4 ms·
Well the implementation is generic to the array type, so if someone creates an array type which builds broadcast kernels on with OpenCL then it would work with
by ChrisRackauckas 7y ago
Well the implementation is generic to the array type, so if someone creates an array type which builds broadcast kernels on with OpenCL then it would work with OpenCL. Sadly, CLArrays.jl never got updated to Julia v1.0, so until someone does that it's just a dream. However, DifferentialEquations.jl will be waiting ready for that to happen :).
- dnautics 7y agoI'm honestly surprised that amd doesn't get more aggressive about developing towards Julia as a deep learning target, it would be way easier for their devs to optimize for performance parity than say tensorflows spaghetti code. Iirc, the raw flops performance of AMD gpus is even or better than Nvidia but their unoptimized libraries kick it in the shins (I could be misremembering thus)
- ViralBShah 7y agoThere are folks working on the AMDgpu target for Julia, and I see that some early work has already happened in terms of building LLVM and such. I suspect the AMD toolchains are not as strong - but with the right contributors, it should be possible to leverage the work done on the Nvidia GPUs for the AMD GPUs and even other accelerators.
- celrod 7y agoAs someone who owns a couple of Vega (AMD) GPUs, I'm watching a few of jpsamaroo's GitHub repos. In particular: https://github.com/jpsamaroo/HSAArrays.jl https://github.com/jpsamaroo/HSAArrays.jl I don't know how far along it is (no documentation yet), but those are definite steps.
- microcolonel 7y agoI wonder how much trouble it would be to bring CuArrays to HIP.