[mandelfpga.git] / Main.v

/* 
 * MandelFPGA
 * by Joshua Wise and Chris Lu
 * 
 * An implementation of a pipelined algorithm to calculate the Mandelbrot set
 * in real time on an FPGA.
 */
 
/* verilator lint_off WIDTH */

`define XRES 640
`define YRES 480
`define WHIRRRRR 47

`define TOPBIT 12

module SyncGen(
	input pixclk,
	output reg vs, hs,
	output reg [11:0] xout = `WHIRRRRR, yout = 0,
	output wire [11:0] xoutreal, youtreal,
	output reg border);
	
	reg [11:0] x = 0, y = 0;	// Used for generating border and timing.
	assign xoutreal = x;
	assign youtreal = y;
	
	parameter XFPORCH = 16;
	parameter XSYNC = 96;
	parameter XBPORCH = 48;
	
	parameter YFPORCH = 10;
	parameter YSYNC = 2;
	parameter YBPORCH = 29;

	always @(posedge pixclk)
	begin
		if (x >= (`XRES + XFPORCH + XSYNC + XBPORCH))
		begin
			if (y >= (`YRES + YFPORCH + YSYNC + YBPORCH))
				y <= 0;
			else
				y <= y + 1;
			x <= 0;
		end else
			x <= x + 1;
			
		if (xout >= (`XRES + XFPORCH + XSYNC + XBPORCH))
		begin
			if (yout >= (`YRES + YFPORCH + YSYNC + YBPORCH))
				yout <= 0;
			else
				yout <= yout + 1;
			xout <= 0;
		end else
			xout <= xout + 1;
		hs <= (x >= (`XRES + XFPORCH)) && (x < (`XRES + XFPORCH + XSYNC));
		vs <= (y >= (`YRES + YFPORCH)) && (y < (`YRES + YFPORCH + YSYNC));
		border <= (x > `XRES) || (y > `YRES);
	end
endmodule

// bits: 1.12

module NaiveMultiplier(
	input clk,
	input [`TOPBIT:0] x, y,
	input xsign, ysign,
	output reg [`TOPBIT:0] out,
	output reg sign,
	output reg ovf);

	always @(posedge clk)
	begin
		{ovf,out} <=
			(((y[12] ? (x       ) : 0)   +
			  (y[11] ? (x[`TOPBIT:1]) : 0)   +
			  (y[10] ? (x[`TOPBIT:2]) : 0))  +
			(((y[9]  ? (x[`TOPBIT:3]) : 0)   +
			  (y[8]  ? (x[`TOPBIT:4]) : 0))  +
			 ((y[7]  ? (x[`TOPBIT:5]) : 0)   +
			  (y[6]  ? (x[`TOPBIT:6]) : 0))))+
			(((y[5]  ? (x[`TOPBIT:7]) : 0)   +
			  (y[4]  ? (x[`TOPBIT:8]) : 0)   +
			  (y[3]  ? (x[`TOPBIT:9]) : 0))  +
			 ((y[2]  ? (x[`TOPBIT:10]): 0)   +
			  (y[1]  ? (x[`TOPBIT:11]): 0)   +
			  (y[0]  ? (x[`TOPBIT]): 0)));
		sign <= xsign ^ ysign;
	end

endmodule

module Multiplier(
	input clk,
	input [`TOPBIT:0] x, y,
	input xsign, ysign,
	output wire [`TOPBIT:0] out,
	output wire sign,
	output wire overflow);

	NaiveMultiplier nm(clk, x, y, xsign, ysign, out, sign, overflow);

endmodule

// Yuq.
module MandelUnit(
	input clk,
	input [`TOPBIT:0] x, y,
	input xsign, ysign,
	input [`TOPBIT+2:0] r, i,
	input rsign, isign,
	input [7:0] ibail, icuriter,
	output reg [`TOPBIT:0] xout, yout,
	output reg xsout, ysout,
	output reg [`TOPBIT+2:0] rout, iout,
	output reg rsout, isout,
	output reg [7:0] obail, ocuriter);

	wire [`TOPBIT+1:0] r2, i2;
	wire [`TOPBIT+2:0] ri, diff;
	wire [`TOPBIT+3:0] twocdiff;
	wire r2sign, i2sign, risign, dsign;
	wire [`TOPBIT+2:0] bigsum;
	wire bigsum_ovf;

	reg [`TOPBIT:0] xd, yd;
	reg ineedbaild;
	reg xsd, ysd;
	reg [7:0] ibaild, curiterd;

	assign ri[0] = 0;

