/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");;

you may not use this file except in compliance with the License.

You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software

distributed under the License is distributed on an "AS IS" BASIS,

WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

See the License for the specific language governing permissions and

limitations under the License.

==============================================================================*/

#include <memory>

// Must be included first

// clang-format off

#include "absl/status/status.h"

#include "absl/strings/str_cat.h"

#include "pybind11/attr.h" // from @pybind11

#include "xla/tsl/python/lib/core/numpy.h" //NOLINT

// clang-format on

#include "Python.h"

#include "absl/strings/str_format.h"

#include "absl/strings/str_join.h"

#include "absl/strings/str_split.h"

#include "include/dlpack/dlpack.h" // from @dlpack

#include "pybind11/chrono.h" // from @pybind11

#include "pybind11/complex.h" // from @pybind11

#include "pybind11/functional.h" // from @pybind11

#include "pybind11/pybind11.h" // from @pybind11

#include "pybind11/pytypes.h" // from @pybind11

#include "pybind11/stl.h" // from @pybind11

#include "tensorflow/c/c_api.h"

#include "tensorflow/c/c_api_experimental.h"

#include "tensorflow/c/eager/c_api.h"

#include "tensorflow/c/eager/c_api_experimental.h"

#include "tensorflow/c/eager/c_api_internal.h"

#include "tensorflow/c/eager/dlpack.h"

#include "tensorflow/c/eager/tfe_cancellation_manager_internal.h"

#include "tensorflow/c/eager/tfe_context_internal.h"

#include "tensorflow/c/eager/tfe_tensorhandle_internal.h"

#include "tensorflow/c/safe_ptr.h"

#include "tensorflow/c/tf_status.h"

#include "tensorflow/c/tf_status_helper.h"

#include "tensorflow/compiler/jit/flags.h"

#include "tensorflow/compiler/jit/get_compiler_ir.h"

#include "tensorflow/core/common_runtime/eager/context.h"

#include "tensorflow/python/eager/pywrap_tensor_conversion.h"

#include "tensorflow/python/eager/pywrap_tfe.h"

#include "tensorflow/python/lib/core/py_exception_registry.h"

#include "tensorflow/python/lib/core/pybind11_lib.h"

#include "tensorflow/python/lib/core/pybind11_status.h"

#include "tensorflow/python/lib/core/safe_pyobject_ptr.h"

#include "tensorflow/python/util/util.h"

// TODO(b/309152522): Remove this switch once it works on Windows.

#define IS_OSS true

#if !IS_OSS

#include "pybind11_protobuf/native_proto_caster.h" // from @pybind11_protobuf

#endif

namespace py = pybind11;

PYBIND11_MAKE_OPAQUE(TFE_Executor);

PYBIND11_MAKE_OPAQUE(TFE_ContextOptions);

PYBIND11_MAKE_OPAQUE(tensorflow::CancellationManager);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringCounter0);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringCounter1);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringCounter2);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringStringGauge0);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringStringGauge1);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringStringGauge2);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringStringGauge3);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringStringGauge4);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringIntGauge0);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringIntGauge1);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringIntGauge2);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringBoolGauge0);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringBoolGauge1);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringBoolGauge2);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringSampler0);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringSampler1);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringSampler2);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringCounterCell);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringIntGaugeCell);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringStringGaugeCell);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringBoolGaugeCell);

PYBIND11_MAKE_OPAQUE(TFE_MonitoringSamplerCell);

PYBIND11_MAKE_OPAQUE(TF_DeviceList);

PYBIND11_MAKE_OPAQUE(TF_Function);

PYBIND11_MAKE_OPAQUE(TF_Buffer);

// Eager helper functions migrated from pywrap_tfe.i.

namespace tensorflow {

// We cannot use Context as an opaque type. SWIG also had

// difficult directly passing the pointer around. These

// typemaps are migrated over from pywrap_tfe.i. I tried

// using a custom type caster, but we get segfaults periodically.

// TODO(amitpatankar): Move input and output logic of Context into a

// pybind11 custom type caster.

TFE_Context* InputTFE_Context(const py::handle& ctx) {

return static_cast<TFE_Context*>(PyCapsule_GetPointer(ctx.ptr(), nullptr));

}

PyObject* OutputTFE_Context(TFE_Context* context) {

return PyCapsule_New(context, nullptr, TFE_DeleteContextCapsule);

}

TF_Buffer* ProtoStringToTFBuffer(PyObject* input) {

// Convert a Python string object to TF_Buffer.

char* c_string;

Py_ssize_t py_size;

// PyBytes_AsStringAndSize() does not copy but simply interprets the input

if (PyBytes_AsStringAndSize(input, &c_string, &py_size) == -1) {

// Python has raised an error (likely TypeError or UnicodeEncodeError).

throw py::error_already_set();

}

return TF_NewBufferFromString(static_cast<void*>(c_string),

static_cast<size_t>(py_size));

}

// These functions are typemaps from the Python side. I did not use

// a custom type caster since the logic is slightly harder to follow. This

// converter is also only used once in `TFE_Py_ExecuteCancelable_wrapper`.

