feat(rocshmem): rework the tile-tests by edgargabriel · Pull Request #11770 · ROCm/rocm-systems

edgargabriel · 2026-09-17T15:53:08Z

Motivation

Rework tile tests to cover a wider range of possible configurations, and produce performance numbers that are more easily comparable to the non-tile variants of the tests.

Technical Details

tile_broadcast / tile_allgather / tile_reduce wave & WG tests: enable using multiple waves in a wg and have each wave use a separate context. Matches the changes made to the wave-collectives last months. driver.sh and CTest updated to use --num-wf 4 with ROCSHMEM_MAX_NUM_CONTEXTS = NUM_WGS × NUM_WF.
Tile RMA size-configurable sweeps: Wired up max_msg_size in the tile RMA tester so tile dimensions scale with -s; the message sweep now runs from one column to the full tile (one valid tile shape per data point) with accurate volume and bandwidth reporting. Added 1 MB variants to driver.sh and CTest for all tile put/get tests.
New wave-level row-major and col-major tile RMA tests
New wg-level row-major and col-major tile RMA tests

Update:

a performance fix for tile-rma-wg operations with the IPC backend is also included now as part of the PR. The problem became obvious when doing performance comparisons between the various patterns.
ammend the existing parameter checks to ensure that we raise an error if the dimensionality of the tile(s) exceeds 2, since we only support 1-D and 2-D tiles at the moment.

Issue Tracking

JIRA ID: AIROCSHMEM-517

Submission Checklist

Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.

therock-pr-bot · 2026-09-17T15:53:46Z

✅ All Policy Checks Passed

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	⚠️ Warning	Error: Source/code files changed without an accompanying unit test. Expected: add at least one test file named like `test_<name>.py` / `test_<name>.cpp` (or `<name>_test.`). Current:* code file(s) changed: `projects/rocshmem/src/gda/context_gda_tmpl_device.hpp`, `projects/rocshmem/src/ipc/context_ipc_tmpl_device.hpp`, `projects/rocshmem/src/tile_memcpy_device.hpp`, `projects/rocshmem/tests/functional_tests/tester.cpp`, `projects/rocshmem/tests/functional_tests/tester.hpp` (+8 more); no test file found
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

🎉 All policy checks passed!

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

Copilot

🟡 Changes recommended

Unresolved critical and moderate correctness, timing, sizing, and resource issues remain.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

This PR reworks rocSHMEM tile functional tests with multi-wave collectives, configurable RMA sweeps, additional layouts, and expanded test coverage.

Changes:

Adds wave/WG row-major and column-major RMA variants.
Updates broadcast, allgather, and reduce tests for multiple wave contexts.
Expands CTest and driver coverage, including 1 MiB variants.

File summaries

File	Description and final review comments
`projects/rocshmem/tests/functional_tests/tile_rma_tester.hpp`	Adds configurable tile dimensions.
`projects/rocshmem/tests/functional_tests/tile_rma_tester.cpp`	Implements layouts, size sweeps, verification, and accounting. Critical · 3 votes: Row-major wave slots use an incorrect stride, causing overlapping source/destination slices. Critical · 2 votes: Wave verification uses the service-thread count and can skip verification entirely. Moderate · 2 votes: Arbitrary layouts cannot exercise the advertised 1 MiB payload. Moderate · 1 vote: Requests below one column silently become 16 KiB tests. Moderate · 1 vote: End timestamps can precede completion of all quiet operations.
`projects/rocshmem/tests/functional_tests/tile_reduce_tester.hpp`	Updates the wave reduction interface.
`projects/rocshmem/tests/functional_tests/tile_reduce_tester.cpp`	Adds multi-wave contexts and teams. Moderate · 1 vote: Timing buffers and message counts are not populated, producing invalid performance results.
`projects/rocshmem/tests/functional_tests/tile_broadcast_tester.hpp`	Updates the wave broadcast interface.
`projects/rocshmem/tests/functional_tests/tile_broadcast_tester.cpp`	Adds multi-wave broadcast handling. Moderate · 1 vote: Timing buffers and message counts are not populated, producing invalid performance results.
`projects/rocshmem/tests/functional_tests/tile_allgather_tester.hpp`	Updates the wave allgather interface.
`projects/rocshmem/tests/functional_tests/tile_allgather_tester.cpp`	Adds multi-wave allgather handling. Moderate · 1 vote: Timing buffers and message counts are not populated, producing invalid performance results.
`projects/rocshmem/tests/functional_tests/tester.hpp`	Registers new RMA test types.
`projects/rocshmem/tests/functional_tests/tester.cpp`	Creates testers for the new variants.
`projects/rocshmem/tests/functional_tests/CTestFunctionalTests.cmake`	Registers expanded test configurations. Moderate · 3 votes: WG tests multiply context allocation by `NUM_WF` despite using one context per WG, potentially exhausting backend resources.
`projects/rocshmem/scripts/functional_tests/driver.sh`	Adds driver mappings and test invocations. Moderate · 1 vote: New WG 1 MiB cases can allocate approximately 1–2 GiB per buffer and exhaust the 6 GiB heap.

