[AIROCMLIR-888] Implement per-problem narrowing in the quick-tuning DB by mirza-halilcevic · Pull Request #2361 · ROCm/rocMLIR

mirza-halilcevic · 2026-05-04T23:06:17Z

Motivation

Quick-tuning compile time was dominated by the size of the perfconfig list we swept for each kernel. We previously sent the entire per-(arch, op, dtype) set-cover at every kernel, even when the problem at hand was already represented in our tuning campaigns and a much smaller subset would have sufficed.

This PR addresses that on two fronts:

Known problems: tier-1 problems we have tuning data for now carry an explicit problem → perfconfig-subset mapping. The compiler narrows the sweep to just the configs that cover this exact shape, instead of the full set.
Unknown problems: the returned list is capped with a new env var, ROCMLIR_QUICK_TUNING_LIST_MAX (default 30). The set-cover is already sorted by descending coverage count, so head-truncating preserves the configs that win on the most problems; statistically the right tradeoff when we have no problem-specific signal.

The branch also lands a few quality-of-life cleanups around the quick-tuning plumbing that were prerequisites for the per-problem mapping.

The existing quick-tune lists were migrated to the new per-key .inc format in this PR; the migration carries the set-covers only and does not include problem maps. Regenerating the lists with problem maps populated will land in a follow-up. Until then, all entries take the unknown-problem path (full set-cover, capped to ROCMLIR_QUICK_TUNING_LIST_MAX).

Note on diff size: the diff is inflated by the ParamLookupTable → QuickTuningDb rename and by the migration of the existing perfconfig data from one monolithic .inc into per-key .inc files. The hand-written code change is significantly smaller than the line count suggests.

Technical Details

Quick-tuning DB layout

One .inc file per (arch, op, dtype) key (mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx942ConvF16.inc, etc.). The previous monolithic QuickTuningPerfconfigs.inc is gone. CMake globs the per-key files in sorted order and generates a single QuickTuningDb.inc that the lookup TU includes twice (once with QUICK_TUNING_DB_ARRAYS to emit the static arrays, once with QUICK_TUNING_DB_ENTRIES to emit the table initializer). Sorted filenames ↔ sorted snake_case keys, which is the precondition the binary-search lookup relies on.
ParamLookupTable → QuickTuningDb, restructured from a class into a namespace. The semantics (lookup-by-key with arch-family fallback) are now closer to a curated database than a generic param table, and the new name reflects that. Private helpers moved into anonymous namespaces in the cpp.
quickTuningGen.py auto-detects op type from the input .debug headers (Chip column for arch, dataframe columns for op). One invocation now handles a mixed bag of gemm/conv/attention runs without per-op flags.

The problem map

Each table entry has the shape:

struct QuickTuningDbEntry {
  const char *key;                   // e.g. "gfx942_conv_f16"
  const StringRef *setCover; ...;    // full set-cover (sorted by coverage desc)
  const unsigned *indices;   ...;    // flat run of indices into setCover
  const ProblemRef *problemMap; ...; // sorted by hash
};

struct ProblemRef { uint64_t hash; unsigned offset, count; };

A ProblemRef is a (hash, offset, count) triple. To answer "what configs should we sweep for this problem?" we:

Compute the problem-key hash for the op (see below).
Resolve the table entry for (arch, op, dtype, isAccel) via binary-search with arch-family fallback (e.g. gfx940 → gfx942, bf16 → f16).
Binary-search problemMap for the hash. On hit, take indices[offset .. offset+count) and gather those positions out of setCover. On miss, fall back to the full set-cover.
Truncate to ROCMLIR_QUICK_TUNING_LIST_MAX (default 30).

The flat-indices layout means a single .inc carries one set-cover plus a compact per-problem index ribbon; no duplicated config strings, no per-problem allocations.

Problem keys

The on-disk hashes are xxh3_64 of a '_'-joined column tuple. Python is the authoritative producer (quickTuningGen.py:make_problem_key / hash_problem_key); C++ rebuilds the same byte sequence at runtime so the hashes line up:

Bool columns are coerced to int at DataFrame ingest, so the wire format is 0/1 (was True/False). C++ mirrors this in QuickTuningProblemKey.cpp.
Per-op column lists are pinned to the *_COLUMNS constants in the generator (GEMM_COLUMNS, CONV_COLUMNS, ATTENTION_COLUMNS).
Goldens are pinned in both mlir/utils/performance/tests/test_quickTuningGen.py and mlir/unittests/Dialect/Rock/QuickTuningDbTests.cpp. Drift on either side fails both suites.

