MCRAW SpeedTranscode

English / 简体中文

mcraw-speedtranscode is a MotionCam .mcraw → intermediate-format transcoding tool that balances speed and quality. It retains the CPU general-purpose reference pipeline while incorporating a GPU-acceleration design.

Current Status

• This project is still in early development and is not suitable for production use!

Features

• The official MotionCam decoder as the CPU RAW ground truth for compression 6/7.
• Independent reading and validation of MCRAW frame indexes and compressed payloads.
• RGGB, BGGR, GRBG, and GBRG.
• Black and white levels for each CFA position; negative and above-white values are preserved.
• librtprocess AMaZE, RCD, IGV, DCB, and LMMSE; RCD remains the default.
• Linear DWG Capture Sharpening with a default strength of 0.4.
• DNG CameraNeutral → xy iteration and inverse-CCT dual-matrix interpolation.
• Two color paths: ForwardMatrix, and ColorMatrix + Bradford.
• XYZ D50 → D65 → DaVinci Wide Gamut.
• Exact analytic DaVinci Intermediate OETF/EOTF, plus LUTs generated from analytic formulas and cached per conversion instance.
• BT.2020 non-constant-luminance RGB→YCbCr, left-sited quality 4:2:2, deterministic dithering, and 10-bit video-range packing.
• FFmpeg prores_ks ProRes 422 HQ, source timestamps, PCM audio, and sidecar JSON.
• OpenMP row parallelism, bounded multi-frame parallelism, a multi-context parallel ProRes encoding pipeline, and a CPU/RAM-aware automatic execution plan.
• An FFmpeg-owned Vulkan device, GPU RGB→10-bit 4:2:2, timeline semaphores, bounded frame/job/packet queues, asynchronous Vulkan ProRes encoding, and an independent mux worker.
• Automatic/forced GPU selection, stable UUID/PCI/name selectors, CPU fallback, structured queue/transfer/mux telemetry, and a partial-file strategy that renames files only after successful validation.
• inspect, convert, extract-frame, validate, benchmark, print-effective-config, and list-capabilities.

Environment

• Windows 10 or 11
• An AVX2-capable x86-64 CPU (the default build uses AVX2 as its minimum CPU baseline)
• The Vulkan GPU backend requires support for compute, timeline semaphores, and storage-compatible 10-bit 4:2:2 images; the CPU backend does not require a GPU
• Visual Studio 2022 C++ workload
• CMake 3.25+
• Git
• vcpkg (the repository script can install it into the ignored .deps/ directory, or an existing installation can be used)

Dependencies are pinned and fetched during configuration:

• MotionCam decoder commit 2c49edb17277c07989ff90bd3a3bf557c2f68b4a
• librtprocess 0.12.0
• Catch2 v3.15.0
• FFmpeg 8.1.2 (resolved by the fixed vcpkg baseline)
• nlohmann-json uses the version bundled with the pinned MotionCam decoder to avoid mixing ABIs across DLLs

Build

In “Developer PowerShell for VS 2022”:

# Run this once first if vcpkg is not already installed:
.\scripts\bootstrap-vcpkg.ps1

cmake --preset msvc-release `
  -DCMAKE_TOOLCHAIN_FILE="$env:VCPKG_ROOT/scripts/buildsystems/vcpkg.cmake"
cmake --build --preset msvc-release
ctest --preset msvc-release

The script installs vcpkg 2026.06.24 into .deps/vcpkg and sets VCPKG_ROOT for the current PowerShell process. If you use an existing vcpkg installation, set an environment variable with the same name yourself; the baseline in the manifest will still pin the port versions.

The executable is usually located at:

build/msvc-release/Release/mcraw-speedtranscode.exe

Usage

$exe = '.\build\msvc-release\Release\mcraw-speedtranscode.exe'

& $exe inspect '.\mcraw_sample\sample.mcraw' --raw-json
& $exe validate '.\mcraw_sample\sample.mcraw' --frame 0 --compare-fused
& $exe benchmark '.\mcraw_sample\sample.mcraw' --frames 8
& $exe print-effective-config
& $exe convert '.\mcraw_sample\sample.mcraw' '.\output.mov'
& $exe convert '.\mcraw_sample\sample.mcraw' '.\output-gpu.mov' `
    --config '.\config\vulkan-gpu-pipeline.json'

The conversion atomically renames .partial.mov to the target file only after the trailer has been written successfully. Existing output is not overwritten by default; add --overwrite explicitly when needed.

Start modifying the effective configuration from config/default.json. The Vulkan opt-in presets are config/vulkan-precise.json and config/vulkan-fast.json; the CPU backend remains the default. precise uses an FP32 intermediate, while fast uses FP16 storage + FP32 compute and FP32 analytic DI; both retain precise GPU RCD, quality chroma, and deterministic dithering:

& $exe convert input.mcraw output.mov --config '.\config\default.json'

Quality options and comparison configuration:

{
  "schema_version": 1,
  "demosaic": "rcd",
  "capture_sharpening": 0.4,
  "capture_sharpening_threshold": 0.002
}

demosaic can be set to rcd, amaze, igv, dcb, or lmmse. Capture Sharpening enhances neutral-luminance detail after camera RGB is converted to linear DWG and before the DaVinci Intermediate OETF; the default strength is 0.4, and setting it to 0.0 disables it. The threshold is likewise applied in the linear DWG domain.

The repository's config/compare-amaze.json, config/compare-dcb.json, config/compare-lmmse.json, and config/compare-capture-sharpening.json can be used directly.

cpu_threads is the CPU thread budget for the entire process, while max_parallel_frames is the maximum number of frames in flight. A value of 0 for either selects automatically. Automatic mode reserves at least two threads per frame, processes at most eight frames concurrently, and uses a conservative budget of no more than one-quarter of the currently available physical memory. On low-memory machines, set these values explicitly:

{
  "schema_version": 1,
  "cpu_threads": 4,
  "max_parallel_frames": 1
}

The GPU configuration's backend can be cpu, auto, or vulkan. auto first completes device initialization, a full-size frame context/encoder setup, and a small shader→encoder smoke test; if it fails, the specific reason is written to the sidecar and it falls back to CPU according to fallback. vulkan is forced mode: any preflight or runtime error terminates the current partial file, and the pipeline does not switch to CPU partway through the same MOV. async_depth controls both the GPU slots and the baseline depth of the bounded queues; the currently validated precision mode is fp32, while fp16 is explicitly rejected.

In the GPU sidecar, gpu_resident=true strictly means that the encoder-compatible YUV frame has no CPU upload/readback. rgb_upload_bytes separately reports the FP32 RGB staging traffic still produced by the CPU RAW/color frontend; it must not be interpreted as meaning that the complete RAW pipeline has zero PCIe copies.

--compare-fused runs both the analytic reference path and the LUT/fused path on the specified frames, and reports the number of differing samples and the maximum code-value error across all 10-bit Y/Cb/Cr samples.

Documentation

• See the docs directory for all technical and development documentation.

Current Boundaries

• Only the first DWG/DaVinci Intermediate profile is implemented.
• ProRes packing is fixed to BT.2020 NCL coefficients, video range, and left-sited 4:2:2; primaries/TRC remain unspecified because no standard MOV enumerations exist for DWG/DI, while the sidecar clearly records the actual profile.
• validate uses the official decoder to generate reproducible RAW hashes; a custom/GPU decoder has not yet been added, so there is currently no “official vs. GPU bit-exact” conclusion.
• The transcoder does not perform RAW, chroma, or temporal denoising; noise processing is left to post-production color grading. Bad-pixel correction, lens shading, geometric correction, other Log profiles, and a GUI are outside the v0.1 scope.
• The Vulkan path enters GPU calibration, RCD, camera color, sharpening, DI, YUV, and ProRes processing after CPU official U16 RAW decoding. A specific GPU-AV diagnostic for the FFmpeg ProRes DCT shader has a documented limited waiver; other validation diagnostics still block release.
• mcraw_sample/ and output files are not committed to version control.