junrushao1994 commented on a change in pull request #9871:
URL: https://github.com/apache/tvm/pull/9871#discussion_r780592805
##########
File path: src/tir/schedule/schedule.cc
##########
@@ -183,6 +183,20 @@ TVM_REGISTER_GLOBAL("tir.schedule.ScheduleStorageAlign")
TVM_REGISTER_GLOBAL("tir.schedule.ScheduleSetScope")
.set_body_method<Schedule>(&ScheduleNode::SetScope);
/******** (FFI) Blockize & Tensorize ********/
+TVM_REGISTER_GLOBAL("tir.schedule.ScheduleBlockize")
+ .set_body_method<Schedule>(&ScheduleNode::Blockize);
+TVM_REGISTER_GLOBAL("tir.schedule.ScheduleTensorize")
+ .set_body_typed([](Schedule self, LoopRV loop_rv, ObjectRef intrin) {
+ if (const auto* str = intrin.as<runtime::StringObj>()) {
+ return self->Tensorize(loop_rv, GetRef<String>(str));
+ }
+ if (const auto* p_intrin = intrin.as<TensorIntrinNode>()) {
+ return self->Tensorize(loop_rv, GetRef<TensorIntrin>(p_intrin));
+ }
+ LOG(FATAL) << "TypeError: Cannot handle type: " << intrin->GetTypeKey();
+ throw;
Review comment:
I'm considering...is it reasonable to use only the `String` API and not
expose the `TensorIntrin` one? The reason I'm asking this is that
`TensorIntrin` isn't available for `TracedSchedule` anyway, which is the
default go-to for python frontend users
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
