[TF:XLA] Clean up unused XLA options and functions.

PiperOrigin-RevId: 175217850

tensorflow/compiler/jit/kernels/xla_launch_op.cc

@@ -257,7 +257,6 @@ void XlaLocalLaunchOp::Compute(OpKernelContext* ctx) {
   options.flib_def = ctx->function_library()->GetFunctionLibraryDefinition();
   options.graph_def_version = ctx->function_library()->graph_def_version();
   options.allow_cpu_custom_calls = (platform_id_ == gpu::host::kHostPlatformId);
-  options.local_executable_has_hybrid_result = true;
 
   const XlaCompiler::CompilationResult* kernel;
   xla::LocalExecutable* executable;

tensorflow/compiler/jit/xla_compilation_cache.cc

@@ -227,10 +227,7 @@ Status XlaCompilationCache::BuildExecutable(
   }
   xla::ExecutableBuildOptions build_options;
   build_options.set_device_ordinal(client_->default_device_ordinal());
-  build_options.set_platform(client_->platform());
   build_options.set_result_layout(result.xla_output_shape);
-  build_options.set_has_hybrid_result(
-      options.local_executable_has_hybrid_result);
 
   auto compile_result =
       client_->Compile(*result.computation, argument_layouts, build_options);

tensorflow/compiler/tf2xla/xla_compiler.h

@@ -236,12 +236,6 @@ class XlaCompiler {
     // to the computation.
     bool allow_cpu_custom_calls = false;
 
-    // If 'local_executable_has_hybrid_result', the top-level pointers of the
-    // result tuple of compiled programs are stored in host memory and the
-    // nested buffers in device memory, otherwise the whole result tuple is
-    // stored in device memory.
-    bool local_executable_has_hybrid_result = false;
-
     // If not nullptr, populate_resource_manager is called with the
     // compilation device's resource manager when the compilation
     // device is created, and can be used to create metadata objects

tensorflow/compiler/xla/client/local_client.cc

@@ -27,16 +27,6 @@ namespace se = ::perftools::gputools;
 
 namespace xla {
 
-ExecutableBuildOptions& ExecutableBuildOptions::set_platform(
-    perftools::gputools::Platform* platform) {
-  platform_ = platform;
-  return *this;
-}
-
-perftools::gputools::Platform* ExecutableBuildOptions::platform() const {
-  return platform_;
-}
-
 ExecutableBuildOptions& ExecutableBuildOptions::set_device_ordinal(
     int device_ordinal) {
   device_ordinal_ = device_ordinal;
@@ -56,16 +46,6 @@ const Shape* ExecutableBuildOptions::result_layout() const {
   return result_layout_set_ ? &result_layout_ : nullptr;
 }
 
-ExecutableBuildOptions& ExecutableBuildOptions::set_has_hybrid_result(
-    bool has_hybrid_result) {
-  has_hybrid_result_ = has_hybrid_result;
-  return *this;
-}
-
-bool ExecutableBuildOptions::has_hybrid_result() const {
-  return has_hybrid_result_;
-}
-
 namespace {
 StatusOr<Backend::StreamPtr> BorrowStreamForDevice(int device_ordinal,
                                                    Backend* backend) {

tensorflow/compiler/xla/client/local_client.h

@@ -37,14 +37,6 @@ namespace xla {
 // LocalClient::Compile.
 class ExecutableBuildOptions {
  public:
-  // If set, this is the platform to build the computation for. This must match
-  // the underlying platform of the service. A value of nullptr indicates the
-  // option has not been set.
-  //
-  // TODO(b/28616830): Support multiple platforms.
-  ExecutableBuildOptions& set_platform(perftools::gputools::Platform* platform);
-  perftools::gputools::Platform* platform() const;
-
   // If set, this is the device to build the computation for. Valid
   // device_ordinal values are: 0 to # of devices - 1. These values are
   // identical to the device ordinal values used by StreamExecutor. The built
@@ -61,18 +53,10 @@ class ExecutableBuildOptions {
   ExecutableBuildOptions& set_result_layout(const Shape& shape_with_layout);
   const Shape* result_layout() const;
 
