(vdot4+) ( a b accumulator -- result )

Vocabulary

math.vectors.simd.extensions

Inputs

a	an object
b	an object
accumulator	an object

Outputs

result

an object

Definition

USING: system ;

IN: math.vectors.simd.extensions

HOOK: (vdot4+) cpu ( a b accumulator -- result )

Methods

USING: cpu.arm.64.features generic kernel math.vectors.simd
math.vectors.simd.extensions
math.vectors.simd.extensions.arm.64.private system ;

M:: arm.64 (vdot4+) ( a b c -- result )
    a uchar-16? b uchar-16? and [
        \ udot4 arm64-kernel-supported?
        [ a b c udot4 ]
        [ a b c M\ arm.64 (vdot4+) (call-next-method) ] if
    ] [
        a char-16? b char-16? and [
            \ sdot4 arm64-kernel-supported?
            [ a b c sdot4 ]
            [ a b c M\ arm.64 (vdot4+) (call-next-method) ] if
        ] [
            \ usdot4 arm64-kernel-supported?
            [ a uchar-16? [ a b c ] [ b a c ] if usdot4 ]
            [ a b c M\ arm.64 (vdot4+) (call-next-method) ] if
        ] if
    ] if ;


USING: kernel math math.vectors.simd.extensions sequences ;

M:: object (vdot4+) ( a b accumulator -- result )
    4 <iota> [| lane |
        lane accumulator nth 4 <iota> swap [| value j |
            lane 4 * j + :> index index a nth index b nth *
            value +
        ] reduce
    ] accumulator map-as ;