react-native-executorch 0.5.15 → 0.6.0

package/third-party/include/executorch/runtime/core/portable_type/c10/c10/util/Half.h

@@ -1,421 +1,8 @@
-#pragma once
+#include <torch/headeronly/util/Half.h>
 
-/// Defines the Half type (half-precision floating-point) including conversions
-/// to standard C types and basic arithmetic operations. Note that arithmetic
-/// operations are implemented by converting to floating point and
-/// performing the operation in float32, instead of using CUDA half intrinsics.
-/// Most uses of this type within ATen are memory bound, including the
-/// element-wise kernels, and the half intrinsics aren't efficient on all GPUs.
-/// If you are writing a compute bound kernel, you can use the CUDA half
-/// intrinsics directly on the Half type from device code.
-
-#include <c10/macros/Export.h>
-#include <c10/macros/Macros.h>
-#include <c10/util/bit_cast.h>
-#include <c10/util/floating_point_utils.h>
-#include <type_traits>
-
-#if defined(__cplusplus)
-#include <cmath>
-#elif !defined(__OPENCL_VERSION__)
-#include <math.h>
-#endif
-
-#ifdef _MSC_VER
-#include <intrin.h>
-#endif
-
-#include <cstdint>
-#include <cstring>
-#include <iosfwd>
-#include <limits>
-#include <ostream>
-
-#ifdef __CUDACC__
-#include <cuda_fp16.h>
-#endif
-
-#ifdef __HIPCC__
-#include <hip/hip_fp16.h>
-#endif
-
-#if defined(CL_SYCL_LANGUAGE_VERSION)
-#include <CL/sycl.hpp> // for SYCL 1.2.1
-#elif defined(SYCL_LANGUAGE_VERSION)
-#include <sycl/sycl.hpp> // for SYCL 2020
-#endif
-
-#if defined(__aarch64__) && !defined(__CUDACC__)
-#include <arm_neon.h>
-#endif
-
-#if defined(__GNUC__) || defined(__clang__)
-#if defined(__x86_64__) || defined(_M_X64) || defined(__i386) ||               \
-    defined(_M_IX86)
-#if defined(__F16C__) && !(defined(__CUDA_ARCH__) || defined(__CUDACC__) ||    \
-                           defined(__HIP_DEVICE_COMPILE__))
-#define C10_X86_F16 1
-#include <immintrin.h> // import conversion ops from f16cintrin.h
-#endif // defined(__F16C__) && !(defined(__CUDA_ARCH__) || defined(__CUDACC__)
-       // || defined(__HIP_DEVICE_COMPILE__))
-#endif // __x86_64__ || _M_X64 || __i386 || _M_IX86
-#endif // __GNUC__ || __clang__
-
-namespace c10 {
-
-namespace detail {
-
-/*
- * Convert a 16-bit floating-point number in IEEE half-precision format, in bit
- * representation, to a 32-bit floating-point number in IEEE single-precision
- * format, in bit representation.
- *
- * @note The implementation doesn't use any floating-point operations.
- */
-inline uint32_t fp16_ieee_to_fp32_bits(uint16_t h) {
-  /*
-   * Extend the half-precision floating-point number to 32 bits and shift to the
-   * upper part of the 32-bit word:
-   *      +---+-----+------------+-------------------+
-   *      | S |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
-   *      +---+-----+------------+-------------------+
-   * Bits  31  26-30    16-25            0-15
-   *
-   * S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0
-   * - zero bits.
-   */
-  const uint32_t w = (uint32_t)h << 16;
-  /*
-   * Extract the sign of the input number into the high bit of the 32-bit word:
-   *
-   *      +---+----------------------------------+
-   *      | S |0000000 00000000 00000000 00000000|
-   *      +---+----------------------------------+
-   * Bits  31                 0-31
-   */
-  const uint32_t sign = w & UINT32_C(0x80000000);
-  /*
-   * Extract mantissa and biased exponent of the input number into the bits 0-30
-   * of the 32-bit word:
-   *
-   *      +---+-----+------------+-------------------+
-   *      | 0 |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
-   *      +---+-----+------------+-------------------+
-   * Bits  30  27-31     17-26            0-16
-   */
-  const uint32_t nonsign = w & UINT32_C(0x7FFFFFFF);
-  /*
-   * Renorm shift is the number of bits to shift mantissa left to make the
-   * half-precision number normalized. If the initial number is normalized, some
-   * of its high 6 bits (sign == 0 and 5-bit exponent) equals one. In this case
-   * renorm_shift == 0. If the number is denormalize, renorm_shift > 0. Note
-   * that if we shift denormalized nonsign by renorm_shift, the unit bit of
-   * mantissa will shift into exponent, turning the biased exponent into 1, and
-   * making mantissa normalized (i.e. without leading 1).
