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

#define min(a, b) (((a) < (b)) ? (a) : (b))
#define max(a, b) (((a) > (b)) ? (a) : (b))

#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
#define MAX_NUM_MPI 2
int num_devices = 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 ns[2], ne[2];
static int Nbegin[MAX_NUM_MPI][MAX_NUM_GPU], Nend[MAX_NUM_MPI][MAX_NUM_GPU];
static float *input_d[MAX_NUM_GPU];
static float *filter_d[MAX_NUM_GPU];
static float *output_d[MAX_NUM_GPU];

MPI_Status status;
MPI_Request request;

__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) {
 
  const int globalRow = blockDim.x * blockIdx.x + threadIdx.x;
  const int globalCol = blockDim.y * blockIdx.y + threadIdx.y;
  
  int _OH = (_H + 2 * _pad - _dilation * (_R - 1) - 1) / _stride + 1;
  int _OW = (_W + 2 * _pad - _dilation * (_S - 1) - 1) / _stride + 1;

  if (globalRow >= _OH || globalCol >= _N * _K * _OW) return;

  int n = globalCol / (_K * _OW);
  int cw = globalCol - n * (_K * _OW);
  int k = cw / _OW;
      cw -= k * _OW;

  float o = 0.f;
  int row = globalRow * _stride - _pad;
  int col = cw * _stride - _pad;

  for (int c = 0; c < _C; ++c) {
    for (int r = 0; r < _R; ++r) {
      int h = row + r * _dilation;
      if (h < 0 || h >= _H) continue;
      for (int s = 0; s < _S; ++s) {
        int w = col + s * _dilation;
        if (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;
      } //s
    } //r
  } //c
  
  _output[n * _K * _OH * _OW + k * _OH * _OW + globalRow * _OW + cw] = o;

  // Synchronise before loading the next tile
  __syncthreads();
}

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) {
    for (int i = 1; i < mpi_world_size; i++) {
      MPI_Isend(&input[ns[i]*C*H*W], (ne[i]-ns[i])*C*H*W, MPI_FLOAT, i, 1, MPI_COMM_WORLD, &request);
      MPI_Isend(filter, K*C*R*S, MPI_FLOAT, i, 2, MPI_COMM_WORLD, &request);
    }
  }
  else { 
    alloc_tensor(&input, N, C, H, W);
    alloc_tensor(&filter, K, C, R, S);
    alloc_tensor(&output, N, K, OH, OW);

    MPI_Recv(&input[ns[mpi_rank]*C*H*W], (ne[mpi_rank]-ns[mpi_rank])*C*H*W, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);
    MPI_Recv(filter, K*C*R*S, MPI_FLOAT, 0, 2, MPI_COMM_WORLD, &status);
  }

  // Upload A and B matrix to every GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(input_d[i], input + 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) );
  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }
  
///////////////// Start Calculation ///////////////////
  // Launch kernel on every GPU

  for (int i = 0; i < num_devices; i++) {
    dim3 gridDim((OH+TS-1)/TS,((Nend[mpi_rank][i] - Nbegin[mpi_rank][i])*K*OW+TS-1)/TS, 1); 
    dim3 blockDim(TS, TS, 1);
    
    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(input_d[i], filter_d[i], output_d[i], 
    Nend[mpi_rank][i] - Nbegin[mpi_rank][i], _C, _H, _W, _K, _R, _S, _pad, _dilation, _stride);
  }

  // NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }
}

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);
/////////////////////// Init Cuda ///////////////////////
  CUDA_CALL( cudaGetDeviceCount(&num_devices) );
  
  if(num_devices > MAX_NUM_GPU) num_devices = MAX_NUM_GPU;

  printf("[MPI:%d] Using %d devices\n", mpi_rank, num_devices);
  for (int i = 0; i < num_devices; i++) {
    cudaDeviceProp prop;
    CUDA_CALL( cudaGetDeviceProperties(&prop, i) );

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

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

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

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

  // Download output matrix from GPUs
  for (int i = 0; i < num_devices; i++) {
  CUDA_CALL( cudaMemcpy(output + Nbegin[mpi_rank][i] * K * OH * OW, output_d[i], 
             (Nend[mpi_rank][i] - Nbegin[mpi_rank][i]) * K * OH * OW * sizeof(float),
             cudaMemcpyDeviceToHost) );
  }
    
  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, 1, 1, MPI_COMM_WORLD, &status);
    }
  }
  else { 
    MPI_Isend(&output[ns[mpi_rank]*K*OH*OW], (ne[mpi_rank]-ns[mpi_rank])*K*OH*OW, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &request);
  }

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

}