From 789b02ad766bd22f5a134caed710da0af5777c88 Mon Sep 17 00:00:00 2001
From: Andrey Morozov <no@email>
Date: Wed, 11 Aug 2010 15:52:55 +0000
Subject: [PATCH] added choise in compute capability

---
 modules/gpu/CMakeLists.txt | 65 ++++++++++++++++++++++++++------------
 1 file changed, 44 insertions(+), 21 deletions(-)

diff --git a/modules/gpu/CMakeLists.txt b/modules/gpu/CMakeLists.txt
index f6a5835ee6..1617517c25 100644
--- a/modules/gpu/CMakeLists.txt
+++ b/modules/gpu/CMakeLists.txt
@@ -1,7 +1,8 @@
 
-set(name "gpu")	
+set(name "gpu")
 set(DEPS "opencv_core")
-	
+
+set(OPENCV_LINKER_LIBS ${OPENCV_LINKER_LIBS} opencv_gpu)
 
 set(the_target "opencv_${name}")
 
@@ -15,55 +16,77 @@ include_directories("${CMAKE_CURRENT_SOURCE_DIR}/include"
 					"${CMAKE_CURRENT_BINARY_DIR}")
 
 foreach(d ${DEPS})
-	if(${d} MATCHES "opencv_")			
+	if(${d} MATCHES "opencv_")
 		string(REPLACE "opencv_" "${CMAKE_CURRENT_SOURCE_DIR}/../" d_dir ${d})
-		include_directories("${d_dir}/include")			
+		include_directories("${d_dir}/include")
 	endif()
-endforeach()		
+endforeach()
 
 file(GLOB lib_srcs "src/*.cpp")
 file(GLOB lib_int_hdrs "src/*.h*")
 file(GLOB lib_cuda "src/cuda/*.cu*")
-file(GLOB lib_cuda_hdrs "src/cuda/*.h*")	
+file(GLOB lib_cuda_hdrs "src/cuda/*.h*")
 source_group("Src\\Host" FILES ${lib_srcs} ${lib_int_hdrs})
 source_group("Src\\Cuda" FILES ${lib_cuda} ${lib_cuda_hdrs})
 
-file(GLOB lib_hdrs "include/opencv2/${name}/*.h*")		
+file(GLOB lib_hdrs "include/opencv2/${name}/*.h*")
 source_group("Include" FILES ${lib_hdrs})
 
 if (HAVE_CUDA)
 	include_directories(${CUDA_INCLUDE_DIRS})
 	link_directories(${CUDA_LIBRARIES})
-	
-	set (CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-arch=sm_13")
+
+	set(CUDA_NVCC_FLAGS_ARCH "1.2 1.3 2.0" CACHE STRING "Add or remove compute copability" FORCE)
+
+	set(CUDA_NVCC_FLAGS_NUM "")
+
+	while(NOT ${CUDA_NVCC_FLAGS_ARCH} STREQUAL "")
+	    string(REGEX MATCH "[0-9]+.[0-9]+" RESULT_NUM ${CUDA_NVCC_FLAGS_ARCH})
+	    string(REGEX MATCHALL "[0-9]" RESULT_STR ${RESULT_NUM})
+	    string(REGEX REPLACE ";" "\ " RESULT ${RESULT_STR})
+	    list(APPEND CUDA_NVCC_FLAGS_NUM ${RESULT})
+	    string(REGEX REPLACE "${RESULT_NUM}" "\ " CUDA_NVCC_FLAGS_ARCH_STR ${CUDA_NVCC_FLAGS_ARCH})
+	    string(STRIP ${CUDA_NVCC_FLAGS_ARCH_STR} CUDA_NVCC_FLAGS_ARCH)
+	endwhile()
+
+	set (loop_var "")
+	foreach( loop_var IN LISTS CUDA_NVCC_FLAGS_NUM)
+	    set (CUDA_NVCC_FLAGS
+		"${CUDA_NVCC_FLAGS} -gencode=arch=compute_${loop_var},code=\"sm_${loop_var},compute_${loop_var}\"")
+	endforeach()
+
+	set (CUDA_NVCC_FLAGS_ARCH ${CUDA_NVCC_FLAGS_ARCH_STR} CACHE STRING "Add or remove compute copability" FORCE)
+
+	message(STATUS " CUDA_NVCC_FLAGS: ${CUDA_NVCC_FLAGS} ")
+
 
 	if (UNIX OR APPLE)
 		set (CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS}  "-Xcompiler;-fPIC;")
 		#set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}" "-fPIC")
-			
-	endif() 
-			
+
+	endif()
+
 	#set (CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-keep")
 	#set (CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS}  "-Xcompiler;/EHsc-;")
-	
+
 	if(MSVC)
 		#string(REPLACE "/W4" "/W3" CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
 		#string(REPLACE "/W4" "/W3" CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE}")
-		#string(REPLACE "/W4" "/W3" CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG}")						
-		set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd4211 /wd4201 /wd4100 /wd4505 /wd4408")			
+		#string(REPLACE "/W4" "/W3" CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG}")
+		set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd4211 /wd4201 /wd4100 /wd4505 /wd4408")
 
 		string(REPLACE "/EHsc-" "/EHs" CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
 		string(REPLACE "/EHsc-" "/EHs" CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE}")
 		string(REPLACE "/EHsc-" "/EHs" CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG}")
 		string(REPLACE "/EHsc-" "/EHs" CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
 		string(REPLACE "/EHsc-" "/EHs" CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE}")
-		string(REPLACE "/EHsc-" "/EHs" CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG}")		
+		string(REPLACE "/EHsc-" "/EHs" CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG}")
 	endif()
-			
+
 	CUDA_COMPILE(cuda_objs ${lib_cuda})
 	#CUDA_BUILD_CLEAN_TARGET()
 endif()
-	
+
 
 add_library(${the_target} ${lib_srcs} ${lib_hdrs} ${lib_int_hdrs} ${lib_cuda} ${lib_cuda_hdrs} ${cuda_objs})
 
@@ -73,7 +96,7 @@ if(PCHSupport_FOUND)
 	if(${CMAKE_GENERATOR} MATCHES "Visual*" OR ${CMAKE_GENERATOR} MATCHES "Xcode*")
 		if(${CMAKE_GENERATOR} MATCHES "Visual*")
 			set(${the_target}_pch "src/precomp.cpp")
-		endif()            
+		endif()
 		add_native_precompiled_header(${the_target} ${pch_header})
 	elseif(CMAKE_COMPILER_IS_GNUCXX AND ${CMAKE_GENERATOR} MATCHES ".*Makefiles")
 		add_precompiled_header(${the_target} ${pch_header})
@@ -84,7 +107,7 @@ endif()
 set_target_properties(${the_target} PROPERTIES
 	VERSION ${OPENCV_VERSION}
 	SOVERSION ${OPENCV_SOVERSION}
-	OUTPUT_NAME "${the_target}${OPENCV_DLLVERSION}"	
+	OUTPUT_NAME "${the_target}${OPENCV_DLLVERSION}"
 	)
 
 # Additional target properties
@@ -92,7 +115,7 @@ set_target_properties(${the_target} PROPERTIES
 	DEBUG_POSTFIX "${OPENCV_DEBUG_POSTFIX}"
 	ARCHIVE_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/lib/"
 	RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/bin/"
-	INSTALL_NAME_DIR "${CMAKE_INSTALL_PREFIX}/lib"	
+	INSTALL_NAME_DIR "${CMAKE_INSTALL_PREFIX}/lib"
 	)
 
 # Add the required libraries for linking: