TLOAD

assembly

[
  "TLOAD <bundle operands>"
]

block

[
  "# Local destination",
  "BSTART.TLOAD DataType",
  "optional B.DATR Layout",
  "B.DIM LB0=ValidCol, LB1=ValidRow, LB2=Col",
  "optional B.IOR RegSrc0=base, RegSrc1=row_stride_bytes",
  "one terminating destination B.IOT",
  "BSTOP",
  "# Shared destination",
  "replace B.IOT with one destination B.IOS"
]

catalog_indices

[
  81
]

catalog_records

[
  {
    "arguments": [
      {
        "operand": "destination0"
      },
      {
        "operand": "address"
      },
      {
        "operand": "scalar0"
      }
    ],
    "command_mnemonic": "BSTART.TLOAD",
    "contract_status": "reviewed-complete",
    "datr_contract": {
      "allowed_nonzero_fields": [
        "PadValueOrByteId",
        "Layout"
      ],
      "pad_union": "pad-value"
    },
    "disposition": "accepted-direct-operation",
    "effect_contract": "TLOAD",
    "family": "TLSU",
    "fault_contract": "ExecuteTileInstruction",
    "function": 0,
    "legality_handler": "TileOperandsLegal_TLOAD",
    "name": "TLOAD",
    "operands": [
      {
        "field": "destination0",
        "role": "destination"
      },
      {
        "field": "address",
        "role": "base-address"
      },
      {
        "field": "scalar0",
        "role": "row-stride-bytes"
      }
    ],
    "restart_contract": "CompleteBundleAtWithAcceptedApplicabilityRules",
    "semantic_handler": "TLOAD",
    "state_effects": [
      "operand:destination0:destination",
      "operand:address:base-address",
      "operand:scalar0:row-stride-bytes"
    ]
  }
]

classification

[
  "memory-and-data-movement",
  "regular"
]

