what's the relationship of this cool repos?

[hughperkins]

what's the relationship of this cool repos?

[hughperkins]

Sure:

• DeepCL A convolutional neural network, built from the ground-up, 100% OpenCL
  • people appreciate it for:
    • the convenient commandline syntax for expressing networks like 32c5-mp2-32c5-mp2-50n-10n,
    • and for the ability to run on Windows (it runs on all three major platforms actually: windows, linux, Mac)
    • it was the first OpenCL library to appear on Soumith's convnet-benchmarks page
• CLBlast is not mine. It's Cedric Nugteren's. It's a BLAS implementation for OpenCL. BLAS handles primarily matrix multiplication, also matrix-vector multiplication
• EasyCL is a low-level middleware that makes launching kernels from C++ a lot less painful than the low-level OpenCL API. I think there is a C++ API for OpenCL too actually, but I think EasyCL adds a few more bells and whistles, and plus I've been using it for a while now, across all my projects, seems to do what I need :-)
• cuda-on-cl lets you build NVIDIA® CUDA™ applications for use on OpenCL devices.

You missed one by the way ;-) Actually two:

• https://github.com/hughperkins/distro-cl port of Torch library to OpenCL
• https://github.com/hughperkins/jeigen Java wrappers for the Eigen library, which is the same library underlying Tensorflow. Eigen handles tensor operations, like addition, reduction, sort, and so on.

[alephman]

Thanks a lot hughperkins! I still have 2 questions:

1. Does the Eigen support OpenCL?
   1. Why tensorflow use Eigen instead of BLAS?

[hughperkins]

Does the Eigen support OpenCL?

Not out-of-the-box. But it does build into OpenCL, if you use cuda-on-cl. But that's because Eigen supports CUDA, and cuda-on-cl can compile CUDA applications into OpenCL.

Why tensorflow use Eigen instead of BLAS?

Tensorflow uses Eigen and BLAS

• BLAS is highly optimized for matrix multiplication
• doesnt handle things like: reduction, sorting, per-element addition etc