Inquiry and Suggestions Regarding OpenBLAS Code Flow with OpenMP

Hello,

I've been delving into the OpenBLAS codebase, specifically focusing on the gemm_driver function in level3_thread.c with the #USE_OPENMP=1 flag enabled. I've come across a section in the code where a lock is used in the Pthreads and Win32 backend before initializing and assigning the blasqueue. The lock is released only after the exec_blas call is completed.

My current understanding is :

1. In the case of Pthreads The thread pool is initialized (POSIX), and when a BLAS call is made, the thread pool is utilized to execute it, after which the threads go back to sleep. The use of locks ensures that only one exec_blas call can be executed at a time.

2. If I use a thread pool with No_of_threads < nthreads (number of queue partitions) defined in level3_thread.c, I encounter a deadlock within the inner_threads function.
3. There seems to be significant busy-waiting (NOP) inside the inner_thread function call for thread synchronization.

I have a few questions and want to seek suggestions from the community:

1. I noticed that OpenMP locking mechanisms like omp_set_lock are not used, and instead, busy-waiting is implemented inside the exec_blas function in blas_server_omp.c using max_parallel_number. Could you please provide insights into this choice?

2. The inner_thread function appears to encounter deadlock when used with fewer threads. I tested this in OpenMP by creating a parallel region inside exec_blas in line 437 - blas_server_omp.c with a fixed number of threads less than nthreads. Could you shed some light on the reasons behind this behavior?

3. Considering the busy-waiting in the code, have there been considerations for putting threads to sleep or employing other synchronization methods to enhance efficiency?

I would appreciate clarification on these points.

Start by tracing gemm_driver in driver/level3/level3_thread.c and exec_blas and inner_thread in driver/others/blas_server_omp.c. Reproduce the reported deadlock with fewer OpenMP threads than the queue partitions, then compare the locking and busy-waiting paths. Done would require a confirmed explanation and a clearly scoped synchronization change or documentation outcome.