Merge pull request #104 from ecmwf-ifs/nams-cuda-beyond-k-caching

CUDA variant optimised beyond k-caching
ecmwf-ifs · Dec 20, 2024 · 8cec441 · 8cec441
2 parents 842c1b8 + 6beb1d9
commit 8cec441
Show file tree

Hide file tree

Showing 8 changed files with 3,572 additions and 36 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -54,6 +54,7 @@ ecbuild_add_option( FEATURE CUDA
     CONDITION CMAKE_CUDA_COMPILER )
 if( HAVE_CUDA )
     enable_language( CUDA )
+    find_package(CUDAToolkit)
 endif()
 
 ecbuild_add_option( FEATURE HIP

diff --git a/README.md b/README.md
@@ -80,7 +80,11 @@ Balthasar Reuter (balthasar.reuter@ecmwf.int)
     of CLOUDSC including host side hoisted temporary local variables.
   - **dwarf-cloudsc-cuda-k-caching**: GPU-enabled, further optimized CUDA
     C version of CLOUDSC including loop fusion and temporary local 
-    array demotion.  
+    array demotion. 
+  - **dwarf-cloudsc-cuda-opt**: GPU-enabled, further optimized beyond
+    k-caching CUDA C version that buffers some variables and
+    uses pipelined global-to-shared memory copies that are overlapped
+    with compute (TMA loads). 
 - **dwarf-cloudsc-gpu-scc-field**: GPU-enabled and optimized version of
   CLOUDSC that uses the SCC loop layout, and uses [FIELD API](https://github.com/ecmwf-ifs/field_api) (a Fortran library purpose-built for IFS data-structures that facilitates the
   creation and management of field objects in scientific code) to perform device offload 

diff --git a/arch/toolchains/github-ubuntu-nvhpc.cmake b/arch/toolchains/github-ubuntu-nvhpc.cmake
@@ -32,6 +32,10 @@ set( OpenACC_Fortran_FLAGS "-acc=gpu -mp" CACHE STRING "" )
 # COMMON FLAGS
 ####################################################################
 
+if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+  set(CMAKE_CUDA_ARCHITECTURES 80)
+endif()
+
 set(ECBUILD_Fortran_FLAGS "-fpic")
 set(ECBUILD_Fortran_FLAGS "${ECBUILD_Fortran_FLAGS} -Mframe")
 set(ECBUILD_Fortran_FLAGS "${ECBUILD_Fortran_FLAGS} -Mbyteswapio")

diff --git a/src/cloudsc_cuda/CMakeLists.txt b/src/cloudsc_cuda/CMakeLists.txt
@@ -17,6 +17,17 @@ if( HAVE_CLOUDSC_C_CUDA )
     enable_language(CUDA)
     enable_language(CXX)
 
+    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+      set( CLOUDSC_CUDA_FLAGS "-lineinfo -maxrregcount=128" )
+    else()
+      set( CLOUDSC_CUDA_FLAGS "-lineinfo -maxrregcount=128 -gencode arch=compute_${CMAKE_CUDA_ARCHITECTURES},code=sm_${CMAKE_CUDA_ARCHITECTURES}>" )
+    endif()
+    if(CMAKE_BUILD_TYPE STREQUAL "Debug")
+      set( CLOUDSC_CUDA_OPT_FLAGS "-O0 -g -G" )
+    else()
+      set( CLOUDSC_CUDA_OPT_FLAGS  "--ptxas-options=-O3 -O3 -use_fast_math" )
+    endif()
+
     ###### SCC-CUDA ####
     ecbuild_add_library(
         TARGET dwarf-cloudsc-c-cuda-lib
@@ -41,22 +52,12 @@ if( HAVE_CLOUDSC_C_CUDA )
 	    $<${HAVE_HDF5}:hdf5::hdf5>
 	    $<${HAVE_SERIALBOX}:Serialbox::Serialbox_C>
 	    $<${HAVE_OMP}:OpenMP::OpenMP_C>
+	    CUDA::cudart
 	DEFINITIONS
 	    ${CLOUDSC_DEFINITIONS}
     )
 
-    target_include_directories(
-        dwarf-cloudsc-c-cuda-lib
-        PUBLIC
-        ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}
-    )
-    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-        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}>) 
-    endif()
-
+    target_compile_options(dwarf-cloudsc-c-cuda-lib PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:SHELL:${CLOUDSC_CUDA_OPT_FLAGS} ${CLOUDSC_CUDA_FLAGS}>)
     set_target_properties( dwarf-cloudsc-c-cuda-lib PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
 
     ecbuild_add_executable(
@@ -99,22 +100,12 @@ if( HAVE_CLOUDSC_C_CUDA )
 	    $<${HAVE_HDF5}:hdf5::hdf5>
 	    $<${HAVE_SERIALBOX}:Serialbox::Serialbox_C>
 	    $<${HAVE_OMP}:OpenMP::OpenMP_C>
+	    CUDA::cudart
 	DEFINITIONS
 	    ${CLOUDSC_DEFINITIONS}
     )
 
-    target_include_directories(
-      dwarf-cloudsc-c-cuda-hoist-lib
-      PUBLIC
-      ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}
-    )
-    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-      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}>)
-    endif()
-
+    target_compile_options(dwarf-cloudsc-c-cuda-hoist-lib PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:SHELL:${CLOUDSC_CUDA_OPT_FLAGS} ${CLOUDSC_CUDA_FLAGS}>)
     set_target_properties( dwarf-cloudsc-c-cuda-hoist-lib PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
 
     ecbuild_add_executable(
@@ -157,21 +148,12 @@ if( HAVE_CLOUDSC_C_CUDA )
 	    $<${HAVE_HDF5}:hdf5::hdf5>
 	    $<${HAVE_SERIALBOX}:Serialbox::Serialbox_C>
 	    $<${HAVE_OMP}:OpenMP::OpenMP_C>
+	    CUDA::cudart
 	DEFINITIONS
 	    ${CLOUDSC_DEFINITIONS}
     )
 
-    target_include_directories(
-      dwarf-cloudsc-c-cuda-k-caching-lib
-      PUBLIC
-      ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}
-    )
-    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-      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}>)
-    endif()
+    target_compile_options(dwarf-cloudsc-c-cuda-k-caching-lib PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:SHELL:${CLOUDSC_CUDA_OPT_FLAGS} ${CLOUDSC_CUDA_FLAGS}>)
     set_target_properties( dwarf-cloudsc-c-cuda-k-caching-lib PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
 
     ecbuild_add_executable(
@@ -191,6 +173,54 @@ 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>
+	    CUDA::cudart
+        DEFINITIONS
+            ${CLOUDSC_DEFINITIONS}
+    )
+
+    target_compile_options(dwarf-cloudsc-c-cuda-opt-lib PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:SHELL:${CLOUDSC_CUDA_OPT_FLAGS} ${CLOUDSC_CUDA_FLAGS}>)
+    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