// retroDE_ps2 — tb_top_psmct32_zbuffer_demo (Brick 2b)
//
// TOP-LEVEL Z-buffer (depth test) demo TB for the BRAM BOARD VARIANT.
// Proves the synthesizable FLAT Z-tested PSMCT32 SPRITE path renders
// end-to-end through `top_psmct32_raster_demo_bram` exactly as a board
// load would SHOW on HDMI.
//
//   EE bootlet (bios_zbuffer.mem) + GIF payload (payload_zbuffer.mem):
//     U1 PACKED -> NEAR red SPRITE, Z=0x200, x[0..11] y[0..7],
//                  TEST_1.ZTE=1 ZTST=GEQUAL, ZBUF_1 ZBP=1 PSMZ32 ZMSK=0.
//     U2 PACKED -> FAR blue SPRITE, Z=0x100, x[4..15] y[0..7] (drawn
//                  second; GIF context PRIM/FRAME/TEST/ZBUF persist).
//
// The Z buffer (ZBP=1 -> byte base 0x800) starts cleared (VRAM power-on
// zero). NEAR draws first (Z 0x200 >= 0 everywhere it covers, passes),
// stamping Z=0x200 at x[0..11]. FAR draws second; with GEQUAL its
// Z=0x100 is BEHIND the near Z(0x200) in the overlap x[4..11], so it
// FAILS and does NOT overwrite — near wins regardless of draw order.
//
//   near-only  x[0..3]   -> RED   (0xFF,0x00,0x00)
//   overlap    x[4..11]  -> RED   (near wins; NOT blue last-drawn)
//   far-only   x[12..15] -> BLUE  (0x00,0x00,0xFF; stored Z=0 there)
//
// The overlap guard asserts the overlap is RED, NOT BLUE — proving the
// depth test actually gated the FAR write rather than last-write-wins.

`timescale 1ns/1ps

module tb_top_psmct32_zbuffer_demo;

    localparam int H_ACTIVE = 16;
    localparam int V_ACTIVE = 8;

    logic clk;
    logic rst_n;
    initial clk = 1'b0;
    always #5 clk = ~clk;

    logic        core_go;
    logic [7:0]  r, g, b;
    logic        hsync, vsync, de;
    logic        core_halt;
    logic        dma_done_seen;
    logic        frame_seen;
    logic        raster_overflow;
    logic        frame_toggle;
    logic        dma_done_toggle;

    top_psmct32_raster_demo_bram #(
        .H_ACTIVE       (H_ACTIVE),
        .V_ACTIVE       (V_ACTIVE),
        .PSMCT32_SWIZZLE(1'b0)
    ) dut (
        .clk(clk), .rst_n(rst_n),
        .core_go(core_go),
        .r(r), .g(g), .b(b),
        .hsync(hsync), .vsync(vsync), .de(de),
        .core_halt(core_halt),
        .dma_done_seen(dma_done_seen),
        .frame_seen(frame_seen),
        .raster_overflow(raster_overflow),
        .frame_toggle(frame_toggle),
        .dma_done_toggle(dma_done_toggle),
        .joy_a_pressed_i(1'b0),
        .joy_b_pressed_i(1'b0)
    );

    // ----- Expected colors -----
    localparam logic [7:0] NEAR_R = 8'hFF, NEAR_G = 8'h00, NEAR_B = 8'h00; // red
    localparam logic [7:0] FAR_R  = 8'h00, FAR_G  = 8'h00, FAR_B  = 8'hFF; // blue

    // ----- Frame capture (delayed counters: BRAM read is registered) -----
    logic [7:0] cap_r [0:V_ACTIVE-1][0:H_ACTIVE-1];
    logic [7:0] cap_g [0:V_ACTIVE-1][0:H_ACTIVE-1];
    logic [7:0] cap_b [0:V_ACTIVE-1][0:H_ACTIVE-1];
    logic       cap_de[0:V_ACTIVE-1][0:H_ACTIVE-1];
    bit         capture_armed;

    initial begin
        for (int y = 0; y < V_ACTIVE; y++)
            for (int x = 0; x < H_ACTIVE; x++) begin
                cap_r[y][x]  = 8'd0;
                cap_g[y][x]  = 8'd0;
                cap_b[y][x]  = 8'd0;
                cap_de[y][x] = 1'b0;
            end
        capture_armed = 1'b0;
    end

    logic [31:0] hcnt_d;
    logic [31:0] vcnt_d;
    always_ff @(posedge clk) begin
        if (!rst_n) begin
            hcnt_d <= 32'd0;
            vcnt_d <= 32'd0;
        end else begin
            hcnt_d <= 32'(dut.u_pcrtc.hcnt);
            vcnt_d <= 32'(dut.u_pcrtc.vcnt);
        end
    end

    always_ff @(posedge clk) begin
        if (rst_n && capture_armed && de
            && (vcnt_d < V_ACTIVE) && (hcnt_d < H_ACTIVE)) begin
            cap_r [vcnt_d][hcnt_d] <= r;
            cap_g [vcnt_d][hcnt_d] <= g;
            cap_b [vcnt_d][hcnt_d] <= b;
            cap_de[vcnt_d][hcnt_d] <= 1'b1;
        end
    end

    int errors;
    int near_ok;
    int overlap_ok;
    int far_ok;
    initial begin
        errors = 0; near_ok = 0; overlap_ok = 0; far_ok = 0;
    end


