tvm.te

Namespace for Tensor Expression Language

Classes:

ComputeOp()	Scalar operation.
ExternOp()	External operation.
HybridOp()	Hybrid operation.
PlaceholderOp()	Placeholder operation.
ScanOp()	Scan operation.
Schedule()	Schedule for all the stages.
SpecializedCondition(conditions)	Specialized condition to enable op specialization.
Stage()	A Stage represents schedule for one operation.
Tensor()	Tensor object, to construct, see function.Tensor
TensorComputeOp()	Tensor operation.
TensorSlice(tensor, indices)	Auxiliary data structure for enable slicing syntax from tensor.

Functions:

abs(x[, span])	Get absolute value of the input element-wise.
acos(x)	Take acos of input x.
acosh(x)	Take acos of input x.
add(lhs, rhs[, span])	Generic add operator.
all(*args[, span])	Create a new expression of the intersection of all conditions in the
any(*args[, span])	Create a new experssion of the union of all conditions in the arguments
asin(x)	Take asin of input x.
asinh(x)	Take asinh of input x.
atan(x)	Take atan of input x.
atanh(x)	Take atanh of input x.
ceil(x[, span])	Take ceil of float input x.
comm_reducer(fcombine, fidentity[, name])	Create a commutative reducer for reduction.
compute(shape, fcompute[, name, tag, attrs, ...])	Construct a new tensor by computing over the shape domain.
const(value[, dtype, span])	Create a new constant with specified value and dtype
cos(x)	Take cos of input x.
cosh(x)	Take cosh of input x.
create_prim_func(ops[, index_dtype_override])	Create a TensorIR PrimFunc from tensor expression
create_schedule(ops)	Create a schedule for list of ops
decl_tensor_intrin(op, fcompute[, name, ...])	Declare a tensor intrinsic function.
div(a, b[, span])	Compute a / b as in C/C++ semantics.
erf(x)	Take gauss error function of the input x.
exp(x)	Take exponential of input x.
extern(shape, inputs, fcompute[, name, ...])	Compute several tensors via an extern function.
extern_primfunc(input_tensors, primfunc, ...)	Compute tensors via a schedulable TIR PrimFunc
floor(x[, span])	Take floor of float input x.
floordiv(a, b[, span])	Compute the floordiv of two expressions.
floormod(a, b[, span])	Compute the floormod of two expressions.
fmod(x, y)	Return the remainder of x divided by y with the same sign as x.
gradient(output, inputs[, head])	Perform reverse-mode automatic differentiation.
if_then_else(cond, t, f[, span])	Conditional selection expression.
indexdiv(a, b[, span])	Compute floor(a / b) where a and b are non-negative.
indexmod(a, b[, span])	Compute the remainder of indexdiv.
isfinite(x[, span])	Check if input value is finite.
isinf(x[, span])	Check if input value is infinite.
isnan(x[, span])	Check if input value is Nan.
log(x)	Take log of input x.
log10(x)	Take log10 of input x.
log2(x)	Take log2 of input x.
max(expr, axis[, where, init])	Create a max expression over axis.
max_value(dtype[, span])	maximum value of dtype
min(expr, axis[, where, init])	Create a min expression over axis.
min_value(dtype[, span])	minimum value of dtype
multiply(lhs, rhs[, span])	Generic multiply operator.
nearbyint(x[, span])	Round elements of the array to the nearest integer.
placeholder(shape[, dtype, name])	Construct an empty tensor object.
popcount(x)	Count the number of set bits in input x.
power(x, y[, span])	x power y
reduce_axis(dom[, name, thread_tag, span])	Create a new IterVar for reduction.
round(x[, span])	Round elements of the array to the nearest integer.
rsqrt(x)	Take reciprocal of square root of input x.
scan(init, update, state_placeholder[, ...])	Construct new tensors by scanning over axis.
sigmoid(x)	Quick function to get sigmoid
sin(x)	Take sin of input x.
sinh(x)	Take sinh of input x.
size_var([name, dtype, span])	Create a new variable represents a tensor shape size, which is non-negative.
sqrt(x)	Take square root of input x.
subtract(lhs, rhs[, span])	Generic subtract operator.
sum(expr, axis[, where, init])	Create a sum expression over axis.
tag_scope(tag)	The operator tag scope.
tan(x)	Take tan of input x.
tanh(x)	Take hyperbolic tanh of input x.
thread_axis([dom, tag, name, span])	Create a new IterVar to represent thread index.
trace(args[, trace_action])	Trace tensor data at the runtime.
trunc(x[, span])	Get truncated value of the input.
truncdiv(a, b[, span])	Compute the truncdiv of two expressions.
truncmod(a, b[, span])	Compute the truncmod of two expressions.
var([name, dtype, span])	Create a new variable with specified name and dtype

class tvm.te.ComputeOp

Scalar operation.

class tvm.te.ExternOp

External operation.

class tvm.te.HybridOp

Hybrid operation.

Attributes:

axis	Represent the IterVar axis, also defined when it is a HybridOp

property axis

Represent the IterVar axis, also defined when it is a HybridOp

class tvm.te.PlaceholderOp

Placeholder operation.

class tvm.te.ScanOp

Scan operation.

Attributes:

scan_axis

Represent the scan axis, only defined when it is a ScanOp

property scan_axis

Represent the scan axis, only defined when it is a ScanOp

class tvm.te.Schedule

Schedule for all the stages.