-   */
-#ifdef _MSC_VER
-  unsigned long nonsign_bsr;
-  _BitScanReverse(&nonsign_bsr, (unsigned long)nonsign);
-  uint32_t renorm_shift = (uint32_t)nonsign_bsr ^ 31;
-#else
-  uint32_t renorm_shift = __builtin_clz(nonsign);
-#endif
-  renorm_shift = renorm_shift > 5 ? renorm_shift - 5 : 0;
-  /*
-   * Iff half-precision number has exponent of 15, the addition overflows
-   * it into bit 31, and the subsequent shift turns the high 9 bits
-   * into 1. Thus inf_nan_mask == 0x7F800000 if the half-precision number
-   * had exponent of 15 (i.e. was NaN or infinity) 0x00000000 otherwise
-   */
-  const int32_t inf_nan_mask =
-      ((int32_t)(nonsign + 0x04000000) >> 8) & INT32_C(0x7F800000);
-  /*
-   * Iff nonsign is 0, it overflows into 0xFFFFFFFF, turning bit 31
-   * into 1. Otherwise, bit 31 remains 0. The signed shift right by 31
-   * broadcasts bit 31 into all bits of the zero_mask. Thus zero_mask ==
-   * 0xFFFFFFFF if the half-precision number was zero (+0.0h or -0.0h)
-   * 0x00000000 otherwise
-   */
-  const int32_t zero_mask = (int32_t)(nonsign - 1) >> 31;
-  /*
-   * 1. Shift nonsign left by renorm_shift to normalize it (if the input
-   * was denormal)
-   * 2. Shift nonsign right by 3 so the exponent (5 bits originally)
-   * becomes an 8-bit field and 10-bit mantissa shifts into the 10 high
-   * bits of the 23-bit mantissa of IEEE single-precision number.
-   * 3. Add 0x70 to the exponent (starting at bit 23) to compensate the
-   * different in exponent bias (0x7F for single-precision number less 0xF
-   * for half-precision number).
-   * 4. Subtract renorm_shift from the exponent (starting at bit 23) to
-   * account for renormalization. As renorm_shift is less than 0x70, this
-   * can be combined with step 3.
-   * 5. Binary OR with inf_nan_mask to turn the exponent into 0xFF if the
-   * input was NaN or infinity.
-   * 6. Binary ANDNOT with zero_mask to turn the mantissa and exponent
-   * into zero if the input was zero.
-   * 7. Combine with the sign of the input number.
-   */
-  return sign |
-         ((((nonsign << renorm_shift >> 3) + ((0x70 - renorm_shift) << 23)) |
-           inf_nan_mask) &
-          ~zero_mask);
-}
-
-/*
- * Convert a 16-bit floating-point number in IEEE half-precision format, in bit
- * representation, to a 32-bit floating-point number in IEEE single-precision
- * format.
- *
- * @note The implementation relies on IEEE-like (no assumption about rounding
- * mode and no operations on denormals) floating-point operations and bitcasts
- * between integer and floating-point variables.
- */
-C10_HOST_DEVICE inline float fp16_ieee_to_fp32_value(uint16_t h) {
-#ifdef C10_X86_F16
-  return _cvtsh_ss(h);
-#else
-  /*
-   * Extend the half-precision floating-point number to 32 bits and shift to the
-   * upper part of the 32-bit word:
-   *      +---+-----+------------+-------------------+
-   *      | S |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
-   *      +---+-----+------------+-------------------+
-   * Bits  31  26-30    16-25            0-15
-   *
-   * S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0
-   * - zero bits.
