Add HDR10 GPU rendering on Windows11, HiDPI zoom-box fix, distortion playback, and 16-bit PNG suppor

[supernova-jpg]

Summary

This PR extends YUView with a GPU-accelerated HDR10 viewing pipeline for 10-bit YUV sources, built on Qt 6 RHI (D3D12 on Windows). It also includes several complementary improvements developed alongside the HDR work:

• HDR transfer functions: PQ (HDR10), HLG, and Linear (scRGB) for 10-bit analytical viewing — including a 10-bit SDR / pass-through path without PQ/HLG tone mapping
• Raw YUV frame caching to bypass CPU YUV→RGB during HDR playback
• HiDPI-aware zoom compensation so canvas layout and zoom box magnification stay consistent across Windows display scaling (100% / 125% / 150% / 200%)
• First-level / second-level distortion playback controls for subjective quality assessment
• 16-bit PNG/BMP loading with full sample precision preserved
• YUV pipeline refactor (YUVDataManager, YUVColorConverter, etc.) as the foundation for GPU upload and per-pixel tools

---

Motivation

YUView is widely used for codec development and subjective quality analysis of 10-bit BT.2020 content. The existing CPU QPainter path quantizes to 8-bit SDR and cannot drive a native HDR10 display. This PR adds a parallel HDR overlay window that:

1. Keeps the existing SDR UI and tools intact
2. Uploads raw 10-bit YUV directly to the GPU
3. Applies PQ / HLG / Linear transfer functions in fragment shaders
4. Preserves overlay tools (grid, zoom box, split view, pixel values) in the HDR path

---

HDR Rendering

Supported render modes (QSettings: HDRMode)

Mode	UI label	Transfer function	Exposure controls	Use case
0	PQ (HDR10)	SMPTE ST.2084 PQ EOTF	350–10,000 nits (log slider)	Standard HDR10 BT.2020 content
1	HLG	ARIB STD-B67 (shader OETF⁻¹ + OOTF)	350–1,000 nits	Broadcast HLG content
2	Linear (scRGB)	None — linear light pass-through	Disabled	10-bit SDR / analytical viewing: inspect linear BT.2020 without tone mapping; 1.0 scRGB = 80 nits reference white

All modes share a unified RGBA16F (scRGB linear light) internal rendering path and an HDR10 swap chain (R10G10B10A2 + PQ color space on Windows).

Shader pipeline (hdr_rhi_yuv_fragment.frag)

YUV (10-bit planar / P010) → BT.2020 YCbCr→R'G'B' matrix
  → [PQ EOTF | HLG decode | Linear skip]
  → exposure scaling (PQ/HLG only): display_luminance *= userNits / peakNits
  → BT.2020 → sRGB gamut + scRGB nits scaling
  → RGBA16F output → HDR10 presentation

Supported YUV input formats (HDRTextureUploader):

• YUV420p10le, YUV422p10le, YUV444p10le (planar)
• P010le (semi-planar; MSB-aligned samples normalized to internal 10-bit layout)

Performance optimizations

• Raw YUV cache (videoHandler::CacheMode::RawYUV): when Enable10BitDisplay is on and source is 10-bit planar, frames are cached as raw YUV (~25% smaller than RGBA for 4:2:0) instead of CPU-converted QImage
• Zero-copy GPU upload: updateHDRFrameYUVMove() moves frame buffers into the upload path, avoiding ~50 MB memcpy per 4K frame (~50 ms saved)
• Lazy settings cache: atomic m_enable10BitDisplayCached avoids per-frame QSettings disk I/O

UI controls (videoHandlerYUV.ui)

Control	Setting key	Behavior
Enable HDR display	Enable10BitDisplay	Master toggle; requires app restart (16-bit QSurfaceFormat set at startup in yuviewapp.cpp)
HDR mode combo	HDRMode	PQ / HLG / Linear; restart required
Tone mapping slider	HDRExposureNits	Log-spaced exposure target in nits; live update when HDR is active
Restart now	—	Relaunches app to apply HDR surface format changes

HDR activates only when all of the following are true:

• User enabled Enable10BitDisplay (and restarted)
• Source is 10-bit YUV (isHDR10Candidate())
• Windows DXGI reports HDR-capable display

HDR detection & fallback

• HDRDetectionWorker runs DXGI enumeration off the GUI thread
• Multi-monitor aware: matches display by name; debounced re-check on window move/resize
• On HDR loss (e.g. dragging to SDR monitor): emits signalDisplayHDRSupportChanged → MainWindow auto-disables HDR
• On detection failure: clears setting, shows error, auto-restarts in SDR mode after 2 s

---

Architecture

New module under YUViewLib/src/video/hdr/:

Component	Responsibility
HDRRenderingManager	Singleton orchestrator: detection → window creation → SplitViewWidget attach → frame forwarding
HDR_RhiVideoWindow	QWindow + QRhi render surface; overlay state; tryInitialize() for delayed expose handling
HDR_RhiVideoWindow_Overlays.cpp	Grid, zoom box, split line, loading spinner, pixel values — rendered in HDR color space
HDRRhiPipeline	RHI backend (D3D12 preferred on Windows), swap chain, shader pipelines, texture upload
HDRColorConversion	BT.601/709/2020 limited/full range matrices
HDRTextureUploader	YUV buffer parsing and GPU upload configuration
HDRDetection / HDRDetectionWorker	Windows DXGI HDR capability probe

Integration with SplitViewWidget

• HDR window embedded via QWidget::createWindowContainer() as a full-widget overlay
• Overlay is mouse-transparent; wheel events forwarded to parent for zoom
• paintEvent: when HDR active, pushes overlay state to HDR_RhiVideoWindow and skips SDR base draw to prevent double rendering
• Chicken-and-egg init fix: calls updateHDRVisibilityForSelection() immediately after embedding (not only on widgetInitialized)
• Scheduled tryInitialize() at 50 / 150 / 300 ms as fallback for delayed expose events

---

HiDPI Zoom Box & Canvas Consistency

Problem

On Windows with 125%/150%/200% display scaling, Qt reports devicePixelRatio() > 1. Without compensation, the same logical zoom factor maps to different pixel densities, causing:

• Mismatched canvas size vs. video pixels
• Zoom box magnification drifting relative to the underlying frame
• HDR overlay transform diverging from SDR painter

Solution (splitViewWidget)

Introduces dual zoom semantics:

Variable	Formula	Used for
zoomFactor (raw)	User zoom steps	UI label, zoom-to-fit target, zoom-box hide threshold
effectiveZoomFactor()	zoomFactor / devicePixelRatio	Item drawing, HDR projection matrix (m_parentZoom)

devicePixelRatioForCurrentScreen() resolves DPR via cascade:
windowHandle() → screen() → primary screen → logicalDpiX() / 96

Applied in:

• paintEvent, setMoveOffset, zoomToFitInternal (physical widget size = logical × DPR)
• HDR_RhiVideoWindow::updateTransform(renderZoom, offset, displayZoom) — render uses DPI-compensated zoom; overlay label shows raw zoom

Zoom box behavior (SDR + HDR)

• Samples a 5×5 source region, magnified 32× into a 160×160 px preview window
• Hidden when effectiveZoom >= 32 (already at pixel level)
• HDR optimization: lazy m_overlayPatchProvider calls videoHandlerYUV::getCurrentFramePatchAsImage(region) — converts only the 5×5 patch via YUVPixelRenderer, not the full frame
• Y/U/V readout: m_overlayYuvPixelProvider → getYuvPixelValueAt() for the info panel
• HDR overlay grid lines use 1/DPR snapping for crisp 1-px lines on HiDPI displays

---

Distortion Playback (First-level / Second-level)

Extracted into DistortionPlaybackController (YUViewLib/src/video/yuv/).

Button	Playback rate	Purpose
First-level	30 FPS	Fast subjective sweep — catch brief artifacts, blocking, motion issues
Second-level	1.5 FPS	Slow frame-by-frame inspection — fine structural distortion

Workflow:

1. Click level button → green active style, view forced to 1× zoom on all splitViewWidget instances
2. Captures current frame as revert point; enables Repeat One loop mode
3. Second click → stop, pause, reset repeat mode
4. Revert buttons seek back to captured frame
5. Stall detection: auto-stops after 3 timer ticks with no frame advance

UI located in videoHandlerYUV.ui (pushButtonFirstLevel, pushButtonSecondLevel, revertButton_L1/L2).

Note: ORI reference-file comparison (findOriFile(), toggleOriComparison()) is scaffolded in videoHandlerYUV_UI.cpp but loading is marked TODO.

---

16-bit PNG / BMP Support

Implemented in FrameHandler.cpp:

Function	Description
load16BitPng()	Custom PNG parser (IHDR + IDAT inflate/unfilter) → QImage::Format_RGBA64 or Format_Grayscale16
load16BitBmp()	48/64-bit uncompressed BMP → Format_RGBA64
loadImagePreservingHighBitDepth()	Entry point: tries 16-bit loaders for .png/.bmp, falls back to QImageReader
getImageRgbSample()	Returns native 16-bit samples for pixel-value overlay
configureHighBitDepthRendering()	One-time OpenGL format check for QPainter high bit-depth path

User-visible effect: image sequences and single frames retain full 16-bit sample precision (not clamped to 8-bit). Pixel-value overlay shows extended range. HDR window accepts QImage::Format_RGBA16FPx4 for RGB upload.

