tiny.simd.UnrollerUnit2D

1pub fn UnrollerUnit2D(2    comptime Input0Lane: type,3    comptime Input1Lane: type,4    comptime OutputLane: type,5) type {6    validateLane(Input0Lane);7    validateLane(Input1Lane);8    validateLane(OutputLane);9    const LargerLane = signedLane(@max(10        @sizeOf(Input0Lane),11        @max(@sizeOf(Input1Lane), @sizeOf(OutputLane)),12    ));13    const LargerTag = tag.ScalableTag(LargerLane);1415    return struct {16        pub const input_count: usize = 2;17        pub const Input0: type = Input0Lane;18        pub const Input1: type = Input1Lane;19        pub const Output: type = OutputLane;20        pub const Larger: type = LargerLane;21        pub const Tag: type = LargerTag;22        pub const Input0Tag: type = LargerTag.rebind(Input0Lane);23        pub const Input1Tag: type = LargerTag.rebind(Input1Lane);24        pub const OutputTag: type = LargerTag.rebind(OutputLane);25        pub const Input0Vector: type = Input0Tag.Vector;26        pub const Input1Vector: type = Input1Tag.Vector;27        pub const OutputVector: type = OutputTag.Vector;28        pub const X0Vector: type = Input0Vector;29        pub const X1Vector: type = Input1Vector;30        pub const YVector: type = OutputVector;31        pub const max_unit_lanes: usize = LargerTag.lane_count;32        pub const actual_lanes: usize = LargerTag.lane_count;3334        pub fn maxUnitLanes() usize {35            return max_unit_lanes;36        }3738        pub fn actualLanes() usize {39            return actual_lanes;40        }4142        pub fn x0Init() Input0Vector {43            return @splat(0);44        }4546        pub fn x1Init() Input1Vector {47            return @splat(0);48        }4950        pub fn yInit() OutputVector {51            return @splat(0);52        }5354        pub fn load0(index: isize, input: []const Input0) Input0Vector {55            return loadFull(Input0Tag, index, input);56        }5758        pub fn load1(index: isize, input: []const Input1) Input1Vector {59            return loadFull(Input1Tag, index, input);60        }6162        pub fn maskLoad0(63            index: isize,64            input: []const Input0,65            places: isize,66        ) Input0Vector {67            return maskLoad0Or(@splat(0), index, input, places);68        }6970        pub fn maskLoad0Or(71            inactive: Input0Vector,72            index: isize,73            input: []const Input0,74            places: isize,75        ) Input0Vector {76            return loadPartial(Input0Tag, inactive, index, input, places);77        }7879        pub fn maskLoad1(80            index: isize,81            input: []const Input1,82            places: isize,83        ) Input1Vector {84            return maskLoad1Or(@splat(0), index, input, places);85        }8687        pub fn maskLoad1Or(88            inactive: Input1Vector,89            index: isize,90            input: []const Input1,91            places: isize,92        ) Input1Vector {93            return loadPartial(Input1Tag, inactive, index, input, places);94        }9596        pub fn storeAndShortCircuit(97            index: isize,98            output: []Output,99            value: OutputVector,100        ) bool {101            storeFull(OutputTag, index, output, value);102            return true;103        }104105        pub fn maskStore(106            index: isize,107            output: []Output,108            value: OutputVector,109            places: isize,110        ) usize {111            return storePartial(OutputTag, index, output, value, places);112        }113114        pub fn reduceFinal(_: OutputVector, _: []Output) usize {115            return 0;116        }117118        pub fn reduceUnrolled(119            _: OutputVector,120            _: OutputVector,121            _: OutputVector,122            value: OutputVector,123        ) OutputVector {124            return value;125        }126    };127}

1pub const UnrollerUnit2D = unroller.UnrollerUnit2D;