tiny.simd.lanes

1const std = @import("std");23pub fn lowerHalf(comptime D: type, value: anytype) D.Vector {4    if (D.lane_count == 1) {5        if (@TypeOf(value) != D.Vector) @compileError("scalar lowerHalf requires the same descriptor");6        return value;7    }8    if (@TypeOf(value) != D.twice().Vector) {9        @compileError("lowerHalf source must have twice the destination lanes");10    }11    var result: D.Vector = undefined;12    inline for (0..D.lane_count) |index| result[index] = value[index];13    return result;14}1516pub fn upperHalf(comptime D: type, value: D.twice().Vector) D.Vector {17    var result: D.Vector = undefined;18    inline for (0..D.lane_count) |index| result[index] = value[D.lane_count + index];19    return result;20}2122pub fn zeroExtendVector(comptime D: type, value: D.half().Vector) D.Vector {23    validate(D);24    var result: D.Vector = @splat(0);25    inline for (0..D.lane_count / 2) |index| result[index] = value[index];26    return result;27}2829pub fn combine(30    comptime D: type,31    high: D.half().Vector,32    low: D.half().Vector,33) D.Vector {34    validate(D);35    var result: D.Vector = undefined;36    inline for (0..D.lane_count / 2) |index| {37        result[index] = low[index];38        result[D.lane_count / 2 + index] = high[index];39    }40    return result;41}4243pub fn concatLowerLower(comptime D: type, high: D.Vector, low: D.Vector) D.Vector {44    const H = D.half();45    return combine(D, lowerHalf(H, high), lowerHalf(H, low));46}4748pub fn concatUpperUpper(comptime D: type, high: D.Vector, low: D.Vector) D.Vector {49    const H = D.half();50    return combine(D, upperHalf(H, high), upperHalf(H, low));51}5253pub fn concatLowerUpper(comptime D: type, high: D.Vector, low: D.Vector) D.Vector {54    const H = D.half();55    return combine(D, lowerHalf(H, high), upperHalf(H, low));56}5758pub fn concatUpperLower(comptime D: type, high: D.Vector, low: D.Vector) D.Vector {59    const H = D.half();60    return combine(D, upperHalf(H, high), lowerHalf(H, low));61}6263pub fn concatOdd(comptime D: type, high: D.Vector, low: D.Vector) D.Vector {64    validate(D);65    var result: D.Vector = undefined;66    inline for (0..D.lane_count / 2) |index| {67        result[index] = low[index * 2 + 1];68        result[D.lane_count / 2 + index] = high[index * 2 + 1];69    }70    return result;71}7273pub fn concatEven(comptime D: type, high: D.Vector, low: D.Vector) D.Vector {74    validate(D);75    var result: D.Vector = undefined;76    inline for (0..D.lane_count / 2) |index| {77        result[index] = low[index * 2];78        result[D.lane_count / 2 + index] = high[index * 2];79    }80    return result;81}8283pub fn interleaveWholeLower(comptime D: type, a: D.Vector, b: D.Vector) D.Vector {84    validate(D);85    var result: D.Vector = undefined;86    inline for (0..D.lane_count / 2) |index| {87        result[index * 2] = a[index];88        result[index * 2 + 1] = b[index];89    }90    return result;91}9293pub fn interleaveWholeUpper(comptime D: type, a: D.Vector, b: D.Vector) D.Vector {94    validate(D);95    var result: D.Vector = undefined;96    inline for (0..D.lane_count / 2) |index| {97        result[index * 2] = a[D.lane_count / 2 + index];98        result[index * 2 + 1] = b[D.lane_count / 2 + index];99    }100    return result;101}102103pub fn lowerHalfOfMask(comptime D: type, mask: D.twice().Mask) D.Mask {104    var result: D.Mask = undefined;105    inline for (0..D.lane_count) |index| result[index] = mask[index];106    return result;107}108109pub fn upperHalfOfMask(comptime D: type, mask: D.twice().Mask) D.Mask {110    var result: D.Mask = undefined;111    inline for (0..D.lane_count) |index| result[index] = mask[D.lane_count + index];112    return result;113}114115pub fn combineMasks(116    comptime D: type,117    high: D.half().Mask,118    low: D.half().Mask,119) D.Mask {120    validate(D);121    var result: D.Mask = undefined;122    inline for (0..D.lane_count / 