TFE_InputTensorHandles InputTFE_InputTensorHandles(

const py::handle& input_tensors) {

TFE_InputTensorHandles input_tensor_handles;

if (input_tensors.ptr() != Py_None) {

if (!PyList_Check(input_tensors.ptr())) {

tensorflow::ThrowTypeError("must provide a list of Tensors as inputs");

}

Py_ssize_t len = PyList_Size(input_tensors.ptr());

input_tensor_handles.resize(len);

for (Py_ssize_t i = 0; i < len; ++i) {

PyObject* elem = PyList_GetItem(input_tensors.ptr(), i);

if (!elem) {

tensorflow::ThrowTypeError("Input Tensor does not exist.");

}

if (EagerTensor_CheckExact(elem)) {

(input_tensor_handles)[i] = EagerTensor_Handle(elem);

} else if (tensorflow::swig::IsEagerTensorSlow(elem)) {

// Use equivalent of object.__getattribute__ to get the underlying

// tf wrapped EagerTensor (if there is one).

tensorflow::Safe_PyObjectPtr tf_should_use_attr(

#if PY_MAJOR_VERSION < 3

PyString_InternFromString("_tf_should_use_wrapped_value")

#else

PyUnicode_InternFromString("_tf_should_use_wrapped_value")

#endif

);

tensorflow::Safe_PyObjectPtr value_attr(

PyObject_GenericGetAttr(elem, tf_should_use_attr.get()));

if (value_attr) {

// This is an EagerTensor wrapped inside a TFShouldUse wrapped object.

(input_tensor_handles)[i] = EagerTensor_Handle(value_attr.get());

} else {

// This is a subclass of EagerTensor that we don't support.

PyErr_Clear();

tensorflow::ThrowTypeError(

tensorflow::strings::StrCat(

"Saw an object that is an instance of a strict subclass of "

"EagerTensor, which is not supported. Item ",

i, " is type: ", elem->ob_type->tp_name)

.c_str());

}

} else if (tensorflow::swig::IsTensorProtocol(elem) &&

tensorflow::swig::IsCoreTypeValue(elem)) {

// For WeakTensors, fetch the underlying Tensors.

// This is placed after the branches `IsEagerTensorSlow` and

// `EagerTensor_CheckExact` to ensure those paths are quick.

elem = PyObject_CallMethod(elem, "__tf_tensor__", nullptr);

(input_tensor_handles)[i] = EagerTensor_Handle(elem);

} else if (tensorflow::swig::IsTensor(elem)) {

// If it isnt an EagerTensor, but is still a Tensor, it must be a graph

// tensor.

tensorflow::Safe_PyObjectPtr py_tensor_repr(PyObject_Repr(elem));

std::string tensor_repr =

py_tensor_repr ? TFE_GetPythonString(py_tensor_repr.get())

: "<unknown>";

tensorflow::Safe_PyObjectPtr py_op(PyObject_GetAttrString(elem, "op"));

tensorflow::Safe_PyObjectPtr py_defined_graph(

PyObject_GetAttrString(py_op.get(), "graph"));

tensorflow::Safe_PyObjectPtr py_defined_graph_str(

PyObject_Str(py_defined_graph.get()));

std::string defined_graph_str =

py_defined_graph_str

? TFE_GetPythonString(py_defined_graph_str.get())

: "<unknown>";

tensorflow::Safe_PyObjectPtr c_op(

PyObject_GetAttrString(py_op.get(), "_c_op"));

auto& node = py::cast<TF_Operation*>(c_op.get())->node;

auto node_name_str = node.name();

std::string frame_str, traceback_str;

if (auto stack_trace = node.GetStackTrace()) {

auto frame = stack_trace->LastUserFrame();

frame_str =

absl::StrFormat("File \"%s\", line %d, in %s", frame.file_name,

frame.line_number, frame.function_name);

auto stack_trace_list =

absl::StrSplit(stack_trace->ToString({true}), '\n');

traceback_str = absl::StrJoin(

stack_trace_list, "", [&](std::string* out, const auto line) {

absl::StrAppend(out, " ", line, "\n");

});

} else {

frame_str = "<unknown>";

traceback_str = "<unknown>\n";

}

// Keep in sync with func_graph.py.

// TODO(b/200991648): Unify those two paths.

tensorflow::ThrowTypeError(

tensorflow::strings::StrCat(

tensor_repr,

" is out of scope and cannot be used here. "

"Use return values, explicit Python locals or TensorFlow "

"collections to access it.\n"

"Please see https://www.tensorflow.org/guide/"

"function#all_outputs_of_a_tffunction_must_be_return_values "

"for more information.\n\n",

tensor_repr, " was defined here:\n", traceback_str,

"\nThe tensor ", tensor_repr,

" cannot be accessed from here, because it was "

"defined in ",

defined_graph_str, ", which is out of scope.")

.c_str());

} else {

tensorflow::ThrowTypeError(

tensorflow::strings::StrCat(

"provided list of inputs contains objects other "

"than 'EagerTensor'. Item ",

i, " is type: ", elem->ob_type->tp_name)

.c_str());

}

}

}

return input_tensor_handles;

}

// These functions are typemaps from the Python side. I did not use

// a custom type caster since the logic is slightly harder to follow. This

// converter is also only used once in `TFE_Py_ExecuteCancelable_wrapper`.

