Skip to content

v0.6.0 - Fantastic Fennel

Latest
Latest
Compare
Choose a tag to compare
@psalz psalz released this 12 Aug 12:56
· 96 commits to master since this release
v0.6.0
8341c51

This is release includes major overhauls to many of Celerity's core internals, improving performance, debuggability as well as laying the groundwork for future optimizations.

HIGHLIGHTS

  • Celerity now supports SimSYCL, a SYCL implementation focused on debugging and verification (#238).
  • Multiple devices can now be managed by a single Celerity process, which allows for more efficient device-to-device communication (#265).
  • The Celerity runtime can now be configured to log detailed tracing events for the Tracy hybrid profiler (#267).
  • Reductions are now supported across all SYCL implementations (#265).
  • The new experimental::hints::oversubscribe hint can be used to improve computation-communication overlapping (#249).
  • API documentation is now available, generated by 🥬doc.

Changelog

This release includes changes that may require adjustments when upgrading:

  • A single Celerity process can now manage multiple devices.
    This means that on a cluster with 4 GPUs per node, only a single MPI rank needs to be spawned per node.
  • The previous behavior of having a separate process per device is still supported but discouraged, as it incurs additional overhead.
  • It is no longer possible to assign a device to a Celerity process using the CELERITY_DEVICES environment variable.
    Please use vendor-specific mechanisms (such as CUDA_VISIBLE_DEVICES) for limiting the set of visible devices instead.
  • We recommend performing a clean build when updating Celerity so that updated submodule dependencies are properly propagated.

We recommend using the following SYCL versions with this release:

  • DPC++: 89327e0a or newer
  • AdaptiveCpp (formerly hipSYCL): v24.06
  • SimSYCL: master

See our platform support guide for a complete list of all officially supported configurations.

Added

  • Add support for SimSYCL as a SYCL implementation (#238)
  • Extend compiler support to GCC (optionally with sanitizers) and C++20 code bases (#238)
  • celerity::hints::oversubscribe can be passed to a command group to increase split granularity and improve computation-communication overlap (#249)
  • Reductions are now unconditionally supported on all SYCL implementations (#265)
  • Add support for profiling with Tracy, via CELERITY_TRACY_SUPPORT and environment variable CELERITY_TRACY (#267)
  • The active SYCL implementation can now be queried via CELERITY_SYCL_IS_* macros (#277)

Changed

  • All low-level host / device operations such as memory allocations, copies, and kernel launches are now represented in the single Instruction Graph for improved asynchronicity (#249)
  • Celerity can now maintain multiple disjoint backing allocations per buffer, so disjoint accesses to the same buffer do not trigger bounding-box allocations (#249)
  • The previous implicit size limit of 128 GiB on buffer transfers is lifted (#249, #252)
  • Celerity now manages multiple devices per node / MPI rank. This significantly reduces overhead in multi-GPU setups (#265)
  • Runtime lifetime is extended until destruction of the last queue, buffer, or host object (#265)
  • Host object instances are now destroyed from a runtime background thread instead of the application thread (#265)
  • Collective host tasks in the same collective group continue to execute on the same communicator, but not necessarily on the same background thread anymore (#265)
  • Updated the internal libenvpp dependency to 1.4.1 and use its new features (#271)
  • Celerity's compile-time feature flags and options are now written to version.h instead of being passed on the command line (#277)

Fixed

  • Scheduler tracking structures are now garbage-collected after buffers and host objects go out of scope (#246)
  • The previous requirement to order accessors by access mode is lifted (#265)
  • SYCL reductions to which only some Celerity nodes contribute partial results would read uninitialized data (#265)

Removed

  • Celerity does not attempt to spill device allocations to the host if resizing buffers fails due to an out-of-memory condition (#265)
  • The CELERITY_DEVICES environment variable is removed in favor of platform-specific visibility specifiers such as CUDA_VISIBLE_DEVICES (#265)
  • The obsolete experimental::user_benchmarker infrastructure has been removed (#268).
Assets 2
Loading