	Multiplier r2m(clk, r[`TOPBIT:0], r[`TOPBIT:0], rsign, rsign, r2[`TOPBIT:0], r2sign, r2[`TOPBIT+1]);
	Multiplier i2m(clk, i[`TOPBIT:0], i[`TOPBIT:0], isign, isign, i2[`TOPBIT:0], i2sign, i2[`TOPBIT+1]);
	Multiplier rim(clk, r[`TOPBIT:0], i[`TOPBIT:0], rsign, isign, ri[`TOPBIT+1:1], risign, ri[`TOPBIT+2]);

	assign bigsum = r2[`TOPBIT+1:0] + i2[`TOPBIT+1:0];
	assign bigsum_ovf = bigsum[`TOPBIT+2];
	
	assign twocdiff = r2 - i2;
	assign diff = twocdiff[`TOPBIT+3] ? -twocdiff : twocdiff;
	assign dsign = twocdiff[`TOPBIT+3];
	
	wire [`TOPBIT+3:0] twocrout = xd - diff;
	wire [`TOPBIT+3:0] twociout = yd - ri;

	always @ (posedge clk)
	begin
		xd <= x;
		yd <= y;
		xsd <= xsign;
		ysd <= ysign;
		xout <= xd;
		yout <= yd;
		xsout <= xsd;
		ysout <= ysd;
		ibaild <= ibail;
		curiterd <= icuriter;
		ineedbaild <= r[`TOPBIT+1] | r[`TOPBIT+2] | i[`TOPBIT+1] | i[`TOPBIT+2];

		// r^2 - i^2 + x
		if (xsd ^ dsign) begin
			if (twocrout[`TOPBIT+3]) begin	// diff > xd
				rout <= -twocrout;
				rsout <= dsign;
			end else begin
				rout <= twocrout;
				rsout <= xsd;
			end
		end else begin
			rout <= diff + xd;
			rsout <= xsd;	// xsd == dsign
		end
		
		// 2 * r * i + y
		if (ysd ^ risign) begin
			if (twociout[`TOPBIT+3]) begin	// ri > yd
				iout <= -twociout;
				isout <= risign;
			end else begin
				iout <= twociout;
				isout <= ysd;
			end
		end else begin
			iout <= ri + yd;
			isout <= ysd;
		end
		
		// If we haven't bailed out, and we meet any of the bailout conditions,
		// bail out now.  Otherwise, leave the bailout at whatever it was before.
		if ((ibaild == 255) && (bigsum_ovf | ineedbaild))
			obail <= curiterd;
		else
			obail <= ibaild;
		ocuriter <= curiterd + 8'b1;
	end

endmodule

module Mandelbrot(
	input mclk,
	input pixclk,
	input [11:0] x, y,
	input [`TOPBIT+1:0] xofs, yofs,
	input [7:0] colorofs,
	input [2:0] scale,
	output reg [2:0] red, green, output reg [1:0] blue);

`define MAXOUTN 21
	
	wire [`TOPBIT:0] rx, ry;
	wire [`TOPBIT+1:0] nx, ny;
	wire rxsign, rysign;
	
	assign nx = {2'b0,x} + {2'b0,xofs};
	assign ny = {2'b0,y} + {2'b0,yofs};
	assign rx = (nx[`TOPBIT+1] ? -nx[`TOPBIT:0] : nx[`TOPBIT:0]) << scale;
	assign rxsign = nx[`TOPBIT+1];
	assign ry = (ny[`TOPBIT+1] ? -ny[`TOPBIT:0] : ny[`TOPBIT:0]) << scale;
	assign rysign = ny[`TOPBIT+1];

	wire [`TOPBIT+2:0] mr[`MAXOUTN:0], mi[`MAXOUTN:0];
	wire mrs[`MAXOUTN:0], mis[`MAXOUTN:0];
	wire [7:0] mb[`MAXOUTN:0];
	wire [`TOPBIT:0] xprop[`MAXOUTN:0], yprop[`MAXOUTN:0];
	wire xsprop[`MAXOUTN:0], ysprop[`MAXOUTN:0];
	wire [7:0] curiter[`MAXOUTN:0];
	
	reg [`TOPBIT:0] initx, inity;
	reg [`TOPBIT+2:0] initr, initi;
	reg [7:0] initci, initb;
	reg initxs, initys, initrs, initis;
	
	// Values after the number of iterations denoted by the subscript.
	reg [`TOPBIT:0] stagex [2:1], stagey [2:1];
	reg [`TOPBIT+2:0] stager [2:1], stagei [2:1];
	reg [7:0] stageci [2:1], stageb [2:1];
	reg stagexs [2:1], stageys [2:1], stagers [2:1], stageis [2:1];
	
	reg [2:0] state = 3'b001;	// One-hot encoded state.
	
	// States are advanced one from what they should be, so that they'll
	// get there on the _next_ mclk.
	always @(posedge mclk)
	begin
		initx <= (state[2]) ? rx :
	            (state[0]) ? stagex[1] :
	            (state[1]) ? stagex[2] : 0;
		inity <= (state[2]) ? ry :
	               (state[0]) ? stagey[1] :
	               (state[1]) ? stagey[2] : 0;
		initr <= (state[2]) ? {2'b0,rx} :
	               (state[0]) ? stager[1] :
	               (state[1]) ? stager[2] : 0;
		initi <= (state[2]) ? {2'b0,ry} :
	               (state[0]) ? stagei[1] :
	               (state[1]) ? stagei[2] : 0;
		initxs <= (state[2]) ? rxsign :
	                (state[0]) ? stagexs[1] :
	                (state[1]) ? stagexs[2] : 0;
		initys <= (state[2]) ? rysign :
	                (state[0]) ? stageys[1] :
	                (state[1]) ? stageys[2] : 0;
		initrs <= (state[2]) ? rxsign :
	                (state[0]) ? stagers[1] :
	                (state[1]) ? stagers[2] : 0;
		initis <= (state[2]) ? rysign :
	                (state[0]) ? stageis[1] :
	                (state[1]) ? stageis[2] : 0;
		initb <= (state[2]) ? 8'b11111111 :
	               (state[0]) ? stageb[1] :
	               (state[1]) ? stageb[2] : 0;
		initci <= (state[2]) ? 8'b00000000 :
	                (state[0]) ? stageci[1] : 
	                (state[1]) ? stageci[2] : 0;
	end
	
	reg [7:0] out;
	
