_SHARED_LIBS) || defined(OPENCV_OCL_SHARED)) && (defined WIN32 || defined _WIN32 || defined WINCE)
+#define CL_RUNTIME_EXPORT __declspec(dllimport)
+#else
+#define CL_RUNTIME_EXPORT
+#endif
+#endif
+
+
+@CLAMDBLAS_FN_DECLARATIONS@
+
+#endif
+
+#endif // __OPENCV_OCL_CLAMDBLAS_RUNTIME_HPP__

modules/ocl/src/cl_runtime/generator/template/clamdfft_runtime.cpp.in

@@ -0,0 +1,75 @@
+#include "precomp.hpp"
+
+#ifdef HAVE_CLAMDFFT
+
+#include "opencv2/ocl/cl_runtime/cl_runtime.hpp"
+#include "opencv2/ocl/cl_runtime/clamdfft_runtime.hpp"
+
+#if defined(_WIN32)
+    static void* WinGetProcAddress(const char* name)
+    {
+        static HMODULE opencl_module = NULL;
+        if (!opencl_module)
+        {
+            opencl_module = GetModuleHandleA("clAmdFft.Runtime.dll");
+            if (!opencl_module)
+            {
+                opencl_module = LoadLibraryA("clAmdFft.Runtime.dll");
+                if (!opencl_module)
+                    return NULL;
+            }
+        }
+        return (void*)GetProcAddress(opencl_module, name);
+    }
+    #define CV_CL_GET_PROC_ADDRESS(name) WinGetProcAddress(name)
+#endif // _WIN32
+
+#if defined(linux)
+    #include <dlfcn.h>
+    #include <stdio.h>
+
+    static void* GetProcAddress (const char* name)
+    {
+        static void* h = NULL;
+        if (!h)
+        {
+            h = dlopen("libclAmdFft.Runtime.so", RTLD_LAZY | RTLD_GLOBAL);
+            if (!h)
+                return NULL;
+        }
+
+        return dlsym(h, name);
+    }
+    #define CV_CL_GET_PROC_ADDRESS(name) GetProcAddress(name)
+#endif
+
+#ifndef CV_CL_GET_PROC_ADDRESS
+#define CV_CL_GET_PROC_ADDRESS(name) NULL
+#endif
+
+@CL_FN_ENUMS@
+@CL_FN_NAMES@
+
+static void* openclamdfft_check_fn(int ID)
+{
+    void* func = CV_CL_GET_PROC_ADDRESS(openclamdfft_fn_names[ID]);
+    if (!func)
+    {
+        std::ostringstream msg;
+        msg << "OpenCL AMD FFT function is not available: [" << openclamdfft_fn_names[ID] << "]";
+        CV_Error(CV_StsBadFunc, msg.str());
+    }
+    extern void* openclamdfft_fn_ptrs[];
+    *(void**)(openclamdfft_fn_ptrs[ID]) = func;
+    return func;
+}
+
+namespace {
+@CL_FN_SWITCH@
+}
+
+@CL_FN_DEFINITIONS@
+
+@CL_FN_PTRS@
+
+#endif

modules/ocl/src/cl_runtime/generator/template/clamdfft_runtime.hpp.in

@@ -0,0 +1,25 @@
+#ifndef __OPENCV_OCL_CLAMDFFT_RUNTIME_HPP__
+#define __OPENCV_OCL_CLAMDFFT_RUNTIME_HPP__
+
+#ifdef HAVE_CLAMDFFT
+
+@CLAMDFFT_REMAP_ORIGIN@
+
+#include <clAmdFft.h>
+
+@CLAMDFFT_REMAP_DYNAMIC@
+
+#ifndef CL_RUNTIME_EXPORT
+#if (defined(BUILD_SHARED_LIBS) || defined(OPENCV_OCL_SHARED)) && (defined WIN32 || defined _WIN32 || defined WINCE)
+#define CL_RUNTIME_EXPORT __declspec(dllimport)
+#else
+#define CL_RUNTIME_EXPORT
+#endif
+#endif
+
+
+@CLAMDFFT_FN_DECLARATIONS@
+
+#endif
+
+#endif // __OPENCV_OCL_CLAMDFFT_RUNTIME_HPP__

modules/ocl/src/fft.cpp

@@ -59,7 +59,7 @@ namespace cv { namespace ocl {
 }}
 void cv::ocl::fft_teardown(){}
 #else
-#include "clAmdFft.h"
+#include "opencv2/ocl/cl_runtime/clamdfft_runtime.hpp"
 namespace cv
 {
     namespace ocl

modules/ocl/src/gemm.cpp

@@ -73,7 +73,7 @@ void cv::ocl::clBlasTeardown()
 }
 
 #else
-#include "clAmdBlas.h"
+#include "opencv2/ocl/cl_runtime/clamdblas_runtime.hpp"
 using namespace cv;
 
 static bool clBlasInitialized = false;

modules/ocl/src/precomp.hpp

@@ -52,8 +52,18 @@
 #pragma warning( disable: 4267 4324 4244 4251 4710 4711 4514 4996 )
 #endif
 
+#if defined(_WIN32)
+#include <windows.h>
+#endif
+
 #include "cvconfig.h"
 
+#if defined(BUILD_SHARED_LIBS) && (defined WIN32 || defined _WIN32 || defined WINCE)
+#define CL_RUNTIME_EXPORT __declspec(dllexport)
+#else
+#define CL_RUNTIME_EXPORT
+#endif
+
 #include <map>
 #include <iostream>
 #include <limits>

modules/ocl/src/safe_call.hpp

@@ -46,11 +46,7 @@
 #ifndef __OPENCV_OPENCL_SAFE_CALL_HPP__
 #define __OPENCV_OPENCL_SAFE_CALL_HPP__
 
-#if defined __APPLE__
+#include "opencv2/ocl/cl_runtime/cl_runtime.hpp"
-#include <OpenCL/opencl.h>
-#else
-#include <CL/cl.h>
-#endif
 
 #if defined(__GNUC__)
 #define openCLSafeCall(expr)  ___openCLSafeCall(expr, __FILE__, __LINE__, __func__)