commit b98c1b92cb630bd696992f47df591bb2f247a8d7
author Eric Feng <55723758+efric@users.noreply.github.com> Fri Nov 21 15:44:03 2025 -0800
committer GitHub <noreply@github.com> Fri Nov 21 15:44:03 2025 -0800
tree 93920d7b3916230fda643d245f1f380ce8c4ca92
parent df3d0762b6120e3740f82bbfe28e0130cf668c9b

[LLVMGPU][Codegen] Emit packed chain FMA from select multi_reductions and contracts (#21855)

This patch teaches the vector lowering pipeline to:

1. Rewrite `vector.multi_reduction<add>` whose input is `arith.mulf`
into a `vector.contract` via
`vector::populateVectorReductionToContractPatterns`
2. Lower a restricted set of `vector.contract` into packed FMA chains.

Previously lowering `vector.multi_reduction` of the same form produced
elementwise pack-muls per K-slice and then reduced them with a
left-associated, serial chain of `v_add_f{16, 32}`

`(mul(a0 ,b0) + (mul(a1, b1) + … + acc`

The new lowering emits a single nested FMA chain and folds the
accumulation into the `math.fma` c-operand

`fma(a0 ,b0, fma(a1, b1, fma(a2, b2, fma(a3, b3, acc))))`

To do this, we first permute the reduction and parallel dimensions of
the `LHS` and `RHS` to the order of `[reduction, ..., parallel, ...]`.
The `LHS` and `RHS` are then collapsed to a 2D shape of `{Π reduction
dimensions, Π parallel dimensions}`. Then we form the FMA chain by
iterating backwards, seeded by the accumulator.

Not all forms of `vector.contract` are suitable in the current approach.
For example, when an operand drops a parallel iterator as in matmul-like
contracts. We require both sides to share the same 2D tuple. Unsupported
cases fall back to the existing lowering.

Fixes: #21483 (variant of original issue; for [issue
#21513](https://github.com/iree-org/iree/issues/21513)).

---------

Signed-off-by: Eric Feng <Eric.Feng@amd.com>