You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
One possible solution would be to improve simplify_algebra to add two loops. The first is to check for horizontal fusions, and the second is to rewrite expressions.
The scalar multiply may be standalone after this, so find_unary_shape_transforms would need to be tweaked to support this as well.
And we may need to add an exception to find_mul_add to skip the rewrite if the input is scalar and feeds into a gemm or convolution.
The text was updated successfully, but these errors were encountered:
In SD clip, there is an opportunity to fuse all the add kernels:
Here the
mul_add
kernel is actually a scalar multiply + add:One possible solution would be to improve
simplify_algebra
to add two loops. The first is to check for horizontal fusions, and the second is to rewrite expressions.The scalar multiply may be standalone after this, so
find_unary_shape_transforms
would need to be tweaked to support this as well.And we may need to add an exception to
find_mul_add
to skip the rewrite if the input is scalar and feeds into a gemm or convolution.The text was updated successfully, but these errors were encountered: