Add 16-bit Bayer pixel format support#54
Open
dcliftreaves wants to merge 11 commits intogopro:masterfrom
Open
Conversation
Add RGGB_16 and GBRG_16 pixel format definitions and supporting infrastructure for encoding and decoding 16-bit raw Bayer sensor data. - Add PIXEL_FORMAT_RAW_RGGB_16 and PIXEL_FORMAT_RAW_GBRG_16 enums - Add 16-bit log curve tables (EncoderLogCurve16, DecoderLogCurve16) with 65536-entry LUTs for the larger input domain - Add integer-based uncompanding table (InitUncompandTable) to replace double-precision cubic evaluation in the decoder hot path - Extend image format handling for 16-bit pixel formats - Add 16-bit support in bitstream and wavelet common code Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Extend the VC5 encoder to support 16-bit Bayer input with appropriate quantization scaling and three new high-quality presets for the larger dynamic range. 16-bit encoding: - Add VC5_ENCODER_PIXEL_FORMAT_RGGB_16 and GBRG_16 format handling - Add UnpackImage_16() for 16-bit raw Bayer data - Scale quantization tables by 12/bits for bit depths > 12, protecting the lowpass band (index 0) which must always keep quant=1 - Adjust midpoint_prequant for 16-bit (value 3 for >= 15-bit) - Use dynamic VC5 buffer sizing (1.5x raw size + 1MB) instead of fixed 10MB to handle larger 16-bit payloads Quality presets: - Add Q6 (Filmscan-3/Edit-Safe), Q7 (Filmscan-4/Near-Lossless), Q8 (Filmscan-5/Virtually-Lossless) quality settings Performance: - Parallelize forward wavelet transforms using pthreads (one thread per Bayer channel, ~2x speedup on multi-core) - Link encoder against pthreads Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Extend the VC5 decoder to support 16-bit Bayer output with optimized entropy decoding and parallel inverse wavelet transforms. 16-bit decoding: - Add VC5_DECODER_PIXEL_FORMAT_RGGB_16 and GBRG_16 format handling - Add fast 16-bit lowpass coefficient bulk read path that extracts coefficients directly from the 32-bit bitstream buffer - Pass actual bit depth through to RGB conversion instead of hardcoding 12 or 14 bits - Add 16-bit repacking in decoder raw output VLC prefix lookup table: - Build a 4096-entry prefix table (12-bit peek) from the codebook for O(1) codeword lookup instead of linear search through 264 entries - Thread-safe initialization via pthread_once - Fast combined RLV + sign bit decode (GetRunFast) with automatic fallback to original GetRun for long codewords - Safety valve: set VLC_NO_FAST=1 to disable for verification Optimized highpass run-length decoding: - Fast path for runs that stay within a single row (common case) - Use memset for zero-valued runs instead of per-pixel loop - Batch row-boundary handling for cross-row runs Performance: - Parallelize inverse wavelet transforms using pthreads (one thread per Bayer channel) - Fix VLC table initialization race condition with pthread_once - Check GetBuffer() return value in lowpass band reader - Link decoder against pthreads Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Wire 16-bit Bayer pixel format support through the GPR SDK layer and command-line tools, with several robustness fixes discovered during testing. SDK integration: - Add PIXEL_FORMAT_RGGB_16 and PIXEL_FORMAT_GBRG_16 to gpr_tuning_info - Add quality parameter to gpr_parameters for explicit quality selection - Pass 16-bit pixel formats through set_vc5_encoder_parameters() - Auto-detect bit depth from DNG WhiteLevel during decoding - Add default_crop_size_h/v fields to gpr_tuning_info Bug fixes: - Guard ProfileByIndex() call against empty profile count - Guard GetLinearizationInfo() against null profile pointer - Add HueSatMap 64-bit overflow check (reject > 8M entries) - Deep-copy and properly destroy HueSatMap data in gpr_parameters - Auto-compute input pitch from width when pitch is 0 or unset - Set saturation levels correctly per pixel format bit depth CLI (gpr_tools): - Add --Quality/-q flag for explicit encoder quality selection (0=Low through 8=Virtually-Lossless, -1=auto) - Add rggb16 and gbrg16 to --InputPixelFormat options - Auto-compute input pitch from width (default 0 instead of 8000) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Enable link-time optimization (LTO/IPO) for cross-translation-unit inlining of hot paths (GetBits, GetBuffer, VLC lookup) - Auto-detect ARM NEON on arm64/aarch64 (Apple Silicon M-series) instead of requiring manual -DNEON=1 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Detailed PSNR, ENOB, and editing latitude analysis for all Q0-Q8 presets tested on 100MP 16-bit Hasselblad X2D sensor data. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Collaborator
|
This is great stuff, I have to find the time to try it all out. Thanks. |