Wiring on the consumer side

PopulateParamsInfo carries an optional problemKeyHash, populated via QuickTuningDb::computeProblemKeyHash(op) (returns FailureOr<uint64_t>).
PopulateParams, PopulateParamsXDL, PopulateParamsWmma, and PopulateParamsGemmGemm thread the hash into QuickTuningDb::lookup.
RockTuningImpl::createGemmTuningRangeQuick does the same in the tuning sweep, so tuning runs also benefit from the narrowing.

Test Plan

New gtest suite QuickTuningDbTests covers DB invariants (sorted by key, per-entry problem maps sorted by hash), resolveKey fallback policy (exact / older-relative / younger-relative / equidistant tiebreak / non-accel → gfx1*+f32 / bf16 → f16), and per-op problem-key goldens for gemm/conv/attention.
New pytest module test_quickTuningGen.py covers make_problem_key format, the bool-cast at ingest, and golden hashes shared with the C++ side.

Test Result

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 restructures Rock’s quick-tuning “DB” into a curated QuickTuningDb with optional per-problem narrowing (via a problem-key hash) and a configurable cap (ROCMLIR_QUICK_TUNING_LIST_MAX) to reduce compile-time spent sweeping large perfconfig lists.

Changes:

Introduces QuickTuningDb + runtime problem-key hashing (xxh3_64) to optionally narrow perfconfig sweeps to a per-problem subset.
Threads an optional problemKeyHash through quick-tuning plumbing so both compile-time selection and tuning sweeps can benefit.
Migrates the perfconfig data into per-(arch, op, dtype) auto-generated .inc shards, and adds C++/Python goldens + DB invariant tests.

