对于#include <stdio.h> #include <arm_neon.h> static void matrix_mul_asm(uint16_t **aa,uint16_t **bb,uint16_t **cc) { uint16_t *a=(uint16_t *)aa; uint16_t *b=(uint16_t )bb; uint16_t c=(uint16_t )cc; asm volatile ( "ld4 {v0.4h-v3.4h},[%0]\n" "ld4 {v4.4h,v5.4h,v6.4h,v7.4h},[%1]\n" "mul v3.4h,v3.4h,v7.4h\n" "mul v2.4h,v2.4h,v6.4h\n" "mul v1.4h,v1.4h,v5.4h\n" "mul v0.4h,v0.4h,v4.4h\n" "st4 {v0.4h,v1.4h,v2.4h,v3.4h},[%2]\n" :"+r"(a),"+r"(b),"+r"(c) : :"cc","memory","v0","v1","v2","v3","v4","v5","v6","v7" ); } int main() { uint16_t aa[4][4]={ {1,8,2,4}, //自己学号的高四位，比如学号为19374331，此处应为{1,9,3,7}, {1,0,5,5}, //自己学号的低四位，比如学号为19374331，此处应为{4,3,3,1}, {3,6,8,1}, {2,6,7,1} }; uint16_t bb[4][4]={ {1,3,5,7}, {2,4,6,8}, {2,5,7,9}, {5,2,7,1} }; uint16_t cc[4][4]={0}; int i,j; matrix_mul_asm((uint16_t **)aa,(uint16_t **)bb,(uint16_t **)cc); for(i=0;i<4;i++) for(j=0;j<4;j++) printf("NO 18241055,cc[%1u][ %1u] out is %11u \n",i,j,cc[i][j]); //19374331位置处填写自己学号 return 0; }用NEON SIMD的intrinsics方式重新一遍

下面是使用NEON SIMD的intrinsics方式实现的矩阵乘法代码：

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

static void matrix_mul_simd(uint16_t **aa, uint16_t **bb, uint16_t **cc) {
    uint16x8_t a[4], b[4];
    uint16x8_t c[4][4] = {0};
    
    for (int i = 0; i < 4; i++) {
        a[i] = vld1q_u16(aa[i]);
    }
    for (int i = 0; i < 4; i++) {
        b[i] = vld1q_u16(bb[i]);
        c[0][i] = vmulq_u16(a[0], vdupq_n_u16(bb[i][0]));
        c[1][i] = vmulq_u16(a[1], vdupq_n_u16(bb[i][1]));
        c[2][i] = vmulq_u16(a[2], vdupq_n_u16(bb[i][2]));
        c[3][i] = vmulq_u16(a[3], vdupq_n_u16(bb[i][3]));
    }
    for (int i = 0; i < 4; i++) {
        c[i][0] = vaddq_u16(c[i][0], vmulq_u16(a[i], b[0]));
        c[i][1] = vaddq_u16(c[i][1], vmulq_u16(a[i], b[1]));
        c[i][2] = vaddq_u16(c[i][2], vmulq_u16(a[i], b[2]));
        c[i][3] = vaddq_u16(c[i][3], vmulq_u16(a[i], b[3]));
    }
    
    for (int i = 0; i < 4; i++) {
        vst1q_u16(cc[i], c[i][0]);
        vst1q_u16(cc[i] + 8, c[i][1]);
        vst1q_u16(cc[i] + 16, c[i][2]);
        vst1q_u16(cc[i] + 24, c[i][3]);
    }
}

int main() {
    uint16_t aa[4][4] = {
        {1, 8, 2, 4}, //自己学号的高四位，比如学号为19374331，此处应为{1,9,3,7},
        {1, 0, 5, 5}, //自己学号的低四位，比如学号为19374331，此处应为{4,3,3,1},
        {3, 6, 8, 1},
        {2, 6, 7, 1}
    };
    uint16_t bb[4][4] = {
        {1, 3, 5, 7},
        {2, 4, 6, 8},
        {2, 5, 7, 9},
        {5, 2, 7, 1}
    };
    uint16_t cc[4][4] = {0};
    int i, j;
    
    matrix_mul_simd((uint16_t **) aa, (uint16_t **) bb, (uint16_t **) cc);
    
    for (i = 0; i < 4; i++) {
        for (j = 0; j < 4; j++) {
            printf("NO 18241055,cc[%1u][ %1u] out is %11u \n", i, j, cc[i][j]); //19374331位置处填写自己学号
        }
    }
    
    return 0;
}

在这个实现中，我们使用了NEON SIMD的intrinsics，首先将输入矩阵a和b分别加载到了8个uint16x8_t向量变量a和b中，然后使用vmulq_u16函数分别计算a和b的每一行的乘积，将结果存储到c矩阵中。最后，使用vaddq_u16函数将结果相加，并使用vst1q_u16函数将结果存储到输出矩阵cc中。这种方法比使用汇编代码更加容易理解和维护，并且可以通过编译器自动向量化来进一步优化性能。