// retroDE_ps2 — tb_gs_vram_writeback (Ch89)
//
// White-box TB pinning down the Ch89 VRAM persistence contract:
// every raster_pixel_emit pulse must land 4 bytes of pixel data
// (PSMCT32 ABGR) at raster_pixel_fb_addr_q in vram_stub, and a
// later read-back at the same address must return EXACTLY the
// emitted color.
//
// Setup: gs_stub + vram_stub instantiated together; gif_reg_*
// driven directly (no DMA). Drive a single 4×4 SPRITE so 16
// pixels emit. Capture the (fb_addr, color) of each emit at the
// posedge it fires; after raster completes, walk each captured
// address and assert vram_stub.read_data == captured color.
//
// Stream:
//   PRIM=SPRITE, FRAME_1 (FBP=2 FBW=10 PSMCT32), RGBAQ
//   v1 = (0, 0)
//   v2 = (3, 3)  — close S1, sprite is 4×4 = 16 pixels

`timescale 1ns/1ps

module tb_gs_vram_writeback;

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

    logic        gif_reg_wr_en;
    logic [7:0]  gif_reg_num;
    logic [63:0] gif_reg_data;

    logic [7:0]          bg_r, bg_g, bg_b;
    logic [63:0]         prim_q, rgbaq_q, xyz2_q, xyzf2_q, frame_1_q, zbuf_1_q;
    logic                prim_complete;
    logic [31:0]         prim_complete_count;
    logic [63:0]         prim_v0_q, prim_v1_q, prim_v2_q;
    logic [63:0]         prim_color_q;
    logic [63:0]         prim_color_v0_q, prim_color_v1_q, prim_color_v2_q;
    trace_pkg::vertex_t  prim_v0_decoded_q, prim_v1_decoded_q, prim_v2_decoded_q;
    trace_pkg::color_t   prim_v0_color_decoded_q, prim_v1_color_decoded_q, prim_v2_color_decoded_q;
    logic                pixel_emit;
    logic [31:0]         pixel_emit_count;
    logic [11:0]         pixel_x_q, pixel_y_q;
    logic [63:0]         pixel_color_q;
    logic [8:0]          pixel_fbp_q;
    logic [5:0]          pixel_fbw_q, pixel_psm_q;
    logic [31:0]         pixel_fb_addr_q;
    logic                raster_pixel_emit;
    logic [31:0]         raster_pixel_emit_count;
    logic [11:0]         raster_pixel_x_q, raster_pixel_y_q;
    logic [63:0]         raster_pixel_color_q;
    logic [31:0]         raster_pixel_fb_addr_q;
    logic [3:0]          raster_pixel_be_q;
    logic [5:0]          raster_pixel_psm_q;
    logic                raster_active;
    logic                raster_overflow;
    logic                raster_degenerate;
    logic                ev_valid;
    trace_pkg::subsys_e  ev_subsys;
    trace_pkg::event_e   ev_event;
    logic [63:0]         ev_arg0, ev_arg1, ev_arg2, ev_arg3;
    logic [31:0]         ev_flags;

