YXY-0922 opened a new pull request, #17413:
URL: https://github.com/apache/tvm/pull/17413
When loading from database_tuning_record.json in Meta Schedule (this line:
`B_reindex_pad_shared_dyn[v0, v1] = T.if_then_else(v0 < 1, B[v1, v0],
T.float16(0)))`, the parameter dtype of the primitive pad_einsum is read as
int64, causing a block iterator v0 that should be int32 to be inferred as
int64. This results in an InternalError: Check failed: (ret_ex.dtype() ==
var.dtype()) is false: substituting v0:int32 -> v0:int64. This commit performs
a type conversion within UnpackedApplyToSchedule to fix this bug.
```
for ax0_ax1_fused in range(2048):
with T.block("B_reindex_pad_shared.dyn"):
v0 = T.axis.spatial(16, ax0_ax1_fused // 128)
v1 = T.axis.spatial(4096, ax2_0_0 * 128 + ax0_ax1_fused % 128)
T.reads(B[v1, v0])
T.writes(B_reindex_pad_shared_dyn[v0, v1])
T.block_attr({"buffer_dim_align": [[0, 0, 32, 8]],
"meta_schedule.cooperative_fetch": 1})
B_reindex_pad_shared_dyn[v0, v1] = T.if_then_else(v0 < 1, B[v1, v0],
T.float16(0))
```
--
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]
