Pipeline Half-Buffer

A single pipeline register with ready/valid handshakes. Decouples the input and ouput handshakes (no combinational path), but does not allow concurrent read/write like a full [Pipeline Skid Buffer] (./Pipeline_Skid_Buffer.html). The half-buffer must be read out before you can write into it again, halving the maximum bandwidth (except in Circular Buffer Mode).

However, using a half-buffer can improve the throughput of a long-running module with ready/valid handshakes, where the module input will not accept new data until the module output is read out, and it takes multiple cycles to compute a result. With a half-buffer, the module can immediately dump its output into the half-buffer and then accept new input data, overlapping another computation with the wait time until the final destination reads out the half-buffer.

Circular Buffer Mode

Normally, a Half-Buffer reads in one value and will not complete another input handshake until the data has been read out. You can think of this as buffering the earliest value from the pipeline.

Setting CIRCULAR_BUFFER parameter to a non-zero value changes the behaviour at the input: the input handshake can always complete, discarding the data already in the buffer even if it was never read out. You can think of this as buffering the latest value from the pipeline. This is a one-entry circular buffer.

However, data in the buffer can only be read out once, as usual, until updated again at the input, possibly in the same clock cycle. Simultaneous input and output handshakes are possible in Circular Buffer Mode since input_ready no longer depends on the empty/full state of the buffer (which forces alternation of input and output handshakes), nor on the state of the output handshake (which is disallowed to prevent creating a combinational path between input and output).

`default_nettype none

module Pipeline_Half_Buffer
#(
    parameter WORD_WIDTH            = 0,
    parameter CIRCULAR_BUFFER       = 0     // non-zero to enable
)
(
    input  wire                     clock,
    input  wire                     clear,

    input  wire                     input_valid,
    output reg                      input_ready,
    input  wire [WORD_WIDTH-1:0]    input_data,

    output reg                      output_valid,
    input  wire                     output_ready,
    output wire [WORD_WIDTH-1:0]    output_data
);

    localparam WORD_ZERO = {WORD_WIDTH{1'b0}};

Storage for the data

    reg half_buffer_load = 1'b0;

    Register
    #(
        .WORD_WIDTH     (WORD_WIDTH),
        .RESET_VALUE    (WORD_ZERO)
    )
    half_buffer
    (
        .clock          (clock),
        .clock_enable   (half_buffer_load),
        .clear          (clear),
        .data_in        (input_data),
        .data_out       (output_data)
    );

And an empty/full bit associated with the data storage.

    reg  set_to_empty = 1'b0;
    reg  set_to_full  = 1'b0;
    wire buffer_full;

    Register
    #(
        .WORD_WIDTH     (1),
        .RESET_VALUE    (1'b0)
    )
    empty_full
    (
        .clock          (clock),
        .clock_enable   (set_to_full),
        .clear          (set_to_empty),
        .data_in        (1'b1),
        .data_out       (buffer_full)
    );

Then, from the state of the empty/full bit and the signals local only to each handshake (no path between them), determine when data can transfer. Note that we must be empty before we can set_to_full. Anything else creates a combinational path from the input handshake to the output handshake.

EXCEPTION: In Circular Buffer Mode, input_ready does not depend on any other logic, which enables simultaneous input and output handhsakes without combinational paths between them.

    always @(*) begin
        input_ready      = (buffer_full   == 1'b0) || (CIRCULAR_BUFFER != 0);
        output_valid     = (buffer_full   == 1'b1);
        set_to_full      = (input_valid   == 1'b1) && (input_ready  == 1'b1);
        set_to_empty     = (output_valid  == 1'b1) && (output_ready == 1'b1) && (set_to_full == 1'b0);
        set_to_empty     = (set_to_empty  == 1'b1) || (clear == 1'b1);
        half_buffer_load = (set_to_full   == 1'b1);
    end

endmodule

Back to FPGA Design Elements

fpgacpu.ca