cvw/pipelined/src/fpu/fcvt.sv

373 lines
22 KiB
Systemverilog
Raw Normal View History

2022-05-26 00:10:51 +00:00
`include "wally-config.vh"
// largest length in IEU/FPU
`define LGLEN ((`NF<`XLEN) ? `XLEN : `NF)
module fcvt (
input logic XSgnE, // input's sign
input logic [`NE-1:0] XExpE, // input's exponent
input logic [`NF:0] XManE, // input's fraction
input logic [`XLEN-1:0] ForwardedSrcAE, // integer input - from IEU
input logic [2:0] FOpCtrlE, // choose which opperation (look below for values)
input logic FWriteIntE, // is fp->int (since it's writting to the integer register)
input logic XZeroE, // is the input zero
input logic XOrigDenormE, // is the input denormalized
input logic XInfE, // is the input infinity
input logic XNaNE, // is the input a NaN
input logic XSNaNE, // is the input a signaling NaN
input logic [2:0] FrmE, // rounding mode 000 = rount to nearest, ties to even 001 = round twords zero 010 = round down 011 = round up 100 = round to nearest, ties to max magnitude
input logic [`FPSIZES/3:0] FmtE, // the input's precision (11=quad 01=double 00=single 10=half)
output logic [`FLEN-1:0] CvtResE, // the fp to fp conversion's result
output logic [`XLEN-1:0] CvtIntResE, // the fp to fp conversion's result
output logic [4:0] CvtFlgE // the fp to fp conversion's flags
);
// 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
// (FF) fp -> fp coversion signals
// (IF) int -> fp coversion signals
// (FI) fp -> int coversion signals
logic [`FPSIZES/3:0] OutFmt; // format of the output
logic [`XLEN-1:0] PosInt; // the positive integer input
logic [`LGLEN-1:0] LzcIn; // input to the Leading Zero Counter (priority encoder)
logic [`NE:0] CalcExp; // the calculated expoent
logic [$clog2(`LGLEN):0] ShiftAmt; // how much to shift by
logic [`LGLEN+`NF:0] ShiftIn; // number to be shifted
logic ResDenormUf;// does the result underflow or is denormalized
logic ResUf; // does the result underflow
logic [`LGLEN+`NF:0] Shifted; // the shifted result
logic [`NE-2:0] NewBias; // the bias of the final result
logic [$clog2(`NF):0] ResNegNF; // the result's fraction length negated (-NF)
logic [`NE-1:0] OldExp; // the old exponent
logic ResSgn; // the result's sign
logic Sticky; // sticky bit - for rounding
logic Round; // round bit - for rounding
logic LSBFrac; // the least significant bit of the fraction - for rounding
logic CalcPlus1; // the calculated plus 1
logic Plus1; // add one to the final result?
logic [`FLEN-1:0] ShiftedPlus1; // plus one shifted to the proper position
logic [`NE:0] FullResExp; // the full result exponent (with the overflow bit)
logic [`NE-1:0] ResExp; // the result's exponent (trimmed to the correct size)
logic [`NF-1:0] ResFrac; // the result's fraction
logic [`XLEN+1:0] NegRes; // the negation of the result
logic [`XLEN-1:0] OfIntRes; // the overflow result for integer output
logic Overflow, Underflow, Inexact, Invalid; // flags
logic IntInexact, FpInexact, IntInvalid, FpInvalid; // flags for FP and int outputs
logic [`NE-1:0] MaxExp; // the maximum exponent before overflow
logic [1:0] NegResMSBS; // the negitive integer result's most significant bits
logic [`FLEN-1:0] NaNRes, InfRes, Res, UfRes; //various special results
logic KillRes; // kill the result?
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 ToInt; // is the opperation an fp->int conversion?
