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

// tunable parameter
#define MAX_GPU_NUMBER 4
#define TS 8

#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 float *input_d[MAX_GPU_NUMBER], *output_d[MAX_GPU_NUMBER], *filter_d[MAX_GPU_NUMBER];
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 num_devices = 1;
static int MP_size[2];
static int N_size[MAX_GPU_NUMBER];

__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 globalRow = blockDim.x * blockIdx.x + threadIdx.x;
  const int globalCol = blockDim.y * blockIdx.y + threadIdx.y;
  
  //indexing
  int n, k, w;
  int col = globalCol;
  int row = globalRow;
  int OH = (_H + 2 * _pad - _dilation * (_R - 1) - 1) / _stride + 1;
  int OW = (_W + 2 * _pad - _dilation * (_S - 1) - 1) / _stride + 1;
  
  w = globalCol;
  n = w / (_K * OW);
  w = w - n *(_K * OW);
  k = w / OW;
  w = w - k * OW;
  col = w;

  if (globalRow >= OH || globalCol >= _N *_K * OW) return;

  int start_row = row * _stride - _pad;
  int start_col = col * _stride - _pad;

  float out_buf = 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 = start_row + i * _dilation;
        int w = start_col + j * _dilation;
        
        if (h < 0 || w < 0 || h >= _H || w >= _W) continue;
        float in_buf = _input[n *_C * _W * _H + c * _W * _H + h * _W + w];
	float filt_buf = _filter[k *_C * _R * _S + c * _R * _S + i * _S + j];
        out_buf += in_buf * filt_buf;
      }
    }
  } 
  _output[n * _K * OH * OW + k * OH * OW + row * OW + col] = out_buf;
}

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_Request request;
  MPI_Status status;
  
  if (mpi_rank == 0 && MP_size[1] != 0 && mpi_world_size == 2) {
    MPI_Isend(&input[MP_size[0] * C * H * W], MP_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 (MP_size[mpi_rank] < MAX_GPU_NUMBER) {
      num_devices = MP_size[mpi_rank];
    }
  } 
  else if (mpi_rank == 1 && MP_size[mpi_rank] != 0) {
    alloc_tensor(&input, MP_size[1], C, H, W);
    alloc_tensor(&output, MP_size[1], K, OH, OW);
    alloc_tensor(&filter, K, C, R, S);
    
    MPI_Recv(input, MP_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 (MP_size[mpi_rank] < MAX_GPU_NUMBER) {
      num_devices = MP_size[mpi_rank];
    }
  }

  int offset = 0;
  
  for (int i = 0 ; i < num_devices ; i++) {
    CUDA_CALL(cudaMemcpy(input_d[i], input + offset, N_size[i] * C * H * W * sizeof(float), cudaMemcpyHostToDevice));
    CUDA_CALL(cudaMemcpy(filter_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(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));
    convolution_cuda<<<gridDim, blockDim>>>(input_d[i], output_d[i], filter_d[i], N_size[i], C, H, W, K, R, S, pad, dilation, stride);
  }

  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, output_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(cudaDeviceSynchronize());
    CUDA_CALL(cudaSetDevice(i));
  }

  if (mpi_rank == 0 && MP_size[1] != 0 && mpi_world_size == 2) {
    MPI_Recv(&output[MP_size[0] * K * OH * OW], MP_size[1] * K * OH * OW, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &status);
  } else if(mpi_rank == 1 && MP_size[1] != 0){
    MPI_Isend(output, MP_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) MP_size[1] = _N / 2;
  MP_size[0] = N - MP_size[1];

  if (MP_size[mpi_rank] < MAX_GPU_NUMBER) {
    num_devices = MP_size[mpi_rank];
    for (int i = 0 ; i < MP_size[mpi_rank] ; i++) {
      N_size[i] = 1;
    }
  } 
  else {
    num_devices = MAX_GPU_NUMBER;
    int remainder = MP_size[mpi_rank] % MAX_GPU_NUMBER;
    int quotient = MP_size[mpi_rank] / MAX_GPU_NUMBER;

    for (int i = 0 ; i < MAX_GPU_NUMBER ; i++) {
      N_size[i] = quotient;
      if (i < remainder) N_size[i]++;
    }
  }

  for (int i = 0 ; i < num_devices ; i++) {
    CUDA_CALL(cudaSetDevice(i));
    CUDA_CALL(cudaMalloc(&input_d[i], N_size[i] * C * H * W * sizeof(float)));
    CUDA_CALL(cudaMalloc(&output_d[i], N_size[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) {
}