// This function actually takes a number rather than an output Tensor holder.

TFE_OutputTensorHandles InputTFE_OutputTensorHandles(

const py::handle& num_outputs) {

TFE_OutputTensorHandles output_tensor_handles;

#if PY_MAJOR_VERSION < 3

if (!PyInt_Check(num_outputs.ptr())) {

#else

if (!PyLong_Check(num_outputs.ptr())) {

#endif

PyErr_SetString(PyExc_TypeError,

"expected an integer value (size of the number of "

"outputs of the operation)");

throw py::error_already_set();

}

#if PY_MAJOR_VERSION < 3

long sz = PyInt_AsLong(num_outputs.ptr()); // NOLINT

#else

long sz = PyLong_AsLong(num_outputs.ptr()); // NOLINT

#endif

// PyLong_AsLong might throw an error if an overflow occurs.

if (PyErr_Occurred()) {

PyErr_SetString(PyExc_ValueError, tensorflow::strings::StrCat(

"Number of outputs is too big: ", sz)

.c_str());

throw py::error_already_set();

}

// We can't handle more than int32 sizes for number of outputs.

if (static_cast<long>(static_cast<int32_t>(sz)) != sz) { // NOLINT

PyErr_SetString(PyExc_ValueError, tensorflow::strings::StrCat(

"Number of outputs is too big: ", sz)

.c_str());

throw py::error_already_set();

}

if (sz > 0) {

#if PY_MAJOR_VERSION < 3

output_tensor_handles.resize(PyInt_AsLong(num_outputs.ptr()), nullptr);

#else

output_tensor_handles.resize(PyLong_AsLong(num_outputs.ptr()), nullptr);

#endif

}

return output_tensor_handles;

}

tensorflow::Device* GetMatchedDevice(py::handle& ctx, const char* device_name) {

auto* context = reinterpret_cast<tensorflow::ImmediateExecutionContext*>(

tensorflow::InputTFE_Context(ctx));

tensorflow::DeviceNameUtils::ParsedName input_device_name;

if (!tensorflow::DeviceNameUtils::ParseFullOrLocalName(device_name,

&input_device_name)) {

tensorflow::ThrowValueError(

absl::StrFormat("Failed parsing device name: '%s'. Note a valid device "

"string should at least contain a device type and a "

"device index, like \"GPU:0\".",

device_name)

.c_str());

}

std::vector<tensorflow::Device*> devices = context->ListLocalTfDevices();

tensorflow::Device* matched_device = nullptr;

for (int device_idx = 0; device_idx < devices.size(); device_idx++) {

tensorflow::Device* device = devices[device_idx];

if (tensorflow::DeviceNameUtils::AreCompatibleDevNames(

input_device_name, device->parsed_name())) {

if (matched_device != nullptr) {

tensorflow::ThrowValueError(

absl::StrFormat("Multiple devices match the provided string "

"'%s': '%s' and '%s'.",

device_name, matched_device->name(), device->name())

.c_str());

}

matched_device = device;

}

}

if (matched_device == nullptr) {

tensorflow::ThrowValueError(

absl::StrFormat("No matching devices found for '%s'", device_name)

.c_str());

}

return matched_device;

}

// Packs multiple `EagerTensor`s of the same dtype and shape into one

// `EagerTensor`.

py::object TFE_Py_PackEagerTensors_wrapper(const py::handle& context,

const py::handle& tensors) {

TFE_Context* ctx = tensorflow::InputTFE_Context(context);

TFE_InputTensorHandles handles = InputTFE_InputTensorHandles(tensors);

tensorflow::Safe_TF_StatusPtr status = tensorflow::make_safe(TF_NewStatus());

int size = handles.size();

TFE_TensorHandle* packed_handle =

TFE_CreatePackedTensorHandle(ctx, handles.data(), &size, status.get());

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

PyObject* packed_tensor =

EagerTensorFromHandle(packed_handle, /*is_packed=*/true);

return tensorflow::PyoOrThrow(packed_tensor);

}

// This function was created from fusing the typemap logic in platform/base.i.

py::object TFE_Py_ExecuteCancelable_wrapper(

const py::handle& context, const char* device_name, const char* op_name,

const py::handle& inputs, const py::handle& attrs,

tensorflow::CancellationManager* cancellation_manager,

const py::handle& num_outputs) {

TFE_Context* ctx = tensorflow::InputTFE_Context(context);

TFE_InputTensorHandles input_tensor_handles =

InputTFE_InputTensorHandles(inputs);

TFE_OutputTensorHandles output_tensor_handles =

InputTFE_OutputTensorHandles(num_outputs);

tensorflow::Safe_TF_StatusPtr status = tensorflow::make_safe(TF_NewStatus());

TFE_Py_ExecuteCancelable(ctx, device_name, op_name, &input_tensor_handles,

attrs.ptr(), tensorflow::wrap(cancellation_manager),

&output_tensor_handles, status.get());

int output_len = output_tensor_handles.size();

PyObject* output_list = PyList_New(output_len);

for (int i = 0; i < output_len; ++i) {

PyObject* output;

output = EagerTensorFromHandle(output_tensor_handles.at(i));

PyList_SetItem(output_list, i, output);

}

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

return tensorflow::PyoOrThrow(output_list);

