Shuffle raw 32-bit words across Local CUBE rows with an explicit control GPR.
PTO-TILE-TSHUFTSHUF <LB0, LB1, LB2, DataType, Layout> SrcTile0, SrcTile1, Scalar0 -> DstTile| Group | Operand | Role | ASL source |
|---|---|---|---|
| Parameter | LB0 | dimension | B.DIM LB0 (optional) |
| Parameter | LB1 | dimension | B.DIM LB0/LB1/LB2 (optional) |
| Parameter | LB2 | dimension | B.DIM LB0/LB1/LB2 (optional) |
| Parameter | DataType | data type | BSTART.SFU TSHUF, DataType |
| Parameter | Layout | layout | B.DATR Layout (optional) |
| Input | SrcTile0 | source | contract.operands[1].source0 |
| Input | SrcTile1 | controls | contract.operands[2].source1 |
| Input | Scalar0 | shuffle-control | contract.operands[3].scalar0 |
| Output | DstTile | destination | contract.operands[0].destination0 |
BSTART.SFU TSHUF, DataType
BSTART.SFU TSHUF, DataTypeRequiredB.DATR Layout (optional)OptionalB.DIM LB0 (optional)OptionalB.IOT source, controls, ->destinationRequiredB.IOR shuffle_controlRequiredBSTOPRequiredBSTART.SFU TSHUF, DataTypeB.IOT source, controls, ->destinationB.IOR shuffle_controlBSTOPBSTART.SFU TSHUF, U32; B.DATR Layout; B.IOT source, controls, ->destination; B.IOR a0; BSTOP| Field | Bits | Signedness | Architectural role | Encoded zero |
|---|---|---|---|---|
DataType | 5 | encoding-defined | — | — |
Mode | 2 | encoding-defined | — | — |
Function | 5 | encoding-defined | — | — |
destination0 | — | — | destination | — |
source0 | — | — | source | — |
source1 | — | — | controls | — |
scalar0 | — | — | shuffle-control | — |
Loading WaveDrom diagram… The encoding table and WaveJSON remain available below.
| Decoded item | Bit range | Value |
|---|---|---|
| DataType | 31:27 | variable |
| Mode | 26:25 | variable |
| Function | 24:20 | variable |
| Constant | 19:0 | 20'b00011001000110000001 |
{
"reg": [
{
"bits": 20,
"name": "20'b00011001000110000001"
},
{
"bits": 5,
"name": "Function"
},
{
"bits": 2,
"name": "Mode"
},
{
"bits": 5,
"name": "DataType"
}
],
"config": {
"bits": 32,
"fontsize": 13,
"hspace": 900,
"lanes": 1,
"offset": 0
}
}BSTART.TEPL Mode, Function, DataType
destination0source0source1scalar0This Operation comes directly from the instruction owner; the page does not rewrite its behavior.
readonly func InstructionContractOperation_TSHUF() => TileOperationbegin return TileOperation_TSHUF;end;readonly func InstructionContractHandler_TSHUF() => TileSemanticHandlerbegin return TileHandler_TSHUF;end;
pure func InstructionContractDataTypeLegal_TSHUF( data_type: TileDataType) => booleanbegin return TileCubeDataTypeSupported(data_type) && TileElementBits(data_type) != 64;end;
readonly func InstructionContractOperandsLegal_TSHUF( destination: TileIndex, source: TileIndex, controls: TileIndex, control: Word) => booleanbegin return TileOperandsLegal_TSHUF(destination, source, controls, control);end;
func InstructionContractExecute_TSHUF( destination: TileIndex, source: TileIndex, controls: TileIndex, control: Word)begin assert InstructionContractOperandsLegal_TSHUF( destination, source, controls, control); TSHUF(destination, source, controls, control);end;// PTO-INSTRUCTION: {"assembly":["TSHUF <bundle operands>"],"block":["BSTART.SFU TSHUF, DataType","B.DATR Layout (optional)","B.DIM LB0 (optional)","B.IOT source, controls, ->destination","B.IOR shuffle_control","BSTOP"],"catalog_indices":[104],"catalog_records":[{"arguments":[{"operand":"destination0"},{"operand":"source0"},{"operand":"source1"},{"operand":"scalar0"}],"command_mnemonic":"BSTART.TEPL","contract_status":"reviewed-complete","datr_contract":{"allowed_nonzero_fields":["Layout"],"pad_union":"must-zero"},"disposition":"accepted-direct-operation","effect_contract":"TSHUF","family":"TEPL","fault_contract":"ExecuteTileInstruction","function":22,"legality_handler":"TileOperandsLegal_TSHUF","mode":3,"name":"TSHUF","operands":[{"field":"destination0","role":"destination"},{"field":"source0","role":"source"},{"field":"source1","role":"controls"},{"field":"scalar0","role":"shuffle-control"}],"restart_contract":"CompleteBundleAtWithAcceptedApplicabilityRules","selector":"0x076","semantic_handler":"TSHUF","state_effects":["operand:destination0:destination","operand:source0:source","operand:source1:controls","operand:scalar0:shuffle-control"]}],"classification":["layout-and-rearrangement","layout"],"contract":{"block_composition":["BSTART.SFU