[AIROCMLIR-668] Add tests for gaps in rocmlir-gen LIT coverage

[justinrosner]

Motivation

This is a port of https://github.com/ROCm/rocmlirTriton/pull/250 from rocmlirTriton.

This PR adds tests for existing rocmlir-gen features that were missing direct test coverage in our LIT tests.

Technical Details

The following is a list of what changed relative to the rocmlirTriton version:

• Removed LIT tests for features that don't exist in rocMLIR
  • `--cpu-timers
• Updated verifier ABI differences (gemm-misc-options.mlir)
  • mcpuVerifyFloat in rocMLIR takes two trailing i1 args, so the expected signature is now (memref<?xf32>, memref<?xf32>, f32, f32, f32, i8, i1, i1)
    instead of (..., i8, i1)
  • The integer verifier helper is mcpuVerifyInt32Int64 with an i64 reference operand (memref<?xi32>, memref<?xi64>, i8), not mcpuVerifyInt32Int32 / memref<?xi32>.
• Updated conv kernel signature & op syntax (conv-bwd-data.mlir)
• Updated attention output transpose (attention-kernel.mlir)
  • -transO surfaces a tr prefix on the second-gemm result inside rock.attention (tr %N = softmax(qk) * %M : ... -> memref<1x64x128xf16>), not an oTransposed attribute.

Test Plan

• Run the PR CI

Test Result

• Passing PR CI: https://ml-ci-internal.amd.com/job/MLIR/job/mlir/job/PR-2392/1/pipeline-overview/

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 adds direct LIT coverage for existing rocmlir-gen and driver option behavior, especially validation/error paths and previously untested command-line feature combinations.

Changes:

• Adds negative option-validation tests for rocmlir-gen.
• Expands GEMM, conv, attention, random-data, and padding option coverage.
• Verifies ABI/signature expectations for verifier calls and current Rock op assembly syntax.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
mlir/test/rocmlir-gen/options.mlir	Adds negative tests for invalid/unsupported option combinations and padding warnings.
mlir/test/rocmlir-gen/gemm-misc-options.mlir	Adds tests for tuning keys, print inputs, verifier result levels, and device selection.
mlir/test/rocmlir-gen/conv-bwd-data.mlir	Adds mixed-output dtype and 3-D backward conv coverage.
mlir/test/rocmlir-gen/attention-kernel.mlir	Adds IR checks for attention transpose flags.
mlir/test/rocmlir-gen/attention-kernel-f16.mlir	Adds softmax intermediate dtype coverage.
mlir/test/rocmlir-driver/populate_random_data.mlir	Adds integer random bounds and selective integer-random input tests.
mlir/test/rocmlir-driver/populate_padding.mlir	Adds asymmetric depth-padding coverage for 3-D backward-data conv.

---

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

[rocmlir-pr-reviewer]

Reviewed 7 files, all test-only .mlir additions covering previously untested rocmlir-gen options (asymmetric depth padding for 3-D bwd-data conv; integer random-data bounds and -rand_type_int_for_inputs; --softmax_dtype; per-tensor attention transpose flags; 3-D bwd-data/bwd-weight kernel signatures; forward conv mixed dtypes; bwd-data/bwd-weight tuning keys; -pi, --print-verify-results, --device; and several negative validation cases). Tests follow existing file conventions, use FileCheck prefixes/COUNT/NOT cleanly, and explanatory comments precisely match the asserted IR. CI is green except for the in-progress review check. Posted 0 inline comments (0 critical, 0 major, 0 minor). Verdict: APPROVE.

[rocmlir-pr-reviewer]

Wording overstates the behavior: regular convolution still goes through createCPUConvFunc for --verifier=cpp; only the GEMM and attention paths reject it (and those are what the two RUN lines below actually exercise). Narrow the comment so it doesn't imply conv is rejected too — e.g. "--verifier=cpp is not implemented for the GEMM or attention paths."

Suggested change

	// --verifier=cpp is not implemented for any of the in-tree operations.
	// --verifier=cpp is not implemented for the GEMM or attention paths.

[rocmlir-pr-reviewer]

Verdict: APPROVE -- submitted as COMMENT (automated reviews are advisory)  ·  Findings: 1 (0 Critical, 0 Major, 1 Minor)

---

Scope

Test-only PR porting rocmlir-gen LIT coverage from rocmlirTriton. Adds positive checks for 3-D backward-data/-weight conv kernels, asymmetric depth padding, mixed-dtype forward conv, attention transpose flags (-transQ/-transV/-transO), softmax dtype override, integer rand bounds & selective -rand_type_int_for_inputs, --emit-tuning-key for bwd kernels, -pi print-inputs, --print-verify-results levels (float + int verifier ABI), and --device/-dev. Adds negative tests in options.mlir for --verifier=cpp/clone, missing --arch/-t, missing GEMM dims, mixed-dtype without output dtype, flash-decoding without return_lse, and padding-flag conflicts.

Findings

One minor wording nit on a test-file comment in mlir/test/rocmlir-gen/options.mlir:7. Otherwise no blocking issues; the FileCheck directives, prefixes, and the updated mcpuVerifyFloat/mcpuVerifyInt32Int64 ABIs line up with the changes the PR description calls out.

Notes

• populate_random_data.mlir uses COUNT-N followed by -NOT: for the same call; that idiom asserts "exactly N matches, none after," which is the intended check.
• Negative tests in options.mlir rely on stable diagnostic substrings (e.g. Value for: m not specified, External gemm validator is not available); if those messages are reworded in the driver later, these LIT lines will need to track the new text — worth keeping in mind but not a blocker.
• All new tests use // RUN: + FileCheck prefixes consistently; no mlir/test/.../CMakeLists.txt change is needed since these files are already under lit.cfg-discovered directories.

CI status

Self/Anthropic check aside, all reportable checks are still pending (Jenkins, Build and Test, MIGraphX, etc.); no fail/cancel buckets to flag.

[rocmlir-pr-reviewer]

Verdict: APPROVE -- submitted as COMMENT (automated reviews are advisory)  ·  New findings: 0 (0 Critical, 0 Major, 0 Minor)

---

Scope

Pure test-coverage PR: adds LIT cases for previously untested rocmlir-gen features across attention transposes, conv mixed-dtype / 3-D bwd-data / bwd-weight, integer-randomness bounds and the -rand_type_int_for_inputs selector, --emit-tuning-key for bwd-data / bwd-weight, -pi print-inputs, --print-verify-results levels (float + int), --device / -dev ctor emission, and a batch of negative options.mlir validations. No production code touched.

Findings

No blocking issues. The previously-flagged wording at mlir/test/rocmlir-gen/options.mlir:7 (// --verifier=cpp is not implemented for any of the in-tree operations.) is still inaccurate -- the RUN lines below only exercise the gemm and attention rejection paths, while regular convolution still goes through createCPUConvFunc for --verifier=cpp. Tracking on the existing inline thread (no developer reply yet); not re-posted to avoid duplicate inline comments.

Notes

• The WARN_PADDING_H / WARN_PADDING_W cases at options.mlir:42-45 intentionally run without not because rocmlir-gen only warns; combining 2>&1 with FileCheck on the warning text is the right shape.
• BWD_DATA_3D / BWD_WEIGHT_3D sizes (filter 432, in/out 512) are consistent with g=1,k=4,c=4,fil_d/h/w=3 and n=2,c=4,d/h/w=4; padding-[h,h,w,w,d,d] ordering matches the documented layout convention in the new populate_padding.mlir block.
• RAND_INT_BOUNDS correctly uses i16 for the bound constants (matches existing populate_random_data.mlir patterns above the diff for the 8-bit codepath).

CI status

No CI failures on the pinned SHA. Both Python lint and Python perf-script jobs are still in progress; the auto-review job's own check is expected to be in-progress and is not a CI failure.