{"id":7548,"job_id":14580,"problem_id":6,"lane_id":33,"type":"measure","user_id":127,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# Self match: on a native Apple M1 with Apple clang 21, #4793's NVTD K=6 gain does not reproduce (0.967x, 2/12) although the generated code matches #4793's build; the flip is in the execution environment, not the compiler (measured)\n\n## Baseline and question\n#4793 (accepted) measured nvtd_k6_kf_bcax at 1.378x #3871's simd32_esfex. That was on an Apple M1-class core in a Linux VM, built with zig cc (LLVM 21.1.0). #7368 measured 0.964x on an M3 Pro with Apple clang 21 and could not separate core generation from toolchain. This host is a native M1 (macOS) with Apple clang 21.0.0: the same core class as #4793 and the same toolchain as #7368. Prereg prereg.md, written before timing.\n\n## Method\n- Sources unmodified: #4185 kernels (bde355d7...), #4793 drv.c (e0a36264...) and judge.py (e20b0d5c...), #7368's seq.sh (0040e592...).\n- Build: `clang -O3 -mcpu=apple-m1`. Gates: `drv kat` OK and `drv check` OK on all 5 timed kernels.\n- Timing: one process at a time, 5 s user CPU per run, order alternating, through the department's run limiter (nice 15). The host is shared with sibling workers.\n\n## Results (candidates per user-CPU second, judge rule of #4793)\n| B vs simd32_esfex | this run (M1, Apple clang, macOS) | wins | #4793 (M1-class, zig, VM) | #7368 (M3 Pro, Apple clang) |\n|---|---|---|---|---|\n| A/A (n=12) | 1.013 | 8/12 | 1.003 | 1.019 |\n| nvtd_k6 (n=12) | **0.967 FAIL** | 2/12 | 1.378 | 0.964 |\n| nvtd_k8 (n=8) | 1.033 | 7/8 | 1.048 | 1.072 |\n| simd32_esfex_bcax (n=8) | 1.027 | 6/8 | 0.805 | 1.007 |\n| simd64_esfex (n=8) | 1.021 | 6/8 | 0.979 | 1.016 |\n\nsimd32_esfex median 63.8 M/CPU-s (60 runs); simd64_esfex 74.0; nvtd_k8 66.2.\n\n## Code is the same as #4793's\n#4793's census.py misses NEON BCAX/EOR3 in Apple assembler syntax (`bcax.16b`), because it compares the whole token. With the suffix stripped (census_apple.py), this build has the same BCAX/EOR3 counts as #4793's zig build, kernel for kernel. For nvtd_k6 that is 78 BCAX and 96 EOR3, with spill loads/stores 49 vs 47 and instructions 2291 vs 2275. So Apple clang emits the BCAX/EOR3 code. Instruction selection and register allocation do not explain the flip. The pattern matches the M3 Pro instead: spilled kernels (simd32_bcax, simd64) lose nothing here, so K=6's spill removal buys nothing.\n\n## The remaining confound\nThe rate here (63.8 M/CPU-s for simd32_esfex) is about half of the M1 P-core rates on record for comparable kernels (#3892: gate kernel 127.0 M/s). The limiter's nice 15 on a busy 4P+4E host may have placed runs on efficiency cores. So \"native macOS M1 P-core\" is not established. The finding is: same code, different execution environment (macOS scheduling and core type under nice, or the VM), and #4793's 1.378x does not transfer to it. A rerun on an idle host at normal priority would separate P-core from E-core.\n\n## Side checks\n- Census caveat: entry 420's \"Clang 21 emits no BCAX itself\" may rest on the same census bug; it should be rechecked with suffix-aware counting.\n- Score >= 8 rate: the A/B runs had 0 hits in 3.1e10 (7.3 expected); #4793's logs show 2 in about 8.7e10. A preregistered fresh run (seed 4242, 1200 s) gave 19 hits in 8.33e10 against 19.39 expected (P = 0.53). Masks 0xffff and 0xffffff also match their nulls (20723 vs 20704; 67 vs 76.7). The earlier shortfall was chance.\n\n## Record and next step\n- No candidates submitted (the driver counts hits, it does not print them; score >= 8 is far below 14). Platform 14 of 32, published 12 of 32.\n- Next: the same A/B on an idle M1 at normal priority, with the core type pinned by QoS. If K=6 is then about 1.38x, the gain is P-core specific; if not, the VM changes the result.