#include "convolution.h"
#include "util.h"
#include <mpi.h>
#include <stdio.h>
#include <cstdio>
#include <cuda_runtime.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_CPU 2
#define MAX_NUM_GPU 4
#define TS  8

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 ns[MAX_NUM_CPU], ne[MAX_NUM_CPU];

static float *input_d[MAX_NUM_GPU];
static float *filter_d[MAX_NUM_GPU];
static float *output_d[MAX_NUM_GPU];
static int Nbegin[MAX_NUM_CPU][MAX_NUM_GPU], Nend[MAX_NUM_CPU][MAX_NUM_GPU];

int num_devices = 0;

__global__ void convolution_thread(
    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, int OW) {
  //printf("blockIdx.x:%d\n", blockIdx.x);
  //printf("blockDim.x:%d\n", blockDim.x);
  //printf("threadIdx.x:%d\n", threadIdx.x);
  //printf("blockIdx.y:%d\n", blockIdx.y);
#if 0
  int oh = (blockIdx.x * blockDim.x + threadIdx.x);
  int ow = (blockIdx.y * blockDim.y + threadIdx.y);
  //int k = (blockIdx.z * blockDim.z + threadIdx.z) % K;
  //int n = (blockIdx.z * blockDim.z + threadIdx.z) / K;
  int k = (blockIdx.z) % K;
  int n = (blockIdx.z) / K;

#else
  int ow = threadIdx.x;
  int oh = blockIdx.x;
  int k = blockIdx.y;
  int n = blockIdx.z;
  //int ow = threadIdx.x;
  //int oh = blockIdx.z;
  //int k = blockIdx.y;
  //int n = blockIdx.x;
#endif
  if (oh >= OH || ow >= OW || k >= K || n >= N) return;
  
  //__shared__ int *filter_d;

  //__syncthreads();
  //

  float o = 0.f;
  for (int c = 0; c < C; ++c) {
    for (int r = 0; r < R; ++r) {
      int h = oh * stride - pad + r * dilation;
      for (int s = 0; s < S; ++s) {
        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];
        //printf("input[n * C * H * W + c * H * W + h * W + w]  = %f\n", i);
        //printf("filter[k * C * R * S + c * R * S + r * S + s] = %f\n", f);
        o += i * f;
      }
    }
  }

  output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;
  //printf("output[n * K * OH * OW + k * OH * OW + oh * OW + ow]= %f\n", 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;

  if (mpi_rank != 0) {
    alloc_tensor(&input, N, C, H, W);
    alloc_tensor(&output, N, K, OH, OW);
    alloc_tensor(&filter, K, C, R, S);
  }


  if (mpi_world_size > 1) {
    if (mpi_rank == 0) {
      for (int i = 1; i < mpi_world_size; i++) {
        MPI_Send(input + ns[i] * C * H * W, (ne[i] - ns[i]) * C * H * W, MPI_FLOAT, i, 0,
            MPI_COMM_WORLD);
        MPI_Send(filter, K * C * R * S, MPI_FLOAT, i, 0, MPI_COMM_WORLD);
      }
    } else {
      MPI_Recv(input + ns[mpi_rank] * C * H * W, (ne[mpi_rank] - ns[mpi_rank]) * C * H * W, MPI_FLOAT,
          0, 0, MPI_COMM_WORLD, nullptr);
      MPI_Recv(filter, K * C * R * S, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, nullptr);
    }
  }

  //printf("ns[%d] : %d\n", mpi_rank, ns[mpi_rank]);
  //printf("ne[%d] : %d\n", mpi_rank, ne[mpi_rank]);
  //for(int i = 0; i < num_devices; i++)
  //{
  //  printf("Nbegin[%d][%d] : %d\n", mpi_rank,i,Nbegin[mpi_rank][i]);
  //  printf("Nend[%d][%d] : %d\n", mpi_rank,i,Nend[mpi_rank][i]);
  //}

  // Cuda start
  for (int i = 0; i < num_devices; i++) {
    if(Nend[mpi_rank][i] - Nbegin[mpi_rank][i] != 0)
    {
      //printf("Nend[mpi_rank][%d]:%d\n", i, Nend[mpi_rank][i]);
      //printf("Nbegin[mpi_rank][%d]:%d\n", i, Nbegin[mpi_rank][i]);
      //printf("%d, %d\n",ns[mpi_rank], Nbegin[mpi_rank][i]);
      CUDA_CALL( cudaMemcpy(input_d[i], input + (ns[mpi_rank] + Nbegin[mpi_rank][i]) * C * H * W,
            (Nend[mpi_rank][i] - Nbegin[mpi_rank][i]) * C * H * W * sizeof(float), cudaMemcpyHostToDevice) );
      CUDA_CALL( cudaMemcpy(filter_d[i], filter, K * C * R * S * sizeof(float), cudaMemcpyHostToDevice) );
    }
  }
 