}

static py::object TF_ListPhysicalDevices() {

std::vector<string> devices;

absl::Status s = tensorflow::DeviceFactory::ListAllPhysicalDevices(&devices);

MaybeRaiseRegisteredFromStatus(s);

PyObject* result = PyList_New(devices.size());

int i = 0;

for (auto& dev : devices) {

PyObject* dev_obj = PyBytes_FromStringAndSize(dev.data(), dev.size());

PyList_SetItem(result, i, dev_obj);

++i;

}

return tensorflow::PyoOrThrow(result);

}

static py::object TF_ListPluggablePhysicalDevices() {

std::vector<string> devices;

absl::Status s =

tensorflow::DeviceFactory::ListPluggablePhysicalDevices(&devices);

MaybeRaiseRegisteredFromStatus(s);

Safe_PyObjectPtr result(PyList_New(devices.size()));

int i = 0;

for (auto& dev : devices) {

PyObject* dev_obj = PyBytes_FromStringAndSize(dev.data(), dev.size());

PyList_SetItem(result.get(), i, dev_obj);

++i;

}

return tensorflow::PyoOrThrow(result.release());

}

static std::unordered_map<string, string> TF_GetDeviceDetails(int index) {

tensorflow::Safe_TF_StatusPtr status = tensorflow::make_safe(TF_NewStatus());

std::unordered_map<string, string> device_details;

absl::Status s =

tensorflow::DeviceFactory::GetAnyDeviceDetails(index, &device_details);

tensorflow::Set_TF_Status_from_Status(status.get(), s);

MaybeRaiseRegisteredFromTFStatus(status.get());

return device_details;

}

static py::object TFE_ClearScalarCache() {

tensorflow::TFE_TensorHandleCache::Get()->Clear();

return py::none();

}

static Device* GetDevice(EagerContext* context, const char* device_name,

const char* platform_name,

const std::vector<Device*>& devices) {

auto device_name_str = platform_name != nullptr

? absl::StrCat("/device:", platform_name, ":0")

: std::string(device_name);

DeviceNameUtils::ParsedName input_device_name;

if (!DeviceNameUtils::ParseFullOrLocalName(device_name_str,

&input_device_name)) {

ThrowValueError(absl::StrFormat("Failed parsing derived device name: '%s'",

device_name_str)

.c_str());

}

auto selected_device = absl::c_find_if(devices, [&](const Device* d) {

return DeviceNameUtils::AreCompatibleDevNames(input_device_name,

d->parsed_name());

});

if (selected_device == devices.end()) {

return nullptr;

}

return *selected_device;

}

// Returns compiler IR for a given function.

static py::bytes TFE_GetCompilerIr(py::handle& ctx,

const char* concrete_function_name,

const char* stage, const char* device_name,

py::handle& flat_arg_inputs,

py::handle& captured_inputs,

const char* platform_name) {

EagerContext* context = ContextFromInterface(

reinterpret_cast<ImmediateExecutionContext*>(InputTFE_Context(ctx)));

std::string s_stage(stage);

IrExportStage selected_stage = [&] {

if (s_stage == "stablehlo") {

return IrExportStage::STABLEHLO;

} else if (s_stage == "stablehlo_serialized") {

return IrExportStage::STABLEHLO_SERIALIZED;

} else if (s_stage == "hlo") {

return IrExportStage::HLO;

} else if (s_stage == "hlo_no_metadata") {

return IrExportStage::HLO_NO_METADATA;

} else if (s_stage == "hlo_serialized") {

return IrExportStage::HLO_SERIALIZED;

} else if (s_stage == "optimized_hlo") {

return IrExportStage::OPTIMIZED_HLO;

} else if (s_stage == "optimized_hlo_serialized") {

return IrExportStage::OPTIMIZED_HLO_SERIALIZED;

} else if (s_stage == "optimized_hlo_proto_serialized") {

return IrExportStage::OPTIMIZED_HLO_PROTO_SERIALIZED;

} else if (s_stage == "optimized_hlo_dot") {

return IrExportStage::OPTIMIZED_HLO_DOT;

} else {

ThrowValueError(

absl::StrFormat("Invalid stage selected: '%s'. Valid values are: "

"'hlo', 'hlo_serialized', 'optimized_hlo', "

"'optimized_hlo_serialized', 'optimized_hlo_dot'",

s_stage)

.c_str());

}

}();

CompilerArgSource compiler_arg_source = [&] {

if (PyList_Size(flat_arg_inputs.ptr()) == 0) {

return CompilerArgSource::CONCRETE_INPUT;

}

PyObject* elem = PyList_GetItem(flat_arg_inputs.ptr(), 0);

if (swig::IsTensorSpec(elem)) {

return CompilerArgSource::TENSOR_SPEC;

} else if (swig::IsTensor(elem)) {

return CompilerArgSource::CONCRETE_INPUT;

} else {

ThrowValueError(

tensorflow::strings::StrCat(

"Only accept tf.TensorSpec or tf.Tensor but got type ",

elem->ob_type->tp_name)

.c_str());

}

}();

Py_ssize_t flat_arg_len = PyList_Size(flat_arg_inputs.ptr());

