Fix: Kokkos 5 performances on AMD

CMakeLists.txt

@@ -16,6 +16,7 @@ option(Idefix_DEBUG "Enable Idefix debug features (makes the code very slow)" OF
 option(Idefix_RUNTIME_CHECKS "Enable runtime sanity checks" OFF)
 option(Idefix_WERROR "Treat compiler warnings as errors" OFF)
 option(Idefix_PYTHON "Enable python bindings (requires pybind11)" OFF)
+option(Idefix_SUPPRESS_FMA "Disable FMA (fused multiply-add) contraction/codegen. Useful for code validation across architectures." OFF)
 set(Idefix_PROBLEM_DIR "${CMAKE_BINARY_DIR}" CACHE STRING "Problem directory to build for.")
 set(Idefix_CXX_FLAGS "" CACHE STRING "Additional compiler/linker flag")
 set(Idefix_DEFS "definitions.hpp" CACHE FILEPATH "Problem definition header file")
@@ -43,6 +44,7 @@ include(AddIdefixSource)
 include(SetIdefixProperty)
 include(SetRequiredBuildSettingsForGCC8)
 include(CheckHdf5ParallelSupport)
+include(SuppressFMA)
 
 #Idefix requires Cuda Lambdas (experimental)
 if(Kokkos_ENABLE_CUDA)
@@ -58,6 +60,8 @@ include_directories(${Kokkos_INCLUDE_DIRS_RET})
 # Add Idefix CXX Flags
 add_compile_options(${Idefix_CXX_FLAGS})
 
+
+
 # Add filesystem libraries for GCC8
 set_required_build_settings_for_GCC8()
 
@@ -181,15 +185,15 @@ endif()
 if(Idefix_EVOLVE_VECTOR_POTENTIAL)
   add_compile_definitions("EVOLVE_VECTOR_POTENTIAL")
 endif()
-#update version.hpp if possible
+
+# determine idefix version from git
 git_describe(GIT_SHA1)
 set(Idefix_VERSION ${Idefix_VERSION_MAJOR}.${Idefix_VERSION_MINOR}.${Idefix_VERSION_PATCH}-${GIT_SHA1})
-file(WRITE ${CMAKE_SOURCE_DIR}/src/version.hpp "#define IDEFIX_GIT_COMMIT \"${GIT_SHA1}\"\n")
-file(APPEND ${CMAKE_SOURCE_DIR}/src/version.hpp "#define IDEFIX_VERSION_MAJOR \"${Idefix_VERSION_MAJOR}\"\n")
-file(APPEND ${CMAKE_SOURCE_DIR}/src/version.hpp "#define IDEFIX_VERSION_MINOR \"${Idefix_VERSION_MINOR}\"\n")
-file(APPEND ${CMAKE_SOURCE_DIR}/src/version.hpp "#define IDEFIX_VERSION_PATCH \"${Idefix_VERSION_PATCH}\"\n")
-file(APPEND ${CMAKE_SOURCE_DIR}/src/version.hpp "#define IDEFIX_VERSION \"${Idefix_VERSION}\"\n")
-
+configure_file(
+    ${CMAKE_SOURCE_DIR}/src/version.h.in
+    ${CMAKE_BINARY_DIR}/build/generated/version.h
+    @ONLY
+)
 
 if(NOT ${Idefix_DEFS} STREQUAL "definitions.hpp")
   add_compile_definitions("DEFINITIONS_FILE=\"${Idefix_DEFS}\"")
@@ -269,6 +273,21 @@ target_include_directories(idefix PUBLIC
 
 target_link_libraries(idefix Kokkos::kokkos)
 
