[Opt depthwise_conv2d] Simplify depthwise_conv2d use_cudnn attribute (#…

…48010) * simplify depthwise_conv2d phi kernel selection * fix depthwise_conv2d
PaddlePaddle · Nov 16, 2022 · 7c30458 · 7c30458
1 parent 8e6315e
commit 7c30458
Show file tree

Hide file tree

Showing 3 changed files with 7 additions and 9 deletions.
diff --git a/paddle/phi/api/yaml/legacy_backward.yaml b/paddle/phi/api/yaml/legacy_backward.yaml
@@ -391,7 +391,7 @@
   optional : mask
 
 - backward_op : depthwise_conv2d_double_grad
-  forward : depthwise_conv2d_grad (Tensor input, Tensor filter, Tensor grad_out, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format, bool use_gpudnn) -> Tensor(grad_input), Tensor(grad_filter)
+  forward : depthwise_conv2d_grad (Tensor input, Tensor filter, Tensor grad_out, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(grad_input), Tensor(grad_filter)
   args : (Tensor input, Tensor filter, Tensor grad_out, Tensor grad_input_grad, Tensor grad_filter_grad, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format)
   output : Tensor(input_grad), Tensor(filter_grad), Tensor(grad_out_grad)
   infer_meta :
@@ -402,16 +402,16 @@
   optional : grad_input_grad, grad_filter_grad
 
 - backward_op : depthwise_conv2d_grad
-  forward : depthwise_conv2d (Tensor input, Tensor filter, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format, bool use_gpudnn) -> Tensor(out)
-  args : (Tensor input, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format, bool use_gpudnn)
+  forward : depthwise_conv2d (Tensor input, Tensor filter, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
+  args : (Tensor input, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format)
   output : Tensor(input_grad), Tensor(filter_grad)
   infer_meta :
     func : GeneralBinaryGradInferMeta
     param : [input, filter]
   kernel :
     func : depthwise_conv2d_grad
     param : [input, filter, out_grad, strides, paddings, padding_algorithm, groups, dilations, data_format]
-    use_gpudnn : use_gpudnn
+    use_gpudnn : True
   backward : depthwise_conv2d_double_grad
 
 - backward_op : depthwise_conv2d_transpose_grad

diff --git a/paddle/phi/api/yaml/legacy_ops.yaml b/paddle/phi/api/yaml/legacy_ops.yaml
@@ -541,15 +541,15 @@
   backward : deformable_conv_grad
 
 - op : depthwise_conv2d
-  args : (Tensor x, Tensor filter, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format, bool use_gpudnn)
+  args : (Tensor x, Tensor filter, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format)
   output : Tensor(out)
   infer_meta :
     func : DepthwiseConvInferMeta
     param : [x, filter, strides, paddings, padding_algorithm, groups, dilations, data_format]
   kernel :
     func : depthwise_conv2d
     param : [x, filter, strides, paddings, padding_algorithm, groups, dilations, data_format]
-    use_gpudnn : use_gpudnn
+    use_gpudnn : true
   backward : depthwise_conv2d_grad
 
 - op : depthwise_conv2d_transpose

diff --git a/python/paddle/nn/functional/conv.py b/python/paddle/nn/functional/conv.py
@@ -172,7 +172,6 @@ def _conv_nd(
             groups,
             dilation,
             data_format,
-            use_cudnn,
         )
         if bias is not None:
             channel_dim = (
@@ -484,7 +483,7 @@ def conv1d(
                 conv2d_data_format,
             )
         else:
-            out = getattr(_C_ops, l_type)(
+            out = _C_ops.depthwise_conv2d(
                 x,
                 weight,
                 stride,
@@ -497,7 +496,6 @@ def conv1d(
                 -1,
                 False,
                 False,
-                use_cudnn,
             )
         if bias is not None:
             out = nn.elementwise_add(out, bias, axis=channel_dim)