mirror of
https://github.com/openhwgroup/cvw
synced 2025-02-03 10:15:19 +00:00
cvtfp module documented
This commit is contained in:
Katherine Parry
parent
6b30adb309
commit
33e5a078bf
@ -1,24 +1,42 @@
|
||||
|
||||
// `include "wally-config.vh"
|
||||
module cvtfp (
|
||||
input logic [10:0] XExpE,
|
||||
input logic [52:0] XManE,
|
||||
input logic XSgnE,
|
||||
input logic XZeroE,
|
||||
input logic XDenormE,
|
||||
input logic XInfE,
|
||||
input logic XNaNE,
|
||||
input logic XSNaNE,
|
||||
input logic [2:0] FrmE,
|
||||
input logic FmtE,
|
||||
output logic [63:0] CvtFpResE,
|
||||
output logic [4:0] CvtFpFlgE);
|
||||
input logic [10:0] XExpE, // input's exponent
|
||||
input logic [52:0] XManE, // input's mantissa
|
||||
input logic XSgnE, // input's sign
|
||||
input logic XZeroE, // is the input zero
|
||||
input logic XDenormE, // 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 FmtE, // the input's precision (1 = double 0 = single)
|
||||
output logic [63:0] CvtFpResE, // the fp to fp conversion's result
|
||||
output logic [4:0] CvtFpFlgE); // the fp to fp conversion's flags
|
||||
|
||||
logic [7:0] DExp;
|
||||
logic [51:0] Frac;
|
||||
logic Denorm;
|
||||
logic [12:0] DSExp; // double to single precision exponent
|
||||
logic Denorm; // is the double to single precision result denormalized
|
||||
logic Shift; // do you shift the double precision exponent (if single precision result is denormalized)
|
||||
logic [51:0] SDFrac; // single to double precision fraction
|
||||
logic [25:0] DSFrac; // double to single precision fraction
|
||||
logic [77:0] DSFracShifted; // single precision fraction shifted for double precision
|
||||
logic Sticky, UfSticky, Guard, Round, LSBFrac, UfGuard, UfRound, UfLSBFrac; // rounding bits
|
||||
logic CalcPlus1, UfCalcPlus1, Plus1, UfPlus1; // do you add one to the result
|
||||
logic [12:0] DSExpFull; // full double to single exponent
|
||||
logic [22:0] DSResFrac; // final double to single fraction
|
||||
logic [7:0] DSResExp; // final double to single exponent
|
||||
logic [10:0] SDExp; // final single to double precision exponent
|
||||
logic Overflow, Underflow, Inexact; // flags
|
||||
logic [31:0] DSRes; // double to single precision result
|
||||
|
||||
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
// LZC
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
|
||||
// LZC - find the first 1 in the input's mantissa
|
||||
logic [8:0] i,NormCnt;
|
||||
always_comb begin
|
||||
i = 0;
|
||||
@ -27,42 +45,62 @@ module cvtfp (
|
||||
end
|
||||
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
// Expoents
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
// convert the single precion exponent to single precision.
