From 0ae7844fea8991b3175a655e876cad814cb434c0 Mon Sep 17 00:00:00 2001
From: Oleksandr Pavlyk <oleksandr.pavlyk@intel.com>
Date: Wed, 5 Jan 2022 12:50:42 -0600
Subject: [PATCH] USM alloc. fns declared nogil, used inside `with nogil`
 context

Large USM allocations can be constly, so releasing GIL while
invoking them may be a good idea.
---
 dpctl/_backend.pxd       | 14 ++++++++------
 dpctl/memory/_memory.pyx | 20 +++++++++++---------
 2 files changed, 19 insertions(+), 15 deletions(-)

diff --git a/dpctl/_backend.pxd b/dpctl/_backend.pxd
index 0ba086db79..b4788ab493 100644
--- a/dpctl/_backend.pxd
+++ b/dpctl/_backend.pxd
@@ -394,23 +394,25 @@ cdef extern from "syclinterface/dpctl_sycl_queue_manager.h":
 cdef extern from "syclinterface/dpctl_sycl_usm_interface.h":
     cdef DPCTLSyclUSMRef DPCTLmalloc_shared(
         size_t size,
-        DPCTLSyclQueueRef QRef)
+        DPCTLSyclQueueRef QRef) nogil
     cdef DPCTLSyclUSMRef DPCTLmalloc_host(
         size_t size,
-        DPCTLSyclQueueRef QRef)
-    cdef DPCTLSyclUSMRef DPCTLmalloc_device(size_t size, DPCTLSyclQueueRef QRef)
+        DPCTLSyclQueueRef QRef) nogil
+    cdef DPCTLSyclUSMRef DPCTLmalloc_device(
+        size_t size,
+        DPCTLSyclQueueRef QRef) nogil
     cdef DPCTLSyclUSMRef DPCTLaligned_alloc_shared(
         size_t alignment,
         size_t size,
-        DPCTLSyclQueueRef QRef)
+        DPCTLSyclQueueRef QRef) nogil
     cdef DPCTLSyclUSMRef DPCTLaligned_alloc_host(
         size_t alignment,
         size_t size,
-        DPCTLSyclQueueRef QRef)
+        DPCTLSyclQueueRef QRef) nogil
     cdef DPCTLSyclUSMRef DPCTLaligned_alloc_device(
         size_t alignment,
         size_t size,
-        DPCTLSyclQueueRef QRef)
+        DPCTLSyclQueueRef QRef) nogil
     cdef void DPCTLfree_with_queue(
         DPCTLSyclUSMRef MRef,
         DPCTLSyclQueueRef QRef)
diff --git a/dpctl/memory/_memory.pyx b/dpctl/memory/_memory.pyx
index 76d6255df3..741d72e22d 100644
--- a/dpctl/memory/_memory.pyx
+++ b/dpctl/memory/_memory.pyx
@@ -142,6 +142,7 @@ cdef class _Memory:
     cdef _cinit_alloc(self, Py_ssize_t alignment, Py_ssize_t nbytes,
                       bytes ptr_type, SyclQueue queue):
         cdef DPCTLSyclUSMRef p = NULL
+        cdef DPCTLSyclQueueRef QRef = NULL
 
         self._cinit_empty()
 
@@ -149,27 +150,28 @@ cdef class _Memory:
             if queue is None:
                 queue = dpctl.SyclQueue()
 
+            QRef = queue.get_queue_ref()
             if (ptr_type == b"shared"):
                 if alignment > 0:
-                    p = DPCTLaligned_alloc_shared(
-                        alignment, nbytes, queue.get_queue_ref()
+                    with nogil: p = DPCTLaligned_alloc_shared(
+                        alignment, nbytes, QRef
                     )
                 else:
-                    p = DPCTLmalloc_shared(nbytes, queue.get_queue_ref())
+                    with nogil: p = DPCTLmalloc_shared(nbytes, QRef)
             elif (ptr_type == b"host"):
                 if alignment > 0:
-                    p = DPCTLaligned_alloc_host(
-                        alignment, nbytes, queue.get_queue_ref()
+                    with nogil: p = DPCTLaligned_alloc_host(
+                        alignment, nbytes, QRef
                     )
                 else:
-                    p = DPCTLmalloc_host(nbytes, queue.get_queue_ref())
+                    with nogil: p = DPCTLmalloc_host(nbytes, QRef)
             elif (ptr_type == b"device"):
                 if (alignment > 0):
-                    p = DPCTLaligned_alloc_device(
-                        alignment, nbytes, queue.get_queue_ref()
+                    with nogil: p = DPCTLaligned_alloc_device(
+                        alignment, nbytes, QRef
                     )
                 else:
-                    p = DPCTLmalloc_device(nbytes, queue.get_queue_ref())
+                    with nogil: p = DPCTLmalloc_device(nbytes, QRef)
             else:
                 raise RuntimeError(
                     "Pointer type is unknown: {}".format(