#include "convolution.h"
#include <mpi.h>
#include <stdio.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
#define TS 1
int num_devices[2] = {0,0};


__global__ void conv(float *input, float *filter, float *output, int N, int C, int H, int W, int K, int R, int S, int pad, int dilation, int stride) {

  int OH = (H + 2 * pad - dilation * (R-1) - 1) / stride + 1;
  int OW = (W + 2 * pad - dilation * (S-1) - 1) / stride + 1;

  int n = blockIdx.x;
  int k = blockIdx.y;
  int oh = blockIdx.z;
  int ow = threadIdx.x;

  if(n >= N || k >= K || oh >= OH || ow >= OW ) return;

//  __syncthreads();
  float tmp = 0.0f;
  for(int c = 0; c < C; c++){
    for(int r = 0; r < R; r++){
      for(int s = 0; s < S; s++){
        int h = oh * stride - pad + r * dilation;
        int w = ow * stride - pad + s * dilation;

        if( h < 0 || h >= H || w < 0 || w >= W ) continue;
        float i = input[n* C * H * W + c * H * W + h * W + w];
        float f = filter[k* C * R * S + c * R * S + r * S + s];
        tmp += i * f;
        }
    }
  }
//  __syncthreads();
output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = tmp;

}

// Array of device (GPU) pointers
static float *a1_d[MAX_NUM_GPU];
static float *a2_d[MAX_NUM_GPU];
static float *b1_d[MAX_NUM_GPU];
static float *b2_d[MAX_NUM_GPU];
static float *c1_d[MAX_NUM_GPU];
static float *c2_d[MAX_NUM_GPU];
static int Mbegin1[MAX_NUM_GPU], Mend1[MAX_NUM_GPU];
static int Mbegin2[MAX_NUM_GPU], Mend2[MAX_NUM_GPU];

static float *input, *output, *filter;
static int N, C, H, W;
static int K, R, S;
static int OH, OW;
static int pad;
static int dilation;
static int stride;
static int mpi_rank, mpi_world_size;
int portion, low_bound, upper_bound;
MPI_Status status;
MPI_Request request;

void convolution(
    float *_input, float *_output, float *_filter, 
    int _N, int _C, int _H, int _W,
    int _K, int _R, int _S,
    int _pad, int _dilation, int _stride) {
  input = _input;
  output = _output;
  filter = _filter;
  N = _N, C = _C, H = _H, W = _W;
  K = _K, R = _R, S = _S;
  pad = _pad, dilation = _dilation, stride = _stride;
  
  MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);
  MPI_Comm_size(MPI_COMM_WORLD, &mpi_world_size);
  OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
  size_t bytes_nchw = N*C*H*W*sizeof(float);
  size_t bytes_kcrs = K*C*R*S*sizeof(float);

  if (mpi_rank == 0) {
    portion = (N / mpi_world_size);
    for(int i=1; i < mpi_world_size; i++){
      low_bound = i*portion;
      if( ((i+1)==mpi_world_size) && (( N % mpi_world_size) != 0 ) ){
        upper_bound = N;
      }
      else{
        upper_bound = low_bound + portion;
      }

      MPI_Isend(&low_bound, 1, MPI_INT, i, 1, MPI_COMM_WORLD, &request);
      MPI_Isend(&upper_bound, 1, MPI_INT, i, 2, MPI_COMM_WORLD, &request);

      MPI_Isend(&input[low_bound*C*H*W], ( (upper_bound - low_bound)*C*H*W ), MPI_FLOAT, i, 3, MPI_COMM_WORLD, &request);
      MPI_Isend(&filter[0], (K*C*R*S), MPI_FLOAT, i, 4, MPI_COMM_WORLD, &request);

    }
    low_bound = 0;
    upper_bound = portion;

      int n = ((upper_bound-low_bound)/TS)*TS;
      
        // Setup problem size for each GPU
        for (int i = 0; i < num_devices[0]; i++) {
          Mbegin1[i] = (n / num_devices[0]) * i + low_bound;
          Mend1[i] = (n / num_devices[0]) * (i + 1) + low_bound;
        }
        Mend1[num_devices[0] - 1] = n + low_bound;

      // Upload A and B matrix to every GPU
        for (int i = 0; i < num_devices[0]; i++) {
          CUDA_CALL( cudaMemcpy(a1_d[i], input +  Mbegin1[i] * C * H * W ,
                     (Mend1[i] - Mbegin1[i]) * C * H * W * sizeof(float),
                     cudaMemcpyHostToDevice) );
          CUDA_CALL( cudaMemcpy(b1_d[i], filter, K * C * R * S * sizeof(float), cudaMemcpyHostToDevice) );
        }
      
        // DO NOT REMOVE; NEEDED FOR TIME MEASURE
        for (int i = 0; i < num_devices[0]; i++) {
          CUDA_CALL( cudaDeviceSynchronize() );
        }

