GPU CMake: Create const param files for available default architectures to be loaded by RTC

davidrohr · davidrohr · commit 68c73f4ca4c8 · 2025-04-24T22:44:53.000+02:00
diff --git a/GPU/GPUTracking/CMakeLists.txt b/GPU/GPUTracking/CMakeLists.txt
@@ -407,6 +407,30 @@ target_sources(${targetName}
                FILES ${GENERATED_HEADERS_LIST}
                BASE_DIRS ${CMAKE_CURRENT_BINARY_DIR})
 
+make_directory(${CMAKE_CURRENT_BINARY_DIR}/genGPUArch)
+set(GPU_CONST_PARAM_ARCHITECTUES AMPERE TURING VEGA MI2xx)
+set(GPU_CONST_PARAM_FILES "")
+foreach(GPU_ARCH ${GPU_CONST_PARAM_ARCHITECTUES})
+  set(PARAMFILE ${CMAKE_CURRENT_BINARY_DIR}/genGPUArch/gpu_const_param_${GPU_ARCH}.par)
+  add_custom_command(
+    OUTPUT ${PARAMFILE}
+    COMMAND bash -c
+            "echo -e '#define GPUCA_GPUTYPE_${GPU_ARCH}\\n#define PARAMETER_FILE \"GPUDefParametersDefaults.h\"\\ngInterpreter->AddIncludePath(\"${CMAKE_CURRENT_SOURCE_DIR}/Definitions\");\\ngInterpreter->AddIncludePath(\"${ON_THE_FLY_DIR}\");\\n.x ${CMAKE_CURRENT_SOURCE_DIR}/Standalone/tools/dumpGPUDefParam.C(\"${PARAMFILE}\")\\n.q\\n'"
+            | root -l -b > /dev/null
+    VERBATIM
+    WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/genGPUArch
+    MAIN_DEPENDENCY Standalone/tools/dumpGPUDefParam.C
+    DEPENDS Definitions/GPUDefParametersDefaults.h
+            ${ON_THE_FLY_DIR}/GPUDefParametersLoadPrepare.h
+            ${ON_THE_FLY_DIR}/GPUDefParametersLoad.inc
+    COMMENT "Generating GPU parameter set for architecture ${GPU_ARCH}")
+  LIST(APPEND GPU_CONST_PARAM_FILES ${PARAMFILE})
+endforeach()
+
+add_custom_target(${MODULE}_GPU_CONST_PARAM_ARCHS ALL DEPENDS ${GPU_CONST_PARAM_FILES})
+install(FILES ${GPU_CONST_PARAM_FILES} DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/GPU/arch_param)
+
+
 # Add compile definitions and libraries depending on available optional dependencies
 if(GPUCA_QA)
   message(STATUS "Building GPU QA")
diff --git a/GPU/GPUTracking/Definitions/GPUDefParametersDefaults.h b/GPU/GPUTracking/Definitions/GPUDefParametersDefaults.h
@@ -22,6 +22,7 @@
 
 // GPU Run Configuration
 #if defined(GPUCA_GPUCODE) && !defined(GPUCA_GPUCODE_GENRTC) && !defined(GPUCA_GPUCODE_NO_LAUNCH_BOUNDS) // Avoid including for RTC generation besides normal include protection.
+  #define GPUCA_LB_SCAN 512
   // GPU-architecture-dependent default settings
   #if defined(GPUCA_GPUTYPE_MI2xx)
     #define GPUCA_WARP_SIZE 64
@@ -498,11 +499,11 @@
   #define GPUCA_LB_GPUTPCNNClusterizerKernels_publishClass1Regression GPUCA_LB_GPUTPCNNClusterizerKernels
   #define GPUCA_LB_GPUTPCNNClusterizerKernels_publishClass2Regression GPUCA_LB_GPUTPCNNClusterizerKernels
 
-  #define GPUCA_LB_GPUTPCCFStreamCompaction_scanStart GPUCA_THREAD_COUNT_SCAN
-  #define GPUCA_LB_GPUTPCCFStreamCompaction_scanUp GPUCA_THREAD_COUNT_SCAN
-  #define GPUCA_LB_GPUTPCCFStreamCompaction_scanTop GPUCA_THREAD_COUNT_SCAN
-  #define GPUCA_LB_GPUTPCCFStreamCompaction_scanDown GPUCA_THREAD_COUNT_SCAN
-  #define GPUCA_LB_GPUTPCCFStreamCompaction_compactDigits GPUCA_THREAD_COUNT_SCAN
+  #define GPUCA_LB_GPUTPCCFStreamCompaction_scanStart GPUCA_LB_SCAN
+  #define GPUCA_LB_GPUTPCCFStreamCompaction_scanUp GPUCA_LB_SCAN
+  #define GPUCA_LB_GPUTPCCFStreamCompaction_scanTop GPUCA_LB_SCAN
+  #define GPUCA_LB_GPUTPCCFStreamCompaction_scanDown GPUCA_LB_SCAN
+  #define GPUCA_LB_GPUTPCCFStreamCompaction_compactDigits GPUCA_LB_SCAN
   #define GPUCA_LB_GPUTPCCompressionGatherKernels_unbuffered GPUCA_LB_COMPRESSION_GATHER
   #define GPUCA_LB_GPUTPCCompressionGatherKernels_buffered32 GPUCA_LB_COMPRESSION_GATHER
   #define GPUCA_LB_GPUTPCCompressionGatherKernels_buffered64 GPUCA_LB_COMPRESSION_GATHER
diff --git a/GPU/GPUTracking/Standalone/CMakeLists.txt b/GPU/GPUTracking/Standalone/CMakeLists.txt
@@ -242,3 +242,4 @@ install(TARGETS ca TPCFastTransformation standalone_support)
 install(FILES "cmake/makefile" DESTINATION "${CMAKE_INSTALL_PREFIX}")
 install(CODE "execute_process(COMMAND ${CMAKE_COMMAND} -E create_symlink ${O2_DIR} ${CMAKE_INSTALL_PREFIX}/src)")
 install(CODE "execute_process(COMMAND ${CMAKE_COMMAND} -E create_symlink ${CMAKE_BINARY_DIR}/config.cmake ${CMAKE_INSTALL_PREFIX}/config.cmake)")
+install(DIRECTORY tools DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/GPU)
diff --git a/GPU/GPUTracking/Standalone/tools/dumpGPUDefParam.C b/GPU/GPUTracking/Standalone/tools/dumpGPUDefParam.C
@@ -12,27 +12,24 @@
 /// \file dumpGPUDefParam.C
 /// \author David Rohr
 
