[llvm-dev] [RFC] Goals for VPlan-based cost modelling

Fri Oct 30 04:38:23 PDT 2020

Hi,

> On Oct 30, 2020, at 10:08, Anna Sophia Welker via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> 
> Hi all, 
> 
> I am looking into the benefits of a VPlan-based cost model, and into how such a cost model should be implemented to make the most out of these benefits. 
> 
> Over the last year I have been working with LLVM, mainly focused on the ARM backend, in the course of a one-year internship at Arm Ltd. My main project from December 2019 to August 2020 was to introduce gather/scatters for MVE auto-vectorization. One of the recurring challenges during this work was to get things right in the cost model. 
> For example, gathers can extend the data they load, while scatters can truncate their input, meaning that an extend following the load, or a truncate preceding the store, is for free if it meets certain type conditions. As the current cost model is not designed for context-based analysis, this was a pain to model. 
> 

One goal is a more modular approach to cost-modeling and transformations. The current cost-model is quite monolithic and keeps some context information in various tables, which is hard to reason about and also hard to change and extend. One idea with VPlan is that cost decisions could be materialized in the VPlans directly by transformations working in concert to find the best overall strategy.

I think the example with gathers that extend the loaded value illustrates that quite well. In the legacy cost model, you could maintain a set of values that are free to extend and add the loaded values to it. When visiting a zext, return cost of zero for those.

With VPlan, instead you could add a new VPInstruction opcode ZextLoad and have a transformation that replaces all `zext (load)` instructions with the new ZExtLoad one. The cost-model needs to be taught about this special instruction and how much it costs. Then you could apply this transformation to an existing plan and check if the overall cost improved.

Of course there’s also the issue of how to generalize the TTI APIs to allow for computing the cost of instructions with more context.

> I have done some research and found out that one of the proposed benefits of VPlan is that a new cost model making use of it would be able to better support context-dependent decisions like this. 
> However, there does not exist much specification about what such a cost model should look like. 
> 
> Also, I have read through the respective code to understand how loop vectorization is currently done and how far the introduction of VPlan has progressed and have realised that the only recipe that actually handles more than one instruction from the input IR is the one for interleaved groups. When the VPlan is generated on the VPlan-native path, every IR instruction is considered and transformed into a recipe separately, ignoring its context (to give a code location, I am looking at VPlanTransforms::VPInstructionsToVPRecipes). 
> And maybe there are architectures that for some cases do not have the same vector instructions, so a pattern that works great for one could be useless for others. So I am wondering: Is there any plan to have target-dependent flavours of recipes, or how will those things be handled?

I think we probably want to keep things as generic as possible, try to model generic concepts and use TTI to decide whether to use it or not (e.g. see how masked loads/stores are handled).

> 
> Right now it makes sense that nothing like this has been implemented yet, as there is no cost model that could guide the transformation. But if recipes are a general thing, should the cost model be the component actually providing the target-specific pattern for a recipe, together with its cost? 
> 

I am not sure what patterns specifically you are thinking about, but I think the cost model should just evaluate the cost of a given plan and not provide anything beyond that. Of course, this still can mean that there might be certain recipes/instructions that are not available/profitable on some targets and we decide to never generate them, based on the cost-information.

> I am considering choosing a topic related to VPlan, possibly cost modelling, for my Master thesis, with the goal to present a solution and implement a prototype. 
> 

I am hoping to make some progress on this in the next months (hopefully the work on modeling the def-use chains between recipes in VPlan will be wrapped up soon) and I expect there to be a few moving parts. Not sure what that means for a master thesis in this area.

Cheers,
Florian
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20201030/2590d8c3/attachment.html>