Methods:

cache_read(tensor, scope, readers)	Create a cache read of original tensor for readers.
cache_write(tensor, scope)	Create a cache write of original tensor, before storing into tensor.
create_group(outputs, inputs[, include_inputs])	Create stage group by giving output and input boundary.
normalize()	Build a normalized schedule from the current schedule.
rfactor(tensor, axis[, factor_axis])	Factor a reduction axis in tensor's schedule to be an explicit axis.

cache_read(tensor, scope, readers)

Create a cache read of original tensor for readers.

This will mutate the body of the readers. A new cache stage will be created for the tensor. Call this before doing any split/fuse schedule.

参数:

• tensor (Tensor) -- The tensor to be cached.

• scope (str) -- The scope of cached

• readers (list of Tensor or Operation) -- The readers to read the cache.

返回:

cache -- The created cache tensor.

返回类型:

Tensor

cache_write(tensor, scope)

Create a cache write of original tensor, before storing into tensor.

This will mutate the body of the tensor. A new cache stage will created before feed into the tensor.

This function can be used to support data layout transformation. If there is a split/fuse/reorder on the data parallel axis of tensor before cache_write is called. The intermediate cache stores the data in the layout as the iteration order of leave axis. The data will be transformed back to the original layout in the original tensor. User can further call compute_inline to inline the original layout and keep the data stored in the transformed layout.

参数:

• tensor (Tensor, list or tuple) -- The tensors to be feed to. All the tensors must be produced by one computeOp

• scope (str) -- The scope of cached

返回:

cache -- The created cache tensor.

返回类型:

Tensor

create_group(outputs, inputs, include_inputs=False)

Create stage group by giving output and input boundary.

The operators between outputs and inputs are placed as member of group. outputs are include in the group, while inputs are not included.

参数:

• outputs (list of Tensors) -- The outputs of the group.

• inputs (list of Tensors) -- The inputs of the group.

• include_inputs (boolean, optional) -- Whether include input operations in the group if they are used by outputs.

返回:

group -- A virtual stage represents the group, user can use compute_at to move the attachment point of the group.

返回类型:

Stage

normalize()

Build a normalized schedule from the current schedule.

Insert necessary rebase to make certain iter var to start from 0. This is needed before bound inference and followup step.

返回:

sch -- The normalized schedule.

返回类型:

Schedule

rfactor(tensor, axis, factor_axis=0)

Factor a reduction axis in tensor's schedule to be an explicit axis.

This will create a new stage that generated the new tensor with axis as the first dimension. The tensor's body will be rewritten as a reduction over the factored tensor.

参数:

• tensor (Tensor) -- The tensor to be factored.

• axis (IterVar) -- The reduction axis in the schedule to be factored.

• factor_axis (int) -- The position where the new axis is placed.

返回:

tfactor -- The created factored tensor.

返回类型:

Tensor or Array of Tensor

class tvm.te.SpecializedCondition(conditions)

Specialized condition to enable op specialization.

Methods:

current()

Returns the current specialized condition

static current()

Returns the current specialized condition

class tvm.te.Stage

A Stage represents schedule for one operation.

Methods:

bind(ivar, thread_ivar)	Bind ivar to thread index thread_ivar
compute_at(parent, scope)	Attach the stage at parent's scope
compute_inline()	Mark stage as inline
compute_root()	Attach the stage at parent, and mark it as root
double_buffer()	Compute the current stage via double buffering.
env_threads(threads)	Mark threads to be launched at the outer scope of composed op.
fuse(*args)	Fuse multiple consecutive iteration variables into a single iteration variable.
parallel(var)	Parallelize the iteration.
pragma(var, pragma_type[, pragma_value])	Annotate the iteration with pragma
prefetch(tensor, var, offset)	Prefetch the specified variable
reorder(*args)	reorder the arguments in the specified order.
rolling_buffer()	Compute the current stage via rolling buffering.
set_scope(scope)	Set the thread scope of this stage
set_store_predicate(predicate)	Set predicate under which store to the array can be performed.
split(parent[, factor, nparts, ...])	Split the stage either by factor providing outer scope, or both
storage_align(axis, factor, offset)	Set alignment requirement for specific axis
tensorize(var, tensor_intrin)	Tensorize the computation enclosed by var with tensor_intrin
tile(x_parent, y_parent, x_factor, y_factor)	Perform tiling on two dimensions
transform_layout(mapping_function)	Defines the layout transformation for the current stage's tensor.
unroll(var)	Unroll the iteration.
vectorize(var)	Vectorize the iteration.

bind(ivar, thread_ivar)

Bind ivar to thread index thread_ivar

参数:

• ivar (IterVar) -- The iteration to be binded to thread.

• thread_ivar (IterVar) -- The thread to be binded.

compute_at(parent, scope)

Attach the stage at parent's scope

参数:

• parent (Stage) -- The parent stage

• scope (IterVar) -- The loop scope t be attached to.

compute_inline()

Mark stage as inline

参数:

parent (Stage) -- The parent stage

compute_root()

Attach the stage at parent, and mark it as root

参数:

parent (Stage) -- The parent stage

double_buffer()

Compute the current stage via double buffering.

This can only be applied to intermediate stage. This will double the storage cost of the current stage. Can be useful to hide load latency.

env_threads(threads)

Mark threads to be launched at the outer scope of composed op.

参数:

threads (list of threads) -- The threads to be launched.

fuse(*args)

Fuse multiple consecutive iteration variables into a single iteration variable.

fused = fuse(...fuse(fuse(args[0], args[1]), args[2]),..., args[-1]) The order is from outer to inner.

参数:

args (list of IterVars) -- Itervars that proceeds each other

返回:

fused -- The fused variable of iteration.

返回类型:

IterVar

parallel(var)

Parallelize the iteration.

