[Mlir-commits] [mlir] [MLIR][XeGPU] Scattered ops sg-to-wi distribution (PR #154949)

[Charitha Saumya]

charithaintc wrote:

> > Before.
> > ```
> > %offsets = arith.constant dense<0> : vector<16xindex>
> > // insert any value to this vector (random, linear does not matter)
> > %v = xegpu.load %base [%offset] : i64, vector<16xindex> -> vector<16xf16>
> > ```
> 
> But should we care about the vector producer? Arith distribution is there, but there are more ways to create a vector, we could even receive it as an argument from the runtime. Do I miss the op or distribution logic constraint that the offsets producer must be retrievable?

"even receive it as an argument from the runtime." what does this mean? a func argument?

AFAIK, from the load gather distribution perspective, only thing we need to care about is what are the distributed types for the base, offsets and masks. Everything else should ideally be handled by the framework. 

I suggest, testing it with distributed type as <1xindex> and see what the framework does. 

https://github.com/llvm/llvm-project/pull/154949