+# Generate header with compiler information (name, flags, path)
+configure_file(
+    ${CMAKE_SOURCE_DIR}/src/compiler_info.h.in
+    ${CMAKE_BINARY_DIR}/build/generated/compiler_info.h
+    @ONLY
+)
+# Make sure the generated header is on the include path
+target_include_directories(idefix PRIVATE ${CMAKE_BINARY_DIR}/build/generated)
+
+# disable FMA if needed
+if(Idefix_SUPPRESS_FMA)
+  message(STATUS "FMA (fused multiply-add) contraction/codegen is disabled")
+  target_suppress_fma(idefix)
+endif()
+
 message(STATUS "Idefix final configuration")
 if(Idefix_EVOLVE_VECTOR_POTENTIAL)
   message(STATUS "    MHD:  ${Idefix_MHD} (Vector potential)")

cmake/SuppressFMA.cmake

@@ -0,0 +1,102 @@
+#[[============================================================================
+  SuppressFMA.cmake
+
+  Provides an option and a function to optionally disable fused
+  multiply-add (FMA) code generation / contraction for a Kokkos-based CXX
+  target.
+
+  Kokkos wraps the real device compiler behind nvcc_wrapper (CUDA) or
+  hipcc (HIP), which makes CMAKE_CXX_COMPILER_ID report the *underlying
+  host* compiler (e.g. "GNU" or "Clang") instead of "NVIDIA" or "Clang
+  as HIP". Backend detection therefore relies on the Kokkos_ENABLE_*
+  variables exported by KokkosConfig.cmake / set by the Kokkos build,
+  and only falls back to CMAKE_CXX_COMPILER_ID for the plain host
+  compilers (no CUDA/HIP backend active):
+
+    - Kokkos_ENABLE_HIP  ON -> AMD HIP  (hipcc, clang-based)  : -ffp-contract=off
+    - Kokkos_ENABLE_CUDA ON -> NVIDIA nvcc (via nvcc_wrapper)  : --fmad=false
+    - otherwise, CMAKE_CXX_COMPILER_ID selects among:
+        GNU        : gcc/g++
+        Intel      : classic icc/icpc
+        IntelLLVM  : Intel oneAPI icx/icpx
+        Clang      : LLVM clang++
+        AppleClang : Xcode clang++
+        NVHPC      : NVIDIA HPC SDK (nvc++), e.g. for OpenMPTarget/OpenACC
+
+  Usage (after find_package(Kokkos) so Kokkos_ENABLE_* are defined):
+    include(SuppressFMA.cmake)
+    add_library(mylib source.cpp)
+    target_link_libraries(mylib PUBLIC Kokkos::kokkos)
+    target_suppress_fma(mylib)
+
+============================================================================]]
+
+include_guard(GLOBAL)
+
+# Determine the CXX FMA-suppression flags for the active Kokkos backend /
+# CXX compiler. Returns the list of flags (possibly empty) via out_var.
+function(_fma_suppression_flags_cxx out_var)
+    set(flags "")
+    set(id "${CMAKE_CXX_COMPILER_ID}")
+
+    # Kokkos backend takes priority: nvcc_wrapper/hipcc hide the real
+    # device compiler from CMAKE_CXX_COMPILER_ID.
+    if(Kokkos_ENABLE_HIP)
+        set(flags "-ffp-contract=off")
+
+    elseif(Kokkos_ENABLE_CUDA)
+        set(flags "--fmad=false")
+
+    elseif(id STREQUAL "GNU")
+        include(CheckCXXCompilerFlag)
+        check_cxx_compiler_flag("-mno-fma" IDEFIX_HAS_MNO_FMA)
+        set(flags "-ffp-contract=off")
+        if(IDEFIX_HAS_MNO_FMA)
+            list(APPEND flags "-mno-fma")
+        endif()
+    elseif(id MATCHES "^(Clang|AppleClang|CrayClang)$")
+        set(flags "-ffp-contract=off")
+
+    elseif(id STREQUAL "Intel")
+        # Intel classic compiler
+        set(flags "-fp-model=precise" "-no-fma")
+
+    elseif(id STREQUAL "IntelLLVM")
+        # Intel oneAPI compiler (clang-based)
+        set(flags "-ffp-contract=off" "-fp-model=strict")
+
+    elseif(id STREQUAL "NVHPC")
+        # NVIDIA HPC SDK (formerly PGI), e.g. OpenMPTarget/OpenACC backend
+        set(flags "-Mnofma") # untested
+    endif()
+
+    set(${out_var} "${flags}" PARENT_SCOPE)
+endfunction()
+
+# target_suppress_fma(<target>)
+#
+# Applies compiler-specific FMA-suppression flags to <target>'s CXX
+# sources, but only when the SUPPRESS_FMA option is ON. Safe to call
+# unconditionally.
+function(target_suppress_fma target)
+
+    if(NOT TARGET ${target})
+        message(FATAL_ERROR "target_suppress_fma: '${target}' is not a target")
+    endif()
+
+    _fma_suppression_flags_cxx(cxx_flags)
+
+    if(cxx_flags)
+        foreach(flag IN LISTS cxx_flags)
+            target_compile_options(${target} PRIVATE
+                $<$<COMPILE_LANGUAGE:CXX>:${flag}>
+            )
+        endforeach()
+    else()
+        message(VERBOSE
+            "target_suppress_fma: no FMA-suppression flag known for "
+            "CXX compiler '${CMAKE_CXX_COMPILER_ID}' "
+            "(Kokkos_ENABLE_CUDA=${Kokkos_ENABLE_CUDA}, "
+            "Kokkos_ENABLE_HIP=${Kokkos_ENABLE_HIP}) (target ${target})")
+    endif()
+endfunction()

