Page MenuHomePhabricator

dcaballe (Diego Caballero)
User

Projects

User does not belong to any projects.

User Details

User Since
Sep 16 2016, 11:47 AM (238 w, 1 d)

Recent Activity

Wed, Mar 17

dcaballe accepted D98530: [mlir][Vector] Pass for lowering of transfer_read/write.

Thanks, Sergei! LGTM. @nicolasvasilache, @aartbik, could we please have final approval from anybody from your side? Thanks!

Wed, Mar 17, 9:26 AM · Restricted Project

Tue, Mar 16

dcaballe added a comment to D98742: [mlir][amx] regression test for tile-muli (all zero/sign-extension combinations).

Thanks, Aart! LGTM. Just minor comments.

Tue, Mar 16, 2:34 PM · Restricted Project

Mon, Mar 15

dcaballe added a comment to D98530: [mlir][Vector] Pass for lowering of transfer_read/write.

Thanks for working on this, Sergei! Some comments.

Mon, Mar 15, 6:54 PM · Restricted Project

Fri, Mar 12

dcaballe added a comment to D98485: [mlir] fix a memory leak in NestedPattern.

Thanks, Alex!

Fri, Mar 12, 11:28 AM · Restricted Project
dcaballe accepted D98476: [mlir] Annotate functions used only in debug mode with LLVM_ATTRIBUTE_UNUSED.

Thanks!

Fri, Mar 12, 11:24 AM · Restricted Project
dcaballe added a comment to D98470: [mlir][amx] Add Intel AMX dialect (architectural-specific vector dialect).

Thanks for working on this, Aart! I think the progressive lowering approach that you are taking here is very on point! I’m not working with AMX but it would be great if you could add me as a reviewer to the related code reviews. It would be very educational for me since this approach is also applicable to similar internal problems we have.

Fri, Mar 12, 8:47 AM · Restricted Project
dcaballe added a comment to D98476: [mlir] Annotate functions used only in debug mode with LLVM_ATTRIBUTE_UNUSED.

Thanks for addressing this issue! Would this work if we compiled in release mode and with asserts enabled?

Fri, Mar 12, 8:41 AM · Restricted Project

Mar 11 2021

dcaballe committed rG0fd0fb532993: Reland: [mlir][Affine][Vector] Add initial support for 'iter_args' to Affine… (authored by dcaballe).
Reland: [mlir][Affine][Vector] Add initial support for 'iter_args' to Affine…
Mar 11 2021, 3:10 PM
dcaballe committed rG96891f041850: Reland: [mlir][Vector][Affine] Improve affine vectorizer algorithm (authored by dcaballe).
Reland: [mlir][Vector][Affine] Improve affine vectorizer algorithm
Mar 11 2021, 2:21 PM
dcaballe committed rGed193bce9d3b: [mlir][Vector][Affine] Fix heap-use-after-free in vectorizer (authored by dcaballe).
[mlir][Vector][Affine] Fix heap-use-after-free in vectorizer
Mar 11 2021, 10:51 AM
dcaballe closed D98382: [mlir][Vector][Affine] Fix heap-use-after-free in vectorizer.
Mar 11 2021, 10:51 AM · Restricted Project
dcaballe added a comment to D97304: [mlir] Add polynomial approximation for math::LogOp (using builders API).

They are actually used but only inside assert, is it a release build when it is a warning?

Mar 11 2021, 10:41 AM · Restricted Project
dcaballe added inline comments to D98239: [MLIR][Affine] Add utility to check if the slice is valid.
Mar 11 2021, 9:04 AM · Restricted Project

Mar 10 2021

dcaballe requested review of D98382: [mlir][Vector][Affine] Fix heap-use-after-free in vectorizer.
Mar 10 2021, 6:49 PM · Restricted Project
dcaballe added a comment to rG95db7b4aeaad: [mlir][Vector][Affine] Improve affine vectorizer algorithm.

