Merge pull request #5844 from kumaraditya303/develop

fix data races in blas_server and level3_thread

Author: martin-frbg

.github/workflows/dynamic_arch.yml

@@ -200,7 +200,7 @@ jobs:
           - msystem: MINGW32
             idx: int32
             target-prefix: mingw-w64-i686
-            fc-pkg: 
+            fc-pkg:
           - msystem: CLANG64
             idx: int32
             target-prefix: mingw-w64-clang-x86_64
@@ -460,13 +460,12 @@ jobs:
         run: |
           cd build
           if [ "${{ matrix.backend }}" = "tsan" ]; then
-            export OPENBLAS_NUM_THREADS=2
             export LLVM_SYMBOLIZER_PATH=/usr/bin/llvm-symbolizer
             export TSAN_OPTIONS=halt_on_error=1:exitcode=66:second_deadlock_stack=1
           else
-            export OPENBLAS_NUM_THREADS=8
             export OMP_NUM_THREADS=16
           fi
+          export OPENBLAS_NUM_THREADS=8
           ctest -R 'dgemm_thread_safety|dgemm_thread_safety_mixed|dgemv_thread_safety' --output-on-failure
 
 
@@ -687,7 +686,7 @@ jobs:
       - name: Build OpenBLAS
         run: |
           make -j${nproc} TARGET=NEOVERSEN1 USE_OPENMP=1
-  
+
   neoverse_n1_ilp64_build:
     if: "github.repository == 'OpenMathLib/OpenBLAS'"
     runs-on: ubuntu-24.04-arm

driver/level3/level3_thread.c

@@ -98,7 +98,11 @@
 #endif
 
 typedef struct {
+#ifdef HAVE_C11
+  _Atomic
+#else
   volatile
+#endif
    BLASLONG working[MAX_CPU_NUMBER][CACHE_LINE_SIZE * DIVIDE_RATE_MAX];
 } job_t;
 

driver/others/blas_server.c

@@ -631,7 +631,13 @@ int blas_thread_init(void){
      exec_blas       ... returns after jobs are finished.
 */
 
+#if   defined(USE_PTHREAD_LOCK)
+static pthread_mutex_t   exec_queue_lock = PTHREAD_MUTEX_INITIALIZER;
+#elif defined(USE_PTHREAD_SPINLOCK)
+static pthread_spinlock_t exec_queue_lock = 0;
+#else
 static BLASULONG exec_queue_lock = 0;
+#endif
 
 int exec_blas_async(BLASLONG pos, blas_queue_t *queue){
 
@@ -652,7 +658,7 @@ int exec_blas_async(BLASLONG pos, blas_queue_t *queue){
   fprintf(STDERR, "Exec_blas_async is called. Position = %d\n", pos);
 #endif
 
-  blas_lock(&exec_queue_lock);
+  LOCK_COMMAND(&exec_queue_lock);
 
     while (queue) {
       queue -> position  = pos;
@@ -717,7 +723,7 @@ int exec_blas_async(BLASLONG pos, blas_queue_t *queue){
 
     }
 
-    blas_unlock(&exec_queue_lock);
+    UNLOCK_COMMAND(&exec_queue_lock);
 
 #ifdef SMP_DEBUG
     fprintf(STDERR, "Done(Number of threads = %2ld).\n", exec_count);