#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);                                                            \
  }                                                                     \
}

__global__ void sgemm(float *_input, float *_filter, float *_output, 
    int _N, int _C, int _H, int _W, int _K, int _R, int _S, 
    int _pad, int _dilation, int _stride) {
  int n = blockIdx.x;
  int k = blockIdx.y;
  int oh = blockIdx.z;
  int ow = threadIdx.x;

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

  float o = 0.0f;
  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;
}

#define MAX_NUM_NODE 2
#define MAX_NUM_GPU 4

static int num_devices = 0;

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;

// Array of device (GPU) pointers
static float *input_d[MAX_NUM_GPU];
static float *filter_d[MAX_NUM_GPU];
static float *output_d[MAX_NUM_GPU];
//static int N, C, H, W, K, R, S, pad, dilation, stride;
static int Nbegin[MAX_NUM_NODE], Nend[MAX_NUM_NODE];
static int Gbegin[MAX_NUM_GPU], Gend[MAX_NUM_GPU];

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 >= mpi_world_size) return;

  //OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  //OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;

  //if (mpi_rank == 0) {

  /*---------- seperating for nodes ----------*/
  if (mpi_rank != 0){
    input = (float *) aligned_alloc(32, sizeof(float) * N*C*H*W);
    filter = (float *) aligned_alloc(32, sizeof(float) * K*C*R*S);
    output = (float *) aligned_alloc(32, sizeof(float) * N*K*OH*OW);
  }

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


  /*---------- calculating ----------*/
  // Upload A and B matrix to every GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(input_d[i], input + Gbegin[i] * C*H*W, (Gend[i] - Gbegin[i]) * C*H*W * sizeof(float), cudaMemcpyHostToDevice) );
    CUDA_CALL( cudaMemcpy(filter_d[i], filter, K*C*R*S * sizeof(float), cudaMemcpyHostToDevice) );
  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }

  // Launch kernel on every GPU
  for (int i = 0; i < num_devices; i++) {
    dim3 blockDim( OW, 1, 1);
    dim3 gridDim(Gend[i]-Gbegin[i], K, OH);

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

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }

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

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }

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

  //if (mpi_rank == 0) {

  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("[GPU %d] %s\n", i, prop.name);
  }

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

  // Setup size for each Node
  for (int i=0; i < mpi_world_size; ++i){
    Nbegin[i] = N / mpi_world_size * i;
    Nend[i] = N / mpi_world_size * (i+1);
  }
  Nend[mpi_world_size-1] = N;

  // Setup problem size for each GPU
  for (int i = 0; i < num_devices; i++) {
    Gbegin[i] = Nbegin[mpi_rank] + (Nend[mpi_rank] - Nbegin[mpi_rank]) / num_devices * i;
    Gend[i] = Nbegin[mpi_rank] + (Nend[mpi_rank] - Nbegin[mpi_rank]) / num_devices * (i + 1);
  }
  Gend[num_devices - 1] = Nend[mpi_rank];

  // Allocate device memory for each GPU
  OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;

  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMalloc(&input_d[i], (Gend[i] - Gbegin[i]) * C*H*W * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&filter_d[i], K*C*R*S * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&output_d[i], (Gend[i] - Gbegin[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) {

  if (mpi_rank >= mpi_world_size) return;
  //if (mpi_rank == 0){
  //}
}