tvm.tir.schedule¶

Namespace for the TensorIR schedule API.

class tvm.tir.schedule.BlockRV¶: A random variable that refers to a block

class tvm.tir.schedule.BlockScope¶

An object corresponds to each block sref in the sref tree, which tracks the producer-consumer dependency between blocks.

Glossary:

Block scope: A contiguous subtree of the sref tree, rooted at each block sref, whose components are:
- scope root: a block sref
- internal srefs: loop srefs
- scope leaves: block srefs
Child block: The scope leaf blocks under the scope root or a specific internal sref

get_deps_by_dst(block: StmtSRef) → List[Dependency]¶

Get all dependencies whose dst is the target block.

Parameters:: block (StmtSRef) – The queried block
Returns:: blocks – The dependencies
Return type:: List[Dependency]

get_deps_by_src(block: StmtSRef) → List[Dependency]¶

Get all dependencies whose src is the target`block`.

Parameters:: block (StmtSRef) – The queried block
Returns:: blocks – The dependencies
Return type:: List[Dependency]

class tvm.tir.schedule.DepKind(value)¶

Type of dependency.

RAW¶

Read-after-write dependency

Type:: int = 0

WAW¶

Write-after-write dependency

Type:: int = 1

WAR¶

Write-after-read dependency. Not supported in TensorIR for now.

Type:: int = 2

OPAQUE¶

Opaque dependency

Type:: int = 3

class tvm.tir.schedule.Dependency¶

A tuple (src, dst, kind) representing certain types of dependency. For example, (A, B, kRAW) means block B depends on block A, and the dependency kind is read-after-write, which means block B reads the result written by block A.

Parameters:

src (StmtSRef) – The source of the dependency relation
dst (StmtSRef) – The destination of the dependency relation
kind (DepKind) – The dependency kind

tvm.tir.schedule.ExprRV¶: alias of PrimExpr

class tvm.tir.schedule.Instruction(kind: InstructionKind, inputs: List[Any], attrs: List[Any], outputs: List[Any])¶

Schedule instructions each corresponds to a schedule primitive

kind¶

The kind of the instruction

Type:: InstructionKind

inputs¶

The input random variables of the instruction, and the type of each element can be one of the following: - BlockRV - LoopRV - ExprRV - float - int - str - None

Type:: List[INPUT_RV_TYPE]

attrs¶

The attributes of the instruction. Similar to attributes of an operator, attributes of an instruction are arbitrary constant metadata required by the instructions. For example, the name of the block to be retrieved in GetBlock.

Type:: List[ATTR_TYPE]

outputs¶

The output random variables of the instruction, and the type of each element can be one of the following: - BlockRV - LoopRV - ExprRV, atomic variables only, won’t be constants or composite PrimExpr

Type:: List[OUTPUT_RV_TYPE]

class tvm.tir.schedule.InstructionKind¶

Kind of an instruction, e.g. Split, Reorder, etc. Besides the name, every kind of instruction has its own properties, including: 1) A boolean indicating if the instruction is pure, i.e. change nothing in the schedule state 2) A functor that applies the instruction to a TensorIR schedule 3) A functor that converts the instruction to a statement in python syntax 4) A functor that serialize its attributes to JSON 5) A functor that deserialize its attributes from JSON

Unlike tvm.ir.op, InstructionKind doesn’t support unstructured properties, mainly because there is no such usecase yet to add any other property.

name¶

The name of a kind of instructions

Type:: str

Note

The functor properties are not exposed on python side at the moment

static get(name: str) → InstructionKind¶

Retrieve an InstructionKind using its name

Parameters:: name (str) – The registered name of the InstructionKind
Returns:: kind – The InstructionKind retrieved
Return type:: InstructionKind

property is_pure: bool¶

Indicates if the instruction is pure, i.e. removing it alone doesn’t mutate the schedule state. For example, the instruction GetBlock is pure because it changes nothing, while ComputeInline is not because removing it leads to a different resulting schedule.

Returns:: pure – The boolean flag indicating if the instruction is pure
Return type:: bool

class tvm.tir.schedule.LoopRV¶: A random variable that refers to a loop

class tvm.tir.schedule.Schedule(mod: PrimFunc | IRModule, *, seed: int | None = None, debug_mask: str | int = 'none', error_render_level: str = 'detail', enable_check: bool = True)¶

The user-facing schedule class

A schedule is a set of transformations that change the order of computation but preserve the semantics of computation. Some example of schedules: 1) Split a loop into two; 2) Reorder two loops; 3) Inline the computation of a specific buffer into its consumer

The schedule class stores auxiliary information to schedule correctly and efficiently.

Link to tutorial: https://tvm.apache.org/docs/tutorials/language/schedule_primitives.html

add_unit_loop(block_or_loop: LoopRV | BlockRV) → LoopRV¶

Create a new unit loop on top of the specific block or loop.

Parameters:: block_or_loop (Union[LoopRV, BlockRV]) – The block above which the new loop is created
Returns:: new_loop – The new unit loop
Return type:: LoopRV

Examples

Before add_unit_loop, in TensorIR, the IR is:

@T.prim_func
def before_add_unit_loop(
    A: T.Buffer((), "int32"),
    B: T.Buffer((), "int32"),
    C: T.Buffer((), "int32"),
) -> None:
    with T.block("C"):
        vi = T.axis.spatial(1, 0)
        C[()] = A[()] + B[()]

Create the schedule and do add-unit-loop:

sch = tir.Schedule(before_add_unit_loop)
sch.add_unit_loop(sch.get_block("C"))
print(sch.mod["main"].script())

After applying add-unit-loop, the IR becomes:

@T.prim_func
def after_add_unit_loop(
    A: T.Buffer((), "int32"),
    B: T.Buffer((), "int32"),
    C: T.Buffer((), "int32"),
) -> None:
    for u in T.serial(1):
        with T.block("C"):
            vi = T.axis.spatial(1, 0)
            C[()] = A[()] + B[()]

Annotate a block/loop with a key value pair

Parameters:

block_or_loop (Union[BlockRV, LoopRV]) – The block/loop to be annotated
ann_key (str) – The annotation key
ann_val (AnnotationValueT) – The annotation value

Examples

Before annotate, in TensorIR, the IR is:

@T.prim_func
def before_annotate(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

Create the schedule and do annotate:

sch = tir.Schedule(before_annotate)
sch.annotate(sch.get_block("B"), "ann_key", "ann_value")
print(sch.mod["main"].script())

After applying annotate, the IR becomes:

@T.prim_func
def after_annotate(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])
            T.block_attr({"ann_key", "ann_value"})
            B[vi, vj] = A[vi, vj] * 2.0

bind(loop: LoopRV, thread_axis: str) → None¶

Bind the input loop to the given thread axis. It requires: 1) The scope block that the loop is in should have stage-pipeline property 2) All the blocks under the loop are complete blocks or reduction blocks, and have affine bindings 3) For each block under the loop, if the thread axis starts with “threadIdx`, the loop can only be contained in data-parallel block iter and reduction block iters’ bindings. Otherwise the loop can only be contained in data-parallel block iters’ bindings

Parameters:

loop (LoopRV) – The loop to be bound to the thread axis
thread_axis (str) – The thread axis to be bound to the loop. Possible candidates: - blockIdx.x/y/z - threadIdx.x/y/z - vthread.x/y/z - vthread (It is a legacy behavior that will be deprecated. Please use vthread.x/y/z instead.)

Examples

Before bind, in TensorIR, the IR is:

@T.prim_func
def before_bind(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

Create the schedule and do bind:

sch = tir.Schedule(before_bind)
i, j = sch.get_loops(sch.get_block("B"))
sch.bind(i, "blockIdx.x")
sch.bind(j, "threadIdx.x")

After applying bind, the IR becomes:

@T.prim_func
def after_bind(a: T.handle, b: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    B = T.match_buffer(b, (128, 128))
    for i in T.thread_binding(0, 128, thread = "blockIdx.x"):
        for j in T.thread_binding(0, 128, thread = "threadIdx.x"):
            with T.block("B"):
                vi, vj = T.axis.remap("SS", [i, j])
                B[vi, vj] = A[vi, vj] * 2.0

blockize(target: LoopRV | List[BlockRV], preserve_unit_iters: bool = True) → BlockRV¶

Convert multiple blocks or the subtree rooted at a specific loop into a block.

Parameters:

target (LoopRV or List[BlockRV]) – The root of the subtree or the specified blocks.
preserve_unit_iters (bool) – Whether or not to preserve unit iterators in block bindings

Returns:

result – The new block.

