AI/rocm-systems
2
0
Atdalīts 0
Kods Problēmas Izmaiņu pieprasījumi Darbības Pakotnes Projekti Laidieni Vikivietne Aktivitāte

Merge commit '2305f8ae565389cfc348a747c467c004909d7a90' into develop

Pārlūkot izejas kodu
Šī revīzija ir iekļauta:
systems-assistant[bot]
2025-08-13 00:24:58 +00:00
vecāks 0ba726f8fc 2305f8ae56
revīzija fff63d428e
22 mainīti faili ar 308 papildinājumiem un 122 dzēšanām
Lejupielādēt ielāpa failu Lejupielādēt izmaiņu failu Izvērst visus failus Savērst visus failus
projects/clr/hipamd/include/hip/amd_detail/amd_hip_bf16.h
+60
Parādīt failu
@@ -781,6 +781,16 @@ __BF16_HOST_DEVICE_STATIC__ __hip_bfloat16 __hadd(const __hip_bfloat16 a, const
return (__bf16)a + (__bf16)b;
}
/**
* \ingroup HIP_INTRINSIC_BFLOAT16_ARITH
* \brief Adds two bfloat16 values, will not fuse into fma
*/
__BF16_HOST_DEVICE_STATIC__ __hip_bfloat16 __hadd_rn(const __hip_bfloat16 a,
const __hip_bfloat16 b) {
#pragma clang fp contract(off)
return (__bf16)a + (__bf16)b;
}
/**
* \ingroup HIP_INTRINSIC_BFLOAT16_ARITH
* \brief Subtracts two bfloat16 values
@@ -789,6 +799,16 @@ __BF16_HOST_DEVICE_STATIC__ __hip_bfloat16 __hsub(const __hip_bfloat16 a, const
return (__bf16)a - (__bf16)b;
}
/**
* \ingroup HIP_INTRINSIC_BFLOAT16_ARITH
* \brief Subtracts two bfloat16 values, will not fuse into fma
*/
__BF16_HOST_DEVICE_STATIC__ __hip_bfloat16 __hsub_rn(const __hip_bfloat16 a,
const __hip_bfloat16 b) {
#pragma clang fp contract(off)
return (__bf16)a - (__bf16)b;
}
/**
* \ingroup HIP_INTRINSIC_BFLOAT16_ARITH
* \brief Divides two bfloat16 values
@@ -815,6 +835,16 @@ __BF16_HOST_DEVICE_STATIC__ __hip_bfloat16 __hmul(const __hip_bfloat16 a, const
return (__bf16)a * (__bf16)b;
}
/**
* \ingroup HIP_INTRINSIC_BFLOAT16_ARITH
* \brief Multiplies two bfloat16 values, will not fuse into fma
*/
__BF16_HOST_DEVICE_STATIC__ __hip_bfloat16 __hmul_rn(const __hip_bfloat16 a,
const __hip_bfloat16 b) {
#pragma clang fp contract(off)
return (__bf16)a * (__bf16)b;
}
/**
* \ingroup HIP_INTRINSIC_BFLOAT16_ARITH
* \brief Negate a bfloat16 value
@@ -861,6 +891,16 @@ __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __hadd2(const __hip_bfloat162 a,
return __hip_bfloat162{__bf16_2(a) + __bf16_2(b)};
}
/**
* \ingroup HIP_INTRINSIC_BFLOAT162_ARITH
* \brief Adds two bfloat162 values, will not fuse into fma
*/
__BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __hadd2_rn(const __hip_bfloat162 a,
const __hip_bfloat162 b) {
#pragma clang fp contract(off)
return __hip_bfloat162{__bf16_2(a) + __bf16_2(b)};
}
/**
* \ingroup HIP_INTRINSIC_BFLOAT162_ARITH
* \brief Performs FMA of given bfloat162 values
@@ -879,6 +919,16 @@ __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __hmul2(const __hip_bfloat162 a,
return __hip_bfloat162{__bf16_2(a) * __bf16_2(b)};
}
/**
* \ingroup HIP_INTRINSIC_BFLOAT162_ARITH
* \brief Multiplies two bfloat162 values, will not fuse into fma
*/
__BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __hmul2_rn(const __hip_bfloat162 a,
const __hip_bfloat162 b) {
#pragma clang fp contract(off)
return __hip_bfloat162{__bf16_2(a) * __bf16_2(b)};
}
/**
* \ingroup HIP_INTRINSIC_BFLOAT162_ARITH
* \brief Converts a bfloat162 into negative
@@ -896,6 +946,16 @@ __BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __hsub2(const __hip_bfloat162 a,
return __hip_bfloat162{__bf16_2(a) - __bf16_2(b)};
}
/**
* \ingroup HIP_INTRINSIC_BFLOAT162_ARITH
* \brief Subtracts two bfloat162 values, will not fuse into fma
*/
__BF16_HOST_DEVICE_STATIC__ __hip_bfloat162 __hsub2_rn(const __hip_bfloat162 a,
const __hip_bfloat162 b) {
#pragma clang fp contract(off)
return __hip_bfloat162{__bf16_2(a) - __bf16_2(b)};
}
/**
* \ingroup HIP_INTRINSIC_BFLOAT16_ARITH
* \brief Operator to multiply two __hip_bfloat16 numbers
projects/clr/hipamd/include/hip/amd_detail/amd_hip_fp16.h
+55
Parādīt failu
@@ -1367,6 +1367,15 @@ THE SOFTWARE.
