search

microsoft / knossos-ksc

Compiler with automatic differentiation
Other
45 stars 10 forks source link

WIP: benchmarking raw ks #859

Closed awf closed 3 years ago

awf commented 3 years ago

So now we can compare ks and ks_fast. Note that this one calls sufrev on every element, so essentially the "backward" includes the forward time.

vrelu3 PyTorch fast  1.692170e+04 Runs: 1541 | Inference:      1.182 ms | Forward:      1.123 ms | Backward      0.924 ms | torch.Size([65536])
vrelu3 Knossos fast  1.692170e+04 Runs: 1442 | Inference:      0.858 ms | Forward:      0.858 ms | Backward      1.739 ms | torch.Size([65536])
vrelu3 Knossos       1.692170e+04 Runs:  528 | Inference:      2.715 ms | Forward:      2.766 ms | Backward      3.984 ms | torch.Size([65536])
toelli-msft commented 3 years ago

Thanks, this works nicely.