[AI] GPU acceleration install: manifest-driven scripts

.github/workflows/refresh-ort-gpu.yml

@@ -0,0 +1,31 @@
+name: Refresh ONNX Runtime GPU registry
+
+on:
+  schedule:
+    # 1st of each month at 06:00 UTC. ONNX Runtime ships every 4-6 weeks
+    # and Radeon updates ROCm patches roughly quarterly, so monthly is
+    # frequent enough to catch updates without spamming review
+    - cron: '0 6 1 * *'
+  workflow_dispatch:
+
+permissions:
+  contents: write
+  pull-requests: write
+
+jobs:
+  refresh:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+      - name: Configure git
+        run: |
+          git config user.name 'github-actions[bot]'
+          git config user.email 'github-actions[bot]@users.noreply.github.com'
+      - name: Refresh registry and open PR if changed
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          python3 tools/ai/refresh-ort-gpu.py --update --pr || true

data/CMakeLists.txt

@@ -171,6 +171,8 @@ if(USE_AI)
   #
   FILE(COPY ai_models.json DESTINATION "${DARKTABLE_DATADIR}")
   install(FILES ai_models.json DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/darktable COMPONENT DTApplication)
+  FILE(COPY ort_gpu.json DESTINATION "${DARKTABLE_DATADIR}")
+  install(FILES ort_gpu.json DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/darktable COMPONENT DTApplication)
 endif(USE_AI)
 
