"GPU fault detected" for concurrency::parallel_for_each about hcc HOT 5 CLOSED

@LWisteria From dmesg it seems there's something wrong in the kernel. Could you first check if ROCm stack is working properly on your system?

In section "Verify Installation" at:
https://github.com/RadeonOpenCompute/ROCm#add-the-rocm-apt-repository

there's a "vector_copy" example which only use HSA API and doesn't involve HCC. Please check if this work or not.

from hcc.

@LWisteria I've changed your test case a bit, please check my updated version. It works fine on my system with AMD R9 Nano + Ubuntu 14.04 + ROCm 1.1 stack.

I changed:

1. use "clamp-config", not "hcc-config". for pure C++AMP codes "clamp-config" is preferred.
2. iterate devices from cbegin() to cend(), instead of crbegin() to crend()
3. I changed array_view so they are captured by copy, not captured by reference.

According to C++ AMP spec, concurrency::array_view instances shall be captured by copy, and concurrency::array instances shall be captured by reference. HCC front end doesn't check this yet so you need to be cautious on how you capture variables in your kernel.

from hcc.

@whchung

Thank you for your answer.

I overlooked https://github.com/RadeonOpenCompute/ROCm#verify-installation and I found hsa_signal_wait_acquire() doesn't return.
This means the problem is not on hcc but for ROCm run-time or etc?
Could you tell me where I should report?

And thanks for modifying my code.
I run your code. "GPU fault detected" error is erased but ./vector_add doen't return maybe on concurrency::parallel_for_each.
It seems to be the same problem as pure HSA version.

from hcc.

@LWisteria I noticed you are using an Intel i7-3770k, which is Ivy Bridge. To my knowledge to get ROCm stack working properly you need a Haswell, Skylake or later ones which support PCIe 3.0 atomics.

I would still recommend you raise the question at ROCm git repo so folks there could help you triage your issue better.

from hcc.

@whchung

Thank you for your very helpful advice. I file the problem on the ROCm repo's issue.
And I will try getting Haswell or later CPU and executing on it.

Please close this issue if you don't have any other problem.
Thanks again!

from hcc.