Fix the configuration to find Blas and decide either to build CPU or GPU architecture. Define this in RCondifure.h

Author: lmoneta

cmake/modules/RootBuildOptions.cmake

@@ -167,6 +167,8 @@ ROOT_BUILD_OPTION(table OFF "Build libTable contrib library")
 ROOT_BUILD_OPTION(tcmalloc OFF "Using the tcmalloc allocator")
 ROOT_BUILD_OPTION(thread ON "Using thread library (cannot be disabled)")
 ROOT_BUILD_OPTION(tmva ON "Build TMVA multi variate analysis library")
+ROOT_BUILD_OPTION(tmva-cpu ON "Build TMVA with CPU support for deep learning. Requires BLAS")
+ROOT_BUILD_OPTION(tmva-gpu ON "Build TMVA with GPU support for deep learning. Requries CUDA")
 ROOT_BUILD_OPTION(unuran OFF "UNURAN - package for generating non-uniform random numbers")
 ROOT_BUILD_OPTION(vc OFF "Vc adds a few new types for portable and intuitive SIMD programming")
 ROOT_BUILD_OPTION(vdt OFF "VDT adds a set of fast and vectorisable mathematical functions")

cmake/modules/RootConfiguration.cmake

@@ -565,6 +565,17 @@ if(qt5web)
 else()
   set(hasqt5webengine undef)
 endif()
+if (tmva AND imt AND BLAS_FOUND)
+  set(hastmvacpu define)
+else()
+  set(hastmvacpu undef)
+endif()  
+if (tmva AND CUDA_FOUND)
+  set(hastmvagpu define)
+else()
+  set(hastmvagpu undef)
+endif()  
+
 
 CHECK_CXX_SOURCE_COMPILES("#include <string_view>
   int main() { char arr[3] = {'B', 'a', 'r'}; std::string_view strv(arr, sizeof(arr)); return 0;}" found_stdstringview)

cmake/modules/SearchInstalledSoftware.cmake

@@ -1513,6 +1513,7 @@ if(tmva AND cuda)
       message(STATUS "CUDA not found. Ensure that the installation of CUDA is in the CMAKE_PREFIX_PATH")
       message(STATUS "                For the time being switching OFF 'cuda' option")
       set(cuda OFF CACHE BOOL "" FORCE)
+      set(tmva-cuda OFF CACHE BOOL "" FORCE)
     endif()
   endif()
 endif()
@@ -1522,6 +1523,13 @@ if(tmva AND imt)
   find_package(BLAS)
 endif()
 
+if(NOT BLAS_FOUND)
+  set(tmva-cpu OFF CACHE BOOL "" FORCE)
+endif()
+if(NOT CUDA_FOUND)
+  set(tmva-gpu OFF CACHE BOOL "" FORCE)
+endif()
+
 
 #---Download googletest--------------------------------------------------------------
 if (testing)

config/RConfigure.in

@@ -55,5 +55,7 @@
 #@usezlib@ R__HAS_DEFAULT_ZLIB  /**/
 #@uselzma@ R__HAS_DEFAULT_LZMA  /**/
 
+#@hastmvacpu@ R__HAS_TMVACPU /**/
+#@hastmvagpu@ R__HAS_TMVAGPU /**/
 
 #endif

test/stressTMVA.cxx

@@ -3049,12 +3049,12 @@ void addClassificationTests( UnitTestSuite& TMVA_test, bool full=true)
    TString configCpu      = "Architecture=CPU:" + config;
    TString configGpu      = "Architecture=GPU:" + config;
 
