AliceO2Group · davidrohr · Oct 14, 2021 · Jul 26, 2021 · Jul 26, 2021 · Jul 27, 2021
@@ -9,49 +9,12 @@
 # granted to it by virtue of its status as an Intergovernmental Organization
 # or submit itself to any jurisdiction.
 
-set(HDRS_INSTALL ../Shared/Kernels.h)
-
 if(CUDA_ENABLED)
-  o2_add_executable(gpu-memory-benchmark-cuda
-                  SOURCES benchmark.cxx
-                          cuda/Kernels.cu
-                  PUBLIC_LINK_LIBRARIES Boost::program_options
-                                        ROOT::Tree
-                  TARGETVARNAME targetName)
+  add_subdirectory(cuda)
 endif()
 
 if(HIP_ENABLED)
-  # Hipify-perl
-  set(HIPIFY_EXECUTABLE "/opt/rocm/bin/hipify-perl")
-
-  set(HIP_KERNEL "Kernels.hip.cxx")
-  set(CU_KERNEL ${CMAKE_CURRENT_SOURCE_DIR}/cuda/Kernels.cu)
-  set(HIP_KERNEL_PATH "${CMAKE_CURRENT_SOURCE_DIR}/hip/${HIP_KERNEL}")
-
-  if(EXISTS ${HIPIFY_EXECUTABLE})
-    set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS ${CU_KERNEL})
-    add_custom_command(
-      OUTPUT ${HIP_KERNEL_PATH}
-      COMMAND ${HIPIFY_EXECUTABLE} --quiet-warnings ${CU_KERNEL} | sed '1{/\#include \"hip\\/hip_runtime.h\"/d}' > ${HIP_KERNEL_PATH}
-    )
-    set(CMAKE_CXX_COMPILER ${HIP_HIPCC_EXECUTABLE})
-    set(CMAKE_CXX_LINKER   ${HIP_HIPCC_EXECUTABLE})
-    set(CMAKE_CXX_EXTENSIONS OFF)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${O2_HIP_CMAKE_CXX_FLAGS} -fgpu-rdc")
-
-    o2_add_executable(gpu-memory-benchmark-hip
-                      SOURCES benchmark.cxx
-                              hip/Kernels.hip.cxx
-                      PUBLIC_LINK_LIBRARIES hip::host
-                                            Boost::program_options
-                                            ROOT::Tree
-                      TARGETVARNAME targetName)
-
-    if(HIP_AMDGPUTARGET)
-      # Need to add gpu target also to link flags due to gpu-rdc option
-      target_link_options(${targetName} PUBLIC --amdgpu-target=${HIP_AMDGPUTARGET})
-    endif()
-  endif()
+  add_subdirectory(hip)
 endif()
 
-o2_add_test_root_macro(macro/showBenchmarks.C)
+o2_add_test_root_macro(macro/showBenchmarks.C)
@@ -27,7 +27,7 @@ namespace o2
 namespace benchmark
 {
 
-template <class chunk_type>
+template <class chunk_t>
 class GPUbenchmark final
 {
  public:
@@ -41,13 +41,22 @@ class GPUbenchmark final
 
   // Single stream synchronous (sequential kernels) execution
   template <typename... T>
-  float benchmarkSync(void (*kernel)(T...),
-                      int nLaunches, int blocks, int threads, T&... args);
+  float runSequential(void (*kernel)(chunk_t*, size_t, T...),
+                      std::pair<int, int>& chunkRanges,
+                      int nLaunches,
+                      int dimGrid,
+                      int dimBlock,
+                      T&... args);
 
   // Multi-streams asynchronous executions on whole memory
   template <typename... T>
-  std::vector<float> benchmarkAsync(void (*kernel)(int, T...),
-                                    int nStreams, int nLaunches, int blocks, int threads, T&... args);
+  std::vector<float> runConcurrent(void (*kernel)(chunk_t*, size_t, T...),
+                                   std::vector<std::pair<int, int>>& chunkRanges,
+                                   int nLaunches,
+                                   int dimStreams,
+                                   int nBlocks,
+                                   int nThreads,
+                                   T&... args);
 
   // Main interface
   void globalInit();     // Allocate scratch buffers and compute runtime parameters
@@ -56,27 +65,14 @@ class GPUbenchmark final
   void printDevices();   // Dump info
 
   // Initializations/Finalizations of tests. Not to be measured, in principle used for report
-  void readInit();
-  void readFinalize();
+  void initTest(Test);
+  void finalizeTest(Test);
 
-  void writeInit();
-  void writeFinalize();
-
-  void copyInit();
-  void copyFinalize();
-
-  // Kernel calling wrappers
-  void readSequential(SplitLevel sl);
-  void readConcurrent(SplitLevel sl, int nRegions = 2);
-
-  void writeSequential(SplitLevel sl);
-  void writeConcurrent(SplitLevel sl, int nRegions = 2);
-
-  void copySequential(SplitLevel sl);
-  void copyConcurrent(SplitLevel sl, int nRegions = 2);
+  // Kernel calling wrapper
+  void runTest(Test, Mode, KernelConfig);
 
  private:
-  gpuState<chunk_type> mState;
+  gpuState<chunk_t> mState;
   std::shared_ptr<ResultWriter> mResultWriter;
   benchmarkOpts mOptions;
 };