---

YUV Pipeline Refactor

Monolithic videoHandlerYUV.cpp logic split into focused modules:

Module	Role
YUVDataManager	Raw buffer lifecycle, bytes/frame, shouldUseRawYUVCache() policy
YUVColorConverter	YUV→RGB conversion API
YUVPixelRenderer	Per-pixel readout + zoomed value drawing
YUVDifferenceCalculator	Frame diff, MSE/PSNR/SSIM
YUVConversionCore/Planes/RGB.h	Hot-path math decomposition
videoHandlerYUV_UI.cpp	UI slot logic extracted from handler
DistortionPlaybackController	Distortion playback timing

Enables both the HDR GPU path and efficient zoom-box / pixel-value tools without full-frame CPU conversion.

---

Other Notable Changes

• Video cache thread safety: VideoCache / LoadingWorker guards against races when dragging files during 4K 10-bit buffer loading
• Qt model fixes: PacketItemModel uses valid index ranges (avoids dataChanged() warnings)
• PlaylistTreeWidget: viewport()->update() instead of invalid dataChanged() indices
• Shader build: YUViewLib.pro adds qsb compile rule for .vert/.frag → .qsb (requires Qt shadertools)
• Build deps: gui-private (QRhi), Windows libs dxgi, user32, ole32

---

Platform Support

Platform	HDR rendering	HDR detection	Fallback
Windows 10/11	Full (D3D12 + DXGI HDR10 swap chain)	DXGI async probe	Auto-restart to SDR on failure
Linux	RHI backends present (Vulkan/OpenGL)	Not implemented (isHDRSupported = false)	Graceful SDR
macOS	RHI Metal backend present	Not implemented	Graceful SDR

Requirements: Qt 6.4+ (6.6+ recommended for public QRhi API), shadertools for shader compilation.

---

Settings Reference

Key	Type	Default	Description
Enable10BitDisplay	bool	false	Master HDR / 10-bit startup toggle
HDRMode	int	0	0=PQ, 1=HLG, 2=Linear
HDRExposureNits	int	1000	Exposure target (live when HDR on)
ToneMapTargetNits	int	1000	Legacy fallback for exposure

---

Test Plan

Build

• Windows: Qt 6.6+ with shadertools; MSVC build succeeds; qsb shaders generated
• Linux/macOS: project compiles with HDR code paths disabled at runtime

HDR core

• Enable HDR Display, restart; open 10-bit YUV420 BT.2020 file
• Verify PQ mode on HDR monitor — highlights beyond SDR white visible
• Switch to HLG — correct gamma/brightness relative to PQ
• Switch to Linear (scRGB) — no tone mapping; exposure slider disabled; linear pass-through
• Adjust exposure slider — live brightness change in PQ/HLG
• Non-HDR monitor → graceful fallback message + SDR restart

HiDPI / Zoom box

• Test at Windows scaling 100%, 125%, 150%, 200%
• At same logical zoom, video pixel mapping identical across scaling levels
• Zoom box 32× magnification aligned with cursor pixel in both SDR and HDR
• Zoom box Y/U/V readout matches pixel under cursor

Distortion playback

• First-level: 30 FPS loop, revert works, second click stops
• Second-level: ~1.5 FPS loop, revert works
• View resets to 1× on activation

16-bit images

• Load 16-bit PNG sequence — pixel values show >255 range
• Load 16-bit BMP — no banding to 8-bit

Regression

• 8-bit YUV/RGB playback unchanged in SDR mode
• Seek, frame cache, split view, comparison view, statistics overlay
• Drag new file while caching 4K 10-bit — no crash
• Multi-monitor: move window SDR↔HDR — debounced detection, no black overlay

---

Known Limitations & Future Work

• HDR detection is Windows-only; Linux/macOS need platform-specific probes
• getHDRRenderedImage() returns empty QImage (stub)
• No unit tests for HDR color conversion / texture upload yet
• ORI reference-file distortion comparison not fully wired
• App restart required for HDR enable/disable (by design — avoids runtime GL format switching)
• Happy to split into smaller PRs (YUV refactor → HDR core → UI/integration) if maintainers prefer

---

PR Scope Note

This branch also includes the YUV pipeline refactor and minor Qt warning fixes that were developed as prerequisites for the HDR GPU upload path. Non-HDR debug-macro cleanups in decoder/parser files are included only where they were part of the same development branch; happy to isolate further on request.

[supernova-jpg]

Hi @ChristianFeldmann,
I understand this is a massive PR with a lot of changes. To avoid putting any potential risks or instability onto the develop branch, would it be better to create a dedicated feature branch (e.g., feature/hdr-rendering) on the main repo? We could merge these changes there first for thorough testing.
Also, if you prefer, I can split out the independent fixes (like the HiDPI zoom-box fix and 16-bit PNG support) into smaller, separate PRs targeting develop so they can be reviewed and merged much faster.
Let me know what you think!