#include "mat_mul.h"

#include <cstdio>
#include <cuda_runtime.h>

#define BLOCK_DIM 4 
#define TS 32
#define WPT 16 
#define RTS (TS/WPT)
#define CUDA_CALL(f)                                                           \
  {                                                                            \
    cudaError_t err = (f);                                                     \
    if (err != cudaSuccess) {                                                  \
      fprintf(stderr, "CUDA error at [%s:%d] %d %s\n", __FILE__, __LINE__,     \
              err, cudaGetErrorString(err));                                   \
      exit(1);                                                                 \
    }                                                                          \
  }

#define MAX_NUM_GPU 4
int num_devices = 0;
// n->K p->N m->M
__global__ void sgemm(float *A, float *B, float *C, int M, int N, int K) {
  const int row = threadIdx.x;//); // row index of C
  const int col = threadIdx.y;//); // column index of C
	const int global_row = (blockDim.x*WPT)* blockIdx.x + threadIdx.x;
	const int global_col = blockDim.y * blockIdx.y + threadIdx.y;
	__shared__ float subA[TS][TS];
	__shared__ float subB[TS][TS];

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

	const int num_tiles = (K+TS-1)/TS;
	for (int t = 0; t < num_tiles; t++) {
		for (int w = 0; w < WPT; w++) {
			const int tiledRow = TS*t + row;
			const int tiledCol = TS*t + col;
			/*if 			(M <= global_row) subA[row + w*RTS][col] = 0.0f;
			else*/

			if (K <= tiledCol) subA[row + w*RTS][col] = 0.0f;
			else if (M <= (global_row + w*RTS)) subA[row + w*RTS][col] = 0.0f;

			else 
				subA[row + w*RTS][col] = A[(global_row + w*RTS)*K + tiledCol];

			if			(N <= global_col) subB[row + w*RTS][col] = 0.0f;
			//else if	(K <= tiledRow) subB[row + w*RTS][col] = 0.0f;
			else if	(K <= (tiledRow + w*RTS)) subB[row + w*RTS][col] = 0.0f;

			else 
				subB[row + w*RTS][col] = B[(tiledRow + w*RTS)*N + global_col];
		}
		__syncthreads();
		//barrier(CLK_LOCAL_MEM_FENCE);
		for (int k = 0; k < TS; k++) {
			for (int w = 0; w < WPT; w++) {
				accum[w] += subA[row + w*RTS][k] * subB[k][col];
			}
		}
		//barrier(CLK_LOCAL_MEM_FENCE);
		__syncthreads();
	}
	for (int w = 0; w < WPT; w++) {
		if				(M <= (global_row + w*RTS));// C[(global_row + w*RTS)*N + global_col] = 0.0f;
		else if		(N <= global_col);// C[(global_row + w*RTS)*N + global_col] = 0.0f;
		else 
			C[(global_row + w*RTS)*N + global_col] = accum[w];
	}
	/*
	    __shared__ float A_tile[BLOCK_DIM][BLOCK_DIM];
    __shared__ float B_tile[BLOCK_DIM][BLOCK_DIM];

    float acc_sum{0};

    for (int tile_idx{0};
         tile_idx < ceilf(static_cast<float>(K) / BLOCK_DIM); ++tile_idx)
    {
        int i{blockIdx.y * blockDim.y + threadIdx.y};
        int j{tile_idx * blockDim.x + threadIdx.x};
        if ((i < M) && (j < K))
        {
            A_tile[threadIdx.y][threadIdx.x] = A[i * K + j];
        }
        else
        {
            A_tile[threadIdx.y][threadIdx.x] = 0;
        }
        i = tile_idx * blockDim.y + threadIdx.y;
        j = blockIdx.x * blockDim.x + threadIdx.x;
        if ((i < K) && (j < N))
        {
            B_tile[threadIdx.y][threadIdx.x] = B[i * N + j];
        }
        else
        {
            B_tile[threadIdx.y][threadIdx.x] = 0;
        }
        __syncthreads();
        for (int k{0}; k < BLOCK_DIM; ++k)
        {
            acc_sum += A_tile[threadIdx.y][k] * B_tile[k][threadIdx.x];
						printf("threadidx.y %d k %d threadidx.x %d k %d \n", threadIdx.y, k, threadIdx.x, k);
        }
        __syncthreads();
    }

    // 2D block and 2D thread
    // Each thread computes one cell in C.
    int i{blockIdx.y * blockDim.y + threadIdx.y};
    int j{blockIdx.x * blockDim.x + threadIdx.x};

    if ((i < M) && (j < N))
    {
        C[i * N + j] = acc_sum;
    }
		*/
	/*
  int i = blockDim.x * blockIdx.x + threadIdx.x;
  int j = blockDim.y * blockIdx.y + threadIdx.y;
  if (i >= M || j >= N)
    return;

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

// Array of device (GPU) pointers
static float *a_d[MAX_NUM_GPU];
static float *b_d[MAX_NUM_GPU];
static float *c_d[MAX_NUM_GPU];
static int M, N, K;
static int Mbegin[MAX_NUM_GPU], Mend[MAX_NUM_GPU];

void mat_mul(float *_A, float *_B, float *_C, int _M, int _N, int _K) {
  // Launch kernel on every GPU
	int M_size[MAX_NUM_GPU];
  for (int i = 0; i < num_devices; i++) {
    //dim3 blockDim(1, 1, 1);
    //dim3 gridDim(Mend[i] - Mbegin[i], N, 1);
		dim3 blockDim(TS/WPT, TS, 1);
		dim3 gridDim(((Mend[i] - Mbegin[i] + TS -1)/TS), (N+TS-1)/TS,1);
		M_size[i] = Mend[i] - Mbegin[i];
    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], M_size[i], N, K);
  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }
}

void mat_mul_init(float *A, float *B, float *C, int _M, int _N, int _K) {
  M = _M, N = _N, K = _K;

  CUDA_CALL( cudaGetDeviceCount(&num_devices) );
  printf("Using %d devices\n", num_devices);
  for (int i = 0; i < num_devices; i++) {
    cudaDeviceProp prop;
    CUDA_CALL( cudaGetDeviceProperties(&prop, i) );

    // Try printing more detailed information here
    printf("[GPU %d] %s\n", i, prop.name);
  }

  if (num_devices <= 0) {
    printf("No CUDA device found. Aborting\n");
    exit(1);
  }

  // Setup problem size for each GPU
  for (int i = 0; i < num_devices; i++) {
    Mbegin[i] = (M / num_devices) * i;
    Mend[i] = (M / num_devices) * (i + 1);
  }
  Mend[num_devices - 1] = M;

  // Allocate device memory for each GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMalloc(&a_d[i], (Mend[i] - Mbegin[i]) * K * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&b_d[i], K * N * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&c_d[i], (Mend[i] - Mbegin[i]) * N * sizeof(float)) );
  }

  // Upload A and B matrix to every GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(a_d[i], A + Mbegin[i] * K,
               (Mend[i] - Mbegin[i]) * K * sizeof(float),
               cudaMemcpyHostToDevice) );
    CUDA_CALL( cudaMemcpy(b_d[i], B, K * N * sizeof(float), cudaMemcpyHostToDevice) );
  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }
}

void mat_mul_final(float *A, float *B, float *C, int M, int N, int K) {
  // Do any post-matmul cleanup work here.

  // Download C matrix from GPUs
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(C + Mbegin[i] * N, c_d[i],
               (Mend[i] - Mbegin[i]) * N * sizeof(float),
               cudaMemcpyDeviceToHost) );
  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }
}