    // Launch kernel on every GPU
  for (int i = 0; i < num_devices[0]; i++) {
    dim3 gridDim( Mend1[i] - Mbegin1[i] , K , OH);
    dim3 blockDim(OW, 1);

    CUDA_CALL( cudaSetDevice(i) );
    conv<<<gridDim, blockDim>>>(a1_d[i], b1_d[i], c1_d[i], Mend1[i]-Mbegin1[i], C, H, W , K , R ,S ,pad,dilation, stride );
  }

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

  } // mpi_rank == 0
  else{
      
      MPI_Recv(&low_bound, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &status);
      MPI_Recv(&upper_bound, 1, MPI_INT, 0, 2, MPI_COMM_WORLD, &status);
      input = (float*)malloc(bytes_nchw);
      filter = (float*)malloc(bytes_kcrs);
      MPI_Recv(&input[low_bound*C*H*W], ( (upper_bound - low_bound)*C*H*W ), MPI_FLOAT, 0, 3, MPI_COMM_WORLD, &status);
      MPI_Recv(&filter[0], (K*C*R*S ), MPI_FLOAT, 0, 4, MPI_COMM_WORLD, &status);

      int n2 = ((upper_bound-low_bound)/TS)*TS;
      
        // Setup problem size for each GPU
        for (int i = 0; i < num_devices[1]; i++) {
          Mbegin2[i] = (n2 / num_devices[1]) * i;
          Mend2[i] = (n2 / num_devices[1]) * (i + 1);
        }
        Mend2[num_devices[1] - 1] = n2;
   //   
     // Upload A and B matrix to every GPU
       for (int i = 0; i < num_devices[1]; i++) {
         CUDA_CALL( cudaMemcpy(a2_d[i], input +  (Mbegin2[i] + low_bound )* C * H * W ,
                    (Mend2[i] - Mbegin2[i]) * C * H * W * sizeof(float),
                    cudaMemcpyHostToDevice) );
         CUDA_CALL( cudaMemcpy(b2_d[i], filter, K * C * R * S * sizeof(float), cudaMemcpyHostToDevice) );
       }
      
        // DO NOT REMOVE; NEEDED FOR TIME MEASURE
        for (int i = 0; i < num_devices[1]; i++) {
          CUDA_CALL( cudaDeviceSynchronize() );
        }

    // Launch kernel on every GPU
  for (int i = 0; i < num_devices[1]; i++) {
    dim3 gridDim( Mend2[i] - Mbegin2[i] , K , OH);
    dim3 blockDim(OW, 1);
    CUDA_CALL( cudaSetDevice(i) );
    conv<<<gridDim, blockDim>>>(a2_d[i], b2_d[i], c2_d[i], Mend2[i]-Mbegin2[i], C, H, W , K , R ,S,pad,dilation, stride );
  }

// DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices[1]; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaDeviceSynchronize() );
  }
  free(input);
  free(filter);

  }
}
//
void convolution_init(
    int _N, int _C, int _H, int _W,
    int _K, int _R, int _S,
    int _pad, int _dilation, int _stride) {

  N = _N; C = _C; H = _H; W = _W;
  K = _K; R = _R; S = _S;
  pad = _pad;
  dilation = _dilation;
  stride = _stride;
  OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
  size_t bytes_nchw = N*C*H*W*sizeof(float);
  size_t bytes_kcrs = K*C*R*S*sizeof(float);

  MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);
  MPI_Comm_size(MPI_COMM_WORLD, &mpi_world_size);

  if (mpi_rank == 0) {
    portion = (N / mpi_world_size);
    for(int i=1; i < mpi_world_size; i++){
      low_bound = i*portion;
      if( ((i+1)==mpi_world_size) && (( N % mpi_world_size) != 0 ) ){
        upper_bound = N;
      }
      else{
        upper_bound = low_bound + portion;
      }

      MPI_Isend(&low_bound, 1, MPI_INT, i, 1, MPI_COMM_WORLD, &request);
      MPI_Isend(&upper_bound, 1, MPI_INT, i, 2, MPI_COMM_WORLD, &request);
    }
    low_bound = 0;
    upper_bound = portion;

      int n = ((upper_bound-low_bound)/TS)*TS;
        CUDA_CALL( cudaGetDeviceCount(&num_devices[0]) );
      
        printf("1:Using %d devices\n", num_devices[0]);
        for (int i = 0; i < num_devices[0]; i++) {
          cudaDeviceProp prop;
          CUDA_CALL( cudaGetDeviceProperties(&prop, i) );
      
          // Try printing more detailed information here
          printf("[1:GPU %d] %s\n", i, prop.name);
        }
      
        if (num_devices[0] <= 0) {
          printf("No CUDA device found. Aborting\n");
          exit(1);
        }
      
        // Setup problem size for each GPU
        for (int i = 0; i < num_devices[0]; i++) {
          Mbegin1[i] = (n / num_devices[0]) * i + low_bound;
          Mend1[i] = (n / num_devices[0]) * (i + 1) + low_bound;
        }
        Mend1[num_devices[0] - 1] = n + low_bound;
      
