Much slower assigning with xt::view vs std::copy

[gte620v]

Hi,

I think these two are equivalent, but the copy is about 3x faster. Am I misusing xt:view?

std::copy(source_tensor.begin(), source_tensor.end(), keep_tensor.begin() + start)

xt::view(keep_tensor, xt::range(start, stop)) = source_tensor

Thanks.

[SylvainCorlay]

Hey, thanks for posting here.

These are not quite equivalent except for 1-D arrays.

The version with views assignments will broadcast the assignment to all the "rows" (first index) within that range.

[gte620v]

I gotcha, I have a 1-D array, though. For that case this should be the same, right? Should they also be the same speed?

For a 2-D array, this will just broadcast in row-major order, right? And if I change the layout to column-major, this will just broadcast in that order?

I am curious about speed because I have a latency-sensitive application with a bunch of xt::view calls and I am wondering if i need to rewrite those as std::copys. Any guidance would be appreciated.

Thanks!

[SylvainCorlay]

Yep, you got it except that broadcasting order is independent of the layout.

I am on a train right now. If things are contiguous, it should be as fast, especially if you enable xsimd.

[SylvainCorlay]

@gte620v is there a particular reason for closing? Were you able to reach the expected speed with SIMD enabled?

[gte620v]

No, I just took your word for it that they will be the same. I need to resurrect my profiling code to verify. I'll reopen if/when I get definitive numbers.

[SylvainCorlay]

Thanks! Sorry the whole team is busy at a series of workshops and conferences.