fill_non_native_type Class — pytorch Architecture

Source Code

aten/src/ATen/native/cpu/FillKernel.cpp lines 16–26

template <typename scalar_t>
void fill_non_native_type(TensorIterator& iter, const Scalar& value_scalar) {
  auto value = value_scalar.to<scalar_t>().x;
  using H = typename std::make_signed_t<decltype(value)>;  // Signed type has more acceleration
  // Reserve the representation of value. static_cast<H>(value) is implementation defined.
  H val = *reinterpret_cast<H*>(std::addressof(value));
  cpu_kernel_vec</*check_dynamic_cast=*/false>(
      iter,
      [val]() -> H { return val; },
      [val]() { return Vectorized<H>(val); });
}