Adjusting CPU threads and ORT copmile definitions

Author: ChSonnabend

GPU/GPUTracking/Base/cuda/CMakeLists.txt

@@ -122,6 +122,19 @@ if(ALIGPU_BUILD_TYPE STREQUAL "O2")
       ${CMAKE_CURRENT_SOURCE_DIR}
     TARGETVARNAME targetName)
 
+  message("Compile definitions for ONNX runtime (CUDA):")
+  message(STATUS "ORT_ROCM_BUILD: ${ORT_ROCM_BUILD}")
+  message(STATUS "ORT_CUDA_BUILD: ${ORT_CUDA_BUILD}")
+  message(STATUS "ORT_MIGRAPHX_BUILD: ${ORT_MIGRAPHX_BUILD}")
+  message(STATUS "ORT_TENSORRT_BUILD: ${ORT_TENSORRT_BUILD}")
+
+  target_compile_definitions(${targetName} PRIVATE
+    GPUCA_HAS_ONNX=1
+    ORT_ROCM_BUILD=$<BOOL:${ORT_ROCM_BUILD}>
+    ORT_CUDA_BUILD=$<BOOL:${ORT_CUDA_BUILD}>
+    ORT_MIGRAPHX_BUILD=$<BOOL:${ORT_MIGRAPHX_BUILD}>
+    ORT_TENSORRT_BUILD=$<BOOL:${ORT_TENSORRT_BUILD}>)
+
   install(FILES ${HDRS} DESTINATION include/GPU)
 endif()
 

GPU/GPUTracking/Base/hip/CMakeLists.txt

@@ -160,6 +160,19 @@ if(ALIGPU_BUILD_TYPE STREQUAL "O2")
       ${GPUCA_HIP_SOURCE_DIR}
     TARGETVARNAME targetName)
 
+  message("Compile definitions for ONNX runtime (HIP / ROCM):")
+  message(STATUS "ORT_ROCM_BUILD: ${ORT_ROCM_BUILD}")
+  message(STATUS "ORT_CUDA_BUILD: ${ORT_CUDA_BUILD}")
+  message(STATUS "ORT_MIGRAPHX_BUILD: ${ORT_MIGRAPHX_BUILD}")
+  message(STATUS "ORT_TENSORRT_BUILD: ${ORT_TENSORRT_BUILD}")
+
+  target_compile_definitions(${targetName} PRIVATE
+    GPUCA_HAS_ONNX=1
+    ORT_ROCM_BUILD=$<BOOL:${ORT_ROCM_BUILD}>
+    ORT_CUDA_BUILD=$<BOOL:${ORT_CUDA_BUILD}>
+    ORT_MIGRAPHX_BUILD=$<BOOL:${ORT_MIGRAPHX_BUILD}>
+    ORT_TENSORRT_BUILD=$<BOOL:${ORT_TENSORRT_BUILD}>)
+
   install(FILES ${HDRS} DESTINATION include/GPU)
 
 #  o2_add_test(GPUsortHIP NAME test_GPUsortHIP

GPU/GPUTracking/CMakeLists.txt

@@ -352,19 +352,8 @@ if(ALIGPU_BUILD_TYPE STREQUAL "O2")
     ${targetName}
     PRIVATE $<TARGET_PROPERTY:O2::Framework,INTERFACE_INCLUDE_DIRECTORIES>)
 
-  message("Compile definitions for ONNX runtime:")
-  message(STATUS "ORT_ROCM_BUILD: ${ORT_ROCM_BUILD}")
-  message(STATUS "ORT_CUDA_BUILD: ${ORT_CUDA_BUILD}")
-  message(STATUS "ORT_MIGRAPHX_BUILD: ${ORT_MIGRAPHX_BUILD}")
-  message(STATUS "ORT_TENSORRT_BUILD: ${ORT_TENSORRT_BUILD}")
-
-
   target_compile_definitions(${targetName} PRIVATE
-    GPUCA_O2_LIB GPUCA_TPC_GEOMETRY_O2 GPUCA_HAS_ONNX=1
-    ORT_ROCM_BUILD=$<BOOL:${ORT_ROCM_BUILD}>
-    ORT_CUDA_BUILD=$<BOOL:${ORT_CUDA_BUILD}>
-    ORT_MIGRAPHX_BUILD=$<BOOL:${ORT_MIGRAPHX_BUILD}>
-    ORT_TENSORRT_BUILD=$<BOOL:${ORT_TENSORRT_BUILD}>)
+    GPUCA_O2_LIB GPUCA_TPC_GEOMETRY_O2 GPUCA_HAS_ONNX=1)
 
   o2_target_root_dictionary(${MODULE}
                             HEADERS ${HDRS_CINT_O2} ${HDRS_CINT_O2_ADDITIONAL}

GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx

@@ -626,7 +626,7 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
   if (GetProcessingSettings().nn.applyNNclusterizer) {
     int32_t deviceId = -1;
     int32_t numLanes = GetProcessingSettings().nTPCClustererLanes;
-    int32_t maxThreads = mRec->MemoryScalers()->nTPCdigits / 6000;
+    int32_t maxThreads = mRec->getNKernelHostThreads(true);
     mRec->runParallelOuterLoop(doGPU, numLanes, [&](uint32_t lane) {
       nnApplications[lane].init(nn_settings);
       if (nnApplications[lane].modelsUsed[0]) {