#include <iostream>
#include <ctime>
#define  N 100*1024*256using namespace std;int main(int argc, char* argv[])
{float *a;float *b;a=(float*)malloc(N*sizeof(float));b=(float*)malloc(N*sizeof(float));for (unsigned long i=0;i<N;++i){a[i]=rand();}unsigned long n=N*sizeof(float);void *tbuf;tbuf=(void*)malloc(2048);time_t start,end;start=clock();//    memcpy(b,a,N*sizeof(float));
    __asm {mov esi, [a]mov ecx, nmov ebx, ecxshr ebx, 11 // 2048 bytes at a time
            mov edi, [b]loop2k: // Copy 2k into temporary buffer
            push edimov edi, tbufmov ecx, 2048shr ecx, 6loopMemToL1:prefetchnta 64[esi] // Prefetch next loop, non-temporalprefetchnta 96[esi]movq mm1,  0[esi] // Read in source datamovq mm2,  8[esi]movq mm3, 16[esi]movq mm4, 24[esi]movq mm5, 32[esi]movq mm6, 40[esi]movq mm7, 48[esi]movq mm0, 56[esi]movq  0[edi], mm1 // Store into L1movq  8[edi], mm2movq 16[edi], mm3movq 24[edi], mm4movq 32[edi], mm5movq 40[edi], mm6movq 48[edi], mm7movq 56[edi], mm0add esi, 64add edi, 64dec ecxjnz loopMemToL1pop edi // Now copy from L1 to system memory
            push esimov esi, tbufmov ecx, 2048shr ecx, 6loopL1ToMem:movq mm1, 0[esi] // Read in source data from L1movq mm2, 8[esi]movq mm3, 16[esi]movq mm4, 24[esi]movq mm5, 32[esi]movq mm6, 40[esi]movq mm7, 48[esi]movq mm0, 56[esi]movntq 0[edi], mm1 // Non-temporal storesmovntq 8[edi], mm2movntq 16[edi], mm3movntq 24[edi], mm4movntq 32[edi], mm5movntq 40[edi], mm6movntq 48[edi], mm7movntq 56[edi], mm0add esi, 64add edi, 64dec ecxjnz loopL1ToMempop esi // Do next 2k block
            dec ebxjnz loop2kemms;}end=clock();cout<<end-start<<endl;for (unsigned long i=N-10;i<N;i++){cout<<a[i]<<"  "<<b[i]<<endl;}free(a);free(b);free(tbuf);system("pause");return 0;
}