Return type:

BlockRV

Examples

Before blockize, in TensorIR, the IR is:

@T.prim_func
def before_blockize(
    A: T.Buffer((128, 128), "float32"),
    B: T.Buffer((128, 128), "float32")
) -> None:
    for i_0, j_0, i_1, j_1 in T.grid(8, 8, 16, 16):
        with T.block("B"):
            vi = T.axis.spatial(128, i_0 * 16 + i_1)
            vj = T.axis.spatial(128, j_0 * 16 + j_1)
            T.reads(A[vi, vj])
            T.writes(B[vi, vj])
            B[vi, vj] = A[vi, vj] * T.float32(2)

Create the schedule and do set_scope:

sch = tir.Schedule(before_blockize)
B = sch.get_block("B")
_, _, i1, _ = sch.get_loops(B)
sch.blockize(i1)
print(sch.mod["main"].script())

After applying blockize, the IR becomes:

@T.prim_func
def after_blockize(
    A: T.Buffer((128, 128), "float32"),
    B: T.Buffer((128, 128), "float32")
)-> None:
    for i_0, j_0 in T.grid(8, 8):
        with T.block("B_o"):
            vio, vjo = T.axis.remap("SS", [i_0, j_0])
            T.reads(A[vio * 16 : vio * 16 + 16, vjo * 16 : vjo * 16 + 16])
            T.writes(B[vio * 16 : vio * 16 + 16, vjo * 16 : vjo * 16 + 16])
            for i_1, j_1 in T.grid(16, 16):
                with T.block("B"):
                    vi, vj = T.axis.remap("SS", [i_1, j_1])
                    T.reads(A[vio * 16 + vi, vjo * 16 + vj])
                    T.writes(B[vio * 16 + vi, vjo * 16 + vj])
                    B[vio * 16 + vi, vjo * 16 + vj] = A[vio * 16 + vi, vjo * 16 + vj]                                                                   * T.float32(2)

Note

blockize requires there is exactly one block under the given loop and the bindings of the block are divisible by the subspace represented by the loops starting at the given loop.

cache_index(block: BlockRV | str, storage_scope: str, cse_thresh: int = 0) → List[BlockRV]¶

Create a block to cache precomputed index for later use. if there is no index computation, keep unchanged.

Parameters:

block (Union[BlockRV, str]) – The target block operates on the target buffer.
storage_scope (str) – The storage scope of cached block.
cse_thresh (int) – The repeat threshold that determines a common sub expr, default 0 means cache all index computation.

Returns:

cached_blocks – The blocks of the stage writing the cache buffers

Return type:

List[BlockRV]

Examples

Before cache_inplace, in TensorIR, the IR is:

@T.prim_func
def resize(a: T.handle, b: T.handle) -> None:
    A = T.match_buffer(a, (1, 3, 40, 40))
    B = T.match_buffer(b, (1, 3, 80, 80))
    for i0, i1, i2, i3 in T.grid(1, 3, 80, 80):
        with T.block("A"):
            n, c, vi, vj = T.axis.remap("SSSS", [i0, i1, i2, i3])
            B[n, c, vi, vj] = A[n, c, vi//4 + vj//4, vj//2]

Create the schedule and cache_index:

sch = tir.Schedule(resize)
block_a = sch.get_block("A")
sch.cache_index(block_a, "global", 1)
print(sch.mod["main"].script())

After applying cache_index, the IR becomes:

@T.prim_func
def resize_cache_index(
    A: T.Buffer((1, 3, 40, 40), "float32"), B: T.Buffer((1, 3, 80, 80), "float32")
) -> None:
    index_var_0 = T.alloc_buffer([80, 80], dtype="int32", strides=[1])
    index_var_1 = T.alloc_buffer([80], dtype="int32", strides=[1])
    for ax0, ax1 in T.grid(80, 80):
        with T.block("index_0"):
            v0 = T.axis.spatial(80, ax0)
            v1 = T.axis.spatial(80, ax1)
            T.reads()
            T.writes(index_var_0[v0, v1])
            index_var_0[v0, v1] = v0 // 4 + v1 // 4
    for ax0 in T.serial(80):
        with T.block("index_1"):
            v0 = T.axis.spatial(80, ax0)
            T.reads()
            T.writes(index_var_1[v0])
            index_var_1[v0] = v0 // 2
    for i0, i1, i2, i3 in T.grid(1, 3, 80, 80):
        with T.block("A"):
            n, c, vi, vj = T.axis.remap("SSSS", [i0, i1, i2, i3])
            T.reads(A[n, c, vi // 4 + vj // 4, vj // 2])
            T.writes(B[n, c, vi, vj])
            B[n, c, vi, vj] = A[n, c, index_var_0[vi, vj], index_var_1[vj]]

cache_inplace(block: BlockRV | str, read_buffer_index: int | str | Buffer, storage_scope: str) → List[BlockRV]¶

Create blocks that reads & write a buffer region into a cache block. It requires the target block both read & write the target buffer. Mainly for inplace operation.

Parameters:

block (Union[BlockRV, str]) – The target block operates on the target buffer.
read_buffer_index (int) – The index of the buffer in block’s read region, the unique name of a read buffer in the block, or a Buffer object that is within the blocks read region.
storage_scope (str) – The target storage scope.

Returns:

cached_blocks – The blocks of the cache stage, read cache first, write cache second

Return type:

List[BlockRV]

Examples

Before cache_inplace, in TensorIR, the IR is:

@T.prim_func
def before_cache_inplace(data_io: T.Buffer((64), "int32")):
    for i0 in T.serial(1):
        with T.block("A"):
            T.reads(data_io[:64])
            T.writes(data_io[:64])
            T.evaluate(T.call_extern("call_impl", data_io.data, dtype=""))

Create the schedule and cache_inplace:

sch = tir.Schedule(before_cache_inplace)
block_a = sch.get_block("A")
sch.cache_inplace(block_a, 0, "local")
print(sch.mod["main"].script())

After applying cache_inplace, the IR becomes:

@T.prim_func
def cache_inplace(data_io: T.Buffer(64, "int32")) -> None:
    data_io_local = T.alloc_buffer([64], dtype="int32", scope="local")
    for i0 in T.serial(1):
        for ax0 in T.serial(64):
            with T.block("data_io_local"):
                v0 = T.axis.spatial(64, ax0)
                T.reads(data_io[v0])
                T.writes(data_io_local[v0])
                data_io_local[v0] = data_io[v0]
        with T.block("A"):
            T.reads(data_io_local[0 : 64])
            T.writes(data_io_local[0 : 64])
            T.evaluate(T.call_extern("call_impl", data_io_local.data, dtype=""))
        for ax0 in T.serial(64):
            with T.block("data_io_local"):
                v0 = T.axis.spatial(64, ax0)
                T.reads(data_io_local[v0])
                T.writes(data_io[v0])
                data_io[v0] = data_io_local[v0]

Create a block that reads a buffer region into a read cache. It requires:

There is at most one block who write the buffer in the scope.
The scope block have stage-pipeline property.

Parameters:

block (Union[BlockRV, str]) – The consumer block of the target buffer.
buffer (Union[int, str, Buffer]) – The index of the buffer in block’s read region, the unique name of a read buffer in the block, or a Buffer object that is within the blocks read region.
storage_scope (str) – The target storage scope.
consumer_blocks (Optional[List[Union[BlockRV, str]]]) – An optional list of consumers that should read from the cache. If not specified, all consumers will use the cache.

Returns:

cached_block – The block of the cache stage

Return type:

BlockRV

Examples

Before cache_read, in TensorIR, the IR is:

@T.prim_func
def before_cache_read(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

Create the schedule and cache_read:

sch = tir.Schedule(before_cache_read)
block_b = sch.get_block("B")
sch.cache_read(block_b, 0, "local")
print(sch.mod["main"].script())

After applying cache_read, the IR becomes:

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

Create a block that reads a buffer region into a write cache. It requires:

There is only one block who write the buffer in the scope.
The scope block have stage-pipeline property.

Parameters:

block (Union[BlockRV, str]) – The producer block of the target buffer.
write_buffer_index (int) – The index of the buffer in block’s write region, the unique name of a write buffer in the block, or a Buffer object that is within the blocks write region.
storage_scope (str) – The target storage scope.
consumer_blocks (Optional[List[Union[BlockRV, str]]]) – An optional list of consumers that should read directly from the cache. If not specified, all consumers will read from the original buffer.

