llvm dev - Jan 2021 - [Attributor] Using Attributor framework for AMDGPU backend

Hi Johannes,
I am trying to use the Attributor framework for AMD GPU backend. We have an
attributor called amdgpu-flat-work-group-size which basically specifies the
min and max work group configuration for the kernel. Typically, this
attribute is given on kernels but it can also be present on device
functions. It needs to be propagated across the call graph with a merge
function which does "max" of the all incoming attributes of type
amdgpu-flat-work-group-size.

How do I write an Attributor pass for this?

I do understand that I need to model the attribute as a class (like
AANoUnwind) but I am not sure if I totally understand the significance of
pessimistic/optimistic fixpoint for this attribute.


-- 
*Disclaimer: Views, concerns, thoughts, questions, ideas expressed in this
mail are of my own and my employer has no take in it. *
Thank You.
Madhur D. Amilkanthwar
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210105/e1a545bf/attachment.html>

Hi Madhur,

did you watch the introduction [0] and attributor tutorial [1]?
There the fixpoint stuff is explained.

If you have a single number per function and want to do "max"
propagation,
you should take a look at AANoUnwind (or AAWillReturn) for the function
deduction part and AAAlign for the use of an integer state that is 
decreasing.
Since you want the opposite, an integer state that is increasing as it 
progresses,
you need to use a `DecIntegerState`  where AAAlign is using an
`IncIntegerState`. Also, consider [2] which adds some helpers to the state
you want to use and also uses it for a "max" join operation.

I hope this helps to get you started, feel free to reply if you get stuck.

~ Johannes


[0] https://youtu.be/CzWkc_JcfS0
[1] https://youtu.be/HVvvCSSLiTw
[2] https://reviews.llvm.org/D87978


On 1/4/21 1:03 PM, Madhur Amilkanthwar wrote:
> Hi Johannes,
> I am trying to use the Attributor framework for AMD GPU backend. We have an
> attributor called amdgpu-flat-work-group-size which basically specifies the
> min and max work group configuration for the kernel. Typically, this
> attribute is given on kernels but it can also be present on device
> functions. It needs to be propagated across the call graph with a merge
> function which does "max" of the all incoming attributes of type
> amdgpu-flat-work-group-size.
>
> How do I write an Attributor pass for this?
>
> I do understand that I need to model the attribute as a class (like
> AANoUnwind) but I am not sure if I totally understand the significance of
> pessimistic/optimistic fixpoint for this attribute.
>
>