#include "convolution.h"
#include "util.h"
#include <mpi.h>
#include <stdio.h>
#include <omp.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);                                                                 \
    }                                                                          \
  }

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;

#define MAX_NUM_GPU 4
int num_devices = 0;
int num_devices_change = 4;

// 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 Nbegin[MAX_NUM_GPU], Nend[MAX_NUM_GPU];
static int nDev = 1;
static int kDev = 1;
static int ohDev = 1;

#define MASTER_TO_SLAVE_TAG 1 //tag for messages sent from master to slaves
#define SLAVE_TO_MASTER_TAG 4 //tag for messages sent from slaves to master

MPI_Request request;
MPI_Status status;

static int min_value(int a, int b) { return a < b ? a : b; }

static int multi_s = 2;

__global__ void sgemm_2(float *input2, float *filter2, float *output2, 
						int N, int C, int H, int W, int K, int R, int S, int pad, int dilation, int stride, 
						int dev, int start_n_gpu, int end_n_gpu) {
  //int n = blockDim.x * blockIdx.x + threadIdx.x;  // raw index
  //int k = blockDim.y * blockIdx.y + threadIdx.y;  // column index
  //int oh = blockDim.z * blockIdx.z + threadIdx.z;  // column index
  	
	
	int n = blockIdx.x;
	int k = blockIdx.y;
	int oh = blockIdx.z;
	int ow = threadIdx.x;
				
  if(n >= (end_n_gpu-start_n_gpu)) return;
  if(k > K) return;

  int OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  int OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
 
  if(oh > OH) return;
  if(ow > OW) return;
   
	float o = 0.f;
	for (int c = 0; c < C; ++c) {
	  for (int r = 0; r < R; ++r) {
		  int h = oh * stride - pad + r * dilation;
		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 = input2[n * C * H * W + c * H * W + h * W + w];
		  float f = filter2[k * C * R * S + c * R * S + r * S + s];
		  
		  o += i * f;
		  //o2 += i2 + f2;
		  
		}
	  }
	}

	 output2[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;
		
  //printf("Dev(%d)(%d/%d) StartN(%d), EndN(%d), BDx(%d), BDy(%d), Bx(%d), By(%d), Tx(%d), Ty(%d) \n", dev, i, j, start_n_gpu, end_n_gpu ,
  //                                                      blockDim.x, blockDim.y, blockIdx.x, blockIdx.y, threadIdx.x, threadIdx.y);
}

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 size_n;
	int start_n;
	int end_n;
    
	int start_n_no0;
	int end_n_no0;  
  
  if(mpi_rank == 0){    
      
      for(int node=0; node< mpi_world_size; node++){
        
        size_n = N / (mpi_world_size);
        start_n = (node)*size_n;
        	
        if(node == 0) start_n_no0 = start_n;
        
        if(((node+1) == mpi_world_size) && ((N%(mpi_world_size)) != 0)){
          end_n = N;
          if(node == 0) end_n_no0 = end_n;
        }
        else{
          end_n = start_n + size_n;
          if(node == 0) end_n_no0 = end_n;
        }
        
        if(node >= 1){
          MPI_Send(&start_n, 1, MPI_INT, node, MASTER_TO_SLAVE_TAG, MPI_COMM_WORLD);
          //printf("SEND start_n (%d) to (%d)\n", start_n, node);
             
          MPI_Send(&end_n, 1, MPI_INT, node, MASTER_TO_SLAVE_TAG + 1, MPI_COMM_WORLD);
          //printf("SEND end_n (%d) to (%d)\n", end_n, node);
          
          int sizeBuf = (end_n - start_n) * C * H * W;
          //printf("SEND input Buff Len : (%d) to (%d)\n", sizeBuf, node);
          MPI_Send(&input[start_n*C * H * W], sizeBuf, MPI_FLOAT, node, MASTER_TO_SLAVE_TAG + 2, MPI_COMM_WORLD);
          //printf("SEND input size (%d) to (%d)\n", sizeBuf, node);
        }
      }
    }
    else{
		alloc_tensor(&input, N, C, H, W);			
        
        MPI_Recv(&start_n, 1, MPI_INT, 0, MASTER_TO_SLAVE_TAG, MPI_COMM_WORLD, &status);
        //printf("RECV start_n (%d) at (%d)\n", start_n, mpi_rank);
        
