Orochi VS ROCm-OpenCL-Runtime

While you're waiting for the HIP SDK to release, check out Orochi as an alternative https://github.com/GPUOpen-LibrariesAndSDKs/Orochi

You can't run CUDA binaries directly. But you can use a wrapper library like Orochi to run both CUDA and HIP using a single binary that dynamically links with CUDA/HIP libraries at runtime https://github.com/GPUOpen-LibrariesAndSDKs/Orochi

I stumbled across orochi from AMD while looking for their FSR2.0 implementation, which basically boils down to being a wrapper over Cuda and HIP. I don't know if it is still maintained or functional, but heres the link if anyone is interested: https://github.com/GPUOpen-LibrariesAndSDKs/Orochi

There's https://docs.amd.com/bundle/ROCm-Installation_FAQ/page/Frequently_Asked_Questions.html which leads to a page which doesn't list any gpus that I can see, there's https://rocm.docs.amd.com/en/latest/release/gpu_os_support.html which lists "RDNA2" , there's https://github.com/RadeonOpenCompute/ROCm/issues/1698 which is from last year and mentions changing an env for the RX 6600 XT (navi 23) . Not a lot is mentioned in the readme of https://github.com/RadeonOpenCompute/ROCm-OpenCL-Runtime .

$ dnf info rocm-opencl Installed Packages Name : rocm-opencl Version : 5.4.3 Release : 2.fc38 Architecture : x86_64 Size : 1.7 M Source : rocm-opencl-5.4.3-2.fc38.src.rpm Repository : @System From repo : updates Summary : ROCm OpenCL Runtime URL : https://github.com/RadeonOpenCompute/ROCm-OpenCL-Runtime License : MIT Description : ROCm OpenCL language runtime. : Supports offline and in-process/in-memory compilation.

Eg this bug about shared cl/gl textures with mipmaps being broken has now breached its first birthday, without even acknowledgement - basic cl/gl functionality here. This bug took a year for a fix to make its way into a public driver. And this fairly performance critical bug is just "wontfix", and also a significant downgrade from their old driver stack

Its definitely workable if you're willing to put in the effort (except for things that are straight up broken in some cases, like device side enqueue), but there are some issues that require.. fairly major workarounds

Their implementation is here: https://github.com/RadeonOpenCompute/ROCm-OpenCL-Runtime

After a lot of moderately annoyed testing, I discovered that the AMD OpenCL implementation is.. rather dumb. If any two kernels share any arguments arguments, it inserts a command barrier between the two, hard-stalling the GPU. After filing a bug, it turns out this is wontfix as well, which is doubly bad. There's no set of flags in OpenCL that you can use to fix this either

What's particularly bizarre, is that with one bug report I filed, they claim to have fixed it internally in april, but.. no public driver has ever been released with the fix. For 7 months? Which is just a bizarre software development process

Could you report these issues here https://github.com/RadeonOpenCompute/ROCm-OpenCL-Runtime? Regarding the device side enqueue issue, could you attach a simple test case to the issue that reproduces the crash? The current pastebin link doesn't give enough info.