TMATMUL_ACC

1// PTO-INSTRUCTION: {"assembly":["TMATMUL_ACC <bundle operands>"],"block":["BSTART.TMATMUL.ACC AType","B.DATR BType, RMode, Sat (optional; BType defaults to AType)","B.FPATR PreQuantMode, ReluMode, GroupNCode, RowMaxEn, GroupMaxEn, RowMaxInit, MaxAbsEn, TransA, TransB, CScaleEn (exactly one)","B.DIM LB0 M or cooperative group_M (optional, default 1)","B.DIM LB1 N (optional, default 1)","B.DIM LB2 K (optional, default 1)","B.IOS complete right or both matrix operand groups (optional; cooperative mask 1111)","B.IOT ordered Local mathematical sources: C CUBE_M16/M32 accumulator matching A with encoded selector distinct from DstTile, A CUBE_M16/M32 primary, B CUBE_N8 primary, optional U8 CScale CUBE_M32 [M,1]","B.IOT D matching A's CUBE_M16/M32 layout with a distinct encoded destination index, optional RowMaxOut, optional GroupMaxOut destinations","B.IOT/B.IOR postprocess operands selected by B.FPATR","BSTOP or the next BSTART completion boundary"],"catalog_indices":[93],"catalog_records":[{"arguments":[{"operand":"destination0"},{"operand":"source0"},{"operand":"source1"},{"operand":"source2"}],"command_mnemonic":"BSTART.TMATMUL.ACC","contract_status":"reviewed-complete","datr_contract":{"allowed_nonzero_fields":["DataType","RMode","Sat"],"pad_union":"must-zero"},"disposition":"accepted-direct-operation","effect_contract":"TMATMUL_ACC","family":"CUBE","fault_contract":"ExecuteTileInstruction","function":2,"legality_handler":"TileOperandsLegal_TMATMUL_ACC","name":"TMATMUL_ACC","operands":[{"field":"destination0","role":"destination"},{"field":"source0","role":"accumulator"},{"field":"source1","role":"left"},{"field":"source2","role":"right"}],"restart_contract":"CompleteBundleAtWithAcceptedApplicabilityRules","semantic_handler":"TMATMUL_ACC","state_effects":["operand:destination0:destination","operand:source0:accumulator","operand:source1:left","operand:source2:right"]}],"classification":["matrix-and-matrix-vector","matrix-matrix"],"contract":{"block_composition":["BSTART.TMATMUL.ACC AType","B.DATR BType, RMode, Sat (optional; BType defaults to AType)","B.FPATR PreQuantMode, ReluMode, GroupNCode, RowMaxEn, GroupMaxEn, RowMaxInit, MaxAbsEn, TransA, TransB, CScaleEn (exactly one)","B.DIM LB0 M or cooperative group_M (optional, default 1)","B.DIM LB1 N (optional, default 1)","B.DIM LB2 K (optional, default 1)","B.IOS complete right or both matrix operand groups (optional; cooperative mask 1111)","B.IOT ordered Local mathematical sources: C CUBE_M16/M32 accumulator matching A with encoded selector distinct from DstTile, A CUBE_M16/M32 primary, B CUBE_N8 primary, optional U8 CScale CUBE_M32 [M,1]","B.IOT D matching A's CUBE_M16/M32 layout with a distinct encoded destination index, optional RowMaxOut, optional GroupMaxOut destinations","B.IOT/B.IOR postprocess operands selected by B.FPATR","BSTOP or the next BSTART completion boundary"],"canonical_assembly":["TMATMUL_ACC <bundle operands>"],"defaults":["Encoded DataType is always AType. Omitted B.DATR preserves AType as BType, selects RNE, and disables saturation.","Omitted LB0 defaults Local M or cooperative group_M to one; omitted LB1 and LB2 default N and K to one.","Exactly one all-zero B.FPATR selects no conversion, activation, or reduction; B.IOR and auxiliary B.IOT operands exist only when a selected postprocess mode requires them.","Local A uses persistent CUBE_M16 or CUBE_M32, Local B uses persistent CUBE_N8, and D is newly allocated in A's M layout; Local C also uses A's M layout. M, N, and K are arbitrary positive values independent of per-PE TSize.","TransA=0 and TransB=0 select no logical transpose. Each nonzero control is legal only when the corresponding primary is Shared.","C and D are both mandatory. CScaleEn accepts one final U8 CUBE_M32 [M,1] Local mathematical source only with FP32 C; omission defaults CScaleEn to zero. In decoded blocks C's six-bit relative selector must differ from zero-extended DstTile before