Johnson9009 opened a new issue #8536:
URL: https://github.com/apache/tvm/issues/8536
I found the bug when sync the latest TVM commits into our own repo, after
that lots of our test cases are failed.
The bug will cause the case that all of the operators are annotated to
executed on a custom device fail, because the "target" is set to "llvm"
incorrectly instead of the custom target like "xxpu".
Even through the root cause of the failure is difficult to describe and
understand, but the bug is very obvious.
### Before changes of #7518.
```C++
Target GetTargetFromInteger(int64_t dev_type) {
if (targets_.size() == 1) {
// homogeneous execution.
const auto& it = targets_.begin();
return (*it).second;
} else {
// heterogeneous execution.
std::string call_dev_name;
if (dev_type == 0) {
--> call_dev_name = "llvm"; <--- Here haven't set "dev_type",
"dev_type" is still equals to 0.
} else {
call_dev_name = runtime::DeviceName(dev_type);
}
if (targets_.count(dev_type) == 0) {
LOG(FATAL) << "No target is provided for device " << call_dev_name;
}
--> return targets_[dev_type]; <--- Use 0 get the correct "target".
}
}
...
target = GetTargetFromInteger(call_dev_type);
```
### After changes of #7518.
```C++
std::string call_dev_name;
if (call_dev_type == 0) {
call_dev_name = "llvm";
---> call_dev_type = kDLCPU; <--- Here shouldn't change the value of
"call_dev_type".
} else {
call_dev_name = ::tvm::runtime::DeviceName(call_dev_type);
}
if (targets_.count(call_dev_type) == 0) {
std::stringstream msg;
msg << "No target is specified for provided device name: `" <<
call_dev_name << "`\n\n";
....
}
LOG(FATAL) << msg.str();
}
---> target = targets_[call_dev_type]; <--- The 0th item of "targets_"
is the correct target, but the "call_dev_type" is set to kDLCPU(i.e., 1), so
the target is wrong.
```
I think the authors of #7518 is mislead by the code `call_dev_name =
"llvm";`, if we set the target dictionary to `{"cpu": "llvm", "xxpu": "xxpu"}`
when call "relay.build", then the size of `targets_` at above code is 3 instead
of 2, the 0th item of `targets_` maybe is "xxpu", so here the call_dev_type
can't be changed to kDLCPU.
@csullivan @jroesch @tqchen
--
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]
