-// Copyright (C) 2018 Intel Corporation
+// Copyright (C) 2018-2019 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
// small helper function to represent uint32_t value as float32
inline float asfloat(uint32_t v) {
- return *reinterpret_cast<float *>(&v);
+ // Both type-punning casts and unions are UB per C++ spec
+ // But compilers usually only break code with casts
+ union {
+ float f;
+ uint32_t i;
+ };
+ i = v;
+ return f;
}
// Function to convert F32 into F16
INFERENCE_ENGINE_API_CPP(float) f16tof32(ie_fp16 x) {
// this is storage for output result
- uint32_t u = x;
+ uint32_t u = static_cast<uint32_t>(x);
// get sign in 32bit format
uint32_t s = ((u & 0x8000) << 16);
u <<= (23 - 10);
u |= EXP_MASK_F32;
u |= s;
- } else if ((x & EXP_MASK_F16) == 0) { // check for zero and denormals. both are converted to zero
- u = s;
+ } else if ((u & EXP_MASK_F16) == 0) { // check for zero and denormals.
+ uint16_t h_sig = (u & 0x03ffu);
+ if (h_sig == 0) {
+ /* Signed zero */
+ u = s;
+ } else {
+ /* Subnormal */
+ uint16_t h_exp = (u & EXP_MASK_F16);
+ h_sig <<= 1;
+ while ((h_sig & 0x0400u) == 0) {
+ h_sig <<= 1;
+ h_exp++;
+ }
+ uint32_t f_exp = (static_cast<uint32_t>(127 - 15 - h_exp)) << 23;
+ uint32_t f_sig = (static_cast<uint32_t>(h_sig & 0x03ffu)) << 13;
+ u = s + f_exp + f_sig;
+ }
} else {
// abs
u = (u & 0x7FFF);
}
// finaly represent result as float and return
- return *reinterpret_cast<float *>(&u);
+ return asfloat(u);
}
// This function convert f32 to f16 with rounding to nearest value to minimize error