bhack
bhack
@jdemouth Yes i know that all are working at assembly level, a level that in AMD world we could call GCN. But I hoped that a common target like SPIR-V...
@LunarG has heavily experimented on a two step IR but on shaders. Probably can give us some feedback on backend limits based on LunarGlass experience.
@jdemouth Yes I'm only guessing if exist and what is the vendor neutral "lowest common denominator". This could still give a margin for hardware vendors to compete but also give...
[OpenVX](https://www.khronos.org/openvx/) was a good example for interface collaboration that involved many stakeholders (Nvidia included). But it has totally lost the occasions to cover deep learning needs in the actual release....
The common IR that target both NVIDIA and AMD is SPIR-V (and was co-designed). But seems that it is not enough for achieve this level of optimization. So if the...
@jdemouth If you like the idea you can try to talk about this with Thierry Lepley. He is the Nvidia representative in the Khronos OpenVX standardization group.
And I think that we could start to consider also the interesting results from [binarizzation](http://arxiv.org/abs/1603.05279)
This start to become interesting also on CPU. /cc @xianyi
@xianyi On AVX-512 I think that there is a XNOR instruction __mmask16 _mm512_kxnor? Do you think that this kind of operators could be included in openblas?
A kernel under BSD it is at https://github.com/MatthieuCourbariaux/BinaryNet/blob/master/Run-time/binary_kernels.cu