Line 1... |
Line 1... |
///////////////////////////////////////////////////////////////////////////////
|
////////////////////////////////////////////////////////////////////////////////
|
//
|
//
|
// Filename: div.v
|
// Filename: div.v
|
//
|
//
|
// Project: Zip CPU -- a small, lightweight, RISC CPU soft core
|
// Project: Zip CPU -- a small, lightweight, RISC CPU soft core
|
//
|
//
|
// Purpose: Provide an Integer divide capability to the Zip CPU.
|
// 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.
|
// Creator: Dan Gisselquist, Ph.D.
|
// Gisselquist Technology, LLC
|
// Gisselquist Technology, LLC
|
//
|
//
|
///////////////////////////////////////////////////////////////////////////////
|
////////////////////////////////////////////////////////////////////////////////
|
//
|
//
|
// Copyright (C) 2015-2016, Gisselquist Technology, LLC
|
// Copyright (C) 2015-2017, Gisselquist Technology, LLC
|
//
|
//
|
// This program is free software (firmware): you can redistribute it and/or
|
// 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
|
// 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
|
// by the Free Software Foundation, either version 3 of the License, or (at
|
// your option) any later version.
|
// your option) any later version.
|
Line 22... |
Line 81... |
// This program is distributed in the hope that it will be useful, but WITHOUT
|
// This program is distributed in the hope that it will be useful, but WITHOUT
|
// ANY WARRANTY; without even the implied warranty of MERCHANTIBILITY or
|
// ANY WARRANTY; without even the implied warranty of MERCHANTIBILITY or
|
// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
// for more details.
|
// 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,
|
// License: GPL, v3, as defined and found on www.gnu.org,
|
// http://www.gnu.org/licenses/gpl.html
|
// http://www.gnu.org/licenses/gpl.html
|
//
|
//
|
//
|
//
|
///////////////////////////////////////////////////////////////////////////////
|
////////////////////////////////////////////////////////////////////////////////
|
|
//
|
//
|
//
|
// `include "cpudefs.v"
|
// `include "cpudefs.v"
|
//
|
//
|
module div(i_clk, i_rst, i_wr, i_signed, i_numerator, i_denominator,
|
module div(i_clk, i_rst, i_wr, i_signed, i_numerator, i_denominator,
|
o_busy, o_valid, o_err, o_quotient, o_flags);
|
o_busy, o_valid, o_err, o_quotient, o_flags);
|
Line 54... |
Line 119... |
assign diff = r_dividend - r_divisor[(BW-1):0];
|
assign diff = r_dividend - r_divisor[(BW-1):0];
|
// assign xdiff= r_dividend - { 1'b0, r_divisor[(BW-1):1] };
|
// assign xdiff= r_dividend - { 1'b0, r_divisor[(BW-1):1] };
|
|
|
reg r_sign, pre_sign, r_z, r_c, last_bit;
|
reg r_sign, pre_sign, r_z, r_c, last_bit;
|
reg [(LGBW-1):0] r_bit;
|
reg [(LGBW-1):0] r_bit;
|
|
|
reg zero_divisor;
|
reg zero_divisor;
|
initial zero_divisor = 1'b0;
|
|
always @(posedge i_clk)
|
|
zero_divisor <= (r_divisor == 0)&&(r_busy);
|
|
|
|
|
// 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;
|
initial r_busy = 1'b0;
|
always @(posedge i_clk)
|
always @(posedge i_clk)
|
if (i_rst)
|
if (i_rst)
|
r_busy <= 1'b0;
|
r_busy <= 1'b0;
|
else if (i_wr)
|
else if (i_wr)
|
r_busy <= 1'b1;
|
r_busy <= 1'b1;
|
else if ((last_bit)||(zero_divisor))
|
else if ((last_bit)||(zero_divisor))
|
r_busy <= 1'b0;
|
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;
|
initial o_busy = 1'b0;
|
always @(posedge i_clk)
|
always @(posedge i_clk)
|
if (i_rst)
|
if (i_rst)
|
o_busy <= 1'b0;
|
o_busy <= 1'b0;
|
else if (i_wr)
|
else if (i_wr)
|
Line 80... |
Line 152... |
else if (((last_bit)&&(~r_sign))||(zero_divisor))
|
else if (((last_bit)&&(~r_sign))||(zero_divisor))
|
o_busy <= 1'b0;
|
o_busy <= 1'b0;
|
else if (~r_busy)
|
else if (~r_busy)
|
o_busy <= 1'b0;
|
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)
|
always @(posedge i_clk)
|
if ((i_rst)||(i_wr))
|
if (i_rst)
|
o_valid <= 1'b0;
|
o_valid <= 1'b0;
|
else if (r_busy)
|
else if (r_busy)
|
begin
|
begin
|
if ((last_bit)||(zero_divisor))
|
if ((last_bit)||(zero_divisor))
|
o_valid <= (zero_divisor)||(~r_sign);
|
o_valid <= (zero_divisor)||(!r_sign);
|
end else if (r_sign)
|
end else if (r_sign)
|
begin
|
begin
|
o_valid <= (~zero_divisor); // 1'b1;
|
o_valid <= (!zero_divisor); // 1'b1;
|
end else
|
end else
|
o_valid <= 1'b0;
|
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;
|
initial o_err = 1'b0;
|
always @(posedge i_clk)
|
always @(posedge i_clk)
|
if((i_rst)||(o_valid))
|
if((i_rst)||(o_valid))
|
o_err <= 1'b0;
|
o_err <= 1'b0;
|
else if (((r_busy)||(r_sign))&&(zero_divisor))
|
else if (((r_busy)||(r_sign))&&(zero_divisor))
|
o_err <= 1'b1;
|
o_err <= 1'b1;
|
else
|
else
|
o_err <= 1'b0;
|
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;
|
initial last_bit = 1'b0;
|
always @(posedge i_clk)
|
always @(posedge i_clk)
|
if ((i_wr)||(pre_sign)||(i_rst))
|
if (r_busy)
|
last_bit <= 1'b0;
|
|
else if (r_busy)
|
|
last_bit <= (r_bit == {{(LGBW-1){1'b0}},1'b1});
|
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)
|
always @(posedge i_clk)
|
// if (i_rst) r_busy <= 1'b0;
|
|
// else
|
|
if (i_wr)
|
if (i_wr)
|
begin
|
|
//
|
|
// Set our values upon an initial command. Here's
|
|
// where we come in and start.