-   */
-  const uint32_t w = (uint32_t)h << 16;
-  /*
-   * Extract the sign of the input number into the high bit of the 32-bit word:
-   *
-   *      +---+----------------------------------+
-   *      | S |0000000 00000000 00000000 00000000|
-   *      +---+----------------------------------+
-   * Bits  31                 0-31
-   */
-  const uint32_t sign = w & UINT32_C(0x80000000);
-  /*
-   * Extract mantissa and biased exponent of the input number into the high bits
-   * of the 32-bit word:
-   *
-   *      +-----+------------+---------------------+
-   *      |EEEEE|MM MMMM MMMM|0 0000 0000 0000 0000|
-   *      +-----+------------+---------------------+
-   * Bits  27-31    17-26            0-16
-   */
-  const uint32_t two_w = w + w;
-
-  /*
-   * Shift mantissa and exponent into bits 23-28 and bits 13-22 so they become
-   * mantissa and exponent of a single-precision floating-point number:
-   *
-   *       S|Exponent |          Mantissa
-   *      +-+---+-----+------------+----------------+
-   *      |0|000|EEEEE|MM MMMM MMMM|0 0000 0000 0000|
-   *      +-+---+-----+------------+----------------+
-   * Bits   | 23-31   |           0-22
-   *
-   * Next, there are some adjustments to the exponent:
-   * - The exponent needs to be corrected by the difference in exponent bias
-   * between single-precision and half-precision formats (0x7F - 0xF = 0x70)
-   * - Inf and NaN values in the inputs should become Inf and NaN values after
-   * conversion to the single-precision number. Therefore, if the biased
-   * exponent of the half-precision input was 0x1F (max possible value), the
-   * biased exponent of the single-precision output must be 0xFF (max possible
-   * value). We do this correction in two steps:
-   *   - First, we adjust the exponent by (0xFF - 0x1F) = 0xE0 (see exp_offset
-   * below) rather than by 0x70 suggested by the difference in the exponent bias
-   * (see above).
-   *   - Then we multiply the single-precision result of exponent adjustment by
-   * 2**(-112) to reverse the effect of exponent adjustment by 0xE0 less the
-   * necessary exponent adjustment by 0x70 due to difference in exponent bias.
-   *     The floating-point multiplication hardware would ensure than Inf and
-   * NaN would retain their value on at least partially IEEE754-compliant
-   * implementations.
-   *
-   * Note that the above operations do not handle denormal inputs (where biased
-   * exponent == 0). However, they also do not operate on denormal inputs, and
-   * do not produce denormal results.
-   */
-  constexpr uint32_t exp_offset = UINT32_C(0xE0) << 23;
-  // const float exp_scale = 0x1.0p-112f;
-  constexpr uint32_t scale_bits = (uint32_t)15 << 23;
-  float exp_scale_val = 0;
-#if defined(_MSC_VER) && defined(__clang__)
-  __builtin_memcpy(&exp_scale_val, &scale_bits, sizeof(exp_scale_val));
-#else
-  std::memcpy(&exp_scale_val, &scale_bits, sizeof(exp_scale_val));
-#endif
-
-  const float exp_scale = exp_scale_val;
-  const float normalized_value =
-      fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-  /*
-   * Convert denormalized half-precision inputs into single-precision results
-   * (always normalized). Zero inputs are also handled here.
-   *
-   * In a denormalized number the biased exponent is zero, and mantissa has
-   * on-zero bits. First, we shift mantissa into bits 0-9 of the 32-bit word.
-   *
-   *                  zeros           |  mantissa
-   *      +---------------------------+------------+
-   *      |0000 0000 0000 0000 0000 00|MM MMMM MMMM|
-   *      +---------------------------+------------+
-   * Bits             10-31                0-9
-   *
-   * Now, remember that denormalized half-precision numbers are represented as:
-   *    FP16 = mantissa * 2**(-24).
-   * The trick is to construct a normalized single-precision number with the
-   * same mantissa and thehalf-precision input and with an exponent which would
-   * scale the corresponding mantissa bits to 2**(-24). A normalized
-   * single-precision floating-point number is represented as: FP32 = (1 +
-   * mantissa * 2**(-23)) * 2**(exponent - 127) Therefore, when the biased
-   * exponent is 126, a unit change in the mantissa of the input denormalized
-   * half-precision number causes a change of the constructed single-precision
-   * number by 2**(-24), i.e. the same amount.
