Archives Discussions

[r_potter]

Hi,

I'm trying to get our HSA code running on a Haswell-E/Fury X machine using the ROCK kernel driver. The machine boots correctly using an upstream 4.4 kernel and amdgpu driver. It will also boot with the ROCK kernel and a R9 390 installed. Unfortunately, when I try to boot the machine with the combination of the ROCK kernel and the Fury X installed it deadlocks during boot. The final line of console output is: "fb: switching to amdgpudrmfb from EFI VGA"

Any hints on whats causing it to hang would be greatly appreciated.

Ralph

[bridgman]

Hi Ralph,

Would you be able to attach your entire dmesg output  ("dmesg >x", attach x) ? Might have a few more clues.

Thanks

[r_potter]

Hi,

Log attached. It looks like things go wrong around line 1180.

[bridgman]

Thanks. Looks like it's blowing up in the early powerplay code, will ask someone who knows that code to take a look in the morning.

[r_potter]

Great. Thanks

[gstoner]

Ralph,   there was a drop March 4th that should have addressed your issue.   We have another drop this comming week up on github.  

ROCm  supports the following capabilities

• Compiler/Language Runtime
  • HCC Compiler and Language Runtime  https://gpuopen.com/compute-product/hcc-heterogeneous-compute-compiler
    • Single Source Compiler supports
    • C++ 11/14/17 Support with Parallel STL
    • OpenMP 3.1 and 4.0  C and C++ support for CPU
  • HCC is built on  CLANG and LLVM compiler foundation with Native GCN ISA code generation
  • Static Offline compiler which is separate from the driver so we can be more aggressive with optimizations
  • HIP Runtime https://gpuopen.com/compute-product/hip-convert-cuda%E2%80%A6-portable-c-code/ and HIPIFY which are tool we can use to port CUDA application over our hardware more easily.
  • Continuum  IO Anaconda with NUMBA support for Accelerated Python
  • Math Libraries
    • HcBLAS https://bitbucket.org/multicoreware/hcblas
    • HcFFT https://bitbucket.org/multicoreware/hcfft
    • HcSPARSE https://bitbucket.org/multicoreware/hcsparse
    • HcRNG  https://bitbucket.org/multicoreware/hcrng
• ROCR Runtime ( which based on the HSA runtime API  http://www.hsafoundation.com/html/HSA_Library.htm enhanced for the need for dGPU’s)  for developing tools and getting the lowest level access to the hardware you need it.
  • User Mode Queues  
  • User Mode DMA
  • Advanced Memory Management API supporting need of Multi-GPU systems
  • Coarse-grain Shared Virtual Memory
  • Peer to Peer Multi-GPU supports with Large Bar support
  • GPU enabled Peer to Peer with RDMA support for Mellanox cards
  • Process Concurrency & Preemption
  • HSA Signals and Atomics (Using PCIe Gen3 Platform Atomics) for System level signaling and synchronization
  • Standardized loader and Code Object Format  for Binary GCN ISA   ( ELF & DWARF support)
• The stack is Fully Open Source ussing   MIT and UIUC Source Code licences 

To run this we need:  * Haswell CPU or newer CPU, like Xeon E5 v3 * Fiji GPU R9 Fury Nano, R9 Fury X, R9 Fury * Ubuntu 14.04 or Fedora 23  You can find more information at www.gpuopen.com and at Github.com at the links

https://github.com/RadeonOpenCompute/ROCR-Runtime/tree/dev

https://github.com/RadeonOpenCompute/ROCK-Kernel-Driver/tree/dev

You can also find more info on GPUopen.com