react-native-executorch 0.4.8 → 0.5.1-rc.0

package/third-party/include/c10/util/BFloat16.h

@@ -0,0 +1,125 @@
+#pragma once
+
+// Defines the bloat16 type (brain floating-point). This representation uses
+// 1 bit for the sign, 8 bits for the exponent and 7 bits for the mantissa.
+
+#include <c10/macros/Macros.h>
+#include <cmath>
+#include <cstdint>
+#include <cstring>
+#include <iosfwd>
+#include <ostream>
+
+#if defined(__CUDACC__) && !defined(USE_ROCM)
+#include <cuda_bf16.h>
+#endif
+
+#if defined(SYCL_EXT_ONEAPI_BFLOAT16_MATH_FUNCTIONS)
+#if defined(CL_SYCL_LANGUAGE_VERSION)
+#include <CL/sycl.hpp> // for SYCL 1.2.1
+#else
+#include <sycl/sycl.hpp> // for SYCL 2020
+#endif
+#include <ext/oneapi/bfloat16.hpp>
+#endif
+
+namespace c10 {
+
+namespace detail {
+inline C10_HOST_DEVICE float f32_from_bits(uint16_t src) {
+  float res = 0;
+  uint32_t tmp = src;
+  tmp <<= 16;
+
+#if defined(USE_ROCM)
+  float *tempRes;
+
+  // We should be using memcpy in order to respect the strict aliasing rule
+  // but it fails in the HIP environment.
+  tempRes = reinterpret_cast<float *>(&tmp);
+  res = *tempRes;
+#else
+  std::memcpy(&res, &tmp, sizeof(tmp));
+#endif
+
+  return res;
+}
+
+inline C10_HOST_DEVICE uint16_t bits_from_f32(float src) {
+  uint32_t res = 0;
+
+#if defined(USE_ROCM)
+  // We should be using memcpy in order to respect the strict aliasing rule
+  // but it fails in the HIP environment.
+  uint32_t *tempRes = reinterpret_cast<uint32_t *>(&src);
+  res = *tempRes;
+#else
+  std::memcpy(&res, &src, sizeof(res));
+#endif
+
+  return res >> 16;
+}
+
+inline C10_HOST_DEVICE uint16_t round_to_nearest_even(float src) {
+#if defined(USE_ROCM)
+  if (src != src) {
+#elif defined(_MSC_VER)
+  if (isnan(src)) {
+#else
+  if (std::isnan(src)) {
+#endif
+    return UINT16_C(0x7FC0);
+  } else {
+    // NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
+    union {
+      uint32_t U32; // NOLINT(facebook-hte-BadMemberName)
+      float F32;    // NOLINT(facebook-hte-BadMemberName)
+    };
+
+    F32 = src;
+    uint32_t rounding_bias = ((U32 >> 16) & 1) + UINT32_C(0x7FFF);
+    return static_cast<uint16_t>((U32 + rounding_bias) >> 16);
+  }
+}
+} // namespace detail
+
+struct alignas(2) BFloat16 {
+  uint16_t x;
+
+  // HIP wants __host__ __device__ tag, CUDA does not
+#if defined(USE_ROCM)
+  C10_HOST_DEVICE BFloat16() = default;
+#else
+  BFloat16() = default;
+#endif
+
+  struct from_bits_t {};
+  static constexpr C10_HOST_DEVICE from_bits_t from_bits() {
+    return from_bits_t();
+  }
+
+  constexpr C10_HOST_DEVICE BFloat16(unsigned short bits, from_bits_t)
+      : x(bits) {}
+  /* implicit */ inline C10_HOST_DEVICE BFloat16(float value);
+  inline C10_HOST_DEVICE operator float() const;
+
+#if defined(__CUDACC__) && !defined(USE_ROCM)
+  inline C10_HOST_DEVICE BFloat16(const __nv_bfloat16 &value);
+  explicit inline C10_HOST_DEVICE operator __nv_bfloat16() const;
+#endif
+
+#if defined(SYCL_EXT_ONEAPI_BFLOAT16_MATH_FUNCTIONS)
+  inline C10_HOST_DEVICE BFloat16(const sycl::ext::oneapi::bfloat16 &value);
+  explicit inline C10_HOST_DEVICE operator sycl::ext::oneapi::bfloat16() const;
+#endif
+};
+
+C10_API inline std::ostream &operator<<(std::ostream &out,
+                                        const BFloat16 &value) {
+  out << (float)value;
+  return out;
+}
+
+} // namespace c10
+
+#include <c10/util/BFloat16-inl.h> // IWYU pragma: keep

package/third-party/include/c10/util/Half-inl.h

