Subversion Repositories ft816float

// ============================================================================
//        __
//   \\__/ o\    (C) 2006-2020  Robert Finch, Waterloo
//    \  __ /    All rights reserved.
//     \/_//     robfinch<remove>@finitron.ca
//       ||
//
//      DFPNormalize.sv
//    - decimal floating point normalization unit
//    - eight cycle latency
//    - parameterized width
//
//
// This source file is free software: you can redistribute it and/or modify 
// it under the terms of the GNU Lesser General Public License as published 
// by the Free Software Foundation, either version 3 of the License, or     
// (at your option) any later version.                                      
//                                                                          
// This source file is distributed in the hope that it will be useful,      
// but WITHOUT ANY WARRANTY; without even the implied warranty of           
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the            
// GNU General Public License for more details.                             
//                                                                          
// You should have received a copy of the GNU General Public License        
// along with this program.  If not, see <http://www.gnu.org/licenses/>.    
//                                                                          
//      This unit takes a floating point number in an intermediate
// format and normalizes it. No normalization occurs
// for NaN's or infinities. The unit has a two cycle latency.
//
// The mantissa is assumed to start with two whole bits on
// the left. The remaining bits are fractional.
//
// The width of the incoming format is reduced via a generation
// of sticky bit in place of the low order fractional bits.
//
// On an underflowed input, the incoming exponent is assumed
// to be negative. A right shift is needed.
// ============================================================================

import fp::*;

module DFPNormalize(clk, ce, i, o, under_i, under_o, inexact_o);
parameter N=33;
input clk;
input ce;
input [(N+1)*4*2+16+4-1:0] i;           // expanded format input
output [N*4+16+4-1+4:0] o;              // normalized output + guard, sticky and round bits, + 1 whole digit
input under_i;
output under_o;
output inexact_o;

integer n;
// ----------------------------------------------------------------------------
// No Clock required
// ----------------------------------------------------------------------------
reg [15:0] xo0;
reg so0;
reg sx0;
reg nan0;
reg inf0;

always @*
        xo0 <= i[(N+1)*4*2+15:(N+1)*4*2];
always @*
        so0 <= i[(N+1)*4*2+16+4-2];             // sign doesn't change
always @*
        sx0 <= i[(N+1)*4*2+16+4-4];
always @*
        nan0 <= i[(N+1)*4*2+16+4-1];
always @*
        inf0 <= i[(N+1)*4*2+16+4-3] || xo0==16'h9999 && i[(N+1)*4*2-4];

// ----------------------------------------------------------------------------
// Clock #1
// - Capture exponent information
// ----------------------------------------------------------------------------
reg xInf1a, xInf1b, xInf1c;
wire [(N+1)*4*2+16+4-1:0] i1;
delay #(.WID((N+2)*4*2+16+4),.DEP(1)) u11 (.clk(clk), .ce(ce), .i(i), .o(i1));

always @(posedge clk)
        if (ce) xInf1a <= xo0==16'h9999 & !under_i;
always @(posedge clk)
        if (ce) xInf1b <= xo0==16'h9998 & !under_i;
always @(posedge clk)
        if (ce) xInf1c <= xo0==16'h9999;

// ----------------------------------------------------------------------------
// Clock #2
// - determine exponent increment
// Since the there are *three* whole digits in the incoming format
// the number of whole digits needs to be reduced. If the MSB is
// set, then increment the exponent and no shift is needed.
// ----------------------------------------------------------------------------
wire xInf2c, xInf2b;
wire [15:0] xo2;
reg incExpByOne2, incExpByTwo2;
delay #(.WID(1),.DEP(1)) u21 (.clk(clk), .ce(ce), .i(xInf1c), .o(xInf2c));
delay #(.WID(1),.DEP(1)) u22 (.clk(clk), .ce(ce), .i(xInf1b), .o(xInf2b));
delay #(.WID(16),.DEP(2)) u23 (.clk(clk), .ce(ce), .i(xo0), .o(xo2));
delay #(.WID(1),.DEP(2)) u24 (.clk(clk), .ce(ce), .i(under_i), .o(under2));

always @(posedge clk)
        if (ce) incExpByOne2 <= !xInf1a & i1[(N+1)*4*2-4];

// ----------------------------------------------------------------------------
// Clock #3
// - increment exponent
// - detect a zero mantissa
// ----------------------------------------------------------------------------

wire incExpByOne3;
wire [(N+1)*4*2+16+4-1:0] i3;
reg [15:0] xo3;
reg zeroMan3;
delay #(.WID(1),.DEP(1)) u32 (.clk(clk), .ce(ce), .i(incExpByOne2), .o(incExpByOne3));
delay #(.WID((N+1)*4*2+16+4),.DEP(3)) u33 (.clk(clk), .ce(ce), .i(i[(N+3)*4*2+16+4-1:0]), .o(i3));

wire [15:0] xo2a;
BCDAddN #(.N(4)) ubcdan1
(
        .ci(1'b0),
        .a(xo2),
        .b(16'h0001),
        .o(xo2a),
        .co()
);

always @(posedge clk)
        if (ce) xo3 <= (incExpByOne2 ? xo2a : xo2);

always @(posedge clk)
        if(ce) zeroMan3 <= 1'b0;

