Snapshot: fog implementation + fidelity tooling baseline (pre bilinear-clamp fix)
Per-vertex GS fog end-to-end (gs_stub emit incl. persp_emit5, gs_prim_list_feeder XYZ2->XYZF2 on PRIM.FGE, gs_make_sh3_scheduler_fixture.py F/FGE packing), new fog TBs, fidelity attribution tooling. Functional baseline before removing the dead bilinear lerp8 clamps (Codex: 161-node comb loop -> -0.042ns setup fail). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -0,0 +1,346 @@
|
||||
// retroDE_ps2 — tb_gs_fog_persp
|
||||
//
|
||||
// Focused white-box TB for GS per-vertex FOG on the PERSPECTIVE-textured emit
|
||||
// path (persp_emit5). Forces PERSPECTIVE_CORRECT=1. The SH3 board scene runs
|
||||
// this build (PERSPECTIVE_CORRECT=1, COMBINED_TAZ=0, TILE_LOCAL=0), so its
|
||||
// geometry emits through persp_emit5 — the path round-1 left unfogged.
|
||||
//
|
||||
// A perspective-textured DECAL triangle (ST/Q supplied so ras_persp activates)
|
||||
// samples a SOLID texture (tex_rd_data tied constant → s1_tex_color == TEXEL
|
||||
// for every pixel), so the emitted color isolates the fog blend and its stage
|
||||
// alignment without needing to model the perspective UV walk. F is affine /
|
||||
// screen-linear per vertex (XYZF2[63:56]).
|
||||
//
|
||||
// Checks:
|
||||
// T1p — FGE=1, flat color, DISTINCT per-vertex F. At every persp_emit5:
|
||||
// (a) ALIGNMENT — the white-box +5-aligned s2_fog_f delay tap
|
||||
// (u_gs.persp_fog_f5) matches the barycentric (screen-linear) F at
|
||||
// the emitted pixel (persp_x5,persp_y5) within 1 LSB. A wrong delay
|
||||
// depth would 1-pixel-shift F and fail here.
|
||||
// (b) BLEND — the captured emitted RGB equals fog_blend(TEXEL,
|
||||
// persp_fog_f5) EXACTLY, alpha unchanged.
|
||||
// T2p — FGE=0, same geometry. Emitted color is BYTE-IDENTICAL to TEXEL (no
|
||||
// fog) — the non-negotiable invariant on the perspective path.
|
||||
|
||||
`timescale 1ns/1ps
|
||||
|
||||
module tb_gs_fog_persp;
|
||||
|
||||
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 [31:0] raster_pixel_mask_q;
|
||||
logic [5:0] raster_pixel_psm_q;
|
||||
logic raster_active;
|
||||
logic raster_overflow;
|
||||
logic raster_fifo_full;
|
||||
logic raster_degenerate;
|
||||
logic tex_rd_en; logic [31:0] tex_rd_addr;
|
||||
logic fb_rd_en; logic [31:0] fb_rd_addr;
|
||||
logic z_rd_en; logic [31:0] z_rd_addr;
|
||||
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;
|
||||
|
||||
// Solid texture: any texel address returns this ABGR word.
|
||||
localparam logic [31:0] TEXEL = 32'hFF_B0_60_20; // A=FF B=B0 G=60 R=20
|
||||
logic [31:0] tex_texel;
|
||||
assign tex_texel = TEXEL;
|
||||
|
||||
gs_stub #(.PERSPECTIVE_CORRECT(1'b1)) 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_mask_q(raster_pixel_mask_q),
|
||||
.raster_pixel_psm_q(raster_pixel_psm_q),
|
||||
.raster_active(raster_active),
|
||||
.raster_overflow(raster_overflow),
|
||||
.raster_fifo_full(raster_fifo_full),
|
||||
.raster_degenerate(raster_degenerate),
|
||||
.tex_rd_en(tex_rd_en), .tex_rd_addr(tex_rd_addr), .tex_rd_data(tex_texel),
|
||||
.fb_rd_en(fb_rd_en), .fb_rd_addr(fb_rd_addr), .fb_rd_data(32'd0),
|
||||
.z_rd_en(z_rd_en), .z_rd_addr(z_rd_addr), .z_rd_data(32'd0),
|
||||
.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)
|
||||
);
|
||||
|
||||
// ----- Drive helpers -----
|
||||
task automatic drive_reg(input logic [7:0] num, input logic [63:0] data);
|
||||
@(negedge clk);
|
||||
gif_reg_wr_en = 1'b1; gif_reg_num = num; gif_reg_data = data;
|
||||
@(posedge clk);
|
||||
endtask
|
||||
task automatic drive_idle();
|
||||
@(negedge clk);
|
||||
gif_reg_wr_en = 1'b0; gif_reg_num = 8'd0; gif_reg_data = 64'd0;
|
||||
@(posedge clk);
|
||||
endtask
|
||||
|
||||
// XYZF2 (reg 0x04): X=[15:0] (12.4), Y=[31:16] (12.4), Z=[55:32], F=[63:56].
