verilog编写RoundRobinArbiter的一些小错误摘录

        毕业设计是要做CGRA的共享存储器的仲裁和修复电路，先从仲裁入手，太久没写verilog有点生疏。

        RoundRobin的代码网上有很多。这里主要是看的微信公众号"IC加油站"的老李的资源，copy了一下代码进行调试，代码如下。文章地址：仲裁器设计（二）-- Round Robin Arbiter

`timescale 1ns/1ps
// two paralleled fixed-priority arbiter using mask algorithm
module RR_Arbiter #(
parameter Req_Width = 6
)(
input clk,
input rst,
input [Req_Width-1:0] req,
output [Req_Width-1:0] gnt
);
reg [Req_Width-1:0] Pointer_Req;//denote the next priority
wire [Req_Width-1:0] req_mask;
wire [Req_Width-1:0] priority_mask;
wire [Req_Width-1:0] grant_mask;
wire [Req_Width-1:0] priority_unmask;
wire [Req_Width-1:0] grant_unmask;
wire no_req_mask_label;
//the first FP arbiter with masking
assign req_mask = req & Pointer_Req;
assign priority_mask[0] = 1'b0;
assign priority_mask[Req_Width-1:1] = req_mask[Req_Width-2:0] | priority_mask[Req_Width-2:0];
assign grant_mask[Req_Width-1:0] = req_mask[Req_Width-1:0] & (~priority_mask[Req_Width-1:0]);

//the second FP arbiter without masking
assign priority_unmask[0] = 1'b0;
assign priority_unmask[Req_Width-1:1] = req[Req_Width-2:0] | priority_unmask[Req_Width-2:0];
assign grant_unmask[Req_Width-1:0] = req[Req_Width-1:0] & (~priority_unmask[Req_Width-1:0]);

//Based on the value of req_mask, choose mask or unmask. (if req_mask is none, it means two conditions
//1: no req; 2: req is in the unmask part.
assign no_req_mask_label = ~(|req_mask);
assign gnt = ({Req_Width{no_req_mask_label}} & grant_unmask) | grant_mask;

//Update the Pointer_Reg
always @ (posedge clk) begin
	if (rst) begin
		Pointer_Req <= {Req_Width{1'b1}};// initialize
	end
	else begin
		if (|req_mask) begin // still have req after mask
			Pointer_Req <= priority_mask;
		end
		else begin // no req after mask, so choose the req with no mask
			if (|req) begin //req with no mask have req
				Pointer_Req <= priority_unmask;
			end
			else begin //req with no mask is none, so remain, don't change
				Pointer_Req <= Pointer_Req;
			end
		end
	end
end
endmodule

其实就是一个简单的组合逻辑加上寄存器存储上一个周期的Pointer_Req用于优先级的轮换，随后编写testbench代码如下：

`timescale 1ns/1ps
module tb_RR_Arbiter #(
parameter Req_Width = 5
)(
//input [Req_Width-1:0] req,
//input clk,
//input rst,
//output reg [Req_Width-1:0] gnt
);

reg [Req_Width-1:0] req;
wire [Req_Width-1:0] gnt;
reg clk;
reg rst;

RR_Arbiter #(5) RRA1 (
.clk(clk),
.rst(rst),
.req(req),
.gnt(gnt)
);

parameter ClockPeriod = 10  ;
 
initial
	begin
		clk = 1 ;
    	repeat(20)
    		#(ClockPeriod) clk = ~clk;
	end

initial 
	begin
		rst = 1;
		req = {Req_Width{1'b0}};
		#20 rst = 0; req = 6'b01010;
		#20 req = 6'b01011;
		#20 req = 6'b10011;
		#20 req = 6'b01010;
		$finish;
	end


endmodule

在Modelsim和Vivado中都跑了下代码，仿真会有问题，问题出现在always块中的pointer_req更新后直接被该周期使用了，没有存储给下一周期。因此，给输出加一拍延时缓冲即可，代码改进如下:

`timescale 1ns/1ps
// two paralleled fixed-priority arbiter using mask algorithm
module RR_Arbiter #(
parameter Req_Width = 5
)(
input clk,
input rst,
input [Req_Width-1:0] req,
output [Req_Width-1:0] gnt,
output reg [Req_Width-1:0] Pointer_Req_test, // for test
output reg [Req_Width-1:0] q_test,
output [Req_Width-1:0] req_mask,
output [Req_Width-1:0] priority_mask,
output [Req_Width-1:0] priority_unmask,
output [Req_Width-1:0] grant_unmask,
output [Req_Width-1:0] grant_mask,
output label_req_mask,
output reg [2:0] label_req
);
reg [Req_Width-1:0] Pointer_Req;//denote the next priority
//wire [Req_Width-1:0] req_mask;
//wire [Req_Width-1:0] priority_mask;
//wire [Req_Width-1:0] grant_mask;
//wire [Req_Width-1:0] priority_unmask;
//wire [Req_Width-1:0] grant_unmask;
wire no_req_mask_label;
//the first FP arbiter with masking
assign req_mask = req & Pointer_Req;
assign priority_mask[0] = 1'b0;
assign priority_mask[Req_Width-1:1] = req_mask[Req_Width-2:0] | priority_mask[Req_Width-2:0];//or operation for each bit, from small to big bit
assign grant_mask[Req_Width-1:0] = req_mask[Req_Width-1:0] & (~priority_mask[Req_Width-1:0]);

