Repository navigation
Vectorizing a reduce_axis does not raise an error #1795
Description
In #670 it was said that vectorizing a reduce_axis should raise an error, but it currently just silently gives wrong answers, e.g.:
import tvm
V = tvm.placeholder((128,), name='V')
ax = tvm.reduce_axis((0, 128), name='ax')
O = tvm.compute((1,), lambda _: tvm.sum(V[ax], axis=[ax]))
s = tvm.create_schedule(O.op)
s[O].vectorize(ax) # INVALID, but does not error!
func = tvm.build(s, [V, O])
vv = tvm.nd.array(np.ones((128,), dtype='float32'))
oo = tvm.nd.array(np.empty((1,), dtype='float32'))
func(vv, oo)
print(oo.asnumpy())
# prints: [1.]
# should print: [128.]I just ran into this issue and was confused :)
Activity
tqchen commentedon Oct 2, 2018
Thanks for reporting this!
I think that was something we missed and CHECK guard should be added here https://github.com/dmlc/tvm/blob/master/src/schedule/schedule_lang.cc#L355 to check the type of var(like https://github.com/dmlc/tvm/blob/master/src/schedule/schedule_lang.cc#L46)
Would you be interested in send a PR along with a regression testcase?
kazimuth commentedon Oct 2, 2018
Sure, I'll get to this within the next day or so!
tqchen commentedon Oct 17, 2018
@kazimuth any updates on this ?
tqchen commentedon Oct 19, 2018
close by #1921