[RF] Avoid need for R__HAS_CUDA

In RooFit, there are checks if ROOT is build with CUDA only in the `.cxx` implementation files, not in the public headers. We therefore don't need a `R__HAS_CUDA` in the ROOT configuration. The `R__HAS_CUDA` flag was only introduced in ROOT 6.26 and it was never advocated to users, so it should be safe to remove it.
root-project · Oct 24, 2023 · a113110 · a113110
1 parent c80403e
commit a113110
Show file tree

Hide file tree

Showing 11 changed files with 54 additions and 52 deletions.
diff --git a/cmake/modules/RootConfiguration.cmake b/cmake/modules/RootConfiguration.cmake
@@ -365,11 +365,6 @@ if(CMAKE_USE_PTHREADS_INIT)
 else()
   set(haspthread undef)
 endif()
-if(cuda)
-  set(hascuda define)
-else()
-  set(hascuda undef)
-endif()
 if(x11)
   set(hasxft define)
 else()

diff --git a/config/RConfigure.in b/config/RConfigure.in
@@ -34,7 +34,6 @@
 #@setresuid@ R__HAS_SETRESUID   /**/
 #@hasmathmore@ R__HAS_MATHMORE   /**/
 #@haspthread@ R__HAS_PTHREAD    /**/
-#@hascuda@ R__HAS_CUDA    /**/
 #@hasxft@ R__HAS_XFT    /**/
 #@hascocoa@ R__HAS_COCOA    /**/
 #@hasvc@ R__HAS_VC    /**/

diff --git a/roofit/batchcompute/CMakeLists.txt b/roofit/batchcompute/CMakeLists.txt
@@ -13,6 +13,10 @@ ROOT_LINKER_LIBRARY(RooBatchCompute
     ${CudaDependencies}
 )
 
+if(cuda)
+  target_compile_definitions(RooBatchCompute PUBLIC ROOFIT_CUDA)
+endif()
+
 target_include_directories(RooBatchCompute INTERFACE $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/res>)
 
 if(vdt OR builtin_vdt)
@@ -68,5 +72,3 @@ if (cuda)
   ROOT_LINKER_LIBRARY(RooBatchCompute_CUDA  src/RooBatchCompute.cu src/ComputeFunctions.cu TYPE SHARED DEPENDENCIES RooBatchCompute)
   target_compile_options(RooBatchCompute_CUDA  PRIVATE -DRF_ARCH=CUDA -lineinfo --expt-relaxed-constexpr)
 endif()
-
-ROOT_INSTALL_HEADERS()
diff --git a/roofit/batchcompute/inc/RooBatchCompute.h → roofit/batchcompute/res/RooBatchCompute.h b/roofit/batchcompute/inc/RooBatchCompute.h → roofit/batchcompute/res/RooBatchCompute.h
@@ -17,7 +17,7 @@
 
 #include <RConfig.h>
 
-#ifdef R__HAS_CUDA
+#ifdef ROOFIT_CUDA
 #include <RooFit/Detail/CudaInterface.h>
 #endif
 
@@ -53,17 +53,15 @@ void init();
 /// the RooBatchCompute library.
 class Config {
 public:
-#ifdef R__HAS_CUDA
+#ifdef ROOFIT_CUDA
    bool useCuda() const { return _cudaStream != nullptr; }
    void setCudaStream(RooFit::Detail::CudaInterface::CudaStream *cudaStream) { _cudaStream = cudaStream; }
    RooFit::Detail::CudaInterface::CudaStream *cudaStream() const { return _cudaStream; }
-#else
-   bool useCuda() const { return false; }
-#endif
 
 private:
-#ifdef R__HAS_CUDA
    RooFit::Detail::CudaInterface::CudaStream *_cudaStream = nullptr;
+#else
+   bool useCuda() const { return false; }
 #endif
 };
 

diff --git a/roofit/batchcompute/src/Initialisation.cxx b/roofit/batchcompute/src/Initialisation.cxx
@@ -51,11 +51,11 @@ void init()
 
    const std::string userChoice = gEnv->GetValue("RooFit.BatchCompute", "auto");
 #ifdef R__RF_ARCHITECTURE_SPECIFIC_LIBS
-#ifdef R__HAS_CUDA
+#ifdef ROOFIT_CUDA
    if(gSystem->Load("libRooBatchCompute_CUDA") != 0) {
       RooBatchCompute::dispatchCUDA = nullptr;
    }
-#endif // R__HAS_CUDA
+#endif // ROOFIT_CUDA
 
    __builtin_cpu_init();
 #if __GNUC__ > 5 || defined(__clang__)

diff --git a/roofit/roofitcore/CMakeLists.txt b/roofit/roofitcore/CMakeLists.txt
@@ -488,15 +488,20 @@ ROOT_STANDARD_LIBRARY_PACKAGE(RooFitCore
   ${EXTRA_DICT_OPTS}
 )
 
+# The following definitions are PUBLIC so they can also be used in ROOT-internal tests
+
 if(roofit_legacy_eval_backend)
-  target_compile_definitions(RooFitCore PRIVATE ROOFIT_LEGACY_EVAL_BACKEND)
+  target_compile_definitions(RooFitCore PUBLIC ROOFIT_LEGACY_EVAL_BACKEND)
 endif()
 
 if(roofit_multiprocess)
-  # PUBLIC to it can also be used in tests
   target_compile_definitions(RooFitCore PUBLIC ROOFIT_MULTIPROCESS)
 endif()
 