参数:

var (IterVar) -- The iteration to be parallelized.

pragma(var, pragma_type, pragma_value=None)

Annotate the iteration with pragma

This will translate to a pragma_scope surrounding the corresponding loop generated. Useful to support experimental features and extensions.

参数:

• var (IterVar) -- The iteration to be anotated

• pragma_type (str) -- The pragma string to be annotated

• pragma_value (Expr, optional) -- The pragma value to pass along the pragma

备注

Most pragmas are advanced/experimental features and may subject to change. List of supported pragmas:

• debug_skip_region

  Force skip the region marked by the axis and turn it into no-op. This is useful for debug purposes.

• parallel_launch_point

  Specify to launch parallel threads outside the specified iteration loop. By default the threads launch at the point of parallel construct. This pragma moves the launching point to even outer scope. The threads are launched once and reused across multiple parallel constructs as BSP style program.

• parallel_barrier_when_finish

  Insert a synchronization barrier between working threads after the specified loop iteration finishes.

• parallel_stride_pattern

  Hint parallel loop to execute in strided pattern. for (int i = task_id; i < end; i += num_task)

prefetch(tensor, var, offset)

Prefetch the specified variable

参数:

• tensor (Tensor) -- The tensor to be prefetched

• var (IterVar) -- The loop point at which the prefetching is applied

• offset (Expr) -- The number of iterations to be prefetched before actual execution

reorder(*args)

reorder the arguments in the specified order.

参数:

args (list of IterVar) -- The order to be ordered

rolling_buffer()

Compute the current stage via rolling buffering.

This can only be applied to intermediate stage. This will change the storage cost of the current stage.

set_scope(scope)

Set the thread scope of this stage

参数:

scope (str) -- The thread scope of this stage

set_store_predicate(predicate)

Set predicate under which store to the array can be performed.

Use this when there are duplicated threads doing the same store and we only need one of them to do the store.

参数:

predicate (Expr) -- The guard condition fo store.

split(parent, factor=None, nparts=None, disable_predication=False)

Split the stage either by factor providing outer scope, or both

参数:

• parent (IterVar) -- The parent iter var.

• factor (Expr, optional) -- The splitting factor

• nparts (Expr, optional) -- The number of outer parts.

• disable_predication (bool, optional) --

  If enabled, don't create a predicate for guarding the loop. This can be useful when splitting with scalable factors that the schedule writer knows are divisible by the loop bound.

  Warning: enabling this feature may result in incorrect code generation if not used carefully.

返回:

• outer (IterVar) -- The outer variable of iteration.

• inner (IterVar) -- The inner variable of iteration.

storage_align(axis, factor, offset)

Set alignment requirement for specific axis

This ensures that stride[axis] == k * factor + offset for some k. This is useful to set memory layout to for more friendly memory access pattern. For example, we can set alignment to be factor=2, offset=1 to avoid bank conflict for thread access on higher dimension in GPU shared memory.

参数:

• axis (IterVar) -- The axis dimension to be aligned.

• factor (int) -- The factor in alignment specification.

• offset (int) -- The offset in the alignment specification.

tensorize(var, tensor_intrin)

Tensorize the computation enclosed by var with tensor_intrin

参数:

• var (IterVar) -- The iteration boundary of tensorization.

• tensor_intrin (TensorIntrin) -- The tensor intrinsic used for computation.

tile(x_parent, y_parent, x_factor, y_factor)

Perform tiling on two dimensions

The final loop order from outmost to inner most are [x_outer, y_outer, x_inner, y_inner]

参数:

• x_parent (IterVar) -- The original x dimension

• y_parent (IterVar) -- The original y dimension

• x_factor (Expr) -- The stride factor on x axis

• y_factor (Expr) -- The stride factor on y axis

返回:

• x_outer (IterVar) -- Outer axis of x dimension

• y_outer (IterVar) -- Outer axis of y dimension

• x_inner (IterVar) -- Inner axis of x dimension

• p_y_inner (IterVar) -- Inner axis of y dimension

transform_layout(mapping_function: Callable[[...], List[PrimExpr]])

Defines the layout transformation for the current stage's tensor.

The map from initial_indices to final_indices must be an invertible affine transformation. This method may be called more than once for a given tensor, in which case each transformation is applied sequentially.

If the stage is a ComputeOp, then the iteration order of the compute stage is rewritten to be a row-major traversal of the tensor, and the new loop iteration variables are returned. For all other stages, the loop iteration order is unmodified, and the return value is None.

参数:

mapping_function (Callable[..., List[tvm.tir.PrimExpr]]) -- A callable that accepts N arguments of type tvm.tir.Var, and outputs a list of PrimExpr. The input arguments represent the location of a value in the current stage's tensor, using the pre-transformation layout. The return value of the function gives the location of that value in the current stage's tensor, using the post-transformation layout.

返回:

new_iter_vars -- If the stage is a ComputeOp, then the return will be the updated loop iteration variables over the data array, in the same order as the output values from the mapping_function.

Otherwise, the return value is None.

返回类型:

Optional[List[tvm.tir.IterVar]]

示例

# ``A`` is a tensor whose compute definition is in NHWC
# format, and should be transformed into NCHWc format.

s[A].transform_layout(
    lambda n,h,w,c: [n, c//4, h, w, c%4]
)

# ``A`` is a tensor whose compute definition is in an
# arbitrary format, and should be transformed such that
# the last index is split, with the slower-changing index
# of the split placed at the slowest changing dimension.

s[A].transform_layout(
    lambda *indices, i: [i//4, *indices, i%4]
)

# ``B`` is a tensor defined by te.compute to be a copy of
# ``A`, and should be transformed such that ``B``'s layout
# is a transpose of ``A``'s layout.  The loop iteration
# that computes ``B`` will correspond to ``B``'s memory
# layout.

