#include "SSProc.h"
#include "stdlib.h"
#include <cstdint>

void SSProcAndNF(
	hls::stream<SLRMatchOutput> 			&inMatchPt,
	hls::stream<LRMatchedPoint> 			&outLRPt,
	ap_uint<1> enable_SS_proc
)
{
#if !(VITIS_HLS)
#pragma HLS DATA_PACK variable=inMatchPt
#pragma HLS DATA_PACK variable=outLRPt
#else
#pragma HLS aggregate variable=inMatchPt compact=bit
#pragma HLS aggregate variable=outLRPt compact=bit
#endif
#pragma HLS INTERFACE axis register both port=inMatchPt
#pragma HLS INTERFACE axis register both port=outLRPt
#pragma HLS INTERFACE mode=ap_none port=enable_SS_proc register

	static SRowData left_buf[ROW_COUNT]; //size = 4k * 72 = 288kbit， RAM36 * 8
#pragma HLS BIND_STORAGE variable=left_buf type=ram_2p impl=bram
#pragma HLS RESET variable=left_buf off
	static SRowData right_buf[ROW_COUNT];
#pragma HLS BIND_STORAGE variable=right_buf type=ram_2p impl=bram
#pragma HLS RESET variable=right_buf off
	static SLineMatchInfo matchInfo_buf[ROW_COUNT];
#pragma HLS BIND_STORAGE variable=matchInfo_buf type=ram_2p impl=bram
#pragma HLS RESET variable=matchInfo_buf off

	// No per-frame clear: receive overwrites every row used by either pass.
	// Input protocol: dense rows, valid header bounds, matchNum/ptNum <= 2.
	SLRMatchOutput 	InData;
	ap_uint<2> VSync;
	while(1)  //wait for vSync signal
	{
#pragma HLS LOOP_TRIPCOUNT min=1 max=1
		InData = inMatchPt.read();
		VSync = InData.syncWord;
		if(0b10 == VSync)  //wait for syncFirst
			break;
	}

	uint32_t frmNo, timeStamp, encInfo;
	uint16_t minY, maxY;
	UintFloat inTmp;
	inTmp.fdata = InData.left_pt1.x;
	frmNo = inTmp.unData;
	inTmp.fdata = InData.left_pt2.x;
	timeStamp = inTmp.unData;
	inTmp.fdata = InData.right_pt1.x;
	encInfo = inTmp.unData;
	inTmp.fdata = InData.right_pt2.x;
	minY = inTmp.unData & 0xffff;
	maxY = inTmp.unData >> 16;

	const uint16_t header_max_y = maxY;
	const bool valid_range = minY <= maxY && maxY < ROW_COUNT;
	bool received_rows = false;
	uint16_t total_match_size = 0;
SS_RECEIVE:
	for(uint16_t n = minY; valid_range && n <= header_max_y; n ++)
	{
#pragma HLS LOOP_TRIPCOUNT min=1 max=3072
#pragma HLS PIPELINE II=1

		InData = inMatchPt.read();
		SRowData left_line, right_line;
		left_line.ptNum = InData.left_ptNum;
		left_line.pt1 = InData.left_pt1;
		left_line.pt2 = InData.left_pt2;
		right_line.ptNum = InData.right_ptNum;
		right_line.pt1 = InData.right_pt1;
		right_line.pt2 = InData.right_pt2;

		SLineMatchInfo line_match;
		line_match = InData.matchInfo;
		left_buf[n] = left_line;
		right_buf[n] = right_line;
		matchInfo_buf[n] = line_match;

		total_match_size += line_match.matchNum;
		received_rows = true;
		maxY = n; // Only scan rows received in this invocation.

		if(0b11 == InData.syncWord)
		{
			break;
		}
	}

	//step 4: single side processing, target II = 2 (top-down + bottom-up)
	//single side processing
	if(enable_SS_proc == 1 && received_rows)
	{
		SPointMatchInfo null_match;
		null_match.match_flag = 0;
		null_match.match_seg_id = 0;
		null_match.pt_idx_L = 0;
		null_match.pt_idx_R = 0;
		null_match.rid_L = 0;
		null_match.rid_R = 0;

		//top - down
		SPointMatchInfo match1 = null_match;
		SPointMatchInfo match2 = null_match;

SS_TOP_DOWN:
		for(uint16_t i = minY; i <= maxY; i ++)
		{
#pragma HLS LOOP_TRIPCOUNT min=1 max=3072
#pragma HLS PIPELINE II=2

			SRowData line_L = left_buf[i];
			SRowData line_R = right_buf[i];
			SLineMatchInfo a_lineMatch = matchInfo_buf[i];
			SLineMatchInfo update_match = a_lineMatch;
            ap_uint<2> flag_L1 = line_L.pt1.matchFlag;
            ap_uint<2> flag_L2 = line_L.pt2.matchFlag;
            ap_uint<2> flag_R1 = line_R.pt1.matchFlag;
            ap_uint<2> flag_R2 = line_R.pt2.matchFlag;

