Native ports and measured gains¶
Catalog of optimizations that replace a chunk of translated x86 code with an equivalent native ARM implementation (see the mechanism in winx86's documentation, Extension point page). Each row states the validation level of the cited number (see Three-level validation) — a qemu number is never presented as a console result. Numbers drawn from the project's internal measurement campaigns (dated reports, not published in this repository); what follows is their synthesis.
No absolute value carries across a campaign
The binary changes, the console drifts thermally (1.5-10%). The percentages below come from A/B/A measurements interleaved on an identical binary, never from two different dates compared to each other.
Ports that produced a confirmed gain on console¶
| Port | Measured gain | Level | Detail |
|---|---|---|---|
Texture cache: 2 TMUs announced to the game (D2_GLNUMTMU=2, default) |
~+20% (1 TMU: 20.06 and 19.95 fps; 2 TMUs: 23.98, 23.95, 23.95; cap 25); re-sent textures ~50 → 1 per frame | Console (Act V patrol bench, valid passes only, not interleaved, ~1 h apart) | With 1 TMU its 3 MiB sprite cache overflowed in steady state: ~50 already-known textures re-sent per frame. Needs a large enough atlas: at 32 MiB it filled up in the wilderness (wrong sprites, blinking Death Maulers); fixed by the 48 MiB atlas and failed-upload handling — console confirmation pending. |
Online frame cap off by default (D2_ONLINE_CAP, 2026-09-28) |
24.0 → 25.0 fps; frames over 60 ms: 25 → 3 per 110 s of patrol | Console (Act V patrol bench, route A, 1 pass before, 3 after) | The engine's 25 Hz cap armed whenever D2NET was set, i.e. in solo too. It rounded its wait up to the next ms (~41.6 ms/frame) and drifted against D2's own clock, which then skipped draws. In solo, D2 already limits itself to one draw per 40 ms step. |
Intrinsics served in line from translated code (D2_INTRINLINE, default since 2026-09-28) |
−0.9 to −1.25 ms of guest work per frame (route A 26.8/26.6 → 25.7/25.9 ms; route B 27.1/27.0 → 25.8/25.8 ms); c0 76–77% → 74–75% | Console (Act V patrol bench, 8 interleaved passes, same-route comparisons); qemu: identical FBHASH pixel oracle over 4000 frames (2 controls + 1 leg), D2_CSTEST contention bench PASS |
~1000 trap round trips per frame cost ~4 µs each (JIT exit, lock, re-entry, block lookup: ~10% of frame time, block profile). The slot's stub becomes a box86 native call; direct return to the caller through the CALLRET stack. GetTickCount and critical sections (~650 calls/frame). |
Vertex copy in line in the ring DLL (default since 2026-09-28; D2_GLRUNS=0 for the old path) |
−0.9 ms of guest work per frame (route A: 27.6/27.6/27.2 → 26.3/26.8 ms) | Console (5 passes); qemu in game, 4000 Perspective frames: 1.07 M draws re-read byte for byte, 0 difference, sabotage detected | Each vertex copy (28 bytes, zigzag strips) was a call + rep movs: 4.4% of the time in the copy function. |
File read-ahead 32 KiB on a seek (D2_READAHEAD_JUMP, default 32 since 2026-09-28) |
Card reads in the window 65–71 → 25–27; card time 208–223 → 130–197 ms | Console (4 passes, same route) | No extra RAM (window still 32 KiB). Effect on stutter too small to measure (2–4 loading frames per pass). |
Native F3 optimized and served in line (D2_F3NATIF=1 + D2_INTRINLINE, 2026-09-28) |
A further −1.25 ms (route B 25.6 → 24.4/24.3 ms; c0 74 → 72%) | Console (4 interleaved passes); qemu: D2_F3NATIF=2 oracle 168,503 cells, 0 divergence; D2_F3CHECK 436,538 strips, 0 difference |
Strips written directly (no per-vertex hostptr), culled cells' corners deferred, visibility in integers where exact. Still off by default: the player's call. |
Native F3: Perspective floor tiles, Game+0x10cf70 (D2_F3NATIF=1) |
~−2 ms of guest work per frame (translated 28.0 → ~20.6 ms, native shims ~5.3 ms; c0 −2 points; fps unchanged, already at the 25 cap in town) | Console (Act V patrol bench, 2 valid passes per flavour); qemu oracle D2_F3NATIF=2: 0 divergence over 82,000+ drawn cells |
Ports the caller (~30 small calls per cell absorbed). The remaining cost is memory (projection and LUT tables, ring shared with the flush thread). Off by default pending a decision. |
Native cell loop (NATIVECELLLOOP) |
+16.1% | Console (A/B/A/B, identical binary) | Traversals ÷2, blit ÷7. |
D2_CELLOPT (optimization mask inside the native loop) |
+9.3% | Console | Half the work per cell. |
D2_CELLPAR=1 (fork-join parallelism, 1 worker thread) |
+12% | Console | D2_CELLPAR=2 = 0% (USER_2 shared, no second core available). |
D2_CALLRET |
+6.8% console / 0% qemu | Console (4 interleaved passes, single binary) | Textbook case: a real gain qemu doesn't see at all. |
fastmmu + D2_BUDGETTAIL |
+2.8% | Console | Taking the base ADD out of the dynarec's critical path. |
D2_CSINTRIN + WX86_B5 (clock intrinsic) |
+1.4% | Console | Clock intrinsic rewritten to leave the critical path. |
fog_raise_snap probe disarmed |
+4 to 7% | Console | The probe itself cost what it measured. |
GPU ring — asynchronous submission (D2_GXMASYNC=2) |
+52% (28.8 → 43.7 fps bench) | Console | The GPU takes 11ms/submission; async mode overlaps that time. |
GPU ring — flat clear (D2_GXMCLEAR=plat) |
+10.5% (−2.5ms) | Console | Simplified framebuffer clear. |
| CPU-side ring — host view + fast vertices | +11.5% (43.75 → 48.78 fps bench) | Console | Ring traversal ÷2.3. |
Hoisting TLS resolution out of CpuBox86::run() |
+3.03% at the menu bench, +8.6% in real online play | Console (interleaved A/B/A/B + guard rail: guest work identical within 0.1%) | The gain tracks GIL-take density (190/frame at the menu vs. 2880/frame in play) — a percentage measured at the menu must never be published as an in-game gain. |
Trap entry/exit batching (trap_retaddr+trap_epilogue, commit e13edbd) |
+13.2% as announced in the commit | Commit-time figure, cited as-is in a later fix — no dedicated campaign document found for this exact number, take it with that caveat. | Already in place today (not an optional knob). |
NATIVELIGHTMAP (dynamic light placement) |
−21% on the light pass (1.20 → 0.95ms/frame), i.e. −2.6% of total drawing | qemu (cross-checked oracle D2_LIGHTMAPVERIFY, 51,200 placements, 0 fallback, 0 divergence) |
This isolated qemu figure predates console validation. Since 2026-09-22 it's shipped as part of the compiled-default group below — see that row for the console number; no isolated console figure exists for this lever alone. |
D2_MEMINTRIN=3 (short copies emitted inline by the dynarec, no helper call) |
+4.6% (21.93 vs 20.97 fps) | Console (passe_barb.sh, real player save) |
Different mechanism from the large-transfer/helper path below (refuted): largest single-lever gain of the whole campaign. Arming proven on hardware (helper calls 24,250,123 → 4,248). Oracle oracle_memintrin.sh PASS, 4,816,757 calls cross-checked, 0 divergence. Compiled default since 2026-09-22. |
D2_FLUSHFIL (ring flush moved to a dedicated core) + D2_TEXHASH + NATIVEDCC + NATIVELIGHTMAP — 4 settings that had sat in env_jeu.txt since 06–13/09 but were missed when commit fff1904 froze the other 22 as compiled defaults |
+11.7% at the real 25Hz cap (20.97 → 23.43 fps) | Console (passe_barb.sh, 4 witness passes, 0.3% spread over 8h) |
