Subversion Repositories zipcpu

////////////////////////////////////////////////////////////////////////////////
//
// Filename:	div.v
//
// Project:	Zip CPU -- a small, lightweight, RISC CPU soft core
//
// Purpose:	Provide an Integer divide capability to the Zip CPU.  Provides
//		for both signed and unsigned divide.
//
// Steps:
//	i_rst	The DIVide unit starts in idle.  It can also be placed into an
//	idle by asserting the reset input.
//
//	i_wr	When i_rst is asserted, a divide begins.  On the next clock:
//
//	  o_busy is set high so everyone else knows we are at work and they can
//		wait for us to complete.
//
//	  pre_sign is set to true if we need to do a signed divide.  In this
//		case, we take a clock cycle to turn the divide into an unsigned
//		divide.
//
//	  o_quotient, a place to store our result, is initialized to all zeros.
//
//	  r_dividend is set to the numerator
//
//	  r_divisor is set to 2^31 * the denominator (shift left by 31, or add
//		31 zeros to the right of the number.
//
//	pre_sign When true (clock cycle after i_wr), a clock cycle is used
//		to take the absolute value of the various arguments (r_dividend
//		and r_divisor), and to calculate what sign the output result
//		should be.
//
//
//	At this point, the divide is has started.  The divide works by walking
//	through every shift of the
//
//		    DIVIDEND	over the
//		DIVISOR
//
//	If the DIVISOR is bigger than the dividend, the divisor is shifted
//	right, and nothing is done to the output quotient.
//
//		    DIVIDEND
//		 DIVISOR
//
//	This repeats, until DIVISOR is less than or equal to the divident, as in
//
//		DIVIDEND
//		DIVISOR
//
//	At this point, if the DIVISOR is less than the dividend, the
//	divisor is subtracted from the dividend, and the DIVISOR is again
//	shifted to the right.  Further, a '1' bit gets set in the output
//	quotient.
//
//	Once we've done this for 32 clocks, we've accumulated our answer into
//	the output quotient, and we can proceed to the next step.  If the
//	result will be signed, the next step negates the quotient, otherwise
//	it returns the result.
//
//	On the clock when we are done, o_busy is set to false, and o_valid set
//	to true.  (It is a violation of the ZipCPU internal protocol for both
//	busy and valid to ever be true on the same clock.  It is also a
//	violation for busy to be false with valid true thereafter.)
//
//
// Creator:	Dan Gisselquist, Ph.D.
//		Gisselquist Technology, LLC
//
////////////////////////////////////////////////////////////////////////////////
//
// Copyright (C) 2015-2017, Gisselquist Technology, LLC
//
// This program is free software (firmware): you can redistribute it and/or
// modify it under the terms of  the GNU General Public License as published
// by the Free Software Foundation, either version 3 of the License, or (at
// your option) any later version.
//
// This program is distributed in the hope that it will be useful, but WITHOUT
// ANY WARRANTY; without even the implied warranty of MERCHANTIBILITY or
// FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
// for more details.
//
// You should have received a copy of the GNU General Public License along
// with this program.  (It's in the $(ROOT)/doc directory.  Run make with no
// target there if the PDF file isn't present.)  If not, see
// <http://www.gnu.org/licenses/> for a copy.
//
// License:	GPL, v3, as defined and found on www.gnu.org,
//		http://www.gnu.org/licenses/gpl.html
//
//
////////////////////////////////////////////////////////////////////////////////
//
//
// `include "cpudefs.v"
//
module	div(i_clk, i_rst, i_wr, i_signed, i_numerator, i_denominator,
		o_busy, o_valid, o_err, o_quotient, o_flags);
	parameter		BW=32, LGBW = 5;
	input			i_clk, i_rst;
	// Input parameters
	input			i_wr, i_signed;
	input	[(BW-1):0]	i_numerator, i_denominator;
	// Output parameters
	output	reg		o_busy, o_valid, o_err;
	output	reg [(BW-1):0]	o_quotient;
	output	wire	[3:0]	o_flags;
 
	// r_busy is an internal busy register.  It will clear one clock
	// before we are valid, so it can't be o_busy ...
	//
	reg			r_busy;
	reg	[(2*BW-2):0]	r_divisor;
	reg	[(BW-1):0]	r_dividend;
	wire	[(BW):0]	diff; // , xdiff[(BW-1):0];
	assign	diff = r_dividend - r_divisor[(BW-1):0];
	// assign	xdiff= r_dividend - { 1'b0, r_divisor[(BW-1):1] };
 
	reg		r_sign, pre_sign, r_z, r_c, last_bit;
	reg	[(LGBW-1):0]	r_bit;
	reg	zero_divisor;
 
