#include "util.h"
#include "convolution.h"
#include <mpi.h>
#include <stdio.h>
#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_NODES       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 num_devices;

static int Nstart[MAX_NUM_GPU], Nsize[MAX_NUM_GPU];
static int NN[MAX_NODES];

MPI_Status  status;
MPI_Request request;

static float *input_d[MAX_NUM_GPU];
static float *output_d[MAX_NUM_GPU];
static float *filter_d[MAX_NUM_GPU];

__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) {
  input = _input;
  output = _output;
  filter = _filter;

  // Scatter Input
  if (mpi_rank == 0 && NN[1] != 0) {
    MPI_Isend(&input[NN[0] * C * H * W], (NN[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_rank == 1 && NN[mpi_rank] != 0) {
    alloc_tensor(&input, NN[1], C, H, W);
    alloc_tensor(&output, NN[1], K, OH, OW);
    alloc_tensor(&filter, K, C, R, S);
    MPI_Recv(input, (NN[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(NN[mpi_rank] < MAX_NUM_GPU) num_devices = NN[mpi_rank];	  

  // Upload input and filter to every GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(input_d[i], input + Nstart[i] * C * H * W, Nsize[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( cudaSetDevice(i) );
    CUDA_CALL( cudaDeviceSynchronize() );
  }

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

    CUDA_CALL( cudaSetDevice(i) );
    conv<<<gridDim, blockDim>>>(input_d[i], output_d[i], filter_d[i], Nsize[i], C, H, W, K, R, S, pad, dilation, stride);
  }

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

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

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

  // Gather Output
  if (mpi_rank == 0) {
    MPI_Recv(&output[NN[0] * K * OH * OW], (NN[1] * K * OH * OW), MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &status);
  } else {
    MPI_Isend(output, NN[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);
  //printf("\nNode[%d] mpi_ramk = %d, mpi_world_size = %d\n", mpi_rank, mpi_rank, 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) );

  printf("Using %d devices\n", num_devices);

  for (int i = 0; i < num_devices; i++) {
    cudaDeviceProp prop;
    CUDA_CALL( cudaGetDeviceProperties(&prop, i) );

    // Try printing more detailed information here
    printf("Node[%d] [GPU %d] %s\n", mpi_rank, i, prop.name);
  }

  if (num_devices <= 0) {
    printf("No CUDA device found. Aborting\n");
    exit(1);
  }

  if(mpi_world_size == 2) {
    NN[0] = N - (N / 2);
    NN[1] = N / 2;
  } else {
    NN[0] = N;
    NN[1] = 0;
  }

  // Setup problem size for each GPU
  if(NN[mpi_rank] < MAX_NUM_GPU) { 
    num_devices = NN[mpi_rank];
    for(int i = 0; i < NN[mpi_rank]; i++) {
      Nstart[i] = i;
      Nsize[i] = 1;      
    }      
  } else {
    for (int i = 0; i < num_devices; i++) {
      Nstart[i] = (NN[mpi_rank] / num_devices) * i;
      Nsize[i]  = ((NN[mpi_rank] / num_devices) * (i + 1)) - Nstart[i];
    }
    Nsize[num_devices - 1] = NN[mpi_rank] - Nstart[num_devices - 1];
  }

  // Allocate device memory for each GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMalloc(&input_d[i],  Nsize[i] * C * H * W * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&filter_d[i], K * C * R * S * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&output_d[i], Nsize[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) {
}