Description

Hi.
It seems that SoftmaxActivation(cudnn version) has a problem in MXNet 1.2.0.
It will cause the error when using SoftmaxActivation.
In the issue(#9823), chinakook considers the bug produced after PR#9677

I found the function of mx.nd.Softmax and mx.nd.SoftmaxActivation seems to be the same. The difference is that Softmax uses pure CUDA, and SoftmaxActivation uses CUDNN.
Is it necessary to merge them?

Environment info (Required)

----------Python Info----------
('Version      :', '2.7.12')
('Compiler     :', 'GCC 5.4.0 20160609')
('Build        :', ('default', 'Dec  4 2017 14:50:18'))
('Arch         :', ('64bit', 'ELF'))
------------Pip Info-----------
('Version      :', '10.0.1')
('Directory    :', '/usr/local/lib/python2.7/dist-packages/pip')
----------MXNet Info-----------
('Version      :', '1.2.0')
('Directory    :', '/usr/local/lib/python2.7/dist-packages/mxnet')
('Commit Hash   :', '5088ca9a65641ddf905b60deae00fa6006f5e431')
----------System Info----------
('Platform     :', 'Linux-4.13.9-coreos-x86_64-with-Ubuntu-16.04-xenial')
('system       :', 'Linux')
('release      :', '4.13.9-coreos')
('version      :', '#1 SMP Thu Oct 26 03:21:00 UTC 2017')

Package used (Python/R/Scala/Julia):
Python

Build info (Required if built from source)

Installed by pip:
pip install mxnet-cu80 --pre

Error Message:

Traceback (most recent call last):
File "/home/wkcn/proj/incubator-mxnet/test_softmaxat.py", line 12, in
print (a.grad.asnumpy())
File "/usr/local/lib/python2.7/dist-packages/mxnet/ndarray/ndarray.py", line 1894, in asnumpy
ctypes.c_size_t(data.size)))
File "/usr/local/lib/python2.7/dist-packages/mxnet/base.py", line 149, in check_call
raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: [07:18:14] src/operator/nn/./cudnn/cudnn_softmax_activation-inl.h:154: Check failed: e == CUDNN_STATUS_SUCCESS (3 vs. 0) cuDNN: CUDNN_STATUS_BAD_PARAM

Minimum reproducible example

import mxnet as mx
ctx = mx.gpu(0)

a = mx.nd.array([[1,2,3]], ctx = ctx)

a.attach_grad()
with mx.autograd.record():
    y = mx.nd.SoftmaxActivation(data = a)

y.backward()
mx.nd.waitall()
print (a.grad.asnumpy())