ld1         {{ v9.4s, v10.4s, v11.4s }}, [x1], #48

    fmla        v16.8h, v0.8h, v8.h[0]
    ldr         w8, [x2], #4
    fmla        v17.8h, v1.8h, v8.h[0]
    ldr         d12, [x1], #8
    fmla        v18.8h, v2.8h, v8.h[0]
    ldr         x12, [x1], #8
    fmla        v19.8h, v3.8h, v8.h[0]
    ldr         d13, [x1], #8
    fmla        v20.8h, v4.8h, v8.h[0]
    ldr         x13, [x1], #8
    fmla        v21.8h, v5.8h, v8.h[0]
    ldr         d14, [x1], #8
    fmla        v22.8h, v6.8h, v8.h[0]
    ldr         x14, [x1], #8
    fmla        v23.8h, v7.8h, v8.h[0]
    ldr         d15, [x1], #8
    fmla        v24.8h, v9.8h, v8.h[0]
    ldr         x15, [x1], #8

    ld1         {{ v0.8h, v1.8h, v2.8h, v3.8h }}, [x1], #64
    ins         v8.s[1], w8
    ld1         {{ v4.8h, v5.8h, v6.8h, v7.8h }}, [x1], #64

    fmla        v25.8h, v10.8h, v8.h[0]
    ins         v12.d[1], x12
    fmla        v26.8h, v11.8h, v8.h[0]
    ins         v13.d[1], x13
    fmla        v27.8h, v12.8h, v8.h[0]
    ins         v14.d[1], x14
    fmla        v28.8h, v13.8h, v8.h[0]
    ins         v15.d[1], x15

    ld1         {{ v9.8h, v10.8h, v11.8h, v12.8h }}, [x1], #64

    fmla        v29.8h, v14.8h, v8.h[0]
    ldr         d13, [x1], #8
    fmla        v30.8h, v15.8h, v8.h[0]
    ldr         x13, [x1], #8
    fmla        v31.8h, v0.8h, v8.h[0]
    ldr         d14, [x1], #8

    fmla        v16.8h, v1.8h, v8.h[2]
    ldr         x14, [x1], #8
    fmla        v17.8h, v2.8h, v8.h[2]
    ldr         d15, [x1], #8
    fmla        v18.8h, v3.8h, v8.h[2]
    ldr         x15, [x1], #8
    fmla        v19.8h, v4.8h, v8.h[2]

    ld1         {{ v0.8h }}, [x1], #16

    fmla        v20.8h, v5.8h, v8.h[2]
    ldr         d1, [x1], #8
    fmla        v21.8h, v6.8h, v8.h[2]
    ldr         x10, [x1], #8

    fmla        v22.8h, v7.8h, v8.h[2]

    fmla        v23.8h, v9.8h, v8.h[2]
    ins         v13.d[1], x13
    fmla        v24.8h, v10.8h, v8.h[2]
    ins         v14.d[1], x14
    fmla        v25.8h, v11.8h, v8.h[2]
    ins         v15.d[1], x15

    fmla        v26.8h, v12.8h, v8.h[2]
    prfm        pldl1keep, [x1, #1024]
    fmla        v27.8h, v13.8h, v8.h[2]
    ins         v1.d[1], x10
    fmla        v28.8h, v14.8h, v8.h[2]
    prfm        pldl1keep, [x1, #1088]
    fmla        v29.8h, v15.8h, v8.h[2]
    prfm        pldl1keep, [x1, #1152]
    fmla        v30.8h, v0.8h, v8.h[2]
    prfm        pldl1keep, [x1, #1216]
    fmla        v31.8h, v1.8h, v8.h[2]
    prfm        pldl1keep, [x2, #256]

    ld1         {{ v0.4s, v1.4s, v2.4s, v3.4s }}, [x1], #64
    ins         v8.h[0], v8.h[3]
    ld1         {{ v4.4s, v5.4s, v6.4s, v7.4s }}, [x1], #64