OpenCores
URL https://opencores.org/ocsvn/ft816float/ft816float/trunk

Subversion Repositories ft816float

Compare Revisions

  • This comparison shows the changes necessary to convert path
    /ft816float/trunk/rtl/verilog2
    from Rev 73 to Rev 74
    Reverse comparison

Rev 73 → Rev 74

/fpFMA32combo.sv
0,0 → 1,644
// ============================================================================
// __
// \\__/ o\ (C) 2019-2022 Robert Finch, Waterloo
// \ __ / All rights reserved.
// \/_// robfinch<remove>@finitron.ca
// ||
//
// fpFMA32combo.sv
// - floating point fused multiplier + adder
// - combinational logic only
// - IEEE 754 representation
//
//
// BSD 3-Clause License
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are met:
//
// 1. Redistributions of source code must retain the above copyright notice, this
// list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// 3. Neither the name of the copyright holder nor the names of its
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
// DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
// FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
// SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
// OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//
// ============================================================================
 
import fp32Pkg::*;
 
module fpFMA32combo (op, rm, a, b, c, o, under, over, inf, zero);
input op; // operation 0 = add, 1 = subtract
input [2:0] rm;
input FP32 a, b, c;
output FP32X o;
output under;
output over;
output inf;
output zero;
 
// constants
wire [fp32Pkg::EMSB:0] infXp = {fp32Pkg::EMSB+1{1'b1}}; // infinite / NaN - all ones
// The following is the value for an exponent of zero, with the offset
// eg. 8'h7f for eight bit exponent, 11'h7ff for eleven bit exponent, etc.
wire [fp32Pkg::EMSB:0] bias = {1'b0,{fp32Pkg::EMSB{1'b1}}}; //2^0 exponent
// The following is a template for a quiet nan. (MSB=1)
wire [fp32Pkg::FMSB:0] qNaN = {1'b1,{fp32Pkg::FMSB{1'b0}}};
 
// -----------------------------------------------------------
// Clock #1
// - decode the input operands
// - derive basic information
// -----------------------------------------------------------
 
wire sa1, sb1, sc1; // sign bit
wire [fp32Pkg::EMSB:0] xa1, xb1, xc1; // exponent bits
wire [fp32Pkg::FMSB+1:0] fracta1, fractb1, fractc1; // includes unhidden bit
wire a_dn1, b_dn1, c_dn1; // a/b is denormalized
wire aNan1, bNan1, cNan1;
wire az1, bz1, cz1;
wire aInf1, bInf1, cInf1;
reg op1;
 
fpDecomp32 u1a (.i(a), .sgn(sa1), .exp(xa1), .fract(fracta1), .xz(a_dn1), .vz(az1), .inf(aInf1), .nan(aNan1) );
fpDecomp32 u1b (.i(b), .sgn(sb1), .exp(xb1), .fract(fractb1), .xz(b_dn1), .vz(bz1), .inf(bInf1), .nan(bNan1) );
fpDecomp32 u1c (.i(c), .sgn(sc1), .exp(xc1), .fract(fractc1), .xz(c_dn1), .vz(cz1), .inf(cInf1), .nan(cNan1) );
 
always_comb
op1 <= op;
 
// -----------------------------------------------------------
// Clock #2
// Compute the sum of the exponents.
// correct the exponent for denormalized operands
// adjust the sum by the exponent offset (subtract 127)
// mul: ex1 = xa + xb, result should always be < 1ffh
// Form partial products (clocks 2 to 5)
// -----------------------------------------------------------
 
reg abz2;
reg [fp32Pkg::EMSB+2:0] ex2;
reg [fp32Pkg::EMSB:0] xc2;
reg realOp2;
reg xcInf2;
 
always_comb
abz2 <= az1|bz1;
always_comb
ex2 <= (xa1|(a_dn1&~az1)) + (xb1|(b_dn1&~bz1)) - bias;
always_comb
xc2 <= (xc1|(c_dn1&~cz1));
always_comb
xcInf2 = &xc1;
 
// Figure out which operation is really needed an add or
// subtract ?
// If the signs are the same, use the orignal op,
// otherwise flip the operation
// a + b = add,+
// a + -b = sub, so of larger
// -a + b = sub, so of larger
// -a + -b = add,-
// a - b = sub, so of larger
// a - -b = add,+
// -a - b = add,-
// -a - -b = sub, so of larger
always_comb
realOp2 <= (sa1 ^ sb1) ^ sc1 ? ~op1 : op1;
 
reg [fp32Pkg::FX:0] fract5;
wire [63:0] fractoo;
mult32x32combo umul1 (.a({9'd0,fracta1}), .b({9'd0,fractb1}), .o(fractoo));
always_comb
fract5 <= fractoo[fp32Pkg::FX:0];
 
// -----------------------------------------------------------
// Clock #3
// Select zero exponent
// -----------------------------------------------------------
 
reg [fp32Pkg::EMSB+2:0] ex3;
reg [fp32Pkg::EMSB:0] xc3;
always_comb
ex3 <= abz2 ? 1'd0 : ex2;
always_comb
xc3 <= xc2;
 
