Compute one same-type reciprocal-square-root operation for every valid Local Tile element.
PTO-TILE-TRSQRTTRSQRT <LB0:ValidCol, LB1:ValidRow, LB2:Col, DataType, PadValue> SrcTile0 -> DstTile| 分组 | Operand | 含义 | ASL 来源 |
|---|---|---|---|
| 参数 | LB0:ValidCol | dimension | B.DIM LB0=ValidCol |
| 参数 | LB1:ValidRow | dimension | B.DIM LB1=ValidRow (optional) |
| 参数 | LB2:Col | dimension | B.DIM LB2=Col (optional) |
| 参数 | DataType | data type | BSTART.SFU TRSQRT, DataType |
| 参数 | PadValue | padding value | B.DATR PadValue (optional) |
| 输入 | SrcTile0 | persistent Local floating source | contract.operands[1].source0 |
| 输出 | DstTile | new Local floating destination | contract.operands[0].destination0 |
BSTART.SFU TRSQRT, DataType
BSTART.SFU TRSQRT, DataType必选B.DATR PadValue (optional)可选B.DIM LB0=ValidCol必选可重复B.DIM LB1=ValidRow (optional)可选可重复B.DIM LB2=Col (optional)可选可重复B.IOT SrcTile, mask=PE_MASK, <last>, ->DstTile<TSize>必选BSTOP必选BSTART.SFU TRSQRT, DataTypeB.DIM LB0=ValidColB.IOT SrcTile, mask=PE_MASK, <last>, ->DstTile<TSize>BSTOPBSTART.SFU TRSQRT, DataType; B.DATR PadValue (optional); B.DIM LB0=ValidCol; B.DIM LB1=ValidRow (optional); B.DIM LB2=Col (optional); B.IOT SrcTile, mask=PE_MASK, <last>, ->DstTile<TSize>; 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 | — | — | new Local floating destination | — |
source0 | — | — | persistent Local floating source | — |
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
destination0source0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_TRSQRT() => TileOperationbegin return TileOperation_TRSQRT;end;pure func InstructionContractDataTypeLegal_TRSQRT( data_type: TileDataType) => booleanbegin return TileUnaryDataTypeSupported( TileUnary_RSQRT, data_type);end;
readonly func InstructionContractOperandsLegal_TRSQRT( destination: TileIndex, source: TileIndex) => booleanbegin return TileOperandsLegal_ExecuteTileUnary( TileUnary_RSQRT, destination, source);end;
readonly func InstructionContractHandler_TRSQRT() => TileSemanticHandlerbegin return TileHandler_ExecuteTileUnary;end;
func InstructionContractExecute_TRSQRT( destination: TileIndex, source: TileIndex)begin assert InstructionContractOperandsLegal_TRSQRT( destination, source); ExecuteTileUnary( TileUnary_RSQRT, destination, source);end;// PTO-INSTRUCTION: {"assembly":["TRSQRT <bundle operands>"],"block":["BSTART.SFU TRSQRT, DataType","B.DATR PadValue (optional)","B.DIM LB0=ValidCol","B.DIM LB1=ValidRow (optional)","B.DIM LB2=Col (optional)","B.IOT SrcTile, mask=PE_MASK, <last>, ->DstTile<TSize>","BSTOP"],"catalog_indices":[20],"catalog_records":[{"arguments":[{"constant":"TileUnary_RSQRT"},{"operand":"destination0"},{"operand":"source0"}],"command_mnemonic":"BSTART.TEPL","contract_status":"reviewed-complete","datr_contract":{"allowed_nonzero_fields":["PadValueOrByteId"],"pad_union":"pad-value"},"disposition":"accepted-direct-operation","effect_contract":"ExecuteTileUnary","family":"TEPL","fault_contract":"ExecuteTileInstruction","function":22,"legality_handler":"TileOperandsLegal_ExecuteTileUnary","mode":0,"name":"TRSQRT","operands":[{"field":"destination0","role":"new Local floating