CUDA variant optimised beyond k-caching, e.g., pipelined global-to-sh…

…ared mem copies overlaped with compute
ecmwf-ifs · Dec 11, 2024 · 99afc47 · 99afc47
1 parent 82fdf4b
commit 99afc47
Show file tree

Hide file tree

Showing 5 changed files with 3,559 additions and 3 deletions.
diff --git a/src/cloudsc_cuda/CMakeLists.txt b/src/cloudsc_cuda/CMakeLists.txt
@@ -54,7 +54,7 @@ if( HAVE_CLOUDSC_C_CUDA )
         target_compile_options(dwarf-cloudsc-c-cuda-lib PRIVATE $<$<COMPILE_LANGUAGE:CUDA>>)
     else()
         target_compile_options(dwarf-cloudsc-c-cuda-lib PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:
-            --ptxas-options=-O3 -use_fast_math -maxrregcount=128 -gencode arch=compute_${CMAKE_CUDA_ARCHITECTURES},code=sm_${CMAKE_CUDA_ARCHITECTURES}>) 
+            --ptxas-options=-O3 -use_fast_math -lineinfo -maxrregcount=128 -gencode arch=compute_${CMAKE_CUDA_ARCHITECTURES},code=sm_${CMAKE_CUDA_ARCHITECTURES}>) 
     endif()
 
     set_target_properties( dwarf-cloudsc-c-cuda-lib PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
@@ -112,7 +112,7 @@ if( HAVE_CLOUDSC_C_CUDA )
       target_compile_options(dwarf-cloudsc-c-cuda-hoist-lib PRIVATE $<$<COMPILE_LANGUAGE:CUDA>>)
     else()
       target_compile_options(dwarf-cloudsc-c-cuda-hoist-lib PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:
-          --ptxas-options=-O3 -use_fast_math -maxrregcount=128 -gencode arch=compute_${CMAKE_CUDA_ARCHITECTURES},code=sm_${CMAKE_CUDA_ARCHITECTURES}>)
+          --ptxas-options=-O3 -use_fast_math -lineinfo -maxrregcount=128 -gencode arch=compute_${CMAKE_CUDA_ARCHITECTURES},code=sm_${CMAKE_CUDA_ARCHITECTURES}>)
     endif()
 
     set_target_properties( dwarf-cloudsc-c-cuda-hoist-lib PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
@@ -170,7 +170,8 @@ if( HAVE_CLOUDSC_C_CUDA )
       target_compile_options(dwarf-cloudsc-c-cuda-k-caching-lib PRIVATE $<$<COMPILE_LANGUAGE:CUDA>>)
     else()
       target_compile_options(dwarf-cloudsc-c-cuda-k-caching-lib PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:
-          --ptxas-options=-O3 -use_fast_math -maxrregcount=128 -gencode arch=compute_${CMAKE_CUDA_ARCHITECTURES},code=sm_${CMAKE_CUDA_ARCHITECTURES}>)
+	  -O3 -use_fast_math -lineinfo -maxrregcount=128 -gencode arch=compute_${CMAKE_CUDA_ARCHITECTURES},code=sm_${CMAKE_CUDA_ARCHITECTURES}>)
+          # -O0 -g -G -maxrregcount=128 -gencode arch=compute_${CMAKE_CUDA_ARCHITECTURES},code=sm_${CMAKE_CUDA_ARCHITECTURES}>)
     endif()
     set_target_properties( dwarf-cloudsc-c-cuda-k-caching-lib PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
 
@@ -191,6 +192,64 @@ if( HAVE_CLOUDSC_C_CUDA )
     )
     ###
 
+    ###### SCC-CUDA-OPT ####
+    ecbuild_add_library(
+        TARGET dwarf-cloudsc-c-cuda-opt-lib
+        INSTALL_HEADERS LISTED
+        SOURCES
+            cloudsc/yoecldp_c.h
+            cloudsc/load_state.h
+            cloudsc/load_state.cu
+            cloudsc/cloudsc_c_opt.h
+            cloudsc/cloudsc_c_opt.cu
+            cloudsc/cloudsc_driver_opt.h
+            cloudsc/cloudsc_driver_opt.cu
+            cloudsc/cloudsc_validate.h
+            cloudsc/cloudsc_validate.cu
+            cloudsc/mycpu.h
+            cloudsc/mycpu.cu
+        PUBLIC_INCLUDES
+            $<INSTALL_INTERFACE:include>
+            $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/cloudsc>
+        PUBLIC_LIBS
+            $<${HAVE_HDF5}:hdf5::hdf5>
+            $<${HAVE_SERIALBOX}:Serialbox::Serialbox_C>
+            $<${HAVE_OMP}:OpenMP::OpenMP_C>
+        DEFINITIONS
+            ${CLOUDSC_DEFINITIONS}
+    )
+
+    target_include_directories(
+      dwarf-cloudsc-c-cuda-opt-lib
+      PUBLIC
+      ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}
+    )
+    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+      target_compile_options(dwarf-cloudsc-c-cuda-opt-lib PRIVATE $<$<COMPILE_LANGUAGE:CUDA>>)
+    else()
+      target_compile_options(dwarf-cloudsc-c-cuda-opt-lib PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:
+          -O3 -use_fast_math -lineinfo -maxrregcount=128 -gencode arch=compute_${CMAKE_CUDA_ARCHITECTURES},code=sm_${CMAKE_CUDA_ARCHITECTURES}>)
+          # -O0 -g -G -maxrregcount=128 -gencode arch=compute_${CMAKE_CUDA_ARCHITECTURES},code=sm_${CMAKE_CUDA_ARCHITECTURES}>)
+    endif()
+    set_target_properties( dwarf-cloudsc-c-cuda-opt-lib PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
+
+    ecbuild_add_executable(
+        TARGET  dwarf-cloudsc-c-cuda-opt
+        SOURCES dwarf_cloudsc.cpp
+        LIBS dwarf-cloudsc-c-cuda-opt-lib
+    )
+
+    target_link_libraries(dwarf-cloudsc-c-cuda-opt dwarf-cloudsc-c-cuda-opt-lib)
+
+    ecbuild_add_test(
+        TARGET dwarf-cloudsc-c-cuda-opt-serial
+        COMMAND bin/dwarf-cloudsc-c-cuda-opt
+        ARGS 1 1000 128
+        WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/../../..
+        OMP 1
+    )
+    ###
+
  # Create symlink for the input data
  if( HAVE_SERIALBOX )
    execute_process(COMMAND ${CMAKE_COMMAND} -E create_symlink