#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 TS_X    2 		
//#define TS_Y    2 		
//#define TS_Z    32	
#define TS_X    2 		
#define TS_Y    2 		
#define TS_Z    32	

#define MAX_NUM_GPU 4
int num_devices = 0;

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];

inline void cuda_conv();
inline void cuda_conv_init(int,int);
inline void cuda_conv_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 p_n, remain, offset;

MPI_Status status;
MPI_Request request;

__global__ void conv(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 k = blockDim.x * blockIdx.x + threadIdx.x;
  int oh = blockDim.y * blockIdx.y + threadIdx.y;
  int ow = blockDim.z * blockIdx.z + threadIdx.z;

    //__shared__ float s_filter[K*C*R*S];
    //for (int i =0; i<K*C*R*S; i++) s_filter[i] = filter[i];
      
  if (k >= K || oh>= OH || ow >= OW) return;

  int CHW = C*H*W;
  int CRS = C*R*S;
  int HW = H*W;
  int RS = R*S;

if (C%8==0 && R==16 && S==16 && dilation==1 && pad==0 && stride==1)
{
  for(int n = 0; n < N; n++)
  {
    float o = 0.f;
    //for (int c = 0; c < C; c+8)
    //for (int c = 0; c < C; c++)
    //for (int c = 0; c < C; c=c+2)
    for (int c = 0; c < C; c=c+4)
    {
      for (int r = 0; r < R; r++)
      {
        for (int s = 0; s < S; s++)
	{
          int h = oh + r ;
	  int w = ow + s ;
          //if (h < 0 || h >= H || w < 0 || w >= W) continue;
	  //float i = input[n * C * H * W + c * H * W + (oh+r) * W + ow+s];
	  //float f = filter[k * C * R * S + c * R * S + r * S + s];
	  //o += i * f;

          //float i0 = input[n * C * H * W + c * H * W + h * W + w];
          //float f0 = filter[k * C * R * S + c * R * S + r * S + s];
          //float i1 = input[n * C * H * W + (c+1) * H * W + h * W + w];
          //float f1 = filter[k * C * R * S + (c+1) * R * S + r * S + s];
          //o += i0*f0 + i1*f1;

          //float i0 = input[n * C * H * W + c * H * W + h * W + w];
          //float f0 = filter[k * C * R * S + c * R * S + r * S + s];
          //float i1 = input[n * C * H * W + (c+1) * H * W + h * W + w];
          //float f1 = filter[k * C * R * S + (c+1) * R * S + r * S + s];
          //o += i0*f0 + i1*f1;
	  
          float i0 = input[n * CHW + c * HW + h * W + w];
          float f0 = filter[k * CRS + c * RS + r * S + s];
          float i1 = input[n * CHW + (c+1) * HW + h * W + w];
          float f1 = filter[k * CRS + (c+1) * RS + r * S + s];
          float i2 = input[n * CHW + (c+2) * HW + h * W + w];
          float f2 = filter[k * CRS + (c+2) * RS + r * S + s];
          float i3 = input[n * CHW + (c+3) * HW + h * W + w];
          float f3 = filter[k * CRS + (c+3) * RS + r * S + s];
          o += i0*f0 + i1*f1 + i2*f2 + i3*f3 ;

          //float i0 = input[n * C * H * W + c * H * W + h * W + w];
          //float f0 = filter[k * C * R * S + c * R * S + r * S + s];
          //float i1 = input[n * C * H * W + (c+1) * H * W + h * W + w];
          //float f1 = filter[k * C * R * S + (c+1) * R * S + r * S + s];
          //float i2 = input[n * C * H * W + (c+2) * H * W + h * W + w];
          //float f2 = filter[k * C * R * S + (c+2) * R * S + r * S + s];
          //float i3 = input[n * C * H * W + (c+3) * H * W + h * W + w];
          //float f3 = filter[k * C * R * S + (c+3) * R * S + r * S + s];
          //float i4 = input[n * C * H * W + (c+4) * H * W + h * W + w];
          //float f4 = filter[k * C * R * S + (c+4) * R * S + r * S + s];
          //float i5 = input[n * C * H * W + (c+5) * H * W + h * W + w];
          //float f5 = filter[k * C * R * S + (c+5) * R * S + r * S + s];
          //float i6 = input[n * C * H * W + (c+6) * H * W + h * W + w];
          //float f6 = filter[k * C * R * S + (c+6) * R * S + r * S + s];
          //float i7 = input[n * C * H * W + (c+7) * H * W + h * W + w];
          //float f7 = filter[k * C * R * S + (c+7) * R * S + r * S + s];
          //o += i0*f0 + i1*f1 + i2*f2 + i3*f3 + i4*f4 + i5*f5 + i6*f6 + i7*f7;
        }
      }
    }
    output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;
  }
}
else
{
  for(int n = 0; n < N; n++)
  {
    float o = 0.f;
    //for (int c = 0; c < C; c+8)
    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;

