Permute raw bytes from two Local CUBE sources by a Local U8 index Tile.
PTO-TILE-TPERMUTETPERMUTE <LB0, LB1, LB2, DataType, Layout> SrcTile0, SrcTile1, SrcTile2 -> DstTile| 分组 | Operand | 含义 | ASL 来源 |
|---|---|---|---|
| 参数 | LB0 | dimension | B.DIM LB0 (optional) |
| 参数 | LB1 | dimension | B.DIM LB0/LB1/LB2 (optional) |
| 参数 | LB2 | dimension | B.DIM LB0/LB1/LB2 (optional) |
| 参数 | DataType | data type | BSTART.SFU TPERMUTE, DataType |
| 参数 | Layout | layout | B.DATR Layout (optional) |
| 输入 | SrcTile0 | source0 | contract.operands[1].source0 |
| 输入 | SrcTile1 | source1 | contract.operands[2].source1 |
| 输入 | SrcTile2 | indices | contract.operands[3].source2 |
| 输出 | DstTile | destination | contract.operands[0].destination0 |
BSTART.SFU TPERMUTE, DataType
BSTART.SFU TPERMUTE, DataType必选B.DATR Layout (optional)可选B.DIM LB0 (optional)可选B.IOT source0, source1必选可重复B.IOT indices, ->destination必选可重复BSTOP必选BSTART.SFU TPERMUTE, DataTypeB.IOT source0, source1B.IOT indices, ->destinationBSTOPBSTART.SFU TPERMUTE, U32; B.DATR Layout; B.IOT source0, source1; B.IOT indices, ->destination; BSTOP| 字段 | 位宽 | 有符号性 | 架构角色 | 编码零 |
|---|---|---|---|---|
DataType | 5 | encoding-defined | — | — |
Mode | 2 | encoding-defined | — | — |
Function | 5 | encoding-defined | — | — |
destination0 | — | — | destination | — |
source0 | — | — | source0 | — |
source1 | — | — | source1 | — |
source2 | — | — | indices | — |
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
destination0source0source1source2下面是该指令所有者中的 Operation；页面没有重写这段行为。
readonly func InstructionContractOperation_TPERMUTE() => TileOperationbegin return TileOperation_TPERMUTE;end;readonly func InstructionContractHandler_TPERMUTE() => TileSemanticHandlerbegin return TileHandler_TPERMUTE;end;
pure func InstructionContractDataTypeLegal_TPERMUTE( data_type: TileDataType) => booleanbegin return TileCubeDataTypeSupported(data_type) && TileElementBits(data_type) != 64;end;
readonly func InstructionContractOperandsLegal_TPERMUTE( destination: TileIndex, source0: TileIndex, source1: TileIndex, indices: TileIndex) => booleanbegin return TileOperandsLegal_TPERMUTE(destination, source0, source1, indices);end;
func InstructionContractExecute_TPERMUTE( destination: TileIndex, source0: TileIndex, source1: TileIndex, indices: TileIndex)begin assert InstructionContractOperandsLegal_TPERMUTE( destination, source0, source1, indices); TPERMUTE(destination, source0, source1, indices);end;// PTO-INSTRUCTION: {"assembly":["TPERMUTE <bundle operands>"],"block":["BSTART.SFU TPERMUTE, DataType","B.DATR Layout (optional)","B.DIM LB0 (optional)","B.IOT source0, source1","B.IOT indices, ->destination","BSTOP"],"catalog_indices":[103],"catalog_records":[{"arguments":[{"operand":"destination0"},{"operand":"source0"},{"operand":"source1"},{"operand":"source2"}],"command_mnemonic":"BSTART.TEPL","contract_status":"reviewed-complete","datr_contract":{"allowed_nonzero_fields":["Layout"],"pad_union":"must-zero"},"disposition":"accepted-direct-operation","effect_contract":"TPERMUTE","family":"TEPL","fault_contract":"ExecuteTileInstruction","function":21,"legality_handler":"TileOperandsLegal_TPERMUTE","mode":3,"name":"TPERMUTE","operands":[{"field":"destination0","role":"destination"},{"field":"source0","role":"source0"},{"field":"source1","role":"source1"},{"field":"source2","role":"indices"}],"restart_contract":"CompleteBundleAtWithAcceptedApplicabilityRules","selector":"0x075","semantic_handler":"TPERMUTE","state_effects":["operand:destination0:destination","operand:source0:source0","operand:source1:source1","operand:source2:indices"]}],"classification":["layout-and-rearrangement","layout"],"contract":{"block_composition":["BSTART.SFU TPERMUTE, DataType","B.DATR Layout (optional)","B.DIM LB0/LB1/LB2 (optional)","B.IOT source0, source1","B.IOT indices, ->destination","BSTOP"],"canonical_assembly":["TPERMUTE <bundle operands>"],"defaults":["B.DATR has no effect other than selecting CUBE_M16 or CUBE_M32; padding and numeric fields remain zero.","A nonzero PE mask requires two ordered B.IOT bindings and no B.IOR."],"encoding_class":"selector-encoded-block-operation","examples":["BSTART.SFU TPERMUTE, U32; B.DATR Layout; B.IOT source0, source1; B.IOT indices, ->destination; BSTOP"],"exceptions":["Illegal raw indices reject before any destination effect with Fault_TileLegality.","CompleteBundleAtWithAcceptedApplicabilityRules supplies restart and completion behavior."],"field_contracts":{},"field_zero_meanings":{},"legality":["TPERMUTE accepts only Local CUBE_M16 or CUBE_M32 data Tiles with matching dtype and geometry.","indices is Local U8 with the same CUBE layout and supplies one byte index for every valid destination byte.","The destination is fresh; source0 and source1 may alias, while indices is distinct from both sources.","Raw bytes are rearranged without numerical conversion."],"memory_effects":["none"],"operands":[{"field":"destination0","role":"destination"},{"field":"source0","role":"source0"},{"field":"source1","role":"source1"},{"field":"source2","role":"indices"}],"ordering":["All index legality and source reads precede destination publication."],"standalone_opcode":false,"state_effects":["Perform per-row two-source raw-byte table lookup and publish only the destination valid region."]},"depends_on":["PTO-TILE-MODEL-EXECUTION-REARRANGEMENT","PTO-TILE-MODEL-LEGALITY-LAYOUT-REARRANGEMENT"],"engine":"SFU","id":"PTO-TILE-TPERMUTE","mnemonic":"TPERMUTE","summary":"Permute raw bytes from two Local CUBE sources by a Local U8 index Tile.","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-TPERMUTE-CONTRACT-001// ndf: kind=contract level=L1 layer=tile status=accepted// TPERMUTE performs a two-source per-row raw-byte lookup over Local CUBE_M16// or CUBE_M32 Cells. Every valid destination byte and its index are checked// before the fresh destination is published; invalid indices reject atomically.// NDF-END: PTO-TPERMUTE-CONTRACT-001// DOC-BEGIN: decodereadonly func InstructionContractOperation_TPERMUTE() => TileOperationbegin return TileOperation_TPERMUTE;end;// DOC-END: decode// DOC-BEGIN: operationreadonly func InstructionContractHandler_TPERMUTE() => TileSemanticHandlerbegin return TileHandler_TPERMUTE;end;
