added NVIDIA kernel compatible build environment

DepthDeluxe
1 parent 97c9bd44
Showing 11 changed files with 35 additions and 141 deletions
CMakeLists.txt
app/br/CMakeLists.txt
openbr/CMakeLists.txt
openbr/cuda-plugins/cglg/customthreshold.cu
openbr/cuda-plugins/cuda-plugins.cmake
openbr/cuda-plugins/cglg/copyfrom.cpp → openbr/plugins/cuda/copyfrom.cpp
openbr/cuda-plugins/cglg/copyto.cpp → openbr/plugins/cuda/copyto.cpp
openbr/cuda-plugins/cglg/passthrough.cpp → openbr/plugins/cuda/passthrough.cpp
openbr/plugins/cuda/passthrough.cu
openbr/cuda-plugins/cglg/threshold.cpp → openbr/plugins/cuda/threshold.cpp
openbr/plugins/plugins.cmake
@@ -108,7 +108,6 @@ endif()
 if(UNIX)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wextra -Wno-strict-overflow -Wno-comment -Wno-unknown-pragmas -fvisibility=hidden -fno-omit-frame-pointer")
   set(CMAKE_INSTALL_RPATH ${CMAKE_INSTALL_PREFIX}/lib ${_qt5Core_install_prefix}/lib)
-  set(CUDA_NVCC_FLAGS "") # put NVCC compiler flags here
   if(NOT APPLE)
     if(${CMAKE_CXX_COMPILER} STREQUAL "/opt/intel/bin/icpc")
       set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -static-intel -wd2196")
@@ -8,4 +8,3 @@ qt5_use_modules(br ${QT_DEPENDENCIES})
 install(TARGETS br RUNTIME DESTINATION bin)
 add_test(NAME br_initialize WORKING_DIRECTORY ${PROJECT_SOURCE_DIR} COMMAND br)
-
@@ -8,7 +8,12 @@ set(SRC openbr.cpp
         universal_template.cpp)
 aux_source_directory(core BR_CORE)
 include(plugins/plugins.cmake)
-include(cuda-plugins/cuda-plugins.cmake)
+
+# CUDA shit
+FIND_PACKAGE(CUDA REQUIRED)
+SET(CUDA_SEPARABLE_COMPILATION ON)
+SET(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -arch=sm_20;--compiler-options "-fPIC")
+SET(CUDA_VERBOSE_BUILD ON)
 # Janus API
 option(BR_WITH_JANUS "Build IARPA Janus related applications." ON)
@@ -29,25 +34,18 @@ if(NOT BR_EMBEDDED)
   install(FILES ${HEADERS} DESTINATION include/openbr/gui)
 endif()
-# normal BR library declaration
-add_library(openbr SHARED ${SRC} ${BR_CORE} ${BR_JANUS} ${BR_GUI} ${BR_ICONS} ${BR_THIRDPARTY_SRC} ${BR_RESOURCES} ${NATURALSTRINGCOMPARE_SRC})
+# normal BR library declaration - added openbr-cuda library
+message(STATUS "BR_THIRDPARTY_SRC")
+message(STATUS ${BR_THIRDPARTY_SRC})
+cuda_add_library(openbr SHARED ${SRC} ${BR_CORE} ${BR_JANUS} ${BR_GUI} ${BR_ICONS} ${BR_THIRDPARTY_SRC} ${BR_RESOURCES} ${NATURALSTRINGCOMPARE_SRC})
 qt5_use_modules(openbr ${QT_DEPENDENCIES})
 set_target_properties(openbr PROPERTIES
                              DEFINE_SYMBOL BR_LIBRARY
                              VERSION ${CPACK_PACKAGE_VERSION_MAJOR}.${CPACK_PACKAGE_VERSION_MINOR}.${CPACK_PACKAGE_VERSION_PATCH}
                              SOVERSION ${CPACK_PACKAGE_VERSION_MAJOR}.${CPACK_PACKAGE_VERSION_MINOR})
-target_link_libraries(openbr ${BR_THIRDPARTY_LIBS})
+target_link_libraries(openbr ${BR_THIRDPARTY_LIBS} ${CUDA_LIBRARIES})
 add_cppcheck(openbr)
-# CUDA portion of BR
-
