Tetris® Forever (PPSA25646)

[nekotaru-retro]

Existing thread

• I searched this category for the title ID and no thread exists for this title on any platform.

Title

Tetris® Forever

Content version tested

01.007.000

AnyPS5 commit

Source build, compiled version -> main (102b7a5) , WinLibs GCC 15.2.0 r7.

System tested

Windows 10 22H2 , Intel Core i7-13700KF, 32GB DDR4, NVIDIA GeForce RTX 4070 (617.42)

Furthest point

In game

What is shown

The game works very well in the modes and scenarios I tested (just some 55fps slowdowns on initial load, first 10 secs aprox) . I tried several games in the collection and various options, without encountering gameplay problems or visible graphical glitches.

Working in my tests:

• Playing Tetris Time Warp and several older Tetris titles.
• Pausing and resuming.
• Leaving one Tetris title and starting another without restarting the application.
• Vibration and visual effects.
• Video playback.
• Watched-video progress is remembered, so save functionality appears to be working.
• Music without shuttering
• Tested many languages and no problems shown.

Approximate memory readings:

• VRAM use : around 2 GB
• Tetris Time Warp: around 2,270 MB RAM.
• Older Tetris titles: around 1,000 MB RAM.

I also tested running the game from an HDD insted SSD and did not notice any performance difference once the game starts.

No problems encountered so far, although this was a selection of mini-games and options rather than an exhaustive test of the entire collection.

Current blocker

None encountered during the tested gameplay.

How it was run

C:\Anyps5\relinker.exe --windows --to-intel --windows-diagnostics --registry C:\tetris\eboot.elf C:\tetris\Tetris.exe

AI-assisted: yes, ChatGPT Codex

[JJenkx]

Tetris Forever (PPSA25646). Linux 7.2.11, Ryzen 9 7950X, RX 6650 XT, RADV Mesa 26.2.6, Release.

Setup:

• Relinked with relinker --registry --rpath <build>/core/libs/libs eboot.elf app.elf.
• Run with APS5_PLAYGO_UNLISTED_CHUNKS=installed.
• Input: Cross every 12 s from 30 s.

On main (b415263451 plus my #14298, which only adds a crash report) it reaches the "Chapter 1: The Perfect Game" museum timeline correctly, exit 0. It runs at only about 3 FPS once the timeline is up.

Why it is slow.

• CPU, not GPU: APS5_PROFILE_GPU=1 shows about 50 ms of GPU time per 10 s, the draw cache hits 100%, and there are 3 presents per second.
• Copies: perf shows about 90% of CPU time in glibc's large memcpy, called from ReadColorTarget and WriteColorTarget in Graphics::Draw. Every draw copies its linear colour target in and back out: up to 3840×2160 RGBA, about 32 MiB, each way.
• The targets: a local diagnostic print showed six linear colour targets from 3840×2160 down to 240×135. All lie in one 1.5 GiB direct-memory range that is registered and watched, but none is ever imported. The shutdown counters show color target 0/0/0.
• The refusal: the colour-target import gate in importRunAt refuses them through importStaysWatched. On this host the start-up probe reports that host-pointer imports are "compared", meaning they lose write protection, while dma-buf imports "stay watched". The gate requires a watched host-pointer import, and a shared-backed range is tried as a host-pointer import first.

A local experiment, not a PR. It changes two things for shared-backed ranges when only dma-buf imports stay watched:

• import them through dma-buf directly;
• let the gate accept them.

Same input and run length:

Build	FPS after the first minute	Colour-target imports
main	about 3	0
experiment	59.94 throughout	3 made, 0 refused or failed

The timeline screenshots match, apart from the animated background. That is one run per build; the difference is far beyond run-to-run variation.

#2434 (dma-buf before host pointer when host imports lose write tracking) covers the import-order part of this. The colour-target gate (importStaysWatched) was added to main after it.

Not verified: the classic games inside the collection, audio and other GPUs.

AI-assisted analysis: yes.

[JJenkx]

Follow-up Linux performance validation, PPSA25646 v01.007.000, RX6650XT/RADV Mesa26.2.6, Ryzen7950X, Linux7.2.11.

Pristine main 26fa1212b82f and integration 6384740a031c both reached the same 1920×1080 PLAY TETRIS! / Chapter1 The Perfect Game / explored0% chapter-selection menu. The integration refreshes #2434 onto main and adds a generic watched-DMA-buf ColorTarget creation/refusal fix.

Official 30-second traces at160–190 seconds: main 5.509FPS (166 presents/30130.067ms); integration 59.940FPS (1799 presents/30013.322ms), with no trace drops. Both completed ordinary close with exit0. Reviewed captures at150/195 show the same menu; animated background video phase differs. These are menu results, not a gameplay or audio verification. Separate fresh saves/caches were used, with no input during the window.