The optimized fast path in DecodeBandRuns consumed all run pixels but never zeroed run.count, causing the post-loop assertion `assert(data_count == 0 && run.count == 0)` to fail on every decode. The slow path correctly decremented run.count to zero, but the fast path (which handles the common case of runs within a single row) skipped this step. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Author
Test Plan VerificationAll 5 checklist items have been verified. One critical bug was found and fixed (commit 28dd522). Results
Bug Fixed:
|
| Quality | GPR Size | PSNR |
|---|---|---|
| Q0 | 3.2 MB | 70.91 dB |
| Q2 | 4.5 MB | 75.21 dB |
| Q4 | 5.4 MB | 79.34 dB |
🤖 Tested with Claude Code
…o DNGs Two fixes discovered while testing with a Hasselblad X2D 100C DNG: 1. gpr.cpp: EXIF software_version and user_comment fields used assert + unchecked memcpy, crashing on DNGs with strings longer than 32 bytes. Now truncates safely with null termination. 2. syntax.c: PutTagPair/PutTagPairOptional asserted that tag values fit in 16 bits, but TAGWORD (int16_t) sign-extends to int when the high bit is set. The packed prescale shift for 16-bit data (0xBC00) became 0xFFFFBC00 (-17408), failing the assertion. Now masks through uint16_t before the check. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Author
Additional Fixes and Verified PSNR DataNew fixes (8ec08a0)
Verified 16-bit PSNR (Hasselblad X2D 100C, 11664×8750)Measured via RAW round-trip on the actual X2D test image:
All values match the PR description within rounding. Note on 12-bit output differences vs masterThe widening of 🤖 Tested with Claude Code |
- VLC: Assert single-codebook assumption instead of silently overwriting the global codebook pointer on every call to GetRunFast - raw.c: Add explicit GBRG format cases and assert on unknown pixel formats instead of silently defaulting to GBRG order Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…erflow - vc5_encoder.c: Add lower-bound check on quality_setting to prevent negative index into quant_table array (buffer underflow) - gpr.cpp: Add comment clarifying GBRG 14-bit maps to GBRG_16 (no GBRG_14 enum exists); zero HueSatMap dims when sanity check triggers to prevent downstream out-of-bounds read from stale dimensions Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Author
Full PR Review and RetestCode ReviewThree parallel review agents audited the decoder, encoder, and SDK/common changes (~4K lines across 43 files). Findings triaged and addressed: Bugs fixed (451c2bf, 221968f):
Reviewed and dismissed:
Test Results
Commits added since last update
🤖 Reviewed and tested with Claude Code |
The COMPONENT_VALUE type was widened from uint16_t (2 bytes) to int32_t (4 bytes), but the pitch-to-element-count conversion in the GPR_RGB_RESOLUTION_HALF path still divided by 2 instead of sizeof(COMPONENT_VALUE). This caused WaveletToRGB to use a stride of 2*width instead of width, reading past allocated buffer boundaries and producing corrupted RGB output at half resolution. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Author
Systematic Codebase Audit (10 Risk Areas)Five parallel audit agents reviewed the entire PR diff across 10 risk areas. One critical bug found and fixed. Critical Bug Fixed (90bf199)WaveletToRGB pitch/sizeof mismatch — Audit Summary
All Commits in PR (10 total)
🤖 Audited with Claude Code |
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
3 participants
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.
Summary
Add support for encoding and decoding 16-bit raw Bayer sensor data (RGGB_16, GBRG_16) in the VC5/GPR codec, along with performance optimizations and robustness fixes.
Quality Analysis — 16-Bit Sensor Data
Tested on a 100-megapixel 16-bit Hasselblad X2D image (11664×8750, 194.7 MB raw). This sensor captures ~14 bits of real photographic data at base ISO — typical of modern medium-format and high-end full-frame sensors.
Raw Bayer Domain PSNR
This is the domain RAW editors (Lightroom, Capture One, RawTherapee, etc.) operate in. Errors here directly affect all downstream edits.
ENOB = Effective Number of Bits = PSNR / 6.02. DN = out of 65535 for 16-bit.
Q6-Q8 (bold) are the new presets added in this PR.
What This Means for Editing Latitude
Each stop (EV) of exposure adjustment doubles or halves pixel values, which effectively costs ~1 bit of precision. The practical editing headroom is roughly:
ENOB − display bits = available editing stops
For an 8-bit final output (web, prints, most displays):
14-Bit vs 16-Bit Context
Most current cameras output 14-bit raw data. 16-bit sensors (like the tested X2D) add 2 more bits of dynamic range (~12 additional dB), which is critical for:
For 14-bit sensor data (e.g., GoPro HERO, most DSLRs/mirrorless), the quantization tables already scale automatically (
scale = 12/bits), meaning:For 16-bit sensor data, the new Q6-Q8 presets are essential:
Recommended Presets
Commits
Backwards Compatibility
VLC_NO_FAST=1)Test plan
VLC_NO_FAST=1— verify bit-exact match with fast path🤖 Generated with Claude Code