Add support for 3D and 2D grouped conolutions #33
Conversation
nithinsubbiah
force-pushed
the
conv_3d
branch
from
a31f546 to
0bdd955
Compare
| @@ -33,30 +35,36 @@ class ConvConfig: | |||
| Q: int | |||
| F: int | |||
| S: int | |||
| is_grouped_conv: bool | |||
| G: int # group count | |||
| is_3D_conv: bool | |||
There was a problem hiding this comment.
possible to make another class instead for conv3d? That way every time we try to add a problem, we don't need the False, -1, -1, -1 part.
There was a problem hiding this comment.
I'll suggest another option, which would be to give these parameters some default values and put them at the end of the __init__ function. Then we don't need an extra class and the conv2d configs don't need to add these fields.
It's probably a good idea to get rid of these classes in favor of using python bindings at some point anyway, so I see the config classes a temporary implementation. Linalg ops are relatively stable, so it has not caused us any maintenance issues yet, but the bindings are much more resilient to changes to IR assembly format. We are making the same transition on the tuner side, right now, because it is very dependent on codegen dialects (which are very in flux right now).
Eventually we can get rid of these separate config classes, and just have a single config class that is used across all kernels (gemm, attention, conv, etc.) that just has functions to build the desired kernel types, and track things like peak flops, arithmetic intensity, etc. Then we can get rid of all these classes, and move everything to a shared benchmarking implementation.
| @@ -33,30 +35,36 @@ class ConvConfig: | |||
| Q: int | |||
| F: int | |||
| S: int | |||
| is_grouped_conv: bool | |||
| G: int # group count | |||
| is_3D_conv: bool | |||
There was a problem hiding this comment.
I'll suggest another option, which would be to give these parameters some default values and put them at the end of the __init__ function. Then we don't need an extra class and the conv2d configs don't need to add these fields.
It's probably a good idea to get rid of these classes in favor of using python bindings at some point anyway, so I see the config classes a temporary implementation. Linalg ops are relatively stable, so it has not caused us any maintenance issues yet, but the bindings are much more resilient to changes to IR assembly format. We are making the same transition on the tuner side, right now, because it is very dependent on codegen dialects (which are very in flux right now).
Eventually we can get rid of these separate config classes, and just have a single config class that is used across all kernels (gemm, attention, conv, etc.) that just has functions to build the desired kernel types, and track things like peak flops, arithmetic intensity, etc. Then we can get rid of all these classes, and move everything to a shared benchmarking implementation.
No description provided.