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

#define MAX_NUM_GPU 4

#define TS 8
int num_devices = 0;

int num_threads = 40;

static float *input, *output, *filter;
static int N, C, H, W;
static int K, R, S;
static int outH, outW;
static int pad;
static int dilation;
static int stride;
static int mpi_rank, mpi_world_size;

//static float *input, *output, *filter;

static float *input_d[MAX_NUM_GPU];
static float *output_d[MAX_NUM_GPU];
static float *filter_d[MAX_NUM_GPU];

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

#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 cuda_conv(
 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){
 
 const int globalRow = blockDim.x * blockIdx.x + threadIdx.x;
 const int globalCol = blockDim.y * blockIdx.y + threadIdx.y;
 
 int OH, OW;
 
 OH = (_H + 2 *_pad - _dilation * (_R - 1) -1) / _stride + 1;
 OW = (_W + 2 *_pad - _dilation * (_S - 1) -1) / _stride + 1;

 int n, k , w;
 w= globalCol;
 n = w/ (_K * OW);
 w= w- n *(_K * OW);
 k= w/ OW;
 w= w-k*OW;
 int col = w;
 int row = globalRow;

 if(globalRow>=OH || globalCol >= _N*_K*OW) return;

 int start_row = row * _stride - _pad;
 int start_col = col * _stride - _pad;

float o =0.0f;

for (int c = 0; c<_C;c++){
    for(int i=0; i<_R;i++){
        for(int j=0;j<_S;j++){
            int h = start_row + i *_dilation;
            int w = start_col + j *_dilation;
            if(h<0|| w<0 || h>=_H || w>=_W) continue;
            float in = _input[n*_C*_W*_H + c*_W*_H + h*_W+w];
            float fil = _filter[k*_C*_R*_S + c*_R*_S + i*_S+j];
            o += in * fil;
        }
    }
}
_output[n*_K*OH*OW + k*OH*OW + row*OW+col]=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) {
  

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

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



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

  if (mpi_world_size == 2) size[1] = _N / 2;
  else size[1] = 0;
  size[0] = N - size[1];

  outH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  outW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
  
  if (mpi_rank == 0 && mpi_world_size == 2) {
    MPI_Isend(&input[size[1]*C*H*W], size[1]*C*H*W, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request);
    MPI_Isend(filter, K*C*R*S, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request);
  } else if (mpi_world_size == 2) {
    alloc_tensor(&input, size[1], C, H, W);
    alloc_tensor(&output, size[1], K, outH, outW);
    alloc_tensor(&filter, K, C, R, S);
    MPI_Recv(input, size[1]*C*H*W, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
    MPI_Recv(filter, K*C*R*S, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
  }



    for (int i = 0; i < num_devices; i++) {
        CUDA_CALL( cudaSetDevice(i) );//jjlee
        printf("%s %d num_devices %d n (begin %d end %d)%d \n",__func__,__LINE__,num_devices,Nbegin[i],Nend[i],Nend[i] - Nbegin[i]);
        CUDA_CALL( cudaMemcpy(input_d[i], input + Nbegin[i] * _C*_H*_W,
                (Nend[i] - Nbegin[i]) * _C*_H*_W * sizeof(float),
                  cudaMemcpyHostToDevice) );
        CUDA_CALL( cudaMemcpy(filter_d[i], filter, _K * _C*_R*_S * sizeof(float), cudaMemcpyHostToDevice) );
    }
printf("%s %d \n",__func__,__LINE__);






  for (int i = 0; i < num_devices; i++) {
    dim3 gridDim((OH+TS-1)/TS, ( (Nend[i] - Nbegin[i]) *K*OW + TS - 1)/TS, 1);
    dim3 blockDim(TS, TS, 1);

    CUDA_CALL( cudaSetDevice(i) );
    cuda_conv<<<gridDim, blockDim>>>(input_d[i], output_d[i], filter_d[i],  (Nend[i] - Nbegin[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( cudaSetDevice(i) );//jjlee
    CUDA_CALL( cudaDeviceSynchronize() );
  }