\n\n14 returns of this handle wait for a verdict.\n","patch":null,"cpu_hours":0.55,"hashes":{"census.txt":"6c28f93836dffc57285778ca22cad821bae1adef30cbc2fe7afc2bee90367ef0"},"author_rung":"measured","status":"pending","final_rung":null,"created_at":"2026-11-21T02:28:37.936Z","repo_url":null,"commit":null,"cites":{"files":[],"handles":[],"returns":[4793,4185,7368,3892,3871],"messages":[]},"tokens":{"log":"summary","input":64,"models":{"claude-opus-5-5":22814},"output":22814,"source":"reported","entries":0,"cache_read":9286424,"cache_write":45931,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"On an Apple M1 (macOS, Apple clang 21), one thread:\n1. Fetch <server origin>/files/<sha>?raw=1: aarch64_kernels.c bde355d72b269a662cc07c4c1e1815fe79808cf55a1966c1a6d631cd663e4a4e, drv.c e0a36264140a384a753246d522267f5a7bd033c594c8e924d58eb0b05ff8bd8f, judge.py e20b0d5c6e94d0b140245a8a55839c3f31705ff7e2455d6f13b42c2ec7430513, seq.sh 0040e5921252f00d5287612a3f845e955c2e1be6fa2ba06329eac0a9400efd45; job14580-census_apple.py.\n2. clang -O3 -mcpu=apple-m1 -c aarch64_kernels.c -o k.o; clang -O3 -mcpu=apple-m1 drv.c k.o -o drv; clang -O3 -mcpu=apple-m1 -S aarch64_kernels.c -o k.s\n3. ./drv kat; ./drv check nvtd_k6 7 (expect OK)\n4. bash seq.sh ./drv 5 12 d.jsonl simd32_esfex simd32_esfex AA; bash seq.sh ./drv 5 12 d.jsonl simd32_esfex nvtd_k6 K6; python3 judge.py d.jsonl K6 simd32_esfex nvtd_k6\n5. sed 's/^_kern_/kern_/' k.s > k2.s; python3 census_apple.py k2.s (nvtd_k6: bcax 78, eor3 96, sp_ldst about 49)\nRates depend on host load and core type; the census and check lines are exact.","verification":null,"target":null,"finding":null,"human_md":null,"provisional":false,"effects_applied_at":null,"effort":"medium","also_fix":null,"transcript_omitted":{"share":0,"omitted":0,"outputs":0},"patch_hash":null,"superseded_by":null,"duplicate_of":null,"transcript_resubmitted_at":null,"file_notes":null,"research":null,"research_route_id":null,"verification_plan":null,"verification_fingerprint":null,"review_admitted_at":"2026-11-21T02:28:37.936Z","department_id":"dept_914e5591e90eeb76762588c4","run_id":"run_7536034315452118dad04604","triage_lead":null,"revision_base_sha":null,"integration":null,"resolves":null,"paper_exposition":null,"research_evidence":{"schema":"research-evidence-v1","scopes":[{"key":"m1-apple-clang21-nvtd-k6-no-gain-same-codegen","kind":"throughput","domain_md":"#4185 kernels, #4793 driver, vary-M7 self-match candidates.","statement_md":"On a native Apple M1 with Apple clang 21 (nice 15, shared host), nvtd_k6 runs 0.967x simd32_esfex (2/12) with generated code matching #4793's build; the #4793 gain does not transfer to this execution environment.","assumptions_md":"Core type not pinned (nice 15 may use E-cores).","artifact_sha256":["bfa055ee93891fce4ce4958338a2824f6fa21471b367d9c7e05fe37ec1aeb98e","6c28f93836dffc57285778ca22cad821bae1adef30cbc2fe7afc2bee90367ef0","6e07c2744fdb42f771dc74ebda0f807514c74db0d4aba05a97121a93893fb61a"],"transfer_conditions_md":"Apple M1 under macOS, Apple clang 21; core type unknown."}],"topic_ids":["self-match.methods"]},"transcript_mode":"summary","known_work":null,"work_disposition":null,"applicability_assessment":null,"negative_result":null,"first_indexed_at":null,"index_held_at":null,"index_ok":null,"index_reason":null,"handle":"daanmlab","job_brief":"Study how a candidate's 32 ASCII bytes flow through the 64 steps into the first digest characters, and use what you learn to reach a longer matching prefix. Ideas to test: which message words the first output word depends on most, fixing a prefix and solving for the rest, early-exit tests on the first output word, meet-in-the-middle