Py_ssize_t captured_input_len = PyList_Size(captured_inputs.ptr());

std::vector<ArgShapeAndDType> flat_args;

std::vector<const TensorHandle*> captured_input_handles;

if (compiler_arg_source == CompilerArgSource::TENSOR_SPEC) {

flat_args.resize(flat_arg_len);

captured_input_handles.reserve(captured_input_len);

for (Py_ssize_t i = 0; i < flat_arg_len; ++i) {

PyObject* elem_ptr = PyList_GetItem(flat_arg_inputs.ptr(), i);

py::object elem = py::reinterpret_borrow<py::object>(elem_ptr);

py::object py_dtype = elem.attr("dtype");

py::object py_shape = elem.attr("shape");

int dtype = py::cast<int>(py_dtype.attr("_type_enum"));

auto shape = py::cast<std::vector<int64_t>>(py_shape);

flat_args[i].dtype = DataType(dtype);

flat_args[i].shape = TensorShape(shape);

}

} else if (compiler_arg_source == CompilerArgSource::CONCRETE_INPUT) {

captured_input_handles.reserve(flat_arg_len + captured_input_len);

TFE_InputTensorHandles handles =

InputTFE_InputTensorHandles(flat_arg_inputs);

for (TFE_TensorHandle* tensor_handle : handles) {

AbstractTensorHandle* abstract_tensor_handle = unwrap(tensor_handle);

captured_input_handles.push_back(

TensorHandleFromInterface(abstract_tensor_handle));

}

}

TFE_InputTensorHandles handles = InputTFE_InputTensorHandles(captured_inputs);

for (TFE_TensorHandle* tensor_handle : handles) {

AbstractTensorHandle* abstract_tensor_handle = unwrap(tensor_handle);

captured_input_handles.push_back(

TensorHandleFromInterface(abstract_tensor_handle));

}

absl::StatusOr<std::string> hlo_str;

std::vector<Device*> devices = context->local_device_mgr()->ListDevices();

Device* selected_device =

GetDevice(context, device_name, platform_name, devices);

if (selected_device != nullptr) {

hlo_str =

GetCompilerIr(selected_stage, context->pflr(), concrete_function_name,

selected_device, context, flat_args,

captured_input_handles, compiler_arg_source);

} else if (platform_name != nullptr) {

hlo_str = GetCompilerIr(

selected_stage, context->pflr(), concrete_function_name, platform_name,

context, flat_args, captured_input_handles, compiler_arg_source);

} else {

ThrowValueError(

absl::StrFormat("No matching device found for '%s'", device_name)

.c_str());

}

if (!hlo_str.ok()) {

ThrowValueError(absl::StrFormat("Failed getting HLO text: '%s'",

hlo_str.status().message())

.c_str());

}

return py::bytes(*hlo_str);

}

} // namespace tensorflow

namespace {

// Wrapper around the EagerContextThreadLocalData struct (defined in

// pywrap_tfe.h), so it can be accessed from Python.

//

// For PyObject* fields, the get_*() methods return a new reference; and the

// set_*() methods create a new reference (i.e., they do not steal a reference).

class EagerContextThreadLocalDataWrapper {

public:

explicit EagerContextThreadLocalDataWrapper(py::handle py_eager_context,

py::handle is_eager,

py::handle device_spec)

: py_eager_context_(py_eager_context.ptr()) {

tensorflow::MakeEagerContextThreadLocalData(

py_eager_context.ptr(), is_eager.ptr(), device_spec.ptr());

}

~EagerContextThreadLocalDataWrapper() {

tensorflow::DestroyEagerContextThreadLocalData(py_eager_context_);

}

bool get_is_eager() const { return GetData()->is_eager; }

void set_is_eager(bool v) { GetData()->is_eager = v; }

bool get_invoking_op_callbacks() const {

return GetData()->invoking_op_callbacks;

}

void set_invoking_op_callbacks(bool v) {

GetData()->invoking_op_callbacks = v;

}

py::object get_device_name() const {

return GetPyObject(&GetData()->device_name);

}

void set_device_name(py::handle v) {

SetPyObject(v, &GetData()->device_name);

}

py::object get_scope_name() const {

return GetPyObject(&GetData()->scope_name);

}

void set_scope_name(py::handle v) { SetPyObject(v, &GetData()->scope_name); }

py::object get_device_spec() const {

return GetPyObject(&GetData()->device_spec);

}

void set_device_spec(py::handle v) {

SetPyObject(v, &GetData()->device_spec);

}

py::object get_function_call_options() const {

return GetPyObject(&GetData()->function_call_options);

}

void set_function_call_options(py::handle v) {

SetPyObject(v, &GetData()->function_call_options);

}

py::handle get_executor() const { return GetPyObject(&GetData()->executor); }

void set_executor(py::handle v) { SetPyObject(v, &GetData()->executor); }

py::object get_op_callbacks() const {

return GetPyObject(&GetData()->op_callbacks);

}

void set_op_callbacks(py::handle v) {

SetPyObject(v, &GetData()->op_callbacks);

}

private:

tensorflow::EagerContextThreadLocalData* GetData() const {

auto* result =

tensorflow::GetEagerContextThreadLocalData(py_eager_context_);

if (!result) {

throw py::error_already_set();

}

return result;

}

py::object GetPyObject(tensorflow::Safe_PyObjectPtr* obj) const {

return pybind11::reinterpret_borrow<py::object>(obj->get());

}

void SetPyObject(py::handle value, tensorflow::Safe_PyObjectPtr* ptr) {

Py_INCREF(value.ptr());

ptr->reset(value.ptr());

}

PyObject* py_eager_context_; // not owned (borrowed reference).

