Announcement

Collapse
No announcement yet.

AMDKFD GPUVM Support Updated For Radeon Discrete GPUs

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • pete910
    replied
    Originally posted by aufkrawall View Post
    On Arch, you can simply install opencl-amd from the AUR, works with newest mainline kernel (e.g. [email protected] works here on Polaris).
    Don't work on vega cards though.

    Leave a comment:


  • bridgman
    replied
    OpenCL on ROCM is just another ROC application - it can't come much earlier than the rest of the stack unfortunately.

    On the other hand once the kernel bits are upstream the rest is already in place.
    Last edited by bridgman; 02-08-2018, 05:05 PM.

    Leave a comment:


  • phred14
    replied
    Originally posted by Peter Fodrek View Post

    AMDKFD is not about openCL support. It is HSA

    HSA in extreme simplification means, that there is no need of data transfer form CPU RAM into GPU RAM and vice versa.

    The result is speedup as of
    https://img.purch.com/w/600/aHR0cDov...NlLUhTQS5wbmc=

    form
    http://www.tomshardware.com/reviews/...cy,3899-5.html

    where execution of same code on the CPU lasts 855.4% time of HSA code execution and pure openCL code runs 643.5% of execution time for HSA code
    but it only runs on APUs at least for open source driver.

    AMDKFD enables HSA (GPU's capability of read form CPU RAM) for discrete AMD GPUs
    All true. However in the closer term, I want to just enable the "opencl" USE flag in Gentoo and see a performance boost. I agree that it won't be as good as native HSA, but simply finding much in the way of native HSA applications is not likely to happen any time soon. A better job of OpenCL-on-HSA will show more benefit, sooner. It's a bit like "WinOS2" in the old days, where the better WinOS2 applications worked, the less the pressure for native OS/2 applications. In this case perhaps the better course would be to help define future OpenCL features to be able to better take advantage of HSA, if present. Then eventually you may get HSA in every respect except name. But you'd still have a compelling hardware sales feature.

    Leave a comment:


  • aufkrawall
    replied
    On Arch, you can simply install opencl-amd from the AUR, works with newest mainline kernel (e.g. [email protected] works here on Polaris).

    Leave a comment:


  • Brisse
    replied
    Is there any way to get around this except downgrading the kernel?

    Code:
    $ sudo dkms --force install rock/1.7.60-ubuntu
    
    Kernel preparation unnecessary for this kernel.  Skipping...
    
    Running the pre_build script:
    
    Building module:
    cleaning build area...
    make -j16 KERNELRELEASE=4.15.2-041502-generic -j16 -C /lib/modules/4.15.2-041502-generic/build M=/var/lib/dkms/rock/1.7.60-ubuntu/build...(bad exit status: 2)
    ERROR (dkms apport): kernel package linux-headers-4.15.2-041502-generic is not supported
    Error! Bad return status for module build on kernel: 4.15.2-041502-generic (x86_64)
    Consult /var/lib/dkms/rock/1.7.60-ubuntu/build/make.log for more information.

    Leave a comment:


  • bridgman
    replied
    Originally posted by gsedej View Post
    What will AMDKFD enable on dGPU? my rx480 already works fine with some opencl tasks.
    This is about taking functionality that is already available out-of-tree (via ROCm stack for several GPUs and AMDGPU-PRO stack for Vega) and getting closer to making it available in upstream kernels.

    Leave a comment:


  • Brisse
    replied
    Marc Driftmeyer Cool! Might give that a try later.

    Leave a comment:


  • Peter Fodrek
    replied
    Originally posted by gsedej View Post
    What will AMDKFD enable on dGPU? my rx480 already works fine with some opencl tasks.
    AMDKFD is not about openCL support. It is HSA

    HSA in extreme simplification means, that there is no need of data transfer form CPU RAM into GPU RAM and vice versa.

    The result is speedup as of
    https://img.purch.com/w/600/aHR0cDov...NlLUhTQS5wbmc=

    form
    http://www.tomshardware.com/reviews/...cy,3899-5.html

    where execution of same code on the CPU lasts 855.4% time of HSA code execution and pure openCL code runs 643.5% of execution time for HSA code
    but it only runs on APUs at least for open source driver.

    AMDKFD enables HSA (GPU's capability of read form CPU RAM) for discrete AMD GPUs

    Leave a comment:


  • Marc Driftmeyer
    replied
    Originally posted by Brisse View Post

    Also works perfectly fine on Ubuntu 17.10 for me even though it's not officially supported

    Doesn't work on kernel 4.14 or 4.15 though, but works fine on the 4.13 kernel which 17.10 is running as default.
    I'm running Debian Sid 4-125-rc8 with AMDGPU PRO OpenCL Stack for Blender. I can tell you which drivers to install and to --force-install the dkms module and ignore its ensuing warning message which will result in it running, but one caveat: you have to be on Mesa 17.2.5 from your distro until they re-roll a new AMDGPU PRO revision for Mesa 18.

    Leave a comment:


  • Brisse
    replied
    Originally posted by ernstp View Post
    For example running a vanilla Ubuntu 16.04.3 with rocm-dkms
    Also works perfectly fine on Ubuntu 17.10 for me even though it's not officially supported

    Doesn't work on kernel 4.14 or 4.15 though, but works fine on the 4.13 kernel which 17.10 is running as default.

    Leave a comment:

Working...
X