quic-sanirudh opened a new issue, #13957:
URL: https://github.com/apache/tvm/issues/13957
When we apply `transform_layout` on a write buffer generated by
`cache_write` with axis separators, we see a crash where the type_index for the
`BlockNode` stmt in its corresponding sref gets corrupted. The below code
reproduces the error:
### Expected behavior
The code should work fine without errors and apply the layout transform with
axis separators
### Actual behavior
Code crashes with different errors like `Unknown type index 2106063344` or
`TypeError: Expects StmtSRef 'block_sref' points to 'Block', but gets:
runtime.Object`
### Environment
Tested on Upstream TVM main with Ubuntu 18.04
### Steps to reproduce
```python
import tvm
import tvm.contrib.hexagon
import tvm.testing
from tvm.script import tir as T
@T.prim_func
def add(
p0: T.Buffer((T.int64(33), T.int64(128)), "float32"),
p1: T.Buffer((T.int64(33), T.int64(128)), "float32"),
T_add: T.Buffer((T.int64(33), T.int64(128)), "float32"),
):
T.func_attr({"global_symbol": "main", "tir.noalias": True})
# with T.block("root"):
for ax0, ax1 in T.grid(T.int64(33), T.int64(128)):
with T.block("T_add"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(p0[v_ax0, v_ax1], p1[v_ax0, v_ax1])
T.writes(T_add[v_ax0, v_ax1])
T_add[v_ax0, v_ax1] = p0[v_ax0, v_ax1] + p1[v_ax0, v_ax1]
def transform_fn(x, y):
return [x // 32, y, tvm.te.AXIS_SEPARATOR, x % 32]
def main():
sch = tvm.tir.Schedule(add)
block_rv = sch.get_block("T_add")
sch.cache_write(block_rv, 0, "global")
sch.transform_layout(block_rv, ("write", 0), transform_fn, pad_value=0.0)
if __name__ == "__main__":
main()
```
### Triage
* [needs-triage](tir:schedule)
--
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]