};

} // namespace

// py::return_value_policy::reference is defined as specified by the

// pybind11 documents listed here.

// https://pybind11.readthedocs.io/en/stable/advanced/functions.html#return-value-policies

// This means that C++ maintains ownership of the object. We

// are only assigning this to functions that return opaque types.

PYBIND11_MODULE(_pywrap_tfe, m) {

// Numpy initialization code for array functions.

tsl::ImportNumpy();

py::class_<TFE_Executor> TFE_Executor_class(m, "TFE_Executor");

py::class_<TFE_ContextOptions> TFE_ContextOptions_class(m,

"TFE_ContextOptions");

py::class_<TFE_MonitoringCounter0> TFE_MonitoringCounter0_class(

m, "TFE_MonitoringCounter0");

py::class_<TFE_MonitoringCounter1> TFE_MonitoringCounter1_class(

m, "TFE_MonitoringCounter1");

py::class_<TFE_MonitoringCounter2> TFE_MonitoringCounter2_class(

m, "TFE_MonitoringCounter2");

py::class_<TFE_MonitoringStringGauge0> TFE_MonitoringStringGauge0_class(

m, "TFE_MonitoringStringGauge0");

py::class_<TFE_MonitoringStringGauge1> TFE_MonitoringStringGauge1_class(

m, "TFE_MonitoringStringGauge1");

py::class_<TFE_MonitoringStringGauge2> TFE_MonitoringStringGauge2_class(

m, "TFE_MonitoringStringGauge2");

py::class_<TFE_MonitoringStringGauge3> TFE_MonitoringStringGauge3_class(

m, "TFE_MonitoringStringGauge3");

py::class_<TFE_MonitoringStringGauge4> TFE_MonitoringStringGauge4_class(

m, "TFE_MonitoringStringGauge4");

py::class_<TFE_MonitoringIntGauge0> TFE_MonitoringIntGauge0_class(

m, "TFE_MonitoringIntGauge0");

py::class_<TFE_MonitoringIntGauge1> TFE_MonitoringIntGauge1_class(

m, "TFE_MonitoringIntGauge1");

py::class_<TFE_MonitoringIntGauge2> TFE_MonitoringIntGauge2_class(

m, "TFE_MonitoringIntGauge2");

py::class_<TFE_MonitoringBoolGauge0> TFE_MonitoringBoolGauge0_class(

m, "TFE_MonitoringBoolGauge0");

py::class_<TFE_MonitoringBoolGauge1> TFE_MonitoringBoolGauge1_class(

m, "TFE_MonitoringBoolGauge1");

py::class_<TFE_MonitoringBoolGauge2> TFE_MonitoringBoolGauge2_class(

m, "TFE_MonitoringBoolGauge2");

py::class_<TFE_MonitoringCounterCell> TFE_MonitoringCounterCell_class(

m, "TFE_MonitoringCounterCell");

py::class_<TFE_MonitoringIntGaugeCell> TFE_MonitoringIntGaugeCell_class(

m, "TFE_MonitoringIntGaugeCell");

py::class_<TFE_MonitoringStringGaugeCell> TFE_MonitoringStringGaugeCell_class(

m, "TFE_MonitoringStringGaugeCell");

py::class_<TFE_MonitoringBoolGaugeCell> TFE_MonitoringBoolGaugeCell_class(

m, "TFE_MonitoringBoolGaugeCell");

py::class_<TFE_MonitoringSamplerCell> TFE_MonitoringSamplerCell_class(

m, "TFE_MonitoringSamplerCell");

py::class_<TFE_MonitoringBuckets> TFE_MonitoringBuckets_class(

m, "TFE_MonitoringBuckets");

py::class_<TFE_MonitoringSampler0> TFE_MonitoringSampler0_class(

m, "TFE_MonitoringSampler0");

py::class_<TFE_MonitoringSampler1> TFE_MonitoringSampler1_class(

m, "TFE_MonitoringSampler1");

py::class_<TFE_MonitoringSampler2> TFE_MonitoringSampler2_class(

m, "TFE_MonitoringSampler2");

py::class_<tensorflow::CancellationManager> TFE_CancellationManager_class(

m, "TFE_CancellationManager");

py::class_<TF_DeviceList> TF_DeviceList_class(m, "TF_DeviceList");

py::class_<TF_Function> TF_Function_class(m, "TF_Function");

py::class_<TF_Buffer> TF_Buffer_class(m, "TF_Buffer", py::module_local());

m.def("TFE_Py_RegisterExceptionClass", [](const py::handle& e) {

return tensorflow::PyoOrThrow(TFE_Py_RegisterExceptionClass(e.ptr()));

});

m.def("TFE_Py_RegisterFallbackExceptionClass", [](const py::handle& e) {

return tensorflow::PyoOrThrow(

TFE_Py_RegisterFallbackExceptionClass(e.ptr()));

