lightzhan-intellif opened a new pull request, #16104:
URL: https://github.com/apache/tvm/pull/16104
This PR tries to fix a bug of the pass LoopPartiton. When there are one or
more tensors containing a shape 1 in the concat dim, the pass will unroll the
loops wrongly after partitioning. For example:
```python
@T.prim_func
def concat_func_single_point(
placeholder: T.Buffer((28, 64), "int8"),
placeholder_1: T.Buffer((28, 1), "int8"),
placeholder_2: T.Buffer((28, 63), "int8"),
T_concat: T.Buffer((28, 128), "int8"),
) -> None:
for i0 in range(28):
for i1 in T.serial(128, annotations={"pragma_loop_partition_hint":
1}):
if i1 > 63:
T_concat[i0, i1] = placeholder[i0, i1 - 64]
elif i1 == 63:
T_concat[i0, i1] = placeholder_1[i0, i1 - 63]
else:
T_concat[i0, i1] = placeholder_2[i0, i1]
```
after LoopPartition:
```python
@T.prim_func
def expected_partitioned_concat_single_point(
placeholder: T.Buffer((28, 64), "int8"),
placeholder_1: T.Buffer((28, 1), "int8"),
placeholder_2: T.Buffer((28, 63), "int8"),
T_concat: T.Buffer((28, 128), "int8"),
):
for i0 in range(28):
T_concat_1 = T.Buffer((3584,), "int8", data=T_concat.data)
for i1 in T.unroll(63): # Note here, it is unrolled.
placeholder_2_1 = T.Buffer((1764,), "int8",
data=placeholder_2.data)
T_concat_1[i0 * 128 + i1] = placeholder_2_1[i0 * 63 + i1]
placeholder_1_1 = T.Buffer((28,), "int8", data=placeholder_1.data)
T_concat_1[i0 * 128 + 63] = placeholder_1_1[i0]
for i1 in T.unroll(64): # here too.
placeholder_3 = T.Buffer((1792,), "int8", data=placeholder.data)
T_concat_1[i0 * 128 + i1 + 64] = placeholder_3[i0 * 64 + i1]
```
cc @wrongtest-intellif @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]