TSHUF, DataType","B.DATR Layout (optional)","B.DIM LB0/LB1/LB2 (optional)","B.IOT source, controls, ->destination","B.IOR shuffle_control","BSTOP"],"canonical_assembly":["TSHUF <bundle operands>"],"defaults":["A nonzero PE mask requires exactly one B.IOR control input; RegSrc1, RegSrc2, and RegDst are zero."],"encoding_class":"selector-encoded-block-operation","examples":["BSTART.SFU TSHUF, U32; B.DATR Layout; B.IOT source, controls, ->destination; B.IOR a0; BSTOP"],"exceptions":["Reserved control encodings reject with Fault_TileLegality before effects.","CompleteBundleAtWithAcceptedApplicabilityRules supplies restart and completion behavior."],"field_contracts":{},"field_zero_meanings":{},"legality":["TSHUF accepts Local CUBE_M16 or CUBE_M32 data and U32 control Tiles with matching geometry.","The control word selects UP, DOWN, BFLY, or IDX; segment and boundary fields are checked before execution.","Raw 32-bit words are shuffled without byte permutation."],"memory_effects":["none"],"operands":[{"field":"destination0","role":"destination"},{"field":"source0","role":"source"},{"field":"source1","role":"controls"},{"field":"scalar0","role":"shuffle-control"}],"ordering":["Source and control snapshots precede destination publication."],"standalone_opcode":false,"state_effects":["Perform independent PTX-style word shuffles for each active CUBE row/group."]},"depends_on":["PTO-TILE-MODEL-EXECUTION-REARRANGEMENT","PTO-TILE-MODEL-LEGALITY-LAYOUT-REARRANGEMENT"],"engine":"SFU","id":"PTO-TILE-TSHUF","mnemonic":"TSHUF","summary":"Shuffle raw 32-bit words across Local CUBE rows with an explicit control GPR.","surface":"tile"}// PTO-REVIEW: {"review_method":"formal-definition-read","outcome":"FORMAL-COMPLETE","reviewed_fields":["assembly","encoding","defaults","operation","state","memory","ordering","faults","reserved"]}// NDF-BEGIN: PTO-TSHUF-CONTRACT-001// ndf: kind=contract level=L1 layer=tile status=accepted// TSHUF selects raw 32-bit words inside independent power-of-two row segments.// Control mode, segment, and boundary fields are checked before publication;// SELF and ZERO define out-of-segment behavior without predicates.// NDF-END: PTO-TSHUF-CONTRACT-001// DOC-BEGIN: decodereadonly func InstructionContractOperation_TSHUF() => TileOperationbegin return TileOperation_TSHUF;end;// DOC-END: decode// DOC-BEGIN: operationreadonly func InstructionContractHandler_TSHUF() => TileSemanticHandlerbegin return TileHandler_TSHUF;end;
pure func InstructionContractDataTypeLegal_TSHUF( data_type: TileDataType) => booleanbegin return TileCubeDataTypeSupported(data_type) && TileElementBits(data_type) != 64;end;
readonly func InstructionContractOperandsLegal_TSHUF( destination: TileIndex, source: TileIndex, controls: TileIndex, control: Word) => booleanbegin return TileOperandsLegal_TSHUF(destination, source, controls, control);end;
func InstructionContractExecute_TSHUF( destination: TileIndex, source: TileIndex, controls: TileIndex, control: Word)begin assert InstructionContractOperandsLegal_TSHUF( destination, source, controls, control); TSHUF(destination, source, controls, control);end;// DOC-END: operation
Why use it. TSHUF moves raw 32-bit word groups among rows inside independent power-of-two CUBE segments, avoiding a separate predicate or numerical conversion step.
How it works. The scalar control selects UP, DOWN, BFLY, or IDX; each U32 control Tile word supplies its row operand in bits [4:0], while bits [31:5] are accepted and ignored; the boundary choice keeps the current word (SELF) or writes zero (ZERO) when the selected row is unavailable.
Inputs and result. source and the fresh destination share a supported non-64-bit dtype, Local CUBE_M16 or CUBE_M32 layout, and geometry; the Local U32 controls Tile matches their layout, valid rows, and CUBE CELL count, while its valid columns equal the number of 32-bit word groups in each data row.
Effects. Source and control Tiles are snapshotted before the complete valid destination region is published; both inputs persist, destination padding is Null, and the operation has no memory effect.
What is rejected. Scalar controls with mode > 3, segment_code > 4, boundary > 1, or nonzero bits 63:32 reject, as do unsupported dtype or layout, undefined input data, mismatched geometry, and destination aliasing; segment width 32 is available only with CUBE_M32.
