Home | History | Annotate | Download | only in kernels 
      1 /* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
      2 
      3 Licensed under the Apache License, Version 2.0 (the "License");
      4 you may not use this file except in compliance with the License.
      5 You may obtain a copy of the License at
      6 
      7     http://www.apache.org/licenses/LICENSE-2.0
      8 
      9 Unless required by applicable law or agreed to in writing, software
     10 distributed under the License is distributed on an "AS IS" BASIS,
     11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
     12 See the License for the specific language governing permissions and
     13 limitations under the License.
     14 ==============================================================================*/
     15 
     16 // RpcOp is a TensorFlow op that sends and receives arbitrary messages.
     17 //
     18 // See docs in ../ops/rpc_op.cc.
     19 
     20 #include <memory>
     21 #include <string>
     22 #include <vector>
     23 
     24 #include "third_party/eigen3/Eigen/Core"
     25 #include "tensorflow/core/framework/op_kernel.h"
     26 #include "tensorflow/core/framework/tensor_types.h"
     27 #include "tensorflow/core/lib/core/errors.h"
     28 #include "tensorflow/core/lib/core/refcount.h"
     29 #include "tensorflow/core/lib/gtl/stl_util.h"
     30 #include "tensorflow/core/lib/strings/stringprintf.h"
     31 #include "tensorflow/core/platform/env.h"
     32 #include "tensorflow/core/platform/logging.h"
     33 #include "tensorflow/core/util/rpc/call_container.h"
     34 #include "tensorflow/core/util/rpc/rpc_factory.h"
     35 #include "tensorflow/core/util/rpc/rpc_factory_registry.h"
     36 
     37 namespace tensorflow {
     38 
     39 class RpcOp : public AsyncOpKernel {
     40  public:
     41   explicit RpcOp(OpKernelConstruction* context) : AsyncOpKernel(context) {
     42     OP_REQUIRES_OK(context, context->GetAttr("protocol", &protocol_));
     43     OP_REQUIRES(context, !protocol_.empty(),
     44                 errors::InvalidArgument("protocol must be non-empty."));
     45     bool fail_fast;
     46     OP_REQUIRES_OK(context, context->GetAttr("fail_fast", &fail_fast));
     47     int64 timeout_in_ms;
     48     OP_REQUIRES_OK(context, context->GetAttr("timeout_in_ms", &timeout_in_ms));
     49 
     50     RPCFactoryRegistry::RPCFactoryFn* rpc_factory_fn =
     51         RPCFactoryRegistry::Global()->Get(protocol_);
     52     OP_REQUIRES(context, rpc_factory_fn != nullptr,
     53                 errors::InvalidArgument("The protocol ", protocol_,
     54                                         " was not recognized."));
     55 
     56     rpc_factory_.reset((*rpc_factory_fn)(context, fail_fast, timeout_in_ms));
     57   }
     58 
     59   ~RpcOp() override {}
     60 
     61   void ComputeAsync(OpKernelContext* ctx, DoneCallback done) override {
     62     const Tensor& address_t = ctx->input(0);
     63     const Tensor& method_t = ctx->input(1);
     64     const Tensor& request_t = ctx->input(2);
     65 
     66     OP_REQUIRES_ASYNC(
     67         ctx, address_t.dims() == 0 || address_t.dims() == 1,
     68         errors::InvalidArgument("address must be a scalar or vector."), done);
     69     OP_REQUIRES_ASYNC(
     70         ctx, method_t.dims() == 0 || method_t.dims() == 1,
     71         errors::InvalidArgument("method must be a scalar or vector."), done);
     72     OP_REQUIRES_ASYNC(
     73         ctx, request_t.dims() == 0 || request_t.dims() == 1,
     74         errors::InvalidArgument("request must be a scalar or vector."), done);
     75 
     76     TensorShape output_shape({});
     77     for (const Tensor& t : {address_t, method_t, request_t}) {
     78       if (t.dims() == 1) {
     79         OP_REQUIRES_ASYNC(
     80             ctx,
     81             output_shape.dims() == 0 ||
     82                 output_shape.dim_size(0) == t.dim_size(0),
     83             errors::InvalidArgument(
     84                 "Input vector shapes don't match: ", output_shape.DebugString(),
     85                 " vs. ", t.shape().DebugString()),
     86             done);
     87         output_shape = t.shape();
     88       }
     89     }
     90 
     91     Tensor* response_t;
     92     OP_REQUIRES_OK_ASYNC(
     93         ctx, ctx->allocate_output(0, output_shape, &response_t), done);
     94 
     95     const bool try_rpc = (ctx->num_outputs() > 1);
     96 
     97     Tensor* status_code_t = nullptr;
     98     Tensor* status_message_t = nullptr;
     99     if (try_rpc) {
    100       OP_REQUIRES_OK_ASYNC(
    101           ctx, ctx->allocate_output(1, output_shape, &status_code_t), done);
    102       OP_REQUIRES_OK_ASYNC(
    103           ctx, ctx->allocate_output(2, output_shape, &status_message_t), done);
    104     }
    105 
    106     if (request_t.NumElements() == 0) {
    107       // Special case, we finished early!
    108       done();
    109       return;
    110     }
    111 
    112     int64 num_elements = output_shape.num_elements();
    113 
    114     rpc_factory_->Call(ctx, num_elements, address_t, method_t, request_t,
    115                        try_rpc, response_t, status_code_t, status_message_t,
    116                        std::move(done));
    117   }
    118 
    119  private:
    120   string protocol_;
    121   std::unique_ptr<RPCFactory> rpc_factory_;
    122 
    123   TF_DISALLOW_COPY_AND_ASSIGN(RpcOp);
    124 };
    125 
    126 REGISTER_KERNEL_BUILDER(Name("Rpc").Device(DEVICE_CPU), RpcOp);
    127 REGISTER_KERNEL_BUILDER(Name("TryRpc").Device(DEVICE_CPU), RpcOp);
    128 
    129 }  // namespace tensorflow
    130