Automated g4 rollback of changelist 160182040

PiperOrigin-RevId: 160190881

4 files changed, 16 insertions, 124 deletions

diff --git a/tensorflow/contrib/verbs/rdma.cc b/tensorflow/contrib/verbs/rdma.cc
index 6f3a616fe8..bc687be0ab 100644
--- a/tensorflow/contrib/verbs/rdma.cc
+++ b/tensorflow/contrib/verbs/rdma.cc
@@ -21,7 +21,6 @@ limitations under the License.
 #include "tensorflow/core/common_runtime/device_mgr.h"
 #include "tensorflow/core/common_runtime/dma_helper.h"
 #include "tensorflow/core/common_runtime/gpu/gpu_util.h"
-#include "tensorflow/core/common_runtime/gpu/process_state.h"
 #include "tensorflow/core/distributed_runtime/rendezvous_mgr_interface.h"
 #include "tensorflow/core/distributed_runtime/session_mgr.h"
 #include "tensorflow/core/framework/rendezvous.h"
@@ -684,6 +683,7 @@ void RdmaTensorBuffer::SendNextItem() {
                          << " error message: " << status.error_message();
       size_t buffer_size = RdmaMessage::kMessageTotalBytes;
       size_t tensor_bytes = 0;
+      TensorProto proto;
       // Figures out which device the tensor is hosted on.
       Device* src_dev = nullptr;
       Status s = channel_->adapter_->worker_env_->device_mgr->LookupDevice(
@@ -703,47 +703,21 @@ void RdmaTensorBuffer::SendNextItem() {
       CHECK(s.ok()) << "dst device not found";
       AllocatorAttributes dst_alloc_attr;
       dst_alloc_attr.set_on_host(true);
-
-      bool can_memcpy = DataTypeCanUseMemcpy(in.dtype());
       // string tensor needs to be serialized
-      Tensor copy;
-      StringPiece copy_buf;
-      TensorProto proto;
       if (src_dev->tensorflow_gpu_device_info() &&
           (!send_args.alloc_attrs.on_host())) {
         CHECK(send_args.device_context)
-          << "send dev name: " << src_dev->name()
-          << " gpu_info: " << src_dev->tensorflow_gpu_device_info();
-
-        if (can_memcpy) {
-          AllocatorAttributes host_alloc_attrs;
-          host_alloc_attrs.set_gpu_compatible(true);
-          host_alloc_attrs.set_on_host(true);
-          Allocator* alloc = ProcessState::singleton()->GetCUDAHostAllocator(0);
-          copy = Tensor(alloc, in.dtype(), in.shape());
-          s = VerbsUtil::CopyGPUTensorToCPUSync(
-              src_dev, send_args.device_context, &in, &copy);
-          CHECK(s.ok()) << "copy tensor from gpu sync";
-          copy_buf = copy.tensor_data();
-        } else {
-          // "val" is on a GPU. Uses GPUUtil to fill the proto.
-          s = VerbsUtil::SetProtoFromGPUSync(
-              in, src_dev, send_args.device_context, &proto, is_dead);
-          CHECK(s.ok()) << "set proto from gpu sync";
-        }
+            << "send dev name: " << src_dev->name()
+            << " gpu_info: " << src_dev->tensorflow_gpu_device_info();
+        // "val" is on a GPU. Uses GPUUtil to fill the proto.
+        s = VerbsUtil::SetProtoFromGPUSync(
+            in, src_dev, send_args.device_context, &proto, is_dead);
+        CHECK(s.ok()) << "set proto from gpu sync";
       } else {
         // tensor is in CPU memory.
-        if (can_memcpy) {
-          copy_buf = in.tensor_data();
-        } else {
-          in.AsProtoTensorContent(&proto);
-        }
-      }
-      if (can_memcpy) {
-        tensor_bytes = in.TotalBytes();
-      } else {
-        tensor_bytes = proto.ByteSize();
+        in.AsProtoTensorContent(&proto);
       }
+      tensor_bytes = proto.ByteSize();
       // maybe some margin for string tensor?
       buffer_size += tensor_bytes;
       // prepare message
@@ -797,16 +771,7 @@ void RdmaTensorBuffer::SendNextItem() {
               static_cast<void*>(static_cast<char*>(buffer_) +