    // Returns 1 (via `ok` output) when the pixel matched; iverilog-12
    // doesn't support task reference ports, so the caller increments
    // the per-region counter from `ok`.
    task automatic chk(input int x, input int y,
                       input logic [7:0] er, input logic [7:0] eg, input logic [7:0] eb,
                       input string label, output bit ok);
        ok = 1'b0;
        if (!cap_de[y][x]) begin
            $error("%s (%0d,%0d) DE never asserted", label, x, y);
            errors = errors + 1;
        end else if (cap_r[y][x] !== er || cap_g[y][x] !== eg || cap_b[y][x] !== eb) begin
            $error("[%s] (%0d,%0d) got (%02x,%02x,%02x) expected (%02x,%02x,%02x)",
                   label, x, y, cap_r[y][x], cap_g[y][x], cap_b[y][x], er, eg, eb);
            errors = errors + 1;
        end else begin
            ok = 1'b1;
        end
    endtask

    initial begin
        rst_n   = 1'b0;
        core_go = 1'b0;
        repeat (4) @(posedge clk);
        rst_n = 1'b1;
        repeat (8) @(posedge clk);

        @(negedge clk);
        core_go = 1'b1;
        @(negedge clk);
        core_go = 1'b0;

        wait (core_halt == 1'b1);
        repeat (4) @(posedge clk);
        wait (dma_done_seen == 1'b1);
        repeat (10) @(posedge clk);
        if (dut.xfer_busy == 1'b1)
            wait (dut.xfer_busy == 1'b0);
        if (dut.u_gs.raster_active == 1'b1)
            wait (dut.u_gs.raster_active == 1'b0);
        repeat (10) @(posedge clk);

        @(posedge dut.u_pcrtc.end_of_frame);
        @(posedge clk);
        capture_armed = 1'b1;
        @(posedge dut.u_pcrtc.end_of_frame);
        @(posedge clk);
        capture_armed = 1'b0;

        begin
            bit ok;
            for (int y = 0; y < 8; y++) begin
                // near-only x[0..3] -> RED
                for (int x = 0; x < 4; x++) begin
                    chk(x, y, NEAR_R, NEAR_G, NEAR_B, "near", ok);
                    if (ok) near_ok = near_ok + 1;
                end
                // overlap x[4..11] -> RED (near wins, NOT far/last-drawn)
                for (int x = 4; x < 12; x++) begin
                    chk(x, y, NEAR_R, NEAR_G, NEAR_B, "overlap", ok);
                    if (ok) overlap_ok = overlap_ok + 1;
                    // Guard: overlap must NOT be the FAR (last-drawn) color.
                    if (cap_r[y][x] === FAR_R && cap_g[y][x] === FAR_G && cap_b[y][x] === FAR_B) begin
                        $error("[overlap] (%0d,%0d) is FAR (last-drawn) — depth NOT gated!", x, y);
                        errors = errors + 1;
                    end
                end
                // far-only x[12..15] -> BLUE
                for (int x = 12; x < 16; x++) begin
                    chk(x, y, FAR_R, FAR_G, FAR_B, "far", ok);
                    if (ok) far_ok = far_ok + 1;
                end
            end
        end

        if (!core_halt)      begin $error("core_halt low at end");        errors = errors + 1; end
        if (!dma_done_seen)  begin $error("dma_done_seen never latched"); errors = errors + 1; end
        if (!frame_seen)     begin $error("frame_seen never latched");    errors = errors + 1; end
        if (raster_overflow) begin $error("raster_overflow set");         errors = errors + 1; end

        $display("[tb_top_psmct32_zbuffer_demo] near_ok=%0d/32 overlap_ok=%0d/64 far_ok=%0d/32 raster_emits=%0d errors=%0d",
                 near_ok, overlap_ok, far_ok,
                 dut.u_gs.raster_pixel_emit_count, errors);
        if (errors == 0)
            $display("[tb_top_psmct32_zbuffer_demo] PASS");
        else
            $display("[tb_top_psmct32_zbuffer_demo] FAIL");
        $finish;
    end

    initial begin
        #20000000;
        $error("[tb_top_psmct32_zbuffer_demo] TIMEOUT");
        $finish;
    end

endmodule : tb_top_psmct32_zbuffer_demo