Sorry, Alex. Let me look into that.

Mar 10 2021, 11:31 AM
dcaballe accepted D97822: [mlir][Vector] Lowering of transfer_read/write to vector.load/store.

Let's address this in a separate patch since the vector dialect doesn't currently have the pass infrastructure, so it will be quite a bit of somewhat unrelated code.

Mar 10 2021, 11:13 AM · Restricted Project
dcaballe committed rG77a9d1549fcc: [mlir][Affine][Vector] Add initial support for 'iter_args' to Affine vectorizer. (authored by dcaballe).
[mlir][Affine][Vector] Add initial support for 'iter_args' to Affine vectorizer.
Mar 10 2021, 11:02 AM
dcaballe closed D97892: [mlir][Affine][Vector] Add initial support for 'iter_args' to Affine vectorizer..
Mar 10 2021, 11:02 AM · Restricted Project
dcaballe added a comment to D97304: [mlir] Add polynomial approximation for math::LogOp (using builders API).

It looks like elementType and isI32 methods in PolynomialApproximation.cpp are not used. Would it be possible to remove them to avoid the warnings?

Mar 10 2021, 11:01 AM · Restricted Project
dcaballe committed rG95db7b4aeaad: [mlir][Vector][Affine] Improve affine vectorizer algorithm (authored by dcaballe).
[mlir][Vector][Affine] Improve affine vectorizer algorithm
Mar 10 2021, 10:31 AM
dcaballe closed D97442: [mlir][Vector][Affine] Improve affine vectorizer algorithm.
Mar 10 2021, 10:31 AM · Restricted Project
dcaballe added a reviewer for D98239: [MLIR][Affine] Add utility to check if the slice is valid: sgrechanik.
Mar 10 2021, 10:05 AM · Restricted Project
dcaballe added a comment to D98239: [MLIR][Affine] Add utility to check if the slice is valid.

Thanks for fixing this bug, Vinayaka! The direction looks good to me. Adding some comments.

Mar 10 2021, 10:01 AM · Restricted Project

Mar 9 2021

dcaballe added a comment to D97892: [mlir][Affine][Vector] Add initial support for 'iter_args' to Affine vectorizer..

Thanks! I'll commit this tomorrow if no more comments.

Mar 9 2021, 2:37 PM · Restricted Project
dcaballe updated the diff for D97442: [mlir][Vector][Affine] Improve affine vectorizer algorithm.

Addressed feedback, thanks!
I'll commit this tomorrow if no more comments.

Mar 9 2021, 2:36 PM · Restricted Project

Mar 8 2021

dcaballe updated the diff for D97892: [mlir][Affine][Vector] Add initial support for 'iter_args' to Affine vectorizer..

Addressed feedback. Thanks!

Mar 8 2021, 10:20 PM · Restricted Project
dcaballe added inline comments to D97442: [mlir][Vector][Affine] Improve affine vectorizer algorithm.
Mar 8 2021, 9:38 PM · Restricted Project
dcaballe updated the diff for D97442: [mlir][Vector][Affine] Improve affine vectorizer algorithm.

Addressed feedback. Thanks!

Mar 8 2021, 9:37 PM · Restricted Project
dcaballe added a comment to D97822: [mlir][Vector] Lowering of transfer_read/write to vector.load/store.

Agreed on the blanket. But how about the simple folding away of unmasked (or all-true/false mask). In the long run, it makes a lot more sense to have one place with that logic (folding/canonicalizaiton) and have all other rewriting rule that expect unmasked transfers simply work on the l/s instead?

Mar 8 2021, 12:00 PM · Restricted Project

Mar 7 2021

dcaballe added a comment to D97822: [mlir][Vector] Lowering of transfer_read/write to vector.load/store.

For now, I would say that hiding the patterns behind a populate function (but not a canonicalization) would be the simplest way forward.

Mar 7 2021, 10:11 PM · Restricted Project
dcaballe added inline comments to D97820: [mlir] Add 'Skip' result to Operation visitor.
Mar 7 2021, 9:59 PM · Restricted Project

Mar 5 2021

dcaballe accepted D98056: [mlir][vector] canonicalize unmasked gather/scatter/compress/expand directly into l/s.

More ops going through vector loads/stores! Great! Thanks, Aart. LGTM.

Mar 5 2021, 2:10 PM · Restricted Project
dcaballe committed rG2de6dbda66b3: [mlir] Add 'Skip' result to Operation visitor (authored by dcaballe).
[mlir] Add 'Skip' result to Operation visitor
Mar 5 2021, 2:06 PM
dcaballe committed rG71a86245ca62: [mlir] Extend Operation visitor with pre-order traversal (authored by dcaballe).
[mlir] Extend Operation visitor with pre-order traversal
Mar 5 2021, 2:06 PM
dcaballe committed rGb635492c3f1a: [mlir][Affine][NFC] Return BlockArgument in AffineForOp::getInductionVar (authored by dcaballe).
[mlir][Affine][NFC] Return BlockArgument in AffineForOp::getInductionVar
Mar 5 2021, 2:06 PM
dcaballe closed D97820: [mlir] Add 'Skip' result to Operation visitor.
Mar 5 2021, 2:06 PM · Restricted Project
dcaballe closed D97217: [mlir] Extend Operation visitor with pre-order traversal.
Mar 5 2021, 2:06 PM · Restricted Project
dcaballe closed D97879: [mlir][Affine][NFC] Return BlockArgument in AffineForOp::getInductionVar.
Mar 5 2021, 2:06 PM · Restricted Project
dcaballe added a comment to D97822: [mlir][Vector] Lowering of transfer_read/write to vector.load/store.

In the sense that e.g. unmasked one becomes regular loads, yes. But AFAIK, this is very much in the spirit of MLIR, to canonicalize everything to the simplest form possible, and have patterns in the rewriting rules that expect only the simplest form possible to avoid cluttering the patterns with implicit canonicalization.

Mar 5 2021, 9:27 AM · Restricted Project
dcaballe added a comment to D97822: [mlir][Vector] Lowering of transfer_read/write to vector.load/store.

Don't we want to make some of these folders/canonicalization patterns? That way, everyone doing rewriting benefits automatically, and does not need to explicitly pull in these new specific rewriting patterns?

Mar 5 2021, 8:51 AM · Restricted Project

Mar 4 2021

dcaballe added a comment to D97820: [mlir] Add 'Skip' result to Operation visitor.

Thanks! I will commit it tomorrow if no more comments.

Mar 4 2021, 5:06 PM · Restricted Project
dcaballe updated the diff for D97820: [mlir] Add 'Skip' result to Operation visitor.

Addressed feedback.

Mar 4 2021, 5:06 PM · Restricted Project
dcaballe retitled D97820: [mlir] Add 'Skip' result to Operation visitor from [mlir] Add pruning to Operation visitor to [mlir] Add 'Skip' result to Operation visitor.
Mar 4 2021, 3:56 PM · Restricted Project
dcaballe added a comment to D97820: [mlir] Add 'Skip' result to Operation visitor.

Just to clarify, "prune" is ignored for post order walks?

Mar 4 2021, 3:54 PM · Restricted Project
dcaballe updated the diff for D97820: [mlir] Add 'Skip' result to Operation visitor.

Addressed feedback.