	// The Divide logic begins with r_busy.  We use r_busy to determine
	// whether or not the divide is in progress, vs being complete.
	// Here, we clear r_busy on any reset and set it on i_wr (the request
	// do to a divide).  The divide ends when we are on the last bit,
	// or equivalently when we discover we are dividing by zero.
	initial	r_busy = 1'b0;
	always @(posedge i_clk)
		if (i_rst)
			r_busy <= 1'b0;
		else if (i_wr)
			r_busy <= 1'b1;
		else if ((last_bit)||(zero_divisor))
			r_busy <= 1'b0;
 
	// o_busy is very similar to r_busy, save for some key differences.
	// Primary among them is that o_busy needs to (possibly) be true
	// for an extra clock after r_busy clears.  This would be that extra
	// clock where we negate the result (assuming a signed divide, and that
	// the result is supposed to be negative.)  Otherwise, the two are
	// identical.
	initial	o_busy = 1'b0;
	always @(posedge i_clk)
		if (i_rst)
			o_busy <= 1'b0;
		else if (i_wr)
			o_busy <= 1'b1;
		else if (((last_bit)&&(~r_sign))||(zero_divisor))
			o_busy <= 1'b0;
		else if (~r_busy)
			o_busy <= 1'b0;
 
	// If we are asked to divide by zero, we need to halt.  The sooner
	// we halt and report the error, the better.  Hence, here we look
	// for a zero divisor while being busy.  The always above us will then
	// look at this and halt a divide in the middle if we are trying to
	// divide by zero.
	//
	// Note that this works off of the 2BW-1 length vector.  If we can
	// simplify that, it should simplify our logic as well.
	initial	zero_divisor = 1'b0;
	always @(posedge i_clk)
		// zero_divisor <= (r_divisor == 0)&&(r_busy);
		if (i_rst)
			zero_divisor <= 1'b0;
		else if (i_wr)
			zero_divisor <= (i_denominator == 0);
		else if (!r_busy)
			zero_divisor <= 1'b0;
 
	// o_valid is part of the ZipCPU protocol.  It will be set to true
	// anytime our answer is valid and may be used by the calling module.
	// Indeed, the ZipCPU will halt (and ignore us) once the i_wr has been
	// set until o_valid gets set.
	//
	// Here, we clear o_valid on a reset, and any time we are on the last
	// bit while busy (provided the sign is zero, or we are dividing by
	// zero).  Since o_valid is self-clearing, we don't need to clear
	// it on an i_wr signal.
	initial	o_valid = 1'b0;
	always @(posedge i_clk)
		if (i_rst)
			o_valid <= 1'b0;
		else if (r_busy)
		begin
			if ((last_bit)||(zero_divisor))
				o_valid <= (zero_divisor)||(!r_sign);
		end else if (r_sign)
		begin
			o_valid <= (!zero_divisor); // 1'b1;
		end else
			o_valid <= 1'b0;
 
	// Division by zero error reporting.  Anytime we detect a zero divisor,
	// we set our output error, and then hold it until we are valid and
	// everything clears.
	initial	o_err = 1'b0;
	always @(posedge i_clk)
		if((i_rst)||(o_valid))
			o_err <= 1'b0;
		else if (((r_busy)||(r_sign))&&(zero_divisor))
			o_err <= 1'b1;
		else
			o_err <= 1'b0;
 
	// r_bit
	//
	// Keep track of which "bit" of our divide we are on.  This number
	// ranges from 31 down to zero.  On any write, we set ourselves to
	// 5'h1f.  Otherwise, while we are busy (but not within the pre-sign
	// adjustment stage), we subtract one from our value on every clock.
	always @(posedge i_clk)
		if ((r_busy)&&(!pre_sign))
			r_bit <= r_bit + {(LGBW){1'b1}};
		else
			r_bit <= {(LGBW){1'b1}};
 
	// last_bit
	//
	// This logic replaces a lot of logic that was inside our giant state
	// machine with ... something simpler.  In particular, we'll use this
	// logic to determine we are processing our last bit.  The only trick
	// is, this bit needs to be set whenever (r_busy) and (r_bit == 0),
	// hence we need to set on (r_busy) and (r_bit == 1) so as to be set
	// when (r_bit == 0).
	initial	last_bit = 1'b0;
	always @(posedge i_clk)
		if (r_busy)
			last_bit <= (r_bit == {{(LGBW-1){1'b0}},1'b1});
		else
			last_bit <= 1'b0;
 
