2d-matmul-mixed ( vec1: float-array vec2: float-array res: float-array m: fixnum n: fixnum p: fixnum start: fixnum -- )

Vocabulary

tensors.private

Inputs

vec1	a float-array
vec2	a float-array
res	a float-array
m	a fixnum
n	a fixnum
p	a fixnum
start	a fixnum

Outputs

None

Definition

USING: accessors kernel math math.vectors sequences
sequences.private typed ;

IN: tensors.private

TYPED:: 2d-matmul-mixed
    ( vec1: float-array vec2: float-array res: float-array m: fixnum n: fixnum p: fixnum start: fixnum -- )
    { n p } vec2 <tensor> 2d-transpose vec>> :> vec2 start m n *
    * :> start1 start n p * * :> start2 m [
        :> i i n * :> in 4 4 in start1 + 4 mod - swap mod
        :> in4m i p * :> ip vec1 in n make-subseq
        :> sub1 sub1 in4m <simd-slice> :> slice1 p [
            :> j j n * :> jn 4 4 jn 4 mod - swap mod
            :> jn4m vec2 jn n make-subseq in4m jn4m = [
                jn4m <simd-slice> slice1 swap 2dup
                [ first-slice>> ] bi@ 0.0 [ * + ] 2reduce
                [ 2dup [ simd-slice>> ] bi@ ] dip
                [ vdot + ] 2reduce [ [ end-slice>> ] bi@ ] dip
                [ * + ] 2reduce
            ] [ sub1 swap 0.0 [ * + ] 2reduce ] if ip j +
            res set-nth-unsafe
        ] each-integer
    ] each-integer ;