1 |
258 |
julius |
//////////////////////////////////////////////////////////////////////
|
2 |
|
|
//// ////
|
3 |
|
|
//// or1200_fpu_pre_norm_addsub ////
|
4 |
|
|
//// ////
|
5 |
|
|
//// This file is part of the OpenRISC 1200 project ////
|
6 |
|
|
//// http://opencores.org/project,or1k ////
|
7 |
|
|
//// ////
|
8 |
|
|
//// Description ////
|
9 |
|
|
//// pre-normalization entity for the addition/subtraction unit ////
|
10 |
|
|
//// ////
|
11 |
|
|
//// To Do: ////
|
12 |
|
|
//// ////
|
13 |
|
|
//// ////
|
14 |
|
|
//// Author(s): ////
|
15 |
|
|
//// - Original design (FPU100) - ////
|
16 |
|
|
//// Jidan Al-eryani, jidan@gmx.net ////
|
17 |
|
|
//// - Conv. to Verilog and inclusion in OR1200 - ////
|
18 |
|
|
//// Julius Baxter, julius@opencores.org ////
|
19 |
|
|
//// ////
|
20 |
|
|
//////////////////////////////////////////////////////////////////////
|
21 |
|
|
//
|
22 |
|
|
// Copyright (C) 2006, 2010
|
23 |
|
|
//
|
24 |
|
|
// This source file may be used and distributed without
|
25 |
|
|
// restriction provided that this copyright statement is not
|
26 |
|
|
// removed from the file and that any derivative work contains
|
27 |
|
|
// the original copyright notice and the associated disclaimer.
|
28 |
|
|
//
|
29 |
|
|
// THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY
|
30 |
|
|
// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
|
31 |
|
|
// TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
|
32 |
|
|
// FOR A PARTICULAR PURPOSE. IN NO EVENT SHALL THE AUTHOR
|
33 |
|
|
// OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
|
34 |
|
|
// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
|
35 |
|
|
// (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
|
36 |
|
|
// GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
|
37 |
|
|
// BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
|
38 |
|
|
// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
39 |
|
|
// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
|
40 |
|
|
// OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
|
41 |
|
|
// POSSIBILITY OF SUCH DAMAGE.
|
42 |
|
|
//
|
43 |
|
|
|
44 |
|
|
module or1200_fpu_pre_norm_addsub (
|
45 |
|
|
|
46 |
|
|
clk_i,
|
47 |
|
|
opa_i,
|
48 |
|
|
opb_i,
|
49 |
|
|
fracta_28_o,
|
50 |
|
|
fractb_28_o,
|
51 |
|
|
exp_o
|
52 |
|
|
);
|
53 |
|
|
|
54 |
|
|
parameter FP_WIDTH = 32;
|
55 |
|
|
parameter MUL_SERIAL = 0; // 0 for parallel multiplier, 1 for serial
|
56 |
|
|
parameter MUL_COUNT = 11; //11 for parallel multiplier, 34 for serial
|
57 |
|
|
parameter FRAC_WIDTH = 23;
|
58 |
|
|
parameter EXP_WIDTH = 8;
|
59 |
|
|
parameter ZERO_VECTOR = 31'd0;
|
60 |
|
|