A = te.placeholder([n,m])
B = te.compute(A.shape, lambda i,j: A[i,j])
s = te.create_schedule(B.op)

s[B].transform_layout(lambda i,j: [j,i])

unroll(var)

Unroll the iteration.

参数:

var (IterVar) -- The iteration to be unrolled.

vectorize(var)

Vectorize the iteration.

参数:

var (IterVar) -- The iteration to be vectorize

class tvm.te.Tensor

Tensor object, to construct, see function.Tensor

Attributes:

axis	Axis of the tensor.
ndim	Dimension of the tensor.
op	The corressponding Operation.
shape	The output shape of the tensor.
value_index	The output value index the tensor corresponds to.

property axis

Axis of the tensor.

property ndim

Dimension of the tensor.

property op

The corressponding Operation.

property shape

The output shape of the tensor.

property value_index

The output value index the tensor corresponds to.

class tvm.te.TensorComputeOp

Tensor operation.

class tvm.te.TensorSlice(tensor, indices)

Auxiliary data structure for enable slicing syntax from tensor.

Methods:

asobject()

Convert slice to object.

Attributes:

dtype

Data content of the tensor.

asobject()

Convert slice to object.

property dtype

Data content of the tensor.

tvm.te.abs(x, span=None)

Get absolute value of the input element-wise.

参数:

• x (PrimExpr) -- Input argument.

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.acos(x)

Take acos of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.acosh(x)

Take acos of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.add(lhs, rhs, span=None)

Generic add operator.

参数:

• lhs (object) -- The left operand.

• rhs (object) -- The right operand.

• span (Optional[Span]) -- The location of this operator in the source.

返回:

op -- The result Expr of add operaton.

返回类型:

tvm.Expr

tvm.te.all(*args, span=None)

Create a new expression of the intersection of all conditions in the

arguments

参数:

• args (list) -- List of symbolic boolean expressions

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

expr -- Expression

返回类型:

Expr

tvm.te.any(*args, span=None)

Create a new experssion of the union of all conditions in the arguments

参数:

• args (list) -- List of symbolic boolean expressions

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

expr -- Expression

返回类型:

Expr

tvm.te.asin(x)

Take asin of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.asinh(x)

Take asinh of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.atan(x)

Take atan of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.atanh(x)

Take atanh of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.ceil(x, span=None)

Take ceil of float input x.

参数:

• x (PrimExpr) -- Input argument.

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.comm_reducer(fcombine, fidentity, name='reduce')

Create a commutative reducer for reduction.

参数:

• fcombine (function(Expr -> Expr -> Expr)) -- A binary function which takes two Expr as input to return a Expr.

• fidentity (function(str -> Expr)) -- A function which takes a type string as input to return a const Expr.

返回:

reducer -- A function which creates a reduce expression over axis. There are two ways to use it:

1. accept (expr, axis, where) to produce an Reduce Expr on specified axis;

2. simply use it with multiple Exprs.

返回类型:

function

示例

n = te.var("n")
m = te.var("m")
mysum = te.comm_reducer(lambda x, y: x+y,
    lambda t: tvm.tir.const(0, dtype=t), name="mysum")
A = te.placeholder((n, m), name="A")
k = te.reduce_axis((0, m), name="k")
B = te.compute((n,), lambda i: mysum(A[i, k], axis=k), name="B")

tvm.te.compute(shape, fcompute, name='compute', tag='', attrs=None, varargs_names=None)

Construct a new tensor by computing over the shape domain.

The compute rule is result[axis] = fcompute(axis)

参数:

• shape (Tuple of Expr) -- The shape of the tensor

• fcompute (lambda function of indices-> value) -- Specifies the input source expression

• name (str, optional) -- The name hint of the tensor

• tag (str, optional) -- Additional tag information about the compute.

• attrs (dict, optional) -- The additional auxiliary attributes about the compute.

• varargs_names (list, optional) -- The names to use for each of the varargs. If not supplied, the varargs will be called i1, i2, ...

返回:

tensor -- The created tensor

返回类型:

Tensor

tvm.te.const(value, dtype='int32', span=None)

Create a new constant with specified value and dtype

参数:

• value (Union[bool, int, float, numpy.ndarray, tvm.nd.NDArray]) -- The constant value.

• dtype (str) -- The data type

• span (Optional[Span]) -- The location of this variable in the source.

返回:

const -- The result constant expr.

返回类型:

PrimExpr

tvm.te.cos(x)

Take cos of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.cosh(x)

Take cosh of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.create_prim_func(ops: List[Tensor | Var], index_dtype_override: str | None = None) → PrimFunc

Create a TensorIR PrimFunc from tensor expression

参数:

ops (List[Union[_tensor.Tensor, tvm.tir.Var]]) -- The source expression.

示例

We define a matmul kernel using following code:

import tvm
from tvm import te
from tvm.te import create_prim_func
import tvm.script

A = te.placeholder((128, 128), name="A")
B = te.placeholder((128, 128), name="B")
k = te.reduce_axis((0, 128), "k")
C = te.compute((128, 128), lambda x, y: te.sum(A[x, k] * B[y, k], axis=k), name="C")
func = create_prim_func([A, B, C])
print(func.script())

If we want to use TensorIR schedule to do transformations on such kernel, we need to use create_prim_func([A, B, C]) to create a schedulable PrimFunc. The generated function looks like:

@T.prim_func
def tir_matmul(a: T.handle, b: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    B = T.match_buffer(b, (128, 128))
    C = T.match_buffer(c, (128, 128))

    for i, j, k in T.grid(128, 128, 128):
        with T.block():
            vi, vj, vk = T.axis.remap("SSR", [i, j, k])
            with T.init():
                C[vi, vj] = 0.0
            C[vi, vj] += A[vi, vk] * B[vj, vk]

返回:

func -- The created function.

