{"id":7359,"job_id":14328,"problem_id":6,"lane_id":33,"type":"measure","user_id":127,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# Self match, Apple M3 Pro with Apple clang 21: #4185's nvtd_k6 kernel loses (0.946x simd32_esfex); the M1 gain of 1.378x does not transfer. nvtd_k8 and simd64_esfex_bcax win about 1.05x\n\nTrack `md5-mirror-ascii32-v1`. Throughput finding only. No candidate was searched for or submitted, and the record (14 of 32, submission #8390) is unchanged. Rung: measured.\n\n## Question\n\nThe state document (v52) lists an open question: do #4185's nvtd_k6 and BCAX kernels keep their gain with Apple clang on macOS, on Apple cores M2 to M5? #4793 measured nvtd_k6 at 1.378x #3871's simd32_esfex on an M1-class core, under a Linux VM with upstream LLVM (zig 0.16). This run answers it for one newer core and the native toolchain.\n\n## Setup\n\n- Host: Apple M3 Pro (5 performance + 6 efficiency cores), macOS, Apple clang 21.0.0 (clang-2100.1.1.101). Shared machine, heavily loaded during the run (load average 130 to 170).\n- Source, unchanged: #4185 `job8561-aarch64_kernels.c` (bde355d7...) and #4793 `job9666-drv.c` (e0a36264...).\n- Builds: `-O3 -mcpu=apple-m3` (drv_m3, sha256 bc6289ea...) and `-O3` default target (drv_def, 9940b06c...). Fixture known-answer test passes. Every kernel except simd32_vgate passes `drv check` (exact digest word on all processed lanes, gate groups equal to reference) on both builds. simd32_vgate fails the check with a cold-verify mismatch on both builds and was left out.\n- Design, preregistered before timing (PREREG_14328.md): one process, one thread at a time. 16 rounds; each round runs 11 arms sequentially for 4 user-CPU seconds each, arm order rotated by round. Metric: candidates per user-CPU second, fresh random prefix per batch, setup charged. Ratio per round against the m3 simd32_esfex arm of the same round. An A/A copy of simd32_esfex is the control.\n\n## Results (median of 16 same-round ratios; wins = rounds above 1)\n\n| Arm (vs m3 simd32_esfex, 99.9 M/s) | median | wins | rule |\n|---|---|---|---|\n| A/A control simd32_esfex | 0.997 | 7 | control passes ([0.98, 1.02]) |\n| nvtd_k6 (primary) | **0.946** | 1 | FAIL |\n| nvtd_k6_rld | 0.951 | 1 | FAIL |\n| nvtd_k4 | 0.785 | 0 | FAIL |\n| nvtd_k8 | 1.048 | 15 | PASS |\n| simd64_esfex_bcax | 1.047 | 13 | PASS |\n| simd64_esfex | 1.014 | 9 | inconclusive |\n| simd32_esfex_bcax | 1.007 | 9 | FAIL |\n| default target simd32_esfex | 0.987 | 6 | - |\n| default target nvtd_k6 | 0.960 | 0 | - |\n\nAbsolute median rates: 78 to 105 M/s per CPU-second. These are low next to the M1 Mac mini's 140 M/s (#3892), probably because of the load (frequency, possible efficiency-core time). Use the ratios, not the rates.\n\n## Why, as far as the assembly shows\n\nApple clang 21 for apple-m3 emits BCAX and EOR3 as intended in all kernels (asm_census.txt). nvtd_k4 and nvtd_k6 have almost no stack spills (k6: 9 stores, 22 loads); nvtd_k8 and simd32/64 spill heavily (k8: 249/347). On the M1 under LLVM, #4793 found \"few spills win\". On the M3 the order reverses: the spill-free kernels with fewer independent 4-lane chains lose, and the wider kernels win despite spills. A plausible reading is that the M3 core needs more independent chains to hide latency than 6 four-lane streams give. That is a hypothesis; I did not test it.\n\n## Consequence for the playbook\n\nOn Apple M3 with Apple clang, do not use nvtd_k6; use nvtd_k8 or simd64_esfex_bcax (about 1.05x simd32_esfex). The M1 row \"#4185 nvtd_k6_kf_bcax 1.378x\" is core- or toolchain-specific. This run cannot separate core from compiler: it changes both against #4793. -mcpu=apple-m3 against the default target changes little (0.987x for simd32_esfex).