Exposing use_cudnn #7536
Exposing use_cudnn #7536
Conversation
chengduoZH
force-pushed
the
feature/refine_conv_pool_python
branch
from
839a4ee
to
b7f1a75
Compare
chengduoZH
force-pushed
the
feature/refine_conv_pool_python
branch
from
b7f1a75
to
251c603
Compare
|
@chengduoZH We need an issue to describe the problem to be solved by this PR. |
|
@wangkuiyi Thanks! |
python/paddle/v2/fluid/nets.py
Outdated
| pool_stride=1, | ||
| pool_type=None): | ||
| pool_type=None, | ||
| pool_use_cudnn=True): |
There was a problem hiding this comment.
Just use_cudnn, as the last argument.
typhoonzero
dismissed
their stale review
Agree with @dzhwinter's comments should be fixed.
| @@ -70,6 +70,13 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { | |||
| framework::OpKernelType ConvOp::GetExpectedKernelType( | |||
| const framework::ExecutionContext& ctx) const { | |||
| bool use_cudnn = ctx.Attr<bool>("use_cudnn"); | |||
| use_cudnn &= platform::is_gpu_place(ctx.GetPlace()); | |||
There was a problem hiding this comment.
I think use_cudnn means that user wants to force use cudnn operator, if env does not support cudnn, the framework should throw an error but not implicit change to use other operator.
There was a problem hiding this comment.
Agree with the comment.
But we need a default policy right now, by default, we need to set use_cudnn True.
As we discussed offline, we reach the agreement that place is also a user configuration. we need to overwrite our default policy, not user specified.
There was a problem hiding this comment.
ok, I get the point, seems we still need a priority mechanism to resolve such kind of problem in the future.
fix #7582
What this PR does includes:
use_cudnnis setTrue.use_cudnn = use_cudnn && is_gpu_place(ctx.GetPlace());