			SPointMatchInfo pre_m1 = match1;
			SPointMatchInfo pre_m2 = match2;

			SPointMatchInfo curr_m1 = a_lineMatch.match1;
			SPointMatchInfo curr_m2 = a_lineMatch.match2;

			ap_uint<2> m1_newMatch = 0;
			ap_uint<2> m2_newMatch = 0;
			SPointMatchInfo new_m1 = null_match;
			SPointMatchInfo new_m2 = null_match;

			//single side match
			if(pre_m1.match_flag > 0)
			{
				if( (curr_m1.match_flag > 0) && (pre_m1.rid_L == curr_m1.rid_L) && (pre_m1.rid_R == curr_m1.rid_R)) //m1 <-> left_pt1
				{
					new_m1 = curr_m1;
					m1_newMatch = 2;  //inherit
				}
				else if ( (curr_m1.match_flag > 0) && (pre_m1.rid_L == curr_m1.rid_L) && (pre_m1.rid_R != curr_m1.rid_R))
				{
					m1_newMatch = 0;
				}
				else if ( (curr_m1.match_flag > 0) && (pre_m1.rid_L != curr_m1.rid_L) && (pre_m1.rid_R == curr_m1.rid_R))
				{
					m1_newMatch = 0;
				}
				else if( (curr_m2.match_flag > 0) && (pre_m1.rid_L == curr_m2.rid_L) && (pre_m1.rid_R == curr_m2.rid_R)) //m1 <-> left_pt1
				{
					new_m1 = curr_m2;
					m1_newMatch = 2; //inherit
				}
				else if( (curr_m2.match_flag > 0) && (pre_m1.rid_L == curr_m2.rid_L) && (pre_m1.rid_R != curr_m2.rid_R)) //m1 <-> left_pt1
				{
					m1_newMatch = 0;
				}
				else if( (curr_m2.match_flag > 0) && (pre_m1.rid_L != curr_m2.rid_L) && (pre_m1.rid_R == curr_m2.rid_R)) //m1 <-> left_pt1
				{
					m1_newMatch = 0;
				}
				//start single side match check
				else if( (line_L.ptNum >= 1) && (pre_m1.rid_L != 0) && (pre_m1.rid_L == line_L.pt1.rid) && (line_L.pt1.matchFlag == 0))
				{
					m1_newMatch = 1; //new single side
					new_m1.match_flag = 2;
					new_m1.match_seg_id = pre_m1.match_seg_id;
					new_m1.pt_idx_L = 1;
					new_m1.pt_idx_R = 0;
					new_m1.rid_L = pre_m1.rid_L;
					new_m1.rid_R = pre_m1.rid_R;
				}
				else if( (line_L.ptNum >= 2) && (pre_m1.rid_L != 0) && (pre_m1.rid_L == line_L.pt2.rid) && (line_L.pt2.matchFlag == 0))
				{
					m1_newMatch = 1; //new single side
					new_m1.match_flag = 2;
					new_m1.match_seg_id = pre_m1.match_seg_id;
					new_m1.pt_idx_L = 2;
					new_m1.pt_idx_R = 0;
					new_m1.rid_L = pre_m1.rid_L;
					new_m1.rid_R = pre_m1.rid_R;
				}
				else if( (line_R.ptNum >= 1) && (pre_m1.rid_R != 0) && (pre_m1.rid_R == line_R.pt1.rid) && (line_R.pt1.matchFlag == 0))
				{
					m1_newMatch = 1; //new single side
					new_m1.match_flag = 3;
					new_m1.match_seg_id = pre_m1.match_seg_id;
					new_m1.pt_idx_L = 0;
					new_m1.pt_idx_R = 1;
					new_m1.rid_L = pre_m1.rid_L;
					new_m1.rid_R = pre_m1.rid_R;
				}
				else if( (line_R.ptNum >= 2) && (pre_m1.rid_R != 0) && (pre_m1.rid_R == line_R.pt2.rid) && (line_R.pt2.matchFlag == 0))
				{
					m1_newMatch = 1; //new single side
					new_m1.match_flag = 3;
					new_m1.match_seg_id = pre_m1.match_seg_id;
					new_m1.pt_idx_L = 0;
					new_m1.pt_idx_R = 2;
					new_m1.rid_L = pre_m1.rid_L;
					new_m1.rid_R = pre_m1.rid_R;
				}
			}

