1 |
29 |
robfinch |
// ============================================================================
|
2 |
|
|
// __
|
3 |
|
|
// \\__/ o\ (C) 2006-2019 Robert Finch, Waterloo
|
4 |
|
|
// \ __ / All rights reserved.
|
5 |
|
|
// \/_// robfinch<remove>@finitron.ca
|
6 |
|
|
// ||
|
7 |
|
|
//
|
8 |
|
|
// fpRound.v
|
9 |
|
|
// - floating point rounding unit
|
10 |
|
|
// - parameterized FPWIDth
|
11 |
|
|
// - IEEE 754 representation
|
12 |
|
|
//
|
13 |
|
|
//
|
14 |
|
|
// This source file is free software: you can redistribute it and/or modify
|
15 |
|
|
// it under the terms of the GNU Lesser General Public License as published
|
16 |
|
|
// by the Free Software Foundation, either version 3 of the License, or
|
17 |
|
|
// (at your option) any later version.
|
18 |
|
|
//
|
19 |
|
|
// This source file is distributed in the hope that it will be useful,
|
20 |
|
|
// but WITHOUT ANY WARRANTY; without even the implied warranty of
|
21 |
|
|
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
22 |
|
|
// GNU General Public License for more details.
|
23 |
|
|
//
|
24 |
|
|
// You should have received a copy of the GNU General Public License
|
25 |
|
|
// along with this program. If not, see <http://www.gnu.org/licenses/>.
|
26 |
|
|
//
|
27 |
|
|
// ============================================================================
|
28 |
|
|
|
29 |
|
|
`include "fpConfig.sv"
|
30 |
|
|
|
31 |
|
|
module fpRound(clk, ce, rm, i, o);
|
32 |
|
|
parameter FPWID = 128;
|
33 |
|
|
`include "fpSize.sv"
|
34 |
|
|
input clk;
|
35 |
|
|
input ce;
|
36 |
|
|
input [2:0] rm; // rounding mode
|
37 |
|
|
input [MSB+3:0] i; // intermediate format input
|
38 |
|
|
output [MSB:0] o; // rounded output
|
39 |
|
|
|
40 |
|
|
//------------------------------------------------------------
|
41 |
|
|
// variables
|
42 |
|
|
wire so;
|
43 |
|
|
wire [EMSB:0] xo;
|
44 |
|
|
reg [FMSB:0] mo;
|
45 |
|
|
reg [EMSB:0] xo1;
|
46 |
|
|
reg [FMSB+3:0] mo1;
|
47 |
|
|
wire xInf = &i[MSB+2:FMSB+4];
|
48 |
|
|
wire so0 = i[MSB+3];
|
49 |
|
|
assign o = {so,xo,mo};
|
50 |
|
|
|
51 |
|
|
wire g = i[2]; // guard bit: always the same bit for all operations
|
52 |
|
|
wire r = i[1]; // rounding bit
|
53 |
|
|
wire s = i[0]; // sticky bit
|
54 |
|
|
reg rnd;
|
55 |
|
|
|
56 |
|
|
//------------------------------------------------------------
|
57 |
|
|
// Clock #1
|
58 |
|
|
// - determine round amount (add 1 or 0)
|
59 |
|
|
//------------------------------------------------------------
|
60 |
|
|
|
61 |
|
|
`ifdef MIN_LATENCY
|
62 |
|
|
always @*
|
63 |
|
|
`else
|
64 |
|
|
always @(posedge clk)
|
65 |
|
|
`endif
|
66 |
|
|
if (ce) xo1 <= i[MSB+2:FMSB+4];
|
67 |
|
|
`ifdef MIN_LATENCY
|
68 |
|
|
always @*
|
69 |
|
|
`else
|
70 |
|
|
always @(posedge clk)
|
71 |
|
|
`endif
|
72 |
|
|
if (ce) mo1 <= i[FMSB+3:0];
|
73 |
|
|
|
74 |
|
|
// Compute the round bit
|
75 |
|
|
// Infinities and NaNs are not rounded!
