/////////////////////////////////////////// // // Written: Katherine Parry, David Harris // Modified: 6/23/2021 // // Purpose: Floating point multiply-accumulate of configurable size // // A component of the Wally configurable RISC-V project. // // Copyright (C) 2021 Harvey Mudd College & Oklahoma State University // // 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 A 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" // `include "../../../config/rv64icfd/wally-config.vh" module fma( input logic clk, input logic reset, input logic FlushM, // flush the memory stage input logic StallM, // stall memory stage input logic FmtE, FmtM, // precision 1 = double 0 = single input logic [2:0] FOpCtrlE, // 000 = fmadd (X*Y)+Z, 001 = fmsub (X*Y)-Z, 010 = fnmsub -(X*Y)+Z, 011 = fnmadd -(X*Y)-Z, 100 = fmul (X*Y) input logic [2:0] FrmM, // 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 XSgnE, YSgnE, ZSgnE, // input signs - execute stage input logic [`NE-1:0] XExpE, YExpE, ZExpE, // input exponents - execute stage input logic [`NF:0] XManE, YManE, ZManE, // input mantissa - execute stage input logic XSgnM, YSgnM, // input signs - memory stage input logic [`NE-1:0] XExpM, YExpM, ZExpM, // input exponents - memory stage input logic [`NF:0] XManM, YManM, ZManM, // input mantissa - memory stage input logic XDenormE, YDenormE, ZDenormE, // is denorm input logic XZeroE, YZeroE, ZZeroE, // is zero - execute stage input logic XNaNM, YNaNM, ZNaNM, // is NaN input logic XSNaNM, YSNaNM, ZSNaNM, // is signaling NaN input logic XZeroM, YZeroM, ZZeroM, // is zero - memory stage input logic XInfM, YInfM, ZInfM, // is infinity input logic [10:0] BiasE, // bias (max exponent/2) ***parameterize in unpacking unit output logic [`FLEN-1:0] FMAResM, // FMA result output logic [4:0] FMAFlgM); // FMA flags //fma/mult/add // fmadd = 000 // fmsub = 001 // fnmsub = 010 -(a*b)+c // fnmadd = 011 -(a*b)-c // fmul = 100 // fadd = 110 // fsub = 111 // signals transfered between pipeline stages logic [3*`NF+5:0] SumE, SumM; logic [`NE+1:0] ProdExpE, ProdExpM; logic AddendStickyE, AddendStickyM; logic KillProdE, KillProdM; logic InvZE, InvZM; logic NegSumE, NegSumM; logic ZSgnEffE, ZSgnEffM; logic PSgnE, PSgnM; logic [8:0] NormCntE, NormCntM; fma1 fma1 (.XSgnE, .YSgnE, .ZSgnE, .XExpE, .YExpE, .ZExpE, .XManE, .YManE, .ZManE, .BiasE, .XDenormE, .YDenormE, .ZDenormE, .XZeroE, .YZeroE, .ZZeroE, .FOpCtrlE, .FmtE, .SumE, .NegSumE, .InvZE, .NormCntE, .ZSgnEffE, .PSgnE, .ProdExpE, .AddendStickyE, .KillProdE); // E/M pipeline registers flopenrc #(3*`NF+6) EMRegFma2(clk, reset, FlushM, ~StallM, SumE, SumM); flopenrc #(13) EMRegFma3(clk, reset, FlushM, ~StallM, ProdExpE, ProdExpM); flopenrc #(15) EMRegFma4(clk, reset, FlushM, ~StallM, {AddendStickyE, KillProdE, InvZE, NormCntE, NegSumE, ZSgnEffE, PSgnE}, {AddendStickyM, KillProdM, InvZM, NormCntM, NegSumM, ZSgnEffM, PSgnM}); fma2 fma2(.XSgnM, .YSgnM, .XExpM, .YExpM, .ZExpM, .XManM, .YManM, .ZManM, .FrmM, .FmtM, .ProdExpM, .AddendStickyM, .KillProdM, .SumM, .NegSumM, .InvZM, .NormCntM, .ZSgnEffM, .PSgnM, .XZeroM, .YZeroM, .ZZeroM, .XInfM, .YInfM, .ZInfM, .XNaNM, .YNaNM, .ZNaNM, .XSNaNM, .YSNaNM, .ZSNaNM, .FMAResM, .FMAFlgM); endmodule module fma1( input logic XSgnE, YSgnE, ZSgnE, // input's signs input logic [`NE-1:0] XExpE, YExpE, ZExpE, // biased exponents in B(NE.0) format input logic [`NF:0] XManE, YManE, ZManE, // fractions in U(0.NF) format input logic XDenormE, YDenormE, ZDenormE, // is the input denormal input logic XZeroE, YZeroE, ZZeroE, // is the input zero input logic [`NE-1:0] BiasE, // bias (max exponent/2) input logic [2:0] FOpCtrlE, // 000 = fmadd (X*Y)+Z, 001 = fmsub (X*Y)-Z, 010 = fnmsub -(X*Y)+Z, 011 = fnmadd -(X*Y)-Z, 100 = fmul (X*Y) input logic FmtE, // precision 1 = double 0 = single output logic [`NE+1:0] ProdExpE, // X