#include "convolution.h"
#include "util.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;

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;
static int size[2];

__global__ void 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) {

	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 = blockIdx.x;
	int k = blockIdx.y;
	int oh = blockIdx.z;
	int ow = threadIdx.x;

	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;
	    } // s
	   } // r
	  } // c
	output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;
}

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];

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) {

	MPI_Request request;
	MPI_Status status;

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

	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, OH, OW);
		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);
	}
  
  // Upload input and filter to every GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpyAsync(input_d[i], input + Nbegin[i]*C*H*W,
													(Nend[i] - Nbegin[i])*C*H*W*sizeof(float),cudaMemcpyHostToDevice) );
    CUDA_CALL( cudaMemcpyAsync(filter_d[i], filter, K*C*R*S*sizeof(float), cudaMemcpyHostToDevice) );
  }

	for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaDeviceSynchronize() );
  }

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

	// Download output from GPUs
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpyAsync(output + Nbegin[i] * K * OH * OW , output_d[i],
               (Nend[i] - Nbegin[i]) * K * OH * OW * sizeof(float),
               cudaMemcpyDeviceToHost) );
  }

  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaDeviceSynchronize() );
  }
 	
	if (mpi_rank == 0 && mpi_world_size == 2) {
		MPI_Recv(&output[size[0]*K*OH*OW], size[1]*K*OH*OW, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &status);		
	}
	else if (mpi_world_size == 2) {
		MPI_Isend(output, size[1]*K*OH*OW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request);
	}	
}


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);
  }

  // Setup problem size for each Node
	if (mpi_world_size == 2) 
		size[1] = N / 2;
	else 
		size[1] = 0;
	size[0] = N - size[1];

  // Setup problem size for each GPU per Node
	if (mpi_rank == 1 && mpi_world_size == 2) {
	  for (int i = 0; i < num_devices; i++) {
			Nbegin[i] = (size[1] / num_devices) * i;
			Nend[i] = (size[1] / num_devices) * (i + 1);
		}
		Nend[num_devices - 1] = size[1];
	} 
	else { // mpi_rank == 0
	  for (int i = 0; i < num_devices; i++) {
			Nbegin[i] = (size[0] / num_devices) * i;
			Nend[i] = (size[0] / num_devices) * (i + 1);
		}
		Nend[num_devices - 1] = size[0];
	}

	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(&output_d[i], (Nend[i] - Nbegin[i])*K*OH*OW*sizeof(float)) );
    CUDA_CALL( cudaMalloc(&filter_d[i], K*C*R*S*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) {
}