ROCM-21372: clamp rocSOLVER reset_info launch block counts

[Copilot]

Some rocSOLVER paths compute reset_info grid dimensions directly from problem size or batch count. A few of those launches can produce invalid HIP grid/workgroup configurations on AMD GPUs when the derived block count falls outside the valid range.

• What changed

  • Added a shared helper in lib_device_helpers.hpp to compute reset_info block counts with bounded thread/block dimensions:
    • thread count clamped to [32, 1024]
    • block count clamped to [1, 64*1024 - 1]
  • Switched rocSOLVER reset_info launch sites to use the helper instead of open-coded ceil-division formulas.

• Where this is applied

  • lapack_device_functions.hpp (rocsolver_lacn2_template)
  • rocSOLVER LAPACK, auxiliary, and refactor paths that launch reset_info
  • existing launch thread sizes remain unchanged; only the derived block count is normalized

• Effect

  • avoids zero-block launches
  • avoids oversized grid dimensions for reset_info
  • keeps launch sizing consistent across rocSOLVER call sites

• Example

  auto const nblocks = reset_info_nblocks(batch_count, BS1);
  ROCSOLVER_LAUNCH_KERNEL(reset_info, dim3(nblocks, 1, 1), dim3(BS1, 1, 1), 0, stream,
                          info, batch_count, 0);

Original prompt

Create a pull request in ROCm/rocm-libraries against the develop branch to address ROCM-21372: ensure kernel launches that use reset_info do not use an invalid workgroup/grid configuration.

Context:

• The concern is that some launch configurations for reset_info may produce an invalid workgroup size in HIP on AMD GPUs.
• Relevant files include:
  • projects/rocsolver/library/src/include/lib_device_helpers.hpp (contains the reset_info kernel)
  • projects/rocsolver/library/src/include/lapack_device_functions.hpp (contains at least one caller in rocsolver_lacn2_template that computes blocks for reset_info)
• The repository branch to target is develop.

Requested change:

• Add an adjustment/helper so the number of thread blocks used for launches is always clamped to a valid range.
• The user suggested a helper similar to:

template<typename I>
static inline I ceildiv_block(I const batch_count, I const nthreads_per_block) {
    I const MinThreads = 32;
    I const MaxThreads = 1024;
    I const nthreads = std::max(MinThreads, std::min(MaxThreads, nthreads_per_block));
    I const blocks = (batch_count <= 0) ? 1 : (batch_count - 1) / nthreads + 1;
    I const MinBlocks = 1;
    I const MaxBlocks = 64 * 1024 - 1;
    return std::max(MinBlocks, std::min(MaxBlocks, blocks));
}

and usage like:

auto const nthreads_per_block = BS1;
auto const nblocks = ceildiv_block(batch_count, nthreads_per_block);
ROCSOLVER_LAUNCH_KERNEL(reset_info, dim3(1,1,nblocks), dim3(nthreads_per_block,1,1), 0, stream,
                        info, batch_count, 0);

Implementation guidance:

• Inspect all reset_info launch sites in the repo, especially rocSOLVER paths, and update them to use a valid, clamped block count computation.
• Prefer introducing a reusable helper in an appropriate shared header if that fits project style, rather than duplicating logic at each call site.
• Keep the change minimal and consistent with existing coding patterns.
• Ensure the launch dimensions passed to HIP are valid and avoid zero blocks.
• If there are existing utilities for ceil division or launch-dimension clamping, reuse them instead of introducing redundant helpers.
• Update only what is necessary to fix the invalid launch-configuration concern.

Deliverables:

• Code changes implementing the launch-dimension adjustment.
• A pull request with a concise summary referencing ROCM-21372.

The following is the prior conversation context from the user's chat exploration (may be truncated):

User: Suppose you are an expert GPU programmer for ROCM HIP on AMD GPU, there is a concern from jira ticket ROCM-21372 that there may be invalid workgroup size in launch configuration for kernel reset_info.

Please create a new PR to add an adjustment to the number of thread blocks to be a valid value.

