Module slipstream::mask

Bool-like types used for masked operations.

With multi-lane vectors, it is sometimes useful to do a lane-wise comparison or to disable some of the lanes for a given operation. Naturally, one would express this using a correctly sized bool array.

Nevertheless, the CPU SIMD instructions don't use bools, but signal true/false with a full-sized type with either all bits set to 1 or 0 (TODO: this is not true for AVX-512, what do we want to do about it?). Therefore, we define our own types that act like bools, but are represented in the above way. The comparison operators return vectors of these base mask types. The selection operations accept whatever mask vector with the same number of lanes, but they are expected to act fastest with the correct sized ones.

For the purpose of input, bool is also considered a mask type.

The interesting operations are:

• Comparisons (lt, le, eq, ge, gt)
• The blend method.
• Masked loading and storing of vectors.

The number in the type name specifies the number of bits. Therefore, for the u16x4, the natural mask type is a vector of 4 m16, which is m16x4.

While it is possible to operate with the bools (by converting them), it is more common to simply pipe the masks back into the vectors. Note that they do implement the usual boolean operators (however, only the non-shortcircuiting/bitwise variants). These work lane-wise.

Examples

fn abs(vals: &mut [i32]) {
    let zeroes = i32x8::default();
    for mut v in vals.vectorize_pad(i32x8::default()) {
        // Type of this one is m32x8 and is true whereever the lane isnegative.
        let negative = v.lt(zeroes);
        // Pick lanes from v where non-negative, pick from -v where negative.
        *v = v.blend(-*v, negative);
    }
}
let mut data = [1, -2, 3];
abs(&mut data);
assert_eq!(data, [1, 2, 3]);

Structs

MaskWrapper

Inner implementation of the mask types.

Traits

Mask	The trait implemented by all the mask types.

Type Definitions

m8
m16
m32
m64
m128
msize