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

#define M  27
#define N  25
#define Q  12
//A(M*N),B(N*Q),C(M*Q),transposed(Q*N)对于矩阵维度的说明

void transposeMatrix(float** matrix, float** transposed) {
    for (int i = 0; i < N; i++) {
        for (int j = 0; j < Q; j++) {
            transposed[j][i] = matrix[i][j];
        }
    }
}

void matmul_optimized(float** A, float** B, float** C) {
    for (int i = 0; i < M; i++) {
        for (int j = 0; j < Q; j++) {
            float sum = 0.0f;
            for (int k = 0; k < N; k += 4) 
            {
                float32x4_t vecA, vecB, vecC;
                if (k + 4 <= N) 
                {
                    // 加载A和B的4个元素，进行向量化计算
                    vecA = vld1q_f32(&A[i][k]);
                    vecB = vld1q_f32(&B[j][k]);
                    // 向量化乘法并累加结果
                    vecC = vmulq_f32(vecA, vecB);
                    sum += vgetq_lane_f32(vecC, 0) + vgetq_lane_f32(vecC, 1) +
                           vgetq_lane_f32(vecC, 2) + vgetq_lane_f32(vecC, 3);
                } 
                else 
                {
                    // 处理剩余的元素
                    for (int m = k; m < N; m++) 
                    {
                        sum += A[i][m] * B[j][m];
                    }
                }
            }
            C[i][j] = sum;
        }
    }
}
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; 


	//动态分配内存
    float** denseMatrixA = (float**)malloc(M * sizeof(float*));
    float** denseMatrixB = (float**)malloc(N * sizeof(float*));
    float** C = (float**)malloc(M * sizeof(float*));
    float** transposed = (float**)malloc(Q * sizeof(float*));

    for (int i = 0; i < M; i++) {
        denseMatrixA[i] = (float*)malloc(N * sizeof(float));
        C[i] = (float*)malloc(Q * sizeof(float));
    }
    for (int i = 0; i < N; ++i) 
    {
        denseMatrixB[i] = (float*)malloc(Q * sizeof(float));
    }
    for (int i = 0; i < Q; ++i) 
    {
        transposed[i] = (float*)malloc(N * sizeof(float));
    }

	// 	实现稀疏矩阵转换为普通矩阵
    for(int i=0;i<M;i++)
    {
        for(int j=0;j<N;j++)
        {
            denseMatrixA[i][j]=0;
        }
    }

    for(int i=0;i<N;i++)
    {
        for(int j=0;j<Q;j++)
        {
            denseMatrixB[i][j]=0;
        }
    }
	for(int i = 0; i < A_nonZeroCount; i++) 
	{
		int row = A_rowIndex[i];
		int col = A_colIndex[i];
		denseMatrixA[row][col] = A_values[i];
	}
    	for(int i = 0; i < B_nonZeroCount; i++) 
	{
		int row = B_rowIndex[i];
		int col = B_colIndex[i];
		denseMatrixB[row][col] = B_values[i];
	}
	 
     //打印常规矩阵
     printf("A的常规矩阵为：\n");
    for(int i=0;i<M;i++)
    {
        for(int j=0;j<N;j++)
        {
            printf("%.1f ",denseMatrixA[i][j]);
        }
        printf("\n");
    }
     printf("B的常规矩阵为：\n");
    for(int i=0;i<N;i++)
    {
        for(int j=0;j<Q;j++)
        {
            printf("%.1f ",denseMatrixB[i][j]);
        }
        printf("\n");
    }

    //为了之后的运用NEON来相乘，需要对B进行转置
    transposeMatrix(denseMatrixB,transposed) ;

	clock_t start = clock();
    matmul_optimized(denseMatrixA,transposed,C);
    clock_t end = clock();

    // 计算并输出优化的稀疏矩阵乘法的时间
    double time_spent = double(end - start) / CLOCKS_PER_SEC;
    printf("当矩阵A的维度为%d*%d,矩阵B的维度为%d*%d时，优化的稀疏矩阵乘法的用时：%lf秒\n", M,N,N,Q,time_spent);

    // 释放动态分配的内存
    for (int i = 0; i < M; ++i) {
        free(denseMatrixA[i]);free(C[i]);
    }
    for (int i = 0; i < N ;++i) {
        free(denseMatrixB[i]);
    }
    for (int i = 0; i < Q ;++i) {
        free(transposed[i]);
    }
    free(denseMatrixA); free(denseMatrixB); free(C);free(transposed);
}