https://bugs.freedesktop.org/show_bug.cgi?id=104604

Bug ID: 104604 Summary: [amdgpu/radeon][regression, CIK] Prefetch the compute shader to TC L2 (4a4ff66dbe) causes GPU VM errors when running OpenCL kernels on Hawaii Product: Mesa Version: 17.1 Hardware: All OS: All Status: NEW Severity: critical Priority: medium Component: Drivers/Gallium/radeonsi Assignee: dri-devel@lists.freedesktop.org Reporter: vedran@miletic.net QA Contact: dri-devel@lists.freedesktop.org

Commit 4a4ff66dbe radeonsi: also prefetch compute shaders enabled prefetching the compute shader to TC L2 for CIK+. This causes GPU VM errors on Hawaii, Kabini, and likely other CIK GPUs when running certain OpenCL programs (it seems that the key is to call clCreateContext() twice or more, see below) with Clover, both with amdgpu and radeon kernel drivers. I have tested with

01:00.0 VGA compatible controller [0300]: Advanced Micro Devices, Inc. [AMD/ATI] Hawaii XT GL [FirePro W9100] [1002:67a0]

on Fedora rawhide (28) with kernel 4.15.0-0.rc7.git2.1.fc28.x86_64.

I'll minimize the example program that causes this (right now using GROMACS tests, but that's unnecessarily complex), and provide the details (grepped + full dmesg) in the comments for both amdgpu and radeon kernel drivers.