| 1 |
6 |
robfinch |
/* ===============================================================
|
| 2 |
|
|
(C) 2006 Robert Finch
|
| 3 |
|
|
All rights reserved.
|
| 4 |
|
|
rob@birdcomputer.ca
|
| 5 |
|
|
|
| 6 |
|
|
fpDiv.v
|
| 7 |
|
|
- floating point divider
|
| 8 |
|
|
- parameterized width
|
| 9 |
|
|
- IEEE 754 representation
|
| 10 |
|
|
|
| 11 |
|
|
This source code is free for use and modification for
|
| 12 |
|
|
non-commercial or evaluation purposes, provided this
|
| 13 |
|
|
copyright statement and disclaimer remains present in
|
| 14 |
|
|
the file.
|
| 15 |
|
|
|
| 16 |
|
|
If you do modify the code, please state the origin and
|
| 17 |
|
|
note that you have modified the code.
|
| 18 |
|
|
|
| 19 |
|
|
NO WARRANTY.
|
| 20 |
|
|
THIS Work, IS PROVIDEDED "AS IS" WITH NO WARRANTIES OF
|
| 21 |
|
|
ANY KIND, WHETHER EXPRESS OR IMPLIED. The user must assume
|
| 22 |
|
|
the entire risk of using the Work.
|
| 23 |
|
|
|
| 24 |
|
|
IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
|
| 25 |
|
|
ANY INCIDENTAL, CONSEQUENTIAL, OR PUNITIVE DAMAGES
|
| 26 |
|
|
WHATSOEVER RELATING TO THE USE OF THIS WORK, OR YOUR
|
| 27 |
|
|
RELATIONSHIP WITH THE AUTHOR.
|
| 28 |
|
|
|
| 29 |
|
|
IN ADDITION, IN NO EVENT DOES THE AUTHOR AUTHORIZE YOU
|
| 30 |
|
|
TO USE THE WORK IN APPLICATIONS OR SYSTEMS WHERE THE
|
| 31 |
|
|
WORK'S FAILURE TO PERFORM CAN REASONABLY BE EXPECTED
|
| 32 |
|
|
TO RESULT IN A SIGNIFICANT PHYSICAL INJURY, OR IN LOSS
|
| 33 |
|
|
OF LIFE. ANY SUCH USE BY YOU IS ENTIRELY AT YOUR OWN RISK,
|
| 34 |
|
|
AND YOU AGREE TO HOLD THE AUTHOR AND CONTRIBUTORS HARMLESS
|
| 35 |
|
|
FROM ANY CLAIMS OR LOSSES RELATING TO SUCH UNAUTHORIZED
|
| 36 |
|
|
USE.
|
| 37 |
|
|
|
| 38 |
|
|
This multiplier/divider handles denormalized numbers.
|
| 39 |
|
|
The output format is of an internal expanded representation
|
| 40 |
|
|
in preparation to be fed into a normalization unit, then
|
| 41 |
|
|
rounding. Basically, it's the same as the regular format
|
| 42 |
|
|
except the mantissa is doubled in size, the leading two
|
| 43 |
|
|
bits of which are assumed to be whole bits.