-  // If set, the executable will be built to output a hybrid
-  // ShapedBuffer with top-level tuple pointers in host memory and
-  // result buffers in device memory.
-  ExecutableBuildOptions& set_has_hybrid_result(bool has_hybrid_result);
-  bool has_hybrid_result() const;
-
  private:
-  perftools::gputools::Platform* platform_ = nullptr;
   int device_ordinal_ = -1;
   Shape result_layout_;
   bool result_layout_set_ = false;
-  bool has_hybrid_result_ = true;
 };
 
 class LocalExecutable {

tensorflow/compiler/xla/service/hlo_module_config.cc

@@ -39,8 +39,8 @@ void HloModuleConfig::SetDefaultComputationLayout(
 }
 
 string HloModuleConfig::compilation_cache_key() const {
-  string key = tensorflow::strings::StrCat("profiling=", hlo_profiling_enabled_,
-                                           "::hybrid=", has_hybrid_result_);
+  string key =
+      tensorflow::strings::StrCat("profiling=", hlo_profiling_enabled_);
   StrAppend(&key, "::(");
   std::vector<string> params;
   for (const ShapeLayout& param_layout :

tensorflow/compiler/xla/service/hlo_module_config.h

@@ -104,16 +104,6 @@ class HloModuleConfig {
   // Whether to enable HLO-level profiling.
   bool hlo_profiling_enabled_ = false;
 
-  // If this flag is true, the generated executable will return a ShapedBuffer
-  // holding the result of the computation. In a ShapedBuffer, tuples have their
-  // structure held in host memory and the element arrays (leaves of the tuple
-  // structure) stored in device memory. The ShapedBuffer is considered "hybrid"
-  // because its leaves are on device but its structure is stored on
-  // host. Otherwise, if this flag is false, the generated executable will
-  // return a DeviceMemoryBase where the result is held entirely in device
-  // memory.
-  bool has_hybrid_result_ = false;
-
   // Module/graph-level seed handle.
   uint64 seed_ = 0;
 

tensorflow/compiler/xla/service/local_service.cc

@@ -68,26 +68,6 @@ LocalService::LocalService(const ServiceOptions& options,
                            std::unique_ptr<Backend> execute_backend)
     : Service(options, std::move(execute_backend)) {}
 
-namespace {
-// Returns the space required to allocate a shape. If
-// allocate_space_for_deep_copy the space includes all sub-buffers of
-// a tuple.
-int64 RequiredSpace(const Shape& shape, bool allocate_space_for_deep_copy,
-                    TransferManager* transfer_manager) {
-  int64 size = 0;
-  // TODO(b/33492279) remove once no devices represent result tuples as
-  // contiguous buffers.
-  if (allocate_space_for_deep_copy) {
-    ShapeUtil::ForEachSubshape(
-        shape, [&size, transfer_manager](const Shape& subshape,
-                                         const ShapeIndex& /*index*/) {
-          size += transfer_manager->GetByteSizeRequirement(subshape);
-        });
-  }
-  return size;
-}
-}  // namespace
-
 StatusOr<std::unique_ptr<Executable>> LocalService::CompileExecutable(
     const ComputationHandle& computation,
     const tensorflow::gtl::ArraySlice<const Shape*> argument_layouts,

tensorflow/compiler/xla/service/service.h

@@ -272,8 +272,6 @@ class Service : public ServiceInterface {
 
   // Create a Hlo module config for the given program shape and arguments.
   // execution_options is optional; if not given a default is used.
-  // has_hybrid_result is used to initialize the same-named field in
-  // HloModuleConfig -- see that class for documentation.
   StatusOr<std::unique_ptr<HloModuleConfig>> CreateModuleConfig(
       const ProgramShape& program_shape,
       tensorflow::gtl::ArraySlice<const Shape*> argument_shapes,