GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-04-20 08:39:37 +08:00
Code Issues Packages Projects Releases Wiki Activity

Make bfloat16(float(-nan)) produce -nan, not nan.

Browse Source
This commit is contained in:
Tim Shen 2020-09-14 17:32:56 -07:00
parent 3012e755e9
commit bb56a62582
2 changed files with 13 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
Eigen/src/Core/arch/Default/BFloat16.h
View File

@ -256,7 +256,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 operator / (const bfloat16& a, In
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __bfloat16_raw truncate_to_bfloat16(const float v) { EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __bfloat16_raw truncate_to_bfloat16(const float v) {
__bfloat16_raw output; __bfloat16_raw output;
if (Eigen::numext::isnan EIGEN_NOT_A_MACRO(v)) { if (Eigen::numext::isnan EIGEN_NOT_A_MACRO(v)) {
output.value = 0x7FC0; output.value = std::signbit(v) ? 0xFFC0: 0x7FC0;
return output; return output;
} else if (std::fabs(v) < std::numeric_limits<float>::min EIGEN_NOT_A_MACRO()) { } else if (std::fabs(v) < std::numeric_limits<float>::min EIGEN_NOT_A_MACRO()) {
// Flush denormal to +/- 0. // Flush denormal to +/- 0.
@ -293,7 +293,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __bfloat16_raw float_to_bfloat16_rtne<fals
// //
// qNaN magic: All exponent bits set + most significant bit of fraction // qNaN magic: All exponent bits set + most significant bit of fraction
// set. // set.
output.value = 0x7fc0; output.value = std::signbit(ff) ? 0xFFC0: 0x7FC0;
} else if (std::fabs(ff) < std::numeric_limits<float>::min EIGEN_NOT_A_MACRO()) { } else if (std::fabs(ff) < std::numeric_limits<float>::min EIGEN_NOT_A_MACRO()) {
// Flush denormal to +/- 0.0 // Flush denormal to +/- 0.0
output.value = std::signbit(ff) ? 0x8000 : 0; output.value = std::signbit(ff) ? 0x8000 : 0;

11
test/bfloat16_float.cpp
View File

@ -230,6 +230,17 @@ void test_conversion()
VERIFY((numext::isnan)(bfloat16(__bfloat16_raw(0xffc0)))); VERIFY((numext::isnan)(bfloat16(__bfloat16_raw(0xffc0))));
VERIFY((numext::isinf)(bfloat16(__bfloat16_raw(0x7f80)))); VERIFY((numext::isinf)(bfloat16(__bfloat16_raw(0x7f80))));
VERIFY((numext::isnan)(bfloat16(__bfloat16_raw(0x7fc0)))); VERIFY((numext::isnan)(bfloat16(__bfloat16_raw(0x7fc0))));
VERIFY_IS_EQUAL(bfloat16(BinaryToFloat(0x0, 0xff, 0x40, 0x0)).value, 0x7fc0);
VERIFY_IS_EQUAL(bfloat16(BinaryToFloat(0x1, 0xff, 0x40, 0x0)).value, 0xffc0);
VERIFY_IS_EQUAL(Eigen::bfloat16_impl::truncate_to_bfloat16(
BinaryToFloat(0x0, 0xff, 0x40, 0x0))
.value,
0x7fc0);
VERIFY_IS_EQUAL(Eigen::bfloat16_impl::truncate_to_bfloat16(
BinaryToFloat(0x1, 0xff, 0x40, 0x0))
.value,
0xffc0);
} }
void test_numtraits() void test_numtraits()