gopro · dcliftreaves · Feb 12, 2026 · Feb 12, 2026 · Feb 12, 2026 · Feb 12, 2026
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -8,6 +8,32 @@ set(CMAKE_C_FLAGS "-std=c99")
 # project name
 project( gpr )
 
+# Enable link-time optimization for cross-TU inlining (GetBits, GetBuffer, etc.)
+cmake_policy(SET CMP0069 NEW)
+include(CheckIPOSupported)
+check_ipo_supported(RESULT ipo_supported OUTPUT ipo_output)
+if(ipo_supported)
+  set(CMAKE_INTERPROCEDURAL_OPTIMIZATION TRUE)
+  message(STATUS "LTO/IPO enabled")
+else()
+  message(STATUS "LTO/IPO not supported: ${ipo_output}")
+endif()
+
+# Enable NEON on arm64 (Apple Silicon defaults to ON)
+set(NEON_DEFAULT OFF)
+if(APPLE AND CMAKE_SYSTEM_PROCESSOR MATCHES "^(arm64|aarch64)")
+  set(NEON_DEFAULT ON)
+endif()
+option(NEON "Enable ARM NEON intrinsics" ${NEON_DEFAULT})
+
+if(NEON)
+  add_compile_definitions(NEON=1)
+else()
+  add_compile_definitions(NEON=0)
+endif()
+
+message(STATUS "NEON enabled: ${NEON}")
+
 if(EXISTS "${CMAKE_SOURCE_DIR}/.git")
   execute_process(
     COMMAND git rev-parse --abbrev-ref HEAD

