`pl.tensor`¶

张量级算子 —— 接受并返回 pl.Tensor 值，因此属于编排层，或任何以整张张量为工作单位的地方。见选择命名空间。

Tensor operations for PyPTO Language DSL.

This module provides type-safe wrappers around pypto.ir.op.tensor operations that accept and return Tensor types instead of raw Expr/Call objects.

`create_tensor = create` `module-attribute` ¶

`create(shape, dtype, layout=TensorLayout.ND, manual_dep=False, init_value=None)` ¶

Create a new tensor with specified shape and dtype.

Parameters:

Name	Type	Description	Default
`shape`	`Sequence[IntLike]`	List of dimension sizes (int or Expr)	required
`dtype`	`DataType`	Data type of tensor elements	required
`layout`	`TensorLayout`	Tensor layout (default: ND)	`ND`
`init_value`	`int \| float \| None`	Removed. Passing anything but `None` raises `ValueError`. The runtime dropped its create-info fill, so a runtime-allocated buffer can no longer be pre-filled from orchestration. Seed it with a kernel — `pl.full` materializes a constant tensor, or write the buffer from your own kernel — and order every reader after that kernel with an explicit dependency.	`None`
`manual_dep`	`bool`	Opt this tensor out of OverlapMap auto-dep tracking for its entire lifetime. When True, every task that reads or writes this tensor skips OverlapMap lookup and insert, so the runtime neither makes the task wait on prior writers nor registers it as a producer for later readers. Creator retention (the original `tensor.create`'s owner_task_id) still applies. This is the tensor-lifetime granularity of opting out of auto-dep tracking. The other two granularities, both orthogonal to this one, are: `with pl.manual_scope():` — scope-wide opt-out. `pl.no_dep(t)` at a kernel-call arg position — single-task opt-out for one arg. All three opt-outs compose with the orthogonal explicit edges mechanism (`pl.submit(..., deps=[...])` / `pl.at(..., deps=)`): the final task fanin is auto-tracked deps ∪ explicit deps, so you can use `manual_dep=True` to silence auto-tracking on a scratch buffer while still pinning its ordering with a `deps=[tid]` on the consumer. Internally also used by `InjectGMPipeBuffer` to mark the ring-buffer slots it synthesises.	`False`

Returns:

Type	Description
`Tensor`	Tensor wrapping the create operation

`no_dep(tensor)` ¶

Mark a kernel-call argument as no-dependency (caller-site override).

This is a parser-recognized marker — at runtime it returns the wrapped tensor unchanged. The parser detects pl.no_dep(t) at kernel call arg positions and threads the per-arg ArgDirection.NoDep override into the IR Call's attrs. DeriveCallDirections then overwrites the auto-derived direction at that slot to NoDep.

Effect at runtime: the simpler runtime skips both the OverlapMap dependency lookup and the producer insert for this argument. The marker is legal regardless of whether the callee declares the param as In (read) or Out / InOut (write): the caller is asserting out-of-band that there is no RaW / WaW / WaR conflict on the slot — for example, paged-attention writes whose target offset is data-dependent (so the compiler cannot prove disjointness) but are guaranteed disjoint by the runtime allocation protocol.

Only valid as a direct argument to a kernel call::

# Read-side override: shared_input is read-only at the callee, but
# auto-dep would otherwise serialise sibling fan-out reads.
result = self.my_kernel(pl.no_dep(shared_input), output)

# Write-side override: the callee writes into k_cache via
# ``pl.assemble`` at a data-dependent offset; the user knows the
# offsets are disjoint across the parallel fan-out.
self.rope_kv_cache(q_proj, pl.no_dep(k_cache), pl.no_dep(v_cache))

Outside a kernel call argument list it is a no-op (returns tensor unchanged); the parser only injects the override at recognized call sites.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	The tensor to pass through. Must be a `Tensor` value.	required

Returns:

Type	Description
`Tensor`	The tensor unchanged. The marker is consumed at parse time.

`dump_tag(tensor)` ¶

Mark a tensor for selective dump within the enclosing orchestration.

Declarative per-tensor dump marker (simpler#844 selective tensor dump). Writing pl.dump_tag(q) as a standalone statement records q so that every subsequent kernel dispatch consuming that exact value gets q merged into the dispatch's dump_vars — whether that dispatch lowers to a plain ir.Call (the typical @pl.jit / tensor-op path) or an ir.Submit. The runtime then marks those Arg slots via Arg::dump(...).

This is the declarative counterpart to the explicit dumps=[...] kwarg on pl.submit(...) / pl.at(...) — both feed the same dump_vars set (mirroring how a scope's auto-inferred and explicit deps= edges both feed manual_dep_edges). Use dumps= when you want to list the targets explicitly at a single task launch; use pl.dump_tag when one declaration should stick across every subsequent consumer.

Use this to keep tensor dump viable on large workloads (e.g. paged-attention 64bat/8192ctx) where full dump (enable_dump_args=2) saturates the host-side dump collector (~42 MB/s drain rate) by dumping every binding, eventually triggering a STARS op-timeout kill on the AICPU side. Run partial dump (enable_dump_args=1) and tag only the tensors of interest, so the runtime filters out large bindings (1 GB kv-cache, output buffers, etc.) from the collector queue.

Semantics:

Forward-sticky over the orch scope — one pl.dump_tag(q) statement affects all subsequent kernel calls in the same orch that consume the tagged value. Tracked by Var identity, never by name: reassigning q (e.g. q = self.foo(q)) produces a new value that the prior tag does not cover — re-tag it if needed.
Only effective under partial dump (RunConfig.enable_dump_args == 1) — selective dump filters within the partial pipeline. A no-op when dump is off (0); under full dump (2) every binding is captured, so the tag has nothing to narrow.
Consumed at parse time — recorded into the consuming dispatch's dump_vars and emits no IR statement of its own.

Valid as a standalone statement inside an Orchestration function or an Inline helper (@pl.jit.inline / FunctionType.Inline) that the orchestration inlines. The parser rejects pl.dump_tag written in any other function type (AIV / AIC / Mix kernel bodies) with a clear error::

@pl.function(type=pl.FunctionType.Orchestration)
def orch(self, q: pl.Tensor[...], k_cache: pl.Tensor[...], out: pl.Out[...]):
    pl.dump_tag(q)           # mark q for selective dump
    pl.dump_tag(out)         # mark out for selective dump
    s = self.qk_matmul(q, k_cache, scratch)  # q is dumped here
    out = self.pv_matmul(s, k_cache, out)    # out is dumped here

For a single task launch, list the targets explicitly with the dumps=[...] kwarg: out, tid = pl.submit(self.qk_matmul, q, k_cache, scratch, dumps=[q]) or with pl.at(..., dumps=[q]) as tid:.

Inside an Inline helper, the recorded dump_vars ride on the inline body's kernel calls; the InlineFunctions pass splices those calls into the caller and the mutator substitutes the caller's arg for each inline parameter, so tags on both inline parameters and inline body-local pl.create_tensor(...) results take effect at the inlined call sites. No tag migration is needed; multi-level inlining works at the pass's fixpoint.

Limitations (MVP):

A tag fires only when the tagged Var reaches a kernel dispatch as a static, whole-tensor Arg. Values that never reach such an Arg are silently not dumped:

Dynamic-offset reads — a value read only through a data-dependent offset (q_flat[runtime_row : runtime_row + N, ...]) lowers to a gather / dynamic-address load, not a whole-tensor Arg, so the tag does not attach. Stage it through a buffer read with static, compile-time tiled offsets and tag that buffer.
Orch-level pl.assemble buffers — y = pl.assemble(y, tile, off) lowers to a pure name-alias and emits no kernel dispatch, so there is no Arg to mark. Use a static in-place slice store y[off_slice] = tile and tag y, or dump the producer kernels' output Args.
Orch-tier scalar reads — a tensor consumed only by orchestration-level pl.read(...) (e.g. block tables read to compute offsets) never enters a device kernel as a Tensor Arg; the MVP runtime path covers per-task device Args only.
Distributed L3+ programs: only chip-level orchestration tasks honour the tag; HOST-tier Python SubWorker tensors are not covered by the runtime's selective dump path.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	The tensor to mark. Must be a `Tensor` value bound in the enclosing orchestration's scope.	required

Returns:

Type	Description
`Tensor`	The tensor unchanged. The marker is consumed at parse time.

`set_cache_policy(tensor, policy)` ¶

Declare the GM cache-access policy for every read of tensor in this scope.

The scope-level surface of the cache policy, for tensor programming — where the GM reads are implicit and there is no pl.load call to annotate. The per-access counterpart is the cache= kwarg on pl.load, which names one read instead of all of them. pl.slice deliberately takes no cache=: a slice computes an address descriptor, it moves no data.

