bhack
bhack
Kronos maintain a bidirectional [LLVM to SPIR-V translator](https://github.com/KhronosGroup/SPIRV-LLVM). There are also some efforts inside Kronos on the OpenVX roadmap to support dnn.
It will translate also compatibile LLVM bytecode in SPIR-V. SPIR-V is also the machine model for Vulkan.
@hughperkins If you want to generally take an overview of these differents IR you can see https://www.linkedin.com/pulse/era-intermediate-languages-vincent-hindriksen. I think that sass it is more at gcn level in the AMD...
Take a look at https://github.com/KaimingHe/deep-residual-networks/
And http://torch.ch/blog/2016/02/04/resnets.html
There are interesting performance on a different Atom with https://github.com/IntelLabs/SkimCaffe
/cc @Edgarriba. Some solution could be very useful cause Travis testing time it is very slow when we need to enable NNPACK backend. P.s. Can you start to tag some...
/cc @edgarriba
@Maratyszcza Also I See that you have started to introducing neon and fp16. Do you plan to go also on lower precision like gemmlowp?
We are also supporting 8bit in tinydnn with the @wangyida effort.