#include "convolution.h"
#include <mpi.h>
#include <stdio.h>
#include "util.h"
#include <cstdio>
#include <cuda_runtime.h>
#define MASTER 0
#define SLAVE 1
#define TS 8
#define MAX_NUM_GPU 4

#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);                                                                 \
    }                                                                          \
  }

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;
static int N_Size[MAX_NUM_GPU];
int num_devices = 1; 

// Array of device (GPU) pointers
static float *in_d[MAX_NUM_GPU];
static float *out_d[MAX_NUM_GPU];
static float *fil_d[MAX_NUM_GPU];
//static int Nbegin[MAX_NUM_GPU], Nend[MAX_NUM_GPU];

__global__ void conv_kernel(
    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) 
{ 
  const int globalRow = blockDim.x * blockIdx.x + threadIdx.x; 
  const int globalCol = blockDim.y * blockIdx.y + threadIdx.y;

  int OH, OW;

  OH = (_H + 2 * _pad - _dilation * (_R - 1) - 1) / _stride + 1;
  OW = (_W + 2 * _pad - _dilation * (_S - 1) - 1) / _stride + 1;
 
  int n,k,w;
  w = globalCol;
  n = w / (_K * OW);
  w = w - n *(_K * OW);
  k = w / OW;
  w = w -k * OW;

  int col = w;
  int row = globalRow;
   
  if (globalRow >= OH || globalCol >= _N * _K * OW) return;
   
  int start_row = row * _stride - _pad;
  int start_col = col * _stride - _pad;
 
  float o = 0.f;
  for (int c = 0; c < _C; ++c) {
    for (int r = 0; r < _R; ++r) {
      for (int s = 0; s < _S; ++s) {
        int h = start_row + r * _dilation;
        int w = start_col + 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];
        o += i * f;
      } 
    } 
  } 
  _output[n * _K * OH * OW + k * OH * OW + row * OW + col] = o; 
} 

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) {
		
  N = _N; C = _C; H = _H; W = _W;
  K = _K; R = _R; S = _S;
  pad = _pad;
  dilation = _dilation;
  stride = _stride;
  input = _input;
  output = _output;
  filter = _filter;
  
  MPI_Status status;
  MPI_Request request;
  int size[2];
  int offset;
  
  if (mpi_world_size == 2 && N > 4) size[1] = N / 2;
  else size[1] = 0;
  size[0] = N - size[1];

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

  if(size[mpi_rank] < MAX_NUM_GPU) {
    num_devices = size[mpi_rank];
    for (int i = 0; i <size[mpi_rank]; i++) {
      N_Size[i]=1;
    } 
  }
  else {
    num_devices = MAX_NUM_GPU;
    int qq = size[mpi_rank] / MAX_NUM_GPU;
    int rr = size[mpi_rank] % MAX_NUM_GPU;
    for (int i = 0; i < MAX_NUM_GPU; i++) {
      N_Size[i] = qq;
      if (i < rr) N_Size[i]++;
    }
  }
		  
  // Allocate device memory for each GPU
  for (int i = 0; i <num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) ); 
    CUDA_CALL( cudaMalloc(&in_d[i], N_Size[i]*C*H*W*sizeof(float)));
    CUDA_CALL( cudaMalloc(&out_d[i], N_Size[i]*K*OH*OW*sizeof(float)));
    CUDA_CALL( cudaMalloc(&fil_d[i], K*C*R*S*sizeof(float)));
  } 
  
  if (mpi_rank == 0 && mpi_world_size == 2 && size[1] !=0) {
    MPI_Isend(&input[size[0]*C*H*W], size[1]*C*H*W, MPI_FLOAT, SLAVE, 0, MPI_COMM_WORLD, &request);
    MPI_Isend(filter, K*C*R*S, MPI_FLOAT, SLAVE, 0, MPI_COMM_WORLD, &request);

    if(size[mpi_rank] < MAX_NUM_GPU) {
      num_devices = size[mpi_rank];
    }
  }  
  else if (mpi_rank == 1 && size[mpi_rank] != 0) {
    alloc_tensor(&input, size[1], C, H, W);
    alloc_tensor(&output, size[1], K, OH, OW);
    alloc_tensor(&filter, K, C, R, S);
    MPI_Recv(input, size[1]*C*H*W, MPI_FLOAT, MASTER, 0, MPI_COMM_WORLD, &status);
    MPI_Recv(filter, K*C*R*S, MPI_FLOAT, MASTER, 0, MPI_COMM_WORLD, &status);

    if(size[mpi_rank] < MAX_NUM_GPU) {
      num_devices = size[mpi_rank];
    }
  }
   
  offset =0;
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(in_d[i], input + offset, N_Size[i]*C*H*W * sizeof(float), cudaMemcpyHostToDevice));
    CUDA_CALL( cudaMemcpy(fil_d[i], filter, K*C*R*S*sizeof(float), cudaMemcpyHostToDevice) ); 
    offset += N_Size[i]*C*H*W;	
  }
  
   
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) ); 
    CUDA_CALL( cudaDeviceSynchronize() );
  }
    
	
  for (int i = 0; i <num_devices; i++) {
    dim3 gridDim((OH+TS-1)/TS, (N_Size[i]*K*OW + TS - 1)/TS, 1);
    dim3 blockDim(TS,TS,1);
    CUDA_CALL( cudaSetDevice(i) );
    conv_kernel<<<gridDim, blockDim>>>(in_d[i], out_d[i], fil_d[i], N_Size[i], C, H, W, K, R, S, pad, dilation, stride);
  }
   
   
  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) ); 
    CUDA_CALL( cudaDeviceSynchronize() );
  }
  
   offset = 0;
   for (int i = 0; i < num_devices; i++) {	
     CUDA_CALL( cudaSetDevice(i) );
     CUDA_CALL( cudaMemcpy(output + offset, out_d[i],N_Size[i]*K*OH*OW*sizeof(float),cudaMemcpyDeviceToHost));   
     offset += N_Size[i]*K*OH*OW;
   }
   
   for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) ); 
    CUDA_CALL( cudaDeviceSynchronize() );
   }

	
  if (mpi_rank == 0 && mpi_world_size == 2 && size[1] !=0) {
     MPI_Recv(&output[size[0]*K*OH*OW], size[1]*K*OH*OW, MPI_FLOAT, SLAVE, 0, MPI_COMM_WORLD, &status);
  } 
  else if(mpi_rank == 1 && size[1] !=0) {
    MPI_Isend(output, size[1]*K*OH*OW, MPI_FLOAT, MASTER, 0, MPI_COMM_WORLD, &request);
  }

}


void convolution_init(
    int _N, int _C, int _H, int _W,
    int _K, int _R, int _S,
    int _pad, int _dilation, int _stride) 
{
  
  MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);
  MPI_Comm_size(MPI_COMM_WORLD, &mpi_world_size);
  
  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);
  }  
}

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

}