Write it as a standalone statement directly inside a pl.at(...) / pl.spmd(...) scope body. Anywhere else — nested in an if / for inside the scope, or outside a scope altogether — the parser rejects it, because the declaration attaches to the scope, and a conditionally-executed declaration would be a promise the compiler cannot check.

Semantics:

A contract, not a hint. CachePolicy.BYPASS asserts two things about tensor: that the kernel streams it with no reuse worth caching, and that nothing writes those bytes while the kernel runs. Mixing a cached write and a bypassing read of the same bytes is a coherency bug the compiler cannot detect, so coherency is the author's contract. This is why the policy is never a default and never inferred. Declaring BYPASS on a tensor the scope itself writes is rejected at outlining.
Tracked by Var identity, never by name. The declaration names the binding live at the scope, so rebinding the name afterwards (b = self.foo(b)) yields a new value the declaration does not cover.
Consumed at parse time. It emits no IR statement of its own: the parser records it on the enclosing scope, the scope outliner resolves it to the outlined kernel's parameters, and ConvertTensorToTileOps turns it into a cache kwarg on each tile.load that reads a declared parameter.
Explicit wins. An explicit pl.load(..., cache=...) overrides the scope declaration for that one access, in both directions — so cache=pl.CachePolicy.DEFAULT opts a single read back into the cache inside a bypassing scope.

Current status: PTOAS has no L2-bypass path yet (https://github.com/hw-native-sys/PTOAS/issues/1356). The declaration is carried all the way to codegen, but codegen emits a warning and compiles it as an ordinary cached access, so generated code is unchanged today. Writing the declaration now is what makes the kernel pick the bypass up for free once that lands.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	The tensor whose reads the policy applies to. Must be a `Tensor` value bound outside the scope and read inside it.	required
`policy`	`CachePolicy`	`CachePolicy.BYPASS` to declare a streaming, non-cached read; `CachePolicy.DEFAULT` for an ordinary cached read.	required

Returns:

Type	Description
`None`	Nothing. The marker is consumed at parse time and produces no value.

Example

with pl.at(level=pl.Level.CORE_GROUP, name_hint="mm"): ... pl.set_cache_policy(b, pl.CachePolicy.BYPASS) ... c = pl.matmul(a, b, out_dtype=pl.FP32) ... out = pl.assemble(out, c, [0, 0])

`read(tensor, indices)` ¶

Read a scalar value from a tensor at given indices.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor	required
`indices`	`IntLike \| Sequence[IntLike]`	A single index expression (for 1-D flat access) or a list of index expressions (one per tensor dimension)	required

Returns:

Type	Description
`Scalar`	Scalar wrapping the read operation

`write(tensor, indices, value)` ¶

Write a scalar value into a tensor at given indices.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Destination tensor	required
`indices`	`IntLike \| Sequence[IntLike]`	A single index expression (for 1-D flat access) or a list of index expressions (one per tensor dimension)	required
`value`	`Scalar \| Expr`	Scalar value to write (DSL Scalar or raw Expr)	required

Returns:

Type	Description
`Expr`	The underlying `tensor.write` call expression. Direct callers
`Expr`	typically ignore it; the DSL parser surfaces it as an `EvalStmt`.

`dim(tensor, axis)` ¶

Extract a shape dimension from a tensor as a scalar value.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor	required
`axis`	`int \| ConstInt`	Dimension index (supports negative indexing). Accepts either a Python `int` or a `ConstInt` (parser-shape).	required

Returns:

Type	Description
`Scalar`	Scalar wrapping the dim operation (INT64)

`slice(tensor, shape, offset, valid_shape=None, drop_dims=None, pad_value=None, clamp=False)` ¶

Create a slice of a tensor with new shape and optional valid shape.

The slice is never valid where the source is not: the source's valid region, shifted by offset and cut to the window, bounds the result.

Parameters:

Name	Type	Description	Default
`tensor`	`_TensorT`	Input tensor	required
`shape`	`Sequence[IntLike]`	New shape dimensions. Always full-rank — a scalar-indexed axis contributes a unit dim here and is listed in `drop_dims`.	required
`offset`	`Sequence[IntLike]`	Offset dimensions for the slice	required
`valid_shape`	`Sequence[IntLike] \| None`	Valid shape dimensions. When omitted, the source's validity under the window is used. Narrows the result; cannot widen it.	`None`
`drop_dims`	`Sequence[int \| Expr] \| None`	Optional axes to erase from the result type (numpy-style rank reduction). Each listed axis must be a static unit dim of `shape` and must still be fully valid after the intersection above. `None` / `[]` drops nothing (fully backward compatible).	`None`
`pad_value`	`PadValue \| int \| float \| None`	Optional padding mode for out-of-valid-shape elements. `None` means the source's padding mode carries through. Accepts `PadValue.zero` / `PadValue.max` / `PadValue.min`, or the literal sugars `0`, `math.inf`, `-math.inf` (same spelling as `tensor.fillpad`). Only meaningful when the effective valid region is smaller than `shape` — which an explicit `valid_shape`, a partially-valid source, or `clamp=True` can each bring about.	`None`
`clamp`	`bool`	Sanction a window that runs off the end of the source. By default a slice asserts `offset + shape` stays inside the source and is rejected when that provably fails; `clamp=True` lets the window overhang and cuts the valid region back to the source edge. Use it for a fixed-width tail read whose overhang is never addressed.	`False`

Returns:

Type	Description
`_TensorT`	Tensor wrapping the slice operation

`fillpad(tensor, pad_value=PadValue.zero)` ¶

Fill invalid tensor view elements with the specified padding value.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor	required
`pad_value`	`PadValue \| int \| float`	`PadValue` enum (`zero` / `max` / `min`), or one of the literal sugars `0`, `math.inf`, `-math.inf`. Default is `PadValue.zero`. Other values raise — the hardware only supports the three padding modes.	`zero`

Returns:

Type	Description
`Tensor`	Tensor wrapping the fillpad operation

`fillpad_expand(tensor, shape, pad_value=PadValue.zero)` ¶

Copy a smaller source tensor into a larger destination tensor, padding the rest.

Unlike fillpad (which keeps the same shape and only fills the invalid view region), the destination shape may be larger than the source in either dimension. The source's valid region is copied into the top-left of the destination and every other element is filled with pad_value.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Source tensor	required
`shape`	`Sequence[IntLike]`	Destination shape; each dimension must be >= the source dimension	required
`pad_value`	`PadValue \| int \| float`	`PadValue` enum (`zero` / `max` / `min`), or one of the literal sugars `0`, `math.inf`, `-math.inf`. Default is `PadValue.zero`. Other values raise — the hardware only supports the three padding modes.	`zero`

Returns:

Type	Description
`Tensor`	Tensor wrapping the fillpad_expand operation (a new, larger tensor).

`full(shape, dtype, value)` ¶

Create a tensor of specified shape filled with a constant value.

Parameters:

Name	Type	Description	Default
`shape`	`Sequence[IntLike]`	Shape of the tensor	required
`dtype`	`DataType`	Data type of tensor elements	required
`value`	`int \| float`	Filling scalar value (int or float)	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the full operation

`ci(start, shape, dtype=DataType.INT32, descending=False)` ¶

Generate a contiguous integer sequence into a tensor.

Equivalent to numpy.arange / torch.arange. Lowers to tile.ci → pto.tci.

Note

pto.tci only populates the first row. Leading dimensions must be 1 — prefer shapes of the form [1, N].

Parameters:

Name	Type	Description	Default
`start`	`int \| Scalar`	Starting integer (plain int or Scalar). Must match `dtype`.	required
`shape`	`Sequence[IntLike]`	Destination tensor shape (innermost dim != 1).	required
`dtype`	`DataType`	Destination dtype. One of {INT16, INT32}. Defaults to INT32.	`INT32`
`descending`	`bool`	If True, generate a descending sequence.	`False`

Returns:

Type	Description
`Tensor`	Tensor wrapping the ci operation.

`arange = ci` `module-attribute` ¶

`random(key0, key1, counter0, counter1, counter2, counter3, shape, dtype=DataType.UINT32, rounds=10)` ¶

Generate counter-based pseudo-random values into a tensor.

Implements a counter-based (Philox/ChaCha-style) RNG. Each element is derived deterministically from the 64-bit key (key0, key1) and 128-bit counter (counter0..counter3) plus the element position, so the same seeds always reproduce the same tensor. Lowers to tile.random → pto.trandom.

