PTO-TILE-MODEL-EXECUTION-FUSED-MULTIPLY-ADD

1// PTO-UNIT: {"id":"PTO-TILE-MODEL-EXECUTION-FUSED-MULTIPLY-ADD","surface":"tile","classification":["model","execution","fused-multiply-add"],"depends_on":["PTO-TILE-MODEL-EXECUTION-ELEMENTWISE","PTO-TILE-MODEL-LEGALITY-INDEXED-LAYOUT","PTO-SCALAR-MODEL-FSU-PROFILE"]}2impdef func TileProfileFusedMultiplyAdd(3    data_type: TileDataType,4    addend: Word,5    left: Word,6    right: Word) => (Word, bits(5))7begin8    return ScalarFPFusedProfile(9        FloatingFused_MADD,10        DefaultNumericExecutionControl().rounding_mode,11        TileDataTypeToEncoding(data_type),12        addend,13        left,14        right);15end;16
17impdef func TileProfileFusedInvalidResult(18    data_type: TileDataType,19    left: Word,20    right: Word,21    addend: Word) => (Word, bits(5))22begin23    let (available, quiet_nan) =24        HardwareNumericCanonicalNaNResult(data_type);25    assert available;26    return (quiet_nan, Zeros{5} + 1);27end;28
29pure func TileNumericClassIsNegative(30    value_class: NumericValueClass) => boolean31begin32    return value_class == NumericValue_NegativeZero ||33           value_class == NumericValue_NegativeSubnormal ||34           value_class == NumericValue_NegativeNormal ||35           value_class == NumericValue_NegativeInfinity;36end;37
38func TileFixedFusedMultiplyAddValue(39    data_type: TileDataType,40    left: Word,41    right: Word,42    addend: Word) => (Word, bits(5))43begin44    assert TileFusedMultiplyAddDataTypeSupported(data_type);45    if TileDataTypeIsInteger(data_type) then46        let left_element = TileUnsignedElementValue(left, data_type);47        let right_element = TileUnsignedElementValue(right, data_type);48        let addend_element = TileUnsignedElementValue(addend, data_type);49        return (50            TileUnsignedElementValue(51                MultiplyWord(left_element, right_element) + addend_element,52                data_type),53            Zeros{5});54    end;55
56    assert TileNumericEncodingValid(data_type, left);57    assert TileNumericEncodingValid(data_type, right);58    assert TileNumericEncodingValid(data_type, addend);59    let left_class = TileNumericValueClass(data_type, left);60    let right_class = TileNumericValueClass(data_type, right);61    let addend_class = TileNumericValueClass(data_type, addend);62    let signaling_nan =63        left_class == NumericValue_SignalingNaN ||64        right_class == NumericValue_SignalingNaN ||65        addend_class == NumericValue_SignalingNaN;66    let zero_times_infinity =67        (NumericValueClassIsZero(left_class) &&68         NumericValueClassIsInfinity(right_class)) ||69        (NumericValueClassIsInfinity(left_class) &&70         NumericValueClassIsZero(right_class));71    let product_is_infinite =72        NumericValueClassIsInfinity(left_class) ||73        NumericValueClassIsInfinity(right_class);74    let product_is_negative =75        TileNumericClassIsNegative(left_class) !=76        TileNumericClassIsNegative(right_class);77    let opposite_infinities =78        product_is_infinite &&79        NumericValueClassIsInfinity(addend_class) &&80        product_is_negative != TileNumericClassIsNegative(addend_class);81    if signaling_nan || zero_times_infinity || opposite_infinities then82        return TileProfileFusedInvalidResult(83            data_type,84            left,85            right,86            addend);87    end;88    return TileProfileFusedMultiplyAdd(89        data_type,90        addend,91        left,92        right);93end;94
95// PTO-REQ-TFMA-001: complete preflight precedes three source snapshots. The96// valid payload, padding definedness, descriptor, and accumulated flags are97// computed privately and become visible only through the final publication.98func TFMA(99    destination: TileIndex,100    source_left: TileIndex,101    source_right: TileIndex,102    addend: TileIndex)103begin104    assert TileOperandsLegal_TFMA(105        destination,106        source_left,107        source_right,108        addend);109    let destination_tile = _Tiles[[destination]];110    let left_tile = _Tiles[[source_left]];111    let right_tile = _Tiles[[source_right]];112    let addend_tile = _Tiles[[addend]];113    var result_tile = destination_tile;114    var flags = Zeros{5};115    for row = 0 to destination_tile.valid_rows - 1 looplimit 65536 do116        for column = 0 to destination_tile.valid_columns - 1 looplimit 65536 do117            let element = TileLogicalLinearIndex(118                destination_tile,119                row as integer {0..65535},120                column as integer {0..65535});121            let (result, element_flags) = TileFixedFusedMultiplyAddValue(122                destination_tile.data_type,123                TileReadLogicalElement(left_tile, element),124                TileReadLogicalElement(right_tile, element),125                TileReadLogicalElement(addend_tile, element));126            result_tile = TileInfoWithLogicalElement(result_tile, element,127                result);128            flags = flags OR element_flags;129        end;130    end;131    result_tile = TileWithValidRegionDefined(result_tile);132    result_tile = TileWithPadding(133        result_tile,134        CurrentBundlePadValue());135    _Tiles[[destination]] = result_tile;136    ScalarFPRecordFlags(flags);137end;138