|
||||
// - subtract the double precision exponent (1023) and add the
|
||||
// single precsision exponent (127)
|
||||
// - if the input is zero then kill the exponent
|
||||
|
||||
assign DSExp = ({2'b0,XExpE}-13'd1023+13'd127)&{13{~XZeroE}};
|
||||
|
||||
// is the converted double to single precision exponent in the denormalized range
|
||||
assign Denorm = $signed(DSExp) <= 0 & $signed(DSExp) > $signed(-(13'd23));
|
||||
|
||||
|
||||
// caluculate the final single to double precsion exponent
|
||||
// - subtract the single precision bias (127) and add the double
|
||||
// precision bias (127)
|
||||
// - if the result is zero or denormalized, kill the exponent
|
||||
assign SDExp = XExpE-({2'b0,NormCnt&{9{~XZeroE}}})+({11{XDenormE}}&1024-127); //*** seems ineffecient
|
||||
|
||||
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
// Fraction
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
|
||||
// normalize the single precision fraction for double precsion
|
||||
// - needed for denormal single precsion values
|
||||
assign SDFrac = XManE[51:0] << NormCnt;
|
||||
|
||||
// check if the double precision mantissa needs to be shifted
|
||||
// - the mantissa needs to be shifted if the single precision result is denormal
|
||||
assign Shift = Denorm | (($signed(DSExp) > $signed(-(13'd25))) & DSExp[12]);
|
||||
// shift the mantissa
|
||||
assign DSFracShifted = {XManE, 25'b0} >> ((-DSExp+1)&{13{Shift}}); //***might be some optimization here
|
||||
assign DSFrac = DSFracShifted[76:51];
|
||||
|
||||
|
||||
|
||||
logic [12:0] DExpCalc;
|
||||
// logic Overflow, Underflow;
|
||||
assign DExpCalc = ({2'b0,XExpE}-13'd1023+13'd127)&{13{~XZeroE}};
|
||||
assign Denorm = $signed(DExpCalc) <= 0 & $signed(DExpCalc) > $signed(-(13'd23));
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
// Rounder
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
logic [12:0] ShiftCnt;
|
||||
logic [51:0] SFrac;
|
||||
logic [25:0] DFrac;
|
||||
logic [77:0] DFracTmp;
|
||||
//assign ShiftCnt = FmtE ? -DExpCalc&{13{Denorm}} : NormCnt;
|
||||
assign SFrac = XManE[51:0] << NormCnt;
|
||||
logic Shift;
|
||||
assign Shift = Denorm | (($signed(DExpCalc) > $signed(-(13'd25))) & DExpCalc[12]);
|
||||
assign DFracTmp = {XManE, 25'b0} >> ((-DExpCalc+1)&{13{Shift}});
|
||||
assign DFrac = DFracTmp[76:51];
|
||||
|
||||
logic Sticky, UfSticky, Guard, Round, LSBFrac, UfGuard, UfRound, UfLSBFrac;
|
||||
logic CalcPlus1, UfCalcPlus1;
|
||||
logic Plus1, UfPlus1;
|
||||
// used to determine underflow flag
|
||||
assign UfSticky = |DFracTmp[50:0];
|
||||
assign UfGuard = DFrac[1];
|
||||
assign UfRound = DFrac[0];
|
||||
assign UfLSBFrac = DFrac[2];
|
||||
assign UfSticky = |DSFracShifted[50:0];
|
||||
assign UfGuard = DSFrac[1];
|
||||
assign UfRound = DSFrac[0];
|
||||
assign UfLSBFrac = DSFrac[2];
|
||||
|
||||
|
||||
assign Sticky = UfSticky | UfRound;
|
||||
assign Guard = DFrac[2];
|
||||
assign Round = DFrac[1];
|
||||
assign LSBFrac = DFrac[3];
|
||||
assign Guard = DSFrac[2];
|
||||
assign Round = DSFrac[1];
|
||||
assign LSBFrac = DSFrac[3];
|
||||
|
||||
|
||||
always_comb begin
|
||||
@ -87,31 +125,47 @@ assign DFrac = DFracTmp[76:51];