rename; direct Tile calls require destination TileIndex to differ from accumulator TileIndex."],"encoding_class":"selector-encoded-block-operation","examples":["BSTART.TMATMUL.ACC AType; B.DATR BType, RMode, Sat (optional; BType defaults to AType); B.FPATR PreQuantMode, ReluMode, GroupNCode, RowMaxEn, GroupMaxEn, RowMaxInit, MaxAbsEn, TransA, TransB, CScaleEn (exactly one); B.DIM LB0 M or cooperative group_M (optional, default 1); B.DIM LB1 N (optional, default 1); B.DIM LB2 K (optional, default 1); B.IOS complete right or both matrix operand groups (optional; cooperative mask 1111); B.IOT ordered Local mathematical sources: C CUBE_M16/M32 accumulator matching A with encoded selector distinct from DstTile, A CUBE_M16/M32 primary, B CUBE_N8 primary; B.IOT D matching A's CUBE_M16/M32 layout with a distinct encoded destination index, optional RowMaxOut, optional GroupMaxOut destinations; B.IOT/B.IOR postprocess operands selected by B.FPATR; BSTOP or the next BSTART completion boundary"],"exceptions":["A reserved DataType or fixed-bit mismatch raises Fault_IllegalInstruction before block state changes.","Missing, duplicate, or non-Matrix B.FPATR use raises Fault_BundleControl before allocation or payload effects.","Illegal types, dimensions, masks, binding streams, descriptors, shapes, capacities, aliases, readiness, or postprocess values raise Fault_TileLegality before source snapshots and effects."],"field_contracts":{},"field_zero_meanings":{},"legality":["The carrier selects exactly CUBE Function 2 and TileOperation_TMATMUL_ACC.","Local A uses persistent CUBE_M16 or CUBE_M32, Local B uses persistent CUBE_N8, and D is newly allocated in A's M layout; Local C also uses A's M layout. M, N, and K are arbitrary positive values independent of per-PE TSize.","C and D are both mandatory. CScaleEn accepts one final U8 CUBE_M32 [M,1] Local mathematical source only with FP32 C; omission defaults CScaleEn to zero. In decoded blocks C's six-bit relative selector must differ from zero-extended DstTile before rename; direct Tile calls require destination TileIndex to differ from accumulator TileIndex.","A Shared primary must satisfy hardware-maintained whole-parent readiness and publication before payload access; fixed-quarter allocation or initialization masks are not a prerequisite. Any cooperative Local-A/Shared-B or Shared-A/Shared-B TMATMUL interprets LB0 as Core-total group_M in 1..128; Shared A has shape group_MxK, Shared B has shape KxN, and PE i uses valid_M=clamp(group_M-i*M_per_PE,0,M_per_PE) with M_per_PE 16 for group_M<=64 and 32 for group_M>=65. TransA and TransB apply only to their corresponding Shared primary. Right-only Shared inherits Local A layout; all-Shared ACC inherits C layout; all-Shared non-ACC selects M16 through M=16 and M32 through M=32.","AType and BType must be supported ordinary Matrix types from one numeric class. C is one explicit Local MxN accumulator source and D is a distinct newly published destination; C's encoded relative selector and D's zero-extended DstTile hand must differ before rename. Published Shared operands may replace the right group or both matrix groups; supplementary operands and destinations remain Local.","Every cooperative nonzero PE_MASK must be 1111; all four PEs complete Shared readiness, while zero-row PEs suppress every compute-only Local resolution and effect. Mask zero is a strict no-op before descriptor reads, faults, allocation, readiness checks, or lifetime effects.","B.DATR permits only BType, RMode, and Sat. Exactly one B.FPATR is mandatory and closes the conditional postprocess schema."],"memory_effects":["none"],"operands":[{"field":"destination0","role":"destination"},{"field":"source0","role":"accumulator"},{"field":"source1","role":"left"},{"field":"source2","role":"right"}],"ordering":["Complete schema, field, type, dimension, descriptor, shape, capacity, readiness, alias, and allocation preflight precedes every source snapshot and