diff --git a/docs/quality-analysis.md b/docs/quality-analysis.md
@@ -0,0 +1,224 @@
+# GPR Quality Analysis — 16-Bit
+
+Test image: Hasselblad X2D 100C, 11664x8750, 16-bit Bayer (194.7 MB raw)
+Source: `2024_12_Dec_Austin_0186.dng` (138.2 MB lossless DNG)
+
+## File Size
+
+| Q | GPR Size | Compression | vs DNG | Space Saving |
+|---|----------|-------------|--------|--------------|
+| 0 | 21.4 MB | 9.1x | 6.4x smaller | 84.5% |
+| 1 | 28.7 MB | 6.8x | 4.8x smaller | 79.2% |
+| 2 | 39.2 MB | 5.0x | 3.5x smaller | 71.6% |
+| 3 | 46.0 MB | 4.2x | 3.0x smaller | 66.7% |
+| 4 | 53.0 MB | 3.7x | 2.6x smaller | 61.6% |
+| 5 | 66.9 MB | 2.9x | 2.1x smaller | 51.6% |
+| 6 | 89.8 MB | 2.2x | 1.5x smaller | 35.0% |
+| 7 | 90.9 MB | 2.1x | 1.5x smaller | 34.2% |
+| 8 | 92.0 MB | 2.1x | 1.5x smaller | 33.4% |
+
+Compression ratio is vs uncompressed 16-bit raw (194.7 MB).
+
+## Raw Bayer Domain (pre-demosaic)
+
+This is the domain RAW editors (Lightroom, Capture One, etc.) operate in.
+Errors here directly affect all downstream edits.
+
+| Q | PSNR | RMS Error | MAE | MAE % | Max Error | ENOB |
+|---|------|-----------|-----|-------|-----------|------|
+| 0 | 40.0 dB | 655 DN | 334 DN | 0.509% | 11454 DN | 6.6 bits |
+| 1 | 44.6 dB | 386 DN | 209 DN | 0.319% | 6682 DN | 7.4 bits |
+| 2 | 48.5 dB | 246 DN | 136 DN | 0.208% | 4437 DN | 8.1 bits |
+| 3 | 50.0 dB | 207 DN | 116 DN | 0.176% | 4115 DN | 8.3 bits |
+| 4 | 51.2 dB | 181 DN | 100 DN | 0.153% | 4179 DN | 8.5 bits |
+| 5 | 52.1 dB | 163 DN | 89 DN | 0.135% | 3749 DN | 8.7 bits |
+| 6 | 56.5 dB | 98 DN | 53 DN | 0.081% | 2190 DN | 9.4 bits |
+| 7 | 58.6 dB | 77 DN | 41 DN | 0.062% | 1597 DN | 9.7 bits |
+| 8 | 58.9 dB | 74 DN | 39 DN | 0.060% | 1566 DN | 9.8 bits |
+
+- ENOB = Effective Number of Bits = PSNR / 6.02
+- DN = Digital Number (out of 65535 for 16-bit)
+- This sensor captures ~14 bits of real data at ISO 100
+
+## Rendered RGB (16-bit sRGB linear)
+
+After demosaicing + white balance + color matrix. This is what you see on screen.
+
+| Q | PSNR | R PSNR | G PSNR | B PSNR | MAE | >0.5% | >1% |
+|---|------|--------|--------|--------|-----|-------|-----|
+| 0 | 36.9 dB | 35.2 dB | 39.6 dB | 37.1 dB | 468 DN | 36.8% | 20.4% |
+| 1 | 40.5 dB | 38.7 dB | 43.4 dB | 40.7 dB | 317 DN | 28.7% | 13.0% |
+| 2 | 43.5 dB | 41.6 dB | 46.4 dB | 43.7 dB | 224 DN | 20.4% | 7.4% |
+| 3 | 44.5 dB | 42.6 dB | 47.5 dB | 44.7 dB | 196 DN | 17.2% | 5.8% |
+| 4 | 45.4 dB | 43.4 dB | 48.5 dB | 45.6 dB | 176 DN | 14.9% | 4.7% |
+| 5 | 46.1 dB | 44.1 dB | 49.4 dB | 46.4 dB | 160 DN | 13.1% | 3.9% |
+| 6 | 49.5 dB | 47.7 dB | 52.2 dB | 49.7 dB | 95 DN | 5.4% | 1.2% |
+| 7 | 50.9 dB | 49.3 dB | 53.2 dB | 51.2 dB | 72 DN | 3.2% | 0.7% |
+| 8 | 51.1 dB | 49.5 dB | 53.4 dB | 51.4 dB | 69 DN | 3.0% | 0.6% |
+
+- Red channel is weakest (lowest PSNR) — typical for Bayer demosaic
+- ">0.5%" / ">1%" = % of pixel values with error exceeding that fraction of full scale
+
+## Practical Editing Guidance
+
+Context: This sensor (X2D 100C) captures ~14 bits of real data at ISO 100.
+Compression ENOB ranges from 6.6 bits (Q0) to 9.8 bits (Q8).
+Compression error is always above sensor noise, so it is the limiting factor.
+
+### Q0 — Aggressive Compression (9.1x, 21 MB)
+
+Raw ENOB: 6.6 bits | PSNR: 40.0 dB | RMS error: 655 DN
+
+- OK: Export to JPEG/web at 4K or smaller — artifacts masked by downsampling + 8-bit
+- OK: Basic white balance and exposure +/- 0.5 stop
+- OK: Light color grading (LUT application, saturation +/- 20%)
+- NO: Shadow recovery — pushing shadows reveals quantization banding
+- NO: Heavy curves / local contrast — amplifies per-band wavelet artifacts
+- NO: Large prints from full resolution — artifacts visible at ~200 PPI+
+
+Best for: Preview/proxy files, web gallery, social media, quick turnaround
+
+### Q1 — High Compression (6.8x, 29 MB)
+
+Raw ENOB: 7.4 bits | PSNR: 44.6 dB | RMS error: 386 DN
+
+- OK: Exposure adjustment +/- 1 stop
+- OK: Moderate white balance correction
+- OK: Saturation and vibrance adjustments
+- MAYBE: Shadow push up to +1 stop — minor artifacts in deep shadows
+- NO: Heavy shadow recovery (more than +1 stop)
+
+Best for: Event photography, documentary, editorial with light edits
+
+### Q2-Q3 — Moderate Compression (5.0-4.2x, 39-46 MB)
+
+Raw ENOB: 8.1-8.3 bits | PSNR: 48.5-50.0 dB | RMS error: 207-246 DN
+
+- OK: Exposure +/- 1.5 stops
+- OK: Moderate shadow recovery (+1.5 stops)
+- OK: Color grading and creative white balance shifts
+- OK: Clarity / texture / dehaze adjustments
+- OK: Print up to 24x36" at 300 PPI
+- MAYBE: Aggressive curves — some wavelet ringing in smooth gradients
+- NO: Extreme shadow push (> +2 stops)
+
+Best for: Professional editorial, portrait, landscape with moderate edits
+
+### Q4-Q5 — Balanced (3.7-2.9x, 53-67 MB)
+
+Raw ENOB: 8.5-8.7 bits | PSNR: 51.2-52.1 dB | RMS error: 163-181 DN
+
+- OK: Exposure +/- 2 stops
+- OK: Shadow recovery up to +2 stops
+- OK: Aggressive color grading (split toning, cross-process, film emulation)
+- OK: Local adjustments (dodge/burn, graduated filters, radial masks)
+- OK: Sharpening and noise reduction