Review details

Suppressed comments (6)

projects/rocshmem/scripts/functional_tests/driver.sh:1045

The new WG 1 MiB row-major cases pass -z $((WAVE_SIZE * 16)); TileRMATester allocates args.num_threads slots, so on a 64-lane GPU each row-major slot is 2 MiB and each of the symmetric source/destination buffers is about 2 GiB even though a WG collective uses one tile. The corresponding col-major/contiguous cases are also about 1 GiB per buffer, which can exhaust the 6 GiB heap and makes the new driver sweep unnecessarily large. Allocate by collective count or use a wave-count path that does not inflate the host allocation count.

  ExecTest  "tile_put_wg_rowmajor"      2       1            $((WAVE_SIZE * 16))
  ExecTest  "tile_put_wg_rowmajor"      2       1            $((WAVE_SIZE * 16)) 1048576

projects/rocshmem/tests/functional_tests/tile_allgather_tester.cpp:348

This new wave/WG launch path still does not populate start_time/end_time or num_timed_msgs. Tester::print() subsequently reads the uninitialized timer buffers and computes zero message rate/bandwidth, so these tests cannot produce the comparable performance numbers described by the PR. Populate the tester timing/count contract or use the HIP event timing for these kernels.

      hipLaunchKernelGGL(TileAllgatherWaveTest, gridSize, blockSize,
                         wave_shared, stream,
                         teams, source, dest, tile_extent_0, tile_extent_1,
                         args.myid, n_pes, _shmem_context,
                         wf_size, num_warps, error_flag);

projects/rocshmem/tests/functional_tests/tile_broadcast_tester.cpp:329

This new wave/WG launch path still does not populate start_time/end_time or num_timed_msgs. Tester::print() subsequently reads the uninitialized timer buffers and computes zero message rate/bandwidth, so these tests cannot produce the comparable performance numbers described by the PR. Populate the tester timing/count contract or use the HIP event timing for these kernels.

      hipLaunchKernelGGL(TileBroadcastWaveTest, gridSize, blockSize,
                         wave_shared, stream,
                         teams, source, dest, tile_extent_0, tile_extent_1,
                         args.myid, pe_root, _shmem_context,
                         wf_size, num_warps, error_flag);

projects/rocshmem/tests/functional_tests/tile_reduce_tester.cpp:526

This new wave/WG launch path still does not populate start_time/end_time or num_timed_msgs. Tester::print() subsequently reads the uninitialized timer buffers and computes zero message rate/bandwidth, so these tests cannot produce the comparable performance numbers described by the PR. Populate the tester timing/count contract or use the HIP event timing for these kernels.

      hipLaunchKernelGGL(TileReduceWaveTest, gridSize, blockSize,
                         wave_shared, stream,
                         teams, source, sum_dest, max_dest, min_dest,
                         short_source, short_sum_dest, short_max_dest,
                         short_min_dest, int_source, int_sum_dest,

projects/rocshmem/tests/functional_tests/tile_rma_tester.cpp:359

When -s is smaller than one 64-row column (256 bytes), derived is zero and this leaves tile_extent_1 at the default 64. The tester then changes the requested cap into a 16 KiB test instead of honoring it or using the smallest valid 256-byte tile. Clamp the extent to one column (or reject the invalid request) rather than silently falling back to 64.

