64-BitBrainstorm_2026/AMSS-NCKU
8
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Switch build system from Intel oneAPI to GCC + OpenMPI

Browse Source 
- Replace compilers: ifx→gfortran, icx→gcc, icpx→g++, mpiicpx→mpicxx
- Replace flags: -xHost→-march=x86-64-v4, -ipo→-flto, -fpp→-cpp
- Replace flags: -fp-model fast=2→-ffast-math, -fma→-mfma
- Replace flags: -qopenmp→-fopenmp
- Remove Intel-specific: -align array64byte, -liomp5, -lifcore, -limf
- Switch MKL interface: -lmkl_intel_lp64→-lmkl_gf_lp64 (gfortran)
- Replace TBB malloc with optional jemalloc (default off)
- Disable PGO entirely (was already marked negative optimization)
- TwoPunctureABE and ABE both verified to build successfully

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
CGH0S7
2026-04-28 22:00:58 +08:00
parent f669180572
commit 9687d9a3dd
2 changed files with 137 additions and 174 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
AMSS_NCKU_source/makefile
View File

@@ -40,30 +40,13 @@ POLINT6_FLAG = -DPOLINT6_USE_BARYCENTRIC=$(POLINT6_USE_BARY)
TRANSFER_CACHE_FLAG = -DBSSN_USE_TRANSFER_CACHE=$(EFFECTIVE_USE_TRANSFER_CACHE) TRANSFER_CACHE_FLAG = -DBSSN_USE_TRANSFER_CACHE=$(EFFECTIVE_USE_TRANSFER_CACHE)
ESCALAR_KERNEL_FLAG = -DBSSN_USE_ESCALAR_C_KERNEL=$(EFFECTIVE_USE_CXX_ESCALAR_KERNEL) ESCALAR_KERNEL_FLAG = -DBSSN_USE_ESCALAR_C_KERNEL=$(EFFECTIVE_USE_CXX_ESCALAR_KERNEL)
## ABE build flags selected by PGO_MODE (set in makefile.inc, default: opt) ## GCC build flags (optimized for x86-64-v4)
## make -> opt (PGO-guided, maximum performance) ## PGO disabled (used negative optimization on Intel; not tested on GCC)
## make PGO_MODE=instrument -> instrument (Phase 1: collect fresh profile data) CXXAPPFLAGS = -O3 -march=x86-64-v4 -ffast-math -mfma -flto \
PROFDATA = /home/$(shell whoami)/AMSS-NCKU/pgo_profile/default.profdata
ifeq ($(PGO_MODE),instrument)
## Phase 1: instrumentation — omit -ipo/-fp-model fast=2 for faster build and numerical stability
CXXAPPFLAGS = -O3 -xHost -fma -fprofile-instr-generate -ipo \
-Dfortran3 -Dnewc -I${MKLROOT}/include $(INTERP_LB_FLAGS) \ -Dfortran3 -Dnewc -I${MKLROOT}/include $(INTERP_LB_FLAGS) \
$(TRANSFER_CACHE_FLAG) $(ESCALAR_KERNEL_FLAG) $(TRANSFER_CACHE_FLAG) $(ESCALAR_KERNEL_FLAG)
f90appflags = -O3 -xHost -fma -fprofile-instr-generate -ipo \ f90appflags = -O3 -march=x86-64-v4 -ffast-math -mfma -flto \
-align array64byte -fpp -I${MKLROOT}/include $(POLINT6_FLAG) -cpp -I${MKLROOT}/include $(POLINT6_FLAG)
else
## opt (default): maximum performance with PGO profile data -fprofile-instr-use=$(PROFDATA) \
## PGO has been turned off, now tested and found to be negative optimization
## INTERP_LB_FLAGS has been turned off too, now tested and found to be negative optimization
CXXAPPFLAGS = -O3 -xHost -fp-model fast=2 -fma -ipo \
-Dfortran3 -Dnewc -I${MKLROOT}/include $(INTERP_LB_FLAGS) \
$(TRANSFER_CACHE_FLAG) $(ESCALAR_KERNEL_FLAG)
f90appflags = -O3 -xHost -fp-model fast=2 -fma -ipo \
-align array64byte -fpp -I${MKLROOT}/include $(POLINT6_FLAG)
endif
.SUFFIXES: .o .f90 .C .for .cu .SUFFIXES: .o .f90 .C .for .cu
@@ -104,17 +87,15 @@ z4c_rhs_c.o: z4c_rhs_c.C
#interp_lb_profile.o: interp_lb_profile.C interp_lb_profile.h #interp_lb_profile.o: interp_lb_profile.C interp_lb_profile.h
# ${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@ # ${CXX} $(CXXAPPFLAGS) -c $< $(filein) -o $@
## TwoPunctureABE uses fixed optimal flags with its own PGO profile, independent of CXXAPPFLAGS ## TwoPunctureABE flags (no PGO; PGO was negative optimization)
TP_PROFDATA = /home/$(shell whoami)/AMSS-NCKU/pgo_profile/TwoPunctureABE.profdata TP_OPTFLAGS = -O3 -march=x86-64-v4 -ffast-math -mfma -flto \
TP_OPTFLAGS = -O3 -xHost -fp-model fast=2 -fma -ipo \
-fprofile-instr-use=$(TP_PROFDATA) \
-Dfortran3 -Dnewc -I${MKLROOT}/include -Dfortran3 -Dnewc -I${MKLROOT}/include
TwoPunctures.o: TwoPunctures.C TwoPunctures.o: TwoPunctures.C
${CXX} $(TP_OPTFLAGS) -qopenmp -c $< -o $@ ${CXX} $(TP_OPTFLAGS) -fopenmp -c $< -o $@
TwoPunctureABE.o: TwoPunctureABE.C TwoPunctureABE.o: TwoPunctureABE.C
${CXX} $(TP_OPTFLAGS) -qopenmp -c $< -o $@ ${CXX} $(TP_OPTFLAGS) -fopenmp -c $< -o $@
# Input files # Input files
@@ -232,7 +213,7 @@ ABEGPU: $(C++FILES_GPU) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(CUDAFILE
$(CLINKER) $(CXXAPPFLAGS) -o $@ $(C++FILES_GPU) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(CUDAFILES) $(LDLIBS) $(CLINKER) $(CXXAPPFLAGS) -o $@ $(C++FILES_GPU) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(CUDAFILES) $(LDLIBS)
TwoPunctureABE: $(TwoPunctureFILES) TwoPunctureABE: $(TwoPunctureFILES)
$(CLINKER) $(TP_OPTFLAGS) -qopenmp -o $@ $(TwoPunctureFILES) $(LDLIBS) $(CLINKER) $(TP_OPTFLAGS) -fopenmp -o $@ $(TwoPunctureFILES) $(LDLIBS)
clean: clean:
rm *.o ABE ABEGPU TwoPunctureABE make.log -f rm *.o ABE ABEGPU TwoPunctureABE make.log -f