        MPI_Recv(&end_n, 1, MPI_INT, 0, MASTER_TO_SLAVE_TAG + 1, MPI_COMM_WORLD, &status);
        //printf("RECV end_n (%d) at (%d)\n", end_n, mpi_rank);
        
        int sizeBuf = (end_n - start_n) * C * H * W;
        //printf("RECV input Buff Len : %d \n", sizeBuf);
        MPI_Recv(&input[0], sizeBuf, MPI_FLOAT, 0, MASTER_TO_SLAVE_TAG + 2, MPI_COMM_WORLD, &status);       
        //printf("RECV input size (%d) at (%d)\n", sizeBuf, mpi_rank);        
    }

	OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
	OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
	  
	if(mpi_rank > 0){
		alloc_tensor(&filter, K, C, R, S);
    }
    MPI_Bcast(&filter[0], K*C*R*S, MPI_FLOAT, 0, MPI_COMM_WORLD);
    //printf("Bcast filter (%d) to (%d) \n", K*C*R*S, mpi_rank);
    
    if(mpi_rank > 0){
	  alloc_tensor(&output, N, K, OH, OW);
    }
	
    //MPI_Bcast(&output[0], N*K*OH*OW, MPI_FLOAT, 0, MPI_COMM_WORLD);
    //printf("Bcast output (%d) to (%d) \n", N*K*OH*OW, mpi_rank);

    
    MPI_Barrier(MPI_COMM_WORLD); 

	//timer_start(1);
	if(mpi_rank == 0){
		
		int size_n_gpu = end_n_no0 - start_n_no0;
		
		// N Devide with GPU
		// Setup problem size for each GPU
		for (int i = 0; i < num_devices; i++) {
			Nbegin[i] = (size_n_gpu / num_devices) * i;
			Nend[i] = (size_n_gpu / num_devices) * (i + 1);
		}
		int mod = (size_n_gpu % num_devices);
		int small_size = size_n_gpu / num_devices;
		Nend[num_devices - 1] = small_size * num_devices + mod;
		  
		// 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() );
		}
		
		
		/**************Start GPU Cal*************************/
		/*printf("Start GPU Cal\n");
 		for (int i = 0; i < num_devices; i++) {
			printf("Rank (%d), StartN(%d), EndN(%d)\n",  mpi_rank, Nbegin[i], Nend[i]);
		} */
		
		// Launch kernel on every GPU
		for (int i = 0; i < num_devices; i++) {
			//dim3 gridDim((Nend[i] - Nbegin[i])*C*H*W, 1, 1);    // grid : raw (Mend[i] - Mbegin[i]) , column (W)
			//dim3 blockDim(1, 1, 1);                    // block : 1 x 32
			//dim3 gridDim((Nend[i] - Nbegin[i])*C*H*W, 1, 1);    // grid : raw (Mend[i] - Mbegin[i]) , column (W)
			dim3 gridDim((Nend[i] - Nbegin[i]), K, OH);    // grid : raw (Mend[i] - Mbegin[i]) , column (W)
			//dim3 blockDim(nDev, kDev, ohDev);                    // block : 1 x 32
			dim3 blockDim(OW, 1);
			
			// 4600 GF
			/*
				dim3 blockDim(OW, 1);
				dim3 gridDim(Mend[i] - Mbegin[i], K, OH);
			
				int n = blockIdx.x;
				int k = blockIdx.y;
				int oh = blockIdx.z;
				int ow = threadIdx.x;
			*/

			CUDA_CALL( cudaSetDevice(i) );
			sgemm_2<<<gridDim, blockDim>>>(input_d[i], filter_d[i], output_d[i], N, C, H, W, K, R, S, pad, dilation, stride, i, Nbegin[i], Nend[i]);
		}

		// DO NOT REMOVE; NEEDED FOR TIME MEASURE
/* 		for (int i = 0; i < num_devices; i++) {
			CUDA_CALL( cudaDeviceSynchronize() );
		}   */
		//printf("End GPU Cal\n");
		/**************End GPU Cal*************************/
		
		// integrate ouput
		// Download C matrix from GPUs
		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() );
		}		
    }
     else{	 
		 
		int size_n_gpu = end_n - start_n;
		
		// N Devide with GPU
		// Setup problem size for each GPU
		for (int i = 0; i < num_devices; i++) {
			Nbegin[i] = (size_n_gpu / num_devices) * i;
			Nend[i] = (size_n_gpu / num_devices) * (i + 1);
		}
		int mod = (size_n_gpu % num_devices);
		int small_size = size_n_gpu / num_devices;
		Nend[num_devices - 1] = small_size * num_devices + mod;
		  	