Parameters:

Name	Type	Description	Default
`key0`	`int \| Scalar`	Low INT32 key word (plain int or Scalar).	required
`key1`	`int \| Scalar`	High INT32 key word (plain int or Scalar).	required
`counter0`	`int \| Scalar`	First INT32 counter word.	required
`counter1`	`int \| Scalar`	Second INT32 counter word.	required
`counter2`	`int \| Scalar`	Third INT32 counter word.	required
`counter3`	`int \| Scalar`	Fourth INT32 counter word.	required
`shape`	`Sequence[IntLike]`	Destination tensor shape (static).	required
`dtype`	`DataType`	Destination dtype. One of {INT32, UINT32}. Defaults to UINT32.	`UINT32`
`rounds`	`int`	Cipher round count, 7 or 10. Defaults to 10.	`10`

Returns:

Type	Description
`Tensor`	Tensor wrapping the random operation.

`matmul(lhs, rhs, out_dtype=None, a_trans=False, b_trans=False, c_matrix_nz=False)` ¶

Matrix multiplication with optional transpose.

A transpose flag swaps its own operand's two trailing axes, so that operand must be at least 2D: a_trans with a 1D lhs (or b_trans with a 1D rhs) raises rather than being ignored. On the mixed mat-vec / vec-mat forms the flag applies to the matrix side, so a lhs stored [K, M] with a_trans=True against a [K] rhs deduces [M].

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`Tensor`	Right-hand side tensor	required
`out_dtype`	`int \| DataType \| None`	Output data type. Optional; inferred from the operands when omitted. When given it must be one of the dtypes the Cube writeback can produce: FP32/FP16/BF16 for float operands, INT32 for int operands (see `pl.matmul`)	`None`
`a_trans`	`bool`	Whether to transpose lhs (requires a 2D+ lhs)	`False`
`b_trans`	`bool`	Whether to transpose rhs (requires a 2D+ rhs)	`False`
`c_matrix_nz`	`bool`	C matrix non-zero flag	`False`

Returns:

Type	Description
`Tensor`	Tensor wrapping the matmul operation

`matmul_acc(acc, lhs, rhs, a_trans=False, b_trans=False, init_cond=None)` ¶

Matrix multiplication with accumulation: acc += lhs @ rhs.

init_cond makes the accumulator's initial value conditional: on the steps where it holds, acc is overwritten with lhs @ rhs rather than accumulated into. This is the split-K idiom, and it removes the need to zero the accumulator or to peel the first K step::

for k0 in pl.pipeline(0, K, K_TILE, stage=2):
    acc[t0 : t0 + R, :] = pl.matmul_acc(
        acc[t0 : t0 + R, :], x_k, w_k, b_trans=True, init_cond=(k0 == 0)
    )

Only 2D operands support the predicate; loop over the batch dimension instead of passing higher-rank operands alongside init_cond.

Parameters:

Name	Type	Description	Default
`acc`	`Tensor`	Accumulator tensor	required
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`Tensor`	Right-hand side tensor	required
`a_trans`	`bool`	Whether to transpose lhs	`False`
`b_trans`	`bool`	Whether to transpose rhs	`False`
`init_cond`	`BoolLike \| None`	Optional predicate selecting overwrite over accumulate	`None`

Returns:

Type	Description
`Tensor`	Tensor wrapping the matmul_acc operation

`mul(lhs, rhs)` ¶

Element-wise multiplication of tensor and tensor or scalar.

Automatically selects between tensor.mul (tensor x tensor) and tensor.muls (tensor x scalar) based on the rhs type.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`int \| float \| Tensor \| Scalar \| Expr`	Right-hand side tensor or scalar (int/float/Tensor/Scalar)	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the mul operation

`muls(lhs, rhs)` ¶

Element-wise multiplication of tensor and scalar.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`int \| float \| Expr \| Scalar`	Right-hand side scalar (int/float/Expr with ScalarType)	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the muls operation

`add(lhs, rhs)` ¶

Element-wise addition of tensor and tensor or scalar.

Automatically selects between tensor.add (tensor + tensor) and tensor.adds (tensor + scalar) based on the rhs type.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`int \| float \| Tensor \| Scalar \| Expr`	Right-hand side tensor or scalar (int/float/Tensor/Scalar)	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the add operation

`adds(lhs, rhs)` ¶

Element-wise addition of tensor and scalar.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`int \| float \| Expr \| Scalar`	Right-hand side scalar (int/float/Expr with ScalarType)	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the adds operation

`sub(lhs, rhs)` ¶

Element-wise subtraction of tensor and tensor or scalar.

Automatically selects between tensor.sub (tensor - tensor) and tensor.subs (tensor - scalar) based on the rhs type.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`int \| float \| Tensor \| Scalar \| Expr`	Right-hand side tensor or scalar (int/float/Tensor/Scalar)	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the sub operation

`subs(lhs, rhs)` ¶

Element-wise subtraction of tensor and scalar.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`int \| float \| Expr \| Scalar`	Right-hand side scalar (int/float/Expr with ScalarType)	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the subs operation

`div(lhs, rhs, high_precision=False)` ¶

Element-wise division of tensor and tensor or scalar.

Automatically selects between tensor.div (tensor / tensor) and tensor.divs (tensor / scalar) based on the rhs type.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`int \| float \| Tensor \| Scalar \| Expr`	Right-hand side tensor or scalar (int/float/Tensor/Scalar)	required
`high_precision`	`bool`	Whether to select PTOAS's high-precision division mode. Only available when `rhs` is a Tensor.	`False`

Returns:

Type	Description
`Tensor`	Tensor wrapping the div operation

`divs(lhs, rhs)` ¶

Element-wise division of tensor and scalar.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`int \| float \| Expr \| Scalar`	Right-hand side scalar (int/float/Expr/Scalar)	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the divs operation

`part_add(lhs, rhs)` ¶

Partial element-wise add of two tensors.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	First source tensor	required
`rhs`	`Tensor`	Second source tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the part_add operation

`part_mul(lhs, rhs)` ¶

Partial element-wise multiply of two tensors.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	First source tensor	required
`rhs`	`Tensor`	Second source tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the part_mul operation

`part_max(lhs, rhs)` ¶

Partial element-wise max of two tensors.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	First source tensor	required
`rhs`	`Tensor`	Second source tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the part_max operation

`part_min(lhs, rhs)` ¶

Partial element-wise min of two tensors.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	First source tensor	required
`rhs`	`Tensor`	Second source tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the part_min operation

`fmod(lhs, rhs)` ¶

Element-wise truncating remainder of tensor and tensor or scalar.

Automatically selects between tensor.fmod (tensor, tensor) and tensor.fmods (tensor, scalar) based on the rhs type. The result matches torch.fmod (the remainder takes the sign of the dividend).

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`int \| float \| Tensor \| Scalar \| Expr`	Right-hand side tensor or scalar (int/float/Tensor/Scalar)	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the fmod operation

`fmods(lhs, rhs)` ¶

Element-wise truncating remainder of tensor and scalar.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`int \| float \| Expr \| Scalar`	Right-hand side scalar (int/float/Expr/Scalar)	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the fmods operation

`maximum(lhs, rhs)` ¶

Element-wise maximum of tensor and tensor or scalar.

The conversion pass handles the tensor-vs-tensor / tensor-vs-scalar dispatch internally — there is no separate maximums front-end op.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`int \| float \| Tensor \| Scalar \| Expr`	Right-hand side tensor or scalar	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the maximum operation

`minimum(lhs, rhs)` ¶

Element-wise minimum of tensor and tensor or scalar.

The conversion pass handles the tensor-vs-tensor / tensor-vs-scalar dispatch internally — there is no separate minimums front-end op.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`int \| float \| Tensor \| Scalar \| Expr`	Right-hand side tensor or scalar	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the minimum operation

`cmp(lhs, rhs, cmp_type=0)` ¶

Element-wise comparison of tensor and tensor or scalar (returns 0/1 tensor).

The conversion pass handles the tensor-vs-tensor / tensor-vs-scalar dispatch internally — there is no separate cmps front-end op.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor	required
`rhs`	`int \| float \| Tensor \| Scalar \| Expr`	Right-hand side tensor or scalar	required
`cmp_type`	`int`	Comparison type code (0=eq, 1=ne, 2=lt, 3=le, 4=gt, 5=ge)	`0`

Returns:

Type	Description
`Tensor`	Tensor of 0/1 with the same shape and dtype as `lhs`

`and_(lhs, rhs)` ¶

Element-wise bitwise AND of tensor and tensor or scalar.

Automatically selects between tensor.and (tensor & tensor) and tensor.ands (tensor & scalar) based on the rhs type.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor (integer dtype)	required
`rhs`	`int \| Tensor \| Scalar \| Expr`	Right-hand side tensor or integer scalar	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the and operation

`ands(lhs, rhs)` ¶

Element-wise bitwise AND of tensor and scalar.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor (integer dtype)	required
`rhs`	`int \| Expr \| Scalar`	Right-hand side integer scalar	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the ands operation

`or_(lhs, rhs)` ¶

Element-wise bitwise OR of tensor and tensor or scalar.

