cubecl-hip 0.9.0

AMD ROCm HIP runtime for CubeCL
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

# ROCm HIP runtime

Runtime that runs on ROCm HIP supported AMD GPUs.

Matrix multiplication acceleration is based on rdna3 by default and requires [rocwmma][] for other architectures.
Note that kernel compilation time with [rocwmma][] might be slow.

[rocwmma]: https://github.com/ROCm/rocWMMA
[rdna3]: https://gpuopen.com/learn/wmma_on_rdna3/

## RDNA 2

RDNA 2 should work with cubecl-hip, however due RDNA2 not officially being supported by AMD, errors might be encountered. In order for an RDNA 2 GPU to work, the environment variable `HSA_OVERRIDE_GFX_VERSION` must be set to `10.3.0`.

It is known that some errors might occur during initial startup with an RDNA 2 GPU, but if the programs gets through startup then no further errors should be encountered. Re-run the program until it works. This is mostly due to a limitation of the support from AMD.