|
|
//
|
|
// r_busy <= 1'b1;
|
|
//
|
|
o_quotient <= 0;
|
|
r_bit <= {(LGBW){1'b1}};
|
|
r_divisor <= { i_denominator, {(BW-1){1'b0}} };
|
|
r_dividend <= i_numerator;
|
|
r_sign <= 1'b0;
|
|
pre_sign <= i_signed;
|
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_z <= 1'b1;
|
end else if (pre_sign)
|
|
|
// 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
|
begin
|
//
|
// If we are doing a signed divide, then take the
|
// Note that we only come in here, for one clock, if
|
// absolute value of the dividend
|
// our initial value may have been signed. If we are
|
|
// doing an unsigned divide, we then skip this step.
|
|
//
|
|
r_sign <= ((r_divisor[(2*BW-2)])^(r_dividend[(BW-1)]));
|
|
// Negate our dividend if necessary so that it becomes
|
|
// a magnitude only value
|
|
if (r_dividend[BW-1])
|
if (r_dividend[BW-1])
|
r_dividend <= -r_dividend;
|
r_dividend <= -r_dividend;
|
// Do the same with the divisor--rendering it into
|
// The begin/end block is important so we don't lose
|
// a magnitude only.
|
// 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)])
|
if (r_divisor[(2*BW-2)])
|
r_divisor[(2*BW-2):(BW-1)] <= -r_divisor[(2*BW-2):(BW-1)];
|
r_divisor[(2*BW-2):(BW-1)]
|
//
|
<= -r_divisor[(2*BW-2):(BW-1)];
|
// We only do this stage for a single clock, so go on
|
|
// with the rest of the divide otherwise.
|
|
pre_sign <= 1'b0;
|
|
end else if (r_busy)
|
end else if (r_busy)
|
begin
|
|
// While the divide is taking place, we examine each bit
|
|
// in turn here.
|
|
//
|
|
r_bit <= r_bit + {(LGBW){1'b1}}; // r_bit = r_bit - 1;
|
|
r_divisor <= { 1'b0, r_divisor[(2*BW-2):1] };
|
r_divisor <= { 1'b0, r_divisor[(2*BW-2):1] };
|
if (|r_divisor[(2*BW-2):(BW)])
|
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
|
begin
|
end else if (diff[BW])
|
o_quotient <= { o_quotient[(BW-2):0], 1'b0 };
|
|
if ((r_divisor[(2*BW-2):(BW)] == 0)&&(!diff[BW]))
|
begin
|
begin
|
//
|
o_quotient[0] <= 1'b1;
|
// diff = r_dividend - r_divisor[(BW-1):0];
|
|
//
|
|
// If this value was negative, there wasn't
|
|
// enough value in the dividend to support
|
|
// pulling off a bit. We'll move down a bit
|
|
// therefore and try again.
|
|
//
|
|
end else begin
|
|
//
|
|
// Put a '1' into our output accumulator.
|
|
// Subtract the divisor from the dividend,
|
|
// and then move on to the next bit
|
|
//
|
|
r_dividend <= diff[(BW-1):0];
|
|
o_quotient[r_bit[(LGBW-1):0]] <= 1'b1;
|
|
r_z <= 1'b0;
|
|
end
|
end
|
r_sign <= (r_sign)&&(~zero_divisor);
|
|
end else if (r_sign)
|
end else if (r_sign)
|
begin
|
|
r_sign <= 1'b0;
|
|
o_quotient <= -o_quotient;
|
o_quotient <= -o_quotient;
|
end
|
else
|
|
o_quotient <= 0;
|
|
|
// Set Carry on an exact divide
|
// Set Carry on an exact divide
|
wire w_n;
|
// Perhaps nothing uses this, but ... well, I suppose we could remove
|
|
// this logic eventually, just ... not yet.
|
always @(posedge i_clk)
|
always @(posedge i_clk)
|
r_c <= (r_busy)&&((diff == 0)||(r_dividend == 0));
|
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 w_n = o_quotient[(BW-1)];
|
|
|
assign o_flags = { 1'b0, w_n, r_c, r_z };
|
assign o_flags = { 1'b0, w_n, r_c, r_z };
|
endmodule
|
endmodule
|
|
|
No newline at end of file
|
No newline at end of file
|