inter_pred_pipeline.v

a H.264/AVC Baseline Decoder

			end
		else
			yOffsetL <= 0;
	
	reg [2:0] xOffsetC,yOffsetC;
	always @ (IsInterChroma or mv_below8x8_curr or blk4x4_rec_counter[0] or Inter_chroma2x2_counter[0])
		if (IsInterChroma)
			begin
				if (mv_below8x8_curr == 1'b0)
					xOffsetC <= (blk4x4_rec_counter[0] == 1'b0)? 3'd0:3'd4;
				else 
					case (blk4x4_rec_counter[0])
						1'b0:xOffsetC <= (Inter_chroma2x2_counter[0] == 1'b1)? 3'd0:3'd2;
						1'b1:xOffsetC <= (Inter_chroma2x2_counter[0] == 1'b1)? 3'd4:3'd6;
					endcase
			end
		else
			xOffsetC <= 0; 
			
	always @ (IsInterChroma or mv_below8x8_curr or blk4x4_rec_counter[1] or Inter_chroma2x2_counter[1])
		if (IsInterChroma)
			begin
				if (mv_below8x8_curr == 1'b0)
					yOffsetC <= (blk4x4_rec_counter[1] == 1'b0)? 3'd0:3'd4;
				else 
					case (blk4x4_rec_counter[1])
						1'b0:yOffsetC <= (Inter_chroma2x2_counter[1] == 1'b1)? 3'd0:3'd2;
						1'b1:yOffsetC <= (Inter_chroma2x2_counter[1] == 1'b1)? 3'd4:3'd6;
					endcase
			end
		else
			yOffsetC <= 3'd0;
	//----------------------------------------------------------------------------------------
	//Integer position of each left-up-most pixel  of a 8x8/4x4/2x2 blk
	//----------------------------------------------------------------------------------------		
	wire [8:0] xIntL_unclip,yIntL_unclip;	// 2's complement,bit[8] is the sign bit
	wire [7:0] xIntC_unclip,yIntC_unclip;	// 2's complement,bit[7] is the sign bit
	assign xIntL_unclip = (IsInterLuma)?   ({1'b0,mb_num_h,4'b0} + xOffsetL + {{3{Inter_blk_mvx[7]}},Inter_blk_mvx[7:2]}):0;
	assign yIntL_unclip = (IsInterLuma)?   ({1'b0,mb_num_v,4'b0} + yOffsetL + {{3{Inter_blk_mvy[7]}},Inter_blk_mvy[7:2]}):0;
	assign xIntC_unclip = (IsInterChroma)? ({1'b0,mb_num_h,3'b0} + xOffsetC + {{3{Inter_blk_mvx[7]}},Inter_blk_mvx[7:3]}):0;
	assign yIntC_unclip = (IsInterChroma)? ({1'b0,mb_num_v,3'b0} + yOffsetC + {{3{Inter_blk_mvy[7]}},Inter_blk_mvy[7:3]}):0; 
	
	wire [8:0] xInt_org_unclip;
	wire [8:0] yInt_org_unclip;
	assign xInt_org_unclip = (IsInterLuma)? xIntL_unclip:{xIntC_unclip[7],xIntC_unclip};
	assign yInt_org_unclip = (IsInterLuma)? yIntL_unclip:{yIntC_unclip[7],yIntC_unclip};
	assign xInt_org_unclip_1to0 = xInt_org_unclip[1:0];
	//----------------------------------------------------------------------------------------
	//Fractional motion vector for both luma and chroma
	//----------------------------------------------------------------------------------------		
	wire [3:0] pos_FracL;
	wire Is_InterChromaCopy;//If chroma is predicted by direct copy,calculate cycle would reduce
							//from 16 cycles to 4 cycles
	
	assign xFracL = (IsInterLuma)?   Inter_blk_mvx[1:0]:0;
	assign yFracL = (IsInterLuma)?   Inter_blk_mvy[1:0]:0;
	assign xFracC = (IsInterChroma)? Inter_blk_mvx[2:0]:0;
	assign yFracC = (IsInterChroma)? Inter_blk_mvy[2:0]:0;
	assign pos_FracL = {xFracL,yFracL};
	assign Is_InterChromaCopy = (IsInterChroma && xFracC == 0 && yFracC == 0)? 1'b1:1'b0;
	
