harrisonGPU wrote:
> The solution looks alright to me, but I wonder if this could be implemented
> by hoisting the accumulator instead and then relying on the existing constant
> folding to bring the constants down. That is:
>
> ```
> x := amdgcn.{s,u}dot{2,4,8}(a, b, C)
> y := x +/- K
> ->
> x := C +/- K
> y := dot(a, b, x)
> ```
>
> This would avoid us needing to compute +/- and instead just hoist for all
> OpCodes we know are well-behaved w.r.t. this, so it would be extendible in
> the future (if it would ever make sense.)
>
> I suppose the biggest question (aside from correctness) is whether this would
> make sense to do always, or only if we know both sides are constants.
Yes, I’m planning to implement this optimization as well. For this first step,
I wanted to limit the change to constant folding and keep the patch small and
easy to review. As a follow up, I plan to handle non constant cases when the
transformation is profitable, such as folding an add into the accumulator when
the original accumulator is zero.
https://github.com/llvm/llvm-project/pull/225002
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
