-
released this
2026-10-09 22:34:04 +00:00 | -21 commits to master since this releaseThis release is the camera-ready version for SSCAD-WIC: the paper was tightened to the 8-page body limit while answering the reviews, supporting numbers that did not fit were moved into report documentation, and a latexrun hang under TinyTeX was fixed.
What Changed
- Camera-ready manuscript (
report/main.tex): rewritten abstract, related work (added Dave & Kotak 2024 and Saha 2026; softened the gap claim), methodology, results, and conclusion to address reviews. Corrected analytical model (Amdahl form of Eq. 1 now uses the properf_{32} = 2f/(1+f); residual for ATAX decomposed into latency rise + sub-linear miss reduction), added predicted-vs-measured line-size column, clarified core-model realism and determinism, made the warm-start explicit (caches are not flushed at ROI entry) and used it to attribute ATAX’s capacity gain to L3 ≥ 32 MiB, replaced the unmeasured “compulsory and capacity misses” claim with the actual L1D miss counts from the associativity sweep, and fixed access-pattern / live-data occupancy descriptions. - New report notes (
report/README.md): records what the page limit forced us to compress or omit (dropped Table 4, full f_stall sweep, original model error, associativity/capacity per-preset numbers, determinism details, etc.), the corrections to the accepted text (arithmetic fixes, attribution of capacity, warm-start mechanism, residual of Eq. 1, etc.), and deferred reviewer requests. - Compact line-size figure:
visualize_results.pygainscreate_cache_line_compact_figure()so L1D miss rate and IPC fit a single short two-panel figure in the report. - latexrun freeze under TinyTeX: pin
max_print_line=79inreport/.justfileandreport/Makefile, and guard the unwrap loop in the vendoredlatexrunso a non-79 wrap no longer spins forever. - Bibliography: added Smith (1987), Hill & Smith (1989), Cantin & Hill (2001), Williams et al. (2009), Hennessy & Patterson (2017), Gholami et al. (2024), Dave & Kotak (2024), Saha (2026); repository citation moved to a footnote and the old
repo:2026entry removed. - Tag / clone instructions: default branch tag is now
v0.1.2(README and demo script).
Setup/Docs Improvements
- Top-level README points readers at
report/README.mdfor paper-specific constraints, omissions, and how to rebuild only the PDF. - Report build docs: one-command
zig build report(orjust allinsidereport/) when figures are already present.
Downloads
-
Source code (ZIP)
0 downloads
-
Source code (TAR.GZ)
0 downloads
-
A_Quantitative_Cache_Evaluation_of_Select_PolyBench_Kernels_-_Andrade_Canizela_Dimas_Jordão_Pellegrini_0.1.2.r0.g6954d3c.pdf
0 downloads ·
2026-10-10 16:12:47 +00:00 · 157 KiB -
results.tar.zst
0 downloads ·
2026-10-10 16:13:53 +00:00 · 12 MiB -
workloads.tar.zst
0 downloads ·
2026-10-10 16:13:09 +00:00 · 1.4 MiB
- Camera-ready manuscript (