Skip to main content

ADR 0097: Local and Shared capacity pools with cooperative M-sharding

Context​

Local Tile state belongs to one PE, while Shared Tile state belongs to one Core. Using one combined capacity counter makes one PE's Local allocation consume another PE's capacity and makes a Shared allocation scale with the number of participating PEs. Cooperative Matrix execution also requires one explicit rule that maps per-PE M/N/K dimensions onto the Core-wide Shared-A matrix.

Decision​

Each of the four PEs owns an independent 256 KiB Local capacity pool. The Core owns a separate 256 KiB Shared capacity pool. Local and Shared allocations do not compete for one combined budget.

B.IOT and B.IOS use the same SizeCode byte table. Code zero is the source role, codes 1 through 12 encode 128 B through 256 KiB, and codes 13 through 15 are reserved. A Local B.IOT destination charges the selected size once in each PE named by decoded PEMode. A Shared B.IOS destination charges the complete encoded size once in the Core-wide Shared pool; decoded PEMode controls participation and payload-quarter updates but does not multiply the Shared allocation.

A single Local object or destination is capped at 64 KiB (SizeCode 1..10). The independent Local pool remains 256 KiB per PE, so multiple Local objects may jointly consume that aggregate pool. A Shared parent remains capped at 256 KiB per Core.

B.IOS renames the identifier field to SharedTileID. The field is six bits at instruction bits 25:20 and names the absolute Core-private registers S0..S63. Bits 27:26 are reserved zero. A nonzero reserved bit rejects as Fault_IllegalInstruction before identifier, binder, descriptor, allocation, payload, or memory effects.

For cooperative four-PE TMATMUL with a Shared left primary, LB0, LB1, and LB2 denote per-PE M, N, and K. Shared A has logical shape (4 * LB0) x LB2; Shared B has logical shape LB2 x LB1; PE i consumes Shared-A rows [i * LB0, (i + 1) * LB0) and publishes one Local result of shape LB0 x LB1. The same rule applies to ordinary, ACC, BIAS, and MX forms whenever their left matrix primary is Shared. A Shared-B-only form retains its PE-local A and common Shared B. TGEMV remains Local-only.

Every encoding, capacity, group-shape, descriptor, readiness, alias, and destination check completes before a source is consumed or any output becomes visible. Shared sources remain persistent and the existing complete-output atomicity and restart rules remain in force.

Consequences​

  • Multiple Local destinations may jointly consume one PE's complete 256 KiB pool, while each individual Local destination remains capped at SizeCode 10 (64 KiB); other PEs retain their independent capacity.
  • A Shared SizeCode-12 destination consumes the complete 256 KiB Shared pool.
  • Reducing the Shared namespace changes the accepted B.IOS encoding mask and requires decoder, assembler, disassembler, compiler, model, and test updates.
  • Cooperative Shared-A tests must distinguish the four fixed PE row slices; a broadcast of the first LB0 rows is not conforming behavior.
  • The architecture and encoding ABI are incompatible with prior Shared capacity and identifier semantics. The eventual release name is assigned only when the maintainers combine this decision with the remaining queued architecture changes; no release is created by this implementation.

Supersession​

This decision supersedes the per-PE Shared-capacity, combined-capacity-budget, eight-bit Shared identifier, and cooperative Shared-A broadcast behavior in the active model. The B.IOT/B.IOS SizeCode and PEMode bit positions, PEMode decode table, absolute Shared naming, strict zero-mode no-op, immutable first allocation mask, and precise preflight/rollback rules remain in force.

Amendment for 0.58.4.1​

ADR-0102 owns the Shared whole-parent readiness and single-issuer publication closure. ADR-0106 owns per-PE Shared source range derivation. The independent Local object cap is 64 KiB while the aggregate Local pool remains 256 KiB per PE.

Last updated on by Kevin Zhou