AI/rocm-systems
2
0
Bifurcation 0
Code Tickets Demandes d'ajout Actions Paquets Projets Publications Wiki Activité

enable building rccl for gfx950 (#1571)

Parcourir la source 
[ROCm/rccl commit: f268553ee4]
Cette révision appartient à :
Pedram Alizadeh
2025-02-25 16:13:48 -05:00
révisé par GitHub
Parent c363118589
révision acf5822a6c
24 fichiers modifiés avec 66 ajouts et 47 suppressions
Télécharger le Fichier Patch Télécharger le Fichier des Différences Développer tous les fichiers Réduire tous les fichiers
projects/rccl/src/include/rccl_float8.h
+7 -7
Voir le fichier
@@ -344,7 +344,7 @@ struct rccl_float8
// default constructor
HIP_HOST_DEVICE rccl_float8() = default;
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
// device specific optimized F8 down-conversion code
template <bool stochastic_rounding = false>
@@ -384,7 +384,7 @@ struct rccl_float8
#endif // __gfx940__
// constructor from float
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
// NOTE: ON-DEVICE... always optimal bias
explicit HIP_DEVICE rccl_float8(float v,
@@ -446,7 +446,7 @@ struct rccl_float8
}
// convert to float
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
// upcast using device specific intrinsic
explicit inline HIP_DEVICE operator float() const
{
@@ -511,7 +511,7 @@ struct rccl_bfloat8
// default constructor
HIP_HOST_DEVICE rccl_bfloat8() = default;
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
// device specific optimized F8 down-conversion code
template <bool stochastic_rounding = false>
@@ -551,7 +551,7 @@ struct rccl_bfloat8
#endif // __gfx940__
// constructor from float
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
// NOTE: ON-DEVICE... always optimal bias
explicit HIP_DEVICE rccl_bfloat8(float v,
@@ -613,7 +613,7 @@ struct rccl_bfloat8
}
// convert to float
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
// upcast using device specific intrinsic
explicit inline HIP_DEVICE operator float() const
{
@@ -980,7 +980,7 @@ template <
= 0>
inline __host__ __device__ T explicit_downcast(Ta a, uint32_t rng)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
// NOTE: we are directly calling cast_to_f8_from_f32 instead of constructor to optimize away one runtime branch
T val;
if(std::is_same<T, rccl_float8>::value)