});

m.def("TFE_GetMemoryInfo", [](py::handle& ctx, const char* device_name) {

tensorflow::Device* matched_device =

tensorflow::GetMatchedDevice(ctx, device_name);

tensorflow::AllocatorAttributes attrs;

tensorflow::Allocator* allocator = matched_device->GetAllocator(attrs);

if (absl::optional<tensorflow::AllocatorStats> stats =

allocator->GetStats()) {

return std::map<std::string, int64_t>{{"current", stats->bytes_in_use},

{"peak", stats->peak_bytes_in_use}};

}

tensorflow::ThrowValueError(

absl::StrFormat("Allocator stats not available for device '%s'",

device_name)

.c_str());

});

m.def("TFE_ResetMemoryStats", [](py::handle& ctx, const char* device_name) {

tensorflow::Device* matched_device =

tensorflow::GetMatchedDevice(ctx, device_name);

tensorflow::AllocatorAttributes attrs;

tensorflow::Allocator* allocator = matched_device->GetAllocator(attrs);

if (!allocator->ClearStats()) {

tensorflow::ThrowValueError(

absl::StrFormat("Cannot reset memory stats for device '%s'",

device_name)

.c_str());

}

});

// XLA Eager Logic

m.def("TF_SetXlaEnableLazyCompilation", &TF_SetXlaEnableLazyCompilation);

m.def("TF_SetTfXlaCpuGlobalJit", &TF_SetTfXlaCpuGlobalJit);

m.def("TF_SetXlaAutoJitMode", &TF_SetXlaAutoJitMode);

m.def("TF_SetXlaConstantFoldingDisabled", &TF_SetXlaConstantFoldingDisabled);

m.def("TF_GetXlaConstantFoldingDisabled", &TF_GetXlaConstantFoldingDisabled);

m.def("TF_SetXlaMinClusterSize", &TF_SetXlaMinClusterSize);

m.def("TF_GetCompilerIr", &tensorflow::TFE_GetCompilerIr);

// MLIR Logic

m.def("TF_IsMlirBridgeEnabled", [] {

// Since python protobuf enums are integers, cast to an integer before

// returning the enum to python.

return static_cast<int32_t>(

tensorflow::GetMlirCommonFlags()->tf_mlir_enable_mlir_bridge);

});

m.def("TF_EnableMlirBridge", [](bool enabled) {

tensorflow::GetMlirCommonFlags()->tf_mlir_enable_mlir_bridge =

enabled

? tensorflow::ConfigProto::Experimental::MLIR_BRIDGE_ROLLOUT_ENABLED

: tensorflow::ConfigProto::Experimental::

MLIR_BRIDGE_ROLLOUT_DISABLED;

});

m.def("TF_EnableXlaDevices", [] {

tensorflow::GetXlaDeviceFlags()->tf_xla_enable_xla_devices = true;

});

m.def("TF_ResetJitCompilerFlags",

[] { tensorflow::ResetJitCompilerFlags(); });

// TFE_Context Logic

m.def(

"TFE_NewContext",

[](const TFE_ContextOptions* opts) {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

TFE_Context* context = TFE_NewContext(opts, status.get());

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

return tensorflow::PyoOrThrow(tensorflow::OutputTFE_Context(context));

},

py::return_value_policy::reference);

m.def("TFE_DeleteContext", [](py::handle& o) {

TFE_DeleteContext(tensorflow::InputTFE_Context(o));

});

m.def(

"TFE_ContextListDevices",

[](py::handle& o) {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

auto output = TFE_ContextListDevices(tensorflow::InputTFE_Context(o),

status.get());

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

return output;

},

py::return_value_policy::reference);

m.def(

"TFE_SetLogicalCpuDevices",

[](py::handle& ctx, int num_cpus, const char* prefix) {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

TFE_SetLogicalCpuDevices(tensorflow::InputTFE_Context(ctx), num_cpus,

prefix, status.get());

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

},

py::return_value_policy::reference);

m.def("TFE_HostAddressSpace", [](py::handle& o, TF_Buffer& buf) {

TFE_HostAddressSpace(tensorflow::InputTFE_Context(o), &buf);

});

m.def("TFE_ContextAddFunction", [](py::handle& ctx, TF_Function* func) {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

TFE_ContextAddFunction(tensorflow::InputTFE_Context(ctx), func,

status.get());

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

});

m.def("TFE_ContextAddFunctionDef",

[](py::handle& ctx, const char* serialized_function_def, size_t size) {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

TFE_ContextAddFunctionDef(tensorflow::InputTFE_Context(ctx),

serialized_function_def, size,

status.get());

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

});

m.def(

"TFE_ContextGetFunction",

[](py::handle& ctx, const char* function_name) {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

TF_Function* tf_function = TFE_ContextGetFunction(

tensorflow::InputTFE_Context(ctx), function_name, status.get());

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

return tf_function;

},

py::return_value_policy::reference);

m.def("TFE_ContextGetFunctionDef",

[](py::handle& ctx, const char* function_name, TF_Buffer& buf) {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

TFE_ContextGetFunctionDef(tensorflow::InputTFE_Context(ctx),

function_name, &buf, status.get());

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

});

// TODO(b/309152522): Remove the switch once it works on Windows.