|
||||
function automatic logic [63:0] xyzf2(input int x, input int y,
|
||||
input int z, input int f);
|
||||
return {8'(f), 24'(z), 12'(y), 4'd0, 12'(x), 4'd0};
|
||||
endfunction
|
||||
// RGBAQ with Q in [55:32] (24-bit, FRAC=12).
|
||||
function automatic logic [63:0] rgbaq_q_val(input int r, input int g, input int b,
|
||||
input int a, input int q_fp);
|
||||
return {8'd0, 24'(q_fp), 8'(a), 8'(b), 8'(g), 8'(r)};
|
||||
endfunction
|
||||
// ST (reg 0x02): S in [23:0], T in [55:32] (24-bit, FRAC=12).
|
||||
function automatic logic [63:0] st_val(input int s_fp, input int t_fp);
|
||||
return {8'd0, 24'(t_fp), 8'd0, 24'(s_fp)};
|
||||
endfunction
|
||||
|
||||
localparam logic [7:0] R_PRIM = 8'h00;
|
||||
localparam logic [7:0] R_RGBAQ = 8'h01;
|
||||
localparam logic [7:0] R_ST = 8'h02;
|
||||
localparam logic [7:0] R_XYZF2 = 8'h04;
|
||||
localparam logic [7:0] R_TEX0_1 = 8'h06;
|
||||
localparam logic [7:0] R_FOGCOL = 8'h3D;
|
||||
localparam logic [7:0] R_FRAME_1 = 8'h4C;
|
||||
|
||||
// PRIM: TRIANGLE(3) + TME(bit4) + FGE(bit5).
|
||||
localparam logic [63:0] PRIM_TRI_TEX_FGE = 64'd3 | (64'd1 << 4) | (64'd1 << 5);
|
||||
localparam logic [63:0] PRIM_TRI_TEX = 64'd3 | (64'd1 << 4); // FGE=0
|
||||
localparam logic [63:0] FRAME_1_VAL = 64'h0000_0000_0001_0000; // FBW=1, PSMCT32
|
||||
// TEX0_1: TBP0=8, TBW=1, PSM=PSMCT32(0), TW=TH=0, TCC=0, TFX=DECAL(1)@[36:35].
|
||||
localparam logic [63:0] TEX0_VAL = 64'd8 | (64'd1 << 14) | (64'd1 << 35);
|
||||
localparam int Q_ONE = 24'h001000; // 1.0 in FRAC=12
|
||||
|
||||
localparam int FOG_R = 8'h20, FOG_G = 8'h40, FOG_B = 8'h60;
|
||||
localparam logic [63:0] FOGCOL_VAL = {40'd0, 8'(FOG_B), 8'(FOG_G), 8'(FOG_R)};
|
||||
|
||||
function automatic int fog_ch(input int c, input int f, input int fc);
|
||||
return (c * f + fc * (255 - f)) >> 8;
|
||||
endfunction
|
||||
function automatic int absdiff(input int a, input int b);
|
||||
return (a > b) ? (a - b) : (b - a);
|
||||
endfunction
|
||||
|
||||
// ----- Triangle reference (edge fn + fill rule + barycentric for F) -----
|
||||
int vx0, vy0, vx1, vy1, vx2, vy2, vf0, vf1, vf2;
|
||||
int ref_det, sx1, sy1, sx2, sy2, sf1, sf2;
|
||||
|
||||
function automatic int edge_f(input int px, input int py,
|
||||
input int ax, input int ay,
|
||||
input int bx, input int by);
|
||||
return (px - ax) * (by - ay) - (py - ay) * (bx - ax);
|
||||
endfunction
|
||||
task automatic setup_ref();
|