parameter INF = 31'b1111111100000000000000000000000;
|
61 |
|
|
parameter QNAN = 31'b1111111110000000000000000000000;
|
62 |
|
|
parameter SNAN = 31'b1111111100000000000000000000001;
|
63 |
|
|
|
64 |
|
|
|
65 |
|
|
input clk_i;
|
66 |
|
|
input [FP_WIDTH-1:0] opa_i;
|
67 |
|
|
input [FP_WIDTH-1:0] opb_i;
|
68 |
|
|
// carry(1) & hidden(1) & fraction(23) & guard(1) & round(1) & sticky(1)
|
69 |
|
|
output reg [FRAC_WIDTH+4:0] fracta_28_o;
|
70 |
|
|
output reg [FRAC_WIDTH+4:0] fractb_28_o;
|
71 |
|
|
output reg [EXP_WIDTH-1:0] exp_o;
|
72 |
|
|
|
73 |
|
|
reg [EXP_WIDTH-1 : 0] s_exp_o ;
|
74 |
|
|
wire [FRAC_WIDTH+4 : 0] s_fracta_28_o, s_fractb_28_o ;
|
75 |
|
|
wire [EXP_WIDTH-1 : 0] s_expa;
|
76 |
|
|
wire [EXP_WIDTH-1 : 0] s_expb ;
|
77 |
|
|
wire [FRAC_WIDTH-1 : 0] s_fracta;
|
78 |
|
|
wire [FRAC_WIDTH-1 : 0] s_fractb ;
|
79 |
|
|
wire [FRAC_WIDTH+4 : 0] s_fracta_28;
|
80 |
|
|
|
81 |
|
|
wire [FRAC_WIDTH+4 : 0] s_fractb_28 ;
|
82 |
|
|
|
83 |
|
|
wire [FRAC_WIDTH+4 : 0] s_fract_sm_28;
|
84 |
|
|
wire [FRAC_WIDTH+4 : 0] s_fract_shr_28 ;
|
85 |
|
|
|
86 |
|
|
reg [EXP_WIDTH-1 : 0] s_exp_diff ;
|
87 |
|
|
reg [5 : 0] s_rzeros ;
|
88 |
|
|
wire s_expa_eq_expb;
|
89 |
|
|
wire s_expa_gt_expb;
|
90 |
|
|
wire s_fracta_1;
|
91 |
|
|
wire s_fractb_1;
|
92 |
|
|
wire s_op_dn,s_opa_dn, s_opb_dn;
|
93 |
|
|
wire [1 : 0] s_mux_diff ;
|
94 |
|
|
wire s_mux_exp;
|
95 |
|
|
wire s_sticky;
|
96 |
|
|
|
97 |
|
|
|
98 |
|
|
assign s_expa = opa_i[30:23];
|
99 |
|
|
assign s_expb = opb_i[30:23];
|
100 |
|
|
assign s_fracta = opa_i[22:0];
|
101 |
|
|
assign s_fractb = opb_i[22:0];
|
102 |
|
|
|
103 |
|
|
always @(posedge clk_i)
|
104 |
|
|
begin
|
105 |
|
|
exp_o <= s_exp_o;
|
106 |
|
|
fracta_28_o <= s_fracta_28_o;
|
107 |
|
|
fractb_28_o <= s_fractb_28_o;
|
108 |
|
|
end
|
109 |
|
|
|
110 |
|
|
assign s_expa_eq_expb = (s_expa == s_expb);
|
111 |
|
|
|
112 |
|
|
assign s_expa_gt_expb = (s_expa > s_expb);
|
113 |
|
|
|
114 |
|
|
// '1' if fraction is not zero
|
115 |
|
|
assign s_fracta_1 = |s_fracta;
|
116 |
|
|
assign s_fractb_1 = |s_fractb;
|
117 |
|
|
|
118 |
|
|
// opa or Opb is denormalized
|
119 |
|
|
assign s_opa_dn = !(|s_expa);
|
120 |
|
|
assign s_opb_dn = !(|s_expb);
|
121 |
|
|
assign s_op_dn = s_opa_dn | s_opb_dn;
|
122 |
|
|
|
123 |
|
|
// Output larger exponent
|
124 |
|
|
assign s_mux_exp = s_expa_gt_expb;
|
125 |
|
|
|
126 |
|
|
always @(posedge clk_i)
|
127 |
|
|
s_exp_o <= s_mux_exp ? s_expa : s_expb;
|
128 |
|
|
|
129 |
|
|
// convert to an easy to handle floating-point format
|
130 |
|
|
assign s_fracta_28 = s_opa_dn ?
|
131 |
|
|
{2'b00, s_fracta, 3'b000} : {2'b01, s_fracta, 3'b000};
|
132 |
|
|
assign s_fractb_28 = s_opb_dn ?
