search

Maratyszcza / NNPACK

Acceleration package for neural networks on multi-core CPUs
BSD 2-Clause "Simplified" License
1.68k stars 316 forks source link

How to reduce transform time? #109

Closed dzyong closed 7 years ago

dzyong commented 7 years ago

When I use NNPACK with caffe's convolution layer, I find it cost almost half of time to transform output data: profile = {total = 0.35187557300014305, input_transform = 0.032353789985791082, kernel_transform = 0.00099588599914568476, output_transform = 0.14845979897654615, block_multiplication = 0.1696860040538013}

Is there any way to reduce this time?

Maratyszcza commented 7 years ago

It looks like you have small input image with few input channels. Output transform time can't be reduces. You can get rid of kernel transform time, through, by pre-computing kernel transforms as described in #82