why bias cost so much ?

[guoxuesong]

I found some examples of neon not add bias to Affine, I tried to add it back, It becomes much slower. for example mnist_mlp.py use 3.76s/epoch without bias, but 5.30s/epoch with bias.

The difference is HUGE !

convnet-benchmarks/googlenet_v1.py, also huge difference.

I tried lasagne, with bias or without bias, nearly same.

[wei-v-wang]

Hi @guoxuesong Thanks for your report! This is a known issue to us and we have been trying to fix it. Can you test with this commit id to see if you still have the issue with adding bias?

https://github.com/NervanaSystems/neon/commit/199f215aeac1b699a63cb808ada88ece180c46f0

[wei-v-wang]

Hi @guoxuesong We have done fusion of convolution + bias to improve the performance for convolution with bias. Can you please verify if you observe big perf. improvement?

https://github.com/NervanaSystems/neon/tree/v2.3.0 (neon v2.3.0)

[wei-v-wang]

The reason adding bias was slower was because Bias layer was a non-MKL op and data conversion needed to happen. With neon v2.3.0. Convolution +Bias layers become Convolution_Bias layers.

[wei-v-wang]

Please re-open if issues still exist. Thanks for reporting!