Add supports for AMD GPU #6161
Add supports for AMD GPU #6161
Conversation
…is based on Dockerfile for NVIDIA GPU.
1. replace __shfl_down_sync(0xffffffff, ...) with __shfl_down(...)
__shfl_xor_sync(0xffffffff,...) with __shfl_xor(...)
for AMD ROCm HIP
2. replace dim3 integer vector initialization { , , } with dim3( , , )
for AMD ROCm HIP
for more information, see https://pre-commit.ci
|
thanks for the PR, we don't currently have the ci/cd resources to verify/test the changes, is there any ci runners we can use to test and for future maintenance? |
|
Thanks for the quick response. Currently we are considering different options to run the ci/cd on AMD hardware. We would like to discuss with you at your convenience, so that we can setup ci to fit current ci pipeline of MONAI project better. |
HI @wyli Would it be possible for us to setup a call to discuss how we can integrate this PR ? We would like to understand how we can help in setting up the CI/CD resources for validating the changes. Thank you. |
|
/black
sure @yaomingamd @japarada, let's have a call, please send your contact details to wenqil@nvidia.com I'll make a meeting time poll and invite colleagues. |
|
As we discussed let's remove the Dockerfile from this PR for now since we'll refactor the existing file to allow for selection of image by variable which will let us choose AMD images. The changes otherwise are fine since they pass our tests and we'll have to assume for now our notional AMD tests will pass also. One thing that does have to be done is to sign your commits to get DCO to pass, clicking on details will explain how to do a remedial commit for that. Thanks! |
I, Yaoming Mu <yaoming.mu@amd.com>, hereby add my Signed-off-by to this commit: 018609c I, Yaoming Mu <yaoming.mu@amd.com>, hereby add my Signed-off-by to this commit: a8cfcf6 As discussed with MONAI team, moves DOckerfile.amd from this PR. The existing file "Dockerfile" will be refactored to allow build images for different GPUs. Signed-off-by: Yaoming Mu <yaoming.mu@amd.com>
…duce the amount of codes. Currently AMD HIP does not support __shfl_xxx_sync functions as NVIDIA CUDA, but when mask is 0xffffffff, __shfl_xxx_sync would be replaced by __shfl_xxx for the codes in gmm_cuda.cu run unittest tests/test_gmm.py for NVIDIA and AMD GPUs and tests passed. Signed-off-by: Yaoming Mu <yaoming.mu@amd.com>
Signed-off-by: monai-bot <monai.miccai2019@gmail.com>
|
/build |
Fixes # .
Description
Add supports for AMD GPU:
1. add docker file "Dockerfile.amd " for creating docker image to support AMD GPU. The file is based on "Dockerfile" for NVIDIA GPU2. In monai/_extensions/gmm/gmm_cuda.cu , replaces __shfl_down_sync and __shfl_xor_sync with __shfl_down and __shfl_xor when build for AMD GPU.
Types of changes
./runtests.sh -f -u --net --coverage../runtests.sh --quick --unittests --disttests.make htmlcommand in thedocs/folder.