// -----------------------------------------------------------
// Clock #4
// Generate partial products.
// -----------------------------------------------------------
 
reg [fp32Pkg::EMSB+2:0] ex4;
reg [fp32Pkg::EMSB:0] xc4;
 
always_comb
ex4 <= ex3;
always_comb
xc4 <= xc3;
 
// -----------------------------------------------------------
// Clock #5
// Sum partial products (above)
// compute multiplier overflow and underflow
// -----------------------------------------------------------
 
// Status
reg under5;
reg over5;
reg [fp32Pkg::EMSB+2:0] ex5;
reg [fp32Pkg::EMSB:0] xc5;
reg aInf5, bInf5;
reg aNan5, bNan5;
reg qNaNOut5;
 
always_comb
under5 <= ex4[fp32Pkg::EMSB+2];
always_comb
over5 <= (&ex4[fp32Pkg::EMSB:0] | ex4[fp32Pkg::EMSB+1]) & !ex4[fp32Pkg::EMSB+2];
always_comb
ex5 <= ex4;
always_comb
xc5 <= xc4;
always_comb
aInf5 <= aInf1;
always_comb
bInf5 <= bInf1;
 
// determine when a NaN is output
reg [fp32Pkg::MSB:0] a5,b5;
always_comb
qNaNOut5 <= (aInf1&bz1)|(bInf1&az1);
always_comb
aNan5 <= aNan1;
always_comb
bNan5 <= bNan1;
always_comb
a5 <= a;
always_comb
b5 <= b;
 
// -----------------------------------------------------------
// Clock #6
// - figure multiplier mantissa output
// - figure multiplier exponent output
// - correct xponent and mantissa for exceptional conditions
// -----------------------------------------------------------
 
reg [fp32Pkg::FX:0] mo6;
reg [fp32Pkg::EMSB+2:0] ex6;
reg [fp32Pkg::EMSB:0] xc6;
reg [fp32Pkg::FMSB+1:0] fractc6;
reg under6;
always_comb
fractc6 <= fractc1;
always_comb
under6 <= under5;
 
always_comb
xc6 <= xc5;
 
always_comb
casez({aNan5,bNan5,qNaNOut5,aInf5,bInf5,over5})
6'b1?????: mo6 <= {1'b1,1'b1,a5[fp32Pkg::FMSB-1:0],{fp32Pkg::FMSB+1{1'b0}}};
6'b01????: mo6 <= {1'b1,1'b1,b5[fp32Pkg::FMSB-1:0],{fp32Pkg::FMSB+1{1'b0}}};
6'b001???: mo6 <= {1'b1,qNaN|3'd4,{fp32Pkg::FMSB+1{1'b0}}}; // multiply inf * zero
6'b0001??: mo6 <= 0; // mul inf's
6'b00001?: mo6 <= 0; // mul inf's
6'b000001: mo6 <= 0; // mul overflow
default: mo6 <= fract5;
endcase
 
always_comb
casez({qNaNOut5|aNan5|bNan5,aInf5,bInf5,over5,under5})
5'b1????: ex6 <= infXp; // qNaN - infinity * zero
5'b01???: ex6 <= infXp; // 'a' infinite
5'b001??: ex6 <= infXp; // 'b' infinite
5'b0001?: ex6 <= infXp; // result overflow
5'b00001: ex6 <= ex5; //0; // underflow
default: ex6 <= ex5; // situation normal
endcase
 
// -----------------------------------------------------------
// Clock #7
// - prep for addition, determine greater operand
// -----------------------------------------------------------
reg ex_gt_xc7;
reg xeq7;
reg ma_gt_mc7;
reg meq7;
reg az7, bz7, cz7;
reg realOp7;
 
// which has greater magnitude ? Used for sign calc
always_comb
ex_gt_xc7 <= xc6=='d0 ? |ex6 : $signed(ex6) > $signed({2'b0,xc6});
always_comb
xeq7 <= (ex6=={2'b0,xc6});
always_comb
ma_gt_mc7 <= mo6 > {fractc6,{fp32Pkg::FMSB+1{1'b0}}};
always_comb
meq7 <= mo6 == {fractc6,{fp32Pkg::FMSB+1{1'b0}}};
always_comb
az7 <= az1;
always_comb
bz7 <= bz1;
always_comb
cz7 <= cz1;
always_comb
realOp7 <= realOp2;
 
