search

NVIDIA / cutlass

CUDA Templates for Linear Algebra Subroutines
Other
5.49k stars 930 forks source link

[QST] Integrating Cutlass EVT to MLIR-Based compiler problems #1807

Open Xinyu302 opened 1 month ago

Xinyu302 commented 1 month ago

Hi! I am currently integrating Cutlass EVT into an MLIR-based deep learning compiler to address arbitrary epilogue fusion issues. The deep learning compiler uses stablehlo as the frontend. I am attempting to translate stablehlo ops into a Python Ast frontend, integrating the Cutlass-generated kernel as a custom call into the compiler. I have encountered the following issues during this process and would like to hear some suggestions:

  1. In stablehlo, activation functions need to be decomposed into basic operator representations, so the ReLU activation function becomes stablehlo.max. However, the operations supported by the EVT python ast frontend are limited; to my knowledge, it only supports addition, subtraction, multiplication, division, and some activation functions. Do you have any good solutions? Do you have a plan to extend the python ast frontend?
  2. Constructing EVTD::arguments is quite challenging. I am currently deducing the construction method for EVTD::arguments by parsing ctypes.structure in Python. Do you have any good solutions?
github-actions[bot] commented 6 hours ago

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.