Synchronization function release GIL

DPCTLQueue_Wait and friends are declared nogil, and their use
is moved into `with nogil` context.

This allows us to use host_task to decrement reference count on
a Python object without introducing a deadlock.

Author: oleksandr-pavlyk

dpctl/_backend.pxd

@@ -227,8 +227,8 @@ cdef extern from "syclinterface/dpctl_sycl_device_selector_interface.h":
 cdef extern from "syclinterface/dpctl_sycl_event_interface.h":
     cdef DPCTLSyclEventRef DPCTLEvent_Create()
     cdef DPCTLSyclEventRef DPCTLEvent_Copy(const DPCTLSyclEventRef ERef)
-    cdef void DPCTLEvent_Wait(DPCTLSyclEventRef ERef)
-    cdef void DPCTLEvent_WaitAndThrow(DPCTLSyclEventRef ERef)
+    cdef void DPCTLEvent_Wait(DPCTLSyclEventRef ERef) nogil
+    cdef void DPCTLEvent_WaitAndThrow(DPCTLSyclEventRef ERef) nogil
     cdef void DPCTLEvent_Delete(DPCTLSyclEventRef ERef)
     cdef _event_status_type DPCTLEvent_GetCommandExecutionStatus(DPCTLSyclEventRef ERef)
     cdef _backend_type DPCTLEvent_GetBackend(DPCTLSyclEventRef ERef)
@@ -356,7 +356,7 @@ cdef extern from "syclinterface/dpctl_sycl_queue_interface.h":
         size_t NDims,
         const DPCTLSyclEventRef *DepEvents,
         size_t NDepEvents)
-    cdef void DPCTLQueue_Wait(const DPCTLSyclQueueRef QRef)
+    cdef void DPCTLQueue_Wait(const DPCTLSyclQueueRef QRef) nogil
     cdef DPCTLSyclEventRef DPCTLQueue_Memcpy(
         const DPCTLSyclQueueRef Q,
         void *Dest,

dpctl/_sycl_event.pyx

@@ -98,7 +98,8 @@ cdef class _SyclEvent:
 
     def __dealloc__(self):
         if (self._event_ref):
-            DPCTLEvent_Wait(self._event_ref)
+            with nogil:
+                DPCTLEvent_Wait(self._event_ref)
             DPCTLEvent_Delete(self._event_ref)
         self._event_ref = NULL
         self.args = None
@@ -224,7 +225,8 @@ cdef class SyclEvent(_SyclEvent):
 
     @staticmethod
     cdef void _wait(SyclEvent event):
-        DPCTLEvent_WaitAndThrow(event._event_ref)
+        with nogil:
+            DPCTLEvent_WaitAndThrow(event._event_ref)
 
     @staticmethod
     def wait_for(event):
@@ -370,4 +372,5 @@ cdef class SyclEvent(_SyclEvent):
 
     cpdef void wait(self):
         "Synchronously wait for completion of this event."
-        DPCTLEvent_Wait(self._event_ref)
+        with nogil:
+            DPCTLEvent_Wait(self._event_ref)

dpctl/_sycl_queue.pyx

@@ -824,7 +824,8 @@ cdef class SyclQueue(_SyclQueue):
         return SyclEvent._create(Eref, args)
 
     cpdef void wait(self):
-        DPCTLQueue_Wait(self._queue_ref)
+        with nogil:
+            DPCTLQueue_Wait(self._queue_ref)
 
     cpdef memcpy(self, dest, src, size_t count):
         cdef void *c_dest
@@ -846,7 +847,8 @@ cdef class SyclQueue(_SyclQueue):
             raise RuntimeError(
                 "SyclQueue.memcpy operation encountered an error"
             )
-        DPCTLEvent_Wait(ERef)
+        with nogil:
+            DPCTLEvent_Wait(ERef)
         DPCTLEvent_Delete(ERef)
 
     cpdef prefetch(self, mem, size_t count=0):
@@ -866,7 +868,8 @@ cdef class SyclQueue(_SyclQueue):
             raise RuntimeError(
                 "SyclQueue.prefetch encountered an error"
             )
-        DPCTLEvent_Wait(ERef)
+        with nogil:
+            DPCTLEvent_Wait(ERef)
         DPCTLEvent_Delete(ERef)
 
     cpdef mem_advise(self, mem, size_t count, int advice):
@@ -886,7 +889,8 @@ cdef class SyclQueue(_SyclQueue):
             raise RuntimeError(
                 "SyclQueue.mem_advise operation encountered an error"
             )
-        DPCTLEvent_Wait(ERef)
+        with nogil:
+            DPCTLEvent_Wait(ERef)
         DPCTLEvent_Delete(ERef)
 
     @property

dpctl/memory/_memory.pyx

@@ -94,15 +94,17 @@ cdef void copy_via_host(void *dest_ptr, SyclQueue dest_queue,
         src_ptr,
         nbytes
     )
-    DPCTLEvent_Wait(E1Ref)
+    with nogil:
+        DPCTLEvent_Wait(E1Ref)
 
     E2Ref = DPCTLQueue_Memcpy(
         dest_queue.get_queue_ref(),
         dest_ptr,
         <void *>&host_buf[0],
         nbytes
     )
-    DPCTLEvent_Wait(E2Ref)
+    with nogil:
+        DPCTLEvent_Wait(E2Ref)
     DPCTLEvent_Delete(E1Ref)
     DPCTLEvent_Delete(E2Ref)
 
@@ -398,7 +400,8 @@ cdef class _Memory:
             <void *>self.memory_ptr,  # source
             <size_t>self.nbytes
         )
-        DPCTLEvent_Wait(ERef)
+        with nogil:
+            DPCTLEvent_Wait(ERef)
         DPCTLEvent_Delete(ERef)
 
         return obj
@@ -423,7 +426,8 @@ cdef class _Memory:
             <void *>&host_buf[0],     # source
             <size_t>buf_len
         )
-        DPCTLEvent_Wait(ERef)
+        with nogil:
+            DPCTLEvent_Wait(ERef)
         DPCTLEvent_Delete(ERef)
 
     cpdef copy_from_device(self, object sycl_usm_ary):
@@ -465,7 +469,8 @@ cdef class _Memory:
                     <void *>src_buf.p,
                     <size_t>src_buf.nbytes
                 )
-                DPCTLEvent_Wait(ERef)
+                with nogil:
+                    DPCTLEvent_Wait(ERef)
                 DPCTLEvent_Delete(ERef)
             else:
                 copy_via_host(