fix potential hang in capi initialization

dpctl_capi singleton intiailization could cause deadlocks with updated order manager

Author: ndgrigorian

dpctl/apis/include/dpctl4pybind11.hpp

@@ -26,9 +26,11 @@
 #pragma once
 
 #include "dpctl_capi.h"
+#include <atomic>
 #include <complex>
 #include <cstddef> // for std::size_t for C++ linkage
 #include <memory>
+#include <mutex>
 #include <pybind11/pybind11.h>
 #include <stddef.h> // for size_t for C linkage
 #include <stdexcept>
@@ -142,8 +144,32 @@ class dpctl_capi
 
     static auto &get()
     {
-        static dpctl_capi api{};
-        return api;
+        static std::atomic<dpctl_capi *> global_capi{nullptr};
+        dpctl_capi *capi_ptr = global_capi.load(std::memory_order_acquire);
+
+        // fast path
+        if (capi_ptr) {
+            return *capi_ptr;
+        }
+
+        // double-checked lock
+        // can't use std::call_once here as it still may deadlock
+        static std::mutex init_mtx;
+
+        py::gil_scoped_release release;
+        std::lock_guard<std::mutex> lock(init_mtx);
+
+        capi_ptr = global_capi.load(std::memory_order_relaxed);
+
+        if (!capi_ptr) {
+            py::gil_scoped_acquire acquire;
+
+            capi_ptr = new dpctl_capi();
+
+            global_capi.store(capi_ptr, std::memory_order_release);
+        }
+
+        return *capi_ptr;
     }
 
     py::object default_sycl_queue_pyobj() { return *default_sycl_queue_; }