||||
int sa;
|
||||
sa = (vx1 - vx0) * (vy2 - vy0) - (vy1 - vy0) * (vx2 - vx0);
|
||||
if (sa < 0) begin
|
||||
sx1 = vx2; sy1 = vy2; sf1 = vf2;
|
||||
sx2 = vx1; sy2 = vy1; sf2 = vf1;
|
||||
ref_det = -sa;
|
||||
end else begin
|
||||
sx1 = vx1; sy1 = vy1; sf1 = vf1;
|
||||
sx2 = vx2; sy2 = vy2; sf2 = vf2;
|
||||
ref_det = sa;
|
||||
end
|
||||
endtask
|
||||
function automatic int ref_f(input int px, input int py);
|
||||
int L0, L1, L2, num;
|
||||
L0 = -edge_f(px, py, sx1, sy1, sx2, sy2);
|
||||
L1 = -edge_f(px, py, sx2, sy2, vx0, vy0);
|
||||
L2 = -edge_f(px, py, vx0, vy0, sx1, sy1);
|
||||
num = L0 * vf0 + L1 * sf1 + L2 * sf2;
|
||||
if (ref_det == 0) return 0;
|
||||
return num / ref_det;
|
||||
endfunction
|
||||
|
||||
int errors, align_checks, blend_checks;
|
||||
|
||||
// Emit capture: expected color computed at persp_emit5 from the RTL's own
|
||||
// +5-aligned F (persp_fog_f5), plus the alignment check vs barycentric F.
|
||||
logic [31:0] exp_c [0:31][0:31];
|
||||
bit exp_set [0:31][0:31];
|
||||
bit fge_on;
|
||||
|
||||
task automatic clear_exp();
|
||||
for (int y = 0; y < 32; y++)
|
||||
for (int x = 0; x < 32; x++) begin
|
||||
exp_c[y][x] = 32'd0; exp_set[y][x] = 1'b0;
|
||||
end
|
||||
endtask
|
||||
|
||||
// persp_emit5 monitor: alignment + record expected fogged color.
|
||||
always_ff @(posedge clk) begin
|
||||
if (rst_n && u_gs.ras_persp && u_gs.persp_emit5) begin
|
||||
int px, py, ef, af;
|
||||
logic [31:0] want;
|
||||
px = int'(u_gs.persp_x5);
|
||||
py = int'(u_gs.persp_y5);
|
||||
af = int'(u_gs.persp_fog_f5);
|
||||
if (px < 32 && py < 32) begin
|
||||
// (a) alignment: RTL +5-aligned F vs barycentric F at this pixel.
|
||||
ef = ref_f(px, py);
|
||||
if (absdiff(af, ef) > 1) begin
|
||||
$error("[persp align] (%0d,%0d) persp_fog_f5=%0d expected ~%0d", px, py, af, ef);
|
||||
errors = errors + 1;
|
||||
end else begin
|
||||
align_checks = align_checks + 1;
|
||||
end
|
||||
// (b) expected emitted color = fog(TEXEL, af) if FGE, else TEXEL.
|
||||
if (fge_on)
|
||||
want = {TEXEL[31:24],
|
||||
8'(fog_ch(int'(TEXEL[23:16]), af, FOG_B)),
|
||||
8'(fog_ch(int'(TEXEL[15:8]), af, FOG_G)),
|
||||
8'(fog_ch(int'(TEXEL[7:0]), af, FOG_R))};
|
||||
else
|
||||
want = TEXEL;
|
||||
exp_c[py][px] <= want;
|
||||
exp_set[py][px] <= 1'b1;
|
||||
end
|
||||
end
|
||||
end
|
||||
|
||||
// Capture the actual emitted pixel and compare to the expected recorded above.