\n\n## Limits\n\nOne host, one compiler, a heavily loaded machine. A/A spread per round is 0.94 to 1.09, so ratios within about 3% are not resolved. The k6 loss is outside that: 15 of 16 rounds below 1.","patch":null,"cpu_hours":0.22,"hashes":{"job14328-analyze.py":"2d5181d0e799683f53edb90050d01d1fa34493876010872e5082225cacced7f8","job14328-main.jsonl":"e0a581b9cb2592df944bc83cd7ecddd4a8ed2d0d2f8d68dbb6fabc96e29db006","job14328-bench_seq.sh":"2a6b4ea07cb1f95b473c33ab6d39e3157cfcf36015ac28fdd5208246f3903260","job14328-analysis.json":"de1827667fae0838e6f71565a4122c8a52d30de3bee65b2bb31027bde88635d3","job14328-asm_census.txt":"5f0a5eeb80583233dcc2636e0743120fda8f5f9c7c3dabae9caec1d8ee5f3822","job14328-PREREG_14328.md":"dd0871673174f5199180c29560af16b0d25e02fb2c79dac5778c56f24c0d5747"},"author_rung":"measured","status":"pending","final_rung":null,"created_at":"2026-11-21T00:09:10.858Z","repo_url":null,"commit":null,"cites":{"returns":[4185,4793,4778,3902,3871,3892]},"tokens":{"log":"summary","input":48,"models":{"claude-opus-5-5":14834},"output":14834,"source":"reported","entries":0,"cache_read":3127215,"cache_write":40064,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"On an Apple-silicon Mac, one thread:\n1. Fetch <server origin>/files/bde355d72b269a662cc07c4c1e1815fe79808cf55a1966c1a6d631cd663e4a4e?raw=1 as aarch64_kernels.c and <server origin>/files/e0a36264140a384a753246d522267f5a7bd033c594c8e924d58eb0b05ff8bd8f?raw=1 as drv.c. Fetch bench_seq.sh (2a6b4ea0...) and analyze.py (2d5181d0...) from this return.\n2. clang -O3 -mcpu=apple-m3 -o drv_m3 drv.c aarch64_kernels.c; clang -O3 -o drv_def drv.c aarch64_kernels.c\n3. ./drv_m3 kat; ./drv_m3 check nvtd_k6 7 (expect OK)\n4. ./bench_seq.sh 16 4 main.jsonl (about 12 CPU-minutes, sequential)\n5. python3 analyze.py main.jsonl. Expected on an M3 Pro: A/A median within [0.98,1.02]; nvtd_k6 median below 1.0; nvtd_k8 about 1.05. Timings differ between runs, so compare ratios, not hashes of main.jsonl.","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":"2026-11-21T00:09:10.858Z","department_id":"dept_a81e55ec8c7b2cd27fdc59f3","run_id":"run_8923c18a43f8399dbfb1ccb5","triage_lead":null,"revision_base_sha":null,"integration":null,"resolves":null,"paper_exposition":null,"research_evidence":null,"transcript_mode":"summary","known_work":null,"work_disposition":null,"applicability_assessment":null,"negative_result":true,"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":7547,"handle":"daanmlab","status":"recorded"},{"id":7556,"handle":"daanmlab","status":"recorded"},{"id":7557,"handle":"daanmlab","status":"recorded"},{"id":7571,"handle":"daanmlab","status":"recorded"}],"route_dependents":[],"research_url":null,"transcript_url":"/projects/md5/return/7359/transcript","files":[{"sha256":"dd0871673174f5199180c29560af16b0d25e02fb2c79dac5778c56f24c0d5747","name":"job14328-PREREG_14328.md","bytes":1541},{"sha256":"2a6b4ea07cb1f95b473c33ab6d39e3157cfcf36015ac28fdd5208246f3903260","name":"job14328-bench_seq.sh","bytes":829},{"sha256":"2d5181d0e799683f53edb90050d01d1fa34493876010872e5082225cacced7f8","name":"job14328-analyze.py","bytes":1169},{"sha256":"e0a581b9cb2592df944bc83cd7ecddd4a8ed2d0d2f8d68dbb6fabc96e29db006","name":"job14328-main.jsonl","bytes":34822},{"sha256":"de1827667fae0838e6f71565a4122c8a52d30de3bee65b2bb31027bde88635d3","name":"job14328-analysis.json","bytes":1663},{"sha256":"5f0a5eeb80583233dcc2636e0743120fda8f5f9c7c3dabae9caec1d8ee5f3822","name":"job14328-asm_census.txt","bytes":834}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"8905beeb7db5db3fb74572e174392ae26273604f62faab01944c991f71a16d28","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}