Switch to NVIDIA HPC Toolchain

2026-04-29 08:30:47 +08:00
parent c689cc8dc9
commit a0dab90bcb
2 changed files with 75 additions and 46 deletions
--- a/AMSS_NCKU_source/makefile
+++ b/AMSS_NCKU_source/makefile
@@ -13,12 +13,15 @@ POLINT6_FLAG = -DPOLINT6_USE_BARYCENTRIC=$(POLINT6_USE_BARY)
 ##   make PGO_MODE=instrument    -> instrument (Phase 1: collect fresh profile data)
 PROFDATA = /home/$(shell whoami)/AMSS-NCKU/pgo_profile/default.profdata

+ifeq ($(TOOLCHAIN),intel)
+OMP_FLAG = -qopenmp
+
 ifeq ($(PGO_MODE),instrument)
-## Phase 1: instrumentation — omit -ipo/-fp-model fast=2 for faster build and numerical stability
+## Intel Phase 1: instrumentation — omit -ipo/-fp-model fast=2 for faster build and numerical stability
 CXXAPPFLAGS = -O3 -xHost -fma -fprofile-instr-generate -ipo \
-              -Dfortran3 -Dnewc -I${MKLROOT}/include $(INTERP_LB_FLAGS)
+              -Dfortran3 -Dnewc $(MKL_INC) $(INTERP_LB_FLAGS)
 f90appflags = -O3 -xHost -fma -fprofile-instr-generate -ipo \
-              -align array64byte -fpp -I${MKLROOT}/include $(POLINT6_FLAG)
+              -align array64byte -fpp $(MKL_INC) $(POLINT6_FLAG)
 else
 ## opt (default): maximum performance with PGO profile data -fprofile-instr-use=$(PROFDATA) \
 ## PGO has been turned off, now tested and found to be negative optimization
@@ -26,9 +29,24 @@ else


 CXXAPPFLAGS = -O3 -xHost -fp-model fast=2 -fma -ipo \
-              -Dfortran3 -Dnewc -I${MKLROOT}/include $(INTERP_LB_FLAGS)
+              -Dfortran3 -Dnewc $(MKL_INC) $(INTERP_LB_FLAGS)
 f90appflags = -O3 -xHost -fp-model fast=2 -fma -ipo \
-              -align array64byte -fpp -I${MKLROOT}/include $(POLINT6_FLAG)
+              -align array64byte -fpp $(MKL_INC) $(POLINT6_FLAG)
+endif
+
+TP_OPTFLAGS = -O3 -xHost -fp-model fast=2 -fma -ipo \
+              -fprofile-instr-use=$(TP_PROFDATA) \
+              -Dfortran3 -Dnewc $(MKL_INC)
+else
+## NVHPC defaults: mpicc/mpicxx/mpifort wrappers
+## PGO_MODE is ignored in this branch.
+OMP_FLAG = -mp
+CXXAPPFLAGS = -O3 -tp=host -Mcache_align -Mfma \
+              -Dfortran3 -Dnewc $(MKL_INC) $(INTERP_LB_FLAGS)
+f90appflags = -O3 -tp=host -Mcache_align -Mfma -Mpreprocess \
+              $(MKL_INC) $(POLINT6_FLAG)
+TP_OPTFLAGS = -O3 -tp=host -Mcache_align -Mfma \
+              -Dfortran3 -Dnewc $(MKL_INC)
 endif

 .SUFFIXES: .o .f90 .C .for .cu
@@ -78,17 +96,11 @@ z4c_rhs_c.o: z4c_rhs_c.C
 #interp_lb_profile.o: interp_lb_profile.C interp_lb_profile.h
 #	${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@

-## TwoPunctureABE uses fixed optimal flags with its own PGO profile, independent of CXXAPPFLAGS
-TP_PROFDATA = /home/$(shell whoami)/AMSS-NCKU/pgo_profile/TwoPunctureABE.profdata
-TP_OPTFLAGS = -O3 -xHost -fp-model fast=2 -fma -ipo \
-              -fprofile-instr-use=$(TP_PROFDATA) \
-              -Dfortran3 -Dnewc -I${MKLROOT}/include
-
 TwoPunctures.o: TwoPunctures.C
-	${CXX} $(TP_OPTFLAGS) -qopenmp -c $< -o $@
+	${CXX} $(TP_OPTFLAGS) $(OMP_FLAG) -c $< -o $@