Automatically selects between tensor.or (tensor | tensor) and tensor.ors (tensor | scalar) based on the rhs type.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor (integer dtype)	required
`rhs`	`int \| Tensor \| Scalar \| Expr`	Right-hand side tensor or integer scalar	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the or operation

`ors(lhs, rhs)` ¶

Element-wise bitwise OR of tensor and scalar.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor (integer dtype)	required
`rhs`	`int \| Expr \| Scalar`	Right-hand side integer scalar	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the ors operation

`xor(lhs, rhs)` ¶

Element-wise bitwise XOR of tensor and tensor or scalar.

Automatically selects between tensor.xor (tensor ^ tensor) and tensor.xors (tensor ^ scalar) based on the rhs type.

Unlike pl.tile.xor, there is no tmp parameter: the scratch buffer that pto.txor needs is allocated during Tensor-to-Tile lowering.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor (integer dtype)	required
`rhs`	`int \| Tensor \| Scalar \| Expr`	Right-hand side tensor or integer scalar	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the xor operation

`xors(lhs, rhs)` ¶

Element-wise bitwise XOR of tensor and scalar.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor (integer dtype)	required
`rhs`	`int \| Expr \| Scalar`	Right-hand side integer scalar	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the xors operation

`not_(input)` ¶

Element-wise bitwise NOT of a tensor.

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor; int16 or uint16, matching the TNOT instruction	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the not operation

`shl(lhs, rhs)` ¶

Element-wise bitwise left shift of tensor by tensor or scalar.

Automatically selects between tensor.shl (tensor << tensor) and tensor.shls (tensor << scalar) based on the rhs type.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor (integer dtype)	required
`rhs`	`int \| Tensor \| Scalar \| Expr`	Shift amount as a tensor or integer scalar	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the shl operation

`shls(lhs, rhs)` ¶

Element-wise bitwise left shift of tensor by scalar.

Note

The shift amount must be zero or positive. A negative constant is rejected when the op is built; a negative value only known at runtime is undefined behaviour on the hardware.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor (integer dtype)	required
`rhs`	`int \| Expr \| Scalar`	Shift amount; must be >= 0	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the shls operation

`shr(lhs, rhs)` ¶

Element-wise bitwise right shift of tensor by tensor or scalar.

Automatically selects between tensor.shr (tensor >> tensor) and tensor.shrs (tensor >> scalar) based on the rhs type. The shift is arithmetic for signed dtypes and logical for unsigned ones, matching the tile ops and the underlying ISA.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor (integer dtype)	required
`rhs`	`int \| Tensor \| Scalar \| Expr`	Shift amount as a tensor or integer scalar	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the shr operation

`shrs(lhs, rhs)` ¶

Element-wise bitwise right shift of tensor by scalar.

Note

The shift amount must be zero or positive. A negative constant is rejected when the op is built; a negative value only known at runtime is undefined behaviour on the hardware.

Parameters:

Name	Type	Description	Default
`lhs`	`Tensor`	Left-hand side tensor (integer dtype)	required
`rhs`	`int \| Expr \| Scalar`	Shift amount; must be >= 0	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the shrs operation

`row_max(input)` ¶

Row-wise max reduction (reduces along last axis, keeps dim).

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_max operation

`row_sum(input)` ¶

Row-wise sum reduction (reduces along last axis, keeps dim).

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_sum operation

`row_min(input)` ¶

Row-wise min reduction (reduces along last axis, keeps dim).

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_min operation

`row_prod(input)` ¶

Row-wise product reduction (reduces along last axis, keeps dim).

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_prod operation

`col_sum(input)` ¶

Column-wise sum reduction (reduces along axis=-2, keeps dim).

Output shape is [..., 1, N] for an input of shape [..., M, N].

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_sum operation

`col_max(input)` ¶

Column-wise max reduction (reduces along axis=-2, keeps dim).

Output shape is [..., 1, N] for an input of shape [..., M, N].

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_max operation

`col_min(input)` ¶

Column-wise min reduction (reduces along axis=-2, keeps dim).

Output shape is [..., 1, N] for an input of shape [..., M, N].

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_min operation

`col_prod(input)` ¶

Column-wise product reduction (reduces along axis=-2, keeps dim).

Output shape is [..., 1, N] for an input of shape [..., M, N].

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_prod operation

`row_argmax(input)` ¶

Row-wise argmax: index of the per-row maximum (int32, reduces along last axis).

Output shape is [..., M, 1] for an input of shape [..., M, N].

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_argmax operation

`row_argmin(input)` ¶

Row-wise argmin: index of the per-row minimum (int32, reduces along last axis).

Output shape is [..., M, 1] for an input of shape [..., M, N].

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_argmin operation

`col_argmax(input)` ¶

Column-wise argmax: index of the per-column maximum (int32, reduces along axis=-2).

Output shape is [..., 1, N] for an input of shape [..., M, N].

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_argmax operation

`col_argmin(input)` ¶

Column-wise argmin: index of the per-column minimum (int32, reduces along axis=-2).

Output shape is [..., 1, N] for an input of shape [..., M, N].

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_argmin operation

`row_expand(target, row_vec)` ¶

Row-wise expansion: expand row_vec [M, 1] to target shape [M, N].

Parameters:

Name	Type	Description	Default
`target`	`Tensor`	Target tensor defining output shape (TensorType [M, N])	required
`row_vec`	`Tensor`	Row vector to expand (TensorType [M, 1])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_expand operation

`row_expand_mul(tensor, row_vec)` ¶

Row-wise broadcast multiplication: tensor[i,:] * row_vec[i,0].

Multiplies each row of the tensor by the corresponding row vector value, for all i.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`row_vec`	`Tensor`	Row vector (TensorType [M, 1])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_expand_mul operation

`row_expand_div(tensor, row_vec)` ¶

Row-wise broadcast division: tensor[i,:] / row_vec[i,0].

Divides each row of the tensor by the corresponding row vector value, for all i.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`row_vec`	`Tensor`	Row vector (TensorType [M, 1])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_expand_div operation

`row_expand_add(tensor, row_vec)` ¶

Row-wise broadcast addition: tensor[i,:] + row_vec[i,0].

Adds a row vector to each row of the tensor, for all i.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`row_vec`	`Tensor`	Row vector (TensorType [M, 1])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_expand_add operation

`row_expand_sub(tensor, row_vec)` ¶

Row-wise broadcast subtraction: tensor[i,:] - row_vec[i,0].

Subtracts a row vector from each row of the tensor, for all i.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`row_vec`	`Tensor`	Row vector (TensorType [M, 1])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_expand_sub operation

`row_expand_max(tensor, row_vec)` ¶

Row-wise broadcast maximum: max(tensor[i,:], row_vec[i,0]).

Takes the element-wise maximum of each row and the row vector value, for all i.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`row_vec`	`Tensor`	Row vector (TensorType [M, 1])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_expand_max operation

`row_expand_min(tensor, row_vec)` ¶

Row-wise broadcast minimum: min(tensor[i,:], row_vec[i,0]).

Takes the element-wise minimum of each row and the row vector value, for all i.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`row_vec`	`Tensor`	Row vector (TensorType [M, 1])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_expand_min operation

`row_expand_expdif(tensor, row_vec)` ¶

Row-wise exp-diff: exp(tensor[i,:] - row_vec[i,0]).

Computes the exponential of the per-row difference, for all i.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`row_vec`	`Tensor`	Row vector providing per-row scalar (TensorType [M, 1])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the row_expand_expdif operation

`col_expand_mul(tensor, col_vec)` ¶

Column-wise broadcast multiplication: tensor[:,j] * col_vec[0,j].

Multiplies each column of the tensor by the corresponding column vector value, for all j.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`col_vec`	`Tensor`	Column vector (TensorType [1, N])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_expand_mul operation

`col_expand(tensor, col_vec)` ¶

Column-wise expansion: expand col_vec [1, N] to target shape [M, N].

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Target tensor defining output shape (TensorType [M, N])	required
`col_vec`	`Tensor`	Column vector to expand (TensorType [1, N])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_expand operation

`col_expand_div(tensor, col_vec)` ¶

Column-wise broadcast division: tensor[:,j] / col_vec[0,j].

Divides each column of the tensor by the corresponding column vector value, for all j.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`col_vec`	`Tensor`	Column vector (TensorType [1, N])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_expand_div operation

`col_expand_sub(tensor, col_vec)` ¶

Column-wise broadcast subtraction: tensor[:,j] - col_vec[0,j].

