#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 MAX_NUM_GPU 4
int num_devices = 0;

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();
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;
static int stride_N[4];
static int offset[4];

MPI_Status status;
MPI_Request request;

__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 k = blockDim.x * blockIdx.x + threadIdx.x;
    int oh = blockDim.y * blockIdx.y + threadIdx.y;
    int ow = blockDim.z * blockIdx.z + threadIdx.z;
      
    if (k >= K || oh>= OH || ow >= OW) return;

     for(int n = 0; n < N; n++){
	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;
     }
}

void convolution_cuda() {

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

void convolution_cuda_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++) {
    Mbegin[i] = (stride_N[mpi_rank] / num_devices) * i;
    Mend[i] = (stride_N[mpi_rank] / num_devices) * (i + 1); 
  }
  Mend[num_devices - 1] = stride_N[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], (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] * 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) );
  }

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

void convolution_cuda_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 + Mbegin[i] * 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( 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;
 
  int divided_N = N / mpi_world_size;
  int modular_N = N % mpi_world_size;

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

  offset[0] = 0;
  stride_N[0] = divided_N;

  for (int i = 1; i < mpi_world_size; i++) {
  	if (i <= modular_N) stride_N[i] = divided_N + 1;
  	else stride_N[i] = divided_N;
  	offset[1] = divided_N;
  	if (i > 1) offset[i] = offset[i-1] + stride_N[i-1]; 
  }

  if (mpi_rank != 0) {
	alloc_tensor(&input, stride_N[mpi_rank], C, H, W);
	alloc_tensor(&output, stride_N[mpi_rank], K, OH, OW);
	alloc_tensor(&filter, K, C, R, S);
  }

  MPI_Bcast(filter, K*C*R*S, MPI_FLOAT, 0, MPI_COMM_WORLD);

  if (mpi_rank == 0) {
	for (int i = 1; i < mpi_world_size; i++)
  		MPI_Isend(&input[offset[i]*C*H*W], stride_N[i]*C*H*W, MPI_FLOAT, i, 1, MPI_COMM_WORLD, &request);
  }
  else {
		MPI_Recv(input, stride_N[mpi_rank]*C*H*W, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);
  }

  convolution_cuda_init();
  convolution_cuda();
  convolution_cuda_final();

  if (mpi_rank == 0) {
	for (int i = 1; i < mpi_world_size; i++)
  		MPI_Recv(&output[offset[i]*K*OH*OW], stride_N[i]*K*OH*OW, MPI_FLOAT, i, 2, MPI_COMM_WORLD, &status);
  } 
  else {
  	MPI_Send(output, stride_N[mpi_rank]*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) {
}