tiny.simd.per4LaneBlockShuffle

1pub fn per4LaneBlockShuffle(2    comptime D: type,3    comptime index3: usize,4    comptime index2: usize,5    comptime index1: usize,6    comptime index0: usize,7    value: D.Vector,8) D.Vector {9    inline for (.{ index0, index1, index2, index3 }) |index| {10        if (comptime index >= 4) @compileError("per-four-lane shuffle index must be below four");11    }12    const group = @min(D.lane_count, 4);13    const indices = [4]usize{ index0, index1, index2, index3 };14    const lanes: [D.lane_count]D.Lane = @bitCast(value);15    var result: [D.lane_count]D.Lane = undefined;16    inline for (0..D.lane_count) |lane| {17        const source = indices[lane % group];18        result[lane] = if (source < group) lanes[lane / group * group + source] else lanes[lane];19    }20    return @bitCast(result);21}

1pub const per4LaneBlockShuffle = block.per4LaneBlockShuffle;