Reviewed changes

Copilot reviewed 121 out of 122 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
mlir/utils/performance/tests/test_quickTuningGen.py	Adds pytest goldens for problem-key formatting, bool-cast behavior at ingest, and pinned xxh3_64 hashes.
mlir/unittests/Dialect/Rock/ParamLookupTableTests.cpp	Removes legacy ParamLookupTable fallback-policy tests (superseded by QuickTuningDb tests).
mlir/unittests/Dialect/Rock/CMakeLists.txt	Switches unit tests to `QuickTuningDbTests.cpp` and adds `MLIRParser` for parsing MLIR test snippets.
mlir/unittests/Dialect/Rock/QuickTuningDbTests.cpp	Adds gtests for DB invariants, resolveKey fallback policy, and problem-key hash goldens.
mlir/lib/Dialect/Rock/Tuning/RockTuningImpl.cpp	Threads `problemKeyHash` into quick tuning range creation so tuning sweeps can narrow lists too.
mlir/lib/Dialect/Rock/Tuning/GridwiseGemmParams.cpp	Populates `PopulateParamsInfo::problemKeyHash` from the op (when computable).
mlir/lib/Dialect/Rock/Tuning/GridwiseGemmGemmParams.cpp	Switches GemmGemm quick-tuning lookup to `QuickTuningDb` and computes optional problem hash.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb.cpp	Implements key resolution + lookup (including per-problem selection + list cap).
mlir/lib/Dialect/Rock/Tuning/QuickTuningProblemKey.cpp	Implements op→problem-key string formatting and xxh3_64 hashing to match Python generator.
mlir/lib/Dialect/Rock/Tuning/CMakeLists.txt	Replaces ParamLookupTable sources with QuickTuningDb sources; generates `QuickTuningDb.inc` by globbing/sorting per-key `.inc` files.
mlir/lib/Dialect/Rock/Tuning/ParamLookupTable.cpp	Removes legacy ParamLookupTable implementation.
mlir/include/mlir/Dialect/Rock/Tuning/QuickTuningDb.h	Adds the public QuickTuningDb API (lookup/resolveKey/problem hashing + invariants for tests).
mlir/include/mlir/Dialect/Rock/Tuning/ParamLookupTable.h	Removes legacy ParamLookupTable API.
mlir/include/mlir/Dialect/Rock/Tuning/QuickTuningPerfconfigs.inc	Removes the previous monolithic quick-tuning perfconfig include.
mlir/include/mlir/Dialect/Rock/Tuning/GridwiseGemmParams.h	Switches tuning parameter plumbing to QuickTuningDb and adds `problemKeyHash` to `PopulateParamsInfo`.
mlir/include/mlir/Dialect/Rock/Tuning/GridwiseGemmGemmParams.h	Switches GemmGemm tuning to QuickTuningDb.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1000AttentionF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1000ConvF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1000ConvF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1000ConvFp8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1000ConvI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1000GemmF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1000GemmF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1000GemmFp8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1000GemmI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1100ConvF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1100ConvF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1100ConvI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1100GemmF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1100GemmF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1100GemmI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1101AttentionBf16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1101AttentionF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1101AttentionI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1103AttentionBf16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1103AttentionF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1103AttentionI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1103ConvF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1103ConvF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1103ConvI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1103GemmF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1103GemmF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1103GemmI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1150AttentionBf16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1150AttentionF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1150AttentionI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1150ConvF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1150ConvF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1150ConvI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1150GemmF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1150GemmF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1150GemmI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1151ConvF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1151ConvF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1151ConvI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1151GemmF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1151GemmF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1151GemmI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1152AttentionBf16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1152AttentionF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1152AttentionI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1152ConvF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1152ConvF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1152ConvI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1152GemmF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1152GemmF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1152GemmI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1201AttentionBf16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1201AttentionF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1201AttentionI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1201ConvF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1201ConvF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1201ConvI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1201GemmF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1201GemmF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx1201GemmI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx900AttentionF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx900AttentionF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx900ConvFp8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx900GemmFp8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx908AttentionBf16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx908AttentionF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx908AttentionF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx908AttentionI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx908ConvF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx908ConvF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx908ConvI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx908GemmF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx908GemmF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx908GemmI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx90aAttentionBf16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx90aAttentionF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx90aAttentionF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx90aAttentionI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx90aConvF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx90aConvF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx90aConvI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx90aGemmF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx90aGemmF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx90aGemmI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx942AttentionBf16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx942AttentionF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx942AttentionF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx942AttentionI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx942ConvF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx942ConvF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx942ConvI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx942GemmF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx942GemmF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx942GemmI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx950AttentionBf16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx950AttentionF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx950AttentionF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx950AttentionI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx950ConvF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx950ConvF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx950ConvI8.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx950GemmF16.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx950GemmF32.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx950GemmFp4.inc	Auto-generated per-key set-cover shard.
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb/Gfx950GemmI8.inc	Auto-generated per-key set-cover shard.

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

codecov · 2026-06-02T03:53:21Z

Codecov Report

❌ Patch coverage is 81.71429% with 64 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb.cpp	74.74%	35 Missing and 13 partials ⚠️
.../lib/Dialect/Rock/Tuning/QuickTuningProblemKey.cpp	88.28%	5 Missing and 10 partials ⚠️
...lib/Dialect/Rock/Tuning/GridwiseGemmGemmParams.cpp	87.50%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #2361      +/-   ##
===========================================
+ Coverage    82.57%   82.66%   +0.09%     
===========================================
  Files          120      119       -1     
  Lines        42852    43025     +173     
  Branches      7110     7152      +42     
===========================================
+ Hits         35381    35563     +182     
- Misses        4815     4823       +8     
+ Partials      2656     2639      -17

Flag	Coverage Δ
gfx120x	`82.51% <81.43%> (-0.02%)`	⬇️
gfx950	`82.50% <81.43%> (+0.15%)`	⬆️
mfma	`82.40% <81.14%> (?)`

