Merge pull request #555 from IntelPython/dpctl-sycl-timer

Moved sycl timer into dpctl.SyclTimer

Author: oleksandr-pavlyk

dpctl/__init__.py

@@ -64,6 +64,7 @@
     set_global_queue,
 )
 
+from ._sycl_timer import SyclTimer
 from ._version import get_versions
 from .enum_types import backend_type, device_type, event_status_type
 
@@ -88,7 +89,7 @@
 ]
 __all__ += [
     "SyclEvent",
-    "SyclEventRaw",
+    "SyclTimer",
 ]
 __all__ += [
     "get_platforms",

examples/python/sycl_timer.py renamed to dpctl/_sycl_timer.py

@@ -17,27 +17,65 @@
 
 import timeit
 
-import dpctl
+from . import SyclQueue
 
 
 class SyclTimer:
-    def __init__(self, host_time=timeit.default_timer, time_scale=1):
-        self.timer = host_time
+    """
+    SyclTimer(host_timer=timeit.default_timer, time_scale=1)
+    Python class to measure device time of execution of commands submitted to
+    :class:`dpctl.SyclQueue` as well as the wall-time.
+
+    :Example:
+        .. code-block:: python
+
+            import dpctl
+
+            # Create a default SyclQueue
+            q = dpctl.SyclQueue(property='enable_profiling')
+
+            # create the timer
+            miliseconds_sc = 1e-3
+            timer = dpctl.SyclTimer(time_scale = miliseconds_sc)
+
+            # use the timer
+            with timer(queue=q):
+                code_block
+
+            # retrieve elapsed times in miliseconds
+            sycl_dt, wall_dt = timer.dt
+
+    Remark:
+       The timer synchronizes the queue at the entrance and the
+       exit of the context.
+
+    Args:
+       host_timer (callable): A callable such that host_timer() returns current
+       host time in seconds.
+       time_scale (int, float): Ratio of the unit of time of interest and
+       one second.
+    """
+
+    def __init__(self, host_timer=timeit.default_timer, time_scale=1):
+        self.timer = host_timer
         self.time_scale = time_scale
+        self.queue = None
 
     def __call__(self, queue=None):
-        if isinstance(queue, dpctl.SyclQueue):
+        if isinstance(queue, SyclQueue):
             if queue.has_enable_profiling:
                 self.queue = queue
             else:
                 raise ValueError(
-                    "The queue does not contain the enable_profiling property"
+                    "The given queue was not created with the "
+                    "enable_profiling property"
                 )
         else:
-            raise ValueError(
-                "The passed queue must be <class 'dpctl._sycl_queue.SyclQueue'>"
+            raise TypeError(
+                "The passed queue must have type dpctl.SyclQueue, "
+                "got {}".format(type(queue))
             )
-        return self.__enter__()
+        return self
 
     def __enter__(self):
         self.event_start = self.queue.submit_barrier()
@@ -48,6 +86,7 @@ def __exit__(self, *args):
         self.event_finish = self.queue.submit_barrier()
         self.host_finish = self.timer()
 
+    @property
     def dt(self):
         self.event_start.wait()
         self.event_finish.wait()

dpctl/tests/test_sycl_event.py

@@ -157,3 +157,26 @@ def test_profiling_info():
         assert event.profiling_info_end
     else:
         pytest.skip("No OpenCL CPU queues available")
+
+
+def test_sycl_timer():
+    try:
+        q = dpctl.SyclQueue(property="enable_profiling")
+    except dpctl.SyclQueueCreationError:
+        pytest.skip("Queue creation of default device failed")
+    timer = dpctl.SyclTimer()
+    m1 = dpctl_mem.MemoryUSMDevice(256 * 1024, queue=q)
+    m2 = dpctl_mem.MemoryUSMDevice(256 * 1024, queue=q)
+    with timer(q):
+        # device task
+        m1.copy_from_device(m2)
+        # host task
+        [x ** 2 for x in range(1024)]
+    host_dt, device_dt = timer.dt
+    assert host_dt > device_dt
+    q_no_profiling = dpctl.SyclQueue()
+    assert q_no_profiling.has_enable_profiling is False
+    with pytest.raises(ValueError):
+        timer(queue=q_no_profiling)
+    with pytest.raises(TypeError):
+        timer(queue=None)