Returns:

cached_block – The block of the cache stage

Return type:

BlockRV

Examples

Before cache_write, in TensorIR, the IR is:

@T.prim_func
def before_cache_write(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

Create the schedule and cache_write:

sch = tir.Schedule(before_cache_write)
block_b = sch.get_block("B")
sch.cache_write(block_b, 0, "local")
print(sch.mod["main"].script())

After applying cache_write, the IR becomes:

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

can_decompose_padding(block: BlockRV | str, loop: LoopRV) → bool¶: Check whether the block match padding pattern and can be decomposed.

compute_at(block: BlockRV | str, loop: LoopRV, preserve_unit_loops: bool = False, index: int = -1) → None¶

Compute-At. Move a producer block under the specific loop, and regenerate the loops induced by the block so that the buffer region produced by the producer block could cover those regions consumed by its consumer blocks under the given loop. It requires:

block and loop are under the same scope, loop is not the ancestor of block
The scope block has stage-pipeline property

3) The subtree of the scope block, where the given block is in, satisfies the compact dataflow condition. i.e. all the blocks in the scope block’s subtree must be either complete block or reduction block

4) The block is not an output block with regard to the scope block, i.e. the buffers written by the block are allocated under the scope block

All the consumers of the block are under the given loop

Parameters:

block (Union[BlockRV, str]) – The block to be moved
loop (LoopRV) – The loop where the block to be moved under
preserve_unit_loops (bool) – Whether to keep the trivial loops whose extents are 1
index (int) – The block index of the loop body subtree blocks: - index = -1 means inserted into the last possible insertion point; - index = -2 means inserted into the first possible insertion point; - Otherwise, index is a nonnegative number that indicates the insertion point

Examples

Before compute-at, in TensorIR, the IR is:

@T.prim_func
def before_compute_at(a: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (128, 128), "float32")
    B = T.alloc_buffer((128, 128), "float32")
    C = T.match_buffer(c, (128, 128), "float32")
    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
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = B[vi, vj] + 1.0

Create the schedule and do compute-at:

sch = tir.Schedule(before_compute_at)
block = sch.get_block("B")
loop, _ = sch.get_loops(sch.get_block("C"))
sch.compute_at(block, loop, preserve_unit_loops=False)
print(sch.mod["main"].script())

After applying compute-at, the IR becomes:

@T.prim_func
def after_compute_at(a: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (128, 128), "float32")
    B = T.alloc_buffer((128, 128), "float32")
    C = T.match_buffer(c, (128, 128), "float32")
    for i in T.serial(0, 128):
        for j in T.serial(0, 128):
            with T.block("B"):
                vi, vj = T.axis.remap("SS", [i, j])
                B[vi, vj] = A[vi, vj] * 2.0
        for j in T.serial(0, 128):
            with T.block("C"):
                vi, vj = T.axis.remap("SS", [i, j])
                C[vi, vj] = B[vi, vj] + 1.0

compute_inline(block: BlockRV | str) → None¶

Inline a block into its consumer(s). It requires:

The block is a complete non-root block, which only produces one buffer
The block must not be the only leaf in the scope.
The body of the block must be a BufferStore statement in the form of, A[i, j, k, ...] = ... where the indices of the LHS are all distinct atomic variables, and no variables other than those indexing variables are allowed in the statement.

Parameters:: block (Union[BlockRV, str]) – The block to be inlined to its consumer(s)

Examples

Before compute-inline, in TensorIR, the IR is:

@T.prim_func
def before_inline(a: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    B = T.alloc_buffer((128, 128))
    C = T.match_buffer(c, (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
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = B[vi, vj] + 1.0

Create the schedule and do compute-inline:

sch = tir.Schedule(before_inline)
sch.compute_inline(sch.get_block("B"))
print(sch.mod["main"].script())

After applying compute-inline, the IR becomes:

@T.prim_func
def after_inline(a: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    C = T.match_buffer(c, (128, 128))
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = A[vi, vj] * 2.0 + 1.0

copy() → Schedule¶

Returns a copy of the schedule, including both the state and the symbol table, * guaranteeing that * 1) SRef tree is completely reconstructed; * 2) The IRModule being scheduled is untouched; * 3) All the random variables are valid in the copy, pointing to the corresponding sref * reconstructed

Returns:: copy – A new copy of the schedule
Return type:: Schedule

decompose_padding(block: BlockRV | str, loop: LoopRV) → BlockRV¶

Decompose a block of padding computation pattern into two separate blocks.

The block which fill const pad values into full write region;
The block which fill in-bound values into region where pad predicate is true.

The pad value filling block is inserted right before the given loop.

The schedule primitive requires:

The input block is a complete block.
The input loop is the ancestor of the block.
The input block is a block which match padding pattern.

Parameters:

block (Union[BlockRV, str]) – The padding block to be decomposed.
loop (LoopRV) – The loop above which the pad value filling block is inserted before.

Returns:

pad_value_block – The block filling const pad values.

Return type:

BlockRV

Examples

Before decompose-padding, in TensorIR, the IR is:

@T.prim_func
def before_decompose(x: T.Buffer(128, "int32"), y: T.Buffer(140, "int32")):
    for i in range(140):
        with T.block("block"):
            vi = T.axis.remap("S", [i])
            y[vi] = T.if_then_else(vi >= 6 and vi < 134, x[vi - 6], 0, dtype="int32")

Create the schedule and do decompose-padding with specified loop:

sch = tir.Schedule(before_decompose, debug_mask="all")
block = sch.get_block("block")
sch.decompose_padding(block, sch.get_loops(block)[0])
print(sch.mod["main].script())

After applying decompose-padding, the IR becomes:

@T.prim_func
def after_decompose(x: T.Buffer(128, "int32"), y: T.Buffer(140, "int32")):
    for i in T.serial(140):
        with T.block("block_pad_const"):
            vi = T.axis.spatial(140, i)
            y[vi] = 0
    for i in T.serial(128):
        with T.block("block"):
            vi = T.axis.spatial(128, i)
            y[vi + 6] = x[vi]

decompose_reduction(block: BlockRV | str, loop: LoopRV) → BlockRV¶

Decompose a reduction block into two separate blocks.

The init block, which is translated from the init statement of the reduction block;
The update block, which is the original block without init statement.

The init block is inserted right before the given loop.

The schedule primitive requires:

The input block is a reduction block.
The input loop is the ancestor of the block.
The input loop is not lower than all the loops related to reduce block var.

Parameters:

block (Union[BlockRV, str]) – The reduction block to be decomposed
loop (LoopRV) – The loop above which the init block is inserted before.

Returns:

init_block – The init block

Return type:

BlockRV

Examples

Before decompose-reduction, in TensorIR, the IR is:

@T.prim_func
def before_decompose(a: ty.handle, c: ty.handle) -> None:
    A = tir.match_buffer(a, [128, 128])
    B = tir.match_buffer(b, [128, 128])
    C = tir.match_buffer(c, [128, 128])
    for i, j, k in tir.grid(128, 128, 128):
        with tir.block([128, 128, tir.reduce_axis(0, 128)], "C") as [vi, vj, vk]:
            with tir.init():
                C[vi, vj] = 0.0
            C[vi, vj] = C[vi, vj] + A[vi, vk] * B[vj, vk]

Create the schedule and do decompose-reduction with specified loop:

sch = tir.Schedule(before_decompose)
C = sch.get_block("C")
i, j, k = sch.get_loops(C)
sch.decompose_reduction(C, i)
print(sch.mod["main"].script())

After applying decompose-reduction, the IR becomes:

@T.prim_func
def after_decompose(a: ty.handle, c: ty.handle) -> None:
    A = tir.match_buffer(a, [128, 128])
    B = tir.match_buffer(b, [128, 128])
    C = tir.match_buffer(c, [128, 128])
    for i in tir.serial(128):
        for j in tir.serial(128):
            with tir.block([128, 128]) as [vi, vj]:
                C[vi, vj] = 0.0
    for i, j, k in tir.grid(128, 128, 128):
        with tir.block([128, 128, tir.reduce_axis(0, 128)], "C") as [vi, vj, vk]:
            C[vi, vj] = C[vi, vj] + A[vi, vk] * B[vj, vk]

enter_postproc() → None¶: A no-op that marks the start of postprocessing phase of scheduling

fork_seed() → int¶

Returns a forked random state as seed for new schedules

Returns:: seed – The forked random state, not the same as the current random state
Return type:: int

property func_working_on: GlobalVar | None¶: Returns the GlobalVar of the func that the schedule is currently working on

fuse(*loops: List[LoopRV], preserve_unit_iters: bool = True) → LoopRV¶

Fuse a list of consecutive loops into one. It requires: 1) The loops can’t have annotations or thread bindings. 2) The (i+1)-th loop must be the only child of the i-th loop. 3) All loops must start with 0. 4) The domain of a loop to be fused cannot depend on another loop to be fused.

Parameters:: *loops (List[LoopRV]) – The loops to be fused
Returns:: fused_loop – The new loop after fusion
Return type:: LoopRV

Examples

Before applying fuse, in TensorIR, the IR is:

@T.prim_func
def before_fuse(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

Create the schedule and do fuse:

sch = tir.Schedule(before_fuse)
i, j = sch.get_loops(sch.get_block("B"))
sch.fuse(i, j)
print(sch.mod["main"].script())

After applying fuse, the IR becomes:

@T.prim_func
def after_fuse(a: T.handle, b: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    B = T.match_buffer(b, (128, 128))
    # the 2 loops are fused into 1
    for i_j_fused in T.serial(0, 16384):
        with T.block("B"):
            vi = T.axis.S(128, T.floordiv(i_j_fused, 128))
            vj = T.axis.S(128, T.floormod(i_j_fused, 128))
            B[vi, vj] = A[vi, vj] * 2.0

Returns: - the corresponding Block that a BlockRV evaluates to; - the corresponding For that a LoopRV evaluates to; - the corresponding integer that a ExprRV evaluates to; - the corresponding Block that a block sref points to; - the corresponding For that a loop sref points to;

Parameters:: rand_var_or_sref (Union[ExprRV, BlockRV, LoopRV, StmtSRef]) – The random variable / sref to be evaluated
Returns:: result – The corresponding result
Return type:: Optional[Union[int, Block, For]]

get_block(name: str, func_name: str | None = None) → BlockRV¶

Retrieve a block in a specific function with its name

By default, if func_name is not specified, the schedule will search for the block in the function that is currently being “worked on”. To switch the function to be worked on, use work_on before calling this method.

