Add GPU capability (#106 , #120)

The work of many years by many people finally contributed.

---------

Co-authored-by: Andreas Mueller <andreas.mueller@ecmwf.int>
Co-authored-by: Nils Wedi <nils.wedi@ecmwf.int>
Co-authored-by: Willem Deconinck <willem.deconinck@ecmwf.int>
Co-authored-by: Olivier Marsden <olivier.marsden@ecmwf.int>
Co-authored-by: Lukas Mosimann <lmosimann@nvidia.com>

AUTHORS

@@ -1,20 +1,23 @@
 Authors and Contributors
 ========================
 
+- P. Courtier (ECMWF)
 - W. Deconinck (ECMWF)
+- D. Degrauwe (RMI)
 - D. Dent (ECMWF)
 - P. Dueben (ECMWF)
 - R. El Khatib (Meteo France)
+- D. Giard (Meteo France)
 - J. Hague (ECMWF)
 - M. Hamrud (ECMWF)
+- S. Hatfield (ECMWF)
+- M. Hortal (ECMWF)
 - L. Isaksen (ECMWF)
-- G. Mozdzynski (ECMWF)
 - P. Marguinaud (Meteo France)
+- O. Marsden (ECMWF)
+- L. Mosimann (NVIDIA)
+- G. Mozdzynski (ECMWF)
 - A. Mueller (ECMWF)
-- M. Hortal (ECMWF)
-- P. Courtier (ECMWF)
-- D. Degrauwe (RMI)
-- D. Giard (Meteo France)
 - G. Radnoti (ECMWF)
 - D. Salmond (ECMWF)
 - Y. Seity (Meteo France)

CMakeLists.txt

@@ -6,32 +6,50 @@
 # granted to it by virtue of its status as an intergovernmental organisation
 # nor does it submit to any jurisdiction.
 
-cmake_minimum_required( VERSION 3.12 FATAL_ERROR )
-
-# Once we update the minimum required version to 3.20,
-# we can use the command `cmake_path` instead of `get_filename_component`.
-# For discussion, see https://github.com/ecmwf-ifs/ectrans/pull/61#discussion_r1494362092
+cmake_minimum_required( VERSION 3.18 FATAL_ERROR )
+# CMake 3.17 adds INTERFACE link options which get propagated to the link stage,
+#   even if the target is linked in privately
+# CMake 3.18 allows "LINK_LANG_AND_ID" generator expression.
 
 find_package( ecbuild 3.4 REQUIRED HINTS ${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/../ecbuild )
 
-project( ectrans LANGUAGES C Fortran )
+project( ectrans LANGUAGES C CXX Fortran )
 include( ectrans_macros )
 
+set(CMAKE_CXX_STANDARD 17)
+
 ecbuild_enable_fortran( REQUIRED NO_MODULE_DIRECTORY )
 
-### Find (optional) dependencies 
+### Find (optional) dependencies
 
 ecbuild_find_package( NAME fiat REQUIRED )
 
-ecbuild_add_option( FEATURE MPI 
+ecbuild_add_option( FEATURE MPI
                     DESCRIPTION "Support for MPI distributed memory parallelism"
+                    REQUIRED_PACKAGES "MPI COMPONENTS Fortran CXX"
                     CONDITION fiat_HAVE_MPI )
 
 ecbuild_add_option( FEATURE OMP
                     DEFAULT ON
                     DESCRIPTION "Support for OpenMP shared memory parallelism"
                     REQUIRED_PACKAGES "OpenMP COMPONENTS Fortran" )
 
+if( ${CMAKE_VERSION} VERSION_LESS "3.25" AND (NOT DEFINED ENABLE_ACC OR ENABLE_ACC ) )
+  # See https://gitlab.kitware.com/cmake/cmake/-/issues/23691, fixed in CMake 3.25
+  # (TL;DR: FindOpenACC sets OpenACC_<LANG>_FOUND correctly but does not set
+  #  OpenACC_FOUND unless all three C, CXX, and Fortran have been found - even if
+  #  only one language has been requested via COMPONENTS)
+  find_package( OpenACC COMPONENTS Fortran )
+  if( OpenACC_Fortran_FOUND )
+    set( OpenACC_FOUND ON )
+  endif()
+endif()
+
+ecbuild_add_option( FEATURE ACC
+                    DEFAULT OFF
+                    DESCRIPTION "Support for using GPUs with OpenACC"
+                    REQUIRED_PACKAGES "OpenACC COMPONENTS Fortran" )
+
 ecbuild_add_option( FEATURE DOUBLE_PRECISION
                     DEFAULT ON
                     DESCRIPTION "Support for Double Precision" )
@@ -55,15 +73,90 @@ if( NOT HAVE_MKL )
     option( FFTW_ENABLE_MKL OFF )
 endif()
 
