Implementation details and scope for performance improvements about caffe-nnpack

[Maratyszcza]

1. cxgemm is a tuple-based complex matrix-matrix multiplication micro-kernel. Please see slides from my BLIS retreat 2016 talk for details. cxgemm is only used with FFT-based algorithm, Winograd-based algorithm does sxgemm instead.
2. NNPACK uses NCHW format - the same as Caffe, Torch & Theano (but different from TensorFlow and Neon).
3. nnp_convolution_inference is close to optimum, at least in single-threaded execution. Other convolution functions are behind in optimizations. No plans to implement smaller Winograd tiles.

[anijain2305]

Thanks for the comments and the slides.

Also, my characterization was based on the convolution-output implementation as caffe-nnpack directly links to output functions. So, I think I went the wrong path.

Can you tell me how to link to convolution-interference? Is my current understanding that caffe-nnpack links to convolution-output incorrect?

[Maratyszcza]

@anijain2305 Please see nnpack-pr branch in ajtulloch/caffe, which is a newer version of Caffe bindings. caffe-nnpack wasn't updated for a long time.