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

#define TS 8 
#define MAX_NUM_GPU 4
#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 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;
int num_devices = 0;
__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) {
	const int globalRow = blockDim.x * blockIdx.x + threadIdx.x;
	const int globalCol = 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 = globalCol;
	n = w / (_K*OW);
	w = w - n * (_K*OW);
	k = w / OW;
	w = w - k*OW;
	int col = w;
	int row = globalRow;
	if (globalRow >= OH || globalCol >= _N*_K*OW) return;

	int start_row = row * _stride - _pad;
	int start_col = col * _stride - _pad;
	float o = 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 = _input[n*_C*_W*_H + c*_W*_H + h*_W + w];
				float fil = _filter[k*_C*_R*_S + c*_R*_S+ i*_S +j];
				o += in * fil;
			}
		}
	}
	_output[n*_K*OH*OW + k*OH*OW + row*OW + col] = o;

}

static float *in_d[MAX_NUM_GPU];
static float *out_d[MAX_NUM_GPU];
static float *fil_d[MAX_NUM_GPU];
static int Nbegin[MAX_NUM_GPU], Nend[MAX_NUM_GPU];
	int size[2];
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) {
  //CUDA_CALL( cudaGetDeviceCount(&num_devices) );
	MPI_Request request;
	MPI_Status status;

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

	if (mpi_world_size == 2) size[1] = _N/2;
	else size[1] = 0;
	size[0] = _N-size[1];

  OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
  
  if (mpi_rank == 0 && mpi_world_size == 2) {
		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);
	} else if (mpi_world_size == 2) {
		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);
	}
	int n = size[mpi_rank];
  CUDA_CALL( cudaGetDeviceCount(&num_devices) );
  for (int i = 0; i < num_devices; i++) {
    Nbegin[i] = (n/ num_devices) * i;
    Nend[i] = (n/ num_devices) * (i + 1);
  }
  Nend[num_devices - 1] = n;
  for (int i = 0; i < num_devices; i++) {
	}
  
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMalloc(&in_d[i], (Nend[i] - Nbegin[i]) * _C * _H * _W * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&out_d[i], (Nend[i] - Nbegin[i]) * _K * OH * OW * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&fil_d[i], _K * _C * _R * _S * sizeof(float)) );
  }
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(in_d[i], input + Nbegin[i] * _C*_H*_W,
               (Nend[i] - Nbegin[i]) * _C*_H*_W * sizeof(float),
               cudaMemcpyHostToDevice) );
    CUDA_CALL( cudaMemcpy(fil_d[i], filter, _K * _C*_R*_S * sizeof(float), cudaMemcpyHostToDevice) );
  }
  for (int i = 0; i < num_devices; i++) {
    dim3 gridDim((OH+TS-1)/TS, ((Nend[i] - Nbegin[i])*_K*OW+TS-1)/TS,1);
		dim3 blockDim(TS,TS,1);
		CUDA_CALL( cudaSetDevice(i) );
		conv<<<gridDim, blockDim>>>(in_d[i], out_d[i], fil_d[i], Nend[i] - Nbegin[i], _C,_H,_W,_K,_R,_S,_pad,_dilation,_stride);
	}

}
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) {
	MPI_Request request;
	MPI_Status status;
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(output + (Nbegin[i] * _K*OH*OW), out_d[i],
               (Nend[i] - Nbegin[i]) * _K*OH*OW * sizeof(float),
               cudaMemcpyDeviceToHost) );
  }
		if (mpi_rank == 0 && mpi_world_size == 2) {
			MPI_Recv(&output[size[0]*K*OH*OW], size[1]*K*OH*OW, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &status);
		} else if (mpi_world_size == 2) {
			MPI_Isend(output, size[1]*K*OH*OW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request);
		}
}