(dialects and transforms): csl stencil lowering #2747
Draft
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
n-io
added
documentation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #2747 +/- ##
==========================================
+ Coverage 89.79% 89.85% +0.05%
==========================================
Files 394 408 +14
Lines 48747 51021 +2274
Branches 7471 7912 +441
==========================================
+ Hits 43774 45843 +2069
- Misses 3793 3928 +135
- Partials 1180 1250 +70 ☔ View full report in Codecov by Sentry. |
PapyChacal
reviewed
Jun 19, 2024
Comment on lines
19
to
22
| func.func @gauss_seidel(%a : memref<1024x512xtensor<512xf32>>, %b : memref<1024x512xtensor<512xf32>>) { | ||
| %0 = stencil.external_load %a : memref<1024x512xtensor<512xf32>> -> !stencil.field<[-1,1023]x[-1,511]xtensor<512xf32>> | ||
| %1 = stencil.load %0 : !stencil.field<[-1,1023]x[-1,511]xtensor<512xf32>> -> !stencil.temp<[-1,1023]x[-1,511]xtensor<512xf32>> | ||
| %2 = stencil.external_load %b : memref<1024x512xtensor<512xf32>> -> !stencil.field<[-1,1023]x[-1,511]xtensor<512xf32>> |
There was a problem hiding this comment.
Unless I missed something specific to the tensorized aspect of it, you could write the function as taking the !stencil.fields directly here, they would get lowered to the memref you are converting from anyway 🙂
There was a problem hiding this comment.
Always happy to work with as many kernels as I can get my hands on and any variety of representations we might want to support. IIRC, this particular one was provided by @mesham, the tensorisation pass applied to it should only rewrite types and accesses.
AntonLydike
reviewed
Jun 24, 2024
AntonLydike
reviewed
Jun 24, 2024
n-io
added a commit
that referenced
this pull request
Jun 24, 2024
…#2766) Introduces an intermediate dialect with ops: * `csl_stencil.prefetch` to indicate prefetched buffer transfers * `csl_stencil.access` performs a `stencil.access` to a prefetched buffer The `stencil-to-csl-stencil` transform: * lowers `dmp.swap` to `csl_stencil.prefetch` * adds prefetched buffers to signature of `stencil.apply` * lowers `stencil.access` to `csl_stencil.access` iff they are accesses to prefetched buffers For a more detailed description see the document in #2747. This PR implements Step 1 outlined there. --------- Co-authored-by: n-io <[email protected]>
n-io
added a commit
that referenced
this pull request
Jun 28, 2024
This PR implements the `csl_stencil.apply` op as outlined in Step 2 of #2747 This operation combines a `csl_stencil.prefetch` (symmetric buffer communication across a given stencil shape) with a `stencil.apply`. Please see the doc string of the op for a detailed description. --------- Co-authored-by: n-io <[email protected]>
n-io
added a commit
that referenced
this pull request
Jul 5, 2024
This PR implements the conversion to csl_stencil.apply op as outlined in Step 2 of #2747 The `csl_stencil.apply` op combines a csl_stencil.prefetch (symmetric buffer communication across a given stencil shape) with a stencil.apply. The transformation consists of several steps: * When rewriting a `stencil.apply`, select the `csl_stencil.prefetch` with the biggest memory overhead (if several are present) to be fused with the apply op into a `csl_stencil.apply` * Find a suitable split of ops to be divided across the two code blocks * Re-order arithmetic e.g. `(a+b)+c -> (c+a)+b` to access, consume, and reduce data of neighbours (for input stencil only) * Move this into first code block, move all other ops into second code block * Fallback strategy: Move everything into first code block * Add `tensor.InsertSliceOp` to insert computed chunk into returned z-value tensor * Set up code block args * Move ops into new regions according to determined split * Translate arg usage to match new block args * Set up yield ops * Run tensor update shape on chunk region --------- Co-authored-by: n-io <[email protected]>
|
Is this branch still useful? |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR proposes a
csl_stencildialect that manages stencil accesses across the stencil pattern, including communication between PEs, with the goal of preparing the stencil dialect for lowering to CSL.