OpenCL support

[MikeInnes]

Tracking the issues we have getting OpenCL support.

• adapt integration: Added support for Adapt.jl JuliaGPU/CLArrays.jl#16
• Dual number support: Custom types seem pretty broken (particularly dual numbers) JuliaGPU/CLArrays.jl#17
• Integration with Flux.gpu
• GPU-compatible softmax: Generic softmax JuliaGPU/GPUArrays.jl#96
• Convolutions: Convolutions and Pooling JuliaGPU/GPUArrays.jl#102

Worth checking out ROCm and MIOpen as part of this.

[ViralBShah]

@SimonDanisch

[SimonDanisch]

on it :) most is done, just waiting for tags & tests to turn green

[SimonDanisch]

Allirght: JuliaLang/METADATA.jl#13419

[yblainm]

Hello! Does anyone know what the status of OpenCL support is?

[jpsamaroo]

CLArrays is broken due to Transpiler not being updated for Julia 1.0 (which is unlikely to happen unless some adventurous individual cares to take that on), so automatic conversion of Julia code (i.e. Flux models and functions) is not currently possible.

[dented42]

Is ArrayFire a potential alternative that could be used as an underlying representation for OpenCL support?

[MikeInnes]

Yes, supporting ArrayFire arrays may be a good idea as well. I'm not sure what they're like convolutions wise but the basics should be straightforward.

[jpsamaroo]

FYI for anyone still watching this thread, #938 will address this for AMD GPUs by skipping OpenCL and using ROCArrays instead.

[MikeInnes]

I'll close this issue since it was pretty tied to CLArrays, and that work isn't moving any more. We can track ROCm support in #938.

[hzurt]

I see this has been closed in favor of ROCm support, however ROCm only works with a few of the newest amd gpus. Is there any way to use older AMD gpus that are not supported by ROCm with flux?