GPU deployment using ROCmlink

IREE can accelerate model execution on AMD GPUs using ROCm.

Prerequisiteslink

In order to use ROCm to drive the GPU, you need to have a functional ROCm environment. It can be verified by the following steps:

rocm-smi | grep rocm

If rocm-smi does not exist, you will need to install the latest ROCm Toolkit SDK for Windows or Linux.

Get the IREE compilerlink

Download the compiler from a releaselink

Python packages are regularly published to PyPI. See the Python Bindings page for more details. The core iree-base-compiler package includes the ROCm compiler:

Stable releases Nightly releases

Stable release packages are published to PyPI.

python -m pip install iree-base-compiler

Nightly pre-releases are published on GitHub releases.

python -m pip install \
  --find-links https://iree.dev/pip-release-links.html \
  --upgrade --pre iree-base-compiler

Tip

iree-compile and other tools are installed to your python module installation path. If you pip install with the user mode, it is under ${HOME}/.local/bin, or %APPDATA%Python on Windows. You may want to include the path in your system's PATH environment variable:

export PATH=${HOME}/.local/bin:${PATH}

Build the compiler from sourcelink

Please make sure you have followed the Getting started page to build the IREE compiler, then enable the ROCm compiler target with the IREE_TARGET_BACKEND_ROCM option.

Tip

iree-compile will be built under the iree-build/tools/ directory. You may want to include this path in your system's PATH environment variable.

Get the IREE runtimelink

Next you will need to get an IREE runtime that includes the HIP HAL driver.

You can check for HIP support by looking for a matching driver and device:

$ iree-run-module --list_drivers

        cuda: NVIDIA CUDA HAL driver (via dylib)
         hip: HIP HAL driver (via dylib)
  local-sync: Local execution using a lightweight inline synchronous queue
  local-task: Local execution using the IREE multithreading task system
      vulkan: Vulkan 1.x (dynamic)

$ iree-run-module --list_devices

  hip://GPU-00000000-1111-2222-3333-444444444444
  local-sync://
  local-task://

Download the runtime from a releaselink

Python packages are regularly published to PyPI. See the Python Bindings page for more details. The core iree-base-runtime package includes the HIP HAL driver:

Stable releases Nightly releases

Stable release packages are published to PyPI.

python -m pip install iree-base-runtime

Nightly pre-releases are published on GitHub releases.

python -m pip install \
  --find-links https://iree.dev/pip-release-links.html \
  --upgrade --pre iree-base-runtime

Build the runtime from sourcelink

Please make sure you have followed the Getting started page to build IREE from source, then enable the HIP HAL driver with the IREE_HAL_DRIVER_HIP option.

Compile and run a program modellink

With the compiler and runtime ready, we can now compile programs and run them on GPUs.

Compile a programlink

The IREE compiler transforms a model into its final deployable format in many sequential steps. A model authored with Python in an ML framework should use the corresponding framework's import tool to convert into a format (i.e., MLIR) expected by the IREE compiler first.

Using MobileNet v2 as an example, you can download the SavedModel with trained weights from TensorFlow Hub and convert it using IREE's TensorFlow importer. Then run one of the following commands to compile:

iree-compile \
    --iree-hal-target-backends=rocm \
    --iree-hip-target=<...> \
    mobilenet_iree_input.mlir -o mobilenet_rocm.vmfb

Note that IREE comes with bundled bitcode files, which are used for linking certain intrinsics on AMD GPUs. These will be used automatically or if the --iree-hip-bc-dir is empty. As additional support may be needed for different chips, users can use this flag to point to an explicit directory. For example, in ROCm installations on Linux, this is often found under /opt/rocm/amdgcn/bitcode.

Canonically a HIP target (iree-hip-target) matching the LLVM AMDGPU backend of the form gfx<arch_number> is needed to compile towards each GPU chip. If no target is specified then we will default to gfx908.

Here is a table of commonly used architectures:

AMD GPU	Target Chip	Architecture Code Name
AMD MI100	`gfx908`	`cdna1`
AMD MI210	`gfx90a`	`cdna2`
AMD MI250	`gfx90a`	`cdna2`
AMD MI300X (early units)	`gfx940`	`cdna3`
AMD MI300A (early units)	`gfx941`	`cdna3`
AMD MI300A	`gfx942`	`cdna3`
AMD MI300X	`gfx942`	`cdna3`
AMD RX7900XTX	`gfx1100`	`rdna3`
AMD RX7900XT	`gfx1100`	`rdna3`
AMD RX7800XT	`gfx1101`	`rdna3`
AMD RX7700XT	`gfx1101`	`rdna3`

For a more comprehensive list of prior GPU generations, you can refer to the LLVM AMDGPU backend.

In addition to the canonical gfx<arch_number> scheme, iree-hip-target also supports two additonal schemes to make a better developer experience:

Architecture code names like cdna3 or rdna3
GPU product names like mi300x or rx7900xtx

These two schemes are translated into the canonical form under the hood. We add support for common code/product names without aiming to be exhaustive. If the ones you want are missing, please use the canonical form.

Run a compiled programlink

Run the following command:

iree-run-module \
    --device=hip \
    --module=mobilenet_rocm.vmfb \
    --function=predict \
    --input="1x224x224x3xf32=0"

The above assumes the exported function in the model is named as predict and it expects one 224x224 RGB image. We are feeding in an image with all 0 values here for brevity, see iree-run-module --help for the format to specify concrete values.