[PATCH 00/24] KFD fixes, cleanups, features to make the runtime happy

[Felix Kuehling]

Update on KV testing with upstream kernel code:

I was able to make OpenCL work with a simple change. I compiled it from
the open-source version available on GitHub
(https://github.com/RadeonOpenCompute/ROCm-OpenCL-Runtime) and applied
the change below.

I'm running the SHOC benchmark right now, and it's looking good so far.
I can easily get it to hang with OpenCL conformance tests, so it's not
entirely stable. But this is progress and I think a reasonable baseline
for validating further KFD changes. Most importantly, it's a real
open-source user mode GPU compute stack using KFD, now running on an
upstream kernel on both CZ and KV.

Regards,
  Felix

commit 2761ff8bb444cf93de4d3d182b38d1da140df986
Author: Felix Kuehling <Felix.Kuehling at amd.com>
Date:   Fri Sep 8 16:14:23 2017 -0400

    Enable Kaveri
    
    Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>

diff --git a/runtime/device/rocm/rocdevice.cpp b/runtime/device/rocm/rocdevice.cpp
index 16db78a..edd3cbd 100644
--- a/runtime/device/rocm/rocdevice.cpp
+++ b/runtime/device/rocm/rocdevice.cpp
@@ -76,6 +76,8 @@ static HsaDeviceId getHsaDeviceId(hsa_agent_t device, uint32_t& pci_id) {
   }
 
   switch (gfxipVersion) {
+    case 700:
+      return HSA_SPECTRE_ID;
     case 701:
       return HSA_HAWAII_ID;
     case 801:


On 2017-08-15 10:59 PM, Felix Kuehling wrote:
> On KV, the current ROCm OpenCL driver doesn't work. This is a
> limitation of the OpenCL driver. I'm trying to find out how hard it
> would be to change that. In the mean time, some hsaconformance tests
> can be run on KV. It passes most tests up to
> 162:code_recursive_kernel_function, where it hangs.