#include <arm_neon.h>
#include <stdio.h>
#include <time.h>
#include <stdlib.h>
#include <string.h>
#define SIZE 1024
#define DENSE_MATRIX_SIZE 5

float** sparce_matmul_coo(float* A_values, int* A_rowIndex, int* A_colIndex, int A_nonZeroCount,
                          float* B_values, int* B_rowIndex, int* B_colIndex, int B_nonZeroCount);
void matmul_optimized(float** A, float** B, float** C, int n);
void print_matrix(float** m, int rows, int cols);
int main() {
	// 矩阵 A 的 COO 格式
	float A_values[] = {1, 2, 3, 4, 5}; 
	int A_rowIndex[] = {0, 0, 1, 2, 2}; 
	int A_colIndex[] = {0, 2, 1, 0, 2}; 
	int A_nonZeroCount = 5;
	// 矩阵 B 的 Coo 格式
	float B_values[] = {6, 8, 7, 9}; 
	int B_rowIndex[] = {0, 2, 1, 2}; 
	int B_colIndex[] = {0, 0, 1, 2}; 
	int B_nonZeroCount = 4;// 结果矩阵 C 的 Coo 格式
	float C_values[SIZE]; 
	int C_rowIndex[SIZE]; 
	int C_colIndex[SIZE]; 
	int C_nonZeroCount = 0;

	clock_t start = clock();
	float** ans = sparce_matmul_coo(A_values, A_rowIndex, A_colIndex, A_nonZeroCount,
	                  B_values, B_rowIndex, B_colIndex, B_nonZeroCount);
	clock_t end = clock();
	printf("优化的稀疏矩阵乘法时间：%lf\n", (double)(end-start) / CLOCKS_PER_SEC);
	print_matrix(ans, DENSE_MATRIX_SIZE, DENSE_MATRIX_SIZE);
}

float** sparce_matmul_coo(float* A_values, int* A_rowIndex, int* A_colIndex, int A_nonZeroCount,
                          float* B_values, int* B_rowIndex, int* B_colIndex, int B_nonZeroCount) {
	// 分配内存空间
	float** A = (float**)malloc(sizeof(float*) * DENSE_MATRIX_SIZE);
	float** B = (float**)malloc(sizeof(float*) * DENSE_MATRIX_SIZE);
	float** C = (float**)malloc(sizeof(float*) * DENSE_MATRIX_SIZE);
	int i;
	for(i=0; i<DENSE_MATRIX_SIZE; i++) {
		A[i] = (float*)malloc(sizeof(float) * DENSE_MATRIX_SIZE);
		B[i] = (float*)malloc(sizeof(float) * DENSE_MATRIX_SIZE);
		C[i] = (float*)malloc(sizeof(float) * DENSE_MATRIX_SIZE);
		memset(A[i], 0, sizeof(float) * DENSE_MATRIX_SIZE);
		memset(B[i], 0, sizeof(float) * DENSE_MATRIX_SIZE);
		memset(C[i], 0, sizeof(float) * DENSE_MATRIX_SIZE);
	}
	for (i=0; i<A_nonZeroCount; i++) A[A_rowIndex[i]][A_colIndex[i]] = A_values[i];
	for (i=0; i<B_nonZeroCount; i++) B[B_rowIndex[i]][B_colIndex[i]] = B_values[i];
	matmul_optimized(A, B, C, DENSE_MATRIX_SIZE);

	// 释放内存空间
	for(i=0; i<DENSE_MATRIX_SIZE; i++) {
		free(A[i]);
		free(B[i]);
	}
	free(A); free(B);
	return C;
}

void matmul_optimized(float** A, float** B, float** C, int n) {
	int i, j, k;
	float32x4_t vecA, vecB, vecC;
	for(i=0; i<n; i++)
	for(j=0; j<n; j++) {
		vecC = vdupq_n_f32(0.0);
		for(k=0; k<n; k+=4) {
			vecA = vld1q_f32(&A[i][k]);
			vecB = vld1q_f32(&A[k][j]);
			vecC = vmlaq_f32(vecC, vecA, vecB);
		}
		C[i][j] = vgetq_lane_f32(vecC, 0) + vgetq_lane_f32(vecC, 1) + 
		          vgetq_lane_f32(vecC, 2) + vgetq_lane_f32(vecC, 3);
	}
}
void print_matrix(float** m, int rows, int cols) {
	int i, j;
	for(i=0; i<rows; i++) {
		for(j=0; j<cols; j++) {
			printf("%5.0f ", m[i][j]);
		}
		printf("\n");
	}
}