Parameters:

name (str) – The name of the block
func_name (Optional[str] = None) – The name of the function

Returns:

block – The block retrieved IndexError is raised if 0 or multiple blocks exist with the specific name.

Return type:

BlockRV

get_child_blocks(block_or_loop: BlockRV | LoopRV) → List[BlockRV]¶

Get the leaf blocks of a specific block/loop

Parameters:: block_or_loop (Union[BlockRV, LoopRV]) – The query block/loop
Returns:: blocks – A list of leaf blocks inside a specific block/loop
Return type:: List[LoopRV]

get_consumers(block: BlockRV | str) → List[BlockRV]¶

Get the consumers of a specific block

Parameters:: block (Union[BlockRV, str]) – The block in the query
Returns:: consumers – A list of consumers of the given block
Return type:: List[BlockRV]

get_loops(block: BlockRV | str) → List[LoopRV]¶

Get the parent loops of the block in its scope, from outer to inner

Parameters:: block (Union[BlockRV, str]) – The query block
Returns:: loops – A list of loops above the given block in its scope, from outer to inner
Return type:: List[LoopRV]

get_output_blocks(scope_block: BlockRV | str) → List[BlockRV]¶

Get the list of output blocks within the given scope An output block is a block which has atleast one buffer being written to, but is not allocated within the PrimFunc

Parameters:: scope_block (Union[BlockRV, str],) – The scope block from which output blocks are collected
Returns:: output_blocks – A list of all blocks that write to some output buffer
Return type:: List[BlockRV]

get_producers(block: BlockRV | str) → List[BlockRV]¶

Get the producers of a specific block

Parameters:: block (Union[BlockRV, str]) – The block in the query
Returns:: producers – A list of producers of the given block
Return type:: List[BlockRV]

get_sref(rand_var_or_stmt: BlockRV | LoopRV | Block | For) → StmtSRef | None¶

Returns the corresponding sref to the given 1) LoopRV 2) BlockRV 3) Block 4) For

Parameters:: rand_var_or_stmt (Union[BlockRV, LoopRV, Block, For]) – The random variable / sref to be evaluated
Returns:: result – The corresponding result
Return type:: Optional[StmtSRef]

loop_partition(loop: LoopRV, factors: List[int | PrimExpr | None], preserve_unit_iters: bool = True) → List[LoopRV]¶

Partition a loop into a list of consecutive loops. It requires: 1) The loop can’t have annotation or thread binding. Predicates may be added to ensure the total loop numbers keeps unchanged. In factors, at most one of the factors can be None, which will be automatically inferred.

Parameters:

loop (LoopRV) – The loop to be partition
factors (List[Union[int, ExprRV, None]]) – The partitioning factors Potential inputs are: - None - ExprRV - Positive constant integers
preserve_unit_iters (bool) – Whether or not to preserve unit iterators in block bindings

Returns:

partition_loops – The new loops after partition

Return type:

List[LoopRV]

Examples

Before partition, in TensorIR, the IR is:

@T.prim_func
def before_partition(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

Create the schedule and do partition:

sch = tir.Schedule(before_partition)
i, j = sch.get_loops(sch.get_block("B"))
sch.partition(i, factors=[2, 64])
print(sch.mod["main"].script())

After applying partition, the IR becomes:

def after_partition(a: T.handle, b: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    B = T.match_buffer(b, (128, 128))
    # the original loop is partition into 3 loops
    with T.block("root"):
        T.reads()
        T.writes()
        with T.block("B_i_common"):
            T.reads()
            T.writes()
            with T.block("B_i0_partition"):
                T.reads()
                T.writes()
                for i0, j in T.grid(2, 128):
                    with T.block("B_i0"):
                        vi, vj = T.axis.remap("SS", [i0, j])
                        T.reads(A[0:2, 0:128])
                        T.writes(B[0:2, 0:128])
                        B[vi, vj] = A[vi, vj] * T.float32(2)
            with T.block("B_i1_partition"):
                T.reads()
                T.writes()
                for i1 in range(2, 66):
                    for j in range(128):
                        with T.block("B_i1"):
                            vi, vj = T.axis.remap("SS", [i1, j])
                            T.reads(A[2:66, 0:128])
                            T.writes(B[2:66, 0:128])
                            B[vi, vj] = A[vi, vj] * T.float32(2)
            with T.block("B_partition_2"):
                T.reads()
                T.writes()
                for i2 in range(66, 128):
                    for j in range(128):
                        with T.block("B_i2"):
                            vi, vj = T.axis.remap("SS", [i2, j])
                            T.reads(A[66:128, 0:128])
                            T.writes(B[66:128, 0:128])
                            B[vi, vj] = A[vi, vj] * T.float32(2)

merge(*loops: List[LoopRV]) → LoopRV¶

Merge a list of loops into one. The loops under their LCA requires: 1) Under the same scope. 2) Can’t have annotations or thread bindings. 3) Start with 0 and have same extent and same nesting depth. 4) From target loop to their LCA, The inner loop must be the only child of the outer loop.

Parameters:: *loops (List[LoopRV]) – The loops to be merged
Returns:: fused_loop – The new loop after merge
Return type:: LoopRV

Examples

Before applying merge, in TensorIR, the IR is:

@T.prim_func
def before_merge(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 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
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = A[vi, vj] * 2.0

Create the schedule and do fuse:

sch = tir.Schedule(before_fuse)
i1, _ = sch.get_loops(sch.get_block("B"))
i2, _ = sch.get_loops(sch.get_block("C"))
sch.merge(i1, i2)
print(sch.mod["main"].script())

After applying fuse, the IR becomes:

@T.prim_func
def after_fuse(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))
    # the 2 loops are merged into 1
    for i_m in range(128):
        for j in range(128):
            with T.block("B"):
                vi, vj = T.axis.remap("SS", [i_m, j])
                T.reads(A[vi, vj])
                T.writes(B[vi, vj])
                B[vi, vj] = A[vi, vj] * T.float32(2)
        for j in range(128):
            with T.block("C"):
                vi, vj = T.axis.remap("SS", [i_m, j])
                T.reads(A[vi, vj])
                T.writes(C[vi, vj])
                C[vi, vj] = A[vi, vj] * T.float32(2)

property mod: IRModule¶: Returns the AST of the module being scheduled

pad_einsum(block: BlockRV | str, padding: List[int]) → None¶

Pad the computation of Einsum.

On a block with trivial binding, this primitive pads the iteration domain of the block by the given padding factors, for example, 127 -> 128, 132 -> 144 when padding factor is 16. Extra producer and consumer padding blocks will be generated to avoid out-of-bound buffer access.

Einsum pattern means all the indices on the buffer access are either by constants (e.g. B[0]) or by variables (e.g. B[i]), but not by composite expressions (e.g. B[i + 1]).

Parameters:

block (Union[BlockRV, str]) – The block that matches the Einsum pattern.
padding (List[int]) – The padding for each block iter.

Examples

Before applying pad-einsum, in TensorIR, the IR is:

@T.prim_func
def before_pad_einsum(
    A: T.Buffer((127, 127), "float32"),
    B: T.Buffer((127, 127), "float32"),
    C: T.Buffer((127, 127), "float32"),
) -> None:
    for i0, i1, i2 in T.grid(127, 127, 127):
        with T.block("C_shared"):
            i, j, k = T.axis.remap("SSR", [i0, i1, i2])
            with T.init():
                C[i, j] = T.float32(0)
            C[i, j] = C[i, j] + A[i, k] * B[k, j]

Create the schedule and do pad-einsum with specified block:

sch = tir.Schedule(before_pad_einsum, debug_mask="all")
block = sch.get_block("C_shared")
sch.pad_einsum(block, [32, 32, 32])
print(sch.mod["main"].script())

After applying decompose-padding, the IR becomes:

@T.prim_func
def main(
    A: T.Buffer((127, 127), "float32"),
    B: T.Buffer((127, 127), "float32"),
    C: T.Buffer((127, 127), "float32"),
):
    # with T.block("root"):
    A_pad = T.alloc_buffer((128, 128))
    B_pad = T.alloc_buffer((128, 128))
    C_pad = T.alloc_buffer((128, 128))
    for i0, i1 in T.grid(128, 128):
        with T.block("A_pad"):
            v0, v1 = T.axis.remap("SS", [i0, i1])
            A_pad[v0, v1] = T.if_then_else(
                v0 < 127 and v1 < 127,
                A[v0, v1],
                T.float32(0),
            )
    for i0, i1 in T.grid(128, 128):
        with T.block("B_pad"):
            v0, v1 = T.axis.remap("SS", [i0, i1])
            B_pad[v0, v1] = T.if_then_else(
                v0 < 127 and v1 < 127,
                B[v0, v1],
                T.float32(0),
            )
    for i0, i1, i2 in T.grid(128, 128, 128):
        with T.block("C_shared"):
            i, j, k = T.axis.remap("SSR", [i0, i1, i2])
            with T.init():
                C_pad[i, j] = T.float32(0)
            C_pad[i, j] = C_pad[i, j] + A_pad[i, k] * B_pad[k, j]
    for i0, i1 in T.grid(127, 127):
        with T.block("C_pad"):
            v0, v1 = T.axis.remap("SS", [i0, i1])
            C[v0, v1] = C_pad[v0, v1]

parallel(loop: LoopRV) → None¶

Parallelize the input loop. It requires: 1) The scope block that the loop is in should have stage-pipeline property 2) All the blocks under the loop are complete blocks or reduction blocks, and have affine bindings 3) For each block under the loop, the loop can only be contained in data-parallel block iters’ bindings

Parameters:: loop (LoopRV) – The loop to be parallelized

Examples

Before parallel, in TensorIR, the IR is:

@T.prim_func
def before_parallel(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

Create the schedule and do parallel:

sch = tir.Schedule(before_parallel)
i, j = sch.get_loops(sch.get_block("B"))
sch.parallel(i)

After applying parallel, the IR becomes:

@T.prim_func
def after_parallel(a: T.handle, b: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    B = T.match_buffer(b, (128, 128))
    for i in T.parallel(0, 128):
        for j in T.serial(0, 128):
            with T.block("B"):
                vi, vj = T.axis.remap("SS", [i, j])
                B[vi, vj] = A[vi, vj] * 2.0

reindex(block: BlockRV | str, buffer: Tuple[str, int] | str | Buffer) → BlockRV¶

Create a block that read/write a buffer region into a read/write cache with reindexing. The layout of the cache will be the same as by the iterators of the block that reads/writes the buffer. It requires: 1) There is only one block who reads/writes the target buffer 2) There is only one buffer load/store of this buffer in the block

Parameters:

block (Union[BlockRV, str]) – The block that accesses the target buffer. If a string, this must uniquely identify a block.
buffer (Union[Tuple[str,int], Buffer, str]) –
The buffer to be transformed, or a specification of how to identify the buffer to be transformed.

If buffer if a tuple of (str,int), the first item should be either “read” or “write”, and the second item is an index into the block’s read or write regions.

If buffer is a string, it is the name of the buffer, which must exist within the reads/writes of the block. In addition, the reads/writes of the block may not contain more than one buffer with this name.

If buffer is a Buffer object, it must exist within the reads/writes of the block.

Returns:

reindex_block – The block of the reindex stage

Return type:

BlockRV

Examples

Before reindex, in TensorIR, the IR is:

@T.prim_func
def before_reindex(
    A: T.Buffer((128, 128), "float32"),
    B: T.Buffer((128, 128), "float32")
) -> None:
    for i, j in T.grid(128, 128):
        with T.block("B"):
            vi, vj = T.axis.remap("SS", [i, j])
            B[vi, vj] = A[vj, vi] * 2.0

Create the schedule and do reindex:

sch = tir.Schedule(before_reindex)
block = sch.get_block("B")
sch.reindex(block, ("read", 0))

After applying reindex, the IR becomes:

@T.prim_func
def after_reindex(
    A: T.Buffer((128, 128), "float32"),
    B: T.Buffer((128, 128), "float32")
) -> None:
    A_reindex = T.alloc_buffer((128, 128), "float32")
    for i, j in T.grid(128, 128):
        with T.block("A_reindex"):
            vi, vj = T.axis.remap("SS", [i, j])
            A_reindex[vi, vj] = A[vj, vi]
    for i, j in T.grid(128, 128):
        with T.block("B"):
            vi, vj = T.axis.remap("SS", [i, j])
            B[vi, vj] = A_reindex[vi, vj] * 2.0

reindex_cache_read(block: BlockRV | str, read_buffer_index: int, storage_scope: str, index_map: IndexMap | Callable) → BlockRV¶

Create a block that reads a buffer region into a read cache using customized indices specified by index map. The read region of the buffer must be a single point.

The cache stage block follows the original order of loops and block itervars in the block. If a block itervar does not appear in the buffer access region, it and its corresponding loop variables will be omitted. User can then use transform_block_layout primitive to reorder the block itervars and surrounding loops of the cache read/write block.

Unlike cache_read, reindex_cache_read only supports single consumer, please use cache_read when there are multiple consumers.

Parameters:

block (BlockRV) – The consumer block of the target buffer.
read_buffer_index (int) – The index of the buffer in block’s read region.
storage_scope (str) – The target storage scope.
index_map (Union[IndexMap, Callable]) – User defined indices to access allocated cache buffer, maps from block iter vars.

Returns:

cached_block – The block of the cache stage

Return type:

BlockRV

Examples

Before reindex_cache_read, in TensorIR, the IR is:

@T.prim_func
def before_reindex_cache_read(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

Create the schedule and reindex_cache_read:

sch = tir.Schedule(before_cache_read)
block_b = sch.get_block("B")
sch.reindex_cache_read(block_b, 0, "local", lambda vi, vj: (vj, vi))
print(sch.mod["main"].script())

After applying reindex_cache_read, the IR becomes:

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

See also

reorder

reverse_compute_at(block: BlockRV | str, loop: LoopRV, preserve_unit_loops: bool = False, index: int = -1) → None¶

Reverse-Compute-At. Move a consumer block under the specific loop, and regenerate the loops induced by the block so that the buffer region consumed by the consumer block could cover those regions produced by its producer blocks under the given loop. It requires:

block and loop are under the same scope, loop is not the ancestor of block
The scope block has stage-pipeline property

3) The subtree of the scope block, where the given block is in, satisfies the compact dataflow condition. i.e. all the blocks in the scope block’s subtree must be either complete block or reduction block

All the producers of the block are under the given loop

Parameters:

block (Union[BlockRV, str]) – The block to be moved
loop (LoopRV) – The loop where the block to be moved under
preserve_unit_loops (bool) – Whether to keep the trivial loops whose extents are 1
index (int) – The block index of the loop body subtree blocks: - index = -1 means inserted into the last possible insertion point; - index = -2 means inserted into the first possible insertion point; - Otherwise, index is a nonnegative number that indicates the insertion point

Examples

Before reverse-compute-at, in TensorIR, the IR is:

@T.prim_func
def before_reverse_compute_at(a: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (128, 128), "float32")
    B = T.alloc_buffer((128, 128), "float32")
    C = T.match_buffer(c, (128, 128), "float32")
    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
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = B[vi, vj] + 1.0

Create the schedule and do reverse-compute-at:

sch = tir.Schedule(before_reverse_compute_at)
block = sch.get_block("C")
loop, _ = sch.get_loops(sch.get_block("B"))
sch.reverse_compute_at(block, loop, preserve_unit_loops=False)
print(sch.mod["main"].script())

After applying reverse-compute-at, the IR becomes:

@T.prim_func
def after_reverse_compute_at(a: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (128, 128), "float32")
    B = T.alloc_buffer((128, 128), "float32")
    C = T.match_buffer(c, (128, 128), "float32")
    for i in T.serial(0, 128):
        for j in T.serial(0, 128):
            with T.block("B"):
                vi, vj = T.axis.remap("SS", [i, j])
                B[vi, vj] = A[vi, vj] * 2.0
        for j in T.serial(0, 128):
            with T.block("C"):
                vi, vj = T.axis.remap("SS", [i, j])
                C[vi, vj] = B[vi, vj] + 1.0

reverse_compute_inline(block: BlockRV | str) → None¶

Inline a block into its only producer. It requires:

The block is a complete non-root block, which only produces and consumes one buffer
The block must not be the only leaf in the scope.
The only producer of the block is a read-after-write producer and a complete non-root block
The body of the block must be a BufferStore statement in the form of, B[f(i, j, k, ...)] = g(i, j, k, A[i, j, k, ...] ...) where the indices of each BufferLoad on the RHS are all distinct atomic variables, and no variables other than those indexing variables are allowed in the statement.

Parameters:: block (Union[BlockRV, str]) – The block to be inlined to its producer

Examples

Before reverse-compute-inline, in TensorIR, the IR is:

@T.prim_func
def before_inline(a: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    B = T.alloc_buffer((128, 128))
    C = T.match_buffer(c, (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
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = B[vi, vj] + 1.0

Create the schedule and do reverse-compute-inline:

sch = tir.Schedule(before_inline)
sch.reverse_compute_inline(sch.get_block("C"))
print(sch.mod["main"].script())

After applying reverse-compute-inline, the IR becomes:

@T.prim_func
def after_inline(a: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    C = T.match_buffer(c, (128, 128))
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = A[vi, vj] * 2.0 + 1.0

rfactor(loop: LoopRV, factor_axis: int) → BlockRV¶

Factorize an associative reduction block by the specified loop.