// -----------------------------------------------------------
// Clock #8
// - prep for addition, determine greater operand
// - determine if result will be zero
// -----------------------------------------------------------
 
reg a_gt_b8;
reg resZero8;
reg ex_gt_xc8;
reg [fp32Pkg::EMSB+2:0] ex8;
reg [fp32Pkg::EMSB:0] xc8;
reg xcInf8;
reg [2:0] rm8;
reg op8;
reg sa8, sc8;
 
always_comb
ex8 <= ex6;
always_comb
xc8 <= xc6;
always_comb
xcInf8 <= xcInf2;
always_comb
rm8 <= rm;
always_comb
op8 <= op1;
always_comb
sa8 <= sa1 ^ sb1;
always_comb
sc8 <= sc1;
 
always_comb
ex_gt_xc8 <= ex_gt_xc7;
always_comb
a_gt_b8 <= ex_gt_xc7 || (xeq7 && ma_gt_mc7);
 
// Find out if the result will be zero.
always_comb
resZero8 <= (realOp7 & xeq7 & meq7) || // subtract, same magnitude
((az7 | bz7) & cz7); // a or b zero and c zero
 
// -----------------------------------------------------------
// CLock #9
// Compute output exponent and sign
//
// The output exponent is the larger of the two exponents,
// unless a subtract operation is in progress and the two
// numbers are equal, in which case the exponent should be
// zero.
// -----------------------------------------------------------
 
reg so9;
reg [fp32Pkg::EMSB+2:0] ex9;
reg [fp32Pkg::EMSB+2:0] ex9a;
reg ex_gt_xc9;
reg [fp32Pkg::EMSB:0] xc9;
reg a_gt_c9;
reg [fp32Pkg::FX:0] mo9;
reg [fp32Pkg::FMSB+1:0] fractc9;
reg under9;
reg xeq9;
 
always_comb
ex_gt_xc9 <= ex_gt_xc8;
always_comb
a_gt_c9 <= a_gt_b8;
always_comb
xc9 <= xc8;
always_comb
ex9a <= ex8;
always_comb
mo9 <= mo6;
always_comb
fractc9 <= fractc6;
always_comb
under9 <= under6;
always_comb
xeq9 <= xeq7;
 
always_comb
ex9 <= resZero8 ? 1'd0 : ex_gt_xc8 ? ex8 : {2'b0,xc8};
 
// Compute output sign
always_comb
case ({resZero8,sa8,op8,sc8}) // synopsys full_case parallel_case
4'b0000: so9 <= 0; // + + + = +
4'b0001: so9 <= !a_gt_b8; // + + - = sign of larger
4'b0010: so9 <= !a_gt_b8; // + - + = sign of larger
4'b0011: so9 <= 0; // + - - = +
4'b0100: so9 <= a_gt_b8; // - + + = sign of larger
4'b0101: so9 <= 1; // - + - = -
4'b0110: so9 <= 1; // - - + = -
4'b0111: so9 <= a_gt_b8; // - - - = sign of larger
4'b1000: so9 <= 0; // A + B, sign = +
4'b1001: so9 <= rm8==3; // A + -B, sign = + unless rounding down
4'b1010: so9 <= rm8==3; // A - B, sign = + unless rounding down
4'b1011: so9 <= 0; // +A - -B, sign = +
4'b1100: so9 <= rm8==3; // -A + B, sign = + unless rounding down
4'b1101: so9 <= 1; // -A + -B, sign = -
4'b1110: so9 <= 1; // -A - +B, sign = -
4'b1111: so9 <= rm8==3; // -A - -B, sign = + unless rounding down
endcase
 
// -----------------------------------------------------------
// Clock #10
// Compute the difference in exponents, provides shift amount
// Note that ex9a will be negative for an underflow condition
// so it's added rather than subtracted from xc9 as -(-num)
// is the same as an add. The underflow is tracked rather than
// using extra bits in the exponent.
// -----------------------------------------------------------
reg [fp32Pkg::EMSB+2:0] xdiff10;
reg [fp32Pkg::FX:0] mfs;
reg ops10;
 
// If the multiplier exponent was negative (underflowed) then
// the mantissa needs to be shifted right even more (until
// the exponent is zero. The total shift would be xc9-0-
// amount underflows which is xc9 + -ex9a.
 
always_comb
xdiff10 <= ex_gt_xc9 ? ex9a - xc9
: ex9a[fp32Pkg::EMSB+2] ? xc9 + (~ex9a+2'd1)
: xc9 - ex9a;
 
// Determine which fraction to denormalize (the one with the
// smaller exponent is denormalized). If the exponents are equal
// denormalize the smaller fraction.
always_comb
mfs <=
xeq9 ? (a_gt_c9 ? {4'b0,fractc9,{fp32Pkg::FMSB+1{1'b0}}} : mo9)
: ex_gt_xc9 ? {4'b0,fractc9,{fp32Pkg::FMSB+1{1'b0}}} : mo9;
 
always_comb
ops10 <= xeq9 ? (a_gt_c9 ? 1'b1 : 1'b0)
: (ex_gt_xc9 ? 1'b1 : 1'b0);
 
// -----------------------------------------------------------
// Clock #11
// Limit the size of the shifter to only bits needed.
// -----------------------------------------------------------
reg [7:0] xdif11;
 
always_comb
xdif11 <= xdiff10 > fp32Pkg::FX+3 ? fp32Pkg::FX+3 : xdiff10;
 