                                  RdmaMessage::kTensorBufferStartIndex);
           CHECK(tensor_bytes + RdmaMessage::kTensorBufferStartIndex <= size_);
-          if (can_memcpy) {
-            CHECK(copy_buf.size() == tensor_bytes)
-               << "unexpected tensor size: "
-               << copy_buf.size()
-               << " != "
-               << tensor_bytes;
-            memcpy(output, copy_buf.data(), tensor_bytes);
-          } else {
-            proto.SerializeToArray(output, tensor_bytes);
-          }
+          proto.SerializeToArray(output, tensor_bytes);
         } else {
           buffer_size = RdmaMessage::kMessageTotalBytes;
         }
diff --git a/tensorflow/contrib/verbs/rdma_rendezvous_mgr.cc b/tensorflow/contrib/verbs/rdma_rendezvous_mgr.cc
index 9ea696589a..5871400f26 100644
--- a/tensorflow/contrib/verbs/rdma_rendezvous_mgr.cc
+++ b/tensorflow/contrib/verbs/rdma_rendezvous_mgr.cc
@@ -21,7 +21,6 @@ limitations under the License.
 #include "tensorflow/core/common_runtime/device.h"
 #include "tensorflow/core/common_runtime/device_mgr.h"
 #include "tensorflow/core/common_runtime/dma_helper.h"
-#include "tensorflow/core/common_runtime/gpu/process_state.h"
 #include "tensorflow/core/lib/core/errors.h"
 #include "tensorflow/core/lib/strings/numbers.h"
 #include "tensorflow/core/lib/strings/str_util.h"
@@ -100,40 +99,12 @@ void RdmaRemoteRendezvous::RecvFromRemoteAsync(
     if (!rm.is_dead_) {
       void* input = static_cast<char*>(rb->buffer_) +
                     RdmaMessage::kTensorBufferStartIndex;
-      bool can_memcpy = DataTypeCanUseMemcpy(rm.data_type_);
-      if (can_memcpy) {
-        if (dst_dev->tensorflow_gpu_device_info() &&
-            (!recv_args.alloc_attrs.on_host())) {
-          CHECK(recv_args.device_context)
-            << "send dev name: " << src_dev->name()
-            << " gpu_info: " << src_dev->tensorflow_gpu_device_info();
-          Allocator* alloc = ProcessState::singleton()->GetCUDAHostAllocator(0);
-          Tensor copy(alloc, rm.data_type_, rm.tensor_shape_);
-          memcpy(DMAHelper::base(&copy), input, rm.tensor_bytes_);
-
-          Allocator* dst_alloc = dst_dev->GetAllocator(recv_args.alloc_attrs);
-          Tensor gpu_copy(dst_alloc, rm.data_type_, rm.tensor_shape_);
-          s = VerbsUtil::CopyCPUTensorToGPUSync(&copy, recv_args.device_context,
-                                                dst_dev, &gpu_copy);
-          CHECK(s.ok()) << "copy tensor to gpu sync";
-          val = std::move(gpu_copy);
-        } else {
-          AllocatorAttributes host_alloc_attrs;
-          host_alloc_attrs.set_gpu_compatible(true);
-          host_alloc_attrs.set_on_host(true);
-          Allocator* alloc = dst_dev->GetAllocator(host_alloc_attrs);
-          Tensor copy(alloc, rm.data_type_, rm.tensor_shape_);
-          memcpy(DMAHelper::base(&copy), input, rm.tensor_bytes_);
-          val = std::move(copy);
-        }
-      } else {
-        TensorProto proto;
-        CHECK(rm.tensor_bytes_ + RdmaMessage::kTensorBufferStartIndex <=
-              rb->size_);
-        CHECK(ParseProtoUnlimited(&proto, input, rm.tensor_bytes_))
-            << "fail to parse proto from array";
-        s = dst_dev->MakeTensorFromProto(proto, recv_args.alloc_attrs, &val);
-      }
+      TensorProto proto;
+      CHECK(rm.tensor_bytes_ + RdmaMessage::kTensorBufferStartIndex <=
+            rb->size_);
+      CHECK(ParseProtoUnlimited(&proto, input, rm.tensor_bytes_))
+          << "fail to parse proto from array";
+      s = dst_dev->MakeTensorFromProto(proto, recv_args.alloc_attrs, &val);
     }
 