An associative reduction cannot be parallelized directly, because it leads to potential race condition during accumulation. Alternatively, the reduction could be factorized on a loop with the following steps: - Step 1: evenly slice the reduction into n separate chunks, where n is the loop extent - Step 2: compute the chunks separately and write the result into n intermediate buffers; - Step 3: accumulate the n separate buffer into the result buffer. Note that the Step 2 above introduces opportunities for parallelization.

RFactor is a schedule primitive that implements the transformation described above: Given a block that writes to buffer B, it factorizes a loop of extent n.

For example, the pseudocode below accumulates B[i] = sum(A[i, : , : ]):

for i in range(128):                    # loop i is a data parallel loop
    for j in range(128):                # loop j is a reduction loop
        for k in range(128):            # loop k is a reduction loop
            B[i] = B[i] + A[i, j, k]

Suppose RFactor is applied on the innermost loop k and factor_axis = 1. RFactor then creates an intermediate buffer and two blocks.

1. The intermediate buffer, or “rf-buffer” is a buffer of rank ndim(B) + 1 and size size(B) * n, whose shape expands from shape(B) by adding an axis of n at the position specified by factor_axis. For example,

shape(B) = [1, 2, 3], factor_axis = 0 => shape(B_rf) = [n, 1, 2, 3]

shape(B) = [1, 2, 3], factor_axis = 1 => shape(B_rf) = [1, n, 2, 3]

shape(B) = [1, 2, 3], factor_axis = 2 => shape(B_rf) = [1, 2, n, 3]

shape(B) = [1, 2, 3], factor_axis = 3 => shape(B_rf) = [1, 2, 3, n]

2. The rfactor block, or “rf-block”, is a block that writes to the rf-buffer without accumulating over the loop k, i.e. the loop k is converted from a reduction loop to a data parallel loop. In our example, the rf-block is:

B_rf = np.zeros((128, 128))     # the rf-buffer
for k in range(128):            # loop k is converted to a data parallel loop
    for i in range(128):        # loop i is a data parallel loop (unchanged)
        for j in range(128):    # loop j is a reduction loop (unchanged)
            B_rf[i, k] = B_rf[i, k] + A[i, j, k]

3. The write-back block, or wb-block, is a block that accumulates the rf-buffer into the result buffer. All the reduction loops are removed except the loop k for accumulation. In our example, the wb-block is:

for i in range(128):            # loop i is a data parallel loop (unchanged)
                                # loop j is removed because it is a reduction loop
    for k in range(128):        # loop k is a reduction loop (unchanged)
        B[i] = B[i] + B_rf[i, k]

Parameters:

loop (LoopRV) – The loop outside block for which we want to do rfactor
factor_axis (int) – The position where the new dimension is placed in the new introduced rfactor buffer

Returns:

rf_block – The block which computes partial results over each slices (i.e., the first block as described in the above illustration)

Return type:

BlockRV

Examples

Before rfactor, in TensorIR, the IR is:

@T.prim_func
def before_rfactor(a: T.handle, b: T.handle) -> None:
    A = T.match_buffer(a, (128, 128, 128))
    B = T.match_buffer(b, (128,))
    for ii, i, j in T.grid(128, 128, 128):
    with T.block("B"):
        vii, vi, vj = T.axis.remap("SRR", [ii, i, j])
        with T.init():
            B[vii] = 0.0
        B[vii] = B[vii] + A[vii, vi, vj]

Create the schedule and do rfactor:

sch = tir.Schedule(before_rfactor)
_, _, k = sch.get_loops(sch.get_block("B"))
sch.rfactor(k, 0)
print(sch.mod["main"].script())

After applying rfactor, the IR becomes:

@T.prim_func
def after_rfactor(a: T.handle, b: T.handle) -> None:
    A = T.match_buffer(a, [128, 128, 128])
    B = T.match_buffer(b, [128])
    B_rf = T.alloc_buffer([128, 128])
    for i2, ii, i in T.grid(128, 128, 128):
        with T.block("B_rf"):
            vi2, vii, vi = T.axis.remap("SSR", [i2, ii, i])
            with T.init():
                B_rf[vi2, vii] = 0.0
            B_rf[vi2, vii] = (B_rf[vi2, vii] + A[vii, vi, vi2])
    for ii, i2 in T.grid(128, 128):
        with T.block("B"):
            vii, vi2 = T.axis.remap("SR", [ii, i2])
            with T.init():
                B[vii] = 0.0
            B[vii] = B[vii] + B_rf[vi2, vii]

Note

Rfactor requires: 1) loop has only one child block, and it is a reduction block; 2) loop is a reduction loop, i.e. the loop variable is bound to only reduction variables in the block binding; 3) loop is not parallelized, vectorized, unrolled or bound to any thread axis; 4) The block scope that loop is in is a staged-pipeline; 5) The outermost loop outside the reduction block should has the reduction block as its first child block; 6) The outermost reduction loop should have only one child block; 7) An unary extent loop that is not bound to any reduction or data parallel variables in the block binding should not appear under some reduction loop; 8) The reduction block should write to only one buffer, and its init and body are both simple BufferStore`s, and the pattern is registered as an associative reducer. The pre-defined patterns include: plus, multiplication, min and max; 9) Each of the loops on top of the block cannot be bound to a data parallel and a reduction block binding at the same time; 10) `factor_axis should be in range [-ndim(B) - 1, ndim(B)], where B is the buffer that the reduction block writes to. Negative indexing is normalized according to numpy convention.

rolling_buffer(block: BlockRV | str, write_buffer_index: int) → None¶

Compute the target buffer via rolling buffering, select the outermost rollable axis with a positive bound overlap that appears in the block’s ancestor loops as rolling axis, fold and circularize the buffer along the rolling dimension, append block predicate to avoid recomputing overlapping elements. It requires:

The block is not an output block and has only RAW dependencies.
The buffer to be an intermediate buffer defined via alloc_buffer.

3) The LCA of the producer and consumer of the buffer is a for loop, typically, the producer and consumer of the buffer are cascaded through compute_at.

4) The access region of the buffer has at least one dimension that contains a positive bound overlap.

Parameters:

block (Union[BlockRV, str]) – The producer block of the buffer.
write_buffer_index (int) – The index of the buffer in block’s write region.

Examples

Before rolling_buffer, in TensorIR, the IR is:

@T.prim_func
def before_rolling_buffer(
    A: T.Buffer((12, 12), "int8"), C: T.Buffer((8, 8), "int8")
) -> None:
    # body
    # with T.block("root")
    B = T.alloc_buffer([10, 10], dtype="int8")
    for i0, i1 in T.grid(2, 2):
        for ax0, ax1, ax2, ax3 in T.grid(6, 6, 3, 3):
            with T.block("B"):
                ax0_1 = T.axis.spatial(10, i0 * 4 + ax0)
                ax1_1 = T.axis.spatial(10, i1 * 4 + ax1)
                rv0, rv1 = T.axis.remap("RR", [ax2, ax3])
                B[ax0_1, ax1_1] = T.max(
                    B[ax0_1, ax1_1], A[ax0_1 + rv0, ax1_1 + rv1]
                )
        for ax0, ax1, ax2, ax3 in T.grid(4, 4, 3, 3):
            with T.block("C"):
                ax0_1 = T.axis.spatial(8, i0 * 4 + ax0)
                ax1_1 = T.axis.spatial(8, i1 * 4 + ax1)
                rv0, rv1 = T.axis.remap("RR", [ax2, ax3])
                C[ax0_1, ax1_1] = T.max(
                    C[ax0_1, ax1_1], B[ax0_1 + rv0, ax1_1 + rv1]
                )

Create the schedule and do rolling_buffer:

sch = tir.Schedule(before_rolling_buffer)
sch.rolling_buffer(sch.get_block("B"), write_buffer_index=0)
print(sch.mod["main"].script())

After applying rolling_buffer, the IR becomes:

@T.prim_func
def after_rolling_buffer(
    A: T.Buffer((12, 12), "int8"),
    C: T.Buffer((8, 8), "int8")
) -> None:
    # body
    # with T.block("root")
    B = T.alloc_buffer([6, 10], dtype="int8")
    for i0, i1 in T.grid(2, 2):
        for ax0, ax1, ax2, ax3 in T.grid(6, 6, 3, 3):
            with T.block("B"):
                T.where((i0 < 1 or 2 <= ax0) and (i1 < 1 or 2 <= ax1))
                ax0_1 = T.axis.spatial(10, i0 * 4 + ax0)
                ax1_1 = T.axis.spatial(10, i1 * 4 + ax1)
                rv0, rv1 = T.axis.remap("RR", [ax2, ax3])
                B[ax0_1 % 6, ax1_1] = T.max(
                    B[ax0_1 % 6, ax1_1], A[ax0_1 + rv0, ax1_1 + rv1]
                )
        for ax0, ax1, ax2, ax3 in T.grid(4, 4, 3, 3):
            with T.block("C"):
                ax0_1 = T.axis.spatial(8, i0 * 4 + ax0)
                ax1_1 = T.axis.spatial(8, i1 * 4 + ax1)
                rv0, rv1 = T.axis.remap("RR", [ax2, ax3])
                C[ax0_1, ax1_1] = T.max(
                    C[ax0_1, ax1_1], B[ax0_1 % 6 + rv0, ax1_1 + rv1]
                )