Exact source branch, complete official summary lines, traces, screenshots, GPU test logs, negative controls, and compatibility patch. The same4 existing tests pass on main/integration;3 import tests pass on integration,0skips/VUIDs. Full AGC scope, Intel, Windows, and further gameplay have not been validated by this run. The import-selection PR remains unmerged; these changes are not in main yet.

AI-assisted:yes.

[JJenkx]

Linux/RADV current-main performance follow-up for #2434.

Current-main refresh of the earlier Tetris Forever menu result: 5.58 FPS on main 1c8c0efc8c4 → 59.94 FPS with the refreshed integration 0df9bc849a4 (10.74×).

PPSA25646 v01.007.000, 1920×1080, PLAY TETRIS / CHAPTER 1 THE PERFECT GAME / EXPLORED 0% chapter-selection menu. Directly reviewed captures before and after both trace windows show that selection; all flagged alternate readbacks were reviewed and show the same menu with near-time background motion. The background clip phases differ between arms. Gameplay, audio and other-title performance remain unmeasured.

Build	Presents / measured milliseconds	Mean FPS	Frame p50 / p99
main 1c8c0efc8c4	168 / 30113.392	5.579	179.179 / 183.979 ms
integration 0df9bc849a4	1799 / 30013.316	59.940	16.683 / 16.698 ms

Official tools/trace_digest.py summary lines:

Main

**seconds 160:30** from present 2001: 168 presents, frame p50 179.2 / p90 181.3 / p99 184.0 / max 184.3 ms, GPU busy 4.6 ms per present
AMD Radeon RX 6650 XT (RADV NAVI23), radv Mesa 26.2.6-arch1.1, Linux 7.2.11-arch1-1, 1920x1080, commit 1c8c0efc8c41, dropped: no

Integration

**seconds 160:30** from present 9579: 1799 presents, frame p50 16.7 / p90 16.7 / p99 16.7 / max 16.7 ms, GPU busy 10.3 ms per present
AMD Radeon RX 6650 XT (RADV NAVI23), radv Mesa 26.2.6-arch1.1, Linux 7.2.11-arch1-1, 1920x1080, commit 0df9bc849a46, dropped: no

Same machine: Ryzen 9 7950X (32 threads), 127901 MiB RAM, RX 6650 XT / RADV NAVI23, Mesa 26.2.6-arch1.1, Linux 7.2.11-arch1-1, GCC 16.2.1 Release. Both captures use APS5_TRACE_SECONDS=160:30, 2097152 events, fresh private saves and fresh shader/Mesa caches, APS5_HOST_IMPORT_MIB=16384, and a temporary udmabuf size limit of 16384 MiB. The original 64 MiB limit is restored after each complete arm. The paired comparison above uses those explicit import settings. A separate integration-only stock-import follow-up is recorded below. Neither trace drops events. Both native runs close normally (exit 0), leave no private native process, and preserve executable/SDK/input hashes.

Navigation: integration Cross at80s reaches the menu by100s; corrected main Cross at80s plus image-confirmed Cross at100s reaches it by120s. No input follows. The initial main run is excluded because missing observer image-reference fixtures caused late navigation. The retained integration arm logs two optional classifier file errors at100/120s before the trace; its scheduled navigation had already reached the menu, directly checked at100/150/195s. Corrected main has the references and no observer errors. These details and the excluded-run receipt are retained in the bundle.

Official traces, summaries, digests, screenshots/readbacks, commands, configuration, hashes and run receipts — SHA-256 47247674f83c4863abe0556e97a8858faf5791817d19f6f1ff8724e8066a1bc9. Tested source: contrib/pr2434-main1c8c-20261103. The preceding GPU-only report covers 1293 common registrations plus six added physical regressions, with no new validation IDs/counts.

Main menu capture · Integration menu capture

Stock import settings follow-up (integration only): 0df9bc849a4, with APS5_HOST_IMPORT_MIB unset and the normal 64 MiB udmabuf limit, reaches the same Chapter 1 menu at 1920×1080. The official 160:30 trace records 1799 presents / 30013.317 ms = 59.9401 FPS, frame p50 16.683 ms and p99 16.699 ms, with no trace drops. The pre/post captures and flagged alternate were directly reviewed. The run has no observer errors, closes normally (exit 0), preserves SDK/input/executable hashes, leaves no private process and restores the 64 MiB setting. This establishes the integration's menu frame rate without either import override; it is a standalone check, not an additional main-versus-change comparison. Gameplay/audio remain unmeasured.

Stock-import trace, digest, captures/readbacks, commands and receipt — SHA-256 61b0d1669b847f3821a8fcf7c6116189f862b4e8a44dfdac77df542e1b484ea1.

AI-assisted: yes (Codex).

