Replace platform::XPUDeviceGuard phi::backends::xpu::XPUDeviceGuard (#65802)

co63oc · web-flow · commit d7bdfc1a30fa · 2024-07-12T14:18:01.000+08:00
diff --git a/paddle/fluid/distributed/collective/bkcl_tools.h b/paddle/fluid/distributed/collective/bkcl_tools.h
@@ -39,7 +39,7 @@ class XPUEventManager {
 
   ~XPUEventManager() {
     if (is_created_) {
-      platform::XPUDeviceGuard guard(device_index_);
+      phi::backends::xpu::XPUDeviceGuard guard(device_index_);
       xpu_event_destroy(event_);
     }
   }
@@ -77,7 +77,7 @@ class XPUEventManager {
                                         device_index,
                                         device_index_));
 
-    platform::XPUDeviceGuard guard(device_index_);
+    phi::backends::xpu::XPUDeviceGuard guard(device_index_);
     // TODO(zhangxiaoci) temporary solution: xpu::event seems buggy
     PADDLE_ENFORCE_XPU_SUCCESS(xpu_wait(ctx.stream()));
   }
@@ -92,7 +92,7 @@ class XPUEventManager {
  private:
   void CreateEvent(int device_index) {
     device_index_ = device_index;
-    platform::XPUDeviceGuard guard(device_index);
+    phi::backends::xpu::XPUDeviceGuard guard(device_index);
 
     PADDLE_ENFORCE_XPU_SUCCESS(xpu_event_create(&event_));
 
diff --git a/paddle/fluid/distributed/collective/process_group_bkcl.cc b/paddle/fluid/distributed/collective/process_group_bkcl.cc
@@ -57,7 +57,7 @@ bool ProcessGroupBKCL::BKCLTask::Wait(std::chrono::milliseconds timeout) {
     // for xpu (for now), so all we can do is sync whatever stream that we know
     // and hope for the best. Note that for correctness the communication stream
     // needs to be in sync mode.
-    platform::XPUDeviceGuard guard(place_.GetDeviceId());
+    phi::backends::xpu::XPUDeviceGuard guard(place_.GetDeviceId());
     xpu_wait();
     calc_ctx->Wait();
   }
@@ -192,7 +192,7 @@ void ProcessGroupBKCL::BroadcastUniqueBKCLID(BKCLUniqueId* bkcl_id) {
 
 void ProcessGroupBKCL::CreateBKCLEnvCache(const Place& place,
                                           const std::string& place_key) {
-  platform::XPUDeviceGuard guard(place.GetDeviceId());
+  phi::backends::xpu::XPUDeviceGuard guard(place.GetDeviceId());
 
   VLOG(3) << "init bkcl rank: " << rank_ << ", nranks: " << size_
           << ", place: " << place_key;
@@ -245,7 +245,7 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::Collective(
   const auto& place = tensor.place();
   const auto& key = GetKeyFromPlace(place);
 
-  platform::XPUDeviceGuard xpu_guard(place);
+  phi::backends::xpu::XPUDeviceGuard xpu_guard(place);
 
   if (!calc_event_ ||
       (place_to_comm_ctx_.find(key) == place_to_comm_ctx_.end())) {
@@ -302,7 +302,7 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::Point2Point(
   int p2p_target_rank = peer;
   std::string key = GetKeyFromPlace(place);
 
-  platform::XPUDeviceGuard xpu_guard(place);
+  phi::backends::xpu::XPUDeviceGuard xpu_guard(place);
 
   if (place_to_comm_ctx_.find(key) == place_to_comm_ctx_.end()) {
     CreateBKCLEnvCache(place, key);
diff --git a/paddle/fluid/framework/fleet/heter_ps/heter_comm_inl.h b/paddle/fluid/framework/fleet/heter_ps/heter_comm_inl.h
@@ -462,7 +462,8 @@ void HeterComm<KeyType, ValType, GradType, GPUAccessor>::create_storage(
 #elif defined(PADDLE_WITH_XPU_KP)
   auto &nodes = path_[start_index][end_index].nodes_;
   for (size_t i = 0; i < nodes.size(); ++i) {
-    platform::XPUDeviceGuard guard(resource_->dev_id(nodes[i].dev_num));
+    phi::backends::xpu::XPUDeviceGuard guard(
+        resource_->dev_id(nodes[i].dev_num));
     auto place = DevPlace(resource_->dev_id(nodes[i].dev_num));
     if (keylen > 0) {
       auto node_keys_mem = MemoryAlloc(place, keylen);
@@ -494,7 +495,7 @@ void HeterComm<KeyType, ValType, GradType, GPUAccessor>::create_tmp_storage(
       resource_->remote_stream(end_index, start_index)));
 
