|
|
#include <stdio.h>
|
|
|
#include <stdlib.h>
|
|
|
#include <ctime>
|
|
|
// 包含NEON头文件，启用NEON指令
|
|
|
#include <arm_neon.h>
|
|
|
// 定义矩阵大小
|
|
|
#define SIZE 1024
|
|
|
// 基础的矩阵乘法函数
|
|
|
void matmul(float** A, float** B, float** C, int n) {
|
|
|
for (int i = 0; i < n; i++) {
|
|
|
for (int j = 0; j < n; j++) {
|
|
|
float sum = 0;
|
|
|
for (int k = 0; k < n; k++) {
|
|
|
sum += A[i][k] * B[k][j];
|
|
|
}
|
|
|
C[i][j] = sum;
|
|
|
}
|
|
|
}
|
|
|
}
|
|
|
// 使用NEON指令优化的矩阵乘法函数
|
|
|
void matmul_optimized(float** A, float** B, float** C, int n) {
|
|
|
for (int i = 0; i < n; i++) {
|
|
|
for (int j = 0; j < n; j++) {
|
|
|
float32x4_t vecC = vdupq_n_f32(0); // 初始化结果向量为0
|
|
|
for (int k = 0; k < n; k += 4) {
|
|
|
// 向量加载，将A和B的4个连续元素加载到float32x4_t类型的向量中
|
|
|
float32x4_t a_vec = vld1q_f32(&A[i][k]);
|
|
|
float32x4_t b_vec = vld1q_f32(&B[k][j]);
|
|
|
// 向量乘法和累加，使用vmlaq_f32完成对应元素相乘并累加到vecC中
|
|
|
vecC = vmlaq_f32(vecC, a_vec, b_vec);
|
|
|
}
|
|
|
// 向量还原，提取累加结果并存储到C[i][j]中
|
|
|
C[i][j] = vgetq_lane_f32(vecC, 0) + vgetq_lane_f32(vecC, 1) +
|
|
|
vgetq_lane_f32(vecC, 2) + vgetq_lane_f32(vecC, 3);
|
|
|
}
|
|
|
}
|
|
|
}
|
|
|
int main() {
|
|
|
// 动态分配两个输入矩阵A和B，以及结果矩阵C的内存
|
|
|
float** A = (float**)malloc(SIZE * sizeof(float*));
|
|
|
for (int i = 0; i < SIZE; i++) {
|
|
|
A[i] = (float*)malloc(SIZE * sizeof(float));
|
|
|
}
|
|
|
float** B = (float**)malloc(SIZE * sizeof(float*));
|
|
|
for (int i = 0; i < SIZE; i++) {
|
|
|
B[i] = (float*)malloc(SIZE * sizeof(float));
|
|
|
}
|
|
|
float** C = (float**)malloc(SIZE * sizeof(float*));
|
|
|
for (int i = 0; i < SIZE; i++) {
|
|
|
C[i] = (float*)malloc(SIZE * sizeof(float));
|
|
|
}
|
|
|
float** C_optimized = (float**)malloc(SIZE * sizeof(float*));
|
|
|
for (int i = 0; i < SIZE; i++) {
|
|
|
C_optimized[i] = (float*)malloc(SIZE * sizeof(float));
|
|
|
}
|
|
|
// 初始化矩阵数据，将A和B矩阵的每个元素随机初始化
|
|
|
for (int i = 0; i < SIZE; i++) {
|
|
|
for (int j = 0; j < SIZE; j++) {
|
|
|
A[i][j] = (float)(rand() % 100);
|
|
|
B[i][j] = (float)(rand() % 100);
|
|
|
}
|
|
|
}
|
|
|
// 测试基础矩阵乘法函数的运行时间
|
|
|
clock_t start_time_original = clock();
|
|
|
matmul(A, B, C, SIZE);
|
|
|
clock_t end_time_original = clock();
|
|
|
double elapsed_time_original = (double)(end_time_original - start_time_original) / CLOCKS_PER_SEC;
|
|
|
// 测试NEON优化后的矩阵乘法函数的运行时间
|
|
|
clock_t start_time_optimized = clock();
|
|
|
matmul_optimized(A, B, C_optimized, SIZE);
|
|
|
clock_t end_time_optimized = clock();
|
|
|
double elapsed_time_optimized = (double)(end_time_optimized - start_time_optimized) / CLOCKS_PER_SEC;
|
|
|
// 输出基础矩阵乘法的运行时间
|
|
|
printf("original time: %lf s\n", elapsed_time_original);
|
|
|
// 输出NEON优化后的矩阵乘法的运行时间
|
|
|
printf("NEON optimized time: %lf s\n", elapsed_time_optimized);
|
|
|
// 释放动态分配的内存空间
|
|
|
for (int i = 0; i < SIZE; i++) {
|
|
|
free(A[i]);
|
|
|
}
|
|
|
free(A);
|
|
|
for (int i = 0; i < SIZE; i++) {
|
|
|
free(B[i]);
|
|
|
}
|
|
|
free(B);
|
|
|
for (int i = 0; i < SIZE; i++) {
|
|
|
free(C[i]);
|
|
|
}
|
|
|
free(C);
|
|
|
for (int i = 0; i < SIZE; i++) {
|
|
|
free(C_optimized[i]);
|
|
|
}
|
|
|
free(C_optimized);
|
|
|
return 0;
|
|
|
} |
