Remove Views Pass by ThrudPrimrose · Pull Request #2335 · spcl/dace

ThrudPrimrose · 2026-04-15T13:42:30Z

This PR introduces a pass that removes views from the SDFG by folding them into the access patterns of the arrays they view.

Context: Reduced SDFG IR
This is the first PR toward a Reduced IR. A simplified subset of the full SDFG IR that limits the number of features transformations that need support. The Reduced IR eliminates views, WCR edges, other_subset on edges, dynamic thread-block maps, and persistent device maps by lowering each to a smaller set of primitives (e.g., WCR becomes a reduction library node, other_subset requiring copies becomes explicit copy nodes).

The goal of this PR is to be part of many to provide a lower pipeline from the full IR to the Reduced IR, removing as many of these features as possible. Alongside this pass, I have also extended the Copy library node and implemented a pass that inserts explicit copies where needed. (Separate PR)

What this pass does
The pass tries to remove as many views as possible by absorbing them into surrounding memlets. For example, a view that selects a strided column from a row-major array:

is removed with the following subset:

(It supports many other patterns as well).

tbennun · 2026-04-15T23:49:34Z

What cases does your pass remove that the existing view removal transformations in array elimination doesn’t cover?

ThrudPrimrose · 2026-04-16T09:33:23Z

What cases does your pass remove that the existing view removal transformations in array elimination doesn’t cover?

I compared ArrayElimination against RemoveViews. Remove views handles interstate edges, and some other complicated patterns.

test	before	after AE Only	AE correct
test_view_array_array	1	1	skip
test_view_slice_detect_simple	1	0	OK
test_read_slice	3	0	OK
test_simple_slice_view	1	0	OK
test_reshape_view	2	1	OK
test_squeeze_view	1	0	OK
test_view_chain	2	0	OK
test_noop_no_views	0	0	OK
test_unsqueeze_view	1	0	OK
test_multiple_views_same_state	2	0	OK
test_write_view	1	0	OK
test_column_view	1	0	OK
test_column_view_w_offset	1	0	OK
test_strided_column_view	1	1	OK
test_flatten_view	1	1	OK
test_view_in_interstate_edge	1	1	OK

tbennun

Please add to simplification pipeline

tbennun · 2026-04-18T05:41:15Z

+Two strategies are used depending on the view kind:
+
+Four strategies are tried in order until one succeeds:


is it two or four :)

tbennun · 2026-04-30T15:21:03Z

@@ -206,9 +201,6 @@ def _delinearize_flat(flat, astrides: List[int], array_shape: List[int]):
    return [(flat // astr) % shp for astr, shp in zip(astrides, array_shape)]


can you explain how this works, especially with complex permutations?

tbennun · 2026-05-12T16:38:22Z

+from dace.sdfg import nodes as nd, utils as sdutil, graph as gr
+from dace.transformation import pass_pipeline as ppl, transformation
+
+_PASS = 'RemoveViews'


unnecessary

tbennun · 2026-05-12T16:39:19Z

+from dace.transformation import pass_pipeline as ppl, transformation
+
+_PASS = 'RemoveViews'
+_DEBUGPRINT = config.Config.get('debugprint') in (True, '1', 'true', 'yes')


this does not work. Also this does not conform to our API.

It will only read it on import, so someone who sets the config temporarily will not observe this change.

Call config.Config.get_bool('debugprint')

So just call this function in the only place it is used.