You may consider something like the following

template
static inline I ceildiv_block( I const batch_count, I const nthreads_per_block) {
I const MinThreads = 32; I const MaxThreads = 1024;
I const nthreads = std::max( MinThreads, std::min( MaxThreads, nthreads_per_block));
I const blocks = (batch_count <= 0) ? 1 : (batch_count-1)/nthreads + 1;
I const MinBlocks = 1; I const MaxBlocks = 64 * 1024 - 1;
return std::max( MinBlocks, std::min( MaxBlocks, blocks ));
}

then use as something like

auto const nthreads_per_block = BS1;
auto const nblocks = ceildiv( batch_count, nthreads_per_block);

ROCSOLVER_LAUNCH_KERNEL( reset_info, dim3(1,1,nblocks), dim3(nthreads_per_block,1,1),0,stream, info, batch_count, 0);

Assistant: I’m waiting for your confirmation on the pull request creation request, and once you approve it I’ll continue with the change.

This pull request was created from Copilot chat.

[Copilot]

Pull request overview

Addresses ROCM-21372 by introducing a shared reset_info_nblocks helper that clamps derived grid x-dimension to [1, 65535] (and threads to [32, 1024]) for all reset_info kernel launches in rocSOLVER, eliminating the possibility of zero-block or oversized grid configurations. Because reset_info is implemented as a strided 1D loop, the clamp is correctness-preserving.

Changes:

• Added reset_info_nblocks helper (and <algorithm> include) in lib_device_helpers.hpp.
• Replaced open-coded ceil-division block counts at ~37 reset_info launch sites across LAPACK, auxiliary, and refactor paths.
• Removed the per-launch hipGetDeviceProperties clamp in rocsolver_lacn2_template in favor of the shared helper.

Reviewed changes

Copilot reviewed 49 out of 49 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
library/src/include/lib_device_helpers.hpp	Adds reset_info_nblocks helper and <algorithm> include.
library/src/include/lapack_device_functions.hpp	lacn2 switches from device-property clamp to helper.
library/src/auxiliary/rocauxiliary_bdsqr.hpp	Replaces batch_count/BS1 + 1 with reset_info_nblocks(batch_count + 1, BS1).
library/src/auxiliary/rocauxiliary_bdsvdx.hpp	Uses helper for both info and Dtgk resets.
library/src/auxiliary/rocauxiliary_gecon.hpp	Uses helper for rcond reset.
library/src/auxiliary/rocauxiliary_lasyf.hpp	Uses helper for kb/info reset.
library/src/auxiliary/rocauxiliary_stebz.hpp	Uses helper for reset grid.
library/src/auxiliary/rocauxiliary_stedc.hpp	Uses helper for reset grid.
library/src/auxiliary/rocauxiliary_stedcj.hpp	Uses helper for reset grid.
library/src/auxiliary/rocauxiliary_stedcx.hpp	Uses helper for reset grid.
library/src/auxiliary/rocauxiliary_stein.hpp	Uses helper for reset grid.
library/src/auxiliary/rocauxiliary_steqr.hpp	Uses helper for reset grid.
library/src/auxiliary/rocauxiliary_sterf.hpp	Uses helper for reset grid.
library/src/lapack/roclapack_gels.hpp / _outofplace.hpp	Uses helper for reset grid.
library/src/lapack/roclapack_gesdd.hpp	Uses helper in quick-return path.
library/src/lapack/roclapack_gesv.hpp / _outofplace.hpp	Uses helper for reset grid.
library/src/lapack/roclapack_gesvdj.hpp / _notransv.hpp	Uses helper in quick-return path.
library/src/lapack/roclapack_gesvdx.hpp / _notransv.hpp	Uses helper for info/nsv reset.
library/src/lapack/roclapack_getf2.hpp	Uses helper with static_cast<I>(256).
library/src/lapack/roclapack_getrf.hpp	Uses helper in quick-return path.
library/src/lapack/roclapack_getri.hpp / _outofplace.hpp	Uses helper in quick-return path.
library/src/lapack/roclapack_posv.hpp	Uses helper for reset grid.
library/src/lapack/roclapack_potf2.hpp / roclapack_potrf.hpp	Uses helper for reset grid.
library/src/lapack/roclapack_potri.hpp	Uses helper in quick-return path.
library/src/lapack/roclapack_syev_heev.hpp and syevd/_dx/_dx_inplace/_dj/_jx_jx variants	Uses helper for reset grid.
library/src/lapack/roclapack_syevj_heevj.hpp	Uses helper at two sites; + 1 preserves prior batch_count/BS1 + 1 semantics.
library/src/lapack/roclapack_sygv_hegv.hpp and sygvd/_dj/_dx/_dx_inplace/_j/_x variants	Uses helper for reset grid.
library/src/lapack/roclapack_sytf2.hpp / roclapack_sytrf.hpp	Uses helper in quick-return path.
library/src/lapack/roclapack_trtri.hpp	Uses helper for info reset.
library/src/refact/rocrefact_csrrf_splitlu.hpp	Uses helper for ptrU/ptrL/... resets.
library/src/refact/rocrefact_csrrf_sumlu.hpp	Uses helper with n + 1 to match prior n/BS1 + 1 semantics.