	//----------------------------------------------------------------------------------------
	//Inter prediction step control counter
	//---------------------------------------------------------------------------------------- 
	//1.Preload integer pels counter
	//	If block partition equals 8x8 or above,preload only at first 4x4 block of each 8x8block
	//  If block partition is 8x4,4x8 or 4x4,  preload at each 4x4 block
	always @ (posedge clk)
		if (reset_n == 1'b0)
			blk4x4_inter_preload_counter <= 0;
		//luma
		else if (trigger_blk4x4_inter_pred && IsInterLuma)
			begin
				if (!mv_below8x8_curr && blk4x4_rec_counter[1:0] == 2'b00)
					case (pos_FracL)
						`pos_Int                          :blk4x4_inter_preload_counter <= (xInt_org_unclip[1:0] == 2'b00)? 6'd17:6'd25;
						`pos_f,`pos_q,`pos_i,`pos_k,`pos_j:blk4x4_inter_preload_counter <= 6'd53;
						`pos_d,`pos_h,`pos_n              :blk4x4_inter_preload_counter <= (xInt_org_unclip[1:0] == 2'b00)? 6'd27:6'd40;
						`pos_a,`pos_b,`pos_c              :blk4x4_inter_preload_counter <= 6'd33;
						`pos_e,`pos_g,`pos_p,`pos_r       :blk4x4_inter_preload_counter <= 6'd49;
					endcase
				else if (mv_below8x8_curr)	//partition below 8x8block
					case (pos_FracL)
						`pos_Int						              :blk4x4_inter_preload_counter <= (xInt_org_unclip[1:0] == 2'b00)? 6'd5:6'd9;
						`pos_f,`pos_q,`pos_i,`pos_k,`pos_j:blk4x4_inter_preload_counter <= 6'd28;
						`pos_d,`pos_h,`pos_n			        :blk4x4_inter_preload_counter <= (xInt_org_unclip[1:0] == 2'b00)? 6'd10:6'd19;
						`pos_a,`pos_b,`pos_c			        :blk4x4_inter_preload_counter <= 6'd13;
						`pos_e,`pos_g,`pos_p,`pos_r		    :blk4x4_inter_preload_counter <= 6'd24;
					endcase	
			end
		//chroma
		else if (trigger_blk4x4_inter_pred && IsInterChroma && mv_below8x8_curr == 1'b0)
			begin
				if (xFracC == 0 && yFracC == 0)
					blk4x4_inter_preload_counter <= (xInt_org_unclip[1:0] == 2'b00)? 6'd5:6'd9;
				else
					blk4x4_inter_preload_counter <= 6'd11;
			end
		else if (trigger_blk2x2_inter_pred && IsInterChroma && mv_below8x8_curr == 1'b1)
			begin
				if (xFracC == 0 && yFracC == 0)
					blk4x4_inter_preload_counter <= (xInt_org_unclip[1:0] == 2'b11)? 6'd5:6'd3;
				else
					blk4x4_inter_preload_counter <= (xInt_org_unclip[1]   == 1'b0 )? 6'd4:6'd7;
			end
		else if (blk4x4_inter_preload_counter != 0)
			blk4x4_inter_preload_counter <= blk4x4_inter_preload_counter - 1;
		
	//2.Calculate counter
	always @ (posedge clk)
		if (reset_n == 1'b0)
			blk4x4_inter_calculate_counter <= 0;
		//luma
		else if (IsInterLuma && ((!mv_below8x8_curr && (
										(blk4x4_rec_counter[1:0] == 2'b00 && blk4x4_inter_preload_counter == 1) || 
										(blk4x4_rec_counter[1:0] != 2'b00 && trigger_blk4x4_inter_pred))) ||
								(mv_below8x8_curr && blk4x4_inter_preload_counter == 1))) 
			case (pos_FracL)
				`pos_j,`pos_f,`pos_q:blk4x4_inter_calculate_counter <= 4'd5;
				`pos_i,`pos_k       :blk4x4_inter_calculate_counter <= 4'd8;
				default             :blk4x4_inter_calculate_counter <= 4'd4;
			endcase
		//chroma
		else if (blk4x4_inter_preload_counter == 1 && IsInterChroma == 1'b1)
			case (mv_below8x8_curr)
				1'b0:blk4x4_inter_calculate_counter <= 4'd4;
				1'b1:blk4x4_inter_calculate_counter <= 4'd1;
			endcase
		else if (blk4x4_inter_calculate_counter != 0)
			blk4x4_inter_calculate_counter <= blk4x4_inter_calculate_counter - 1;
	