44
AMSS_NCKU_source/makefile.inc
View File

@@ -1,33 +1,16 @@
## GCC version (commented out) ## GCC version with OpenMPI and oneMKL
## 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${MKLROOT}/include filein = -I/usr/include/ -I${MKLROOT}/include
## Using sequential MKL (OpenMP disabled for better single-threaded performance) ## Using MKL with gfortran interface (-lmkl_gf_lp64 instead of -lmkl_intel_lp64)
## Added -lifcore for Intel Fortran runtime and -limf for Intel math library LDLIBS = -L${MKLROOT}/lib -lmkl_gf_lp64 -lmkl_sequential -lmkl_core -lgfortran -lpthread -lm -ldl -lgomp
LDLIBS = -L${MKLROOT}/lib -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lifcore -limf -lpthread -lm -ldl -liomp5
## Memory allocator switch ## Memory allocator switch
## 1 (default) : link Intel oneTBB allocator (libtbbmalloc) ## 0 (default) : use system default allocator (ptmalloc)
## 0 : use system default allocator (ptmalloc) ## 1 : use jemalloc (install jemalloc-devel first)
USE_TBBMALLOC ?= 1 USE_JEMALLOC ?= 0
TBBMALLOC_SO ?= /home/intel/oneapi/2025.3/lib/libtbbmalloc.so ifeq ($(USE_JEMALLOC),1)
ifneq ($(wildcard $(TBBMALLOC_SO)),) LDLIBS := -ljemalloc $(LDLIBS)
TBBMALLOC_LIBS = -Wl,--no-as-needed $(TBBMALLOC_SO) -Wl,--as-needed
else
TBBMALLOC_LIBS = -Wl,--no-as-needed -ltbbmalloc -Wl,--as-needed
endif endif
ifeq ($(USE_TBBMALLOC),1)
LDLIBS := $(TBBMALLOC_LIBS) $(LDLIBS)
endif
## PGO build mode switch (ABE only; TwoPunctureABE always uses opt flags)
## opt : (default) maximum performance with PGO profile-guided optimization
## instrument : PGO Phase 1 instrumentation to collect fresh profile data
PGO_MODE ?= opt
## Interp_Points load balance profiling mode ## Interp_Points load balance profiling mode
## off : (default) no load balance instrumentation ## off : (default) no load balance instrumentation
@@ -70,13 +53,12 @@ USE_TRANSFER_CACHE ?= auto
## 0 : use original Fortran rungekutta4_rout.o ## 0 : use original Fortran rungekutta4_rout.o
USE_CXX_RK4 ?= 1 USE_CXX_RK4 ?= 1
f90 = ifx f90 = gfortran
f77 = ifx f77 = gfortran
CXX = icpx CXX = g++
CC = icx CC = gcc
CLINKER = mpiicpx CLINKER = mpicxx
Cu = nvcc Cu = nvcc
CUDA_LIB_PATH = -L/usr/lib/cuda/lib64 -I/usr/include -I/usr/lib/cuda/include 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_APP_FLAGS = -c -g -O3 --ptxas-options=-v -Dfortran3 -Dnewc