last edited: 2023-08-21 17:31:36 +0000


Table of Contents

  1. Using the model
  2. ROCm
  3. Documentation and Tutorials

The GCN3 GPU is a model that simulates a GPU at the ISA level, as opposed to the intermediate language level. This page will give you a general overview of how to use this model, the software stack the model uses, and provide resources that detail the model and how it is implemented.

Using the model

Currently, the GCN3 GPU model in gem5 is supported on the stable and develop branch.

The gem5 repository comes with a dockerfile located in util/dockerfiles/gcn-gpu/. This dockerfile contains the drivers and libraries needed to run the GPU model. A pre-built version of the docker image is hosted at

The gem5-resources repository also comes with a number of sample applications that can be used to verify that the model runs correctly. We recommend users start with square, as it is a simple, heavily tested application that should run relatively quickly.

Using the image

The docker image can either be built or pulled from

To build the docker image from source:

# Working directory: gem5/util/dockerfiles/gcn-gpu
docker build -t <image_name> .

To pull the pre-built docker image (Note the v22-0 tag, to get the correct image for this release):

docker pull

You can also put as the image in the docker run command without pulling beforehand and it will be pulled automatically.

Building gem5 using the image

See square in gem5 resources for an example of how to build gem5 in the docker. Note: these directions assume you are pulling the latest image automatically.

Building & running a GPU application using the image

See gem5 resources for examples of how to build and run GPU applications in the docker.


The GCN3 model was designed with enough fidelity to not require an emulated runtime. Instead, the GCN3 model uses the Radeon Open Compute platform (ROCm). ROCm is an open platform from AMD that implements Heterogeneous Systems Architecture (HSA) principles. More information about the HSA standard can be found on the HSA Foundation’s website. More information about ROCm can be found on the ROCm website

Simulation support for ROCm

The model currently only works with system-call emulation (SE) mode, therefore all kernel level driver functionality is modeled entirely within the SE mode layer of gem5. In particular, the emulated GPU driver supports the necessary ioctl() commands it receives from the userspace code. The source for the emulated GPU driver can be found in:

The HSA driver code models the basic functionality for an HSA agent, which is any device that can be targeted by the HSA runtime and accepts Architected Query Language (AQL) packets. AQL packets are a standard format for all HSA agents, and are used primarily to initiate kernel launches on the GPU. The base HSADriver class holds a pointer to the HSA packet processor for the device, and defines the interface for any HSA device. An HSA agent does not have to be a GPU, it could be a generic accelerator, CPU, NIC, etc.

The GPUComputeDriver derives from HSADriver and is a device-specific implementation of an HSADriver. It provides the implementation for GPU-specific ioctl() calls.

The src/dev/hsa/kfd_ioctl.h header must match the kfd_ioctl.h header that comes with ROCt. The emulated driver relies on that file to interpret the ioctl() codes the thunk uses.

ROCm toolchain and software stack

The GCN3 model supports ROCm version 4.0.

The following ROCm components are required:

The following additional components are used to build and run machine learning programs:

For information about installing these components locally, the commands in the GCN3 dockerfile (util/dockerfiles/gcn-gpu/) can be followed on an Ubuntu 16 machine.

Documentation and Tutorials

GCN3 Model

Describes the GCN3 model

gem5 GCN3 ISCA tutorial

Covers information about the GPU architecture, GCN3 ISA and HW-SW interfaces in gem5. Also provides an introduction to ROCm.


ROCm Documentation

Contains further documentation about the ROCm stack, as well as programming guides for using ROCm.

AMDGPU LLVM Information