r-barnes / reproducible_floating_sums Goto Github PK

Thank you for this work, it is a great demonstration of the reproducible summation algorithms.

The license of this source code doesn't seem to be specified. Would the author consider making this work available under a BSD / MIT / Apache license?

Thanks

Hello, I'm just curious if there are any plans to replace PyTorch's CPU SumKernel implementation for floating point types (which uses pairwise summation) with one of these?

I'm yet to test them but am curious as to how well they perform with multiple threads, and if they have better accuracy than pairwise summation.

Thank you!

Having finally gotten around to hacking on this code, to port it to CUDA, while I've got an initial single thread version that passes the tests ok, in moving to a multi-thread reduction, I've found that the binned_dmdmadd function, used to sum two instances of ReproducibleFloatingAccumulator is buggy.

Specifically, when instantiated, it doesn't compile because:

• The function attempts to modify the other reference, which is const qualified, e.g.

priY[i*incpriY] = priX[i*incpriX];

• These lines are erroneous, since they attempt to pass an argument to a function that expects none

const auto X_index = binned_index(priX);
const auto Y_index = binned_index(priY);