Subtracts a column vector from each column of the tensor, for all j.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`col_vec`	`Tensor`	Column vector (TensorType [1, N])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_expand_sub operation

`col_expand_add(tensor, col_vec)` ¶

Column-wise broadcast addition: tensor[:,j] + col_vec[0,j].

Adds a column vector to each column of the tensor, for all j.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`col_vec`	`Tensor`	Column vector (TensorType [1, N])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_expand_add operation

`col_expand_max(tensor, col_vec)` ¶

Column-wise broadcast maximum: max(tensor[:,j], col_vec[0,j]).

Takes the element-wise maximum of each column and the column vector value, for all j.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`col_vec`	`Tensor`	Column vector (TensorType [1, N])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_expand_max operation

`col_expand_min(tensor, col_vec)` ¶

Column-wise broadcast minimum: min(tensor[:,j], col_vec[0,j]).

Takes the element-wise minimum of each column and the column vector value, for all j.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`col_vec`	`Tensor`	Column vector (TensorType [1, N])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_expand_min operation

`col_expand_expdif(tensor, col_vec)` ¶

Column-wise exp-diff: exp(tensor[:,j] - col_vec[0,j]).

Computes the exponential of the per-column difference, for all j.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (TensorType [M, N])	required
`col_vec`	`Tensor`	Column vector providing per-column scalar (TensorType [1, N])	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the col_expand_expdif operation

`expands(target, scalar)` ¶

Expand scalar to target tensor shape.

Parameters:

Name	Type	Description	Default
`target`	`Tensor`	Target tensor defining output shape	required
`scalar`	`int \| float \| Scalar`	Scalar value to expand	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the expands operation

`expand_clone(src, target)` ¶

Clone and expand input to target shape.

Parameters:

Name	Type	Description	Default
`src`	`Tensor`	Source tensor	required
`target`	`Tensor`	Target tensor defining output shape	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the expand_clone operation

`exp(input)` ¶

Element-wise exponential operation.

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the exp operation

`log(input, high_precision=False)` ¶

Element-wise natural logarithm operation.

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required
`high_precision`	`bool`	Whether to select PTOAS's high-precision logarithm mode	`False`

Returns:

Type	Description
`Tensor`	Tensor wrapping the log operation

`sin(input)` ¶

Element-wise sine operation (input in radians). FP32-only.

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor (must be FP32; cast explicitly otherwise)	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the sin operation

`cos(input)` ¶

Element-wise cosine operation (input in radians). FP32-only.

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor (must be FP32; cast explicitly otherwise)	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the cos operation

`neg(input)` ¶

Element-wise negation operation.

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the neg operation

`abs(input)` ¶

Element-wise absolute value operation.

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the abs operation

`recip(input, high_precision=False)` ¶

Element-wise reciprocal (1/x) operation.

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required
`high_precision`	`bool`	Whether to select PTOAS's high-precision reciprocal mode (FP16/FP32 only)	`False`

Returns:

Type	Description
`Tensor`	Tensor wrapping the recip operation

`sqrt(input)` ¶

Element-wise square root operation.

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the sqrt operation

`rsqrt(input, high_precision=False)` ¶

Element-wise reciprocal square root operation.

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required
`high_precision`	`bool`	If True, lower to the higher-precision PTO path. The compiler allocates a scratch buffer during tensor-to-tile conversion.	`False`

Returns:

Type	Description
`Tensor`	Tensor wrapping the rsqrt operation

`cast(input, target_type, mode='round', *, saturation_mode=None)` ¶

Type casting operation.

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Input tensor	required
`target_type`	`int \| DataType`	Target data type	required
`mode`	`str \| int`	Rounding mode — string name ("none", "rint", "round", "floor", "ceil", "trunc", "odd") or int (0–6)	`'round'`
`saturation_mode`	`str \| int \| None`	Destination saturation — `"on"` (1) clamps a rounded value that falls outside the destination range to that range; `"off"` (0) keeps the target's non-saturating conversion, including its overflow and non-finite behavior. Defaults to `"on"` for an integer destination: nothing standard fixes what an overflowing conversion to an integer produces, clamping is the safer of the two to get by accident, and it is what the hardware converts natively. A float destination keeps the target's own IEEE behavior (an out-of-range narrowing yields an infinity) unless you ask otherwise. When the cast lowers to a chain of native conversions, the mode applies to the final hop.	`None`

Returns:

Type	Description
`Tensor`	Tensor wrapping the cast operation

`assemble(target, source, offset, *, atomic=AtomicType.None_)` ¶

Write/update tensor values at specified offset.

Parameters:

Name	Type	Description	Default
`target`	`_TensorT`	Target tensor to update	required
`source`	`Tensor`	Source tensor to write	required
`offset`	`Sequence[IntLike]`	Offset dimensions for where to write	required
`atomic`	`AtomicType`	Combine mode for the write. `AtomicType.None_` (default) overwrites; `AtomicType.Add` atomically adds `source` into the target at `offset` — used for split-K, where several cores accumulate partial products into one output. Only valid inside an InCore function — typically a `pl.at(level=pl.Level.CORE_GROUP, ...)` scope — where `source` lowers to an on-chip tile (a compute result) and the write lowers to an atomic-add store into a global-memory target (a function output tensor). Every other form is rejected at compile time: a tensor-to-tensor assemble has no store to carry the combine, an assemble into an on-chip tile has no global-memory destination, and at the orchestration level no atomic-combine instruction exists at all. NOTE: atomic-add accumulation order across cores is not fixed, so floating-point results are non-deterministic. The target must be zero-initialised before the kernel runs. Supported dtypes: fp32 / bf16 / fp16 / int32 / int16 / int8. bf16 atomic-add is available on the Ascend910B (A2/A3) profile; it is not supported on A5, where an fp32 accumulator + cast is required instead.	`None_`

Returns:

Type	Description
`_TensorT`	Tensor wrapping the assemble operation

`concat(src0, src1)` ¶

Concatenate two tensors along the column dimension.

Parameters:

Name	Type	Description	Default
`src0`	`Tensor`	First source tensor	required
`src1`	`Tensor`	Second source tensor	required

Returns:

Type	Description
`Tensor`	Tensor with concatenated columns

`reshape(tensor, shape)` ¶

Reshape tensor to new shape.

The valid region is carried through, never widened: the result holds real data in exactly the cells the input did. See pl.reshape for the cases that always map.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor	required
`shape`	`Sequence[IntLike]`	New shape dimensions	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the reshape operation

Raises:

Type	Description
`ValueError`	If the element count changes, or if the input holds real data in only part of its buffer and no origin-anchored region of `shape` describes those same cells.

`reinterpret_view(data, dtype, *, shape=None)` ¶

Reinterpret a tensor over the same bytes with a different dtype.

Parameters:

Name	Type	Description	Default
`data`	`Tensor`	Input tensor.	required
`dtype`	`DataType`	Target element dtype, which must differ from the source dtype.	required
`shape`	`Sequence[IntLike] \| None`	Optional byte-equivalent target shape. When omitted, the physically contiguous dimension is scaled according to the source/target dtype byte ratio.	`None`

Returns:

Type	Description
`Tensor`	Tensor wrapping the zero-copy reinterpret-view operation.

`transpose(tensor, axis1, axis2)` ¶

Transpose tensor by swapping two axes.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor	required
`axis1`	`int`	First axis to swap (supports negative indexing)	required
`axis2`	`int`	Second axis to swap (supports negative indexing)	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the transpose operation

`view(tensor, shape=None, valid_shape=None, *, layout=None)` ¶

Reinterpret a tensor over the same physical memory.

At least one of shape or layout must be provided: shape derives canonical strides for the requested shape, layout derives the canonical ND/DN layout view. The result is a zero-copy view over the same physical memory, and its target shape must have rank at least 1.

Parameters:

Name	Type	Description	Default
`tensor`	`_TensorT`	Source tensor.	required
`shape`	`Sequence[IntLike] \| None`	New shape for the view. Must be product-preserving unless symbolic dimensions are present. Rank-zero views are not supported.	`None`
`valid_shape`	`Sequence[IntLike] \| None`	Explicit valid dimensions for a packed ND leading-dimension collapse to 2D or contiguous-prefix linear collapse to `[1, product(shape)]`. Required when either supported collapse reinterprets a source with partial validity.	`None`
`layout`	`TensorLayout \| None`	Target `TensorLayout` (ND or DN); DN requires rank at least 2. Orchestration supports ND shape reinterprets and shaped ND/MX_A_ZZ/MX_B_NN backing and consumer views for dynamic FP8E8M0 MX scales. Other layout-changing shape views are limited to in-core lowering.	`None`

Returns:

Type	Description
`_TensorT`	Tensor wrapping the view operation.

Raises:

Type	Description
`ValueError`	If the requested shape/layout is missing, unsupported, or inconsistent with the source tensor metadata.

`scatter_update(input, *args, **kwargs)` ¶

Update input tensor rows at positions specified by 2D index with values from src.

Supports two rank variants:

2D: input [rows, d], src [b*s, d], index [b, s]
4D: input [blockNum, blockSize, 1, d], src [b, s, 1, d], index [b, s]

For each (i, j), row input[index[i*s + j]] receives row src[i*s + j] (linear layout).

Accepts the same flexible call shapes as the IR builder pypto.ir.op.tensor.scatter_update:

scatter_update(input, dim, index, src)
scatter_update(input, index, src, dim=-2)
scatter_update(input, dim, index=..., src=...)

Tensor / Scalar wrappers are unwrapped before forwarding so the IR builder receives raw Expr operands.

`set_validshape(tensor, valid_rows, valid_cols)` ¶

Update valid-shape metadata of a tensor without data movement.

.. note:: Prefer expressing the extent at its source where possible — pl.load(..., valid_shape=...) or a slice's valid_shape= — and use this to pin an extent the type deducer cannot infer.

Parameters:

Name	Type	Description	Default
`tensor`	`Tensor`	Input tensor (must be 2D)	required
`valid_rows`	`IntLike`	Number of valid rows (int or Scalar[INDEX])	required
`valid_cols`	`IntLike`	Number of valid columns (int or Scalar[INDEX])	required

Returns:

Type	Description
`Tensor`	Tensor with updated valid_shape metadata

`sort32(src, idx)` ¶

Sort fixed 32-element blocks with explicit index tensor (tensor-level).

Tensor-level counterpart of pl.tile.sort32. Sorts 32-element blocks in src, permuting idx alongside. Returns an 8-byte value-index-pair tensor; its last dimension is 2x the input width for FP32 and 4x for FP16.

For FP16 src: initialize idx with [0, 1, 2, ..., 31] per block. For FP32 src: initialize idx with [0, 2, 4, ..., 62] per block.

Parameters:

Name	Type	Description	Default
`src`	`Tensor`	Input value tensor (FP16 or FP32)	required
`idx`	`Tensor`	Input index tensor with sequential offsets	required

Returns:

Type	Description
`Tensor`	Tensor wrapping the dtype-dependent expanded sort32 output

`mrgsort(src0, src1=None, src2=None, src3=None, *, exhausted=False, block_len=None)` ¶

mrgsort(src0: Tensor, *, block_len: int | Scalar) -> Tensor

mrgsort(src0: Tensor, src1: Tensor, *, exhausted: bool = ...) -> Tensor

mrgsort(src0: Tensor, src1: Tensor, src2: Tensor, *, exhausted: bool = ...) -> Tensor

mrgsort(src0: Tensor, src1: Tensor, src2: Tensor, src3: Tensor, *, exhausted: bool = ...) -> Tensor

Merge sort — format1 (single-list) or format2 (2-4 way merge), tensor-level.

Tensor-level counterpart of pl.tile.mrgsort. The scratch tmp and executed tiles required by the tile-level op are synthesized automatically during conversion as local Vec tiles — users do not pass them.

Format1 usage (keyword block_len): out = mrgsort(src, block_len=64)

Format2 usage

out = mrgsort(src0, src1) # 2-way out = mrgsort(src0, src1, src2) # 3-way out = mrgsort(src0, src1, src2, src3) # 4-way out = mrgsort(src0, src1, exhausted=True)

Parameters:

Name	Type	Description	Default
`src0`	`Tensor`	For format1: input tensor with pre-sorted runs (FP16 or FP32). For format2: first sorted input tensor.	required
`src1`	`Tensor \| None`	(format2) Second sorted input tensor.	`None`
`src2`	`Tensor \| None`	(format2, optional) Third sorted input tensor (3-way or 4-way).	`None`
`src3`	`Tensor \| None`	(format2, optional) Fourth sorted input tensor (4-way only).	`None`
`exhausted`	`bool`	(format2) If True, marks inputs as exhausted (default: False).	`False`
`block_len`	`int \| Scalar \| None`	(format1, keyword-only) Run length, must be multiple of 64.	`None`

Returns:

Type	Description
`Tensor`	Tensor with merged sorted elements

`gather(input, dim=None, index=None, *, mask_pattern=None, output_dtype=None, kvalue=None, cmp_mode=None, out_cols=None, offset=0, count_dtype=None)` ¶

gather(input: Tensor, dim: int, index: Tensor) -> Tensor

gather(input: Tensor | Tile, dim: Tensor | Tile) -> Tensor

gather(input: Tensor | Tile, *, index: Tensor | Tile) -> Tensor

gather(input: Tensor, *, mask_pattern: int, output_dtype: int | DataType | None = None) -> Tensor

gather(input: Tensor, *, kvalue: int | Scalar | Expr, cmp_mode: str | int, out_cols: int, offset: int = 0, count_dtype: int | DataType | None = None) -> tuple[Tensor, Tensor]

Gather elements of input — flat / axis / mask / compare form.

The tensor layer exposes a single unified gather. Based on the arguments you pass, it selects the matching form and lowering:

Flat form (index, no dim)::

output = input.reshape(-1)[index]

Also accepts ``pl.gather(input, index)``. Runtime indices are 2D INT32;
shape and valid shape follow ``index``, dtype follows ``input``
(FP16/FP32/INT16/INT32). Indices must address valid source elements;
negative indexing and bounds checking are unsupported.
Contiguous ND GM sources use [`pl.tile.mgather`][pypto.language.tile.mgather];
static 2D unboxed row-major Vec sources use
[`pl.tile.gather`][pypto.language.tile.gather] with managed packing/scratch.
On-chip source rows must be 32-byte aligned unless there is only one row.
GM operands accept local distributed windows; tile indices must be
unboxed row-major Vec. Physical index columns must be positive static
multiples of 16 for FP16/INT16, or 8 for FP32/INT32, including single-row
tiles. Pad physical storage and use ``set_validshape`` for narrower
valid regions, which need not be aligned.

Axis form (dim + index) → pl.tile.gather, for example dim=1::

output[b, k] = input[b, index[b, k]]

Lowering supports rank-2/rank-3 inputs and any axis, including negative axes.
``index`` must be an INT32 tensor, or INT16 when ``input`` is a 16-bit
dtype (FP16/INT16; INT16 indices require A5). Its rank matches ``input``;
non-gather extents cannot exceed the source. Output shape follows ``index``
and dtype follows ``input``.

Mask form (mask_pattern=<int>) → pl.tile.gather_mask: Selects columns of each row by a fixed hardware mask pattern. Last-dim shrinks by 2 (P0101/P1010) or 4 (P0001..P1000), or stays the same for P1111.

Compare form (kvalue + cmp_mode + out_cols) → pl.tile.gather_compare: Scalar threshold compare (applied to every row). Returns (dst, cdst) — gathered indices [rows, out_cols] INT32 and per-row match counts [1, rows] count_dtype.

Parameters:

Name	Type	Description	Default
`input`	`Tensor \| Tile`	Source tensor (FP16/FP32/INT16/INT32); flat form also accepts a Tile.	required
`dim`	`int \| Tensor \| Tile \| None`	Axis to gather along; omit for flat indexing. A tensor/tile in this positional slot is interpreted as the flat index.	`None`
`index`	`Tensor \| Tile \| None`	Flat form: 2D INT32 tensor/tile. Axis form: tensor with the same rank as `input` and the index dtype constraints above.	`None`
`mask_pattern`	`int \| None`	(mask form, keyword-only) Mask pattern selector (1-7). 1=P0101, 2=P1010, 3=P0001, 4=P0010, 5=P0100, 6=P1000, 7=P1111.	`None`
`output_dtype`	`int \| DataType \| None`	(mask form, keyword-only) Optional output dtype with the same bit width as `input.dtype` (e.g. FP32 → UINT32 for sort32 index bits).	`None`
`kvalue`	`int \| Scalar \| Expr \| None`	(compare form, keyword-only) Scalar threshold (dtype must match `input`).	`None`
`cmp_mode`	`str \| int \| None`	(compare form, keyword-only) `"eq"`/`"ne"`/`"lt"`/`"le"`/ `"gt"`/`"ge"` or int 0..5.	`None`
`out_cols`	`int \| None`	(compare form, keyword-only) Output column count per row.	`None`
`offset`	`int`	(compare form, keyword-only) Starting index offset (default 0).	`0`
`count_dtype`	`int \| DataType \| None`	(compare form, keyword-only) Per-row count dtype, INT32 or UINT32 (defaults to INT32).	`None`

Returns:

Type	Description
`Tensor \| tuple[Tensor, Tensor]`	Tensor (index/mask form) or `(dst, cdst)` tuple (compare form).