	// We detect when the state should be poked by a high negedge followed
	// by a high posedge -- if that happens, then we're guaranteed that the
	// state following the current state will be 3'b100.
	reg lastneg;
	always @(negedge mclk)
		lastneg <= pixclk;
	
	always @(posedge mclk)
	begin
		if (lastneg && pixclk)	// If a pixclk has happened, the state should be reset.
			state <= 3'b100;
		else						// Otherwise, just poke it forward.
			case(state)
			3'b001: state <= 3'b010;
			3'b010: state <= 3'b100;
			3'b100: state <= 3'b001;
		`ifdef isim
			default: begin $display("invalid state"); $finish; end
		`endif
			endcase
	
		// Data output handling
		if (state[0]) begin
			{red, green, blue} <= {out[0],out[3],out[6],out[1],out[4],out[7],out[2],out[5]};
		end
		if (state[1]) begin
			out <= ~mb[`MAXOUTN] + colorofs;
		end
		
		if (state[0]) begin		// PnR0 in, PnR2 out
			stagex[2] <= xprop[`MAXOUTN];
			stagey[2] <= yprop[`MAXOUTN];
			stager[2] <= mr[`MAXOUTN];
			stagei[2] <= mi[`MAXOUTN];
			stagexs[2] <= xsprop[`MAXOUTN];
			stageys[2] <= ysprop[`MAXOUTN];
			stagers[2] <= mrs[`MAXOUTN];
			stageis[2] <= mis[`MAXOUTN];
			stageb[2] <= mb[`MAXOUTN];
			stageci[2] <= curiter[`MAXOUTN];
		end
		
		if (state[2]) begin	// PnR2 in, PnR1 out
			stagex[1] <= xprop[`MAXOUTN];
			stagey[1] <= yprop[`MAXOUTN];
			stager[1] <= mr[`MAXOUTN];
			stagei[1] <= mi[`MAXOUTN];
			stagexs[1] <= xsprop[`MAXOUTN];
			stageys[1] <= ysprop[`MAXOUTN];
			stagers[1] <= mrs[`MAXOUTN];
			stageis[1] <= mis[`MAXOUTN];
			stageb[1] <= mb[`MAXOUTN];
			stageci[1] <= curiter[`MAXOUTN];
		end
	end

	MandelUnit mu0(
		mclk,
		initx, inity, initxs, initys,
		initr, initi, initrs, initis,
		initb, initci,
		xprop[0], yprop[0], xsprop[0], ysprop[0],
		mr[0], mi[0], mrs[0], mis[0],
		mb[0], curiter[0]);
		
`define MAKE_UNIT(name, num) \
	MandelUnit name(mclk, \
		xprop[(num)], yprop[(num)], xsprop[(num)], ysprop[(num)], mr[(num)], mi[(num)], mrs[(num)], mis[(num)], mb[(num)], curiter[(num)], \
		xprop[(num)+1], yprop[(num)+1], xsprop[(num)+1], ysprop[(num)+1], mr[(num)+1], mi[(num)+1], mrs[(num)+1], mis[(num)+1], mb[(num)+1], curiter[(num)+1])

	`MAKE_UNIT(mu1, 0);
	`MAKE_UNIT(mu2, 1);
	`MAKE_UNIT(mu3, 2);
	`MAKE_UNIT(mu4, 3);
	`MAKE_UNIT(mu5, 4);
	`MAKE_UNIT(mu6, 5);
	`MAKE_UNIT(mu7, 6);
	`MAKE_UNIT(mu8, 7);
	`MAKE_UNIT(mu9, 8);
	`MAKE_UNIT(mua, 9);
	`MAKE_UNIT(mub, 10);
	`MAKE_UNIT(muc, 11);
	`MAKE_UNIT(mud, 12);
	`MAKE_UNIT(mue, 13);
	`MAKE_UNIT(muf, 14);
	`MAKE_UNIT(mug, 15);
	`MAKE_UNIT(muh, 16);
	`MAKE_UNIT(mui, 17);
	`MAKE_UNIT(muj, 18);
	`MAKE_UNIT(muk, 19);
	`MAKE_UNIT(mul, 20);
endmodule

module Logo(
	input pixclk,
	input [11:0] x, y,
	output wire enb,
	output wire [2:0] red, green, output wire [1:0] blue);
	
	reg [1:0] logo[8191:0];
	initial $readmemb("logo.readmemb", logo);
	
	assign enb = (x < 96) && (y < 64);
	wire [12:0] addr = {y[5:0], x[6:0]};
	wire [1:0] data = logo[addr];
	assign {red, green, blue} = 
		 (data == 2'b00) ? 8'b00000000 :
		((data == 2'b01) ? 8'b00011100 :
		((data == 2'b10) ? 8'b11100000 :
		                    8'b11111111));
endmodule

module MandelTop(
`ifdef verilator
	input pixclk, mclk,
`else
	input gclk, output wire dcmok,
`endif
	output wire vs, hs,
	output wire [2:0] red, green, output [1:0] blue,
	input left, right, up, down, rst, cycle, logooff,
	input [2:0] scale);
	
