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

#define MAX_NODE (8)
#define MAX_THREADS (100)

#define MATRIX_SEND_DATA_MSG_ID 1000
#define MATRIX_SEND_RESULT_MSG_ID 1001

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 num_threads;

void run_convolution(
    float *_input, float *_output,
    int endM) {

    for (int n = 0; n < endM; ++n) {
      #pragma omp parallel for num_threads(MAX_THREADS) collapse(3) schedule(dynamic)
      for (int k = 0; k < K; ++k) {
        for (int oh = 0; oh < OH; ++oh) {
          for (int ow = 0; ow < OW; ++ow) {
            register 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 = oh * stride - pad + r * dilation;
                  int w = ow * stride - pad + s * dilation;
                  if (h < 0 || h >= H || w < 0 || w >= W) continue;
                  register float i = _input[n * C * H * W + c * H * W + h * W + w];
                  register float f = filter[k * C * R * S + c * R * S + r * S + s];
                  o += i * f;
                }
              }
            }
            _output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = 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) {

  int input_size, filter_size, output_size;
  int i, slice = 0, m_size, startM, endM;
  MPI_Status status;
  MPI_Request req1[MAX_NODE], req2[MAX_NODE], req3[MAX_NODE], req4[MAX_NODE];

  if (mpi_world_size <= mpi_rank)
    return;

  input_size = C * H * W;
  filter_size = K * C * R * S;
  output_size = K * OH * OW;
  
  if (mpi_rank == 0) {
    input = _input;
    output = _output;
    filter = _filter;

    slice = N / mpi_world_size;

    // Send Matrix Information
    for (i = 1; i < mpi_world_size; i++) {
      startM = i * slice;
      endM = (i == mpi_world_size - 1) ? N : (i + 1) * slice;
      m_size = (endM - startM);
      MPI_Isend(&m_size, 1, MPI_INT, i, MATRIX_SEND_DATA_MSG_ID, MPI_COMM_WORLD, &req1[i]);
      MPI_Isend(&input[startM * input_size], m_size * input_size, MPI_FLOAT, i, MATRIX_SEND_DATA_MSG_ID, MPI_COMM_WORLD, &req2[i]);
      MPI_Isend(&filter[0], filter_size, MPI_FLOAT, i, MATRIX_SEND_DATA_MSG_ID, MPI_COMM_WORLD, &req3[i]);
    }
    startM = 0;
    endM = slice;
  } else {
    startM = 0;
    MPI_Recv(&endM, 1, MPI_INT, 0, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
    m_size = (endM - startM);
    MPI_Irecv(&input[0], m_size * input_size, MPI_FLOAT, 0, MPI_ANY_TAG, MPI_COMM_WORLD, &req2[0]);
    MPI_Irecv(&filter[0], filter_size, MPI_FLOAT, 0, MPI_ANY_TAG, MPI_COMM_WORLD, &req3[0]);
    //zero_tensor(output, m_size, K, OH, OW);
    // wait for receiving
    MPI_Wait(&req2[0], &status);
    MPI_Wait(&req3[0], &status);
  }
  
  run_convolution(input, output, endM);

  if (mpi_rank == 0) {
    // wait for sending
    for (i = 1; i < mpi_world_size; i++) {
      MPI_Wait(&req1[i], &status);
      MPI_Wait(&req2[i], &status);
      MPI_Wait(&req3[i], &status);
    }

    // receiving the result
    for (i = 1; i < mpi_world_size; i++) {
      startM = i * slice;
      endM = (i == mpi_world_size - 1) ? N : (i + 1) * slice;
      m_size = (endM - startM) * output_size;
      MPI_Irecv(&output[startM * output_size], m_size, MPI_FLOAT, i, MPI_ANY_TAG, MPI_COMM_WORLD, &req4[i]);
    }

    // wait for receiving
    for (i = 1; i < mpi_world_size; i++) {
      MPI_Wait(&req4[i], &status);
    }
  } else {
    // sending the result
    m_size = (endM - startM) * output_size;
    MPI_Isend(&output[0], m_size, MPI_FLOAT, 0, MATRIX_SEND_RESULT_MSG_ID, MPI_COMM_WORLD, &req4[0]);
    MPI_Wait(&req4[0], &status);
  }
}

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;
  int i, slice = 0, m_size, startM, endM;

  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;

  num_threads = MAX_THREADS;
  slice = N / mpi_world_size;
  //if (num_threads > K) {
  //  num_threads = K;
  //}

  if (N < 2) {
    mpi_world_size = 1;
  }
  
  for (i = 1; i < mpi_world_size; i++) {
    startM = i * slice;
    endM = (i == mpi_world_size - 1) ? N : (i + 1) * slice;
	m_size = (endM - startM);
    alloc_tensor(&input, m_size, C, H, W);
	alloc_tensor(&output, m_size, K, OH, OW);
	alloc_tensor(&filter, K, C, R, S);
  }
}

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