Note

The region_cover property of the consumer block of the target buffer will become false.

sample_categorical(candidates: List[int], probs: List[float], decision: int | None = None) → PrimExpr¶

Sample an integer given the probability distribution

Parameters:

candidates (List[int]) – The candidates to be sampled from
probs (List[float]) – The probability of each candidate
decision (Optional[int]) – The sampling decision, if any

Returns:

result – The random variable sampled from candidates

Return type:

ExprRV

sample_compute_location(block: BlockRV | str, decision: int | None = None) → LoopRV¶

Sample a compute-at location of the given block

Parameters:

block (Union[BlockRV, str]) – The block whose compute-at location is to be sampled
decision (Optional[int]) – The sampling decision

Returns:

result – The sampled loop where the input block is to be computed at

Return type:

LoopRV

sample_partitioned_tile(loop: LoopRV, n: int, partition_pos: int = 0, innerpart_factor: int = 1, decision: List[int] | None = None) → List[PrimExpr]¶

Sample the factors to a partitioned tile for a specific loop

Parameters:

loop (LoopRV) – The loop to be tiled
n (int) – The number of tiles to be sampled
partition_pos (int) – The position to partition tiles to two parts
innerpart_factor (int) – The factor of the second part
decision (Optional[List[int]]) – The sampling decision, if any

Returns:

result – A list of length n, the random partitioned tile sizes sampled

Return type:

List[ExprRV]

sample_perfect_tile(loop: LoopRV, n: int, max_innermost_factor: int = 16, decision: List[int] | None = None) → List[PrimExpr]¶

Sample the factors to perfect tile a specific loop

Parameters:

loop (LoopRV) – The loop to be tiled
n (int) – The number of tiles to be sampled
max_innermost_factor (int) – The maximum tile size allowed to be sampled in the innermost loop
decision (Optional[List[int]]) – The sampling decision, if any

Returns:

result – A list of length n, the random perfect tile sizes sampled

Return type:

List[ExprRV]

seed(seed: int) → None¶

Seed the randomness

Parameters:: seed (int) – The new random seed, -1 if use device random, otherwise non-negative

set_axis_separator(block: BlockRV | str, buffer: Tuple[str, int] | str | Buffer, axis_separators: List[int] | None) → None¶

Set the axis separator of a buffer, where the buffer is specified by a block and a read or write index.

Parameters:

block (Union[BlockRV, str]) – The block that accesses the target buffer. If a string, this must uniquely identify a block.
buffer (Union[Tuple[str,int], Buffer, str]) –
The buffer to be transformed, or a specification of how to identify the buffer to be transformed.

If buffer if a tuple of (str,int), the first item should be either “read” or “write”, and the second item is an index into the block’s read or write regions.

If buffer is a string, it is the name of the buffer, which must exist within the reads/writes of the block. In addition, the reads/writes of the block may not contain more than one buffer with this name.

If buffer is a Buffer object, it must exist within the reads/writes of the block.
axis_separators (Optional[List[int]]) – The axis separators.

Examples

Before set_axis_separator, in TensorIR, the IR is:

@T.prim_func
def before_set_axis_separator(
    A: T.Buffer((128, 128), "float32"), C: T.Buffer((128, 128), "float32")
) -> None:
    B = T.alloc_buffer((128, 128), dtype="float32")

    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
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = B[vi, vj] + 1.0

Create the schedule and do set_axis_separator:

sch = tir.Schedule(before_set_axis_separator)
sch.set_axis_separators(sch.get_block("B"), buffer_index=0, buffer_index_type="write",
                        axis_separators=[1])
print(sch.mod["main"].script())

After applying set_axis_separator, the IR becomes:

@T.prim_func
def after_set_axis_separators(
    A: T.Buffer((128, 128), "float32"), C: T.Buffer((128, 128), "float32")
) -> None:
    B = T.alloc_buffer([128, 128], dtype="float32", axis_separators=[1])

    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] * T.float32(2)
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = B[vi, vj] + T.float32(1)

set_scope(block: BlockRV | str, buffer_index: int | str | Buffer, storage_scope: str) → None¶

Set the storage scope of a buffer, where the buffer is specified by the a block and a write-index.

Parameters:

block (Union[BlockRV, str]) – The producer block of the buffer
buffer_index (int) – The index of the buffer in block’s write region
storage_scope (str) – The storage scope to be set

Examples

Before set_scope, in TensorIR, the IR is:

@T.prim_func
def before_set_scope(
    A: T.Buffer((128, 128), "float32"), C: T.Buffer((128, 128), "float32")
) -> None:
    B = T.alloc_buffer((128, 128), dtype="float32")

    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
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = B[vi, vj] + 1.0

Create the schedule and do set_scope:

sch = tir.Schedule(before_set_scope)
sch.set_scope(sch.get_block("B"), buffer_index=0, storage_scope="shared")
print(sch.mod["main"].script())

After applying set_scope, the IR becomes:

@T.prim_func
def after_set_scope(
    A: T.Buffer((128, 128), "float32"), C: T.Buffer((128, 128), "float32")
) -> None:
    B_shared = T.alloc_buffer([128, 128], dtype="float32", scope="shared")

    for i, j in T.grid(128, 128):
        with T.block("B"):
            vi, vj = T.axis.remap("SS", [i, j])
            B_shared[vi, vj] = A[vi, vj] * T.float32(2)
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = B_shared[vi, vj] + T.float32(1)

Note

set_scope requires the buffer to be an intermediate buffer defined via alloc_buffer.

show(*args, **kwargs) → None¶

A sugar for print highlighted TVM script.

All parameters are forwarded to the underlying Module.show and Trace.show methods.

split(loop: LoopRV, factors: List[int | PrimExpr | None], preserve_unit_iters: bool = True) → List[LoopRV]¶

Split a loop into a list of consecutive loops. It requires: 1) The loop can’t have annotation or thread binding. 2) The loop must start with 0. Predicates may be added to ensure the total loop numbers keeps unchanged. In factors, at most one of the factors can be None, which will be automatically inferred.

Parameters:

loop (LoopRV) – The loop to be split
factors (List[Union[int, ExprRV, None]]) – The splitting factors Potential inputs are: - None - ExprRV - Positive constant integers
preserve_unit_iters (bool) – Whether or not to preserve unit iterators in block bindings

Returns:

split_loops – The new loops after split

Return type:

List[LoopRV]

Examples

Before split, in TensorIR, the IR is:

@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

Create the schedule and do split:

sch = tir.Schedule(before_split)
i, j = sch.get_loops(sch.get_block("B"))
sch.split(i, factors=[2, 64])
print(sch.mod["main"].script())

After applying split, the IR becomes:

@T.prim_func
def after_split(a: T.handle, b: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    B = T.match_buffer(b, (128, 128))
    # the original loop is split into 2 loops
    for i0, i1, j in T.grid(2, 64, 128):
        with T.block("B"):
            vi = T.axis.S(128, i0 * 64 + i1)
            vj = T.axis.S(128, j)
            B[vi, vj] = A[vi, vj] * 2.0

property state: ScheduleState¶: Returns the ScheduleState in the current schedule class

storage_align(block: BlockRV | str, buffer_index: int, axis: int, factor: int, offset: int) → None¶

Set alignment requirement for specific dimension such that stride[axis] == k * factor + offset for some k. This is useful to set memory layout 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.

Parameters:

block (Union[BlockRV, str]) – The producer block of the buffer.
buffer_index (int) – The index of the buffer in block’s write region.
axis (int) – The dimension to be specified for alignment.
factor (int) – The factor multiple of alignment.
offset (int) – The required offset factor.

Examples

Before storage_align, in TensorIR, the IR is:

@T.prim_func
def before_storage_align(a: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    B = T.alloc_buffer((128, 128))
    C = T.match_buffer(c, (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
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = B[vi, vj] + 1.0

Create the schedule and do storage_align:

sch = tir.Schedule(before_storage_align)
sch.storage_align(sch.get_block("B"), buffer_index=0, axis=0, factor=128, offset=1)
print(sch.mod["main"].script())

After applying storage_align, the IR becomes:

@T.prim_func
def after_storage_align(a: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    B = T.alloc_buffer((128, 128))
    C = T.match_buffer(c, (128, 128))
    for i, j in T.grid(128, 128):
        with T.block("B"):
            T.block_attr({"buffer_dim_align": [[[0, 128, 1]]]})
            vi, vj = T.axis.remap("SS", [i, j])
            B[vi, vj] = A[vi, vj] * 2.0
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = B[vi, vj] + 1.0

After lowering passes, buffer B will have strides as [129, 1].

Note

Storage_align requires the buffer to be an intermediate buffer defined via alloc_buffer.

tensorize(block_or_loop: BlockRV | LoopRV, tensor_intrin: str, preserve_unit_iters: bool = True) → None¶

Tensorize the computation enclosed by loop with the tensor intrinsic.

