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

Switch back to Intel toolchain as the default option

Browse Source 
Seems that Intel MPI also supports CUDA-aware by setting I_MPI_OFFLOAD to 1. Besides, I_MPI_OFFLOAD_IPC=0 is needed to avoid segfaults.
This commit is contained in:
ianchb
2026-05-01 21:59:13 +08:00
parent 35b6ceff02
commit 06f62dee36
2 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
AMSS_NCKU_source/makefile.inc
View File

@@ -1,7 +1,7 @@
## Toolchain selection ## Toolchain selection
## nvhpc : NVIDIA HPC SDK + CUDA-aware MPI (default) ## nvhpc : NVIDIA HPC SDK + CUDA-aware MPI (default)
## intel : Intel oneAPI toolchain (legacy path) ## intel : Intel oneAPI toolchain (legacy path)
TOOLCHAIN ?= nvhpc TOOLCHAIN ?= intel
## PGO build mode switch (ABE only; TwoPunctureABE always uses opt flags) ## PGO build mode switch (ABE only; TwoPunctureABE always uses opt flags)
## opt : (default) maximum performance with PGO profile-guided optimization ## opt : (default) maximum performance with PGO profile-guided optimization

2
makefile_and_run.py
View File

@@ -266,7 +266,7 @@ def run_ABE():
#mpi_command = " mpirun -np " + str(input_data.MPI_processes) + " ./ABE" #mpi_command = " mpirun -np " + str(input_data.MPI_processes) + " ./ABE"
mpi_command_outfile = "ABE_out.log" mpi_command_outfile = "ABE_out.log"
elif (input_data.GPU_Calculation == "yes"): elif (input_data.GPU_Calculation == "yes"):
mpi_command = NUMACTL_CPU_BIND + " mpirun -np " + str(input_data.MPI_processes) + " ./ABE_CUDA" mpi_command = NUMACTL_CPU_BIND + " I_MPI_OFFLOAD=1 I_MPI_OFFLOAD_IPC=0 mpirun -np " + str(input_data.MPI_processes) + " ./ABE_CUDA"
mpi_command_outfile = "ABEGPU_out.log" mpi_command_outfile = "ABEGPU_out.log"
mpi_env = _gpu_runtime_env() mpi_env = _gpu_runtime_env()
started_mps = _start_cuda_mps_if_requested(mpi_env) started_mps = _start_cuda_mps_if_requested(mpi_env)