			//2nd
			if(pre_m2.match_flag > 0)
			{
				if( (curr_m1.match_flag > 0) && (pre_m2.rid_L == curr_m1.rid_L) && (pre_m2.rid_R == curr_m1.rid_R)) //m1 <-> left_pt1
				{
					new_m2 = curr_m1;
					m2_newMatch = 2;  //inherit
				}
				else if ( (curr_m1.match_flag > 0) && (pre_m2.rid_L == curr_m1.rid_L) && (pre_m2.rid_R != curr_m1.rid_R))
				{
					m2_newMatch = 0;
				}
				else if ( (curr_m1.match_flag > 0) && (pre_m2.rid_L != curr_m1.rid_L) && (pre_m2.rid_R == curr_m1.rid_R))
				{
					m2_newMatch = 0;
				}
				else if( (curr_m2.match_flag > 0) && (pre_m2.rid_L == curr_m2.rid_L) && (pre_m2.rid_R == curr_m2.rid_R)) //m1 <-> left_pt1
				{
					new_m2 = curr_m2;
					m2_newMatch = 2;  //inherit
				}
				else if((curr_m2.match_flag > 0) && (pre_m2.rid_L == curr_m2.rid_L) && (pre_m2.rid_R != curr_m2.rid_R)) //m1 <-> left_pt1
				{
					m2_newMatch = 0;
				}
				else if((curr_m2.match_flag > 0) && (pre_m2.rid_L != curr_m2.rid_L) && (pre_m2.rid_R == curr_m2.rid_R)) //m1 <-> left_pt1
				{
					m2_newMatch = 0;
				}
				//start single side match check
				else if( (line_L.ptNum >= 1) && (pre_m2.rid_L != 0) && (pre_m2.rid_L == line_L.pt1.rid) && (line_L.pt1.matchFlag == 0))
				{
					m2_newMatch = 1; //new single side
					new_m2.match_flag = 2;
					new_m2.match_seg_id = pre_m2.match_seg_id;
					new_m2.pt_idx_L = 1;
					new_m2.pt_idx_R = 0;
					new_m2.rid_L = pre_m2.rid_L;
					new_m2.rid_R = pre_m2.rid_R;
				}
				else if( (line_L.ptNum >= 2) && (pre_m2.rid_L != 0) && (pre_m2.rid_L == line_L.pt2.rid) && (line_L.pt2.matchFlag == 0))
				{
					m2_newMatch = 1; //new single side
					new_m2.match_flag = 2;
					new_m2.match_seg_id = pre_m2.match_seg_id;
					new_m2.pt_idx_L = 2;
					new_m2.pt_idx_R = 0;
					new_m2.rid_L = pre_m2.rid_L;
					new_m2.rid_R = pre_m2.rid_R;
				}
				else if( (line_R.ptNum >= 1) && (pre_m2.rid_R != 0) && (pre_m2.rid_R == line_R.pt1.rid) && (line_R.pt1.matchFlag == 0))
				{
					m2_newMatch = 1; //new single side
					new_m2.match_flag = 3;
					new_m2.match_seg_id = pre_m2.match_seg_id;
					new_m2.pt_idx_L = 0;
					new_m2.pt_idx_R = 1;
					new_m2.rid_L = pre_m2.rid_L;
					new_m2.rid_R = pre_m2.rid_R;
				}
				else if( (line_R.ptNum >= 2) && (pre_m2.rid_R != 0) && (pre_m2.rid_R == line_R.pt2.rid) && (line_R.pt2.matchFlag == 0))
				{
					m2_newMatch = 1; //new single side
					new_m2.match_flag = 3;
					new_m2.match_seg_id = pre_m2.match_seg_id;
					new_m2.pt_idx_L = 0;
					new_m2.pt_idx_R = 2;
					new_m2.rid_L = pre_m2.rid_L;
					new_m2.rid_R = pre_m2.rid_R;
				}
			}

            match1 = (m1_newMatch != 0) ? new_m1 : null_match;
            match2 = (m2_newMatch != 0) ? new_m2 : null_match;
			//update LR match for next using
			if( (m1_newMatch == 0) && (m2_newMatch == 0))//m1 and m2 have no match
			{
				if( (curr_m1.match_flag == 1) && (curr_m2.match_flag == 1))
				{
					//a_lineMatch.match1 and a_lineMatch.match2 are new L-R match
					match1 = curr_m1;
					match2 = curr_m2;
				}
				else if (curr_m1.match_flag == 1)
				{
					//a_lineMatch.match1 is a new L-R match
					match1 = curr_m1;
					match2 = null_match;
				}
				else if(curr_m2.match_flag == 1)
				{
					//a_lineMatch.match2 is a new L-R match
					match1 = curr_m2;
					match2 = null_match;
				}
				else
				{
					match1 = null_match;
					match2 = null_match;
				}
			}
			else if((m1_newMatch == 0) && (m2_newMatch > 0)) //m1 has no match, but m2 is matched
			{
				if((curr_m1.match_flag == 1) && ((new_m2.rid_L != curr_m1.rid_L)||(new_m2.rid_R != curr_m1.rid_R)))
				{
					//a_lineMatch.match1 is a new L-R match
					match1 = curr_m1;
				}
				else if((curr_m2.match_flag == 1) && ((new_m2.rid_L != curr_m2.rid_L)||(new_m2.rid_R != curr_m2.rid_R)))
				{
					//a_lineMatch.match2 is a new L-R match
					match1 = curr_m2;
				}
				else
				{
					//only one L-R match, move m2 to m1
					match1 = match2;
					match2 = null_match;
				}
			}
			else if((m1_newMatch > 0) && (m2_newMatch == 0))//m2 has no match, but m1 is matched
			{
				if((curr_m1.match_flag == 1) && ((new_m1.rid_L != curr_m1.rid_L)||(new_m1.rid_R != curr_m1.rid_R)))
				{
					//a_lineMatch.match1 is a new L-R match
					match2 = curr_m1;
				}
				else if((curr_m2.match_flag == 1) && ((new_m1.rid_L != curr_m2.rid_L)||(new_m1.rid_R != curr_m2.rid_R)))
				{
					//a_lineMatch.match2 is a new L-R match
					match2 = curr_m2;
				}
				else
				{
					match2 = null_match;
				}
			}