 TwoPunctureABE.o: TwoPunctureABE.C
-	${CXX} $(TP_OPTFLAGS) -qopenmp -c $< -o $@
+	${CXX} $(TP_OPTFLAGS) $(OMP_FLAG) -c $< -o $@

 # Input files

@@ -242,7 +254,7 @@ ABE_CUDA: $(C++FILES) $(ABE_CUDA_CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS)
 #	$(CLINKER) $(CXXAPPFLAGS) -o $@ $(C++FILES_GPU) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(CUDAFILES) $(LDLIBS)

 TwoPunctureABE: $(TwoPunctureFILES)
-	$(CLINKER) $(TP_OPTFLAGS) -qopenmp -o $@ $(TwoPunctureFILES) $(LDLIBS)
+	$(CLINKER) $(TP_OPTFLAGS) $(OMP_FLAG) -o $@ $(TwoPunctureFILES) $(LDLIBS)

 clean:
 	rm *.o ABE ABE_CUDA ABEGPU TwoPunctureABE make.log -f
--- a/AMSS_NCKU_source/makefile.inc
+++ b/AMSS_NCKU_source/makefile.inc
@@ -1,28 +1,7 @@
-## GCC version (commented out)
-## filein  = -I/usr/include -I/usr/lib/x86_64-linux-gnu/mpich/include -I/usr/lib/x86_64-linux-gnu/openmpi/lib/ -I/usr/lib/gcc/x86_64-linux-gnu/11/ -I/usr/include/c++/11/
-## filein  = -I/usr/include/ -I/usr/include/openmpi-x86_64/ -I/usr/lib/x86_64-linux-gnu/openmpi/include/ -I/usr/lib/x86_64-linux-gnu/openmpi/lib/ -I/usr/lib/gcc/x86_64-linux-gnu/11/ -I/usr/include/c++/11/
-## LDLIBS  = -L/usr/lib/x86_64-linux-gnu -L/usr/lib64 -L/usr/lib/gcc/x86_64-linux-gnu/11 -lgfortran -lmpi -lgfortran
-
-## Intel oneAPI version with oneMKL (Optimized for performance)
-filein  = -I/usr/include/ -I${MKLROOT}/include
-
-## Using sequential MKL (OpenMP disabled for better single-threaded performance)
-## Added -lifcore for Intel Fortran runtime and -limf for Intel math library
-LDLIBS  = -L${MKLROOT}/lib -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lifcore -limf -lpthread -lm -ldl -liomp5
-
-## Memory allocator switch
-##   1 (default) : link Intel oneTBB allocator (libtbbmalloc)
-##   0           : use system default allocator (ptmalloc)
-USE_TBBMALLOC ?= 1
-TBBMALLOC_SO ?= /home/intel/oneapi/2025.3/lib/libtbbmalloc.so
-ifneq ($(wildcard $(TBBMALLOC_SO)),)
-TBBMALLOC_LIBS = -Wl,--no-as-needed $(TBBMALLOC_SO) -Wl,--as-needed
-else
-TBBMALLOC_LIBS = -Wl,--no-as-needed -ltbbmalloc -Wl,--as-needed
-endif
-ifeq ($(USE_TBBMALLOC),1)
-LDLIBS := $(TBBMALLOC_LIBS) $(LDLIBS)
-endif
+## Toolchain selection
+##   nvhpc : NVIDIA HPC SDK + CUDA-aware MPI (default)
+##   intel : Intel oneAPI toolchain (legacy path)
+TOOLCHAIN ?= nvhpc

 ## PGO build mode switch (ABE only; TwoPunctureABE always uses opt flags)
 ##   opt        : (default) maximum performance with PGO profile-guided optimization
@@ -43,6 +22,14 @@ else
 INTERP_LB_FLAGS =
 endif