返回类型:

tir.PrimFunc

tvm.te.create_schedule(ops)

Create a schedule for list of ops

参数:

ops (list of Operations) -- The source expression.

返回:

sch -- The created schedule.

返回类型:

schedule.Schedule

tvm.te.decl_tensor_intrin(op, fcompute, name='tensor_intrin', binds=None, scalar_params=None, default_buffer_params=None)

Declare a tensor intrinsic function.

参数:

• op (Operation) -- The symbolic description of the intrinsic operation

• fcompute (lambda function of inputs, outputs-> stmt) --

  Specifies the IR statement to do the computation. See the following note for function signature of fcompute

  备注

  Parameters

  • ins (list of tvm.tir.Buffer) - Placeholder for each inputs

  • outs (list of tvm.tir.Buffer) - Placeholder for each outputs

  Returns

  • stmt (tvm.tir.Stmt, or tuple of three stmts)

  • If a single stmt is returned, it represents the body

  • If tuple of three stmts are returned they corresponds to body, reduce_init, reduce_update

• name (str, optional) -- The name of the intrinsic.

• binds (dict of Tensor to tvm.tir.Buffer, optional) -- Dictionary that maps the Tensor to Buffer which specified the data layout requirement of the function. By default, a new compact buffer is created for each tensor in the argument.

• scalar_params (a list of variables used by op, whose values will be passed) -- as scalar_inputs when the tensor intrinsic is called.

• default_buffer_params (Optional[dict]) -- Dictionary of buffer arguments to be passed when constructing a buffer.

返回:

intrin -- A TensorIntrin that can be used in tensorize schedule.

返回类型:

TensorIntrin

tvm.te.div(a, b, span=None)

Compute a / b as in C/C++ semantics.

参数:

• a (PrimExpr) -- The left hand operand, known to be non-negative.

• b (PrimExpr) -- The right hand operand, known to be non-negative.

• span (Optional[Span]) -- The location of this operator in the source.

返回:

res -- The result expression.

返回类型:

PrimExpr

备注

When operands are integers, returns truncdiv(a, b, span).

tvm.te.erf(x)

Take gauss error function of the input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.exp(x)

Take exponential of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.extern(shape, inputs, fcompute, name='extern', dtype=None, in_buffers=None, out_buffers=None, tag='', attrs=None)

Compute several tensors via an extern function.

参数:

• shape (tuple or list of tuples.) -- The shape of the outputs.

• inputs (list of Tensor) -- The inputs

• fcompute (lambda function of inputs, outputs-> stmt) --

  Specifies the IR statement to do the computation. See the following note for function signature of fcompute

  备注

  Parameters

  • ins (list of tvm.tir.Buffer) - Placeholder for each inputs

  • outs (list of tvm.tir.Buffer) - Placeholder for each outputs

  Returns

  • stmt (tvm.tir.Stmt) - The statement that carries out array computation.

• name (str, optional) -- The name hint of the tensor

• dtype (str or list of str, optional) -- The data types of outputs, by default dtype will be same as inputs.

• in_buffers (tvm.tir.Buffer or list of tvm.tir.Buffer, optional) -- Input buffers.

• out_buffers (tvm.tir.Buffer or list of tvm.tir.Buffer, optional) -- Output buffers.

tag: str, optional

Additonal tag information about the compute.

attrs: dict, optional

The additional auxiliary attributes about the compute.

返回:

tensor -- The created tensor or tuple of tensors contains multiple outputs.

返回类型:

Tensor or list of Tensors

示例

In the code below, C is generated by calling external PackedFunc tvm.contrib.cblas.matmul

A = te.placeholder((n, l), name="A")
B = te.placeholder((l, m), name="B")
C = te.extern((n, m), [A, B],
               lambda ins, outs: tvm.tir.call_packed(
                  "tvm.contrib.cblas.matmul",
                    ins[0], ins[1], outs[0], 0, 0), name="C")

tvm.te.extern_primfunc(input_tensors: List[Tensor], primfunc: PrimFunc, **kwargs)

Compute tensors via a schedulable TIR PrimFunc

参数:

• input_tensors (list of Tensor) -- Input tensors that map to the corresponding primfunc input params.

• primfunc (PrimFunc) -- The TIR PrimFunc

返回:

tensor -- The created tensor or tuple of tensors if it contains multiple outputs.

返回类型:

Tensor or list of Tensors

示例

In the code below, a TVMScript defined TIR PrimFunc is inlined into a TE ExternOp. Applying te.create_prim_func on this

A = te.placeholder((128, 128), name="A")
B = te.placeholder((128, 128), name="B")

@T.prim_func
def before_split(a: T.handle, b: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    B = T.match_buffer(b, (128, 128))
    for i, j in T.grid(128, 128):
        with T.block("B"):
            vi, vj = T.axis.remap("SS", [i, j])
            B[vi, vj] = A[vi, vj] * 2.0

C = te.extern_primfunc([A, B], func)

tvm.te.floor(x: PrimExprWithOp, span=None)

Take floor of float input x.

参数:

• x (PrimExpr) -- Input argument.

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.floordiv(a, b, span=None)

Compute the floordiv of two expressions.

参数:

• a (PrimExpr) -- The left hand operand

• b (PrimExpr) -- The right hand operand

• span (Optional[Span]) -- The location of this operator in the source.

返回:

res -- The result expression.

返回类型:

PrimExpr

tvm.te.floormod(a, b, span=None)

Compute the floormod of two expressions.

参数:

• a (PrimExpr) -- The left hand operand

• b (PrimExpr) -- The right hand operand

• span (Optional[Span]) -- The location of this operator in the source.

返回:

res -- The result expression.

返回类型:

PrimExpr

tvm.te.fmod(x, y)

Return the remainder of x divided by y with the same sign as x.

参数:

• x (PrimExpr) -- Input argument.

• y (PrimExpr) -- Input argument.

返回:

z -- The result.

返回类型:

PrimExpr

tvm.te.gradient(output, inputs, head=None)

Perform reverse-mode automatic differentiation.

参数:

• output (Tensor) -- The tensor to differentiate.

• inputs (List[Tensor]) -- The list of input tensors to be differentiated wrt.

• head (Tensor) -- The adjoint of the output, in other words, some tensor, by which the Jacobians will be multiplied. Its shape must be of the form prefix + output.shape. If None is passed, the identity tensor of shape output.shape + output.shape will be used.

返回:

tensors -- The result gradient, in the same order as the inputs

返回类型:

List[Tensor]

示例

x = tvm.placeholder((32, 3, 28, 28), name='x')
w1 = tvm.placeholder((10, 3, 3, 3), name='w1')
w2 = tvm.placeholder((10, 10, 3, 3), name='w2')
z1 = topi.nn.conv2d(x, w1, 1, 1, 1)
z2 = topi.nn.conv2d(z1, w2, 1, 1, 1)
y = topi.sum(z2)

# produce gradients
[dw1, dw2] = tvm.gradient(y, [w1, w2])

# produce Jacobians
[jw1, jw2] = tvm.gradient(z2, [w1, w2])

# produce gradients, the head adjoint for z2 is provided manually
[dw1, dw2] = tvm.gradient(z2, [w1, w2], topi.full_like(z2, 1.0))

tvm.te.if_then_else(cond, t, f, span=None)

Conditional selection expression.

参数:

• cond (PrimExpr) -- The condition

• t (PrimExpr) -- The result expression if cond is true.

• f (PrimExpr) -- The result expression if cond is false.

• span (Optional[Span]) -- The location of this operator in the source.

返回:

result -- The result of conditional expression.

返回类型:

Node

备注

Unlike Select, if_then_else will not execute the branch that does not satisfy the condition. You can use it to guard against out of bound access. Unlike Select, if_then_else cannot be vectorized if some lanes in the vector have different conditions.

tvm.te.indexdiv(a, b, span=None)

Compute floor(a / b) where a and b are non-negative.

参数:

• a (PrimExpr) -- The left hand operand, known to be non-negative.

• b (PrimExpr) -- The right hand operand, known to be non-negative.

• span (Optional[Span]) -- The location of this operator in the source.

返回:

res -- The result expression.

返回类型:

PrimExpr

备注

Use this function to split non-negative indices. This function may take advantage of operands' non-negativeness.

tvm.te.indexmod(a, b, span=None)

Compute the remainder of indexdiv. a and b are non-negative.

参数:

• a (PrimExpr) -- The left hand operand, known to be non-negative.

• b (PrimExpr) -- The right hand operand, known to be non-negative.

• span (Optional[Span]) -- The location of this operator in the source.

返回:

res -- The result expression.

返回类型:

PrimExpr

备注

Use this function to split non-negative indices. This function may take advantage of operands' non-negativeness.

tvm.te.isfinite(x, span=None)

Check if input value is finite.

参数:

• x (PrimExpr) -- Input argument.

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.isinf(x, span=None)

Check if input value is infinite.

参数:

• x (PrimExpr) -- Input argument.

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.isnan(x, span=None)

Check if input value is Nan.

参数:

• x (PrimExpr) -- Input argument.

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.log(x)

Take log of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.log10(x)

Take log10 of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.log2(x)

Take log2 of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.max(expr, axis, where=None, init=None, *args)

Create a max expression over axis.

参数:

• expr (PrimExpr) -- The source expression.

• axis (IterVar) -- The reduction IterVar axis

• where (optional, Expr) -- Filtering predicate of the reduction.

返回:

value -- The result value.

返回类型:

PrimExpr

示例

m = te.var("m")
n = te.var("n")
A = te.placeholder((m, n), name="A")
k = te.reduce_axis((0, n), name="k")

# there are two way to use this max reducer:
# mode 1, accept (expr, axis, where) to produce an Reduce Expr
# tvm.max represents tvm.te.max or tvm.tir.max.
B = te.compute((m,), lambda i: tvm.max(A[i, k], axis=k), name="B")

# mode 2, simply use it with multiple Exprs:
max_res = tvm.max(m, n)

tvm.te.max_value(dtype: str, span: Span | None = None) → Any

maximum value of dtype

参数:

• dtype (str) -- The data type.

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

value -- The maximum value of dtype.

返回类型:

tvm.Expr

tvm.te.min(expr, axis, where=None, init=None, *args)

Create a min expression over axis.

参数:

• expr (PrimExpr) -- The source expression.

• axis (IterVar) -- The reduction IterVar axis

• where (optional, Expr) -- Filtering predicate of the reduction.

返回:

value -- The result value.

返回类型:

PrimExpr

示例

m = te.var("m")
n = te.var("n")
A = te.placeholder((m, n), name="A")
k = te.reduce_axis((0, n), name="k")

# there are two way to use this min reducer:
# mode 1, accept (expr, axis, where) to produce an Reduce Expr
# tvm.min represents tvm.te.min or tvm.tir.min.
B = te.compute((m,), lambda i: tvm.min(A[i, k], axis=k), name="B")

# mode 2, simply use it with multiple Exprs:
min_res = tvm.min(m, n)

tvm.te.min_value(dtype, span=None)

minimum value of dtype

参数:

• dtype (str) -- The data type.

