Repository navigation
Build and run instructions for CUDA devices #338
Description
This is a request to add documentation for CUDA devices. For example, it would be great to see:
- Instructions for setting up the Conda environment and a
.yamlfile for CUDA devices - Instructions for building DPBench (with support for the SYCL implementations)
- Run instructions to target CUDA devices with the benchmarks
- Which implementations of which benchmarks can we run on CUDA devices (CuPy and numba_cuda are expected, but it would be great to be able to run the SYCL implementations too)
Thanks!
Activity
ZzEeKkAa commentedon May 3, 2024
Thank you @rickybalin for the feedback!
All the cuda implementations are experimental. It should be as easy to get it work as install of all required packages from the product documentation: https://numba.readthedocs.io/en/stable/cuda/overview.html#requirements
Regarding cuda support in sycl - you need to use sycl build with cuda support. The easiest way is to use nightly build: https://github.com/intel/llvm/releases . Please be couches: if you are using nightly release from outside of conda packages, you must not have dpcpp related packages inside conda environment. It may be tricky, since
dpctl,dpnpandnumba-dpexrequires them. You would need to compile those packages using nightly build and update requirements to ignore dpcpp conda dependencies.