// seperate OpCtrl for code readability
assign Signed = FOpCtrlE[0];
assign Int64 = FOpCtrlE[1];
assign IntToFp = FOpCtrlE[2];
assign ToInt = FWriteIntE;
// choose the ouptut format depending on the opperation
// - fp -> fp: OpCtrl contains the percision of the output
// - int -> fp: FmtE contains the percision of the output
assign OutFmt = IntToFp ? FmtE : (FOpCtrlE[1:0] == `FMT);
///////////////////////////////////////////////////////////////////////////
// negation
///////////////////////////////////////////////////////////////////////////
// negate the input if the input is a negitive singed integer
// - remove leading ones if the input is a unsigned 32-bit integer
//
// Negitive input
// 64-bit input : negate the input
// 32-bit input : trim to 32-bits and negate the input
// Positive input
// 64-bit input : do nothing
// 32-bit input : trim to 32-bits
assign PosInt = ResSgn ? Int64 ? -ForwardedSrcAE : {{`XLEN-32{1'b0}}, -ForwardedSrcAE[31:0]} :
Int64 ? ForwardedSrcAE : {{`XLEN-32{1'b0}}, ForwardedSrcAE[31:0]};
///////////////////////////////////////////////////////////////////////////
// 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 LzcIn = IntToFp ? {PosInt, {`LGLEN-`XLEN{1'b0}}} : // I->F
{XManE[`NF-1:0], {`LGLEN-`NF{1'b0}}}; // F->F
// lglen is the largest possible value of ZeroCnt (NF or XLEN) hence normcnt must be log2(lglen) bits
logic [$clog2(`LGLEN):0] i, ZeroCnt;
always_comb begin
i = 0;
while (~LzcIn[`LGLEN-1-i] & i <= `LGLEN-1) i = i+1; // search for leading one
ZeroCnt = i;
end
///////////////////////////////////////////////////////////////////////////
// shifter
///////////////////////////////////////////////////////////////////////////
// F->F shift so the fraction is not denormalized
// Large->Small Denrom -> Norm Frac:
//
// | Frac | `NF zeros| << ShiftCnt
//
// Small->Large Norm -> Denorm Frac:
// - shift right so that the new-bias exponet = 1
// - so shift right by new-bias - 1 exponent
// - ie shift left by NF - 1 + new-bias exponent (if this is negitive then 0 is selected as a result later)
// - new-bias exponent is negitive
//
// | `NF-1 zeros |1| Frac | << NF + new-bias exponent
// | keep |
//
// Int -> Fp :
// | Int | `NF zeros| << ShiftCnt
// Fp -> Int :
// | `XLEN zeros | Man | << CalcExp
// seclect the input to the shifter
// fp -> int:
// | `XLEN zeros | Mantissa | 0's if nessisary |
// Other problems:
// - if shifting to the right (neg CalcExp) then don't a 1 in the round bit (to prevent an incorrect plus 1 later durring rounding)
// - we do however want to keep the one in the sticky bit so set one of bits in the sticky bit area to 1
// - ex: for the case 0010000.... (double)
// ??? -> fp:
// - if result is denormalized or underflowed then we want to normalize the result:
// | `NF zeros | Mantissa | 0's if nessisary |
// - otherwise:
// | lzcIn | 0's if nessisary |
assign ShiftIn = ToInt ? {{`XLEN{1'b0}}, XManE[`NF]&~CalcExp[`NE], XManE[`NF-1]|(CalcExp[`NE]&XManE[`NF]), XManE[`NF-2:0], {`LGLEN-`XLEN{1'b0}}} :
ResDenormUf ? {{`NF-1{1'b0}}, XManE, {`LGLEN-`NF+1{1'b0}}} : {LzcIn, {`NF+1{1'b0}}};
// 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 ZeroCnt+1 - 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?