@@ -0,0 +1,347 @@
+#pragma once
+
+#include <c10/macros/Macros.h>
+#include <c10/util/bit_cast.h>
+
+#include <cstring>
+#include <limits>
+
+#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(CPU_CAPABILITY_AVX2) || defined(CPU_CAPABILITY_AVX512)) &&        \
+    !defined(__APPLE__)
+#include <ATen/cpu/vec/vec_half.h>
+#endif
+
+C10_CLANG_DIAGNOSTIC_PUSH()
+#if C10_CLANG_HAS_WARNING("-Wimplicit-int-float-conversion")
+C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-int-float-conversion")
+#endif
+
+namespace c10 {
+
+#if defined(__aarch64__) && !defined(__CUDACC__)
+/// Constructors
+inline Half::Half(float16_t value) : x(detail::fp16_to_bits(value)) {}
+inline Half::operator float16_t() const { return detail::fp16_from_bits(x); }
+#else
+
+inline C10_HOST_DEVICE Half::Half(float value)
+    :
+#if defined(__CUDA_ARCH__) || defined(__HIP_DEVICE_COMPILE__)
+      x(__half_as_short(__float2half(value)))
+#elif defined(__SYCL_DEVICE_ONLY__)
+      x(c10::bit_cast<uint16_t>(sycl::half(value)))
+#elif (defined(CPU_CAPABILITY_AVX2) || defined(CPU_CAPABILITY_AVX512)) &&      \
+    !defined(__APPLE__)
+      x(at::vec::float2half_scalar(value))
+#else
+      x(detail::fp16_ieee_from_fp32_value(value))
+#endif
+{
+}
+
+/// Implicit conversions
+
+inline C10_HOST_DEVICE Half::operator float() const {
+#if defined(__CUDA_ARCH__) || defined(__HIP_DEVICE_COMPILE__)
+  return __half2float(*reinterpret_cast<const __half *>(&x));
+#elif defined(__SYCL_DEVICE_ONLY__)
+  return float(c10::bit_cast<sycl::half>(x));
+#elif (defined(CPU_CAPABILITY_AVX2) || defined(CPU_CAPABILITY_AVX512)) &&      \
+    !defined(__APPLE__)
+  return at::vec::half2float_scalar(x);
+#elif defined(__aarch64__) && !defined(__CUDACC__)
+  return detail::native_fp16_to_fp32_value(x);
+#else
+  return detail::fp16_ieee_to_fp32_value(x);
+#endif
+}
+
+#endif /* !defined(__aarch64__) || defined(__CUDACC__)                         \
+        */
+
+#if defined(__CUDACC__) || defined(__HIPCC__)
+inline C10_HOST_DEVICE Half::Half(const __half &value) {
+  x = *reinterpret_cast<const unsigned short *>(&value);
+}
+inline C10_HOST_DEVICE Half::operator __half() const {
+  return *reinterpret_cast<const __half *>(&x);
+}
+#endif
+
+#ifdef SYCL_LANGUAGE_VERSION
+inline C10_HOST_DEVICE Half::Half(const sycl::half &value) {
+  x = *reinterpret_cast<const unsigned short *>(&value);
+}
+inline C10_HOST_DEVICE Half::operator sycl::half() const {
+  return *reinterpret_cast<const sycl::half *>(&x);
+}
+#endif
+
+// CUDA intrinsics
+
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 350)) ||                      \
+    (defined(__clang__) && defined(__CUDA__))
+inline __device__ Half __ldg(const Half *ptr) {
+  return __ldg(reinterpret_cast<const __half *>(ptr));
+}
+#endif
+
+/// Arithmetic
+
+inline C10_HOST_DEVICE Half operator+(const Half &a, const Half &b) {
+  return static_cast<float>(a) + static_cast<float>(b);
+}
+
+inline C10_HOST_DEVICE Half operator-(const Half &a, const Half &b) {
+  return static_cast<float>(a) - static_cast<float>(b);
+}
+
+inline C10_HOST_DEVICE Half operator*(const Half &a, const Half &b) {
+  return static_cast<float>(a) * static_cast<float>(b);
+}
+
+inline C10_HOST_DEVICE Half operator/(const Half &a, const Half &b)
+    __ubsan_ignore_float_divide_by_zero__ {
+  return static_cast<float>(a) / static_cast<float>(b);
+}
+
+inline C10_HOST_DEVICE Half operator-(const Half &a) {
