#define  row_col_size_local  64
 #define  rows_per_local  32

__kernel void sgemm(__global float *A, __global float *B, __global float *C, int M, int N, int K) 
{
 const int local_row = get_local_id(0); // local row ID 
 const int local_col = get_local_id(1); // local col ID

 const int global_row = row_col_size_local*get_group_id(0) + local_row;  // c행렬의 row
 const int global_col = row_col_size_local*get_group_id(1) + local_col;   // c행렬의 col
 //printf("local_id0: %d, local_id1: %d, global_id0: %d, global_id1:%d\n",get_local_id(0),get_local_id(1),get_group_id(0),get_group_id(1));

 const int RTS = row_col_size_local/rows_per_local;

 //int rows_per_local_var = rows_per_local;

 __local float Asub[row_col_size_local][row_col_size_local];
 __local float Bsub[row_col_size_local][row_col_size_local];

//int global_row_end = global_row+ ((rows_per_local-1)*RTS) ;
//int global_col_end = global_col;
 

 	float temp_val[rows_per_local];
 	for(int w=0; w<rows_per_local; w++)
 	{
		temp_val[w]=0.0f;
 	}

 	const int num_tiles = ((K+row_col_size_local-1)/row_col_size_local) ;  //32개씩 몇번 돌건지 정의
 	for(int t=0;t<num_tiles;t++)
 	{
		const int t_row = row_col_size_local * t  + local_row;
		const int t_col = row_col_size_local * t  + local_col;
		
		
		for(int w=0; w<rows_per_local; w++)
		{
			if(global_row + (w*RTS) >=M || t_col >=K)
			{
				Asub[local_row+w*RTS][local_col] = 0.0f; 
			}
			else
			{
				Asub[local_row + w*RTS][local_col] = A[(global_row + w*RTS)*K + t_col];
			}
		
			if(t_row + (w*RTS) >=K || global_col >=N)
			{
				Bsub[local_row+w*RTS][local_col] = 0.0f; 
			}
			else
			{
				Bsub[local_row + w*RTS][local_col] = B[(t_row + w*RTS)*N + global_col];
			}
			//Asub[local_row+ w*RTS][local_col] = A[(global_row+ w*RTS) * K + t_col];
			//Bsub[local_row+ w*RTS][local_col] = B[(t_row+ w*RTS) * N + global_col];
		}
		barrier(CLK_LOCAL_MEM_FENCE);

		for(int k=0; k<row_col_size_local; k++)
		{
			for(int w=0; w<rows_per_local; w++)
			{
				temp_val[w] += Asub[local_row+w*RTS][k] * Bsub[k][local_col];
			}
		}
		barrier(CLK_LOCAL_MEM_FENCE);
	}


 	for(int w=0; w<rows_per_local; w++)
	{
		if(global_row + w*RTS < M && global_col < N)
		{
			C[(global_row+ w*RTS)*N + global_col] = temp_val[w];
 		}
	}
}

 
/* 8192 8192 8192 는 15000 달성
 #define  row_col_size_local  64
 #define  rows_per_local  32

__kernel void sgemm(__global float *A, __global float *B, __global float *C, int M, int N, int K) 
{
 const int local_row = get_local_id(0); // local row ID 
 const int local_col = get_local_id(1); // local col ID

 const int global_row = row_col_size_local*get_group_id(0) + local_row;  // c행렬의 row
 const int global_col = row_col_size_local*get_group_id(1) + local_col;   // c행렬의 col
 //printf("local_id0: %d, local_id1: %d, global_id0: %d, global_id1:%d\n",get_local_id(0),get_local_id(1),get_group_id(0),get_group_id(1));

 const int RTS = row_col_size_local/rows_per_local;

 //int rows_per_local_var = rows_per_local;

 __local float Asub[row_col_size_local][row_col_size_local];
 __local float Bsub[row_col_size_local][row_col_size_local];

//int global_row_end = global_row+ ((rows_per_local-1)*RTS) ;
//int global_col_end = global_col;
 

 float temp_val[rows_per_local];
 for(int w=0; w<rows_per_local; w++)
 {
	temp_val[w]=0.0f;
 }

 const int num_tiles = (K/row_col_size_local) ;  //32개씩 몇번 돌건지 정의
 for(int t=0;t<num_tiles;t++)
 {
	const int t_row = row_col_size_local * t  + local_row;
	const int t_col = row_col_size_local * t  + local_col;
	
	for(int w=0; w<rows_per_local; w++)
	{
		Asub[local_row+ w*RTS][local_col] = A[(global_row+ w*RTS) * K + t_col];
		Bsub[local_row+ w*RTS][local_col] = B[(t_row+ w*RTS) * N + global_col];
	}
	barrier(CLK_LOCAL_MEM_FENCE);

	for(int k=0; k<row_col_size_local; k++)
	{
		for(int w=0; w<rows_per_local; w++)
		{
			temp_val[w] += Asub[local_row+w*RTS][k] * Bsub[k][local_col];
		}
	}
	barrier(CLK_LOCAL_MEM_FENCE);
 }


 for(int w=0; w<rows_per_local; w++)
 {
	C[(global_row+ w*RTS)*N + global_col] = temp_val[w];
 }
}
*/


/* 679 성능
__kernel void sgemm(__global float *A, __global float *B, __global float *C, int M, int N, int K) 
{
 const int row = get_local_id(0); // local row ID 
 const int col = get_local_id(1); // local col ID
 const int global_row = 32*get_group_id(0) + row;  // c행렬의 row
 const int global_col = 32*get_group_id(1) + col;   // c행렬의 col

 __local float Asub[32][32];
 __local float Bsub[32][32];

 float temp_val = 0.0f;
 const int num_tiles = K/32;
 
 for(int t=0; t<num_tiles; t++)
 {
	const int t_row = 32*t + row;
	const int t_col = 32*t + col;
	
	Asub[row][col] = A[global_row * K + t_col];
	Bsub[row][col] = B[t_row*N + global_col];

	barrier(CLK_LOCAL_MEM_FENCE);

 	for(int k=0; k<32; k++)
 	{
		temp_val += Asub[row][k] * Bsub[k][col];
	}
	barrier(CLK_LOCAL_MEM_FENCE);
 }

 C[global_row*N + global_col]= temp_val;
}
*/	


/* 최초 코드
// super super slow sgemm kernel by heehoon
__kernel void sgemm(__global float *A, __global float *B, __global float *C, int M, int N, int K) {
  int i = get_global_id(0); // row index of C
  int j = get_global_id(1); // column index of C
  if (i >= M || j >= N) return; // boundary check

  C[i * N + j] = 0;
  for (int k = 0; k < K; k++) {
    C[i * N + j] += A[i * K + k] * B[k * N + j];
  }
}
*/