OpenCL issues - RTX 4070 x64

[peardox]

With whisper-bench I'm getting...

ggml_opencl: selecting platform: 'NVIDIA CUDA'
ggml_opencl: selecting device: 'NVIDIA GeForce RTX 4070 Laptop GPU (OpenCL 3.0 CUDA)'
Unsupported GPU: NVIDIA GeForce RTX 4070 Laptop GPU

Compiled with MSVC Win11 ...

-DGGML_CUDA=1 -DGGML_OPENCL=1 -DGGML_OPENCL_USE_ADRENO_KERNELS=0

Is OpenCL only for phones or something? I've run some OpenCL tests in the past so know OpenCL works, just GGML has problems

OK, CUDA is going to be faster - I just want to run comparisons. I've also got Vulkan support compiled in and that's only slightly slower than CUDA

The Adreno flag sort hints at a possible issue I guess but can't find anything saying it WONT work.

Possibly vckpg's opencl is an issue? [it's openvino was an issue]

[ggerganov]

More info: ggml-org/llama.cpp#10693

[peardox]

Thanks - what about Kompute? Any limitations there?

[ggerganov]

The kompute backend has not been maintained for a very long time, so it's not relevant atm.

[peardox]

It would be highly beneficial if there was a public table of platform status by platform

I mean, ggml_backend_load_all in ggml-backend-reg.cpp tries this...

ggml_backend_load_best("blas", silent, dir_path);
ggml_backend_load_best("cann", silent, dir_path);
ggml_backend_load_best("cuda", silent, dir_path);
ggml_backend_load_best("hip", silent, dir_path);
ggml_backend_load_best("kompute", silent, dir_path);
ggml_backend_load_best("metal", silent, dir_path);
ggml_backend_load_best("rpc", silent, dir_path);
ggml_backend_load_best("sycl", silent, dir_path);
ggml_backend_load_best("vulkan", silent, dir_path);
ggml_backend_load_best("opencl", silent, dir_path);
ggml_backend_load_best("musa", silent, dir_path);
ggml_backend_load_best("cpu", silent, dir_path);

Plus an optional ggml_backend_load(backend_path);

Here's what I currently know

I've got blas, cuda, vulkan and rpc ATM on Windows (+ todo on Linux)
OpenCL is only for Mobile by the look of it
metal is Apple (tried Apple Silicon, optionally CoreML - x64 is an experiment-in-waiting)
Kompute = out of date
cpu = fine of course (but not GGML_CPU_ALL_VARIANTS on Apple Silicon, probably ok on Apple x64)

Not touched cann (Huawei Mobile?), hip, sycl or musa (looks old) yet - any of these good (and on what platforms?)

Raspberry Pi = cpu only (in tests)
Jetson Nano (2019 version) = too old to do CUDA - bad nvcc (CUDA 10.2 - reports CUDA17?)
Jetson Nano Orin (2025) = Can't buy one til August - should be fine

Planning AMD + CUDA tests on cheaper AWS Linux instances (once Windows/Mac via Pascal all good)