[Bugfix] Conv3Dtranspose default kernel layout should be IODHW

[rebel-jangys]

Fix the issue #14326
@masahi

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• No users to tag found in teams: bugfix _{See #10317 for details}

_{Generated by tvm-bot}

[masahi]

Can you look at the failed test?

[rebel-jangys]

SInce this PR corrects the conv3d transpose kernel layout, many frontend relay converters and test cases have to be fixed as well.

For example, the below test case intentionally set wrong kernel_layout to fit in original wrong kernel layout implementation.

tvm/tests/python/contrib/test_dnnl.py

Lines 479 to 490 in 46fb2ff

	def get_conv3d_transpose(
	x_shape=(1, 32, 8, 8, 8),
	k_shape=(32, 16, 3, 3, 3),
	groups=1,
	padding=(0, 0, 0),
	strides=(1, 1, 1),
	output_padding=(0, 0, 0),
	activation=None,
	dtype="float32",
	data_layout="NCDHW",
	kernel_layout="OIDHW",
	):

I’ve fixed pytorch relay converter and some test cases, but I would greatly appreciate any assistance with fix.

[rebel-jangys]

cc @apeskov , this PR is the one you commented as TODO

tvm/src/runtime/contrib/dnnl/dnnl_json_runtime.cc

Line 427 in cbe068c

// TODO(@apeskov): WA. conv3dTranspose uses wrong layout specifier. IO instead of OI.

[masahi]

@rebel-jangys Thanks for your hard work, it's merged finally!