SebastianBoblest commented on code in PR #12353:
URL: https://github.com/apache/tvm/pull/12353#discussion_r942613452
##########
python/tvm/relay/op/contrib/cmsisnn.py:
##########
@@ -122,6 +130,25 @@ def check_qnn_conv2d(pattern):
conv2d_input = conv2d.args[0]
conv2d_weight = conv2d.args[1]
+ # check if sum of paddings from pad() and conv2d() satisfies CMSIS-NN
constraints
+ can_pad_be_fused = True
+ if isinstance(conv2d_input, tvm.relay.expr.Call) and
str(conv2d_input.op.name) == "nn.pad":
+ data_layout = conv2d.attrs.data_layout
+ conv2d_padding = conv2d.attrs.padding # (top, left, bottom, right)
+ pad = conv2d_input
+ pad_width = pad.attrs.pad_width
+ pad_width_w = pad_width[data_layout.find("W")] # (left, right)
+ pad_width_h = pad_width[data_layout.find("H")] # (top, bottom)
+ # calculate effective padding post pad fusion
+ pad_top = pad_width_h[0] + conv2d_padding[0]
+ pad_bottom = pad_width_h[1] + conv2d_padding[2]
+ pad_left = pad_width_w[0] + conv2d_padding[1]
+ pad_right = pad_width_w[1] + conv2d_padding[3]
+ # check if difference in the side paddings is 1 along each
dimension
+ pad_w_diff = int(pad_right - pad_left)
Review Comment:
> They are IntImm. When they are used in conditional checks, it returns
IntImm in place of boolean which is what we'd expect normally. We have seen
this in the past too:
>
>
https://github.com/apache/tvm/blob/22ba659438a317ca59c8201430c662f86e2550fd/python/tvm/relay/op/contrib/cmsisnn.py#L134
Thanks for the explanation!
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