 #elif defined(PADDLE_WITH_XPU_KP)
-  platform::XPUDeviceGuard guard(resource_->dev_id(end_index));
+  phi::backends::xpu::XPUDeviceGuard guard(resource_->dev_id(end_index));
   auto place = DevPlace(resource_->dev_id(end_index));
   auto node_vals_mem = MemoryAlloc(place, vallen);
   dest = reinterpret_cast<void *>(node_vals_mem->ptr());
diff --git a/paddle/fluid/framework/fleet/heter_ps/heter_resource.cc b/paddle/fluid/framework/fleet/heter_ps/heter_resource.cc
@@ -72,7 +72,7 @@ XPUResource::XPUResource(std::vector<int> &dev_ids, int index) {
   dev_ids_ = dev_ids;
   dev_id_ = dev_ids_[index];
 
-  platform::XPUDeviceGuard guard(dev_id_);
+  phi::backends::xpu::XPUDeviceGuard guard(dev_id_);
   local_streams_.resize(dev_ids_.size());
 
   comm_streams_.resize(dev_ids_.size(), NULL);
@@ -86,7 +86,7 @@ XPUResource::XPUResource(std::vector<int> &dev_ids, int index) {
 }
 
 XPUResource::~XPUResource() {
-  platform::XPUDeviceGuard guard(dev_id_);
+  phi::backends::xpu::XPUDeviceGuard guard(dev_id_);
   for (size_t i = 0; i < local_streams_.size(); ++i) {
     PADDLE_ENFORCE_XPU_SUCCESS(xpu_stream_destroy(local_streams_[i]));
   }
diff --git a/paddle/fluid/framework/fleet/heter_ps/heter_resource.h b/paddle/fluid/framework/fleet/heter_ps/heter_resource.h
@@ -94,7 +94,7 @@ using AnyDeviceGuard = platform::CUDADeviceGuard;
 #elif defined(PADDLE_WITH_XPU_KP)
 using DevResource = XPUResource;
 using DevPlace = phi::XPUPlace;
-using AnyDeviceGuard = platform::XPUDeviceGuard;
+using AnyDeviceGuard = phi::backends::xpu::XPUDeviceGuard;
 #endif
 
 #if defined(PADDLE_WITH_CUDA)
diff --git a/paddle/fluid/framework/heterxpu_trainer.cc b/paddle/fluid/framework/heterxpu_trainer.cc
@@ -125,7 +125,7 @@ void HeterXpuTrainer::CreateThreadParam(const ProgramDesc& program, int num) {
 
 #ifdef PADDLE_WITH_XPU
   auto dev_id = place.device;
-  platform::XPUDeviceGuard guard(dev_id);
+  phi::backends::xpu::XPUDeviceGuard guard(dev_id);
 #endif
 
   auto& block = program.Block(0);
@@ -361,7 +361,7 @@ int HeterXpuTrainer::EndPass(const HeterRequest* request,
 #ifdef PADDLE_WITH_XPU
         auto place = thread_tensor->place();
         auto dev_id = place.device;
-        platform::XPUDeviceGuard guard(dev_id);
+        phi::backends::xpu::XPUDeviceGuard guard(dev_id);
         platform::DeviceContextPool& pool =
             platform::DeviceContextPool::Instance();
         platform::DeviceContext* dev_ctx = pool.Get(place);
@@ -398,7 +398,7 @@ int HeterXpuTrainer::EndPass(const HeterRequest* request,
 #ifdef PADDLE_WITH_XPU
       auto place = root_tensor->place();
       auto dev_id = place.device;
-      platform::XPUDeviceGuard guard(dev_id);
+      phi::backends::xpu::XPUDeviceGuard guard(dev_id);
       platform::DeviceContextPool& pool =
           platform::DeviceContextPool::Instance();
       platform::DeviceContext* dev_ctx = pool.Get(place);
diff --git a/paddle/fluid/memory/allocation/naive_best_fit_allocator.cc b/paddle/fluid/memory/allocation/naive_best_fit_allocator.cc
@@ -145,7 +145,7 @@ void *Alloc<phi::XPUPlace>(const phi::XPUPlace &place, size_t size) {
   VLOG(10) << "Allocate " << size << " bytes on " << phi::Place(place);
   void *p = nullptr;
 
-  platform::XPUDeviceGuard guard(place.device);
+  phi::backends::xpu::XPUDeviceGuard guard(place.device);
   int ret = xpu_malloc(reinterpret_cast<void **>(&p), size);
   if (ret != XPU_SUCCESS) {
     VLOG(10) << "xpu memory malloc(" << size << ") failed, try again";
@@ -176,7 +176,7 @@ void Free<phi::XPUPlace>(const phi::XPUPlace &place, void *p, size_t size) {
   VLOG(10) << "Free " << size << " bytes on " << phi::Place(place);
   VLOG(10) << "Free pointer=" << p << " on " << phi::Place(place);
 
-  platform::XPUDeviceGuard guard(place.device);
+  phi::backends::xpu::XPUDeviceGuard guard(place.device);
   xpu_free(p);
 #else
   PADDLE_THROW(
diff --git a/paddle/fluid/operators/reader/buffered_reader.cc b/paddle/fluid/operators/reader/buffered_reader.cc
@@ -270,7 +270,7 @@ void BufferedReader::ReadAsync(size_t i) {
         xpu_ptrs.emplace_back(xpu[i].mutable_data(place_, cpu[i].type()));
       }
 