The game thread no longer touches the flush at all (gxm-flush: total-us=3 vs 9,600). Moving ~12–13ms of memory-bound work off-core only recovers ~4.7ms of frame time (L2/bandwidth shared across cores) — don't size further work on the un-offloaded figure. |
All five settings above + NATIVELIGHTOCC (light-occlusion field that feeds NATIVELIGHTMAP, Game+0x750f0) — all six compiled defaults together |
+13.0% at the real 25Hz cap (20.97 → 23.69 fps) | Console (same campaign) | 50–100ms frame bucket: 1,474 → 51 (÷29). 100–250ms bucket (the real 16-17fps drops): 41 → 28 (−32%) — moves only in this full combination; neither D2_MEMINTRIN=3 nor the flush group alone touches it, because NATIVELIGHTOCC produces the field NATIVELIGHTMAP consumes. |
| Cumulative, night of 04-05/09, uncapped bench | Control 22.5-23.1 → 28.1-28.6 fps (+22%) | Console, interleaved campaign of 20 passes | Cumulative effect of several of the ports above. |
| Cumulative GDI → asynchronous Glide | 22.5 → 55.7 fps (+96%) bench; 25 steps/s + 60 fps smoothed in real play | Console | Cumulative effect of moving to the asynchronous GPU rendering path (ring + async submission). |
Native F3 by default (D2_F3NATIF, Perspective floor tiles, 10/06) |
22.27 → 23.68 fps (+6.3 %) | Console (Frozen Orb bench, 2 baseline passes, 4 F3 passes, interleaved) | The port existed since 09/28 but was opt-in. |
Missile light branches (NATIVELIGHTDISC, NATIVELIGHTDYN, Game+0x4748d0 / 0x474d70, 10/06) |
23.76 → 24.66 fps (+3.8 %), run 36.3 → 33.5 ms |
Console (Frozen Orb bench, 2+2 interleaved passes); console cross-oracles 133,000 + 14,000 calls, 0 divergence; qemu pixel-identical | The per-light pass (0x4755a0) has three branches; only the kind-2 one was ported. Orbs and their shards take the other two. |
Per-call room cache in the light-grid rebuild (D2_GRIDCACHE, 10/06) |
Time share 3.74 % → 1.68 % (~−0.9 ms/frame) | Console (BLKSAMP profile, Frozen Orb bench); qemu: same room decisions cell by cell, pixel-identical, D2_GRIDVERIFY 0 divergence |
Every cell re-read guest memory through a checked access; the ~5 % of cells that change room redid the whole search. |
F3's texture bind served natively (D2_F3TEXNAT, Game+0x50fbd0 resident branch, 10/06) |
Free time per frame 0.3-0.7 → 3.2-3.3 ms, c0 96 → 90 % (at the 25 fps cap) | Console (Frozen Orb bench, 2+2 interleaved passes); qemu: identical ring fingerprint (3.27 M records, states included), sabotage detected | ~400 round trips per frame into game code. The host writes grTexSource/grTexCombine through the DLL's own dedup cache, published in the ring header (new glide3x.dll required). |
Ports refuted or with no measurable effect — kept as a warning¶
| Port | Result | Level | Why it's instructive |
|---|---|---|---|
D2_MEMINTRIN=1 (native memcpy/memset served via a helper call, large transfers) |
Mechanism proven correct (4.4M calls, 0 divergence) but 0% on console | Console | The guest CRT already vectorizes its large transfers (NEON); the Glide path makes 90 calls/frame vs. ~1,100 under GDI — the lever had disappeared with the renderer change. Still true for this specific path; D2_MEMINTRIN=3 (short copies, inline, no helper) is a different mechanism — see the confirmed-gains table, now the compiled default. |
D2_FORWARD |
Inert (+0.6% = noise) | Console | "Shorter blocks" hypothesis never actually tested. |
D2_NOPEND (lever L4) |
Refuted, ceiling 0.002% | Console | The original document targeted the wrong code site (0 real calls instead of the assumed site). |
| Offloading RLE to a dedicated thread | FAIL on three modes; the mode that passed three days earlier no longer passes | qemu (oracle) | "Prove the oracle before concluding" — the founding PASS had never been replayed since. |