Mar 4 2021, 3:54 PM · Restricted Project

Mar 3 2021

dcaballe requested review of D97892: [mlir][Affine][Vector] Add initial support for 'iter_args' to Affine vectorizer..
Mar 3 2021, 3:16 PM · Restricted Project
dcaballe added inline comments to D97442: [mlir][Vector][Affine] Improve affine vectorizer algorithm.
Mar 3 2021, 12:13 PM · Restricted Project
dcaballe updated the diff for D97442: [mlir][Vector][Affine] Improve affine vectorizer algorithm.
  • Addressed Sergei's feedback.
  • Minor fix to 'iter_args' tests.
  • Minor changes to registration APIs.
Mar 3 2021, 12:13 PM · Restricted Project
dcaballe requested review of D97879: [mlir][Affine][NFC] Return BlockArgument in AffineForOp::getInductionVar.
Mar 3 2021, 12:01 PM · Restricted Project
dcaballe added a comment to D97822: [mlir][Vector] Lowering of transfer_read/write to vector.load/store.

Thanks for working on this, Sergei! Some initial comments!

Mar 3 2021, 10:00 AM · Restricted Project

Mar 2 2021

dcaballe added a reviewer for D97820: [mlir] Add 'Skip' result to Operation visitor: sgrechanik.
Mar 2 2021, 8:58 PM · Restricted Project
dcaballe requested review of D97820: [mlir] Add 'Skip' result to Operation visitor.
Mar 2 2021, 5:23 PM · Restricted Project
dcaballe updated the diff for D97217: [mlir] Extend Operation visitor with pre-order traversal.

Restoring previous approved version. I'll commit it if no more comments.
I'll create a separate review for the prune changes.

Mar 2 2021, 3:20 PM · Restricted Project
dcaballe added a comment to D97217: [mlir] Extend Operation visitor with pre-order traversal.

Can you separate out the prune functionality into a different commit?

Mar 2 2021, 8:42 AM · Restricted Project

Mar 1 2021

dcaballe added a comment to D97442: [mlir][Vector][Affine] Improve affine vectorizer algorithm.

Kind ping :)

Mar 1 2021, 9:06 PM · Restricted Project
dcaballe requested review of D97217: [mlir] Extend Operation visitor with pre-order traversal.
Mar 1 2021, 8:59 PM · Restricted Project
dcaballe updated the diff for D97217: [mlir] Extend Operation visitor with pre-order traversal.

This new diff introduces important changes to the previous version.
Hopefully, @rriddle/@frgossen/@bondhugula can take another look.

Mar 1 2021, 8:59 PM · Restricted Project

Feb 26 2021

dcaballe planned changes to D97217: [mlir] Extend Operation visitor with pre-order traversal.

Let me revisit a few things in this patch and come back.

Feb 26 2021, 7:00 PM · Restricted Project
dcaballe added a comment to D97422: [mlir][vector] add higher dimensional support to gather/scatter.

Thanks, Aart! I just replied to some comments. Nothing else from my side!
Maybe it would be good that somebody else from your side took a look, as well.

Feb 26 2021, 12:02 PM · Restricted Project

Feb 25 2021

dcaballe added a comment to D97422: [mlir][vector] add higher dimensional support to gather/scatter.

(sorry, replace "first dimension" with "last dimension" in my prev. comment :))

Feb 25 2021, 4:03 PM · Restricted Project
dcaballe added a comment to D97422: [mlir][vector] add higher dimensional support to gather/scatter.

Hey Aart!

Feb 25 2021, 11:20 AM · Restricted Project
dcaballe committed rGebca222b65cb: [mlir] Check 'iter_args' in 'isLoopParallel' utility (authored by dcaballe).
[mlir] Check 'iter_args' in 'isLoopParallel' utility
Feb 25 2021, 8:24 AM
dcaballe committed rG203d5eeec55b: [MLIR][affine-loop-fusion] Handle defining ops between the source and dest loops (authored by tungld).
[MLIR][affine-loop-fusion] Handle defining ops between the source and dest loops
Feb 25 2021, 8:24 AM
dcaballe closed D97347: [mlir] Check 'iter_args' in 'isLoopParallel' utility.
Feb 25 2021, 8:23 AM · Restricted Project
dcaballe closed D97030: [MLIR][affine-loop-fusion] Handle defining ops between the source and dest loops.
Feb 25 2021, 8:23 AM · Restricted Project

Feb 24 2021

dcaballe requested review of D97442: [mlir][Vector][Affine] Improve affine vectorizer algorithm.
Feb 24 2021, 10:47 PM · Restricted Project
dcaballe accepted D97030: [MLIR][affine-loop-fusion] Handle defining ops between the source and dest loops.

