MKLDNN: SoftmaxGrad Op #11519
MKLDNN: SoftmaxGrad Op #11519
Conversation
| @@ -99,5 +99,124 @@ inline mkldnn::memory::format GetMKLDNNFormat(const mkldnn::memory memory) { | |||
| memory.get_primitive_desc().desc().data.format); | |||
| } | |||
|
|
|||
| class MKLDNNHandler { | |||
There was a problem hiding this comment.
Could you give more introductions about MKLDNNHandler？
Is this specifically added for softmax grad since this.
When adding functions at mkldnn_helper.h, it should be something common I suppose.
There was a problem hiding this comment.
@tensor-tang You are right that softmax functionality should go to into MKLDNNHandler. MKLDNNHandler will agregate common functionality as for softmax/conv/pool primitive specific functionality there will be designated derived classes eg. SoftmaxMKLDNNHandler , ConvMKLDNNHandler etc. I updated the code adding SoftmaxMKLDNNHandler. Please take a look
jczaja
force-pushed
the
prv-softmax-mkldnn-grad-operator
branch
5 times, most recently
from
8a571ef
to
6a2a282
Compare
|
Hi, I answered on @tensor-tang question within relevant disscussion. My changes were updated |
|
There was a merge conflict but it is resolved now. |
- Added hash function inside of MKLDNN softmax op to be used as handle for primitives stroing in a context - Style fixes to softmax mkldnn op - Fixes after review - Coding style - Fix to style - style fixes - style fix - style fixes - Fix to cody style check - Rephrasing a comment fix t obroken merge Fixes to rebase Conflicts: benchmark/fluid/models/machine_translation.py cmake/external/mkldnn.cmake paddle/fluid/operators/softmax_mkldnn_op.cc - Bumped revision of MKL-DNN up to have softmax backward primitive - Added choosing MKLDNN softmax grad operator - First reuse of softmax backward - Reinvented reusing for softmax - Fix to crash in reinvented reuse - Clang format fixes - Clang format fixes - Improved softmax mkldnn reuse mechanism - clang format fixes - Fix to broken merge - Fix
jczaja
force-pushed
the
prv-softmax-mkldnn-grad-operator
branch
from
be4a58c
to
98f3ad3
Compare
With this PR machine translation when MKLDNN is used (use_mkldnn=True, shall be added in fc layer of machine_translation.py) is speeding up significantly