Examples:

out = gather(input, index=flat_idx) out = gather(input, flat_idx) out = gather(input, dim=-1, index=idx) out = gather(input, mask_pattern=1) out = gather(input, mask_pattern=pl.tile.MaskPattern.P1010, output_dtype=pl.UINT32) dst, cdst = gather(input, kvalue=kv, cmp_mode="eq", out_cols=8)

`paged_gather(src, indices, block_table, block_size, size, max_indices, *, space=MemorySpace.Mat, col_off=0, is_trans=False, is_b_matrix=False)` ¶

Paged gather directly into an on-chip buffer (L1 by default, or UB).

Gathers scattered rows of a 2D paged KV pool src selected by indices, translated through a paged block_table, directly into an L1 (space=Mat, default) or UB (space=Vec) tile — so a subsequent matmul reads from L1 without a GM round-trip. The lowering is a fully-scalar per-row GM -> on-chip load loop on the Cube core (the bulk KV never touches UB).

Physical row per logical index idx::

phys = block_table[idx // block_size] * block_size + idx % block_size

Parameters:

Name	Type	Description	Default
`src`	`Tensor`	Paged KV pool in GM (2D; FP16/BF16/FP32/INT8).	required
`indices`	`Tensor`	Logical row indices (INT32; 1D `[n]` or 2D `[1, n]`).	required
`block_table`	`Tensor`	Page table mapping logical block -> physical block (INT32).	required
`block_size`	`int`	Number of tokens per page block.	required
`size`	`int`	Number of elements gathered per row (<= src columns).	required
`max_indices`	`int`	Static upper bound on gathered rows; sizes the on-chip tile.	required
`space`	`MemorySpace`	Destination space — `MemorySpace.Mat` (L1, default) or `MemorySpace.Vec` (UB).	`Mat`
`col_off`	`int`	Column start offset within each src row (default 0).	`0`
`is_trans`	`bool`	Transpose for matmul B-operand layout (requires `space=Mat`).	`False`
`is_b_matrix`	`bool`	Hint that the result feeds matmul as the B matrix.	`False`

Returns:

Type	Description
`Tensor`	Tensor of shape `[max_indices, size]` (or transposed) in the chosen space.

Examples:

kv_l1 = pl.paged_gather(kv_pool, topk_idx, block_table, block_size=128, size=head_dim, max_indices=256) out = pl.matmul(q, kv_l1)

`create_l1(shape, dtype, transpose=False)` ¶

Create an on-chip (L1/Mat) accumulator for a kernel-driven paged gather.

Companion of gather_row. Returns a tensor-typed value that composes with pl.matmul / softmax but lowers to an L1 (MemorySpace.Mat) tile, so a kernel can build a matmul operand directly on-chip — no GM round-trip.

Parameters:

Name	Type	Description	Default
`shape`	`Sequence[IntLike]`	Accumulator shape (static dims). For a matmul B-operand use `[rows, size]`; for a `b_trans` B-operand filled by transposing gathers use the transposed shape `[size, rows]` with `transpose=True`.	required
`dtype`	`DataType`	Element dtype (matches the gathered `src` pool).	required
`transpose`	`bool`	Allocate the transposed Mat (ZN) layout — required when filling with `pl.gather_row(..., transpose=True)` (a `DN2ZN` per-row load) and consuming as a `pl.matmul(..., b_trans=True)` B-operand.	`False`

Returns:

Type	Description
`Tensor`	Tensor of shape `shape` backed by L1.

Examples:

kv = pl.create_l1([ATTN_K_TILE, HEAD_DIM], pl.BF16) for r in pl.range(ATTN_K_TILE): kv = pl.gather_row(kv, kv_pool, [r, 0], [phys, 0], [1, HEAD_DIM]) oi = pl.matmul(probs, kv)

`gather_row(acc, src, dst_offset, src_offset, shapes, transpose=False, *, valid_shape=None)` ¶

Gather one GM row into a sub-region of an on-chip accumulator (DPS).

Per-row primitive for a kernel-driven paged gather into L1 — the flexible counterpart to paged_gather: the caller computes the physical src_offset (block-table lookup, multi-source selection, invalid clamping) and the dst_offset slot itself, so arbitrary gather logic stays in the kernel. DMAs src straight into acc (GM -> L1, no tmov); the returned tensor feeds pl.matmul directly.

Parameters:

Name	Type	Description	Default
`acc`	`Tensor`	On-chip accumulator from `create_l1` (loop-carried).	required
`src`	`Tensor`	Source pool in GM.	required
`dst_offset`	`Sequence[IntLike]`	`[row, col]` slot within `acc` to write.	required
`src_offset`	`Sequence[IntLike]`	`[row, col]` physical offset within the GM `src`.	required
`shapes`	`Sequence[IntLike]`	GM row window `[r, c]` (typically `[1, size]`). Must be compile-time constant.	required
`valid_shape`	`Sequence[IntLike] \| None`	How much of that window to actually transfer, defaulting to all of it. May hold runtime `Scalar` values, so a dynamic row count leaves the accumulator's allocation and layout untouched. Not supported together with `transpose=True`.	`None`
`transpose`	`bool`	Place the GM row `[r, c]` as an L1 column `[c, r]` — use for a matmul B-operand whose consumer would otherwise need `b_trans`.	`False`

Returns:

Type	Description
`Tensor`	Tensor aliasing `acc` (written in place).

`scatter(input, dim=None, index=None, src=None, *, mask_pattern=None, dst=None)` ¶

scatter(input: Tensor, dim: int, index: Tensor, src: Tensor) -> Tensor

scatter(input: Tensor, *, mask_pattern: int, dst: Tensor) -> Tensor

Scatter elements of src into input (tensor-level) — index or mask form.

The tensor layer exposes a single unified scatter. Based on the arguments you pass, it lowers to one of two tile-level ops:

Index form (dim + index + src) → pl.tile.scatter — the column-wise inverse of gather, so index has the same shape as src (just like gather's index matches its output)::

output = input
output[b, index[b, k]] = src[b, k]   # for all b, k

MVP: rank-2 input with ``dim == -1``. ``src``/``index`` are ``[rows, K]``;
``input``/output are ``[rows, S]`` with ``K <= S``. ``index`` element
width must match ``input``: 4-byte input → INT32, 2-byte → INT16,
1-byte → INT16.

Mask form (mask_pattern=<int> + dst) → pl.tile.scatter_mask: Writes each row of input into the columns of dst selected by the hardware mask pattern. dst.cols equals input.cols * stride (stride = 2 for P0101/P1010, 4 for P0001..P1000, 1 for P1111). Unlike the gather mask form (a real pto.tgather ISA op on A2/A3 and A5), mask-pattern scatter is not a distinct pto-isa instruction — PyPTO emits it as a pto.tscatter mask-form construct for A2/A3 / CPU-sim style lowering paths.

Parameters:

Name	Type	Description	Default
`input`	`Tensor`	Base tensor (FP16/FP32/BF16/INT8/INT16/INT32, 2D).	required
`dim`	`int \| None`	(index form) Axis along which to scatter. MVP accepts -1.	`None`
`index`	`Tensor \| None`	(index form) Per-element destination column indices, same shape as `src` (INT16/INT32).	`None`
`src`	`Tensor \| None`	(index form) Source values tensor (same dtype as `input`).	`None`
`mask_pattern`	`int \| None`	(mask form, keyword-only) Mask pattern selector (1-7). 1=P0101, 2=P1010, 3=P0001, 4=P0010, 5=P0100, 6=P1000, 7=P1111.	`None`
`dst`	`Tensor \| None`	(mask form, keyword-only) Destination tensor; `dst.cols == input.cols * stride`.	`None`

Returns:

Type	Description
`Tensor`	Tensor representing the post-scatter result.

Examples:

out = scatter(input, dim=-1, index=idx, src=src_vals) out = scatter(input, mask_pattern=pl.tile.MaskPattern.P0101, dst=dst)

`alloc(memory_space, size)` ¶

Stub for the internal tensor.alloc IR operation.

This function is never called in user-written DSL code. It is emitted by the C++ python-printer after the InitMemRef pass and must be importable so that the printed source is valid Python.

The result is a base Ptr (allocation identity token): the printer annotates the assignment target as pl.Ptr, matching the IR design where tensor.alloc Calls carry PtrType.

Parameters:

Name	Type	Description	Default
`memory_space`	`MemorySpace`	Space the allocation lives in, as resolved by InferTileMemorySpace.	required
`size`	`int`	Allocation size in bytes.	required

Returns:

Type	Description
`Ptr`	A `Ptr` standing for the allocation, carrying no address of its own.

`get_block_idx()` ¶

Get the current block index (tensor-scope alias of pl.tile.get_block_idx).

Lowers to tile.get_block_idx in ConvertTensorToTileOps.

Returns:

Type	Description
`Scalar`	Scalar wrapping the `tensor.get_block_idx` operation (INDEX type)

Example

block_idx = pl.tensor.get_block_idx()

`get_subblock_idx()` ¶

Get the current sub-block (vector core) index (tensor-scope alias of pl.tile.get_subblock_idx).

Lowers to tile.get_subblock_idx in ConvertTensorToTileOps.

Returns:

Type	Description
`Scalar`	Scalar wrapping the `tensor.get_subblock_idx` operation (INDEX type)

`get_block_num()` ¶

Get the total number of blocks in the current SPMD task.

Tensor-scope alias of pl.tile.get_block_num; lowers to tile.get_block_num in ConvertTensorToTileOps.

Returns:

Type	Description
`Scalar`	Scalar wrapping the `tensor.get_block_num` operation (INDEX type)

Example

block_idx = pl.tensor.get_block_idx() block_num = pl.tensor.get_block_num()

pl.tensor¶

create_tensor = create module-attribute ¶

create(shape, dtype, layout=TensorLayout.ND, manual_dep=False, init_value=None) ¶

no_dep(tensor) ¶

dump_tag(tensor) ¶

set_cache_policy(tensor, policy) ¶

read(tensor, indices) ¶

write(tensor, indices, value) ¶

dim(tensor, axis) ¶

slice(tensor, shape, offset, valid_shape=None, drop_dims=None, pad_value=None, clamp=False) ¶

fillpad(tensor, pad_value=PadValue.zero) ¶

fillpad_expand(tensor, shape, pad_value=PadValue.zero) ¶

full(shape, dtype, value) ¶

ci(start, shape, dtype=DataType.INT32, descending=False) ¶

arange = ci module-attribute ¶

random(key0, key1, counter0, counter1, counter2, counter3, shape, dtype=DataType.UINT32, rounds=10) ¶

matmul(lhs, rhs, out_dtype=None, a_trans=False, b_trans=False, c_matrix_nz=False) ¶

matmul_acc(acc, lhs, rhs, a_trans=False, b_trans=False, init_cond=None) ¶

mul(lhs, rhs) ¶

muls(lhs, rhs) ¶

add(lhs, rhs) ¶

adds(lhs, rhs) ¶

sub(lhs, rhs) ¶

subs(lhs, rhs) ¶

div(lhs, rhs, high_precision=False) ¶

divs(lhs, rhs) ¶

part_add(lhs, rhs) ¶

part_mul(lhs, rhs) ¶

part_max(lhs, rhs) ¶

part_min(lhs, rhs) ¶

fmod(lhs, rhs) ¶

fmods(lhs, rhs) ¶

maximum(lhs, rhs) ¶

minimum(lhs, rhs) ¶

cmp(lhs, rhs, cmp_type=0) ¶

and_(lhs, rhs) ¶

ands(lhs, rhs) ¶

or_(lhs, rhs) ¶

ors(lhs, rhs) ¶

xor(lhs, rhs) ¶

xors(lhs, rhs) ¶

not_(input) ¶

shl(lhs, rhs) ¶

shls(lhs, rhs) ¶

shr(lhs, rhs) ¶

shrs(lhs, rhs) ¶

row_max(input) ¶

row_sum(input) ¶

row_min(input) ¶

row_prod(input) ¶

col_sum(input) ¶

col_max(input) ¶

col_min(input) ¶

col_prod(input) ¶

row_argmax(input) ¶

row_argmin(input) ¶

col_argmax(input) ¶

col_argmin(input) ¶

row_expand(target, row_vec) ¶

row_expand_mul(tensor, row_vec) ¶

row_expand_div(tensor, row_vec) ¶

row_expand_add(tensor, row_vec) ¶

row_expand_sub(tensor, row_vec) ¶

row_expand_max(tensor, row_vec) ¶

row_expand_min(tensor, row_vec) ¶

row_expand_expdif(tensor, row_vec) ¶

col_expand_mul(tensor, col_vec) ¶

col_expand(tensor, col_vec) ¶

col_expand_div(tensor, col_vec) ¶

col_expand_sub(tensor, col_vec) ¶

col_expand_add(tensor, col_vec) ¶

col_expand_max(tensor, col_vec) ¶

col_expand_min(tensor, col_vec) ¶

col_expand_expdif(tensor, col_vec) ¶

expands(target, scalar) ¶

expand_clone(src, target) ¶

exp(input) ¶

log(input, high_precision=False) ¶

sin(input) ¶

cos(input) ¶

`pl.tensor`¶

`create_tensor = create` `module-attribute` ¶

`create(shape, dtype, layout=TensorLayout.ND, manual_dep=False, init_value=None)` ¶

`no_dep(tensor)` ¶

`dump_tag(tensor)` ¶

`set_cache_policy(tensor, policy)` ¶

`read(tensor, indices)` ¶

`write(tensor, indices, value)` ¶

`dim(tensor, axis)` ¶

`slice(tensor, shape, offset, valid_shape=None, drop_dims=None, pad_value=None, clamp=False)` ¶

`fillpad(tensor, pad_value=PadValue.zero)` ¶

`fillpad_expand(tensor, shape, pad_value=PadValue.zero)` ¶

`full(shape, dtype, value)` ¶

`ci(start, shape, dtype=DataType.INT32, descending=False)` ¶

`arange = ci` `module-attribute` ¶

`random(key0, key1, counter0, counter1, counter2, counter3, shape, dtype=DataType.UINT32, rounds=10)` ¶

`matmul(lhs, rhs, out_dtype=None, a_trans=False, b_trans=False, c_matrix_nz=False)` ¶

`matmul_acc(acc, lhs, rhs, a_trans=False, b_trans=False, init_cond=None)` ¶

`mul(lhs, rhs)` ¶

`muls(lhs, rhs)` ¶

`add(lhs, rhs)` ¶

`adds(lhs, rhs)` ¶

`sub(lhs, rhs)` ¶

`subs(lhs, rhs)` ¶

`div(lhs, rhs, high_precision=False)` ¶

`divs(lhs, rhs)` ¶

`part_add(lhs, rhs)` ¶

`part_mul(lhs, rhs)` ¶

`part_max(lhs, rhs)` ¶

`part_min(lhs, rhs)` ¶

`fmod(lhs, rhs)` ¶

`fmods(lhs, rhs)` ¶

`maximum(lhs, rhs)` ¶

`minimum(lhs, rhs)` ¶

`cmp(lhs, rhs, cmp_type=0)` ¶

`and_(lhs, rhs)` ¶

`ands(lhs, rhs)` ¶

`or_(lhs, rhs)` ¶

`ors(lhs, rhs)` ¶

`xor(lhs, rhs)` ¶

`xors(lhs, rhs)` ¶

`not_(input)` ¶

`shl(lhs, rhs)` ¶

`shls(lhs, rhs)` ¶

`shr(lhs, rhs)` ¶

`shrs(lhs, rhs)` ¶

`row_max(input)` ¶

`row_sum(input)` ¶

`row_min(input)` ¶

`row_prod(input)` ¶

`col_sum(input)` ¶

`col_max(input)` ¶

`col_min(input)` ¶

`col_prod(input)` ¶

`row_argmax(input)` ¶

`row_argmin(input)` ¶

`col_argmax(input)` ¶

`col_argmin(input)` ¶

`row_expand(target, row_vec)` ¶

`row_expand_mul(tensor, row_vec)` ¶

`row_expand_div(tensor, row_vec)` ¶

`row_expand_add(tensor, row_vec)` ¶

`row_expand_sub(tensor, row_vec)` ¶

`row_expand_max(tensor, row_vec)` ¶

`row_expand_min(tensor, row_vec)` ¶

`row_expand_expdif(tensor, row_vec)` ¶

`col_expand_mul(tensor, col_vec)` ¶

`col_expand(tensor, col_vec)` ¶

`col_expand_div(tensor, col_vec)` ¶

`col_expand_sub(tensor, col_vec)` ¶

`col_expand_add(tensor, col_vec)` ¶

`col_expand_max(tensor, col_vec)` ¶

`col_expand_min(tensor, col_vec)` ¶

`col_expand_expdif(tensor, col_vec)` ¶

`expands(target, scalar)` ¶

`expand_clone(src, target)` ¶

`exp(input)` ¶

`log(input, high_precision=False)` ¶

`sin(input)` ¶

`cos(input)` ¶