// ----------------------------------------------------------------------------
// Clock #4
// - Shift mantissa left
// - If infinity is reached then set the mantissa to zero
//   shift mantissa left to reduce to a single whole digit
// - create sticky bit
// ----------------------------------------------------------------------------

reg [(N+2)*4-1:0] mo4;
reg inexact4;

always @(posedge clk)
if(ce)
casez({zeroMan3,incExpByOne3})
2'b1?:  mo4 <= 1'd0;
2'b01:  mo4 <= {i3[(N+1)*4*2-1:(N+1)*4],3'b0,|i3[(N+1)*4-1:0]};
default:        mo4 <= {i3[(N+1)*4*2-1-4:N*4],3'b0,|i3[N*4-1:0]};
endcase

always @(posedge clk)
if(ce)
casez({zeroMan3,incExpByOne3})
2'b1?:  inexact4 <= 1'd0;
2'b01:  inexact4 <= |i3[(N+1)*4-1:0];
default:        inexact4 <= |i3[N*4-1:0];
endcase

// ----------------------------------------------------------------------------
// Clock edge #5
// - count leading zeros
// ----------------------------------------------------------------------------
reg [7:0] leadingZeros5;
wire [15:0] xo5;
wire xInf5;
delay #(.WID(16),.DEP(2)) u51 (.clk(clk), .ce(ce), .i(xo3), .o(xo5));
delay #(.WID(1),.DEP(3)) u52 (.clk(clk), .ce(ce), .i(xInf2c), .o(xInf5) );

/* Lookup table based leading zero count modules give slightly better
   performance but cases must be coded.
generate
begin
if (FPWID <= 32) begin
cntlz32Reg clz0 (.clk(clk), .ce(ce), .i({mo4,4'b0}), .o(leadingZeros5) );
assign leadingZeros5[7:6] = 2'b00;
end
else if (FPWID<=64) begin
assign leadingZeros5[7] = 1'b0;
cntlz64Reg clz0 (.clk(clk), .ce(ce), .i({mo4,7'h0}), .o(leadingZeros5) );
end
else if (FPWID<=80) begin
assign leadingZeros5[7] = 1'b0;
cntlz80Reg clz0 (.clk(clk), .ce(ce), .i({mo4,11'b0}), .o(leadingZeros5) );
end
else if (FPWID<=84) begin
assign leadingZeros5[7] = 1'b0;
cntlz96Reg clz0 (.clk(clk), .ce(ce), .i({mo4,23'b0}), .o(leadingZeros5) );
end
else if (FPWID<=96) begin
assign leadingZeros5[7] = 1'b0;
cntlz96Reg clz0 (.clk(clk), .ce(ce), .i({mo4,11'b0}), .o(leadingZeros5) );
end
else if (FPWID<=128)
cntlz128Reg clz0 (.clk(clk), .ce(ce), .i({mo4,11'b0}), .o(leadingZeros5) );
end
endgenerate
*/

// Sideways add.
// Normally there would be only one to two leading zeros. It is tempting then
// to check for only one or two. But, denormalized numbers might have more
// leading zeros. If denormals were not supported this could be made smaller
// and faster.
`ifdef SUPPORT_DENORMALS
reg [7:0] lzc;
reg got_one;
always @*
begin
  got_one = 1'b0;
  lzc = 8'h00;
  for (n = (N+2)*4-1; n >= 0; n = n - 4) begin
    if (!got_one) begin
      if (mo4[n]|mo4[n-1]|mo4[n-2]|mo4[n-3])
        got_one = 1'b1;
      else
        lzc = lzc + 1'b1;
    end
  end
end      
always @(posedge clk)
  if (ce) leadingZeros5 <= lzc;
`else
always @(posedge clk)
if (ce)
casez(mo4[(N+1)*4-1:(N-1)*4-1])
8'b00000000:    leadingZeros5 <= 8'd2;
8'b0000????:    leadingZeros5 <= 8'd1;
default:                        leadingZeros5 <= 8'd0;
endcase
`endif


