From 5ba2cdcac3b4f6b78f3060b8db7d3c72aff31931 Mon Sep 17 00:00:00 2001
From: vlad-perevezentsev <vladislav.perevezentsev@intel.com>
Date: Thu, 12 Aug 2021 06:51:25 -0500
Subject: [PATCH] Add 3 methods of SyclEventRaw class for profiling

---
 dpctl/_backend.pxd             |  5 +++-
 dpctl/_sycl_event.pyx          | 23 ++++++++++++++
 dpctl/tests/test_sycl_event.py | 55 ++++++++++++++++++++++------------
 3 files changed, 63 insertions(+), 20 deletions(-)

diff --git a/dpctl/_backend.pxd b/dpctl/_backend.pxd
index 03ec7f21d1..30701b44ca 100644
--- a/dpctl/_backend.pxd
+++ b/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
 
 
@@ -238,6 +238,9 @@ cdef extern from "dpctl_sycl_event_interface.h":
         size_t index)
     cdef DPCTLEventVectorRef DPCTLEvent_GetWaitList(
         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":
diff --git a/dpctl/_sycl_event.pyx b/dpctl/_sycl_event.pyx
index 919e44152c..499949eff6 100644
--- a/dpctl/_sycl_event.pyx
+++ b/dpctl/_sycl_event.pyx
@@ -24,6 +24,7 @@
 import logging
 
 from cpython cimport pycapsule
+from libc.stdint cimport uint64_t
 
 from ._backend cimport (  # noqa: E211
     DPCTLEvent_Copy,
@@ -31,6 +32,9 @@ from ._backend cimport (  # noqa: E211
     DPCTLEvent_Delete,
     DPCTLEvent_GetBackend,
     DPCTLEvent_GetCommandExecutionStatus,
+    DPCTLEvent_GetProfilingInfoEnd,
+    DPCTLEvent_GetProfilingInfoStart,
+    DPCTLEvent_GetProfilingInfoSubmit,
     DPCTLEvent_GetWaitList,
     DPCTLEvent_Wait,
     DPCTLEventVector_Delete,
@@ -260,3 +264,22 @@ cdef class SyclEventRaw(_SyclEventRaw):
             events.append(SyclEventRaw._create(ERef))
         DPCTLEventVector_Delete(EVRef)
         return events
+
+    def profiling_info_submit(self):
+        cdef uint64_t profiling_info_submit = 0
+        profiling_info_submit = DPCTLEvent_GetProfilingInfoSubmit(
+                                self._event_ref
+        )
+        return profiling_info_submit
+
+    @property
+    def profiling_info_start(self):
+        cdef uint64_t profiling_info_start = 0
+        profiling_info_start = DPCTLEvent_GetProfilingInfoStart(self._event_ref)
+        return profiling_info_start
+
+    @property
+    def profiling_info_end(self):
+        cdef uint64_t profiling_info_end = 0
+        profiling_info_end = DPCTLEvent_GetProfilingInfoEnd(self._event_ref)
+        return profiling_info_end
diff --git a/dpctl/tests/test_sycl_event.py b/dpctl/tests/test_sycl_event.py
index f298405743..4e0fe30f49 100644
--- a/dpctl/tests/test_sycl_event.py
+++ b/dpctl/tests/test_sycl_event.py
@@ -28,6 +28,32 @@
 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()
@@ -37,25 +63,7 @@ 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:
@@ -132,5 +140,14 @@ def test_get_wait_list():
                 "Failed to get a list of waiting events from SyclEventRaw"
             )
         assert len(wait_list)
+
+
+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")