|
||||
always_ff @(posedge clk) begin
|
||||
if (rst_n && raster_pixel_emit
|
||||
&& raster_pixel_x_q < 32 && raster_pixel_y_q < 32) begin
|
||||
if (!exp_set[raster_pixel_y_q][raster_pixel_x_q]) begin
|
||||
$error("[persp emit] (%0d,%0d) emitted but no persp_emit5 expected color recorded",
|
||||
raster_pixel_x_q, raster_pixel_y_q);
|
||||
errors = errors + 1;
|
||||
end else if (raster_pixel_color_q[31:0] !== exp_c[raster_pixel_y_q][raster_pixel_x_q]) begin
|
||||
$error("[persp emit] (%0d,%0d) got %08x expected %08x",
|
||||
raster_pixel_x_q, raster_pixel_y_q,
|
||||
raster_pixel_color_q[31:0], exp_c[raster_pixel_y_q][raster_pixel_x_q]);
|
||||
errors = errors + 1;
|
||||
end else begin
|
||||
blend_checks = blend_checks + 1;
|
||||
end
|
||||
end
|
||||
end
|
||||
|
||||
task automatic drive_persp_tri(input logic [63:0] prim_word);
|
||||
drive_reg(R_FOGCOL, FOGCOL_VAL);
|
||||
drive_reg(R_PRIM, prim_word);
|
||||
drive_reg(R_FRAME_1, FRAME_1_VAL);
|
||||
drive_reg(R_TEX0_1, TEX0_VAL);
|
||||
drive_reg(R_RGBAQ, rgbaq_q_val(8'hFF, 8'hFF, 8'hFF, 8'hFF, Q_ONE));
|
||||
drive_reg(R_ST, st_val(0, 0));
|
||||
drive_reg(R_XYZF2, xyzf2(vx0, vy0, 0, vf0));
|
||||
drive_reg(R_RGBAQ, rgbaq_q_val(8'hFF, 8'hFF, 8'hFF, 8'hFF, Q_ONE));
|
||||
drive_reg(R_ST, st_val(0, 0));
|
||||
drive_reg(R_XYZF2, xyzf2(vx1, vy1, 0, vf1));
|
||||
drive_reg(R_RGBAQ, rgbaq_q_val(8'hFF, 8'hFF, 8'hFF, 8'hFF, Q_ONE));
|
||||
drive_reg(R_ST, st_val(0, 0));
|
||||
drive_reg(R_XYZF2, xyzf2(vx2, vy2, 0, vf2)); // closes
|
||||
drive_idle();
|
||||
repeat (900) @(posedge clk);
|
||||
endtask
|
||||
|
||||
initial begin
|
||||
errors = 0; align_checks = 0; blend_checks = 0;
|
||||
rst_n = 1'b0; gif_reg_wr_en = 1'b0; gif_reg_num = 8'd0; gif_reg_data = 64'd0;
|
||||
clear_exp();
|
||||
|
||||
repeat (4) @(posedge clk);
|
||||
rst_n = 1'b1;
|
||||
repeat (2) @(posedge clk);
|
||||
|
||||
// ============================================================
|
||||
// T1p — FGE=1, distinct per-vertex F, solid texel. Alignment + blend.
|
||||
// ============================================================
|
||||
vx0=2; vy0=2; vf0=8'h20;
|
||||
vx1=12; vy1=3; vf1=8'hE0;
|
||||
vx2=4; vy2=9; vf2=8'h80;
|
||||
setup_ref();
|
||||
clear_exp(); fge_on = 1'b1;
|
||||
drive_persp_tri(PRIM_TRI_TEX_FGE);
|
||||
if (align_checks == 0) begin
|
||||
$error("[T1p] no persp_emit5 pixels observed (persp path did not fire)");
|
||||
errors = errors + 1;
|
||||
end
|
||||
if (raster_overflow || raster_degenerate) begin
|
||||
$error("[T1p] raster anomaly"); errors = errors + 1;
|
||||
end
|
||||
|
||||
// ============================================================
|
||||
// T2p — FGE=0, same geometry. Emitted MUST be byte-identical to TEXEL.
|
||||
// ============================================================
|
||||
clear_exp(); fge_on = 1'b0;
|
||||
drive_persp_tri(PRIM_TRI_TEX);
|
||||
|
||||
$display("[tb_gs_fog_persp] align_checks=%0d blend_checks=%0d errors=%0d",
|
||||
align_checks, blend_checks, errors);
|
||||
if (errors == 0) $display("[tb_gs_fog_persp] PASS");
|
||||
else $display("[tb_gs_fog_persp] FAIL");
|
||||
$finish;
|
||||
end
|
||||
|
||||
initial begin
|
||||
#8000000;
|
||||
$error("[tb_gs_fog_persp] timeout");
|
||||
$finish;
|
||||
end
|
||||
|
||||
endmodule : tb_gs_fog_persp
|
||||
Reference in New Issue
Block a user