#include "convolution.h"

#include <mpi.h>
#include <stdio.h>
#include <cuda_runtime.h>
#include "util.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 8

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 float *in_d[MAX_NUM_GPU], *out_d[MAX_NUM_GPU], *filter_d[MAX_NUM_GPU];
static int num_devices = 1;
static int Size[2];
static int NN[MAX_NUM_GPU];

__global__ void Convolution_cuda(
    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 global_row = blockDim.x * blockIdx.x + threadIdx.x ;
  const int global_col = 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 = global_col;//
  n = w / (_K * OW);//
  w = w - n *(_K * OW);//
  k = w / OW;//
  w = w - k * OW;//

  int col = w;
  int row = global_row;
  if (global_row >= OH || global_col >= _N*_K*OW) return;

  int row_st = row * _stride - _pad;
  int col_st = col * _stride - _pad;
  float temp_out = 0.0f;
  for (int c = 0 ; c < _C ; c++) {
    for (int i = 0 ; i < _R ; i++) {
      for (int j = 0 ; j < _S ; j++) {
	int h = row_st + i * _dilation;
	int w = col_st + j * _dilation;
	if (h < 0 || w < 0 || h >= _H || w >= _W) continue;
        float temp_input  = _input[n*_C*_W*_H + c*_W*_H + h*_W + w];
	float temp_filter = _filter[k*_C*_R*_S + c*_R*_S + i*_S + j];
	temp_out += temp_input * temp_filter;
      }
    }
  } 
  _output[n*_K*OH*OW + k*OH*OW + row*OW + col] = temp_out;
}

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) {

  int offset = 0;

  input = _input;
  output = _output;
  filter = _filter;

  MPI_Request request;
  MPI_Status status;

  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, 1, 0, MPI_COMM_WORLD, &request);
    MPI_Isend(filter, _K*_C*_R*_S, MPI_FLOAT, 1, 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, 0, 0, MPI_COMM_WORLD, &status);
    MPI_Recv(filter, _K*_C*_R*_S, MPI_FLOAT, 0, 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, NN[i]*C*H*W*sizeof(float), cudaMemcpyHostToDevice) );
    CUDA_CALL( cudaMemcpy(filter_d[i], filter, K*C*R*S*sizeof(float),cudaMemcpyHostToDevice) );
    offset += NN[i] * C * H * W;
  }

  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }

  for (int i = 0; i < num_devices; i++) {
    dim3 gridDim((OH+TS-1)/TS, (NN[i]*K*OW + TS - 1)/TS, 1);
    dim3 blockDim(TS, TS, 1);

    CUDA_CALL( cudaSetDevice(i) );
    Convolution_cuda<<<gridDim, blockDim>>>(in_d[i], out_d[i], filter_d[i], NN[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],
               NN[i]*K*OH*OW * sizeof(float),
               cudaMemcpyDeviceToHost) );
    offset += NN[i]*K*OH*OW;
  }

  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
    CUDA_CALL( cudaSetDevice(i) );
  }

  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, 1, 0, MPI_COMM_WORLD, &status);
  } else if(mpi_rank == 1 && Size[1] != 0){
    MPI_Isend(output, Size[1]*K*OH*OW, MPI_FLOAT, 0, 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) {
  N = _N; C = _C; H = _H; W = _W; K = _K; R = _R; S = _S;
  pad = _pad;
  dilation = _dilation;
  stride = _stride;

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

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

  if (mpi_world_size == 2) {
	  Size[1] = _N / 2;
	 }
  Size[0] = N - Size[1];

  /* MP size of MPI  MAX GPU */
  if (Size[mpi_rank] < MAX_NUM_GPU) {
    num_devices = Size[mpi_rank];
    for (int i = 0 ; i < Size[mpi_rank] ; i++) {
      NN[i] = 1;
    }
  } else {
    num_devices = MAX_NUM_GPU;
    int remain = Size[mpi_rank] % MAX_NUM_GPU;
    int quot = Size[mpi_rank] / MAX_NUM_GPU;
    for (int i = 0 ; i < MAX_NUM_GPU ; i++) {
      NN[i] = quot;
      if (i < remain) { 
	      NN[i]++;
      }
    }
  }

  for (int i = 0 ; i < num_devices ; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMalloc(&in_d[i], NN[i]*C*H*W*sizeof(float)) );
    CUDA_CALL( cudaMalloc(&out_d[i], NN[i]*K*OH*OW*sizeof(float)) );
    CUDA_CALL( cudaMalloc(&filter_d[i], K*C*R*S*sizeof(float)) );
  }
}

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