#include "convolution.h"
#include "util.h"
#include <mpi.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 8
#define TS 32

int num_devices = 0;
static int mpi_world_size, mpi_rank;
static float *input, *output, *filter;
//static int OH, OW;

__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) {
  
    int n = blockIdx.x;
    int k = blockIdx.y;
    int oh = blockIdx.z;
    int ow = threadIdx.x;

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

   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*OH*OW + k*OH*OW + oh*OW + ow] = o;
    
}

// 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 N, C, H, W, K, R, S, pad, dilation, stride, OH, OW;
static int Nbegin[MAX_NUM_GPU], Nend[MAX_NUM_GPU];
static int is[2], ie[2];

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;

 if(mpi_world_size > 1) {
  for (int i = 0; i < mpi_world_size; i++) {
            is[i] = N / mpi_world_size * i;
            ie[i] = N / mpi_world_size * (i + 1);
          }
          ie[mpi_world_size - 1] = N;
  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);
   }
  if (mpi_rank == 0) {
            for (int i = 1; i < mpi_world_size; i++) {
              MPI_Send(input+is[i]*C*H*W, (ie[i]-is[i])*C*H*W, MPI_FLOAT, i, 0, MPI_COMM_WORLD);
            }
          }
                else {
            MPI_Recv(input+is[mpi_rank]*C*H*W, (ie[mpi_rank]-is[mpi_rank])*C*H*W, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, nullptr);
          }

  // Broadcast B
  MPI_Bcast(filter, K*C*R*S, MPI_FLOAT, 0, MPI_COMM_WORLD); 
}
//if(mpi_rank == 0){
/*  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);
  }


if(mpi_world_size > 1){
  // Setup problem size for each GPU
  if(mpi_rank == 0){	  
    for (int i = 0; i < num_devices; i++) {
      Nbegin[i] = (N/mpi_world_size) / num_devices * i;
      Nend[i] = (N/mpi_world_size) / num_devices * (i + 1);
    }
    Nend[num_devices - 1] = N/mpi_world_size;
  }
  else {
     for (int i = 0; i < num_devices; i++) {
      Nbegin[i] = N/mpi_world_size + (N/mpi_world_size + N%mpi_world_size) / num_devices * i;
      Nend[i] = N/mpi_world_size + (N/mpi_world_size + N%mpi_world_size) / num_devices * (i+1);
    }
    Nend[num_devices - 1] = N;
  }
}
else {
  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)) );
  }*/
 // Upload A and B matrix to every GPU
  for (int i = 0; i < num_devices; 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) );
  }

  // 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);

    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<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( cudaDeviceSynchronize() );
  }
//}
/*   // Gather C
   if (mpi_rank == 0) {
     for (int i = 1; i < mpi_world_size; i++) {
        MPI_Recv(output+is[i]*K*OH*OW, (ie[i]-is[i])*K*OH*OW, MPI_FLOAT, i, 0, MPI_COMM_WORLD, nullptr);
     }
   }
   else {
     MPI_Send(output+is[mpi_rank]*K*OH*OW, (ie[mpi_rank]-is[mpi_rank])*K*OH*OW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD);
   }*/

// }
  for (int i = 0; i < num_devices; 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_world_size > 1){
   // Gather C
   if (mpi_rank == 0) {
     for (int i = 1; i < mpi_world_size; i++) {
        MPI_Recv(output+is[i]*K*OH*OW, (ie[i]-is[i])*K*OH*OW, MPI_FLOAT, i, 0, MPI_COMM_WORLD, nullptr);
     }
   }
   else {
     MPI_Send(output+is[mpi_rank]*K*OH*OW, (ie[mpi_rank]-is[mpi_rank])*K*OH*OW, 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);

   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);
  }


if(mpi_world_size > 1){
  // Setup problem size for each GPU
  if(mpi_rank == 0){
    for (int i = 0; i < num_devices; i++) {
      Nbegin[i] = (N/mpi_world_size) / num_devices * i;
      Nend[i] = (N/mpi_world_size) / num_devices * (i + 1);
    }
    Nend[num_devices - 1] = N/mpi_world_size;
  }
  else {
     for (int i = 0; i < num_devices; i++) {
      Nbegin[i] = N/mpi_world_size + (N/mpi_world_size + N%mpi_world_size) / num_devices * i;
      Nend[i] = N/mpi_world_size + (N/mpi_world_size + N%mpi_world_size) / num_devices * (i+1);
    }
    Nend[num_devices - 1] = N;
  }
}
else {
  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;
}

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


  // 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)) );
  }


}

void convolution_final(int _N, int _C, int _H, int _W, int _K, int _R, int _S, int _pad, int _dilation, int _stride) {
  // Do any post-matmul cleanup work here.

  //int OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  //int OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
  //float *input, *output, *filter;
//if(mpi_rank == 0) {
  // Download C matrix from GPUs
//if(mpi_rank == 0){
/*  for (int i = 0; i < num_devices; 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_world_size > 1){
   // Gather C
   if (mpi_rank == 0) {
     for (int i = 1; i < mpi_world_size; i++) {
        MPI_Recv(output+is[i]*K*OH*OW, (ie[i]-is[i])*K*OH*OW, MPI_FLOAT, i, 0, MPI_COMM_WORLD, nullptr);
     }
   }
   else {
     MPI_Send(output+is[mpi_rank]*K*OH*OW, (ie[mpi_rank]-is[mpi_rank])*K*OH*OW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD);
   }
} */

//}
}