Rename unroll_n_to_subgroups to subgroups_n

[bjacob]

"Unroll" usually means "generate more instructions", so the terminology being changed here, unroll_n_to_subgroups, created confusion.

[qedawkins]

distribute_to_subgroups might be more canonical w.r.t. naming in the rest of the compiler.

[bjacob]

Interesting. I parsed "distribute to subgroups" as the action of distributing to the preestablished number of subgroups, while here, the effect is to set/increase the number of subgroups. WDYT?

[qedawkins]

Hmmm I parsed the field as "this dimension will be distributed to subgroups" as an analog to "this dimension will be unrolled to more instructions"

[bjacob]

Yeah but the nuance is that this field is what causes this dimension to be created in the first place.

[qedawkins]

I see, so really we have two distinct names occupying the same field of the attribute. Before MaterializeEncodings it is a yet-to-be introduced dimension and is expanding to subgroups like you're saying. After MaterializeEncodings this becomes a concrete dimensions that is going to be distributed.

I would think that the name of the field in the DataTiledMMAAttribute, since that attribute is only materialized in IR after MaterializeEncodings, should be distribute_to_subgroups, whereas if we want we can make the names of C++ referencing subgroups expand_to_subgroups. That's my intuition at least, but I won't be picky here if you insist on this name (because it's better than unroll).

[bjacob]

Before MaterializeEncoding, the attribute does not exist yet. MaterializeEncoding is where it is originally created.

iree/compiler/src/iree/compiler/Codegen/Common/GPU/GPUMaterializeEncoding.cpp

Line 210 in 915b06b

return DataTiledMMAAttr::get(ctx, intrinsicMma.getIntrinsic(), unrollM,

. That is also when the extra tensor dimensions are created: these two aspects are simultaneous, as MaterializeEncoding is a type-conversion pass, and that is necessary anyway to have well-formed multi_mma ops.

From MaterializeEncoding up until DistributeMmaToLanes, the DataTiledMMAAttr attribute carries those unroll_{m,n}_to_subgroups parameters. These are dropped at DistributeMmaToLanes, as the IR after that point is single-thread IR where these subgroup parameters are no longer relevant.

[bjacob]

Just discussed online with @qedawkins and we agreed on subgroups_m, subgroups_n.

[hanhanW]

Just discussed online with @qedawkins and we agreed on subgroups_m, subgroups_n.

Do you also need to update the PR description?