CanonicalizeTileSlice Pass

Lowers a tile.slice into the canonical tile.extract form so that movement is unified on pto.textract — Mat-resident slices (folded into matmul / tile.extract consumers), Vec slices whose lazy materialization would corrupt their source (materialized for the tile.col_expand_* family, issues #1640 and #2010), and unaligned Vec subviews (issue #1789).

Overview

A tile.slice whose result tile is Mem.Mat is a legal high-level "sub-window of a Mat tile" construct. FlattenTileNdTo2D emits one per batch page when it unrolls a tile.batch_matmul: the page offset is batch_index * page_rows, and for a leading-dim-1 batch that offset is 0 and the window covers the whole tile — but it is still a tile.slice.

PTO ISA supports pto.subview on Mat as a zero-copy alias (no data movement), so a standalone Mat slice is valid when its consumer accepts the subview SSA directly. However, consumers that trigger lazy materialization (via MaterializeSubviewOperandIfNeeded) would attempt a loc=mat → loc=mat pto.textract — an unsupported L1→L1 DMA path on targets such as Ascend 910C. This pass eliminates Mat-resident tile.slice nodes whose consumers it can canonicalize (extract/matmul) by folding the offset into each consumer for efficiency, then drops the now-dead slice. A Mat slice with a consumer that is not canonicalized (e.g. tile.move) is left intact — it lowers to a valid pto.subview.

The pass also canonicalizes a Vec tile.slice consumed by the tile.col_expand_* family (issues #1640, #2010). Those ops cannot read a pto.subview operand, so codegen lazily materializes the slice via pto.textract into the slice's own result buffer — and because tile.slice inherits its source's memory, that buffer sits inside the still-live source. The extract therefore runs in place over its own input, which is only safe when it is an identity copy. Two conditions must hold:

Condition	Why it can fail
The destination address is right	AllocateMemoryAddr folds a ConstInt offset into base + off, but a dynamic offset cannot be encoded as a ConstInt address and falls back to the bare source base — the extracted window lands on the source's row 0 (#1640).
The destination layout matches	The slice's buffer is dense (row pitch = slice cols) while the source window is strided (row pitch = source cols). These coincide only for a contiguous window: a single row, or one spanning every column. A column slice of a multi-row tile (t[:, a:b]) repacks strided → dense on top of its own source and destroys it — only row 0 survives, because its dense destination happens to equal its source address (#2010).

When either condition fails, the operand is replaced by a fresh tile.extract(..., target_memory=Vec), whose result gets its own non-inherited allocation. tile.extract is registered not_inplace_safe(), so MemoryReuse cannot place that fresh buffer back onto the source either. A slice whose materialization is an identity copy is left untouched, so it keeps sharing the source buffer rather than paying for a duplicate allocation.

Independently, PTO vector instructions require tile operand base addresses to be 32-byte aligned. A zero-copy Vec slice starts at

base + (off_row * base_cols + off_col) * storage_bits

so an FP32 column slice at [:, 1:2] starts only 4 bytes past an aligned allocation. Feeding that subview directly to an ordinary vector op such as tile.muls can hang the device (#1789). The pass therefore replaces an unaligned Vec slice operand with a fresh tile.extract(..., target_memory=Vec). The new allocation is aligned; provably aligned slices remain zero-copy. Dynamic offsets are also kept zero-copy when scalar SSA arithmetic proves that their known multiple produces a 32-byte-aligned row or column displacement.

Pipeline position: After AutoTileMatmulL0 (so the per-iter tile.extracts that read the batch-page slices already exist), before InferTileMemorySpace.

Requirements: SSAForm, SplitIncoreOrch, IncoreTileOps, TileOps2D, NormalizedStmtStructure.

Produces: same as required (property-preserving rewrite).

Invalidates: nothing.

When to use: Always, as part of the default tile-stage pipeline. The pass is a no-op when no canonical tile.slice exists.

API

C++	Python	Level
pass::CanonicalizeTileSlice()	passes.canonicalize_tile_slice()	Function-level

from pypto.pypto_core import passes

program_canon = passes.canonicalize_tile_slice()(program)

Algorithm

For each InCore-typed function, in three phases:

1. Collect — index every AssignStmt whose value is a tile.slice(src, shape, offset) in canonical 3-argument form. A slice whose src is itself a recorded slice is peeled, accumulating the offset, so each entry resolves to a non-slice base tile plus a total (off_row, off_col). Direct ConstInt SSA definitions and their plain aliases are resolved before this analysis, so a literal offset does not become artificially dynamic after ConvertToSSA. Scalar SSA definitions are also retained for modular alignment proofs: for example, block_idx * 32 is dynamic but has a statically known 32-element multiple. Slices carrying valid_shape / drop_dims (4–5 arguments) are not plain windows and are skipped.