          //float i0 = input[n * C * H * W + c * H * W + h * W + w];
          //float f0 = filter[k * C * R * S + c * R * S + r * S + s];
          //float i1 = input[n * C * H * W + (c+1) * H * W + h * W + w];
          //float f1 = filter[k * C * R * S + (c+1) * R * S + r * S + s];
          //float i2 = input[n * C * H * W + (c+2) * H * W + h * W + w];
          //float f2 = filter[k * C * R * S + (c+2) * R * S + r * S + s];
          //float i3 = input[n * C * H * W + (c+3) * H * W + h * W + w];
          //float f3 = filter[k * C * R * S + (c+3) * R * S + r * S + s];
          //float i4 = input[n * C * H * W + (c+4) * H * W + h * W + w];
          //float f4 = filter[k * C * R * S + (c+4) * R * S + r * S + s];
          //float i5 = input[n * C * H * W + (c+5) * H * W + h * W + w];
          //float f5 = filter[k * C * R * S + (c+5) * R * S + r * S + s];
          //float i6 = input[n * C * H * W + (c+6) * H * W + h * W + w];
          //float f6 = filter[k * C * R * S + (c+6) * R * S + r * S + s];
          //float i7 = input[n * C * H * W + (c+7) * H * W + h * W + w];
          //float f7 = filter[k * C * R * S + (c+7) * R * S + r * S + s];
          //o += i0*f0 + i1*f1 + i2*f2 + i3*f3 + i4*f4 + i5*f5 + i6*f6 + i7*f7;
        }
      }
    }
    output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;
  }
}



}

inline void cuda_conv() {

  // Launch kernel on every GPU
  for (int i = 0; i < num_devices; i++) {
    dim3 blockDim(TS_X, TS_Y, TS_Z);
    dim3 gridDim((K+TS_X-1)/TS_X, (OH+TS_Y-1)/TS_Y, (OW+TS_Z-1)/TS_Z);

    CUDA_CALL( cudaSetDevice(i) );
    conv<<<gridDim, blockDim>>>(in_d[i], fil_d[i], out_d[i], Nend[i] - Nbegin[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( cudaSetDevice(i) );
    CUDA_CALL( cudaDeviceSynchronize() );
  }
}

inline void cuda_conv_init() {
  
  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] = (p_n / num_devices) * i;
    Nend[i] = (p_n / num_devices) * (i + 1); 
  }
  
  Nend[num_devices - 1] = p_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)) );
  }

  // 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( cudaSetDevice(i) );
	CUDA_CALL( cudaDeviceSynchronize() );
  }
}

inline void cuda_conv_final() {
  // Do any post-matmul cleanup work here.

  // 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( cudaSetDevice(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;
    
  p_n = N/mpi_world_size;
  remain= N%mpi_world_size;
  offset = remain;

  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 == 2)	  
    {
      offset += p_n;
      MPI_Isend(&offset, 1, MPI_INT, 1 , 1, MPI_COMM_WORLD,&request);
      MPI_Isend(&p_n, 1, MPI_INT, 1 , 1, MPI_COMM_WORLD,&request);
      MPI_Isend(&input[offset*C*H*W], p_n*C*H*W, MPI_FLOAT, 1 , 1, MPI_COMM_WORLD,&request);
      MPI_Isend(filter, K*C*R*S, MPI_FLOAT, 1, 1, MPI_COMM_WORLD,&request);
    }
	
    p_n += remain;
    cuda_conv_init();
    cuda_conv();
    cuda_conv_final();

    if ( mpi_world_size ==2)
    {
      MPI_Recv(&offset, 1, MPI_INT, 1, 2, MPI_COMM_WORLD, &status);
      MPI_Recv(&p_n, 1, MPI_INT, 1, 2, MPI_COMM_WORLD, &status);
      MPI_Recv(&output[offset*K*OH*OW], p_n*K*OH*OW, MPI_FLOAT, 1, 2, MPI_COMM_WORLD, &status);
    } 
  }
  
  if(mpi_rank > 0)
  {
    alloc_tensor(&input, p_n, C,H,W);
    alloc_tensor(&filter, K,C,R,S);
    alloc_tensor(&output, p_n, K,OH,OW);
	
    MPI_Recv(&offset, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &status);
    MPI_Recv(&p_n, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &status);
    MPI_Recv(input, p_n*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);

    cuda_conv_init();
    cuda_conv();
    cuda_conv_final();
    MPI_Send(&offset, 1, MPI_INT, 0, 2, MPI_COMM_WORLD);
    MPI_Send(&p_n, 1, MPI_INT, 0, 2, MPI_COMM_WORLD);
    MPI_Send(output, p_n*K*OH*OW, MPI_FLOAT, 0, 2, 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) {
}