LGTM! Just minor comments. I think Uday planned to commit it. Otherwise, I could commit it tomorrow together with one of mine.

Feb 24 2021, 6:00 PM · Restricted Project
dcaballe added inline comments to D97252: [MLIR][affine] Certain Call Ops to prevent fusion.
Feb 24 2021, 11:25 AM · Restricted Project

Feb 23 2021

dcaballe updated the diff for D97347: [mlir] Check 'iter_args' in 'isLoopParallel' utility.

Addressing feedback.

Feb 23 2021, 10:50 PM · Restricted Project
dcaballe added a comment to D97347: [mlir] Check 'iter_args' in 'isLoopParallel' utility.

Does this patch handle iter_args in an inner loop? For example:

Feb 23 2021, 10:49 PM · Restricted Project
dcaballe added inline comments to D97217: [mlir] Extend Operation visitor with pre-order traversal.
Feb 23 2021, 10:34 PM · Restricted Project
dcaballe updated the diff for D97217: [mlir] Extend Operation visitor with pre-order traversal.
  • Passing order as an extra parameter.
  • Fixing typo.
Feb 23 2021, 10:33 PM · Restricted Project
dcaballe added a reviewer for D97347: [mlir] Check 'iter_args' in 'isLoopParallel' utility: sgrechanik.
Feb 23 2021, 4:54 PM · Restricted Project
dcaballe added a reviewer for D97217: [mlir] Extend Operation visitor with pre-order traversal: sgrechanik.
Feb 23 2021, 4:53 PM · Restricted Project
dcaballe requested review of D97347: [mlir] Check 'iter_args' in 'isLoopParallel' utility.
Feb 23 2021, 4:51 PM · Restricted Project
dcaballe updated the summary of D97217: [mlir] Extend Operation visitor with pre-order traversal.
Feb 23 2021, 3:48 PM · Restricted Project
dcaballe updated the diff for D97217: [mlir] Extend Operation visitor with pre-order traversal.

Addressing feedback:

  • Use enum to encode walk order.
  • Set region, block and operation walks to post-order by default.
  • Adjust walks in Liveness and NumberOfExecutions to align with the new default traversal.
Feb 23 2021, 3:47 PM · Restricted Project
dcaballe added inline comments to D97252: [MLIR][affine] Certain Call Ops to prevent fusion.
Feb 23 2021, 9:36 AM · Restricted Project

Feb 22 2021

dcaballe added inline comments to D97252: [MLIR][affine] Certain Call Ops to prevent fusion.
Feb 22 2021, 11:32 PM · Restricted Project
dcaballe added a comment to D97217: [mlir] Extend Operation visitor with pre-order traversal.

Please also see: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/tensorflow/utils/visitor_util.h

Feb 22 2021, 10:07 PM · Restricted Project
dcaballe accepted D97252: [MLIR][affine] Certain Call Ops to prevent fusion.

It looks great! Thank you so much!

Feb 22 2021, 9:59 PM · Restricted Project
dcaballe added inline comments to D97217: [mlir] Extend Operation visitor with pre-order traversal.
Feb 22 2021, 12:30 PM · Restricted Project
dcaballe requested review of D97217: [mlir] Extend Operation visitor with pre-order traversal.
Feb 22 2021, 12:27 PM · Restricted Project
dcaballe accepted D97030: [MLIR][affine-loop-fusion] Handle defining ops between the source and dest loops.

I think the add between affine.load and affine.for was handled by hasNonAffineUsersOnThePath.

Feb 22 2021, 10:09 AM · Restricted Project
dcaballe accepted D97032: [MLIR][affine] Prevent fusion when ops with memory effect free are present between producer and consumer.