}

#if 0
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) {
  int size[2];
  MPI_Request request;
  MPI_Status status;

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

  if (mpi_world_size == 2) size[1] = _N / 2;
  else size[1] = 0;
  size[0] = N - size[1];

  outH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  outW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
  
  if (mpi_rank == 0 && mpi_world_size == 2) {
    MPI_Isend(&input[size[0]*C*H*W], size[1]*C*H*W, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request);
    MPI_Isend(filter, K*C*R*S, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request);
  } else if (mpi_world_size == 2) {
    alloc_tensor(&input, size[1], C, H, W);
    alloc_tensor(&output, size[1], K, outH, outW);
    alloc_tensor(&filter, K, C, R, S);
    MPI_Recv(input, size[1]*C*H*W, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
    MPI_Recv(filter, K*C*R*S, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
  }
  
/*  
#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
    for (int n = 0; n < size[mpi_rank]; ++n) {
#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
      for (int k = 0; k < K; ++k) {
#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
        for (int oh = 0; oh < outH; ++oh) {

*/
#pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
    for (int n = 0; n < size[mpi_rank]; ++n) {
      for (int k = 0; k < K; ++k) {
        for (int oh = 0; oh < outH; ++oh) {          
#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
          for (int ow = 0; ow < outW; ++ow) {
            float o = 0.f;
            for (int c = 0; c < C; ++c) {// channel 
              for (int r = 0; r < R; ++r) { //filter width
                for (int s = 0; s < S; ++s) { //filter height
                  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;
                } 
              } // r
            } // c
            
            output[n * K * outH * outW + k * outH * outW + oh * outW + ow] = o;
            if(n==0&&k==0&&oh==0&&ow==0);
            printf("output[%0][%0][%0][%0] : _value = %f\n",o);
          } //ow
        } // oh
      } // k
    } // n

  if (mpi_rank == 0 && mpi_world_size == 2) {
    MPI_Recv(&output[size[0]*K*outH*outW], size[1]*K*outH*outW, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &status);
  } else if(mpi_world_size == 2){
    MPI_Isend(output, size[1]*K*outH*outW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request);
  }
 

}
#endif

#if 0
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;

  outH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  outW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
  
  if (mpi_rank == 0) {

    for (int n = 0; n < N; ++n) {
      for (int k = 0; k < K; ++k) {
        for (int oh = 0; oh < outH; ++oh) {
          for (int ow = 0; ow < outW; ++ow) {
            float o = 0.f;
            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 * outH * outW + k * outH * outW + oh * outW + ow] = o;
          }
        }
      }
    }

  }
}
#endif


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;

  int size[2];
  MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);
  MPI_Comm_size(MPI_COMM_WORLD, &mpi_world_size);

  if (mpi_world_size == 2) size[1] = _N / 2;
  else size[1] = 0;
  size[0] = N - size[1];

  N=size[0];

 int OH, OW;
 
 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) );
  //num_devices=2;

  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(&input_d[i], (Nend[i] - Nbegin[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[i] - Nbegin[i]) * K*OH*OW * sizeof(float)) );
 }

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


}

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

 int OH, OW;
  int size[2];
  MPI_Request request;
  MPI_Status status;

  if (mpi_world_size == 2) size[1] = _N / 2;
  else size[1] = 0;
  size[0] = _N - size[1];

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

  // Download C matrix from GPUs
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMemcpy(output + Nbegin[i] * _K*OH*OW, output_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() );
  }



  if (mpi_rank == 0 && mpi_world_size == 2) {
    printf("%s %d IRecv size[1] %d  %d \n",__func__,__LINE__,size[1],size[1]*K*outH*outW);
    MPI_Recv(&output[size[1]*K*outH*outW], size[1]*K*outH*outW, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &status);
  } else if(mpi_world_size == 2){
    printf("%s %d ISend size[1] %d %d \n",__func__,__LINE__,size[1],size[1]*K*outH*outW);
    MPI_Isend(output, size[1]*K*outH*outW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request);
  }
 


}