// mul a: v0, v1, v2, v3, v4, v5 b: v7
// load a: v5(d5/x5), v6(d6,x6), v7(d7,x7), v8(d8, x8)
// load b: v9(d9/x9)

fmla        v16.4s, v0.4s, v4.s[0]
ldr         d5, [x1], #8
fmla        v17.4s, v1.4s, v4.s[0]
ldr         d9, [x2], #8
fmla        v18.4s, v2.4s, v4.s[0]
ldr         x5, [x1], #8
fmla        v19.4s, v3.4s, v4.s[0]
ldr         x9, [x2], #8
fmla        v20.4s, v0.4s, v4.s[1]
ldr         d6, [x1], #8
fmla        v21.4s, v1.4s, v4.s[1]
ldr         x6, [x1], #8
fmla        v22.4s, v2.4s, v4.s[1]
ldr         d7, [x1], #8
fmla        v23.4s, v3.4s, v4.s[1]
ldr         x7, [x1], #8

fmla        v24.4s, v0.4s, v4.s[2]
ldr         d8, [x1], #8
fmla        v25.4s, v1.4s, v4.s[2]
ldr         x8, [x1], #8
fmla        v26.4s, v2.4s, v4.s[2]
ins         v5.d[1], x5
fmla        v27.4s, v3.4s, v4.s[2]
ins         v6.d[1], x6
fmla        v28.4s, v0.4s, v4.s[3]
ins         v7.d[1], x7
fmla        v29.4s, v1.4s, v4.s[3]
ins         v8.d[1], x8
fmla        v30.4s, v2.4s, v4.s[3]
ins         v9.d[1], x9
fmla        v31.4s, v3.4s, v4.s[3]

// mul a: v5, v6, v7, v8 b: v9
// load a: v0(d0/x5), v1(d1,x6), v2(d2,x7), v3(d3, x8)
// load b: v4(d4/x9)

fmla        v16.4s, v5.4s, v9.s[0]
ldr         d0, [x1], #8
fmla        v17.4s, v6.4s, v9.s[0]
ldr         d4, [x2], #8
fmla        v18.4s, v7.4s, v9.s[0]
ldr         x5, [x1], #8
fmla        v19.4s, v8.4s, v9.s[0]
ldr         x9, [x2], #8
fmla        v20.4s, v5.4s, v9.s[1]
ldr         d1, [x1], #8
fmla        v21.4s, v6.4s, v9.s[1]
ldr         x6, [x1], #8
fmla        v22.4s, v7.4s, v9.s[1]
ldr         d2, [x1], #8
fmla        v23.4s, v8.4s, v9.s[1]
ldr         x7, [x1], #8

fmla        v24.4s, v5.4s, v9.s[2]
ldr         d3, [x1], #8
fmla        v25.4s, v6.4s, v9.s[2]
ldr         x8, [x1], #8
fmla        v26.4s, v7.4s, v9.s[2]
ins         v0.d[1], x5
fmla        v27.4s, v8.4s, v9.s[2]
ins         v1.d[1], x6
fmla        v28.4s, v5.4s, v9.s[3]
ins         v2.d[1], x7
fmla        v29.4s, v6.4s, v9.s[3]
ins         v3.d[1], x8
fmla        v30.4s, v7.4s, v9.s[3]
ins         v4.d[1], x9
fmla        v31.4s, v8.4s, v9.s[3]