laba_5_EVM

;asm lib sse
public maximum
public Summ
public Vpir

_TEXT segment


DllMain proc

mov rax, 2
ret 12
DllMain endp


maximum proc export


movups xmm0, [rcx]
movups xmm1, [rdx]
maxps xmm0, xmm1

movups [r9], xmm0;
sub rcx, 16
add r9, 16
movups [r9], xmm0;
;movups [rax], xmm0

ret
maximum endp


Summ proc export
movups xmm0, [rcx]
movups xmm1, [rdx]
mov rax, 8

  add_loop:
    sub rax, 4
    addps xmm0, xmm1
    sub rcx, 16
    sub rdx, 16
    movups [r9], xmm0
    add r9, 16

    cmp rax, 0
    jnle add_loop
  ;movups [r9], xmm0
ret
Summ endp

Vpir proc export

movups xmm0, [rcx] ;S
movups xmm1, [rdx] ;V
movups xmm2, [r8]; 3


mov rax, 8

  loop2:
    sub rax, 4
    mulps xmm0, xmm1
    divps xmm0, xmm2


    sub rcx, 16
    sub rdx, 16

    movups [r9], xmm0
    add r9, 16

    cmp rax, 0
    jnle loop2

ret
Vpir endp


_TEXT ends
end


//client c++
#include <Windows.h>
#include <iostream>
using namespace std;

typedef float* (WINAPI* Summ)(float*, float*, int, float*);
typedef float* (WINAPI* Vpd)(float*, float*, float*, float*);

// v = S * H / 3
float* Vpiramid(float* square, float* height, float* three, float* res) {
    for (int i = 0; i < 3; i++)
    {
        res[i] = square[i] * height[i] / three[i];
    }
    return res;
}

int main()
{

    HMODULE hModule, hModule1;
    hModule = LoadLibrary(TEXT("asm_sse.dll"));
    if (hModule == NULL) return -1;

    Summ sum = (Summ)GetProcAddress(hModule, "Summ");

    Vpd Vpir = (Vpd)GetProcAddress(hModule, "Vpir");

    int size = 8;
    float square[] = { 3.5, 2.2, 1.6, 2.2, 8.1, 4.1, 9.4, 10.2};
    float height[] = { 6.2, 7.6, 9.1, 2.2, 8.1, 4.1, 9.4, 10.2 };
    float three[] = { 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0 };
    float norm[] = { 0,0,0,0,0,0,0,0 };



    float result[] = { 2.1, 1.2,   2.3, 4.4,   6.5, 10.6, 2.7, 8.8};
    float notres[] = { 3.1, 11.2, 11.3, 11.4, 11.5, 11.6, 11.7, 11.8 };
    float n[] = {0, 0, 0, 0, 0, 0, 0, 0};

    /*sum(result, notres, 8, n);
    for (int i = 0; i < 8; i++)
    {
        cout << result[i] << endl;
    }
    cout << "\n\n\n" << endl;*/

    int time_start_1 = clock();
    //Vpir(square, height, three, norm);
    for (int i = 0; i < 100000000; i++)
    {
        Vpir(square, height, three, norm);
    }
    /*for (int i = 0; i < size; i++)
    {
        cout << norm[i] << endl;
    }*/

    int time_end_1 = clock();
    cout << "asm time: " << time_end_1 - time_start_1 << endl;


    int time_start_2 = clock();
    for (int i = 0; i < 100000000; i++)
    {
        Vpiramid(square, height, three, norm);
    }
    int time_end_2 = clock();
    cout << "c++ time: " << time_end_2 - time_start_2 << endl;

    FreeLibrary(hModule);
    //FreeLibrary(hModule1);
}