+if(cuda)
+  target_compile_definitions(RooFitCore PUBLIC ROOFIT_CUDA)
+endif()
+
 target_include_directories(RooFitCore INTERFACE $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/res>)
 
 # For recent clang, this can facilitate auto-vectorisation.

diff --git a/roofit/roofitcore/inc/RooFit/Evaluator.h b/roofit/roofitcore/inc/RooFit/Evaluator.h
@@ -50,11 +50,9 @@ class Evaluator {
 private:
    void processVariable(NodeInfo &nodeInfo);
    void setClientsDirty(NodeInfo &nodeInfo);
-#ifdef R__HAS_CUDA
    std::span<const double> getValHeterogeneous();
    void markGPUNodes();
    void assignToGPU(NodeInfo &info);
-#endif
    void computeCPUNode(const RooAbsArg *node, NodeInfo &info);
    void setOperMode(RooAbsArg *arg, RooAbsArg::OperMode opMode);
    void syncDataTokens();
@@ -66,9 +64,7 @@ class Evaluator {
    int _nEvaluations = 0;
    bool _needToUpdateOutputSizes = false;
    RooFit::Detail::DataMap _dataMapCPU;
-#ifdef R__HAS_CUDA
    RooFit::Detail::DataMap _dataMapCUDA;
-#endif
    std::vector<NodeInfo> _nodes; // the ordered computation graph
    std::stack<std::unique_ptr<ChangeOperModeRAII>> _changeOperModeRAIIs;
 };

diff --git a/roofit/roofitcore/src/Buffers.cxx b/roofit/roofitcore/src/Buffers.cxx
@@ -17,7 +17,7 @@
 #include <queue>
 #include <map>
 
-#ifdef R__HAS_CUDA
+#ifdef ROOFIT_CUDA
 namespace CudaInterface = RooFit::Detail::CudaInterface;
 using CudaInterface::CudaStream;
 #endif
@@ -68,7 +68,7 @@ class CPUBufferContainer {
    std::vector<double> _vec;
 };
 
-#ifdef R__HAS_CUDA
+#ifdef ROOFIT_CUDA
 class GPUBufferContainer {
 public:
    GPUBufferContainer(std::size_t size) : _arr(size) {}
@@ -140,7 +140,7 @@ class PinnedBufferContainer {
    CudaStream *_cudaStream = nullptr;
    mutable LastAccessType _lastAccess = LastAccessType::CPU_READ;
 };
-#endif // R__HAS_CUDA
+#endif // ROOFIT_CUDA
 
 template <class Container>
 class BufferImpl : public AbsBuffer {
@@ -175,15 +175,15 @@ class BufferImpl : public AbsBuffer {
 
 using ScalarBuffer = BufferImpl<ScalarBufferContainer>;
 using CPUBuffer = BufferImpl<CPUBufferContainer>;
-#ifdef R__HAS_CUDA
+#ifdef ROOFIT_CUDA
 using GPUBuffer = BufferImpl<GPUBufferContainer>;
 using PinnedBuffer = BufferImpl<PinnedBufferContainer>;
 #endif
 
 struct BufferQueuesMaps {
    ScalarBuffer::QueuesMap scalarBufferQueuesMap;
    CPUBuffer::QueuesMap cpuBufferQueuesMap;
-#ifdef R__HAS_CUDA
+#ifdef ROOFIT_CUDA
    GPUBuffer::QueuesMap gpuBufferQueuesMap;
    PinnedBuffer::QueuesMap pinnedBufferQueuesMap;
 #endif
@@ -204,7 +204,7 @@ std::unique_ptr<AbsBuffer> BufferManager::makeCpuBuffer(std::size_t size)
 {
    return std::make_unique<CPUBuffer>(size, _queuesMaps->cpuBufferQueuesMap);
 }
-#ifdef R__HAS_CUDA
+#ifdef ROOFIT_CUDA
 std::unique_ptr<AbsBuffer> BufferManager::makeGpuBuffer(std::size_t size)
 {
    return std::make_unique<GPUBuffer>(size, _queuesMaps->gpuBufferQueuesMap);
@@ -215,7 +215,7 @@ std::unique_ptr<AbsBuffer> BufferManager::makePinnedBuffer(std::size_t size, Cud
    out->vec().setCudaStream(stream);
    return out;
 }
-#endif // R__HAS_CUDA
+#endif // ROOFIT_CUDA
 
 } // end namespace Detail
 } // end namespace RooFit
diff --git a/roofit/roofitcore/src/RooFit/Detail/Buffers.h b/roofit/roofitcore/src/RooFit/Detail/Buffers.h
@@ -15,7 +15,7 @@
 
 #include <RConfig.h>
 
-#ifdef R__HAS_CUDA
+#ifdef ROOFIT_CUDA
 #include <RooFit/Detail/CudaInterface.h>
 #endif
 
@@ -48,7 +48,7 @@ class BufferManager {
 
    std::unique_ptr<AbsBuffer> makeScalarBuffer();
    std::unique_ptr<AbsBuffer> makeCpuBuffer(std::size_t size);
-#ifdef R__HAS_CUDA
+#ifdef ROOFIT_CUDA
    std::unique_ptr<AbsBuffer> makeGpuBuffer(std::size_t size);
    std::unique_ptr<AbsBuffer>
    makePinnedBuffer(std::size_t size, RooFit::Detail::CudaInterface::CudaStream *stream = nullptr);