#include "convolution.h"
#include <mpi.h>
#include <stdio.h>
#include "util.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 NODE 2
#define MAX_NUM_GPU 4

void convolution_cuda();
void convolution_cuda_init(int,int);
void convolution_cuda_final(int);

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 num_devices = 0;



__global__ void conv_kernel(float *input, float *filter, float *output,  int N, int C, int H, int W,
    int K, int R, int S, int OH, int OW, int pad, int dilation, int stride) {
    int m = blockDim.x * blockIdx.x + threadIdx.x;
    int j = blockDim.y * blockIdx.y + threadIdx.y;
    int k = blockDim.z * blockIdx.z + threadIdx.z;

    //int m = threadIdx.x;
    //int j = threadIdx.y;
    //int k = threadIdx.z;
      
   //printf("kernel test %d %d %d\n", blockDim.x, blockIdx.x, threadIdx.x);
   
   
   if (m >= K || j>= OH || k >= OW) return;

     for(int n=0;n<N;n++){
      //for (int k = 0; k < K; ++k) {
        //for (int oh = 0; oh < OH; ++oh) {
          //for (int ow = 0; ow < OW; ++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 = j * stride - pad + r * dilation;
                  int w = k * 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[m * C * R * S + c * R * S + r * S + s];
                  o += i * f;
                }
              }
            }
            output[n * K * OH * OW + m * OH * OW + j * OW + k] = o;
            //printf("kernel test = %f ",o);
          //}
          //printf("\n");
        //}
      //}
      }
}


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_Status status;
  MPI_Request request1,request2;
  
  int rows;
  int half = N/mpi_world_size;
  if(mpi_world_size == 2){
    rows = N-half;
  }else{
    rows = N;
  }
  
  int offset = N/mpi_world_size; 

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

  
  if (mpi_rank == 0) {
  
    if(mpi_world_size != 1){
      MPI_Isend(&input[(offset)*C*H*W], (rows)*C*H*W, MPI_FLOAT, 1 , 1, MPI_COMM_WORLD,&request1);
      MPI_Isend(filter, K*C*R*S, MPI_FLOAT, 1, 1, MPI_COMM_WORLD,&request2);
    }
      
      //printf("\n test %d %d\n",rows,offset);
    convolution_cuda_init(rows,0);
    convolution_cuda();
    convolution_cuda_final(0);

    if(mpi_world_size != 1){
      MPI_Recv(&output[(offset)*K*OH*OW], (rows)*K*OH*OW, MPI_FLOAT, 1, 2, MPI_COMM_WORLD, &status);
    }
    

  }else if(mpi_rank > 0){
    alloc_tensor(&input, rows, C,H,W);
    alloc_tensor(&filter, K,C,R,S);
    alloc_tensor(&output, rows, K,OH,OW);

     //printf("\n test11 %d %d %f\n",i,offset,*(input + (Mbegin[i]+offset) * C * H * W));
    MPI_Recv(input, (rows)*C*H*W, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);
    MPI_Recv(filter,  K*C*R*S, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);

    //printf("\n test %d %d\n",rows,offset);  
    convolution_cuda_init(rows,0);
    convolution_cuda();
    convolution_cuda_final(0);
   
    MPI_Send(output, (rows)*K*OH*OW, MPI_FLOAT, 0, 2, MPI_COMM_WORLD);

  }

}




// Array of device (GPU) pointers
static float *input_d[MAX_NUM_GPU];
static float *filter_d[MAX_NUM_GPU];
static float *output_d[MAX_NUM_GPU];
static int Mbegin[MAX_NUM_GPU], Mend[MAX_NUM_GPU];

void convolution_cuda() {

  // Launch kernel on every GPU
  for (int i = 0; i < num_devices; i++) {
    dim3 blockDim(4, 4, 64);
    dim3 gridDim((K+3)/4, (OH+3)/4, (OW+63)/64);

    CUDA_CALL( cudaSetDevice(i) );
    conv_kernel<<<gridDim, blockDim>>>(input_d[i], filter_d[i], output_d[i], Mend[i] - Mbegin[i], C, H ,W , K, R, S, OH, OW, pad, dilation, stride );
  }

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


void convolution_cuda_init(int rows,int offset) {
  
  //printf("\n test %d %d\n",rows,offset);
  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++) {
    Mbegin[i] = (rows / num_devices) * i;
    Mend[i] = (rows / num_devices) * (i + 1);
    //printf("\n test %d %d\n",Mbegin[i],Mend[i]);
    
  }
  
  Mend[num_devices - 1] = rows;

  // Allocate device memory for each GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMalloc(&input_d[i], (Mend[i] - Mbegin[i]) * C * H * W * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&filter_d[i], K * C * R * S * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&output_d[i], (Mend[i] - Mbegin[i]) * K* OH * OW * sizeof(float)) );
  }



  // Upload A and B matrix to every GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(input_d[i], input + (Mbegin[i]+offset) * C * H * W,
               (Mend[i] - Mbegin[i]) * C * H * W * sizeof(float),
               cudaMemcpyHostToDevice) );
    CUDA_CALL( cudaMemcpy(filter_d[i], filter, K * C * R * S * sizeof(float), cudaMemcpyHostToDevice) );
    //printf("\n test %d %d %f\n",i,offset,*(input + (Mbegin[i]+offset) * C * H * W));
  }

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

void convolution_cuda_final(int offset) {
  // Do any post-matmul cleanup work here.

  // Download C matrix from GPUs
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(output + (Mbegin[i]+offset) * K * OH * OW, output_d[i],
               (Mend[i] - Mbegin[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) {
}