Migrate build system from Intel oneAPI to AMD AOCC/AOCL/OpenMPI

Replace Intel compilers (ifx/icpx/icx) with AOCC (flang/clang++/clang),
Intel MPI (mpiicpx) with AOCC-built OpenMPI (mpicxx), and Intel MKL
with AOCL BLIS/libFLAME. Replace -xHost with -march=znver4, -ipo with
-flto, -fp-model fast=2 with -ffast-math, -qopenmp with -fopenmp.
Remove PGO, TBB allocator, and Intel-specific runtime libraries.
Fix MKL-specific includes in TwoPunctures.C and gaussj.C to use
standard CBLAS/LAPACKE headers from AOCL.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

AMSS_NCKU_source/TwoPunctures.C

@@ -27,7 +27,7 @@ using namespace std;
 #endif
 
 #include "TwoPunctures.h"
-#include <mkl_cblas.h>
+#include <cblas.h>
 
 TwoPunctures::TwoPunctures(double mp, double mm, double b,
                            double P_plusx, double P_plusy, double P_plusz,

AMSS_NCKU_source/gaussj.C

@@ -18,7 +18,7 @@ using namespace std;
 #endif
 
 // Intel oneMKL LAPACK interface
-#include <mkl_lapacke.h>
+#include <lapacke.h>
 /* Linear equation solution using Intel oneMKL LAPACK.
 a[0..n-1][0..n-1] is the input matrix. b[0..n-1] is input
 containing the right-hand side vectors. On output a is

AMSS_NCKU_source/makefile

@@ -40,30 +40,12 @@ POLINT6_FLAG = -DPOLINT6_USE_BARYCENTRIC=$(POLINT6_USE_BARY)
 TRANSFER_CACHE_FLAG = -DBSSN_USE_TRANSFER_CACHE=$(EFFECTIVE_USE_TRANSFER_CACHE)
 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)
+## AMD AOCC build flags optimized for EPYC Zen 4 (-march=znver4)
-##   make                        -> opt  (PGO-guided, maximum performance)
+CXXAPPFLAGS = -O3 -march=znver4 -ffast-math -flto \
-##   make PGO_MODE=instrument    -> instrument (Phase 1: collect fresh profile data)
+              -Dfortran3 -Dnewc -I$(AOCL_ROOT)/include $(INTERP_LB_FLAGS) \
-PROFDATA = /home/$(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) \
               $(TRANSFER_CACHE_FLAG) $(ESCALAR_KERNEL_FLAG)
-f90appflags = -O3 -xHost -fma -fprofile-instr-generate -ipo \
+f90appflags = -O3 -march=znver4 -ffast-math -flto \
-              -align array64byte -fpp -I${MKLROOT}/include $(POLINT6_FLAG)
+              -cpp -I$(AOCL_ROOT)/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
 
@@ -101,17 +83,15 @@ lopsided_kodis_c.o: lopsided_kodis_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
+## TwoPunctureABE uses fixed optimal flags (AMD AOCC, no PGO)
-TP_PROFDATA = /home/$(shell whoami)/AMSS-NCKU/pgo_profile/TwoPunctureABE.profdata
+TP_OPTFLAGS = -O3 -march=znver4 -ffast-math -flto \
-TP_OPTFLAGS = -O3 -xHost -fp-model fast=2 -fma -ipo \
+              -Dfortran3 -Dnewc -I$(AOCL_ROOT)/include
-              -fprofile-instr-use=$(TP_PROFDATA) \
-              -Dfortran3 -Dnewc -I${MKLROOT}/include
 
 TwoPunctures.o: TwoPunctures.C
-	${CXX} $(TP_OPTFLAGS) -qopenmp -c $< -o $@
+	${CXX} $(TP_OPTFLAGS) -fopenmp -c $< -o $@
 
 TwoPunctureABE.o: TwoPunctureABE.C
-	${CXX} $(TP_OPTFLAGS) -qopenmp -c $< -o $@
+	${CXX} $(TP_OPTFLAGS) -fopenmp -c $< -o $@
 
 # Input files
 
@@ -222,7 +202,7 @@ ABEGPU: $(C++FILES_GPU) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(CUDAFILE
 	$(CLINKER) $(CXXAPPFLAGS) -o $@ $(C++FILES_GPU) $(CFILES) $(F90FILES) $(F77FILES) $(AHFDOBJS) $(CUDAFILES) $(LDLIBS)
 
 TwoPunctureABE: $(TwoPunctureFILES)
-	$(CLINKER) $(TP_OPTFLAGS) -qopenmp -o $@ $(TwoPunctureFILES) $(LDLIBS)
+	$(CLINKER) $(TP_OPTFLAGS) -fopenmp -o $@ $(TwoPunctureFILES) $(LDLIBS)
 
 clean:
 	rm *.o ABE ABEGPU TwoPunctureABE make.log -f

AMSS_NCKU_source/makefile.inc

@@ -1,33 +1,17 @@
-## GCC version (commented out)
+## AMD AOCC version with AOCL (Optimized for AMD EPYC Zen 4)
-## 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)
+## AOCL root path for includes and libraries
-filein  = -I/usr/include/ -I${MKLROOT}/include
+AOCL_ROOT ?= /home/gh0s7/AOCC/aocl/5.2.0/aocc
 
-## Using sequential MKL (OpenMP disabled for better single-threaded performance)
+## AOCC-built OpenMPI prefix
-## Added -lifcore for Intel Fortran runtime and -limf for Intel math library
+OMPI_PREFIX ?= /home/gh0s7/AOCC/aocc-openmpi
-LDLIBS  = -L${MKLROOT}/lib -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lifcore -limf -lpthread -lm -ldl -liomp5
 
-## Memory allocator switch
+filein  = -I/usr/include/ -I$(AOCL_ROOT)/include
-##   1 (default) : link Intel oneTBB allocator (libtbbmalloc)
-##   0           : use system default allocator (ptmalloc)
-USE_TBBMALLOC ?= 1
-TBBMALLOC_SO ?= /home/intel/oneapi/2025.3/lib/libtbbmalloc.so
-ifneq ($(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
 
-## PGO build mode switch (ABE only; TwoPunctureABE always uses opt flags)
+## Using AOCL BLIS + libFLAME for BLAS/LAPACK
-##   opt        : (default) maximum performance with PGO profile-guided optimization
+## AOCC Fortran runtime: -lflang (includes FortranRuntime)
-##   instrument : PGO Phase 1 instrumentation to collect fresh profile data
+## AOCC OpenMP runtime: -lomp (LLVM OpenMP)
-PGO_MODE ?= opt
+LDLIBS  = -L$(AOCL_ROOT)/lib -lblis -lflame -lamdlibm -lflang -lpgmath -lpthread -lm -ldl -lomp
 
 ## Interp_Points load balance profiling mode
 ##   off        : (default) no load balance instrumentation
@@ -65,11 +49,11 @@ USE_TRANSFER_CACHE ?= auto
 ##   0           : use original Fortran rungekutta4_rout.o
 USE_CXX_RK4 ?= 1
 
-f90          = ifx
+f90          = flang
-f77          = ifx
+f77          = flang
-CXX          = icpx
+CXX          = clang++
-CC           = icx
+CC           = clang
-CLINKER      = mpiicpx
+CLINKER      = $(OMPI_PREFIX)/bin/mpicxx
 
 Cu = nvcc
 CUDA_LIB_PATH = -L/usr/lib/cuda/lib64 -I/usr/include -I/usr/lib/cuda/include