torch/csrc/utils.cpp

#include <fmt/core.h>
#include <torch/csrc/DynamicTypes.h>
#include <torch/csrc/THP.h>
#include <torch/csrc/autograd/variable.h>
#include <torch/csrc/python_headers.h>
#include <torch/csrc/utils/invalid_arguments.h>
#include <torch/csrc/utils/python_strings.h>
#include <torch/csrc/utils/python_symnode.h>
#include <torch/csrc/utils/python_tuples.h>

#include <torch/csrc/Export.h>

#include <algorithm>
#include <cstdarg>
#include <cstring>
#include <iterator>
#include <sstream>
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>

int THPUtils_getCallable(PyObject* arg, PyObject** result) {
  if (!PyCallable_Check(arg))
    return 0;
  *result = arg;
  return 1;
}

bool THPUtils_checkIndex(PyObject* obj) {
  if (PyBool_Check(obj)) {
    return false;
  }
  if (THPUtils_checkLong(obj)) {
    return true;
  }
  // Avoid poking __index__ early as that will immediately cause a guard
  if (torch::is_symint(py::handle(obj))) {
    return true;
  }
  torch::jit::tracer::NoWarn no_warn_guard;
  auto index = THPObjectPtr(PyNumber_Index(obj));
  if (!index) {
    PyErr_Clear();
    return false;
  }
  return true;
}

std::vector<int64_t> THPUtils_unpackLongs(PyObject* arg) {
  bool tuple = PyTuple_Check(arg);
  bool list = PyList_Check(arg);
  if (tuple || list) {
    // NOLINTNEXTLINE(bugprone-branch-clone)
    const auto nDim = tuple ? PyTuple_GET_SIZE(arg) : PyList_GET_SIZE(arg);
    std::vector<int64_t> sizes(nDim);
    for (int i = 0; i != nDim; ++i) {
      PyObject* item =
          tuple ? PyTuple_GET_ITEM(arg, i) : PyList_GET_ITEM(arg, i);
      if (!THPUtils_checkLong(item)) {
        std::ostringstream oss;
        oss << "expected int at position " << i
            << ", but got: " << THPUtils_typename(item);
        throw std::runtime_error(oss.str());
      }
      sizes[i] = THPUtils_unpackLong(item);
    }
    return sizes;
  }
  throw std::runtime_error("Expected tuple or list");
}

bool THPUtils_checkIntTuple(PyObject* arg) {
  if (!PyTuple_Check(arg)) {
    return false;
  }
  for (Py_ssize_t i = 0; i < PyTuple_GET_SIZE(arg); ++i) {
    if (!THPUtils_checkLong(PyTuple_GET_ITEM(arg, i))) {
      return false;
    }
  }
  return true;
}

std::vector<int> THPUtils_unpackIntTuple(PyObject* arg) {
  if (!THPUtils_checkIntTuple(arg)) {
    throw std::runtime_error("Couldn't unpack int tuple");
  }
  std::vector<int> values(PyTuple_GET_SIZE(arg));
  for (Py_ssize_t i = 0; i < PyTuple_GET_SIZE(arg); ++i) {
    values[i] = (int)THPUtils_unpackLong(PyTuple_GET_ITEM(arg, i));
  }
  return values;
}

void THPUtils_setError(const char* format, ...) {
  static const size_t ERROR_BUFFER_SIZE = 1000;
  // NOLINTNEXTLINE(cppcoreguidelines-avoid-c-arrays,modernize-avoid-c-arrays)
  char buffer[ERROR_BUFFER_SIZE];
  va_list fmt_args;

  va_start(fmt_args, format);
  vsnprintf(buffer, ERROR_BUFFER_SIZE, format, fmt_args);
  va_end(fmt_args);
  PyErr_SetString(PyExc_RuntimeError, buffer);
}

void THPUtils_addPyMethodDefs(
    std::vector<PyMethodDef>& vector,
    PyMethodDef* methods) {
  if (!vector.empty()) {
    // remove nullptr terminator
    vector.pop_back();
  }
  while (true) {
    vector.push_back(*methods);
    if (!methods->ml_name) {
      break;
    }
    methods++;
  }
}

static const char* classOrTypename(PyObject* obj) {
  if (PyType_Check(obj)) {
    return ((PyTypeObject*)obj)->tp_name;
  }
  return Py_TYPE(obj)->tp_name;
}

PyObject* THPUtils_dispatchStateless(
    PyObject* tensor,
    const char* name,
    PyObject* args,
    PyObject* kwargs) {
  THPObjectPtr methods(
      PyObject_GetAttrString(tensor, THP_STATELESS_ATTRIBUTE_NAME));
  if (!methods) {
    return PyErr_Format(
        PyExc_TypeError,
        "Type %s doesn't implement stateless methods",
        classOrTypename(tensor));
  }
  THPObjectPtr method(PyObject_GetAttrString(methods, name));
  if (!method) {
    return PyErr_Format(
        PyExc_TypeError,
        "Type %s doesn't implement stateless method %s",
        classOrTypename(tensor),
        name);
  }
  return PyObject_Call(method.get(), args, kwargs);
}

void THPUtils_invalidArguments(
    PyObject* given_args,
    PyObject* given_kwargs,
    const char* function_name,
    size_t num_options,
    ...) {
  std::vector<std::string> option_strings;
  va_list option_list;
  va_start(option_list, num_options);
  std::generate_n(
      std::back_inserter(option_strings), num_options, [&option_list] {
        return va_arg(option_list, const char*);
      });
  va_end(option_list);

