#include <stdio.h>
#include <cstdio>

#include <cuda_runtime.h>
#include "convolution.h"
#include "util.h"
#include <mpi.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
#define TS 8

// Array of device (GPU) pointers
static float *input, *output, *filter;
static float *in_d[MAX_NUM_GPU], *out_d[MAX_NUM_GPU], *fil_d[MAX_NUM_GPU];

static int N, C, H, W;
static int K, R, S;
static int pad, dilation, stride;
static int mpi_rank, mpi_world_size;

static int num_devices =1;
static int Chunk_size[2];
static int NN[MAX_NUM_GPU];
static int OH, OW;



__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){

	const int globalRow = blockDim.x * blockIdx.x + threadIdx.x;
	const int globalCol = blockDim.y * blockIdx.y + threadIdx.y;
	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, k, w;
	w = globalCol;
	n = w/(_K*OW);
	w = w - n*(_K*OW);
	k = w/OW;
	w = w - k*OW;

	int col = w;
	int row = globalRow;

	if(globalRow >= OH || globalCol >= _N*_K*OW) return;

	int start_row = row*_stride - _pad;
	int start_col = col*_stride - _pad;

	float o = 0.0f;
	for(int c=0; c<_C; c++){
		for(int i=0; i<_R; i++){
			for(int j=0; j<_S; j++){
				int h = start_row + i*_dilation;
				int w = start_col + j*_dilation;
				if(h<0 || w<0 || h>= _H || w>= _W) continue;
				float in = _input[n*_C*_W*_H + c*_W*_H + h*_W + w];
				float fil = _filter[k*_C*_R*_S + c*_R*_S + i*_S + j];
				o+= in*fil;
			}
		}
	}
	_output[n*_K*OH*OW + k*OH*OW + row*OW + col] = o;
}




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){

	int offset = 0;
	MPI_Request request;
	MPI_Status status;

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

	if(mpi_rank == 0 && mpi_world_size ==2 && Chunk_size[1] != 0){
		MPI_Isend(&input[Chunk_size[0]*C*H*W], Chunk_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);
		if(Chunk_size[mpi_rank] < MAX_NUM_GPU){
			num_devices = Chunk_size[mpi_rank];
		}
	}
	else if(mpi_rank ==1 && Chunk_size[mpi_rank] != 0){
		alloc_tensor(&input, Chunk_size[1], C, H, W);
		alloc_tensor(&output, Chunk_size[1], K, OH, OW);
		alloc_tensor(&filter, _K, _C, _R, _S);

		MPI_Recv(input, Chunk_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);
		if(Chunk_size[mpi_rank] < MAX_NUM_GPU){
			num_devices = Chunk_size[mpi_rank];
		}
	}

	offset = 0;
        for(int i=0; i<num_devices; i++){
		CUDA_CALL( cudaMemcpy(in_d[i], input+offset, NN[i]*C*H*W*sizeof(float), cudaMemcpyHostToDevice) );
    		CUDA_CALL( cudaMemcpy(fil_d[i], filter, K*C*R*S*sizeof(float), cudaMemcpyHostToDevice) );
  		offset += NN[i] * C * H * W;
	}

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

        // Launch kernel on every GPU
  	for (int i = 0; i < num_devices; i++) { 	//최적화 필요
    		dim3 blockDim(TS, TS, 1);
    		dim3 gridDim((OH+TS-1)/TS, (NN[i]*K*OW+TS-1)/TS,1);

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

	offset = 0;
	for(int i=0; i<num_devices; i++){
		CUDA_CALL( cudaSetDevice(i) );
		CUDA_CALL( cudaMemcpy(output + offset, out_d[i],
					NN[i]*K*OH*OW*sizeof(float), cudaMemcpyDeviceToHost) );
		offset += NN[i]*K*OH*OW;
	}

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

	if(mpi_rank==0 && mpi_world_size==2 && Chunk_size[1]!=0){
		MPI_Recv(&output[Chunk_size[0]*K*OH*OW], Chunk_size[1]*K*OH*OW, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &status);
	}
	else if(mpi_rank==1 && Chunk_size[1]!=0){
		MPI_Isend(output, Chunk_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);

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

	if(mpi_world_size ==2 && _N>4) Chunk_size[1] = _N/2;
	else Chunk_size[1]=0;
	
	Chunk_size[0] = N- Chunk_size[1];

	if(Chunk_size[mpi_rank] < MAX_NUM_GPU){
		num_devices = Chunk_size[mpi_rank];
		for(int i=0; i<Chunk_size[mpi_rank]; i++)
			NN[i] = 1;
	}
	else{
		num_devices = MAX_NUM_GPU;
		int quotient = Chunk_size[mpi_rank]/MAX_NUM_GPU;
		int remain = Chunk_size[mpi_rank]%MAX_NUM_GPU;
		for(int i=0; i<MAX_NUM_GPU; i++){
			NN[i] = quotient;
			if(i<remain) NN[i]++;
		}
	}

	for(int i=0; i<num_devices; i++){
		CUDA_CALL( cudaSetDevice(i) );
		CUDA_CALL( cudaMalloc(&in_d[i], NN[i]*C*H*W*sizeof(float)) );
		CUDA_CALL( cudaMalloc(&out_d[i], NN[i]*K*OH*OW*sizeof(float)) );
		CUDA_CALL( cudaMalloc(&fil_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){
}