Clpeak compute-sp freezes with mailbox

[doe300]

When executed via the mailbox-interface, the clpeak --compute-sp benchmark freezes (the kernel?) in the second execution. The --global-bandwidth benchmark runs fine, so do both benchmarks when run via register-poking.

More exact error description:

• according to htop, the VC4CL Queue Handler thread moves into status D, which means uninterruptible sleep (usually IO) (see here), which looks like it hangs in the mailbox-syscall
• forcibly stopping the process is unable to kill the VC4CL Queue Handler thread, even sudo reboot does nothing.

Side note: When this error is fixed, re-examine the delay in src/executor.cpp:108, since at least the --global-bandwidth benchmark seems not to require it (anymore).

[ThijsWithaar]

Maybe related: since two weeks or so, clinfo freezes after: Max work group size 12 This is with raspbian stretch, so gcc 6.3.0 and clang 3.9 installed. It used to work.

[doe300]

@ThijsWithaar, do you use clinfo from the repository or a manually compiled version? The version in the repository has an error, because of which I started using the github version. I just currently don't remember what the error was.

[ThijsWithaar]

I use the one from raspbian (stretch)

[doe300]

On my Raspberry Pi, the clinfo from the repository fails (before any output) with

terminate called after throwing an instance of unsigned int

The manually compiled version runs fine, though I am not sure, this is related.