2) |index| {123        result[index] = low[index];124        result[D.lane_count / 2 + index] = high[index];125    }126    return result;127}128129fn validate(comptime D: type) void {130    if (D.lane_count < 2) @compileError("lane combination requires at least two lanes");131}132133fn verifyLaneType(comptime T: type) !void {134    const simd = @import("root.zig");135    const D = simd.FixedTag(T, 8);136    const H = D.half();137    const value: D.Vector = @splat(0);138    const low = lowerHalf(H, value);139    const high = upperHalf(H, value);140    try std.testing.expect(@reduce(.And, combine(D, high, low) == value));141    try std.testing.expect(@reduce(.And, concatOdd(D, value, value) == value));142    try std.testing.expect(@reduce(.And, concatEven(D, value, value) == value));143    try std.testing.expect(@reduce(.And, interleaveWholeLower(D, value, value) == value));144    try std.testing.expect(@reduce(.And, interleaveWholeUpper(D, value, value) == value));145}146147test "Highway lane geometry instantiates every lane type" {148    inline for (.{ u8, i8, u16, i16, u32, i32, u64, i64, f16, f32, f64 }) |T| {149        try verifyLaneType(T);150    }151    const D = @import("root.zig").FixedTag(u32, 1);152    try std.testing.expect(@reduce(.And, lowerHalf(D, @as(D.Vector, .{9})) ==153        @as(D.Vector, .{9})));154}155156test "Highway vector halves combine and zero extend in lane order" {157    const simd = @import("root.zig");158    const D = simd.FixedTag(u32, 8);159    const H = D.half();160    const value: D.Vector = .{ 0, 1, 2, 3, 4, 5, 6, 7 };161    const low = lowerHalf(H, value);162    const high = upperHalf(H, value);163    try std.testing.expect(@reduce(.And, low == @as(H.Vector, .{ 0, 1, 2, 3 })));164    try std.testing.expect(@reduce(.And, high == @as(H.Vector, .{ 4, 5, 6, 7 })));165    try std.testing.expect(@reduce(.And, combine(D, high, low) == value));166    try std.testing.expect(@reduce(.And, zeroExtendVector(D, low) ==167        @as(D.Vector, .{ 0, 1, 2, 3, 0, 0, 0, 0 })));168}169170test "Highway whole-vector concatenations select exact quarters" {171    const simd = @import("root.zig");172    const D = simd.FixedTag(u16, 8);173    const low: D.Vector = .{ 0, 1, 2, 3, 4, 5, 6, 7 };174    const high: D.Vector = .{ 10, 11, 12, 13, 14, 15, 16, 17 };175    try std.testing.expect(@reduce(.And, concatLowerLower(D, high, low) ==176        @as(D.Vector, .{ 0, 1, 2, 3, 10, 11, 12, 13 })));177    try std.testing.expect(@reduce(.And, concatUpperUpper(D, high, low) ==178        @as(D.Vector, .{ 4, 5, 6, 7, 14, 15, 16, 17 })));179    try std.testing.expect(@reduce(.And, concatLowerUpper(D, high, low) ==180        @as(D.Vector, .{ 4, 5, 6, 7, 10, 11, 12, 13 })));181    try std.testing.expect(@reduce(.And, concatUpperLower(D, high, low) ==182        @as(D.Vector, .{ 0, 1, 2, 3, 14, 15, 16, 17 })));183    try std.testing.expect(@reduce(.And, concatOdd(D, high, low) ==184        @as(D.Vector, .{ 1, 3, 5, 7, 11, 13, 15, 17 })));185    try std.testing.expect(@reduce(.And, concatEven(D, high, low) ==186        @as(D.Vector, .{ 0, 2, 4, 6, 10, 12, 14, 16 })));187}188189test "Highway whole interleave and mask halves retain ordering" {190    const simd = @import("root.zig");191    const D = simd.FixedTag(i32, 8);192    const H = D.half();193    const a: D.Vector = .{ 0, 1, 2, 3, 4, 5, 6, 7 };194    const b: D.Vector = .{ 10, 11, 12, 13, 14, 15, 16, 17 };195    try std.testing.expect(@reduce(.And, interleaveWholeLower(D, a, b) ==196        @as(D.Vector, .{ 0, 10, 1, 11, 2, 12, 3, 13 })));197    try std.testing.expect(@reduce(.And, interleaveWholeUpper(D, a, b) ==198        @as(D.Vector, .{ 4, 14, 5, 15, 6, 16, 7, 17 })));199    const mask: D.Mask = .{ true, false, true, false, false, true, false, true };200    const low = lowerHalfOfMask(H, mask);201    const high = upperHalfOfMask(H, mask);202    try std.testing.expect(@reduce(.And, combineMasks(D, high, low) == mask));203}

1pub const lanes = @import("lanes.zig");