exponent + Y exponent - bias in B(NE+2.0) format; adds 2 bits to allow for size of number and negative sign output logic AddendStickyE, // sticky bit that is calculated during alignment output logic KillProdE, // set the product to zero before addition if the product is too small to matter output logic [3*`NF+5:0] SumE, // the positive sum output logic NegSumE, // was the sum negitive output logic InvZE, // intert Z output logic ZSgnEffE, // the modified Z sign output logic PSgnE, // the product's sign output logic [8:0] NormCntE // normalization shift cnt ); logic [`NE-1:0] Denorm; // value of a denormaized number based on precision logic [`NE-1:0] XExpVal, YExpVal; // Exponent value after taking into account denormals logic [2*`NF+1:0] ProdManE; // 1.X frac * 1.Y frac in U(2.2Nf) format logic [3*`NF+5:0] AlignedAddendE; // Z aligned for addition in U(NF+5.2NF+1) /////////////////////////////////////////////////////////////////////////////// // Calculate the product // - When multipliying two fp numbers, add the exponents // - Subtract the bias (XExp + YExp has two biases, one from each exponent) // - If the product is zero then kill the exponent - this is a problem /////////////////////////////////////////////////////////////////////////////// // denormalized numbers have diffrent values depending on which precison it is. // double - 1 // single - 1024-128+1 = 897 assign Denorm = FmtE ? 1 : 897; assign XExpVal = XDenormE ? Denorm : XExpE; assign YExpVal = YDenormE ? Denorm : YExpE; // take into account if the product is zero, the product's exponent does not compute properly if X or Y is zero assign ProdExpE = (XExpVal + YExpVal - BiasE)&{`NE+2{~(XZeroE|YZeroE)}}; // multiplication of the mantissa's mult mult(.XManE, .YManE, .ProdManE); /////////////////////////////////////////////////////////////////////////////// // Alignment shifter /////////////////////////////////////////////////////////////////////////////// alignshift alignshift(.ZExpE, .ZManE, .ZDenormE, .XZeroE, .YZeroE, .ZZeroE, .ProdExpE, .Denorm, .AlignedAddendE, .AddendStickyE, .KillProdE); // Calculate the product's sign // Negate product's sign if FNMADD or FNMSUB assign PSgnE = XSgnE ^ YSgnE ^ (FOpCtrlE[1]&~FOpCtrlE[2]); assign ZSgnEffE = ZSgnE^FOpCtrlE[0]; // Swap sign of Z for subtract // /////////////////////////////////////////////////////////////////////////////// // // Addition/LZA // /////////////////////////////////////////////////////////////////////////////// fmaadd fmaadd(.AlignedAddendE, .ProdManE, .PSgnE, .ZSgnEffE, .KillProdE, .SumE, .NegSumE, .InvZE, .NormCntE, .XZeroE, .YZeroE); endmodule module fma2( input logic XSgnM, YSgnM, // input signs input logic [`NE-1:0] XExpM, YExpM, ZExpM, // input exponents input logic [`NF:0] XManM, YManM, ZManM, // input mantissas input logic [2:0] FrmM, // 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 FmtM, // precision 1 = double 0 = single input logic [`NE+1:0] ProdExpM, // X exponent + Y exponent - bias input logic AddendStickyM, // sticky bit that is calculated during alignment input logic KillProdM, // set the product to zero before addition if the product is too small to matter input logic XZeroM, YZeroM, ZZeroM, // inputs are zero input logic XInfM, YInfM, ZInfM, // inputs are infinity input logic XNaNM, YNaNM, ZNaNM, // inputs are NaN input logic XSNaNM, YSNaNM, ZSNaNM, // inputs are signaling NaNs input logic [3*`NF+5:0] SumM, // the positive sum input logic NegSumM, // was the sum negitive input logic InvZM, // do you invert Z input logic ZSgnEffM, // the modified Z sign - depends on instruction input logic PSgnM, // the product's sign input logic [8:0] NormCntM, // the normalization shift count output logic [`FLEN-1:0] FMAResM, // FMA final result output logic [4:0] FMAFlgM); // FMA flags {invalid, divide by