// -----------------------------------------------------------
// Clock #12
// Determine the sticky bit
// -----------------------------------------------------------
 
wire sticky;
reg sticky12;
reg [fp32Pkg::FX:0] mfs12;
reg [7:0] xdif12;
 
redorN #(.BSIZE(fp32Pkg::FX+1)) uredor1 (.a({1'b0,xdif11+fp32Pkg::FMSB}), .b(mfs), .o(sticky));
/*
generate
begin
if (FPWID==128)
redor128 u121 (.a(xdif11), .b({mfs,2'b0}), .o(sticky) );
else if (FPWID==96)
redor96 u121 (.a(xdif11), .b({mfs,2'b0}), .o(sticky) );
else if (FPWID==84)
redor84 u121 (.a(xdif11), .b({mfs,2'b0}), .o(sticky) );
else if (FPWID==80)
redor80 u121 (.a(xdif11), .b({mfs,2'b0}), .o(sticky) );
else if (FPWID==64)
redor64 u121 (.a(xdif11), .b({mfs,2'b0}), .o(sticky) );
else if (FPWID==32)
redor32 u121 (.a(xdif11), .b({mfs,2'b0}), .o(sticky) );
else begin
always @* begin
$display("redor operation needed in fpFMA");
$finish;
end
end
end
endgenerate
*/
 
// register inputs to shifter and shift
always_comb
sticky12 <= sticky;
always_comb
xdif12 <= xdif11;
always_comb
mfs12 <= mfs;
 
// -----------------------------------------------------------
// Clock #13
// - denormalize operand (shift right)
// -----------------------------------------------------------
reg [fp32Pkg::FX+2:0] mfs13;
reg [fp32Pkg::FX:0] mo13;
reg ex_gt_xc13;
reg [fp32Pkg::FMSB+1:0] fractc13;
reg ops13;
 
always_comb
mo13 <= mo9;
always_comb
ex_gt_xc13 <= ex_gt_xc9;
always_comb
fractc13 <= fractc9;
always_comb
ops13 <= ops10;
 
always_comb
mfs13 <= ({mfs12,2'b0} >> xdif12)|sticky12;
 
// -----------------------------------------------------------
// Clock #14
// Sort operands
// -----------------------------------------------------------
reg [fp32Pkg::FX+2:0] oa, ob;
reg a_gt_b14;
 
always_comb
a_gt_b14 <= a_gt_b8;
 
always_comb
oa <= ops13 ? {mo13,2'b00} : mfs13;
always_comb
ob <= ops13 ? mfs13 : {fractc13,{fp32Pkg::FMSB+1{1'b0}},2'b00};
 
// -----------------------------------------------------------
// Clock #15
// - Sort operands
// -----------------------------------------------------------
reg [fp32Pkg::FX+2:0] oaa, obb;
reg realOp15;
reg [fp32Pkg::EMSB:0] ex15;
reg underflow15;
 
//wire [fp32Pkg::EMSB:0] ex9c = ex9[fp32Pkg::EMSB+1] ? infXp : ex9[fp32Pkg::EMSB:0];
wire [fp32Pkg::EMSB:0] ex9c = (&ex9[fp32Pkg::EMSB:0] | ex9[fp32Pkg::EMSB+1]) & !ex9[fp32Pkg::EMSB+2] ? infXp : ex9[fp32Pkg::EMSB:0];
reg overflow15;
always_comb
realOp15 <= realOp7;
always_comb
ex15 <= ex9c;
always_comb
overflow15 <= (ex9[fp32Pkg::EMSB+1]| &ex9[fp32Pkg::EMSB:0]) & !ex9[fp32Pkg::EMSB+2];
always_comb
underflow15 = ex9[fp32Pkg::EMSB+2];
always_comb
oaa <= a_gt_b14 ? oa : ob;
always_comb
obb <= a_gt_b14 ? ob : oa;
 
// -----------------------------------------------------------
// Clock #16
// - perform add/subtract
// - addition can generate an extra bit, subtract can't go negative
// -----------------------------------------------------------
reg [fp32Pkg::FX+3:0] mab;
reg [fp32Pkg::FX:0] mo16;
reg [fp32Pkg::FMSB+1:0] fractc16;
reg Nan16;
reg cNan16;
reg aInf16, cInf16;
reg op16;
reg exinf16;
 
always_comb
Nan16 <= qNaNOut5|aNan5|bNan5;
always_comb
cNan16 <= cNan1;
always_comb
aInf16 <= &ex6;
always_comb
cInf16 <= cInf1;
always_comb
op16 <= op1;
always_comb
mo16 <= mo13;
always_comb
fractc16 <= fractc9;
always_comb
exinf16 <= &ex15;
 
always_comb
mab <= realOp15 ? oaa - obb : oaa + obb;
 
