Thrsu opened a new issue, #17487:
URL: https://github.com/apache/tvm/issues/17487
The below code defines a custom TIR function that computes the atan of each
element in a buffer of shape (20,) and then uses it within a relax function.
When trying to build the module using relax.build targeting llvm, it raises an
error: TVMError: unknown intrinsic Op(tir.atan).
### Expected behavior
The tir.atan operation should be recognized and compiled correctly without
throwing this error, as it is a common mathematical operation.
### Actual behavior
```
File "/software/tvm/src/target/llvm/codegen_llvm.cc", line 1491
TVMError: unknown intrinsic Op(tir.atan)
```
### Steps to reproduce
```python
import tvm
from tvm import relax
from tvm.script import ir as I
from tvm.script import tir as T
from tvm.script import relax as R
@I.ir_module
class Module:
@T.prim_func(private=True)
def tir_atan(x: T.Buffer((T.int64(20),), "float16"), compute:
T.Buffer((T.int64(20),), "float16")):
T.func_attr({"tir.noalias": T.bool(True)})
for i0 in range(T.int64(20)):
with T.block("compute"):
v_i0 = T.axis.spatial(T.int64(20), i0)
T.reads(x[v_i0])
T.writes(compute[v_i0])
compute[v_i0] = T.atan(x[v_i0])
@R.function
def main(x: R.Tensor((20,), dtype="float16")) -> R.Tensor((20,),
dtype="float16"):
R.func_attr({"num_input": 1})
cls = Module
with R.dataflow():
gv = R.call_tir(cls.tir_atan, (x,), out_sinfo=R.Tensor((20,),
dtype="float16"))
R.output(gv)
return gv
mod = Module
ex = relax.build(mod, target='llvm')
```
It is unclear if this is due to a missing intrinsic support for atan in TIR
or if there is an issue with registering this intrinsic in the target. Any
guidance or fixes to resolve this issue would be appreciated.
--
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]