+#if (defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530) ||                        \
+    defined(__HIP_DEVICE_COMPILE__)
+  return __hneg(a);
+#elif defined(__SYCL_DEVICE_ONLY__)
+  return -c10::bit_cast<sycl::half>(a);
+#else
+  return -static_cast<float>(a);
+#endif
+}
+
+inline C10_HOST_DEVICE Half &operator+=(Half &a, const Half &b) {
+  a = a + b;
+  return a;
+}
+
+inline C10_HOST_DEVICE Half &operator-=(Half &a, const Half &b) {
+  a = a - b;
+  return a;
+}
+
+inline C10_HOST_DEVICE Half &operator*=(Half &a, const Half &b) {
+  a = a * b;
+  return a;
+}
+
+inline C10_HOST_DEVICE Half &operator/=(Half &a, const Half &b) {
+  a = a / b;
+  return a;
+}
+
+/// Arithmetic with floats
+
+inline C10_HOST_DEVICE float operator+(Half a, float b) {
+  return static_cast<float>(a) + b;
+}
+inline C10_HOST_DEVICE float operator-(Half a, float b) {
+  return static_cast<float>(a) - b;
+}
+inline C10_HOST_DEVICE float operator*(Half a, float b) {
+  return static_cast<float>(a) * b;
+}
+inline C10_HOST_DEVICE float
+operator/(Half a, float b) __ubsan_ignore_float_divide_by_zero__ {
+  return static_cast<float>(a) / b;
+}
+
+inline C10_HOST_DEVICE float operator+(float a, Half b) {
+  return a + static_cast<float>(b);
+}
+inline C10_HOST_DEVICE float operator-(float a, Half b) {
+  return a - static_cast<float>(b);
+}
+inline C10_HOST_DEVICE float operator*(float a, Half b) {
+  return a * static_cast<float>(b);
+}
+inline C10_HOST_DEVICE float
+operator/(float a, Half b) __ubsan_ignore_float_divide_by_zero__ {
+  return a / static_cast<float>(b);
+}
+
+inline C10_HOST_DEVICE float &operator+=(float &a, const Half &b) {
+  return a += static_cast<float>(b);
+}
+inline C10_HOST_DEVICE float &operator-=(float &a, const Half &b) {
+  return a -= static_cast<float>(b);
+}
+inline C10_HOST_DEVICE float &operator*=(float &a, const Half &b) {
+  return a *= static_cast<float>(b);
+}
+inline C10_HOST_DEVICE float &operator/=(float &a, const Half &b) {
+  return a /= static_cast<float>(b);
+}
+
+/// Arithmetic with doubles
+
+inline C10_HOST_DEVICE double operator+(Half a, double b) {
+  return static_cast<double>(a) + b;
+}
+inline C10_HOST_DEVICE double operator-(Half a, double b) {
+  return static_cast<double>(a) - b;
+}
+inline C10_HOST_DEVICE double operator*(Half a, double b) {
+  return static_cast<double>(a) * b;
+}
+inline C10_HOST_DEVICE double
+operator/(Half a, double b) __ubsan_ignore_float_divide_by_zero__ {
+  return static_cast<double>(a) / b;
+}
+
+inline C10_HOST_DEVICE double operator+(double a, Half b) {
+  return a + static_cast<double>(b);
+}
+inline C10_HOST_DEVICE double operator-(double a, Half b) {
+  return a - static_cast<double>(b);
+}
+inline C10_HOST_DEVICE double operator*(double a, Half b) {
+  return a * static_cast<double>(b);
+}
+inline C10_HOST_DEVICE double
+operator/(double a, Half b) __ubsan_ignore_float_divide_by_zero__ {
+  return a / static_cast<double>(b);
+}
+
+/// Arithmetic with ints
+
+inline C10_HOST_DEVICE Half operator+(Half a, int b) {
+  return a + static_cast<Half>(b);
+}
+inline C10_HOST_DEVICE Half operator-(Half a, int b) {
+  return a - static_cast<Half>(b);
+}
+inline C10_HOST_DEVICE Half operator*(Half a, int b) {
+  return a * static_cast<Half>(b);
+}
+inline C10_HOST_DEVICE Half operator/(Half a, int b) {
+  return a / static_cast<Half>(b);
+}
+
+inline C10_HOST_DEVICE Half operator+(int a, Half b) {
+  return static_cast<Half>(a) + b;
+}
