Repository navigation
Reduce kernel launch and first-use latency - #3237
Merged
Merged
Conversation
Thread-safe launch tracking added a set allocation, a sort, and one stream-capture query per managed allocation. This cost about 300 ns, or 4-7% on small-kernel benchmarks. Sort a copy and deduplicate adjacent entries, bypass that work for a single allocation, and query capture state once per launch. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A runtime-length argument splat leaves the tuple type uninferred at the @cuda call site, causing dynamic dispatch at each conversion, compilation, and launch stage. Put the pipeline behind one function barrier so it specializes on the concrete tuple after a single dispatch. This recovers the 4% regression in the by-value slices benchmark. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The device-free precompile workload cannot reach context creation or reflection, leaving about 150 ms of compilation on the first kernel. Move context creation out of the memoization closure into a named function, and precompile it together with compiler configuration and the common code_native/code_llvm signatures. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Driver and toolkit changes on the benchmark agent have caused dashboard steps that were otherwise difficult to attribute. Print CUDA.versioninfo() before the suite so the job log records the environment. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #3237 +/- ##
==========================================
- Coverage 92.38% 92.35% -0.04%
==========================================
Files 184 184
Lines 16084 16084
==========================================
- Hits 14860 14855 -5
- Misses 1224 1229 +5 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
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.
Reduce overhead in common kernel compilation and launch paths:
Locally, type-unstable launch overhead dropped from 2625 ns and 21 allocations to 1653 ns and 6 allocations. First-kernel latency improved by roughly 150 ms, while package import time remained unchanged.