-ecbuild_find_package( NAME FFTW REQUIRED COMPONENTS double ${single} )
+ecbuild_add_option( FEATURE CPU
+                    DEFAULT ON
+                    DESCRIPTION "Compile CPU version of ectrans"
+                   )
+
+if( HAVE_CPU )
+  ecbuild_find_package( NAME FFTW REQUIRED COMPONENTS double ${single} )
+endif()
 
 ecbuild_add_option( FEATURE TRANSI
                     DEFAULT ON
                     DESCRIPTION "Compile TransI C-interface to trans"
-                    CONDITION HAVE_DOUBLE_PRECISION )
+                    CONDITION HAVE_DOUBLE_PRECISION AND HAVE_CPU )
+
+# Search for available GPU runtimes, searching for CUDA first and, if not found,
+# attempt to find HIP
+if( ECTRANS_ENABLE_GPU OR (NOT DEFINED ECTRANS_ENABLE_GPU AND ENABLE_GPU))
+  set(HAVE_CUDA 0)
+  set(HAVE_HIP 0)
+  ectrans_find_cuda() # sets "HAVE_CUDA"
+  if( NOT HAVE_CUDA )
+    if(CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
+        ectrans_find_hip() # sets "HAVE_HIP"
+    else()
+        message("Cannot enable for HIP language, require CMake >= 3.24")
+    endif()
+  endif()
+endif()
+
+ecbuild_add_option( FEATURE GPU
+                    DEFAULT OFF
+                    DESCRIPTION "Compile GPU version of ectrans (Requires OpenACC or sufficient OpenMP offloading support and MPI)"
+                    CONDITION (HAVE_HIP OR HAVE_CUDA) AND (HAVE_ACC OR HAVE_OMP) AND HAVE_MPI )
+
+if( HAVE_GPU )
+  if( HAVE_ACC )
+    set( GPU_OFFLOAD "ACC" )
+  elseif( HAVE_OMP )
+    set( GPU_OFFLOAD "OMP" )
+  else()
+    ecbuild_error("Could not enable GPU as OMP or ACC were not enabled")
+  endif()
+endif()
+
+ecbuild_add_option( FEATURE CUTLASS
+                    DEFAULT OFF
+                    CONDITION HAVE_GPU AND HAVE_CUDA AND CMAKE_Fortran_COMPILER_ID MATCHES "NVHPC"
+                    DESCRIPTION "Support for Cutlass BLAS operations"
+                    REQUIRED_PACKAGES "NvidiaCutlass VERSION 2.11" )
+
+# following also needs cuda arch sm80 to be effective
+ecbuild_add_option( FEATURE CUTLASS_3XTF32
+                    DEFAULT ON
+                    CONDITION HAVE_SINGLE_PRECISION AND HAVE_CUTLASS
+                    DESCRIPTION "Support for 3xTF32 with Cutlass (>= 2.8) and CUDA_ARCHITECTURES >= 80" )
+
+ecbuild_add_option( FEATURE GPU_AWARE_MPI
+                    DEFAULT ON
+                    CONDITION HAVE_GPU
+                    REQUIRED_PACKAGES "MPI COMPONENTS CXX Fortran"
+                    DESCRIPTION "Enable CUDA-aware MPI" )
+
+ecbuild_add_option( FEATURE GPU_GRAPHS_GEMM
+                    DEFAULT ON
+                    CONDITION HAVE_GPU
+                    DESCRIPTION "Enable graph-based optimisation of Legendre transform GEMM kernel" )
+
+if( BUILD_SHARED_LIBS )
+  set( GPU_STATIC_DEFAULT OFF )
+else()
+  set( GPU_STATIC_DEFAULT ON )
+endif()
+ecbuild_add_option( FEATURE GPU_STATIC
+                    DEFAULT ${GPU_STATIC_DEFAULT}
+                    DESCRIPTION "Compile GPU library as static library")
 
 ectrans_find_lapack()
 