pytools/idfx_test.py

@@ -225,21 +225,19 @@ def _genCmakeCommand(self, definitionFile=""):
 
         if self.cuda:
             comm.append("-DKokkos_ENABLE_CUDA=ON")
-            # disable fmad operations on Cuda to make it compatible with CPU arithmetics
-            comm.append("-DIdefix_CXX_FLAGS=--fmad=false")
             # disable Async cuda malloc for tests performed on old UCX implementations
             comm.append("-DKokkos_ENABLE_IMPL_CUDA_MALLOC_ASYNC=OFF")
 
         if self.intel:
             # disable fmad operations on Cuda to make it compatible with CPU arithmetics
-            comm.append("-DIdefix_CXX_FLAGS=-fp-model=strict")
             comm.append("-DCMAKE_CXX_COMPILER=icpx")
             comm.append("-DCMAKE_C_COMPILER=icx")
 
         if self.hip:
             comm.append("-DKokkos_ENABLE_HIP=ON")
-            # disable fmad operations on HIP to make it compatible with CPU arithmetics
-            comm.append("-DIdefix_CXX_FLAGS=-ffp-contract=off")
+
+        # disable FMA for testing so that we have the same results on CPU and GPU (otherwise, the results are not bitwise identical)
+        comm.append("-DIdefix_SUPPRESS_FMA=ON")
 
         # if we use single precision
         if self.single:

src/compiler_info.h.in

@@ -0,0 +1,17 @@
+// ***********************************************************************************
+// Idefix MHD astrophysical code
+// Copyright(C) Geoffroy R. J. Lesur <geoffroy.lesur@univ-grenoble-alpes.fr>
+// and other code contributors
+// Licensed under CeCILL 2.1 License, see COPYING for more information
+// ***********************************************************************************
+
+#ifndef COMPILER_INFO_H
+#define COMPILER_INFO_H
+
+namespace CompilerInfo {
+    inline constexpr const char* name    = "@CMAKE_CXX_COMPILER_ID@";
+    inline constexpr const char* version = "@CMAKE_CXX_COMPILER_VERSION@";
+    inline constexpr const char* path    = "@CMAKE_CXX_COMPILER@";
+}
+
+#endif

src/dataBlock/dumpToFile.cpp

@@ -9,8 +9,8 @@
 #include <cstdio>
 #include "../idefix.hpp"
 #include "dataBlock.hpp"
-#include "version.hpp"
 #include "fluid.hpp"
+#include "version.h"
 