			//update current new match. Parallel compare
			if(a_lineMatch.matchNum == 0)
			{
				if( (m1_newMatch == 1) && (m2_newMatch == 1))
				{
					update_match.matchNum = 2;
					update_match.match1 = new_m1;
					update_match.match2 = new_m2;

					if(new_m1.pt_idx_L == 1)
						flag_L1 = 2;//new_m1.match_flag;
					else if(new_m1.pt_idx_L == 2)
						flag_L2 = 2;//new_m1.match_flag;
					else if(new_m1.pt_idx_R == 1)
						flag_R1 = 3;//new_m1.match_flag;
					else if(new_m1.pt_idx_R == 2)
						flag_R2 = 3;//new_m1.match_flag;

					if(new_m2.pt_idx_L == 1)
						flag_L1 = 2;//new_m2.match_flag;
					else if(new_m2.pt_idx_L == 2)
						flag_L2 = 2;//new_m2.match_flag;
					else if(new_m2.pt_idx_R == 1)
						flag_R1 = 3;//new_m2.match_flag;
					else if(new_m2.pt_idx_R == 2)
						flag_R2 = 3;//new_m2.match_flag;
				}
				else if(m1_newMatch == 1)
				{
					update_match.matchNum = 1;
					update_match.match1 = new_m1;
					update_match.match2 = null_match;
					if(new_m1.pt_idx_L == 1)
						flag_L1 = 2;//new_m1.match_flag;
					else if(new_m1.pt_idx_L == 2)
						flag_L2 = 2;//new_m1.match_flag;
					else if(new_m1.pt_idx_R == 1)
						flag_R1 = 3;//new_m1.match_flag;
					else if(new_m1.pt_idx_R == 2)
						flag_R2 = 3;//new_m1.match_flag;
				}
				else if(m2_newMatch == 1)
				{
					update_match.matchNum = 1;
					update_match.match1 = new_m2;
					update_match.match2 = null_match;
					if(new_m2.pt_idx_L == 1)
						flag_L1 = 2;//new_m2.match_flag;
					else if(new_m2.pt_idx_L == 2)
						flag_L2 = 2;//new_m2.match_flag;
					else if(new_m2.pt_idx_R == 1)
						flag_R1 = 3;//new_m2.match_flag;
					else if(new_m2.pt_idx_R == 2)
						flag_R2 = 3;//new_m2.match_flag;
				}
			}
			else if(a_lineMatch.matchNum == 1)
			{
				if(m1_newMatch == 1)
				{
					update_match.matchNum = 2;
					update_match.match2 = new_m1;
					if(new_m1.pt_idx_L == 1)
						flag_L1 = 2;//new_m1.match_flag;
					else if(new_m1.pt_idx_L == 2)
						flag_L2 = 2;//new_m1.match_flag;
					else if(new_m1.pt_idx_R == 1)
						flag_R1 = 3;//new_m1.match_flag;
					else if(new_m1.pt_idx_R == 2)
						flag_R2 = 3;//new_m1.match_flag;
				}
				else if(m2_newMatch == 1)
				{
					update_match.matchNum = 2;
					update_match.match2 = new_m2;
					if(new_m2.pt_idx_L == 1)
						flag_L1 = 2;//new_m2.match_flag;
					else if(new_m2.pt_idx_L == 2)
						flag_L2 = 2;//new_m2.match_flag;
					else if(new_m2.pt_idx_R == 1)
						flag_R1 = 3;//new_m2.match_flag;
					else if(new_m2.pt_idx_R == 2)
						flag_R2 = 3;//new_m2.match_flag;
				}
			}
            // One write per flag field per row; all branch selection is local.
            left_buf[i].pt1.matchFlag = flag_L1;
            left_buf[i].pt2.matchFlag = flag_L2;
            right_buf[i].pt1.matchFlag = flag_R1;
            right_buf[i].pt2.matchFlag = flag_R2;
            total_match_size += (unsigned int)update_match.matchNum
                              - (unsigned int)a_lineMatch.matchNum;
			matchInfo_buf[i] = update_match;  //read - modify - write operation
		}

