tree 87300461be652460866e2c27f7618efd451387fb
parent a937cf9adc8886732970a849491bc35f6341f991
author Benoit Jacob <jacob.benoit.1@gmail.com> 1781612803 -0400
committer GitHub <noreply@github.com> 1781612803 +0200
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJqMUEDCRC1aQ7uu5UhlAAAskgQADj6OSwPdlbyHziS5kqthHth
 B6fO8CZKmnKzbDnlnkMmzVgMRSBExhHCFRH0c1DHrPV52+0YskMq5nN4bIt5GXPC
 XAuWkVJToMnQpflyTXw81UzHXgYDbb2vTDH9k6ACqqo5C6YifVH81ABvP7m2uIwv
 J1Qjrmn9BCEC6hnROZW4zUUhfV1koaS1FJlXszwGVSN/RXUEBc2pzA/PyZ3t7tJX
 ysMWe9donJHctMvVTLs5GKbFjWQGpObPXpzfR5IFhe13KE67J6VPUlfMyTI8TN9P
 BzA9muGgolsuTQ0e2+Ap/0jIHb2f0nQqLjQbSzRARJ/y4KzaNXSynMAKXX28Ogka
 Wt9Aao7SUAsj0vRTe+aNumLUWeHN3jWbubuSc0cubyPTrvS8UTj5PdtmpU/FbXbM
 gQoYMOreg5TVmRIhLKELS6BrELPTV4MJDeJoFU6+n8wFOZhcjAm6aVOZNhkeIlIg
 pruZIHgN/jKPMWw6rwxldrImJW4RxBTOYYN8Nk7nGHWJ0dlRrhdX144MNoXnurzH
 mmLY8YxGPxEguh8mqRPwctoVaIdhQ5cGAZcBgZ6MhC0T0Y42ETIM6luPXoFQJ8Sj
 PhqjKSgypL38TCmnb+TFwNSpYSiENcmZAPl1mptKKlR/Y2MXbLe4WYPpML9YwYno
 0RyA0ZrpnG2gtDbarl0K
 =+4Fk
 -----END PGP SIGNATURE-----
 

[Codegen][CPU] Register embedded ukernel bitcode + provider lookup. (#24568)

Wires the LLVMCPU plugin and the `#iree_cpu.ukernel_provider` attribute
to actually resolve ukernel bitcode:

* On plugin init, the embedded `iree_uk_cpu_bitcode` TOC is iterated and
each `.bc` is added to the global `EmbeddedDataDirectory` (mirrors
`addAMDGPUUkernelBitcodeToGlobalEmbeddedDataDirectory` in
`ROCMTarget.cpp`).

* `UKernelProviderAttr::createAndReplaceWithUkernelOp` now resolves the
bitcode by name: user-supplied `hal.executable.objects` ancestors first
(BYO and BYO-override paths), then the embedded directory. If found and
not already on the op, the attribute is attached to the source op; the
default `LowerBitcodeUKernelsPass` rewrite then preserves it on the
resulting `ukernel.generic`. The provider still returns `std::nullopt`
for the rewrite itself — specialized `inner_tiled` handling (scalar
`intrinsics_{m,n,k}` + outer K count + ACC inner stride) lands with
SelectUKernels in a follow-up.

* Filename match is `<ukernelName>.<features>.bc` (or bare
`<ukernelName>.bc`), matching the build rules under
`compiler/plugins/target/LLVMCPU/builtins/ukernel/`. Multi-variant
disambiguation by target feature set also lands with SelectUKernels.

BUILD.bazel: LLVMCPU plugin gains a dep on

`//compiler/plugins/target/LLVMCPU/builtins/ukernel:iree_uk_cpu_bitcode`,
and the `iree_cpu` dialect picks up HAL IR + `EmbeddedDataDirectory` so
the provider implementation can live with the attribute (consistent with
how `ROCMUkernelBitcodeSupport.cpp` sits inside the ROCM dialect).

Tests: a three-way lit test (`lookup_builtin_and_byo.mlir`) covers each
branch of the lookup:
1. Built-in only — no user objects; expect a `dense_resource<…>` with
the embedded bf16 bitcode (2732 bytes) attached.
2. BYO override — user supplies bitcode under the built-in's filename;
expect the user's bytes to flow through unchanged (and CHECK-NOT against
`dense_resource` to catch a regression where the provider silently swaps
in the embedded copy).
3. Pure BYO — user-only name with no matching built-in; expect the
user's bitcode to surface on the rewritten op.

Progress towards https://github.com/iree-org/iree/issues/24574.

Signed-off-by: Benoit Jacob <jacob.benoit.1@gmail.com>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>