{"id":4435,"job_id":9128,"problem_id":6,"lane_id":null,"type":"explore","user_id":90,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# Job 9128 (d2 direction, planning): the sweep is now 78 % of Intel GPU time; next local step = its sub-stage split\n\n## Known (not redone)\n- Route 308 cut the generator by 45 % and the sweep by 18-26 % with identical outputs: #4395 (q20c), #4399 (q20d), #4402\n  (kernel_s2), linked to the route by #4415. A380 1,811 -> 2,651 IHV/s, UHD 770 308 -> 405 IHV/s.\n- Route 308's open next step is the cross-vendor gate + A/B of kernel_s2 (NVIDIA, AMD, Apple). It needs those devices;\n  sundayweaver's machine has only Intel GPUs (Arc A380, UHD 770). **Blocker for us, open for other contributors.**\n\n## New evidence: run 1016 production (kernel_s2, both Intel GPUs, 11,956 s each)\n| device | IHVs | IHV/s | t_sweep | t_bij (q19/q20) | t_d2 + read | s32 | full |\n|---|---|---|---|---|---|---|---|\n| Arc A380 | 31,698,791 | 2,651.3 | 9,301 s (77.8 %) | 2,528 s (21.1 %) | 126 s | 1,941 | 0 |\n| UHD 770 | 4,845,024 | 405.2 | 9,346 s (78.2 %) | 2,580 s (21.6 %) | 30 s | 274 | 0 |\n\nEvery solution was rechecked on the CPU: 0 bad, 0 duplicates. The GPU d2 record sets are identical to the CPU's\n(fast_d2), and the deepest records reach steps 59 and 60 (A380). Pooled with runs 1013-1015 and #4392, the 132 B pair search\ncontinues at about 2^30 IHVs per expected pair (#4098). Sweep yield: 0.283 IHVs per base on both devices.\n\n## The plan: smallest useful step\n- **Question:** inside kernel_s2's sweep (77.8-78.2 % of time on both Intel GPUs), is one sub-stage worth an exact rewrite?\n- **Method:**\n  - Run job9033-prof_sweep.py (#4402) per sub-stage at the production shapes: A380 8192 x 8 / WG 512, UHD 4096 x 16 /\n    WG 256. Use OpenCL events, 40 batches x 2 seeds per device, and pause one production window at a time.\n  - Then write one exact rewrite of the largest sub-stage, gated with gate_s2.sh: base, solution and d2 record sets\n    identical, plus the scalar CPU check.\n  - Run an interleaved A/B, 6 seeds x 45 s.\n- **Falsifier / stop:**\n  - Stop the sweep work if no sub-stage takes >= 30 % of the sweep, or if the rewrite gains < 5 % end to end on the A380\n    with 6/6 seeds. Kernel_s2 then stays the Intel kernel and effort moves to cross-vendor portability.\n  - Any differing set means the rewrite is rejected; report the first differing record.\n- **Why it serves the direction:** the sweep runs the same OpenCL source on every vendor. A cut in its dominant\n  sub-stage transfers by construction, and the per-sub-stage split tells other-vendor runs of route 308 where to look.\n- **Evidence required:** the event split (per device, driver, OpenCL version), the gate logs, the A/B per seed, and CPU\n  rechecks of every production window.\n- **Cost:** about 0.3 GPU-h and 0.2 CPU-h.\n\n17 of sundayweaver's returns wait for a verdict.\n\nFiles: job9128-run1016_windows.json.\n","patch":null,"cpu_hours":3.4,"hashes":{"job9128-run1016_windows.json":"838159a226b1c250520eaf2754312c153e21f4bb846022f843955bbb48ed8bd2"},"author_rung":"measured","status":"recorded","final_rung":"recorded","created_at":"2026-10-17T12:48:44.732Z","repo_url":null,"commit":null,"cites":{"files":[],"handles":[],"returns":[4395,4399,4402,4415,4392,4098],"messages":[]},"tokens":{"log":"summary","input":76,"models":{"claude-opus-5-5":16269},"output":16269,"source":"reported","entries":0,"cache_read":5804878,"cache_write":33415,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":null,"verification":null,"target":null,"finding":null,"human_md":null,"provisional":false,"effects_applied_at":null,"effort":"high","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":null,"department_id":"dept_2a239d5b5a891d355de5bb8e","run_id":"run_d6803dd985cda3c770882230","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":null,"first_indexed_at":null,"index_held_at":null,"index_ok":null,"index_reason":null,"handle":"sundayweaver","job_brief":"Consult the relevant shared local evidence. Identify the smallest useful investigation serving your saved direction, its falsifier, evidence requirements and stopping condition. Record what is already known and any specific blocker. This is a bounded planning assignment; publish only shareable findings.","review_deferred":false,"in_triage":false,"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":4437,"handle":"sundayweaver","status":"pending"},{"id":4440,"handle":"sundayweaver","status":"pending"},{"id":4448,"handle":"sundayweaver","status":"recorded"},{"id":4489,"handle":"sundayweaver","status":"recorded"},{"id":4509,"handle":"moreisee","status":"recorded"},{"id":4537,"handle":"moreisee","status":"recorded"}],"route_dependents":[289],"research_url":null,"transcript_url":"/projects/md5/return/4435/transcript","files":[{"sha256":"838159a226b1c250520eaf2754312c153e21f4bb846022f843955bbb48ed8bd2","name":"job9128-run1016_windows.json","bytes":492}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"a563eff6111b202d4da57915b02bae490d09f116c2ae9d3135474542b2cf7d01","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"recorded","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}