	assign end_of_one_blk4x4_inter = (blk4x4_inter_calculate_counter == 4'd1 &&
	((IsInterChroma && mv_below8x8_curr && Inter_chroma2x2_counter == 2'b00) ||
	!(IsInterChroma && mv_below8x8_curr)));
	//----------------------------------------------------------------------------------------
	//Inter prediction reference frame RAM read control
	//----------------------------------------------------------------------------------------
	assign ref_frame_RAM_rd = ((IsInterLuma || IsInterChroma) && blk4x4_inter_preload_counter != 6'd0 && blk4x4_inter_preload_counter != 6'd1);
	
	//compared with blk4x4_inter_preload_counter,blk4x4_inter_preload_counter_m2 has some advantages
	//during some pos_FracL for vertical memory address decoding
	wire [5:0] blk4x4_inter_preload_counter_m2;	
	assign blk4x4_inter_preload_counter_m2 = (blk4x4_inter_preload_counter == 6'd0 || blk4x4_inter_preload_counter == 6'd1)?
												6'd0:(blk4x4_inter_preload_counter - 2);
				
	//xInt_curr_offset: offset from the left-upper most pixel of current block,ranging -2 ~ +10.
	//After each preload cycle,xInt_curr_offset will increase 4
	reg [4:0] xInt_curr_offset;
	always @ (IsInterLuma or mv_below8x8_curr or pos_FracL or xFracC or yFracC 
		or xInt_org_unclip[1:0] or blk4x4_inter_preload_counter_m2 or blk4x4_inter_preload_counter)
		if (blk4x4_inter_preload_counter != 6'd0 && blk4x4_inter_preload_counter != 6'd1)
			begin
				if (IsInterLuma)
					begin
						if (!mv_below8x8_curr)
							case (pos_FracL)
								`pos_f,`pos_q,`pos_i,`pos_k,`pos_j:
								case (blk4x4_inter_preload_counter_m2[1:0])
									2'b00:xInt_curr_offset <= 5'b01010; //+10
									2'b01:xInt_curr_offset <= 5'b00110; //+6
									2'b10:xInt_curr_offset <= 5'b00010; //+2
									2'b11:xInt_curr_offset <= 5'b11110; //-2
								endcase
								`pos_d,`pos_h,`pos_n:
								if (xInt_org_unclip[1:0] == 2'b00)
									xInt_curr_offset <= (blk4x4_inter_preload_counter_m2[0])? 4'b0:4'b0100; //+0 or +4
								else
									case (blk4x4_inter_preload_counter_m2)
										6'd38,6'd35,6'd32,6'd29,6'd26,6'd23,6'd20,6'd17,6'd14,6'd11,6'd8,6'd5,6'd2:
										xInt_curr_offset <= 5'b0; 			//+0
										6'd37,6'd34,6'd31,6'd28,6'd25,6'd22,6'd19,6'd16,6'd13,6'd10,6'd7,6'd4,6'd1:
										xInt_curr_offset <= 5'b00100;		//+4
										default:xInt_curr_offset <= 5'b01000;//+8
									endcase
								`pos_a,`pos_b,`pos_c:
								case (blk4x4_inter_preload_counter_m2[1:0])
									2'b00:xInt_curr_offset <= 5'b01010; //+10
									2'b01:xInt_curr_offset <= 5'b00110; //+6
									2'b10:xInt_curr_offset <= 5'b00010; //+2
									2'b11:xInt_curr_offset <= 5'b11110; //-2
								endcase
								`pos_Int:
								if (xInt_org_unclip[1:0] == 2'b00)
									xInt_curr_offset <= (blk4x4_inter_preload_counter_m2[0])? 5'b0:5'b0100; //+0 or +4
								else
									case (blk4x4_inter_preload_counter_m2)
										6'd23,6'd20,6'd17,6'd14,6'd11,6'd8,6'd5,6'd2:
										xInt_curr_offset <= 5'b00000;	   	//+0