Concrete example. In BFLY mode with segment width 16 and per-row operand 1, source rows [1, 2, 3, 4] publish [2, 1, 4, 3] for the corresponding word group.
Bodies come from owning ASL. Dragging or buttons change only this page-session view order.
TSHUF selects raw 32-bit words inside independent power-of-two row segments. Control mode, segment, and boundary fields are checked before publication; SELF and ZERO define out-of-segment behavior without predicates.
PTO-TSHUF-CONTRACT-001asl/tile/layout-and-rearrangement/layout/TSHUF.asl33ff7ddba75203adb60983616329e8e294b786b70714ea622c0aee936045855cbefc6c4d2776fb9147088996b72eb628639b91725d01acb806545d756e5b79a410 matching entries
PTO-AVS-BLOCK-CELL-REARRANGE-DECODED-001tests/asl/block/model/dispatch/cell-rearrangement-schema/block-exec-cell-rearrangement-decoded-001.asl4991c63e3d0381992b52d28493920d050c24584b2ab3c3dd1c4cfced9032ff1fPTO-AVS-TILE-TSHUF-BFLY-001tests/asl/tile/layout-and-rearrangement/layout/TSHUF/tile-exec-tshuf-bfly-001.asl148eb832de230512d9a9e2a72c2a749ec0335db251b88cd98cfa6ea61373af24PTO-AVS-TILE-TSHUF-STATIC-001tests/asl/tile/layout-and-rearrangement/layout/TSHUF/tile-static-tshuf-contract-001.asl445a8e6ad22ef7f0cf59785a7fc34aad69588db5e61bcf53be27cc41ffa49f0aPTO-EVIDENCE-RELEASE-TRACEABILITYPTO-EVIDENCE-RELEASE-TRACEABILITYspec/evidence/release-traceability-readiness.jsonc7327021d39dc67ac5564bc55073b3870a397d79ac8d9648284d56e33bc14a3ePTO-EVIDENCE-INSTRUCTION-CONTRACT-CLOSUREPTO-EVIDENCE-INSTRUCTION-CONTRACT-CLOSUREspec/evidence/instruction-contract-closure.json3ef2bb62421c79dff8fa77a1c7983923b523244b8090812883ef81286ca8106aPTO-EVIDENCE-ARCHITECTURE-READINESSPTO-EVIDENCE-ARCHITECTURE-READINESSspec/evidence/architecture-readiness.json4b0b85199101251bea744e0f3591cc31906909dc80d5ab651c417a936036a004PTO-EVIDENCE-RELEASE-GATE-READINESSPTO-EVIDENCE-RELEASE-GATE-READINESSspec/evidence/release-gate-readiness.jsona0f4d2b6920c08981ea55fd8ef820708a40d4feb5402c5150e8e9ab532d84ce0PTO-EVIDENCE-RELEASE-MANIFESTPTO-EVIDENCE-RELEASE-MANIFESTspec/release-manifest.json1a64c109ed7a90351c41e2a418b3c0ebaf8ad975838986d2101385186b85c0d8Loading ADR-0103…
ADR-0103docs/status/decisions/0103-private-cube-vector-and-cell-rearrangement.md3b3d67683fb0987dc85ba20c2d2effef9e7e1b170d3950cb505cc1f3ab6109d7Loading ADR-0108…
ADR-0108docs/status/decisions/0108-0584-1-to-0585-compatibility.md5689fa78a7291b6ae88d55197205e4cf79e9719ee5ba18ebf186f5c6871c350d{
"classification": [
"layout-and-rearrangement",
"layout"
],
"documentation": "docs/tile/layout-and-rearrangement/layout/TSHUF.md",
"id": "PTO-TILE-TSHUF",
"instruction_contract": {
"artifact": "spec/evidence/instruction-contract-closure.json",
"mnemonic": "TSHUF",
"ndf_clause": "PTO-INST-TILE-TSHUF"
},
"mnemonic": "TSHUF",
"readiness_subjects": [
"ADR-0103",
"ADR-0108"
],
"semantic_tests": [
"PTO-AVS-TILE-TSHUF-BFLY-001"
],
"source": "asl/tile/layout-and-rearrangement/layout/TSHUF.asl",
"surface": "tile",
"tests": [
"PTO-AVS-TILE-TSHUF-BFLY-001",
"PTO-AVS-TILE-TSHUF-STATIC-001"
]
}0.58.5 · Release candidate7dc8b7e5b121d2b2499a2273bebff29e2cd8681233ff7ddba75203adb60983616329e8e294b786b70714ea622c0aee936045855c65b8360bcc35b692142dc17a5f5320aed09a1656340efed0f0634970e3fcbf98asl/tile/layout-and-rearrangement/layout/TSHUF.aslasl/tile/layout-and-rearrangement/layout/TSHUF.aslasl/tile/layout-and-rearrangement/layout/TSHUF.asl