// -----------------------------------------------------------
// Clock #17
// - adjust for Nans
// -----------------------------------------------------------
reg [fp32Pkg::EMSB:0] ex17;
reg [fp32Pkg::FX:0] mo17;
reg so17;
reg exinf17;
reg overflow17;
 
always_comb
so17 <= so9;
always_comb
ex17 <= ex15;
always_comb
exinf17 <= exinf16;
always_comb
overflow17 <= overflow15;
 
always_comb
casez({aInf16&cInf16,Nan16,cNan16,exinf16})
4'b1???: mo17 <= {1'b0,op16,{fp32Pkg::FMSB-1{1'b0}},op16,{fp32Pkg::FMSB{1'b0}}}; // inf +/- inf - generate QNaN on subtract, inf on add
4'b01??: mo17 <= {1'b0,mo16};
4'b001?: mo17 <= {1'b1,1'b1,fractc16[fp32Pkg::FMSB-1:0],{fp32Pkg::FMSB+1{1'b0}}};
4'b0001: mo17 <= 1'd0;
default: mo17 <= mab[fp32Pkg::FX+3:2]; // mab has two extra lead bits and two trailing bits
endcase
 
assign o.sign = so17;
assign o.exp = ex17;
assign o.sig = mo17;
 
assign zero = {ex17,mo17}==1'd0;
assign inf = exinf17;
assign under = underflow15;//ex17==1'd0;
assign over = overflow17;
 
endmodule
 
 
// Multiplier with normalization and rounding.
 
module fpFMA32nrCombo(op, rm, a, b, c, o, inf, zero, overflow, underflow, inexact);
input op;
input [2:0] rm;
input FP32 a, b, c;
output FP32 o;
output zero;
output inf;
output reg overflow;
output reg underflow;
output reg inexact;
 
wire FP32X fma_o;
wire fma_underflow;
wire fma_overflow;
wire norm_underflow;
wire norm_inexact;
wire sign_exe1, inf1, overflow1, underflow1;
wire FP32N fpn0;
 
fpFMA32combo u1
(
.op(op),
.rm(rm),
.a(a),
.b(b),
.c(c),
.o(fma_o),
.under(fma_underflow),
.over(fma_overflow),
.zero(),
.inf()
);
fpNormalize32combo u2
(
.i(fma_o),
.o(fpn0),
.under_i(fma_underflow),
.under_o(norm_underflow),
.inexact_o(norm_inexact)
);
fpRound32combo u3(.rm(rm), .i(fpn0), .o(o) );
fpDecomp32 u4(.i(o), .xz(), .vz(zero), .inf(inf));
always_comb
underflow <= fma_underflow;
always_comb
overflow <= fma_overflow;
always_comb
inexact <= norm_inexact;
//assign overflow = inf;
 
endmodule
 
/fpNormalize32combo.sv
0,0 → 1,339
// ============================================================================
// __
// \\__/ o\ (C) 2006-2022 Robert Finch, Waterloo
// \ __ / All rights reserved.
// \/_// robfinch<remove>@finitron.ca
// ||
//
// fpNormalize32combo.sv
// - floating point normalization unit
// - combinational logic only
// - IEEE 754 representation
//
//
// BSD 3-Clause License
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are met:
//
// 1. Redistributions of source code must retain the above copyright notice, this
// list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// 3. Neither the name of the copyright holder nor the names of its
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
// DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
// FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
// SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
// OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//
// 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 fp32Pkg::*;
 
module fpNormalize32combo(i, o, under_i, under_o, inexact_o);
input FP32X i; // expanded format input
output FP32N o; // normalized output + guard, sticky and round bits, + 1 whole digit
input under_i;
output reg under_o;
output reg inexact_o;
 
integer n;
// ----------------------------------------------------------------------------
// No Clock required
// ----------------------------------------------------------------------------
reg [fp32Pkg::EMSB+1:0] xo0;
reg so0;
 
always_comb
xo0 <= {under_i,i.exp};
always_comb
so0 <= i.sign; // sign doesn't change
 
// ----------------------------------------------------------------------------
// Clock #1
// - Capture exponent information
// ----------------------------------------------------------------------------
reg xInf1a, xInf1b, xInf1c;
FP32X i1;
 
always_comb
i1 <= i;
always_comb
xInf1a <= &xo0 & !under_i;
always_comb
xInf1b <= &xo0[fp32Pkg::EMSB:1] & !under_i;
always_comb
xInf1c <= &xo0[fp32Pkg::EMSB:0] & !under_i;
 
// ----------------------------------------------------------------------------
// 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.
// ----------------------------------------------------------------------------
reg xInf2c, xInf2b;
reg [fp32Pkg::EMSB:0] xo2;
reg incExpByOne2, incExpByTwo2;
reg under2;
always_comb
xInf2c <= xInf1c;
always_comb
xInf2b <= xInf1b;
always_comb
xo2 <= xo0;
always_comb
under2 <= under_i;
always_comb
incExpByTwo2 <= !xInf1b & i1[fp32Pkg::FX];
always_comb
incExpByOne2 <= !xInf1a & i1[fp32Pkg::FX-1];
 