-   *
-   * The last step is to adjust the bias of the constructed single-precision
-   * number. When the input half-precision number is zero, the constructed
-   * single-precision number has the value of FP32 = 1 * 2**(126 - 127) =
-   * 2**(-1) = 0.5 Therefore, we need to subtract 0.5 from the constructed
-   * single-precision number to get the numerical equivalent of the input
-   * half-precision number.
-   */
-  constexpr uint32_t magic_mask = UINT32_C(126) << 23;
-  constexpr float magic_bias = 0.5f;
-  const float denormalized_value =
-      fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-  /*
-   * - Choose either results of conversion of input as a normalized number, or
-   * as a denormalized number, depending on the input exponent. The variable
-   * two_w contains input exponent in bits 27-31, therefore if its smaller than
-   * 2**27, the input is either a denormal number, or zero.
-   * - Combine the result of conversion of exponent and mantissa with the sign
-   * of the input number.
-   */
-  constexpr uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-  const uint32_t result =
-      sign | (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value)
-                                          : fp32_to_bits(normalized_value));
-  return fp32_from_bits(result);
-#endif // C10_X86_F16
-}
-
-/*
- * Convert a 32-bit floating-point number in IEEE single-precision format to a
- * 16-bit floating-point number in IEEE half-precision format, in bit
- * representation.
- *
- * @note The implementation relies on IEEE-like (no assumption about rounding
- * mode and no operations on denormals) floating-point operations and bitcasts
- * between integer and floating-point variables.
- */
-inline uint16_t fp16_ieee_from_fp32_value(float f) {
-#ifdef C10_X86_F16
-  return _cvtss_sh(f, _MM_FROUND_TO_NEAREST_INT);
-#else
-  // const float scale_to_inf = 0x1.0p+112f;
-  // const float scale_to_zero = 0x1.0p-110f;
-  constexpr uint32_t scale_to_inf_bits = (uint32_t)239 << 23;
-  constexpr uint32_t scale_to_zero_bits = (uint32_t)17 << 23;
-  float scale_to_inf_val = 0, scale_to_zero_val = 0;
-  std::memcpy(&scale_to_inf_val, &scale_to_inf_bits, sizeof(scale_to_inf_val));
-  std::memcpy(&scale_to_zero_val, &scale_to_zero_bits,
-              sizeof(scale_to_zero_val));
-  const float scale_to_inf = scale_to_inf_val;
-  const float scale_to_zero = scale_to_zero_val;
-
-#if defined(_MSC_VER) && _MSC_VER == 1916
-  float base = ((signbit(f) != 0 ? -f : f) * scale_to_inf) * scale_to_zero;
-#else
-  float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-#endif
-
-  const uint32_t w = fp32_to_bits(f);
-  const uint32_t shl1_w = w + w;
-  const uint32_t sign = w & UINT32_C(0x80000000);
-  uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-  if (bias < UINT32_C(0x71000000)) {
-    bias = UINT32_C(0x71000000);
-  }
-
-  base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-  const uint32_t bits = fp32_to_bits(base);
-  const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-  const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-  const uint32_t nonsign = exp_bits + mantissa_bits;
-  return static_cast<uint16_t>(
-      (sign >> 16) |
-      (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign));
-#endif // C10_X86_F16
-}
-
-#ifdef C10_X86_F16
-#undef C10_X86_F16
-#endif // C10_X86_F16
-
-#if defined(__aarch64__) && !defined(__CUDACC__)
-inline float16_t fp16_from_bits(uint16_t h) {
-  return c10::bit_cast<float16_t>(h);
-}
-
-inline uint16_t fp16_to_bits(float16_t f) { return c10::bit_cast<uint16_t>(f); }
-
