#include "mat_mul.h"

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

#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 = 4;
#define TS  32
#define WPT 16
#define RTSF (TS/WPT)

//int MM;
//int KK;
//int NN;

// A: M x K
// B: K x N
// C: M x N
__global__ void sgemm(float *A, float *B, float *C, int M, int N, int K) {
 //int i = blockDim.x * blockIdx.x + threadIdx.x;
 //int j = blockDim.y * blockIdx.y + threadIdx.y;

  int row = threadIdx.x;
  int col = threadIdx.y;
  int globalRow = (blockDim.x * WPT) * blockIdx.x + threadIdx.x;
  int globalCol = blockDim.y * blockIdx.y + threadIdx.y;

  //f (i >= M || j >= N)
  // return;

  __shared__ float Asub[TS][TS];
  __shared__ float Bsub[TS][TS];

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

  int num_tiles = (K+TS-1)/TS;

  for(int t=0;t<num_tiles;t++){
	  for(int w=0;w<WPT;w++){
	    int tiledRow = TS*t+row;
	    int tiledCol = TS*t+col;

	    if(globalRow + w*RTSF >= M || tiledCol >= K) {
		Asub[row+w*RTSF][col]=0.0f;
	    }
	    else {
	    	Asub[row+w*RTSF][col]=A[(globalRow+w*RTSF)*K+tiledCol];
	    }
	    if(tiledRow + w*RTSF >= K || globalCol >= N) {
		Bsub[row+w*RTSF][col]=0.0f;
	    }
	    else {
	    	Bsub[row+w*RTSF][col]=B[(tiledRow+w*RTSF)*N+globalCol];
	    }
	 }
	 __syncthreads();

  	for (int k = 0; k < TS; k++) {
	   for(int w=0;w<WPT;w++) { 
    		acc[w] += Asub[row+w*RTSF][k]*Bsub[k][col];
	   }
	}

	 __syncthreads();
  }

	for(int w=0;w<WPT;w++) {
	  if(globalRow + w*RTSF >= M || globalCol >= N) continue;
	  C[(globalRow+w*RTSF)*N+globalCol]=acc[w];
	}

}

// 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) {

// printf("M: %d, K: %d, N: %d\n",M, K, N);
 //printf("MM: %d, KK: %d, NN: %d\n",MM, KK, NN);
  // Launch kernel on every 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);

    CUDA_CALL( cudaSetDevice(i) );
// if(K%TS==0&&K==N) 
//   sgemmO<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], Mend[i] - Mbegin[i], N, K);
// else
//   sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], Mend[i] - Mbegin[i], N, K);
    sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], Mend[i] - Mbegin[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
// if(M%8)	MM=(((M/8)+1)*8);
// else		MM=M;
// if(K%8)	KK=(((K/8)+1)*8);
// else		KK=K;
// if(N%8)	NN=(((N/8)+1)*8);
// else		NN=N;

  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;
// for (int i = 0; i < num_devices; i++) {
//   Mbegin[i] = (MM / num_devices) * i;
//   Mend[i] = (MM / num_devices) * (i + 1);
// }
// Mend[num_devices - 1] = MM;

  // 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)) );
   //CUDA_CALL( cudaMalloc(&a_d[i], (Mend[i] - Mbegin[i]) * KK * sizeof(float)) );
   //CUDA_CALL( cudaMalloc(&b_d[i], KK * NN * sizeof(float)) );
   //CUDA_CALL( cudaMalloc(&c_d[i], (Mend[i] - Mbegin[i]) * NN * 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() );
  }
}