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

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;

#define min(a,b) (a>b?b:a)
#define HSLICE OH
#define WSLICE 128
int num_threads = 100;

void convolution_omp_slc()
{
  if((dilation == 1) && (stride == 1) && (pad == 0) && (W%16 == 0) && (S%16  == 0))
  {
    //printf("AVX\n");
    #pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
    for (int n = 0; n < N; ++n) {
      for (int k = 0; k < K; ++k) {
        //#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
        for (int oh = 0; oh < OH; ++oh) {
        //#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
          for (int ow = 0; ow < OW; ++ow) {
            //float o = 0.f;
	    __m512 vo = {0.0f};
            for (int c = 0; c < C; ++c) {
              for (int r = 0; r < R; ++r) {
                //for (int s = 0; s < S; ++s) {
                for (int s = 0; s < S; s+=16) {
                  int h = oh * stride - pad + r;
                  int w = ow * stride - pad + s;
                  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];
		  __m512 vi = _mm512_loadu_ps(&input[n * C * H * W + c * H * W + h * W + w]);
		  __m512 vf = _mm512_loadu_ps(&filter[k * C * R * S + c * R * S + r * S + s]);

                  //o += i * f;
		  vo = _mm512_fmadd_ps(vi, vf, vo);
                }
              }
            }
            //_mm512_storeu_ps(&output[n * K * OH * OW + k * OH * OW + oh * OW + ow], vo);
	    float o = _mm512_reduce_add_ps(vo);
            output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;
          }
        }
      }
    }
  }
  else
  {
    #pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
    for (int n = 0; n < N; ++n) {
      for (int k = 0; k < K; ++k) {
        for (int ohs = 0; ohs<OH; ohs += HSLICE) {
        for (int ows = 0; ows<OW; ows += WSLICE) {
        //#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
        for (int oh = ohs; oh < min(ohs + HSLICE, OH); ++oh) {
        //#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
          for (int ow = ows; ow < min(ows + WSLICE, OW); ++ow) {
            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;
                  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 + oh * OW + ow] = o;
          }
        }
        //Slice
        }}
      }
    }
  }

}

void convolution_omp()
{
    #pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
    for (int n = 0; n < N; ++n) {
      for (int k = 0; k < K; ++k) {
        //#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
        for (int oh = 0; oh < OH; ++oh) {
        //#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
          for (int ow = 0; ow < OW; ++ow) {
            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;
                  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 + 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) {
  input = _input;
  output = _output;
  filter = _filter;

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

  int default_div_size = N/mpi_world_size;
  MPI_Status status;
  //MPI_Request request;

  if (mpi_rank == 0) {
	  // 1. Distribute batch to the other nodes
	  //timer_start(1);

          MPI_Request arrA_req[4];
          MPI_Status arrA_status[4];
	  for(int target_rank = 1; target_rank < mpi_world_size; target_rank++){
                  int div_start, div_size;
                  div_start = target_rank * default_div_size;
                  div_size = default_div_size;
                  if(target_rank == (mpi_world_size - 1))
                          div_size += N - (default_div_size * mpi_world_size);

                  //printf("send. target_rank=%d, div_start=%d, div_size=%d, tot_size=%d\n", target_rank, div_start, div_size, div_size * C*H*W);
                  MPI_Isend(input + (div_start * C*H*W), div_size * C*H*W, MPI_FLOAT, target_rank, 0, MPI_COMM_WORLD, &arrA_req[target_rank-1]);

                  //MPI_Isend(B, K * N, MPI_FLOAT, target_rank, 0, MPI_COMM_WORLD, &request);
          }
	  // 2. Broadcase all Filters
          MPI_Bcast(filter, K*C*R*S, MPI_FLOAT, 0, MPI_COMM_WORLD);

	  //double elapsed_time = timer_stop(1);
          //printf("[rank %d] scatter time: %f sec\n", mpi_rank, elapsed_time);

	  int original_N = N;
	  N = default_div_size;

	  // 3. Do Convolution
	  //timer_start(1);
	  convolution_omp_slc();
	  //elapsed_time = timer_stop(1);
          //printf("[rank %d] time: %f sec\n", mpi_rank, elapsed_time);
	  N = original_N;

	  //timer_start(1);
	  // 4. Receive result from the other node

          MPI_Request arrC_req[4];
          MPI_Status arrC_status[4];
          for(int target_rank = 1; target_rank < mpi_world_size; target_rank++){
                  int div_start, div_size;
                  div_start = target_rank * default_div_size;
                  div_size = default_div_size;
                  if(target_rank == (mpi_world_size - 1))
                          div_size += N - (default_div_size * mpi_world_size);

		  //printf("wait div_size=%d\n", div_size);
                  MPI_Irecv(output + (div_start * K*OH*OW), div_size * K*OH*OW, MPI_FLOAT, target_rank, 0, MPI_COMM_WORLD, &arrC_req[target_rank-1]);
                  //MPI_Recv(output + (div_start * K*OH*OW), div_size * K*OH*OW, MPI_FLOAT, target_rank, 0, MPI_COMM_WORLD, &arrC_status[target_rank-1]);
          }
          //MPI_Waitall(mpi_world_size-1, arrA_req, arrA_status);
          MPI_Waitall(mpi_world_size-1, arrC_req, arrC_status);

	  //elapsed_time = timer_stop(1);
          //printf("[rank %d] collect time: %f sec\n", mpi_rank, elapsed_time);
  }else{
	  //0. alloc local memory
          int div_size;

          div_size = default_div_size;
          if(mpi_rank == (mpi_world_size - 1))
                  div_size += N - (default_div_size * mpi_world_size);

	  int original_N = N;
          N = div_size; // Adjust N size
	  //printf("defulat div size=%d\n", default_div_size);

          alloc_tensor(&input, N, C, H, W);
          alloc_tensor(&filter, K, C, R, S);
	  alloc_tensor(&output, N, K, OH, OW);

	  // 1. Recv part of A
          //printf("sub. rank=%d, div_size=%d, Recv start, tot_size=%d\n", mpi_rank, div_size, N*C*H*W);
          MPI_Recv(input, N*C*H*W, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
          //printf("sub. rank=%d, div_size=%d, Recv end\n", mpi_rank, div_size);

          // 2. Recv full Filter
          MPI_Bcast(filter, K*C*R*S, MPI_FLOAT, 0, MPI_COMM_WORLD);

	  // 3. Do Convolution
	  //timer_start(1);
	  convolution_omp();
	  //double elapsed_time = timer_stop(1);
          //printf("[rank %d] time: %f sec\n", mpi_rank, elapsed_time);

	  // 4. Send C to rank 0 node.
          //printf("sub. end. my rank=%d, div_size=%d\n", mpi_rank, div_size);
          MPI_Send(output, N*K*OH*OW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD);

	  N = original_N;

	  //free
	  free(input);
	  free(filter);
	  free(output);
  }
}

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_rank(MPI_COMM_WORLD, &mpi_rank);
  MPI_Comm_size(MPI_COMM_WORLD, &mpi_world_size);
}

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