Back to Gap Map Infrastructure / Compilers & Model Optimization

Composable Kernel

AMD
open source / Overall score: 2.5

Composable Kernel provides a programming model and template library for performance-critical machine learning kernels, decomposing GEMM, convolution and attention operations into tile-level pieces that recombine and fuse. It targets AMD GPUs through ROCm and includes a code generator for instance selection. AMD develops it in the ROCm/rocm-libraries monorepo, with the standalone ROCm/composable_kernel repository kept as a read-only mirror.

The rocm-libraries repository holds many AMD libraries, so any figure read from it measures the whole monorepo unless it is scoped to the composable_kernel path.

Openness

5 high confidence
5.0
license
MIT(OSI)
source
public
core features withheld
no

The LICENSE file contains the MIT text. The repository is public and the README describes no paid tier, enterprise edition or license-gated build, so the core is ungated.

Adoption

1 low confidence
1.0

GitHub stars are the only usable adoption signal, since Composable Kernel is vendored into ROCm libraries and built from source rather than distributed as its own package. The count comes from the now-frozen ROCm/composable_kernel repository, not the ROCm/rocm-libraries monorepo it moved into, which bundles many AMD libraries under one star count.

Capability

4 medium confidence
4.0

Composable Kernel is vendored as a git submodule by PyTorch and FlashAttention on their AMD paths. It covers one layer of the kernel stack rather than the full framework-to-target compilation path, which keeps it one step below Apache TVM.

Verified 2026-09-27