Rearchitect and Reinstate GPU

[onurulgen]

• Reinstated GPU into the project by fixing or implementing missing functions.
• Rearchitected the project to combine GPU and CPU implementations by using virtual functions.
• Increased the performance up to 11 times by using GPU.
• Implemented unit and regression tests for GPU and CPU by using Catch2 framework.
• Developed GitHub Actions for tests, coverage, static code analysis, and producing CPU and GPU executables for Windows, Linux and macOS.

To fix the issue, we will add a permissions block at the root level of the workflow. This block will restrict permissions to contents: read, which is sufficient for most workflows that do not require write access. If the workflow requires additional permissions in the future, they can be added explicitly and scoped to specific jobs.

The permissions block will be added immediately after the name field (line 1) to apply to all jobs in the workflow.

---

To fix the issue, we will add a permissions block at the root level of the workflow file. This block will specify the minimum permissions required for the workflow to function correctly. Based on the current workflow steps, the workflow only needs to read repository contents and upload coverage reports to Codecov. Therefore, we will set contents: read and id-token: write (required for the Codecov action). This ensures that no unnecessary write permissions are granted.

---

To fix the issue, we will add a permissions block at the root of the workflow file. This block will specify the minimal permissions required for the workflow to function. Since the workflow only checks out the repository and runs tests, it likely only needs contents: read permissions. This ensures that the workflow cannot perform any unintended write operations.

The permissions block will be added at the top level of the workflow, just below the name field, so it applies to all jobs in the workflow.

---

[kaczmarj]

hi @onurulgen - thank you for your tremendous effort in getting CUDA support back into niftyreg for a variety of realignment methods.

is this branch OK to try out? i am currently using version 1.3.9 from 2012 for CUDA-enabled reg_f3d.

[onurulgen]

hi @onurulgen - thank you for your tremendous effort in getting CUDA support back into niftyreg for a variety of realignment methods.

is this branch OK to try out? i am currently using version 1.3.9 from 2012 for CUDA-enabled reg_f3d.

Yes, it's stable, you may go ahead and test it. I'd like to hear how it goes as well. Could you please give some feedback if you can?

To fix the problem, add a permissions section to the workflow. This can be done either globally at the root level, or per job (here, just the build job). For maximal clarity and future-proofing, it's best to add at the root, unless different jobs require different permissions. The workflow uploads release assets but does not modify repository contents; thus, the minimal required permissions should be contents: read, which allows reading release info but not writing to the repository. The upload-release-action uses repo_token—this is only for uploading assets to a release, which needs contents: read for release info, and requires contents: write to upload assets. To function correctly, contents: write may be needed; but to adhere to least privilege, we can set contents: write. So, add a block:

permissions:
  contents: write

directly under the name: Release line (before on:).

[kaczmarj]

@onurulgen - niftyreg v2.0.0 is working great so far. thanks again for all of your work.

i compiled it in a docker container. i copied the dockerfile below if anyone is interested. i got errors with CUDA 12.x so i stuck with 11.8

FROM nvcr.io/nvidia/cuda:11.8.0-devel-ubuntu22.04 AS builder

ARG DEBIAN_FRONTEND=noninteractive
RUN apt-get update && apt-get install -y \
        build-essential \
        cmake \
        git \
        ninja-build \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /work
RUN git clone https://github.com/KCL-BMEIS/niftyreg . \
    && git checkout v2.0.0

RUN cmake -S . -B build -G "Ninja" \
        -DCMAKE_BUILD_TYPE=Release \
        -DBUILD_ALL_DEP=ON \
        -DUSE_CUDA=ON \
        -DCHECK_GPU=OFF \
        -DUSE_OPENMP=ON \
        -DUSE_SSE=ON \
        -DCMAKE_INSTALL_PREFIX=/opt/niftyreg \
        -DCMAKE_INSTALL_RPATH="\$ORIGIN/../lib" \
        # Support many different NVIDIA GPUs.
        -DCMAKE_CUDA_ARCHITECTURES="75-real;80-real;86-real;89-real;90" \
    && cmake --build build --parallel \
    && cmake --install build

FROM nvcr.io/nvidia/cuda:11.8.0-runtime-ubuntu22.04
COPY --from=builder /opt/niftyreg /opt/niftyreg
ENV PATH="/opt/niftyreg/bin:${PATH}"
WORKDIR /work

i built with this (i'm on an arm-based macbook)

docker buildx build --platform=linux/amd64 --load -t kaczmarj/niftyreg .

then I got the binaries out with this:

docker run --rm \
  -v "$(pwd)":/out \
  kaczmarj/niftyreg \
  bash -c "cd /opt && tar -czf /out/niftyreg.tar.gz niftyreg"

[onurulgen]

Thank you! You may use precompiled binaries; and if you'd like to compile yourself, you may follow release.yml file. The code is compatible up to CUDA 12.4.x. It might've failed if you tried later versions.

[kaczmarj]

fantastic, i will use the precompiled binaries in that case. thanks kindly