#include "convolution.h"
#include <mpi.h>
#include <stdio.h>
#include "util.h"

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 N_MPI;

static int p_n, remain;
MPI_Status status;
MPI_Request request;


static void convolution_omp() {
	#pragma omp parallel for collapse(3) num_threads(100) schedule(dynamic)
//	#pragma omp parallel for collapse(2) num_threads(40)
	for (int n = 0; n < p_n; ++n) {
		for (int k = 0; k < K; ++k) {
			for (int oh = 0; oh < OH; ++oh) {
				for (int ow = 0; ow < OW; ++ow) {
					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(
		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 == 0) {

		p_n = N/(N_MPI);
		remain = N%(N_MPI);

		for(int dest=1; dest< N_MPI;dest++){
			MPI_Isend(&p_n, 1, MPI_INT, dest, 1, MPI_COMM_WORLD, &request);
			MPI_Isend(&input[(p_n+remain)*C*H*W], p_n*C*H*W, MPI_FLOAT, dest, 1, MPI_COMM_WORLD, &request);
			MPI_Isend(filter, K*C*R*S, MPI_FLOAT, dest, 1, MPI_COMM_WORLD, &request);
		}

		p_n += remain; 
		convolution_omp();

		for(int source=1; source< N_MPI;source++){	
			MPI_Recv(&p_n, 1, MPI_INT, source, 2, MPI_COMM_WORLD, &status);
			MPI_Recv(&output[(p_n+remain)*K*OH*OW], p_n*K*OH*OW, MPI_FLOAT, source, 2, MPI_COMM_WORLD, &status);
		}

	}
	if (mpi_rank > 0){


		MPI_Recv(&p_n, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &status);	


		alloc_tensor(&input, p_n, C, H, W);
		alloc_tensor(&output, p_n, K, OH, OW);
		alloc_tensor(&filter, K, C, R, S);
		zero_tensor(output, p_n, K, OH, OW);


		MPI_Recv(input, p_n*C*H*W, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);
		MPI_Recv(filter, K*C*R*S, MPI_FLOAT,0, 1, MPI_COMM_WORLD, &status);


		convolution_omp();

		MPI_Isend(&p_n, 1, MPI_INT, 0, 2, MPI_COMM_WORLD, &request);
		MPI_Isend(output, p_n*K*OH*OW, MPI_FLOAT, 0, 2, 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);

	if(N < mpi_world_size) N_MPI = N;
	else N_MPI = 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) {
}