		// bottoom - up
		match1 = null_match;
		match2 = null_match;
SS_BOTTOM_UP:
		for(int i = maxY; i >= (int)minY; i --)
		{
#pragma HLS LOOP_TRIPCOUNT min=1 max=3072
#pragma HLS PIPELINE II=2

			SRowData line_L = left_buf[i];
			SRowData line_R = right_buf[i];
			SLineMatchInfo a_lineMatch = matchInfo_buf[i];
			SLineMatchInfo update_match = a_lineMatch;
            ap_uint<2> flag_L1 = line_L.pt1.matchFlag;
            ap_uint<2> flag_L2 = line_L.pt2.matchFlag;
            ap_uint<2> flag_R1 = line_R.pt1.matchFlag;
            ap_uint<2> flag_R2 = line_R.pt2.matchFlag;

			SPointMatchInfo pre_m1 = match1;
			SPointMatchInfo pre_m2 = match2;

			SPointMatchInfo curr_m1 = a_lineMatch.match1;
			SPointMatchInfo curr_m2 = a_lineMatch.match2;

			ap_uint<2> m1_newMatch = 0;
			ap_uint<2> m2_newMatch = 0;
			SPointMatchInfo new_m1 = null_match;
			SPointMatchInfo new_m2 = null_match;

			//single side match
			if(pre_m1.match_flag > 0)
			{
				if( (curr_m1.match_flag > 0) && (pre_m1.rid_L == curr_m1.rid_L) && (pre_m1.rid_R == curr_m1.rid_R)) //m1 <-> left_pt1
				{
					new_m1 = curr_m1;
					m1_newMatch = 2;  //inherit
				}
				else if ( (curr_m1.match_flag > 0) && (pre_m1.rid_L == curr_m1.rid_L) && (pre_m1.rid_R != curr_m1.rid_R))
				{
					m1_newMatch = 0;
				}
				else if ( (curr_m1.match_flag > 0) && (pre_m1.rid_L != curr_m1.rid_L) && (pre_m1.rid_R == curr_m1.rid_R))
				{
					m1_newMatch = 0;
				}
				else if( (curr_m2.match_flag > 0) && (pre_m1.rid_L == curr_m2.rid_L) && (pre_m1.rid_R == curr_m2.rid_R)) //m1 <-> left_pt1
				{
					new_m1 = curr_m2;
					m1_newMatch = 2; //inherit
				}
				else if( (curr_m2.match_flag > 0) && (pre_m1.rid_L == curr_m2.rid_L) && (pre_m1.rid_R != curr_m2.rid_R)) //m1 <-> left_pt1
				{
					m1_newMatch = 0;
				}
				else if( (curr_m2.match_flag > 0) && (pre_m1.rid_L != curr_m2.rid_L) && (pre_m1.rid_R == curr_m2.rid_R)) //m1 <-> left_pt1
				{
					m1_newMatch = 0;
				}
				//start single side match check
				else if( (line_L.ptNum >= 1) && (pre_m1.rid_L != 0) && (pre_m1.rid_L == line_L.pt1.rid) && (line_L.pt1.matchFlag == 0))
				{
					m1_newMatch = 1; //new single side
					new_m1.match_flag = 2;
					new_m1.match_seg_id = pre_m1.match_seg_id;
					new_m1.pt_idx_L = 1;
					new_m1.pt_idx_R = 0;
					new_m1.rid_L = pre_m1.rid_L;
					new_m1.rid_R = pre_m1.rid_R;
				}
				else if( (line_L.ptNum >= 2) && (pre_m1.rid_L != 0) && (pre_m1.rid_L == line_L.pt2.rid) && (line_L.pt2.matchFlag == 0))
				{
					m1_newMatch = 1; //new single side
					new_m1.match_flag = 2;
					new_m1.match_seg_id = pre_m1.match_seg_id;
					new_m1.pt_idx_L = 2;
					new_m1.pt_idx_R = 0;
					new_m1.rid_L = pre_m1.rid_L;
					new_m1.rid_R = pre_m1.rid_R;
				}
				else if( (line_R.ptNum >= 1) && (pre_m1.rid_R != 0) && (pre_m1.rid_R == line_R.pt1.rid) && (line_R.pt1.matchFlag == 0))
				{
					m1_newMatch = 1; //new single side
					new_m1.match_flag = 3;
					new_m1.match_seg_id = pre_m1.match_seg_id;
					new_m1.pt_idx_L = 0;
					new_m1.pt_idx_R = 1;
					new_m1.rid_L = pre_m1.rid_L;
					new_m1.rid_R = pre_m1.rid_R;
				}
				else if( (line_R.ptNum >= 2) && (pre_m1.rid_R != 0) && (pre_m1.rid_R == line_R.pt2.rid) && (line_R.pt2.matchFlag == 0))
				{
					m1_newMatch = 1; //new single side
					new_m1.match_flag = 3;
					new_m1.match_seg_id = pre_m1.match_seg_id;
					new_m1.pt_idx_L = 0;
					new_m1.pt_idx_R = 2;
					new_m1.rid_L = pre_m1.rid_L;
					new_m1.rid_R = pre_m1.rid_R;
				}
			}

