search

microsoft / knossos-ksc

Compiler with automatic differentiation
Other
45 stars 10 forks source link

sqrl upper bounds #968

Closed toelli-msft closed 2 years ago

toelli-msft commented 2 years ago

The handwritten C++ is beating or roughly matching PyTorch in all cases except backwards on large sizes.

image

awf commented 2 years ago

Thanks! For sqrl, the 4x4 size is the targetted one, so this is great.

toelli-msft commented 2 years ago

Thanks! For sqrl, the 4x4 size is the targetted one, so this is great.

On that point, do we really want to benchmark "vsqrl"? https://github.com/microsoft/knossos-ksc/issues/966

awf commented 2 years ago

Shall we work to merge this?

toelli-msft commented 2 years ago

Shall we work to merge this?

Yes, the current version is a good start. We can increase the problem sizes during or after the implementation of vsqrl.

toelli-msft commented 2 years ago

What we actually want here is vsqrl so I suggest we wait until vsqrl can be implemented via https://github.com/microsoft/knossos-ksc/pull/1010 and then implement vsqrl in C++ rather than just sqrl.

awf commented 2 years ago

vsqrl is now in -- what should next steps be here?

toelli-msft commented 2 years ago

The next step is to write vsqrl in C++. I'm not sure if I'll get to that, I'm afraid.

awf commented 2 years ago

No problem, this should go in anyway.