1 |
75 |
robfinch |
// ============================================================================
|
2 |
|
|
// __
|
3 |
|
|
// \\__/ o\ (C) 2006-2022 Robert Finch, Waterloo
|
4 |
|
|
// \ __ / All rights reserved.
|
5 |
|
|
// \/_// robfinch@finitron.ca
|
6 |
|
|
// ||
|
7 |
|
|
//
|
8 |
|
|
// DFPRound96.sv
|
9 |
|
|
// - decimal floating point rounding unit
|
10 |
|
|
// - parameterized width
|
11 |
|
|
//
|
12 |
|
|
//
|
13 |
|
|
// BSD 3-Clause License
|
14 |
|
|
// Redistribution and use in source and binary forms, with or without
|
15 |
|
|
// modification, are permitted provided that the following conditions are met:
|
16 |
|
|
//
|
17 |
|
|
// 1. Redistributions of source code must retain the above copyright notice, this
|
18 |
|
|
// list of conditions and the following disclaimer.
|
19 |
|
|
//
|
20 |
|
|
// 2. Redistributions in binary form must reproduce the above copyright notice,
|
21 |
|
|
// this list of conditions and the following disclaimer in the documentation
|
22 |
|
|
// and/or other materials provided with the distribution.
|
23 |
|
|
//
|
24 |
|
|
// 3. Neither the name of the copyright holder nor the names of its
|
25 |
|
|
// contributors may be used to endorse or promote products derived from
|
26 |
|
|
// this software without specific prior written permission.
|
27 |
|
|
//
|
28 |
|
|
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
29 |
|
|
// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
30 |
|
|
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
31 |
|
|
// DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
32 |
|
|
// FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
33 |
|
|
// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
34 |
|
|
// SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
35 |
|
|
// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
36 |
|
|
// OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
37 |
|
|
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
38 |
|
|
//
|
39 |
|
|
// ============================================================================
|
40 |
|
|
|
41 |
|
|
import DFPPkg::*;
|
42 |
|
|
|
43 |
|
|
`ifdef MIN_LATENCY
|
44 |
|
|
`define PIPE_ADV *
|
45 |
|
|
`else
|
46 |
|
|
`define PIPE_ADV (posedge clk)
|
47 |
|
|
`endif
|
48 |
|
|
|
49 |
|
|
module DFPRound96(clk, ce, rm, i, o);
|
50 |
|
|
parameter N=25;
|
51 |
|
|
input clk;
|
52 |
|
|
input ce;
|
53 |
|
|
input [2:0] rm; // rounding mode
|
54 |
|
|
input DFP96UN i; // intermediate format input
|
55 |
|
|
output DFP96 o; // packed rounded output
|
56 |
|
|
|
57 |
|
|
parameter ROUND_CEILING = 3'd0;
|
58 |
|
|
parameter ROUND_FLOOR = 3'd1;
|
59 |
|
|
parameter ROUND_HALF_UP = 3'd2;
|
60 |
|
|
parameter ROUND_HALF_EVEN = 3'd3;
|
61 |
|
|
parameter ROUND_DOWN = 3'd4;
|
62 |
|
|
|
63 |
|
|
//------------------------------------------------------------
|
64 |
|
|
// variables
|
65 |
|
|
wire nano, qnano, snano;
|
66 |
|
|
wire infinity;
|
67 |
|
|
wire so;
|
68 |
|
|
wire [11:0] xo;
|
69 |
|
|
reg [N*4-1:0] mo;
|
70 |
|
|
reg [11:0] xo1;
|
71 |
|
|
reg [N*4-1:0] mo1;
|
72 |
|
|
wire xInf = i.exp==12'hBFF;
|
73 |
|
|
wire so0 = i.sign;
|
74 |
|
|
|
75 |
|
|
wire [3:0] l = i.sig[7:4];
|
76 |
|
|
wire [3:0] r = i.sig[3:0];
|
77 |
|
|
|
78 |
|
|
reg rnd;
|
79 |
|
|
|
80 |
|
|
//------------------------------------------------------------
|
81 |
|
|
// Clock #1
|
82 |
|
|
// - determine round amount (add 1 or 0)
|
83 |
|
|
//------------------------------------------------------------
|
84 |
|
|
|
85 |
|
|
always @`PIPE_ADV
|
86 |
|
|
if (ce) xo1 <= i.exp;
|
87 |
|
|
always @`PIPE_ADV
|
88 |
|
|
if (ce) mo1 <= i.sig[(N+1)*4-1:4];
|
89 |
|
|
|
90 |
|
|
// Compute the round bit
|
91 |
|
|
// Infinities and NaNs are not rounded!
