1 |
6 |
robfinch |
// ============================================================================
|
2 |
|
|
// __
|
3 |
|
|
// \\__/ o\ (C) 2006-2016 Robert Finch, Stratford
|
4 |
|
|
// \ __ / All rights reserved.
|
5 |
|
|
// \/_// robfinch<remove>@finitron.ca
|
6 |
|
|
// ||
|
7 |
|
|
//
|
8 |
|
|
// This source file is free software: you can redistribute it and/or modify
|
9 |
|
|
// it under the terms of the GNU Lesser General Public License as published
|
10 |
|
|
// by the Free Software Foundation, either version 3 of the License, or
|
11 |
|
|
// (at your option) any later version.
|
12 |
|
|
//
|
13 |
|
|
// This source file is distributed in the hope that it will be useful,
|
14 |
|
|
// but WITHOUT ANY WARRANTY; without even the implied warranty of
|
15 |
|
|
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
16 |
|
|
// GNU General Public License for more details.
|
17 |
|
|
//
|
18 |
|
|
// You should have received a copy of the GNU General Public License
|
19 |
|
|
// along with this program. If not, see <http://www.gnu.org/licenses/>.
|
20 |
|
|
//
|
21 |
|
|
// fpDiv.v
|
22 |
|
|
// - floating point divider
|
23 |
|
|
// - parameterized width
|
24 |
|
|
// - IEEE 754 representation
|
25 |
|
|
//
|
26 |
|
|
// Floating Point Divider
|
27 |
|
|
//
|
28 |
|
|
// Properties:
|
29 |
|
|
// +-0 / +-0 = QNaN
|
30 |
|
|
//
|
31 |
|
|
// ============================================================================
|
32 |
|
|
//
|
33 |
|
|
module fpDiv(clk, ce, ld, a, b, o, done, sign_exe, overflow, underflow);
|
34 |
|
|
|
35 |
|
|
parameter WID = 32;
|
36 |
|
|
localparam MSB = WID-1;
|
37 |
|
|
localparam EMSB = WID==80 ? 14 :
|
38 |
|
|
WID==64 ? 10 :
|
39 |
|
|
WID==52 ? 10 :
|
40 |
|
|
WID==48 ? 10 :
|
41 |
|
|
WID==44 ? 10 :
|
42 |
|
|
WID==42 ? 10 :
|
43 |
|
|
WID==40 ? 9 :
|
44 |
|
|
WID==32 ? 7 :
|
45 |
|
|
WID==24 ? 6 : 4;
|
46 |
|
|
localparam FMSB = WID==80 ? 63 :
|
47 |
|
|
WID==64 ? 51 :
|
48 |
|
|
WID==52 ? 39 :
|
49 |
|
|
WID==48 ? 35 :
|
50 |
|
|
WID==44 ? 31 :
|
51 |
|
|
WID==42 ? 29 :
|
52 |
|
|
WID==40 ? 28 :
|
53 |
|
|
WID==32 ? 22 :
|
54 |
|
|
WID==24 ? 15 : 9;
|
55 |
|
|
|
56 |
|
|
localparam WX = 3;
|
57 |
|
|
localparam FX = (FMSB+1)*2-1; // the MSB of the expanded fraction
|
58 |
|
|
localparam EX = FX + WX + EMSB + 1;
|
59 |
|
|
|
60 |
|
|
input clk;
|
61 |
|
|
input ce;
|
62 |
|
|
input ld;
|
63 |
|
|
input [MSB:0] a, b;
|
64 |
|
|
output [EX+1:0] o;
|
65 |
|
|
output done;
|
66 |
|
|
output sign_exe;
|
67 |
|
|
output overflow;
|
68 |
|
|
output underflow;
|
69 |
|
|
|
70 |
|
|
// registered outputs
|
71 |
|
|
reg sign_exe;
|
72 |
|
|
reg inf;
|
73 |
|
|
reg overflow;
|
74 |
|
|
reg underflow;
|
75 |
|
|
|
76 |
|
|
reg so;
|
77 |
|
|
reg [EMSB:0] xo;
|
78 |
|
|
reg [FX+WX:0] mo;
|
79 |
|
|
assign o = {so,xo,mo};
|
80 |
|
|
|
81 |
|
|
// constants
|
82 |
|
|
wire [EMSB:0] infXp = {EMSB+1{1'b1}}; // infinite / NaN - all ones
|
83 |
|
|
// The following is the value for an exponent of zero, with the offset
|
84 |
|
|
// eg. 8'h7f for eight bit exponent, 11'h7ff for eleven bit exponent, etc.