    gs_stub u_gs (
        .clk(clk), .rst_n(rst_n),
        .reg_wr_en(1'b0), .reg_wr_addr(16'd0), .reg_wr_data(64'd0),
        .gif_reg_wr_en(gif_reg_wr_en),
        .gif_reg_num(gif_reg_num),
        .gif_reg_data(gif_reg_data),
        .bg_r(bg_r), .bg_g(bg_g), .bg_b(bg_b),
        .prim_q(prim_q), .rgbaq_q(rgbaq_q),
        .xyz2_q(xyz2_q), .xyzf2_q(xyzf2_q),
        .frame_1_q(frame_1_q), .zbuf_1_q(zbuf_1_q),
        .prim_complete(prim_complete),
        .prim_complete_count(prim_complete_count),
        .prim_v0_q(prim_v0_q), .prim_v1_q(prim_v1_q), .prim_v2_q(prim_v2_q),
        .prim_color_q(prim_color_q),
        .prim_color_v0_q(prim_color_v0_q),
        .prim_color_v1_q(prim_color_v1_q),
        .prim_color_v2_q(prim_color_v2_q),
        .prim_v0_decoded_q(prim_v0_decoded_q),
        .prim_v1_decoded_q(prim_v1_decoded_q),
        .prim_v2_decoded_q(prim_v2_decoded_q),
        .prim_v0_color_decoded_q(prim_v0_color_decoded_q),
        .prim_v1_color_decoded_q(prim_v1_color_decoded_q),
        .prim_v2_color_decoded_q(prim_v2_color_decoded_q),
        .pixel_emit(pixel_emit),
        .pixel_emit_count(pixel_emit_count),
        .pixel_x_q(pixel_x_q), .pixel_y_q(pixel_y_q),
        .pixel_color_q(pixel_color_q),
        .pixel_fbp_q(pixel_fbp_q),
        .pixel_fbw_q(pixel_fbw_q),
        .pixel_psm_q(pixel_psm_q),
        .pixel_fb_addr_q(pixel_fb_addr_q),
        .raster_pixel_emit(raster_pixel_emit),
        .raster_pixel_emit_count(raster_pixel_emit_count),
        .raster_pixel_x_q(raster_pixel_x_q),
        .raster_pixel_y_q(raster_pixel_y_q),
        .raster_pixel_color_q(raster_pixel_color_q),
        .raster_pixel_fb_addr_q(raster_pixel_fb_addr_q),
        .raster_pixel_be_q(raster_pixel_be_q),
        .raster_pixel_psm_q(raster_pixel_psm_q),
        .raster_active(raster_active),
        .raster_overflow(raster_overflow),
        .raster_degenerate(raster_degenerate),
        .ev_valid(ev_valid), .ev_subsys(ev_subsys), .ev_event(ev_event),
        .ev_arg0(ev_arg0), .ev_arg1(ev_arg1),
        .ev_arg2(ev_arg2), .ev_arg3(ev_arg3),
        .ev_flags(ev_flags)
    );

    logic [31:0] vram_read_addr;
    logic [31:0] vram_read_data;

    vram_stub u_vram (
        .clk(clk), .rst_n(rst_n),
        .write_en  (raster_pixel_emit),
        .write_addr(raster_pixel_fb_addr_q),
        .write_data(raster_pixel_color_q[31:0]),
        .write_be  (raster_pixel_be_q),
        .write_mask(32'hFFFF_FFFF),
        .read_addr (vram_read_addr),
        .read_data (vram_read_data),
        .read2_addr(32'd0),
        .read2_data()
    );

    // Capture the (fb_addr, color, x, y) of every emit pulse.
    int          emit_count;
    logic [31:0] emit_fb_addr [0:31];
    logic [31:0] emit_color   [0:31];
    logic [11:0] emit_x       [0:31];
    logic [11:0] emit_y       [0:31];

    initial begin
        emit_count = 0;
    end

    always_ff @(posedge clk) begin
        if (rst_n && raster_pixel_emit && emit_count < 32) begin
            emit_fb_addr[emit_count] <= raster_pixel_fb_addr_q;
            emit_color  [emit_count] <= raster_pixel_color_q[31:0];
            emit_x      [emit_count] <= raster_pixel_x_q;
            emit_y      [emit_count] <= raster_pixel_y_q;
            emit_count               <= emit_count + 1;
        end
    end

    task automatic step_drive(input logic        wr_en,
                              input logic [7:0]  num,
                              input logic [63:0] data);
        @(negedge clk);
        gif_reg_wr_en = wr_en;
        gif_reg_num   = num;
        gif_reg_data  = data;
        @(posedge clk);
    endtask

    task automatic drive_reg(input logic [7:0] num, input logic [63:0] data);
        step_drive(1'b1, num, data);
    endtask

    task automatic drive_idle();
        step_drive(1'b0, 8'd0, 64'd0);
    endtask

    function automatic logic [63:0] xyz2_data(input logic [11:0] x_int,
                                              input logic [11:0] y_int);
        return {32'd0, y_int, 4'd0, x_int, 4'd0};
    endfunction

