Repository navigation
CUDA: CPU fallback for GPUs the build has no code for (#1315), plus fixes from #1316-#1321 - #1324
Conversation
Release binaries built with the default CMAKE_CUDA_ARCHITECTURES=native carry native code only for the GPU of the build machine (v2.4.0: sm_89, no PTX), so on any other GPU the first kernel or __constant__ access failed mid-run with "named symbol not found (code 500)" and the application exited. - CMake resolves the architectures the CUDA code is compiled for (native, all, all-major or an explicit list) and passes them to initDevices(), which skips a device it cannot run on: native code needs the same major version and an equal or lower minor one, PTX an equal or lower version. When no device is left the log says why and how to rebuild, and densification continues on the CPU. - OpenMVS_MAX_CUDA_COMPATIBILITY is removed: distributed binaries pass the standard -DCMAKE_CUDA_ARCHITECTURES=all-major (or set CUDAARCHS) instead. The SiftGPU port defaults to all-major, as vcpkg builds it once for every machine the binaries run on. - No usable CUDA device is an expected condition, not an error: initDevices() asks the runtime for the device count instead of the asserting cuInit() (the runtime initializes the driver), and AllocateCudaPool() initializes the devices itself instead of probing through a PatchMatch constructor that then asserts a device exists. - CUDA_CHECK() reports through the log, also from .cu files (their fprintf to stderr never reached the log, so develop exited silently), with the location and the expression of the failed call instead of the header's. The driver-API reporter initializes the strings cuGetErrorName/String leave untouched for an unknown code. - PatchMatch keeps its cameras in __constant__ byte storage in every build (Release declared a Camera array, which nvcc 13.4 + GCC 16 reject as dynamic initialization); the SASS only differs in the geometric-consistency kernels, and the last geometric pass is faster (fountain-P11, full resolution, 6 interleaved runs: median 3.13 s vs 3.77 s, other passes equal). - Remove the helpers left from the runtime-loaded PTX modules: the arch hint block, ensureCudaResult, align, MemDevicePtr/MemDeviceArr, stream smart pointers and the unused ArrayRT accessors and typedefs. Includes the fixes proposed in #1316. Co-authored-by: Connor Baker <ConnorBaker01@gmail.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
UseImage() loads a depth-map, then ejects the least recently used ones until the cache fits its budget; with a budget smaller than that depth-map the ejection reached the one just loaded, so the caller got a released map (a crash in fusion on 2.4.0, a silently skipped reference since the IsEmpty() check). The ejection now stops at the requested depth-map, which stays loaded even if it alone exceeds the budget, and UseImage() asserts it on return, so every caller is safe without protecting the image beforehand. Adds DMapCacheKeepsRequestedImageTest. Supersedes #1321, which reported it. Co-authored-by: Tobias Fischer <info@tobiasfischer.info> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With Clang the executable linker flags were replaced by "-stdlib=libc++", dropping any the user or the toolchain had set (ex. conda's), so two C++ runtimes could end up in the same process. The flag is redundant there: it is already in CMAKE_CXX_FLAGS, which CMake passes on every C++ link. Supersedes #1318, which reported it. Co-authored-by: Tobias Fischer <tobias.fischer@qut.edu.au> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- OpenCV 5 adds the unsigned depths CV_32U and CV_64U and defines cv::DataType<unsigned> and cv::DataType<uint64_t> itself; the OpenMVS specializations (mapping them onto CV_32S and CV_64F) are now defined only for OpenCV < 5. - OpenCV 5 moved DistanceTypes (cv::DIST_L2 used by ImportROMA2) into the new geometry module, which opencv2/opencv.hpp does not include. Checked against the OpenCV 5.0.0 headers (core/traits.hpp, core/hal/interface.h, geometry/2d.hpp, opencv.hpp); nothing changes for OpenCV 4. Note that OpenCV 5 widens CV_CN_SHIFT from 3 to 5, so the type code of a multi-channel cv::Mat differs between the two versions; the only untyped cv::Mat OpenMVS archives are the single-channel SfM descriptors, whose codes are the same in both, so the scene files stay compatible. Supersedes #1317. Co-authored-by: Tobias Fischer <tobias.fischer@qut.edu.au> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A project using an installed OpenMVS (build/PackageTest, now built by the CI against the install tree) failed at every step: - `cmake --install --prefix` was ignored for the libraries: the install directories were made absolute from CMAKE_INSTALL_PREFIX at configure time. They are now relative to the prefix, resolved at install time. - find_package(OpenMVS) did not find the config on Windows, installed in <prefix>/CMake/OpenMVS; it is now in lib/cmake/OpenMVS on every platform, and OpenMVSConfig.cmake computes its paths from its own location. - The config looked up none of the packages the exported targets link against. Each such lookup now goes through find_exported_dependency(), which records it (with the <Package>_ROOT hint given to the build) for the config, so the consumer finds exactly what OpenMVS was built with, whatever the options, instead of a hand-maintained list that drifts; the gflags binding the ceres target needs is recorded the same way. PNG, JPEG and TIFF are linked through their imported targets instead of absolute paths. - The exported targets carried no include directory, did not require the C++ standard the headers were compiled with (the BS::thread_pool layout depends on it) and did not define OPENMVS_SHARED, which the headers need to import the symbols of the shared libraries; Common now exports all three. - The headers were incomplete: *.hpp (BS_thread_pool.hpp, json.hpp) and the MVS/CUDA headers were not installed. - The Tests helper is no longer exported: installers leaving it out made the imported targets fail to load. OpenMVS_LIBRARIES and OpenMVS_BINARIES now name the namespaced targets. Supersedes #1319, which reported the missing dependencies and include directory. Co-authored-by: Tobias Fischer <tobias.fischer@qut.edu.au> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
🟡 Changes recommended
Installed targets omit required libc++ flags, risking consumer ABI and link failures, while stripped CUDA architecture qualifiers can admit incompatible GPUs and cause aborts instead of CPU fallback.
2 open issues
What changed in this PR
This PR improves OpenMVS portability and reliability by adding CPU fallback for unsupported CUDA devices and addressing the build, packaging, and depth-map cache issues reported in #1315–#1321.
Changes:
- Detect CUDA architecture compatibility before device selection, improve error logging, and simplify camera constant storage.
- Keep requested depth maps loaded under cache pressure, with a regression test.
- Add OpenCV 5 compatibility, preserve linker flags, and improve installed CMake packages with a CI consumer build.
| File | Description |
|---|---|
ports/siftgpu/source/CMakeLists.txt |
Default distributed SiftGPU builds to all-major. |
libs/SFM/CMakeLists.txt |
Record exported SfM dependencies. |
libs/MVS/SceneDensify.h |
Update CUDA pool documentation. |
libs/MVS/SceneDensify.cpp |
Initialize CUDA before pool creation and log CPU fallback. |
libs/MVS/PatchMatchCUDA.cu |
Use aligned byte storage for constant cameras in every build. |
libs/MVS/PatchMatchCUDA.cpp |
Require CUDA initialization before construction. |
libs/MVS/DMapCache.h |
Export the cache class and declare eviction protection. |
libs/MVS/DMapCache.cpp |
Prevent eviction of the requested depth map. |
libs/MVS/CMakeLists.txt |
Record dependencies and install CUDA headers. |
libs/Math/CMakeLists.txt |
Record optional SuiteSparse dependencies. |
libs/IO/CMakeLists.txt |
Use imported image-library targets and install .hpp headers. |
libs/Common/UtilCUDADevice.h |
Route runtime errors through centralized logging. |
libs/Common/UtilCUDA.h |
Initialize error strings and remove unused helpers. |
libs/Common/UtilCUDA.cpp |
Filter incompatible GPUs and handle unavailable devices. |
libs/Common/Types.inl |
Avoid conflicting OpenCV 5 type specializations. |
libs/Common/Types.h |
Include OpenCV 5 geometry definitions. |
libs/Common/CMakeLists.txt |
Export consumer requirements and embed CUDA architecture metadata. |
docs/wiki/Building.md |
Document CUDA targeting and installed-package usage. |
CMakeLists.txt |
Resolve CUDA targets and generate relocatable package configuration. |
build/Utils.cmake |
Preserve linker flags and record exported dependencies. |
build/Templates/OpenMVSConfig.cmake.in |
Resolve dependencies and expose namespaced targets. |
build/PackageTest/main.cpp |
Exercise installed MVS and SfM libraries. |
build/PackageTest/CMakeLists.txt |
Add a minimal installed-package consumer. |
apps/Tests/TestsMVS.h |
Declare the cache regression test. |
apps/Tests/TestsMVS.cpp |
Test requested-map retention under cache pressure. |
apps/Tests/Tests.cpp |
Register the cache regression test. |
apps/Tests/CMakeLists.txt |
Stop exporting the test executable. |
.github/workflows/continuous_integration.yml |
Build the package consumer against the install tree. |
Review effort: Balanced
| # BS::thread_pool layout) and on the library type: tell them we are building shared | ||
| # libraries, so they pick dllimport and the cross-DLL-friendly path for things like | ||
| # boost exception-unwinding (see Types.h) | ||
| TARGET_COMPILE_FEATURES(Common PUBLIC cxx_std_${CMAKE_CXX_STANDARD}) |
There was a problem hiding this comment.
Fixed in 56495bf: when built with Clang, the installed Common target now carries -stdlib=libc++ as an INSTALL_INTERFACE compile option (C++ only) and link option, so the projects using the package pick the same standard library. The build tree already gets it from CMAKE_CXX_FLAGS.
| if(NOT CMAKE_MATCH_3 STREQUAL "-virtual") | ||
| LIST(APPEND _CUDA_ARCHS_SASS ${CMAKE_MATCH_1}) |
There was a problem hiding this comment.
Fixed in 40df18d: architecture-specific targets (ex. 90a-real) now go to a separate _CUDA_ARCHS_SASS_EXACT list and _isComputeCapabilitySupported() accepts them only for the exact compute capability. Family-specific f targets keep the same-major, equal-or-higher-minor rule, and suffixed PTX was already excluded from the forward-compatible PTX.
An `a` target (ex. 90a-real) was recorded as generic native code, so initDevices() accepted any GPU of the same major version with an equal or higher minor, which cannot load it. Architecture-specific SASS is now listed separately and matched exactly; family-specific (`f`) targets keep the generic same-major rule. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PYTHON_INSTALL_PATH was never set, so `cmake --install` tried to write the bindings to /openmvs. It now defaults to the site-packages folder of a Python installation or virtual environment, relative to the prefix. Clang builds use libc++, so the installed Common target now tells the projects using OpenMVS to compile and link against it as well. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
libs/IO/json.hpp was nlohmann JSON 3.5.0 from 2018, while vcpkg already builds 3.12.0 for tinygltf. IO now links nlohmann_json::nlohmann_json (recorded as a dependency of the installed package) and the bundled copy is removed, so package managers use their own copy with no option to set. Co-Authored-By: Tobias Fischer <info@tobiasfischer.info> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
GLAD 1 is unmaintained and missing on some platforms (ex. conda-forge linux-aarch64), while vcpkg ships only GLAD 1. The Viewer now prefers a GLAD 2 CMake package (glad_add_library, gl:compatibility=4.6 so the GL 4.3 debug output stays available) and falls back to GLAD 1. Common.h is the only place choosing the header and the loader call; the other Viewer headers get OpenGL through it. Co-Authored-By: Tobias Fischer <info@tobiasfischer.info> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The install(CODE) script put execute_process() and if() on one line, which CMake rejects, so `cmake --install` failed on every Linux build with the Viewer. It also ran in the configure-time prefix; it now uses the prefix and DESTDIR given at install time, and belongs to the desktop component. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tall prefix The .desktop Exec path was configured with the configure-time CMAKE_INSTALL_PREFIX, so installing with --prefix left the menu entry pointing to a binary that is not there. The file is now configured at install time, with the prefix given then. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
bash drops the backslashes of an unquoted Windows path, so the package test could not find the vcpkg toolchain on the Windows runner; pass the paths through quoted shell variables instead. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The user's request of 2026-11-05: "next merge latest remote develop, remove all chached results / features / matches, etc, and rerun all test datasets as there were made improvements to SiftGPU" (T/briefs-20261017/develop-merge-20261106.md). develop brings 17 commits since 9fabec1: SiftGPU parallel decoding, weak-texture detector, moment orientation and adaptive grid selection (#1320), the CUDA CPU fallback (#1324), MVS refine/texture/split and mesh algorithms moved to halfmesh (#1314, #1299, #1300, #1302), viewer (#1310), densify (#1309, #1298), SGM (#1306, #1305), scene split (#1304), the MSVC build split of Scene.cpp (#1303), InterfacePolycam output paths (#1245), TIFF write (#1246), and C++17 minimum with vcpkg tinyxml2/nlohmann-json (#1301). Conflicts (one line each in T/records-2026-11-06-merge/resolution.md): - vcpkg.json: tinyxml2 added; tinynpy stays out (this branch removed the NPZ import at 1380970). - libs/SFM/CMakeLists.txt: develop's find_exported_dependency for TinyEXIF, PoseLib, Ceres; no TinyNPY. - libs/SFM/StarInitializer.cpp: this branch's ComputePairScales kept (develop edited the inline trimmed mean that ComputePairScales replaced; same statistic, so the result is unchanged). - libs/SFM/FeaturesExtractor.cpp: develop's adaptive SiftGPU selection (SelectFeatures), followed by this branch's ReleaseFeatures (clears the dense-keypoint boundary). - libs/Common/List.h: this branch's std::vector constructor without the removed _SUPPORT_CPP11 guard; both cListTest additions. - apps/CreateStructure/CreateStructure.cpp: develop's --siftgpu-darkness-adaptivity and --siftgpu-cuda options with this branch's --match-mode 4 / --match-sequence-overlap text; --siftgpu-cuda gates SiftGPU only, and RoMa2's GPU switch stays as it was. - apps/InterfacePolycam/InterfacePolycam.cpp: develop's output-path handling (a superset of this branch's ensureValidPath fix). - apps/Tests/Tests.cpp, TestsMVS.cpp: both sides' tests (PlatformArrGrowOwnershipTest and develop's MeshRefineStep, window-stats, clean-per-vertex and image-resize tests). - docs/design/HierarchicalSFM.md: this branch's content placed in develop's 7-section design-record layout. - docs/design/SphericalCameraSupport.md: develop's rewrite. - docs/pipeline-traces.md: deleted, as develop did (this branch's edits to it are already in docs/pipelines.md). Clean merges that needed a fix because develop removed a definition this branch's new code uses: MAYBEUNUSED -> [[maybe_unused]] (libs/SFM/SceneBlocks.cpp); ../IO/json.hpp -> <nlohmann/json.hpp> (libs/SFM/RoMa2Matcher.cpp, apps/Tests/TestsSFM.cpp). Verification: Release build (Ninja Multi-Config; the reconfigure installed tinyxml2, nlohmann-json, halfmesh 0.5.1 and develop's siftgpu port) has no warning in libs/SFM (40 of 40 SFM objects rebuilt; the 4 warnings in the build are develop's -Wclass-memaccess in libs/MVS/MeshHalfMesh.cpp). Release ctest passes 3 of 3 (SFMPipelineTest 313.7 s). Debug build with ASSERTs on (a dev copy of the merge tree with a temporary test dispatcher, never committed; T/records-2026-11-06-merge/debug-tests.txt): 39 of 40 selected SFM tests pass with no ASSERT (StarEvidence, StarReferenceView, TripletStarInit, NearDuplicateYardstick, PairScales, DenseKeypointBoundary, FundamentalPoseTwoCameras, the ROMA2/RoMa2, spherical, CubeMap and SceneBlocks tests, ...), and so do the unit tests (cListTest, PlatformArrGrowOwnership, develop's MVS tests). The one failure, ROMA2ReconstructTest, ran out of GPU memory while another job held 35 GB of the GPU; its rerun and the smoke run are in T/records-2026-11-06-merge/smoke.log. The SceneAnalyzeSFM binary, run against the merged libraries, writes byte-identical CSVs for b68e320e dup-20261105-base. Snapshot T6/bin-merge-4eaec0b-develop. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>


Fixes #1315. Also supersedes #1316, #1317, #1318, #1319, #1321, #1322 and #1323, which reported the problems fixed here; thanks to @ConnorBaker and @Tobias-Fischer, who are credited as co-authors of the matching commits.
There is one commit per problem.
CUDA: use the CPU when the build has no code for the GPU (#1315, #1316)
Cause. The v2.4.0 Windows CUDA binaries contain native code only for
sm_89, the build machine's GPU, and no PTX (cuobjdump). The default becameCMAKE_CUDA_ARCHITECTURES=nativein 2.4.0, and it was75in 2.3.0. On any other GPU, the first__constant__upload fails withnamed symbol not found (code 500)and the application exits.CUDA_FORCE_PTX_JIT=1reproduces it exactly on an RTX 4070. Ondevelopthe same run exits silently: errors raised in.cufiles went tofprintf(stderr), which never reaches the log.Fix.
native,all,all-majoror a list).initDevices()skips a device the build cannot run on, logs the rebuild command, and densification continues on the CPU.OpenMVS_MAX_CUDA_COMPATIBILITYis removed. Distributed binaries pass-DCMAKE_CUDA_ARCHITECTURES=all-majorinstead. The SiftGPU port defaults toall-major.AllocateCudaPool()initializes the devices itself instead of probing through thePatchMatchconstructor.CUDA_CHECKlogs the caller's file, line and expression to the log file.PatchMatchstores its cameras in__constant__byte storage in every build. Release used aCameraarray, which nvcc 13.4 + GCC 16 reject. This was checked for speed: only the geometric-consistency kernels' SASS changes, and the last geometric pass got faster (fountain-P11, full resolution, 6 interleaved runs: median 3.13 s vs 3.77 s).Verified on an RTX 4070:
sm_89buildsm_75-only buildCUDA device 0 (NVIDIA GeForce RTX 4070) has compute capability 8.9, but this build contains CUDA code only for sm_75; rebuild with -DCMAKE_CUDA_ARCHITECTURES=89 (or all-major), then the CPU densifies (exit 0)CUDA_VISIBLE_DEVICES=-1CUDA_FORCE_PTX_JIT=1PatchMatchCUDA.cu:715: cudaMemcpyToSymbolAsync(...)Tests.exe0, 1 and 2 pass, test 2 also with no device.Review follow-up. An architecture-specific target (ex.
90a-real) now matches only its exact compute capability; family-specificftargets keep the same-major rule.densify:
DMapCache::UseImage()never ejects the depth-map it provides (#1321)The ejection now stops at the requested depth-map, and
UseImage()asserts it on return. This makes every caller safe without having to protect the image before loading it. The PR addsDMapCacheKeepsRequestedImageTest.build: stop overwriting
CMAKE_EXE_LINKER_FLAGSwith-stdlib=libc++(#1318)The line is deleted: the flag is already in
CMAKE_CXX_FLAGS, which CMake passes on every C++ link.common: support OpenCV 5 (#1317)
These are #1317's changes, checked against the OpenCV 5.0.0 headers. OpenCV 5 also widens
CV_CN_SHIFTfrom 3 to 5, so multi-channelcv::Mattype codes differ between OpenCV 4 and 5. The only untypedcv::MatOpenMVS archives are the single-channel SfM descriptors, whose codes are the same in both versions, so scene files stay compatible.build: make the installed CMake package usable and relocatable (#1319)
build/PackageTestis a minimal project using an installed OpenMVS, and CI now builds it against the install tree on all three platforms. It hit these problems, all fixed:--prefixwas ignored for the libraries: install directories are now relative.The config was not found on Windows: it now installs to
lib/cmake/OpenMVSeverywhere.No dependency was looked up. Each exported lookup now goes through
find_exported_dependency(), which records it for the config, so the list cannot drift from the build.The targets were missing the include directory, the C++ standard and
OPENMVS_SHARED.*.hppandMVS/CUDAheaders were not installed.The
Testshelper was exported.The Python bindings were installed to
/openmvs: they now go to the prefix'ssite-packages(Lib/site-packageson Windows,lib/pythonX.Y/site-packageselsewhere).Clang builds use libc++: the installed
Commontarget now passes-stdlib=libc++to the projects using it.Verified locally: a consumer built against
cmake --install --prefix <scratch>loads a scene.build: use the nlohmann-json package instead of a bundled copy (#1323)
libs/IO/json.hppwas nlohmann JSON 3.5.0 (2018), while vcpkg already builds 3.12.0 for tinygltf. IO now linksnlohmann_json::nlohmann_json, also recorded in the installed package, and the bundled copy is deleted. Package managers such as conda-forge use their own copy with no option to set, and only one JSON version is ever compiled in.viewer: support GLAD 2 alongside GLAD 1 (#1322)
GLAD 1 is unmaintained and missing on some platforms (ex. conda-forge linux-aarch64), but vcpkg ships only GLAD 1, so both are supported. A GLAD 2 CMake package (
glad_add_library,gl:compatibility=4.6so the GL 4.3 debug output stays available) is preferred when found.Common.his the only place choosing the header and the loader call; the other Viewer headers get OpenGL through it. The GLAD 1 path is built and started locally and in CI; the GLAD 2 path follows #1322, which was tested in a stand-alone project.viewer: Linux install fixes
update-mime-databaseinstall step brokecmake_install.cmake(missing newline between commands); it now also runs in$DESTDIRand the prefix given at install time..desktopentry'sExecpath is configured at install time, so it points at the binary under the actual--prefix.ci
🤖 Generated with Claude Code