From 059b05a72e275c402f179ffc9f542dd0ccf04818 Mon Sep 17 00:00:00 2001
From: chujinjin <chujinjin52@huawei.com>
Date: Mon, 17 May 2021 15:23:41 +0800
Subject: [PATCH] fix memcopy async error

---
 .../runtime/device/ascend/ascend_device_address.cc    |  2 +-
 .../runtime/device/ascend/ascend_kernel_runtime.cc    | 11 +++++------
 .../ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc    |  8 +-------
 mindspore/ccsrc/runtime/device/kernel_runtime.h       |  3 ---
 4 files changed, 7 insertions(+), 17 deletions(-)

diff --git a/mindspore/ccsrc/runtime/device/ascend/ascend_device_address.cc b/mindspore/ccsrc/runtime/device/ascend/ascend_device_address.cc
index f4368970018..909ecb5d909 100644
--- a/mindspore/ccsrc/runtime/device/ascend/ascend_device_address.cc
+++ b/mindspore/ccsrc/runtime/device/ascend/ascend_device_address.cc
@@ -113,7 +113,7 @@ void SyncMemory(void *dst, const void *src, uint64_t size, rtMemcpyKind_t kind)
       MS_EXCEPTION(DeviceProcessError) << "rtMemcpy failed";
     }
   } else {
-    auto ret = runtime_instance->MemcpyAsync(dst, src, size, static_cast<int32_t>(kind));
+    auto ret = runtime_instance->MemcpyAsync(dst, src, size, static_cast<int32_t>(RT_MEMCPY_HOST_TO_DEVICE_EX));
     if (!ret) {
       MS_EXCEPTION(DeviceProcessError) << "MemcpyAsync failed";
     }
diff --git a/mindspore/ccsrc/runtime/device/ascend/ascend_kernel_runtime.cc b/mindspore/ccsrc/runtime/device/ascend/ascend_kernel_runtime.cc
index 690638762f8..5b26fc06f53 100644
--- a/mindspore/ccsrc/runtime/device/ascend/ascend_kernel_runtime.cc
+++ b/mindspore/ccsrc/runtime/device/ascend/ascend_kernel_runtime.cc
@@ -703,7 +703,6 @@ bool AscendKernelRuntime::SyncStream() {
     MS_LOG(ERROR) << "Call runtime rtStreamSynchronize error.";
     return false;
   }
-  FreeAndClearBufferPtrs();
   return true;
 }
 
@@ -714,11 +713,11 @@ bool AscendKernelRuntime::MemcpyAsync(void *dst, const void *src, uint64_t size,
     return false;
   }
 
-  std::shared_ptr<char[]> buffer(new char[size]());
-  MS_EXCEPTION_IF_NULL(buffer);
-  std::copy(reinterpret_cast<const char *>(src), reinterpret_cast<const char *>(src) + size, buffer.get());
-  AddBufferPtr(buffer);
-  if (RT_ERROR_NONE != rtMemcpyAsync(dst, size, buffer.get(), size, static_cast<rtMemcpyKind_t>(kind), stream_)) {
+  auto copy_kind = static_cast<rtMemcpyKind_t>(kind);
+  if (copy_kind != RT_MEMCPY_HOST_TO_DEVICE_EX) {
+    MS_LOG(EXCEPTION) << "Memory copy async not support cache host buffer in kind: " << kind;
+  }
+  if (RT_ERROR_NONE != rtMemcpyAsync(dst, size, src, size, static_cast<rtMemcpyKind_t>(kind), stream_)) {
     MS_LOG(ERROR) << "Call runtime rtMemcpyAsync error.";
     return false;
   }
diff --git a/mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc b/mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc
index ab432c080e3..58c027a804c 100644
--- a/mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc
+++ b/mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc
@@ -65,7 +65,6 @@ bool GPUKernelRuntime::SyncStream() {
     MS_LOG(ERROR) << "Call SyncStream error.";
     return false;
   }
-  FreeAndClearBufferPtrs();
   return true;
 }
 
@@ -198,14 +197,9 @@ void LoadKernelData(Debugger *debugger, const CNodePtr &kernel,
 }  // namespace
 
 bool GPUKernelRuntime::MemcpyAsync(void *dst, const void *src, uint64_t size, int32_t kind) {
-  std::shared_ptr<char[]> buffer(new char[size]());
-  MS_EXCEPTION_IF_NULL(buffer);
-  std::copy(reinterpret_cast<const char *>(src), reinterpret_cast<const char *>(src) + size, buffer.get());
-  AddBufferPtr(buffer);
-
   auto &stream = GPUDeviceManager::GetInstance().default_stream();
   MS_EXCEPTION_IF_NULL(stream);
-  auto ret = GPUDeviceManager::GetInstance().CopyHostMemToDeviceAsync(dst, buffer.get(), size, stream);
+  auto ret = GPUDeviceManager::GetInstance().CopyHostMemToDeviceAsync(dst, src, size, stream);
   if (!ret) {
     MS_LOG(ERROR) << "CopyHostMemToDeviceAsync failed";
     return false;
diff --git a/mindspore/ccsrc/runtime/device/kernel_runtime.h b/mindspore/ccsrc/runtime/device/kernel_runtime.h
index 562bc407b3d..de7cfb4be9c 100644
--- a/mindspore/ccsrc/runtime/device/kernel_runtime.h
+++ b/mindspore/ccsrc/runtime/device/kernel_runtime.h
@@ -105,8 +105,6 @@ class KernelRuntime {
 
   virtual void PreInit() {}
   virtual uint64_t GetAvailableMemMaxSize() const { return 0; }
-  void AddBufferPtr(std::shared_ptr<char[]> ptr) { buffer_ptrs_.push_back(ptr); }
-  void FreeAndClearBufferPtrs() { buffer_ptrs_.clear(); }
   void GenKernelEvents(const session::KernelGraph *graph);
   virtual std::shared_ptr<DeviceEvent> CreateDeviceEvent() { return nullptr; }
   virtual DeviceAddressType GetTargetDeviceAddressType() const = 0;
@@ -162,7 +160,6 @@ class KernelRuntime {
   std::map<uint32_t,
            std::pair<std::vector<std::vector<std::function<void()>>>, std::vector<std::vector<std::function<void()>>>>>
     graph_kernel_events_map_;
-  std::vector<std::shared_ptr<char[]>> buffer_ptrs_ = {};
 };
 using KernelRuntimePtr = std::shared_ptr<KernelRuntime>;
 }  // namespace device