`ifdef verilator
`else
	wire pixclk, mclk, clk;
	wire dcm1ok, dcm2ok;
	assign dcmok = dcm1ok && dcm2ok;
	
	IBUFG iclkbuf(.O(clk), .I(gclk));
	
	pixDCM dcm(					// CLKIN is 50MHz xtal, CLKFX_OUT is 25MHz
		.CLKIN_IN(clk), 
		.CLKFX_OUT(pixclk),
		.LOCKED_OUT(dcm1ok)
		);
	
	mandelDCM dcm2(
		.CLKIN_IN(clk),
		.CLKFX_OUT(mclk),
		.LOCKED_OUT(dcm2ok)
		);
`endif

	wire border;
	wire [11:0] x, y;
	reg [`TOPBIT+1:0] xofs = -`XRES/2, yofs = -`YRES/2;
	reg [5:0] slowctr = 0;
	reg [7:0] colorcycle = 0;
	wire [11:0] realx, realy;
	
	wire logoenb;
	wire [2:0] mandelr, mandelg, logor, logog;
	wire [1:0] mandelb, logob;
	
	SyncGen sync(pixclk, vs, hs, x, y, realx, realy, border);
	Mandelbrot mandel(mclk, pixclk, x, y, xofs, yofs, cycle ? colorcycle : 8'b0, scale, mandelr, mandelg, mandelb);
	Logo logo(pixclk, realx, realy, logoenb, logor, logog, logob);
	
	assign {red,green,blue} =
		border ? 8'b00000000 :
		(!logooff && logoenb) ? {logor, logog, logob} : {mandelr, mandelg, mandelb};
	
	always @(posedge vs)
	begin
		if (rst)
		begin
			xofs <= -`XRES/2;
			yofs <= -`YRES/2;
			colorcycle <= 0;
		end else begin
			if (up) yofs <= yofs + 1;
			else if (down) yofs <= yofs - 1;
			
			if (left) xofs <= xofs + 1;
			else if (right) xofs <= xofs - 1;
			
			if (slowctr == 0)
				colorcycle <= colorcycle + 1;
		end
		
		if (slowctr == 12)
			slowctr <= 0;
		else
			slowctr <= slowctr + 1;
	end
endmodule
Commit	Line	Data
cf1ae842	1	/*
05c0805b JW	2	* MandelFPGA
	3	* by Joshua Wise and Chris Lu
	4	*
	5	* An implementation of a pipelined algorithm to calculate the Mandelbrot set
	6	* in real time on an FPGA.
	7	*/
772b2413 JW	8
772b2413 JW	9	/* verilator lint_off WIDTH */
05c0805b JW	10
	11	`define XRES 640
	12	`define YRES 480
fd517766 JW	13	`define WHIRRRRR 47
	14
	15	`define TOPBIT 12
05c0805b JW	16
	17	module SyncGen(
	18	input pixclk,
	19	output reg vs, hs,
	20	output reg [11:0] xout = `WHIRRRRR, yout = 0,
	21	output wire [11:0] xoutreal, youtreal,
	22	output reg border);
	23
	24	reg [11:0] x = 0, y = 0; // Used for generating border and timing.
	25	assign xoutreal = x;
	26	assign youtreal = y;
	27
	28	parameter XFPORCH = 16;
	29	parameter XSYNC = 96;
	30	parameter XBPORCH = 48;
	31
	32	parameter YFPORCH = 10;
	33	parameter YSYNC = 2;
	34	parameter YBPORCH = 29;
	35
	36	always @(posedge pixclk)
	37	begin
	38	if (x >= (`XRES + XFPORCH + XSYNC + XBPORCH))
	39	begin
	40	if (y >= (`YRES + YFPORCH + YSYNC + YBPORCH))
	41	y <= 0;
	42	else
	43	y <= y + 1;
	44	x <= 0;
	45	end else
	46	x <= x + 1;
	47
	48	if (xout >= (`XRES + XFPORCH + XSYNC + XBPORCH))
	49	begin
	50	if (yout >= (`YRES + YFPORCH + YSYNC + YBPORCH))
	51	yout <= 0;
	52	else
	53	yout <= yout + 1;
	54	xout <= 0;
	55	end else
	56	xout <= xout + 1;
	57	hs <= (x >= (`XRES + XFPORCH)) && (x < (`XRES + XFPORCH + XSYNC));
	58	vs <= (y >= (`YRES + YFPORCH)) && (y < (`YRES + YFPORCH + YSYNC));
	59	border <= (x > `XRES) \|\| (y > `YRES);
	60	end
	61	endmodule
	62
	63	// bits: 1.12
	64
	65	module NaiveMultiplier(
	66	input clk,
fd517766	67	input [`TOPBIT:0] x, y,
05c0805b	68	input xsign, ysign,
fd517766	69	output reg [`TOPBIT:0] out,
05c0805b	70	output reg sign,
fb8d158b	71	output reg ovf);
05c0805b JW	72
	73	always @(posedge clk)
	74	begin
	75	{ovf,out} <=
b7403498	76	(((y[12] ? (x ) : 0) +
fd517766 JW	77	(y[11] ? (x[`TOPBIT:1]) : 0) +
	78	(y[10] ? (x[`TOPBIT:2]) : 0)) +
	79	(((y[9] ? (x[`TOPBIT:3]) : 0) +
	80	(y[8] ? (x[`TOPBIT:4]) : 0)) +
	81	((y[7] ? (x[`TOPBIT:5]) : 0) +
	82	(y[6] ? (x[`TOPBIT:6]) : 0))))+
	83	(((y[5] ? (x[`TOPBIT:7]) : 0) +
	84	(y[4] ? (x[`TOPBIT:8]) : 0) +
	85	(y[3] ? (x[`TOPBIT:9]) : 0)) +
	86	((y[2] ? (x[`TOPBIT:10]): 0) +
	87	(y[1] ? (x[`TOPBIT:11]): 0) +
	88	(y[0] ? (x[`TOPBIT]): 0)));
05c0805b JW	89	sign <= xsign ^ ysign;
	90	end
	91
	92	endmodule
	93
	94	module Multiplier(
	95	input clk,
fd517766	96	input [`TOPBIT:0] x, y,
05c0805b	97	input xsign, ysign,
fd517766	98	output wire [`TOPBIT:0] out,
05c0805b	99	output wire sign,
fb8d158b	100	output wire overflow);
05c0805b JW	101
	102	NaiveMultiplier nm(clk, x, y, xsign, ysign, out, sign, overflow);
	103
	104	endmodule
	105
fb8d158b	106	// Yuq.
05c0805b JW	107	module MandelUnit(
05c0805b JW	108	input clk,
fd517766	109	input [`TOPBIT:0] x, y,
05c0805b	110	input xsign, ysign,
fd517766	111	input [`TOPBIT+2:0] r, i,
05c0805b JW	112	input rsign, isign,
05c0805b JW	113	input [7:0] ibail, icuriter,
fd517766	114	output reg [`TOPBIT:0] xout, yout,
05c0805b	115	output reg xsout, ysout,
fd517766	116	output reg [`TOPBIT+2:0] rout, iout,
05c0805b JW	117	output reg rsout, isout,
	118	output reg [7:0] obail, ocuriter);
	119
fd517766 JW	120	wire [`TOPBIT+1:0] r2, i2;
	121	wire [`TOPBIT+2:0] ri, diff;
	122	wire [`TOPBIT+3:0] twocdiff;
05c0805b	123	wire r2sign, i2sign, risign, dsign;
fd517766	124	wire [`TOPBIT+2:0] bigsum;
6cdf39e2	125	wire bigsum_ovf;
05c0805b	126
fd517766	127	reg [`TOPBIT:0] xd, yd;
2afeab21	128	reg ineedbaild;
05c0805b JW	129	reg xsd, ysd;
	130	reg [7:0] ibaild, curiterd;
	131