+- OK: Large prints at full resolution
+- MAYBE: Extreme shadow push (+3 stops) — faint banding in deepest shadows
+
+Best for: Professional photography with full editing workflow
+
+### Q6-Q8 — Near-Lossless (2.2-2.1x, 90-92 MB)
+
+Raw ENOB: 9.4-9.8 bits | PSNR: 56.5-58.9 dB | RMS error: 74-98 DN
+
+- OK: Shadow recovery up to +3 stops
+- OK: Extreme exposure adjustments (+/- 3 stops)
+- OK: Aggressive local contrast / clarity / HDR tone mapping
+- OK: Fine art printing at maximum size
+- OK: Pixel-level compositing and retouching
+- MAYBE: Extreme shadow push (+4 stops) — may show faint artifacts
+
+Note: Q7 and Q8 are nearly identical (58.6 vs 58.9 dB) because the
+quantization divisor hits the noise floor. Q6 offers the best
+quality/size tradeoff in this tier.
+
+Best for: Archival, fine art, maximum editing flexibility
+
+## Key Insight
+
+This is a 16-bit sensor with ~14 bits of real data. Even Q8 (best quality)
+only preserves ~10 bits of the 14 available. The remaining ~4 bits of sensor
+precision are lost to wavelet quantization. For truly lossless archival,
+keep the original DNG.
+
+However: for practical photography, 10 bits of ENOB in the raw domain
+translates to 8+ bits of final output precision after demosaic + color
+processing — more than sufficient for any 8-bit display or print.
+
+The compression artifacts are in the wavelet domain (smooth spatial patterns
+across 8x8 or 16x16 pixel blocks), not in the pixel domain. They are less
+visually objectionable than JPEG artifacts at the same PSNR.
+
+## Per-Channel Quantization Scaling
+
+VC5 encodes Bayer data in color-difference space (not raw RGGB):
+
+| Channel | Name | Content |
+|---------|------|---------|
+| 0 (GS) | Green Sum | (Gr+Gb)/2 — luminance |
+| 1 (RG) | Red-Green | R-G — chrominance |
+| 2 (BG) | Blue-Green | B-G — chrominance |
+| 3 (GD) | Green Diff | (Gr-Gb)/2 — detail/noise |
+
+The `--ChannelScale` flag (or `channel_quant_scale` API) applies per-channel
+multipliers to quantization. Scale > 1.0 = more compression, < 1.0 = less.
+
+### Test Results (Q4 base, 100MP 16-bit)
+
+| Config | Scale (GS,RG,BG,GD) | Size | Raw PSNR | RGB PSNR | G PSNR | R PSNR | B PSNR |
+|--------|---------------------|------|----------|----------|--------|--------|--------|
+| Uniform Q4 | 1,1,1,1 | 53.0 MB | 51.2 dB | 44.6 dB | 49.2 dB | 42.0 dB | 45.3 dB |
+| A | 1.0, 2.0, 2.0, 3.0 | 39.9 MB | 46.2 dB | 39.9 dB | 45.0 dB | 37.2 dB | 40.7 dB |
+| B | 1.0, 3.0, 3.0, 4.0 | 35.3 MB | 43.7 dB | 37.4 dB | 43.2 dB | 34.5 dB | 38.4 dB |
+| C | 0.5, 2.0, 2.0, 2.0 | 47.4 MB | 47.5 dB | 40.4 dB | 46.3 dB | 37.5 dB | 41.4 dB |
+| D | 1.0, 4.0, 4.0, 6.0 | 31.9 MB | 41.3 dB | 35.5 dB | 41.2 dB | 32.6 dB | 36.5 dB |
+
+### Size-Matched Comparison (per-channel vs uniform)
+
+| Per-Channel | Size | RGB PSNR | Uniform Match | Size | RGB PSNR | Delta |
+|-------------|------|----------|---------------|------|----------|-------|
+| Q4+A(1,2,2,3) | 39.9 MB | 39.9 dB | Uniform Q2 | 39.2 MB | 42.8 dB | -2.9 dB |
+| Q4+C(.5,2,2,2) | 47.4 MB | 40.4 dB | Uniform Q3 | 46.0 MB | 43.8 dB | -3.4 dB |
+
+### Analysis
+
+At the same file size, uniform quality presets outperform per-channel scaling
+by 2-3 dB. The reason: VC5's quality tables are already carefully tuned for
+the color-difference space. Each preset has per-band ratios (e.g., lower Q for
+low-frequency, higher Q for high-frequency) that are perceptually meaningful.
+A uniform multiplier across all bands within a channel is less efficient than
+using a different preset designed as a coherent whole.
+
+### When Per-Channel Scaling IS Useful
+
+- **Chroma-insensitive workflows**: Machine vision, scientific imaging where
+  only luminance matters — compress RG/BG/GD aggressively
+- **Noise-dominated detail channel**: At high ISO, GD is mostly noise —
+  compressing it with scale 1.5-2.0 saves space with no visible loss
+- **Fine-tuning between presets**: When Q3 is too large and Q2 is too lossy,
+  `Q3 -c "1.0,1.0,1.0,1.3"` gives a size between Q2 and Q3
+- **Artistic intent**: Deliberately softening color detail while preserving
+  luminance sharpness
+
+### CLI Usage
+
+```
+# Compress chroma 2x, detail 3x (GS untouched)
+gpr_tools -i input.dng -o output.gpr -q 4 -c "1.0,2.0,2.0,3.0"
+
+# Boost luminance quality, compress chroma mildly
+gpr_tools -i input.dng -o output.gpr -q 3 -c "0.7,1.0,1.0,1.3"
+```
+
+## Recommended Presets
+
+| Use Case | Recommended | Size (100MP) | PSNR |
+|----------|-------------|-------------|------|
+| Proxy / preview / web gallery | Q0 | 21 MB | 40.0 dB |
+| Event / documentary (light edit) | Q1 | 29 MB | 44.6 dB |
+| Editorial / portrait | Q2 or Q3 | 39-46 MB | 48-50 dB |
+| Professional (full edit workflow) | Q4 or Q5 | 53-67 MB | 51-52 dB |
+| Archival / fine art | Q6 | 90 MB | 56.5 dB |
+| Maximum quality (diminishing returns) | Q7 or Q8 | 91-92 MB | 59 dB |
+| Truly lossless | Original DNG | 138 MB | infinite |
diff --git a/source/app/gpr_tools/main.cpp b/source/app/gpr_tools/main.cpp
@@ -65,11 +65,13 @@ class my_argument_parser : public argument_parser
     string  rgb_file_resolution;
 
