This is an old revision of the document!
Howto Optimize Cuda Kernels for Libcusmm
Step 1: Go to the directory libcusmm directory
$ cd $CP2K_ROOT/src/dbcsr/cuda/libcusmm
Step 2: Adopt tune.py for your Environment
The tune.py
script generates job files. You have to adopt the script to the environment of your supercomputer and your personal settings.
... def gen_jobfile(outdir, m, n, k): t = "/tune_%dx%dx%d"%(m,n,k) all_exe_src = [basename(fn) for fn in glob(outdir+t+"_*_main.cu")] all_exe = sorted([fn.replace("_main.cu", "") for fn in all_exe_src]) output = "#!/bin/bash -l\n" output += "#SBATCH --nodes=%d\n"%len(all_exe) output += "#SBATCH --time=0:30:00\n" output += "#SBATCH --account=s441\n" output += "\n" output += "source ${MODULESHOME}/init/sh;\n" output += "module unload PrgEnv-cray\n" output += "module load cudatoolkit PrgEnv-gnu\n" output += "module list\n" output += "cd $SLURM_SUBMIT_DIR \n" output += "\n" output += "date\n" for exe in all_exe: output += "aprun -b -n 1 -N 1 -d 8 make -j 16 %s &\n"%exe ...
Step 3: Run the script tune.py
The script takes as arguments the blocksizes you want to add to libcusmm. For example, if your system contains blocks of size 5 and 8 type:
$ ./tune.py 5 8 Found 23 parameter sets for 5x5x5 Found 31 parameter sets for 5x5x8 Found 107 parameter sets for 5x8x5 Found 171 parameter sets for 5x8x8 Found 75 parameter sets for 8x5x5 Found 107 parameter sets for 8x5x8 Found 248 parameter sets for 8x8x5 Found 424 parameter sets for 8x8x8
The script will create a directory for each combination of the blocksizes:
$ ls -d tune_* tune_5x5x5 tune_5x5x8 tune_5x8x5 tune_5x8x8 tune_8x5x5 tune_8x5x8 tune_8x8x5 tune_8x8x8
Each directory contains a number of files:
$ ls -1 tune_8x8x8/ Makefile tune_8x8x8_exe0_main.cu tune_8x8x8_exe0_part0.cu tune_8x8x8_exe0_part1.cu tune_8x8x8_exe0_part2.cu tune_8x8x8_exe0_part3.cu tune_8x8x8_exe0_part4.cu tune_8x8x8.job
For each possible parameter-set a launcher is generated. A launcher is a small snipped of C code, which launches the kernel by using the cuda specific <<< >>>
-notation. It also instantiates the C++ template which contains the actual kernel code.
In order to parallelize the benchmarking the launchers are distributed over multiple executables.
Currently, up to 10000 launchers are benchmarked by one executable. Each executable is linked together from several tune_*_part???.o
and a tune_*_main.o
. Each part-files contains up to 100 launchers. This allows to parallelize the compilation over multiple CPU cores.
Step 4: Adopt submit.py for your Environment
The script submit.py
was written for the slurm batch system as used e.g. by CRAY supercomputers. If your computer runs a different batch system you have to adopt submit.py
accordingly.
Step 5: Submit Jobs
Each tune-directory contains a job file.
Since, there might be many tune-directories the convenience script submit.py
can be used. It will go through all the tune_*
-directories and check if it has already been submitted or run. For this the script calls squeue
in the background and it searches for slurm-*.out
files.
When submit.py
is called without arguments it will just list the jobs that could be submitted:
$ ./submit.py tune_5x5x5: Would submit, run with "doit!" tune_5x5x8: Would submit, run with "doit!" tune_5x8x5: Would submit, run with "doit!" tune_5x8x8: Would submit, run with "doit!" tune_8x5x5: Would submit, run with "doit!" tune_8x5x8: Would submit, run with "doit!" tune_8x8x5: Would submit, run with "doit!" tune_8x8x8: Would submit, run with "doit!" Number of jobs submitted: 8
Only when submit.py
is called with doit!
as its first argument it will actually submit jobs:
$ ./submit.py doit! tune_5x5x5: Submitting Submitted batch job 277987 tune_5x5x8: Submitting Submitted batch job 277988 tune_5x8x5: Submitting Submitted batch job 277989 tune_5x8x8: Submitting Submitted batch job 277990 tune_8x5x5: Submitting Submitted batch job 277991 tune_8x5x8: Submitting Submitted batch job 277992 tune_8x8x5: Submitting Submitted batch job 277993 tune_8x8x8: Submitting Submitted batch job 277994 Number of jobs submitted: 8
Step 5: Collect Results
Run collect.py
to parse all log files and to determine the best kernel for each blocksize:
$ ./collect.py Reading: tune_5x5x5/tune_5x5x5_exe0.log Reading: tune_5x5x8/tune_5x5x8_exe0.log Reading: tune_5x8x5/tune_5x8x5_exe0.log Reading: tune_5x8x8/tune_5x8x8_exe0.log Reading: tune_8x5x5/tune_8x5x5_exe0.log Reading: tune_8x5x8/tune_8x5x8_exe0.log Reading: tune_8x8x5/tune_8x8x5_exe0.log Reading: tune_8x8x8/tune_8x8x8_exe0.log Kernel_dnt_tiny(m=5, n=5, k=5, split_thread=32, threads=64, grouping=16, minblocks=1) , # 27.9623 GFlops Kernel_dnt_tiny(m=5, n=5, k=8, split_thread=32, threads=96, grouping=16, minblocks=1) , # 37.8978 GFlops Kernel_dnt_medium(m=5, n=8, k=5, tile_m=1, tile_n=1, threads=96, grouping=16, minblocks=8) , # 32.9231 GFlops Kernel_dnt_tiny(m=5, n=8, k=8, split_thread=32, threads=96, grouping=16, minblocks=1) , # 47.0366 GFlops Kernel_dnt_medium(m=8, n=5, k=5, tile_m=1, tile_n=1, threads=96, grouping=16, minblocks=12) , # 33.1999 GFlops Kernel_dnt_medium(m=8, n=5, k=8, tile_m=1, tile_n=1, threads=96, grouping=16, minblocks=12) , # 49.3499 GFlops Kernel_dnt_tiny(m=8, n=8, k=5, split_thread=32, threads=96, grouping=16, minblocks=1) , # 62.8469 GFlops Kernel_dnt_tiny(m=8, n=8, k=8, split_thread=32, threads=128, grouping=16, minblocks=1) , # 90.7763 GFlops