cvw/pipelined/src/fpu/fcvt.sv

237 lines
12 KiB
Systemverilog

///////////////////////////////////////////
// fcvt.sv
//
// Written: me@KatherineParry.com
// Modified: 7/5/2022
//
// Purpose: Floating point conversions of configurable size
//
// Int component of the Wally configurable RISC-V project.
//
// Copyright (C) 2021 Harvey Mudd College & Oklahoma State University
//
// MIT LICENSE
// Permission is hereby granted, free of charge, to any person obtaining a copy of this
// software and associated documentation files (the "Software"), to deal in the Software
// without restriction, including without limitation the rights to use, copy, modify, merge,
// publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons
// to whom the Software is furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in all copies or
// substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED,
// INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR Int PARTICULAR
// PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
// BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
// TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE
// OR OTHER DEALINGS IN THE SOFTWARE.
////////////////////////////////////////////////////////////////////////////////////////////////
`include "wally-config.vh"
module fcvt (
input logic Xs, // input's sign
input logic [`NE-1:0] Xe, // input's exponent
input logic [`NF:0] Xm, // input's fraction
input logic [`XLEN-1:0] Int, // integer input - from IEU
input logic [2:0] OpCtrl, // choose which opperation (look below for values)
input logic ToInt, // is fp->int (since it's writting to the integer register)
input logic XZero, // is the input zero
input logic XDenorm, // is the input denormalized
input logic [`FMTBITS-1:0] Fmt, // the input's precision (11=quad 01=double 00=single 10=half)
output logic [`NE:0] Ce, // the calculated expoent
output logic [`LOGCVTLEN-1:0] ShiftAmt, // how much to shift by
output logic ResDenormUf,// does the result underflow or is denormalized
output logic Cs, // the result's sign
output logic IntZero, // is the integer zero?
output logic [`CVTLEN-1:0] LzcIn // input to the Leading Zero Counter (priority encoder)
);
// OpCtrls:
// fp->fp conversions: {0, output precision} - only one of the operations writes to the int register
// half - 10
// single - 00
// double - 01
// quad - 11
// int<->fp conversions: {is int->fp?, is the integer 64-bit?, is the integer signed?}
// bit 2 bit 1 bit 0
// for example: signed long -> single floating point has the OpCode 101
logic [`FMTBITS-1:0] OutFmt; // format of the output
logic [`XLEN-1:0] PosInt; // the positive integer input
logic [`XLEN-1:0] TrimInt; // integer trimmed to the correct size
logic [`NE-2:0] NewBias; // the bias of the final result
logic [`NE-1:0] OldExp; // the old exponent
logic Signed; // is the opperation with a signed integer?
logic Int64; // is the integer 64 bits?
logic IntToFp; // is the opperation an int->fp conversion?
logic [`CVTLEN:0] LzcInFull; // input to the Leading Zero Counter (priority encoder)
logic [`LOGCVTLEN-1:0] LeadingZeros; // output from the LZC
// seperate OpCtrl for code readability
assign Signed = OpCtrl[0];
assign Int64 = OpCtrl[1];
assign IntToFp = OpCtrl[2];
// choose the ouptut format depending on the opperation
// - fp -> fp: OpCtrl contains the percision of the output
// - int -> fp: Fmt contains the percision of the output
if (`FPSIZES == 2)
assign OutFmt = IntToFp ? Fmt : (OpCtrl[1:0] == `FMT);
else if (`FPSIZES == 3 | `FPSIZES == 4)
assign OutFmt = IntToFp ? Fmt : OpCtrl[1:0];
///////////////////////////////////////////////////////////////////////////
// negation
///////////////////////////////////////////////////////////////////////////
// 1) negate the input if the input is a negitive singed integer
// 2) trim the input to the proper size (kill the 32 most significant zeroes if needed)
assign PosInt = Cs ? -Int : Int;
assign TrimInt = {{`XLEN-32{Int64}}, {32{1'b1}}} & PosInt;
assign IntZero = ~|TrimInt;
///////////////////////////////////////////////////////////////////////////
// lzc
///////////////////////////////////////////////////////////////////////////
// choose the input to the leading zero counter i.e. priority encoder
// int -> fp : | positive integer | 00000... (if needed) |
// fp -> fp : | fraction | 00000... (if needed) |
assign LzcInFull = IntToFp ? {TrimInt, {`CVTLEN-`XLEN+1{1'b0}}} :
{Xm, {`CVTLEN-`NF{1'b0}}};
assign LzcIn = LzcInFull[`CVTLEN-1:0];
lzc #(`CVTLEN+1) lzc (.num(LzcInFull), .ZeroCnt(LeadingZeros));
///////////////////////////////////////////////////////////////////////////
// shifter
///////////////////////////////////////////////////////////////////////////
// kill the shift if it's negitive
// select the amount to shift by
// fp -> int:
// - shift left by CalcExp - essentially shifting until the unbiased exponent = 0
// - don't shift if supposed to shift right (underflowed or denorm input)
// denormalized/undeflowed result fp -> fp:
// - shift left by NF-1+CalcExp - to shift till the biased expoenent is 0
// ??? -> fp:
// - shift left by LeadingZeros - to shift till the result is normalized
// - only shift fp -> fp if the intital value is denormalized
// - this is a problem because the input to the lzc was the fraction rather than the mantissa
// - rather have a few and-gates than an extra bit in the priority encoder??? *** is this true?
