#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

static float *input, *output, *filter;
static float *in_d[MAX_NUM_GPU], *out_d[MAX_NUM_GPU], *fil_d[MAX_NUM_GPU];
static int N, C, H, W;
static int K, R, S;
static int pad;
static int dilation;
static int stride;
static int mpi_rank, mpi_world_size;
static int num_devices = 1;
static int size[2];
static int Nsize[MAX_NUM_GPU];
static int OH, OW;

static MPI_Request request;
static MPI_Status status;

__global__ void conv(
    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 OH = (_H + 2 * _pad - _dilation * (_R - 1) - 1) / _stride + 1;
  int OW = (_W + 2 * _pad - _dilation * (_S - 1) - 1) / _stride + 1;

  int n = blockIdx.x;
  int k = blockIdx.y;
  int oh = blockIdx.z;
  int ow = threadIdx.x;

  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) {

    int offset = 0;

		if (mpi_rank == 0) {
			input  = _input;
			output = _output;
			filter = _filter;
		}

    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) {
      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, Nsize[i] * C * H * W * sizeof(float),
					       cudaMemcpyHostToDevice) );
      CUDA_CALL( cudaMemcpy(fil_d[i], filter, K * C * R * S * sizeof(float),
								 cudaMemcpyHostToDevice) );
      offset += Nsize[i] * C * H * W;
    }

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

    for (int i = 0; i < num_devices; i++) {
	    dim3 blockDim(OW, 1, 1);
  	  dim3 gridDim(Nsize[i], K, OH);

      CUDA_CALL( cudaSetDevice(i) );
      conv<<<gridDim, blockDim>>>(in_d[i], out_d[i], fil_d[i], Nsize[i], 
					                        _C, _H, _W, _K, _R, _S, _pad, _dilation, _stride);
    }

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

		// Download output from GPUs
    offset = 0;
    for (int i=0; i<num_devices ; ++i) {
      CUDA_CALL( cudaMemcpy(output + offset, out_d[i], Nsize[i] * K * OH * OW * sizeof(float), 
						     cudaMemcpyDeviceToHost) );
      offset += Nsize[i] * K * OH * OW;
    }

		// DO NOT REMOVE; NEEDED FOR TIME MEASURE
    for (int i = 0; i < num_devices; 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, 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_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;

#if 0 
	size[0] = (mpi_world_size == 2 && N > MAX_NUM_GPU)? N / 2 + 1 : N;
	size[1] = N - size[0];
#else
	size[0] = (mpi_world_size == 2 && N > MAX_NUM_GPU)? (int)(N * 0.60) : N;
	size[1] = N - size[0];
#endif
        
  if (size[mpi_rank] < MAX_NUM_GPU) {
    num_devices = size[mpi_rank];
    for (int i = 0; i < size[mpi_rank]; i++) {
		 	Nsize[i] = 1;
		}
  }
	else {
    num_devices = MAX_NUM_GPU;
    for (int i = 0; i < MAX_NUM_GPU; i++) { 
      Nsize[i] = size[mpi_rank] / MAX_NUM_GPU;
      if(i < size[mpi_rank] % MAX_NUM_GPU) {
			 	Nsize[i]++;
			}
    }
  }

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

	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);
	}
}

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