-cuda_add_library(openbr-cuda SHARED ${BR_THIRDPARTY_CUDA_SRC})
-qt5_use_modules(openbr-cuda ${QT_DEPENDENCIES})
-set_target_properties(openbr-cuda PROPERTIES
-                                  DEFINE_SYMBOL BR_CUDA_LIBRARY
-                                  VERSION ${CPACK_PACKAGE_VERSION_MAJOR}.${CPACK_PACKAGE_VERSION_MINOR}.${CPACK_PACKAGE_VERSION_PATCH}
-                                SOVERSION ${CPACK_PACKAGE_VERSION_MAJOR}.${CPACK_PACKAGE_VERSION_MINOR})
-
 # Janus implementation
 if(BR_WITH_JANUS)
   set(JANUS_BUILD_PP5_WRAPPER ${BR_WITH_PP5} CACHE BOOL "Build Janus implementation using PittPatt 5")
-/*
-#include <iostream>
-
-// external opencv CUDA interface
-#include <opencv2/opencv.hpp>
-#include <opencv2/gpu/gpu.hpp>
-
-// internal CUDA stuff
-#include <opencv2/core/cuda/common.hpp>
-#include <opencv2/core/cuda/emulation.hpp>
-#include <opencv2/core/cuda/transform.hpp>
-#include <opencv2/core/cuda/functional.hpp>
-#include <opencv2/core/cuda/utility.hpp>
-
-#include <openbr/plugins/openbr_internal.h>
-
-using namespace std;
-
-using namespace cv;
-using namespace cv::gpu;
-using namespace cv::cuda;
-using namespace cv::cuda::device;
-
-namespace br
-{
-  class CUDACustomThresholdTransform : public UntrainableTransform
-  {
-    Q_OBJECT
-
-private:
-    void project(const Template &src, Template &dst) const
-    {
-      // get the mat to send to the GPU
-      GpuMat gpuMat_src, gpuMat_dst;
-
-      try
-      {
-        // copy the contents to the GPU
-        gpuMat_src.upload(src.m());
-
-        threshold(gpuMat_src, gpuMat_dst, 128.0, 255.0, CV_THRESH_BINARY);
-
-        gpuMat_dst.download(dst.m());
-      }
-      catch(const cv::Exception& ex)
-      {
-        cout << "Error: " << ex.what() << endl;
-      }
-    }
-  };
-
-  BR_REGISTER(Transform, CUDACustomThresholdTransform);
-
-  namespace cuda { namespace customthreshold {
-    texture<uchar, cudaTextureType2D, cudaReadModeElementType> tex_src(false, cudaFilterModePoint, cudaAddressModeClamp);
-    struct SrcTex {
-      __host__ SrcTex(int _xoff, int _yoff) : xoff(_xoff), yoff(_yoff) {}
-      __device__ __forceinline__ int operator ()(int y, int x) const {
-        return tex2D(tex_src, x + xoff, y + yoff);
-      }
-    }
-    __global__ void testKernel(const SrcTex src) {
-      const int x = blockIdx.x * blockDim.x + threadIdx.x;
-      const int y = blockIdx.y * blockDim.y + threadIdx.y;
-
-      src(x, y) = 1;sajfflksajlkfjdsalkfjsadjflkdsaf
-    }
-  }
-}
-
-#include "cglg/customthreshold.moc"
-*/
-# Optional Appendable CMake Variables:
-# BR_THIRDPARTY_CUDA_PLUGINS     - Additional plugins
-# BR_THIRDPARTY_CUDA_PLUGINS_DIR - Additional folder(s) of plugins
-# BR_EXCLUDED_CUDA_PLUGINS       - Plugins that should not be built
-# BR_THIRDPARTY_SRC         - Additional source code needed by a plugin
-# BR_THIRDPARTY_LIBS        - Additional libaries needed by a plugin
-
