Merge remote-tracking branch 'rccl/develop' into 2.19.4

[ROCm/rccl commit: 1a134b283b]
Este commit está contenido en:
Wenkai Du
2024-01-31 11:08:54 -06:00
Se han modificado 14 ficheros con 31 adiciones y 30 borrados
+2 -4
Ver fichero
@@ -372,7 +372,7 @@ __device__ __forceinline__ void mscclRunInterpreter(
return;
}
if (t->hasDependence && tid == nthreads-1)
__atomic_store_n(&mscclFlags[bid].flag, (uint64_t) COMPUTE_FLAG(workIndex, iter, step), t->type == MSCCL_REDUCE ? __ATOMIC_RELEASE : __ATOMIC_RELAXED);
__atomic_store_n(&mscclFlags[bid].flag, (uint64_t) COMPUTE_FLAG(workIndex, iter, step), (t->type == MSCCL_REDUCE || t->type == MSCCL_RECV) ? __ATOMIC_RELEASE : __ATOMIC_RELAXED);
step++;
}
}
@@ -427,8 +427,6 @@ __global__ void MSCCL_KERNEL_ENTRY_NAME(devredop, type, Simple, fullOps)(struct
#define MSCCL_IMPL_KERNEL_ENTRY_FUNC() \
MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP(Sum, false) \
MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP(Prod, false) \
MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP(MinMax, false) \
MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP(PreMulSum, false) \
MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_NOFLOAT(SumPostDiv, false)
MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP(MinMax, false)
#endif