2. Rewrite consumers — for each slice:

3. tile.extract(slice, ir, ic, shape) → tile.extract(base, ir + off_row, ic + off_col, shape). The extract reads the slice's source directly; the index add is constant-folded when both terms are ConstInt.
4. tile.matmul / tile.matmul_acc / tile.matmul_bias operand (Mat slices only) → the operand is replaced by a fresh tile.extract(base, off_row, off_col, slice_shape, target_memory=Left|Right) — Left for the lhs operand, Right for the rhs. (The tile.matmul_acc accumulator operand is Acc-resident and never a Mat slice, so it is not rewritten here; it is instead checked for L0C contiguity — see Acc accumulator windows.)
5. tile.col_expand_* operand (Vec slices only) → when the lazy pto.textract would not be an identity copy — a dynamic offset, or a window that is not contiguous in the base tile (more than one row and narrower than the base) — the operand is replaced by a fresh tile.extract(base, off_row, off_col, slice_shape, target_memory=Vec). Both operands are checked. Contiguous const-offset windows are left untouched.
6. Ordinary call operand (Vec slices only, in either an AssignStmt or an EvalStmt) → compute (base_byte_offset * 8 + (off_row * base_cols + off_col) * storage_bits) mod 256. A known concrete MemRef byte offset is included before the modulo calculation; the allocation-planning sentinel is treated as an aligned root, while a non-constant base offset is not statically provable. Scalar SSA definitions are followed through aliases, addition/subtraction, and multiplication to prove aligned dynamic multiples. If the result is nonzero or cannot be proved, replace the operand by a fresh tile.extract(base, off_row, off_col, slice_shape, target_memory=Vec). tile.slice itself is skipped so chained views can be peeled, and tile.extract uses the direct folding rule above.

7. SSA escape (Vec slices only) → an unaligned slice assigned through a plain alias is materialized at the alias definition. An unaligned loop initializer is materialized before the loop and substituted through its IterArg; an unaligned value carried by yield is materialized before the yield. This prevents aliases and loop-carried identities from bypassing the ordinary-call lookup.

8. Drop dead slices — a tile.slice whose result no longer has any use is removed. A chained slice only becomes dead once the slice consuming it is dropped, so this iterates to a fixpoint (bounded by the slice count). A slice still used at the end had a consumer this pass does not canonicalize; it is left intact — no regression versus the pre-pass IR.

The pass is a FunctionPass; functions are returned unchanged when no canonical tile.slice is present.

Examples

Slice folded into tile.extract

The offset-0 full-shape slice FlattenTileNdTo2D emits for a leading-dim-1 batch operand:

Before:

lhs_slice: pl.Tile[[32, 512], pl.INT8, pl.Mem.Mat] = pl.tile.slice(x_mat, [32, 512], [0, 0])
a:         pl.Tile[[32, 256], pl.INT8, pl.Mem.Left] = pl.tile.extract(
    lhs_slice, 0, ko, shape=[32, 256], target_memory=pl.Mem.Left)

After (slice dropped; extract reads the loaded Mat tile directly):

a: pl.Tile[[32, 256], pl.INT8, pl.Mem.Left] = pl.tile.extract(
    x_mat, 0, ko, shape=[32, 256], target_memory=pl.Mem.Left)

A non-zero page offset folds into the extract index — e.g. a slice at [32, 0] turns extract(slice, 0, ko, ...) into extract(x_mat, 32, ko, ...).

Slice folded into a tile.matmul operand

When AutoTileMatmulL0 leaves a matmul untiled (already L0-sized), its Mat-slice operands are converted directly:

Before:

lhs_slice: pl.Tile[[16, 256], pl.BF16, pl.Mem.Mat] = pl.tile.slice(lhs_mat, [16, 256], [0, 0])
rhs_slice: pl.Tile[[256, 64], pl.BF16, pl.Mem.Mat] = pl.tile.slice(rhs_mat, [256, 64], [0, 0])
c:         pl.Tile[[16, 64],  pl.FP32, pl.Mem.Acc] = pl.tile.matmul(lhs_slice, rhs_slice)