-# Also look for CMake modules in the thirdparty plugins folder(s)
-set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} ${BR_THIRDPARTY_CUDA_PLUGINS_DIR})
-
-# Gather all of the plugin subdirectories
-file(GLOB SUBFILES cuda-plugins/*)
-foreach(FILE ${SUBFILES})
-  if(IS_DIRECTORY ${FILE})
-    set(BR_CUDA_PLUGINS_DIR ${BR_CUDA_PLUGINS_DIR} ${FILE})
-  endif()
-endforeach()
-set(BR_CUDA_PLUGINS_DIR ${BR_CUDA_PLUGINS_DIR} cuda-plugins/) # Remove this when finished with reorg
-
-# Exclude pertinent plugins based on .cmake files
-mark_as_advanced(BR_EXCLUDED_CUDA_PLUGINS)
-foreach(DIR cuda-plugins/cmake ${BR_THIRDPARTY_CUDA_PLUGINS_DIR})
-  file(GLOB CMAKE_FILES ${DIR}/*.cmake)
-  foreach(CMAKE_FILE ${CMAKE_FILES})
-    if (NOT ${CMAKE_FILE} MATCHES "Find.*cmake")
-      include(${CMAKE_FILE})
-    endif()
-  endforeach()
-endforeach()
-
-# Collect all source files except for excluded plugins
-foreach(DIR ${BR_CUDA_PLUGINS_DIR} ${BR_THIRDPARTY_CUDA_PLUGINS_DIR})
-  get_filename_component(DIR_NAME ${DIR} NAME)
-  file(GLOB CUDA_PLUGINS ${DIR}/*.cu ${DIR}/*.cpp ${DIR}/*.h)
-  foreach(CUDA_PLUGIN ${CUDA_PLUGINS})
-    get_filename_component(CUDA_PLUGIN_NAME ${CUDA_PLUGIN} NAME)
-    set(EXCLUDE FALSE)
-    foreach(EXCLUDED_CUDA_PLUGIN ${BR_EXCLUDED_CUDA_PLUGINS})
-      get_filename_component(EXCLUDED_CUDA_PLUGIN_NAME ${EXCLUDED_CUDA_PLUGIN} NAME)
-      if (${CUDA_PLUGIN_NAME} STREQUAL ${EXCLUDED_CUDA_PLUGIN_NAME})
-        set(EXCLUDE TRUE)
-      endif()
-    endforeach()
-    if(NOT ${EXCLUDE})
-      set(BR_THIRDPARTY_CUDA_PLUGINS ${BR_THIRDPARTY_CUDA_PLUGINS} ${CUDA_PLUGIN})
-    endif()
-  endforeach()
-endforeach()
-
-set(BR_THIRDPARTY_CUDA_SRC ${BR_THIRDPARTY_CUDA_SRC} ${BR_THIRDPARTY_CUDA_PLUGINS})
-#include <opencv2/imgproc/imgproc.hpp>
-
 #include <openbr/plugins/openbr_internal.h>
+#include <opencv2/imgproc/imgproc.hpp>
+#include <opencv2/gpu/gpu.hpp>
+
 using namespace cv;
+using namespace cv::gpu;
+
+extern void br_cuda_device_wrapper();
 namespace br
 {
@@ -13,7 +17,16 @@ namespace br
 private:
     void project(const Template &src, Template &dst) const
     {
-      dst = src;
+      // upload the src mat to the GPU
+      GpuMat srcGpuMat, dstGpuMat;
+      srcGpuMat.upload(src.m());
+      dstGpuMat.upload(src.m());
+
+      br_cuda_device_wrapper();
+
+      dstGpuMat.download(dst.m());
+
+      // TODO(colin): add delete code
     }
   };
+__global__ void br_cuda_device_kernel() {
+
+}
+
+void br_cuda_device_wrapper() {
+  br_cuda_device_kernel<<<1,1>>>();
+}
@@ -31,7 +31,7 @@ endforeach()
 # Collect all source files except for excluded plugins
 foreach(DIR ${BR_PLUGINS_DIR} ${BR_THIRDPARTY_PLUGINS_DIR})
   get_filename_component(DIR_NAME ${DIR} NAME)
-  file(GLOB PLUGINS ${DIR}/*.cpp ${DIR}/*.h)
+  file(GLOB PLUGINS ${DIR}/*.cpp ${DIR}/*.cu ${DIR}/*.h)
   foreach(PLUGIN ${PLUGINS})
     get_filename_component(PLUGIN_NAME ${PLUGIN} NAME)
     set(EXCLUDE FALSE)