Lunderberg commented on PR #14320:
URL: https://github.com/apache/tvm/pull/14320#issuecomment-1473951122
@junrushao Thank you, and I can work around the issue by using an explicit
`T.FloorDiv`, so this isn't by any means a blocking issue for me. It's more
that it was an unexpected change, such as in the example below, and I wanted to
avoid it becoming a similar issue for others.
@cyx-6 Certainly. The example below shows the weird interaction between the
dynamic parameter and the variable name, which is what led to my initial
investigation.
```python
import tvm
from tvm.script import tir as T
@pytest.mark.parametrize("split_factor", [1, 2, 4])
def test_function(split_factor):
@T.prim_func
def explicit_tir_node(A: T.Buffer(16, "int32"), B: T.Buffer(16,
"int32")):
for i in T.serial(T.FloorDiv(16, split_factor)):
chunk_start = T.Mul(i, split_factor)
for j in T.vectorized(split_factor):
B[chunk_start + j] = A[chunk_start + j]
@T.prim_func
def using_operator(A: T.Buffer(16, "int32"), B: T.Buffer(16, "int32")):
for i in T.serial(16 // split_factor):
chunk_start = i * split_factor
for j in T.vectorized(split_factor):
B[chunk_start + j] = A[chunk_start + j]
using_operator.show()
tvm.ir.assert_structural_equal(explicit_tir_node, using_operator)
```
This passes when the dynamic parameter `split_factor` is 2 or 4, but fails
when `split_factor == 1`. In that case, it generates the PrimFunc shown below.
For this case, the `chunk_start = i * 1` was constant-folded to `chunk_start =
i`, which was then used to rename `i` to `chunk_start` in the loop iterator.
```python
@T.prim_func
def main(A: T.Buffer((16,), "int32"), B: T.Buffer((16,), "int32")):
for chunk_start in range(16):
for j in T.vectorized(1):
B[chunk_start + j] = A[chunk_start + j]
```
--
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: commits-unsubscr...@tvm.apache.org
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