-#ifdef DNNCPU
+#ifdef R__HAS_TMVACPU
    TMVA_test.addTest(new MethodUnitTestWithROCLimits(
                          TMVA::Types::kDNN, "DNN CPU", configCpu, 0.85, 0.98)
                      );
 #endif
-#ifdef DNNCUDA
+#ifdef R__HAS_TMVAGPU
    TMVA_test.addTest(new MethodUnitTestWithROCLimits(
                          TMVA::Types::kDNN, "DNN GPU", configGpu, 0.85, 0.98)
                      );

tmva/tmva/CMakeLists.txt

@@ -73,14 +73,14 @@ endif()
 #---Handle BLAS dependent code. -----------------
 if(BLAS_FOUND AND imt)
     message(STATUS "Using TMVA-DNN with BLAS installation") 
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DDNNCPU")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
   set(DNN_CPU_LIBRARIES MathCore Matrix ${BLAS_LINKER_FLAGS} ${BLAS_LIBRARIES} ${TBB_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
   include_directories(SYSTEM ${TBB_INCLUDE_DIRS})
 else()
 if (mathmore AND imt)
 #use GSL cblas installation
   message(STATUS "Using TMVA-DNN with gslcblas installation") 
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DDNNCPU -DDNN_USE_CBLAS")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DDNN_USE_CBLAS")
   set(DNN_CPU_LIBRARIES MathCore Matrix ${TBB_LIBRARIES} ${GSL_LIBRARIES}  ${CMAKE_THREAD_LIBS_INIT})
   include_directories(SYSTEM ${TBB_INCLUDE_DIRS} ${GSL_INCLUDE_DIR} )
 else()

tmva/tmva/inc/TMVA/MethodDL.h

@@ -43,11 +43,11 @@
 
 #include "TMVA/DNN/Architectures/Reference.h"
 
-#ifdef DNNCPU
+#ifdef R__HAS_TMVACPU
 #include "TMVA/DNN/Architectures/Cpu.h"
 #endif
 
-#ifdef DNNCUDA
+#ifdef R__HAS_TMVACUDA
 #include "TMVA/DNN/Architectures/Cuda.h"
 #endif
 
@@ -77,13 +77,13 @@ class MethodDL : public MethodBase {
 private:
    // Key-Value vector type, contining the values for the training options
    using KeyValueVector_t = std::vector<std::map<TString, TString>>;
-#ifdef DNNCPU
-   using ArchitectureCpu_t = TMVA::DNN::TCpu<Double_t>;
+#ifdef R__HAS_TMVACPU
+   using ArchitectureImpl_t = TMVA::DNN::TCpu<Double_t>;
 #else
-   using ArchitectureCpu_t = TMVA::DNN::TReference<Double_t>;
+   using ArchitectureImpl_t = TMVA::DNN::TReference<Double_t>;
 #endif  
-   using DeepNetCpu_t = TMVA::DNN::TDeepNet<ArchitectureCpu_t>;
-   std::unique_ptr<DeepNetCpu_t> fNet;
+   using DeepNetImpl_t = TMVA::DNN::TDeepNet<ArchitectureImpl_t>;
+   std::unique_ptr<DeepNetImpl_t> fNet;
 
    /*! The option handling methods */
    void DeclareOptions();
@@ -199,6 +199,8 @@ class MethodDL : public MethodBase {
    size_t GetBatchHeight() const { return fBatchHeight; }
    size_t GetBatchWidth() const { return fBatchWidth; }
 
+   const DeepNetImpl_t & GetDeepNet() const { return *fNet; }
+
    DNN::EInitialization GetWeightInitialization() const { return fWeightInitialization; }
    DNN::EOutputFunction GetOutputFunction() const { return fOutputFunction; }
    DNN::ELossFunction GetLossFunction() const { return fLossFunction; }

tmva/tmva/inc/TMVA/MethodDNN.h

@@ -52,6 +52,13 @@
 #include "TMVA/DNN/Minimizers.h"
 #include "TMVA/DNN/Architectures/Reference.h"
 
+#ifdef R__HAS_TMVACPU
+#define DNNCPU
+#endif
+#ifdef R__HAS_TMVAGPU
+#define DNNCUDA
+#endif
+
 #ifdef DNNCPU
 #include "TMVA/DNN/Architectures/Cpu.h"
 #endif