thejayman77/retroDE_ps2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Ch443b: revert texture 4-bank split; 2-cycle drain (F_SETTLE + multicycle)

Browse Source 
The four-bank tex_mem split (27dfd0b) closed the -0.370 EMIF drain
write-address fanout but the owner GUI fit showed the fitter SCATTERED
the banks, pushing the DESIGN-clock sampler read cone
(ras_v0_x -> perspective-UV -> texel addr -> tex_mem portbaddr, the
design's fundamental ~40ns critical path) to -2.208 ns. Net worse.

Codex's call (Option 1 + honest write-side multicycle), implemented:
- Restore the MONOLITHIC 65536x32 tex_mem, recovering the clean 25 MHz
  read-cone placement. Sampler/read-address path stays fully timed
  (Ch439g); nothing about it is relaxed.
- Make the EMIF drain write genuinely two-cycle: new F_SETTLE state
  between F_DRAIN and F_WRITE. drain_idx_q/drain_word_q are loaded in
  F_DRAIN, HELD unchanged through F_SETTLE (the load block gates on
  F_DRAIN), and the RAM write + CRC happen at the later F_WRITE edge.
- SDC: fail-closed 2-cycle-setup / 1-cycle-hold multicycle from ONLY
  u_texcache|drain_idx_q[*] to tex_mem (a 6.45 ns EMIF window for the
  drain write-address). Scoped -from the drain regs, so the sampler
  read path (different launch regs) is untouched. HALTs if tex_mem is
  present but drain_idx_q renamed.

The AW buffer (gs_axi_aw_regbuf) and the tile-CDC max-skew 2.5 relax
from 27dfd0b are KEPT unchanged (both closed their families in the fit).

Verified: tb_gs_texture_cache (monolithic + F_SETTLE, distinct-per-byte-
lane + full-word, 0 errors), aw/w regbuf, texture_psmt8_clut,
scanout_diag, ps2_hps_bridge, and the complete f52 replay BYTE-IDENTICAL
(Z 0/307200, COLOR 0/245760). No Quartus/board/push from here.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
thejayman77
2026-07-22 15:11:46 -04:00
parent 27dfd0b0cf
commit 4358bc328b
3 changed files with 55 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rtl/gif_gs/gs_texture_cache.sv
+27 -37
View File
@@ -85,20 +85,22 @@ module gs_texture_cache #(
// 1x 65536x32 (HERE) -> latch each AXI beat, drain 8 lanes over 8 axi_clk cycles.
// Serializing the fill removes the multi-bank/multi-write geometry while preserving the sampler's
// one-cycle registered 32-bit read. The one-shot fill is still tiny compared with board startup.
// Ch443 (Codex): four 65536x8 byte-width banks with INDEPENDENT preserved write-
// address launch registers, replacing the single 65536x32 logical RAM. The single
// drain_idx_q fanning across the whole 128-M20K macro was the -0.370 ns setup family
// (drain_idx_q[4] -> tex_mem ram_block*~reg0). Splitting the WIDTH gives each address
// tree ~32 M20Ks; total M20K count is unchanged (4x32 == 128). Selector structure,
// one-shot fill sequence, and the 1-cycle registered read latency are all preserved;
// the read reconstructs the 32-bit word by concatenation.
(* ramstyle = "M20K" *) logic [7:0] tex_mem_b0 [0:TEX_WORDS-1];
(* ramstyle = "M20K" *) logic [7:0] tex_mem_b1 [0:TEX_WORDS-1];
(* ramstyle = "M20K" *) logic [7:0] tex_mem_b2 [0:TEX_WORDS-1];
(* ramstyle = "M20K" *) logic [7:0] tex_mem_b3 [0:TEX_WORDS-1];
// Ch443 (Codex): MONOLITHIC 65536x32 texture RAM. A four-bank width split was tried
// to relieve the -0.370 ns EMIF drain write-address fanout, but it scattered the banks
// and pushed the DESIGN-clock sampler read cone (ras_v0_x -> perspective-UV -> texel
// addr -> tex_mem portbaddr) to -2.208 ns. The monolithic RAM restores that clean 25MHz
// placement; the drain write-address is instead given a FUNCTIONALLY HONEST 2-cycle
// window (F_SETTLE state + a drain-only multicycle SDC exception), never touching the
// sampler-facing read port.
