-
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: introduce --bia-dt option for drivers supporting bias (fixed MFDNN-3031, fixed MFDNN-12936) #2335
base: main
Are you sure you want to change the base?
Conversation
make test |
@@ -21,7 +21,7 @@ where *matmul-knobs* are: | |||
specification for `src`, `weights`, and `dst` tensors through | |||
strides values. Refer to [option documentation](knob_strides.md) | |||
for details. | |||
- `--bia_dt={undef [default], f32, s32, s8, u8}` -- bias data type. | |||
- `--bia-dt={undef [default], f32, s32, s8, u8}` -- bias 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.
Is there any particular reason we aren't just fully spelling out this argument as --bias-dt
? Shortening bias
to bia
just seems like it creates naming complexity for minimal benefit.
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.
The path of respecting the legacy... :)
But I'm fine to rename it in case we get some up votes.
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 believe we historically used fully spelled names in API (except for src/dst for some reason), and 3 letter shortening in benchdnn (src/dst/wei/bia). So unless we want to fully spell other names as well in benchdnn (e.g. wei -> weights
), I would vote to keep it as is.
d766266
to
befeb28
Compare
make test |
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.
The .md files look good to me, thanks!
befeb28
to
c1228bd
Compare
make test |
3100fd1
to
4d6c004
Compare
make test |
752361e
to
fa70bfb
Compare
make test |
fa70bfb
to
25804ee
Compare
This file is a standard for Python tools. Described here: https://peps.python.org/pep-0518/#file-format
25804ee
to
c907431
Compare
make test |
MFDNN-3031
MFDNN-12936
To keep the compatibility, the default bias data type is
undef
, which means no bias is specified. This is aligned with currentFWD_D
andFWD_I
settings.FWD_B
value will update thebia-dt
value tof32
which is also compatible with previous behavior.At some point in v3.8 both
FWD_B
andBWD_WB
values for the--dir
option will be marked as deprecated and issue the corresponding warning.To specify some data type for bias,
FWD_D
orFWD_I
should be used:Feel free to ignore matmul input files changes as it's a straight rename from
--bia_dt
to--bia-dt
.