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
...lude/mlir/Dialect/Rock/Tuning/GridwiseGemmParams.h	`100.00% <ø> (+4.76%)`	⬆️
...lir/lib/Dialect/Rock/Tuning/GridwiseGemmParams.cpp	`77.08% <100.00%> (+1.05%)`	⬆️
mlir/lib/Dialect/Rock/Tuning/RockTuningImpl.cpp	`59.41% <100.00%> (+0.09%)`	⬆️
...lib/Dialect/Rock/Tuning/GridwiseGemmGemmParams.cpp	`88.51% <87.50%> (-0.52%)`	⬇️
.../lib/Dialect/Rock/Tuning/QuickTuningProblemKey.cpp	`88.28% <88.28%> (ø)`
mlir/lib/Dialect/Rock/Tuning/QuickTuningDb.cpp	`74.74% <74.74%> (ø)`

... and 25 files with indirect coverage changes

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

rocmlir-pr-reviewer

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

Scope

Reworks the quick-tuning database: renames ParamLookupTable → QuickTuningDb (class → namespace), splits the monolithic QuickTuningPerfconfigs.inc into per-(arch,op,dtype) .inc files globbed+sorted by CMake, adds a per-problem narrowing path (xxh3_64 problem-key hashing mirrored between quickTuningGen.py and QuickTuningProblemKey.cpp), and caps results via ROCMLIR_QUICK_TUNING_LIST_MAX (default 30). Consumer wiring threads an optional problemKeyHash through PopulateParams* and RockTuningImpl. Bulk of the diff is migrated/auto-generated data plus the rename.

Findings

No blocking issues found. The one prior inline comment (Copilot: std::mismatch 3-iterator overload reading past itKey/prevKey end) is already addressed — QuickTuningDb.cpp now uses the bounded 4-iterator overload at both call sites.

Notes

Non-accel fallback now restricted to gfx1*. getRelatives forces archPrefix="gfx1" and f32 when !isAccel (QuickTuningDb.cpp), whereas the old ParamLookupTable used a 3-char gfx prefix (any family). This is intentional and covered by ResolveKeyTest.NonAccelRestrictsToGfx1Family, but it is a behavior change: a non-accel f32 query on a non-MFMA gfx9 part (e.g. gfx900, which has no *GemmF32/*ConvF32 data here) now resolves to a gfx1* entry rather than the nearest gfx9 sibling. Worth a sanity check that no supported flow regresses; configs remain valid (vetted by paramsProbablyValid), so worst case is sub-optimal, not incorrect.
Conv problem-key for grouped/asymmetric cases. makeConvKey encodes C as in[ci]*in[gi], K as fil[k]*fil[g], and padding as pad[0]/pad[2] only. Goldens only exercise G=1 and symmetric padding=1; when problem maps are populated in the follow-up, extend goldens to G>1 / asymmetric padding to confirm byte-parity with the Python C/PaddingH/PaddingW columns. No current effect since the migrated data carries no problem maps (all queries take the full set-cover path).
Minor/non-blocking: lookup() calls normalizeArch and then resolveEntry normalizes again (idempotent); pip_requirements.txt correctly gains pulp/xxhash; new files carry the correct Part of the MLIR Project SPDX header.

CI status

No genuine CI failures. Jenkins, C/C++ premerge, Python format/lint, and Python performance script tests all pass; many hardware jobs are still pending. The failing review check is this auto-review pipeline's own job and is expected.

Implement per-problem narrowing in the quick-tuning DB.

882f9cf

mirza-halilcevic requested review from Copilot, dhernandez0 and umangyadav May 4, 2026 23:09

Copilot started reviewing on behalf of mirza-halilcevic May 4, 2026 23:12 View session

Copilot AI reviewed May 4, 2026

View reviewed changes

Comment thread mlir/lib/Dialect/Rock/Tuning/QuickTuningDb.cpp Outdated

mirza-halilcevic and others added 2 commits May 11, 2026 14:42

Merge branch 'develop' into quick-tune-prob-map

0fd78fb

Address code review.

a516f62

mirza-halilcevic marked this pull request as ready for review May 11, 2026 13:46

mirza-halilcevic requested a review from causten as a code owner May 11, 2026 13:46

mirza-halilcevic and others added 6 commits May 11, 2026 20:27

Apply clang-format to .inc files.