#if !IS_OSS

pybind11_protobuf::ImportNativeProtoCasters();

m.def(

"TFE_ContextAddFunctionDefNoSerialization",

[](py::handle& ctx, tensorflow::FunctionDef function_def) {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

// Annotate eager runtime construction context to the given

// `function_def` as an attribute.

tensorflow::AttrValue value;

SetAttrValue("kEagerRuntime", &value);

(*function_def.mutable_attr())["_construction_context"] = value;

status->status = tensorflow::unwrap(tensorflow::InputTFE_Context(ctx))

->AddFunctionDef(function_def);

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

return;

},

pybind11::arg("ctx"), pybind11::arg("function_def"));

m.def("TFE_ContextGetFunctionDefNoSerialization",

[](py::handle& ctx,

const char* function_name) -> tensorflow::FunctionDef {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

const tensorflow::FunctionDef* ctx_function_def =

tensorflow::unwrap(tensorflow::InputTFE_Context(ctx))

->FindFunctionDef(function_name);

if (ctx_function_def == nullptr) {

status->status = absl::NotFoundError(absl::StrCat(

"Unable to find FunctionDef with name: ", function_name));

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

tensorflow::FunctionDef function_def;

return function_def;

}

status->status = absl::OkStatus();

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

return *ctx_function_def;

});

#else

// Defining this function to make type checker happy, as there's an entry in

// _pywrap_tfe.pyi.

m.def("TFE_ContextGetFunctionDefNoSerialization",

[](py::handle& ctx, const char* function_name) -> int {

LOG(FATAL) << "This function cannot be called.";

return -1;

});

m.def("TFE_ContextAddFunctionDefNoSerialization",

// Opensource fails whenever a protobuf is used as argument. The

// disrepency in the type is to make opensource tests pass.

[](py::handle& ctx, int function_def) {

LOG(FATAL) << "This function cannot be called.";

return -1;

});

#endif

m.def("TFE_ContextGetGraphDebugInfo",

[](py::handle& ctx, const char* function_name, TF_Buffer& buf) {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

TFE_ContextGetGraphDebugInfo(tensorflow::InputTFE_Context(ctx),

function_name, &buf, status.get());

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

});

m.def("TFE_ContextRemoveFunction", [](py::handle& ctx, const char* name) {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

TFE_ContextRemoveFunction(tensorflow::InputTFE_Context(ctx), name,

status.get());

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

});

m.def("TFE_ContextHasFunction", [](py::handle& ctx, const char* name) {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

auto output =

TFE_ContextHasFunction(tensorflow::InputTFE_Context(ctx), name);

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

return output;

});

m.def("TFE_ContextListFunctionNames", [](py::handle& ctx) {

return tensorflow::unwrap(tensorflow::InputTFE_Context(ctx))

->ListFunctionNames();

});

m.def("TFE_ContextEnableRunMetadata", [](py::handle& ctx) {

TFE_ContextEnableRunMetadata(tensorflow::InputTFE_Context(ctx));

});

m.def("TFE_ContextDisableRunMetadata", [](py::handle& ctx) {

TFE_ContextEnableRunMetadata(tensorflow::InputTFE_Context(ctx));

});

m.def("TFE_ContextEnableGraphCollection", [](py::handle& ctx) {

TFE_ContextEnableGraphCollection(tensorflow::InputTFE_Context(ctx));

});

m.def("TFE_ContextDisableGraphCollection", [](py::handle& ctx) {

TFE_ContextDisableGraphCollection(tensorflow::InputTFE_Context(ctx));

});

m.def("TFE_ContextExportRunMetadata", [](py::handle& ctx, TF_Buffer& buf) {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

TFE_ContextExportRunMetadata(tensorflow::InputTFE_Context(ctx), &buf,

status.get());

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

});

m.def("TFE_ContextClearCaches", [](py::handle& o) {

TFE_ContextClearCaches(tensorflow::InputTFE_Context(o));

});

m.def("TFE_GetContextId", [](py::handle& ctx) {

return TFE_GetContextId(tensorflow::InputTFE_Context(ctx));

});

m.def("TFE_ContextGetDevicePlacementPolicy", [](py::handle& ctx) {

return TFE_ContextGetDevicePlacementPolicy(

tensorflow::InputTFE_Context(ctx));

});

m.def("TFE_ContextSetThreadLocalDevicePlacementPolicy",

[](py::handle& ctx, TFE_ContextDevicePlacementPolicy policy) {

TFE_ContextSetThreadLocalDevicePlacementPolicy(

tensorflow::InputTFE_Context(ctx), policy);

});

m.def("TFE_ContextSetServerDef", [](py::handle& ctx, int keep_alive_secs,

py::bytes proto) {

tensorflow::Safe_TF_StatusPtr status =

tensorflow::make_safe(TF_NewStatus());

tensorflow::Safe_TF_BufferPtr buf =

tensorflow::make_safe(tensorflow::ProtoStringToTFBuffer(proto.ptr()));

TFE_ContextSetServerDef(tensorflow::InputTFE_Context(ctx), keep_alive_secs,

buf.get()->data, buf.get()->length, status.get());

tensorflow::MaybeRaiseRegisteredFromTFStatus(status.get());

});

m.def("TFE_ContextSetServerDefWithTimeoutAndRetries",