After:

lhs_left:  pl.Tile[[16, 256], pl.BF16, pl.Mem.Left]  = pl.tile.extract(
    lhs_mat, 0, 0, shape=[16, 256], target_memory=pl.Mem.Left)
rhs_right: pl.Tile[[256, 64], pl.BF16, pl.Mem.Right] = pl.tile.extract(
    rhs_mat, 0, 0, shape=[256, 64], target_memory=pl.Mem.Right)
c:         pl.Tile[[16, 64],  pl.FP32, pl.Mem.Acc]   = pl.tile.matmul(lhs_left, rhs_right)

Vec slice materialized into a tile.col_expand_mul operand (#1640)

A dynamic-offset slice of a local tile feeding col_expand_mul (the same rewrite applies to col_expand_add):

Before:

row:    pl.Tile[[1, 256], pl.FP32, pl.Mem.Vec] = pl.tile.slice(local, [1, 256], [row_off, 0])
scaled: pl.Tile[[1, 256], pl.FP32, pl.Mem.Vec] = pl.tile.col_expand_mul(row, gamma_t)

After (slice dropped; the operand is materialized into a fresh, non-aliasing tile):

row_ext: pl.Tile[[1, 256], pl.FP32, pl.Mem.Vec] = pl.tile.extract(
    local, row_off, 0, shape=[1, 256], target_memory=pl.Mem.Vec)
scaled:  pl.Tile[[1, 256], pl.FP32, pl.Mem.Vec] = pl.tile.col_expand_mul(row_ext, gamma_t)

Column slice of a multi-row Vec tile (#2010)

t[:, 64:128] on a [16, 128] tile is a static-offset slice, but its window is not contiguous — 16 rows, 64 of the source's 128 columns — so it is materialized too:

Before:

hi:     pl.Tile[[16, 64], pl.FP32, pl.Mem.Vec] = pl.tile.slice(t, [16, 64], [0, 64])
scaled: pl.Tile[[16, 64], pl.FP32, pl.Mem.Vec] = pl.tile.col_expand_mul(hi, gamma_t)

After:

hi_ext: pl.Tile[[16, 64], pl.FP32, pl.Mem.Vec] = pl.tile.extract(
    t, 0, 64, shape=[16, 64], target_memory=pl.Mem.Vec)
scaled: pl.Tile[[16, 64], pl.FP32, pl.Mem.Vec] = pl.tile.col_expand_mul(hi_ext, gamma_t)

Without the rewrite, hi allocates a dense [16, 64] buffer at t + 256 B — inside t — and the lazy pto.textract repacks t's strided columns into it, overwriting t as it reads it. Only row 0 comes back correct, which is why the same construct is harmless on a single-row tile.

Unaligned Vec column slice (#1789)

Column 1 of an FP32 tile begins four bytes past its aligned source allocation, so it cannot be used directly by tile.muls:

head:   pl.Tile[[16, 1], pl.FP32, pl.Mem.Vec] = pl.tile.slice(local, [16, 1], [0, 1])
scaled: pl.Tile[[16, 1], pl.FP32, pl.Mem.Vec] = pl.tile.muls(head, 0.5)

The pass gives the vector operation an aligned, independently allocated tile:

head_ext: pl.Tile[[16, 1], pl.FP32, pl.Mem.Vec] = pl.tile.extract(
    local, 0, 1, shape=[16, 1], target_memory=pl.Mem.Vec)
scaled:   pl.Tile[[16, 1], pl.FP32, pl.Mem.Vec] = pl.tile.muls(head_ext, 0.5)

By contrast, FP32 column 8 is 32 bytes from the source base and remains a zero-copy slice. A dynamic row offset also remains zero-copy when the source row stride is 32-byte aligned.

Implementation

Header: include/pypto/ir/transforms/passes.h

Properties: include/pypto/ir/transforms/pass_properties.h (kCanonicalizeTileSliceProperties)

Implementation: src/ir/transforms/canonicalize_tile_slice_pass.cpp

Python binding: python/bindings/modules/passes.cpp

Tests: tests/ut/ir/transforms/test_canonicalize_tile_slice.py

Acc accumulator windows

The pass also rejects one shape it cannot repair. L0C is NZ: block (r_b, c_b) of an [M, N] tile sits at (c_b * M/16 + r_b) * fractal. A window is therefore contiguous only when it spans the parent's full row extent, or lies inside a single 16-column block. A row window of a multi-block-column Acc tile is strided, and the MAD writes its destination compactly from a bare pointer with no destination stride, so every block column past the first lands in the wrong row tile — silently, with only the first 16 columns of each row tile correct.

Unlike the Vec cases above there is no repair available: an Acc window cannot be copied out and back, because nothing in the memory graph points into Acc. So a matmul accumulator operand that is a non-contiguous Acc window raises a ValueError naming the working spelling, which differs only by the sliced axis — pack the accumulator as [rows, N * tiles] and slice columns.

The check is scoped by the op registry's set_output_reuses_input, so it covers tile.matmul_acc / tile.gemv_acc / tile.matmul_mx_acc without naming them. It stays silent on anything it cannot prove: a symbolic extent, a non-Acc layout, or a dynamic column offset it cannot show stays inside one block. It applies to the accumulator whether the op is bound to a result (acc = pl.tile.matmul_acc(...)) or written as a bare statement (pl.tile.matmul_acc(...)).

Resolving the accumulator operand back to its window

The window an accumulator names is rarely the tile.slice result Var. Hoisting the slice out of the K loop — the natural spelling, since the window does not change per K step — makes the in-body operand a loop IterArg, a different Var from the slice result; a shape-preserving tile.reshape is another SSA name for the same bytes. So the pass records, for every Var, which window that Var names, not merely "is this Var a slice", and the check looks the operand up in that map.

A Var inherits another Var's window along these identity-preserving edges, each recorded during the same single program-order sweep that collects the slices:

Edge	Recorded from
v = tile.slice(src, shape, offset)	The window itself (seed); chained slices are peeled, offsets accumulated
v = w (plain SSA alias)	w's window
v = <op>(src, ...) where the op's output lands in a source operand's storage	The source operand's window. The relation is a registry read (OutputMemoryInheritsInput() && IsInplaceSafe(), or set_output_reuses_input), never an op list — so tile.reshape / tile.set_validshape and a chained tile.matmul_acc are covered, while tile.transpose and tile.extract (fresh buffers) are excluded
Loop IterArg	Its initializer's window (ForStmt / WhileStmt)
ForStmt / WhileStmt return_vars_[i]	The window the trailing pl.yield_ names for iter_args_[i] — only when that is the same window the initializer names, which is the canonical carry. A zero-trip loop returns the initializer instead, so binding a different window would claim one that a path does not name
IfStmt::return_vars_[i]	The window both arms yield, when they provably yield the same one (same parent Var, equal offsets)

An edge is recorded only when both ends provably describe the same window extent — same rank, same dtype, and every dimension a compile-time constant of the same value. Landing in the source's storage says the two Vars name the same bytes; the extent test says the shape the check reads off the operand's own TileType is still the extent those bytes form. Together they are a proof, and anything unproven is left unrecorded — which reproduces the previous silence rather than guessing:

• a shape-changing tile.reshape fails the test, so a reshaped extent is never compared against the parent's rows;
• tile.reinterpret_view fails it on dtype — the 16-column-block arithmetic assumes a 4-byte accumulator element;
• tile.transpose_view swaps the trailing two extents, so it fails it for every non-square window. A square window passes and is harmless: the transposed view spans the identical bytes with identical row and column extents, so both accept rules reach the same verdict as for the untransposed window;
• a symbolic extent fails it, matching the check's own refusal to reason about symbolic shapes.

Memory space is deliberately not compared: the verdict comes from the parent tile's space and the operand's extents, never from the operand's own space — and the tile view ops legitimately deduce a result type that leaves it unset for InferTileMemorySpace to fill in later.

The loop back edge is checked, not recorded

A carried accumulator names two windows: the initializer's on iteration 0, and whatever the body's trailing pl.yield_ names on every iteration after. A Var cannot be recorded as naming two windows, and recording either one would make the other iteration's destination invisible — so the yielded window is checked against the same rule instead, with the destination extent taken from the IterArg (the type the loop gives the carry).

Three conditions keep that from over-rejecting:

• it fires only for carries the body actually uses as an accumulator destination, so a carry the body merely reads is never rejected for a shape no MAD writes;
• it fires only when the loop provably runs more than once — constant start / stop / step with start + step < stop. A single-trip loop never takes its back edge; a symbolic trip count and every while count as unprovable and stay silent;
• it is a check, never a binding, so the map stays acyclic. No fixpoint is needed and the sweep stays one forward pass.

The canonical carry (yield matmul_acc(iter_arg, ...)) resolves straight back to the initializer's window, so this re-reaches the verdict the body already passed and costs nothing on the shapes the pipeline generates.

Known limits — all of them silent, never a false rejection:

• An unprovable trip count leaves the back edge unchecked. A symbolic loop bound or a while cannot be shown to take the back edge, so a carry whose yield names a worse window than its initializer goes unchecked there.
• An if-merge whose arms name different windows is not bound. Binding either one would be a claim the other path contradicts — and this map also feeds the chained-slice offset arithmetic, so a wrong binding could move the verdict for an unrelated slice.
• Interprocedural routes are out of reach. This is a function pass, so a window arriving as an InCore function parameter carries no window information.

