mirror of
https://github.com/allaunthefox/Research-Stack.git
synced 2026-07-31 03:05:21 +00:00
allaun
475f6319ea
This squashes all local history (768 commits) onto the scrubbed PR #90 baseline. Individual commits were lost during filter-repo corruption; the working tree content is preserved intact. Build: N/A (working tree state only)
57 lines
2.3 KiB
Markdown
57 lines
2.3 KiB
Markdown
# QR Decomposition → 8×8 Braid Bridge
|
||
|
||
**Source:** `c0rmac/qr-apple-silicon` — Metal/AMX QR factorization
|
||
**Relevance:** QR's Householder reflections are structurally isomorphic to braid crossings
|
||
|
||
## The isomorphism
|
||
|
||
| QR decomposition | Braid pipeline | Our module |
|
||
|-----------------|---------------|------------|
|
||
| Householder reflector H = I - τvvᵀ | Braid crossing operator | `BraidEigensolid.lean` |
|
||
| Compact WY representation (b=32) | Blocked braid crossings (b=8) | `crossStep` / sidon addressing |
|
||
| AMX 8×8 simdgroup_matrix tiles | Matrix8 type (8×8 Fin 8) | `AdjugateMatrix.lean` |
|
||
| QR factorization A = QR | Braid factorization of shock | `BurgersPDE.lean` |
|
||
| Thin QR (M×K, K×N) | Dimensional Shock Trim | `DST(M_D) → (A_r, ε)` |
|
||
|
||
## What the AMX tells us about our braid
|
||
|
||
The Apple AMX coprocessor accelerates 8×8 matrix tiles in hardware. This is
|
||
not a coincidence — the 8×8 tile size corresponds to the 8-strand braid
|
||
topology (`Fin 8 → BraidStrand`). Every matrix operation on the AMX is
|
||
a batch of 8-dimensional braid crossings.
|
||
|
||
The `qr-streaming-amx` kernel's pipeline:
|
||
|
||
```
|
||
Panel factorisation (b=32 columns)
|
||
→ T-matrix construction (Compact WY)
|
||
→ Trailing matrix update (grid-parallel, AMX tiles)
|
||
→ Q accumulation (WY update)
|
||
```
|
||
|
||
This maps to our braid pipeline as:
|
||
|
||
```
|
||
Braid panel factorisation (N=8 strands × b=8 crossings)
|
||
→ Sidon T-matrix (sumset collision table)
|
||
→ DualQuaternion trailing update (advection + viscosity)
|
||
→ Receipt accumulation (energy convergence)
|
||
```
|
||
|
||
## Potential integration
|
||
|
||
The QR library provides reference implementations of:
|
||
- **Householder reflection** — the 8×8 Householder matrix is exactly a braid
|
||
crossing operator in matrix form
|
||
- **Compact WY batching** — batches of 32 Householder reflectors = batches of
|
||
32 braid crossings, which is 4× the sidon address budget (128/32)
|
||
- **AMX tile programming** — the Metal shaders show exactly how to program
|
||
8×8 matrix tiles, which is the native hardware format for our braid pipeline
|
||
|
||
## Reference
|
||
|
||
- Golub & Van Loan, Matrix Computations §5.1 — Householder QR
|
||
- Schreiber & Van Loan, 1989 — Compact WY representation
|
||
- `c0rmac/qr-apple-silicon` — Metal/AMX QR implementation
|
||
- `AdjugateMatrix.lean` — 8×8 matrix operations
|
||
- `BraidEigensolid.lean` — braid crossing operators
|