GemmKernels

Flexible and performant GEMM kernels in Julia

Julia	CI
1.6
1.7
nightly

This package contains a framework to instantiate flexible, performant GEMM (General Matrix Multiplication) kernels.

It decomposes GEMM kernels into orthogonal components:

• Params determine the tiling size and launch configuration of the GEMM kernel. The tiling sizes are specified in logical coordinates, i.e. with a meaning specified by the user.
• Layouts convert the logical coordinates of tiles to physical offsets in memory.
• Transforms are used to apply any arbitrary Julia functor to the GEMM's inputs or outputs. They are applied after every load, and before every store.
• Operators are responsible to perform the matrix multiplication itself. They load tiles from shared memory, perform the matrix multiplication, and store the resultant tile back to shared memory.
• Epilogues copy tiles of the resultant matrix to global memory, and can be used to implement arbitrary post-processing, such as adding a bias vector to the resultant matrix.

Each of these components corresponds to a set of functions with a predetermined interface. These functions can be customised by the user through Julia's multiple dispatch functionality.

The package includes 2 user-facing interfaces: a fully-featured interface (see e.g. test/matmul.jl) and a BLAS-like interface (see e.g. test/blas.jl).

The documentation is still a WIP, but you can get an idea of the usage of this package using the examples in test/ and benchmark/.

Quick Start

The package can be installed using Julia's built-in package manager. Open the Julia REPL, type ] to enter Pkg-mode, and run:

pkg> add GemmKernels

Note that you need a sufficiently recent NVIDIA GPU (Volta or later) that contains Tensor Cores to use this package.

Project Status

At the moment, the package only contains GEMM kernels for CUDA-enabled NVIDIA GPUs and targets Tensor Cores exclusively.

It contains the necessary components for mixed-precision GEMMs using WMMA, GEMMs exploiting operation fusion with elementwise operations or bias vectors, diagonal matrices, and matrices of complex/dual numbers.

Performance

The above figure shows the performance of a mixed-precision multiplication of two FP16 matrices, resulting in an FP32 resultant matrix, for different memory layouts. We compare our kernels with the state-of-the-art libraries cuBLAS and CUTLASS on an RTX 2080 Ti.

Citation

For more details on the implementation and performance results, please see our accompanying paper (pre-print available on arXiv). The CITATION.bib file in the root of this repository contains a citation in BibTeX format.