#include "convolution.h"
#include <stdio.h>
#include <mpi.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 TS 8
#define MAX_NUM_GPU 4

static float *input, *output, *filter;
static float *in_d[MAX_NUM_GPU], *out_d[MAX_NUM_GPU], *filter_d[MAX_NUM_GPU];
static int N, C, H, W;
static int K, R, S;
static int OH, OW;
static int pad;
static int stride;
static int dilation;
static int mpi_rank, mpi_world_size;

static int n_devices = 1;

static int MP_size[2];
static int NUM_OFN[MAX_NUM_GPU];


__global__ void cuda_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 global_row = blockDim.x * blockIdx.x + threadIdx.x ;
  const int global_col = blockDim.y * blockIdx.y + threadIdx.y;
  int OH, OW;
  int NN, KK, WW;

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

  WW = global_col;
  NN = WW / (_K * OW);
  WW = WW - NN *(_K * OW);
  KK = WW / OW;
  WW = WW - KK * OW;

  int col = WW;
  int row = global_row;

  if (global_row >= OH || global_col >= _N*_K*OW) return;

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

  float OUT = 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 WW = start_col + j * _dilation;
	if (h < 0 || WW < 0 || h >= _H || WW >= _W) continue;
        float in = _input[NN*_C*_W*_H + c*_W*_H + h*_W + WW];
	float filt = _filter[KK*_C*_R*_S + c*_R*_S + i*_S + j];
	OUT += in * filt;
      }
    }
  } 

  _output[NN*_K*OH*OW + KK*OH*OW + row*OW + col] = OUT;
}

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;

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

    MPI_Request request;
    MPI_Status status;

    if (mpi_rank == 0 && mpi_world_size == 2 && MP_size[1] != 0) {
    	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_NUM_GPU) {
      	n_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_NUM_GPU) {
    	  n_devices = MP_size[mpi_rank];
      }
    }

    offset = 0;
    for (int i = 0 ; i < n_devices ; i++) {
    	CUDA_CALL( cudaMemcpy(in_d[i], input + offset, NUM_OFN[i]*C*H*W*sizeof(float), cudaMemcpyHostToDevice) );
    	CUDA_CALL( cudaMemcpy(filter_d[i], filter, K*C*R*S*sizeof(float),cudaMemcpyHostToDevice) );
    	offset += NUM_OFN[i] * C * H * W;
    }

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

    for (int i = 0; i < n_devices; i++) {
    	dim3 gridDim((OH+TS-1)/TS, (NUM_OFN[i]*K*OW + TS - 1)/TS, 1);
    	dim3 blockDim(TS, TS, 1);

    	CUDA_CALL( cudaSetDevice(i) );
    	cuda_conv<<<gridDim, blockDim>>>(in_d[i], out_d[i], filter_d[i], NUM_OFN[i], _C, _H, _W, _K, _R, _S, _pad, _dilation, _stride);
    }
    
  
    for (int i = 0; i < n_devices; i++) {
    	CUDA_CALL( cudaSetDevice(i) );
    	CUDA_CALL( cudaDeviceSynchronize() );
    }

  	offset = 0;
  for (int i = 0; i < n_devices; i++) {
  	CUDA_CALL( cudaSetDevice(i) );
  	CUDA_CALL( cudaMemcpy(output + offset, out_d[i], NUM_OFN[i]*K*OH*OW * sizeof(float), cudaMemcpyDeviceToHost) );
  	offset += NUM_OFN[i]*K*OH*OW;
  }

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

  if (mpi_rank == 0 && mpi_world_size == 2 && MP_size[1] != 0) {
  	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_NUM_GPU) {
    n_devices = MP_size[mpi_rank];
    for (int i = 0 ; i < MP_size[mpi_rank] ; i++) {
      NUM_OFN[i] = 1;
    }
  } else {
    n_devices = MAX_NUM_GPU;
    int remain = MP_size[mpi_rank] % MAX_NUM_GPU;
    int quot = MP_size[mpi_rank] / MAX_NUM_GPU;
    for (int i = 0 ; i < MAX_NUM_GPU ; i++) {
      NUM_OFN[i] = quot;
      if (i < remain) { 
	      NUM_OFN[i]++;
      }
    }
  }

  for (int i = 0 ; i < n_devices ; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMalloc(&in_d[i], NUM_OFN[i]*C*H*W*sizeof(float)) );
    CUDA_CALL( cudaMalloc(&out_d[i], NUM_OFN[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) {
}