  PyErr_SetString(
      PyExc_TypeError,
      torch::format_invalid_args(
          given_args, given_kwargs, function_name, option_strings)
          .c_str());
}

template <>
void THPPointer<THPGenerator>::free() {
  if (ptr)
    Py_DECREF(ptr);
}

template class THPPointer<THPGenerator>;

static bool backCompatBroadcastWarn = false;

void setBackCompatBroadcastWarn(bool warn) {
  backCompatBroadcastWarn = warn;
}

bool getBackCompatBroadcastWarn() {
  return backCompatBroadcastWarn;
}

static bool backCompatKeepdimWarn = false;

void setBackCompatKeepdimWarn(bool warn) {
  backCompatKeepdimWarn = warn;
}

bool getBackCompatKeepdimWarn() {
  return backCompatKeepdimWarn;
}

bool maybeThrowBackCompatKeepdimWarn(char* func) {
  if (getBackCompatKeepdimWarn()) {
    std::ostringstream ss;
    ss << "backwards compatibility: call to \"" << func
       << "\" uses default value for keepdim which has changed default to False.  Consider passing as kwarg.",
        PyErr_WarnEx(PyExc_UserWarning, ss.str().c_str(), 1);
  }
  return true;
}

template <>
void THPPointer<THPStorage>::free() {
  if (ptr)
    Py_DECREF(ptr);
}

void storage_fill(const at::Storage& self, uint8_t value) {
  auto options = c10::TensorOptions().device(self.device()).dtype(at::kByte);
  auto self_t = at::empty({0}, options).set_(self);
  self_t.fill_(value);
}

void storage_set(const at::Storage& self, ptrdiff_t idx, uint8_t value) {
  TORCH_CHECK(
      (idx >= 0) && (idx < static_cast<ptrdiff_t>(self.nbytes())),
      "out of bounds");
  auto options = c10::TensorOptions().device(self.device()).dtype(at::kByte);
  auto self_t = at::empty({0}, options).set_(self);
  self_t[idx].fill_(value);
}

uint8_t storage_get(const at::Storage& self, ptrdiff_t idx) {
  TORCH_CHECK(
      (idx >= 0) && (idx < static_cast<ptrdiff_t>(self.nbytes())),
      "out of bounds");
  auto options = c10::TensorOptions().device(self.device()).dtype(at::kByte);
  auto self_t = at::empty({0}, options).set_(self);
  return self_t[idx].item<uint8_t>();
}

template class THPPointer<THPStorage>;

namespace torch::gdb {
/* ~~~ misc debugging utilities ~~~
 *
 * torch::gdb::* functions are NOT meant to be called by general pytorch code,
 * but only from within a gdb session. As such, utils.h does not contain any
 * declaration for those.
 */

// This is a helper needed by the torch-tensor-repr gdb command.
// Return an human-readable representation of the given Tensor. The resulting
// string is stored into a malloc()ed buffer. The caller is responsible to
// free() it. We use malloc() instead of new[] because it's much easier to
// call free than delete[] from withing gdb.
// Currently the code for computing the repr of a tensor is written in Python,
// so we need to wrap the Tensor into a Python object first.
char* tensor_repr(at::Tensor tensor) {
  PyGILState_STATE gil = PyGILState_Ensure();
  PyObject* pytensor = nullptr;
  PyObject* repr = nullptr;
  Py_ssize_t bufsize = 0;
  const char* buf = nullptr;
  char* result = nullptr;

  // NB: It's important not to move the tensor into THPVariable_Wrap,
  // because this function is only called from our gdb macros, and
  // we want to avoid accidentally moving out the tensor.  In principle,
  // the Tensor signature above should induce a copy, but we've
  // observed that sometimes gdb passes the outer Tensor address exactly as is
  // into this function.
  // See https://github.com/pytorch/pytorch/issues/134762
  pytensor = THPVariable_Wrap(tensor);
  if (!pytensor)
    // NOLINTNEXTLINE(cppcoreguidelines-avoid-goto,hicpp-avoid-goto)
    goto error;
  repr = PyObject_Repr(pytensor);
  if (!repr)
    // NOLINTNEXTLINE(cppcoreguidelines-avoid-goto,hicpp-avoid-goto)
    goto error;
  buf = PyUnicode_AsUTF8AndSize(repr, &bufsize);
  if (!buf)
    // NOLINTNEXTLINE(cppcoreguidelines-avoid-goto,hicpp-avoid-goto)
    goto error;
  // account for the trailing \0
  // NOLINTNEXTLINE(cppcoreguidelines-no-malloc)
  result = static_cast<char*>(malloc(bufsize + 1));
  if (!result) {
    fmt::print(stderr, "cannot allocate memory for the result\n");
    // NOLINTNEXTLINE(cppcoreguidelines-avoid-goto,hicpp-avoid-goto)
    goto error;
  }
  std::strncpy(result, buf, bufsize);
  result[bufsize] = '\0';
  Py_XDECREF(pytensor);
  Py_XDECREF(repr);
  PyGILState_Release(gil);
  return result;