assign ShiftAmt = ToInt ? CalcExp[$clog2(`LGLEN):0]&{$clog2(`LGLEN)+1{~CalcExp[`NE]}} :
ResDenormUf&~IntToFp ? ($clog2(`LGLEN)+1)'(`NF-1)+CalcExp[$clog2(`LGLEN):0] : (ZeroCnt+1)&{$clog2(`LGLEN)+1{XOrigDenormE|IntToFp}};
// shift
assign Shifted = ShiftIn << ShiftAmt;
///////////////////////////////////////////////////////////////////////////
// exp calculations
///////////////////////////////////////////////////////////////////////////
// fp -> int
// CalcExp = 1 - largest bias + 1 -
// *** 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
assign NewBias = ToInt ? (`NE-1)'(1) : OutFmt ? (`NE-1)'(`BIAS) : (`NE-1)'(`BIAS1);
// select the old exponent
// int -> fp : largest bias + XLEN
// fp -> ??? : XExp
assign OldExp = IntToFp ? (`NE)'(`BIAS)+(`NE)'(`XLEN) : XExpE;
// calculate CalcExp
// fp -> fp :
// - XExp - Largest bias + new bias
// fp -> int : XExp
// int -> fp : largest bias + XLEN
// the -XOrigDenorm is to take into account the correction (which had a plus 1)
assign CalcExp = {1'b0, OldExp} - (`NE+1)'(`BIAS) + {2'b0, NewBias} - {{`NE{1'b0}}, XOrigDenormE|IntToFp} - {{`NE-$clog2(`LGLEN){1'b0}}, (ZeroCnt&{$clog2(`LGLEN)+1{XOrigDenormE|IntToFp}})};
// if result is 0 or negitive
assign ResDenormUf = (~|CalcExp | CalcExp[`NE])&~XZeroE;
assign ResNegNF = (FOpCtrlE[1:0] == `FMT) ? -`NF : -`NF1;
// if the reuslt underflows and somthing is shifted out set the sticky bit
assign ResUf = ($signed(CalcExp) < $signed({{`NE-$clog2(`NF){1'b1}}, ResNegNF}))&~XZeroE;
///////////////////////////////////////////////////////////////////////////
// sign
///////////////////////////////////////////////////////////////////////////
assign ResSgn = IntToFp ? Int64 ? ForwardedSrcAE[`XLEN-1]&Signed : ForwardedSrcAE[31]&Signed : XSgnE;
///////////////////////////////////////////////////////////////////////////
// rounding
///////////////////////////////////////////////////////////////////////////
// round to nearest even
// {Round, Sticky}
// 0x - do nothing
// 10 - tie - Plus1 if result is odd (LSBNormSum = 1)
// 11 - Plus1
// round to zero - do nothing
// round to -infinity - Plus1 if negative
// round to infinity - Plus1 if positive
// round to nearest max magnitude
// {Guard, Round, Sticky}
// 0x - do nothing
// 1x - Plus1
// ResUf is used when a fp->fp result underflows but all the bits get shifted out, leaving nothing for the sticky bit
assign Sticky = ToInt ? |Shifted[`LGLEN+`NF-`XLEN-1:0] :
(OutFmt ? |Shifted[`LGLEN+`NF-`NF-1:0] : |Shifted[`LGLEN+`NF-`NF1-1:0])|ResUf;
assign Round = ToInt ? Shifted[`LGLEN+`NF-`XLEN] :
OutFmt ? Shifted[`LGLEN+`NF-`NF] : |Shifted[`LGLEN+`NF-`NF1];
assign LSBFrac = ToInt ? Shifted[`LGLEN+`NF-`XLEN+1] :
OutFmt ? Shifted[`LGLEN+`NF-`NF+1] : Shifted[`LGLEN+`NF-`NF1+1];
always_comb begin // ***remove guard bit
// Determine if you add 1
case (FrmE)
3'b000: CalcPlus1 = Round & (Sticky | LSBFrac);//round to nearest even
3'b001: CalcPlus1 = 0;//round to zero
3'b010: CalcPlus1 = ResSgn;//round down
3'b011: CalcPlus1 = ~ResSgn;//round up
3'b100: CalcPlus1 = Round;//round to nearest max magnitude