  if (args.max_msg_size_set) {
    int derived = static_cast<int>(args.max_msg_size / (tile_extent_0 * sizeof(float)));
    if (derived >= 1) {
      tile_extent_1 = derived;
    }

projects/rocshmem/tests/functional_tests/tile_rma_tester.cpp:709

These new wave cases report one message per wave, but TileRMATest records end_time[wg_id] from every thread immediately after only thread 0 calls rocshmem_ctx_quiet; there is no post-quiet barrier. A nonzero thread can therefore publish an end time before the quiet has completed all wave operations, overstating bandwidth for these newly reported multi-wave counts. Record the end timestamp after a WG-wide barrier following quiet, from one thread.

    case TilePutWaveRowMajorTestType:
    case TilePutWaveColumnMajorTestType:
    case TileGetWaveRowMajorTestType:
    case TileGetWaveColumnMajorTestType:
      tiles_per_loop = gridSize.x * num_warps;

Files reviewed: 12/12 changed files
Comments generated: 4
Review effort level: Lite

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

ddebonis-amd

Missing tile reduce tests for types (long / int / short) and ops (min / max). Only testing float / sum.

ddebonis-amd

Looks good.

…ams and contexts ## Motivation Apply the same per-wave context pattern introduced in dd5b69b for the wave-level collective testers to the tile_broadcast wave and wg tests. ## Technical Details tile_broadcast_tester.cpp / .hpp: - TileBroadcastWaveTest: replaced single __shared__ rocshmem_ctx_t ctx + single team with extern __shared__ rocshmem_ctx_t ctx_array[] sized to num_waves_per_wg; each wave creates its own context from a dedicated team via rocshmem_wg_team_create_ctx and operates on an independent source/dest buffer slice indexed by flat_wf_id (wg_id * num_waves_per_wg + wf_id); all waves now execute concurrently - TileBroadcastTest (WG-level): migrated from __shared__ ctx to extern __shared__ ctx_array[0] for consistency with the pattern - Constructor sizes the teams array to num_wgs * num_warps (one team per wave, the maximum needed across all test types) - preLaunchKernel/postLaunchKernel: create and destroy the correct number of teams per test type (wave: num_wgs * num_warps, WG: num_wgs, thread: 1) - launchKernel: passes num_warps * sizeof(rocshmem_ctx_t) as dynamic shared memory for the wave test; sizeof(rocshmem_ctx_t) for the WG test driver.sh: - tile_broadcast_wave runs with --num-wf 4 (4 waves per WG) and ROCSHMEM_MAX_NUM_CONTEXTS = NUM_WGS * 4, matching the per-wave context requirement; the 4-PE variant now also uses 4 WGs - export/unset pattern mirrors the existing wave collective tests CTestFunctionalTests.cmake: - _add_single_rocshmem_test: compute ROCSHMEM_MAX_NUM_CONTEXTS as WORKGROUPS * NUM_WF when NUM_WF is set, instead of just WORKGROUPS; this correctly handles all current and future wave-level tests - tile_broadcast_wave entries updated to NUM_WF 4, with the 4-PE variant using WORKGROUPS 4 (yielding MAX_NUM_CONTEXTS=16) Co-Authored-By: Claude <noreply@anthropic.com>

…ams and contexts Apply the same per-wave context pattern from dd5b69b and the tile_broadcast rework to the tile_allgather wave and wg tests. tile_allgather_tester.cpp / .hpp: - TileAllgatherWaveTest: replaced single __shared__ rocshmem_ctx_t ctx + single team with extern __shared__ rocshmem_ctx_t ctx_array[] sized to num_waves_per_wg; each wave creates its own context from a dedicated team and operates on independent buffer slices indexed by flat_wf_id: source at flat_wf_id * tile_size, dest at flat_wf_id * tile_size * n_pes - TileAllgatherTest (WG-level): migrated from __shared__ ctx to extern __shared__ ctx_array[0] - Constructor sizes teams array to num_wgs * num_warps; allocates source and dest with num_teams slots to cover all wave slots - preLaunchKernel/postLaunchKernel: create/destroy teams per test type (wave: num_wgs * num_warps, WG: num_wgs, thread: 1) - launchKernel: passes dynamic shared memory sized to num_warps * sizeof(rocshmem_ctx_t) for wave test; sizeof(rocshmem_ctx_t) for WG test - resetBuffers: initializes source as myid * 100 + slot * 1000 + i so each wave slot carries a distinct value; verification checks the same driver.sh: - tile_allgather_wave runs with --num-wf 4 and ROCSHMEM_MAX_NUM_CONTEXTS = NUM_WGS * 4; 4-PE variant uses 4 WGs CTestFunctionalTests.cmake: - tile_allgather_wave entries updated to NUM_WF 4, 4-PE variant uses WORKGROUPS 4 (MAX_NUM_CONTEXTS computed as 4 and 16 respectively by the existing WORKGROUPS * NUM_WF formula) Co-Authored-By: Claude <noreply@anthropic.com>

… and contexts Apply the same per-wave context pattern from dd5b69b and the previous tile_broadcast/allgather reworks to the tile_reduce wave and wg tests. tile_reduce_tester.cpp / .hpp: - TileReduceWaveTest: replaced single __shared__ rocshmem_ctx_t ctx + single team with extern __shared__ rocshmem_ctx_t ctx_array[] sized to num_waves_per_wg; each wave creates its own context and operates on an independent buffer slice (offset = flat_wf_id * matrix_size) for all four data types (float, short, int, long); verification runs per-wave on the root PE using t_id % wf_size == 0 - TileReduceTest (WG-level): migrated from __shared__ ctx to extern __shared__ ctx_array[0] - Constructor: sizes num_teams to num_wgs * num_warps for wave tests, num_wgs for WG tests, 1 for thread tests, so one allocation covers all - preLaunchKernel/postLaunchKernel: simplified to a single loop over num_teams (already set correctly per test type in the constructor) - launchKernel: passes num_warps * sizeof(rocshmem_ctx_t) dynamic shared memory for wave test; sizeof(rocshmem_ctx_t) for WG test; wave test now launches gridSize blocks instead of dim3(1) driver.sh: - tile_reduce_wave runs with --num-wf 4 and ROCSHMEM_MAX_NUM_CONTEXTS = NUM_WGS * 4; 4-PE variant uses 4 WGs CTestFunctionalTests.cmake: - tile_reduce_wave entries updated to NUM_WF 4, 4-PE variant uses WORKGROUPS 4; tile_reduce_wg entries updated to NUM_WF 1 WORKGROUPS 4 Co-Authored-By: Claude <noreply@anthropic.com>

…gurable sweeps ## Motivation The tile RMA tests (tile_put/get_*) previously used hardcoded 64×64 tile dimensions and ignored the -s / MAX_MSG_SIZE argument entirely, making it impossible to sweep data sizes from the driver or CTest. Reported bandwidths were meaningless for sub-tile message sizes since the kernel always transferred the full tile regardless. ## Technical Details tile_rma_tester.hpp: - Added tile_extent_0 and tile_extent_1 as protected member fields tile_rma_tester.cpp: - Constructor derives tile_extent_1 from args.max_msg_size when set (max_msg_size = tile_extent_0 * tile_extent_1 * sizeof(float)) - tile_extent_0 is fixed at 64 rows; tile_extent_1 scales with message size - Arbitrary-stride layout clamps tile_extent_1 to 85 (floor(257/3)) to keep column accesses within the fixed row_stride=257 - Sets this->args.min_msg_size = one column (tile_extent_0 * sizeof(float)) and max_msg_size = full tile bytes, so the framework sweep runs from one column up to the full tile — every data point is a valid tile shape - launchKernel: derives ke_tile_extent_1 from the current iteration's size so each sweep point transfers the correctly-sized tile and reports accurate bandwidth - resetBuffers: re-initializes the source buffer using the current iteration's tile dimensions (ke_tile_extent_1-based offsets) so that source layout matches exactly what the kernel expects for each sweep point - verifyResults: re-derives tile dimensions from size using the same logic as launchKernel so only the actually-transferred elements are checked - Buffer allocation in the constructor uses the max tile_extent_1 to cover all sweep points within a single allocation driver.sh: - Added a 1MB-max variant alongside each existing tile RMA test; sweeps from 256B (1 column) to 1048576B (4096 columns) CTestFunctionalTests.cmake: - Added MAX_MSG_SIZE 1048576 variants for all tile put/get test groups Co-Authored-By: Claude <noreply@anthropic.com>

…t and tile_get The tile RMA wave-level tests previously covered only the contiguous layout. Row-major and col-major variants exist at the thread level but were missing for the wave-collective operations, preventing performance comparisons across memory layouts at the wave level. tester.hpp: - Added 4 new enum entries appended after the last existing entry (SdmaPutNbi=161): TilePutWaveRowMajor=162, TilePutWaveColumnMajor=163, TileGetWaveRowMajor=164, TileGetWaveColumnMajor=165 tile_rma_tester.cpp: - Offset calculation: new types added to the wave-collective branch (offset = matrix_size * flat_wave_id) - Kernel branches: 4 new if-constexpr blocks using rocshmem_ctx_tile_put_wave / rocshmem_ctx_tile_get_wave with row-major (row_stride=2*t1) and col-major (row_stride=1, col_stride=t0) tensor layouts - Constructor: new types routed to correct buffer_elements_per_thread cases, source/dest assignment (put→local/remote, get→remote/local) - resetBuffers: new types added to the appropriate stride/slot-size cases - launchKernel: 4 new LAUNCH_TILE_RMA_TEST dispatch cases; new types added to the wave tiles_per_loop branch (gridSize.x * num_warps) - verifyResults: new types added to check_id, row/col_stride, num_tiles_transferred, and buffer_elements_per_thread switches tester.cpp: - 4 new TileRMATester instantiation cases in Tester::create() driver.sh: - Added TEST_NUMBERS entries (162-165) - Added ExecTest calls with default and 1MB size variants alongside the existing wave contiguous tests CTestFunctionalTests.cmake: - Added set(TEST_tile_*) entries for the 4 new test numbers - Added add_rocshmem_functional_test entries (default + MAX_MSG_SIZE 1048576) in the wave put and get test groups Co-Authored-By: Claude <noreply@anthropic.com>

…wg and tile_get_wg Apply the same pattern from the wave variant commit (ddde4d522e) to WG-collective tile RMA operations. 4 new enum entries appended after TileGetWaveColumnMajor (165): TilePutWGRowMajor=166, TilePutWGColumnMajor=167, TileGetWGRowMajor=168, TileGetWGColumnMajor=169 tile_rma_tester.cpp: offset uses wg_id; kernel branches use rocshmem_ctx_tile_put_wg / rocshmem_ctx_tile_get_wg with row-major (row_stride=2*t1) and col-major (row_stride=1, col_stride=t0) layouts; num_timed_msgs uses gridSize.x (one tile per WG per loop); all switch statements updated consistently across offset, buffer sizing, source/dest assignment, resetBuffers, launchKernel, and verifyResults. tester.cpp: 4 new TileRMATester instantiation cases. driver.sh: TEST_NUMBERS entries 166-169; ExecTest calls with default and 1MB size variants using $((WAVE_SIZE * 16)) threads (matching existing tile_put_wg_contiguous pattern). CTestFunctionalTests.cmake: set() entries and add_rocshmem_functional_test entries (NUM_WF 16, default + MAX_MSG_SIZE 1048576) in the WG put/get groups. Co-Authored-By: Claude <noreply@anthropic.com>

tile_memcpy_contig<Kind, TileScope::Wg> was calling memcpy_lane through thread 0 only, leaving the other 1023 threads idle for the contiguous fast path. Replace with memcpy_wg, which distributes the transfer across all threads in the workgroup — matching the pattern already used by the Wave scope (memcpy_wave) and the RowContig/ColContig strided paths. Observed impact on tile_put_wg_contiguous / tile_get_wg_contiguous at 1MB with 1 WG / 1024 threads: Before: ~6000 µs / 0.16 GB/s (serialized through thread 0) After: ~44 µs / 22 GB/s (all 1024 threads cooperate) Co-Authored-By: Claude <noreply@anthropic.com>