|
133 |
|
|
{2'b00, s_fractb, 3'b000} : {2'b01, s_fractb, 3'b000};
|
134 |
|
|
|
135 |
|
|
assign s_mux_diff = {s_expa_gt_expb, s_opa_dn ^ s_opb_dn};
|
136 |
|
|
|
137 |
|
|
// calculate howmany postions the fraction will be shifted
|
138 |
|
|
always @(posedge clk_i)
|
139 |
|
|
begin
|
140 |
|
|
case(s_mux_diff)
|
141 |
|
|
2'b00: s_exp_diff <= s_expb - s_expa;
|
142 |
|
|
2'b01: s_exp_diff <= s_expb - (s_expa + 8'd1);
|
143 |
|
|
2'b10: s_exp_diff <= s_expa - s_expb;
|
144 |
|
|
2'b11: s_exp_diff <= s_expa - (s_expb + 8'd1);
|
145 |
|
|
endcase
|
146 |
|
|
end
|
147 |
|
|
|
148 |
|
|
assign s_fract_sm_28 = s_expa_gt_expb ? s_fractb_28 : s_fracta_28;
|
149 |
|
|
|
150 |
|
|
// shift-right the fraction if necessary
|
151 |
|
|
assign s_fract_shr_28 = s_fract_sm_28 >> s_exp_diff;
|
152 |
|
|
|
153 |
|
|
// count the zeros from right to check if result is inexact
|
154 |
|
|
always @(s_fract_sm_28)
|
155 |
|
|
casex(s_fract_sm_28) // synopsys full_case parallel_case
|
156 |
|
|
28'b???????????????????????????1: s_rzeros <= 0;
|
157 |
|
|
28'b??????????????????????????10: s_rzeros <= 1;
|
158 |
|
|
28'b?????????????????????????100: s_rzeros <= 2;
|
159 |
|
|
28'b????????????????????????1000: s_rzeros <= 3;
|
160 |
|
|
28'b???????????????????????10000: s_rzeros <= 4;
|
161 |
|
|
28'b??????????????????????100000: s_rzeros <= 5;
|
162 |
|
|
28'b?????????????????????1000000: s_rzeros <= 6;
|
163 |
|
|
28'b????????????????????10000000: s_rzeros <= 7;
|
164 |
|
|
28'b???????????????????100000000: s_rzeros <= 8;
|
165 |
|
|
28'b??????????????????1000000000: s_rzeros <= 9;
|
166 |
|
|
28'b?????????????????10000000000: s_rzeros <= 10;
|
167 |
|
|
28'b????????????????100000000000: s_rzeros <= 11;
|
168 |
|
|
28'b???????????????1000000000000: s_rzeros <= 12;
|
169 |
|
|
28'b??????????????10000000000000: s_rzeros <= 13;
|
170 |
|
|
28'b?????????????100000000000000: s_rzeros <= 14;
|
171 |
|
|
28'b????????????1000000000000000: s_rzeros <= 15;
|
172 |
|
|
28'b???????????10000000000000000: s_rzeros <= 16;
|
173 |
|
|
28'b??????????100000000000000000: s_rzeros <= 17;
|
174 |
|
|
28'b?????????1000000000000000000: s_rzeros <= 18;
|
175 |
|
|
28'b????????10000000000000000000: s_rzeros <= 19;
|
176 |
|
|
28'b???????100000000000000000000: s_rzeros <= 20;
|
177 |
|
|
28'b??????1000000000000000000000: s_rzeros <= 21;
|
178 |
|
|
28'b?????10000000000000000000000: s_rzeros <= 22;
|
179 |
|
|
28'b????100000000000000000000000: s_rzeros <= 23;
|
180 |
|
|
28'b???1000000000000000000000000: s_rzeros <= 24;
|
181 |
|
|
28'b??10000000000000000000000000: s_rzeros <= 25;
|
182 |
|
|
28'b?100000000000000000000000000: s_rzeros <= 26;
|
183 |
|
|
28'b1000000000000000000000000000: s_rzeros <= 27;
|
184 |
|
|
28'b0000000000000000000000000000: s_rzeros <= 28;
|
185 |
|
|
endcase // casex (s_fract_sm_28)
|
186 |
|
|
|
187 |
|
|
assign s_sticky = (s_exp_diff > s_rzeros) & (|s_fract_sm_28);
|
188 |
|
|
|
189 |
|
|
assign s_fracta_28_o = s_expa_gt_expb ?
|
190 |
|
|
s_fracta_28 :
|
191 |
|
|
{s_fract_shr_28[27:1],(s_sticky|s_fract_shr_28[0])};
|
192 |
|
|
|
193 |
|
|
assign s_fractb_28_o = s_expa_gt_expb ?
|
194 |
|
|
{s_fract_shr_28[27:1],(s_sticky|s_fract_shr_28[0])} :
|
195 |
|
|
s_fractb_28;
|
196 |
|
|
|
197 |
|
|
endmodule // or1200_fpu_pre_norm_addsub
|
198 |
|
|
|
199 |
|
|
|