//the second FP arbiter without masking
assign priority_unmask[0] = 1'b0;
assign priority_unmask[Req_Width-1:1] = req[Req_Width-2:0] | priority_unmask[Req_Width-2:0];//?????????
assign grant_unmask[Req_Width-1:0] = req[Req_Width-1:0] & (~priority_unmask[Req_Width-1:0]);

//Based on the value of req_mask, choose mask or unmask. (if req_mask is none, it means two conditions
//1: no req; 2: req is in the unmask part.
assign no_req_mask_label = ~(|req_mask);
assign gnt = ({Req_Width{no_req_mask_label}} & grant_unmask) | grant_mask;

//test
assign label_req_mask = |req_mask;
//assign label_req = |req;
//reg [Req_Width-1:0] q;
//Update the Pointer_Reg

always @ (posedge clk) begin
	#1
	if (rst) begin
		q_test <= {Req_Width{1'b1}};// initialize
		//q_test <= q;
		//Pointer_Req <= q_test;
		//Pointer_Req_test <= q_test;
		label_req <= 1;
		//Pointer_Req_test <= {Req_Width{1'b1}};
	end
	else begin
		if (label_req_mask) begin // still have req after mask
			q_test <= priority_mask;
			//Pointer_Req <= q_test;
			//Pointer_Req_test <= q_test;
			label_req <= 2;
			//Pointer_Req_test <= priority_mask;
		end
		else begin // no req after mask, so choose the req with no mask
			if (|req) begin //req with no mask have req
				q_test <= priority_unmask;
				//Pointer_Req <= q_test;// delay one clock, otherwise the result will be wrong, since the comb circuit will immediately use the pointer-reg
				//Pointer_Req_test <= q_test;
				label_req <= 3;
				//Pointer_Req_test <= priority_unmask;
			end
			else begin //req with no mask is none, so remain, don't change
				q_test <= Pointer_Req;
				//Pointer_Req <= q_test;
				//Pointer_Req_test <= q_test;
				label_req <= 4;
				//Pointer_Req_test <= Pointer_Req;
			end
		end
	end
	
end

always @ (posedge clk) begin
	Pointer_Req <= q_test;
	Pointer_Req_test <= q_test;
end
endmodule

同时在Always块中加入少许延时，确保每次计算该周期的gnt的时候用的是上一周期的Pointer_Req而不是本周期。随后，tb代码如下：

`timescale 1ns/1ps
module tb_RR_Arbiter #(
parameter Req_Width = 10
)(
//input [Req_Width-1:0] req,
//input clk,
//input rst,
//output reg [Req_Width-1:0] gnt
);

reg [Req_Width-1:0] req;
wire [Req_Width-1:0] gnt;
reg clk;
reg rst;
wire [Req_Width-1:0] Pointer_Req_test;
wire [Req_Width-1:0] q_test;
wire [Req_Width-1:0] req_mask;
wire [Req_Width-1:0] priority_mask;
wire [Req_Width-1:0] priority_unmask;
wire [Req_Width-1:0] grant_mask;
wire [Req_Width-1:0] grant_unmask;
wire label_req_mask;
wire [2:0] label_req;
RR_Arbiter #(Req_Width) RRA1 (
.clk(clk),
.rst(rst),
.req(req),
.gnt(gnt),
.Pointer_Req_test(Pointer_Req_test),
.q_test(q_test),
.req_mask(req_mask),
.priority_mask(priority_mask),
.priority_unmask(priority_unmask),
.grant_mask(grant_mask),
.grant_unmask(grant_unmask),
.label_req_mask(label_req_mask),
.label_req(label_req)
);

parameter ClockPeriod = 10  ;
 
initial
	begin
		clk = 1 ;
    	repeat(40)
    		#(ClockPeriod) clk = ~clk;
	end

initial 
	begin
		rst = 1;
		req = {Req_Width{1'b0}};
		#40 rst = 0; req = 10'b0000000000;//0
		#20 req = 10'b0000001101;//2
		#20 req = 10'b0000001110;//cha dui
		#20 req = 10'b0001101101;//4
		#20 req = 10'b0011101001;
		#20 req = 10'b0011100101;
		#20 req = 10'b0011000110;
		#20 req = 10'b0010000111;//again
		#20 req = 10'b0000000111;
		#20 req = 10'b1000000110;
		#20 req = 10'b1000000110;
		#20 req = 10'b1000000011;
		$finish;
	end


endmodule

得到的仿真图如下：

 总结：太久没有编写verilog，过于生疏，需要熟练。时序电路的阻塞和非阻塞赋值需要注意，有时候设计时钟上升沿触发，可能输入信号的值变为了这一周期而不是上一周期的，学会用延时去卡拍。