[oneDNN] GRU BF16 kernel #27731
[oneDNN] GRU BF16 kernel #27731
Conversation
|
Thanks for your contribution! |
test=develop
jczaja
force-pushed
the
prv-gru-bf16-pr
branch
from
a07a253
to
af73601
Compare
|
@grygielski Could you please inspect fusion_gru_mkldnn_op.cc changes ? |
| "place does not support BF16 evaluation") | ||
| class TestFusionGRUBF16MKLDNNOp(OpTest): | ||
| def set_confs(self): | ||
| self.mkldnn_data_type = False |
There was a problem hiding this comment.
The mkldnn_data_type attribute is most commonly used in cpu_bfloat16 passes. In the GRU kernel it is not used, but the type of this attribute is std::string. So you can set self.mkldnn_data_type = "bfloat16" or remove this assignment.
There was a problem hiding this comment.
ok, will do in next PR
|
@luotao1 Could you please start your review? Your approval is needed to pass CI. |
|
Hi, @luotao1 This PR has only APPROVAL CI failing, please help approve. Thanks |
There was a problem hiding this comment.
LGTM
0. Unittest is not allowed to be disabled.
2020-10-01 20:54:05 You must have one RD (kolinwei(Recommend), or luotao1) approval for the usage of @unittest.skip or @unittest.skipIf.
You can add a new approval for the usage of @unittest.skipIf(not core.supports_bfloat16() in tools/check_file_diff_approvals.sh and skip the original one. The one RD (jczaja, xxx, xxx, or luotao1).
Or just skip the approval for the usage of @unittest.skip or @unittest.skipIf when not core.supports_bfloat16()
PR types
New features
PR changes
OPs
Describe
Introducing GRU BF16 (brain float 16) kernel implemented via oneDNN library.