	132	assign ri[0] = 0;
	133
fd517766 JW	134	Multiplier r2m(clk, r[`TOPBIT:0], r[`TOPBIT:0], rsign, rsign, r2[`TOPBIT:0], r2sign, r2[`TOPBIT+1]);
	135	Multiplier i2m(clk, i[`TOPBIT:0], i[`TOPBIT:0], isign, isign, i2[`TOPBIT:0], i2sign, i2[`TOPBIT+1]);
	136	Multiplier rim(clk, r[`TOPBIT:0], i[`TOPBIT:0], rsign, isign, ri[`TOPBIT+1:1], risign, ri[`TOPBIT+2]);
05c0805b	137
fd517766 JW	138	assign bigsum = r2[`TOPBIT+1:0] + i2[`TOPBIT+1:0];
fd517766 JW	139	assign bigsum_ovf = bigsum[`TOPBIT+2];
2afeab21	140
9032b2b5	141	assign twocdiff = r2 - i2;
fd517766 JW	142	assign diff = twocdiff[`TOPBIT+3] ? -twocdiff : twocdiff;
fd517766 JW	143	assign dsign = twocdiff[`TOPBIT+3];
2afeab21	144
fd517766 JW	145	wire [`TOPBIT+3:0] twocrout = xd - diff;
fd517766 JW	146	wire [`TOPBIT+3:0] twociout = yd - ri;
05c0805b JW	147
	148	always @ (posedge clk)
	149	begin
	150	xd <= x;
	151	yd <= y;
	152	xsd <= xsign;
	153	ysd <= ysign;
	154	xout <= xd;
	155	yout <= yd;
	156	xsout <= xsd;
	157	ysout <= ysd;
	158	ibaild <= ibail;
	159	curiterd <= icuriter;
fd517766	160	ineedbaild <= r[`TOPBIT+1] \| r[`TOPBIT+2] \| i[`TOPBIT+1] \| i[`TOPBIT+2];
05c0805b	161
2afeab21	162	// r^2 - i^2 + x
05c0805b	163	if (xsd ^ dsign) begin
fd517766	164	if (twocrout[`TOPBIT+3]) begin // diff > xd
2afeab21	165	rout <= -twocrout;
05c0805b JW	166	rsout <= dsign;
05c0805b JW	167	end else begin
2afeab21	168	rout <= twocrout;
05c0805b JW	169	rsout <= xsd;
	170	end
	171	end else begin
	172	rout <= diff + xd;
2afeab21	173	rsout <= xsd; // xsd == dsign
05c0805b JW	174	end
05c0805b JW	175
2afeab21	176	// 2 * r * i + y
05c0805b	177	if (ysd ^ risign) begin
fd517766	178	if (twociout[`TOPBIT+3]) begin // ri > yd
2afeab21	179	iout <= -twociout;
05c0805b JW	180	isout <= risign;
05c0805b JW	181	end else begin
2afeab21	182	iout <= twociout;
05c0805b JW	183	isout <= ysd;
	184	end
	185	end else begin
	186	iout <= ri + yd;
	187	isout <= ysd;
	188	end
	189
	190	// If we haven't bailed out, and we meet any of the bailout conditions,
	191	// bail out now. Otherwise, leave the bailout at whatever it was before.
2afeab21	192	if ((ibaild == 255) && (bigsum_ovf \| ineedbaild))
05c0805b JW	193	obail <= curiterd;
	194	else
	195	obail <= ibaild;
	196	ocuriter <= curiterd + 8'b1;
	197	end
	198
	199	endmodule
	200
	201	module Mandelbrot(
	202	input mclk,
	203	input pixclk,
	204	input [11:0] x, y,
fd517766	205	input [`TOPBIT+1:0] xofs, yofs,
05c0805b JW	206	input [7:0] colorofs,
	207	input [2:0] scale,
	208	output reg [2:0] red, green, output reg [1:0] blue);
281eac32	209
fd517766	210	`define MAXOUTN 21
05c0805b	211
fd517766 JW	212	wire [`TOPBIT:0] rx, ry;
fd517766 JW	213	wire [`TOPBIT+1:0] nx, ny;
05c0805b JW	214	wire rxsign, rysign;
05c0805b JW	215
772b2413 JW	216	assign nx = {2'b0,x} + {2'b0,xofs};
772b2413 JW	217	assign ny = {2'b0,y} + {2'b0,yofs};
fd517766 JW	218	assign rx = (nx[`TOPBIT+1] ? -nx[`TOPBIT:0] : nx[`TOPBIT:0]) << scale;
	219	assign rxsign = nx[`TOPBIT+1];
	220	assign ry = (ny[`TOPBIT+1] ? -ny[`TOPBIT:0] : ny[`TOPBIT:0]) << scale;
	221	assign rysign = ny[`TOPBIT+1];
05c0805b	222
fd517766	223	wire [`TOPBIT+2:0] mr[`MAXOUTN:0], mi[`MAXOUTN:0];
281eac32 JW	224	wire mrs[`MAXOUTN:0], mis[`MAXOUTN:0];
281eac32 JW	225	wire [7:0] mb[`MAXOUTN:0];
fd517766	226	wire [`TOPBIT:0] xprop[`MAXOUTN:0], yprop[`MAXOUTN:0];
281eac32 JW	227	wire xsprop[`MAXOUTN:0], ysprop[`MAXOUTN:0];
281eac32 JW	228	wire [7:0] curiter[`MAXOUTN:0];
05c0805b	229
fd517766 JW	230	reg [`TOPBIT:0] initx, inity;
fd517766 JW	231	reg [`TOPBIT+2:0] initr, initi;
79af494a JW	232	reg [7:0] initci, initb;
79af494a JW	233	reg initxs, initys, initrs, initis;
05c0805b	234
534b3903	235	// Values after the number of iterations denoted by the subscript.
