#include "tensorflow/compiler/tf2xla/xla_compiler.h"
#include "tensorflow/compiler/tf2xla/xla_helpers.h"
#include "xla/client/local_client.h"
#include "xla/executable_run_options.h"
#include "xla/hlo/ir/hlo_input_output_alias_config.h"
#include "xla/pjrt/pjrt_client.h"
#include "xla/pjrt/tf_pjrt_client.h"
#include "xla/service/executable.h"
#include "xla/service/gpu/gpu_executable_run_options.h"
#include "tensorflow/core/framework/function.h"

           "Minimum number of operators in an XLA compilation. Ignored for "
           "operators placed on an XLA device or operators explicitly marked "
           "for compilation."),
      Flag("tf_xla_max_cluster_size",
           &mark_for_compilation_flags->tf_xla_max_cluster_size,
           "Maximum number of operators in an XLA compilation."),
      Flag(
          "tf_xla_ops_to_cluster",

        "@local_xla//xla:xla_data_proto_cc",
        "@local_xla//xla/client:client_library",
        "@local_xla//xla/client:compile_only_client",
        "@local_xla//xla/client:xla_computation",
        "@local_xla//xla/service:compiler",
        "@local_xla//xla/service/cpu:buffer_info_util",
        "@local_xla//xla/service/cpu:cpu_compiler",
        "@local_xla//xla/stream_executor:platform_manager",
    ],
)

tf_cc_test(

        "@local_tsl//tsl/platform:protobuf",
        "@local_xla//xla/client:client_library",
        "@local_xla//xla/client:compile_only_client",
        "@local_xla//xla/service/cpu:cpu_compiler",
        "@local_xla//xla/service/cpu:cpu_transfer_manager",
        "@local_xla//xla/stream_executor",
        "@local_xla//xla/stream_executor/host:host_platform",
        "@local_xla//xla/stream_executor/host:host_platform_id",
    ],

#include "tensorflow/compiler/jit/variable_info.h"
#include "tensorflow/compiler/jit/xla_tensor.h"
#include "tensorflow/compiler/tf2xla/xla_compiler.h"
#include "xla/client/local_client.h"
#include "xla/pjrt/pjrt_client.h"
#include "xla/service/shaped_buffer.h"
#include "xla/stream_executor/device_memory_allocator.h"
#include "tensorflow/core/framework/allocation_description.pb.h"
#include "tensorflow/core/framework/tensor.h"

      $resource,
      (TF_SubOp
        (CreateTFReadVariableOp $src_op, $value, $resource),
        $value
      ),
      (CreateConstBoolAttrFalse)
    )
  >;

// This decomposition is only correct inside XLA as it ignores use_locking
// attribute.
def DecomposeResourceApplyGradientDescentOp :
  Pat<
    (TF_ResourceApplyGradientDescentOp:$src_op $resource, $alpha, $delta, BoolAttr:$_),
    (TF_AssignVariableOp
      $resource,

// The XlaDevice executes a TensorFlow graph using the XLA linear algebra
// runtime.
//
// Operators assigned to an XlaDevice are compiled into XLA computations.
// Tensors on an XlaDevice are thin wrappers around XLA ScopedShapedBuffers.
//
// XlaDevice is instantiated separately for each XLA backend (e.g., CPU or GPU),
// under different names (e.g., XLA_CPU or XLA_GPU).

  EXPECT_EQ(exec_args.size(), 2);

  std::shared_ptr<xla::Literal> literal1 = *exec_args[0]->ToLiteralSync();
  EXPECT_TRUE(xla::LiteralTestUtil::Equal(
      *literal1, xla::LiteralUtil::CreateR2<int32_t>({{1, 2, 3}})));

  std::shared_ptr<xla::Literal> literal2 = *exec_args[1]->ToLiteralSync();
  EXPECT_TRUE(xla::LiteralTestUtil::Equal(
      *literal2, xla::LiteralUtil::CreateR2<int32_t>({{4, 5, 6}})));
}

#include "tensorflow/compiler/jit/xla_platform_info.h"
#include "tensorflow/compiler/tf2xla/xla_compiler.h"
#include "xla/client/executable_build_options.h"
#include "xla/client/local_client.h"
#include "xla/service/hlo_graph_dumper.h"
#include "xla/status_macros.h"
#include "xla/stream_executor/host/host_platform_id.h"
#include "xla/stream_executor/platform.h"
#include "tensorflow/core/common_runtime/eager/tensor_handle.h"

//     eliminates the use of resource variables.
//   . Legalizes the operations to XLA HLO operations.
//   . Canonicalizes the XLA HLO operations.
//
// device_type: XLA JIT device to use for compilation such as "XLA_CPU_JIT",
//   "XLA_GPU_JIT" or "XLA_TPU_JIT".
// use_tuple_args: when this is true, always create a tuple argument for the
//   entry computation.