threadpool: make ggml_barrier play nice with the thread sanitizer

ggml/src/ggml.c

@@ -3198,16 +3198,27 @@ static void ggml_barrier(struct ggml_threadpool * threadpool) {
 
     int passed_old = atomic_load_explicit(n_barrier_passed, memory_order_relaxed);
 
-    if (atomic_fetch_add(n_barrier, 1) == n_threads - 1) {
+    // All threads go through the full fence (memory barrier) operation once to ensure
+    // that all previos updates have completed.
+    // The rest of the reads and writes can be relaxed, but the thread sanitizer wants
+    // to see an explicit acquire / release sequence to declare all futher accesses
+    // as safe.
+
+    #if defined(__has_feature) && __has_feature(thread_sanitizer)
+    #define passed_acquire memory_order_acquire
+    #define passed_release memory_order_release
+    #else
+    #define passed_acquire memory_order_relaxed
+    #define passed_release memory_order_relaxed
+    #endif
+
+    if (atomic_fetch_add_explicit(n_barrier, 1, memory_order_seq_cst) == n_threads - 1) {
         // last thread
-        atomic_store(n_barrier, 0);
-        atomic_fetch_add_explicit(n_barrier_passed, 1, memory_order_relaxed);
+        atomic_store_explicit(n_barrier, 0, memory_order_relaxed);
+        atomic_fetch_add_explicit(n_barrier_passed, 1, passed_release);
     } else {
         // wait for other threads
-        while (true) {
-            if (atomic_load_explicit(n_barrier_passed, memory_order_relaxed) != passed_old) {
-                return;
-            }
+        while (atomic_load_explicit(n_barrier_passed, passed_acquire) == passed_old) {
             ggml_thread_cpu_relax();
         }
     }