fd517766 JW	236	reg [`TOPBIT:0] stagex [2:1], stagey [2:1];
fd517766 JW	237	reg [`TOPBIT+2:0] stager [2:1], stagei [2:1];
534b3903 JW	238	reg [7:0] stageci [2:1], stageb [2:1];
534b3903 JW	239	reg stagexs [2:1], stageys [2:1], stagers [2:1], stageis [2:1];
05c0805b	240
534b3903	241	reg [2:0] state = 3'b001; // One-hot encoded state.
05c0805b	242
79af494a JW	243	// States are advanced one from what they should be, so that they'll
	244	// get there on the _next_ mclk.
	245	always @(posedge mclk)
	246	begin
	247	initx <= (state[2]) ? rx :
	248	(state[0]) ? stagex[1] :
	249	(state[1]) ? stagex[2] : 0;
	250	inity <= (state[2]) ? ry :
	251	(state[0]) ? stagey[1] :
	252	(state[1]) ? stagey[2] : 0;
772b2413	253	initr <= (state[2]) ? {2'b0,rx} :
79af494a JW	254	(state[0]) ? stager[1] :
79af494a JW	255	(state[1]) ? stager[2] : 0;
772b2413	256	initi <= (state[2]) ? {2'b0,ry} :
79af494a JW	257	(state[0]) ? stagei[1] :
	258	(state[1]) ? stagei[2] : 0;
	259	initxs <= (state[2]) ? rxsign :
	260	(state[0]) ? stagexs[1] :
	261	(state[1]) ? stagexs[2] : 0;
	262	initys <= (state[2]) ? rysign :
	263	(state[0]) ? stageys[1] :
	264	(state[1]) ? stageys[2] : 0;
	265	initrs <= (state[2]) ? rxsign :
	266	(state[0]) ? stagers[1] :
	267	(state[1]) ? stagers[2] : 0;
	268	initis <= (state[2]) ? rysign :
	269	(state[0]) ? stageis[1] :
	270	(state[1]) ? stageis[2] : 0;
	271	initb <= (state[2]) ? 8'b11111111 :
	272	(state[0]) ? stageb[1] :
	273	(state[1]) ? stageb[2] : 0;
	274	initci <= (state[2]) ? 8'b00000000 :
	275	(state[0]) ? stageci[1] :
	276	(state[1]) ? stageci[2] : 0;
	277	end
05c0805b JW	278
05c0805b JW	279	reg [7:0] out;
251788d8 JW	280
251788d8 JW	281	// We detect when the state should be poked by a high negedge followed
fb8d158b JW	282	// by a high posedge -- if that happens, then we're guaranteed that the
fb8d158b JW	283	// state following the current state will be 3'b100.
251788d8	284	reg lastneg;
265061f2	285	always @(negedge mclk)
251788d8	286	lastneg <= pixclk;
05c0805b JW	287
	288	always @(posedge mclk)
	289	begin
251788d8 JW	290	if (lastneg && pixclk) // If a pixclk has happened, the state should be reset.
	291	state <= 3'b100;
	292	else // Otherwise, just poke it forward.
a3a4354b JW	293	case(state)
	294	3'b001: state <= 3'b010;
	295	3'b010: state <= 3'b100;
	296	3'b100: state <= 3'b001;
cf311cdb	297	`ifdef isim
772b2413	298	default: begin $display("invalid state"); $finish; end
cf311cdb	299	`endif
a3a4354b	300	endcase
251788d8	301
534b3903 JW	302	// Data output handling
534b3903 JW	303	if (state[0]) begin
05c0805b	304	{red, green, blue} <= {out[0],out[3],out[6],out[1],out[4],out[7],out[2],out[5]};
05c0805b	305	end
3068fa61	306	if (state[1]) begin
534b3903 JW	307	out <= ~mb[`MAXOUTN] + colorofs;
	308	end
	309
3068fa61	310	if (state[0]) begin // PnR0 in, PnR2 out
534b3903 JW	311	stagex[2] <= xprop[`MAXOUTN];
	312	stagey[2] <= yprop[`MAXOUTN];
	313	stager[2] <= mr[`MAXOUTN];
	314	stagei[2] <= mi[`MAXOUTN];
	315	stagexs[2] <= xsprop[`MAXOUTN];
	316	stageys[2] <= ysprop[`MAXOUTN];
	317	stagers[2] <= mrs[`MAXOUTN];
	318	stageis[2] <= mis[`MAXOUTN];
	319	stageb[2] <= mb[`MAXOUTN];
	320	stageci[2] <= curiter[`MAXOUTN];
	321	end
	322
3068fa61	323	if (state[2]) begin // PnR2 in, PnR1 out
534b3903 JW	324	stagex[1] <= xprop[`MAXOUTN];
	325	stagey[1] <= yprop[`MAXOUTN];
	326	stager[1] <= mr[`MAXOUTN];
	327	stagei[1] <= mi[`MAXOUTN];
	328	stagexs[1] <= xsprop[`MAXOUTN];
	329	stageys[1] <= ysprop[`MAXOUTN];
	330	stagers[1] <= mrs[`MAXOUTN];
	331	stageis[1] <= mis[`MAXOUTN];
	332	stageb[1] <= mb[`MAXOUTN];
	333	stageci[1] <= curiter[`MAXOUTN];
	334	end
05c0805b JW	335	end
	336
	337	MandelUnit mu0(
	338	mclk,
	339	initx, inity, initxs, initys,
	340	initr, initi, initrs, initis,
	341	initb, initci,
