%global githash 98e6b36
%global gitdate 20210913

Name:           mxnet
Version:        2.0.0.alpha.rc3
Release:        %{gitdate}.git%{githash}%{?dist}
Summary:        Lightweight, Portable, Flexible Distributed/Mobile Deep Learning
License:        Apache
URL:            https://github.com/apache/incubator-mxnet

Source0:        mxnet.tar.xz
Patch0:         mxnet-sover.patch
Patch1:         mxnet-cmake.patch
Patch2:         mxnet-gcc11.patch
Patch3:         mxnet-cuda11.patch

BuildRequires:  python3-setuptools perl-Encode perl-Data-Dumper dmlc-core-devel
BuildRequires:  gperftools-devel openblas-devel python3-devel onednn-devel tbb-devel
BuildRequires:  doxygen cmake gcc-c++ lapack-devel gtest-devel libjpeg-turbo-devel ocl-icd-devel
BuildRequires:  opencv-devel


%global have_cuda 1

%global vcu_maj 11
%global vcu_min 4
%global gpu_target_vers "3.5+PTX 5.2 6.1 7.5 8.6"

%global __cmake_in_source_build 1

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

%if %{have_cuda}
%if 0%{?fedora}
BuildRequires:  cuda-gcc-c++
%endif
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-nvprof-%{vcu_maj}-%{vcu_min}
BuildRequires:  libcublas-devel-%{vcu_maj}-%{vcu_min}
BuildRequires:  libcufft-devel-%{vcu_maj}-%{vcu_min}
BuildRequires:  libcurand-devel-%{vcu_maj}-%{vcu_min}
BuildRequires:  libcusparse-devel-%{vcu_maj}-%{vcu_min}
BuildRequires:  libcusolver-devel-%{vcu_maj}-%{vcu_min}
BuildRequires:  libnccl-devel
BuildRequires:  libcudnn8-devel
BuildRequires:  libcutensor-devel
Requires:  cuda-cudart-%{vcu_maj}-%{vcu_min}
Requires:  cuda-nvrtc-%{vcu_maj}-%{vcu_min}
Requires:  libcublas-%{vcu_maj}-%{vcu_min}
Requires:  libcufft-%{vcu_maj}-%{vcu_min}
Requires:  libcurand-%{vcu_maj}-%{vcu_min}
Requires:  libcusparse-%{vcu_maj}-%{vcu_min}
Requires:  libcusolver-%{vcu_maj}-%{vcu_min}
%endif

%description
Open deep learning compiler stack for cpu, gpu and specialized accelerators.

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

%description devel
This package contains development files for mxnet.

%package python3
Summary: Python files for mxnet
Requires: %{name} = %{version}-%{release}
Requires: python3-graphviz

%description python3
This package contains python files for mxnet.


%prep
%setup -q -n %{name}
%patch0 -p0 -b .sover~
#%patch1 -p0 -b .cmake~
%patch2 -p0 -b .gcc11~
%patch3 -p0 -b .cuda11~

%build
rm -rf include/onednn
rm -rf 3rdparty/onednn
rm -rf 3rdparty/openmp
#rm -rf 3rdparty/dlpack
rm -rf 3rdparty/dmlc-core
rm -rf 3rdparty/googletest
rm -rf 3rdparty/nvidia_cub

# graphviz
sed -i -e 's|graphviz<0.9.0,>=0.8.1|graphviz>=0.8.1|g' python/setup.py
# external gtest
sed -i -e '/set(GTEST_ROOT/d' CMakeLists.txt
sed -i -e '/add_subdirectory(${GTEST_ROOT})/d' CMakeLists.txt
sed -i -e 's|${GTEST_ROOT}/include|${GTEST_ROOT}|g' CMakeLists.txt
sed -i -e '/set_target_properties(gtest/d' CMakeLists.txt
sed -i -e '/set_target_properties(gtest_main/d' CMakeLists.txt
# external mkl-dnn
sed -i -e '/add_custom_command(TARGET mxnet/,/)/d'  CMakeLists.txt
sed -i -e 's/load_onednn()/find_package(DNNL REQUIRED)/' CMakeLists.txt
sed -i -e '/set_target_properties(dnnl/d' CMakeLists.txt
sed -i -e 's|add_subdirectory(3rdparty/onednn)|find_package(DNNL REQUIRED)|' CMakeLists.txt
# shared openblas (openmp variant)
sed -i -e 's|libopenblas.a|libopenblaso.so|' cmake/Modules/FindOpenBLAS.cmake
# disable test
sed -i -e 's|add_subdirectory(tests)|#add_subdirectory(tests)|g' CMakeLists.txt
# add flag
sed -i -e 's| MSHADOW_INT64_TENSOR_SIZE| 0|' include/mxnet/libinfo.h
# relax warn-errors
sed -i -e '/-Werror/d' CMakeLists.txt

