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

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 input_line;
static int filter_size;
static int output_line;

#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
int num_devices = 0;

// 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 int Nbegin[MAX_NUM_GPU], Nend[MAX_NUM_GPU];

__global__ void conv4(float *input, float* filter, float* output,
        int Nsplit, int C, int H, int W, int K, int R, int S, int OH, int OW,
        int pad, int dilation, int stride)
{
  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) { // # of input channel
    for (int r = 0; r < R; ++r) { // filter height (=row)
      for (int s = 0; s < S; s+=4) { // filter width (=column)
        int h = oh * stride - pad + r * dilation;
        int w = ow * stride - pad + s * dilation;
        if (h < 0 || h >= H || w < 0 || w >= W) continue;
        int idx_i = (n * C * H * W) + (c * H * W) + (h * W) + w;
        int idx_f = (k * C * R * S) + (c * R * S) + (r * S) + s;
        float4 i = make_float4(input[idx_i + 0], input[idx_i + 1], input[idx_i + 2], input[idx_i + 3]);
        float4 f = reinterpret_cast<float4*>(filter)[idx_f/4];
        o = o
          + i.x * f.x
          + i.y * f.y
          + i.z * f.z
          + i.w * f.w
          ;
      }
    }
  }
  output[(n * K * OH * OW) + (k * OH * OW) + (oh * OW) + ow] = o;

}

__global__ void conv(float *input, float* filter, float* output,
        int Nsplit, int C, int H, int W, int K, int R, int S, int OH, int OW,
        int pad, int dilation, int stride)
{
  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) { // # of input channel
    for (int r = 0; r < R; ++r) { // filter height (=row)
      for (int s = 0; s < S; ++s) { // filter width (=column)
        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 cuda_init(int ns, int ne)
{
  int N_ = ne - ns;
  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 problem size for each GPU
  for (int i = 0; i < num_devices; i++) {
    Nbegin[i] = ns + (N_ / num_devices) * i;
    Nend[i]   = ns + (N_ / num_devices) * (i + 1);
    /* debug */// printf("%d: Nbegin[%d] = %d, Nend[%d] = %d\n", mpi_rank, i, Nbegin[i], i, Nend[i]);
  }
  Nend[num_devices - 1] = ns + N_;

  // Allocate device memory for each GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMalloc(&i_d[i], (Nend[i] - Nbegin[i]) * input_line * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&f_d[i], filter_size * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&o_d[i], (Nend[i] - Nbegin[i]) * output_line * sizeof(float)) );
  }

  // Upload A and B matrix to every GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(i_d[i], input + Nbegin[i] * input_line,
          (Nend[i] - Nbegin[i]) * input_line * sizeof(float),
          cudaMemcpyHostToDevice) );
    CUDA_CALL( cudaMemcpy(f_d[i], filter, filter_size * sizeof(float), cudaMemcpyHostToDevice) );
  }

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

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

    if (S==16 && R==16 && dilation==1) {
      CUDA_CALL( cudaSetDevice(i) );
      conv4<<<gridDim, blockDim>>>(i_d[i], f_d[i], o_d[i],
          Nsplit, C, H, W, K, R, S, OH, OW,
          pad, dilation, stride);
    } else {
      // printf("\nconv4!\n");
      CUDA_CALL( cudaSetDevice(i) );
      conv<<<gridDim, blockDim>>>(i_d[i], f_d[i], o_d[i],
          Nsplit, C, H, W, K, R, S, OH, OW,
          pad, dilation, stride);
    }
  }

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

void cuda_final()
{
  // Download C matrix from GPUs
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(output + Nbegin[i] * output_line, o_d[i],
          (Nend[i] - Nbegin[i]) * output_line * sizeof(float),
          cudaMemcpyDeviceToHost) );
  }

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

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;

  OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
  input_line = C*H*W;
  filter_size = K*C*R*S;
  output_line = K*OH*OW;

  // Allocate
  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);
  }

  // N split
  int ns[mpi_world_size], ne[mpi_world_size];
  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;

  // Scatter Input
  if (mpi_rank == 0) {
    for (int i = 1; i < mpi_world_size; i++) {
      MPI_Send(input + ns[i] * input_line, (ne[i] - ns[i]) * input_line, MPI_FLOAT, i, 0,
               MPI_COMM_WORLD);
    }
  } else {
    MPI_Recv(input + ns[mpi_rank] * input_line, (ne[mpi_rank] - ns[mpi_rank]) * input_line, MPI_FLOAT,
             0, 0, MPI_COMM_WORLD, nullptr);
  }

  // Broadcast Filter
  MPI_Bcast(filter, filter_size, MPI_FLOAT, 0, MPI_COMM_WORLD);

  cuda_init(ns[mpi_rank], ne[mpi_rank]);
  cuda_run();
  cuda_final();

  // Gather Output
  if (mpi_rank == 0) {
    for (int i = 1; i < mpi_world_size; i++) {
      MPI_Recv(output + ns[i] * output_line, (ne[i] - ns[i]) * output_line, MPI_FLOAT, i, 0,
               MPI_COMM_WORLD, nullptr);
    }
  } else {
    MPI_Send(output + ns[mpi_rank] * output_line, (ne[mpi_rank] - ns[mpi_rank]) * output_line, 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);
}

void convolution_final(
    int _N, int _C, int _H, int _W,
    int _K, int _R, int _S,
    int _pad, int _dilation, int _stride) {
}