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
|
||||
|
||||
// opt into maximum shared memory for the kernel if necessary
|
||||
#if defined(XGBOOST_USE_CUDA)
|
||||
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 =
|
||||
sizeof(GradientPairInt64) * feature_groups.max_group_bins;
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user