return __half_raw{
static_cast<__half_raw>(x).data +
static_cast<__half_raw>(y).data};
}
inline
__HOST_DEVICE__
__half __hadd_rn(__half x, __half y)
{
#pragma clang fp contract(off)
return __half_raw{
static_cast<__half_raw>(x).data +
static_cast<__half_raw>(y).data};
}
inline
__HOST_DEVICE__
@@ -1390,6 +1399,15 @@ THE SOFTWARE.
}
inline
__HOST_DEVICE__
__half __hsub_rn(__half x, __half y)
{
#pragma clang fp contract(off)
return __half_raw{
static_cast<__half_raw>(x).data -
static_cast<__half_raw>(y).data};
}
inline
__HOST_DEVICE__
__half __hmul(__half x, __half y)
{
return __half_raw{
@@ -1398,6 +1416,15 @@ THE SOFTWARE.
}
inline
__HOST_DEVICE__
__half __hmul_rn(__half x, __half y)
{
#pragma clang fp contract(off)
return __half_raw{
static_cast<__half_raw>(x).data *
static_cast<__half_raw>(y).data};
}
inline
__HOST_DEVICE__
__half __hadd_sat(__half x, __half y)
{
return __clamp_01(__hadd(x, y));
@@ -1446,6 +1473,16 @@ THE SOFTWARE.
static_cast<__half2_raw>(x).data +
static_cast<__half2_raw>(y).data};
}
inline
__HOST_DEVICE__
__half2 __hadd2_rn(__half2 x, __half2 y)
{
#pragma clang fp contract(off)
return __half2{
static_cast<__half2_raw>(x).data +
static_cast<__half2_raw>(y).data};
}
inline
__HOST_DEVICE__
__half2 __habs2(__half2 x)
@@ -1462,6 +1499,15 @@ THE SOFTWARE.
}
inline
__HOST_DEVICE__
__half2 __hsub2_rn(__half2 x, __half2 y)
{
#pragma clang fp contract(off)
return __half2{
static_cast<__half2_raw>(x).data -
static_cast<__half2_raw>(y).data};
}
inline
__HOST_DEVICE__
__half2 __hmul2(__half2 x, __half2 y)
{
return __half2{
@@ -1470,6 +1516,15 @@ THE SOFTWARE.
}
inline
__HOST_DEVICE__
__half2 __hmul2_rn(__half2 x, __half2 y)
{
#pragma clang fp contract(off)
return __half2{
static_cast<__half2_raw>(x).data *
static_cast<__half2_raw>(y).data};
}
inline
__HOST_DEVICE__
__half2 __hadd2_sat(__half2 x, __half2 y)
{
auto r = static_cast<__half2_raw>(__hadd2(x, y));
projects/clr/hipamd/include/hip/amd_detail/amd_hip_vector_types.h
+2 -2
Parādīt failu
@@ -51,8 +51,8 @@ THE SOFTWARE.
#include <type_traits>
#endif // defined(__HIPCC_RTC__)
template <class T, int N> constexpr __hip_internal::size_t __hip_vec_align_v() {
return (N == 4 && alignof(T) == 8) ? 16 : N * alignof(T);
template <class T, int _hip_N> constexpr __hip_internal::size_t __hip_vec_align_v() {
return (_hip_N == 4 && alignof(T) == 8) ? 16 : _hip_N * alignof(T);
}
template <typename T, unsigned int n> struct HIP_vector_base;
projects/clr/hipamd/include/hip/amd_detail/host_defines.h
+4 -4
Parādīt failu
@@ -199,16 +199,16 @@ struct integer_sequence {
template<size_t... Ints>
using index_sequence = integer_sequence<size_t, Ints...>;
template<size_t N, size_t... Ints>
struct make_index_sequence_impl : make_index_sequence_impl<N - 1, N - 1, Ints...> {};
template <size_t _hip_N, size_t... Ints>
struct make_index_sequence_impl : make_index_sequence_impl<_hip_N - 1, _hip_N - 1, Ints...> {};
template<size_t... Ints>
struct make_index_sequence_impl<0, Ints...> {
using type = index_sequence<Ints...>;
};
template<size_t N>
using make_index_sequence = typename make_index_sequence_impl<N>::type;
template <size_t _hip_N>
using make_index_sequence = typename make_index_sequence_impl<_hip_N>::type;
template <size_t... Ints>
constexpr index_sequence<Ints...> make_index_sequence_value(index_sequence<Ints...>) {