test: add gemv per-dispatch overhead micro-bench#23
Merged
Conversation
Measures the gemv dispatch overhead two ways on the same Apple GPU: one command buffer per op (per-commit) versus many ops batched into a single command buffer. Quantifies the cost the batched-on-command-buffer Ops API saves on the decode hot path (per-commit ~275 us/dispatch vs batched ~22 us/op at 2048x2048). No correctness threshold, prints the numbers and passes; runs in under a second.
This was referenced Jun 23, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Adds a single micro-bench test (the last functional leftover from the combined backends branch #20) measuring gemv dispatch overhead two ways on the same Apple GPU:
This quantifies what the batched-on-command-buffer
Ops.gemv(weight:input:on:)API saves on the decode hot path. No correctness threshold, prints the numbers and passes in under a second, so it is CI-safe.Context
The bulk of #20 already landed via #15-#22. The remaining unique source was already on dev under relocated paths (e.g.
Stats/Perplexity.swiftis the same as the mergedTelemetry/Perplexity.swift); this bench was the one genuinely new functional file. Per scoping, docs and Nemotron research notes are intentionally left out, and the staleIndirectDispatchTests(which targeted a kernel variant removed in the reorg) is dropped.Test
swift test --filter DispatchOverheadBenchpasses locally (M5, 0.88s).