|
||||
|
||||
end
|
||||
|
||||
// If an answer is exact don't round
|
||||
// if an answer is exact don't round
|
||||
assign Plus1 = CalcPlus1 & (Sticky | UfGuard | Guard | Round);
|
||||
assign UfPlus1 = UfCalcPlus1 & (Sticky | UfGuard);
|
||||
logic [12:0] DExpFull;
|
||||
logic [22:0] DResFrac;
|
||||
logic [7:0] DResExp;
|
||||
assign {DExpFull, DResFrac} = {DExpCalc&{13{~Denorm}}, DFrac[25:3]} + {35'b0,Plus1};
|
||||
assign DResExp = DExpFull[7:0];
|
||||
|
||||
logic [10:0] SExp;
|
||||
assign SExp = XExpE-({2'b0,NormCnt&{9{~XZeroE}}})+({11{XDenormE}}&1024-127);
|
||||
|
||||
logic Overflow, Underflow, Inexact;
|
||||
assign Overflow = $signed(DExpFull) >= $signed({5'b0, {8{1'b1}}}) & ~(XNaNE|XInfE);
|
||||
assign Underflow = (($signed(DExpFull) <= 0) & ((Sticky|Guard|Round) | (XManE[52]&~|DFrac) | (|DFrac&~Denorm)) | ((DExpFull == 1) & Denorm & ~(UfPlus1&UfLSBFrac))) & ~(XNaNE|XInfE);
|
||||
|
||||
// round the double to single precision result
|
||||
assign {DSExpFull, DSResFrac} = {DSExp&{13{~Denorm}}, DSFrac[25:3]} + {35'b0,Plus1};
|
||||
assign DSResExp = DSExpFull[7:0];
|
||||
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
// Flags
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
// calculate the flags
|
||||
// - overflow, underflow and inexact can only be set by the double to single precision opperation
|
||||
// - don't set underflow or overflow if the input is NaN or Infinity
|
||||
// - don't set the inexact flag if the input is NaN
|
||||
assign Overflow = $signed(DSExpFull) >= $signed({5'b0, {8{1'b1}}}) & ~(XNaNE|XInfE);
|
||||
assign Underflow = (($signed(DSExpFull) <= 0) & ((Sticky|Guard|Round) | (XManE[52]&~|DSFrac) | (|DSFrac&~Denorm)) | ((DSExpFull == 1) & Denorm & ~(UfPlus1&UfLSBFrac))) & ~(XNaNE|XInfE);
|
||||
assign Inexact = (Sticky|Guard|Round|Underflow|Overflow) &~(XNaNE);
|
||||
|
||||
logic [31:0] DRes;
|
||||
assign DRes = XNaNE ? {XSgnE, {8{1'b1}}, 1'b1, XManE[50:29]} :
|
||||
// pack the flags together and choose the result based on the opperation
|
||||
assign CvtFpFlgE = FmtE ? {XSNaNE, 1'b0, Overflow, Underflow, Inexact} : {XSNaNE, 4'b0};
|
||||
|
||||
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
// Result Selection
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
// select the double to single precision result
|
||||
assign DSRes = XNaNE ? {XSgnE, {8{1'b1}}, 1'b1, XManE[50:29]} :
|
||||
Underflow & ~Denorm ? {XSgnE, 30'b0, CalcPlus1&(|FrmE[1:0]|Shift)} :
|
||||
Overflow | XInfE ? ((FrmE[1:0]==2'b01) | (FrmE[1:0]==2'b10&~XSgnE) | (FrmE[1:0]==2'b11&XSgnE)) & ~XInfE ? {XSgnE, 8'hfe, {23{1'b1}}} :
|
||||
{XSgnE, 8'hff, 23'b0} :
|
||||
{XSgnE, DResExp, DResFrac};
|
||||
assign CvtFpResE = FmtE ? {{32{1'b1}},DRes} : {XSgnE, SExp, SFrac[51]|XNaNE, SFrac[50:0]};
|
||||
assign CvtFpFlgE = FmtE ? {XSNaNE, 1'b0, Overflow, Underflow, Inexact} : {XSNaNE, 4'b0};
|
||||
{XSgnE, DSResExp, DSResFrac};
|
||||
|
||||
// select the final result based on the opperation
|
||||
assign CvtFpResE = FmtE ? {{32{1'b1}},DSRes} : {XSgnE, SDExp, SDFrac[51]|XNaNE, SDFrac[50:0]};
|
||||
|
||||
endmodule // fpadd
|
||||
|
||||
|
||||
Loading…
Reference in New Issue
Block a user