Skip to content

[SYCL] Add bfloat16 generic implementation for fmax, fmin #7732

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 4 commits into from
Dec 15, 2022
Merged

[SYCL] Add bfloat16 generic implementation for fmax, fmin #7732

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
f7d6d06
[SYCL] Add bfloat16 generic implementation for fmax, fmin
jinge90 Dec 12, 2022
293fb42
1.Fix return type of isnan(bfloat16)
jinge90 Dec 12, 2022
ea8be62
Apply suggestions from code review
bader Dec 15, 2022
1bc646b
Fix formatting.
bader Dec 15, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
78 changes: 51 additions & 27 deletions sycl/include/sycl/ext/oneapi/experimental/bfloat16_math.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,14 @@ uint32_t to_uint32_t(sycl::marray<bfloat16, N> x, size_t start) {
}
} // namespace detail

// According to bfloat16 format, NAN value's exponent field is 0xFF and
// significand has non-zero bits.
template <typename T>
std::enable_if_t<std::is_same<T, bfloat16>::value, bool> isnan(T x) {
oneapi::detail::Bfloat16StorageT XBits = oneapi::detail::bfloat16ToBits(x);
return (((XBits & 0x7F80) == 0x7F80) && (XBits & 0x7F)) ? true : false;
}

template <typename T>
std::enable_if_t<std::is_same<T, bfloat16>::value, T> fabs(T x) {
#if defined(__SYCL_DEVICE_ONLY__) && defined(__NVPTX__)
Expand Down Expand Up @@ -74,20 +82,31 @@ std::enable_if_t<std::is_same<T, bfloat16>::value, T> fmin(T x, T y) {
oneapi::detail::Bfloat16StorageT YBits = oneapi::detail::bfloat16ToBits(y);
return oneapi::detail::bitsToBfloat16(__clc_fmin(XBits, YBits));
#else
std::ignore = x;
std::ignore = y;
throw runtime_error(
"bfloat16 math functions are not currently supported on the host device.",
PI_ERROR_INVALID_DEVICE);
static const oneapi::detail::Bfloat16StorageT CanonicalNan = 0x7FC0;
oneapi::detail::Bfloat16StorageT XBits = oneapi::detail::bfloat16ToBits(x);
oneapi::detail::Bfloat16StorageT YBits = oneapi::detail::bfloat16ToBits(y);
if (isnan(x) && isnan(y))
return oneapi::detail::bitsToBfloat16(CanonicalNan);

if (isnan(x))
return y;
if (isnan(y))
return x;
if (((XBits | YBits) ==
static_cast<oneapi::detail::Bfloat16StorageT>(0x8000)) &&
!(XBits & YBits))
return oneapi::detail::bitsToBfloat16(
static_cast<oneapi::detail::Bfloat16StorageT>(0x8000));

return (x < y) ? x : y;
#endif // defined(__SYCL_DEVICE_ONLY__) && defined(__NVPTX__)
}

template <size_t N>
sycl::marray<bfloat16, N> fmin(sycl::marray<bfloat16, N> x,
sycl::marray<bfloat16, N> y) {
#if defined(__SYCL_DEVICE_ONLY__) && defined(__NVPTX__)
sycl::marray<bfloat16, N> res;

#if defined(__SYCL_DEVICE_ONLY__) && defined(__NVPTX__)
for (size_t i = 0; i < N / 2; i++) {
auto partial_res = __clc_fmin(detail::to_uint32_t(x, i * 2),
detail::to_uint32_t(y, i * 2));
Expand All @@ -101,15 +120,12 @@ sycl::marray<bfloat16, N> fmin(sycl::marray<bfloat16, N> x,
oneapi::detail::bfloat16ToBits(y[N - 1]);
res[N - 1] = oneapi::detail::bitsToBfloat16(__clc_fmin(XBits, YBits));
}

