From 80c4199cb3a99eb7242dbc90f155fac45a0fe195 Mon Sep 17 00:00:00 2001 From: pypzov7ui <2137546866@qq.com> Date: Fri, 29 Nov 2024 11:11:25 +0800 Subject: [PATCH] ADD file via upload --- step4.c | 56 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 56 insertions(+) create mode 100644 step4.c diff --git a/step4.c b/step4.c new file mode 100644 index 0000000..b9cc827 --- /dev/null +++ b/step4.c @@ -0,0 +1,56 @@ +#include +#include +#include +#include + +#define SIZE 1024 + +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 sum = vdupq_n_f32(0); + for (int k = 0; k < n; k += 4) { + float32x4_t vecA = vld1q_f32(&A[i][k]); + float32x4_t vecB = vld1q_f32(&B[k][j]); + sum = vmlaq_f32(sum, vecA, vecB); + } + C[i][j] = vgetq_lane_f32(sum, 0) + vgetq_lane_f32(sum, 1) + vgetq_lane_f32(sum, 2) + vgetq_lane_f32(sum, 3); // 提取结果 + } + } +} + +int main() { + float** A = (float**)malloc(SIZE * sizeof(float*)); + float** B = (float**)malloc(SIZE * sizeof(float*)); + float** C = (float**)malloc(SIZE * sizeof(float*)); + for (int i = 0; i < SIZE; i++) { + A[i] = (float*)malloc(SIZE * sizeof(float)); + B[i] = (float*)malloc(SIZE * sizeof(float)); + C[i] = (float*)malloc(SIZE * sizeof(float)); + } + + for (int i = 0; i < SIZE; i++) { + for (int j = 0; j < SIZE; j++) { + A[i][j] = rand() % 100; + B[i][j] = rand() % 100; + } + } + + clock_t start = clock(); + matmul_optimized(A, B, C, SIZE); + clock_t end = clock(); + + double time_taken = (double)(end - start) / CLOCKS_PER_SEC; + printf("优化后的稠密矩阵乘法: %f seconds\n", time_taken); + + for (int i = 0; i < SIZE; i++) { + free(A[i]); + free(B[i]); + free(C[i]); + } + free(A); + free(B); + free(C); + + return 0; +}