ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add switch between compilers, make 9110 compiler default, add full QA scripts. (#322)

Browse Source 
* adding scripts for full perf test suite

* uncomment the sql queries

* fix typo and chmod a+x for scripts

* dos2unix for all new scripts

* disable verification in full performance test

* fix reduction scripts, add gfrouped_gemm hotfix

* fix the grouped_gemm hotfix and only run reduction for fp16

* change compiler flag syntax

* fix syntax

* add predefinition of dockerArgs

* avoid redefinitions of dockerArgs

* add blank space at the end of dockerArgs

* try to build with release compiler

* adding spaces inside if condition

* limit the number of threads for building 9110 compiler

* change the way HIP_CLANG_PATH is set

* remove the export command

* change the conditional ENV syntax

* set HIP_CLANG_PATH at docker run time

* update scripts for full qa

* enable the sql write query

* fix typo

* remove a comment from a script
This commit is contained in:
Illia Silin
2022-07-13 07:27:43 -07:00
committed by GitHub
parent 639147432b
commit 39acaea36d
12 changed files with 921 additions and 352 deletions
Download Patch File Download Diff File Expand all files Collapse all files

54
script/run_performance_tests.sh
View File

@@ -10,17 +10,27 @@ pip3 install sqlalchemy pymysql pandas sshtunnel
# post your new test results to the database and compare them to the baseline
# please contact Illia.Silin@amd.com for more details
#
# run the script as "./run_performance_tests.sh <tag for your test environment>
#get the test environment type:
export env_type=$1
echo 'Environment type ' $env_type
function print_log_header(){
rm -f $1;
git status | grep -e 'On branch' > $1;
echo -n 'Node name: ' >>$1; hostname >> $1;
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> $1; rocminfo | grep "Name:" | grep "gfx" >> $1;
rocminfo | grep "Compute Unit:" >> $1;
hipcc --version | grep -e 'HIP version' >> $1;
echo 'Environment type: ' $2 >>$1;
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> $1;
}
#run gemm tests
export gemm_log="perf_gemm.log"
rm -f $gemm_log
git status | grep -e 'On branch' > ${gemm_log}
echo -n 'Node name: ' >>${gemm_log}; hostname >> ${gemm_log}
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> ${gemm_log}; rocminfo | grep "Name:" | grep "gfx" >> ${gemm_log}
rocminfo | grep "Compute Unit:" >> ${gemm_log}
hipcc --version | grep -e 'HIP version' >> ${gemm_log}
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${gemm_log}
./profile_gemm.sh gemm 0 0 0 1 0 5 | tee -a ${gemm_log}
print_log_header $gemm_log $env_type
./profile_gemm.sh gemm 0 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 1 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 0 0 1 0 5 | tee -a $gemm_log
@@ -36,22 +46,14 @@ hipcc --version | grep -e 'HIP version' >> ${gemm_log}
./profile_gemm.sh gemm 1 3 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 3 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 3 0 1 0 5 | tee -a $gemm_log
python3 parse_perf_data.py ${gemm_log}
python3 process_perf_data.py $gemm_log
#run resnet50 test
export resnet_log="perf_resnet50.log"
rm -f $resnet_log
git status | grep -e 'On branch' > ${resnet_log}
echo -n 'Node name: '>>${resnet_log}; hostname >>${resnet_log}
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> ${resnet_log}; rocminfo | grep "Name:" | grep "gfx" >> ${resnet_log}
rocminfo | grep "Compute Unit:" >> ${resnet_log}
hipcc --version | grep -e 'HIP version' >> ${resnet_log}
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${resnet_log}
#first run tests with N=256
./profile_conv.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 256 | tee -a ${resnet_log}
#then run with N=4
./profile_conv.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 4 | tee -a ${resnet_log}
#the script will put the results from N=256 and N=4 runs into separate tables
python3 parse_perf_data.py ${resnet_log}
export resnet256_log="perf_resnet50_N256.log"
print_log_header $resnet256_log $env_type
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 256 | tee -a $resnet256_log
python3 process_perf_data.py $resnet256_log
export resnet4_log="perf_resnet50_N4.log"
print_log_header $resnet4_log $env_type
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 4 | tee -a $resnet4_log
python3 process_perf_data.py $resnet4_log