GPU: propagate architecture name and LTO for ITS

Signed-off-by: Felix Schlepper <felix.schlepper@cern.ch>

Author: f3sch

Common/DCAFitter/GPU/cuda/CMakeLists.txt

@@ -20,6 +20,9 @@ o2_add_library(DCAFitterCUDA
                                      O2::DetectorsBase
                PRIVATE_LINK_LIBRARIES O2::GPUTrackingCUDAExternalProvider)
 set_property(TARGET ${targetName} PROPERTY CUDA_SEPARABLE_COMPILATION ON)
+# Device LTO, so that the device link can inline across the
+# O2::GPUTrackingCUDAExternalProvider objects, which are compiled to LTO IR.
+set_property(TARGET ${targetName} PROPERTY INTERPROCEDURAL_OPTIMIZATION ON)
 # add_compile_options(-lineinfo)
 
 #o2_add_test(DCAFitterNCUDA

Detectors/ITSMFT/ITS/tracking/GPU/cuda/CMakeLists.txt

@@ -35,6 +35,8 @@ if(CUDA_ENABLED)
     #     $<$<COMPILE_LANGUAGE:CUDA>:-G;-O0;-Xptxas=-O0>
     #     $<$<COMPILE_LANGUAGE:CXX>:-O0;-g>
     )
+    # -dlto is incompatible with -G, so device debugging needs it switched off.
+    set_property(TARGET ${targetName} PROPERTY INTERPROCEDURAL_OPTIMIZATION ON)
     # target_compile_definitions(${targetName} PRIVATE ITS_MEASURE_GPU_TIME ITS_GPU_LOG)
     target_compile_definitions(${targetName} PRIVATE $<TARGET_PROPERTY:O2::ITStracking,COMPILE_DEFINITIONS>)
 endif()

GPU/GPUTracking/Base/cuda/CMakeLists.txt

@@ -204,4 +204,9 @@ add_library(O2::GPUTrackingCUDAExternalProvider ALIAS GPUTrackingCUDAExternalPro
 set_property(TARGET GPUTrackingCUDAExternalProvider PROPERTY CUDA_SEPARABLE_COMPILATION ON)
 target_compile_definitions(GPUTrackingCUDAExternalProvider PRIVATE $<TARGET_PROPERTY:${targetName},COMPILE_DEFINITIONS>)
 target_include_directories(GPUTrackingCUDAExternalProvider PRIVATE $<TARGET_PROPERTY:${targetName},INCLUDE_DIRECTORIES>)
+# Emit LTO IR (code=[compute_XX,lto_XX]) instead of SASS, so that consumers
+# device-linking with -dlto can inline across the provider. Consumers doing a
+# plain device link still work: nvlink falls back to compiling the IR, it just
+# does not get the cross-module inlining.
+set_property(TARGET GPUTrackingCUDAExternalProvider PROPERTY INTERPROCEDURAL_OPTIMIZATION ON)
 add_dependencies(GPUTrackingCUDAExternalProvider O2::GPUTracking) # must not depend on GPU backend to avoid cyclic dependencies

cmake/O2AddHipifiedExecutable.cmake

@@ -78,4 +78,9 @@ function(o2_add_hipified_executable baseTargetName)
   o2_add_executable("${baseTargetName}"
                     SOURCES ${HIP_SOURCES}
                     ${FORWARD_ARGS})
+
+  # Export architecture name
+  if(A_TARGETVARNAME)
+    set(${A_TARGETVARNAME} ${${A_TARGETVARNAME}} PARENT_SCOPE)
+  endif()
 endfunction()

cmake/O2AddHipifiedLibrary.cmake

@@ -72,4 +72,9 @@ function(o2_add_hipified_library baseTargetName)
   o2_add_library("${baseTargetName}"
                  SOURCES ${HIP_SOURCES}
                  ${FORWARD_ARGS})
+
+  # Export architecture name
+  if(A_TARGETVARNAME)
+    set(${A_TARGETVARNAME} ${${A_TARGETVARNAME}} PARENT_SCOPE)
+  endif()
 endfunction()