react-native-executorch 0.5.5 → 0.5.7

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

@@ -14,14 +14,11 @@
 #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
+#elif defined(SYCL_LANGUAGE_VERSION)
 #include <sycl/sycl.hpp> // for SYCL 2020
 #endif
-#include <ext/oneapi/bfloat16.hpp>
-#endif
 
 namespace c10 {
 
@@ -31,7 +28,7 @@ inline C10_HOST_DEVICE float f32_from_bits(uint16_t src) {
   uint32_t tmp = src;
   tmp <<= 16;
 
-#if defined(USE_ROCM)
+#if defined(USE_ROCM) && defined(__HIPCC__)
   float *tempRes;
 
   // We should be using memcpy in order to respect the strict aliasing rule
@@ -48,7 +45,7 @@ inline C10_HOST_DEVICE float f32_from_bits(uint16_t src) {
 inline C10_HOST_DEVICE uint16_t bits_from_f32(float src) {
   uint32_t res = 0;
 
-#if defined(USE_ROCM)
+#if defined(USE_ROCM) && defined(__HIPCC__)
   // 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);
@@ -61,7 +58,7 @@ inline C10_HOST_DEVICE uint16_t bits_from_f32(float src) {
 }
 
 inline C10_HOST_DEVICE uint16_t round_to_nearest_even(float src) {
-#if defined(USE_ROCM)
+#if defined(USE_ROCM) && defined(__HIPCC__)
   if (src != src) {
 #elif defined(_MSC_VER)
   if (isnan(src)) {
@@ -87,7 +84,7 @@ struct alignas(2) BFloat16 {
   uint16_t x;
 
   // HIP wants __host__ __device__ tag, CUDA does not
-#if defined(USE_ROCM)
+#if defined(USE_ROCM) && defined(__HIPCC__)
   C10_HOST_DEVICE BFloat16() = default;
 #else
   BFloat16() = default;

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

@@ -242,7 +242,12 @@ C10_HOST_DEVICE inline float fp16_ieee_to_fp32_value(uint16_t h) {
   // 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;

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

@@ -3,7 +3,7 @@
 #include <cstring>
 #include <type_traits>
 
-#if __has_include(<bit>) && (__cplusplus >= 202002L || (defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L))
+#if __has_include(<bit>) && (defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L)
 #include <bit>
 #define C10_HAVE_STD_BIT_CAST 1
 #else