 #define  NAMESIZE     16
 #define  HEADERSIZE 128
@@ -78,7 +78,7 @@ void DataBlock::DumpToFile(std::string filebase)  {
 
   // Write Header
   char header[HEADERSIZE];
-  std::snprintf(header, HEADERSIZE, "Idefix %s Debug DataBlock", IDEFIX_VERSION);
+  std::snprintf(header, HEADERSIZE, "Idefix %s Debug DataBlock", VersionInfo::version);
   fwrite (header, sizeof(char), HEADERSIZE, fileHdl);
 
   // Write Vc

src/input.cpp

@@ -18,8 +18,9 @@
 
 #include "idefix.hpp"
 #include "input.hpp"
-#include "version.hpp"
 #include "profiler.hpp"
+#include "version.h"
+#include "compiler_info.h"
 
 // Flag will be set if a signal has been received
 bool Input::abortRequested = false;
@@ -415,7 +416,10 @@ void Input::PrintOptions() {
 }
 
 void Input::PrintVersion() {
-  idfx::cout << "              Idefix version " << IDEFIX_VERSION << std::endl;
-  idfx::cout << "              Built against Kokkos " << KOKKOS_VERSION << std::endl;
-  idfx::cout << "              Compiled on " << __DATE__ <<  " at " << __TIME__ << std::endl;
+  idfx::cout << "Idefix version " << VersionInfo::version << std::endl;
+  idfx::cout << "Built against Kokkos " << KOKKOS_VERSION << std::endl;
+  idfx::cout << "Compiled on " << __DATE__ <<  " at " << __TIME__ << std::endl;
+  idfx::cout << "Compiler name: " << CompilerInfo::name    << std::endl;
+  idfx::cout << "Compiler version: " << CompilerInfo::version   << std::endl;
+  idfx::cout << "Compiler path:  " << CompilerInfo::path    << std::endl;
 }

src/loop.hpp

@@ -85,7 +85,7 @@ inline void idefix_for(const std::string & NAME,
   idfx::pushRegion("idefix_for("+NAME+")");
   #endif
   const int NI = IE - IB;
-  Kokkos::parallel_for(NAME, NI,
+  Kokkos::parallel_for(NAME, Kokkos::RangePolicy<Kokkos::LaunchBounds<256>>(0, NI),
     KOKKOS_LAMBDA (const int& IDX) {
       int i = IDX;
       i += IB;
@@ -112,7 +112,7 @@ inline void idefix_for(const std::string & NAME,
     const int NJ = JE - JB;
     const int NI = IE - IB;
     const int NJNI = NJ * NI;
-    Kokkos::parallel_for(NAME, NJNI,
+    Kokkos::parallel_for(NAME, Kokkos::RangePolicy<Kokkos::LaunchBounds<256>>(0, NJNI),
       KOKKOS_LAMBDA (const int& IDX) {
         int j = IDX  / NI;
         int i = IDX - j*NI;
@@ -172,7 +172,7 @@ inline void idefix_for(const std::string & NAME,
     const int NI = IE - IB;
     const int NKNJNI = NK*NJ*NI;
     const int NJNI = NJ * NI;
-    Kokkos::parallel_for(NAME,NKNJNI,
+    Kokkos::parallel_for(NAME,Kokkos::RangePolicy<Kokkos::LaunchBounds<256>>(0, NKNJNI),
       KOKKOS_LAMBDA (const int& IDX) {
         int k = IDX / NJNI;
         int j = (IDX - k*NJNI) / NI;
@@ -259,7 +259,7 @@ inline void idefix_for(const std::string & NAME,
     const int NNNKNJNI = NN*NK*NJ*NI;
     const int NKNJNI = NK*NJ*NI;
     const int NJNI = NJ * NI;
-    Kokkos::parallel_for(NAME,NNNKNJNI,
+    Kokkos::parallel_for(NAME,Kokkos::RangePolicy<Kokkos::LaunchBounds<256>>(0, NNNKNJNI),
       KOKKOS_LAMBDA (const int& IDX) {
         int n = IDX / NKNJNI;
         int k = (IDX - n*NKNJNI) / NJNI;