zero, overflow, underflow, inexact} logic [`NF-1:0] ResultFrac; // Result fraction logic [`NE-1:0] ResultExp; // Result exponent logic ResultSgn; // Result sign logic [`NE+1:0] SumExp; // exponent of the normalized sum logic [`NE+1:0] FullResultExp; // ResultExp with bits to determine sign and overflow logic [`NF+2:0] NormSum; // normalized sum logic NormSumSticky; // sticky bit calulated from the normalized sum logic SumZero; // is the sum zero logic ResultDenorm; // is the result denormalized logic Sticky, UfSticky; // Sticky bit logic Plus1, Minus1, CalcPlus1; // do you add or subtract one for rounding logic UfPlus1; // do you add one (for determining underflow flag) logic Invalid,Underflow,Overflow; // flags logic ZeroSgn; // the result's sign if the sum is zero logic ResultSgnTmp; // the result's sign assuming the result is not zero logic Guard, Round; // bits needed to determine rounding logic UfRound, UfLSBNormSum; // bits needed to determine rounding for underflow flag logic [`FLEN-1:0] XNaNResult, YNaNResult, ZNaNResult, InvalidResult, OverflowResult, KillProdResult, UnderflowResult; // possible results /////////////////////////////////////////////////////////////////////////////// // Normalization /////////////////////////////////////////////////////////////////////////////// normalize normalize(.SumM, .ZExpM, .ProdExpM, .NormCntM, .FmtM, .KillProdM, .AddendStickyM, .NormSum, .SumZero, .NormSumSticky, .UfSticky, .SumExp, .ResultDenorm); /////////////////////////////////////////////////////////////////////////////// // Rounding /////////////////////////////////////////////////////////////////////////////// // round to nearest even // round to zero // round to -infinity // round to infinity // round to nearest max magnitude fmaround fmaround(.FmtM, .FrmM, .Sticky, .UfSticky, .NormSum, .AddendStickyM, .NormSumSticky, .ZZeroM, .InvZM, .ResultSgn, .SumExp, .CalcPlus1, .Plus1, .UfPlus1, .Minus1, .FullResultExp, .ResultFrac, .ResultExp, .Round, .Guard, .UfRound, .UfLSBNormSum); /////////////////////////////////////////////////////////////////////////////// // Sign calculation /////////////////////////////////////////////////////////////////////////////// // Determine the sign if the sum is zero // if cancelation then 0 unless round to -infinity // otherwise psign assign ZeroSgn = (PSgnM^ZSgnEffM)&~Underflow ? FrmM[1:0] == 2'b10 : PSgnM; // is the result negitive // if p - z is the Sum negitive // if -p + z is the Sum positive // if -p - z then the Sum is negitive assign ResultSgnTmp = InvZM&(ZSgnEffM)&NegSumM | InvZM&PSgnM&~NegSumM | ((ZSgnEffM)&PSgnM); assign ResultSgn = SumZero ? ZeroSgn : ResultSgnTmp; /////////////////////////////////////////////////////////////////////////////// // Flags /////////////////////////////////////////////////////////////////////////////// fmaflags fmaflags(.XSNaNM, .YSNaNM, .ZSNaNM, .XInfM, .YInfM, .ZInfM, .XZeroM, .YZeroM, .XNaNM, .YNaNM, .ZNaNM, .FullResultExp, .SumExp, .ZSgnEffM, .PSgnM, .Round, .Guard, .UfRound, .UfLSBNormSum, .Sticky, .UfPlus1, .FmtM, .Invalid, .Overflow, .Underflow, .FMAFlgM); /////////////////////////////////////////////////////////////////////////////// // Select the result /////////////////////////////////////////////////////////////////////////////// assign XNaNResult = FmtM ? {XSgnM, XExpM, 1'b1, XManM[`NF-2:0]} : {{32{1'b1}}, XSgnM, XExpM[7:0], 1'b1, XManM[50:29]}; assign YNaNResult = FmtM ? {YSgnM, YExpM, 1'b1, YManM[`NF-2:0]} : {{32{1'b1}}, YSgnM, YExpM[7:0], 1'b1, YManM[50:29]}; assign ZNaNResult = FmtM ? {ZSgnEffM, ZExpM, 1'b1, ZManM[`NF-2:0]} : {{32{1'b1}}, ZSgnEffM, ZExpM[7:0], 1'b1, ZManM[50:29]}; assign OverflowResult = FmtM ? ((FrmM[1:0]==2'b01) | (FrmM[1:0]==2'b10&~ResultSgn) | (FrmM[1:0]==2'b11&ResultSgn)) ? {ResultSgn, {`NE-1{1'b1}}, 