Pre Merge pull request !1313 from 储星/memcpy_task

ge/CMakeLists.txt

@@ -349,6 +349,7 @@ set(TRAIN_SRC_LIST
     "single_op/task/op_task.cc"
     "single_op/task/tbe_task_builder.cc"
     "single_op/task/aicpu_task_builder.cc"
+    "single_op/task/rts_kernel_task_builder.cc"
     "single_op/task/aicpu_kernel_task_builder.cc"
     "hybrid/common/tensor_value.cc"
     "hybrid/common/npu_memory_allocator.cc"

ge/single_op/single_op_model.cc

@@ -30,6 +30,7 @@
 #include "runtime/rt.h"
 #include "task/aicpu_task_builder.h"
 #include "task/aicpu_kernel_task_builder.h"
+#include "task/rts_kernel_task_builder.h"
 #include "task/tbe_task_builder.h"
 #include "hybrid/executor/hybrid_model_executor.h"
 #include "hybrid/node_executor/node_executor.h"
@@ -248,7 +249,9 @@ Status SingleOpModel::ParseInputsAndOutputs() {
   for (auto &op_desc : data_ops_) {
     GE_CHK_STATUS_RET_NOLOG(ParseInputNode(op_desc));
   }
-  ParseOutputNode(netoutput_op_);
+  if (netoutput_op_ != nullptr) {
+    ParseOutputNode(netoutput_op_);
+  }
   return SUCCESS;
 }
 
@@ -330,6 +333,17 @@ Status SingleOpModel::BuildTaskList(StreamResource *stream_resource, SingleOp &s
       aicpu_task->SetModelArgs(model_name_, model_id_);
       ParseArgTable(aicpu_task, single_op);
       single_op.tasks_.emplace_back(aicpu_task);
+    } else if (task_type == RT_MODEL_TASK_MEMCPY_ASYNC || task_type == RT_MODEL_TASK_MEMCPY_ADDR_ASYNC) {
+      auto kernel_def = task_def.memcpy_async();
+      auto node = op_list_[kernel_def.op_index()];
+      GE_CHECK_NOTNULL(node);
+      auto op_desc = node->GetOpDesc();
+      GE_CHECK_NOTNULL(op_desc);
+      std::unique_ptr<MemcpyAsyncTask> task;
+      GE_CHK_STATUS_RET_NOLOG(RtsKernelTaskBuilder::BuildMemcpyAsyncTask(op_desc, kernel_def, model_params_, task));
+      task->SetModelArgs(model_name_, model_id_);
+      ParseArgTable(task.get(), single_op);
+      single_op.tasks_.emplace_back(task.release());
     } else {
       // skip
       GELOGD("Skip task type: %d", static_cast<int>(task_type));

ge/single_op/single_op_model.h

@@ -22,6 +22,7 @@
 #include <memory>
 #include <string>
 #include <vector>
+#include <single_op/task/op_task.h>
 
 #include "common/helper/model_helper.h"
 #include "single_op/single_op.h"

ge/single_op/task/op_task.cc

@@ -35,6 +35,7 @@ namespace ge {
 namespace {
 constexpr int kLaunchRetryTimes = 1000;
 constexpr int kSleepTime = 10;
+constexpr size_t kMemcpyArgCount = 2;
 constexpr uint64_t kReleaseFlag = 1;
 constexpr int kCopyNum = 2;
 constexpr uint64_t kInferSessionId = 0;
@@ -911,4 +912,16 @@ void AiCpuCCTask::GetIoAddr(uintptr_t *&arg_base, size_t &arg_count) {
   arg_base = io_addr_;
   arg_count = io_addr_num_;
 }
+
+Status MemcpyAsyncTask::LaunchKernel(rtStream_t stream) {
+  auto src_addr = reinterpret_cast<void *>(addresses_[0]);
+  auto dst_addr = reinterpret_cast<void *>(addresses_[1]);
+  GE_CHK_RT_RET(rtMemcpyAsync(dst_addr, dst_max_, src_addr, count_, kind_, stream));
+  return SUCCESS;
+}
+
+void MemcpyAsyncTask::GetIoAddr(uintptr_t *&arg_base, size_t &arg_count) {
+  arg_base = addresses_;
+  arg_count = kMemcpyArgCount;
+}
 }  // namespace ge

ge/single_op/task/op_task.h

@@ -44,6 +44,9 @@ class OpTask {
   virtual Status UpdateArgTable(const SingleOpModelParam &param);
   void SetModelArgs(std::string model_name, uint32_t model_id);
   Status GetProfilingArgs(TaskDescInfo &task_desc_info, uint32_t &model_id);
+  void SetOpDesc(const OpDescPtr &op_desc) {
+    op_desc_ = op_desc;
+  }
   const OpDescPtr &GetOpdesc() const {return op_desc_;}
   Status OpenDump(rtStream_t stream);
   virtual void GetIoAddr(uintptr_t *&arg_base, size_t &arg_count) = 0;
@@ -242,6 +245,22 @@ private:
   std::string op_type_;
   uint64_t kernel_id_ = 0;
 };
+
+class MemcpyAsyncTask : public OpTask {
+ public:
+  Status LaunchKernel(rtStream_t stream) override;
+  void GetIoAddr(uintptr_t *&arg_base, size_t &arg_count) override;
+
+ private:
+  friend class SingleOpModel;
+  friend class RtsKernelTaskBuilder;
+
+  uintptr_t addresses_[2];
+  size_t dst_max_;
+  size_t count_;
+  rtMemcpyKind_t  kind_;
+  NodePtr node_;
+};
 }  // namespace ge
 
 #endif  // GE_SINGLE_OP_TASK_OP_TASK_H_

ge/single_op/task/rts_kernel_task_builder.cc

@@ -0,0 +1,45 @@
+/**
+ * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "rts_kernel_task_builder.h"
+#include "build_task_utils.h"
+
+namespace ge {
+namespace {
+const size_t kNumAddresses = 2;
+}  // namespace
+
+Status RtsKernelTaskBuilder::BuildMemcpyAsyncTask(const OpDescPtr &op_desc,
+                                                  const domi::MemcpyAsyncDef &kernel_def,
+                                                  const SingleOpModelParam &param,
+                                                  std::unique_ptr<MemcpyAsyncTask> &task) {
+  task.reset(new(std::nothrow)MemcpyAsyncTask());
+  GE_CHECK_NOTNULL(task);
+  task->SetOpDesc(op_desc);
+  task->dst_max_ = kernel_def.dst_max();
+  task->count_ = kernel_def.count();
+  task->kind_ = static_cast<rtMemcpyKind_t>(kernel_def.kind());
+  auto addresses = BuildTaskUtils::JoinAddresses(BuildTaskUtils::GetAddresses(op_desc, param, false));
+  if (addresses.size() != kNumAddresses) {
+    GELOGE(INTERNAL_ERROR, "[Build][MemcpyAsyncTask] Invalid address count: %zu", addresses.size());
+    return INTERNAL_ERROR;
+  }
+
+  task->addresses_[0] = reinterpret_cast<uintptr_t>(addresses[0]);
+  task->addresses_[1] = reinterpret_cast<uintptr_t>(addresses[1]);
+  return SUCCESS;
+}
+}  // namespace ge

ge/single_op/task/rts_kernel_task_builder.h

@@ -0,0 +1,34 @@
+/**
+ * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef GE_SINGLE_OP_TASK_RTS_KERNEL_TASK_BUILDER_H_
+#define GE_SINGLE_OP_TASK_RTS_KERNEL_TASK_BUILDER_H_
+
+#include <vector>
+#include "graph/op_desc.h"
+#include "single_op/single_op.h"
+#include "single_op/single_op_model.h"
+
+namespace ge {
+class RtsKernelTaskBuilder {
+ public:
+  static Status BuildMemcpyAsyncTask(const OpDescPtr &op_desc,
+                                     const domi::MemcpyAsyncDef &kernel_def,
+                                     const SingleOpModelParam &param,
+                                     std::unique_ptr<MemcpyAsyncTask> &task);
+};
+}  // namespace ge
+#endif  // GE_SINGLE_OP_TASK_RTS_KERNEL_TASK_BUILDER_H_

tests/ut/ge/CMakeLists.txt

@@ -581,6 +581,7 @@ set(SINGLE_OP_SRC_FILES
     "${GE_CODE_DIR}/ge/single_op/single_op_manager.cc"
     "${GE_CODE_DIR}/ge/single_op/task/aicpu_task_builder.cc"
     "${GE_CODE_DIR}/ge/single_op/task/aicpu_kernel_task_builder.cc"
+    "${GE_CODE_DIR}/ge/single_op/task/rts_kernel_task_builder.cc"
     "${GE_CODE_DIR}/ge/hybrid/common/tensor_value.cc"
     "${GE_CODE_DIR}/ge/hybrid/common/npu_memory_allocator.cc"
     "${GE_CODE_DIR}/ge/hybrid/executor/rt_callback_manager.cc"

tests/ut/ge/single_op/single_op_model_unittest.cc

@@ -25,6 +25,7 @@
 #define private public
 #include "single_op/single_op_model.h"
 #include "single_op/task/tbe_task_builder.h"
+#include "single_op/task/rts_kernel_task_builder.h"
 #undef private
 #undef protected
 
@@ -223,3 +224,33 @@ TEST_F(UtestSingleOpModel, test_build_dynamic_op) {
   model.BuildDynamicOp(res, dynamic_single_op);
 }
 
+TEST_F(UtestSingleOpModel, test_build_memcpy_task) {
+  auto op_desc = std::make_shared<OpDesc>(MEMCPYASYNC, MEMCPYASYNC);
+  domi::MemcpyAsyncDef kernel_def;
+  kernel_def.set_dst_max(8);
+  kernel_def.set_count(8);
+  kernel_def.set_kind(2);
+  GeShape shape({2});
+  GeTensorDesc tensor_desc(shape);
+  op_desc->AddInputDesc("x", tensor_desc);
+  op_desc->AddOutputDesc("y", tensor_desc);
+  op_desc->SetInputOffset({0});
+  op_desc->SetOutputOffset({0});
+  std::unique_ptr<MemcpyAsyncTask> task;
+  SingleOpModelParam param{};
+  param.memory_size = 1024;
+  auto ret = RtsKernelTaskBuilder::BuildMemcpyAsyncTask(op_desc, kernel_def, param, task);
+  ASSERT_EQ(ret, SUCCESS);
+  uintptr_t *addresses = nullptr;
+  size_t arg_count = 0;
+  task->GetIoAddr(addresses, arg_count);
+  ASSERT_EQ(arg_count, 2);
+  ASSERT_TRUE(addresses != nullptr);
+  ASSERT_EQ(task->LaunchKernel(nullptr), SUCCESS);
+
+  op_desc->SetInputOffset({0, 0});
+  op_desc->AddOutputDesc("y2", tensor_desc);
+  ret = RtsKernelTaskBuilder::BuildMemcpyAsyncTask(op_desc, kernel_def, param, task);
+  ASSERT_EQ(ret, INTERNAL_ERROR);
+}
+