FSM vs Pipeline

Question

I am new to the world of HDL. I am currently working on implementing AES in Verilog code.

I manage the flow of my logic using varying FSMs. Given this approach I currently have to wait for the first block to finish processing before inputting the next one.

To improve throughput, I want to perform pipelining so I can input a new block every N clock cycles. However, I can't fathom how I would use both FSM and pipelining. Given I use a register to track the current state, if I pass a new input in this will overwrite the state for the previous input.

Is it possible to use an FSM and pipelining in unison?

Given the below code, I cannot generate a new output every clock cycle. This is because the second input cannot go through the FSM before until the first completes. This is a simplified example of what I am trying to achieve.

`timescale 1ns / 1ps

module pipeline_multiplier(
    input [15:0] a,
    input [15:0] b,
    input clk,
    input rst,
    output reg [31:0] c
    );
    
    reg [15:0] a_temp, b_temp;
    reg [31:0] c_temp;
    
    parameter [1:0] 
        Pipe1=2'b00, 
        Pipe2=2'b01, 
        Pipe3=2'b10;
    reg [1:0] state = Pipe1;
    
    always @(posedge clk)
    begin
        if(rst)
        begin
            state = Pipe1;            
        end
        else
        begin
            case(state)
                Pipe1:
                begin
                    a_temp <= a;
                    b_temp <= b;
                    state <= Pipe2;
                end
                Pipe2:
                begin
                    c_temp <= a_temp * b_temp;
                    state <= Pipe3;
                end
                Pipe3:
                begin
                    c <= c_temp;
                    state <= Pipe1;
                end
            endcase
        end
    end
   
    
endmodule

Answer 1

Yes, it is possible to use an FSM and pipelining "in unison". This is how the instruction execution pipelines on processors are controlled. An FSM of some fashion is needed for situations such as branching, where the pipeline may need to be flushed.

Answer 2

Good luck with learning verilog :). This seems like a simple multiplier and doesn't seem to require an FSM. I would code the module this way which will implicitly be pipelined. Also, I wouldn't registers the inputs as they will be flopped (registered) in the previous module of the AES pipeline.

`timescale 1ns / 1ps

module pipeline_multiplier(
input [15:0] a,
input [15:0] b,
input clk,
input rst,
output reg [31:0] c
);

reg [15:0] a_temp, b_temp;
reg [31:0] c_temp; // not needed as c is already a register 

always @(posedge clk)
begin
    if(rst)
    begin
        a_temp <= 16'd0;
        b_temp <= 16'd0;
        c <= 16'd0;
    end
    else
    begin
         a_temp <= a;  // This isn't typically done as inputs are registered 
                       // in previous module of the pipeline
         b_temp <= b;
         c <= a_temp * b_temp;
    end
end


endmodule