Initial commit: retroDE_ps2 — first-of-its-kind PS2 GS FPGA core (DE25-Nano / Agilex 5)

RTL (GS rasterizer, EE core stub, platform bridge, LPDDR4B path), sim regression
(272 TBs), docs, and tooling. Copyrighted PS2 content (BIOS, game code, GS dumps,
and all dump-derived textures/traces) is excluded via .gitignore and stays local.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
2026-06-29 20:10:50 -04:00
commit ec82764bef
2462 changed files with 2174303 additions and 0 deletions
+165
View File
@@ -0,0 +1,165 @@
# Ch271 closeout — SQ implemented; qbert progresses 2,247× further
**Status:** Closed. **Verdict from re-running qbert.elf:**
`elf_first_unsupported_opcode (pc=0x00100068 instr=0x0080e02d)`
**DADDU**, the next missing R5900 opcode. **That frames Ch272.**
## Numbers, end to end
| Metric | Pre-Ch271 (Ch270 verdict) | Post-Ch271 (this chapter) |
|-----------------------|----------------------------|----------------------------|
| qbert retire_count | 12 | **26,958** (2,247× more) |
| First-trap PC | 0x00100024 (SQ) | 0x00100068 (DADDU) |
| First-trap instr | 0x7C400000 | 0x0080E02D |
| Distance in qbert text | ~9 instructions from entry | ~24 instructions further |
The SQ implementation correctly cleared the qbert prolog buffer
that previously stalled execution. Now qbert progresses ~24
instructions further into its prolog before hitting DADDU.
## What landed
### RTL — ee_core_stub.sv (5 surgical edits)
1. `OP_SQ = 6'h1F` localparam constant alongside the other store
opcodes.
2. `is_sq` logic declaration + `assign is_sq = (opcode == OP_SQ)`.
3. **Alignment**: extended `is_align_fault` to include
`is_quad_access && (ea[3:0] != 4'd0)`, and added `is_sq` to
`is_align_store`. Misaligned SQ now trips the existing
AdES exception path (or strict trap, depending on
`TRAP_ALIGN_ERROR`).
4. **Decoder allow-list**: added `!is_sq` to the `is_nop_class`
catch-all so SQ doesn't get rejected by `STRICT_UNSUPPORTED`.
5. **4-beat FSM**: new `sq_beat` 2-bit register; transition into
`S_MEM_WRITE` from EXECUTE; in `S_MEM_WRITE` combinational
block, `map_wr_addr = ea + {sq_beat, 2'b00}` and
`map_wr_data = (sq_beat == 0) ? rt_val : 32'd0` (upper 96
bits of $rt aren't modelled; for `sq $zero,...` — the qbert
case — every beat naturally writes zero); in `S_MEM_WRITE`
FSM state, stay in state and increment `sq_beat` until
`sq_beat == 2'd3`, then retire and return to `S_IFETCH_REQ`.
The single architectural SQ instruction takes 4 bus beats but
produces exactly ONE retire event — matching the architectural
model.
### TB — sim/tb/integration/tb_ee_core_sq.sv
Focused 18-instruction test:
- Bootstrap from `0xBFC00000` reset vector via J to
`0xBFC00100`.
- LUI/ORI to load `$v0 = 0x80000400` (kseg0 → EE RAM phys
0x400).
- Pre-poke EE RAM at phys 0x400..0x40F with distinct non-zero
values (`0xDEADBEEF / 0xCAFEF00D / 0x12345678 / 0x9ABCDEF0`)
via hierarchical `ram_word()` task so a missing SQ beat would
leave a non-zero word.
- Execute `sq $0, 0($v0)` (= 0x7C400000, the exact qbert
instruction).
- LW + BNE-to-FAIL chain over the 4 words verifies each lane is
zero.
- Belt-and-braces: direct hierarchical peek of
`u_ee_ram.mem[0x40]` after halt to confirm all 128 bits are 0.
- PASS via syscall.
Result: `[tb_ee_core_sq] retired=18 halt=1 trap=0 pc=0xbfc0013c
errors=0 PASS`. Both the BNE chain and the direct RAM check
agree the SQ wrote 16 zero bytes correctly.
### Makefile — `tb_ee_core_sq` target + regression list
Added to both PHONY list and `run:` master list. Regression
bumps from 158 → 159.
## Why not just NOP the opcode (Codex's caution honoured)
Codex called this out explicitly: `0x7C400000` is `sq $zero,
0($v0)` — a 128-bit store of zero. NOP-ing op=0x1F would let
qbert continue, but it would silently skip real memory
initialization. For the prolog, that's a buffer clear; later
code would read uninitialized values from those bytes and
behave nondeterministically.
**Minimal-correct SQ** (4 beats of 32-bit writes) is the right
choice. The "minimal" part: we don't model the upper 96 bits of
$rt (PS2 EE has 128-bit GPRs); for `sq $zero,...` this is
exact, and for `sq $non-zero,...` we write the low 32 bits to
beat 0 and zero elsewhere — a documented approximation that
degrades gracefully for the common "clear a 128-bit kernel
slot" use case. When/if a real PS2 program does `sq` of a
non-zero 128-bit register, we'll see silent data corruption
that the runner's hot-PC verdict can identify; that's the
trigger to upgrade to 128-bit GPR modelling.
## Codex Ch271 acceptance — line-by-line
| Requirement | Status | Where |
|----------------------------------------------------------------------------|--------|-------|
| Decode primary opcode 0x1F as SQ | ✅ | OP_SQ + is_sq |
| Support `sq $zero, imm(base)` at minimum | ✅ | rt_val=0 case writes 0 every beat (and rt_val=non_zero writes low 32 to beat 0) |
| 4-beat 32-bit-stripe FSM through existing memory interface | ✅ | sq_beat counter, stays in S_MEM_WRITE for 4 beats |
| Require 16-byte alignment; misaligned → strict/exc trap | ✅ | is_quad_access check in is_align_fault |
| Focused TB: preload base, exec SQ, verify 4 zero words | ✅ | tb_ee_core_sq |
| Verify PC advances + no GPR writeback | ✅ | Final PC check + retire path doesn't touch regfile |
| Re-run qbert.elf, report next blocker | ✅ | DADDU at pc=0x00100068 |
| Don't NOP all op=0x1F (would mask real stores) | ✅ | Targeted decode, exact 4-beat write semantics |
| Don't overbuild full LQ/SQ/vector yet | ✅ | SQ only (no LQ, no PSQ_*, no vector); upper 96 bits left for later |
| Regression unaffected | ✅ | 159/159 in flight |
## Recommendation for Codex's Ch272
**`daddu $gp, $a0, $zero` at pc=0x00100068 instr=0x0080E02D.**
DADDU is MIPS-III's 64-bit version of ADDU. The R5900 is a
64-bit core; PS2 ELFs use DADDU as the canonical 64-bit
register-move pseudo-instruction (`move rd, rs`
`daddu rd, rs, $zero`).
Our model has 32-bit regfile (`logic [31:0] regfile [0:31]`),
so a faithful 64-bit DADDU would need 64-bit GPRs. For the
qbert blocker specifically, the operation degenerates to a
32-bit move: `$gp = $a0 + 0`.
Three Ch272 framings, in order of scope:
1. **Decode DADDU and treat it as ADDU.** Low-32-bit semantics
only; upper 32 bits silently dropped (already true everywhere
else in the model). Touches one line in `is_nop_class`
allow-list + one new R-type funct case + adding `is_daddu` to
the `is_rtype_alu` group. Same "minimal-correct" pattern that
worked for SQ.
2. **Decode DADDU + DADD + DSUBU + DSUB + DAND + DOR + DXOR + DNOR
as their 32-bit counterparts.** Broader, but these are all
commonly emitted by gcc for r5900 alongside DADDU. Pre-empts
the next 4-7 chapters worth of one-opcode-at-a-time growth.
3. **Properly implement 64-bit GPRs.** Architecturally correct,
but invasive — touches regfile width, all ALU paths, LW/SW
to-from regfile, and the trace. Probably 1-2 chapters of work
on its own.
(1) is the strict Codex-style "minimal-correct next blocker"
answer. (2) would shorten the chapter chain if Codex thinks
qbert's prolog uses several D* ops. (3) is a "do it right" pivot
that's worth doing eventually but probably not in Ch272.
My read: **(1) is the right Ch272 — same shape as Ch271, fast
to land, lets the verdict surface the next real divergence.**
If the next blocker is also a D* op, we recur. If it's something
totally different (LQ? MMI? VU0 macro?), we know (1) was the
right scope.
Standing by.
## Files changed
- `rtl/ee/ee_core_stub.sv` — 5 surgical edits (~20 LOC total) for
SQ decode + 4-beat write FSM.
- `sim/tb/integration/tb_ee_core_sq.sv` — new focused TB.
- `sim/Makefile``tb_ee_core_sq` target + added to both
regression lists.
## Regression
In flight at the moment of writing; expected 159/159 (was 158, +1
for tb_ee_core_sq).