+MKLROOT ?= /home/intel/oneapi/mkl/latest
+MKL_LIBDIR ?= $(MKLROOT)/lib/intel64
+MKL_INC ?= -I$(MKLROOT)/include
+
+NVHPC_ROOT ?= /home/nvidia/hpc_sdk/Linux_x86_64/25.11
+CUDA_HOME  ?= $(NVHPC_ROOT)/cuda
+CUDA_ARCH  ?= sm_80
+
 ## Kernel implementation switch
 ##   1 (default) : use C++ rewrite of bssn_rhs and helper kernels (faster)
 ##   0           : fall back to original Fortran kernels
@@ -58,17 +45,47 @@ USE_CXX_Z4C_KERNELS ?= 1
 ##   0           : use original Fortran rungekutta4_rout.o
 USE_CXX_RK4 ?= 1

+## Memory allocator switch
+##   1 (default) : link Intel oneTBB allocator (libtbbmalloc)
+##   0           : use system default allocator (ptmalloc)
+USE_TBBMALLOC ?= 1
+TBBMALLOC_SO ?= /home/intel/oneapi/2025.3/lib/libtbbmalloc.so
+ifneq ($(wildcard $(TBBMALLOC_SO)),)
+TBBMALLOC_LIBS = -Wl,--no-as-needed $(TBBMALLOC_SO) -Wl,--as-needed
+else
+TBBMALLOC_LIBS = -Wl,--no-as-needed -ltbbmalloc -Wl,--as-needed
+endif
+
+ifeq ($(TOOLCHAIN),intel)
 f90          = ifx
 f77          = ifx
 CXX          = icpx
 CC           = icx
 CLINKER      = mpiicpx
+filein  = -I/usr/include/ $(MKL_INC) -I$(CUDA_HOME)/include
+LDLIBS       = -L$(MKL_LIBDIR) -Wl,-rpath,$(MKL_LIBDIR) \
+               -lmkl_intel_lp64 -lmkl_sequential -lmkl_core \
+               -lifcore -limf -liomp5 -lpthread -lm -ldl \
+               -L$(CUDA_HOME)/lib64 -Wl,-rpath,$(CUDA_HOME)/lib64 -lcuda -lcudart
+else ifeq ($(TOOLCHAIN),nvhpc)
+f90          = mpifort
+f77          = mpifort
+CXX          = mpicxx
+CC           = mpicc
+CLINKER      = mpicxx

-Cu = nvcc
-CUDA_LIB_PATH = -L/usr/lib/cuda/lib64 -I/usr/include -I/usr/lib/cuda/include
-#CUDA_APP_FLAGS = -c -g -O3 --ptxas-options=-v -arch compute_13 -code compute_13,sm_13 -Dfortran3 -Dnewc
-CUDA_APP_FLAGS = -c -g -O3 --ptxas-options=-v -Dfortran3 -Dnewc
-CUDA_ARCH ?= sm_80
-ifneq ($(strip $(CUDA_ARCH)),)
-CUDA_APP_FLAGS += -arch=$(CUDA_ARCH)
+filein       = -I/usr/include/ $(MKL_INC) -I$(CUDA_HOME)/include
+LDLIBS       = -L$(MKL_LIBDIR) -Wl,-rpath,$(MKL_LIBDIR) \
+               -lmkl_intel_lp64 -lmkl_sequential -lmkl_core \
+               -lpthread -lm -ldl \
+               -L$(CUDA_HOME)/lib64 -Wl,-rpath,$(CUDA_HOME)/lib64 -lcuda -lcudart \
+               -fortranlibs
 endif
+
+ifeq ($(USE_TBBMALLOC),1)
+LDLIBS := $(TBBMALLOC_LIBS) $(LDLIBS)
+endif
+
+Cu = $(NVHPC_ROOT)/compilers/bin/nvcc
+CUDA_LIB_PATH = -L$(CUDA_HOME)/lib64 -I$(CUDA_HOME)/include
+CUDA_APP_FLAGS = -c -g -O3 --ptxas-options=-v -Dfortran3 -Dnewc -arch=$(CUDA_ARCH)