Orion34C commented on a change in pull request #4546: [CODEGEN] Support cuda
tensorcore subbyte int data type in auto tensorcore
URL: https://github.com/apache/incubator-tvm/pull/4546#discussion_r362204788
##########
File path: src/pass/arg_binder.cc
##########
@@ -184,7 +184,9 @@ void ArgBinder::BindDLTensor(const Buffer& buffer,
UIntImm::make(UInt(8), dtype.bits()) &&
TVMArrayGet(UInt(16), handle, intrinsic::kArrTypeLanes) ==
UIntImm::make(UInt(16), dtype.lanes()));
- asserts_.emplace_back(AssertStmt::make(cond, type_err_msg.str(), nop));
+ if (!(dtype == Int(4) || dtype == UInt(4) || dtype == Int(1))) {
Review comment:
This hack is added because int4/int1 is not supported in basic data types as
in numpy or c, we can only feed int4/int1 data in int8 or int32，which is not
aligned with tvm ir's data type Int(4) or Int(1) etc. I wonder if you guys have
a better solution than this hacky one.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
