0a494822fd
Previously, we were using ROCm 1.6.2 as there were issues with some of the machine learning applications that weren't present on 1.6.2. However, after re-running them we've found that they, and all other applications previously tested, run to completion. Additionally, there have been patches to enable BLIT kernels which made it so we no longer need to build HIP and MIOpen differently for APU and DGPU code. This allows us to install HIP directly from the .deb packages instead of from source. Installing from the .deb packages also avoid the hipDeviceSynchronize() bug. Finally, this makes it so most GPU programs can be run as-is without modifications to remove hipMalloc/hipMemcpy calls as was done previously. Change-Id: Ic61b09ed200b19f759d891487cde874abd607537 Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/37675 Reviewed-by: Bobby R. Bruce <bbruce@ucdavis.edu> Maintainer: Bobby R. Bruce <bbruce@ucdavis.edu> Tested-by: kokoro <noreply+kokoro@google.com>
gcn3-gpu dockerfile
This dockerfile contains all the dependences necessary to run GPU applications in gem5 using the gcn3 APU model
Building the image
docker build -t <image_name> .
Building gem5 using the image
The following command assumes the gem5 directory is a subdirectory of your current directory
docker run --rm -v $PWD/gem5:/gem5 -w /gem5 <image_name> scons -sQ -j$(nproc) build/GCN3_X86/gem5.opt
Test gem5 using a prebuilt application
wget http://dist.gem5.org/dist/current/test-progs/hip_sample_bins/MatrixTranspose
docker run --rm -v $PWD/MatrixTranspose:/MatrixTranspose -v $PWD/public_gem5:/gem5 -w /gem5 \
<image_name> build/GCN3_X86/gem5.opt configs/example/apu_se.py -n2 --benchmark-root=/ -cMatrixTranspose
Notes
- When using the
-vflag, the path to the input file/directory needs to be the absolute path; symlinks don't work - Currently linking in an AFS volume is not supported, as it uses ACLs instead of owner/group IDs
ToDo
- Add square to gem5-resources github, add directions for building and running an application