// ----------------------------------------------------------------------------
// Clock #3
// - increment exponent
// - detect a zero mantissa
// ----------------------------------------------------------------------------
 
reg incExpByTwo3;
reg incExpByOne3;
FP32X i3;
reg [fp32Pkg::EMSB+1:0] xo3;
reg zeroMan3;
always_comb
incExpByTwo3 <= incExpByTwo2;
always_comb
incExpByOne3 <= incExpByOne2;
always_comb
i3 <= i;
wire [fp32Pkg::EMSB+1:0] xv3a = xo2 + {incExpByTwo2,1'b0};
wire [fp32Pkg::EMSB+1:0] xv3b = xo2 + incExpByOne2;
 
always_comb
xo3 <= xo2 + (incExpByTwo2 ? 2'd2 : incExpByOne2 ? 2'd1 : 2'd0);
 
always_comb
zeroMan3 <= ((xv3b[fp32Pkg::EMSB+1]|| &xv3b[fp32Pkg::EMSB:0])||(xv3a[fp32Pkg::EMSB+1]| &xv3a[fp32Pkg::EMSB:0]))
&& !under2 && !xInf2c;
 
// ----------------------------------------------------------------------------
// 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 [fp32Pkg::FMSB+5:0] mo4;
reg inexact4;
 
always_comb
casez({zeroMan3,incExpByTwo3,incExpByOne3})
3'b1??: mo4 <= 1'd0;
3'b01?: mo4 <= {i3[fp32Pkg::FX:fp32Pkg::FMSB],|i3[fp32Pkg::FMSB-1:0]};
3'b001: mo4 <= {i3[fp32Pkg::FX-1:fp32Pkg::FMSB-1],|i3[fp32Pkg::FMSB-2:0]};
default: mo4 <= {i3[fp32Pkg::FX-2:fp32Pkg::FMSB-2],|i3[fp32Pkg::FMSB-3:0]};
endcase
 
always_comb
casez({zeroMan3,incExpByTwo3,incExpByOne3})
3'b1??: inexact4 <= 1'd0;
3'b01?: inexact4 <= |i3[fp32Pkg::FMSB+1:0];
3'b001: inexact4 <= |i3[fp32Pkg::FMSB:0];
default: inexact4 <= |i3[fp32Pkg::FMSB-1:0];
endcase
 
// ----------------------------------------------------------------------------
// Clock edge #5
// - count leading zeros
// ----------------------------------------------------------------------------
reg [7:0] leadingZeros5;
reg [fp32Pkg::EMSB+1:0] xo5;
reg xInf5;
always_comb
xo5 <= xo3;
always_comb
xInf5 <= xInf2c;
 
/* 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<=32) begin
assign leadingZeros5[7] = 1'b0;
cntlz32Reg 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_comb
begin
got_one = 1'b0;
lzc = 8'h00;
for (n = fp32Pkg::FMSB+5; n >= 0; n = n - 1) begin
if (!got_one) begin
if (mo4[n])
got_one = 1'b1;
else
lzc = lzc + 1'b1;
end
end
end
always_comb
leadingZeros5 <= lzc;
`else
always_comb
casez(mo4[fp32Pkg::FMSB+5:fp32Pkg::FMSB+4])
2'b1?: leadingZeros5 <= 8'd0;
2'b01: leadingZeros5 <= 8'd1;
2'b00: leadingZeros5 <= 8'd2;
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;
reg rightOrLeft6; // 0=left,1=right
reg xInf6;
reg [fp32Pkg::EMSB+1:0] xo6;
reg [fp32Pkg::FMSB+5:0] mo6;
reg zeroMan6;
always_comb
rightOrLeft6 <= under_i;
always_comb
xo6 <= xo5;
always_comb
mo6 <= mo4;
always_comb
xInf6 <= xInf5;
always_comb
zeroMan6 <= zeroMan3;
 
always_comb
lshiftAmt6 <= leadingZeros5 > xo5 ? xo5 : leadingZeros5;
 
always_comb
rshiftAmt6 <= xInf5 ? 1'd0 : $signed(xo5) > 1'd0 ? 1'd0 : ~xo5+2'd1; // xo2 is negative !
 
// ----------------------------------------------------------------------------
// Clock edge #7
// - figure exponent
// - shift mantissa
// - figure sticky bit
// ----------------------------------------------------------------------------
 
reg [fp32Pkg::EMSB:0] xo7;
reg rightOrLeft7;
reg [fp32Pkg::FMSB+5:0] mo7l, mo7r;
reg St6,St7;
always_comb
rightOrLeft7 <= rightOrLeft6;
 
always_comb
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
xo6 - lshiftAmt6; // on a left shift, the exponent can't be decremented below zero
 
always_comb
mo7r <= mo6 >> rshiftAmt6;
always_comb
mo7l <= mo6 << lshiftAmt6;
 