… MapToForLoop reparent Bundle of related canonicalize-pipeline improvements: * **`merge` -> `ITE` global rename**: the 3-input ternary blend used by branch_normalization, vectorization, and (now) `EarlyExitToFindIndex` is renamed throughout. `class ITE(sympy.Function)` is the sole class in `dace.symbolic`; `dace/runtime/include/dace/merge.h` is renamed to `ITE.h` and the C++ template is `ITE(c, a, b)`. Files moved: `lower_merge_to_fp_factor.py` -> `lower_ite_to_fp_factor.py`, `same_write_set_if_else_to_merge_cfg.py` -> `same_write_set_if_else_to_ite_cfg.py`, plus the matching test files under `tests/`. All call sites updated to emit `ITE(...)`. Tests: test_split_tasklets / test_ite_function / test_lower_ite_to_fp_factor / test_same_write_set_if_else_to_ite_cfg / test_branch_normalization 43/43 focused tests pass. * **`SplitTasklets`: lower Python `IfExp` to `ITE(c, t, e)`** in ASTSplitter so the tasklet body never carries a Python ternary and cppunparse never has to emit a `c ? t : e` ternary. * **`LowerITEToFpFactor` wired after `SplitTasklets` in 'clean'** so every `ITE(c, t, e)` call lowers to the FP-factor form `c * t + (1 - c) * e` at canonicalize time. * **EarlyExitToFindIndex** indicator-Map tasklet now emits `__out = ITE((cond), i, N)` (was `__out = i if (cond) else N`); also adds `_trace_transient_to_source_arrays` so the soundness check walks transient gather chains (`d_index = d[i]`) back to the underlying non-transient array. Three new tests: combo body_pre + body_post, cond/body overlap refusal pre+post, and the s481/s482 contracts. * **Scalar-slice fold passes wired into `_structural_cleanup`**: `CleanAccessNodeToScalarSliceToTaskletPattern` and its inverse `CleanTaskletToScalarSliceToAccessNodePattern` collapse the gather bridges so downstream matchers see the underlying array names. Wired between `RemoveViews` and `EmptyStateElimination`. Known regression set on the TSVC corpus (s314-s316, s3110/s3111/s31111, s352/s353, s421-s443, s4112-s4117) is being hardened separately at the LoopToReduce / LoopToScan / ArgMaxLift matchers; this commit lands the wiring + accepts the regressions. * **RemoveViews pass (PR #2335)** wired into `_structural_cleanup` between InlineSDFG and the scalar-slice folds; 16/16 new `tests/remove_views_test.py` tests pass. * **MapToForLoop `inline_after` successor-edge reparent**: when the inline migration is requested, the placeholder graph's pre-existing `graph -> *` interstate edges are now reparented onto the new `target_state` so the placeholder has exactly one out-edge. Without this, `control_flow_raising` later lifted the placeholder into a `ConditionalBlock` with the `else` branch not in the last position, and `DeadStateElimination._find_dead_branches` aborted. Three previously-failing tests now pass: `test_thomas_solve_value_preserving`, `test_thomas_solve_keeps_vertical_axis_sequential`, and `test_canonicalization_stages_apply_to_cloudsc`. Regression test `test_a6_multi_successor_reparent_on_inline` covers the contract. * **`control_flow_raising`** stable-sorts out-edges so unconditional edges fall at the tail before populating the `ConditionalBlock` branches, keeping the lift robust against future producers that emit out-edges in arbitrary order. Regression test `test_unconditional_edge_lifted_as_last_branch`. * **`tasklet_utils.classify_tasklet`**: unknown shapes (unknown function call, RHS not a recognised pattern) now report `type = TaskletType.UNKNOWN` instead of raising `NotImplementedError`. Adds `TaskletType.UNKNOWN` enum member. Existing callers using `info.get("type") in {...}` keep working; callers that want to explicitly skip unclassified tasklets now branch on `ttype == TaskletType.UNKNOWN`.

ThrudPrimrose added 2 commits April 15, 2026 15:40

Add remove views PR

247dd95

Refactor

2f1ef46

ThrudPrimrose marked this pull request as ready for review April 15, 2026 20:51

ThrudPrimrose requested review from phschaad and tbennun April 15, 2026 21:18

ThrudPrimrose added 2 commits April 16, 2026 11:34

Support interstate edges

20137fc

Refactor comments

f70904d

romanc mentioned this pull request Apr 17, 2026

ci: find and automatically remove unused imports #2318

Merged

tbennun requested changes Apr 30, 2026

View reviewed changes

tbennun requested changes May 12, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Remove Views Pass#2335

Remove Views Pass#2335
ThrudPrimrose wants to merge 4 commits into
mainfrom
clean_views

ThrudPrimrose commented Apr 15, 2026 •

edited

Loading

Uh oh!

tbennun commented Apr 15, 2026

Uh oh!

ThrudPrimrose commented Apr 16, 2026 •

edited

Loading

Uh oh!

tbennun left a comment

Uh oh!

tbennun Apr 18, 2026

Uh oh!

tbennun Apr 30, 2026

Uh oh!

tbennun May 12, 2026

Uh oh!

tbennun May 12, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		Two strategies are used depending on the view kind:

		Four strategies are tried in order until one succeeds:

		@@ -206,9 +201,6 @@ def _delinearize_flat(flat, astrides: List[int], array_shape: List[int]):
		return [(flat // astr) % shp for astr, shp in zip(astrides, array_shape)]

Uh oh!

Conversation

ThrudPrimrose commented Apr 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

tbennun commented Apr 15, 2026

Uh oh!

ThrudPrimrose commented Apr 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

tbennun left a comment

Choose a reason for hiding this comment

Uh oh!

tbennun Apr 18, 2026

Choose a reason for hiding this comment

Uh oh!

tbennun Apr 30, 2026

Choose a reason for hiding this comment

Uh oh!

tbennun May 12, 2026

Choose a reason for hiding this comment

Uh oh!

tbennun May 12, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

ThrudPrimrose commented Apr 15, 2026 •

edited

Loading

ThrudPrimrose commented Apr 16, 2026 •

edited

Loading