-// According to https://godbolt.org/z/frExdbsWG it would translate to single
-// fcvt s0, h0
-inline float native_fp16_to_fp32_value(uint16_t h) {
-  return static_cast<float>(fp16_from_bits(h));
-}
-
-inline uint16_t native_fp16_from_fp32_value(float f) {
-  return fp16_to_bits(static_cast<float16_t>(f));
-}
-#endif
-
-} // namespace detail
-
-struct alignas(2) Half {
-  unsigned short x;
-
-  struct from_bits_t {};
-  C10_HOST_DEVICE static constexpr from_bits_t from_bits() {
-    return from_bits_t();
-  }
-
-  // HIP wants __host__ __device__ tag, CUDA does not
-#if defined(USE_ROCM)
-  C10_HOST_DEVICE Half() = default;
-#else
-  Half() = default;
-#endif
-
-  constexpr C10_HOST_DEVICE Half(unsigned short bits, from_bits_t) : x(bits) {}
-#if defined(__aarch64__) && !defined(__CUDACC__)
-  inline Half(float16_t value);
-  inline operator float16_t() const;
-#else
-  inline C10_HOST_DEVICE Half(float value);
-  inline C10_HOST_DEVICE operator float() const;
+// need to keep the following for BC because the APIs in here were exposed
+// before migrating Half to torch/headeronly
+#if (defined(CPU_CAPABILITY_AVX2) || defined(CPU_CAPABILITY_AVX512)) &&        \
+    !defined(__APPLE__)
+#include <ATen/cpu/vec/vec_half.h>
 #endif
-
-#if defined(__CUDACC__) || defined(__HIPCC__)
-  inline C10_HOST_DEVICE Half(const __half &value);
-  inline C10_HOST_DEVICE operator __half() const;
-#endif
-#ifdef SYCL_LANGUAGE_VERSION
-  inline C10_HOST_DEVICE Half(const sycl::half &value);
-  inline C10_HOST_DEVICE operator sycl::half() const;
-#endif
-};
-
-C10_API inline std::ostream &operator<<(std::ostream &out, const Half &value) {
-  out << (float)value;
-  return out;
-}
-
-} // namespace c10
-
-#include <c10/util/Half-inl.h> // IWYU pragma: keep

package/third-party/include/executorch/runtime/core/portable_type/c10/c10/util/TypeSafeSignMath.h

@@ -1,133 +1 @@
-#pragma once
-
-#include <c10/macros/Macros.h>
-#include <limits>
-#include <type_traits>
-
-C10_CLANG_DIAGNOSTIC_PUSH()
-#if C10_CLANG_HAS_WARNING("-Wstring-conversion")
-C10_CLANG_DIAGNOSTIC_IGNORE("-Wstring-conversion")
-#endif
-#if C10_CLANG_HAS_WARNING("-Wimplicit-int-float-conversion")
-C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-int-float-conversion")
-#endif
-
-namespace c10 {
-
-/// Returns false since we cannot have x < 0 if x is unsigned.
-template <typename T>
-inline constexpr bool is_negative(const T & /*x*/,
-                                  std::true_type /*is_unsigned*/) {
-  return false;
-}
-
-/// Returns true if a signed variable x < 0
-template <typename T>
-inline constexpr bool is_negative(const T &x, std::false_type /*is_unsigned*/) {
-  return x < T(0);
-}
-
-/// Returns true if x < 0
-/// NOTE: Will fail on an unsigned custom type
-///       For the most part it's possible to fix this if
-///       the custom type has a constexpr constructor.
-///       However, notably, c10::Half does not :-(
-template <typename T> inline constexpr bool is_negative(const T &x) {
-  return is_negative(x, std::is_unsigned<T>());
-}
-
-/// Returns the sign of an unsigned variable x as 0, 1
-template <typename T>
-inline constexpr int signum(const T &x, std::true_type /*is_unsigned*/) {
-  return T(0) < x;
-}
-
-/// Returns the sign of a signed variable x as -1, 0, 1
-template <typename T>
-inline constexpr int signum(const T &x, std::false_type /*is_unsigned*/) {
-  return (T(0) < x) - (x < T(0));
-}
-
-/// Returns the sign of x as -1, 0, 1
-/// NOTE: Will fail on an unsigned custom type
-///       For the most part it's possible to fix this if
-///       the custom type has a constexpr constructor.
