Arch Linux User Repository

The package is currently broken

[trougnouf@sd darknet-hankai-git-2]$ makepkg /orb/Dev/aur/darknet-hankai-git-2/PKGBUILD: line 16: syntax error near unexpected token >>' /orb/Dev/aur/darknet-hankai-git-2/PKGBUILD: line 16:>>>>>>> fccdcf6 (Fix CUDA/ROCm detection and Qt6 dependencies)' /orb/Dev/aur/darknet-hankai-git-2/PKGBUILD: line 17: syntax error near unexpected token )' /orb/Dev/aur/darknet-hankai-git-2/PKGBUILD: line 17:)' ==> ERROR: Failed to source /orb/Dev/aur/darknet-hankai-git-2/PKGBUILD

Line 16 (in the middle of optdepends) is ">>>>>>> fccdcf6 (Fix CUDA/ROCm detection and Qt6 dependencies)"

It still fails on ROCm after removing the problematic "">>>>>>>": https://pastebin.com/raw/tsydCUiB

Could you add provides=('darknet-hankai') conflicts=('darknet-hankai')?

Thanks for making this. It compiles, but doesn't detect cuDNN. Can I specify the location somehow, like:

-DCMAKE_CUDA_COMPILER=/opt/cuda/bin/nvcc

nvidia-smi and nvcc show no issues, seems to be the makefile looking in the wrong place.

darknet reports:

Darknet V4 "Slate" v4.0-51-g53faaf9c CUDA runtime version 12080 (v12.8), driver version 12080 (v12.8) cuDNN is DISABLED => 0: NVIDIA GeForce RTX 4060 Ti [#8.9], 15.6 GiB

The ROCm issue is fixed as follow:

build() {
  cd "$srcdir/darknet/build"
  export CFLAGS="${CFLAGS// -fcf-protection/}"
  export CXXFLAGS="${CXXFLAGS// -fcf-protection/}"


cmake -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_INSTALL_PREFIX=/usr \
    -DCMAKE_HIP_COMPILER=/opt/rocm/lib/llvm/bin/clang \
    -DCMAKE_HIP_COMPILER_ROCM_ROOT=/opt/rocm -DDARKNET_TRY_CUDA=OFF -DCMAKE_HIP_ARCHITECTURES=gfx1100 \
    ..

  make -j"$(nproc)"
}

For AMD you need to disable CUDA and detect the architecture.

Something like cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/usr -DCMAKE_HIP_COMPILER=/opt/rocm/lib/llvm/bin/clang -DCMAKE_HIP_COMPILER_ROCM_ROOT=/opt/rocm -DDARKNET_TRY_CUDA=OFF -DCMAKE_HIP_ARCHITECTURES=gfx1100 .. works manually, but for some reason I get an error when trying to compile in AUR. I guess some makepkg flag. ( https://github.com/hank-ai/darknet/issues/116 )

Either way neither AMD nor NVidia work in your current package, and I would recommend doing a split-package (-cuda, -rocm, -cpu) rather than the current detection attempt.

For AMD, I was able to get the compilation to detect HIP and go much farther with -DCMAKE_HIP_COMPILER=/opt/rocm/lib/llvm/bin/clang -DCMAKE_HIP_COMPILER_ROCM_ROOT=/opt/rocm, and installing amdsmi which should be added to the optional dependencies.

Unfortunately the built stopped due to an incompatible target (?)

[ 18%] Building CXX object src-lib/CMakeFiles/darknetobjlib.dir/convolutional_layer.cpp.o
clang-18: error: unsupported CUDA gpu architecture: gfx1032
clang-18: error: unsupported CUDA gpu architecture: gfx1035
make[2]: *** [src-lib/CMakeFiles/darknetobjlib.dir/build.make:202: src-lib/CMakeFiles/darknetobjlib.dir/blas_kernels.cu.o] Error 1

No luck there. Tried on two computers (NVidia / AMD),

-- Looking for a CUDA compiler
-- Looking for a CUDA compiler - NOTFOUND
CMake Warning at CM_dependencies.cmake:52 (MESSAGE):
  Support for NVIDIA CUDA not found.
Call Stack (most recent call first):
  CMakeLists.txt:34 (INCLUDE)


-- Looking for a HIP compiler
-- Looking for a HIP compiler - NOTFOUND
CMake Warning at CM_dependencies.cmake:135 (MESSAGE):
  Support for AMD/ROCm/HIP not found.
Call Stack (most recent call first):
  CMakeLists.txt:34 (INCLUDE)


CMake Warning at CM_dependencies.cmake:146 (MESSAGE):
  Neither NVIDIA CUDA nor AMD ROCm detected.  Darknet will be CPU-only.
Call Stack (most recent call first):
  CMakeLists.txt:34 (INCLUDE)


[...]

CMake Warning:
  Manually-specified variables were not used by the project:

    ROCM_PATH

(or)

CMake Warning:
  Manually-specified variables were not used by the project:

    CUDA_TOOLKIT_ROOT_DIR

For NVidia it seems only nvcc matters (works with -DCMAKE_CUDA_COMPILER=/opt/cuda/bin/nvcc).

Can you try again? I can't test NVIDIA or AMD at the moment, but that issue should be fixed and ROCM and CPU building should also work.

Probably the same issue with ROCm

CUDA is not detected, please change the cmake line to cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/usr -DCMAKE_CUDA_COMPILER=/opt/cuda/bin/nvcc ..