Repository navigation
beta.1: tested profiles, headset settings menu, standby fix, frame budget and image quality (.66-.134) - #23
Merged
Merged
Conversation
Constant bitrate was spread over the refresh rate. A game presenting 150-180 fps on a 207 Hz panel then got bitrate/207 per frame and used only 77-87% of the bitrate. adapt_to_framerate is now on by default and follows an exponential average of the present interval: rising rates apply at once, falling rates at the 1 s update, and a frame is at most twice the refresh-rate frame size.
When no head sample reaches the PC between two vsyncs, the server extrapolates the newest head pose by its velocity to a marked timestamp and renders from that instead of re-rendering the previous pose. The client keeps the motion it sent and resolves the marked timestamp with the same extrapolation for reprojection. Head velocity is now sent (was zeroed). Setting: headset.extrapolate_late_head_poses (default on). ALVR stream socket and PyroWave transport default to UDP (Wi-Fi only; USB stays TCP). Wi-Fi 6E, 120 Hz / 1250 Mbps: UDP 2.9% vs TCP 4.2% repeated poses. Scene: Q3PW_SCENE_OVERDRAW for a GPU-bound test scene.
125 % (2592x2784 after alignment) holds 115.5 fresh FPS at 120 Hz wired with a 7.7 ms decode fence; 3072x3216 reaches 86. Offline at 1500 Mbps, 2560x2720 keeps 1.3 dB more PSNR-HVS-M against a 3072x3216 render than 2080x2208.
ABBA, wired CDF 5/3: network p99 7.8-7.9 -> 5.1-5.2 ms at 207 Hz/1000 Mbps and 16.7-18.2 -> 14.8-14.9 ms at 120 Hz/2000 Mbps (stream 125 %), same fresh FPS.
JMS1717
changed the title
…text tools/downsample/entropy_model.py costs the same quantized CDF 5/3 coefficients as PyroWave's raw bit-planes, a neighbour/parent-context coder and a static per-group code, and reports PSNR at the live byte caps. docs/ENTROPY.md records the results.
debug.q3pw.sharpen=1..100 applies FidelityFX CAS (cross neighbourhood, luma only) in the packed-YCbCr eye shader. Off by default; not applied with light foveated encoding.
… (.72) The headset's eye shader applies CAS at the configured sharpness; debug.q3pw.sharpen still overrides it for tests (0 forces off).
The rest is drawn with the plain program, as foveation draws its bands. debug.q3pw.sharpen_area (10-100) sets the share for tests.
Centre-only CAS at 50: free at 120 Hz/100 %, about -5 FPS at 207 Hz and -8 at 120 Hz/125 %; whole-image cost 19 FPS in both.
…ass (.74) Luma, Cb and Cr are written as 2x2 quads side by side in one RGBA8 AHB 3/2 the frame width; ALVR's eye shader reads full-resolution chroma at the luma pixel. 4:4:4 with CDF 5/3 asks for mode 7 by default and falls back to mode 3.
…rpening recovers half
Offline at 120 Hz/1500 (2080) linear 0.15/neighbour keeps +1.69 dB PSNR-HVS-M over no sharpening against +1.23 for CAS 50, with fewer overshooting pixels (8.3 % vs 9.8 %). debug.q3pw.sharpen_kernel=cas keeps CAS for A/B.
…decode 2.65 ms tools/entropy: coefficient dump, QR coder (CPU encode/decode), Vulkan decode shader and a bench that checks every coefficient. On the Quest the decode reaches 2.65 ms with lane-interleaved stores, longest blocks first and branch-uniform steps; it is close to throughput-bound. Results, rejected variants and what building it would take are in docs/ENTROPY.md; parked until the owner's 120 Hz tests.
… (.121) HorizonOS clears debug.oculus.gpuLevel whenever the client restarts. The display helper rewrote it with a wake and a client restart and, after two clears, stopped forcing it until the server restarted, so the GPU fell to level 4 (640 MHz) in long sessions. The runtime applies the level live (cleared: level 4 within 3 s; 7: 690 MHz within 6 s), so a GPU-only change is now written directly. Reverts are counted only when a display change restarts the client.