62d60aa

Merge branch 'develop' into quick-tune-prob-map

f9ad230

Merge branch 'develop' into quick-tune-prob-map

5c5e6ee

Merge branch 'develop' into quick-tune-prob-map

b922c5c

Merge branch 'develop' into quick-tune-prob-map

2e2e92f

Merge branch 'develop' into quick-tune-prob-map

d7cff2b

mirza-halilcevic changed the title ~~[AIROCMLIR-560] Implement per-problem narrowing in the quick-tuning DB~~ [AIROCMLIR-888] Implement per-problem narrowing in the quick-tuning DB May 19, 2026

mirza-halilcevic and others added 3 commits May 22, 2026 07:57

Merge branch 'develop' into quick-tune-prob-map

c968a7b

Merge branch 'develop' into quick-tune-prob-map

5f5b71e

Merge branch 'develop' into quick-tune-prob-map

46571a4

mirza-halilcevic added the claude-review Trigger automated PR review by claude[bot]; auto-removed after the run. label Jun 1, 2026

rocmlir-pr-reviewer Bot removed the claude-review Trigger automated PR review by claude[bot]; auto-removed after the run. label Jun 1, 2026

Merge branch 'develop' into quick-tune-prob-map

a90de04

mirza-halilcevic added the claude-review Trigger automated PR review by claude[bot]; auto-removed after the run. label Jun 1, 2026

rocmlir-pr-reviewer Bot removed the claude-review Trigger automated PR review by claude[bot]; auto-removed after the run. label Jun 1, 2026

mirza-halilcevic and others added 4 commits June 2, 2026 09:34

Merge branch 'develop' into quick-tune-prob-map

eaf6adb

Merge branch 'develop' into quick-tune-prob-map

9e7a29a

Merge branch 'develop' into quick-tune-prob-map

3b21d37

Fix flake8 checks.

e4bb8a8

mirza-halilcevic added the claude-review Trigger automated PR review by claude[bot]; auto-removed after the run. label Jun 3, 2026

rocmlir-pr-reviewer Bot removed the claude-review Trigger automated PR review by claude[bot]; auto-removed after the run. label Jun 3, 2026

mirza-halilcevic added 5 commits June 3, 2026 19:39

Merge branch 'develop' into quick-tune-prob-map

4dd3891

Merge branch 'develop' into quick-tune-prob-map

e973cb1

Merge branch 'develop' into quick-tune-prob-map

d7fa5be

Merge branch 'develop' into quick-tune-prob-map

a033b2e

Merge branch 'develop' into quick-tune-prob-map

938215c

mirza-halilcevic requested review from justinrosner and pabloantoniom June 15, 2026 09:40

mirza-halilcevic added the claude-review Trigger automated PR review by claude[bot]; auto-removed after the run. label Jun 16, 2026

rocmlir-pr-reviewer Bot removed the claude-review Trigger automated PR review by claude[bot]; auto-removed after the run. label Jun 16, 2026

mirza-halilcevic added the claude-review Trigger automated PR review by claude[bot]; auto-removed after the run. label Jun 16, 2026

rocmlir-pr-reviewer Bot reviewed Jun 16, 2026

View reviewed changes

rocmlir-pr-reviewer Bot removed the claude-review Trigger automated PR review by claude[bot]; auto-removed after the run. label Jun 16, 2026

mirza-halilcevic added 2 commits June 17, 2026 00:04

Merge branch 'develop' into quick-tune-prob-map

14e27d6

Merge branch 'develop' into quick-tune-prob-map

6bcf1b0

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[AIROCMLIR-888] Implement per-problem narrowing in the quick-tuning DB#2361

[AIROCMLIR-888] Implement per-problem narrowing in the quick-tuning DB#2361
mirza-halilcevic wants to merge 24 commits into
developfrom
quick-tune-prob-map

mirza-halilcevic commented May 4, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

codecov Bot commented Jun 2, 2026 •

edited

Loading

Uh oh!

rocmlir-pr-reviewer Bot left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

mirza-halilcevic commented May 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Technical Details

Quick-tuning DB layout

The problem map

Problem keys

Wiring on the consumer side

Test Plan

Test Result

Submission Checklist

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

codecov Bot commented Jun 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

rocmlir-pr-reviewer Bot left a comment

Choose a reason for hiding this comment

Scope

Findings

Notes

CI status

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

mirza-halilcevic commented May 4, 2026 •

edited

Loading

codecov Bot commented Jun 2, 2026 •

edited

Loading