The 1MB sweep (min_msg_size=256B, max_msg_size=1MB) already covers all data points of the default no-size run (max_msg_size=16384B), making the latter redundant. Remove the no-size ExecTest calls from driver.sh and the corresponding add_rocshmem_functional_test entries from CTestFunctionalTests.cmake for all tile put/get tests. Co-Authored-By: Claude <noreply@anthropic.com>

…testers ## tile_broadcast and tile_allgather testers - Constructor: derive tile_extent_1 from max_msg_size; set min_msg_size = tile_extent_0 * sizeof(float) (one column) and max_msg_size = full tile so the framework sweeps from one column to the configured maximum - Constructor: fix num_timers for wave variant (num_wgs * num_warps) - Kernels: add loop/skip timing with wall_clock64; start_time/end_time arrays now instrumented for all three variants (thread, wave, WG) - launchKernel: derive ke_tile_extent_1 from current iteration's size and pass it to the kernel; set num_msgs/num_timed_msgs correctly - resetBuffers: reinitialize source using ke_tile_extent_1-sized slot strides so source layout matches what the kernel expects each sweep step - verifyResults: print PASSED only at the last sweep iteration (size == max_msg_size); exit on first failure covers all prior iterations ## tile_reduce tester — full template refactor Replaced the monolithic TileReduceTester (4 types, 3 ops, all timed together) with TileReduceTester<T, Op> templated on element type and reduction operation, matching the pattern used by TeamReductionTester: - TILE_REDUCE_DEF_GEN macro specialises tile_reduce_lane/wave/wg<T,Op> for float/short/int/long × SUM/MAX/MIN dispatching to the typed tile reduce API - Three kernel templates (TileReduceThreadTest, TileReduceWaveTest, TileReduceWGTest) each time exactly one type+op in a loop - Host class: same constructor sweep-pin and ke_tile_extent_1 pattern as broadcast/allgather; verifyResults checks result on root PE host-side - tester.cpp: instantiates TileReduceTester<float, ROCSHMEM_SUM> for all three test types with init (s=1, r=0) and verify (v == n_pes) lambdas - CMakeLists.txt: tile_reduce_tester.cpp removed from direct compilation (it is included by the header as a template implementation) ## driver.sh and CTestFunctionalTests.cmake - All tile collective tests (broadcast, allgather, reduce) now run up to 1MB message size - Wave collective variants use 1 WG with NUM_WF=4 for both 2-PE and 4-PE cases (was 4 WGs for 4-PE case); ROCSHMEM_MAX_NUM_CONTEXTS = 1 * 4 = 4 - Stale old-style reduce size entries (262144B, 2048B) replaced uniformly with 1048576B across all three reduce variants Co-Authored-By: Claude <noreply@anthropic.com>

Both the IPC and GDA backends previously silently ignored ndim values outside {1, 2}: tile_memcpy_rma and tile_make_view have no else branch for ndim > 2, resulting in a no-op (no data transferred, no error). The reduce inner impls rejected ndim <= 0 but not ndim > 2. Add explicit ndim < 1 || ndim > 2 guards at the entry point of all six RMA functions in each backend (tile_put, tile_put_wave, tile_put_wg, tile_get, tile_get_wave, tile_get_wg), returning ROCSHMEM_ERROR immediately. Collectives (allgather, broadcast) propagate naturally since they delegate to the guarded RMA functions internally. Extend the existing ndim <= 0 check in both backends' inner reduce implementations to ndim < 1 || ndim > 2, covering the previously unhandled ndim > 2 case. Co-Authored-By: Claude <noreply@anthropic.com>

…nsts Replace bare integer literals used as tile dimensions with named constexpr members in each tester class: - TileRMATester: DEFAULT_TILE_ROWS=64, DEFAULT_TILE_COLS=64, ARBITRARY_MAX_COLS=85 (floor(257/3), the max col count for the arbitrary-stride layout where row_stride=257 and col_stride=3) - TileBroadcastTester: DEFAULT_TILE_ROWS=8, DEFAULT_TILE_COLS=8 - TileAllgatherTester: DEFAULT_TILE_ROWS=8, DEFAULT_TILE_COLS=8 - TileReduceTester: DEFAULT_TILE_ROWS=8, DEFAULT_TILE_COLS=8 Co-Authored-By: Claude <noreply@anthropic.com>

