[cfe-dev] [OpenCL] Representing kernel attributes by LLVM target-dependent attribute-value pairs instead of metadata

Thu May 12 15:47:13 PDT 2016

Actually provided that we can have more than one metadata attachment on a function it should not be different from attributes.
But I have also noticed dropUnknownMetadata() implementation in r235783. I believe we need to be sure this info is not just dropped like it can now happen with metadata attachments on instruction. In this respect attributes seem better, or we will need to make our specific metadata known to llvm.

Stas

From: Anastasia Stulova [mailto:Anastasia.Stulova at arm.com]
Sent: Thursday, May 12, 2016 8:31 AM
To: Mekhanoshin, Stanislav <Stanislav.Mekhanoshin at amd.com>; Liu, Yaxun (Sam) <Yaxun.Liu at amd.com>; cfe-dev (cfe-dev at lists.llvm.org) <cfe-dev at lists.llvm.org>; Pan, Xiuli <xiuli.pan at intel.com>; Bader, Alexey (alexey.bader at intel.com) <alexey.bader at intel.com>
Cc: Stellard, Thomas <Tom.Stellard at amd.com>; Sumner, Brian <Brian.Sumner at amd.com>; nd <nd at arm.com>
Subject: RE: [OpenCL] Representing kernel attributes by LLVM target-dependent attribute-value pairs instead of metadata

I feel that both solutions offer similar functionality with slightly different APIs. Just my general feeling is that these are rather properties of an application and have nothing to do with target specific details.

> The only problem I see here if someone else would want to use function metadata attachment for some other reason, so we will need to search and/or merge it.
@Stas, would it not be the same if other target-dependent attributes are attached too?

> The function metadata also requires a commonly agreed attribute name. In this aspect it is similar to target-dependent attribute-value pair. However it may be more difficult to manipulate.
@Sam, could you provide more details where attributes would be more beneficial to use?

Thanks,
Anastasia

From: Mekhanoshin, Stanislav [mailto:Stanislav.Mekhanoshin at amd.com]
Sent: 11 May 2016 19:39
To: Liu, Yaxun (Sam); Anastasia Stulova; cfe-dev (cfe-dev at lists.llvm.org<mailto:cfe-dev at lists.llvm.org>); Pan, Xiuli; Bader, Alexey (alexey.bader at intel.com<mailto:alexey.bader at intel.com>)
Cc: Stellard, Thomas; Sumner, Brian; nd
Subject: RE: [OpenCL] Representing kernel attributes by LLVM target-dependent attribute-value pairs instead of metadata

Sam,

I guess metadata attached right to the function should work as well. Nothing would hold us from forming the same list of strings in a single metadata entry. The only problem I see here if someone else would want to use function metadata attachment for some other reason, so we will need to search and/or merge it.
Given that consideration attributes looks better to me, they are already a top-level list.

Stas

From: Liu, Yaxun (Sam)
Sent: 11 мая 2016 г. 10:18
To: Anastasia Stulova <Anastasia.Stulova at arm.com<mailto:Anastasia.Stulova at arm.com>>; Mekhanoshin, Stanislav <Stanislav.Mekhanoshin at amd.com<mailto:Stanislav.Mekhanoshin at amd.com>>; cfe-dev (cfe-dev at lists.llvm.org<mailto:cfe-dev at lists.llvm.org>) <cfe-dev at lists.llvm.org<mailto:cfe-dev at lists.llvm.org>>; Pan, Xiuli <xiuli.pan at intel.com<mailto:xiuli.pan at intel.com>>; Bader, Alexey (alexey.bader at intel.com<mailto:alexey.bader at intel.com>) <alexey.bader at intel.com<mailto:alexey.bader at intel.com>>
Cc: Stellard, Thomas <Tom.Stellard at amd.com<mailto:Tom.Stellard at amd.com>>; Sumner, Brian <Brian.Sumner at amd.com<mailto:Brian.Sumner at amd.com>>; nd <nd at arm.com<mailto:nd at arm.com>>
Subject: RE: [OpenCL] Representing kernel attributes by LLVM target-dependent attribute-value pairs instead of metadata

Hi Anastasia,

Target-dependent attribute-value pair is not limited to one target. It is just a pair of strings attached to a function. It can be used by different targets as far as they accept a commonly agreed attribute name.

The function metadata also requires a commonly agreed attribute name. In this aspect it is similar to target-dependent attribute-value pair. However it may be more difficult to manipulate.

Stas, any comments on this?

Thanks.