		// 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() );
		}
		
		/**************Start GPU Cal*************************/		
/* 		printf("Start GPU Cal\n");
		for (int i = 0; i < num_devices; i++) {
			printf("Rank (%d), StartN(%d), EndN(%d)\n",  mpi_rank, Nbegin[i], Nend[i]);
		}
		 */
		// Launch kernel on every GPU
		for (int i = 0; i < num_devices; i++) {
			//dim3 gridDim((Nend[i] - Nbegin[i])*C*H*W, 1, 1);    // grid : raw (Mend[i] - Mbegin[i]) , column (W)
			//dim3 blockDim(1, 1, 1);                    // block : 1 x 32
			//dim3 gridDim((Nend[i] - Nbegin[i])*C*H*W, 1, 1);    // grid : raw (Mend[i] - Mbegin[i]) , column (W)
			dim3 gridDim((Nend[i] - Nbegin[i]), K, OH);    // grid : raw (Mend[i] - Mbegin[i]) , column (W)
			//dim3 blockDim(nDev, kDev, ohDev);                    // block : 1 x 32
			dim3 blockDim(OW, 1);
			
			CUDA_CALL( cudaSetDevice(i) );
			sgemm_2<<<gridDim, blockDim>>>(input_d[i], filter_d[i], output_d[i], N, C, H, W, K, R, S, pad, dilation, stride, i, Nbegin[i], Nend[i]);
		}

		// DO NOT REMOVE; NEEDED FOR TIME MEASURE
/* 		for (int i = 0; i < num_devices; i++) {
			CUDA_CALL( cudaDeviceSynchronize() );
		}   */
		//printf("End GPU Cal\n");
		/**************End GPU Cal*************************/
		
		// integrate ouput
		// Download C matrix from GPUs
		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_rank >= 1){
          MPI_Send(&start_n, 1, MPI_INT, 0, SLAVE_TO_MASTER_TAG, MPI_COMM_WORLD);
          //printf("SEND output start_n (%d) to (%d)\n", start_n, 0);
             
          MPI_Send(&end_n, 1, MPI_INT, 0, SLAVE_TO_MASTER_TAG + 1, MPI_COMM_WORLD);
          //printf("SEND output end_n (%d) to (%d)\n", end_n, 0);

          int sizeBuf = (end_n - start_n) * K*OH*OW;
          //printf("SEND output Buff Len : (%d) to (%d)\n", sizeBuf, 0);
          MPI_Send(&output[0], sizeBuf, MPI_FLOAT, 0, SLAVE_TO_MASTER_TAG + 2, MPI_COMM_WORLD);
          //printf("SEND output size (%d) to (%d)\n", sizeBuf, 0);
        }
    }
    //double elapsed_time = timer_stop(1);
    //double flops = (double) 2.0 * N * OH * OW * K * C * R * S / elapsed_time;
    //printf("Rank (%d), %f sec\n", mpi_rank, elapsed_time);
	
	
	
    if(mpi_rank == 0){
	  OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
	  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
      for (int node = 1; node < mpi_world_size; node++) {// untill all slaves have handed back the processed data
            //receive low bound from a slave
            MPI_Recv(&start_n, 1, MPI_INT, node, SLAVE_TO_MASTER_TAG, MPI_COMM_WORLD, &status);
            //receive upper bound from a slave
            MPI_Recv(&end_n, 1, MPI_INT, node, SLAVE_TO_MASTER_TAG + 1, MPI_COMM_WORLD, &status);
            //receive processed data from a slave
            MPI_Recv(&output[start_n*K*OH*OW], (end_n - start_n) * K*OH*OW, MPI_FLOAT, node, SLAVE_TO_MASTER_TAG + 2, MPI_COMM_WORLD, &status);
        }
    } 
}

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 init
	CUDA_CALL( cudaGetDeviceCount(&num_devices) );

	printf("Using %d devices\n", num_devices);
	printf("Change Device Num (%d)->(%d)\n", num_devices, num_devices_change);
	num_devices = num_devices_change;
	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);
	}	

}

void convolution_final(
    int _N, int _C, int _H, int _W,
    int _K, int _R, int _S,
    int _pad, int _dilation, int _stride) {

}