Repository navigation
D-RMAP-2: count the row-major B blocks in the BF16 byte total - #1625
Conversation
Bf16Staged::bytes omitted b_rows, which stage() keeps for the polyfill entry, so the cost example reported 256 KiB for an object holding 384 KiB (matrix, packed blocks, row-major blocks, 128 KiB each). Adds a test pinning the total; plan §5a and the board entry corrected. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VyWZYKycy6MoYDxHRVnV5K
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars. ⚙️ Run configuration
Warning Billing warning: we have not been able to collect payment for this subscription for more than 72 hours. Please update the payment method or pay any pending invoices in Billing to avoid service interruption. Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
2 participants
Follow-up to #1623. Lab probe only, nothing admitted to production.
Bf16Staged::bytes()left outb_rows, the row-major B blocksstage()keeps for the polyfill entry. Sotile_gemm_costreportedbf16_bytesas 256 KiB for an object that holds 384 KiB (the BF16 matrix, packed blocks and row-major blocks, 128 KiB each). Codex found this on #1623 at23f0998. The fix was pushed after #1623 merged and did not make it into the merge, so this PR carries it ontomain(one commit,74eb6cf7, cut from8b570ed3).bytes()now countsb_rows.bf16_bytes_count_every_retained_bufferpins the total to3 * 256 * 256 * 2. With theb_rowsterm removed it fails (262 144 vs 393 216).Not in this PR: the cost example reported tier
AMX TDPBF16PSon one run of the session host, where §5a's numbers come from the F32x16 polyfill tier. One run is not a measurement, so §5a's results and its "AMX hosts not measured" line are unchanged.Checks:
cargo test --release --test tile_gemm(5 passed), clippy-D warningsand fmt clean,entries_index.py --checkpasses. CI on74eb6cf7: all 13 checks green.🤖 Generated with Claude Code
https://claude.ai/code/session_01VyWZYKycy6MoYDxHRVnV5K