default: CalcPlus1 = 1'bx;
endcase
end
assign Plus1 = CalcPlus1&(Round|Sticky);
assign ShiftedPlus1 = OutFmt ? {{`FLEN-1{1'b0}},Plus1} : {{`NE+`NF1{1'b0}}, Plus1, {`FLEN-`NE-`NF1-1{1'b0}}};
// kill calcExp if the result is denormalized
assign {FullResExp, ResFrac} = {CalcExp&{`NE+1{~ResDenormUf}}, Shifted[`LGLEN+`NF:`LGLEN+`NF+1-`NF]} + ShiftedPlus1;
assign ResExp = FullResExp[`NE-1:0];
///////////////////////////////////////////////////////////////////////////
// flags
///////////////////////////////////////////////////////////////////////////
// calculate the flags
// dont set underflow overflow or inexact flags if result is NaN
assign MaxExp = ToInt ? Int64 ? 65 : 33 :
OutFmt ? {`NE{1'b1}} : {{`NE-`NE1{1'b0}}, {`NE1{1'b1}}};
// if the exponent is lager or equal to the maximum and it's not negitive
// F->F if the input is inf then the output is also Inf ie exact, so dont set the underflow flag
// if the result exponent is larger then the maximum possible exponent
// | and the exponent is positive
// | | and the input is not NaN or Infinity
// | | |
assign Overflow = ((ResExp >= MaxExp)&~CalcExp[`NE]&~(XNaNE|XInfE));
// only set the underflow flag if not-exact
// set the underflow flag if the result is denomal or underflowed
// can't underflow durring to integer conversions
// if the result is denormalized or underflowed
// | and the result did not round into normal values
// | | and the result is not exact
// | | | and the result isn't NaN
// | | | |
assign Underflow = ResDenormUf & ~(ResExp==1 & CalcExp == 0) & (Sticky|Round)&~(XNaNE);
// we are using the IEEE convertToIntegerExact opperations (rather then the exact ones) which do singal the inexact flag
// if there were bits thrown away
// | if overflowed or underflowed
// | | and if not a NaN
// | | |
assign FpInexact = (Sticky|Round|Underflow|Overflow)&(~XNaNE|IntToFp);
// if the result is too small to be represented and not 0
// | and if the result is not invalid (outside the integer bounds)
// | |
assign IntInexact = ((CalcExp[`NE]&~XZeroE)|Sticky|Round)&~Invalid;
assign Inexact = ToInt ? IntInexact : FpInexact;
// if an input was a singaling NaN(and we're using a FP input)
assign FpInvalid = (XSNaNE&~IntToFp);
assign NegResMSBS = Signed ? Int64 ? NegRes[`XLEN:`XLEN-1] : NegRes[32:31] :
Int64 ? NegRes[`XLEN+1:`XLEN] : NegRes[33:32];
// if the input is NaN or infinity
// | if the integer result overflows (out of range)
// | | if the input was negitive but ouputing to a unsigned number
// | | | the result doesn't round to zero
// | | | | or the result rounds up out of bounds
// | | | | and the result didn't underflow
// | | | | |
assign IntInvalid = XNaNE|XInfE|Overflow|((XSgnE&~Signed)&(~((CalcExp[`NE]|(~|CalcExp))&~Plus1)))|(NegResMSBS[1]^NegResMSBS[0]);
// |
// or when the positive result rounds up out of range
assign Invalid = ToInt ? IntInvalid : FpInvalid;
// pack the flags together and choose the result based on the opperation
// don't set the overflow or underfolw flags if converting to integer
