Rounding to Even (Verilog) - 2024.1 English

Vivado Design Suite User Guide: Synthesis (UG901)

Document ID
UG901
Release Date
2024-05-30
Version
2024.1 English

Filename: convergentRoundingEven.v

// Convergent rounding(Even) Example which makes use of pattern detect
// File: convergentRoundingEven.v
module convergentRoundingEven (
input clk,
input [23:0] a,
input [15:0] b,
output reg signed [23:0] zlast
);

reg signed [23:0] areg;
reg signed [15:0] breg;
reg signed [39:0] z1;

reg pattern_detect;
wire [15:0] pattern = 16'b0000000000000000;
wire [39:0] c = 40'b0000000000000000000000000111111111111111; // 15 ones

wire signed [39:0] multadd;
wire signed [15:0] zero;
reg signed [39:0] multadd_reg;

// Convergent Rounding: LSB Correction Technique
// ---------------------------------------------
// For static convergent rounding, the pattern detector can be used
// to detect the midpoint case. For example, in an 8-bit round, if
// the decimal place is set at 4, the C input should be set to
// 0000.0111. Round to even rounding should use CARRYIN = "1" and
// check for PATTERN "XXXX.0000" and replace the units place with 0
// if the pattern is matched. See UG193 for more details.

assign multadd = z1 + c + 1'b1;

always @(posedge clk)
begin
areg <= a;
breg <= b;
z1 <= areg * breg;
pattern_detect <= multadd[15:0] == pattern ? 1'b1 : 1'b0;
multadd_reg <= multadd;
end

// Unit bit replaced with 0 if pattern is detected
always @(posedge clk)
zlast <= pattern_detect ? {multadd_reg[39:17],1'b0} : multadd_reg[39:16];

endmodule // convergentRoundingEven