|
76 |
|
|
`ifdef MIN_LATENCY
|
77 |
|
|
always @*
|
78 |
|
|
`else
|
79 |
|
|
always @(posedge clk)
|
80 |
|
|
`endif
|
81 |
|
|
if (ce)
|
82 |
|
|
casez ({xInf,rm})
|
83 |
|
|
4'b0000: rnd <= (g & r) | (r & s); // round to nearest even
|
84 |
|
|
4'b0001: rnd <= 1'd0; // round to zero (truncate)
|
85 |
|
|
4'b0010: rnd <= (r | s) & !so0; // round towards +infinity
|
86 |
|
|
4'b0011: rnd <= (r | s) & so0; // round towards -infinity
|
87 |
|
|
4'b0100: rnd <= (r | s); // round to nearest away from zero
|
88 |
|
|
4'b1???: rnd <= 1'd0; // no rounding if exponent indicates infinite or NaN
|
89 |
|
|
default: rnd <= 0;
|
90 |
|
|
endcase
|
91 |
|
|
|
92 |
|
|
//------------------------------------------------------------
|
93 |
|
|
// Clock #2
|
94 |
|
|
// round the number, check for carry
|
95 |
|
|
// note: inf. exponent checked above (if the exponent was infinite already, then no rounding occurs as rnd = 0)
|
96 |
|
|
// note: exponent increments if there is a carry (can only increment to infinity)
|
97 |
|
|
//------------------------------------------------------------
|
98 |
|
|
|
99 |
|
|
reg [MSB:0] rounded2;
|
100 |
|
|
reg carry2;
|
101 |
|
|
reg rnd2;
|
102 |
|
|
reg dn2;
|
103 |
|
|
wire [EMSB:0] xo2;
|
104 |
|
|
wire [MSB:0] rounded1 = {xo1,mo1[FMSB+3:2]} + rnd;
|
105 |
|
|
`ifdef MIN_LATENCY
|
106 |
|
|
always @*
|
107 |
|
|
`else
|
108 |
|
|
always @(posedge clk)
|
109 |
|
|
`endif
|
110 |
|
|
if (ce) rounded2 <= rounded1;
|
111 |
|
|
`ifdef MIN_LATENCY
|
112 |
|
|
always @*
|
113 |
|
|
`else
|
114 |
|
|
always @(posedge clk)
|
115 |
|
|
`endif
|
116 |
|
|
if (ce) carry2 <= mo1[FMSB+3] & !rounded1[FMSB+1];
|
117 |
|
|
`ifdef MIN_LATENCY
|
118 |
|
|
always @*
|
119 |
|
|
`else
|
120 |
|
|
always @(posedge clk)
|
121 |
|
|
`endif
|
122 |
|
|
if (ce) rnd2 <= rnd;
|
123 |
|
|
`ifdef MIN_LATENCY
|
124 |
|
|
always @*
|
125 |
|
|
`else
|
126 |
|
|
always @(posedge clk)
|
127 |
|
|
`endif
|
128 |
|
|
if (ce) dn2 <= !(|xo1);
|
129 |
|
|
assign xo2 = rounded2[MSB:FMSB+2];
|
130 |
|
|
|
131 |
|
|
//------------------------------------------------------------
|
132 |
|
|
// Clock #3
|
133 |
|
|
// - shift mantissa if required.
|
134 |
|
|
//------------------------------------------------------------
|
135 |
|
|
`ifdef MIN_LATENCY
|
136 |
|
|
assign so = i[MSB+3];
|
137 |
|
|
assign xo = xo2;
|
138 |
|
|
`else
|
139 |
|
|
delay3 #(1) u21 (.clk(clk), .ce(ce), .i(i[MSB+3]), .o(so));
|
140 |
|
|
delay1 #(EMSB+1) u22 (.clk(clk), .ce(ce), .i(xo2), .o(xo));
|
141 |
|
|
`endif
|
142 |
|
|
|
143 |
|
|
`ifdef MIN_LATENCY
|
144 |
|
|
always @*
|
145 |
|
|
`else
|
146 |
|
|
always @(posedge clk)
|
147 |
|
|
`endif
|
148 |
|
|
casez({rnd2,&xo2,carry2,dn2})
|
149 |
|
|
4'b0??0: mo <= mo1[FMSB+2:2]; // not rounding, not denormalized, => hide MSB
|
150 |
|
|
4'b0??1: mo <= mo1[FMSB+3:3]; // not rounding, denormalized
|
151 |
|
|
4'b1000: mo <= rounded2[FMSB :0]; // exponent didn't change, number was normalized, => hide MSB,
|
152 |
|
|
4'b1001: mo <= rounded2[FMSB+1:1]; // exponent didn't change, but number was denormalized, => retain MSB
|
153 |
|
|
4'b1010: mo <= rounded2[FMSB+1:1]; // exponent incremented (new MSB generated), number was normalized, => hide 'extra (FMSB+2)' MSB
|
154 |
|
|
4'b1011: mo <= rounded2[FMSB+1:1]; // exponent incremented (new MSB generated), number was denormalized, number became normalized, => hide 'extra (FMSB+2)' MSB
|
155 |
|
|
4'b11??: mo <= 1'd0; // number became infinite, no need to check carry etc., rnd would be zero if input was NaN or infinite
|
156 |
|
|
endcase
|
157 |
|
|
|
158 |
|
|
endmodule
|
159 |
|
|
|
160 |
|
|
|
161 |
|
|
// Round and register the output
|
162 |
|
|
/*
|
163 |
|
|
module fpRoundReg(clk, ce, rm, i, o);
|
164 |
|
|
parameter FPWID = 128;
|
165 |
|
|
`include "fpSize.sv"
|
166 |
|
|
|
167 |
|
|
input clk;
|
168 |
|
|
input ce;
|
169 |
|
|
input [2:0] rm; // rounding mode
|
170 |
|
|
input [MSB+3:0] i; // expanded format input
|
171 |
|
|
output reg [FPWID-1:0] o; // rounded output
|
172 |
|
|
|
173 |
|
|
wire [FPWID-1:0] o1;
|
174 |
|
|
fpRound #(FPWID) u1 (.rm(rm), .i(i), .o(o1) );
|
175 |
|
|
|
176 |
|
|
always @(posedge clk)
|
177 |
|
|
if (ce)
|
178 |
|
|
o <= o1;
|
179 |
|
|
|
180 |
|
|
endmodule
|
181 |
|
|
*/
|