#include "convolution.h"
#include "util.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_NUM_GPU (4)
#define BLOCK_SIZE (1024)

int num_devices = 0;
static int Nbegin[MAX_NUM_GPU], Nend[MAX_NUM_GPU], Nsize[MAX_NUM_GPU], Nslice;
// Array of device (GPU) pointers
static float *i_d[MAX_NUM_GPU];
static float *f_d[MAX_NUM_GPU];
static float *o_d[MAX_NUM_GPU];

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 sn, en, firstSize, modN;

__global__ void convOpt(
    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 now = blockDim.x * blockIdx.x + threadIdx.x;
  int k = blockIdx.y;
  int oh = blockIdx.z;
  int n = now / OW;
  int ow = now % OW;
  if (oh >= OH || ow >= OW || k >= K || n >= N) return;
  //for (int n = 0; n < N; ++n)
  {
    //for (int k = 0; k < K; ++k)
    {
      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;
    }
  }
}

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;
  MPI_Request request[2];
  MPI_Status status[2];

  if (N > 1)
  {
    if (mpi_rank == 0)
    {
      int snTmp = firstSize;
      int enTmp = N;
      MPI_Isend(input+snTmp*C*H*W, (enTmp-snTmp)*C*H*W, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request[0]);
      MPI_Isend(filter, K*C*R*S, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request[1]);
    }
    else
    {
      if (modN > 0)
      {
        alloc_tensor(&input, modN, C, H, W);
        alloc_tensor(&output, modN, K, OH, OW);
        alloc_tensor(&filter, K, C, R, S);
      }
      MPI_Irecv(input, modN*C*H*W, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request[0]);
      MPI_Irecv(filter, K*C*R*S, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request[1]);
    }
    MPI_Waitall(2,request, status);
  }
  if (modN > 0) {
    // Allocate device memory for each GPU
    for (int i = 0; i < num_devices; i++) {
      CUDA_CALL( cudaSetDevice(i) );
      CUDA_CALL( cudaMalloc(&i_d[i], Nsize[i]*C*H*W * sizeof(float)) );
      CUDA_CALL( cudaMalloc(&f_d[i], K*C*R*S * sizeof(float)) );
      CUDA_CALL( cudaMalloc(&o_d[i], Nsize[i]*K*OH*OW * sizeof(float)) );
    }
    // Upload A and B matrix to every GPU
    for (int i = 0; i < num_devices; i++) {
      CUDA_CALL( cudaSetDevice(i) );
      CUDA_CALL( cudaMemcpy(i_d[i], input + Nbegin[i]*C*H*W,
                 Nsize[i]*C*H*W * sizeof(float),
                 cudaMemcpyHostToDevice) );
      CUDA_CALL( cudaMemcpy(f_d[i], filter,
                 K*C*R*S * sizeof(float),
                 cudaMemcpyHostToDevice) );
    }

    // Launch kernel on every GPU
    for (int i = 0; i < num_devices; i++) {
      CUDA_CALL( cudaSetDevice(i) );
      dim3 blockDim(BLOCK_SIZE, 1, 1);
      dim3 gridDim((Nsize[i]*OW+BLOCK_SIZE-1)/BLOCK_SIZE, K, OH);
      convOpt<<<gridDim, blockDim>>>(i_d[i], o_d[i], f_d[i], Nsize[i],
                                  C, H, W, K, R, S,
                                  pad, dilation, stride);
    }

    // Download output matrix from GPUs
    for (int i = 0; i < num_devices; i++) {
      CUDA_CALL( cudaSetDevice(i) );
      CUDA_CALL( cudaMemcpy(output + Nbegin[i]*K*OH*OW, o_d[i],
                 Nsize[i]*K*OH*OW * sizeof(float),
                 cudaMemcpyDeviceToHost) );
    }
    // DO NOT REMOVE; NEEDED FOR TIME MEASURE
    for (int i = 0; i < num_devices; i++) {
      CUDA_CALL( cudaSetDevice(i) );
      CUDA_CALL( cudaDeviceSynchronize() );
    }
  }

  if (N > 1)
  {
    if (mpi_rank == 0)
    {
      int snTmp = firstSize;
      int enTmp = N;
      MPI_Irecv(output+snTmp*K*OH*OW, (enTmp-snTmp)*K*OH*OW, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request[0]);
    }
    else
    {
      MPI_Isend(output, modN*K*OH*OW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request[0]);
    }
    MPI_Wait(&request[0], &status[0]);
  }
}

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;

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

  MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);
  MPI_Comm_size(MPI_COMM_WORLD, &mpi_world_size);
  firstSize = (N + 1) / mpi_world_size;
  if (mpi_rank == 0)
  {
    sn = 0;
    en = firstSize;
    modN = firstSize;
  }
  else
  {
    sn = firstSize;
    en = N;
    modN = N - firstSize;
  }

  CUDA_CALL( cudaGetDeviceCount(&num_devices) );
  //printf("Using %d devices\n", num_devices);
  Nslice = (modN + num_devices - 1) / 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 problem size for each GPU
  for (int i = 0; i < num_devices; i++) {
    Nbegin[i] = Nslice * i;
    if (Nbegin[i] + Nslice < modN)
      Nend[i] = Nbegin[i] + Nslice;
    else if (Nbegin[i] < modN)
      Nend[i] = modN;
    else
      Nbegin[i]=0, Nend[i] = 0;
    Nsize[i] = Nend[i] - Nbegin[i];
  }
}

void convolution_final(
    int _N, int _C, int _H, int _W,
    int _K, int _R, int _S,
    int _pad, int _dilation, int _stride) {
  if (N > 1 && mpi_rank != 0)
  {
    free(input);
    free(output);
    free(filter);
  }
}