(* ramstyle = "M20K" *) logic [31:0] tex_mem [0:TEX_WORDS-1];
// ================= fill side (axi_clk) =================
typedef enum logic [2:0] { F_IDLE, F_AR, F_R, F_DRAIN, F_WRITE, F_DONE } fstate_t;
// Ch443: F_SETTLE inserted between F_DRAIN and F_WRITE. drain_idx_q/drain_word_q are
// loaded at the F_DRAIN edge and held unchanged through F_SETTLE; the RAM write-enable
// asserts only at the later F_WRITE edge. The drain address is thus functionally
// required 2 cycles after launch -> a fail-closed 2-cycle-setup/1-cycle-hold multicycle
// (drain_idx_q -> tex_mem only) is honest, giving the EMIF write-address a 6.45 ns window.
typedef enum logic [2:0] { F_IDLE, F_AR, F_R, F_DRAIN, F_SETTLE, F_WRITE, F_DONE } fstate_t;
fstate_t fst;
logic [$clog2(N_BEATS):0] beat; // 0..N_BEATS
logic [255:0] fill_data_q;
@@ -113,10 +115,7 @@ module gs_texture_cache #(
// so reset values are unobservable; keeping them out of the 4k-fanout EMIF calibration reset removes that reset
// from the duplicated RAM-address launch registers at 310 MHz.
logic [31:0] drain_word_q;
// Ch443: four independent, preserved, non-merged write-address launch registers —
// one per byte bank. dont_merge is REQUIRED: they hold identical values, so without
// it Quartus would re-merge them into a single high-fanout register, undoing the fix.
(* preserve, dont_merge *) logic [WIDX_BITS-1:0] drain_idx_q0, drain_idx_q1, drain_idx_q2, drain_idx_q3;
logic [WIDX_BITS-1:0] drain_idx_q;
// fill_start is an EDGE/TOGGLE (bridge toggles it on each arm), CDC-synced here so the
// cache is RE-FILLABLE: each arm reloads the texture (lets the HPS re-stage a different
// texture without a board reset). 3-FF sync + edge-detect, like the read/write probes.
@@ -164,16 +163,19 @@ module gs_texture_cache #(
end
end
F_DRAIN: begin
// LOAD half only (see the separate load block). Advance to F_SETTLE so
// the just-loaded address/word are held one extra cycle before the write.
fst <= F_SETTLE;
end
F_SETTLE: begin
// HOLD: drain_idx_q/drain_word_q are NOT reloaded (load block gates on
// F_DRAIN), so they stay stable across this edge. No RAM write here.
fst <= F_WRITE;
end
F_WRITE: begin
// COMMIT half: registered word -> four byte banks (each with its own
// preserved address reg); CRC over the FULL word actually committed.
tex_mem_b0[drain_idx_q0] <= drain_word_q[7:0];
tex_mem_b1[drain_idx_q1] <= drain_word_q[15:8];
tex_mem_b2[drain_idx_q2] <= drain_word_q[23:16];
tex_mem_b3[drain_idx_q3] <= drain_word_q[31:24];
fill_crc <= fill_crc + drain_word_q; // sum32 over the words written (unchanged)
// COMMIT half: registered word -> M20K; CRC over the word actually committed.
tex_mem[drain_idx_q] <= drain_word_q;
fill_crc <= fill_crc + drain_word_q; // sum32 over the words written
if (fill_lane == 3'd7) begin
fill_beats <= fill_beats + 32'd1;
fill_bytes <= fill_bytes + 32'd32;
@@ -215,10 +217,7 @@ module gs_texture_cache #(
always_ff @(posedge axi_clk) begin
if (fst == F_DRAIN) begin
drain_word_q <= fill_data_q[fill_lane*32 +: 32];
drain_idx_q0 <= fill_word_idx; // four identical loads; kept separate by dont_merge
drain_idx_q1 <= fill_word_idx;
drain_idx_q2 <= fill_word_idx;
drain_idx_q3 <= fill_word_idx;
drain_idx_q <= fill_word_idx; // held through F_SETTLE (block gates on F_DRAIN), written at F_WRITE
end
end
@@ -227,18 +226,9 @@ module gs_texture_cache #(
// present (tex_rd_addr) when tex_rd_en, data lands next cycle.
wire [31:0] word_off = (tex_rd_addr - TEX_VRAM_BASE) >> 2;
wire [WIDX_BITS-1:0] rd_word = word_off[WIDX_BITS-1:0];
// Ch443: read each byte bank, register the four bytes, reconstruct the 32-bit word
// by concatenation. Identical 1-cycle registered latency to the pre-split read.
logic [7:0] tex_rd_b0, tex_rd_b1, tex_rd_b2, tex_rd_b3;
always_ff @(posedge sample_clk) begin
if (tex_rd_en) begin
tex_rd_b0 <= tex_mem_b0[rd_word];
tex_rd_b1 <= tex_mem_b1[rd_word];
tex_rd_b2 <= tex_mem_b2[rd_word];
tex_rd_b3 <= tex_mem_b3[rd_word];
if (tex_rd_en) tex_rd_data <= tex_mem[rd_word];
end
end
assign tex_rd_data = {tex_rd_b3, tex_rd_b2, tex_rd_b1, tex_rd_b0};
// fill_done -> sample_clk (2-FF). The read mux only goes live once warm.
logic [1:0] done_sync;
sim/tb/gif_gs/tb_gs_texture_cache.sv
+6 -5
View File
@@ -131,16 +131,17 @@ module tb_gs_texture_cache;
check(tex_ready, "tex_ready never synced");
// ---- verify every word via the sampler 1-cycle read port ----
// Full reconstructed word AND per-bank byte pinpoint (Ch443 four-bank split).