1'b0, {`NF{1'b1}}} : {ResultSgn, {`NE{1'b1}}, {`NF{1'b0}}} : ((FrmM[1:0]==2'b01) | (FrmM[1:0]==2'b10&~ResultSgn) | (FrmM[1:0]==2'b11&ResultSgn)) ? {{32{1'b1}}, ResultSgn, 8'hfe, {23{1'b1}}} : {{32{1'b1}}, ResultSgn, 8'hff, 23'b0}; assign InvalidResult = FmtM ? {ResultSgn, {`NE{1'b1}}, 1'b1, {`NF-1{1'b0}}} : {{32{1'b1}}, ResultSgn, 8'hff, 1'b1, 22'b0}; assign KillProdResult = FmtM ? {ResultSgn, {ZExpM, ZManM[`NF-1:0]} - (Minus1&AddendStickyM) + (Plus1&AddendStickyM)} : {{32{1'b1}}, ResultSgn, {ZExpM[`NE-1],ZExpM[6:0], ZManM[51:29]} - {30'b0, (Minus1&AddendStickyM)} + {30'b0, (Plus1&AddendStickyM)}}; assign UnderflowResult = FmtM ? {ResultSgn, {`FLEN-1{1'b0}}} + (CalcPlus1&(AddendStickyM|FrmM[1])) : {{32{1'b1}}, {ResultSgn, 31'b0} + {31'b0, (CalcPlus1&(AddendStickyM|FrmM[1]))}}; assign FMAResM = XNaNM ? XNaNResult : YNaNM ? YNaNResult : ZNaNM ? ZNaNResult : Invalid ? InvalidResult : XInfM ? FmtM ? {PSgnM, XExpM, XManM[`NF-1:0]} : {{32{1'b1}}, PSgnM, XExpM[7:0], XManM[51:29]} : YInfM ? FmtM ? {PSgnM, YExpM, YManM[`NF-1:0]} : {{32{1'b1}}, PSgnM, YExpM[7:0], YManM[51:29]} : ZInfM ? FmtM ? {ZSgnEffM, ZExpM, ZManM[`NF-1:0]} : {{32{1'b1}}, ZSgnEffM, ZExpM[7:0], ZManM[51:29]} : KillProdM ? KillProdResult : Overflow ? OverflowResult : Underflow & ~ResultDenorm & (ResultExp!=1) ? UnderflowResult : FmtM ? {ResultSgn, ResultExp, ResultFrac} : {{32{1'b1}}, ResultSgn, ResultExp[7:0], ResultFrac[51:29]}; // *** use NF where needed endmodule module mult( input logic [`NF:0] XManE, YManE, output logic [2*`NF+1:0] ProdManE ); assign ProdManE = XManE * YManE; endmodule module alignshift( input logic [`NE-1:0] ZExpE, // biased exponents in B(NE.0) format input logic [`NF:0] ZManE, // fractions in U(0.NF) format] input logic ZDenormE, // is the input denormal input logic XZeroE, YZeroE, ZZeroE, // is the input zero input logic [`NE+1:0] ProdExpE, // the product's exponent input logic [`NE-1:0] Denorm, // the biased value of a denormalized number output logic [3*`NF+5:0] AlignedAddendE, // Z aligned for addition in U(NF+5.2NF+1) output logic AddendStickyE, // Sticky bit calculated from the aliged addend output logic KillProdE // should the product be set to zero ); logic [`NE+1:0] AlignCnt; // how far to shift the addend to align with the product in Q(NE+2.0) format logic [4*`NF+5:0] ZManShifted; // output of the alignment shifter including sticky bits U(NF+5.3NF+1) logic [4*`NF+5:0] ZManPreShifted; // input to the alignment shifter U(NF+5.3NF+1) logic [`NE-1:0] ZExpVal; // Exponent value after taking into account denormals /////////////////////////////////////////////////////////////////////////////// // Alignment shifter /////////////////////////////////////////////////////////////////////////////// // determine the shift count for alignment // - negitive means Z is larger, so shift Z left // - positive means the product is larger, so shift Z right // - Denormal numbers have a diffrent exponent value depending on the precision assign ZExpVal = ZDenormE ? Denorm : ZExpE; assign AlignCnt = ProdExpE - ZExpVal + (`NF+3); // Defualt Addition without shifting // | 54'b0 | 106'b(product) | 2'b0 | // | addnend | // the 1'b0 before the added is because the product's mantissa has two bits before the binary point (xx.xxxxxxxxxx...) assign ZManPreShifted = {ZManE,(3*`NF+5)'(0)}; always_comb begin // If the product is too small to effect the sum, kill the product // | 54'b0 | 106'b(product) | 2'b0 | // | addnend | if ($signed(AlignCnt) < $signed(0)) begin KillProdE = 1; ZManShifted = ZManPreShifted; AddendStickyE = ~(XZeroE|YZeroE); // If the Addend is shifted right // | 54'b0 | 106'b(product) | 2'b0 | // | addnend | end else if ($signed(AlignCnt)<=$signed(3*`NF+4)) begin KillProdE = 0; ZManShifted = ZManPreShifted >> AlignCnt; AddendStickyE = |(ZManShifted[`NF-1:0]); // If the addend is too small to effect the addition // - The