			//2nd
			if(pre_m2.match_flag > 0)
			{
				if( (curr_m1.match_flag > 0) && (pre_m2.rid_L == curr_m1.rid_L) && (pre_m2.rid_R == curr_m1.rid_R)) //m1 <-> left_pt1
				{
					new_m2 = curr_m1;
					m2_newMatch = 2;  //inherit
				}
				else if ( (curr_m1.match_flag > 0) && (pre_m2.rid_L == curr_m1.rid_L) && (pre_m2.rid_R != curr_m1.rid_R))
				{
					m2_newMatch = 0;
				}
				else if ( (curr_m1.match_flag > 0) && (pre_m2.rid_L != curr_m1.rid_L) && (pre_m2.rid_R == curr_m1.rid_R))
				{
					m2_newMatch = 0;
				}
				else if( (curr_m2.match_flag > 0) && (pre_m2.rid_L == curr_m2.rid_L) && (pre_m2.rid_R == curr_m2.rid_R)) //m1 <-> left_pt1
				{
					new_m2 = curr_m2;
					m2_newMatch = 2;  //inherit
				}
				else if((curr_m2.match_flag > 0) && (pre_m2.rid_L == curr_m2.rid_L) && (pre_m2.rid_R != curr_m2.rid_R)) //m1 <-> left_pt1
				{
					m2_newMatch = 0;
				}
				else if((curr_m2.match_flag > 0) && (pre_m2.rid_L != curr_m2.rid_L) && (pre_m2.rid_R == curr_m2.rid_R)) //m1 <-> left_pt1
				{
					m2_newMatch = 0;
				}
				//start single side match check
				else if( (line_L.ptNum >= 1) && (pre_m2.rid_L != 0) && (pre_m2.rid_L == line_L.pt1.rid) && (line_L.pt1.matchFlag == 0))
				{
					m2_newMatch = 1; //new single side
					new_m2.match_flag = 2;
					new_m2.match_seg_id = pre_m2.match_seg_id;
					new_m2.pt_idx_L = 1;
					new_m2.pt_idx_R = 0;
					new_m2.rid_L = pre_m2.rid_L;
					new_m2.rid_R = pre_m2.rid_R;
				}
				else if( (line_L.ptNum >= 2) && (pre_m2.rid_L != 0) && (pre_m2.rid_L == line_L.pt2.rid) && (line_L.pt2.matchFlag == 0))
				{
					m2_newMatch = 1; //new single side
					new_m2.match_flag = 2;
					new_m2.match_seg_id = pre_m2.match_seg_id;
					new_m2.pt_idx_L = 2;
					new_m2.pt_idx_R = 0;
					new_m2.rid_L = pre_m2.rid_L;
					new_m2.rid_R = pre_m2.rid_R;
				}
				else if( (line_R.ptNum >= 1) && (pre_m2.rid_R != 0) && (pre_m2.rid_R == line_R.pt1.rid) && (line_R.pt1.matchFlag == 0))
				{
					m2_newMatch = 1; //new single side
					new_m2.match_flag = 3;
					new_m2.match_seg_id = pre_m2.match_seg_id;
					new_m2.pt_idx_L = 0;
					new_m2.pt_idx_R = 1;
					new_m2.rid_L = pre_m2.rid_L;
					new_m2.rid_R = pre_m2.rid_R;
				}
				else if( (line_R.ptNum >= 2) && (pre_m2.rid_R != 0) && (pre_m2.rid_R == line_R.pt2.rid) && (line_R.pt2.matchFlag == 0))
				{
					m2_newMatch = 1; //new single side
					new_m2.match_flag = 3;
					new_m2.match_seg_id = pre_m2.match_seg_id;
					new_m2.pt_idx_L = 0;
					new_m2.pt_idx_R = 2;
					new_m2.rid_L = pre_m2.rid_L;
					new_m2.rid_R = pre_m2.rid_R;
				}
			}