+inline C10_HOST_DEVICE Half operator-(int a, Half b) {
+  return static_cast<Half>(a) - b;
+}
+inline C10_HOST_DEVICE Half operator*(int a, Half b) {
+  return static_cast<Half>(a) * b;
+}
+inline C10_HOST_DEVICE Half operator/(int a, Half b) {
+  return static_cast<Half>(a) / b;
+}
+
+//// Arithmetic with int64_t
+
+inline C10_HOST_DEVICE Half operator+(Half a, int64_t b) {
+  return a + static_cast<Half>(b);
+}
+inline C10_HOST_DEVICE Half operator-(Half a, int64_t b) {
+  return a - static_cast<Half>(b);
+}
+inline C10_HOST_DEVICE Half operator*(Half a, int64_t b) {
+  return a * static_cast<Half>(b);
+}
+inline C10_HOST_DEVICE Half operator/(Half a, int64_t b) {
+  return a / static_cast<Half>(b);
+}
+
+inline C10_HOST_DEVICE Half operator+(int64_t a, Half b) {
+  return static_cast<Half>(a) + b;
+}
+inline C10_HOST_DEVICE Half operator-(int64_t a, Half b) {
+  return static_cast<Half>(a) - b;
+}
+inline C10_HOST_DEVICE Half operator*(int64_t a, Half b) {
+  return static_cast<Half>(a) * b;
+}
+inline C10_HOST_DEVICE Half operator/(int64_t a, Half b) {
+  return static_cast<Half>(a) / b;
+}
+
+/// NOTE: we do not define comparisons directly and instead rely on the implicit
+/// conversion from c10::Half to float.
+
+} // namespace c10
+
+namespace std {
+
+template <> class numeric_limits<c10::Half> {
+public:
+  static constexpr bool is_specialized = true;
+  static constexpr bool is_signed = true;
+  static constexpr bool is_integer = false;
+  static constexpr bool is_exact = false;
+  static constexpr bool has_infinity = true;
+  static constexpr bool has_quiet_NaN = true;
+  static constexpr bool has_signaling_NaN = true;
+  static constexpr auto has_denorm = numeric_limits<float>::has_denorm;
+  static constexpr auto has_denorm_loss =
+      numeric_limits<float>::has_denorm_loss;
+  static constexpr auto round_style = numeric_limits<float>::round_style;
+  static constexpr bool is_iec559 = true;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo = false;
+  static constexpr int digits = 11;
+  static constexpr int digits10 = 3;
+  static constexpr int max_digits10 = 5;
+  static constexpr int radix = 2;
+  static constexpr int min_exponent = -13;
+  static constexpr int min_exponent10 = -4;
+  static constexpr int max_exponent = 16;
+  static constexpr int max_exponent10 = 4;
+  static constexpr auto traps = numeric_limits<float>::traps;
+  static constexpr auto tinyness_before =
+      numeric_limits<float>::tinyness_before;
+  static constexpr c10::Half min() {
+    return c10::Half(0x0400, c10::Half::from_bits());
+  }
+  static constexpr c10::Half lowest() {
+    return c10::Half(0xFBFF, c10::Half::from_bits());
+  }
+  static constexpr c10::Half max() {
+    return c10::Half(0x7BFF, c10::Half::from_bits());
+  }
+  static constexpr c10::Half epsilon() {
+    return c10::Half(0x1400, c10::Half::from_bits());
+  }
+  static constexpr c10::Half round_error() {
+    return c10::Half(0x3800, c10::Half::from_bits());
+  }
+  static constexpr c10::Half infinity() {
+    return c10::Half(0x7C00, c10::Half::from_bits());
+  }
+  static constexpr c10::Half quiet_NaN() {
+    return c10::Half(0x7E00, c10::Half::from_bits());
+  }
+  static constexpr c10::Half signaling_NaN() {
+    return c10::Half(0x7D00, c10::Half::from_bits());
+  }
+  static constexpr c10::Half denorm_min() {
+    return c10::Half(0x0001, c10::Half::from_bits());
+  }
+};
+
+} // namespace std
+
+C10_CLANG_DIAGNOSTIC_POP()