NVIDIA / MatX

An efficient C++17 GPU numerical computing library with Python-like syntax

Home Page:https://nvidia.github.io/MatX

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

[BUG] Compilation error with MATX_ENABLE_CUTLASS in transforms/matmul.h

bk8190 opened this issue · comments

Describe the Bug
Compilation error in MatX/include/matx/transforms/matmul.h when building with MATX_EN_CUTLASS=on

Regression seems to have be introduced by d75d702
The commit replaces "idx" with "a_idx" "b_idx" and "c_idx". The variable "idx" likely needs to be updated in the #if MATX_ENABLE_CUTLASS block starting on line 948

Hi @bk8190 , we were not aware of anyone using the cutlass interface since it's untested. We will take a look and make sure it compiles. Can I ask why you're using this over the default cublas interface?

Hi @cliffburdick - I am experimenting with the library and saw the option in the main CMakeLists.txt so I figured I would give it a try, but if it's untested I will stay clear

@bk8190 it was working several months back, but because of the compile times it was taking far too long to test. We had to put it on hold in favor of cuBLAS because of the compile times. CUTLASS does have a few more capabilities that cuBLAS did not have at the time, like non-affine indexing, but the compile times were ultimately not something we could deal with for now. If you have a reason to use it over cuBLAS please let us know.