Bundled-game follow-up (2026-11-04): possible rendering regression. After selecting Tetris Time Warp, plain main 1c8c0efc8c4 renders its title/logo correctly, while this private 0df9bc849a4 refresh showed a font/controller texture sheet over that screen. Directly reviewed both primary and X11 readbacks. Main title screen / changed-build corruption. Both used default import settings and 64 MiB udmabuf and exited normally. Navigation timing differed to accommodate loading; this is visual evidence, not a matched gameplay benchmark. Root cause is under investigation. The chapter-menu FPS result above remains scoped to that menu; broader graphical correctness and gameplay are not established. A subsequent run of the unchanged 0df9bc849a4 build rendered the title correctly in direct X11 readbacks at 280/310 seconds, so the corruption is intermittent and its cause remains unproven.

Capture/replay follow-up: the atlas corruption reproduced in both desktop capture methods, but starting the driver frame capture (which waits for GPU idle) was followed by a correct title screen. All three saved driver reference buffers are correct. Main 1c8c0efc8c4 and changed 0df9bc849a4 each replay those buffers pixel-for-pixel with real Vulkan validation, zero VUIDs, no failed registrations, unfinished work or synthesized waits. Thus this capture is a correct replay control, not a reproducer of the corruption. The intermittent live defect remains unlocalized; capture synchronization changing its timing is an inference. Commands, replay logs, reviewed results and reference image, SHA-256 a8adb63b9e0ffb223223aed2a848adfe3a32d392f31e38f5fb6b24eb7acd2804. Raw game memory/call captures remain local. No additional performance or gameplay claim.

Default-settings gameplay follow-up: private refresh 0df9bc849a4 reached Tetris Time Warp → Single Player → Time Warp Score Attack and ran an active level-1 board through Game Over at 1:35. Falling/locked pieces, timer and next-piece queue advanced; score went from 0 to 24 after an up/drop pulse. The preceding left pulse is not independently verified because the helper writes commands by replacement and the two writes were only 34 ms apart. No synchronous-flip switch or frame recorder was enabled; import budget was default and udmabuf 64 MiB.

The official 390:30 gameplay trace records 1799 presents / 30013.320 ms = 59.94005 FPS, p50 16.683 ms, p99 16.704 ms, no trace drops, at 1920×1080 on the same RADV RX 6650 XT host. This is a single-build gameplay observation, not a main-versus-change gameplay speedup measurement.

The graphical limitation remains: at 400 seconds the primary window capture showed a keyboard/texture sheet while its near-time X11 readback showed the correct board; both channels showed a correct board at 430 seconds and Game Over at 455. Consequently this does not establish graphics correctness throughout. Audio, line clearing, other modes and long sessions were not tested. Native ordinary close returned 0; protected input, SDK and executable hashes remained unchanged, kernel settings were restored and no private native process remained. Trace, official digest, full observer/game logs, input record and both capture channels, SHA-256 af26acc27d6ee73d6bc2938dd1a518c2493998c58f206b25bfd8608b65b66d9e.

Native screenshot/input follow-up: private 5bee47efb96 combines 0df9bc849a4 with the source-identical screenshot trigger from #15671 (d0d37670d29). With APS5_NO_PAD_INPUT=1 and the game window unfocused, all six native 1920×1080 PNG requests completed. Five gameplay PNGs are clear. Separated controller pulses now verify movement and drop: the blue piece is centered at the 391-second capture, at the far left at 397, and locked at bottom-left with score 12 at 401. The window capture at 455 seconds still shows an active board at 1:45.

The title-screen native PNG is nearly black with a small white mark, while the nearby window capture shows the glyph/controller sheet. Thus the discrepancy also reaches the built-in screenshot path; rendering, presentation and screenshot readback are not yet distinguished, and no graphical fix is claimed. Native ordinary close returned 0; protected input/SDK/executable hashes and kernel settings were preserved, no private process remained. This is an instrumented private-integration observation, not a new performance comparison. All six native PNGs, nearby window captures, exact input/request timing, integration-source continuity, logs and receipts, SHA-256 f647ce94927af726f8ce92d9dcd33dde4912789d4ed8cbef9c280ddb7cf1847c.

Main-feature native control: exact #15671 head d0d37670d29 (main 1c8c0efc8c4 plus the same host screenshot feature, without this import integration) produced five completed native PNGs: all three sampled Time Warp title images and both sampled gameplay boards are clear. Native exit 0, protected input/SDK/executable hashes unchanged, kernel restored, no remaining private process. This keeps the import integration under investigation, but one clean arm versus an intermittent fault does not establish causality. Navigation timing and sampled game states differ; this is not a matched gameplay performance comparison. All five native PNGs, window captures, input/request timing, logs and cleanup receipts, SHA-256 b8ac6b51b392a983880a46197e3818bce893c43c86be10dbbd98818351cf7c44.