Repository navigation
Conversation
This was referenced Sep 10, 2026
This comment was marked as resolved.
This comment was marked as resolved.
giordano
force-pushed
the
mg/ka-0.10
branch
from
September 12, 2026 22:51
a263c5e to
9e35af3
Compare
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
giordano
force-pushed
the
mg/ka-0.10
branch
3 times, most recently
from
October 4, 2026 12:39
0fe1598 to
688ba4b
Compare
giordano
commented
Oct 6, 2026
giordano
commented
Oct 6, 2026
giordano
commented
Oct 6, 2026
| @inline kernel_time_step(arch, grid, Δt) = convert(eltype(grid), Δt) | ||
|
|
||
| using Adapt: Adapt | ||
| using Base: @pure |
This was referenced Oct 11, 2026
giordano
changed the title
giordano
force-pushed
the
mg/ka-0.10
branch
7 times, most recently
from
October 16, 2026 19:01
beb2700 to
a300fa2
Compare
giordano
force-pushed
the
mg/ka-0.10
branch
3 times, most recently
from
October 17, 2026 18:59
e4d2d73 to
08d1641
Compare
…ange mapping The active cells map was implemented by wrapping the kernel function in a `MappedFunction` and overriding KernelAbstractions 0.9 internals (`__validindex`, `expand` with the map stored in place of the workitems), plus `@device_override`s of `__validindex` in the CUDA, AMDGPU and Metal extensions. KernelAbstractions 0.10 validates work items differently, as `expand(iterspace, group, item) in ndrange` for iteration spaces with a custom mapping, so every kernel launched over a map failed to compile on the POCL CPU backend (a `MethodError` in `in`). The adaptation of the context also dropped its new launch configuration. KernelAbstractions 0.10 supports custom mappings in `NDRange` through `expand`, `in` and `linear_index`, so the map now uses that hook: `launch!` passes an `IndexMap` as the `ndrange` of the kernel, `partition` stores it as the `NDRange` mapping, `expand` looks up the index of each work item, and the context's `ndrange` is a `MappedIndices` containing every index but the one returned past the end of the map. The linear index of a mapped work item, its position in the map, is defined through `NDIteration.linear_index(iterspace, ndrange, groupidx, idx)` (JuliaGPU/KernelAbstractions.jl#848), KernelAbstractions' hook for custom `NDRange` mappings, rather than by extending the internal `__index_Global_Linear`, which KernelAbstractions documents as not to be extended. KernelAbstractions adapts the mapping and keeps the launch configuration itself, and the backend overrides are no longer needed. All the new methods are on Oceananigans types. The split-explicit substep loops, including the multi-region one, configure their kernels once with `configure_kernel` and then launch them directly, many times. They discarded the work size it returns and launched without an `ndrange`, which with a kernel configured over a map fails ("Can not partition kernel!"). They now keep the work size and launch with `ndrange = launch_ndrange(worksize)`, which is `nothing` for the launches without a map and the `IndexMap` with one. There was no direct test of launches over a map, so add one checking that each listed cell is visited exactly once over several workgroups, unlisted cells aren't touched, and the linear index is the position in the map. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EtLmZUPY4g9hi8aJsM4Wrs Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PrdqyLiieQJGV868Vz4zvt
`convert_to_device(::CPU, args)` adapted kernel arguments to `CPU()`, so that fields reached CPU kernels as their data and kernels did not specialize on boundary conditions (#6068). With KernelAbstractions 0.10 CPU kernels run on the POCL backend, which converts every argument with its own adaptor when launching and compiles the kernel for the converted types, as on GPUs: kernels no longer specialize on field metadata whether or not the arguments are converted beforehand. Converting them beforehand also adapted them twice, and objects already adapted to `CPU()` are not necessarily adaptable again: a `FieldTimeSeries` became a `GPUAdaptedFieldTimeSeries`, which KernelAbstractions passed through with its CPU `Array`, so every kernel taking a `FieldTimeSeries` (in forcings, boundary conditions, `PrescribedVelocityFields`, ...) failed to compile. Converting them with the POCL adaptor instead, as GPUs do with `cudaconvert`, isn't an option: it turns arrays into raw pointers, and only the arguments passed to the launch are preserved from garbage collection while the kernel runs. So `convert_to_device` is now the identity on the CPU, and the `CPU()` adaptation and its pass-throughs for grids and functions are removed. The test now checks that the argument types KernelAbstractions compiles CPU kernels for don't depend on boundary conditions, and that converting the arguments, which happens at every launch, doesn't allocate. On a split-explicit hydrostatic model with CATKE on an immersed grid (32×32×8), allocations (19520 B) and run time (~50 ms) per time step are unchanged, while the first time step takes ~3.5 s longer (18.4–21.9 s against 15.0–16.9 s, alternating runs on a loaded machine), as the host side of the launches is compiled for the unconverted argument types. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EtLmZUPY4g9hi8aJsM4Wrs
In KernelAbstractions 0.10, `ndevices` and `zeros` are owned by KernelInterface, a separate package that KernelAbstractions loads and doesn't re-export them from, so accessing them as `KernelAbstractions.ndevices` and `KernelAbstractions.zeros` failed the ExplicitImports check that qualified accesses go through the owner. Add KernelInterface as a dependency, from the same KernelAbstractions revision as in `test/Project.toml`, and use `KernelInterface.ndevices` and `KernelInterface.zeros`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EtLmZUPY4g9hi8aJsM4Wrs
With KernelAbstractions 0.10, CPU kernels are launched on the POCL backend, and every launch allocates, so a time step on the CPU allocates more in the allocation regression tests. On the upside, newer CUDA.jl versions reduced allocations. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EtLmZUPY4g9hi8aJsM4Wrs
Kernels launched over offset index ranges (`KernelParameters`, and the interior work layouts that exclude peripheral nodes) used a custom `OffsetStaticSize` ndrange, a subtype of KernelAbstractions' internal `_Size`. To make KernelAbstractions launch over it, Oceananigans extended the internal `__ndrange` and `__groupsize` for its contexts, stored the offsets in the `workitems` slot of the `NDRange`, and replaced `partition` and `expand` for such kernels. KernelAbstractions 0.10 supports index ranges in a static ndrange: `StaticSize((2:5, 1:3))` launches over those indices, with the offsets as the `NDRange` mapping. Build the work sizes with `StaticSize` of the ranges instead, and remove `OffsetStaticSize` with the methods extending KernelAbstractions' internals, and the imports only they used. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EtLmZUPY4g9hi8aJsM4Wrs
This branch has not been deployed
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
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.
No description provided.