Michael Larabel writes via Phoronix: AMD has been working for years to port codebases targeting NVIDIA's CUDA API to make it easier to run on HIP/ROCm, but it still requires work on the part of developers. Tools such as his HIPIFY that assist with automatic generation are improving, but they are not a simple and instant guaranteed solution, especially when aiming for optimal performance. For the past two years, AMD has been quietly funding efforts to achieve binary compatibility. This allows many NVIDIA CUDA applications to run at the library level on the AMD ROCm stack. This can be replaced with a drop-in without the need to adjust the source code. In fact, for many real-world workloads, this is a solution for an end user to run her CUDA-enabled software without developer intervention. Here I'll explain more about this “skunkworks” project, which is open sourced, and some of my own tests and performance benchmarks of this his CUDA implementation built for Radeon GPUs. […]
For those wondering about open source code, it is dual licensed with either Apache 2.0 or MIT. Rust fans will be excited to know that this Radeon implementation leverages the Rust programming language. […] If you want to check out the new ZLUDA open source code for Radeon GPUs, you can do so via GitHub.