-      platform::XPUDeviceGuard guard(place_.device);
+      phi::backends::xpu::XPUDeviceGuard guard(place_.device);
       int r = xpu_event_record(events_[i].get(), compute_stream_);
       PADDLE_ENFORCE_XDNN_SUCCESS(r, "xpu_event_record");
       r = xpu_stream_wait_event(stream_.get(), events_[i].get());
diff --git a/paddle/fluid/platform/device/xpu/xpu_resource_pool.cc b/paddle/fluid/platform/device/xpu/xpu_resource_pool.cc
@@ -22,14 +22,14 @@ XpuStreamResourcePool::XpuStreamResourcePool() {
   pool_.reserve(dev_cnt);
   for (int dev_idx = 0; dev_idx < dev_cnt; ++dev_idx) {
     auto creator = [dev_idx] {
-      platform::XPUDeviceGuard guard(dev_idx);
+      phi::backends::xpu::XPUDeviceGuard guard(dev_idx);
       xpuStream stream;
       xpu_stream_create(&stream);
       return stream;
     };
 
     auto deleter = [dev_idx](xpuStream stream) {
-      platform::XPUDeviceGuard guard(dev_idx);
+      phi::backends::xpu::XPUDeviceGuard guard(dev_idx);
       xpu_stream_destroy(stream);
     };
 
@@ -63,14 +63,14 @@ XpuEventResourcePool::XpuEventResourcePool() {
   pool_.reserve(dev_cnt);
   for (int dev_idx = 0; dev_idx < dev_cnt; ++dev_idx) {
     auto creator = [dev_idx] {
-      platform::XPUDeviceGuard guard(dev_idx);
+      phi::backends::xpu::XPUDeviceGuard guard(dev_idx);
       xpuEventHandle event;
       xpu_event_create(&event);
       return event;
     };
 
     auto deleter = [dev_idx](xpuEventHandle event) {
-      platform::XPUDeviceGuard guard(dev_idx);
+      phi::backends::xpu::XPUDeviceGuard guard(dev_idx);
       xpu_event_destroy(event);
     };
 
diff --git a/paddle/fluid/pybind/tensor_py.h b/paddle/fluid/pybind/tensor_py.h
@@ -439,7 +439,7 @@ void SetTensorFromPyArrayT(
     // NOTE(wangxi): When copying data to the accelerator card,
     // we need set_device(dev_id) first.
     phi::Place tmp_place = place;
-    platform::XPUDeviceGuard guard(tmp_place.device);
+    phi::backends::xpu::XPUDeviceGuard guard(tmp_place.device);
     auto dst = self->mutable_data<T>(place);
     memory::Copy(tmp_place,
                  static_cast<void *>(dst),

Original file line number	Diff line number	Diff line change
`@@ -72,7 +72,7 @@ XPUResource::XPUResource(std::vector<int> &dev_ids, int index) {`
`72`	`72`	`dev_ids_ = dev_ids;`
`73`	`73`	`dev_id_ = dev_ids_[index];`
`74`	`74`
`75`		`- platform::XPUDeviceGuard guard(dev_id_);`
	`75`	`+ phi::backends::xpu::XPUDeviceGuard guard(dev_id_);`
`76`	`76`	`local_streams_.resize(dev_ids_.size());`
`77`	`77`
`78`	`78`	`comm_streams_.resize(dev_ids_.size(), NULL);`
`@@ -86,7 +86,7 @@ XPUResource::XPUResource(std::vector<int> &dev_ids, int index) {`
`86`	`86`	`}`
`87`	`87`
`88`	`88`	`XPUResource::~XPUResource() {`
`89`		`- platform::XPUDeviceGuard guard(dev_id_);`
	`89`	`+ phi::backends::xpu::XPUDeviceGuard guard(dev_id_);`
`90`	`90`	`for (size_t i = 0; i < local_streams_.size(); ++i) {`
`91`	`91`	`PADDLE_ENFORCE_XPU_SUCCESS(xpu_stream_destroy(local_streams_[i]));`
`92`	`92`	`}`
Original file line number	Diff line number	Diff line change
`@@ -270,7 +270,7 @@ void BufferedReader::ReadAsync(size_t i) {`
`270`	`270`	`xpu_ptrs.emplace_back(xpu[i].mutable_data(place_, cpu[i].type()));`
`271`	`271`	`}`
`272`	`272`
`273`		`- platform::XPUDeviceGuard guard(place_.device);`
	`273`	`+ phi::backends::xpu::XPUDeviceGuard guard(place_.device);`
`274`	`274`	`int r = xpu_event_record(events_[i].get(), compute_stream_);`
`275`	`275`	`PADDLE_ENFORCE_XDNN_SUCCESS(r, "xpu_event_record");`
`276`	`276`	`r = xpu_stream_wait_event(stream_.get(), events_[i].get());`