|
| 44 |
|
|
|
| 45 |
|
|
|
| 46 |
|
|
Floating Point Multiplier / Divider
|
| 47 |
|
|
|
| 48 |
|
|
Properties:
|
| 49 |
|
|
+-inf * +-inf = -+inf (this is handled by exOver)
|
| 50 |
|
|
+-inf * 0 = QNaN
|
| 51 |
|
|
+-0 / +-0 = QNaN
|
| 52 |
|
|
|
| 53 |
|
|
Ref: Webpack8.2i Spartan3-4 xc3s1000-4ft256
|
| 54 |
|
|
316 LUTS / 174 slices / 49.7 MHz
|
| 55 |
|
|
=============================================================== */
|
| 56 |
|
|
|
| 57 |
|
|
module fpDiv(clk, ce, ld, a, b, o, done, sign_exe, overflow, underflow);
|
| 58 |
|
|
|
| 59 |
|
|
parameter WID = 32;
|
| 60 |
|
|
localparam MSB = WID-1;
|
| 61 |
|
|
localparam EMSB = WID==80 ? 14 :
|
| 62 |
|
|
WID==64 ? 10 :
|
| 63 |
|
|
WID==52 ? 10 :
|
| 64 |
|
|
WID==48 ? 10 :
|
| 65 |
|
|
WID==44 ? 10 :
|
| 66 |
|
|
WID==42 ? 10 :
|
| 67 |
|
|
WID==40 ? 9 :
|
| 68 |
|
|
WID==32 ? 7 :
|
| 69 |
|
|
WID==24 ? 6 : 4;
|
| 70 |
|
|
localparam FMSB = WID==80 ? 63 :
|
| 71 |
|
|
WID==64 ? 51 :
|
| 72 |
|
|
WID==52 ? 39 :
|
| 73 |
|
|
WID==48 ? 35 :
|
| 74 |
|
|
WID==44 ? 31 :
|
| 75 |
|
|
WID==42 ? 29 :
|
| 76 |
|
|
WID==40 ? 28 :
|
| 77 |
|
|
WID==32 ? 22 :
|
| 78 |
|
|
WID==24 ? 15 : 9;
|
| 79 |
|
|
|
| 80 |
|
|
localparam FX = (FMSB+2)*2-1; // the MSB of the expanded fraction
|
| 81 |
|
|
localparam EX = FX + 1 + EMSB + 1 + 1 - 1;
|
| 82 |
|
|
|
| 83 |
|
|
input clk;
|
| 84 |
|
|
input ce;
|
| 85 |
|
|
input ld;
|
| 86 |
|
|
input [MSB:0] a, b;
|
| 87 |
|
|
output [EX:0] o;
|
| 88 |
|
|
output done;
|
| 89 |
|
|
output sign_exe;
|
| 90 |
|
|
output overflow;
|
| 91 |
|
|
output underflow;
|
| 92 |
|
|
|
| 93 |
|
|
// registered outputs
|
| 94 |
|
|
reg sign_exe;
|
| 95 |
|
|
reg inf;
|
| 96 |
|
|
reg overflow;
|
| 97 |
|
|
reg underflow;
|
| 98 |
|
|
|
| 99 |
|
|
reg so;
|
| 100 |
|
|
reg [EMSB:0] xo;
|
| 101 |
|
|
reg [FX:0] mo;
|
| 102 |
|
|
assign o = {so,xo,mo};
|
| 103 |
|
|
|
| 104 |
|
|
// constants
|
| 105 |
|
|
wire [EMSB:0] infXp = {EMSB+1{1'b1}}; // infinite / NaN - all ones
|
| 106 |
|
|
// The following is the value for an exponent of zero, with the offset
|
| 107 |
|
|
// eg. 8'h7f for eight bit exponent, 11'h7ff for eleven bit exponent, etc.