	// pre_sign
	//
	// This is part of the state machine.  pre_sign indicates that we need
	// a extra clock to take the absolute value of our inputs.  It need only
	// be true for the one clock, and then it must clear itself.
	initial	pre_sign = 1'b0;
	always @(posedge i_clk)
		if (i_wr)
			pre_sign <= i_signed;
		else
			pre_sign <= 1'b0;
 
	// As a result of our operation, we need to set the flags.  The most
	// difficult of these is the "Z" flag indicating that the result is
	// zero.  Here, we'll use the same logic that sets the low-order
	// bit to clear our zero flag, and leave the zero flag set in all
	// other cases.  Well ... not quite.  If we need to flip the sign of
	// our value, then we can't quite clear the zero flag ... yet.
	always @(posedge i_clk)
		if((r_busy)&&(r_divisor[(2*BW-2):(BW)] == 0)&&(!diff[BW]))
			// If we are busy, the upper bits of our divisor are
			// zero (i.e., we got the shift right), and the top
			// (carry) bit of the difference is zero (no overflow),
			// then we could subtract our divisor from our dividend
			// and hence we add a '1' to the quotient, while setting
			// the zero flag to false.
			r_z <= 1'b0;
		else if ((!r_busy)&&(!r_sign))
			r_z <= 1'b1;
 
	// r_dividend
	// This is initially the numerator.  On a signed divide, it then becomes
	// the absolute value of the numerator.  We'll subtract from this value
	// the divisor shifted as appropriate for every output bit we are
	// looking for--just as with traditional long division.
	always @(posedge i_clk)
		if (pre_sign)
		begin
			// If we are doing a signed divide, then take the
			// absolute value of the dividend
			if (r_dividend[BW-1])
				r_dividend <= -r_dividend;
			// The begin/end block is important so we don't lose
			// the fact that on an else we don't do anything.
		end else if((r_busy)&&(r_divisor[(2*BW-2):(BW)]==0)&&(!diff[BW]))
			// This is the condition whereby we set a '1' in our
			// output quotient, and we subtract the (current)
			// divisor from our dividend.  (The difference is
			// already kept in the diff vector above.)
			r_dividend <= diff[(BW-1):0];
		else if (!r_busy)
			// Once we are done, and r_busy is no longer high, we'll
			// always accept new values into our dividend.  This
			// guarantees that, when i_wr is set, the new value
			// is already set as desired.
			r_dividend <=  i_numerator;
 
	initial	r_divisor = 0;
	always @(posedge i_clk)
		if (pre_sign)
		begin
			if (r_divisor[(2*BW-2)])
				r_divisor[(2*BW-2):(BW-1)]
					<= -r_divisor[(2*BW-2):(BW-1)];
		end else if (r_busy)
			r_divisor <= { 1'b0, r_divisor[(2*BW-2):1] };
		else
			r_divisor <= {  i_denominator, {(BW-1){1'b0}} };
 
	// r_sign
	// is a flag for our state machine control(s).  r_sign will be set to
	// true any time we are doing a signed divide and the result must be
	// negative.  In that case, we take a final logic stage at the end of
	// the divide to negate the output.  This flag is what tells us we need
	// to do that.  r_busy will be true during the divide, then when r_busy
	// goes low, r_sign will be checked, then the idle/reset stage will have
	// been reached.  For this reason, we cannot set r_sign unless we are
	// up to something.
	initial	r_sign = 1'b0;
	always @(posedge i_clk)
		if (pre_sign)
			r_sign <= ((r_divisor[(2*BW-2)])^(r_dividend[(BW-1)]));
		else if (r_busy)
			r_sign <= (r_sign)&&(!zero_divisor);
		else
			r_sign <= 1'b0;
 
	always @(posedge i_clk)
		if (r_busy)
		begin
			o_quotient <= { o_quotient[(BW-2):0], 1'b0 };
			if ((r_divisor[(2*BW-2):(BW)] == 0)&&(!diff[BW]))
			begin
				o_quotient[0] <= 1'b1;
			end
		end else if (r_sign)
			o_quotient <= -o_quotient;
		else
			o_quotient <= 0;
 
	// Set Carry on an exact divide
	// Perhaps nothing uses this, but ... well, I suppose we could remove
	// this logic eventually, just ... not yet.
	always @(posedge i_clk)
		r_c <= (r_busy)&&((diff == 0)||(r_dividend == 0));
 
	// The last flag: Negative.  This flag is set assuming that the result
	// of the divide was negative (i.e., the high order bit is set).  This
	// will also be true of an unsigned divide--if the high order bit is
	// ever set upon completion.  Indeed, you might argue that there's no
	// logic involved.
	wire	w_n;
	assign w_n = o_quotient[(BW-1)];
 
	assign o_flags = { 1'b0, w_n, r_c, r_z };
endmodule