peterbell10 comments

Results 134 comments of


                                            peterbell10

[FRONTEND][BACKEND] Add `tl.atomic_load` and `tl.atomic_store`

> could we have sem param to load/store similar to e.g., atomic_add to enforce ordering? That is one option, but I prefer this form for a number of reasons: 1....

[FRONTEND][BACKEND] Add `tl.atomic_load` and `tl.atomic_store`

For some additional context, I was hoping to extend this to support vectorized loads and stores. So I could write a pair of values in a single atomic operation. For...

[FRONTEND][BACKEND] Add `tl.atomic_load` and `tl.atomic_store`

So from the above, I'm taking that we should support tensors of size 1, 2 or 4 elements where all the elements will be loaded/stored from a single thread. I'd...

[FRONTEND][BACKEND] Add `tl.atomic_load` and `tl.atomic_store`

I'd be happy to rebase if @ptillet is okay with adding `tl.atomic_{load,store}` as scalar-only for now.

Calling .to(...) on scalar arg breaks when specialization kicks in

It would be nice if `constexpr` values had the same interface as `tensor`, but in the mean time you can use `tl.full` to turn a constant into a scalar tensor...

Teach dynamo about torch.func.jvp

@pytorchbot revert -m "broken mac jobs on main" -c nosignal

Fix FakeTensorUpdater logic for updating fake tensors

@pytorchbot merge -f "No macos runners available"

Refactor instance_descriptor for new triton version

@pytorchbot merge

Refactor instance_descriptor for new triton version

@pytorchbot merge

Refactor instance_descriptor for new triton version

@pytorchbot merge -i