|
92 |
|
|
always @`PIPE_ADV
|
93 |
|
|
if (ce)
|
94 |
|
|
if (i.nan | i.infinity)
|
95 |
|
|
rnd = 1'b0;
|
96 |
|
|
else
|
97 |
|
|
case (rm)
|
98 |
|
|
ROUND_CEILING: rnd <= (r == 4'd0 || i.sign==1'b1) ? 1'b0 : 1'b1;
|
99 |
|
|
ROUND_FLOOR: rnd <= (r == 4'd0 || i.sign==1'b0) ? 1'b0 : 1'b1;
|
100 |
|
|
ROUND_HALF_UP: rnd <= r >= 4'h5;
|
101 |
|
|
ROUND_HALF_EVEN: rnd <= r==4'h5 ? l[0] : r > 4'h5 ? 1'b1 : 1'b0;
|
102 |
|
|
ROUND_DOWN: rnd <= 1'b0;
|
103 |
|
|
default: rnd <= 1'b0;
|
104 |
|
|
endcase
|
105 |
|
|
|
106 |
|
|
//------------------------------------------------------------
|
107 |
|
|
// Clock #2
|
108 |
|
|
// round the number, check for carry
|
109 |
|
|
// note: inf. exponent checked above (if the exponent was infinite already, then no rounding occurs as rnd = 0)
|
110 |
|
|
// note: exponent increments if there is a carry (can only increment to infinity)
|
111 |
|
|
//------------------------------------------------------------
|
112 |
|
|
|
113 |
|
|
wire [N*4-1:0] rounded1;
|
114 |
|
|
wire cobcd;
|
115 |
|
|
|
116 |
|
|
BCDAddN #(.N(N)) ubcdan1
|
117 |
|
|
(
|
118 |
|
|
.ci(1'b0),
|
119 |
|
|
.a(mo1),
|
120 |
|
|
.b({{N*4-1{1'd0}},rnd}),
|
121 |
|
|
.o(rounded1),
|
122 |
|
|
.co(cobcd)
|
123 |
|
|
);
|
124 |
|
|
|
125 |
|
|
reg [N*4-1:0] rounded2;
|
126 |
|
|
reg rnd2;
|
127 |
|
|
reg dn2;
|
128 |
|
|
reg [12:0] xo2;
|
129 |
|
|
always @`PIPE_ADV
|
130 |
|
|
if (ce) rounded2 <= rounded1;
|
131 |
|
|
always @`PIPE_ADV
|
132 |
|
|
if (ce) rnd2 <= rnd;
|
133 |
|
|
always @`PIPE_ADV
|
134 |
|
|
if (ce) dn2 <= !(|xo1);
|
135 |
|
|
always @`PIPE_ADV
|
136 |
|
|
if (ce) xo2 <= xo1 + cobcd;
|
137 |
|
|
|
138 |
|
|
//------------------------------------------------------------
|
139 |
|
|
// Clock #3
|
140 |
|
|
// - shift mantissa if required.
|
141 |
|
|
//------------------------------------------------------------
|
142 |
|
|
wire infinity2;
|
143 |
|
|
`ifdef MIN_LATENCY
|
144 |
|
|
assign nano = i.nan;
|
145 |
|
|
assign qnano = i.qnan;
|
146 |
|
|
assign snano = i.snan;
|
147 |
|
|
assign infinity = i.infinity | (rnd2 && xo2[11:0]==12'hBFF);
|
148 |
|
|
assign so = i.sign;
|
149 |
|
|
assign xo = xo2[11:0];
|
150 |
|
|
`else
|
151 |
|
|
delay3 #(1) u21 (.clk(clk), .ce(ce), .i(i.nan), .o(nano));
|
152 |
|
|
delay3 #(1) u22 (.clk(clk), .ce(ce), .i(i.qnan), .o(qnano));
|
153 |
|
|
delay3 #(1) u23 (.clk(clk), .ce(ce), .i(i.snan), .o(snano));
|
154 |
|
|
delay2 #(1) u24 (.clk(clk), .ce(ce), .i(i.infinity), .o(infinity2));
|
155 |
|
|
delay3 #(1) u25 (.clk(clk), .ce(ce), .i(i.sign), .o(so));
|
156 |
|
|
delay1 #(12) u26 (.clk(clk), .ce(ce), .i(xo2[13:0]), .o(xo));
|
157 |
|
|
delay1 #(1) u27 (.clk(clk), .ce(ce), .i(infinity2 | (rnd2 && xo2[11:0]==12'hBFF)), .o(infinity));
|
158 |
|
|
`endif
|
159 |
|
|
|
160 |
|
|
wire carry2 = xo2[12];
|
161 |
|
|
|
162 |
|
|
always @`PIPE_ADV
|
163 |
|
|
if (ce)
|
164 |
|
|
casez({rnd2,xo2[11:0]==12'hBFF,carry2,dn2})
|
165 |
|
|
4'b0??0: mo <= mo1[N*4-1:0]; // not rounding, not denormalized
|
166 |
|
|
4'b0??1: mo <= mo1[N*4-1:0]; // not rounding, denormalized
|
167 |
|
|
4'b1000: mo <= rounded2[N*4-1: 0]; // exponent didn't change, number was normalized
|
168 |
|
|
4'b1001: mo <= rounded2[N*4-1: 0]; // exponent didn't change, but number was denormalized
|
169 |
|
|
4'b1010: mo <= {4'h1,rounded2[N*4-1: 4]}; // exponent incremented (new MSD generated), number was normalized
|
170 |
|
|
4'b1011: mo <= rounded2[N*4-1:0]; // exponent incremented (new MSB generated), number was denormalized, number became normalized
|
171 |
|
|
4'b11??: mo <= {N*4{1'd0}}; // number became infinite, no need to check carry etc., rnd would be zero if input was NaN or infinite
|
172 |
|
|
endcase
|
173 |
|
|
|
174 |
|
|
//------------------------------------------------------------
|
175 |
|
|
// Clock #4
|
176 |
|
|
// - Pack output
|
177 |
|
|
//------------------------------------------------------------
|
178 |
|
|
|
179 |
|
|
DFP96U o1;
|
180 |
|
|
DFP96 o2;
|
181 |
|
|
|
182 |
|
|
assign o1.nan = nano;
|
183 |
|
|
assign o1.qnan = qnano;
|
184 |
|
|
assign o1.snan = snano;
|
185 |
|
|
assign o1.infinity = infinity;
|
186 |
|
|
assign o1.sign = so;
|
187 |
|
|
assign o1.exp = xo;
|
188 |
|
|
assign o1.sig = mo;
|
189 |
|
|
|
190 |
|
|
DFPPack96 u41 (o1, o2);
|
191 |
|
|
always_ff @(posedge clk)
|
192 |
|
|
if (ce) o <= o2;
|
193 |
|
|
|
194 |
|
|
endmodule
|