destination"},{"field":"source0","role":"persistent Local floating source"}],"restart_contract":"CompleteBundleAtWithAcceptedApplicabilityRules","selector":"0x016","semantic_handler":"ExecuteTileUnary","state_effects":["operand:destination0:new-local-floating-destination","operand:source0:persistent-local-floating-source","runtime:CurrentBundlePadValue:numeric-padding","runtime:NumericStatusFlags:profile-status"]}],"classification":["elementwise-tile-tile","transcendental"],"contract":{"block_composition":["BSTART.SFU TRSQRT, DataType","B.DATR PadValue (optional)","B.DIM LB0=ValidCol","B.DIM LB1=ValidRow (optional)","B.DIM LB2=Col (optional)","B.IOT SrcTile, mask=PE_MASK, <last>, ->DstTile<TSize>","BSTOP"],"canonical_assembly":["TRSQRT <bundle operands>"],"defaults":["LB0 is required and supplies nonzero ValidCol. Omitted LB1 selects ValidRow=1. Omitted LB2 selects Col=ValidCol; every explicitly present dimension must be nonzero.","Omitted B.DATR selects PadValue=Null. Explicit PadValue 00, 01, 10, and 11 select Zero, Max, Min, and Null.","The selected numeric profile supplies the operation-fixed approximation, rounding, exceptional result, and exact NV/DZ/OF/UF/NX status vector."],"encoding_class":"selector-encoded-block-operation","examples":["BSTART.SFU TRSQRT, DataType; B.DATR PadValue (optional); B.DIM LB0=ValidCol; B.DIM LB1=ValidRow (optional); B.DIM LB2=Col (optional); B.IOT SrcTile, mask=PE_MASK, <last>, ->DstTile<TSize>; BSTOP"],"exceptions":["Malformed Local bindings, B.IOR or B.IOS presence, missing or zero dimensions, unsupported DataType, undefined or invalid source encoding, descriptor mismatch, invalid capacity, or allocation failure raises the applicable Tile fault before effects.","Floating zero is legal, reports divide-by-zero, and produces signed infinity where representable; positive infinity produces positive zero; negative nonzero values report invalid and produce quiet NaN.","E4M3 has no infinity encoding and TRSQRT does not admit saturation, so either signed zero produces canonical E4M3 quiet NaN 0x7F and records only divide-by-zero without overflow."],"field_contracts":{"B.DATR.PadValueOrByteId":{"ref":"PTO-FIELD-BLOCK-PADVALUE-OR-BYTEID"}},"field_zero_meanings":{"B.DATR.PadValueOrByteId":"Zero padding when B.DATR is present; omission selects Null."},"legality":["TRSQRT retains its TEPL raw Mode 0 carrier and executes canonically on the SFU engine.","Exactly one terminating Local B.IOT supplies one persistent source and one newly allocated destination. B.IOR and B.IOS are illegal.","The selected DataType is exactly FP16, FP32, or BF16; every integer, exponent-only, other compact, packed, assigned-but-inapplicable, or reserved DataType rejects before effects.","Source and destination match physical shape, valid shape, row-major layout, and DataType. Every valid source element is defined and has a valid encoding.","PadValueOrByteId is the only applicable B.DATR field; nondefault CMode, Sat, Canonicalize, secondary DataType, RMode, or Layout is illegal.","Source and destination use one PE_MASK. PE_MASK=0000 is a strict no-op before descriptor reads, allocation, faults, numeric status, or payload