GPUVM Discrete GPU Code For AMDKFD, Radeon Compute Could Be Ready For Linux 4.17

[phoronix]

Phoronix: GPUVM Discrete GPU Code For AMDKFD, Radeon Compute Could Be Ready For Linux 4.17

Sent out Friday night were the latest patches for getting the discrete GPU support within the AMDKFD HSA kernel driver up to scratch for allowing the ROCm compute stack working off a mainline kernel...

http://www.phoronix.com/scan.php?page=news_item&px=AMDKFD-dGPU-GPUVM

[Jumbotron]

So...I would guess sometime in 2019 4.17 will be out ( stable that is ). 8 years. 8 LONG years of waiting for the dream AMD set forth in 2011 with the release of the AMD Fusion APUs and the larger idea of Heterogeneous Computing.

Better late than never. Now if we can only get OpenCL up to snuff and better supported.

[tiwake]

My life will be complete when this happens.

Complete top-down open source... gentoo anyone?

[R41N3R]

Finally all the small puzzles come together... Linux 4.17 will be an awesome release then :-)

[eydee]

Here's a list of cards you can purchase to use with this:

[Brisse]

Here's a list of cards you can purchase to use with this:

Hilarious list.

Good thing I already own a Fiji, and I am actually rocking ROCm with good results. Still need that out of tree kernel module though, but not really a big deal for me. It's already easy to get up and running. This will just make easy even easier.

[Beta Version]

So, with 4.17 mining will work even on open-source AMD drivers? Or not?

[Brisse]

So, with 4.17 mining will work even on open-source AMD drivers? Or not?

It already works and you don't need 4.17 for that. I'm running ethminer as we speak on Ubuntu 17.10 with ROCm installed alongside the default kernel and mesa drivers. ROCm currently comes with a separate out of tree kernel module to make it work. That kernel module will be integrated into 4.17 so there's no longer any need to install it separately like you have to right now.

[Jumbotron]

Here's a list of cards you can purchase to use with this:

Annnnnnnd.....we're STILL waiting.

[bridgman]

So...I would guess sometime in 2019 4.17 will be out ( stable that is ). 8 years. 8 LONG years of waiting for the dream AMD set forth in 2011 with the release of the AMD Fusion APUs and the larger idea of Heterogeneous Computing.

In fairness, KFD code for APUs has been upstream for a couple of years already, although we didn't make it easy to find matching userspace code.

Upstreaming DGPU support (for older AMD CPUs and all Intel CPUs) took longer because we did not have the recoverable page fault support from IOMMUv2 that we could rely on with our APUs, and had to come up with an upstreamable alternative by allowing the controlled eviction of pages which ROC userspace believed to be pinned along with temporarily disabling the affected userspace processes.

Early ROC releases did not include this eviction logic and hard-pinned buffers from userspace, which was fine for compute workloads but did not play sufficiently well with arbitrary graphics workloads to be upstreamable.

Original RFC for eviction with comments is here...

https://lists.freedesktop.org/archiv...ch/078573.html

... and current implementation is in patches 6 & 16 from this set, among others.

Just curious, how do you get to 2019 ? The 4.17 kernel should be released mid-2018 if I'm doing the math right, and I don't think it takes 6 months for stable versions to start.