// The sticky bit is set if the bits shifted out on a right shift are set.
always_comb
begin
St6 = 1'b0;
for (n = 0; n < FMSB+5; n = n + 1)
if (n <= rshiftAmt6 + 1) St6 = St6|mo6[n];
end
always_comb
St7 <= St6;
 
// ----------------------------------------------------------------------------
// Clock edge #8
// - select mantissa
// ----------------------------------------------------------------------------
 
reg so;
reg [fp32Pkg::EMSB:0] xo;
reg [fp32Pkg::FMSB+5:0] mo;
always_comb
so <= so0;
always_comb
xo <= xo7;
always_comb
inexact_o <= inexact4;
always_comb
under_o <= rightOrLeft7;
 
always_comb
mo <= rightOrLeft7 ? mo7r|{St7,2'b0} : mo7l;
 
assign o.sign = so;
assign o.exp = xo;
assign o.sig = mo[FMSB+5:2];
 
endmodule
/fpRound32combo.sv
0,0 → 1,135
// ============================================================================
// __
// \\__/ o\ (C) 2006-2022 Robert Finch, Waterloo
// \ __ / All rights reserved.
// \/_// robfinch<remove>@finitron.ca
// ||
//
// fpRound32combo.sv
// - floating point rounding unit
// - IEEE 754 representation
// - combinational logic only
//
//
// BSD 3-Clause License
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are met:
//
// 1. Redistributions of source code must retain the above copyright notice, this
// list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// 3. Neither the name of the copyright holder nor the names of its
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
// DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
// FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
// SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
// OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//
// ============================================================================
 
import fp32Pkg::*;
 
module fpRound32combo(rm, i, o);
input [2:0] rm; // rounding mode
input FP32N i; // intermediate format input
output FP32 o; // rounded output
 
//------------------------------------------------------------
// variables
wire so;
wire [fp32Pkg::EMSB:0] xo;
reg [fp32Pkg::FMSB:0] mo;
reg [fp32Pkg::EMSB:0] xo1;
reg [fp32Pkg::FMSB+3:0] mo1;
wire xInf = &i[fp32Pkg::MSB+2:fp32Pkg::FMSB+4];
wire so0 = i[fp32Pkg::MSB+3];
 
assign o.sign = so;
assign o.exp = xo;
assign o.sig = mo;
 
 
wire l = i[3];
wire g = i[2]; // guard bit: always the same bit for all operations
wire r = i[1]; // rounding bit
wire s = i[0]; // sticky bit
reg rnd;
 
//------------------------------------------------------------
// Clock #1
// - determine round amount (add 1 or 0)
//------------------------------------------------------------
 
always_comb
xo1 <= i[fp32Pkg::MSB+2:fp32Pkg::FMSB+4];
always_comb
mo1 <= i[fp32Pkg::FMSB+3:0];
 
wire tie = g & ~(r|s);
// Compute the round bit
// Infinities and NaNs are not rounded!
always_comb
casez ({xInf,rm})
4'b0000: rnd <= (g & (r|s)) | (l & tie); // round to nearest ties to even
4'b0001: rnd <= 1'd0; // round to zero (truncate)
4'b0010: rnd <= g & !so0; // round towards +infinity
4'b0011: rnd <= g & so0; // round towards -infinity
4'b0100: rnd <= (g & (r|s)) | tie; // round to nearest ties away from zero
4'b1???: rnd <= 1'd0; // no rounding if exponent indicates infinite or NaN
default: rnd <= 0;
endcase
 
//------------------------------------------------------------
// Clock #2
// round the number, check for carry
// note: inf. exponent checked above (if the exponent was infinite already, then no rounding occurs as rnd = 0)
// note: exponent increments if there is a carry (can only increment to infinity)
//------------------------------------------------------------
 
reg [fp32Pkg::MSB:0] rounded2;
reg carry2;
reg rnd2;
reg dn2;
wire [fp32Pkg::EMSB:0] xo2;
wire [fp32Pkg::MSB:0] rounded1 = {xo1,mo1[fp32Pkg::FMSB+3:3],1'b0} + {rnd,1'b0}; // Add onto LSB, GRS=0
always_comb
rounded2 <= rounded1;
always_comb
carry2 <= mo1[fp32Pkg::FMSB+3] & !rounded1[fp32Pkg::FMSB+1];
always_comb
rnd2 <= rnd;
always_comb
dn2 <= !(|xo1);
assign xo2 = rounded2[fp32Pkg::MSB:fp32Pkg::FMSB+2];
 
