[Rock] Add i4 (iu4 16x16x32) WMMA GEMM support for gfx12

[rhway666]

Adds native i4 (iu4) integer GEMM codegen for gfx12/RDNA4: an i4×i4→i32 GEMM
now lowers through the rock dialect to wmma_i32_16x16x32_iu4 (mirrors the
existing i8 wmma_i32_16x16x16_iu8 path). Distinct from prior i4 work (#1584 is
i4×f16 mixed-precision; #1682 is i4 vector loads) — this is the native
iu4×iu4 integer WMMA path.

Tracking issue for i4 follow-ups (autotuner, verifier, conv/other-arch,
signedness, full tuning short-list): #2418

Assisted-by: Claude (Anthropic)

Motivation

The i4 WMMA type id (I4_To_I32_TyId) and getRetType (i4→i32) already exist
upstream, but i4 was never wired to an instruction. This connects it: the
instruction map, the frontend (-t i4), the type gates (gfx12-only), the
tuning-key string, and a tuning config. iu4 uses K=32 and inputVectorLen=16
(vs i8's 16/8) because a 4-bit element packs at twice the K depth of an 8-bit
element. Gated to gfx12 (gfx11 rejects i4).

Technical Details

Files (8):

• Dialect/Rock/IR/WmmaInsnGroup.cpp — gfx12 insn map {I4_To_I32_TyId, 32} → wmma_i32_16x16x32_iu4; select chooses K=32 for i4.
• tools/rocmlir-gen/rocmlir-gen.cpp — accept -t i4; default i4 output to i32.
• IR/RockOps.td — allow I<4> in GemmInputTypes.
• IR/RockDialect.cpp — verifyGemmTypes accepts i4 only on gfx12 (RDNA4).
• Tuning/ParamLookupTable.cpp — map i4 to "i4" in the tuning key.
• Tuning/QuickTuningPerfconfigs.inc — a gfx1201 i4 tuning config.
• README.md — list i4 in the -t/--dtype examples.
• test/Dialect/Rock/wmma_gemm_i4.mlir — lowering check + gfx11 negative check.

Notes for reviewers:

• Signedness: matches the existing integer WMMA path; this PR does not add a
  signedness selector (tracked as a shared integer-WMMA follow-up in the issue).
• K selection: select picks K=32 for i4 via a ternary since i4 is the only
  K≠16 type on gfx12 today; mirrors the gfx1250 per-type-K style and can be
  refactored to if/else-if if more K variants are added on gfx12.
• Tuning (note on a generated file): QuickTuningPerfconfigs.inc is generated
  by quickTuningGen.py --update from sweep TSVs. The i4 entry is required,
  not a perf nicety: the quick-tuning lookup report_fatal_errors on a miss
  (Tuning parameters not found for key gfx1201_gemm_i4) and falls back only by
  same op+dtype across architectures — so i4 has no i8 fallback and no other arch
  carries gemm_i4. Without it, i4 GEMM does not lower. The autotuner
  (rocmlir-tuning-driver) cannot yet tune i4 (its applicability pipeline rejects
  all i4 configs), so the entry is placed by hand and a future --update
  would drop it (flagged inline). The single config is correctness-verified and
  was chosen as the most robust single config across the i8 tier1-gemm shapes
  (offline rocprofv3 sweep on gfx1201); those numbers are provisional and not
  CI-verified. A full per-shape short-list will be produced data-driven via
  quickTuningGen.py --update (not hand-edited) once i4 is wired into the
  autotuner — tracked in the issue.

Known limitation (out of scope, follow-up): the --verifier cpu/gpu
random-data auto-compare paths still need i4 narrow-type handling — (1)
rock.transform's affine index-map must be rewritten for the packed (i8) layout;
(2) the workgroup-LDS memref<…xi4, #gpu.workgroup> alloc + vector.store of
vector<Nxi4> must be packed consistently. This does not affect the generated
kernel (validated below); it's a separate narrow-type follow-up.

Test Plan

• Lit test (mlir/test/Dialect/Rock/wmma_gemm_i4.mlir, CI-runnable):
  rocmlir-gen … -t i4 -wmma=on --arch gfx1201 | rocmlir-driver … lowers to
  amdgpu.wmma 16x16x32 … vector<16xi4> and
  rocdl.wmma.i32.16x16x32.iu4 … (vector<2xi32>, vector<2xi32>, vector<8xi32>) -> vector<8xi32>; a gfx1100 i4 GEMM is rejected.
• Hardware run on AMD Radeon AI PRO R9700 (gfx1201, ROCm 7.2): the generated
  i4 kernel compiles to a gfx1201 binary and executes.
• iu4 instruction independently checked via a standalone HIP microkernel.

Test Result

• Lit test passes (both lowering checks + the gfx11 negative check).
• Hardware: all-ones inputs give C[m][n] = Σ_k 1·1 = K — verified output == K
  for K ∈ {32, 128, 256} on gfx1201.
• iu4 16x16x32 microkernel: all-ones → 32 across all 256 accumulator slots
  (bit-exact).
• clang-format clean (git clang-format --diff is empty).

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Copilot]

Pull request overview

This PR wires up native i4 (iu4) integer WMMA GEMM codegen on gfx12/RDNA4, enabling i4×i4→i32 GEMM to lower through rock to the wmma_i32_16x16x32_iu4 instruction, and adds the supporting frontend, verification, tuning-key, perfconfig, and test updates.

Changes:

• Add gfx12 WMMA instruction selection/mapping for iu4 16x16x32 (K=32) and gate i4 GEMM types to RDNA4.
• Extend rocmlir-gen dtype handling to accept -t i4 and default i4 output to i32.
• Add tuning-key + quick-tuning entry for gfx1201_gemm_i4 and a Lit test covering gfx12 lowering and gfx11 rejection.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
README.md	Documents -t i4 in rocmlir-gen dtype examples.
mlir/tools/rocmlir-gen/rocmlir-gen.cpp	Adds CLI parsing for i4 and defaults i4 output dtype to i32.
mlir/test/Dialect/Rock/wmma_gemm_i4.mlir	New Lit test for gfx1201 lowering to WMMA iu4 and gfx1100 rejection.
mlir/lib/Dialect/Rock/Tuning/ParamLookupTable.cpp	Adds i4 (bitwidth==4) dtype string normalization for tuning keys.
mlir/lib/Dialect/Rock/IR/WmmaInsnGroup.cpp	Adds gfx12 iu4 WMMA op mapping and selects K=32 for i4.
mlir/lib/Dialect/Rock/IR/RockDialect.cpp	Verifier allows i4 on RDNA4 WMMA GEMM.
mlir/include/mlir/Dialect/Rock/Tuning/QuickTuningPerfconfigs.inc	Adds gfx1201_gemm_i4 quick-tuning perfconfig entry.
mlir/include/mlir/Dialect/Rock/IR/RockOps.td	Allows I<4> in GemmInputTypes (affects GEMM and rock.conv forward).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.