ADR 0072: Cooperative Shared CUBE Inputs and Transpose

Accepted with implementation by pull request 110. It depends on the accepted PTO-CUBE-CELL-STATE-001 and PTO-CUBE-LOCAL-MATRIX-001 contracts.

Decision

Shared Tiles remain ordinary two-dimensional descriptors. They never persist a CUBE CELL layout, ND/DN orientation, transpose state, or per-operation Matrix role.

Cooperative TMATMUL-family operations may bind the right primary operand, or both A and B primary operands, from published Shared Tiles according to the operation's complete-bundle schema. Local C/D and every Local primary operand continue to use the CUBE layout contract. TGEMV remains Local-only and rejects every Shared primary binding.

When only B is Shared, D inherits the CUBE_M16 or CUBE_M32 layout class of Local A. When both A and B are Shared, an ACC form inherits the layout class of Local C. A non-ACC all-Shared form selects CUBE_M16 for 1 <= M <= 16 and CUBE_M32 for 17 <= M <= 32. No additional layout selector is encoded.

Transpose encoding

The existing B.FPATR command assigns two previously reserved bits:

• bit 7 is TransA;
• bit 8 is TransB; and
• bits 9 and 10 remain reserved zero.

The canonical no-transpose command encodes both controls as zero. TransA=1 is legal only when A is Shared. TransB=1 is legal only when B is Shared. A transpose request for a corresponding Local input raises Tile legality before source snapshots, consumption, or destination allocation.

Shared inputs are read through their ordinary logical descriptor. After all four PEs have completed rendezvous and Shared-readiness preflight, the selected transpose is applied to the logical input consumed by CUBE. The transformation does not modify the Shared descriptor, payload, publication state, or lifetime.

PE mask and rendezvous

LB0=M remains the logical row count for one PE. It is not a core-total M. Each participating PE computes M result rows. The group-visible row count is derived from the number and fixed identity of participating PEs.

Any nonzero four-bit PE mask is legal. Multiple set bits are allowed. All four PEs reach the cooperative rendezvous and perform complete Shared source readiness and schema preflight. Only PEs whose bit is set perform the selected computation, allocate Local destinations, and publish output state.

PE mask zero is a strict no-op before descriptor reads, Shared readiness, dimension/layout/type checks, faults, allocation, source effects, or output publication.

Preflight and ordering

The cooperative operation completes these checks before any selected PE snapshots a source or allocates a destination:

• exact Local/Shared operand schema and role ordering;
• B.FPATR presence, field legality, and reserved bits;
• TransA/TransB correspondence to Shared inputs;
• all four PEs' rendezvous participation;
• publication and complete readiness of every referenced Shared Tile;
• selected PEs' Local descriptor, dimension, layout, dtype, capacity, and destination availability; and
• operation-specific auxiliary and post-process schema.

Rejection preserves all Local and Shared state. A successful cooperative read does not create GM ordering and does not define an order among unrelated Shared or Local accesses.

Defaults and protected behavior

• TransA=0 and TransB=0 select no logical transpose.
• Existing B.FPATR post-processing fields and their numeric semantics remain unchanged.
• Fixed PE identities, per-PE TSize, and Shared core-private visibility remain unchanged.
• Partial masks select destination producers; they do not change logical M for an individual selected PE.
• Existing Matrix function numbers and start encodings remain unchanged.

Explicit exclusions

This decision adds no persistent Shared CUBE layout, GM/Shared CUBE conversion, Shared TGEMV form, hidden orientation state, or implicit memory fence.

Acceptance criteria

The accepted ASL, generated documentation, and independent decoded tests prove:

1. B.FPATR 00, 01, 10, and 11 transpose controls and exact logical results;
2. bits 9 and 10 remain decode-reserved;
3. transpose acceptance only for the corresponding Shared input;
4. right-only and two-Shared primary schemas;
5. all single-bit and representative multi-bit nonzero masks;
6. four-PE rendezvous/readiness with selected-PE-only allocation and publish;
7. strict zero-mask no-effect before all legality and readiness checks;
8. unchanged Shared descriptors and payload after successful reads;
9. TGEMV Shared rejection; and
10. rollback on every late cooperative preflight failure.