fmla        v8.4s, v0.4s, v4.s[0]
    ldr         w4, [x1]
fmla        v9.4s, v1.4s, v4.s[0]
        ldr         w20, [x2], #4
fmla        v10.4s, v2.4s, v4.s[0]
    ldr         w5, [x1, #4]

fmla        v11.4s, v0.4s, v4.s[1]
        ldr         w21, [x2], #4
fmla        v12.4s, v1.4s, v4.s[1]
    ldr         w6, [x1, #8]
fmla        v13.4s, v2.4s, v4.s[1]
        ldr         w22, [x2], #4

fmla        v14.4s, v0.4s, v4.s[2]
    ldr         w7, [x1, #12]
fmla        v15.4s, v1.4s, v4.s[2]
        ldr         w23, [x2], #4
fmla        v16.4s, v2.4s, v4.s[2]
    ldr         w8, [x1, #16]
fmla        v17.4s, v0.4s, v4.s[3]
        ldr         w24, [x2], #4
fmla        v18.4s, v1.4s, v4.s[3]
    ldr         w9, [x1, #20]
fmla        v19.4s, v2.4s, v4.s[3]
        ldr         w25, [x2], #4

fmla        v20.4s, v0.4s, v5.s[0]
    ldr         w10, [x1, #24]
fmla        v21.4s, v1.4s, v5.s[0]
        ldr         w26, [x2], #4
fmla        v22.4s, v2.4s, v5.s[0]
    ldr         w11, [x1, #28]
fmla        v23.4s, v0.4s, v5.s[1]
        ldr         w27, [x2], #4
fmla        v24.4s, v1.4s, v5.s[1]
    ldr         w12, [x1, #32]
fmla        v25.4s, v2.4s, v5.s[1]
    ldr         w13, [x1, #36]

fmla        v26.4s, v0.4s, v5.s[2]
    ldr         w14, [x1, #40]
fmla        v27.4s, v1.4s, v5.s[2]
    ldr         w15, [x1, #44]
fmla        v28.4s, v2.4s, v5.s[2]
    prfm        pldl1keep, [x1, #512]
fmla        v29.4s, v0.4s, v5.s[3]
    add         x1, x1, #48
fmla        v30.4s, v1.4s, v5.s[3]
    prfm        pldl1keep, [x2, #384]
fmla        v31.4s, v2.4s, v5.s[3]

    ins         v0.s[0], w4

    ins         v1.s[0], w8
    ins         v2.s[0], w12

        ins         v4.s[0], w20
        ins         v5.s[0], w24

    ins         v0.s[1], w5
    ins         v1.s[1], w9

    ins         v2.s[1], w13
        ins         v4.s[1], w21

        ins         v5.s[1], w25
    ins         v0.s[2], w6

    ins         v1.s[2], w10
    ins         v2.s[2], w14

        ins         v4.s[2], w22
        ins         v5.s[2], w26

    ins         v0.s[3], w7
    ins         v1.s[3], w11

    ins         v2.s[3], w15
        ins         v4.s[3], w23
        ins         v5.s[3], w27