// retroDE_ps2 — tb_top_psmct32_tile2x2_demo (Ch304)
//
// 2x2 MULTI-TILE renderer TB. ONE combined TME+ABE+ZTE triangle spanning the
// 32x32 region (a 2x2 grid of 16x16 tiles), crossing BOTH tile seams (x=16 and
// y=16). The renderer re-tests the triangle against each of the 4 tiles
// (CLEAR -> RENDER-clipped -> FLUSH per tile). Proves:
//   PROOF GRID  : all 4 tiles (col,row in {0,1}x{0,1}) clear independently
//                 (256 tile_color writes each) and flush (1024 FB emits total).
//   PROOF SEAM  : the WHOLE 32x32 scanout matches a SINGLE screen-space reference
//                 (one continuous barycentric function of screen x,y) — so the
//                 image is identical whether a pixel was rendered in tile (0,0)
//                 or its neighbour. Matching across x=15|16 and y=15|16 with NO
//                 discontinuity IS the seam proof. (Plus an explicit seam-pixel
//                 match count.)
//   PROOF DEPTH/BLEND : top half blended (red->orange / blue->teal over green),
//                 bottom half occluded green — same per-pixel rule as Ch302/303.

`timescale 1ns/1ps

module tb_top_psmct32_tile2x2_demo;

    localparam int H_ACTIVE = 32, V_ACTIVE = 32;
    localparam int TP_CLEAR=1, TP_FLUSH=3;
    localparam int ZBG = 'h4000;

    logic clk, 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, dma_done_seen, frame_seen, raster_overflow;
    logic        frame_toggle, dma_done_toggle;

    top_psmct32_raster_demo_bram #(
        .H_ACTIVE(H_ACTIVE), .V_ACTIVE(V_ACTIVE),
        .VRAM_BYTES(16*1024), .PSMCT32_SWIZZLE(1'b0),
        .COMBINED_TAZ(1'b1), .TILE_LOCAL(1'b1), .TILE_COLS(2), .TILE_ROWS(2)
    ) 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)
    );

    // ---- single screen-space reference: triangle v0(3,3) v1(28,3) v2(16,29) ----
    function automatic real bdet(input real ax,input real ay,input real bx,input real by,input real cx,input real cy);
        bdet=(by-cy)*(ax-cx)+(cx-bx)*(ay-cy); endfunction
    function automatic real bwa(input real px,input real py,input real ax,input real ay,input real bx,input real by,input real cx,input real cy);
        bwa=((by-cy)*(px-cx)+(cx-bx)*(py-cy))/bdet(ax,ay,bx,by,cx,cy); endfunction
    function automatic real bwb(input real px,input real py,input real ax,input real ay,input real bx,input real by,input real cx,input real cy);
        bwb=((cy-ay)*(px-cx)+(ax-cx)*(py-cy))/bdet(ax,ay,bx,by,cx,cy); endfunction

    // ---- grid tracers ----
    int clear_cw_tile [0:1][0:1];   // tile_color writes per (row,col) during CLEAR
    int flush_emits;
    initial begin
        for (int rr=0;rr<2;rr++) for (int cc=0;cc<2;cc++) clear_cw_tile[rr][cc]=0;
        flush_emits=0;
    end
    always_ff @(posedge clk) if (rst_n) begin
        if (dut.u_gs.tile_phase==TP_CLEAR && dut.u_gs.tile_color_we) begin
            int cc, rr; cc=int'(dut.u_gs.tile_col_r); rr=int'(dut.u_gs.tile_row_r);
            if (cc<2 && rr<2) clear_cw_tile[rr][cc] <= clear_cw_tile[rr][cc] + 1;
        end
        if (dut.u_gs.raster_pixel_emit) flush_emits <= flush_emits + 1;  // tile mode: only flush emits
    end

    // ---- scanout capture (32x32) ----
    logic [7:0] cap_r[0:V_ACTIVE-1][0:H_ACTIVE-1], cap_g[0:V_ACTIVE-1][0:H_ACTIVE-1], 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]=0;cap_g[y][x]=0;cap_b[y][x]=0;cap_de[y][x]=0; end
        capture_armed=1'b0;
    end
    logic [31:0] hcnt_d, vcnt_d;
    always_ff @(posedge clk) begin
        if (!rst_n) begin hcnt_d<=0; vcnt_d<=0; 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, n_match, n_check, n_seam;

    initial begin
        errors=0; n_match=0; n_check=0; n_seam=0;
        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);
        @(posedge dut.u_gs.raster_active);   // grid render begins
        @(negedge dut.u_gs.raster_active);   // all 4 tiles done (clear+render+flush)
        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;