error:
  fprintf(stderr, "torch::gdb::tensor_repr: unexpected error\n");
  if (PyErr_Occurred())
    PyErr_Print();
  Py_XDECREF(pytensor);
  Py_XDECREF(repr);
  // NOLINTNEXTLINE(cppcoreguidelines-no-malloc)
  free(result);
  PyGILState_Release(gil);
  return nullptr;
}

std::string int_array_ref_string(at::IntArrayRef sizes) {
  std::stringstream ss;
  ss << sizes;
  return ss.str();
}

std::string dispatch_keyset_string(c10::DispatchKeySet keyset) {
  std::stringstream ss;
  ss << keyset;
  return ss.str();
}

} // namespace torch::gdb

namespace pybind11::detail {

bool type_caster<at::Tensor>::load(handle src, bool) {
  PyObject* obj = src.ptr();
  if (THPVariable_Check(obj)) {
    value = THPVariable_Unpack(obj);
    return true;
  }
  return false;
}

handle type_caster<at::Tensor>::cast(
    const at::Tensor& src,
    return_value_policy /* policy */,
    handle /* parent */) {
  return handle(THPVariable_Wrap(src));
}

bool type_caster<at::IntArrayRef>::load(handle src, bool) {
  PyObject* source = src.ptr();
  auto tuple = PyTuple_Check(source);
  if (tuple || PyList_Check(source)) {
    // NOLINTNEXTLINE(bugprone-branch-clone)
    const auto size =
        tuple ? PyTuple_GET_SIZE(source) : PyList_GET_SIZE(source);
    v_value.resize(size);
    for (const auto idx : c10::irange(size)) {
      PyObject* obj =
          tuple ? PyTuple_GET_ITEM(source, idx) : PyList_GET_ITEM(source, idx);
      if (THPVariable_Check(obj)) {
        v_value[idx] = THPVariable_Unpack(obj).item<int64_t>();
      } else if (PyLong_Check(obj)) {
        // use THPUtils_unpackLong after it is safe to include
        // python_numbers.h
        v_value[idx] = THPUtils_unpackLong(obj);
      } else {
        return false;
      }
    }
    value = v_value;
    return true;
  }
  return false;
}
handle type_caster<at::IntArrayRef>::cast(
    at::IntArrayRef src,
    return_value_policy /* policy */,
    handle /* parent */) {
  return handle(THPUtils_packInt64Array(src.size(), src.data()));
}

bool type_caster<at::SymIntArrayRef>::load(handle src, bool) {
  PyObject* source = src.ptr();

  auto tuple = PyTuple_Check(source);
  if (tuple || PyList_Check(source)) {
    // NOLINTNEXTLINE(bugprone-branch-clone)
    const auto size =
        tuple ? PyTuple_GET_SIZE(source) : PyList_GET_SIZE(source);
    v_value.resize(size);
    for (const auto idx : c10::irange(size)) {
      PyObject* obj =
          tuple ? PyTuple_GET_ITEM(source, idx) : PyList_GET_ITEM(source, idx);

      if (THPVariable_Check(obj)) {
        // TODO: this is for consistency with IntArrayRef but arguably
        // we shouldn't really allow this on pybind11 casters
        v_value[idx] = THPVariable_Unpack(obj).item<int64_t>();
      } else if (torch::is_symint(py::handle(obj))) {
        v_value[idx] = py::handle(obj).cast<c10::SymInt>();
      } else if (PyLong_Check(obj)) {
        v_value[idx] = c10::SymInt(THPUtils_unpackIndex(obj));
      } else {
        return false;
      }
    }
    value = v_value;
    return true;
  }
  return false;
}
handle type_caster<at::SymIntArrayRef>::cast(
    at::SymIntArrayRef src,
    return_value_policy /* policy */,
    handle /* parent */) {
  py::list t(src.size());
  for (const auto i : c10::irange(src.size())) {
    t[i] = py::cast(src[i]);
  }
  return t.release();
}

bool type_caster<at::ArrayRef<c10::SymNode>>::load(handle src, bool) {
  TORCH_INTERNAL_ASSERT(0, "NYI");
}
handle type_caster<at::ArrayRef<c10::SymNode>>::cast(
    at::ArrayRef<c10::SymNode> src,
    return_value_policy /* policy */,
    handle /* parent */) {
  py::list t(src.size());
  for (const auto i : c10::irange(src.size())) {
    // TODO: this is terrible but I don't know how to override when
    // the SymNode is also explicitly cast by py::cast
    auto* py_node = dynamic_cast<torch::impl::PythonSymNodeImpl*>(src[i].get());
    if (py_node) {
      // Return the Python directly (unwrap)
      t[i] = py_node->getPyObj();
    } else {
      t[i] = py::cast(src[i]);
    }
  }
  return t.release();
}

} // namespace pybind11::detail