Figure out the future of dynamic_slice vs real_dynamic_slice
#2176
Comments
GleasonK
added a commit
that referenced
this issue
May 13, 2024
A proposal to remove redundant operations from StableHLO before long-term compatibility guarantees go into place. High level summary: - Remove `CreateTokenOp`, `TraceOp`, `BroadcastOp`, `DotOp`, `UnaryEinsumOp`, `RealDynamicSliceOp`. - Enhance `DynamicSliceOp`. - Move `CrossReplicaSumOp` to CHLO. - Hopefully remove/move to CHLO (need feedback) `MapOp`, `RngOp`, `EinsumOp`, `TorchIndexSelectOp`, `GetTupleElementOp`, `tuple` and `tuple` type. OpenXLA Discuss post: https://groups.google.com/a/openxla.org/g/openxla-discuss/c/sBAkvnd2bcA Related tickets: #2176, #3
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Currently we have the specced
dynamic_sliceand the unspeccedreal_dynamic_slicethat is inherited from mhlo.dynamic_sliceis a misnomer and not fully dynamic andreal_dynamic_sliceis the fully dynamic version ofslice.dynamic_sliceonly offers dynamicstart_indices, but the size of the slice still needs to be determined statically. In addition, there is no way to provide slices with strides other than one.real_dynamic_sliceon the other hand provides full capabilities ofslicewith dynamism support for everything:start_indices,limit_indices, andstrides.This is the current state for now, but we should figure out how to move forward with these two ops. Some options are:
real_dynamimc_sliceto the opset and keepdynamic_slice.real_dynamimc_sliceto the opset and deprecatedynamic_slice.Further investigation is needed to decide on which path to move forward.
The text was updated successfully, but these errors were encountered: