Improving linking support for ROCM and ukernels. #19211
Merged
Conversation
This file contains 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
This is already present in CUDATarget/LLVMCPUTarget.
benvanik
added
compiler/dialects
bjacob
approved these changes
Nov 19, 2024
Plugins and other dialects can define serializable attributes. For
example an `#iree_codegen.embedded_bitcode<"some_name.bc">` could
reference an embedded file via `c_embed_data`. The attributes could be
arbitrarily complex, e.g. `#cutlass.kernel<{args}>` that on-demand
generates an external kernel function to link in at compile-time.
This runs nested on variants to find all `hal.executable.objects` attrs nested in the inner module and move them to the parent `hal.executable.variant`. This allows codegen/plugin/etc passes running on executable contents to declare an object they want to include by making only local changes (such as in a pattern rewriter) and then letting the pass move them to the variant where they belong. This only handles arrays of objects as expected after `MaterializeInterfacesPass` runs - target object dictionaries are not very easy to merge and we generally want to run after executable translation/linking anyway where they have already been baked out.
benvanik
force-pushed
the
users/benvanik/rocm-linking
branch
from
bc0d083
to
f510664
Compare
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Outdated suggestions cannot be applied.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.
To support externally-defined ukernels on ROCM the ROCMTarget has been brought in-line with LLVMCPU/CUDA by calling
linkBitcodeObjects. To make authoring passes that include object references#hal.executable.objectnow allows any data type to be associated so long as it is serializable allowing for external resource attrs and other custom attributes that may serialize based on other information. To allow patterns to attach object references all ops within an executable variant can now declare ahal.executable.objectsarray that will be hoisted and merged into the top-level variant objects after our executable linking pass (before serialization where they are used).