دوستان کسی میتونه این کد رو توضیح اجمالی بده

[Iman_a13]

// ConsoleApplication_avx.cpp : This file contains the 'main' function. Program execution begins and ends there.

//





#include <random>       // prng

#include <cstdint>      // uint32_t

#include <iostream>     // std::cout

#include <immintrin.h>  // AVX intrinsics

#include <stdint.h>





// timers distributed with this book





#include "C:\Users\Studio\Desktop\ConsoleApplication_avx\he  lper_HPC.hpp"





void init(float* data, uint64_t length) {





    std::mt19937 engine(42);

    std::uniform_real_distribution<float> density(-1, 1);





    for (uint64_t i = 0; i < length; i++)

        data[i] = density(engine);

}





inline float hsum_sse3(__m128 v) {

    __m128 shuf = _mm_movehdup_ps(v);        // broadcast elements 3,1 to 2,0

    __m128 maxs = _mm_add_ps(v, shuf);

    shuf = _mm_movehl_ps(shuf, maxs); // high half -> low half

    maxs = _mm_add_ss(maxs, shuf);

    return        _mm_cvtss_f32(maxs);

}





inline float hsum_avx(__m256 v) {

    __m128 lo = _mm256_castps256_ps128(v);   // low 128

    __m128 hi = _mm256_extractf128_ps(v, 1); // high 128

    lo = _mm_add_ps(lo, hi);          // max the low 128

    return hsum_sse3(lo);                    // and inline the sse3 version

}





void plain_dmm(float* A,

    float* B,

    float* C,

    uint64_t M,

    uint64_t L,

    uint64_t N,

    bool parallel) {





#pragma omp parallel for collapse(2) if(parallel)

    for (uint64_t i = 0; i < M; i++)

        for (uint64_t j = 0; j < N; j++) {

            float accum = float(0);

            for (uint64_t k = 0; k < L; k++)

                accum += A[i * L + k] * B[j * L + k];

            C[i * N + j] = accum;

        }

}





void avx_dmm(float* A,

    float* B,

    float* C,

    uint64_t M,

    uint64_t L,

    uint64_t N,

    bool parallel) {





#pragma omp parallel for collapse(2) if(parallel)

    for (uint64_t i = 0; i < M; i++)

        for (uint64_t j = 0; j < N; j++) {





            __m256 X = _mm256_setzero_ps();

            for (uint64_t k = 0; k < L; k += 8) {

                const __m256 AV = _mm256_load_ps(A + i * L + k);

                const __m256 BV = _mm256_load_ps(B + j * L + k);

                X = _mm256_add_ps(X, _mm256_mul_ps(AV, BV));

            }





            C[i * N + j] = hsum_avx(X);

        }

}





void avx_dmm_unroll_2(float* A,

    float* B,

    float* C,

    uint64_t M,

    uint64_t L,

    uint64_t N,

    bool parallel) {





#pragma omp parallel for collapse(2) if(parallel)

    for (uint64_t i = 0; i < M; i++)

        for (uint64_t j = 0; j < N; j++) {





            __m256 X = _mm256_setzero_ps();

            __m256 Y = _mm256_setzero_ps();

            for (uint64_t k = 0; k < L; k += 16) {

                const __m256 AVX = _mm256_load_ps(A + i * L + k + 0);

                const __m256 BVX = _mm256_load_ps(B + j * L + k + 0);

                const __m256 AVY = _mm256_load_ps(A + i * L + k + 8);

                const __m256 BVY = _mm256_load_ps(B + j * L + k + 8);

                X = _mm256_add_ps(X, _mm256_mul_ps(AVX, BVX));

                Y = _mm256_add_ps(X, _mm256_mul_ps(AVY, BVY));

            }





            C[i * N + j] = hsum_avx(X) + hsum_avx(Y);

        }

}





int main() {





    const uint64_t M = 1UL << 10;

    const uint64_t L = 1UL << 11;

    const uint64_t N = 1UL << 12;





    TIMERSTART(alloc_memory)

        auto A = static_cast<float*>(_mm_malloc(M * L * sizeof(float), 32));

    auto B = static_cast<float*>(_mm_malloc(N * L * sizeof(float), 32));

    auto C = static_cast<float*>(_mm_malloc(M * N * sizeof(float), 32));

    TIMERSTOP(alloc_memory)





        TIMERSTART(init)

        init(A, M * L);

    init(B, N * L);

    TIMERSTOP(init)





        TIMERSTART(plain_dmm_single)

        plain_dmm(A, B, C, M, L, N, false);

    TIMERSTOP(plain_dmm_single)





        TIMERSTART(plain_dmm_multi)

        plain_dmm(A, B, C, M, L, N, true);

    TIMERSTOP(plain_dmm_multi)





        TIMERSTART(avx_dmm_single)

        avx_dmm(A, B, C, M, L, N, false);

    TIMERSTOP(avx_dmm_single)





        TIMERSTART(avx_dmm_multi)

        avx_dmm(A, B, C, M, L, N, true);

    TIMERSTOP(avx_dmm_multi)





        TIMERSTART(avx_dmm_unroll_2_single)

        avx_dmm_unroll_2(A, B, C, M, L, N, false);

    TIMERSTOP(avx_dmm_unroll_2_single)





        TIMERSTART(avx_dmm_unroll_2_multi)

        avx_dmm_unroll_2(A, B, C, M, L, N, true);

    TIMERSTOP(avx_dmm_unroll_2_multi)





        TIMERSTART(free_memory)

        _mm_free(A);

    _mm_free(B);

    _mm_free(C);

    TIMERSTOP(free_memory)

}