tree 2c20874109b74792eb3d5c618c48aa79d71c4ac3
parent 6ec0a37c41c0dec313b5f69df6238370e07e303a
author Benoit Jacob <jacob.benoit.1@gmail.com> 1782916590 -0400
committer GitHub <noreply@github.com> 1782916590 +0200
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJqRSXuCRC1aQ7uu5UhlAAA/K0QAFOZ3e1xtUo9+yqD2OJubqs1
 5GZRykxc5Y/LMR5bj78BleufUZgyU1lCLLHK4qKtZbNO3/xTSEQmNYAmJKaWcnUe
 jXC1n16xQA0ezZ6u/JlU804L9YvT8b+GcB8m4lKFatolH8k6/iMSywcM9YqvGMg6
 eQKFM/mOQnVwgaFNzufDcG4Pf91HFKyN/G9W+cFaOdOKaSKVCKwuBbMcbU86XrZ+
 YmCzt+RZkE83zoqqDrtMfxB1WGUpEFJE6I6Cvi00BjxKv+lSymQsBqHG55nNl0r/
 kPK9U61Qmk5NnNfxHE44j7D9JTi0tYJnUXkuyi9ceG0swmsBFxly16jwGM50mohD
 DCJIWnMimqRPp/5a/EgkpIyMlGKZVD3OHRRJKFb3msB6U4Hlp4CTUvfWQ7rIBh2J
 NnFCWt6zxWR5hbEfJViAfex6jlWcMa9E9YD4MnNHmY6y5/5NsRfJgOVWPYrvJc29
 6bLaWvCLogTF9dgLezdqEz6u02GGO4FjL5x5QHCg4jJRvujqVFFQcya/OuGl2dIg
 yvQ7gJvPN7HxhWQ6wZgu519EkxUj56lNgKvhg0zwemT7XV0NcJuiKFoCchdC1Jaf
 lhmSq0RbFHWx6cmF4yOi9bbMBsiJWlH5vWpVpmiFybzvbEQzUsH2YVLgtKpMYsZz
 /LVX2kgGTH/BhDtAZE0Z
 =ZeDO
 -----END PGP SIGNATURE-----
 

[Codegen][CPU] Drop the ACC stride from the inner_tiled ukernel ABI. (#24652)

New bottom of the LLVMCPU C-ukernel stack.

The ACC stride threaded to the C ukernels is always a compile-time
constant
equal to one intrinsic's ACC fragment size (the data-tiled ACC tile is
contiguous), so it folds to a constant after inlining and carries no
information the intrinsic-specific ukernel doesn't already encode. This
drops
it:

- `IREECPUAttrs.cpp`: remove `getAccInnermostCrossIntrinsicDim`; build
the
`ukernel.generic` with `num_strided_outer_dims=0`, so all shaped
operands
  (LHS, RHS, ACC) lower to `(base, offset)` only.
- Both seed ukernel `.c` signatures lose the `acc_stride` parameter.

This addresses review feedback on the now-merged #24569 / #24571 (why
pass a
per-operand `strided_dims` array, and why pass an ACC stride that's
statically
known). It makes the three operands uniform and removes the
`getAccInnermostCrossIntrinsicDim` swizzle machinery.

Verified: the lowered `llvm.call` to the ukernel now has no stride
argument,
and the ukernel lit tests (`lower_inner_tiled_to_bitcode_ukernel*`,
`e2e_inner_tiled_pipeline`, `select_ukernel`) pass.

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

Signed-off-by: Benoit Jacob <jacob.benoit.1@gmail.com>