     int     rgb_file_bits;
-    
+
     string  output_file_path;
 
     string  apply_gpr_parameters;
-
+
+    int     quality;
+
 public:
 
     bool get_verbose() { return verbose; }
@@ -98,9 +100,9 @@ class my_argument_parser : public argument_parser
 
         ("InputHeight,h",                                   input_height,                               3000,                   "Input image height in pixel samples [3000]")
 
-        ("InputPitch,p",                                    input_pitch,                                8000,                   "Input image pitch in bytes [8000]")
+        ("InputPitch,p",                                    input_pitch,                                0,                      "Input image pitch in bytes (0 = auto from width)")
 
-        ("InputPixelFormat,x",                              input_pixel_format,                         string("rggb14"),       "Input pixel format \n(rggb12, rggb12p, [rggb14], gbrg12, gbrg12p)")
+        ("InputPixelFormat,x",                              input_pixel_format,                         string("rggb14"),       "Input pixel format \n(rggb12, rggb12p, [rggb14], rggb16, gbrg12, gbrg12p, gbrg16)")
 
         ("ApplyGprParameters,a",                            apply_gpr_parameters,                       string(""),             "Parameters to use for GPR or DNG file.")
 
@@ -109,7 +111,9 @@ class my_argument_parser : public argument_parser
         ("RgbFileResolution,r",                             rgb_file_resolution,                        string(""),             "Output RGB resolution \n[1:1, 2:1, 4:1, 8:1. 16:1]")
         ("RgbFileBits,b",                                   rgb_file_bits,                              8,                      "Output RGB bits [8]")
 
-        ("OutputFilePath,o",                                output_file_path,                           string(""),             "Output file path.\n(files types: GPR, DNG, PPM, RAW, JPG)");
+        ("OutputFilePath,o",                                output_file_path,                           string(""),             "Output file path.\n(files types: GPR, DNG, PPM, RAW, JPG)")
+
+        ("Quality,q",                                       quality,                                    -1,                     "Encoder quality\n(0=Low, 1=Medium, 2=High, 3=FS1, 4=FSX, 5=FS2, 6=FS3, 7=FS4, 8=FS5, -1=auto)");
         ;
     }
 };
@@ -195,7 +199,7 @@ int main(int argc, char *argv [])
     {
         return dng_convert_main(args.input_file_path.c_str(), args.input_width, args.input_height, args.input_pitch, args.input_skip_rows, args.input_pixel_format.c_str(),
                                 args.output_file_path.c_str(), args.apply_gpr_parameters.c_str(), args.gpmf_file_path.c_str(), args.rgb_file_resolution.c_str(), args.rgb_file_bits,
-                                args.jpg_preview_file_path.c_str(), args.jpg_preview_file_width, args.jpg_preview_file_height );
+                                args.jpg_preview_file_path.c_str(), args.jpg_preview_file_width, args.jpg_preview_file_height, args.quality );
     }
 
     return 0;