[LLVMbugs] [Bug 20738] New: Radeon: code generation for GCN GPUs seems to be broken. Clpeak benchmark outputs many errors ("Can't spill VGPR!") and then GPU hangs.

bugzilla-daemon at llvm.org bugzilla-daemon at llvm.org
Sat Aug 23 17:25:02 PDT 2014


http://llvm.org/bugs/show_bug.cgi?id=20738

            Bug ID: 20738
           Summary: Radeon: code generation for GCN GPUs seems to be
                    broken. Clpeak benchmark outputs many errors ("Can't
                    spill VGPR!") and then GPU hangs.
           Product: clang
           Version: trunk
          Hardware: Other
                OS: Linux
            Status: NEW
          Severity: release blocker
          Priority: P
         Component: -New Bugs
          Assignee: unassignedclangbugs at nondot.org
          Reporter: t3st at sogetthis.com
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

Created attachment 12931
  --> http://llvm.org/bugs/attachment.cgi?id=12931&action=edit
LLVM error log.

Configuration: 
 AMD R9 270 GPU. 
 Linux kernel 3.16
 LLVM 3.15-rc3

Problem:
 Code generation for AMD GPUs seems to be broken in LLVM. At very least for GCN
based GPUs.
 When trying to run certain opencl kernels, LLVMS seems to face numerous
troubles and then probably generates bugged code which locks up GPU.

To reproduce:
1) Get "clpeak" tool at https://github.com/krrishnarraj/clpeak.git
2) Build it. 
3) Try to use it with AMD GCN-based GPU. I used R9 270 for this test.
4) Take a look what happens.

Test result: 
1) LLVM outputs about 40Kb of error messages (log attached).
2) After some time of running OpenCL kernels GPU locks up.

Notes: 
1) I was never able to complete single run of clpeak benchmark on opencl GPUs
due to GPU lockups. Bug seems to be dead simple to reproduce.
2) Could also be some kernel or lib issue, let's AMD staff to classify it
correctly.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20140824/47bed256/attachment.html>


More information about the llvm-bugs mailing list