csyby/t2.cpp

#include <arm_neon.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>

#define SIZE 1024

// NEON<4F>Ż<EFBFBD><C5BB><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>ӷ<EFBFBD><D3B7><EFBFBD><EFBFBD><EFBFBD>
void vector_add_optimized(float* A, float* B, float* C, int size) {
    for (int i = 0; i < size; i += 4) {
        // <20><><EFBFBD><EFBFBD>A<EFBFBD><41>B<EFBFBD><42>4<EFBFBD><34><EFBFBD><EFBFBD><EFBFBD><EFBFBD>Ԫ<EFBFBD>ص<EFBFBD>NEON<4F>Ĵ<EFBFBD><C4B4><EFBFBD><EFBFBD><EFBFBD>
        float32x4_t a = vld1q_f32(A + i);
        float32x4_t b = vld1q_f32(B + i);
        // ִ<><D6B4><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>ӷ<EFBFBD>
        float32x4_t c = vaddq_f32(a, b);
        // <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>C
        vst1q_f32(C + i, c);
    }
}

int main() {
    // <20><><EFBFBD><EFBFBD><EFBFBD>ڴ沢<DAB4><E6B2A2>ʼ<EFBFBD><CABC><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
    float *A = (float*)malloc(SIZE * sizeof(float));
    float *B = (float*)malloc(SIZE * sizeof(float));
    float *C = (float*)malloc(SIZE * sizeof(float));

    // <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD>ʼ<EFBFBD><CABC><EFBFBD><EFBFBD><EFBFBD><EFBFBD>A<EFBFBD><41>B
    srand((unsigned int)time(NULL));
    for (int i = 0; i < SIZE; ++i) {
        A[i] = rand() % 100;
        B[i] = rand() % 100;
    }

    // <20><>ʱ<EFBFBD><CAB1>ʼ
    clock_t start = clock();
    // ִ<><D6B4>NEON<4F>Ż<EFBFBD><C5BB><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>ӷ<EFBFBD>
    vector_add_optimized(A, B, C, SIZE);
    // <20><>ʱ<EFBFBD><CAB1><EFBFBD><EFBFBD>
    clock_t end = clock();

    // <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>ʱ<EFBFBD><CAB1>
    printf("NEON<EFBFBD>Ż<EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>ӷ<EFBFBD><EFBFBD><EFBFBD>ʱ: %lf <20><><EFBFBD><EFBFBD>\n", 1000.0 * (end - start) / CLOCKS_PER_SEC);

    // <20>ͷ<EFBFBD><CDB7>ڴ<EFBFBD>
    free(A);
    free(B);
    free(C);

    return 0;
}