|
| 108 |
|
|
wire [EMSB:0] bias = {1'b0,{EMSB{1'b1}}}; //2^0 exponent
|
| 109 |
|
|
// The following is a template for a quiet nan. (MSB=1)
|
| 110 |
|
|
wire [FMSB:0] qNaN = {1'b1,{FMSB{1'b0}}};
|
| 111 |
|
|
|
| 112 |
|
|
// variables
|
| 113 |
|
|
wire [EMSB+2:0] ex1; // sum of exponents
|
| 114 |
|
|
wire [FX:0] divo;
|
| 115 |
|
|
|
| 116 |
|
|
// Operands
|
| 117 |
|
|
wire sa, sb; // sign bit
|
| 118 |
|
|
wire [EMSB:0] xa, xb; // exponent bits
|
| 119 |
|
|
wire [FMSB+1:0] fracta, fractb;
|
| 120 |
|
|
wire a_dn, b_dn; // a/b is denormalized
|
| 121 |
|
|
wire az, bz;
|
| 122 |
|
|
wire aInf, bInf;
|
| 123 |
|
|
|
| 124 |
|
|
|
| 125 |
|
|
// -----------------------------------------------------------
|
| 126 |
|
|
// - decode the input operands
|
| 127 |
|
|
// - derive basic information
|
| 128 |
|
|
// - calculate exponent
|
| 129 |
|
|
// - calculate fraction
|
| 130 |
|
|
// -----------------------------------------------------------
|
| 131 |
|
|
|
| 132 |
|
|
fpDecomp #(WID) u1a (.i(a), .sgn(sa), .exp(xa), .fract(fracta), .xz(a_dn), .vz(az), .inf(aInf) );
|
| 133 |
|
|
fpDecomp #(WID) u1b (.i(b), .sgn(sb), .exp(xb), .fract(fractb), .xz(b_dn), .vz(bz), .inf(bInf) );
|
| 134 |
|
|
|
| 135 |
|
|
// Compute the exponent.
|
| 136 |
|
|
// - correct the exponent for denormalized operands
|
| 137 |
|
|
// - adjust the difference by the bias (add 127)
|
| 138 |
|
|
// - also factor in the different decimal position for division
|
| 139 |
|
|
assign ex1 = (xa|a_dn) - (xb|b_dn) + bias + FMSB-1;
|
| 140 |
|
|
|
| 141 |
|
|
// check for exponent underflow/overflow
|
| 142 |
|
|
wire under = ex1[EMSB+2]; // MSB set = negative exponent
|
| 143 |
|
|
wire over = (&ex1[EMSB:0] | ex1[EMSB+1]) & !ex1[EMSB+2];
|
| 144 |
|
|
|
| 145 |
|
|
// Perform divide
|
| 146 |
|
|
// could take either 1 or 16 clock cycles
|
| 147 |
|
|
fpdivr8 #(WID) u2 (.clk(clk), .ld(ld), .a(fracta), .b(fractb), .q(divo), .r(), .done(done));
|
| 148 |
|
|
|
| 149 |
|
|
// determine when a NaN is output
|
| 150 |
|
|
wire qNaNOut = (az&bz)|(aInf&bInf);
|
| 151 |
|
|
|
| 152 |
|
|
always @(posedge clk)
|
| 153 |
|
|
if (ce) begin
|
| 154 |
|
|
if (done) begin
|
| 155 |
|
|
casex({qNaNOut,bInf,bz})
|
| 156 |
|
|
3'b1xx: xo = infXp; // NaN exponent value
|
| 157 |
|
|
3'bx1x: xo = 0; // divide by inf
|
| 158 |
|
|
3'bxx1: xo = infXp; // divide by zero
|
| 159 |
|
|
default: xo = ex1; // normal or underflow: passthru neg. exp. for normalization
|
| 160 |
|
|
endcase
|
| 161 |
|
|
|
| 162 |
|
|
casex({qNaNOut,bInf,bz})
|
| 163 |
|
|
3'b1xx: mo = {1'b0,qNaN[FMSB:0]|{aInf,1'b0}|{az,bz},{FMSB+1{1'b0}}};
|
| 164 |
|
|
3'bx1x: mo = 0; // div by inf
|
| 165 |
|
|
3'bxx1: mo = 0; // div by zero
|
| 166 |
|
|
default: mo = divo; // plain div
|
| 167 |
|
|
endcase
|
| 168 |
|
|
|
| 169 |
|
|
so = sa ^ sb;
|
| 170 |
|
|
sign_exe = sa & sb;
|
| 171 |
|
|
overflow = over;
|
| 172 |
|
|
underflow = under;
|
| 173 |
|
|
end
|
| 174 |
|
|
end
|
| 175 |
|
|
|
| 176 |
|
|
endmodule
|