18761692867 opened a new issue, #12162:
URL: https://github.com/apache/tvm/issues/12162
I have a problem today, maybe is a bug in tvm 0.6, my english is poor, i
will try my best to describe the problem clearly ^ ^
**the test py :**
`def test_rfactor():
shape_in = (4, 64, 32, 16)
dtype = "float16"
ph_0 = tvm.placeholder(shape_in, dtype=dtype, name="ph_0")
reduce_axis_1 = tvm.reduce_axis((0, 64), name="reduce_axis_1")
reduce_axis_3 = tvm.reduce_axis((0, 16), name="reduce_axis_3")
shape_out = (4, 1, 32, 1)
reduce_0 = tvm.compute(shape_out,
lambda *i: tvm.sum(ph_0[i[0], reduce_axis_1,
i[2], reduce_axis_3],
(reduce_axis_1,
reduce_axis_3)),
name="reduce_0")
sch = tvm.create_schedule([reduce_0.op])
cache_read_buffer = sch.cache_read(ph_0, "local.UB", [reduce_0])
blk_outer, blk_inner = sch[reduce_0].split(reduce_0.op.axis[0], factor=2)
rfactor_buffer = sch.rfactor(reduce_0, reduce_0.op.reduce_axis[-1], -1)
print(tvm.lower(sch, [ph_0], simple_mode=True))`
**the error message:**
terminate called after throwing an instance of 'std::out_of_range'
what(): _Map_base::at
Aborted (core dumped)
**the reason is:**
in rfactor function, we create .rf stage, this stage has origin stage's
root_axis and relations, but not has origin stage's leaf_axis. so, if origin
stage's relations has a SplitNode, rf stage will also has. then, in
inferbound's passup phase, it will coredump. because we use
up_state.at(r->outer), but up_state dont has r->outer.
--
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]