dpctl/tests/test_sycl_kernel_submit.py

@@ -37,19 +37,20 @@ def test_create_program_from_source(self):
             size_t index = get_global_id(0);                                   \
             c[index] = d*a[index] + b[index];                                  \
         }"
-        q = dpctl.SyclQueue("opencl:gpu")
+        q = dpctl.SyclQueue("opencl:gpu", property="enable_profiling")
         prog = dpctl_prog.create_program_from_source(q, oclSrc)
         axpyKernel = prog.get_sycl_kernel("axpy")
 
-        bufBytes = 1024 * np.dtype("i").itemsize
+        n_elems = 1024 * 512
+        bufBytes = n_elems * np.dtype("i").itemsize
         abuf = dpctl_mem.MemoryUSMShared(bufBytes, queue=q)
         bbuf = dpctl_mem.MemoryUSMShared(bufBytes, queue=q)
         cbuf = dpctl_mem.MemoryUSMShared(bufBytes, queue=q)
-        a = np.ndarray((1024), buffer=abuf, dtype="i")
-        b = np.ndarray((1024), buffer=bbuf, dtype="i")
-        c = np.ndarray((1024), buffer=cbuf, dtype="i")
-        a[:] = np.arange(1024)
-        b[:] = np.arange(1024, 0, -1)
+        a = np.ndarray((n_elems,), buffer=abuf, dtype="i")
+        b = np.ndarray((n_elems,), buffer=bbuf, dtype="i")
+        c = np.ndarray((n_elems,), buffer=cbuf, dtype="i")
+        a[:] = np.arange(n_elems)
+        b[:] = np.arange(n_elems, 0, -1)
         c[:] = 0
         d = 2
         args = []
@@ -59,10 +60,17 @@ def test_create_program_from_source(self):
         args.append(c.base)
         args.append(ctypes.c_int(d))
 
-        r = [1024]
+        r = [
+            n_elems,
+        ]
 
-        q.submit(axpyKernel, args, r)
-        self.assertTrue(np.allclose(c, a * d + b))
+        timer = dpctl.SyclTimer()
+        with timer(q):
+            q.submit(axpyKernel, args, r)
+            ref_c = a * d + b
+        host_dt, device_dt = timer.dt
+        self.assertTrue(host_dt > device_dt)
+        self.assertTrue(np.allclose(c, ref_c))
 
 
 if __name__ == "__main__":

examples/python/dppy_kernel.py

@@ -17,9 +17,9 @@
 
 import numba_dppy
 import numpy as np
-from sycl_timer import SyclTimer
 
 import dpctl
+from dpctl import SyclTimer
 
 
 @numba_dppy.kernel
@@ -45,11 +45,12 @@ def dppy_gemm(a, b, c):
 c = np.ones_like(a).reshape(X, X)
 
 q = dpctl.SyclQueue("opencl:gpu", property="enable_profiling")
+timer = SyclTimer(time_scale=1)
 with dpctl.device_context(q):
-    timers = SyclTimer(time_scale=1)
-    with timers(q):
+    with timer(q):
         dppy_gemm[griddim, blockdim](a, b, c)
         cc = np.dot(a, b)
-    host_time, device_time = timers.dt()
-    print("Wall time: ", host_time, "\n", "Device time: ", device_time)
-    print(np.allclose(c, cc))
+    host_time, device_time = timer.dt
+
+print("Wall time: ", host_time, "\nDevice time: ", device_time)
+print(np.allclose(c, cc))