mat-mat-ops

Custom C++ and CUDA operators for Matrix-Matrix Operations in PyTorch. Here I implemented Shared Memory Cache-Blocking and Block-tiling for both forward and backward kernels.

If you want to know how to write your own custom kernel, this PyTorch offical tutorial is all you need :)

Requirements:

CUDA Toolkit 12.4 PyTorch 2.4+

Supported operations so far:

• Mat-Mat Mul
• Mat-Mat L1

To build:

pip install .

To test:

the interactive option : test/test.ipynb or

python test/test_extension.py

Author

Mehdi Moshtaghi