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

value -- The minimum value of dtype.

返回类型:

tvm.Expr

tvm.te.multiply(lhs, rhs, span=None)

Generic multiply operator.

参数:

• lhs (object) -- The left operand.

• rhs (object) -- The right operand.

• span (Optional[Span]) -- The location of this operator in the source.

返回:

op -- The result Expr of multiply operaton.

返回类型:

tvm.Expr

tvm.te.nearbyint(x, span=None)

Round elements of the array to the nearest integer. This intrinsic uses llvm.nearbyint instead of llvm.round which is faster but will results different from te.round. Notably nearbyint rounds according to the rounding mode, whereas te.round (llvm.round) ignores that. For differences between the two see: https://en.cppreference.com/w/cpp/numeric/math/round https://en.cppreference.com/w/cpp/numeric/math/nearbyint

参数:

• x (PrimExpr) -- Input argument.

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.placeholder(shape, dtype=None, name='placeholder')

Construct an empty tensor object.

参数:

• shape (Tuple of Expr) -- The shape of the tensor

• dtype (str, optional) -- The data type of the tensor

• name (str, optional) -- The name hint of the tensor

返回:

tensor -- The created tensor

返回类型:

Tensor

tvm.te.popcount(x)

Count the number of set bits in input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.power(x, y, span=None)

x power y

参数:

• x (PrimExpr) -- Input argument.

• y (PrimExpr) -- The exponent

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

z -- The result.

返回类型:

PrimExpr

tvm.te.reduce_axis(dom, name='rv', thread_tag='', span=None)

Create a new IterVar for reduction.

参数:

• dom (Range) -- The domain of iteration.

• name (str) -- The name of the variable.

• thread_tag (Optional[str]) -- The name of the thread_tag.

• span (Optional[Span]) -- The location of this variable in the source.

返回:

axis -- An iteration variable representing the value.

返回类型:

IterVar

tvm.te.round(x, span=None)

Round elements of the array to the nearest integer.

参数:

• x (PrimExpr) -- Input argument.

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.rsqrt(x)

Take reciprocal of square root of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.scan(init, update, state_placeholder, inputs=None, name='scan', tag='', attrs=None)

Construct new tensors by scanning over axis.

参数:

• init (Tensor or list of Tensor) -- The initial condition of first init.shape[0] timestamps

• update (Tensor or list of Tensor) -- The update rule of the scan given by symbolic tensor.

• state_placeholder (Tensor or list of Tensor) -- The placeholder variables used by update.

• inputs (Tensor or list of Tensor, optional) -- The list of inputs to the scan. This is not required, but can be useful for the compiler to detect scan body faster.

• name (str, optional) -- The name hint of the tensor

• tag (str, optional) -- Additonal tag information about the compute.

• attrs (dict, optional) -- The additional auxiliary attributes about the compute.

返回:

tensor -- The created tensor or tuple of tensors contains multiple outputs.

返回类型:

Tensor or list of Tensors

示例

# The following code is equivalent to numpy.cumsum
m = te.var("m")
n = te.var("n")
X = te.placeholder((m, n), name="X")
s_state = te.placeholder((m, n))
s_init = te.compute((1, n), lambda _, i: X[0, i])
s_update = te.compute((m, n), lambda t, i: s_state[t-1, i] + X[t, i])
res = tvm.te.scan(s_init, s_update, s_state, X)

tvm.te.sigmoid(x)

Quick function to get sigmoid

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.sin(x)

Take sin of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.sinh(x)

Take sinh of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.size_var(name='size', dtype='int32', span=None)

Create a new variable represents a tensor shape size, which is non-negative.

参数:

• name (str) -- The name

• dtype (str) -- The data type

• span (Optional[Span]) -- The location of this variable in the source.

返回:

var -- The result symbolic shape variable.

返回类型:

SizeVar

tvm.te.sqrt(x)

Take square root of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.subtract(lhs, rhs, span=None)

Generic subtract operator.

参数:

• lhs (object) -- The left operand.

• rhs (object) -- The right operand.

• span (Optional[Span]) -- The location of this operator in the source.

返回:

op -- The result Expr of subtract operaton.

返回类型:

tvm.Expr

tvm.te.sum(expr, axis, where=None, init=None, *args)

Create a sum expression over axis.

参数:

• expr (PrimExpr) -- The source expression.

• axis (IterVar) -- The reduction IterVar axis

• where (optional, Expr) -- Filtering predicate of the reduction.

返回:

value -- The result value.

返回类型:

PrimExpr

示例

m = te.var("m")
n = te.var("n")
A = te.placeholder((m, n), name="A")
k = te.reduce_axis((0, n), name="k")

# there are two way to use this sum reducer:
# mode 1, accept (expr, axis, where) to produce an Reduce Expr
# tvm.sum represents tvm.te.sum or tvm.tir.sum.
B = te.compute((m,), lambda i: tvm.sum(A[i, k], axis=k), name="B")

# mode 2, simply use it with multiple Exprs:
sum_res = tvm.sum(m, n)

tvm.te.tag_scope(tag)

The operator tag scope.

参数:

tag (str) -- The tag name.

返回:

tag_scope -- The tag scope object, which can be used as decorator or context manger.

