JuliaDiff / JuliaDiff/ChainRules.jl
Avoid device-to-host copy in `∇getindex!`
Open
Nobody has claimed this yet.
GPU
- Dominant language
- Julia
- Stars
- 475
- Forks
- 98
- PR merge metrics
- No merged PRs in 30d
Description
Can we use custom kernel with atomics for ∇getindex!(dx::AbstractGPUArray, dy, inds...) instead of copying everything to CPU?
This way we'd be able to avoid synchronizations and we can add such kernel via extension
Contributor guide
No contributing guide indexed for this repository
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 src/rulesets/Base/indexing.jl at ∇getindex!(dx::AbstractGPUArray, dy, inds...), especially around line 183. Trace how the current implementation copies data to the CPU, then investigate the proposed custom atomic kernel and extension path. Done means the GPU operation avoids the device-to-host copy and associated synchronizations.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- julia
- Domain
- performance
- Issue type
- Refactor
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100