---

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

[codecov-commenter]

Codecov Report

❌ Patch coverage is 98.36066% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...csolver/library/src/include/lib_device_helpers.hpp	90.00%	0 Missing and 1 partial ⚠️

❌ Your project status has failed because the head coverage (77.83%) is below the target coverage (80.00%). You can increase the head coverage or adjust the target coverage.

Additional details and impacted files

@@           Coverage Diff            @@
##           develop    #7914   +/-   ##
========================================
  Coverage    61.43%   61.43%           
========================================
  Files         2091     2091           
  Lines       358852   358857    +5     
  Branches     54266    54265    -1     
========================================
+ Hits        220451   220457    +6     
  Misses      119672   119672           
+ Partials     18729    18728    -1     

Flag	Coverage Δ		*Carryforward flag
TensileLite	27.41% <ø> (ø)		Carriedforward from 9c82e2e
hipBLAS	90.65% <ø> (ø)		Carriedforward from 9c82e2e
hipBLASLt	41.23% <ø> (ø)		Carriedforward from 9c82e2e
hipCUB	82.21% <ø> (ø)		Carriedforward from 9c82e2e
hipDNN	86.59% <ø> (ø)		Carriedforward from 9c82e2e
hipFFT	51.59% <ø> (ø)		Carriedforward from 9c82e2e
hipRAND	76.12% <ø> (ø)		Carriedforward from 9c82e2e
hipSOLVER	69.24% <ø> (ø)		Carriedforward from 9c82e2e
hipSPARSE	86.54% <ø> (ø)		Carriedforward from 9c82e2e
rocBLAS	48.09% <ø> (ø)		Carriedforward from 9c82e2e
rocFFT	45.24% <ø> (ø)		Carriedforward from 9c82e2e
rocRAND	57.03% <ø> (ø)		Carriedforward from 9c82e2e
rocSOLVER	77.83% <98.36%> (+0.01%)	⬆️
rocSPARSE	72.67% <ø> (ø)		Carriedforward from 9c82e2e

*This pull request uses carry forward flags. Click here to find out more.

Files with missing lines	Coverage Δ
...olver/library/src/auxiliary/rocauxiliary_bdsqr.hpp	76.12% <100.00%> (ø)
...lver/library/src/auxiliary/rocauxiliary_bdsvdx.hpp	86.36% <100.00%> (ø)
...olver/library/src/auxiliary/rocauxiliary_gecon.hpp	85.60% <100.00%> (ø)
...olver/library/src/auxiliary/rocauxiliary_lasyf.hpp	73.91% <100.00%> (ø)
...olver/library/src/auxiliary/rocauxiliary_stebz.hpp	81.58% <100.00%> (ø)
...olver/library/src/auxiliary/rocauxiliary_stedc.hpp	55.56% <100.00%> (ø)
...lver/library/src/auxiliary/rocauxiliary_stedcj.hpp	64.10% <100.00%> (ø)
...lver/library/src/auxiliary/rocauxiliary_stedcx.hpp	66.47% <100.00%> (ø)
...olver/library/src/auxiliary/rocauxiliary_stein.hpp	82.22% <100.00%> (ø)
...olver/library/src/auxiliary/rocauxiliary_steqr.hpp	85.03% <100.00%> (ø)
... and 39 more

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.