 #

data/ort_gpu.json

@@ -0,0 +1,155 @@
+{
+  "version": 1,
+  "install_docs": {
+    "nvidia": "https://developer.nvidia.com/cudnn-downloads",
+    "amd": "https://rocm.docs.amd.com/projects/install-on-linux/en/latest/"
+  },
+  "packages": [
+    {
+      "vendor": "nvidia",
+      "platform": "linux",
+      "arch": "x86_64",
+      "cuda_min": "12.0",
+      "cuda_max": "12.99",
+      "ort_version": "1.25.1",
+      "url": "https://github.com/microsoft/onnxruntime/releases/download/v1.25.1/onnxruntime-linux-x64-gpu-1.25.1.tgz",
+      "sha256": "ddfc4ca4ccc9cd5345d3820edab710ee84e749569d052eed92c42693d3b448a8",
+      "format": "tgz",
+      "lib_pattern": "libonnxruntime",
+      "install_subdir": "onnxruntime-cuda",
+      "size_mb": 250,
+      "requirements": "CUDA 12.x, cuDNN 9.x"
+    },
+    {
+      "vendor": "nvidia",
+      "platform": "linux",
+      "arch": "x86_64",
+      "cuda_min": "13.0",
+      "cuda_max": "13.99",
+      "ort_version": "1.25.1",
+      "url": "https://github.com/microsoft/onnxruntime/releases/download/v1.25.1/onnxruntime-linux-x64-gpu_cuda13-1.25.1.tgz",
+      "sha256": "ebc14e1290db2a30a7bb415bd1c3e1390a7816bb4db87677dc36d071ed22833c",
+      "format": "tgz",
+      "lib_pattern": "libonnxruntime",
+      "install_subdir": "onnxruntime-cuda",
+      "size_mb": 200,
+      "requirements": "CUDA 13.x, cuDNN 9.x"
+    },
+    {
+      "vendor": "nvidia",
+      "platform": "windows",
+      "arch": "x86_64",
+      "cuda_min": "12.0",
+      "cuda_max": "12.99",
+      "ort_version": "1.25.1",
+      "url": "https://github.com/microsoft/onnxruntime/releases/download/v1.25.1/onnxruntime-win-x64-gpu-1.25.1.zip",
+      "sha256": "e72bc4895f84400368382d7b7287c018eabbb2be384e67446f3ab64af7fe5552",
+      "format": "zip",
+      "lib_pattern": "onnxruntime",
+      "install_subdir": "onnxruntime-cuda",
+      "size_mb": 300,
+      "requirements": "CUDA 12.x, cuDNN 9.x"
+    },
+    {
+      "vendor": "nvidia",
+      "platform": "windows",
+      "arch": "x86_64",
+      "cuda_min": "13.0",
+      "cuda_max": "13.99",
+      "ort_version": "1.25.1",
+      "url": "https://github.com/microsoft/onnxruntime/releases/download/v1.25.1/onnxruntime-win-x64-gpu_cuda13-1.25.1.zip",
+      "sha256": "d54b465e5f9300d8928c43772751ac95c123ebd90069a0a4b8ce02e503335469",
+      "format": "zip",
+      "lib_pattern": "onnxruntime",
+      "install_subdir": "onnxruntime-cuda",
+      "size_mb": 300,
+      "requirements": "CUDA 13.x, cuDNN 9.x"
+    },
+    {
+      "vendor": "amd",
+      "platform": "linux",
+      "arch": "x86_64",
+      "rocm_min": "7.0",
+      "rocm_max": "7.0",
+      "ort_version": "1.22.1",
+      "url": "https://repo.radeon.com/rocm/manylinux/rocm-rel-7.0.2/onnxruntime_rocm-1.22.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl",
+      "sha256": "0f9c1555c31cc1bef84504807a0747486017ec4bd9b9f7df0b15e659140a7f7e",
+      "format": "whl",
+      "lib_pattern": "libonnxruntime",
+      "install_subdir": "onnxruntime-migraphx",
+      "size_mb": 200,
+      "requirements": "ROCm 7.0, MIGraphX",
+      "required_libs": [
+        "libmigraphx_c.so.3",
+        "librocm_smi64.so.7"
+      ]
+    },
+    {
+      "vendor": "amd",
+      "platform": "linux",
+      "arch": "x86_64",
+      "rocm_min": "7.1",
+      "rocm_max": "7.1",
+      "ort_version": "1.23.1",
+      "url": "https://repo.radeon.com/rocm/manylinux/rocm-rel-7.1.1/onnxruntime_migraphx-1.23.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl",
+      "sha256": "0bb62f7c2b326b2435396ef90a5f3ad030fe779aa9d0da7013d8b514c27de08f",
+      "format": "whl",
+      "lib_pattern": "libonnxruntime",
+      "install_subdir": "onnxruntime-migraphx",
+      "size_mb": 100,
+      "requirements": "ROCm 7.1, MIGraphX"
+    },
+    {
+      "vendor": "amd",
+      "platform": "linux",
+      "arch": "x86_64",
+      "rocm_min": "7.2",
+      "rocm_max": "7.2",
+      "ort_version": "1.23.2",
+      "url": "https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2.2/onnxruntime_migraphx-1.23.2-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl",
+      "sha256": "76f22bba96991d6653c8169999817203cde72e6703b8211f00dc8dd80ff8e43b",
+      "format": "whl",
+      "lib_pattern": "libonnxruntime",
+      "install_subdir": "onnxruntime-migraphx",
+      "size_mb": 50,
+      "requirements": "ROCm 7.2, MIGraphX"
+    },
+    {
+      "vendor": "intel",
+      "platform": "linux",
+      "arch": "x86_64",
+      "ort_version": "1.24.1",
+      "url": "https://files.pythonhosted.org/packages/50/cf/17ba72de2df0fcba349937d2788f154397bbc2d1a2d67772a97e26f6bc5f/onnxruntime_openvino-1.24.1-cp312-cp312-manylinux_2_28_x86_64.whl",
+      "sha256": "d617fac2f59a6ab5ea59a788c3e1592240a129642519aaeaa774761dfe35150e",
+      "format": "whl",
+      "lib_pattern": "libonnxruntime",
+      "lib_extra_patterns": [
+        "libopenvino",
+        "libtbb"
+      ],
+      "install_subdir": "onnxruntime-openvino",
+      "size_mb": 80,
+      "requirements": "Intel GPU driver (OpenCL)"
+    },
+    {
+      "vendor": "intel",
+      "platform": "windows",
+      "arch": "x86_64",
+      "ort_version": "1.24.1",
+      "url": "https://files.pythonhosted.org/packages/59/37/d301f2c68b19a9485ed5db3047e0fb52478f3e73eb08c7d2a7c61be7cc1c/onnxruntime_openvino-1.24.1-cp312-cp312-win_amd64.whl",
+      "sha256": "f186335a9c9b255633275290da7521d3d4d14c7773fee3127bfa040234d3fa5a",
+      "format": "whl",
+      "lib_pattern": "onnxruntime",
+      "install_subdir": "onnxruntime-openvino",
+      "size_mb": 20,
+      "requirements": "Intel GPU driver",
+      "runtime_url": "https://files.pythonhosted.org/packages/73/cb/07035dcbf60347c0706c4f8d3ffb003c34fe8205e4041699a6655cc858c3/openvino-2026.1.0-21367-cp312-cp312-win_amd64.whl",
+      "runtime_sha256": "21fbc8f9aa765d9ded024311f4e48e38118643cda0f0139dcfa5422c0c457eb6",
+      "runtime_lib_pattern": "openvino",
+      "runtime_extra_patterns": [
+        "tbb"
+      ],
+      "runtime_size_mb": 70
+    }
+  ]
+}

tools/ai/README.md

@@ -0,0 +1,149 @@
+# GPU-Accelerated ONNX Runtime for darktable
+
+darktable bundles a CPU-only ONNX Runtime on Linux, DirectML on Windows,
+and CoreML on macOS. To enable GPU acceleration for AI features (denoise,
+upscale, segmentation), install a GPU-enabled ONNX Runtime build using the
+preferences UI or one of the install scripts in this directory.
+
+## What's bundled by default
+
+| Platform | Bundled ONNX Runtime | GPU support |
+|----------|------------|-------------|
+| Linux | CPU only | None – install GPU ONNX Runtime below |
+| Windows | DirectML | AMD, NVIDIA, Intel via DirectX 12 |
+| macOS | CoreML | Apple Neural Engine |
+
+## Easiest: install from darktable preferences
+
+1. Open darktable preferences (Ctrl+,)
+2. Go to the **AI** tab
+3. Click **install** – darktable detects your GPU and downloads
+   the correct ONNX Runtime package automatically
+4. Restart darktable
+
+Click **detect** instead to find a previously installed or
+system-packaged ONNX Runtime library.
+
+## Installing via script
+
+Linux:
+```bash
+./tools/ai/install-ort-gpu.sh --help    # see all flags
+./tools/ai/install-ort-gpu.sh
+```
+
+Windows (PowerShell):
+```powershell
+.\tools\ai\install-ort-gpu.ps1 -Help    # see all flags
+.\tools\ai\install-ort-gpu.ps1
+```
+
+If Windows blocks the script ("running scripts is disabled on this
+system"), bypass once:
+
+```powershell
+powershell -ExecutionPolicy Bypass -File .\tools\ai\install-ort-gpu.ps1
+```
+
+### Requirements
+
+**NVIDIA (CUDA)** – Pascal-or-newer GPU (compute 6.0+), driver 525+,
+CUDA 12.x or 13.x toolkit, cuDNN 9.x.
+
+**AMD (MIGraphX)** – ROCm-supported GPU (Radeon RX 6000+ / Instinct
+MI100+), ROCm 7.x with MIGraphX. Wheels are manylinux-repaired and
+bundle their own ROCm runtime.
+
+**Intel (OpenVINO)** – Intel iGPU (HD/UHD/Iris Xe) or Arc discrete,
+GPU driver with OpenCL (`intel-opencl-icd`) and/or Level Zero. The
+OpenVINO runtime ships in the package.
+
+### AMD: building from source
+
+If the prebuilt package doesn't work (ABI mismatch, unsupported ROCm
+version), build ONNX Runtime against your installed ROCm:
+
+```bash
+./tools/ai/install-ort-amd-build.sh
+```
+
+Requires cmake 3.26+, gcc/g++, python3, git. Takes 10–20 minutes.
+
+## Enabling the custom ONNX Runtime in darktable
+
+After running the script or built-in installer:
+
+1. Open darktable preferences (Ctrl+,)
+2. Go to the **AI** tab
+3. Click **detect**, or use the browse button to select the library
+   manually
+4. Restart darktable
+
+Or set `DT_ORT_LIBRARY` in the environment:
+
+```bash
+# Linux
+DT_ORT_LIBRARY=~/.local/lib/onnxruntime-cuda/libonnxruntime.so.1.24.4 darktable
+```
+```powershell
+# Windows
+$env:DT_ORT_LIBRARY="$env:LOCALAPPDATA\onnxruntime-cuda\onnxruntime.dll"; darktable
+```
+
+If neither preference nor env var is set, darktable uses the bundled
+ONNX Runtime (CPU on Linux, DirectML on Windows, CoreML on macOS).
+
+## Verifying
+
+```bash
+darktable -d ai
+```
+
+Look for:
+```
+[darktable_ai] loaded ORT 1.24.4 from '/home/user/.local/lib/onnxruntime-cuda/libonnxruntime.so.1.24.4'
+[darktable_ai] execution provider: CUDA
+[darktable_ai] NVIDIA CUDA enabled successfully on device 0: NVIDIA GeForce RTX 4090
+```
+
+## Maintaining the GPU package registry
+
+`data/ort_gpu.json` lists the upstream ONNX Runtime URLs the install scripts and
+preferences UI pull from. It needs to be refreshed whenever Microsoft
+ships a new ONNX Runtime release or AMD ships a new ROCm patch. Use the
+script in this directory:
+
+```bash
+# show what would change (no writes); exits non-zero if updates exist
+python3 tools/ai/refresh-ort-gpu.py --check
+
+# apply the updates in place
+python3 tools/ai/refresh-ort-gpu.py --update
+
+# apply and open a PR via gh CLI (CI mode; needs GITHUB_TOKEN)
+python3 tools/ai/refresh-ort-gpu.py --update --pr
+
+# verbose progress (otherwise quiet by default in --update mode)
+python3 tools/ai/refresh-ort-gpu.py --check -v
+```
+
+What it does:
+
+- Queries `api.github.com/repos/microsoft/onnxruntime/releases` for the
+  latest non-prerelease that has all four expected GPU assets
+  (linux/windows × CUDA 12/13). Updates the four NVIDIA entries.
+- Scrapes `https://repo.radeon.com/rocm/manylinux/`, finds the cp312
+  ONNX Runtime wheel in each `rocm-rel-X.Y.Z/` directory, and keeps the
+  latest patch per ROCm minor. Updates the AMD entries with
+  range-based matching (`rocm_min: "7.2"`, `rocm_max: "7.3"` covers
+  every 7.2.x patch).
+- Computes SHA256 only for wheels whose URL changed since the last
+  refresh — a no-op run does no downloads.
+- Preserves vendors it doesn't manage (e.g. Intel/OpenVINO) and any
+  manual fields (`required_libs`, `lib_pattern`, `install_subdir`).
+
+Stdlib only — no extra Python deps. Network access required.
+
+A weekly CI job (`.github/workflows/refresh-ort-gpu.yml`) runs the
+script in `--update --pr` mode every Monday and opens a PR if anything
+upstream moved. Maintainer reviews and merges; nothing is auto-merged.