Context
When studying the DLPack header I realized that there is an explicit kDLCUDAManaged device type (see include/matx/core/dlpack.h:104) that is not emitted in the MatX repo. MatX still uses this fallback (tensor.h:1509):
// We have a record of this pointer and can map it from the record
switch (kind) {
case MATX_MANAGED_MEMORY:
case MATX_DEVICE_MEMORY:
case MATX_ASYNC_DEVICE_MEMORY:
t->device.device_type = kDLCUDA;
t->device.device_id = dev_ord;
break;
This is not 'wrong' but I wondered if there is any particular reason for doing so, or if it is simply a stale artifact that should be 'fixed'.
If this is intended, it might be helpful to add a short comment.
Context
When studying the DLPack header I realized that there is an explicit kDLCUDAManaged device type (see include/matx/core/dlpack.h:104) that is not emitted in the MatX repo. MatX still uses this fallback (tensor.h:1509):
This is not 'wrong' but I wondered if there is any particular reason for doing so, or if it is simply a stale artifact that should be 'fixed'.
If this is intended, it might be helpful to add a short comment.