URL
https://opencores.org/ocsvn/thor/thor/trunk
Subversion Repositories thor
[/] [thor/] [trunk/] [rtl/] [verilog/] [fpUnit/] [fpNormalize.v] - Rev 25
Go to most recent revision | Compare with Previous | Blame | View Log
/* =============================================================== (C) 2006 Robert Finch All rights reserved. rob@birdcomputer.ca fpNormalize.v - floating point normalization unit - two cycle latency - parameterized width - IEEE 754 representation This source code is free for use and modification for non-commercial or evaluation purposes, provided this copyright statement and disclaimer remains present in the file. If you do modify the code, please state the origin and note that you have modified the code. NO WARRANTY. THIS Work, IS PROVIDEDED "AS IS" WITH NO WARRANTIES OF ANY KIND, WHETHER EXPRESS OR IMPLIED. The user must assume the entire risk of using the Work. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR ANY INCIDENTAL, CONSEQUENTIAL, OR PUNITIVE DAMAGES WHATSOEVER RELATING TO THE USE OF THIS WORK, OR YOUR RELATIONSHIP WITH THE AUTHOR. IN ADDITION, IN NO EVENT DOES THE AUTHOR AUTHORIZE YOU TO USE THE WORK IN APPLICATIONS OR SYSTEMS WHERE THE WORK'S FAILURE TO PERFORM CAN REASONABLY BE EXPECTED TO RESULT IN A SIGNIFICANT PHYSICAL INJURY, OR IN LOSS OF LIFE. ANY SUCH USE BY YOU IS ENTIRELY AT YOUR OWN RISK, AND YOU AGREE TO HOLD THE AUTHOR AND CONTRIBUTORS HARMLESS FROM ANY CLAIMS OR LOSSES RELATING TO SUCH UNAUTHORIZED USE. 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. Ref: Webpack 8.2 Spartan3-4 xc3s1000-4ft256 302 LUTs / 166 slices / 550 LUTs / 291 slices / 89 MHz 163 LUTs / 93 slices / 113.6 MHz? =============================================================== */ module fpNormalize(clk, ce, under, i, o); parameter WID = 32; localparam MSB = WID-1; localparam EMSB = WID==80 ? 14 : WID==64 ? 10 : WID==52 ? 10 : WID==48 ? 10 : WID==44 ? 10 : WID==42 ? 10 : WID==40 ? 9 : WID==32 ? 7 : WID==24 ? 6 : 4; localparam FMSB = WID==80 ? 63 : WID==64 ? 51 : WID==52 ? 39 : WID==48 ? 35 : WID==44 ? 31 : WID==42 ? 29 : WID==40 ? 28 : WID==32 ? 22 : WID==24 ? 15 : 9; localparam FX = (FMSB+2)*2-1; // the MSB of the expanded fraction localparam EX = FX + 1 + EMSB + 1 + 1 - 1; input clk; input ce; input under; input [EX:0] i; // expanded format input output [WID+2:0] o; // normalized output + guard, sticky and round bits, + 1 whole digit // variables wire so; wire so1 = i[EX]; // sign doesn't change // Since the there are *two* 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. wire [EMSB:0] xo; wire [EMSB:0] xo1a = i[EX-1:FX+1]; wire xInf = &xo1a & !under; wire incExp1 = !xInf & i[FX]; wire [EMSB:0] xo1 = xo1a + incExp1; wire [EMSB:0] xo2; wire xInf1 = &xo1; // If infinity is reached then set the mantissa to zero wire gbit = i[FMSB]; wire rbit = i[FMSB-1]; wire sbit = |i[FMSB-2:0]; // shift mantissa left by one to reduce to a single whole digit // if there is no exponent increment wire [FMSB+3:0] mo; wire [FMSB+3:0] mo1 = xInf1 & incExp1 ? 0 : incExp1 ? {i[FX:FMSB+1],gbit,rbit,sbit} : // reduce mantissa size {i[FX-1:FMSB+1],gbit,rbit,sbit,1'b0}; // reduce mantissa size wire [FMSB+3:0] mo2; wire [6:0] leadingZeros2; cntlz64Reg clz0 (.clk(clk), .ce(ce), .i(mo1), .o(leadingZeros2) ); // compensate for leadingZeros delay wire xInf2; delay1 #(EMSB+1) d2(.clk(clk), .ce(ce), .i(xo1), .o(xo2) ); delay1 #(1) d3(.clk(clk), .ce(ce), .i(xInf1), .o(xInf2) ); // 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. wire rightOrLeft2; // 0=left,1=right delay1 #(1) d8(.clk(clk), .ce(ce), .i(under), .o(rightOrLeft2) ); // Compute how much we want to decrement by wire [6:0] lshiftAmt2 = leadingZeros2 > xo2 ? xo2 : leadingZeros2; // compute amount to shift 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 wire [6:0] rshiftAmt2 = xInf2 ? 0 : -xo2 > FMSB+3 ? FMSB+4 : FMSB+4+xo2; // xo2 is negative ! // sign // the output sign is the same as the input sign delay1 #(1) d7(.clk(clk), .ce(ce), .i(so1), .o(so) ); // exponent // always @(posedge clk) // if (ce) assign xo = xInf2 ? xo2 : // an infinite exponent is either a NaN or infinity; no need to change rightOrLeft2 ? 0 : // on a right shift, the exponent was negative, it's being made to zero xo2 - lshiftAmt2; // on a left shift, the exponent can't be decremented below zero // mantissa delay1 #(FMSB+3) d4(.clk(clk), .ce(ce), .i(mo1), .o(mo2) ); wire [FMSB+3:0] mo2a; shiftAndMask #(FMSB+4) u1 (.op({rightOrLeft2,1'b0}), .a(mo2), .b(rightOrLeft2 ? lshiftAmt2 : rshiftAmt2), .mb(6'd0), .me(FMSB+3), .o(mo2a) ); // always @(posedge clk) // if (ce) assign mo = mo2a;//rightOrLeft2 ? mo2 >> rshiftAmt2 : mo2 << lshiftAmt2; assign o = {so,xo,mo}; endmodule
Go to most recent revision | Compare with Previous | Blame | View Log