-///       However, notably, c10::Half does not :-(
-template <typename T> inline constexpr int signum(const T &x) {
-  return signum(x, std::is_unsigned<T>());
-}
-
-/// Returns true if a and b are not both negative
-template <typename T, typename U>
-inline constexpr bool signs_differ(const T &a, const U &b) {
-  return is_negative(a) != is_negative(b);
-}
-
-// Suppress sign compare warning when compiling with GCC
-// as later does not account for short-circuit rule before
-// raising the warning, see https://godbolt.org/z/Tr3Msnz99
-#ifdef __GNUC__
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wsign-compare"
-#endif
-
-/// Returns true if x is greater than the greatest value of the type Limit
-template <typename Limit, typename T>
-inline constexpr bool greater_than_max(const T &x) {
-  constexpr bool can_overflow =
-      std::numeric_limits<T>::digits > std::numeric_limits<Limit>::digits;
-  return can_overflow && x > std::numeric_limits<Limit>::max();
-}
-
-#ifdef __GNUC__
-#pragma GCC diagnostic pop
-#endif
-
-/// Returns true if x < lowest(Limit). Standard comparison
-template <typename Limit, typename T>
-inline constexpr bool less_than_lowest(const T &x,
-                                       std::false_type /*limit_is_unsigned*/,
-                                       std::false_type /*x_is_unsigned*/) {
-  return x < std::numeric_limits<Limit>::lowest();
-}
-
-/// Returns false since all the limit is signed and therefore includes
-/// negative values but x cannot be negative because it is unsigned
-template <typename Limit, typename T>
-inline constexpr bool less_than_lowest(const T & /*x*/,
-                                       std::false_type /*limit_is_unsigned*/,
-                                       std::true_type /*x_is_unsigned*/) {
-  return false;
-}
-
-/// Returns true if x < 0, where 0 is constructed from T.
-/// Limit is not signed, so its lower value is zero
-template <typename Limit, typename T>
-inline constexpr bool less_than_lowest(const T &x,
-                                       std::true_type /*limit_is_unsigned*/,
-                                       std::false_type /*x_is_unsigned*/) {
-  return x < T(0);
-}
-
-/// Returns false sign both types are unsigned
-template <typename Limit, typename T>
-inline constexpr bool less_than_lowest(const T & /*x*/,
-                                       std::true_type /*limit_is_unsigned*/,
-                                       std::true_type /*x_is_unsigned*/) {
-  return false;
-}
-
-/// Returns true if x is less than the lowest value of type T
-/// NOTE: Will fail on an unsigned custom type
-///       For the most part it's possible to fix this if
-///       the custom type has a constexpr constructor.
-///       However, notably, c10::Half does not :
-template <typename Limit, typename T>
-inline constexpr bool less_than_lowest(const T &x) {
-  return less_than_lowest<Limit>(x, std::is_unsigned<Limit>(),
-                                 std::is_unsigned<T>());
-}
-
-} // namespace c10
-
-C10_CLANG_DIAGNOSTIC_POP()
+#include <torch/headeronly/util/TypeSafeSignMath.h>

package/third-party/include/executorch/runtime/core/portable_type/c10/c10/util/bit_cast.h

@@ -1,43 +1 @@
-#pragma once
-
-#include <cstring>
-#include <type_traits>
-
-#if __has_include(<bit>) && (defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L)
-#include <bit>
-#define C10_HAVE_STD_BIT_CAST 1
-#else
-#define C10_HAVE_STD_BIT_CAST 0
-#endif // __has_include(<bit>) && (__cplusplus >= 202002L ||
-       // (defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L))
-
-namespace c10 {
-
-#if C10_HAVE_STD_BIT_CAST
-using std::bit_cast;
-#else
-// Implementations of std::bit_cast() from C++ 20.
-//
-// This is a less sketchy version of reinterpret_cast.
-//
-// See https://en.cppreference.com/w/cpp/numeric/bit_cast for more
-// information as well as the source of our implementations.
-template <class To, class From>
-std::enable_if_t<sizeof(To) == sizeof(From) &&
-                     std::is_trivially_copyable_v<From> &&
-                     std::is_trivially_copyable_v<To>,
-                 To>
-// constexpr support needs compiler magic
-bit_cast(const From &src) noexcept {
-  static_assert(std::is_trivially_constructible_v<To>,
-                "This implementation additionally requires "
-                "destination type to be trivially constructible");
-
-  To dst;
-  std::memcpy(&dst, &src, sizeof(To));
-  return dst;
-}
-#endif // C10_HAVE_STD_BIT_CAST
-#undef C10_HAVE_STD_BIT_CAST
-
-} // namespace c10
+#include <torch/headeronly/util/bit_cast.h>