Repository navigation
[Epic] Reduce CK / CKTile Build Times #4229
Description
[Migrated from original issue] ROCm/composable_kernel#3575
Original issue author: @shumway
We're kicking off a systematic effort to dramatically reduce CK and CK-Tile build times, starting with an intensive three-week research and hackathon-style phase, followed by a sustained multi-month effort.
Why
Current builds take hours, slowing iteration, CI, and developer productivity. Faster build times will be great for users, and kernels that compile in seconds enable more interactive work.
Strategy
Systematic approach:
- Profile builds thoroughly and measure impact of changes.
- Use data-driven prioritization to focus on high-value improvements.
- Document and test the core meta-programming logic in CK and CK Tile.
Three strategies for code changes:
- Rewrite metaprogramming using best practices for C++17, e.g. replace math algorithms implemented with recursive template instantiation with algorithms in
constevalfunctions.. - Remove cruft and unused or obsolete kernel code.
- Simplify code, e.g. by pruning the include dependency trees and reduce duplication, to lighten compiler load.
Complementary approaches
- Bottom-up complexity analysis of compile-time operations.
- Add unit tests for metaprogramming primitives to enable safe refactoring.
- AI-assisted detection of compile-time inefficiencies.
Success Criteria
- Build times significantly reduced.
- Evidence-based improvements tracked over time.
- Processes in place for sustained monitoring and improvement of build times.
Activity
assistant-librarian commentedon Feb 3, 2026
Comment by @shumway on 2026-01-28T05:29:43Z
Build Time Reduction Status Update for 2026.01.28
This week we identified several build-time reductions through template simplification, removed costly header dependencies, and established comprehensive documentation and analysis infrastructure. We documented our optimization strategy in BUILD_TIME_OPTIMIZATION.md and introduced Python analysis tools with Jupyter notebooks for data-driven optimization.
The Jupyter notebooks can be run directly from VSCode, giving developers powerful data analysis capabilities to examine compiler time in detail. This enables systematic investigation of template instantiation costs, compilation phase breakdowns, and identification of optimization opportunities across thousands of compilation units.
Merged PRs:
std::chronodependency from CK Tile to reduce header costs-ftime-tracedata with pandas and Jupyter notebooksUpcoming Priorities
sci-libs/composable-kernel: add 7.2.0
sci-libs/composable-kernel: add 7.2.0
Add multi-file trace parsing and analysis pipeline (#4259)
Add multi-file trace parsing and analysis pipeline (#4259)
Add multi-file trace parsing and analysis pipeline (#4259)
[CK TILE] Refactor sequence_reverse_inclusive_scan (#4355)
[CK TILE] Refactor sequence_reverse_inclusive_scan (#4355)
Add generate_identity_sequences helper and replace lambdas with named…
sci-libs/composable-kernel: add 7.2.0
Add multi-file trace parsing and analysis pipeline (ROCm#4259)
[CK TILE] Refactor sequence_reverse_inclusive_scan (ROCm#4355)
Add generate_identity_sequences helper and replace lambdas with named…
Replace nested static_for lambdas with compile-time search helper (#6696