mkdir build
pushd build
export LD_LIBRARY_PATH="/usr/local/cuda-%{vcu_maj}.%{vcu_min}/%{_lib}/"
%cmake .. -Wno-dev \
       -DCMAKE_SKIP_RPATH=ON \
       -DCMAKE_VERBOSE_MAKEFILE=OFF \
       -DCMAKE_BUILD_TYPE=RelWithDebInfo \
       -DUSE_INT64_TENSOR_SIZE=OFF \
       -DUSE_OPENCV=ON \
       -DUSE_OPENMP=ON \
       -DUSE_BLAS="Open" \
       -DUSE_ONEDNN=ON \
%ifarch x86_64 %{ix86}
       -DUSE_SSE=ON \
       -DUSE_F16C=ON \
       -DCMAKE_C_FLAGS="-mf16c" \
%else
       -DUSE_SSE=OFF \
       -DUSE_F16C=OFF \
%endif
%if %{have_cuda}
       -DUSE_CUDA=ON \
       -DUSE_NCCL=ON \
       -DUSE_CUDNN=ON \
       -DUSE_FATBIN_COMPRESSION=ON \
       -DCUDA_PROPAGATE_HOST_FLAGS=OFF \
       -DMXNET_CUDA_ARCH=%{gpu_target_vers} \
       -DCMAKE_CUDA_FLAGS="--compiler-options -fPIC -Wno-deprecated-gpu-targets" \
%if 0%{?fedora}
       -DCMAKE_CUDA_HOST_COMPILER="%{_bindir}/cuda-gcc" \
%endif
       -DCMAKE_CUDA_COMPILER="/usr/local/cuda-%{vcu_maj}.%{vcu_min}/bin/nvcc" \
%else
       -DUSE_CUDA=OFF \
%endif
       -DUSE_JEMALLOC=OFF \
       -DUSE_GPERFTOOLS=OFF \
       -DUSE_LIBJPEG_TURBO=OFF \
       -DBUILD_CPP_EXAMPLES=OFF \
       -DGTEST_ROOT="%{_includedir}/gtest"

make %{?_smp_mflags}

popd

pushd python
%{__python3} setup.py build
popd


%install
rm -rf %{buildroot}
export LD_LIBRARY_PATH="/usr/local/cuda-%{vcu_maj}.%{vcu_min}/%{_lib}/"
pushd python
%{__python3} setup.py install --root %{buildroot}
popd

# includes
for d in `find ./include/ -type d`; do
  mkdir -p %{buildroot}%{_includedir}/$d
done

for i in `find ./include/ -type f`; do
  install -p -m 0644 $i %{buildroot}%{_includedir}/$i
done


# python
mkdir -p %{buildroot}/%{_libdir}
cp -rl %{buildroot}/usr/lib/* -t %{buildroot}/%{_libdir}/
rm -rf %{buildroot}/usr/lib
# libraries
cp -rl %{buildroot}/usr/%{name}/* -t %{buildroot}/%{_libdir}/
rm -rf %{buildroot}/usr/%{name}

# python3
for f in `find %{buildroot} -name '*.py'`;
do
  sed -i -e 's|#!/usr/bin/env python|#!/usr/bin/python3|' $f
done

# fix requirement
sed -i -e 's|0.9.0|0.15.0|g' %{buildroot}/%{python3_sitearch}/*.egg-info/requires.txt

# fix symlink
ln -sf %{_libdir}/libmxnet.so %{buildroot}/%{python3_sitearch}/mxnet/libmxnet.so


%files
%doc README.md CONTRIBUTORS.md
%license LICENSE
%{_libdir}/lib*.so*

%files devel
%{_includedir}/*
%{_datadir}/*

%files python3
%defattr(-,root,root,-)
%{python3_sitearch}/*


%changelog
* Mon Mar 18 2019 Balint Cristian <cristian.balint@gmail.com>
- github build