        // Allocate device memory for each GPU
        for (int i = 0; i < num_devices[0]; i++) {
          CUDA_CALL( cudaSetDevice(i) );
          CUDA_CALL( cudaMalloc(&a1_d[i], (Mend1[i] - Mbegin1[i]) * C * H * W * sizeof(float)) );
          CUDA_CALL( cudaMalloc(&b1_d[i], K * C * R * S * sizeof(float)) );
          CUDA_CALL( cudaMalloc(&c1_d[i], (Mend1[i] - Mbegin1[i]) * K * OH * OW * sizeof(float)) );
        }
//      
  }
  else{

      MPI_Recv(&low_bound, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &status);
      MPI_Recv(&upper_bound, 1, MPI_INT, 0, 2, MPI_COMM_WORLD, &status);
//
      int n2 = ((upper_bound-low_bound)/TS)*TS;
        CUDA_CALL( cudaGetDeviceCount(&num_devices[1]) );
      
        printf("2:Using %d devices\n", num_devices[1]);
        for (int i = 0; i < num_devices[1]; i++) {
          cudaDeviceProp prop;
          CUDA_CALL( cudaGetDeviceProperties(&prop, i) );
      
          printf("[2:GPU %d] %s\n", i, prop.name);
        }
      
        if (num_devices[1] <= 0) {
          printf("No CUDA device found. Aborting\n");
          exit(1);
        }
      
        // Setup problem size for each GPU
        for (int i = 0; i < num_devices[1]; i++) {
          Mbegin2[i] = (n2 / num_devices[1]) * i;
          Mend2[i] = (n2 / num_devices[1]) * (i + 1);
        }
        Mend2[num_devices[1] - 1] = n2;
      
        // Allocate device memory for each GPU
        for (int i = 0; i < num_devices[1]; i++) {
          CUDA_CALL( cudaSetDevice(i) );
          CUDA_CALL( cudaMalloc(&a2_d[i], (Mend2[i] - Mbegin2[i]) * C * H * W * sizeof(float)) );
          CUDA_CALL( cudaMalloc(&b2_d[i], K * C * R * S * sizeof(float)) );
          CUDA_CALL( cudaMalloc(&c2_d[i], (Mend2[i] - Mbegin2[i]) * K * OH * OW * sizeof(float)) );
        }
//      
  } // else
}
//
void convolution_final(
    int _N, int _C, int _H, int _W,
    int _K, int _R, int _S,
    int _pad, int _dilation, int _stride) {
  N = _N; C = _C; H = _H; W = _W;
  K = _K; R = _R; S = _S;

  MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);
  MPI_Comm_size(MPI_COMM_WORLD, &mpi_world_size);
  OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
  size_t bytes_nkohow = N*K*OH*OW*sizeof(float);

  if (mpi_rank == 0) {

    for(int i = 1; i < mpi_world_size; i++){

      MPI_Recv(&low_bound, 1, MPI_INT, i, 4, MPI_COMM_WORLD, &status);
      MPI_Recv(&upper_bound, 1, MPI_INT, i, 5, MPI_COMM_WORLD, &status);
      MPI_Recv(&output[low_bound*K*OH*OW], ( (upper_bound - low_bound)*K*OH*OW ), MPI_FLOAT, i, 6, MPI_COMM_WORLD, &status);

    }

     for (int i = 0; i < num_devices[0]; i++) {
        CUDA_CALL( cudaMemcpy( output +  Mbegin1[i] * K * OH * OW , c1_d[i],
                   (Mend1[i] - Mbegin1[i]) * K * OH * OW * sizeof(float),
                   cudaMemcpyDeviceToHost) );
      }
      // DO NOT REMOVE; NEEDED FOR TIME MEASURE
      for (int i = 0; i < num_devices[0]; i++) {
        CUDA_CALL( cudaSetDevice(i) );
        CUDA_CALL( cudaDeviceSynchronize() );
      }

   }
  else{

    output = (float*)malloc(bytes_nkohow);
  for (int i = 0; i < num_devices[1]; i++) {
    CUDA_CALL( cudaMemcpy( output +  (Mbegin2[i] + low_bound) * K * OH * OW , c2_d[i],
               (Mend2[i] - Mbegin2[i]) * K * OH * OW * sizeof(float),
               cudaMemcpyDeviceToHost) );
  }
  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices[1]; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaDeviceSynchronize() );
  }

      MPI_Isend(&low_bound, 1, MPI_INT, 0, 4, MPI_COMM_WORLD, &request);
      MPI_Isend(&upper_bound, 1, MPI_INT, 0, 5, MPI_COMM_WORLD, &request);
      MPI_Isend(&output[low_bound*K*OH*OW], ( (upper_bound - low_bound)*K*OH*OW ), MPI_FLOAT, 0, 6, MPI_COMM_WORLD, &request);
    free(output);
  }
  
}