return res;
#else
std::ignore = x;
std::ignore = y;
throw runtime_error(
"bfloat16 math functions are not currently supported on the host device.",
PI_ERROR_INVALID_DEVICE);
for (size_t i = 0; i < N; i++) {
res[i] = fmin(x[i], y[i]);
}
#endif // defined(__SYCL_DEVICE_ONLY__) && defined(__NVPTX__)
return res;
}

template <typename T>
Expand All @@ -119,20 +135,30 @@ std::enable_if_t<std::is_same<T, bfloat16>::value, T> fmax(T x, T y) {
oneapi::detail::Bfloat16StorageT YBits = oneapi::detail::bfloat16ToBits(y);
return oneapi::detail::bitsToBfloat16(__clc_fmax(XBits, YBits));
#else
std::ignore = x;
std::ignore = y;
throw runtime_error(
"bfloat16 math functions are not currently supported on the host device.",
PI_ERROR_INVALID_DEVICE);
static const oneapi::detail::Bfloat16StorageT CanonicalNan = 0x7FC0;
oneapi::detail::Bfloat16StorageT XBits = oneapi::detail::bfloat16ToBits(x);
oneapi::detail::Bfloat16StorageT YBits = oneapi::detail::bfloat16ToBits(y);
if (isnan(x) && isnan(y))
return oneapi::detail::bitsToBfloat16(CanonicalNan);

if (isnan(x))
return y;
if (isnan(y))
return x;
if (((XBits | YBits) ==
static_cast<oneapi::detail::Bfloat16StorageT>(0x8000)) &&
!(XBits & YBits))
return oneapi::detail::bitsToBfloat16(0);

return (x > y) ? x : y;
#endif // defined(__SYCL_DEVICE_ONLY__) && defined(__NVPTX__)
}

template <size_t N>
sycl::marray<bfloat16, N> fmax(sycl::marray<bfloat16, N> x,
sycl::marray<bfloat16, N> y) {
#if defined(__SYCL_DEVICE_ONLY__) && defined(__NVPTX__)
sycl::marray<bfloat16, N> res;

#if defined(__SYCL_DEVICE_ONLY__) && defined(__NVPTX__)
for (size_t i = 0; i < N / 2; i++) {
auto partial_res = __clc_fmax(detail::to_uint32_t(x, i * 2),
detail::to_uint32_t(y, i * 2));
Expand All @@ -146,14 +172,12 @@ sycl::marray<bfloat16, N> fmax(sycl::marray<bfloat16, N> x,
oneapi::detail::bfloat16ToBits(y[N - 1]);
res[N - 1] = oneapi::detail::bitsToBfloat16(__clc_fmax(XBits, YBits));
}
return res;
#else
std::ignore = x;
std::ignore = y;
throw runtime_error(
"bfloat16 math functions are not currently supported on the host device.",
PI_ERROR_INVALID_DEVICE);
for (size_t i = 0; i < N; i++) {
res[i] = fmax(x[i], y[i]);
}
#endif // defined(__SYCL_DEVICE_ONLY__) && defined(__NVPTX__)
return res;
}

template <typename T>
Expand Down
2 changes: 1 addition & 1 deletion sycl/include/sycl/ext/oneapi/experimental/sycl_complex.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1201,7 +1201,7 @@ SYCL_EXTERNAL complex<_Tp> acos(const complex<_Tp> &__x) {
}
if (sycl::isinf(__x.imag()))
return complex<_Tp>(__pi / _Tp(2), -__x.imag());
if (__x.real() == 0 && (__x.imag() == 0 || isnan(__x.imag())))
if (__x.real() == 0 && (__x.imag() == 0 || sycl::isnan(__x.imag())))
return complex<_Tp>(__pi / _Tp(2), -__x.imag());
complex<_Tp> __z = log(__x + sqrt(__sqr(__x) - _Tp(1)));
if (sycl::signbit(__x.imag()))
Expand Down