            match1 = (m1_newMatch != 0) ? new_m1 : null_match;
            match2 = (m2_newMatch != 0) ? new_m2 : null_match;
			//update LR match for next using
			if( (m1_newMatch == 0) && (m2_newMatch == 0))//m1 and m2 have no match
			{
				if( (curr_m1.match_flag == 1) && (curr_m2.match_flag == 1))
				{
					//a_lineMatch.match1 and a_lineMatch.match2 are new L-R match
					match1 = curr_m1;
					match2 = curr_m2;
				}
				else if (curr_m1.match_flag == 1)
				{
					//a_lineMatch.match1 is a new L-R match
					match1 = curr_m1;
					match2 = null_match;
				}
				else if(curr_m2.match_flag == 1)
				{
					//a_lineMatch.match2 is a new L-R match
					match1 = curr_m2;
					match2 = null_match;
				}
				else
				{
					match1 = null_match;
					match2 = null_match;
				}
			}
			else if((m1_newMatch == 0) && (m2_newMatch > 0)) //m1 has no match, but m2 is matched
			{
				if((curr_m1.match_flag == 1) && ((new_m2.rid_L != curr_m1.rid_L)||(new_m2.rid_R != curr_m1.rid_R)))
				{
					//a_lineMatch.match1 is a new L-R match
					match1 = curr_m1;
				}
				else if((curr_m2.match_flag == 1) && ((new_m2.rid_L != curr_m2.rid_L)||(new_m2.rid_R != curr_m2.rid_R)))
				{
					//a_lineMatch.match2 is a new L-R match
					match1 = curr_m2;
				}
				else
				{
					//only one L-R match, move m2 to m1
					match1 = match2;
					match2 = null_match;
				}
			}
			else if((m1_newMatch > 0) && (m2_newMatch == 0))//m2 has no match, but m1 is matched
			{
				if((curr_m1.match_flag == 1) && ((new_m1.rid_L != curr_m1.rid_L)||(new_m1.rid_R != curr_m1.rid_R)))
				{
					//a_lineMatch.match1 is a new L-R match
					match2 = curr_m1;
				}
				else if((curr_m2.match_flag == 1) && ((new_m1.rid_L != curr_m2.rid_L)||(new_m1.rid_R != curr_m2.rid_R)))
				{
					//a_lineMatch.match2 is a new L-R match
					match2 = curr_m2;
				}
				else
				{
					match2 = null_match;
				}
			}

			//update current new match. Parallel compare
			if(a_lineMatch.matchNum == 0)
			{
				if( (m1_newMatch == 1) && (m2_newMatch == 1))
				{
					update_match.matchNum = 2;
					update_match.match1 = new_m1;
					update_match.match2 = new_m2;

					if(new_m1.pt_idx_L == 1)
						flag_L1 = 2;//new_m1.match_flag;
					else if(new_m1.pt_idx_L == 2)
						flag_L2 = 2;//new_m1.match_flag;
					else if(new_m1.pt_idx_R == 1)
						flag_R1 = 3;//new_m1.match_flag;
					else if(new_m1.pt_idx_R == 2)
						flag_R2 = 3;//new_m1.match_flag;

					if(new_m2.pt_idx_L == 1)
						flag_L1 = 2;//new_m2.match_flag;
					else if(new_m2.pt_idx_L == 2)
						flag_L2 = 2;//new_m2.match_flag;
					else if(new_m2.pt_idx_R == 1)
						flag_R1 = 3;//new_m2.match_flag;
					else if(new_m2.pt_idx_R == 2)
						flag_R2 = 3;//new_m2.match_flag;
				}
				else if(m1_newMatch == 1)
				{
					update_match.matchNum = 1;
					update_match.match1 = new_m1;
					update_match.match2 = null_match;
					if(new_m1.pt_idx_L == 1)
						flag_L1 = 2;//new_m1.match_flag;
					else if(new_m1.pt_idx_L == 2)
						flag_L2 = 2;//new_m1.match_flag;
					else if(new_m1.pt_idx_R == 1)
						flag_R1 = 3;//new_m1.match_flag;
					else if(new_m1.pt_idx_R == 2)
						flag_R2 = 3;//new_m1.match_flag;
				}
				else if(m2_newMatch == 1)
				{
					update_match.matchNum = 1;
					update_match.match1 = new_m2;
					update_match.match2 = null_match;
					if(new_m2.pt_idx_L == 1)
						flag_L1 = 2;//new_m2.match_flag;
					else if(new_m2.pt_idx_L == 2)
						flag_L2 = 2;//new_m2.match_flag;
					else if(new_m2.pt_idx_R == 1)
						flag_R1 = 3;//new_m2.match_flag;
					else if(new_m2.pt_idx_R == 2)
						flag_R2 = 3;//new_m2.match_flag;
				}
			}
			else if(a_lineMatch.matchNum == 1)
			{
				if(m1_newMatch == 1)
				{
					update_match.matchNum = 2;
					update_match.match2 = new_m1;
					if(new_m1.pt_idx_L == 1)
						flag_L1 = 2;//new_m1.match_flag;
					else if(new_m1.pt_idx_L == 2)
						flag_L2 = 2;//new_m1.match_flag;
					else if(new_m1.pt_idx_R == 1)
						flag_R1 = 3;//new_m1.match_flag;
					else if(new_m1.pt_idx_R == 2)
						flag_R2 = 3;//new_m1.match_flag;
				}
				else if(m2_newMatch == 1)
				{
					update_match.matchNum = 2;
					update_match.match2 = new_m2;
					if(new_m2.pt_idx_L == 1)
						flag_L1 = 2;//new_m2.match_flag;
					else if(new_m2.pt_idx_L == 2)
						flag_L2 = 2;//new_m2.match_flag;
					else if(new_m2.pt_idx_R == 1)
						flag_R1 = 3;//new_m2.match_flag;
					else if(new_m2.pt_idx_R == 2)
						flag_R2 = 3;//new_m2.match_flag;
				}
			}
            // One write per flag field per row; all branch selection is local.
            left_buf[i].pt1.matchFlag = flag_L1;
            left_buf[i].pt2.matchFlag = flag_L2;
            right_buf[i].pt1.matchFlag = flag_R1;
            right_buf[i].pt2.matchFlag = flag_R2;
            total_match_size += (unsigned int)update_match.matchNum
                              - (unsigned int)a_lineMatch.matchNum;
			matchInfo_buf[i] = update_match;  //read - modify - write operation
		}
	}

