[ROCM] added a cudnn switch of conv2d for rocm platform #31836
Conversation
|
Thanks for your contribution! |
python/paddle/nn/layer/conv.py
Outdated
There was a problem hiding this comment.
这里写成类似这样的形式
use_cudnn = True if (core.is_compiled_with_cuda() and
cudnn_version is not None) else False
python/paddle/nn/layer/conv.py
Outdated
There was a problem hiding this comment.
加一行注释，解释一下这里这么修改的原因，例如
// add environment of ROCM_CONV2D_DISABLE_CUDNN to disable use_cudnn in Conv2D
// If use_cudnn is not defined, default value is True, then self._use_cudnn = self._use_cudnn
qili93
left a comment
qili93
left a comment
There was a problem hiding this comment.
LGTM, api test will be covered by CPU CI and CUDA CI
python/paddle/fluid/layers/nn.py
Outdated
There was a problem hiding this comment.
If ROCM_CONV2D_DISABLE_CUDNN is not a built-in variable in ROCM but created by Paddle, I suggest following the naming rule like other environments like "FLAGS_xxx_xxx".
There was a problem hiding this comment.
Plus, it's better to get the value by fluid.get_flags instead of os.getenv
ronny1996
force-pushed
the
rocm-dev
branch
2 times, most recently
from
e1cf412 to
ae72bcd
Compare
zhiqiu
left a comment
zhiqiu
left a comment
There was a problem hiding this comment.
LGTM for paddle.fluid.layers.conv2d
3 participants
PR types
Others
PR changes
Others
Describe
added a cudnn switch of conv2d for rocm platform