#include "convolution.h"
#include <cstdio>
#include <cuda_runtime.h>
#include <mpi.h>
#include <stdio.h>

#define MAX_NUM_GPU 4
#define MAX_NUM_NODE 2
#define TS 32
#define WPT 8
#define RTS TS/WPT
int num_devices = 0;

#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);                                                                 \
    }                                                                          \
  }

void my_alloc_tensor(float **t, int D0, int D1, int D2, int D3) {
  *t = (float *) aligned_alloc(32, sizeof(float) * D0 * D1 * D2 * D3);
  if (*t == NULL) {
    printf("Failed to allocate memory for matrix.\n");
    exit(0);
  }
}

__global__ void sgemm(float *input, float *filter, float *output, 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;
}

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;

// Array of device (GPU) pointers
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_NODE], Nend[MAX_NUM_NODE];
static int begin[MAX_NUM_GPU], end[MAX_NUM_GPU];
//static int split_N[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) {
	input = _input;
  output = _output;
  filter = _filter;
  
	OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
  
	if (mpi_rank >= mpi_world_size) return;
	
	//if(mpi_rank != 0){
  //  my_alloc_tensor(&input, N, C, H, W);
  //  my_alloc_tensor(&output, N, K, OH, OW);
  //  my_alloc_tensor(&filter, K, C, R, S);
	//}
  
	if (mpi_rank != 0){
    input = (float *) aligned_alloc(32, sizeof(float) * N*C*H*W);
    filter = (float *) aligned_alloc(32, sizeof(float) * K*C*R*S);
    output = (float *) aligned_alloc(32, sizeof(float) * N*K*OH*OW);
  }

	// Scatter A
  if (mpi_rank == 0) {
    for (int i=1; i<mpi_world_size; i++) {
      MPI_Send(&input[Nbegin[i]*C*H*W], (Nend[i]-Nbegin[i])*C*H*W, MPI_FLOAT, i, 0, MPI_COMM_WORLD);
    }
  } 
	else {
    MPI_Recv(&input[Nbegin[mpi_rank]*C*H*W], (Nend[mpi_rank]-Nbegin[mpi_rank])*C*H*W, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, nullptr);
  }
	
	MPI_Bcast(filter, K*C*R*S, MPI_FLOAT, 0, MPI_COMM_WORLD);

  // Upload A and B matrix to every GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(input_d[i], input+begin[i]*C*H*W, (end[i]-begin[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 gridDim(end[i]-begin[i], K, OH);
		dim3 blockDim(OW, 1, 1);

    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(input_d[i], filter_d[i], output_d[i], end[i]-begin[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() );
  }
  
	// Download C matrix from GPUs
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(output+begin[i]*K*OH*OW, output_d[i], (end[i]-begin[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){
		for(int i=1; i<mpi_world_size; ++i){
			MPI_Recv(&output[Nbegin[i]*K*OH*OW], (Nend[i]-Nbegin[i])*K*OH*OW, MPI_FLOAT, i, 0, MPI_COMM_WORLD, nullptr);
		}
	}
	else{
		MPI_Send(&output[Nbegin[mpi_rank]*K*OH*OW], (Nend[mpi_rank]-Nbegin[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;
	
	OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;

  MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);
  MPI_Comm_size(MPI_COMM_WORLD, &mpi_world_size);
  
	if (mpi_rank >= mpi_world_size) return;
	
	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 < mpi_world_size; i++) {
    Nbegin[i] = (N / mpi_world_size) * i;
    Nend[i] = (N / mpi_world_size) * (i + 1);
  }
  Nend[mpi_world_size - 1] = N;
  
	// Seupt problem size for each GPU
	for (int i = 0; i < num_devices; i++) {
    begin[i] = Nbegin[mpi_rank] + (Nend[mpi_rank]-Nbegin[mpi_rank])/num_devices*i;
    end[i] = Nbegin[mpi_rank] + (Nend[mpi_rank]-Nbegin[mpi_rank])/num_devices*(i+1);
  }
  end[num_devices - 1] = Nend[mpi_rank];

  // Allocate device memory for each GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMalloc(&input_d[i], (end[i]-begin[i])*C*H*W*sizeof(float)) );
    CUDA_CALL( cudaMalloc(&filter_d[i], K*C*R*S*sizeof(float)) );
    CUDA_CALL( cudaMalloc(&output_d[i], (end[i]-begin[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.
  if (mpi_rank >= mpi_world_size) return;
}