MSPR 2x
GitHub
Open source · MIT · Swift + Metal

MSPR / MULTISAMPLE PHASE RECONSTRUCTION

Recover the samples
resolve throws away.

A 2x spatial upscaler for Apple GPUs that reconstructs the four color samples already produced by 4x MSAA—before a conventional resolve averages them into one pixel.

Explore the source Read the technical report
  • 01Deterministicsingle-frame output
  • 02Tile-localbuilt for Apple GPUs
  • 03Model-freezero weights or history
2× UPSCALING / SAME FIELD OF VIEW
LOW-RES RASTER1×
4 × 4 source areafour samples per LR pixel
MSPR OUTPUT2×
8 × 8 output areafour reconstructed phases

2× width + 2× height = 4× output pixels, over the same view.

CONVENTIONAL RESOLVE4 samples → 1 averagesubpixel phase is discarded
MSPR PRE-RESOLVE4 samples → 4 output phasesphase evidence becomes 2× detail
ARCHIVED IPHONE 13 EVIDENCE

Quality and full-path GPU time use the same nine-scene iPhone 13 matrix.

30.301dBMSPR-Quality mean
+1.021dBMSPR-Quality vs Bilinear
+0.629msMSPR-Lite full-path time vs Bilinear*

Quality × GPU cost

Higher quality.
Small frame cost.

In one nine-scene iPhone 13 matrix, MSPR-Quality gains 1.021 dB over Bilinear for 1.067 ms of added full-path GPU time; MSPR-Lite gains 0.785 dB for 0.629 ms.

THE RESULT MSPR improves mean quality over MetalFX Spatial, Bilinear 2×, and FSR1 while adding much less full-path time over Bilinear than MetalFX or FSR1.
Archived iPhone 13 nine-scene quality and full-path GPU comparison
Method Mean RGB PSNR ↑ 9-scene full-path GPU mean ↓ Added full-path time vs Bilinear ↓
MSPR-QualityBest mean quality Mean RGB PSNR ↑30.301 dB 9-scene full-path GPU mean ↓3.656 ms Added full-path time vs Bilinear ↓+1.067 ms
MSPR-LiteSecond in quality · small added time Mean RGB PSNR ↑30.065 dB 9-scene full-path GPU mean ↓3.218 ms Added full-path time vs Bilinear ↓+0.629 ms
MetalFX SpatialApple spatial upscaler Mean RGB PSNR ↑29.410 dB 9-scene full-path GPU mean ↓7.606 ms Added full-path time vs Bilinear ↓+5.017 ms
Bilinear 2×Simple spatial baseline Mean RGB PSNR ↑29.280 dB 9-scene full-path GPU mean ↓2.589 ms Added full-path time vs Bilinear ↓0 ms · reference
FSR1 EASU + RCASIn-project FSR1 pipeline Mean RGB PSNR ↑27.367 dB 9-scene full-path GPU mean ↓8.575 ms Added full-path time vs Bilinear ↓+5.986 ms

Scope: PSNR and performance use the same archived nine-scene iPhone 13 matrix. GPU values are arithmetic scene means of full command-buffer duration after 20 warm-up and 120 measured frames per mode. Added time subtracts Bilinear's 2.589 ms mean; it is a full-path comparison, not isolated kernel timing. Two Apple-demo-dependent rows are historical and not reproducible from this package.

Two operating points

One algorithm.
Two practical profiles.

Both paths consume pre-resolve MSAA evidence. Choose the neighborhood-aware Quality path or the fixed six-tap Lite path.

01 / QUALITY Best reconstruction

MSPR-Quality

A 3×3 resolved neighborhood, fixed phase coefficients, bounded detail, and content-adaptive blending for the strongest reconstruction.

Archived mean
30.301 dB
100-frame set
32.934 dB
Added full-path time vs Bilinear
+1.067 ms*
02 / LITE Minimum overhead

MSPR-Lite

A fixed six-tap, barrier-free path using four samples plus two horizontal resolved taps. No threadgroup staging.

Archived mean
30.065 dB
Quality delta
−0.237 dB
Added full-path time vs Bilinear
+0.629 ms*

Metal rendering contract

Keep phase evidence on-chip.

MSPR runs before conventional resolve, keeps working data in tile memory, and persists only the final 2× image.

Input4× MSAA color + local resolve
Temporal historyNot required
Motion / depthNot required
Model weightsNot required

Measured evidence

A clear gain in the documented matrix.

Archived physical-device results use same-frame Direct 2× references. The broader 100-frame set is secondary, project-built evidence.

RGB PSNR / HIGHER IS BETTERIPHONE 13 · 9 SCENES
Archived nine-scene mean PSNR: MSPR-Quality 30.301 dB, MSPR-Lite 30.065 dB, MetalFX 29.410 dB, Bilinear 29.280 dB, FSR1 27.367 dB
Archived nine-scene mean. The chart is evidence for the documented workloads, not a universal ranking.

Run the demo

From clone to pixels.

The repository includes four shared schemes for Quality and Lite on iOS and macOS. A physical Apple GPU is required for the imageblock path and performance conclusions.

  1. 01

    Open Demo/NeuralUpsampling.xcodeproj in Xcode 16 or newer.

  2. 02

    Select an MSPR-Quality or MSPR-Lite shared scheme.

  3. 03

    For an iPhone, select your Development Team and use a bundle identifier you control.

  4. 04

    Build and run. Use NSS_RENDER_MODE to switch comparison paths.

QUICK START
git clone https://github.com/long-0228/mspr2x.git
cd mspr2x
open Demo/NeuralUpsampling.xcodeproj
No package-manager dependencies

Reports and boundaries

The evidence is part of the release.

Read the full method, measurements, limitations, and provenance in English or Chinese.

PDF / ENEnglish technical report PDF / 中文中文技术报告 MIT / 2026License SCOPELimitations & provenance
PROVENANCE NOTE
Historical Apple demo scenes are not distributed.

The archived temple and fullscreen_image rows used assets from Apple's downloadable order-independent-transparency demo for testing only. Those assets are omitted because this project does not grant redistribution rights. Current source reproduction covers seven project-authored procedural scenes.

Apple demo ↗

* Difference between nine-scene full-path GPU means on the archived iPhone 13 matrix. This is not isolated kernel timing; see the technical report for the complete contract.

FULL COMPARISON