										6'd22,6'd19,6'd16,6'd13,6'd10,6'd7,6'd4,6'd1:
										xInt_curr_offset <= 5'b00100;	   	//+4
										default:xInt_curr_offset <= 5'b01000;//+8
									endcase
								`pos_e,`pos_g,`pos_p,`pos_r:
								case (blk4x4_inter_preload_counter_m2)
									6'd47,6'd44,6'd5,6'd2:
									xInt_curr_offset <= 5'b00000;	//+0
									6'd46,6'd43,6'd4,6'd1:
									xInt_curr_offset <= 5'b00100;	//+4
									6'd45,6'd42,6'd3,6'd0:
									xInt_curr_offset <= 5'b01000;	//+8
									default:
									case (blk4x4_inter_preload_counter_m2[1:0])
										2'b00:xInt_curr_offset <= 5'b00010; //+2
										2'b01:xInt_curr_offset <= 5'b11110; //-2
										2'b10:xInt_curr_offset <= 5'b01010; //+10
										2'b11:xInt_curr_offset <= 5'b00110; //+6
									endcase
								endcase
							endcase
						else		//block partition below 8x8
							case (pos_FracL)
								`pos_f,`pos_q,`pos_i,`pos_k,`pos_j:
								case (blk4x4_inter_preload_counter_m2)
									6'd26,6'd23,6'd20,6'd17,6'd14,6'd11,6'd8,6'd5,6'd2:xInt_curr_offset <= 5'b11110;//-2
									6'd25,6'd22,6'd19,6'd16,6'd13,6'd10,6'd7,6'd4,6'd1:xInt_curr_offset <= 5'b00010;//+2
									default:xInt_curr_offset <= 5'b00110;											//+6
								endcase
								`pos_d,`pos_h,`pos_n:
								if (xInt_org_unclip[1:0] == 2'b00)
									xInt_curr_offset <= 5'b0;	//+0
								else
									xInt_curr_offset <= (blk4x4_inter_preload_counter_m2[0])? 5'b0:5'b00100;//+0 or +4
								`pos_a,`pos_b,`pos_c:
								case (blk4x4_inter_preload_counter_m2)
									6'd11,6'd8,6'd5,6'd2:xInt_curr_offset <= 5'b11110;	//-2
									6'd10,6'd7,6'd4,6'd1:xInt_curr_offset <= 5'b00010;	//+2
									default:xInt_curr_offset <= 5'b00110;				//+6
								endcase
								`pos_Int:
								if (xInt_org_unclip[1:0] == 2'b00)
									xInt_curr_offset <= 5'b0;	//+0
								else
									xInt_curr_offset <= (blk4x4_inter_preload_counter_m2[0])? 5'b0:5'b00100;	//+0 or +4
								`pos_e,`pos_g,`pos_p,`pos_r:
								case (blk4x4_inter_preload_counter_m2)
									6'd22,6'd20,6'd3,6'd1:xInt_curr_offset <= 5'b0;			//+0	
									6'd21,6'd19,6'd2,6'd0:xInt_curr_offset <= 5'b00100;		//+4 
									6'd18,6'd15,6'd12,6'd9,6'd6:xInt_curr_offset <= 5'b11110;//-2
									6'd17,6'd14,6'd11,6'd8,6'd5:xInt_curr_offset <= 5'b00010;//+2
									6'd16,6'd13,6'd10,6'd7,6'd4:xInt_curr_offset <= 5'b00110;//+6
									default:xInt_curr_offset <= 5'b0;
								endcase
							endcase
					end
				else	//IsInterChroma
					begin
						if (!mv_below8x8_curr)
							begin
								if (xFracC == 0 && yFracC == 0)
									begin
										if (xInt_org_unclip[1:0] == 2'b00)
											xInt_curr_offset <= 5'b0;
										else
											xInt_curr_offset <= (blk4x4_inter_preload_counter_m2[0] == 1'b1)? 5'b0:5'b0100;
									end
								else
									xInt_curr_offset <= (blk4x4_inter_preload_counter_m2[0] == 1'b1)? 5'b0:5'b0100;
							end
						else //mv_below8x8_curr == 1'b1
							begin
								if (xFracC == 0 && yFracC == 0)