XNNPACK future work

Hi all,

Is this a good place to discuss XNNPACK or is there a more specific forum?

I was reading Faster Quantized Inference with XNNPACK — The TensorFlow Blog and the section on ‘future work’, and I wondered if there was any thinking about whether Winograd or FFT -based convolution algorithms might be useful in XNNPACK? They could produce good benefits for larger models where applicable.

Would love to hear your thoughts.


Winograd and FFT-based convolution algorithms are most suitable for non-grouped convolutions with large kernel sizes, like the ones in VGG or the original U-Net models. Nearly all modern mobile models use depthwise separable convolutions, where Fast Convolution Algorithms are inapplicable (1x1 convolutions) or inefficient (depthwise convolutions).