//------------------------------------------------------------
// Clock #3
// - shift mantissa if required.
//------------------------------------------------------------
assign so = i[fp32Pkg::MSB+3];
assign xo = xo2;
 
always_comb
casez({rnd2,&xo2,carry2,dn2})
4'b0??0: mo <= mo1[fp32Pkg::FMSB+2:2]; // not rounding, not denormalized, => hide MSB
4'b0??1: mo <= mo1[fp32Pkg::FMSB+3:3]; // not rounding, denormalized
4'b1000: mo <= rounded2[fp32Pkg::FMSB :0]; // exponent didn't change, number was normalized, => hide MSB,
4'b1001: mo <= rounded2[fp32Pkg::FMSB+1:1]; // exponent didn't change, but number was denormalized, => retain MSB
4'b1010: mo <= rounded2[fp32Pkg::FMSB+1:1]; // exponent incremented (new MSB generated), number was normalized, => hide 'extra (FMSB+2)' MSB
4'b1011: mo <= rounded2[fp32Pkg::FMSB+1:1]; // exponent incremented (new MSB generated), number was denormalized, number became normalized, => hide 'extra (FMSB+2)' MSB
4'b11??: mo <= 1'd0; // number became infinite, no need to check carry etc., rnd would be zero if input was NaN or infinite
endcase
 
endmodule
/mult16x16combo.sv
0,0 → 1,47
// ============================================================================
// __
// \\__/ o\ (C) 2020 Robert Finch, Waterloo
// \ __ / All rights reserved.
// \/_// robfinch<remove>@finitron.ca
// ||
//
// mult16x16combo.sv
// - this should typically synthesize to a single DSP multiplier
//
// BSD 3-Clause License
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are met:
//
// 1. Redistributions of source code must retain the above copyright notice, this
// list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// 3. Neither the name of the copyright holder nor the names of its
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
// DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
// FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
// SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
// OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//
// ============================================================================
 
module mult16x16combo(a, b, o);
input [15:0] a;
input [15:0] b;
output reg [31:0] o ='d0;
 
always_comb
o <= a * b;
endmodule
/mult32x32combo.sv
0,0 → 1,145
// ============================================================================
// __
// \\__/ o\ (C) 2020-2022 Robert Finch, Waterloo
// \ __ / All rights reserved.
// \/_// robfinch<remove>@finitron.ca
// ||
//
// mult32x32combo.sv
// - Karatsuba multiply
// - combinational logic only
//
// BSD 3-Clause License
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are met:
//
// 1. Redistributions of source code must retain the above copyright notice, this
// list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// 3. Neither the name of the copyright holder nor the names of its
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
// DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
// FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
// SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
// OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//
// ============================================================================
 
`define KARATSUBA 1
 
`ifdef KARATSUBA
 
module mult32x32combo(a, b, o);
input [31:0] a;
input [31:0] b;
output reg [63:0] o='d0;
 
reg [15:0] a2='d0, b2='d0;
reg [16:0] a1='d0, b1='d0;
reg [31:0] z0, z2, z0a, z2a, z0b, z2b, z0c, z2c, z0d, z2d, p3;
reg [32:0] p4;
reg [32:0] z1; // extra bit for carry
reg sgn2, sgn3, sgn4;
 
always_comb
a1 <= a[15: 0] - a[31:16]; // x0-x1
always_comb
b1 <= b[31:16] - b[15: 0]; // y1-y0
always_comb
a2 <= a1[16] ? -a1 : a1;
always_comb
b2 <= b1[16] ? -b1 : b1;
always_comb
sgn2 <= a1[16]^b1[16];
always_comb
sgn3 <= sgn2;
always_comb
sgn4 <= sgn3;
mult16x16combo u1 (
.a(a[31:16]),
.b(b[31:16]),
.o(z2) // z2 = x1 * y1
);
 
mult16x16combo u2 (
.a(a[15:0]),
.b(b[15:0]),
.o(z0) // z0 = x0 * y0
);
 
mult16x16combo u3 (
.a(a2[15:0]),
.b(b2[15:0]),
.o(p3) // p3 = abs(x0-x1) * abs(y1-y0)
);
 
always_comb
p4 <= sgn3 ? -p3 : p3;
 
always_comb
z2a <= z2;
always_comb
z0a <= z0;
always_comb
z2b <= z2a;
always_comb
z0b <= z0a;
always_comb
z2c <= z2b;
always_comb
z0c <= z0b;
always_comb
z1 <= {{32{p4[32]}},p4} + z2c + z0c;
 
always_comb
z2d <= z2c;
always_comb
z0d <= z0c;
always_comb
o <= {z2d,z0d} + {z1,16'd0};
 
endmodule
 
`else
 
// This version of the multiply has a parameterized pipeline depth and allows
// the tools to perform the multiply. Relies on the ability of tools to retime.
 
module mult32x32combo(a, b, o);
parameter DEP = 6;
input [31:0] a;
input [31:0] b;
output reg [63:0] o;
 
reg [63:0] prod [0:DEP-1];
reg [63:0] prd;
integer n;
 
always_comb
prd <= a * b;
always_comb
prod[0] <= prd;
always_comb
for (n = 0; n < DEP - 1; n = n + 1)
prod[n+1] <= prod[n];
 
always_comb
o <= prod[DEP-1];
 
endmodule
 
`endif

powered by: WebSVN 2.1.0

© copyright 1999-2024 OpenCores.org, equivalent to Oliscience, all rights reserved. OpenCores®, registered trademark.