effects."],"memory_effects":["none"],"operands":[{"field":"destination0","role":"new Local floating destination"},{"field":"source0","role":"persistent Local floating source"}],"ordering":["Complete schema, attribute, dimension, type, descriptor, source-definedness, source-encoding, mask, capacity, and allocation preflight precedes the source snapshot and profile evaluation.","The complete source payload is snapshotted before destination publication, so source/destination aliasing observes the old source value."],"standalone_opcode":false,"state_effects":["For each valid element compute the selected profile's same-type reciprocal square root.","Accumulate all element status flags, apply selected physical padding, and publish payload, definedness, numeric status, and destination descriptor atomically; rejection leaves architectural state unchanged."]},"depends_on":["PTO-BLOCK-MODEL-SCHEMA-PROFILE-ENCODING"],"engine":"SFU","id":"PTO-TILE-TRSQRT","mnemonic":"TRSQRT","summary":"Compute one same-type reciprocal-square-root operation for every valid Local Tile element.","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-TRSQRT-CONTRACT-001// ndf: kind=contract level=L1 layer=tile status=accepted// TRSQRT MUST compute one same-type reciprocal square root for every valid// Local floating element. Either E4M3 signed zero MUST produce canonical// quiet NaN 0x7F and MUST set DZ without setting OF.// NDF-END: PTO-TRSQRT-CONTRACT-001// DOC-BEGIN: decodereadonly func InstructionContractOperation_TRSQRT() => TileOperationbegin return TileOperation_TRSQRT;end;// DOC-END: decode// DOC-BEGIN: operationpure func InstructionContractDataTypeLegal_TRSQRT( data_type: TileDataType) => booleanbegin return TileUnaryDataTypeSupported( TileUnary_RSQRT, data_type);end;
readonly func InstructionContractOperandsLegal_TRSQRT( destination: TileIndex, source: TileIndex) => booleanbegin return TileOperandsLegal_ExecuteTileUnary( TileUnary_RSQRT, destination, source);end;
readonly func InstructionContractHandler_TRSQRT() => TileSemanticHandlerbegin return TileHandler_ExecuteTileUnary;end;
func InstructionContractExecute_TRSQRT( destination: TileIndex, source: TileIndex)begin assert InstructionContractOperandsLegal_TRSQRT( destination, source); ExecuteTileUnary( TileUnary_RSQRT, destination, source);end;// DOC-END: operation
TRSQRT 对每个有效元素计算同类型倒数平方根，并发布一个新的 Local 目标。
该操作只在有效矩形内按助记符选定的带类型的元素规则求值。
浮点结果与元素状态遵循当前具名数值配置档；可移植契约拥有选择、形状、发布与故障顺序。
destination0 标识新分配的目的 Tile。source0 提供持久源 Tile。FP16、FP32、BF16。LB0、LB1、LB2 按该助记符契约补全有效形状与物理形状；所有必需有效范围都必须非零。所有源描述符与载荷都会在目标发布前完成验证和快照。
完整目标载荷、描述符、已定义性、填充状态与适用数值状态会原子发布；拒绝路径不发布任何部分。
Null 填充让有效矩形外的物理坐标保持未定义；显式非 Null 填充值会用选定带类型的值定义这些位置。
源 Tile 在成功执行后保持不变。
完整绑定模式、维度、DataType、布局、源已定义性、数值编码、目标容量与分配都会在效果前预检。
合法性或分配检查失败会引发相应 Tile 故障，不留下部分目标、状态或内存效果。
PE_MASK=0000 是严格无操作，发生在操作数读取、分配、故障、数值状态或载荷效果之前。
下面的示例只帮助理解当前 ASL 绑定契约，并不是第二份指令定义。
TRSQRT <bundle operands> 先完成完整预检与源快照，再原子发布助记符定义的结果与填充状态。
正文来自 owning ASL。拖拽或按钮只临时改变当前页面显示顺序。
TRSQRT MUST compute one same-type reciprocal square root for every valid Local floating element. Either E4M3 signed zero MUST produce canonical quiet NaN 0x7F and MUST set DZ without setting OF.
