Repository navigation
Add a device-less backend - #66
Merged
Merged
Conversation
Every backend instantiates the member templates of device::Algorithms explicitly, and each of them spelled out its own list of 36 instantiations. algorithms/Instantiations.h now holds the type lists together with one instantiation macro per member template; the CUDA/HIP and the SYCL backend use them. A backend that is added later thus exports the same symbols by construction. No functional change: compiled against Algorithms.h, the explicit instantiations and the macros emit the same 36 symbols for both backends. AI-generated. Model: Claude Opus 5.5
DeviceInstance::instance() returns the singleton; api() and algorithms() return references to the API and the algorithms of the instance. instance() is defined in device.cpp. All shared objects of a process thus see the same instance, also those compiled with hidden visibility, as Python modules built by pybind11 are. api() and algorithms() are const members that still give mutable access, as a pointer member does; call sites that hold a const reference to the instance keep working. This changes the interface for all consumers: getInstance() becomes instance(), api-> becomes api(). and algorithms. becomes algorithms(). The tests and the examples are migrated accordingly. AI-generated. Model: Claude Opus 5.5
With DEVICE_BACKEND=none, the device library implements its API on the host; code that uses the API thus builds and runs without any GPU toolchain or GPU. DEVICE_ARCH is not needed for it. The host backend has exactly one device, the host itself. Device memory is host memory, aligned to 128 bytes as on the GPU backends, with the same bookkeeping. All work runs synchronously on the calling thread: streams and events only serve as handles, host functions run right away, and streamWaitMemory makes the caller wait. Graph capturing is not available. The algorithms are plain loops with the semantics of the GPU kernels. The complete test suite passes on it, also under AddressSanitizer and UndefinedBehaviorSanitizer, and the library only depends on the C and C++ runtime libraries. AI-generated. Model: Claude Opus 5.5
The common test suite does not cover events, streamWaitMemory, the order of several host functions on one stream, the alignment and the bookkeeping of allocations, or copies with a pitch that differs from the width. tests/host.cpp checks these for the host backend, together with its properties and the neutral elements of the reductions; it is only built with DEVICE_BACKEND=none. Both a reduction that starts the maximum at the smallest positive value and a 2D copy that ignores the pitches make these tests fail, while the common suite still passes. AI-generated. Model: Claude Opus 5.5
Adds DEVICE_BACKEND=none to the build matrix. The job needs no GPU; it runs on a GitHub-hosted runner in the CPU image that SeisSol uses as well, and it runs the complete test suite, including the tests of the host backend. AI-generated. Model: Claude Opus 5.5
The basic example used the real type without including common.h, and it called getMaxSharedMemSize, getMaxThreadBlockSize and checkOffloading, none of which the API provides. It now copies initialized data to the device, within the device and back, and fails if the result differs. The CI job of the host backend builds and runs it. AI-generated. Model: Claude Opus 5.5
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
1 participant
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.
To simplify compiling for CPU and GPU.
Supersedes #48 .