cat tensorflow* > tensorflow-1.12.0rc0.whl
Motivation: [Request] Pre-build support old CPU #18689
No custom build is available for my workstation which has a fairly recent GPU but an old CPU. According to the tensorflow project team a CPU without AVX will not be supported in the official tensorflow-gpu package any time in the future.
"Unfortunately we are unable to provide pre-built binaries without AVX. Unofficial third-party builds may be your best option."
I have built tensorflow from source without issues. It took about 3 hours. I have installed the wheel and tested the library in a keras framework. It works for me :)
Target environment:
Ubuntu 18.04.1 LTS (bionic) Intel® Xeon(R) CPU W3520 @ 2.67GHz × 8 GeForce GTX 1050 Ti/PCIe/SSE2 CUDA 10.0 CUDNN 7.4.1
Bazel build options:
TF_NEED_OPENCL_SYCL="0" TF_NEED_ROCM="0" TF_NEED_CUDA="1" CUDA_TOOLKIT_PATH="/usr/local/cuda" TF_CUDA_VERSION="10.0" CUDNN_INSTALL_PATH="/usr/lib/x86_64-linux-gnu" TF_CUDNN_VERSION="7" TF_NCCL_VERSION="" TF_CUDA_COMPUTE_CAPABILITIES="6.1" TF_CUDA_CLANG="0" GCC_HOST_COMPILER_PATH="/usr/bin/x86_64-linux-gnu-gcc-6" build --config=cuda test --config=cuda build:opt --copt=-march=native build:opt --copt=-mssse3 build:opt --copt=-mcx16 build:opt --copt=-msse4.1 build:opt --copt=-msse4.2 build:opt --copt=-mpopcnt build:opt --host_copt=-march=native build:opt --define with_default_optimizations=true build:v2 --define=tf_api_version=2