        // PROOF GRID — each of the 4 tiles cleared independently + total flushes.
        for (int rr=0;rr<2;rr++) for (int cc=0;cc<2;cc++)
            if (clear_cw_tile[rr][cc] != 256) begin
                $error("[t2x2] tile(col=%0d,row=%0d) cleared %0d color entries (want 256)", cc, rr, clear_cw_tile[rr][cc]); errors++;
            end
        if (flush_emits != 4*256) begin $error("[t2x2] flush emitted %0d (want %0d)", flush_emits, 4*256); errors++; end

        // PROOF SEAM + DEPTH/BLEND — whole 32x32 vs ONE screen-space reference.
        for (int py=0; py<V_ACTIVE; py++) begin
            for (int px=0; px<H_ACTIVE; px++) begin
                real wa, wb, wc, fragz, uu;
                bit clearly_in, clearly_out;
                logic [7:0] er, eg, eb;
                wa = bwa(px,py, 3,3, 28,3, 16,29);
                wb = bwb(px,py, 3,3, 28,3, 16,29);
                wc = 1.0 - wa - wb;
                clearly_in  = (wa>0.06)&&(wb>0.06)&&(wc>0.06);
                clearly_out = (wa<-0.06)||(wb<-0.06)||(wc<-0.06);
                fragz = (wa+wb)*real'('h6000) + wc*real'('h2000);
                // expected color
                if (clearly_out) begin er=8'h00; eg=8'h80; eb=8'h00; end           // green clear
                else if (clearly_in && (fragz <= real'(ZBG)-1024.0)) begin er=8'h00; eg=8'h80; eb=8'h00; end  // occluded green
                else if (clearly_in && (fragz >= real'(ZBG)+1024.0)) begin
                    uu = wb*7.0 + wc*3.0;
                    if (uu < 3.5) begin er=8'd127; eg=8'd64; eb=8'd0;   end
                    else if (uu > 4.5) begin er=8'd0; eg=8'd64; eb=8'd127; end
                    else begin er=8'hxx; eg=8'hxx; eb=8'hxx; end                    // cell-boundary: skip
                end else begin er=8'hxx; eg=8'hxx; eb=8'hxx; end                    // edge/Z-boundary: skip

                if (er!==8'hxx && cap_de[py][px]) begin
                    n_check++;
                    if (cap_r[py][px]===er && cap_g[py][px]===eg && cap_b[py][px]===eb) begin
                        n_match++;
                        if (px==15||px==16||py==15||py==16) n_seam++;   // seam-region pixels that matched
                    end else begin
                        if (errors<12) $error("[t2x2] (%0d,%0d) got(%02x,%02x,%02x) exp(%02x,%02x,%02x)",px,py,cap_r[py][px],cap_g[py][px],cap_b[py][px],er,eg,eb);
                        errors++;
                    end
                end
            end
        end

        if (n_check < 200) begin $error("[t2x2] too few checkable pixels (%0d)", n_check); errors++; end
        if (n_match != n_check) begin $error("[t2x2] image match %0d/%0d (seam/tiling error)", n_match, n_check); errors++; end
        if (n_seam < 20) begin $error("[t2x2] too few SEAM-region matches (%0d) — seam not exercised", n_seam); errors++; end
        if (!core_halt)      begin $error("core_halt low"); errors++; end
        if (raster_overflow) begin $error("raster_overflow"); errors++; end

        $display("[tb_top_psmct32_tile2x2_demo] clears(per-tile)=[%0d %0d %0d %0d] flush=%0d match=%0d/%0d seam_matches=%0d errors=%0d",
                 clear_cw_tile[0][0],clear_cw_tile[0][1],clear_cw_tile[1][0],clear_cw_tile[1][1], flush_emits, n_match, n_check, n_seam, errors);
        if (errors==0) $display("[tb_top_psmct32_tile2x2_demo] PASS");
        else           $display("[tb_top_psmct32_tile2x2_demo] FAIL");
        $finish;
    end

    initial begin #250000000; $error("[tb_top_psmct32_tile2x2_demo] TIMEOUT"); $finish; end
endmodule : tb_top_psmct32_tile2x2_demo