Adding quickTunerStat.py #1583
Adding quickTunerStat.py #1583
Conversation
files from quickTunerGen.py. Can either compare against dataframe created by quickTunerStat.py or by running rocmlir-tuning-driver (currently disabled).
to 'NormalizedTFlops' for extracting performance information when validating with --data flag. Cleaned up --tuning code, still disabled until further testing is done.
|
|
||
| Generated statistics and verify quick tuning configs | ||
|
|
||
| optional arguments: |
There was a problem hiding this comment.
Why's there a copy of usage here? Shouldn't argparse generate it? Some examples might be good
|
|
||
| if match: | ||
| tup = match.groups() | ||
| transA = True if tup[0].lower() == 'true' else False |
There was a problem hiding this comment.
The GemmConfig class over in tuningRunner already has this and it might be less fragile?
There was a problem hiding this comment.
The issue is I still need to create a tuple in order to index the dictionary contained within the class. I switched out the method of converting the transA/B to the same method used in tuningRunner/perfRunner
There was a problem hiding this comment.
I'm more suggesting using the classes in tuningRunner or perfRunner to parse the perf config - I think parameterSweeps.py might be a good place to look or refactor
There was a problem hiding this comment.
Of course I will take a look at that
| tile_params = tile_params.drop(['param8','param9'], axis=1) | ||
|
|
||
| tile_params['performance'] = df['NormalizedTFlops'] | ||
|
|
There was a problem hiding this comment.
Nit: a bunch of extra newlines that don't have semantic meaning, please remove?
| self.output_df = pd.DataFrame(rank_dict) | ||
| print(self.output_df) | ||
|
|
||
| class tunerValidator(perfConfigValidator): |
There was a problem hiding this comment.
Nit: upper case class names
| rocmlir_path, | ||
| rocm_build_script='/share/scripts/build-rocm'): | ||
| """ | ||
| initializer |
|
|
||
| self.cpp_file = os.path.join(os.path.dirname(rocmlir_path), self.gridwise_gemm_params) | ||
| self.rocm_build_script = rocm_build_script | ||
| self.backup = self.cpp_file + ".bu" |
There was a problem hiding this comment.
~ is the usual suffix for backup files, no?
|
|
||
|
|
||
| def __del__(self): | ||
| # copy back original file |
updated column names, correcting previous splitK mistake.
Added cpp file output on invocation of --rank
perfRunner.GemmConfiguration.fromCommandLine()
| from dataclasses import dataclass | ||
| from perfCommonUtils import Operation | ||
| import perfRunner | ||
| from sklearn.preprocessing import MinMaxScaler |
There was a problem hiding this comment.
We might want to add sklearn to a Dockerfile or a requirements.txt somewhere?
| # to validate file we need data already read, | ||
|
|
||
| all_data = [] | ||
| columns = ['M/block', 'N/block', 'K/block', 'M/wave', 'N/wave', 'kPack', 'splitK', 'forceUnroll', 'bCopyMore'] |
There was a problem hiding this comment.
I'll note that, for example, the attention perf configs don't have this structure, and, more interestingl, that f32 gemm/conv on Navi doesn't either. But this is probably fine, just wanted to call out the limitation
| with open(cpp_filename, 'w') as file: | ||
| file.write(cpp_content) | ||
|
|
||
| def buildRocm(self): |
There was a problem hiding this comment.
buildRocmlir, maybe?
Also, why is this here?
Adding statistics and testing script
quickTunerStat.py, accepts*.qtfilesfromquickTunerGen.py. Can either compare against data frame created byquickTunerStat.pyor by runningrocmlir-tuning-driver(currently disabled).