src/output/dump.cpp

@@ -20,11 +20,11 @@
 #include <string>
 #include <cstdio>
 #include "dump.hpp"
-#include "version.hpp"
 #include "dataBlockHost.hpp"
 #include "gridHost.hpp"
 #include "output.hpp"
 #include "fluid.hpp"
+#include "version.h"
 
 // Max size of array name
 #define  NAMESIZE     16
@@ -880,7 +880,7 @@ int Dump::Write(Output& output) {
 
   char header[HEADERSIZE];
   std::snprintf(header, HEADERSIZE, "Idefix %s Dump Data %s endian",
-                IDEFIX_VERSION, endian.c_str());
+                VersionInfo::version, endian.c_str());
   WriteString(fileHdl, header, HEADERSIZE);
 
   for(int dir = 0; dir < 3 ; dir++) {

src/output/vtk.cpp

@@ -20,12 +20,12 @@
 #else
   error "Missing the <filesystem> header."
 #endif
-#include "version.hpp"
 #include "idefix.hpp"
 #include "dataBlock.hpp"
 #include "gridHost.hpp"
 #include "output.hpp"
 #include "fluid.hpp"
+#include "version.h"
 
 #define VTK_RECTILINEAR_GRID    14
 #define VTK_STRUCTURED_GRID     35
@@ -385,7 +385,7 @@ void Vtk::WriteHeader(IdfxFileHandler fvtk, real time) {
   2. Header
   ------------------------------------------- */
 
-  ssheader << "Idefix " << IDEFIX_VERSION << " VTK Data" << std::endl;
+  ssheader << "Idefix " << VersionInfo::version << " VTK Data" << std::endl;
 
   /* ------------------------------------------
   3. File format

src/output/xdmf.cpp

@@ -21,11 +21,11 @@
 #endif
 
 #include "xdmf.hpp"
-#include "version.hpp"
 #include "idefix.hpp"
 #include "dataBlockHost.hpp"
 #include "gridHost.hpp"
 #include "output.hpp"
+#include "version.h"
 
 // Whether or not we write the time in the XDMF file
 #define WRITE_TIME
@@ -526,7 +526,7 @@ void Xdmf::WriteHeader(
 
   dimstr = 1;
 
-  ssheader << "Idefix " << IDEFIX_VERSION << " XDMF Data";
+  ssheader << "Idefix " << VersionInfo::version << " XDMF Data";
   strspace = H5Screate_simple(1, &dimstr, NULL);
   string_type = H5Tcopy(H5T_C_S1);
   H5Tset_size(string_type, strlen( ssheader.str().c_str() ));

src/version.h.in

@@ -0,0 +1,19 @@
+// ***********************************************************************************
+// Idefix MHD astrophysical code
+// Copyright(C) Geoffroy R. J. Lesur <geoffroy.lesur@univ-grenoble-alpes.fr>
+// and other code contributors
+// Licensed under CeCILL 2.1 License, see COPYING for more information
+// ***********************************************************************************
+
+#ifndef VERSION_H
+#define VERSION_H
+
+namespace VersionInfo {
+    inline constexpr const char* gitCommit  = "@GIT_SHA1@";
+    inline constexpr const char* versionMajor  = "@Idefix_VERSION_MAJOR@";
+    inline constexpr const char* versionMinor  = "@Idefix_VERSION_MINOR@";
+    inline constexpr const char* versionPatch  = "@Idefix_VERSION_PATCH@";
+    inline constexpr const char* version = "@Idefix_VERSION@";
+}
+
+#endif