	LRMatchedPoint outWord = { 0.0f, 0.0f, 0, 0, 0, 0, 0, 0	};
	UintFloat outTmp;
	//first
	outTmp.unData = frmNo;
	outWord.xL = outTmp.fdata;
	outTmp.unData = timeStamp;
	outWord.xR = outTmp.fdata;
	outWord.Sync = 0b10;
	outLRPt.write(outWord);
	//second
	outTmp.unData = encInfo;
	outWord.xL = outTmp.fdata;
	outTmp.unData = (uint32_t)minY | ((uint32_t)maxY << 16);
	outWord.xR = outTmp.fdata;
	outWord.Sync = 0b00;
	outLRPt.write(outWord);
	//3rd
	outTmp.unData = total_match_size;
	outWord.xL = outTmp.fdata;
	outWord.xR = 0.0f;
	if(total_match_size == 0)
		outWord.Sync = 0b11;
	else
		outWord.Sync = 0b00;
	outLRPt.write(outWord);

	if(total_match_size > 0)
	{
		//output
		uint16_t counter = 0;
SS_OUTPUT:
		for(uint16_t i = minY; i<= maxY; i ++)
		{
#pragma HLS LOOP_TRIPCOUNT min=1 max=3072
#pragma HLS PIPELINE II=2
			SLineMatchInfo a_lineMatch = matchInfo_buf[i];
			SRowData line_L = left_buf[i];
			SRowData line_R = right_buf[i];
			if(a_lineMatch.matchNum > 0)
			{
				//first
				outWord.matchFlag = a_lineMatch.match1.match_flag;
				outWord.match_seg_id = a_lineMatch.match1.match_seg_id;
				outWord.y = i;
				if(a_lineMatch.match1.pt_idx_L == 1)
				{
					outWord.xL = line_L.pt1.x;
					outWord.valueL = line_L.pt1.value;
				}
				else if(a_lineMatch.match1.pt_idx_L == 2)
				{
					outWord.xL = line_L.pt2.x;
					outWord.valueL = line_L.pt2.value;
				}
				else
				{
					outWord.xL = 0.0f;
					outWord.valueL = 0;
				}

				//right
				if(a_lineMatch.match1.pt_idx_R == 1)
				{
					outWord.xR = line_R.pt1.x;
					outWord.valueR = line_R.pt1.value;
				}
				else if(a_lineMatch.match1.pt_idx_R == 2)
				{
					outWord.xR = line_R.pt2.x;
					outWord.valueR = line_R.pt2.value;
				}
				else
				{
					outWord.xR = 0.0f;
					outWord.valueR = 0;
				}
				counter ++;
				if(counter == total_match_size)
					outWord.Sync = 0b11;
				else
					outWord.Sync = 0b00;
				//output
				outLRPt.write(outWord);
			}

			if(a_lineMatch.matchNum == 2)
			{
				//second
				outWord.matchFlag = a_lineMatch.match2.match_flag;
				outWord.match_seg_id = a_lineMatch.match2.match_seg_id;
				outWord.y = i;
				if(a_lineMatch.match2.pt_idx_L == 1)
				{
					outWord.xL = line_L.pt1.x;
					outWord.valueL = line_L.pt1.value;
				}
				else if(a_lineMatch.match2.pt_idx_L == 2)
				{
					outWord.xL = line_L.pt2.x;
					outWord.valueL = line_L.pt2.value;
				}
				else
				{
					outWord.xL = 0.0f;
					outWord.valueL = 0;
				}

				//right
				if(a_lineMatch.match2.pt_idx_R == 1)
				{
					outWord.xR = line_R.pt1.x;
					outWord.valueR = line_R.pt1.value;
				}
				else if(a_lineMatch.match2.pt_idx_R == 2)
				{
					outWord.xR = line_R.pt2.x;
					outWord.valueR = line_R.pt2.value;
				}
				else
				{
					outWord.xR = 0.0f;
					outWord.valueR = 0;
				}
				counter ++;
				if(counter == total_match_size)
					outWord.Sync = 0b11;
				else
					outWord.Sync = 0b00;
				//output
				outLRPt.write(outWord);
			}
		}
	}
}