pure func InstructionContractDataTypeLegal_TPERMUTE( data_type: TileDataType) => booleanbegin return TileCubeDataTypeSupported(data_type) && TileElementBits(data_type) != 64;end;
readonly func InstructionContractOperandsLegal_TPERMUTE( destination: TileIndex, source0: TileIndex, source1: TileIndex, indices: TileIndex) => booleanbegin return TileOperandsLegal_TPERMUTE(destination, source0, source1, indices);end;
func InstructionContractExecute_TPERMUTE( destination: TileIndex, source0: TileIndex, source1: TileIndex, indices: TileIndex)begin assert InstructionContractOperandsLegal_TPERMUTE( destination, source0, source1, indices); TPERMUTE(destination, source0, source1, indices);end;// DOC-END: operation
使用场景。 TPERMUTE 在每个 CUBE CELL 内提供一个双源字节表，使每个有效目标字节都能独立选择，而无需对载荷作数值解释。
工作方式。 查找在每个 CUBE CELL 处重新开始：对于 CUBE_M32，每个源的表宽为 4 字节；对于 CUBE_M16，表宽为 8 字节。小于该宽度的索引选择 source0 中同一 CELL 的字节，随后相同大小的索引范围选择 source1 中同一 CELL 的字节。
输入与结果。 两个 Local 数据源采用相同的受支持非 64 位 dtype、CUBE 布局和几何形状；匹配的 Local U8 索引 Tile 为每个有效目标字节提供一个索引，目标则是新 Tile。
效果。 所有索引及其选中的源字节均在完整目标有效区域发布前完成校验和读取；源 Tile 与索引 Tile 保持不变，目标填充为 Null，且该操作不产生内存效果。
拒绝条件。 每个索引必须小于每个 CELL 的合并边界：CUBE_M32 为 8，CUBE_M16 为 16。索引越界、选中的字节未定义、布局、dtype 或几何形状不匹配、索引 Tile 与任一源重叠，或目标与任一输入重叠，都会在目标产生任何效果前被拒绝。
具体示例。 对于一个 CUBE_M32 行，源字分别为 0x04030201 和 0x08070605 时，字节索引 [0, 4, 1, 5] 生成目标字 0x06020501。
正文来自 owning ASL。拖拽或按钮只临时改变当前页面显示顺序。
TPERMUTE performs a two-source per-row raw-byte lookup over Local CUBE_M16 or CUBE_M32 Cells. Every valid destination byte and its index are checked before the fresh destination is published; invalid indices reject atomically.
PTO-TPERMUTE-CONTRACT-001asl/tile/layout-and-rearrangement/layout/TPERMUTE.asl455ba580be500adcba1b4597a270d7bb5f583eb7da95462288c5e2e60d18b19bb75b5e2eac189981637fcedc448a7ebf3a111de1e8a50a902eaf1af5ea181e9e10 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-TPERMUTE-BYTE-TABLE-001tests/asl/tile/layout-and-rearrangement/layout/TPERMUTE/tile-exec-tpermute-byte-table-001.asl310c6441f9c7354a8a040e9577437ff333506e4ca555591e970d78d628a75d9aPTO-AVS-TILE-TPERMUTE-STATIC-001tests/asl/tile/layout-and-rearrangement/layout/TPERMUTE/tile-static-tpermute-contract-001.asl99c2f9d9e8381fb57c475452dde327338bd73ba82d1e71e12fa627e5d1daf0c4PTO-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/TPERMUTE.md",
"id": "PTO-TILE-TPERMUTE",
"instruction_contract": {
"artifact": "spec/evidence/instruction-contract-closure.json",
"mnemonic": "TPERMUTE",
"ndf_clause": "PTO-INST-TILE-TPERMUTE"
},
"mnemonic": "TPERMUTE",
"readiness_subjects": [
"ADR-0103",
"ADR-0108"
],
"semantic_tests": [
"PTO-AVS-TILE-TPERMUTE-BYTE-TABLE-001"
],
"source": "asl/tile/layout-and-rearrangement/layout/TPERMUTE.asl",
"surface": "tile",
"tests": [
"PTO-AVS-TILE-TPERMUTE-BYTE-TABLE-001",
"PTO-AVS-TILE-TPERMUTE-STATIC-001"
]
}0.58.5 · 候选发布7dc8b7e5b121d2b2499a2273bebff29e2cd86812455ba580be500adcba1b4597a270d7bb5f583eb7da95462288c5e2e60d18b19bcc8e93840c3e279879b877e54c76d8984d3bd849ff49a580a6fb200b2e53e6b6asl/tile/layout-and-rearrangement/layout/TPERMUTE.aslasl/tile/layout-and-rearrangement/layout/TPERMUTE.aslasl/tile/layout-and-rearrangement/layout/TPERMUTE.asl