     rc->RemoveRecvCallback(key_with_step_id);
diff --git a/tensorflow/contrib/verbs/verbs_util.cc b/tensorflow/contrib/verbs/verbs_util.cc
index 76e44d34a9..c3350f7958 100644
--- a/tensorflow/contrib/verbs/verbs_util.cc
+++ b/tensorflow/contrib/verbs/verbs_util.cc
@@ -21,40 +21,6 @@ limitations under the License.
 namespace tensorflow {
 
 // static sync wrapper:
-Status VerbsUtil::CopyGPUTensorToCPUSync(Device* gpu_device,
-                              const DeviceContext* device_context,
-                              const Tensor* gpu_tensor,
-                              Tensor* cpu_tensor) {
-  Notification n;
-  Status status;
-  GPUUtil::CopyGPUTensorToCPU(gpu_device, device_context,
-                              gpu_tensor, cpu_tensor,
-                              [&n, &status](const Status& s) {
-                                status = s;
-                                n.Notify();
-                              });
-  n.WaitForNotification();
-  return status;
-}
-
-// static sync wrapper:
-Status VerbsUtil::CopyCPUTensorToGPUSync(const Tensor* cpu_tensor,
-                                         const DeviceContext* device_context,
-                                         Device* gpu_device,
-                                         Tensor* gpu_tensor) {
-  Notification n;
-  Status status;
-  GPUUtil::CopyCPUTensorToGPU(cpu_tensor, device_context,
-                              gpu_device, gpu_tensor,
-                              [&n, &status](const Status& s) {
-                                status = s;
-                                n.Notify();
-                              });
-  n.WaitForNotification();
-  return status;
-}
-
-// static sync wrapper:
 Status VerbsUtil::SetProtoFromGPUSync(const Tensor& tensor, Device* dev,
                                       const DeviceContext* device_context,
                                       TensorProto* proto, bool is_dead) {
diff --git a/tensorflow/contrib/verbs/verbs_util.h b/tensorflow/contrib/verbs/verbs_util.h
index d9da396228..cbc01adae4 100644
--- a/tensorflow/contrib/verbs/verbs_util.h
+++ b/tensorflow/contrib/verbs/verbs_util.h
@@ -28,16 +28,6 @@ class TensorProto;
 
 class VerbsUtil {
  public:
-  // synchronous wrapper of CopyGPUTensorToCPU
-  static Status CopyGPUTensorToCPUSync(Device* gpu_device,
-                                       const DeviceContext* device_context,
-                                       const Tensor* gpu_tensor,
-                                       Tensor* cpu_tensor);
-  // synchronous wrapper of CopyCPUTensorToGPU
-  static Status CopyCPUTensorToGPUSync(const Tensor* cpu_tensor,
-                                       const DeviceContext* device_context,
-                                       Device* gpu_device,
-                                       Tensor* gpu_tensor);
   // synchronous wrapper of SetProtoFromGPU
   static Status SetProtoFromGPUSync(const Tensor& tensor, Device* dev,
                                     const DeviceContext* device_context,