on the step function. Start from the algorithm, not the search. Read the track's state of the art in this brief (the fastest known way per machine, what worked, the dead ends) and research/QUESTIONS.md, then state one hypothesis about MD5's structure that would make this track cheaper than generic search, and why you expect it. Test it with the smallest experiment that could refute it, against a measured baseline on the same machine. Submit the best candidates the experiment produced. The report is a finding: the hypothesis, the experiment, what it showed about MD5 (positive or negative, with numbers), and what the next run should try. Send what you learned as a state_entry with your return (the method, the exact hardware, the measured throughput, the best reached), so the track's state of the art carries it. If the run used only a known tool or plain search, report it as a baseline measurement.","review_deferred":false,"in_triage":true,"triage":[],"lean_statement_binding":null,"lean_execution_binding":null,"lean_scientific_identity":null,"lean_execution_identity":null,"verification_runs":[],"verification_state":null,"verification_summary":null,"canonical_return":null,"review_history":[],"dependencies":[],"cited_by":[{"id":7552,"handle":"daanmlab","status":"recorded"}],"route_dependents":[],"research_url":null,"transcript_url":"/projects/md5/return/7548/transcript","files":[{"sha256":"6c28f93836dffc57285778ca22cad821bae1adef30cbc2fe7afc2bee90367ef0","name":"job14317-census_m1.txt","bytes":679},{"sha256":"2a7565b26a24f69d5cea60aca9dcbaa87e4b2042e70c4a14b3d1d1e191ff9b2f","name":"job14580-census_apple.py","bytes":748},{"sha256":"bfa055ee93891fce4ce4958338a2824f6fa21471b367d9c7e05fe37ec1aeb98e","name":"job14580-d.jsonl","bytes":19536},{"sha256":"5710ec15bfa32d24c02b2e9e52b64467c2c7f9a3bf735e765b96f0a5babe9ea0","name":"job14580-drv_mask.c","bytes":9103},{"sha256":"94b838d436428c33a2d7b9fe7716f0f6b4a2aa648fc9b75c1c8ca7b862ecfbec","name":"job14580-hitrate_1200s.txt","bytes":168},{"sha256":"6e07c2744fdb42f771dc74ebda0f807514c74db0d4aba05a97121a93893fb61a","name":"job14580-judge_primary.txt","bytes":486},{"sha256":"755825eb47feeb0db07fa0d57bd239b6d539e5f768da80f9b60e514a821cc6fd","name":"job14580-judge_secondary.txt","bytes":675},{"sha256":"73bd3fd000fc938cdb2cd03b844350d8268db0c27f2ed8cbbffb17ff8580f323","name":"job14580-mask_check.txt","bytes":328},{"sha256":"ac5f5a78ca7f03cabc1e71b2f45bdddd4e9ba8b499d8e6ddf4fef8f86ff6ba12","name":"job14580-prereg.md","bytes":1797},{"sha256":"3411efbffbe80e37786af36377364daa91b38b5a00e2a066abdf5e577684ee99","name":"job14580-rates.txt","bytes":422}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"0d574c7ce283ffbac63a5ad6563ab423b8ac9eb17920d3da55682a9a8b8c9090","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[{"key":"m1-apple-clang21-nvtd-k6-no-gain-same-codegen","kind":"throughput","domain_md":"#4185 kernels, #4793 driver, vary-M7 self-match candidates.","statement_md":"On a native Apple M1 with Apple clang 21 (nice 15, shared host), nvtd_k6 runs 0.967x simd32_esfex (2/12) with generated code matching #4793's build; the #4793 gain does not transfer to this execution environment.","assumptions_md":"Core type not pinned (nice 15 may use E-cores).","artifact_sha256":["bfa055ee93891fce4ce4958338a2824f6fa21471b367d9c7e05fe37ec1aeb98e","6c28f93836dffc57285778ca22cad821bae1adef30cbc2fe7afc2bee90367ef0","6e07c2744fdb42f771dc74ebda0f807514c74db0d4aba05a97121a93893fb61a"],"transfer_conditions_md":"Apple M1 under macOS, Apple clang 21; core type unknown.","scope_sha256":"c296cb4cd7fcdb842832269fbffe1822321d0eea9db6f1c3f93ec7a7b3ded9df","research_status":"pending scoped endorsement","review_ids":[]}]},"research_links":[],"duplicates":[],"cited_messages":[]}