  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }
  
  for (int i = 0; i < num_devices; i++) {
    //dim3 blockDim(OH, OW);
    //dim3 gridDim(Nend[mpi_rank][i] - Nbegin[mpi_rank][i], K);
    //printf("test[%d]\n", mpi_rank);
    if(Nend[mpi_rank][i] - Nbegin[mpi_rank][i] != 0)
    {
#if 0 
      dim3 blockDim(TS, TS);
      dim3 gridDim((OH + TS - 1) / TS, (OW + TS - 1)/ TS, ((Nend[mpi_rank][i] - Nbegin[mpi_rank][i]) * K));
      //printf("Nend[mpi_rank][%d]:%d\n", i, Nend[mpi_rank][i]);
      //printf("Nbegin[mpi_rank][%d]:%d\n", i, Nbegin[mpi_rank][i]);
#else
      dim3 blockDim(OW);
      dim3 gridDim(OH, K, Nend[mpi_rank][i] - Nbegin[mpi_rank][i]);
      //dim3 gridDim(Nend[mpi_rank][i] - Nbegin[mpi_rank][i], K, OH);
#endif
      CUDA_CALL( cudaSetDevice(i) );
      convolution_thread<<<gridDim, blockDim>>>(input_d[i], output_d[i], filter_d[i], (Nend[mpi_rank][i] - Nbegin[mpi_rank][i]),
          _C, _H, _W, _K, _R, _S, _pad, _dilation, _stride, OH, OW);
    }
  }

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

  for (int i = 0; i < num_devices; i++) {
    if(Nend[mpi_rank][i] - Nbegin[mpi_rank][i] != 0)
    {
      CUDA_CALL( cudaMemcpy(output + (ns[mpi_rank] + Nbegin[mpi_rank][i]) * K * OH * OW, output_d[i],
            (Nend[mpi_rank][i] - Nbegin[mpi_rank][i]) * K * OH * OW * sizeof(float),
            cudaMemcpyDeviceToHost) );
    }
  }
  // Cuda end

  if (mpi_world_size > 1) {
    if (mpi_rank == 0) {
      for (int i = 1; i <mpi_world_size; i++) {
        MPI_Recv(output + ns[i] * K * OH * OW, (ne[i] - ns[i]) * K * OH * OW, MPI_FLOAT, i, 0,
            MPI_COMM_WORLD, nullptr);
      }
    } else {
      MPI_Send(output + ns[mpi_rank] * K * OH * OW, (ne[mpi_rank] - ns[mpi_rank]) * K * OH * OW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD);
    }
  }
}

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;

  CUDA_CALL( cudaGetDeviceCount(&num_devices) );

  for (int i = 0; i < mpi_world_size; i++) {
    ns[i] = N / mpi_world_size * i;
    ne[i] = N / mpi_world_size * (i + 1);
    for (int j = 0; j < num_devices; j++) {
      Nbegin[i][j] = ((ne[i] - ns[i]) / num_devices) * j;
      Nend[i][j] = ((ne[i] - ns[i]) / num_devices) * (j + 1);
    }
  }
  ne[mpi_world_size - 1] = N;
  for(int i = 0; i < mpi_world_size; i++) {
    Nend[i][num_devices - 1] = ne[i] - ns[i];
  }

  //for (int i = 0; i< num_devices; i++) {
  //  printf("%d: Nbegin[%d]: %d\n",mpi_rank, i, Nbegin[mpi_rank][i]);
  //  printf("%d: Nend[%d]: %d\n",mpi_rank, i, Nend[mpi_rank][i]);
  //}

  for (int i = 0; i < num_devices; i++) {
    if(Nend[mpi_rank][i] - Nbegin[mpi_rank][i] != 0)
    {
      CUDA_CALL( cudaSetDevice(i) );
      CUDA_CALL( cudaMalloc(&input_d[i], (Nend[mpi_rank][i] - Nbegin[mpi_rank][i]) * C * H * W * sizeof(float)) );
      CUDA_CALL( cudaMalloc(&filter_d[i], K * C * R * S* sizeof(float)) );
      CUDA_CALL( cudaMalloc(&output_d[i], (Nend[mpi_rank][i] - Nbegin[mpi_rank][i]) * K * OH * OW * 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) {
}