	342	xprop[0], yprop[0], xsprop[0], ysprop[0],
	343	mr[0], mi[0], mrs[0], mis[0],
	344	mb[0], curiter[0]);
e03ccef9 JW	345
	346	`define MAKE_UNIT(name, num) \
	347	MandelUnit name(mclk, \
	348	xprop[(num)], yprop[(num)], xsprop[(num)], ysprop[(num)], mr[(num)], mi[(num)], mrs[(num)], mis[(num)], mb[(num)], curiter[(num)], \
	349	xprop[(num)+1], yprop[(num)+1], xsprop[(num)+1], ysprop[(num)+1], mr[(num)+1], mi[(num)+1], mrs[(num)+1], mis[(num)+1], mb[(num)+1], curiter[(num)+1])
05c0805b	350
e03ccef9 JW	351	`MAKE_UNIT(mu1, 0);
	352	`MAKE_UNIT(mu2, 1);
	353	`MAKE_UNIT(mu3, 2);
	354	`MAKE_UNIT(mu4, 3);
	355	`MAKE_UNIT(mu5, 4);
	356	`MAKE_UNIT(mu6, 5);
	357	`MAKE_UNIT(mu7, 6);
	358	`MAKE_UNIT(mu8, 7);
	359	`MAKE_UNIT(mu9, 8);
	360	`MAKE_UNIT(mua, 9);
	361	`MAKE_UNIT(mub, 10);
fd517766 JW	362	`MAKE_UNIT(muc, 11);
	363	`MAKE_UNIT(mud, 12);
	364	`MAKE_UNIT(mue, 13);
	365	`MAKE_UNIT(muf, 14);
	366	`MAKE_UNIT(mug, 15);
	367	`MAKE_UNIT(muh, 16);
	368	`MAKE_UNIT(mui, 17);
	369	`MAKE_UNIT(muj, 18);
	370	`MAKE_UNIT(muk, 19);
	371	`MAKE_UNIT(mul, 20);
05c0805b JW	372	endmodule
	373
	374	module Logo(
	375	input pixclk,
	376	input [11:0] x, y,
	377	output wire enb,
	378	output wire [2:0] red, green, output wire [1:0] blue);
	379
	380	reg [1:0] logo[8191:0];
	381	initial $readmemb("logo.readmemb", logo);
	382
	383	assign enb = (x < 96) && (y < 64);
	384	wire [12:0] addr = {y[5:0], x[6:0]};
	385	wire [1:0] data = logo[addr];
	386	assign {red, green, blue} =
	387	(data == 2'b00) ? 8'b00000000 :
	388	((data == 2'b01) ? 8'b00011100 :
	389	((data == 2'b10) ? 8'b11100000 :
	390	8'b11111111));
	391	endmodule
	392
	393	module MandelTop(
772b2413 JW	394	`ifdef verilator
	395	input pixclk, mclk,
	396	`else
05c0805b	397	input gclk, output wire dcmok,
772b2413	398	`endif
05c0805b JW	399	output wire vs, hs,
	400	output wire [2:0] red, green, output [1:0] blue,
	401	input left, right, up, down, rst, cycle, logooff,
	402	input [2:0] scale);
772b2413 JW	403
	404	`ifdef verilator
	405	`else
11cc2d9b	406	wire pixclk, mclk, clk;
534b3903	407	wire dcm1ok, dcm2ok;
c3ed4329	408	assign dcmok = dcm1ok && dcm2ok;
265061f2	409
e03ccef9	410	IBUFG iclkbuf(.O(clk), .I(gclk));
534b3903	411
c3ed4329 JW	412	pixDCM dcm( // CLKIN is 50MHz xtal, CLKFX_OUT is 25MHz
	413	.CLKIN_IN(clk),
	414	.CLKFX_OUT(pixclk),
	415	.LOCKED_OUT(dcm1ok)
	416	);
534b3903	417
c3ed4329 JW	418	mandelDCM dcm2(
	419	.CLKIN_IN(clk),
	420	.CLKFX_OUT(mclk),
	421	.LOCKED_OUT(dcm2ok)
	422	);
772b2413 JW	423	`endif
772b2413 JW	424
05c0805b	425	wire border;
05c0805b	426	wire [11:0] x, y;
fd517766	427	reg [`TOPBIT+1:0] xofs = -`XRES/2, yofs = -`YRES/2;
05c0805b JW	428	reg [5:0] slowctr = 0;
	429	reg [7:0] colorcycle = 0;
	430	wire [11:0] realx, realy;
	431
	432	wire logoenb;
	433	wire [2:0] mandelr, mandelg, logor, logog;
	434	wire [1:0] mandelb, logob;
	435
05c0805b	436	SyncGen sync(pixclk, vs, hs, x, y, realx, realy, border);
772b2413	437	Mandelbrot mandel(mclk, pixclk, x, y, xofs, yofs, cycle ? colorcycle : 8'b0, scale, mandelr, mandelg, mandelb);
05c0805b JW	438	Logo logo(pixclk, realx, realy, logoenb, logor, logog, logob);
	439
	440	assign {red,green,blue} =
	441	border ? 8'b00000000 :
	442	(!logooff && logoenb) ? {logor, logog, logob} : {mandelr, mandelg, mandelb};
	443
	444	always @(posedge vs)
	445	begin
	446	if (rst)
	447	begin
	448	xofs <= -`XRES/2;
	449	yofs <= -`YRES/2;
	450	colorcycle <= 0;
	451	end else begin
	452	if (up) yofs <= yofs + 1;
	453	else if (down) yofs <= yofs - 1;
	454
	455	if (left) xofs <= xofs + 1;
	456	else if (right) xofs <= xofs - 1;
	457
	458	if (slowctr == 0)
	459	colorcycle <= colorcycle + 1;
	460	end
	461
	462	if (slowctr == 12)
	463	slowctr <= 0;
	464	else
	465	slowctr <= slowctr + 1;
	466	end
	467	endmodule