Skip to content

week1_report.md

Latest commit

 

History

History
120 lines (101 loc) · 5.73 KB

week1_report.md

File metadata and controls

120 lines (101 loc) · 5.73 KB

Week 1 Report — Birthday Marathon (March 26–29, 2026)

Executive Summary

4 days, ~60 hours of human+AI collaboration. From v1.0.0 release to a comprehensive Z80 arithmetic optimization ecosystem. 1500+ verified sequences, 4 CUDA kernels, 3 image search methods, cross-verified on 5 platforms.

Day-by-Day

Day 1 (March 26) — Release & Foundation

  • v1.0.0 released to GitHub with article "The Z80 Compiler That Never Guesses"
  • 500+ arithmetic sequences (254 mul + 246 div + idioms)
  • 83.6M regalloc shapes, 97.7% infeasibility at 7–15v
  • ISA DSL gpugen: 4 backends (CUDA/Metal/OpenCL/Vulkan), 3 ISA definitions
  • Cross-verified: 5 platforms, 4 APIs, 3 GPU vendors
  • MinZ v0.23.0 ships with our tables

Day 2 (March 27) — Arithmetic Expansion

  • FP16 library, BCD arithmetic (GPU-proven with H-flag)
  • Multi-target approximate search: 13-op pool, 15 targets simultaneously
  • 21-instruction universal pool thesis: 2.7% of ISA generates ALL optimal arithmetic
  • MinZ v0.24.0 (VIR). Frill→CUDA PoC for mul8 search
  • Book: Appendices L(FP) + M(BCD) + N(LUT) + O(meta-analysis)

Day 3 (March 28) — Proofs & Architecture

  • 37.6M enriched shapes with 15 op-aware metrics (78MB compressed)
  • O(1) regalloc: signature = (shape, operation_bag) → hash lookup
  • Z flag write-only proof (exhaustive + induction)
  • SBC A,A trick library: CMOV, ABS, MIN/MAX, gray_decode EXACT
  • Register cost graph: 11 regs, full paths
  • Smart CALL save: 17T avg (was 34T) = 50% reduction
  • SDCC analysis: IX/IY = 6–11%, EXX < 1%

Day 4 (March 28–29) — Regalloc Deep Dive

  • Corpus bias discovery: VIR=79% ≤6v, FatFS=ALL >10v
  • Five-level pipeline: cut-vertex → O(1) → EXX → GPU → Z3
  • Backtracking solver for 7–15v (alternative to Z3)
  • Partition optimizer: ≤18v exhaustive (<2 min), ≤20v (~30 min)
  • Phase diagram: feasibility cliff 95.9% (2v) → 0.9% (6v)

Day 5 (March 29) — Image Search & u32

  • 3 CUDA image generators: dual-layer (cat 4.9%), segmented LFSR (Che 15%), Introspec BB port
  • u32 arithmetic library: 13 operations, SHL32/SHR32 proven optimal
  • Z80 intro attempt (che_intro.asm, 351 bytes)
  • SHA-256 feasibility: ~58ms/block @3.5MHz
  • divmod8 strategy decided: analytical multiply-and-shift

Day 6 (March 29 continued) — Division Breakthrough

  • div8 254/254 complete: v1→v2→v3, avg 154T→135T→79T (−49%)
  • carry_compare trick (GPU-discovered): 5 ops, 26T for all K≥128
  • PRESHIFT trick: (A>>P)×M>>S, dominant for K<128
  • sat_add8: 4 ops, 16T — branchless saturating add masterpiece
  • sign8, sat_sub8, abs16, neg16, min16, max16
  • SHA-256 round decomposition with realistic estimate
  • 4× cross-verified: z80-optimizer, MinZ, MinZ-VIR, MinZ-ABAP
  • MinZ-VIR integrated div8 into IntrinsicTable (commit 8cfba219)

By The Numbers

Metric Start (Day 1) End (Day 6) Delta
Arithmetic sequences 501 1500+ +200%
div8 coverage 0/254 254/254 complete
div8 avg T-states N/A (runtime) 79T vs 80-200T SDCC
mul8 coverage 254/254 254/254 maintained
mul16 coverage 254/254 254/254 maintained
Regalloc shapes 83.6M 83.6M enriched +15 metrics
Branchless idioms 15 25+ +67%
CUDA kernels 2 6 +200%
Image search methods 0 3 new capability
u32 operations 0 13 new capability
Cross-verified systems 3 5+ 4 independent verifiers
Data files shipped 3 12 +300%
Peephole rules 739K 739K + 37M partial ongoing

Key Innovations

  1. carry_compare — GPU-discovered division trick, not in any reference
  2. PRESHIFT division — exploit K's factorization for shorter sequences
  3. 3-level validation — analytical → composite → GPU exhaustive
  4. sat_add8 = OR C — 4-instruction branchless saturating add
  5. Segmented hierarchical LFSR — guaranteed-convergence image search
  6. Subtractive carving — AND NOT layers for face features in image search
  7. ADC HL,rr — Z80's hidden gem for 32-bit arithmetic

Cross-Team Impact

Team Integration
MinZ (compiler) mul8/16, div8 v3, sign/sat, u32 ops, SHA-256 primitives
MinZ-VIR (backend) div8 IntrinsicTable, tryConstDiv/Mod, enriched table reader
MinZ-ABAP (frontend) MIR2 crosscheck, Z3 constant folding
antique-toy (book) Appendix K data, branchless idioms, 3-level methodology

Files Shipped This Week

New Data (12 files)

  • data/div8_optimal.json — 254/254 div8 sequences (v3, avg 79T)
  • data/mod8_optimal.json — 254/254 mod8 sequences
  • data/divmod8_optimal.json — 254/254 divmod8 sequences
  • data/sign_sat_ops.json — sign8 (43T), sat_add8 (16T), sat_sub8 (20T)
  • data/arith16_new.json — abs16, neg16, min16, max16, sign16
  • data/u32_ops.json — 13 u32 operations (DEHL convention)
  • data/sha256_round.json — SHA-256 decomposition (58ms/block)
  • data/mulopt8_clobber.json — 254 mul8 with clobber masks
  • data/mulopt16_complete.json — 254 mul16 sequences
  • data/z80_register_graph.json — 11-register cost model
  • data/enriched_*.enr.zst — 37.6M enriched regalloc shapes (78MB)
  • data/bcd_idioms.json — BCD arithmetic

New CUDA Kernels (4)

  • cuda/prng_hybrid_gpu.cu — dual-layer evolutionary image generator
  • cuda/prng_segmented_search.cu — hierarchical segmented LFSR
  • cuda/bb_search.cu — Introspec BB algorithm port
  • cuda/prng_layered_search.cu — layered LFSR search

Documentation

  • TODO.md — comprehensive roadmap (277 lines, priority matrix)
  • contexts/day[3-6]_wisdom.md — daily knowledge dumps
  • contexts/week1_report.md — this file
  • media/prng_images/README.md — 29-experiment gallery