-
Notifications
You must be signed in to change notification settings - Fork 1k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
benchdnn: graph: enhance input displace and shape rewrite for linked attribute and shapes #2354
Conversation
4af44a1
to
d65e846
Compare
c82a899
to
b6a63b5
Compare
make test |
0c5a191
to
6b3be02
Compare
6b3be02
to
fdb56d6
Compare
fdb56d6
to
05022d1
Compare
make test |
05022d1
to
a3d94cd
Compare
a3d94cd
to
f8ef788
Compare
make test |
f8ef788
to
c992118
Compare
else if (base_op_ref.kind_ == "ConvTranspose") | ||
bia_dt = convert_dt(base_op_ref.in_lts_[2].get_data_type()); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think the logic should be a little bit different:
else if (is_integral(dt)) {
// Or maybe just f32 instead?
bia_dt = convert_dt(base_op_ref.in_lts_[2].get_data_type());
} else {
bia_dt = dt;
}
The limitation to not support s8/u8/s32 bias is uniform across the library and not limited to a specific op.
c992118
to
8096ced
Compare
make test |
Description
The PR enhances input displacement and rewrite functionality in benchdnn graph for the following aspects:
MatMul
and scale and zp ofDynamicDequantize
to support SDPA patterns rewriting.group_shape
and scale/zp input ofDynamicDequantize
. If user provides shapes for one of the attributes or input shapes, benchdnn graph will update the other accordingly after performing some checks.For example: