%global pkgvers 0
%global scdate0 20240411
%global schash0 0fd25d6c70629d978fdc7dec80e5299f0164374e
%global branch0 master
%global source0 https://github.com/Tencent/ncnn.git

%global sshort0 %{expand:%%{lua:print(('%{schash0}'):sub(1,8))}}

Name:           ncnn
Version:        20240410
Release:        %{scdate0}.%{pkgvers}.git%{sshort0}%{?dist}
Summary:        Neural network inference computing framework optimized for mobile
License:        BSD

URL:            https://github.com/Tencent/ncnn

BuildRequires:  git doxygen cmake ocl-icd-devel
BuildRequires:  python3-setuptools python3-devel
BuildRequires:  opencv-devel protobuf-devel pybind11-devel
%if 0%{?fedora}
BuildRequires:  glslang-devel gcc-c++
%endif

%undefine _annotated_build
%global _lto_cflags %{nil}
%global __cmake_in_source_build 1

%global have_cuda 1

%global vcu_maj 12
%global vcu_min 3
%global gpu_target_arch "3.5+PTX;5.2;6.1;7.5;8.6"

%bcond_without cuda
%if %{without cuda}
%global have_cuda 0
%endif

%if %{have_cuda}
BuildRequires:  cuda-nvcc-%{vcu_maj}-%{vcu_min}
BuildRequires:  cuda-nvtx-%{vcu_maj}-%{vcu_min}
BuildRequires:  cuda-cudart-devel-%{vcu_maj}-%{vcu_min}
BuildRequires:  cuda-nvml-devel-%{vcu_maj}-%{vcu_min}
BuildRequires:  cuda-nvrtc-devel-%{vcu_maj}-%{vcu_min}
BuildRequires:  cuda-driver-devel-%{vcu_maj}-%{vcu_min}
BuildRequires:  cuda-profiler-api-%{vcu_maj}-%{vcu_min}
BuildRequires:  libnccl-devel
%endif

%description
High-performance neural network inference computing
framework optimized for mobile platforms.

%package        devel
Summary:        Development files for %{name}
Requires: %{name} = %{version}-%{release}

%description    devel
This package contains development files for %{name}.

%package        python3
Summary:        Python files for %{name}
Requires:       %{name} = %{version}-%{release}
Requires:       python3-numpy python3-tqdm python3-requests python3-portalocker opencv-python3

%description    python3
This package contains python files for %{name}.


%prep
%setup -T -c -n %{name}
git clone --depth 1 -n -b %{branch0} %{source0} .
git fetch --depth 1 origin %{schash0}
git reset --hard %{schash0}
#git submodule update --init
git log --format=fuller


%build
rm -rf glslang
sed -i 's|ncnn STATIC|ncnn SHARED|' src/CMakeLists.txt
sed -i '/PRIVATE -fvisibility=hidden/d' src/CMakeLists.txt
sed -i 's|DESTINATION lib|DESTINATION %{_lib}|g' src/CMakeLists.txt
sed -i 's|ncnn PROPERTIES|ncnn PROPERTIES SOVERSION 1 VERSION 1.0|' src/CMakeLists.txt
sed -i 's|add_subdirectory(pybind11)|find_package(pybind11 REQUIRED)|' python/CMakeLists.txt
sed -i '1i #include <cstdint>' tools/darknet/darknet2ncnn.cpp
mkdir build
pushd build
%global optflags %(echo %{optflags} | sed 's|-fno-exceptions||')
%cmake .. -Wno-dev \
       -DCMAKE_SKIP_RPATH=ON \
       -DCMAKE_VERBOSE_MAKEFILE=OFF \
       -DCMAKE_BUILD_TYPE=RelWithDebInfo \
       -DCMAKE_INSTALL_LIBDIR=%{_lib} \
       -DCMAKE_CXX_FLAGS="%{optflags} -fpermissive" \
       -DCUDA_TOOLKIT_ROOT_DIR="/usr/local/cuda-%{vcu_maj}.%{vcu_min}/" \
%if 0%{?fedora}
       -DNCNN_SYSTEM_GLSLANG=ON \
%else
       -DNCNN_SYSTEM_GLSLANG=OFF \
%endif
%if 0%{?rhel} || (0%{?fedora} == 36)
       -DNCNN_AVXVNNI=OFF \
       -DNCNN_AVX512FP16=OFF \
       -DNCNN_AVX512BF16=OFF \
%endif
       -DNCNN_SHARED_LIB=ON \
%if "%{_lto_cflags}" == "%{nil}"
       -DNCNN_ENABLE_LTO=OFF \
%else
       -DNCNN_ENABLE_LTO=ON \
%endif
       -DNCNN_DISABLE_EXCEPTION=OFF \
       -DNCNN_PYTHON=ON

make %{?_smp_mflags}
popd


%install
rm -rf %{buildroot}

pushd build
make install DESTDIR=%{buildroot}
popd

pushd python
%{__python3} setup.py install --root %{buildroot}
popd
rm -rf %{buildroot}/%{python3_sitelib}/*.egg-info


%files
%doc README.md CONTRIBUTING.md
%{_bindir}/*
%license LICENSE.txt
%{_libdir}/lib*.so*

%files devel
%{_includedir}/*
%{_libdir}/cmake/*
%{_libdir}/pkgconfig/*

%files python3
%{python3_sitelib}/*


%changelog
* Fri Mar 29 2019 Balint Cristian <cristian.balint@gmail.com>
- github upstream releases