addend has to shift two past the end of the addend to be considered too small // - The 2 extra bits are needed for rounding // | 54'b0 | 106'b(product) | 2'b0 | // | addnend | end else begin KillProdE = 0; ZManShifted = 0; AddendStickyE = ~ZZeroE; end end assign AlignedAddendE = ZManShifted[4*`NF+5:`NF]; endmodule module fmaadd( input logic [3*`NF+5:0] AlignedAddendE, // Z aligned for addition in U(NF+5.2NF+1) input logic [2*`NF+1:0] ProdManE, // the product's mantissa input logic PSgnE, ZSgnEffE,// the product and modified Z signs input logic KillProdE, // should the product be set to 0 input logic XZeroE, YZeroE, // is the input zero output logic [3*`NF+5:0] SumE, // the positive sum output logic NegSumE, // was the sum negitive output logic InvZE, // do you invert Z output logic [8:0] NormCntE // normalization shift count ); logic [3*`NF+6:0] PreSum, NegPreSum; // possibly negitive sum logic [2*`NF+1:0] ProdMan2; // product being added logic [3*`NF+6:0] AlignedAddend2; // possibly inverted aligned Z logic [3*`NF+6:0] NegProdMan2; // a negated ProdMan2 logic [8:0] PNormCnt, NNormCnt; // results from the LZA /////////////////////////////////////////////////////////////////////////////// // Addition /////////////////////////////////////////////////////////////////////////////// // Negate Z when doing one of the following opperations: // -prod + Z // prod - Z assign InvZE = ZSgnEffE ^ PSgnE; // Choose an inverted or non-inverted addend - the one has to be added now for the LZA assign AlignedAddend2 = InvZE ? -{1'b0, AlignedAddendE} : {1'b0, AlignedAddendE}; // Kill the product if the product is too small to effect the addition (determined in fma1.sv) assign ProdMan2 = ProdManE&{2*`NF+2{~KillProdE}}; // Negate ProdMan for LZA and the negitive sum calculation assign NegProdMan2 = {{`NF+3{~(XZeroE|YZeroE|KillProdE)}}, -ProdMan2, 2'b0}; // LZAs one for the positive result and one for the negitive // - the +1 from inverting causes problems for normalization poslza poslza(AlignedAddend2, ProdMan2, PNormCnt); neglza neglza({1'b0,AlignedAddendE}, NegProdMan2, NNormCnt); // Do the addition // - calculate a positive and negitive sum in parallel assign PreSum = AlignedAddend2 + {ProdMan2, 2'b0}; assign NegPreSum = AlignedAddendE + NegProdMan2; // Is the sum negitive assign NegSumE = PreSum[3*`NF+6]; // Choose the positive sum and accompanying LZA result. assign SumE = NegSumE ? NegPreSum[3*`NF+5:0] : PreSum[3*`NF+5:0]; assign NormCntE = NegSumE ? NNormCnt : PNormCnt; endmodule module poslza( input logic [3*`NF+6:0] A, // addend input logic [2*`NF+1:0] P, // product output logic [8:0] PCnt // normalization shift count for the positive result ); // calculate the propagate (T) and kill (Z) bits logic [3*`NF+6:0] T; logic [3*`NF+5:0] Z; assign T[3*`NF+6:2*`NF+4] = A[3*`NF+6:2*`NF+4]; assign Z[3*`NF+5:2*`NF+4] = A[3*`NF+5:2*`NF+4]; assign T[2*`NF+3:2] = A[2*`NF+3:2]^P; assign Z[2*`NF+3:2] = A[2*`NF+3:2]|P; assign T[1:0] = A[1:0]; assign Z[1:0] = A[1:0]; // Apply function to determine Leading pattern logic [3*`NF+6:0] pf; assign pf = T^{Z[3*`NF+5:0], 1'b0}; logic [8:0] i; always_comb begin i = 0; while (~pf[3*`NF+6-i] && $unsigned(i) <= $unsigned(3*`NF+6)) i = i+1; // search for leading one PCnt = i; end endmodule module neglza( input logic [3*`NF+6:0] A, // addend input logic [3*`NF+6:0] P, // product output logic [8:0] NCnt // normalization shift count for the negitive result ); // calculate the propagate (T) and kill (Z) bits logic [3*`NF+6:0] T; logic [3*`NF+5:0] Z; assign T = A^P; assign Z = ~(A[3*`NF+5:0]|P[3*`NF+5:0]); // Apply function to determine Leading pattern logic [3*`NF+6:0] f; assign f = T^{~Z, 1'b0}; logic [8:0] i; always_comb begin i = 0; while (~f[3*`NF+6-i] && $unsigned(i) <= $unsigned(3*`NF+6)) i = i+1; // search for leading one NCnt = i; end