Add fix & tests for new mapreduce_scalar bypass - #78
Conversation
kshyatt
changed the title
Codecov Report✅ All modified and coverable lines are covered by tests.
... and 2 files with indirect coverage changes 🚀 New features to boost your workflow:
|
|
Do you know if there is any point in trying to schedule the kernel on the GPU and only afterwards doing the scalar indexing on the output? Or am I misreading this |
|
Which kernel do you mean? |
| out = arrays[1] | ||
| iout = ParentIndex(Strided.offset(out) + 1) | ||
| @allowscalar begin | ||
| v = f(map(a -> a[ParentIndex(Strided.offset(a) + 1)], Base.tail(arrays))...) |
There was a problem hiding this comment.
@kshyatt I meant here, since if I understand this correctly this is effectively just first bringing everything to the CPU, and then computing f? I'm not sure it really matters to be honest, since I don't think you ever want to end up in this situation and need it to be performant to begin with
There was a problem hiding this comment.
Yeah I think it doesn't matter either way, this was only being triggered for AMD because I think cuTENSOR bypasses this. But it was getting picked up by one of the amd/planar tests for a previous issue, so there is a way to reach this code.
lkdvos
left a comment
lkdvos
left a comment
There was a problem hiding this comment.
do you want to immediately bump the version as well?
|
Yeah let's, I'll do so now |
Bump patch version
2 participants
Adds a test that was needed but not added to #76, which currently fails. Once this starts working on AMD, TensorKit.jl should un-break.