tests/python/dnnl/subgraphs/test_conv_subgraph.py::test_pos_conv_...[...-data_shape1] fail with oneDNN v2.4+

[bartekkuncer]

Description

tests/python/dnnl/subgraphs/test_conv_subgraph.py::test_pos_conv_...[...-data_shape1] fail with oneDNN v2.4+ when the data_shape1 is (4, 3, 24, 24). After further investigation it turned out that the problem occurs for the amount of input channels (second dim) is lower than 4. Filled ticket against oneDNN pinpointing oneDNNs change that broke our fuse. Unfortunately the reproducer I got turned out not to work properly due to some other kind of error. As we need oneDNN in v2.4+ to be able to introduce new features to MXNet I temporarily changed the amount of input channels to 4. As soon as this bug gets fixed shape it the test should be restored to its original value.

To Reproduce

MXNET_MODULE_SEED=2004460228 py.test tests/python/dnnl/subgraphs/test_conv_subgraph.py::test_pos_conv_act_add[True-leakyrelu-True-data_shape1]
OR
MXNET_TEST_SEED=611986026 py.test tests/python/dnnl/subgraphs/test_conv_subgraph.py::test_pos_conv_act_add[True-gelu-True-data_shape1]

[bartekkuncer]

After further investigation found out that the real problem was reading from neither updated nor cleared register during convolution with weights stored in blocked format. This problem does not occur when the amount of input channels is 4 and higher as full register is being overwritten. Fix for this issue is coming in v2.6 of oneDNN.

Original shapes have been restored in the test. There was a temporary fix in place: #20847 forcing blocking the weights by 8 instead of 16, but it turned out only to make the problem occur less often. Therefore until v2.6 of oneDNN arrives there will be no blocking the weights for convolutions with the amount of input channels lower than 4: #20970.