Quality 120 4:4:4 (CDF 5/3, 2270x2429 per eye, 1500 Mbps, wired) on .121, ABAB: forced LOW 118.6/119.6/119.9 fresh against 115.7/113.6/116.2 at default priority; frames superseded after decode fall from 4-6.5/s to 0-0.4/s. Latency excluding game time is a tie (about 38 ms): decode runs longer under LOW, the vsync queue shortens by about the same amount.
.122 gave 4:4:4 the 4:2:0 limit (2.0 Gpx/s), which put full-size 4:4:4 at 90 Hz (1.79 G) on LOW. There CDF 5/3 4:4:4 decode takes 8.2-8.5 ms of the 11.1 ms period and LOW gave 75.5/75.6 fresh against 83.6/83.4 at default (wired, 1500 Mbps, level 7). Quality 120 4:4:4 (1.33 G) keeps LOW.
4:4:4 with Haar fell back to the old path (R8 planes plus pyroclient's conversion pass). The multilevel Haar pass is per component, so 4:4:4 chroma takes the luma schedule and its final pass stores quads at x = width/2 and width, the layout Decoder V2 mode 7 and ALVR's eye shader already use. The decoder accepts haar32 modes 1-3 and 7 for 4:4:4; pyroclient asks for 7 when 5-6 are requested and steps down to 3 without storage on the AHB. decoder_ab gains AB_CHROMA=444 and haar32m7/cdf53v2m7 arms for the exactness gate.
Writes the next N (at most 16) complete PyroWave codec frames to the app's external files dir, once per new property value, for offline bitstream and entropy-coding analysis. Off by default.
tools/entropy/pw2qcf.py parses a complete PyroWave frame (wavelet_dequant.comp's layout) into QCF1 for qrenc and reports the quant side information's entropy. qrcodec.hpp: a context with one symbol has frequency 4096, which overflowed the 32-bit renormalization bound to 0 and looped forever; use 64 bits.
A stream taller than the panel's 2208 rows loses much of its finest band to the display resample. The CSF's level-0 frequency now scales by rows / 2208: 23.6 cycles/degree at full size instead of 16.2. Offline at panel resolution, full size, 120 Hz: +0.6/+0.7 dB PSNR-HVS-M at 1500/1000 Mbps; neutral at 2784 rows. csf_study.py gains STUDY_FULL / STUDY_STREAM display-referred scoring.
…anel's rows (.130)
…ile streaming (.131) The in-headset settings menu gains a Profile row that applies a whole tested profile (Quality 120, Godlike 120/90, Colour 4:4:4, Wi-Fi Quality 120, Wi-Fi 90, Competitive 207, Starter 72) and shows two lines of help for the selected row. The streamer rejects USB-only profiles on Wi-Fi and profiles that need a refresh rate the headset did not offer. The dashboard list drops the superseded bitrate and latency experiments. The driver blocks SteamVR's motion-idle standby while a client streams: a still head or a headset on a bench no longer drops the stream to 10 Hz.
… streaming (.132) At 120 Hz full size, CDF 5/3 held 112 fresh FPS against Haar's 116-117 and added 5 ms to ALVR's latency estimate with frames queueing for the decoder, so Godlike 120 keeps Haar. Godlike 90 has 11 ms a frame and uses CDF 5/3 for the cleanest full-size image. One of four .131 cells still entered SteamVR standby 12 s into a stream. The driver now logs standby queries, entries and stream start/stop ([Q3PW_STANDBY]), and when SteamVR enters standby while a client streams it re-asserts presence with a proximity change, as a stream start does. The Apply help in the headset menu says a profile can restart SteamVR.
…standby while streaming (.134)
…iles and the headset menu
JMS1717
changed the title
JMS1717
marked this pull request as ready for review
October 9, 2026 15:52
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Everything on
claude/frame-budgetsince alpha.9 (.63), up to.134. Merging this makes main the beta.1 release.For players
.134): the driver counts overlapping connections and keeps SteamVR from idling a streaming headset.Measured on
.134(one Quest 3, USB, short screens, heavy SteamVR Home)No standby while streaming in five cells. Not checked: sustained play, the Wi-Fi profiles live, optical latency.
Release notes: docs/RELEASE-beta.1.md.