…ss tests The template refactor reduced each tile_reduce test type to a single TileReduceTester<float, ROCSHMEM_SUM> instance, dropping the short/int/long types and MAX/MIN operations that the previous monolithic tester covered. Restore full coverage by pushing 12 instances per test type (float/short/int/long × SUM/MAX/MIN) using a TILE_REDUCE_PUSH macro. Verification uses simple constant source values (s=1, r=0 for sum/max; r=2 for min) so expected results are type-safe and unambiguous: SUM → v == n_pes, MAX → v == 1, MIN → v == 1 Co-Authored-By: Claude <noreply@anthropic.com>

Running 12 type+op combinations per test invocation makes the 1MB 4-PE reduce tests significantly slower. Cap them at 65536B to keep the suite runtime reasonable while still sweeping through all message sizes up to 64KB (256B columns → 16KB tile at the maximum). Co-Authored-By: Claude <noreply@anthropic.com>

systems-assistant · 2026-09-29T07:34:33Z

TheRock Submodule Bump Activity

Newest first

2026-09-29 — Included in TheRock via ROCm/TheRock#8567.

Copilot AI lite review requested due to automatic review settings September 17, 2026 15:53

edgargabriel requested a review from a team as a code owner September 17, 2026 15:53

github-actions Bot added the project: rocshmem label Sep 17, 2026

Copilot started reviewing on behalf of edgargabriel September 17, 2026 15:53

View session

Copilot AI reviewed Sep 17, 2026

View reviewed changes

systems-assistant Bot added the organization: ROCm label Sep 17, 2026

edgargabriel force-pushed the users/egabriel/tile-tests-ctx-per-wave-wg branch 6 times, most recently from 3d022f8 to d474016 Compare September 23, 2026 17:11

ddebonis-amd requested changes Sep 24, 2026

View reviewed changes

edgargabriel force-pushed the users/egabriel/tile-tests-ctx-per-wave-wg branch from d474016 to d2172ed Compare September 24, 2026 15:10

ddebonis-amd self-requested a review September 25, 2026 15:43

ddebonis-amd approved these changes Sep 25, 2026

View reviewed changes

edgargabriel and others added 13 commits September 25, 2026 17:17

edgargabriel force-pushed the users/egabriel/tile-tests-ctx-per-wave-wg branch from d2172ed to 4b61dba Compare September 25, 2026 17:46

edgargabriel merged commit c716bb0 into develop Sep 28, 2026
80 of 82 checks passed

edgargabriel deleted the users/egabriel/tile-tests-ctx-per-wave-wg branch September 28, 2026 13:40

kuhar mentioned this pull request Sep 29, 2026

[rocjitsu] Fix floating-point and operand semantics #12467

Merged

Navigation Menu

Repository navigation

feat(rocshmem): rework the tile-tests - #11770

feat(rocshmem): rework the tile-tests#11770
edgargabriel merged 13 commits into
developfrom
users/egabriel/tile-tests-ctx-per-wave-wg

edgargabriel commented Sep 17, 2026 •

edited

Loading

Uh oh!

therock-pr-bot Bot commented Sep 17, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ddebonis-amd left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ddebonis-amd left a comment

Uh oh!

Uh oh!

systems-assistant Bot commented Sep 29, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

edgargabriel commented Sep 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Technical Details

Issue Tracking

Submission Checklist

Uh oh!

therock-pr-bot Bot commented Sep 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

✅ All Policy Checks Passed

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Suppressed comments (6)

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ddebonis-amd left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ddebonis-amd left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

systems-assistant Bot commented Sep 29, 2026

TheRock Submodule Bump Activity

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

edgargabriel commented Sep 17, 2026 •

edited

Loading

therock-pr-bot Bot commented Sep 17, 2026 •

edited

Loading