    localparam logic [7:0] R_PRIM    = 8'h00;
    localparam logic [7:0] R_RGBAQ   = 8'h01;
    localparam logic [7:0] R_XYZ2    = 8'h05;
    localparam logic [7:0] R_FRAME_1 = 8'h4C;

    localparam logic [63:0] PRIM_SPRITE = 64'd6;
    localparam logic [63:0] FRAME_1_VAL = 64'h0000_0000_000A_0002;
    // Color: ABGR = 0x_00_00_FF_30 → A=0x00, B=0x00, G=0xFF, R=0x30.
    // raster_pixel_color_q[31:0] should mirror this.
    localparam logic [63:0] RGBAQ_VAL    = 64'h0000_0000_0000_FF30;
    localparam logic [31:0] EXPECTED_PIX = 32'h0000_FF30;

    int errors;

    initial begin
        rst_n         = 1'b0;
        gif_reg_wr_en = 1'b0;
        gif_reg_num   = 8'd0;
        gif_reg_data  = 64'd0;
        errors        = 0;
        vram_read_addr = 32'd0;

        repeat (4) @(posedge clk);
        rst_n = 1'b1;
        repeat (2) @(posedge clk);

        drive_reg(R_PRIM,    PRIM_SPRITE);
        drive_reg(R_FRAME_1, FRAME_1_VAL);
        drive_reg(R_RGBAQ,   RGBAQ_VAL);

        // 4×4 SPRITE — 16 pixels.
        drive_reg(R_XYZ2, xyz2_data(12'd0, 12'd0));   // v1
        drive_reg(R_XYZ2, xyz2_data(12'd3, 12'd3));   // v2 — close S1

        drive_idle();
        repeat (40) @(posedge clk);

        $display("[tb_gs_vram_writeback] emit_count=%0d raster_pixel_emit_count=%0d raster_overflow=%b",
                 emit_count, raster_pixel_emit_count, raster_overflow);

        if (emit_count != 16) begin
            $error("emit_count=%0d (expected 16)", emit_count);
            errors = errors + 1;
        end
        if (raster_overflow !== 1'b0) begin
            $error("raster_overflow=%b (expected 0)", raster_overflow);
            errors = errors + 1;
        end

        // Walk each captured emit and verify VRAM read-back matches.
        for (int i = 0; i < emit_count; i++) begin
            vram_read_addr = emit_fb_addr[i];
            #1;  // let comb read settle
            $display("[tb_gs_vram_writeback] cap[%0d] (%0d,%0d) fb=0x%08x vram=0x%08x expect=0x%08x",
                     i, emit_x[i], emit_y[i], emit_fb_addr[i],
                     vram_read_data, emit_color[i]);
            if (vram_read_data !== emit_color[i]) begin
                $error("VRAM mismatch @ fb=0x%08x: got 0x%08x, expected 0x%08x",
                       emit_fb_addr[i], vram_read_data, emit_color[i]);
                errors = errors + 1;
            end
            if (emit_color[i] !== EXPECTED_PIX) begin
                $error("emit color mismatch @ fb=0x%08x: emitted 0x%08x, expected 0x%08x",
                       emit_fb_addr[i], emit_color[i], EXPECTED_PIX);
                errors = errors + 1;
            end
        end

        // Sanity probe: an address NOT touched by any emit must
        // still read as zero (initial mem state).
        vram_read_addr = 32'h0000_0000;
        #1;
        if (vram_read_data !== 32'd0) begin
            $error("VRAM addr 0 should be 0 (untouched), got 0x%08x", vram_read_data);
            errors = errors + 1;
        end

        if (errors == 0) $display("[tb_gs_vram_writeback] PASS");
        else             $display("[tb_gs_vram_writeback] FAIL");
        $finish;
    end

    initial begin
        #5000000;
        $error("[tb_gs_vram_writeback] timeout");
        $finish;
    end

endmodule : tb_gs_vram_writeback