For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo function

mask_select8

def mask_select8[byte_idx: Int](s0: Float32, s1: Float32, s2: Float32, s3: Float32, s4: Float32, s5: Float32, s6: Float32, s7: Float32, mask_bits: UInt32) -> _RegisterPackType[Float32, Float32, Float32, Float32, Float32, Float32, Float32, Float32]

Masks 8 contiguous scores against one byte of a 32-column bitmask.

Lane j keeps its score if bit 8*byte_idx + j of mask_bits is set, otherwise it becomes MASK_VALUE (-10000). The 8 and/setp/selp are confined to a single opaque PTX block so the bit-extraction sits adjacent to the selects: the predicate live-set stays bounded (avoiding the wide up-front bit pre-extraction that spills) and the shape stays R2P-eligible.

Parameters:

byte_idx (Int): Which mask byte to apply (0..3), i.e. columns 8*byte_idx .. 8*byte_idx + 7.

Args:

s0 (Float32): Already-scaled score for lane 0.
s1 (Float32): Already-scaled score for lane 1.
s2 (Float32): Already-scaled score for lane 2.
s3 (Float32): Already-scaled score for lane 3.
s4 (Float32): Already-scaled score for lane 4.
s5 (Float32): Already-scaled score for lane 5.
s6 (Float32): Already-scaled score for lane 6.
s7 (Float32): Already-scaled score for lane 7.
mask_bits (UInt32): Packed 32-column visibility mask.

Returns:

_RegisterPackType[Float32, Float32, Float32, Float32, Float32, Float32, Float32, Float32]: The 8 masked scores, register-packed (index [0] .. [7]).