Commit fa399562 authored by Giovanni Lacopo's avatar Giovanni Lacopo
Browse files

CUFFTMP included

parent feb9a976

Build/Makefile.local

+24 −0
Original line number Diff line number Diff line
@@ -13,10 +13,32 @@ CUDA_LIB = -L/opt/nvidia/hpc_sdk/Linux_x86_64/23.1/cuda/12.0/lib64 
FFTW_INCL=

FFTW_LIB=





##########################################################

#NVIDIA CUFFTMP



CUFFTMP_LIB  = -L/.../lib64

CUFFTMP_INCL = -I/.../include/cufftmp

##########################################################





##########################################################

#NVIDIA NCCL REDUCE



NCCL_INC = -I/opt/nvidia/hpc_sdk/Linux_x86_64/23.3/comm_libs/nccl/include

NCCL_LIB = -L/opt/nvidia/hpc_sdk/Linux_x86_64/23.3/comm_libs/nccl/lib

##########################################################



NVC = /opt/nvidia/hpc_sdk/Linux_x86_64/23.1/compilers/bin/nvc 

NVFLAGS = -O4 -fast -march=native $(OMP_GPU) -mavx -mavx2  

NVLIB = $(CUDA_INC) $(CUDA_LIB) -lcuda -lcudart





NVLIB_2 = $(CUDA_INC) $(CUDA_LIB) $(MPI_INC) $(MPI_LIB) $(CUFFT_INC) $(CUFFT_LIB) $(NVSHMEM_INC) $(NVSHMEM_LIB) -lnvshmem_host -lnvshmem_device -lcuda -lcudart -lcufftMp



NVLIB_3 = $(CUDA_INC) $(CUDA_LIB) $(MPI_INC) $(MPI_LIB) $(NCCL_INC) $(NCCL_LIB) -lcuda -lcudart -lnccl





NVCC = /opt/nvidia/hpc_sdk/Linux_x86_64/23.1/cuda/12.0/bin/nvcc

OPT_NVCC   = -std=c++17 --generate-code arch=compute_86,code=sm_86
@@ -25,6 +47,7 @@ CFLAGS += 
MPICHLIB =







##########################################################

#AMD GPUs (DEFAULT = LUMI)
@@ -41,3 +64,4 @@ HIP_LIB= -L/opt/rocm-5.2.3/hip/lib 


AMDLIB = $(HIP_INCL) $(HIP_LIB) $(RCCL_INCL) $(RCCL_LIB) -D__HIP_PLATFORM_AMD__ -lamdhip64 -lrccl

###########################################################

Makefile

+19 −3
Original line number Diff line number Diff line
@@ -110,7 +110,7 @@ DEPS = w-stacking.h main.c allvars.h 


# ----- define which files will be compiled by MPICC

#

# these are the OBJS that will be compiled by C compiler if no acceleration (neither with OpenACC nor with OpenMP) is provided

# these are the OBJS that will be compiled by C compiler if no acceleration (neither with CUDA nor with OpenMP) is provided

CC_OBJ_NOACC = allvars.o main.o init.o gridding.o gridding_cpu.o fourier_transform.o result.o numa.o reduce.o w-stacking.o phase_correction.o



# these are the OBJs that will be compiled by the normal MPICC compiler if GPU acceleration is switched on
@@ -136,6 +136,12 @@ OBJ_NCCL_REDUCE = gridding_nccl.o 
DEPS_RCCL_REDUCE = gridding_rccl.cpp

OBJ_RCCL_REDUCE  = gridding_rccl.o



# ----- define what files will be compiled by NVCC for Nvidia cufftMP implementation of FFT

#

DEPS_ACC_CUFFTMP = w-stacking.h cuda_fft.cu

OBJ_ACC_CUFFTMP  = cuda_fft.o





# -----------------------------------------------------

#

# end of OBJ definition
@@ -240,6 +246,16 @@ $(OBJ_RCCL_REDUCE): $(DEPS_RCCL_REDUCE) 
OBJ += $(OBJ_RCCL_REDUCE)

endif



ifeq (CUFFTMP,$(findstring CUFFTMP,$(OPT)))

EXEC_EXT := $(EXEC_EXT)_acc-fft

LINKER=$(MPIC++)

FLAGS=$(OPTIMIZE)

LIBS=$(NVLIB) $(NVLIB_2)

$(OBJ_ACC_CUFFTMP): $(DEPS_ACC_CUFFTMP)

	$(NVCC) $(OPT) $(OPT_NVCC) $(CFLAGS) -c $^ $(LIBS)

OBJ += $(OBJ_ACC_CUFFTMP)

endif





###################################################################################