#include "mat_mul.h"

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

#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;

__global__ void sgemm(float *A, float *B, float *C, int M, int N, int K) {

    // Thread identifiers
    const int row       = threadIdx.x; // Local row ID (max: TS)
    const int col       = threadIdx.y; // Local col ID (max: TS/WPT == RTS)
    const int globalRow = (blockDim.x * WPT) * blockIdx.x + row; // Row ID of C (0..M)
    const int globalCol =  blockDim.y *        blockIdx.y + col; // Col ID of C (0..N)

    // Local memory to fit a tile of TS*TS elements of A and B
    __shared__ float Asub[TS][TS];
    __shared__ float Bsub[TS][TS];

    // Initialise the accumulation registers
    float tempVal[WPT];
    for (int w=0; w<WPT; w++)
        tempVal[w] = 0.0f;

    // Loop over all tiles
    const int numTiles = (K + TS - 1) / TS  ;   //  Ceiling to Tile Size
  //const int numTiles = (K / TS) + 1 - (K % TS == 0);
    for (int t=0; t<numTiles; t++) {

        // Load one tile of A and B into local memory
        for (int w=0; w<WPT; w++) {
            const int tiledRow  =   TS*t + row;
            const int tiledCol  =   TS*t + col;

            const int orgM      =   globalRow + w*RTS ;
            const int orgK      =   tiledRow  + w*RTS ;

            //  Copy Global A to local
            if ( (orgM < M) && (tiledCol < K) )   
                Asub[row + w*RTS][col] = A[orgM*K + tiledCol]   ;
            else
                Asub[row + w*RTS][col] = 0.0f                   ;
            
            //  Copy Global B to local
            if ( (orgK < K) && (globalCol < N) ) 
                Bsub[row + w*RTS][col] = B[orgK*N + globalCol]  ;
            else
                Bsub[row + w*RTS][col] = 0.0f                   ;
        }

        // Synchronise to make sure the tile is loaded
      //barrier(CLK_LOCAL_MEM_FENCE);
        __syncthreads() ;

        // Perform the computation for a single tile
        for (int k=0; k<TS; k++) {
            for (int w=0; w<WPT; w++) {
                tempVal[w] += Asub[row + w*RTS][k] * Bsub[k][col];
            }
        }

        // Synchronise before loading the next tile
      //barrier(CLK_LOCAL_MEM_FENCE);
        __syncthreads() ;
    }

    // Store the final results in C
    for (int w=0; w<WPT; w++) {
        const int orgM  =   globalRow + w*RTS   ;

        if ( (orgM < M) && (globalCol < N) )
            C[orgM * N + globalCol] = tempVal[w];
    }

//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 M_str[MAX_NUM_GPU], M_siz[MAX_NUM_GPU];
//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
  for (int i = 0; i < num_devices; i++) {
    dim3 blockDim(TS/WPT, TS, 1);
    dim3 gridDim( (M_siz[i] + TS - 1) / TS, (N + TS  - 1) / TS, 1);
  //dim3 gridDim( (Mend[i] - Mbegin[i] +TS - 1) / TS, (N + TS  - 1) / TS, 1);

    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], M_siz[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

  M_str [0] =   0               ; 
  M_siz [0] =   M / num_devices ;
  int mod = M % num_devices     ;
  if (mod != 0)
      M_siz[0] = M_siz[0] + mod ;

  //    Assign Remained Device
  for (int i = 1 ; i < num_devices ; i++){
    M_str[i] = M_str[i - 1] + M_siz[i - 1];

    if (i == num_devices - 1) {    //  Last Device
        M_siz[i] = M - M_str[i] ;
    } else {
        M_siz[i] = M / num_devices                   ;
    }
  }

 //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], M_siz[i] * K * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&b_d[i], K        * N * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&c_d[i], M_siz[i] * N * sizeof(float)) );
  //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[M_str[i] * K], M_siz[i] * K * sizeof(float), cudaMemcpyHostToDevice) );
    CUDA_CALL( cudaMemcpy(b_d[i],  B              , K        * N * sizeof(float), cudaMemcpyHostToDevice) );
  //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[M_str[i] * N], c_d[i], M_siz[i] * N * sizeof(float), cudaMemcpyDeviceToHost) );
  //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() );
  }
}