Sam

From: Anastasia Stulova [mailto:Anastasia.Stulova at arm.com]
Sent: Wednesday, May 11, 2016 12:32 PM
To: Mekhanoshin, Stanislav <Stanislav.Mekhanoshin at amd.com<mailto:Stanislav.Mekhanoshin at amd.com>>; Liu, Yaxun (Sam) <Yaxun.Liu at amd.com<mailto:Yaxun.Liu at amd.com>>; cfe-dev (cfe-dev at lists.llvm.org<mailto:cfe-dev at lists.llvm.org>) <cfe-dev at lists.llvm.org<mailto:cfe-dev at lists.llvm.org>>; Pan, Xiuli <xiuli.pan at intel.com<mailto:xiuli.pan at intel.com>>; Bader, Alexey (alexey.bader at intel.com<mailto:alexey.bader at intel.com>) <alexey.bader at intel.com<mailto:alexey.bader at intel.com>>
Cc: Stellard, Thomas <Tom.Stellard at amd.com<mailto:Tom.Stellard at amd.com>>; Sumner, Brian <Brian.Sumner at amd.com<mailto:Brian.Sumner at amd.com>>; nd <nd at arm.com<mailto:nd at arm.com>>
Subject: RE: [OpenCL] Representing kernel attributes by LLVM target-dependent attribute-value pairs instead of metadata

Hi Sam,

Interesting idea, I am just concerned whether target-dependent attribute-value pairs are really correct for this use case. Because kernel attributes aren't really target dependent.

However, I agree that current approach is painful enough. Do you think switching to function metadata (see llvm at r235783) would help instead? This way we will also have the kernel attributes directly accessible from the function, simplifying querying them (as compared to what we have at the moment).

Thanks,
Anastasia

From: Mekhanoshin, Stanislav [mailto:Stanislav.Mekhanoshin at amd.com]
Sent: 09 May 2016 21:14
To: Liu, Yaxun (Sam); cfe-dev (cfe-dev at lists.llvm.org<mailto:cfe-dev at lists.llvm.org>); Anastasia Stulova; Pan, Xiuli; Bader, Alexey (alexey.bader at intel.com<mailto:alexey.bader at intel.com>)
Cc: Stellard, Thomas; Sumner, Brian
Subject: RE: [OpenCL] Representing kernel attributes by LLVM target-dependent attribute-value pairs instead of metadata

It will make our life much easier I would say.

Stas

From: Liu, Yaxun (Sam)
Sent: Monday, May 09, 2016 1:05 PM
To: cfe-dev (cfe-dev at lists.llvm.org<mailto:cfe-dev at lists.llvm.org>) <cfe-dev at lists.llvm.org<mailto:cfe-dev at lists.llvm.org>>; 'anastasia.stulova at arm.com' <anastasia.stulova at arm.com<mailto:anastasia.stulova at arm.com>>; Pan, Xiuli <xiuli.pan at intel.com<mailto:xiuli.pan at intel.com>>; Bader, Alexey (alexey.bader at intel.com<mailto:alexey.bader at intel.com>) <alexey.bader at intel.com<mailto:alexey.bader at intel.com>>
Cc: Stellard, Thomas <Tom.Stellard at amd.com<mailto:Tom.Stellard at amd.com>>; Sumner, Brian <Brian.Sumner at amd.com<mailto:Brian.Sumner at amd.com>>; Mekhanoshin, Stanislav <Stanislav.Mekhanoshin at amd.com<mailto:Stanislav.Mekhanoshin at amd.com>>
Subject: [OpenCL] Representing kernel attributes by LLVM target-dependent attribute-value pairs instead of metadata

Hi,

When SPIR 1.2/2.0 spec was developed, there was no target-dependent attribute-value pair in LLVM. As such, kernel attributes were represented as metadata.

This caused lots of inconvenience since these metadata are difficult to manipulate/change when kernel functions went through transformations in backends.

Now LLVM supports target-dependent attribute-value pair, I am wondering whether it is time to use that to represent kernel attributes.

e.g. something like (just to give an idea, not exact llvm IR)

kernel void foo(global int*a, sampler_t s); #1
#1 = {
reqd_work_group_size="1 1 1",
kernel_arg_type="int *, sampler_t",
...
}

basically we keep the info conveyed by the original metadata but attach them to the kernel function as attribute/value pair. This will make these attributes much easier to manipulate/change.

Any feedbacks? Thanks.

Sam
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20160512/00ebfdb5/attachment.html>