| Four image-cache hypotheses | All refuted (0 identical frame out of 4,000) | qemu (D2_CACHEPROBE) |
Found something else instead: the Perspective video option missing from the registry cost +14.2% guest blocks. |
| Hardware PMU (Cortex-A9 cycle counters) | Abandoned | Console | The Vita kernel resets the access register on every context switch; would require patching the system scheduler. |
5 Act V intrinsics (D2_INTRIN: projection, color LUT ×2, light grid, rect fill) |
Negative: ~+4.3 ms of guest run per frame, ~−3.2% fps (5 patrol passes, oracles at 0 divergence) | Console | Tiny integer functions (20,000–30,000+ calls per frame) that the dynarec already translates almost 1:1: the native body is no faster, and every call pays the plumbing. Off by default. |
Native draw serialization (D2_GLNATDRAW=1, trap-window intrinsic) |
Negative: ~+5.5 ms per frame, −3.8% fps (4 interleaved passes, v1 bench without route check, window 2600-5600) | Console | ~1,700 exits from translated code per frame cost more than the copy they avoid. Same lesson: an intrinsic only pays when the body it saves is large per call. |
| Fifteen dynarec levers, campaign summary | A single gain in the entire history of the dynarec (D2_CALLRET, +6.8%) |
Console | "The dynarec isn't the lever" — the gain is in native porting of guest code, not in translation itself. |
| Memoizing F3 cell corners within a call (10/06) | Refuted: 0 repeats over 2.16 M cells, even with the reduced key (sub, dy) |
qemu (deterministic Glide, 5,000 frames) | Cells of one call never share corners; measure the hit rate before writing the cache. |
Method for picking a port that pays off¶
What the campaign learned formalizes as:
gain ≈ N × (T_guest − T_native − 69ns) where 69ns is the measured cost
of the x86→native crossing itself (the price of a round trip, even for an
intrinsic recognized at translation time rather than a real trap). The
right target is not the hottest function in the profile — it's a
caller with a large number of calls (N) that absorbs an entire
subtree in a single traversal. That's exactly the cell loop's pattern: a
single call site produces the entirety of the ~1,600 cell-blit calls per
frame.
Chosen game configuration¶
tools/bancs/env_jeu_glide.txt (deployed on console, chosen by the
maintainer) — this is the combination of every confirmed lever above:
D2SCHED=native
D2WRITE=ux0:data/d2vita/save2
NATIVECELLLOOP=1
D2_CALLRET=1
D2_LAZYSEEK=1
WX86_FRAMEPROF=1
D2_FRAMEUS=1
D2_CELLOPT=63
D2_MMUSTACK=3
D2_MMUFOLD=1
D2_BUDGETTAIL=1
D2_CSINTRIN=1
WX86_B5=1
D2ARGS=game.exe -3dfx
D2_GLIDERING=1
D2_GLIDEGXM=1
D2_GXMASYNC=1
D2_GXMCLEAR=plat
D2_GXMRING=4
WX86_YIELD=2
D2_ONEDRAW=1
D2_IOSTAT=1
D2_READAHEAD=128
D2_READAHEAD_MIN=32
D2_READAHEAD_JUMP=8
D2_READAHEAD_WIN=8
D2_READAHEAD_WINKB=32
D2_SON=1
(kept in sync with tools/bancs/env_jeu_glide.txt, the source of truth —
this block is a copy, not the other way around. D2_NOCAP is deliberately
not in it — it's a bench tool that disengages the game's internal frame
limiters, not a setting to keep for real play.)
D2WRITE names the directory the game writes into — saves, crash.log,
D2's own daily log. Up to 0.1.6 a directory named here was not created
on the console: every write then failed, and the game's C runtime killed
the process about 11 s into the session, with no message pointing at the
directory. Since 0.1.7 the boot creates the named root (and its Save
subfolder), probes that it is writable, and falls back to
ux0:data/d2vita/save with a line in boot_progress.txt when it is not.