// Full reconstructed word AND per-byte-lane pinpoint. The distinct-per-byte golden
// exercises all four byte lanes of the monolithic 32-bit word (Ch443).
for (int i=0; i<TEX_WORDS; i++) begin
logic [31:0] exp;
read_word(i, got);
exp = tex_word(i); // temp: iverilog forbids part-select on a function-call result
check(got == exp, $sformatf("word[%0d]=%08x exp %08x", i, got, exp));
check(got[7:0] === exp[7:0], $sformatf("bank0 byte[%0d]=%02x exp %02x", i, got[7:0], exp[7:0]));
check(got[15:8] === exp[15:8], $sformatf("bank1 byte[%0d]=%02x exp %02x", i, got[15:8], exp[15:8]));
check(got[23:16] === exp[23:16], $sformatf("bank2 byte[%0d]=%02x exp %02x", i, got[23:16], exp[23:16]));
check(got[31:24] === exp[31:24], $sformatf("bank3 byte[%0d]=%02x exp %02x", i, got[31:24], exp[31:24]));
check(got[7:0] === exp[7:0], $sformatf("lane0 byte[%0d]=%02x exp %02x", i, got[7:0], exp[7:0]));
check(got[15:8] === exp[15:8], $sformatf("lane1 byte[%0d]=%02x exp %02x", i, got[15:8], exp[15:8]));
check(got[23:16] === exp[23:16], $sformatf("lane2 byte[%0d]=%02x exp %02x", i, got[23:16], exp[23:16]));
check(got[31:24] === exp[31:24], $sformatf("lane3 byte[%0d]=%02x exp %02x", i, got[31:24], exp[31:24]));
end
$display("[texcache] clean: fill_done=%0d beats=%0d bytes=%0d rd_errs=%0d words_checked=%0d errors=%0d",
synth/de25_nano/top_psmct32_raster_demo/de25_nano_psmct32_raster_demo_top.sdc
+22
View File
@@ -491,6 +491,28 @@ if { $tdn == 0 } { error "Ch439g SDC FATAL: tex_mem destination collection empty
set_false_path -from $ch439_bili_taps -to $tex_mem_dst
post_message -type info "Ch439g SDC: impossible bilinear-tap -> tex_mem packed-DSP arc cut ($ch439_bili_taps_n src -> $tdn dst); all non-tap texture-address launches remain timed"
# Ch443 (Codex) — texture-cache EMIF DRAIN write-address multicycle. The fill FSM now holds
# drain_idx_q/drain_word_q stable across F_DRAIN -> F_SETTLE -> F_WRITE and asserts the RAM
# write-enable ONLY at the later F_WRITE edge, so the drain address is functionally required
# 2 EMIF cycles after launch. Give ONLY drain_idx_q -> tex_mem a 2-cycle setup / 1-cycle hold
# multicycle (a 6.45 ns window for the -0.370 ns drain write-address fanout that closing the
# four-bank split had opened). SCOPED -from the drain registers, so it does NOT touch the
# sampler/read-address path (ras_v0_x / walker / UV / perspective -> tex_mem portbaddr) — that
# launches from DIFFERENT registers and stays fully timed (Ch439g above). Reuses the tex_mem
# destination collection $tex_mem_dst / $tdn defined for the Ch357/Ch439g exceptions.
# Fail-closed: if tex_mem exists but drain_idx_q[*] is gone (renamed), HALT — never silently
# omit the exception (which would let the drain path refail setup unconstrained).
set tex_drain_src [get_keepers -nowarn {*u_texcache|drain_idx_q[*]}]
set tex_drain_src_n [get_collection_size $tex_drain_src]
if { $tdn == 0 } {
post_message -type info "Ch443 SDC: texture drain multicycle inactive (no tex_mem in this profile)"
} else {
if { $tex_drain_src_n == 0 } { error "Ch443 SDC FATAL: tex_mem present but u_texcache|drain_idx_q[*] matched 0 keepers (renamed? drain multicycle orphaned -> write-address would silently refail setup)" }
set_multicycle_path -setup -end 2 -from $tex_drain_src -to $tex_mem_dst
set_multicycle_path -hold -end 1 -from $tex_drain_src -to $tex_mem_dst
post_message -type info "Ch443 SDC: texture drain_idx_q -> tex_mem 2-cycle setup / 1-cycle hold multicycle ($tex_drain_src_n src -> $tdn dst)"
}
# Ch358 (Codex) — the SAME quasi-static EMIF calibration-ready signal also reaches the scanout LINE-BUFFER RAMs
# (26.1 fit at 640x480: lock_sync_inst|dreg[1] -> u_lpddr_scan_lb|lb0/lb1, 10 endpoints at -0.010ns). These are
# startup/reset-derived line-buffer control paths, NOT runtime scanout data (the line buffers are inactive until