vfma ( a b c -- d )

Vocabulary

math.vectors

Inputs

a	a sequence
b	a sequence
c	a sequence

Outputs

d	a sequence

Word description

Computes a*b+c per lane with one rounding. SIMD float-4 and double-2 round at their lane precision; half-8 rounds once to binary16. The compiler does not implicitly fuse v* followed by v+.

Definition

IN: math.vectors

GENERIC: vfma ( a b c -- d )

Methods

USING: math.libm math.vectors math.vectors.simd sequences ;

M: bfloat-8 vfma [ ffmaf ] 3map ;


USING: accessors cpu.architecture kernel math.vectors
math.vectors.simd math.vectors.simd.intrinsics ;

M:: double-2 vfma ( a b c -- d )
    a double-2? b double-2? and [
        a underlying>> b underlying>> c underlying>>
        double-2-rep (simd-vfma) double-2 boa
    ] [ b bad-simd-vector ] if ; inline


USING: accessors cpu.architecture kernel math.vectors
math.vectors.simd math.vectors.simd.intrinsics ;

M:: float-4 vfma ( a b c -- d )
    a float-4? b float-4? and [
        a underlying>> b underlying>> c underlying>> float-4-rep
        (simd-vfma) float-4 boa
    ] [ b bad-simd-vector ] if ; inline


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

M: half-8 vfma half-fma ;


USING: kernel math.functions math.vectors sequences ;

M: object vfma [ fma ] 3map ; inline


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

M: simd-128 vfma 2drop bad-simd-vector ;