mov eax, sz
        mov esi, src
        prefetcht1 [esi + eax - 256]
        prefetcht1 [esi + eax - 192]
        sub eax, 128
        mov edi, dest
lab_1: prefetcht1 [esi + eax - 256]
        movaps xmm0, [esi + eax]
        movaps xmm1, [esi + eax + 16]
        movaps [edi + eax], xmm0
        movaps xmm2, [esi + eax + 32]
        movaps [edi + eax + 16], xmm1
        movaps xmm3, [esi + eax + 48]
        movaps [edi + eax + 32], xmm2
        prefetcht1 [esi + eax - 192]
        movaps xmm4, [esi + eax + 64]
        movaps [edi + eax + 48], xmm3
        movaps xmm5, [esi + eax + 80]
        movaps [edi + eax + 64], xmm4
        movaps xmm6, [esi + eax + 96]
        movaps [edi + eax + 80], xmm5
        movaps xmm7, [esi + eax + 112]
        movaps [edi + eax + 96], xmm6
        movaps [edi + eax + 112], xmm7
        sub eax, 128
        jnc lab_1