Thanks for addressing the comments!

Feb 22 2021, 9:39 AM · Restricted Project

Feb 21 2021

dcaballe requested changes to D97030: [MLIR][affine-loop-fusion] Handle defining ops between the source and dest loops.

I would like to know what @bondhugula and @andydavis1 think but I think we might want to add the proper
%0 -> %i1 dependence to the MDG,

Feb 21 2021, 11:43 PM · Restricted Project
dcaballe accepted D97032: [MLIR][affine] Prevent fusion when ops with memory effect free are present between producer and consumer.

If you could improve the lit test so that loads and stores are also checked, that should be all. Thanks!

Feb 21 2021, 10:27 PM · Restricted Project

Feb 19 2021

dcaballe added a comment to D97032: [MLIR][affine] Prevent fusion when ops with memory effect free are present between producer and consumer.

https://reviews.llvm.org/D97030 is somehow related. I think the common solution is to represent these missing dependences in the MDG. We may have to coordinate the fixes.

Feb 19 2021, 11:34 AM · Restricted Project
dcaballe added a comment to D97030: [MLIR][affine-loop-fusion] Handle defining ops between the source and dest loops.

Hey tungld,

Feb 19 2021, 11:31 AM · Restricted Project
dcaballe added a reviewer for D97030: [MLIR][affine-loop-fusion] Handle defining ops between the source and dest loops: bondhugula.
Feb 19 2021, 10:52 AM · Restricted Project
dcaballe added a reviewer for D97032: [MLIR][affine] Prevent fusion when ops with memory effect free are present between producer and consumer: andydavis1.
Feb 19 2021, 10:49 AM · Restricted Project
dcaballe requested changes to D97032: [MLIR][affine] Prevent fusion when ops with memory effect free are present between producer and consumer.

Thank you so much for the fix! Some comments below.

Feb 19 2021, 10:49 AM · Restricted Project

Feb 12 2021

dcaballe committed rGee66e43a96e1: [mlir][Vector] Introduce 'vector.load' and 'vector.store' ops (authored by dcaballe).
[mlir][Vector] Introduce 'vector.load' and 'vector.store' ops
Feb 12 2021, 10:53 AM
dcaballe committed rG656674a7c433: [mlir][Vector] Align gather/scatter/expand/compress API (authored by dcaballe).
[mlir][Vector] Align gather/scatter/expand/compress API
Feb 12 2021, 10:53 AM
dcaballe closed D96396: [mlir][Vector] Align gather/scatter/expand/compress API.
Feb 12 2021, 10:53 AM · Restricted Project
dcaballe closed D96185: [mlir][Vector] Introduce 'vector.load' and 'vector.store' ops.
Feb 12 2021, 10:53 AM · Restricted Project

Feb 10 2021

dcaballe added a comment to D96284: separate AffineMapAccessInterface from AffineRead/WriteOpInterface .

It works after the last update. Thanks!
LGTM. Some minor comments. Please, wait for @ftynse's final approval.

Feb 10 2021, 10:27 AM · Restricted Project
dcaballe added a comment to D96284: separate AffineMapAccessInterface from AffineRead/WriteOpInterface .

Maybe there's something off on my side but I can't see the diff between the last version of the patch (Diff 3) and Base. When I set up that view in "Revision Contents->History", it shows the Diff against a previous version of the patch, which doesn't include all the changes. Something seems to be off.

Feb 10 2021, 10:10 AM · Restricted Project
dcaballe added a comment to D96185: [mlir][Vector] Introduce 'vector.load' and 'vector.store' ops.

If this is a commitment that you are working on starting to connect the pieces next then fair enough :)

Feb 10 2021, 9:51 AM · Restricted Project

Feb 9 2021

dcaballe requested review of D96396: [mlir][Vector] Align gather/scatter/expand/compress API.
Feb 9 2021, 9:38 PM · Restricted Project