The rejected window is not always in the kernel source

A compiler-generated window can reach this check with no tile.slice written anywhere in the kernel, so the diagnostic must not assume the author can edit the slice it names.

FlattenTileNdTo2D used to be the main source of such windows: it unrolled a tile.batch_matmul_acc by stacking the accumulator's batch pages along rows and slicing one page per batch (acc_page_i = tile.slice(acc, [M, N], [i * M, 0])), which is exactly the rejected shape for N > 16. That lowering is gone — it now packs the pages along columns (acc_page_i = tile.slice(acc, [M, N], [0, i * N])), the accepted full-row-extent shape, and rejects the shapes it cannot pack with its own diagnostic naming the DSL workaround. See Batched accumulators pack along columns.

The narrow row-packed form survives there for pages at most 16 columns wide, which this pass whitelists as a single-block window, so a row window still reaches this check — it is simply never one this pass rejects. The rejection therefore stays as written: it is the backstop that would catch any future pass re-introducing a strided accumulator destination, and narrowing it would let a silent miscompute through.

This is a workaround for an upstream defect (hw-native-sys/pto-isa#253), not a property of the DSL — TMATMUL_ACC_IMPL forwards the destination as a bare .data() pointer and takes m from the left operand, so TileRes::Rows is never read. If pto-isa gains a destination stride, this rejection should be relaxed or deleted rather than kept.

Pass Properties

Property	Value
Required	SSAForm, SplitIncoreOrch, IncoreTileOps, TileOps2D, NormalizedStmtStructure
Produced	SSAForm, SplitIncoreOrch, IncoreTileOps, TileOps2D, NormalizedStmtStructure
Invalidated	—

Scope

Op	Action
Mat-resident tile.slice (3-arg) feeding tile.extract	Folded into the extract; slice dropped
Mat-resident tile.slice (3-arg) feeding a matmul-family operand	Replaced by tile.extract(target_memory=Left\|Right); slice dropped
Dynamic-offset Vec tile.slice (3-arg) feeding a tile.col_expand_* op	Replaced by tile.extract(target_memory=Vec); slice dropped (#1640 — the address falls back to the bare source base)
Static-offset non-contiguous Vec tile.slice (multi-row and narrower than its base, e.g. t[:, a:b]) feeding a tile.col_expand_* op	Replaced by tile.extract(target_memory=Vec); slice dropped (#2010 — the dense repack would run on top of its own live source)
Static-offset contiguous Vec tile.slice (single row, or full source width) feeding a tile.col_expand_* op	Untouched (the lazy textract is a safe identity copy; keeps sharing the source buffer)
Vec tile.slice feeding an ordinary call, inherited address not provably 32-byte aligned	Replaced by tile.extract(target_memory=Vec); slice dropped (#1789)
Vec tile.slice feeding an ordinary call, inherited address provably 32-byte aligned	Untouched; keeps the zero-copy subview
Chained Mat tile.slice (slice of a slice)	Peeled; offsets accumulated
tile.slice with valid_shape / drop_dims	Skipped (not a plain window). If such a slice also fails either identity-copy condition above — a dynamic offset (e.g. a rank-reducing t[i]) or a non-contiguous window — while feeding a col-expand op, codegen rejects it with an INTERNAL_CHECK rather than emitting the source-corrupting materialization. The Acc accumulator check above still applies to such a slice: it needs only the physical base and offset, which are recorded for every window regardless of canonicalization eligibility
Acc-resident tile.slice used as a matmul accumulator, window neither spanning the parent's full row extent nor inside one 16-column block	Rejected with a ValueError naming the column-slice spelling — the MAD has no destination stride, so the write would silently land in the wrong row tile (pto-isa#253). The operand does not have to be the slice: a loop IterArg carrying it, a plain SSA alias, or a shape-preserving view / in-place op over it resolves back to the same window (see "Resolving the accumulator operand back to its window")
Other Left/Right/Acc-resident tile.slice, including a contiguous Acc accumulator window	Untouched (no matching consumer)
Functions with no canonical tile.slice	Returned unchanged

See also

• FlattenTileNdTo2D — upstream pass; emits the Mat-resident batch-page tile.slice this pass lowers
• AutoTileMatmulL0 — upstream pass; emits the tile.extracts that consume the batch-page slices