[zacharyvincze]

Summary

Resolves #11894.

Adds projects/rpp/utilities/test/ — a new, self-contained GoogleTest suite that validates RPP tensor ops against independent host-computed reference models. This is purely additive: the existing utilities/test_suite/ is untouched.

The existing test suite compares RPP output against golden image dumps for a fixed set of inputs. That catches regressions but not defects, and only along whichever axis the goldens happened to be captured on. This suite computes the expected result from the operator's documented definition instead, and sweeps every op over the same set of configuration axes, so a bug in one dtype / layout / ROI combination cannot hide behind a passing default.

Scope: ~97 operators, 24,664 test cases, 200 new files.

What's in it

src/
  main.cpp      entry point; installs the suite's console reporter
  framework/    shared harness: config grid, backend memory, comparators, skip list
  reference/    83 independent golden models
  tests/        94 test files — core/, image/, misc/, voxel/

Each op gets one reference/<op>_ref.hpp golden model plus a short TEST_P that builds descriptors, runs the op on both backends, and compares within tolerance over the ROI.

Test grid

Every case is instantiated over the cross product of:

Axis	Values
Backend	HOST, HIP
DType	U8, F16, F32, I8 (+ I16 on the ND grid)
Layout	PKD3, PLN3, PLN1, plus fused layout conversions (e.g. PKD3toPLN1)
ROI	Full, Partial
Size	default, odd-width tail, sub-vector-width, and per-op edge shapes

The shape presets are chosen to land the width at different remainders against SIMD and HIP block widths, so vectorized main loops and scalar tails are both exercised.

Structured, greppable names

{Domain}_{Category}/{Op}Test.{Intent}/{Backend}_{DType}_{Layout}_{Roi}_{Size}[_{OpParams}]

Image_Color/BrightnessTest.Correctness/HIP_U8toU8_PKD3_FullRoi_2x36x48_a1p75_b50

Every axis is independently selectable with --gtest_filter, so "all F32 PKD3 cases" or "HIP only" is a one-liner.

Known-defect skip list

framework/skip_list.hpp holds ~61 filter globs for cases that fail against a real kernel defect (or, in a few cases, a non-reproducible result). Test bodies assert correct behaviour unconditionally — the list is the single place recording which assertions don't hold today, so fixing a kernel means deleting its entry in the same change. RPP_TEST_NO_SKIP_LIST=1 runs them anyway to find entries that have become obsolete.

The suite has already surfaced several genuine defects this way, including a non-deterministic spatter result and layout-specific failures in flip_voxel.

Reporting

A custom console reporter prints one line per suite plus full detail for each failure, which keeps a 24k-case run readable. It replaces GTest's console printer only — --gtest_list_tests and --gtest_output=xml|json are unaffected, so CI and IDE integrations work as with any GTest binary. RPP_TEST_PLAIN_OUTPUT=1 restores stock GTest output.

Building and running

mkdir build && cd build && cmake .. && cmake --build . --parallel
./rpp_tests                     # everything
./rpp_tests --gtest_filter='Image_Color/*'
ctest                           # one process per case; slower, but isolates GPU crashes

Configures standalone against an installed RPP (found under ROCM_PATH) or in-tree as a subdirectory of an RPP build. GoogleTest is fetched at configure time, pinned by tag; -DRPP_TEST_USE_SYSTEM_GTEST=ON uses an installed copy for network-isolated builds. HIP cases are compiled only when the RPP being built against has the HIP backend. The suite's own sources build with -Wall -Wextra -Werror.

Notes for reviewers

• No RPP library code changes — this PR adds tests only.
• The bulk of the line count is the 83 reference models and the per-op test instantiations, which
  are repetitive by design. The parts worth close review are src/framework/ (the grid, memory
  handling, and comparators that every test depends on) and the skip list's coverage.
• Ops not yet covered, and the mixed-precision (U8->F32) dtype axis, are follow-up work; the
  framework already has the seams for both.

[zacharyvincze]

Addressed documentation issues.

[rrawther]

Approving so we can migrate and fix the remaining tests

[spolifroni-amd]

Wondering if this needs a line in the changelog.

[assistant-librarian]

TheRock Submodule Bump Activity

Newest first

• 2026-10-22 — Included in TheRock via ROCm/TheRock#9426.