contract

{
  "block_composition": [
    "Local: BSTART.TLOAD DataType; optional B.DATR Layout; B.DIM defines ValidCol, ValidRow, and physical Col; optional B.IOR defines the per-PE GM base and byte row stride; one terminating destination B.IOT allocates the result; BSTOP commits.",
    "Shared: replace B.IOT with one destination B.IOS carrying absolute S0..S63, SizeCode, and PE_MASK. One issuer loads the complete parent; multiple issuers require B.ASSEMBLE with explicit writer ranges."
  ],
  "canonical_assembly": [
    "TLOAD <bundle operands>"
  ],
  "defaults": [
    "DataType is explicit in BSTART.TLOAD. Omitted B.DATR selects ordinary NORM layout. Explicit CUBE Layout 21 through 23 requires DTYPE_NONE and consumes PadValue for physical CELL tails.",
    "LB0/ValidCol and LB1/ValidRow default through the common destination-shape rules. Omitted LB2/Col defaults to ValidCol. Rows are derived from TSize, Col, and DataType and must contain ValidRow.",
    "Omitted B.IOR supplies base zero. Ordinary forms use resolved Col and CUBE forms use LB0 valid columns to derive dense byte row stride as ceil(columns * element_bits / 8). An encoded zero GPR selector is present and reads zero, so an explicitly encoded zero stride aliases rows rather than selecting the omission default."
  ],
  "encoding_class": "selector-encoded-block-operation",
  "examples": [
    "BSTART.TLOAD U8; B.DIM LB0, 64; B.DIM LB1, 8; B.DIM LB2, 64; B.IOR zero, a0; B.IOT mask=1111, ->T<1>; BSTOP",
    "BSTART.TLOAD FP16; B.DIM LB0, 32; B.DIM LB1, 4; B.IOS mask=0001, ->S7<1>; BSTOP",
    "BSTART.TLOAD FP16; B.DATR {ND2N8, DTYPE_NONE, Max, EQ, Default, 0, 0}; B.DIM LB0=N; B.DIM LB1=K; B.IOT mask=1111, <last>, ->N<3>; BSTOP"
  ],
  "exceptions": [
    "Reserved DataType, unsupported or wrong-direction Layout, operation-inapplicable PadValue, malformed B.IOR/B.IOT/B.IOS schema, invalid dimensions, capacity or shape overflow, allocation failure, or GM translation, permission, or alignment fault rejects before destination publication.",
    "Every selected memory address is preflighted before any destination payload, descriptor, allocation, definedness, or load event becomes visible. A failed Local allocation is rolled back; a failed Shared update preserves the prior Shared record."
  ],
  "field_contracts": {},
  "field_zero_meanings": {},
  "legality": [
    "TLOAD is selected only by TLSU Function 0 through BSTART.TLOAD; it has no standalone opcode.",
    "The completed block has exactly one destination domain: one terminating destination B.IOT for Local or one destination B.IOS for Shared. It has no Tile source and consumes at most one B.IOR.",
    "The BSTART DataType accepts every assigned Tile DataType code and rejects 15, 21..23, and 29..31 before effects. Ordinary and Shared forms permit only Layout and require PadValue zero; Local CUBE codes 21 through 23 additionally permit all four PadValue encodings and require DTYPE_NONE while CMode, RMode, Sat, and Canonicalize retain zero meanings.",
    "ValidCol and ValidRow are nonzero, ValidCol does not exceed physical Col, and the derived Rows and Col are powers of two large enough for the valid rectangle.",
    "PE_MASK=0000 is a strict no-op before GPR reads, allocation, memory access, faults, load events, descriptor changes, or payload changes.",
    "Ordinary forms require nonzero ValidCol and ValidRow, ValidCol not greater than physical Col, and power-of-two physical Rows and Col. CUBE forms require explicit nonzero LB0/LB1, absent LB2, and derive CELL geometry from Layout, dtype, and valid shape."
  ],
  "memory_effects": [
    "For each selected PE and each element in ValidRow x ValidCol, read GM at base + row * row_stride_bytes + column * element_size. Packed four-bit types add floor(column / 2) to each byte-strided row base and select the nibble from column parity.",
    "The accesses participate in PTO-TSO using the block aq/rl attributes and are precise and restartable."
  ],
  "operands": [
    {
      "field": "destination0",
      "role": "new Local destination or absolute Shared destination"
    },
    {
      "field": "address",
      "role": "per-PE private-GPR GM base address"
    },
    {
      "field": "scalar0",
      "role": "per-PE private-GPR byte row stride"
    }
  ],
  "ordering": [
    "Resolve the complete schema, selected PE mask, per-PE GPR inputs, dimensions, destination capacity, and all memory translations before the first architectural load effect.",
    "On success publish the complete Local destination or complete Shared parent atomically at block commit. A multi-PE Shared producer publishes only after complete B.ASSEMBLE.LAST; failure preserves prior state."
  ],
  "standalone_opcode": false,
  "state_effects": [
    "A successful Local form allocates or renames exactly one destination Tile, installs the derived descriptor, loads every selected valid element, and marks the valid region defined.",
    "A successful singleton Shared form loads and publishes the complete parent from that issuer PE. A multi-PE Shared form uses B.ASSEMBLE explicit ranges; TLOAD never modifies source GPRs or GM.",
    "A successful CUBE form writes raw valid values through CUBE storage indices and applies Zero, Max, Min, or undefined Null to physical tail positions without counting tails as valid elements."
  ]
}

depends_on

[
  "PTO-BLOCK-BSTART-TLOAD",
  "PTO-BLOCK-MODEL-SCHEMA-PROFILE-ENCODING"
]

engine

TLSU

id

PTO-TILE-TLOAD

mnemonic

TLOAD

summary

Load one ordinary Local or Shared rectangle, or explicitly convert one GM rectangle into persistent Local CUBE storage.

surface

tile

{
  "classification": [
    "memory-and-data-movement",
    "regular"
  ],
  "documentation": "docs/tile/memory-and-data-movement/regular/TLOAD.md",
  "id": "PTO-TILE-TLOAD",
  "instruction_contract": {
    "artifact": "spec/evidence/instruction-contract-closure.json",
    "mnemonic": "TLOAD",
    "ndf_clause": "PTO-INST-TILE-TLOAD"
  },
  "mnemonic": "TLOAD",
  "readiness_subjects": [
    "ADR-0020",
    "ADR-0033",
    "ADR-0052",
    "ADR-0056",
    "ADR-0057",
    "ADR-0059",
    "ADR-0070",
    "ADR-0074",
    "ADR-0078",
    "ADR-0096",
    "ADR-0108"
  ],
  "semantic_tests": [
    "PTO-AVS-TILE-TLOAD-DIRECT-001",
    "PTO-AVS-TILE-TLOAD-SHARED-PARENT-002"
  ],
  "source": "asl/tile/memory-and-data-movement/regular/TLOAD.asl",
  "surface": "tile",
  "tests": [
    "PTO-AVS-TILE-TLOAD-DIRECT-001",
    "PTO-AVS-TILE-TLOAD-SHARED-PARENT-002",
    "PTO-AVS-TILE-TLOAD-STATIC-001"
  ]
}