Initial implementation of multiplication operator for MKLDNN #9949
Conversation
mozga-intel
force-pushed
the
mozga-intel/Mul_mkldnn
branch
6 times, most recently
from
97843f1
to
f24fec2
Compare
mozga-intel
changed the title
mozga-intel
force-pushed
the
mozga-intel/Mul_mkldnn
branch
from
f24fec2
to
b768319
Compare
mozga-intel
force-pushed
the
mozga-intel/Mul_mkldnn
branch
from
b768319
to
6e7b883
Compare
| "bias_attr": bias_attr | ||
| "x_num_col_dims": num_flatten_dims, | ||
| "y_num_col_dims": 1, | ||
| "use_mkldnn": use_mkldnn |
There was a problem hiding this comment.
why use mul_mkldnn_op to replace fc_mkldnn_op here?
In my mind, the performance of fc_mkldnn_op is better than mul_mkldnn_op + elementwise_add_op in fluid.
There was a problem hiding this comment.
Yes I must agree with you, but the fully connected layer of mkldnn has a little problem with the input of matrices, What means that in the case when the input of layer is greater than one the mkldnn operator doesn't work correctly. Therefore we hit upon a good idea to remove this pitfall and withdraw the fully connected operator. Moreover this simple solution give us an opportunity to run the new neural network models.
There was a problem hiding this comment.
Do you mean that you will not use fc_mkldnn_op in fluid? Maybe you can add a comment in nn.py.
There was a problem hiding this comment.
With a heavy heart, I must say, yes, this operator is removed but for a while. I guess that in the future we will come back to the fully connected layer for a mkldnn.
| auto desc = tf_desc<Type>(args...); | ||
| return tf_pd<Type>(desc, e, p); | ||
| } | ||
|
|
There was a problem hiding this comment.
Are line 36-61 only used in mul_mkldnn_op? If these lines are not used in other ops, they should be moved into mul_mkldnn_op internel.
There was a problem hiding this comment.
At this moment, yes, but this code may be uses by the all mkldnn operators.
|
@tensor-tang Do you have any other comments about this PR? |
| using tf_desc = typename Type::desc; | ||
|
|
||
| template <class Type> | ||
| using tf_pd = typename Type::primitive_desc; |
There was a problem hiding this comment.
A small question, why named tf_pd here?
There was a problem hiding this comment.
tf = type_function.
No description provided.