[XLA] Remove stream priming from Backend

This was originally added a long time ago to help with GPU benchmarks; I think
it adds needless complexity to the Backend class, which we intend to simplify.
If this is still an issue for benchmarks, it should be done on a higher level.

PiperOrigin-RevId: 157037991

2 files changed, 0 insertions, 19 deletions

diff --git a/tensorflow/compiler/xla/service/backend.cc b/tensorflow/compiler/xla/service/backend.cc
index 1913617fec..5fa515f9c2 100644
--- a/tensorflow/compiler/xla/service/backend.cc
+++ b/tensorflow/compiler/xla/service/backend.cc
@@ -99,8 +99,6 @@ struct Backend::EigenThreadPoolWrapper {
   std::unique_ptr<Backend> backend(
       new Backend(replica_count, platform, compiler, stream_executors,
                   transfer_manager, options.intra_op_parallelism_threads()));
-  TF_RETURN_IF_ERROR(backend->PoolStreams(kInitialStreamsToPool,
-                                          backend->default_stream_executor()));
   return std::move(backend);
 }
 
@@ -113,15 +111,6 @@ Backend::CreateDefaultBackend() {
   return CreateBackend(backend_options);
 }
 
-tensorflow::Status Backend::PoolStreams(int n, se::StreamExecutor* executor) {
-  std::vector<StreamPtr> primed;
-  for (int i = 0; i < n; ++i) {
-    TF_ASSIGN_OR_RETURN(auto stream, BorrowStream(executor));
-    primed.emplace_back(std::move(stream));
-  }
-  return tensorflow::Status::OK();
-}
-
 StatusOr<Backend::StreamPtr> Backend::BorrowStream(int device_ordinal) {
   TF_ASSIGN_OR_RETURN(auto exec, stream_executor(device_ordinal));
   return BorrowStream(exec);
diff --git a/tensorflow/compiler/xla/service/backend.h b/tensorflow/compiler/xla/service/backend.h
index 1068bac277..dffb1f910c 100644
--- a/tensorflow/compiler/xla/service/backend.h
+++ b/tensorflow/compiler/xla/service/backend.h
@@ -74,9 +74,6 @@ class Backend {
  public:
   using StreamPtr = Pool<perftools::gputools::Stream>::SmartPtr;
 
-  // The number of streams we create for the pool at initialization time.
-  static constexpr int kInitialStreamsToPool = 8;
-
   // Creates a new backend for the given platform with the given number of
   // replicas.
   static StatusOr<std::unique_ptr<Backend>> CreateBackend(
@@ -133,11 +130,6 @@ class Backend {
     return stream_executors_[0];
   }
 
-  // Primes the internal pool of streams for BorrowStream with n initialized
-  // stream instances.
-  tensorflow::Status PoolStreams(int n,
-                                 perftools::gputools::StreamExecutor* executor);
-
   // Borrows a stream for use by the caller, either by grabbing it from an
   // internal pool, or by constructing/initializating it, and returns the result
   // to the caller.