lowoha_params.num_threads Parameter not properly used in multiple locations

[vishalMCE]

The lowoha_params structure includes a num_threads field, which is used for optimization decisions (like selecting tile sizes and partitioning strategies, ... etc), but it's not passed to the actual thread management in several parallel code paths. When I set params.num_threads = 2, the optimization logic sees and uses this value, but the backend library is still using all available CPU cores

Problem
I'm trying to control thread for LOWOHA matmul, but setting params.num_threads has no effect. The code always falls back to omp_get_max_threads() regardless of what I set

There are many places in the code where thread control is missing. Here are 4 examples:

1. zendnnl_parallel_for utility function
   File: lowoha_matmul_utils.hpp:41
2. Batch parallel partitioning (high FLOPS path)
   File: lowoha_matmul.cpp:253
3. Batch parallel partitioning (low FLOPS path)
   File: lowoha_matmul.cpp:298
4. AOCL-DLP and BLIS calls
   File: aocl_kernel.cpp:1050-1095

AOCL and BLIS libraries use OpenMP internally and respect omp_set_num_threads(), but we never call it before invoking their GEMM functions

// run_blis function - no thread control!
aocl_gemm_f32f32f32of32(...);
aocl_gemm_bf16s4f32of32(...);
aocl_gemm_bf16bf16f32obf16(...);

Suggested Fix

• for OpenMP loops

  #pragma omp parallel for collapse(2) num_threads(num_threads)

• for parallel-for utility

  template <class F>
  inline void zendnnl_parallel_for(const int64_t begin, const int64_t end,
                                   const int64_t grain_size, 
                                   const int64_t max_num_threads,  // Add this
                                   const F &f);

• for AOCL/BLIS

  void run_blis(...) {
    omp_set_num_threads(lowoha_param.num_threads);  // Add this
    ...
  }

Change default behavior
Consider changing the default num_threads value in constructor from 0 to 1:

// in lowoha_common.hpp
lowoha_params() : dtypes(), postop_(), quant_params(), mem_format_a('n'),
    mem_format_b('n'), lowoha_algo(matmul_algo_t::none), num_threads(1) {}  // Changed from 0 to 1

This makes more sense when the thread count isn’t explicitly set; it defaults to single-threaded rather than using all available cores

Additional Info

• Component: LOWOHA matmul operators
• Affects: All backends (AOCL-DLP, BLIS, LibXSMM, oneDNN)

The purpose of this issue is to enable API-level thread control. Currently, threads can only be set via OMP_NUM_THREADS env variable; having params.num_threads work properly would allow per-call thread control

[navaneeth108]

Thanks for the detailed report and suggestions!

Please check out commit b2bb95a3a145db0c6291348b356e78e94971c6c2, which addresses this issue.

With this update, you can now control the number of threads for any LOWOHA operation using the params. num_threads variable, which will override the system default settings.

Key changes:

• The num_threads parameter is now properly propagated to all parallel code paths
• Thread control works consistently across all backends (AOCL-DLP, BLIS, LibXSMM, oneDNN)
• Setting params. num_threads will now take precedence over OMP_NUM_THREADS environment variable

Thanks

[vishalMCE]

Hi @navaneeth108, thanks for the update
It looks like the commit you referenced hasn’t been open-sourced yet. Will wait for it to be available publicly and will be happy to close the issue :)

[navaneeth108]

Hi @vishalMCE ,
I might have previously shared the internal commit ID. Please refer to the public one instead and let me know if everything looks good: 0c0ef7d.
Thanks!

[vishalMCE]

Thanks @navaneeth108 for the fix, now it's working as expected