destination effect.","D and every enabled reduction output publish as one atomic group; rejection publishes none and successful sources persist."],"standalone_opcode":false,"state_effects":["Multiply matrices and accumulate into the supplied accumulator Tile.","After complete preflight, execute TMATMUL_ACC with the operand bindings listed above; destination definedness changes only as specified by that handler.","For Local execution, publish D with A's CUBE_M16 or CUBE_M32 layout and final output dtype; ordinary Bias, MX scales, and enabled reduction auxiliaries keep their operation-owned layouts.","Successful Shared primary reads leave every Shared descriptor, mask, publication state, payload, and lifetime unchanged.","C is snapshotted before multiplication and remains descriptor-and-payload unchanged after success or rejection."]},"depends_on":["PTO-BLOCK-MODEL-SCHEMA-PROFILE-ENCODING"],"engine":"CUBE","id":"PTO-TILE-TMATMUL-ACC","mnemonic":"TMATMUL_ACC","summary":"Multiply matrices and accumulate into the supplied accumulator Tile.","surface":"tile"}2// PTO-REVIEW: {"review_method":"formal-definition-read","outcome":"FORMAL-COMPLETE","reviewed_fields":["assembly","encoding","defaults","operation","state","memory","ordering","faults","reserved"]}3// NDF-BEGIN: PTO-TMATMUL-ACC-CONTRACT-0014// ndf: kind=contract level=L1 layer=tile status=accepted5// TMATMUL_ACC MUST select CUBE Function 2 and TileOperation_TMATMUL_ACC.6// Its encoded DataType is AType; omitted B.DATR MUST preserve AType as7// BType, omitted LB0/LB1/LB2 MUST default M/N/K independently to one,8// Local M, N, and K MUST be positive per-PE dimensions independent of TSize;9// cooperative LB0 MUST be Core-total group_M.10// For every cooperative Shared-input TMATMUL, LB0 MUST be Core-total group_M;11// Shared A MUST have shape group_MxK and Shared B MUST have shape KxN.12// AType and BType must be supported ordinary Matrix types from one numeric13// class. C is one explicit Local MxN accumulator source and D is a newly14// published destination; C and D may use one architectural Tile name.15// Published Shared operands may replace the right group or both matrix groups;16// supplementary operands and destinations remain Local.17// Shared primaries MUST satisfy hardware-maintained whole-parent readiness and publication before payload access; fixed-quarter allocation or initialization masks are not prerequisites.18// TransA/TransB MUST apply only to the corresponding Shared primary. Every19// cooperative nonzero PE mask MUST be 1111; zero-row PEs MUST have no Local effect.20// C and D MUST both be explicit; direct Tile calls MUST also use different21// destination and accumulator TileIndex values. C MUST persist unchanged.22// Exactly one B.FPATR MUST close postprocess defaults and operands.23// Complete schema, type, shape, capacity, definedness, readiness, alias,24// and allocation preflight MUST precede source snapshots and effects;25// successful outputs MUST publish atomically and sources MUST persist.26// NDF-END: PTO-TMATMUL-ACC-CONTRACT-00127// DOC-BEGIN: decode28readonly func InstructionContractOperation_TMATMUL_ACC()29    => TileOperation30begin31    return TileOperation_TMATMUL_ACC;32end;33// DOC-END: decode34// DOC-BEGIN: operation35readonly func InstructionContractCubeFunction_TMATMUL_ACC()36    => integer {0..31}37begin38    return 2;39end;40
41readonly func InstructionContractSharedOperandsAllowed_TMATMUL_ACC()42    => boolean43begin44    return TRUE;45end;46
47readonly func InstructionContractOperandsLegal_TMATMUL_ACC(48    destination: TileIndex,49    accumulator: TileIndex,50    left: TileIndex,51    right: TileIndex) => boolean52begin53    return TileOperandsLegal_TMATMUL_ACC(54        destination,55        accumulator,56        left,57        right);58end;59
60readonly func InstructionContractHandler_TMATMUL_ACC()61    => TileSemanticHandler62begin63    return TileHandler_TMATMUL_ACC;64end;65// DOC-END: operation66