// ----------------------------------------------------------------------------
// Clock edge #6
// - Compute how much we want to decrement exponent by
// - compute amount to shift left and right
// - at infinity the exponent can't be incremented, so we can't shift right
//   otherwise it was an underflow situation so the exponent was negative
//   shift amount needs to be negated for shift register
// If the exponent underflowed, then the shift direction must be to the
// right regardless of mantissa bits; the number is denormalized.
// Otherwise the shift direction must be to the left.
// ----------------------------------------------------------------------------
reg [7:0] lshiftAmt6;
reg [7:0] rshiftAmt6;
wire rightOrLeft6;      // 0=left,1=right
wire xInf6;
wire [15:0] xo6;
wire [(N+2)*4-1:0] mo6;
wire zeroMan6;
vtdl #(1) u61 (.clk(clk), .ce(ce), .a(4'd5), .d(under_i), .q(rightOrLeft6) );
delay #(.WID(16),.DEP(1)) u62 (.clk(clk), .ce(ce), .i(xo5), .o(xo6));
delay #(.WID((N+2)*4),.DEP(2)) u63 (.clk(clk), .ce(ce), .i(mo4), .o(mo6) );
delay #(.WID(1),.DEP(1)) u64 (.clk(clk), .ce(ce), .i(xInf5), .o(xInf6) );
delay #(.WID(1),.DEP(3)) u65 (.clk(clk), .ce(ce),  .i(zeroMan3), .o(zeroMan6));
delay #(.WID(1),.DEP(5)) u66 (.clk(clk), .ce(ce), .i(sx0), .o(sx5) );

wire [13:0] xo5d = xo5[3:0] + xo5[7:4] * 10 + xo5[11:8] * 100 + xo5[15:12] * 1000;

always @(posedge clk)
        if (ce) lshiftAmt6 <= {leadingZeros5 > xo5d ? xo5d : leadingZeros5,2'b0};

always @(posedge clk)
        if (ce) rshiftAmt6 <= xInf5 ? 1'd0 : sx5 ? 1'd0 : xo5d > N ? N*4 : {xo5d[5:0],2'b00};   // xo2 is negative !

// ----------------------------------------------------------------------------
// Clock edge #7
// - figure exponent
// - shift mantissa
// - figure sticky bit
// ----------------------------------------------------------------------------

reg [15:0] xo7;
wire rightOrLeft7;
reg [(N+2)*4-1:0] mo7l, mo7r;
reg St6,St7;
delay #(.WID(1),.DEP(1)) u71 (.clk(clk), .ce(ce), .i(rightOrLeft6), .o(rightOrLeft7));

wire [11:0] lshftAmtBCD;
wire [15:0] xo7d;
BinToBCD ubbcd1 (lshiftAmt6, lshftAmtBCD);
BCDSubN #(.N(4)) ubcdsn1
(
        .ci(1'b0),
        .a(xo6),
        .b({4'h0,lshftAmtBCD}),
        .o(xo7d),
        .co()
);


always @(posedge clk)
if (ce)
        xo7 <= zeroMan6 ? xo6 :
                xInf6 ? xo6 :                                   // an infinite exponent is either a NaN or infinity; no need to change
                rightOrLeft6 ? 1'd0 :   // on a right shift, the exponent was negative, it's being made to zero
                xo7d;                   // on a left shift, the exponent can't be decremented below zero

always @(posedge clk)
        if (ce) mo7r <= mo6 >> rshiftAmt6;
always @(posedge clk)
        if (ce) mo7l <= mo6 << lshiftAmt6;

// The sticky bit is set if the bits shifted out on a right shift are set.
always @*
begin
  St6 = 1'b0;
  for (n = 0; n < (N+2)*4; n = n + 1)
    if (n <= rshiftAmt6 + 1) St6 = St6|mo6[n];
end
always @(posedge clk)
  if (ce) St7 <= St6;

// ----------------------------------------------------------------------------
// Clock edge #8
// - select mantissa
// ----------------------------------------------------------------------------

wire so,sxo,nano,info;
wire [15:0] xo;
reg [(N+2)*4-1:0] mo;
vtdl #(1) u81 (.clk(clk), .ce(ce), .a(4'd7), .d(so0), .q(so) );
delay #(.WID(16),.DEP(1)) u82 (.clk(clk), .ce(ce), .i(xo7), .o(xo));
vtdl #(.WID(1)) u83 (.clk(clk), .ce(ce), .a(4'd3), .d(inexact4), .q(inexact_o));
delay #(.WID(1),.DEP(1)) u84 (.clk(clk), .ce(ce), .i(rightOrLeft7), .o(under_o));
vtdl #(1) u85 (.clk(clk), .ce(ce), .a(4'd7), .d(sx0), .q(sxo) );
vtdl #(1) u86 (.clk(clk), .ce(ce), .a(4'd7), .d(nan0), .q(nano) );
vtdl #(1) u87 (.clk(clk), .ce(ce), .a(4'd7), .d(inf0), .q(info) );

always @(posedge clk)
        if (ce) mo <= rightOrLeft7 ? mo7r|{St7,4'b0} : mo7l;

assign o = {nano,so,info,sxo,xo,mo[(N+2)*4-1:4]};

endmodule