assign CvtFlgE = {Invalid, 1'b0, Overflow&~ToInt, Underflow&~ToInt, Inexact};
///////////////////////////////////////////////////////////////////////////
// result selection
///////////////////////////////////////////////////////////////////////////
// when the input is zero for F->F the exponent is not calulated as 0 so combine with underflow result
//logic [$clog2(`NF)-1:0] MinDenormExp;
//assign MinDenormExp = FOpCtrlE[1:0] == `FMT ? -`NE : -`NE1;
assign KillRes = (ResUf|(XZeroE&~IntToFp)|(~|PosInt&IntToFp));
//assign NaNRes = FOpCtrlE[1:0] == `FMT ? {1'b0, {`NE+1{1'b1}}, (`NF-1)'(0)} : {{`FLEN-`LEN1{1'b1}}, 1'b0, {`NE1+1{1'b1}}, (`NF1-1)'(0)};
if(`IEEE754) begin
assign NaNRes = FOpCtrlE[1:0] == `FMT ? {1'b0, {`NE+1{1'b1}}, XManE[`NF-2:0]} : {{`FLEN-`LEN1{1'b1}}, 1'b0, {`NE1+1{1'b1}}, XManE[`NF-2:`NF-`NF1]};
end else begin
assign NaNRes = FOpCtrlE[1:0] == `FMT ? {1'b0, {`NE+1{1'b1}}, {`NF-1{1'b0}}} : {{`FLEN-`LEN1{1'b1}}, 1'b0, {`NE1+1{1'b1}}, {`NF1-1{1'b0}}};
end
// assign InfRes = FOpCtrlE[1:0] == `FMT ? {ResSgn, {`NE{1'b1}}, (`NF)'(0)} : {{`FLEN-`LEN1{1'b1}}, ResSgn, {`NE1{1'b1}}, (`NF1)'(0)};
// output one less then the maximum value if rounding down (RZ RU RD)
// if infinitiy output infinity
assign InfRes = FOpCtrlE[1:0] == `FMT ? ~XInfE&((FrmE[1:0]==2'b01) | (FrmE[1:0]==2'b10&~ResSgn) | (FrmE[1:0]==2'b11&ResSgn)) ? {ResSgn, {`NE-1{1'b1}}, 1'b0, {`NF{1'b1}}} :
{ResSgn, {`NE{1'b1}}, {`NF{1'b0}}} :
~XInfE&((FrmE[1:0]==2'b01) | (FrmE[1:0]==2'b10&~ResSgn) | (FrmE[1:0]==2'b11&ResSgn)) ? {{`FLEN-`LEN1{1'b1}}, ResSgn, {`NE1-1{1'b1}}, 1'b0, {`NF1{1'b1}}} :
{{`FLEN-`LEN1{1'b1}}, ResSgn, {`NE1{1'b1}}, (`NF1)'(0)};
// if RU/RD then round the underflowed result if needed
// integer zero's exponent is not calculated corresctly so go through underflow result
assign UfRes = OutFmt ? {ResSgn, (`FLEN-2)'(0), Plus1&FrmE[1]} : {{`FLEN-`LEN1{1'b1}}, ResSgn, (`LEN1-2)'(0), Plus1&FrmE[1]};
assign Res = OutFmt ? {ResSgn, ResExp, ResFrac} : {{`FLEN-`LEN1{1'b1}}, ResSgn, ResExp[`NE1-1:0], ResFrac[`NF-1:`NF-`NF1]};
assign CvtResE = XNaNE&~IntToFp ? NaNRes :
(XInfE|Overflow)&~IntToFp ? InfRes :
KillRes ? UfRes :
Res;
// *** probably can optimize the negation
// NaNs sould ouput the same as a positive infinity
// a 32bit unsigend result should be sign extended (as if it is not a unsigned number)
assign OfIntRes = Signed ? XSgnE&~XNaNE ? Int64 ? {1'b1, {`XLEN-1{1'b0}}} : {{`XLEN-32{1'b1}}, 1'b1, {31{1'b0}}} : // signed negitive
Int64 ? {1'b0, {`XLEN-1{1'b1}}} : {{`XLEN-32{1'b0}}, 1'b0, {31{1'b1}}} : // signed positive
XSgnE&~XNaNE ? {`XLEN{1'b0}} : // unsigned negitive
{`XLEN{1'b1}};// unsigned positive
assign NegRes = XSgnE ? -({2'b0, Shifted[`LGLEN+`NF:`LGLEN+`NF+1-`XLEN]}+{{`XLEN+1{1'b0}},Plus1}) : {2'b0, Shifted[`LGLEN+`NF:`LGLEN+`NF+1-`XLEN]}+{{`XLEN+1{1'b0}},Plus1};
assign CvtIntResE = Invalid ? OfIntRes :
CalcExp[`NE] ? XSgnE&Signed&Plus1 ? {{`XLEN{1'b1}}} : {{`XLEN-1{1'b0}}, Plus1} : //CalcExp has to come after invalid ***swap to actual mux at some point??
Int64 ? NegRes[`XLEN-1:0] : {{`XLEN-32{NegRes[31]}}, NegRes[31:0]};
endmodule