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

#include <cstdio>
#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
int num_devices = 0;

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 float *in_d[MAX_NUM_GPU];
static float *fil_d[MAX_NUM_GPU];
static float *out_d[MAX_NUM_GPU];
static int Nbegin[MAX_NUM_GPU], Nend[MAX_NUM_GPU];
int N_dev[MAX_NUM_GPU];

__global__ void sgemm(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) {

    float* input = _input;
    float* output = _output;
    float* filter = _filter;

    int N = _N; int C = _C; int H = _H; int W = _W;
    int K = _K; int R = _R; int S = _S;
    int pad = _pad;
    int dilation = _dilation;
    int stride = _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 = 0.0f;
    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) {

  if ( mpi_rank != 0 ) return;

  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;

  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] = (N / num_devices) * i;
    Nend[i] = (N / num_devices) * (i + 1);
  }
  Nend[num_devices - 1] = N;

  // Allocate device memory for each GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMalloc(&in_d[i], (Nend[i]-Nbegin[i]) * C*H*W * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&fil_d[i], K*C*R*S * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&out_d[i], (Nend[i]-Nbegin[i]) * K*OH*OW * sizeof(float)) );
  }
  
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }

  // Upload A and B matrix to every GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(in_d[i], input + Nbegin[i] * C*H*W, (Nend[i]-Nbegin[i]) * C*H*W * sizeof(float), cudaMemcpyHostToDevice) );
    CUDA_CALL( cudaMemcpy(fil_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((Nend[i]-Nbegin[i]), K, OH);
    
    N_dev[i]=Nend[i]-Nbegin[i];

    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(in_d[i], out_d[i], fil_d[i], N_dev[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 + Nbegin[i] * K*OH*OW, out_d[i], (Nend[i]-Nbegin[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() );
  }

}

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) {
}