+ecbuild_add_option( FEATURE TESTS
+                    DEFAULT ON
+                    DESCRIPTION "Enable unit testing"
+                    REQUIRED_PACKAGES "MPI COMPONENTS Fortran"
+                    CONDITION HAVE_CPU )
+
 ### Add sources and tests
 include( ectrans_compile_options )
 add_subdirectory( src )
@@ -79,4 +172,3 @@ endif()
 ecbuild_install_project( NAME ${PROJECT_NAME} )
 
 ecbuild_print_summary()
-

README.md

@@ -5,14 +5,14 @@ Introduction
 ============
 
 ecTrans is the global spherical Harmonics transforms library, extracted from the IFS.
-It is using a hybrid of MPI and OpenMP parallelisation strategies.
-The package contains both single- and double precision Fortran libraries (trans_sp, trans_dp),
-as well as a C interface to the double-precision version (transi_dp)
+It contains both CPU and GPU (Nvidia) code-paths.
+The CPU version uses a hybrid of MPI and OpenMP parallelisation strategies, while the GPU version combines MPI and OpenACC.
+A default installation builds both CPU libraries (trans_sp, trans_dp) and various flavours of GPU libraries in (trans_gpu_{sp/dp} shared library, trans_gpu_static_{sp/dp} static library, trans_gpu_static_CA_{sp/dp} static library requiring CUDA-aware MPI implementation), as well as a C interface to the double-precision version (transi_dp). A simple benchmark driver is also built against each of these libraries, allowing simple testing of the transforms.
 
 License
 =======
 
-Trans is distributed under the Apache License Version 2.0.
+ecTrans is distributed under the Apache License Version 2.0.
 See `LICENSE` file for details.
 
 Installing ecTrans
@@ -26,6 +26,8 @@ Supported Platforms
 
 Other UNIX-like operating systems may work too out of the box.
 
+The GPU codepath has only been tested with NVHPC compilers on recent Nvidia GPUs.
+
 Requirements
 ------------
 - Fortran compiler with OpenMP support
@@ -38,6 +40,10 @@ Requirements
 Further optional recommended dependencies:
 - FFTW (http://www.fftw.org)
 
+For the GPU libraries :
+- Fortran compiler with OpenACC support
+- CUDA toolkit (compiler, and CUBLAS and CUFFT libraries)
+
 Building ecTrans
 ----------------
 
@@ -69,8 +75,21 @@ Extra options can be added to the `cmake` command to control the build:
  - `-DENABLE_DOUBLE_PRECISION=<ON|OFF>` default=ON
  - `-DENABLE_TRANSI=<ON|OFF>`           default=ON
  - `-DENABLE_MKL=<ON|OFF>`              default=ON
+ - `-DENABLE_FFTW=<ON|OFF>`             default=ON
+ - `-DENABLE_GPU=<ON|OFF>`              default=OFF
  - `-DCMAKE_INSTALL_PREFIX=<install-prefix>`
 
+Specific extra options exist for GPU installation:
+ - `-DENABLE_GPU_AWARE_MPI=<ON|OFF>`    default=OF
+ - `-DENABLE_GPU_GRAPHS_GEMM=<ON|OFF>`  default=ON
+ - `-DENABLE_CUTLASS=<ON|OFF>`          default=OFF
+ - `-DENABLE_3XTF32=<ON|OFF>`           default=OFF
+
+GPU-aware MPI allows buffers residing on GPU to be passed to MPI communication calls directly. This requires a compatible MPI installation.
+Graph work-flows allow a series of GPU operations to be scheduled in an efficient manner. 
+This is useful both for the batched FFTs and the batched GEMMs on which ecTrans relies, although for FFTs this is currently relied upon.
+Cutlass is an Nvidia library of templates for GEMM operations. 3xTF32 is a specific acceleration for single precision operations, enabled by Cutlass.
+
 More options to control compilation flags, only when defaults are not sufficient
 
  - `-DCMAKE_Fortran_FLAGS=<fortran-flags>`
@@ -82,6 +101,10 @@ Optionally, tests can be run to check succesful compilation, when the feature TE
 
     $ ctest
 
+The benchmark drivers are found in the bin directory.
+A brief description of available command-line arguments can be obtained with e.g.
+ectrans-benchmark-cpu-sp --help
+
 Reporting Bugs
 ==============