-// Run e.g. as:
-// ROOT_INCLUDE_PATH="`pwd`/include/GPU" root -l -q -b src/GPU/GPUTracking/Standalone/tools/dumpGPUDefParam.C'()'
+// Run e.g. as (replacing [FILE] and [OUTPUT]:
+// echo -e '#define PARAMETER_FILE "[FILE]]"\ngInterpreter->AddIncludePath("'`pwd`'/include/GPU");\n.x share/GPU/tools/dumpGPUDefParam.C("[OUTPUT]")\n.q\n' | root -l -b
+// To dump the defaults for AMPERE architecture, run
+// echo -e '#define GPUCA_GPUTYPE_AMPERE\n#define PARAMETER_FILE "GPUDefParametersDefaults.h"\ngInterpreter->AddIncludePath("'`pwd`'/include/GPU");\n.x share/GPU/tools/dumpGPUDefParam.C("default_AMPERE.par")\n.q\n' | root -l -b
 
-// Logic for testing to load the default parameters
-/* #define GPUCA_GPUCODE
-#define GPUCA_GPUTYPE_AMPERE
-#define GPUCA_MAXN 40
-#define GPUCA_ROW_COUNT 152
-#define GPUCA_TPC_COMP_CHUNK_SIZE 1024
-#include "GPUDefParametersConstants.h"
-#include "GPUDefParametersDefaults.h" */
+#ifndef PARAMETER_FILE
+#error Must provide the PARAMETER_FILE as preprocessor define, e.g. -DHEADER_TO_INCLUDE='"GPUDefParametersDefaults.h"'
+#endif
 
-// Alternatively, logic to load file that sets GPUDefParameters
-#include "testParam.h"
+#define GPUCA_GPUCODE
+#include PARAMETER_FILE
 
 #include "GPUDefParametersLoad.inc"
-void dumpGPUDefParam()
+void dumpGPUDefParam(const char* outputfile = "parameters.out")
 {
   auto param = o2::gpu::internal::GPUDefParametersLoad();
-  printf("Loaded params:\n%s", o2::gpu::internal::GPUDefParametersExport(param, false).c_str());
-  FILE* fp = fopen("parameters.out", "w+b");
+  printf("Loaded params:\n%s\nWriting them to %s\n", o2::gpu::internal::GPUDefParametersExport(param, false).c_str(), outputfile);
+  FILE* fp = fopen(outputfile, "w+b");
   fwrite(&param, 1, sizeof(param), fp);
   fclose(fp);
 }
diff --git a/GPU/GPUTracking/display/CMakeLists.txt b/GPU/GPUTracking/display/CMakeLists.txt
@@ -149,10 +149,10 @@ if(ALIGPU_BUILD_TYPE STREQUAL "Standalone")
   target_link_libraries(${targetName} PUBLIC O2::GPUTracking)
   install(TARGETS ${MODULE})
 
-  install(DIRECTORY filterMacros/ DESTINATION displayTrackFilter FILES_MATCHING PATTERN "*.C")
+  install(DIRECTORY filterMacros/ DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/GPU/displayTrackFilter FILES_MATCHING PATTERN "*.C")
   get_property(GPU_DISPLAY_INCLUDE_PATH DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR} PROPERTY INCLUDE_DIRECTORIES)
   configure_file(filterMacros/setinclude.sh.in setinclude.sh @ONLY)
-  install(FILES ${CMAKE_CURRENT_BINARY_DIR}/setinclude.sh PERMISSIONS OWNER_READ OWNER_WRITE OWNER_EXECUTE GROUP_READ GROUP_EXECUTE DESTINATION displayTrackFilter)
+  install(FILES ${CMAKE_CURRENT_BINARY_DIR}/setinclude.sh PERMISSIONS OWNER_READ OWNER_WRITE OWNER_EXECUTE GROUP_READ GROUP_EXECUTE DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/GPU/displayTrackFilter)
 endif()
 
 install(FILES ${HDRS} ${HDRS_INSTALL} DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/GPU)
diff --git a/GPU/TPCFastTransformation/CMakeLists.txt b/GPU/TPCFastTransformation/CMakeLists.txt
@@ -114,8 +114,8 @@ if(${ALIGPU_BUILD_TYPE} STREQUAL "O2")
                            LABELS gpu tpc)
   endforeach()
 
-  install (FILES macro/TPCFastTransformInit.C
-           DESTINATION share/macro/)
+  install(FILES macro/TPCFastTransformInit.C
+          DESTINATION share/macro/)
 endif()
 
 if(ALIGPU_BUILD_TYPE STREQUAL "Standalone")
