`include "defs.svh"

module mem_arbiter
    (   input   bit             clock
    ,   input   bit             reset

    ,   output  bit             command_ready
    ,   input   bit             command_valid
    ,   input   arb_to_ram_t    command_data

    ,   output  bit             [`NUM_PDPS-1:0] pdp_ready
    ,   input   bit             [`NUM_PDPS-1:0] pdp_valid
    ,   input   core_to_mem_t   [`NUM_PDPS-1:0] pdp_data

    ,   input   bit             ram_ready
    ,   output  bit             ram_valid
    ,   output  arb_to_ram_t    ram_data
    );

    bit             [`NUM_PDPS:0]   hold_valid;
    arb_to_ram_t    [`NUM_PDPS:0]   hold_data;

    bit [$clog2(`NUM_PDPS+1):0] selector;

    always @(posedge clock) begin
        if (reset) begin
            command_ready = 0;
            for (int i = 0; i < `NUM_PDPS; ++i)
                pdp_ready[i] = 0;
            ram_valid = 0;
            for (int i = 0; i < `NUM_PDPS+1; ++i)
                hold_valid[i] = 0;
            selector = 0;
        end else begin
            automatic bit selector_stale = 0;

            if (ram_ready) ram_valid = 0;

            if (!ram_valid && hold_valid[selector]) begin
                ram_valid = 1;
                ram_data = hold_data[selector];
                hold_valid[selector] = 0;
                selector_stale = 1;
            end

            if (hold_valid == 0) begin
                if (command_ready && command_valid) begin
                    hold_valid[0] = 1;
                    hold_data[0] = command_data;
                    hold_data[0].tag = 0;
                end
                for (int i = 0; i < `NUM_PDPS; ++i) begin
                    if (pdp_ready[i] && pdp_valid[i]) begin
                        hold_valid[i+1] = 1;
                        hold_data[i+1].address[`RAM_ADDRESS_BITS-1:`PDP_ADDRESS_BITS] = i;
                        hold_data[i+1].address[`PDP_ADDRESS_BITS-1:$clog2(`RAM_LINE_WORDS)] = pdp_data[i].address;
                        hold_data[i+1].write = pdp_data[i].write;
                        hold_data[i+1].data = pdp_data[i].data;
                        hold_data[i+1].mask = pdp_data[i].mask;
                        hold_data[i+1].tag = i+1;
                    end
                end
            end

            if (hold_valid != 0 && !ram_valid) begin
                for (int i = 0; i < `NUM_PDPS+1; ++i) begin
                    if (hold_valid == 1 << i) begin
                        ram_valid = 1;
                        ram_data = hold_data[i];
                        hold_valid[i] = 0;
                        break;
                    end
                end
            end

            if (hold_valid != 0 && (selector_stale || !ram_valid)) begin
                for (int i = 0; i < `NUM_PDPS+1; ++i) begin
                    automatic int j = selector + i;
                    if (j > `NUM_PDPS)
                        j -= `NUM_PDPS+1;
                    if (hold_valid[j]) begin
                        selector = j;
                        break;
                    end
                end
            end

            command_ready = hold_valid == 0;
            for (int i = 0; i < `NUM_PDPS; ++i)
                pdp_ready[i] = hold_valid == 0;
        end
    end

endmodule