From 20cfa8ba86841773cc5cf00dcb838b78ee992c63 Mon Sep 17 00:00:00 2001
From: Aurelius84 <liujiezhangbupt@gmail.com>
Date: Mon, 30 Aug 2021 09:45:33 +0800
Subject: [PATCH] Abstract GenerateDeviceEventFlag to shield platforms (#35219)

* Abstract GenerateDeviceEventFlag to shield platforms

* Remove get_cuda_flags
---
 .../framework/new_executor/interpretercore.cc |  2 +-
 paddle/fluid/platform/device_event_base.cc    | 26 +++++++++++++++++++
 paddle/fluid/platform/device_event_base.h     |  4 +++
 paddle/fluid/platform/event.h                 | 25 ------------------
 paddle/fluid/pybind/cuda_streams_py.cc        |  3 ++-
 5 files changed, 33 insertions(+), 27 deletions(-)

diff --git a/paddle/fluid/framework/new_executor/interpretercore.cc b/paddle/fluid/framework/new_executor/interpretercore.cc
index 0f2ad0ff33..c86a16cef0 100644
--- a/paddle/fluid/framework/new_executor/interpretercore.cc
+++ b/paddle/fluid/framework/new_executor/interpretercore.cc
@@ -77,7 +77,7 @@ void AssociateInputWithEvents(
   for (auto var_id : new_event_var_id) {
     if (var_id2event->count(var_id) == 0) {
       auto device_event = std::make_shared<platform::DeviceEvent>(
-          place, platform::get_cuda_flags(false, false, false));
+          place, platform::GenerateDeviceEventFlag());
       var_id2event->emplace(var_id, std::move(device_event));
     }
     // Add events for next_instr.inputs
diff --git a/paddle/fluid/platform/device_event_base.cc b/paddle/fluid/platform/device_event_base.cc
index 288052edcc..0cd1cff556 100644
--- a/paddle/fluid/platform/device_event_base.cc
+++ b/paddle/fluid/platform/device_event_base.cc
@@ -14,6 +14,7 @@
 
 #include "paddle/fluid/platform/device_event_base.h"
 #include "paddle/fluid/platform/device_event_cpu.h"
+#include "paddle/fluid/platform/event.h"
 
 namespace paddle {
 namespace platform {
@@ -25,6 +26,31 @@ EventFinishFunction DeviceEvent::event_finisher_[MaxDeviceTypes];
 EventFinishFunction DeviceEvent::event_finished_setter_[MaxDeviceTypes];
 EventWaitFunction DeviceEvent::event_waiter_[MaxDeviceTypes][MaxDeviceTypes];
 
+/*
+ * Generate flag used to create event on all sorts of equipment.
+ * NOTE: Support CPU/CUDA/ROCM currently.
+ */
+unsigned int GenerateDeviceEventFlag(bool enable_timing, bool blocking,
+                                     bool interprocess) {
+#ifdef PADDLE_WITH_CUDA
+  unsigned int flags =
+      (blocking ? cudaEventBlockingSync : cudaEventDefault) |
+      (enable_timing ? cudaEventDefault : cudaEventDisableTiming) |
+      (interprocess ? cudaEventInterprocess : cudaEventDefault);
+  return flags;
+#endif
+
+#ifdef PADDLE_WITH_HIP
+  unsigned int flags =
+      (blocking ? hipEventBlockingSync : hipEventDefault) |
+      (enable_timing ? hipEventDefault : hipEventDisableTiming) |
+      (interprocess ? hipEventInterprocess : hipEventDefault);
+  return flags;
+#endif
+
+  return 0;
+}
+
 void DeviceEventCreateCPU(DeviceEvent* event, const platform::Place& place,
                           unsigned int flag) {
   event->InitEvent(std::make_shared<CPUDeviceEventWrapper>(place, flag));
diff --git a/paddle/fluid/platform/device_event_base.h b/paddle/fluid/platform/device_event_base.h
index d713a638af..e2e1fdf29d 100644
--- a/paddle/fluid/platform/device_event_base.h
+++ b/paddle/fluid/platform/device_event_base.h
@@ -39,6 +39,10 @@ inline int DeviceTypeToId(const DeviceType& device_type) {
   return static_cast<int>(device_type);
 }
 
+unsigned int GenerateDeviceEventFlag(bool enable_timing = false,
+                                     bool blocking = false,
+                                     bool interprocess = false);
+
 enum EventStatus {
   INITIALIZED = 0,
   SCHEDULED = 1,
diff --git a/paddle/fluid/platform/event.h b/paddle/fluid/platform/event.h
index a79ab22743..2b11de48a1 100644
--- a/paddle/fluid/platform/event.h
+++ b/paddle/fluid/platform/event.h
@@ -195,30 +195,5 @@ class CudaEvent {
 #endif
 };
 
-static unsigned int get_cuda_flags(bool enable_timing, bool blocking,
-                                   bool interprocess) {
-#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
-
-#ifdef PADDLE_WITH_HIP
-  unsigned int flags =
-      (blocking ? hipEventBlockingSync : hipEventDefault) |
-      (enable_timing ? hipEventDefault : hipEventDisableTiming) |
-      (interprocess ? hipEventInterprocess : hipEventDefault);
-  return flags;
-#else
-  unsigned int flags =
-      (blocking ? cudaEventBlockingSync : cudaEventDefault) |
-      (enable_timing ? cudaEventDefault : cudaEventDisableTiming) |
-      (interprocess ? cudaEventInterprocess : cudaEventDefault);
-  return flags;
-#endif
-
-#else
-  PADDLE_THROW(platform::errors::Unavailable(
-      "Paddle is not compiled with CUDA. Cannot get the cuda event flags."));
-  return 0;
-#endif
-}
-
 }  // namespace platform
 }  // namespace paddle
diff --git a/paddle/fluid/pybind/cuda_streams_py.cc b/paddle/fluid/pybind/cuda_streams_py.cc
index 5ea0a2553f..50c6d0e983 100644
--- a/paddle/fluid/pybind/cuda_streams_py.cc
+++ b/paddle/fluid/pybind/cuda_streams_py.cc
@@ -15,6 +15,7 @@
 #include <string>
 #include <vector>
 
+#include "paddle/fluid/platform/device_event_base.h"
 #include "paddle/fluid/platform/event.h"
 #include "paddle/fluid/platform/stream/cuda_stream.h"
 #include "paddle/fluid/pybind/cuda_streams_py.h"
@@ -331,7 +332,7 @@ void BindCudaStream(py::module *m_ptr) {
            [](paddle::platform::CudaEvent &self, bool enable_timing,
               bool blocking, bool interprocess) {
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
-             unsigned int flags = platform::get_cuda_flags(
+             unsigned int flags = platform::GenerateDeviceEventFlag(
                  enable_timing, blocking, interprocess);
              new (&self) paddle::platform::CudaEvent(flags);
 #else
-- 
GitLab