Add 3 methods of SyclEventRaw class for profiling

Author: vlad-perevezentsev

dpctl/_backend.pxd

@@ -21,7 +21,7 @@
 types defined by dpctl's C API.
 """
 
-from libc.stdint cimport int64_t, uint32_t
+from libc.stdint cimport int64_t, uint32_t, uint64_t
 from libcpp cimport bool
 
 
@@ -221,6 +221,9 @@ cdef extern from "dpctl_sycl_event_interface.h":
     cdef DPCTLSyclEventRef DPCTLEvent_Copy(const DPCTLSyclEventRef ERef)
     cdef void DPCTLEvent_Wait(DPCTLSyclEventRef ERef)
     cdef void DPCTLEvent_Delete(DPCTLSyclEventRef ERef)
+    cdef uint64_t DPCTLEvent_GetProfilingInfoSubmit(DPCTLSyclEventRef ERef)
+    cdef uint64_t DPCTLEvent_GetProfilingInfoStart(DPCTLSyclEventRef ERef)
+    cdef uint64_t DPCTLEvent_GetProfilingInfoEnd(DPCTLSyclEventRef ERef)
 
 
 cdef extern from "dpctl_sycl_kernel_interface.h":

dpctl/_sycl_event.pyx

@@ -24,11 +24,15 @@
 import logging
 
 from cpython cimport pycapsule
+from libc.stdint cimport uint64_t
 
 from ._backend cimport (  # noqa: E211
     DPCTLEvent_Copy,
     DPCTLEvent_Create,
     DPCTLEvent_Delete,
+    DPCTLEvent_GetProfilingInfoEnd,
+    DPCTLEvent_GetProfilingInfoStart,
+    DPCTLEvent_GetProfilingInfoSubmit,
     DPCTLEvent_Wait,
     DPCTLSyclEventRef,
 )
@@ -202,3 +206,20 @@ cdef class SyclEventRaw(_SyclEventRaw):
             "SyclEventRef",
             &_event_capsule_deleter
         )
+
+    def profiling_info_submit(self):
+        cdef uint64_t profiling_info_submit = 0
+        profiling_info_submit = DPCTLEvent_GetProfilingInfoSubmit(
+                                self._event_ref
+        )
+        return profiling_info_submit
+
+    def profiling_info_start(self):
+        cdef uint64_t profiling_info_start = 0
+        profiling_info_start = DPCTLEvent_GetProfilingInfoStart(self._event_ref)
+        return profiling_info_start
+
+    def profiling_info_end(self):
+        cdef uint64_t profiling_info_end = 0
+        profiling_info_end = DPCTLEvent_GetProfilingInfoEnd(self._event_ref)
+        return profiling_info_end

dpctl/tests/test_sycl_event.py

@@ -27,6 +27,32 @@
 from ._helper import has_cpu
 
 
+def produce_event(profiling=False):
+    oclSrc = "                                                             \
+            kernel void add(global int* a) {                                   \
+                size_t index = get_global_id(0);                               \
+                a[index] = a[index] + 1;                                       \
+            }"
+    if profiling:
+        q = dpctl.SyclQueue("opencl:cpu", property="enable_profiling")
+    else:
+        q = dpctl.SyclQueue("opencl:cpu")
+    prog = dpctl_prog.create_program_from_source(q, oclSrc)
+    addKernel = prog.get_sycl_kernel("add")
+
+    bufBytes = 1024 * np.dtype("i").itemsize
+    abuf = dpctl_mem.MemoryUSMShared(bufBytes, queue=q)
+    a = np.ndarray((1024), buffer=abuf, dtype="i")
+    a[:] = np.arange(1024)
+    args = []
+
+    args.append(a.base)
+    r = [1024]
+    ev = q.submit(addKernel, args, r)
+
+    return ev
+
+
 def test_create_default_event_raw():
     try:
         dpctl.SyclEventRaw()
@@ -36,25 +62,7 @@ def test_create_default_event_raw():
 
 def test_create_event_raw_from_SyclEvent():
     if has_cpu():
-        oclSrc = "                                                             \
-            kernel void add(global int* a) {                                   \
-                size_t index = get_global_id(0);                               \
-                a[index] = a[index] + 1;                                       \
-            }"
-        q = dpctl.SyclQueue("opencl:cpu")
-        prog = dpctl_prog.create_program_from_source(q, oclSrc)
-        addKernel = prog.get_sycl_kernel("add")
-
-        bufBytes = 1024 * np.dtype("i").itemsize
-        abuf = dpctl_mem.MemoryUSMShared(bufBytes, queue=q)
-        a = np.ndarray((1024), buffer=abuf, dtype="i")
-        a[:] = np.arange(1024)
-        args = []
-
-        args.append(a.base)
-        r = [1024]
-        ev = q.submit(addKernel, args, r)
-
+        ev = produce_event()
         try:
             dpctl.SyclEventRaw(ev)
         except ValueError:
@@ -70,3 +78,14 @@ def test_create_event_raw_from_capsule():
         dpctl.SyclEventRaw(event_capsule)
     except ValueError:
         pytest.fail("Failed to create an event from capsule")
+
+
+def test_profiling_info():
+    if has_cpu():
+        event = produce_event(profiling=True)
+        event_raw = dpctl.SyclEventRaw(event)
+        assert event_raw.profiling_info_submit()
+        assert event_raw.profiling_info_start()
+        assert event_raw.profiling_info_end()
+    else:
+        pytest.skip("No OpenCL CPU queues available")