pymc-devs / pymc-devs/pytensor
Avoid explicit broadcasting of indices in Advanced[Inc]Subtensor
Open
Nobody has claimed this yet.
graph rewriting
indexing
memory opt
vectorization
- Dominant language
- Python
- Stars
- 644
- Forks
- 208
- Avg merge
- 2d 14h
- Merged PRs (30d)
- 16
Description
Description
Advanced indexing broacast indices implicitly, so in the following case there's no reason to allocate several ones:
import pytensor
import pytensor.tensor as pt
x = pt.matrix("x")
out = x[pt.arange(x.shape[0]), pt.ones(x.shape[0], dtype=int)]
fn = pytensor.function([x], out)
fn.dprint()
# AdvancedSubtensor [id A] 3
# ├─ x [id B]
# ├─ ARange{dtype='int64'} [id C] 2
# │ ├─ 0 [id D]
# │ ├─ Shape_i{0} [id E] 0
# │ │ └─ x [id B]
# │ └─ 1 [id F]
# └─ Alloc [id G] 1
# ├─ 1 [id H]
# └─ Shape_i{0} [id E] 0
# └─ ···
We already do this optimization for the y value in IncSubtensor with local_useless_inc_subtensor_alloc:
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start in pytensor/tensor/rewriting/subtensor.py around local_useless_inc_subtensor_alloc, then run the reproducer in the issue and inspect its dprint output. Reuse the existing optimization's context to determine how AdvancedSubtensor and AdvancedIncSubtensor handle implicitly broadcast indices; done means redundant index allocations no longer appear in the graph.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- compilers, performance
- Issue type
- Feature
- Difficulty
- 3/5
- Estimated time
- 1-2 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 45/100