Add examples/stream/10streams.py

examples/stream/10streams.py

@@ -0,0 +1,37 @@
+import cupy
+import time
+
+device = cupy.cuda.Device()
+memory_pool = cupy.cuda.MemoryPool()
+cupy.cuda.set_allocator(memory_pool.malloc)
+rand = cupy.random.generator.RandomState(seed=1)
+
+n = 10
+streams = []
+events = []
+zs = []
+
+start_time = time.time()
+for i in range(n):
+    stream = cupy.cuda.stream.Stream()
+    streams.append(stream)
+    with stream:
+        x = rand.normal(size=(1, 1024 * 256))
+        y = rand.normal(size=(1024 * 256, 1))
+        z = cupy.matmul(x, y)
+        zs.append(z)
+    events.append(stream.record())
+
+reduction_stream = cupy.cuda.stream.Stream()
+# Block the reduction_stream until all event occurs. This does not block host.
+# This is not required when reduction is performed in the default (Stream.null)
+# stream unless streams are created with `non_blocking=True` flag.
+for i in range(n):
+    reduction_stream.wait_event(events[i])
+with reduction_stream:
+    z = sum(zs)
+
+device.synchronize()
+elapsed_time = time.time() - start_time
+print('elapsed time', elapsed_time)
+print('total bytes', memory_pool.total_bytes())