|
85 |
|
|
wire [EMSB:0] bias = {1'b0,{EMSB{1'b1}}}; //2^0 exponent
|
86 |
|
|
// The following is a template for a quiet nan. (MSB=1)
|
87 |
|
|
wire [FMSB:0] qNaN = {1'b1,{FMSB{1'b0}}};
|
88 |
|
|
|
89 |
|
|
// variables
|
90 |
|
|
wire [EMSB+2:0] ex1; // sum of exponents
|
91 |
|
|
wire [FX+WX:0] divo;
|
92 |
|
|
|
93 |
|
|
// Operands
|
94 |
|
|
wire sa, sb; // sign bit
|
95 |
|
|
wire [EMSB:0] xa, xb; // exponent bits
|
96 |
|
|
wire [FMSB+1:0] fracta, fractb;
|
97 |
|
|
wire a_dn, b_dn; // a/b is denormalized
|
98 |
|
|
wire az, bz;
|
99 |
|
|
wire aInf, bInf;
|
100 |
|
|
|
101 |
|
|
|
102 |
|
|
// -----------------------------------------------------------
|
103 |
|
|
// - decode the input operands
|
104 |
|
|
// - derive basic information
|
105 |
|
|
// - calculate exponent
|
106 |
|
|
// - calculate fraction
|
107 |
|
|
// -----------------------------------------------------------
|
108 |
|
|
|
109 |
|
|
fpDecompose #(WID) u1a (.i(a), .sgn(sa), .exp(xa), .fract(fracta), .xz(a_dn), .vz(az), .inf(aInf) );
|
110 |
|
|
fpDecompose #(WID) u1b (.i(b), .sgn(sb), .exp(xb), .fract(fractb), .xz(b_dn), .vz(bz), .inf(bInf) );
|
111 |
|
|
|
112 |
|
|
// Compute the exponent.
|
113 |
|
|
// - correct the exponent for denormalized operands
|
114 |
|
|
// - adjust the difference by the bias (add 127)
|
115 |
|
|
// - also factor in the different decimal position for division
|
116 |
|
|
assign ex1 = (xa|a_dn) - (xb|b_dn) + bias + FMSB + 1;
|
117 |
|
|
|
118 |
|
|
// check for exponent underflow/overflow
|
119 |
|
|
wire under = ex1[EMSB+2]; // MSB set = negative exponent
|
120 |
|
|
wire over = (&ex1[EMSB:0] | ex1[EMSB+1]) & !ex1[EMSB+2];
|
121 |
|
|
|
122 |
|
|
// Perform divide
|
123 |
|
|
// could take either 1 or 16 clock cycles
|
124 |
|
|
fpdivr2 #(FMSB+2) u2 (.clk(clk), .ld(ld), .a(fracta), .b(fractb), .q(divo[(FMSB+1)*2-1:0]), .r(), .done(done));
|
125 |
|
|
assign divo[FX+WX:(FMSB+1)*2] = 0;
|
126 |
|
|
|
127 |
|
|
// determine when a NaN is output
|
128 |
|
|
wire qNaNOut = (az&bz)|(aInf&bInf);
|
129 |
|
|
|
130 |
|
|
always @(posedge clk)
|
131 |
|
|
if (ce) begin
|
132 |
|
|
if (done) begin
|
133 |
|
|
casex({qNaNOut,bInf,bz})
|
134 |
|
|
3'b1xx: xo = infXp; // NaN exponent value
|
135 |
|
|
3'bx1x: xo = 0; // divide by inf
|
136 |
|
|
3'bxx1: xo = infXp; // divide by zero
|
137 |
|
|
default: xo = ex1; // normal or underflow: passthru neg. exp. for normalization