返回类型:

TagScope

示例

n = te.var('n')
m = te.var('m')
l = te.var('l')
A = te.placeholder((n, l), name='A')
B = te.placeholder((m, l), name='B')
k = te.reduce_axis((0, l), name='k')

with tvm.te.tag_scope(tag='matmul'):
    C = te.compute((n, m), lambda i, j: te.sum(A[i, k] * B[j, k], axis=k))

# or use tag_scope as decorator
@tvm.te.tag_scope(tag="conv")
def compute_relu(data):
    return te.compute(data.shape, lambda *i: tvm.tir.Select(data(*i) < 0, 0.0, data(*i)))

tvm.te.tan(x)

Take tan of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.tanh(x)

Take hyperbolic tanh of input x.

参数:

x (PrimExpr) -- Input argument.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.thread_axis(dom=None, tag='', name='', span=None)

Create a new IterVar to represent thread index.

参数:

• dom (Range or str) -- The domain of iteration When str is passed, dom is set to None and str is used as tag

• tag (str, optional) -- The thread tag

• name (str, optional) -- The name of the var.

• span (Optional[Span]) -- The location of this variable in the source.

返回:

axis -- The thread itervar.

返回类型:

IterVar

tvm.te.trace(args, trace_action='tvm.default_trace_action')

Trace tensor data at the runtime.

The trace function allows to trace specific tensor at the runtime. The tracing value should come as last argument. The trace action should be specified, by default tvm.default_trace_action is used.

参数:

• args (list of Expr or Buffers.) -- Positional arguments.

• trace_action (str.) -- The name of the trace action.

返回:

call -- The call expression.

返回类型:

PrimExpr

参见

tvm.tir.call_packed

Creates packed function.

tvm.te.trunc(x, span=None)

Get truncated value of the input.

The truncated value of the scalar x is the nearest integer i which is closer to zero than x is.

参数:

• x (PrimExpr) -- Input argument.

• span (Optional[Span]) -- The location of this operator in the source code.

返回:

y -- The result.

返回类型:

PrimExpr

tvm.te.truncdiv(a, b, span=None)

Compute the truncdiv of two expressions.

参数:

• a (PrimExpr) -- The left hand operand

• b (PrimExpr) -- The right hand operand

• span (Optional[Span]) -- The location of this operator in the source.

返回:

res -- The result expression.

返回类型:

PrimExpr

备注

This is the default integer division behavior in C.

tvm.te.truncmod(a, b, span=None)

Compute the truncmod of two expressions.

参数:

• a (PrimExpr) -- The left hand operand

• b (PrimExpr) -- The right hand operand

• span (Optional[Span]) -- The location of this operator in the source.

返回:

res -- The result expression.

返回类型:

PrimExpr

备注

This is the default integer division behavior in C.

tvm.te.var(name='tindex', dtype='int32', span=None)

Create a new variable with specified name and dtype

参数:

• name (str) -- The name

• dtype (str) -- The data type

• span (Optional[Span]) -- The location of this variable in the source.

返回:

var -- The result symbolic variable.

返回类型:

tir.Var

tvm.te.hybrid

Hybrid Programming APIs of TVM Python Package.

This package maps a subset of python to HalideIR so that: 1. Users can write some preliminary versions of the computation patterns have not been supported yet and verify it across the real execution and python semantic emulation. 2. So far, it is a text format dedicated to HalideIR Phase 0. Refer tvm.lower for more details. A larger ambition of this module is to support all levels of HalideIR.

Classes:

HybridModule([src, name])

The usage of Hybrid Module is very similar to conventional TVM module, but conventional TVM module requires a function body which is already fully lowered.

Functions:

build(sch, inputs, outputs[, name])	Dump the current schedule to hybrid module
decorate(func, fwrapped)	A wrapper call of decorator package, differs to call time
script(pyfunc)	Decorate a python function as hybrid script.
source_to_op(src, args, symbols, closure_vars)	Another level of wrapper

class tvm.te.hybrid.HybridModule(src=None, name=None)

The usage of Hybrid Module is very similar to conventional TVM module, but conventional TVM module requires a function body which is already fully lowered. This contradicts to the fact that Hybrid Module is originally a text format for Phase 0 HalideIR. Thus, a totally separated module is defined.

Methods:

load(path)

Load the module from a python file

load(path)

Load the module from a python file

参数:

path (str) -- Path to the given python file

tvm.te.hybrid.build(sch, inputs, outputs, name='hybrid_func')

Dump the current schedule to hybrid module

参数:

• sch (tvm.te.Schedule) -- The schedule to be dumped

• inputs (An array of Tensors or Vars) -- The inputs of the function body

• outputs (An array of Tensors) -- The outputs of the function body

返回:

module -- The built results is wrapped in a HybridModule. The usage of HybridModule is roughly the same as normal TVM-built modules.

返回类型:

HybridModule

tvm.te.hybrid.decorate(func, fwrapped)

A wrapper call of decorator package, differs to call time

参数:

• func (function) -- The original function

• fwrapped (function) -- The wrapped function

tvm.te.hybrid.script(pyfunc)

Decorate a python function as hybrid script.

The hybrid function support emulation mode and parsing to the internal language IR.

返回:

hybrid_func -- A decorated hybrid script function.

返回类型:

function

tvm.te.hybrid.source_to_op(src, args, symbols, closure_vars)

Another level of wrapper

参数:

• src (ast.node or str) -- If an ast.node, then directly lower it. If a str, then parse it to ast and lower it.

• args (list of Tensors or Vars) -- The argument lists to the function. It is NOT encouraged to write a function without arguments. It is NOT encouraged to write a function with side effect.

• symbols (list of str) -- The symbol list of the global context of the function.

• closure_vars (dict) -- A dict of external name reference captured by this function.

返回:

res -- The result of output tensors of the formed OpNode.

返回类型:

list of output tensors