navi14

navi14

中文版

Please test override hsa version first. Then ROCm will use gfx1030 fatbin to run on gfx1012, which needn't re-compile.

export HSA_OVERRIDE_GFX_VERSION=10.3.0

This is experimental scripts for building navi14 GPU, aka RX5500.

This is NOT offical supporting, Cannot guarantee RX5500 could run successfully on ROCm, even compiling success.

No, I didnot have a navi14 GPU yet, So I cannot test it. Currently I can just confirm there is no compiling problems. Anybody who had navi14 GPU can have a try. Appreciate for any feedback.

---

The codes based on ROCm-5.2.0, please refer https://github.com/xuhuisheng/rocm-build/blob/master/README.md for preparing build environment. OS is Ubuntu-20.04.4.

OK. One thing must clarify that building ROCm will cost lots of time, and huge memory. If your memory less then 32G, please using swap to prevent Out-Of-Memory. This caused compiling even more slower, but it wont break.

---

Now we can start our experiment.

Please follow ROCm official installation guide for installing related dependenies libraries. https://rocmdocs.amd.com/en/latest/Installation_Guide/Installation-Guide.html#ubuntu

First, install rocm-dev. This part already supports navi14, so we neednot re-compiling them. Since llvm-project may cost hours for compiling.

Then clone rocm-build. Switch to master branch. For Example, we clone rocm-build to /home/work/rocm-build

cd /home/work

git clone https://github.com/xuhuisheng/rocm-build
cd rocm-build
git checkout master

source env.sh

Modify env.sh, find AMDGPU_TARGETS, change it to AMDGPU_TARGETS="gfx1012", gfx1012 means navi14, RX5500XT. It will force ROCm to compile for navi14, even there is no matching hardware. Execute source env.sh to initialize environment variables.

The rocBLAS is a little complex, it depends Tensile. Tensile had already included in ROCm source repo. We need use a patch. For Example, ROCm source repo path is /home/work/ROCm/

Then execute bash navi14/22.rocblas.sh to compile rocBLAS, there will use a patch for prevent compiling problems. (Very slow)

Please make sure Tensile_TEST_LOCAL_PATH in navi14/22.rocblas.sh matches Tensile directory.

Other components is more simple, just execute the script to compile and install. Beside rocBLAS, We need 4 more components to re-build for navi14.

1. execute bash 21.rocfft.sh to compile rocFFT. (Extremely slow)
2. execute bash 24.rocrand.sh to compile rocRAND. (Fast)
3. execute bash navi14/25.rocsparse.sh to compile rocSPARSE, there will use a patch for prevent compiling problems. (Slow)
4. execute bash 27.rccl.sh to compile rccl. (Slow)

Final step is Pytorch-1.12.0 (Extremely Slow)

sudo ln -f -s /usr/bin/python3 /usr/bin/python

git clone https://github.com/ROCmSoftwarePlatform/pytorch
cd pytorch
git checkout release/1.12
git submodule update --init --recursive

sudo apt install -y libopencv-highgui-dev libopenblas-dev python3-dev python3-pip cmake ninja-build git
pip3 install -r requirements.txt
export PATH=/opt/rocm/bin:$PATH \
    ROCM_PATH=/opt/rocm \
    HIP_PATH=/opt/rocm/hip 
export PYTORCH_ROCM_ARCH=gfx1012
python3 tools/amd_build/build_amd.py
USE_ROCM=1 USE_NINJA=1 python3 setup.py bdist_wheel

pip3 install dist/torch-1.12.0a0+git67ece03-cp38-cp38-linux_x86_64.whl

Finally we got a pytorch-1.12.0 only can run on navi14.