|
138 |
|
|
endcase
|
139 |
|
|
|
140 |
|
|
casex({qNaNOut,bInf,bz})
|
141 |
|
|
3'b1xx: mo = {1'b0,qNaN[FMSB:0]|{aInf,1'b0}|{az,bz},{FMSB+1{1'b0}}};
|
142 |
|
|
3'bx1x: mo = 0; // div by inf
|
143 |
|
|
3'bxx1: mo = 0; // div by zero
|
144 |
|
|
default: mo = divo; // plain div
|
145 |
|
|
endcase
|
146 |
|
|
|
147 |
|
|
so = sa ^ sb;
|
148 |
|
|
sign_exe = sa & sb;
|
149 |
|
|
overflow = over;
|
150 |
|
|
underflow = under;
|
151 |
|
|
end
|
152 |
|
|
end
|
153 |
|
|
|
154 |
|
|
endmodule
|
155 |
|
|
|
156 |
|
|
module fpDiv_tb();
|
157 |
|
|
reg clk;
|
158 |
|
|
reg ld;
|
159 |
|
|
wire ce = 1'b1;
|
160 |
|
|
wire sgnx1,sgnx2,sgnx3,sgnx4,sgnx5,sgnx6;
|
161 |
|
|
wire inf1,inf2,inf3,inf4,inf5,inf6;
|
162 |
|
|
wire of1,of2,of3,of4,of5,of6;
|
163 |
|
|
wire uf1,uf2,uf3,uf4,uf5,uf6;
|
164 |
|
|
wire [57:0] o1,o2,o3,o4,o5,o6;
|
165 |
|
|
wire [35:0] o11,o12,o13;
|
166 |
|
|
wire [31:0] o21,o22,o23;
|
167 |
|
|
wire done0,done1,done2,done3,done4,done5,done6;
|
168 |
|
|
|
169 |
|
|
initial begin
|
170 |
|
|
clk = 0;
|
171 |
|
|
ld = 0;
|
172 |
|
|
#20 ld = 1;
|
173 |
|
|
#40 ld = 0;
|
174 |
|
|
end
|
175 |
|
|
always #10 clk <= ~clk;
|
176 |
|
|
|
177 |
|
|
fpDiv u1 (.clk(clk), .ce(1'b1), .ld(ld), .a(0), .b(0), .o(o1), .done(done1), .sign_exe(sgnx1), .overflow(of1), .underflow(uf1));
|
178 |
|
|
fpDiv u2 (.clk(clk), .ce(1'b1), .ld(ld), .a(0), .b(0), .o(o2), .done(done2), .sign_exe(sgnx2), .overflow(of2), .underflow(uf2));
|
179 |
|
|
// 10/10
|
180 |
|
|
fpDiv u3 (.clk(clk), .ce(1'b1), .ld(ld), .a(32'h41200000), .b(32'h41200000), .done(done3), .o(o3), .sign_exe(sgnx2), .overflow(of2), .underflow(uf2));
|
181 |
|
|
// 21/-17
|
182 |
|
|
fpDiv u4 (.clk(clk), .ce(1'b1), .ld(ld), .a(32'h41a80000), .b(32'hc1880000), .done(done4), .o(o4), .sign_exe(sgnx2), .overflow(of2), .underflow(uf2));
|
183 |
|
|
// -17/-15
|
184 |
|
|
fpDiv u5 (.clk(clk), .ce(1'b1), .ld(ld), .a(32'hc1880000), .b(32'hc1700000), .done(done5), .o(o5), .sign_exe(sgnx2), .overflow(of2), .underflow(uf2));
|
185 |
|
|
|
186 |
|
|
fpNormalize u11 (clk, ce, 1'b0, o3, o11);
|
187 |
|
|
fpNormalize u12 (clk, ce, 1'b0, o4, o12);
|
188 |
|
|
fpNormalize u13 (clk, ce, 1'b0, o5, o13);
|
189 |
|
|
|
190 |
|
|
fpRound u21 (3'd1, o11, o21); // zero for zero
|
191 |
|
|
fpRound u22 (3'd1, o12, o22); //
|
192 |
|
|
fpRound u23 (3'd1, o13, o23); //
|
193 |
|
|
|
194 |
|
|
endmodule
|