OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity
master
pytorch/torch/_inductor/analysis
History
Nicolas Macchioni 75f798e05b [inductor][mi350] add tech specs for MI350 (#166576) 
Summary:
was digging through matmul padding for other work, and I noticed that the compute bound checking won't work on MI350 since we haven't supplied the tech specs yet.

I added MI350 specs following the predefined format

Test Plan: CI

Differential Revision: D85804980

Pull Request resolved: https://github.com/pytorch/pytorch/pull/166576
Approved by: https://github.com/leitian
2025-10-30 03:46:52 +00:00
..
__init__.py Inductor logging + analysis of torch.profile (#149697) 2025-07-07 22:13:34 +00:00
device_info.py [inductor][mi350] add tech specs for MI350 (#166576) 2025-10-30 03:46:52 +00:00
profile_analysis.py Fix pyrefly ignore syntax in _inductor (#166247) 2025-10-27 02:48:42 +00:00
README.md [BE][2/16] fix typos in torch/ (torch/_*/) (#156312) 2025-07-12 05:47:06 +00:00

README.md

torch._inductor.analysis

Contains scripts for inductor performance analysis.

Analysis

This will analyze a chrome trace to create a table useful for performance work. We mainly care about . Currently, it will add the flops and the memory reads of a kernel via formula (it's not looking at program counters or anything.) These, combined with the kernel duration, can be use to calculate achieved flops, achieved memory bandwidth, and roofline calculations.

Usage

python profile_analysis.py --analysis <input_json_profile> <default_dtype>

Arguments

  • input_json_profile: The json profile files generated by torch.profile.export_chrome_trace().
  • default_dtype: The default dtype of the model. Sometimes the dtypes of the kernel inputs are not available in the profile, so we use the default dtype to infer the dtypes of the inputs.

Diff

This mode will diff two different profiles and output a table of the differences. It groups by kernel name, which can fail to properly match across hardware vendors. More intelligent grouping coming soon.

Usage

python profile_analysis.py --diff <json_profile_1> <profile_name_1> <json_profile_2> <profile_name_2> <default_dtype> --name_limit 50

Arguments

  • json_profile_1 json_profile_2: The json profile files generated by torch.profile.export_chrome_trace().
  • profile_name_1 profile_name_2: The name of the profile. This is used to identify the profile in the output table.
  • default_dtype: The default dtype of the model. Sometimes the dtypes of the kernel inputs are not available in the profile, so we use the default dtype to infer the dtypes of the inputs.
  • name_limit: The maximum number of characters in the kernel name (they can be quite lengthy and hard to read).

Augment

This mode will add post-hoc analysis to a profile. Currently, it will add the flops and the memory reads of a kernel via formula (it's not looking at program counters or anything.) These, combined with the kernel duration, can be use to calculate achieved flops, achieved memory bandwidth, and roofline calculations.

Usage

python profile_analysis.py --augment_trace <input_json_profile> <output_json_profile> <default_dtype>

Arguments

  • input_json_profile: The json profile files generated by torch.profile.export_chrome_trace().
  • output_json_profile: Where the augmented profile is written.
  • default_dtype: The default dtype of the model. Sometimes the dtypes of the kernel inputs are not available in the profile, so we use the default dtype to infer the dtypes of the inputs.