PTO-TRSQRT-CONTRACT-001asl/tile/elementwise-tile-tile/transcendental/TRSQRT.asle4f019f6ec6bd4bd512ce7c72f8d1c7f4afe31e032345aa9060a4b87a37de28c132f5c8c57d51749108a512ffe034cf63d56fe61ac964de4607df3e4d44bcd0813 matching entries
PTO-AVS-TILE-TRSQRT-E4M3ZERO-001tests/asl/tile/elementwise-tile-tile/transcendental/TRSQRT/tile-exec-trsqrt-e4m3-zero-001.aslb5c2e01d6b8dbfcfac25a4c4e04f5e67713d6354fda433662c4ed00632e2e17cPTO-AVS-TILE-TRSQRT-INFINITY-001tests/asl/tile/elementwise-tile-tile/transcendental/TRSQRT/tile-exec-trsqrt-infinity-001.asl9e28bd0331067003679c09338943df8efdd35c0e2bffa0ef0ec5ffd7e8da9ad9PTO-AVS-TILE-TRSQRT-STATIC-001tests/asl/tile/elementwise-tile-tile/transcendental/TRSQRT/tile-static-trsqrt-contract-001.asl5117780b0e76f16173928f798e648097245c621dba31e29ed92c7eeba521c880PTO-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-0035…
ADR-0035docs/status/decisions/0035-vec-sfu-carrier-totality.md41f8b83397773a166978831f314557629abe0a5bad906648456dec78198c910eLoading ADR-0052…
ADR-0052docs/status/decisions/0052-direct-tile-and-bundle-catalog-closure.md5fdb38bf480e37affe5075f93b35a57d32803018c986bfe05160f188a19ccaf6Loading ADR-0057…
ADR-0057docs/status/decisions/0057-tile-classification-and-engine-aliases.mdb812b5d70b7e6749b2772d5b4b23ed790aa62c915452392b4552a5ac056cad78Loading ADR-0059…
ADR-0059docs/status/decisions/0059-mnemonic-field-encoding-closure.mdf18d13f735cfaf9f2803c754c5818f4af33d1cc42c489f0d9ce060b07bb3a196Loading ADR-0080…
ADR-0080docs/status/decisions/0080-tile-elementwise-and-irregular-operations.md3f6aed3e2ac792a75e46787be99eb8c1d5ce204b098f531aa9fcb80b87448d72{
"classification": [
"elementwise-tile-tile",
"transcendental"
],
"documentation": "docs/tile/elementwise-tile-tile/transcendental/TRSQRT.md",
"id": "PTO-TILE-TRSQRT",
"instruction_contract": {
"artifact": "spec/evidence/instruction-contract-closure.json",
"mnemonic": "TRSQRT",
"ndf_clause": "PTO-INST-TILE-TRSQRT"
},
"mnemonic": "TRSQRT",
"readiness_subjects": [
"ADR-0035",
"ADR-0052",
"ADR-0057",
"ADR-0059",
"ADR-0080"
],
"semantic_tests": [
"PTO-AVS-TILE-TRSQRT-E4M3ZERO-001",
"PTO-AVS-TILE-TRSQRT-INFINITY-001"
],
"source": "asl/tile/elementwise-tile-tile/transcendental/TRSQRT.asl",
"surface": "tile",
"tests": [
"PTO-AVS-TILE-TRSQRT-E4M3ZERO-001",
"PTO-AVS-TILE-TRSQRT-INFINITY-001",
"PTO-AVS-TILE-TRSQRT-STATIC-001"
]
}0.58.5 · 候选发布7dc8b7e5b121d2b2499a2273bebff29e2cd86812e4f019f6ec6bd4bd512ce7c72f8d1c7f4afe31e032345aa9060a4b87a37de28cdd7be39d3e17593dafc7dab5437628ff825ca232f9a07c5b8f074b313a1c694easl/tile/elementwise-tile-tile/transcendental/TRSQRT.aslasl/tile/elementwise-tile-tile/transcendental/TRSQRT.aslasl/tile/elementwise-tile-tile/transcendental/TRSQRT.asl