always_comb
if(ToInt) ShiftAmt = Ce[`LOGCVTLEN-1:0]&{`LOGCVTLEN{~Ce[`NE]}};
else if (ResDenormUf&~IntToFp) ShiftAmt = (`LOGCVTLEN)'(`NF-1)+Ce[`LOGCVTLEN-1:0];
else ShiftAmt = LeadingZeros;
///////////////////////////////////////////////////////////////////////////
// exp calculations
///////////////////////////////////////////////////////////////////////////
// *** possible optimizaations:
// - if subtracting exp by bias only the msb needs a full adder, the rest can be HA - dunno how to implement this for synth
// - Smaller exp -> Larger Exp can be calculated with: *** can use in Other units??? FMA??? insert this thing in later
// Exp if in range: {~Exp[SNE-1], Exp[SNE-2:0]}
// Exp in range if: Exp[SNE-1] = 1 & Exp[LNE-2:SNE] = 1111... & Exp[LNE-1] = 0 | Exp[SNE-1] = 0 & Exp[LNE-2:SNE] = 000... & Exp[LNE-1] = 1
// i.e.: &Exp[LNE-2:SNE-1] xor Exp[LNE-1]
// Too big if: Exp[LNE-1] = 1
// Too small if: none of the above
// Select the bias of the output
// fp -> int : select 1
// ??? -> fp : pick the new bias depending on the output format
if (`FPSIZES == 1) begin
assign NewBias = ToInt ? (`NE-1)'(1) : (`NE-1)'(`BIAS);
end else if (`FPSIZES == 2) begin
logic [`NE-2:0] NewBiasToFp;
assign NewBiasToFp = OutFmt ? (`NE-1)'(`BIAS) : (`NE-1)'(`BIAS1);
assign NewBias = ToInt ? (`NE-1)'(1) : NewBiasToFp;
end else if (`FPSIZES == 3) begin
logic [`NE-2:0] NewBiasToFp;
always_comb
case (OutFmt)
`FMT: NewBiasToFp = (`NE-1)'(`BIAS);
`FMT1: NewBiasToFp = (`NE-1)'(`BIAS1);
`FMT2: NewBiasToFp = (`NE-1)'(`BIAS2);
default: NewBiasToFp = {`NE-1{1'bx}};
endcase
assign NewBias = ToInt ? (`NE-1)'(1) : NewBiasToFp;
end else if (`FPSIZES == 4) begin
logic [`NE-2:0] NewBiasToFp;
always_comb
case (OutFmt)
2'h3: NewBiasToFp = (`NE-1)'(`Q_BIAS);
2'h1: NewBiasToFp = (`NE-1)'(`D_BIAS);
2'h0: NewBiasToFp = (`NE-1)'(`S_BIAS);
2'h2: NewBiasToFp = (`NE-1)'(`H_BIAS);
endcase
assign NewBias = ToInt ? (`NE-1)'(1) : NewBiasToFp;
end
// select the old exponent
// int -> fp : largest bias + XLEN
// fp -> ??? : XExp
assign OldExp = IntToFp ? (`NE)'(`BIAS)+(`NE)'(`XLEN-1) : Xe;
// calculate CalcExp
// fp -> fp :
// - XExp - Largest bias + new bias - (LeadingZeros+1)
// only do ^ if the input was denormalized
// - convert the expoenent to the final preciaion (Exp - oldBias + newBias)
// - correct the expoent when there is a normalization shift ( + LeadingZeros+1)
// fp -> int : XExp - Largest Bias + 1 - (LeadingZeros+1)
// | `XLEN zeros | Mantissa | 0's if nessisary | << CalcExp
// process:
// - start
// | `XLEN zeros | Mantissa | 0's if nessisary |
//
// - shift left 1 (1)
// | `XLEN-1 zeros |bit| frac | 0's if nessisary |
// . <- binary point
//
// - shift left till unbiased exponent is 0 (XExp - Largest Bias)
// | 0's | Mantissa | 0's if nessisary |
// | keep |
//
// - if the input is denormalized then we dont shift... so the "- LeadingZeros" is just leftovers from other options
// int -> fp : largest bias + XLEN - Largest bias + new bias - LeadingZeros = XLEN + NewBias - LeadingZeros
// Process:
// - shifted right by XLEN (XLEN)
// - shift left to normilize (-LeadingZeros)
// - newBias to make the biased exponent
// oldexp - biasold +newbias - LeadingZeros&(XDenorm|IntToFp)
assign Ce = {1'b0, OldExp} - (`NE+1)'(`BIAS) + {2'b0, NewBias} - {{`NE-`LOGCVTLEN+1{1'b0}}, (LeadingZeros&{`LOGCVTLEN{XDenorm|IntToFp}})};
// find if the result is dnormal or underflows
// - if Calculated expoenent is 0 or negitive (and the input/result is not exactaly 0)
// - can't underflow an integer to Fp conversion
assign ResDenormUf = (~|Ce | Ce[`NE])&~XZero&~IntToFp;
///////////////////////////////////////////////////////////////////////////
// sign
///////////////////////////////////////////////////////////////////////////
// determine the sign of the result
// - if int -> fp
// - if 64-bit : check the msb of the 64-bit integer input and if it's signed
// - if 32-bit : check the msb of the 32-bit integer input and if it's signed
// - otherwise: the floating point input's sign
always_comb
if(IntToFp)
if(Int64) Cs = Int[`XLEN-1]&Signed;
else Cs = Int[31]&Signed;
else Cs = Xs;
endmodule