add max shared mem workaround
This commit is contained in:
parent
0325ce0bed
commit
595cd81251
@ -278,7 +278,11 @@ void BuildGradientHistogram(CUDAContext const* ctx, EllpackDeviceAccessor const&
|
|||||||
#endif
|
#endif
|
||||||
|
|
||||||
// opt into maximum shared memory for the kernel if necessary
|
// opt into maximum shared memory for the kernel if necessary
|
||||||
|
#if defined(XGBOOST_USE_CUDA)
|
||||||
size_t max_shared_memory = dh::MaxSharedMemoryOptin(device);
|
size_t max_shared_memory = dh::MaxSharedMemoryOptin(device);
|
||||||
|
#elif defined(XGBOOST_USE_HIP)
|
||||||
|
size_t max_shared_memory = dh::MaxSharedMemory(device);
|
||||||
|
#endif
|
||||||
|
|
||||||
size_t smem_size =
|
size_t smem_size =
|
||||||
sizeof(GradientPairInt64) * feature_groups.max_group_bins;
|
sizeof(GradientPairInt64) * feature_groups.max_group_bins;
|
||||||
|
|||||||
Loading…
x
Reference in New Issue
Block a user