endmodule module normalize( input logic [3*`NF+5:0] SumM, // the positive sum input logic [`NE-1:0] ZExpM, // exponent of Z input logic [`NE+1:0] ProdExpM, // X exponent + Y exponent - bias input logic [8:0] NormCntM, // normalization shift count input logic FmtM, // precision 1 = double 0 = single input logic KillProdM, // is the product set to zero input logic AddendStickyM, // the sticky bit caclulated from the aligned addend output logic [`NF+2:0] NormSum, // normalized sum output logic SumZero, // is the sum zero output logic NormSumSticky, UfSticky, // sticky bits output logic [`NE+1:0] SumExp, // exponent of the normalized sum output logic ResultDenorm // is the result denormalized ); logic [`NE+1:0] FracLen; // length of the fraction logic [`NE+1:0] SumExpTmp; // exponent of the normalized sum not taking into account denormal or zero results logic [8:0] DenormShift; // right shift if the result is denormalized //***change this later logic [3*`NF+5:0] CorrSumShifted; // the shifted sum after LZA correction logic [3*`NF+7:0] SumShifted; // the shifted sum before LZA correction logic [`NE+1:0] SumExpTmpTmp; // the exponent of the normalized sum with the `FLEN bias logic PreResultDenorm; // is the result denormalized - calculated before LZA corection logic LZAPlus1; // add one to the sum's exponent due to LZA correction /////////////////////////////////////////////////////////////////////////////// // Normalization /////////////////////////////////////////////////////////////////////////////// // Determine if the sum is zero assign SumZero = ~(|SumM); // determine the length of the fraction based on precision assign FracLen = FmtM ? `NF+1 : 13'd24; // calculate the sum's exponent assign SumExpTmpTmp = KillProdM ? {2'b0, ZExpM} : ProdExpM + -({4'b0, NormCntM} + 1 - (`NF+4)); assign SumExpTmp = FmtM ? SumExpTmpTmp : (SumExpTmpTmp-1023+127)&{`NE+2{|SumExpTmpTmp}}; // Determine if the result is denormal assign PreResultDenorm = $signed(SumExpTmp)<=0 & ($signed(SumExpTmp)>=$signed(-FracLen)) & ~SumZero; // Determine the shift needed for denormal results // - if not denorm add 1 to shift out the leading 1 assign DenormShift = PreResultDenorm ? SumExpTmp[8:0] : 1; //*** change this when changing the size of DenormShift also change to an and opperation // Normalize the sum assign SumShifted = {2'b0, SumM} << NormCntM+DenormShift; //*** fix mux's with constants in them //***NormCnt can be simplified // LZA correction assign LZAPlus1 = SumShifted[3*`NF+7]; assign CorrSumShifted = LZAPlus1 ? SumShifted[3*`NF+6:1] : SumShifted[3*`NF+5:0]; assign NormSum = CorrSumShifted[3*`NF+5:2*`NF+3]; // Calculate the sticky bit assign NormSumSticky = (|CorrSumShifted[2*`NF+2:0]) | (|CorrSumShifted[136:2*`NF+3]&~FmtM); assign UfSticky = AddendStickyM | NormSumSticky; // Determine sum's exponent assign SumExp = (SumExpTmp+LZAPlus1+(~|SumExpTmp&SumShifted[3*`NF+6])) & {`NE+2{~(SumZero|ResultDenorm)}}; // recalculate if the result is denormalized assign ResultDenorm = PreResultDenorm&~SumShifted[3*`NF+6]&~SumShifted[3*`NF+7]; endmodule module fmaround( input logic FmtM, // precision 1 = double 0 = single input logic [2:0] FrmM, // rounding mode input logic UfSticky, // sticky bit for underlow calculation input logic [`NF+2:0] NormSum, // normalized sum input logic AddendStickyM, // addend's sticky bit input logic NormSumSticky, // normalized sum's sticky bit input logic ZZeroM, // is Z zero input logic InvZM, // invert Z input logic [`NE+1:0] SumExp, // exponent of the normalized sum input logic ResultSgn, // the result's sign output logic CalcPlus1, Plus1, UfPlus1, Minus1, // do you add or subtract on from the result output logic [`NE+1:0] FullResultExp, // ResultExp with bits to determine sign and overflow output logic [`NF-1:0] ResultFrac, // Result fraction output logic [`NE-1:0] ResultExp, // Result exponent output logic Sticky, // sticky bit output logic Round, Guard, UfRound, UfLSBNormSum // bits needed to calculate rounding ); logic LSBNormSum; // bit used for rounding - least significant bit of the normalized sum logic SubBySmallNum, UfSubBySmallNum; // was there supposed to be a subtraction by a small number logic UfGuard; // gaurd bit used to caluculate underflow logic UfCalcPlus1, CalcMinus1; // do you add or subtract on from the result logic [`FLEN:0] RoundAdd; // how much to add to the result logic [`NF-1:0] NormSumTruncated; // the normalized sum trimed to fit the mantissa /////////////////////////////////////////////////////////////////////////////// // Rounding /////////////////////////////////////////////////////////////////////////////// // round to nearest even // {Guard, Round, Sticky} // 0xx - do nothing // 100 - tie - Plus1 if result is odd (LSBNormSum = 1) // - don't add 1 if a small number was supposed to be subtracted // 101 - do nothing if a small number was supposed to subtracted (the sticky bit was set by the small number) // 110/111 - Plus1 // round to zero - subtract 1 if a small number was supposed to be subtracted from a positive result with guard and round bits of 0 // round to -infinity // - Plus1 if negative unless a small number was supposed to be subtracted from a result with guard and round bits of 0 // - subtract 1 if a small number was supposed to be subtracted from a positive result with guard and round bits of 0 // round to infinity // - Plus1 if positive unless a small number was supposed to be subtracted from a result with guard and round bits of 0 // - subtract 1 if a small number was supposed to be subtracted from a negative result with guard and round bits of 0 // round to nearest max magnitude // {Guard, Round, Sticky} // 0xx - do nothing // 100 - tie - Plus1 // - don't add 1 if a small number was supposed to be subtracted // 101 - do nothing if a small number was supposed to subtracted (the sticky bit was set by the small number) // 110/111 - Plus1 // determine guard, round, and least significant bit of the result assign Guard = FmtM ? NormSum[2] : NormSum[31]; assign Round = FmtM ? NormSum[1] : NormSum[30]; assign LSBNormSum = FmtM ? NormSum[3] : NormSum[32]; // used to determine underflow flag assign UfGuard = FmtM ? NormSum[1] : NormSum[30]; assign UfRound = FmtM ? NormSum[0] : NormSum[29]; assign UfLSBNormSum = FmtM ? NormSum[2] : NormSum[31]; // determine sticky assign Sticky = UfSticky | NormSum[0]; // Deterimine if a small number was supposed to be subtrated assign SubBySmallNum = AddendStickyM & InvZM & ~(NormSumSticky|UfRound) & ~ZZeroM; //***here assign UfSubBySmallNum = AddendStickyM & InvZM & ~(NormSumSticky) & ~ZZeroM; //***here always_comb begin // Determine if you add 1 case (FrmM) 3'b000: CalcPlus1 = Guard & (Round | ((Sticky)&~(~Round&SubBySmallNum)) | (~Round&~(Sticky)&LSBNormSum&~SubBySmallNum));//round to nearest even 3'b001: CalcPlus1 = 0;//round to zero 3'b010: CalcPlus1 = ResultSgn & ~(SubBySmallNum & ~Guard & ~Round);//round down 3'b011: CalcPlus1 = ~ResultSgn & ~(SubBySmallNum & ~Guard & ~Round);//round up 3'b100: CalcPlus1 = (Guard & (Round | ((Sticky)&~(~Round&SubBySmallNum)) | (~Round&~(Sticky)&~SubBySmallNum)));//round to nearest max magnitude default: CalcPlus1 = 1'bx; endcase // Determine if you add 1 (for underflow flag) case (FrmM) 3'b000: UfCalcPlus1 = UfGuard & (UfRound | (UfSticky&UfRound|~UfSubBySmallNum) | (~Sticky&UfLSBNormSum&~UfSubBySmallNum));//round to nearest even 3'b001: UfCalcPlus1 = 0;//round to zero 3'b010: UfCalcPlus1 = ResultSgn & ~(UfSubBySmallNum & ~UfGuard & ~UfRound);//round down 3'b011: UfCalcPlus1 = ~ResultSgn & ~(UfSubBySmallNum & ~UfGuard & ~UfRound);//round up 3'b100: UfCalcPlus1 = (UfGuard & (UfRound | (UfSticky&~(~UfRound&UfSubBySmallNum)) | (~Sticky&~UfSubBySmallNum)));//round to nearest max magnitude default: UfCalcPlus1 = 1'bx; endcase // Determine if you subtract 1 case (FrmM) 3'b000: CalcMinus1 = 0;//round to nearest even 3'b001: CalcMinus1 = SubBySmallNum & ~Guard & ~Round;//round to zero 3'b010: CalcMinus1 = ~ResultSgn & ~Guard & ~Round & SubBySmallNum;//round down 3'b011: CalcMinus1 = ResultSgn & ~Guard & ~Round & SubBySmallNum;//round up 3'b100: CalcMinus1 = 0;//round to nearest max magnitude default: CalcMinus1 = 1'bx; endcase end // If an answer is exact don't round assign Plus1 = CalcPlus1 & (Sticky | Guard | Round); assign UfPlus1 = UfCalcPlus1 & (Sticky | UfGuard);//UfRound is part of sticky assign Minus1 = CalcMinus1 & (Sticky | Guard | Round); // Compute rounded result assign RoundAdd = FmtM ? Minus1 ? {`FLEN+1{1'b1}} : {{{`FLEN{1'b0}}}, Plus1} : Minus1 ? {{36{1'b1}}, 29'b0} : {35'b0, Plus1, 29'b0}; assign NormSumTruncated = {NormSum[`NF+2:32], NormSum[31:3]&{29{FmtM}}}; assign {FullResultExp, ResultFrac} = {SumExp, NormSumTruncated} + RoundAdd; assign ResultExp = FullResultExp[`NE-1:0]; endmodule module fmaflags( input logic XSNaNM, YSNaNM, ZSNaNM, // inputs are signaling NaNs input logic XInfM, YInfM, ZInfM, // inputs are infinity input logic XZeroM, YZeroM, // inputs are zero input logic XNaNM, YNaNM, ZNaNM, // inputs are NaN input logic [`NE+1:0] FullResultExp, // ResultExp with bits to determine sign and overflow input logic [`NE+1:0] SumExp, // exponent of the normalized sum input logic ZSgnEffM, PSgnM, // the product and modified Z signs input logic Round, Guard, UfRound, UfLSBNormSum, Sticky, UfPlus1, // bits used to determine rounding input logic FmtM, // precision 1 = double 0 = single output logic Invalid, Overflow, Underflow, // flags used to select the result output logic [4:0] FMAFlgM // FMA flags ); logic [`NE+1:0] MaxExp; // maximum value of the exponent logic SigNaN; // is an input a signaling NaN logic UnderflowFlag, Inexact; // flags /////////////////////////////////////////////////////////////////////////////// // Flags /////////////////////////////////////////////////////////////////////////////// // Set Invalid flag for following cases: // 1) any input is a signaling NaN // 2) Inf - Inf (unless x or y is NaN) // 3) 0 * Inf // assign MaxExp = FmtM ? {`NE{1'b1}} : {8{1'b1}}; assign SigNaN = XSNaNM | YSNaNM | ZSNaNM; assign Invalid = SigNaN | ((XInfM || YInfM) & ZInfM & (PSgnM ^ ZSgnEffM) & ~XNaNM & ~YNaNM) | (XZeroM & YInfM) | (YZeroM & XInfM); // Set Overflow flag if the number is too big to be represented // - Don't set the overflow flag if an overflowed result isn't outputed logic LtMaxExp; assign LtMaxExp = FmtM ? &FullResultExp[`NE-1:0] | FullResultExp[`NE] : &FullResultExp[7:0] | FullResultExp[8]; assign Overflow = LtMaxExp & ~FullResultExp[`NE+1]&~(XNaNM|YNaNM|ZNaNM|XInfM|YInfM|ZInfM); // Set Underflow flag if the number is too small to be represented in normal numbers // - Don't set the underflow flag if the result is exact assign Underflow = (SumExp[`NE+1] | ((SumExp == 0) & (Round|Guard|Sticky)))&~(XNaNM|YNaNM|ZNaNM|XInfM|YInfM|ZInfM); assign UnderflowFlag = (FullResultExp[`NE+1] | ((FullResultExp == 0) | ((FullResultExp == 1) & (SumExp == 0) & ~(UfPlus1&UfLSBNormSum)))&(Round|Guard|Sticky))&~(XNaNM|YNaNM|ZNaNM|XInfM|YInfM|ZInfM); // Set Inexact flag if the result is diffrent from what would be outputed given infinite precision // - Don't set the underflow flag if an underflowed result isn't outputed assign Inexact = (Sticky|Overflow|Guard|Round|Underflow)&~(XNaNM|YNaNM|ZNaNM|XInfM|YInfM|ZInfM); // Combine flags // - FMA can't set the Divide by zero flag // - Don't set the underflow flag if the result was rounded up to a normal number assign FMAFlgM = {Invalid, 1'b0, Overflow, UnderflowFlag, Inexact}; endmodule