Parameters:

block_or_loop (Union[BlockRV, LoopRV]) – The loop to be tensorized.
tensor_intrin (str) – The tensor intrin or the name of the tensor intrin.
preserve_unit_iters (bool) – Whether or not to preserve unit iterators in block bindings

Examples

Before tensorize, in TensorIR, the IR is:

@T.prim_func
def before_tensorize(
    A: T.Buffer((128, 128), "float32"),
    B: T.Buffer((128, 128), "float32"),
    C: T.Buffer((128, 128), "float32"),
) -> None:
    # body
    # with T.block("root")
    for i_0, j_0, k_0, i_1, j_1, k_1 in T.grid(8, 8, 8, 16, 16, 16):
        with T.block("update"):
            vi = T.axis.spatial(128, i_0 * 16 + i_1)
            vj = T.axis.spatial(128, j_0 * 16 + j_1)
            vk = T.axis.reduce(128, k_0 * 16 + k_1)
            T.reads(C[vi, vj], A[vi, vk], B[vj, vk])
            T.writes(C[vi, vj])
            C[vi, vj] = C[vi, vj] + A[vi, vk] * B[vj, vk]

Declare and register the tensor intrinsic:

@T.prim_func
def mma_desc(a: T.handle, b: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (16, 16), align=128, offset_factor=1)
    B = T.match_buffer(b, (16, 16), align=128, offset_factor=1)
    C = T.match_buffer(c, (16, 16), align=128, offset_factor=1)

    with T.block("root"):
        T.reads(C[0 : 16, 0 : 16], A[0 : 16, 0 : 16], B[0 : 16, 0 : 16])
        T.writes(C[0 : 16, 0 : 16])
        for i, j, k in T.grid(16, 16, 16):
            with T.block("update"):
                vi, vj, vk = T.axis.remap("SSR", [i, j, k])
                C[vi, vj] = C[vi, vj] + A[vi, vk] * B[vj, vk]

@T.prim_func
def mma_intrin(a: T.handle, b: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (16, 16), align=128, offset_factor=1)
    B = T.match_buffer(b, (16, 16), align=128, offset_factor=1)
    C = T.match_buffer(c, (16, 16), align=128, offset_factor=1)

    with T.block("root"):
        T.reads(C[0 : 16, 0 : 16], A[0 : 16, 0 : 16], B[0 : 16, 0 : 16])
        T.writes(C[0 : 16, 0 : 16])
        T.evaluate(
            T.tvm_mma_sync(
                C.data,
                C.elem_offset // 256,
                A.data,
                A.elem_offset // 256,
                B.data,
                B.elem_offset // 256,
                C.data,
                C.elem_offset // 256,
                dtype="handle",
            )
        )

tir.TensorIntrin.register("test_mma_intrin", mma_desc, mma_intrin)

Create the schedule and do tensorize:

sch = tir.Schedule(before_tensorize)
update = sch.get_block("update")
_, _, _, i1, _, _ = sch.get_loops(update)
sch.tensorize(i1, "test_mma_intrin")
print(sch.mod["main"].script())

After applying tensorize, the IR becomes:

@T.prim_func
def after_tensorize(
    A: T.Buffer((128, 128), "float32"),
    B: T.Buffer((128, 128), "float32"),
    C: T.Buffer((128, 128), "float32"),
) -> None:
    # body
    # with T.block("root")
    for i_0, j_0, k_0 in T.grid(8, 8, 8):
        with T.block("update_o"):
            vio, vjo, vko = T.axis.remap("SSR", [i_0, j_0, k_0])
            T.reads(
                C[vio * 16 : vio * 16 + 16, vjo * 16 : vjo * 16 + 16],
                A[vio * 16 : vio * 16 + 16, vko * 16 : vko * 16 + 16],
                B[vjo * 16 : vjo * 16 + 16, vko * 16 : vko * 16 + 16],
            )
            T.writes(C[vio * 16 : vio * 16 + 16, vjo * 16 : vjo * 16 + 16])
            A_1 = T.match_buffer(
                A[vio * 16 : vio * 16 + 16, vko * 16 : vko * 16 + 16],
                [16, 16],
                dtype="float32",
                offset_factor=1,
            )
            B_1 = T.match_buffer(
                B[vjo * 16 : vjo * 16 + 16, vko * 16 : vko * 16 + 16],
                [16, 16],
                dtype="float32",
                offset_factor=1,
            )
            C_1 = T.match_buffer(
                C[vio * 16 : vio * 16 + 16, vjo * 16 : vjo * 16 + 16],
                [16, 16],
                dtype="float32",
                offset_factor=1,
            )
            T.evaluate(
                T.tvm_mma_sync(
                    C_1.data,
                    C_1.elem_offset // 256,
                    A_1.data,
                    A_1.elem_offset // 256,
                    B_1.data,
                    B_1.elem_offset // 256,
                    C_1.data,
                    C_1.elem_offset // 256,
                    dtype="handle",
                )
            )

property trace: Trace | None¶: Returns the internally maintained trace of scheduling program execution

transform_block_layout(block: BlockRV | str, index_map: IndexMap | Callable) → None¶

Apply a transformation represented by IndexMap to block

Parameters:

block (Union[BlockRV, str]) – The block to be transformed
index_map (Union[IndexMap, Callable]) – The transformation to apply.

Examples

Before transform_block_layout, in TensorIR, the IR is:

@T.prim_func
def before_transform_block_layout(
    A: T.Buffer((16, 16), "float32"),
    B: T.Buffer((16, 16), "float32")
) -> None:
    for i, j in T.grid(16, 16):
        with T.block("B"):
            vi, vj = T.axis.remap("SS", [i, j])
            B[vi, vj] = A[vi, vj] * 2.0

Create the schedule and do transform_block_layout:

sch = tir.Schedule(before_transform_block_layout)
sch.transform_block_layout(sch.get_block("B"), lambda i, j: (i * 16 + j,))
print(sch.mod["main"].script())

After applying transform_block_layout, the IR becomes:

@T.prim_func
def after_transform_block_layout(
    A: T.Buffer((16, 16), "float32"),
    B: T.Buffer((16, 16), "float32")
) -> None:
    for i in range(256):
        with T.block("B"):
            vi, = T.axis.remap("S", [i])
            B[vi // 16, vi % 16] = A[vi // 16, vi % 16] * 2.0

Apply a transformation represented by IndexMap to buffer

Parameters:

block (Union[BlockRV, str]) – The block that accesses the target buffer. If a string, this must uniquely identify a block.
buffer (Union[Tuple[str,int], Buffer, str]) –
The buffer to be transformed, or a specification of how to identify the buffer to be transformed.

If buffer if a tuple of (str,int), the first item should be either “read” or “write”, and the second item is an index into the block’s read or write regions.

If buffer is a string, it is the name of the buffer, which must exist within the reads/writes of the block. In addition, the reads/writes of the block may not contain more than one buffer with this name.

If buffer is a Buffer object, it must exist within the reads/writes of the block.
index_map (Union[IndexMap, Callable]) –
The transformation to apply.

If index_map is a callable, and the returned list contains IndexMap.AXIS_SEPARATOR, the SetAxisSeparators primitive will be called in addition to the TransformLayout primitive.
pad_value (Optional[Union[int, float, PrimExpr, IndexMap, Callable]]) –
The value to be used for any padding introduced by the transformation. If the schedule contains a producer block for the specified buffer, the pad value will be written as part of the producer block if possible, or after the producer block otherwise. Otherwise, if the buffer is an input, will insert an annotation block to state that the padding contains the known value.

The pad value may not contain instances of BufferLoad, except where it loads a value from the buffer being transformed (e.g. to create a circular buffer with padding that consists of repeated elements).

Note: If applied to an input buffer, the calling scope is responsible for ensuring that the pad_value is present. Algebraic symplifications, branch elimination, and other optimizations may assume that this precondition is met, and may result in incorrect results being returned.

If None, the transformation may not introduce padding.

If an int, float or PrimExpr, the transformation is the specific value to be present in the padding.

If an IndexMap or Callable, the transformation is the value to be present in the padding in terms of the transformed index.
assume_injective_transform (bool) – If set to true, the schedule primitive will assume the index_map is injective and skip checking overlapping of the mapped indices. This can be useful for complicated index_map that the analysis does not cover. It is the callers’ responsibility to ensure the index map is injective, otherwise, the correctness of the schedule is not guaranteed.

Examples

Before transform_layout, in TensorIR, the IR is:

@T.prim_func
def before_transform_layout(a: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (128, 128), "float32")
    B = T.alloc_buffer((128, 128), "float32")
    C = T.match_buffer(c, (128, 128), "float32")
    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
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = B[vi, vj] + 1.0

Create the schedule and do transform_layout:

sch = tir.Schedule(before_storage_align)
sch.transform_layout(sch.get_block("B"), buffer=("write",0),
                     index_map=lambda m, n: (m // 16, n // 16, m % 16, n % 16))
print(sch.mod["main"].script())

After applying transform_layout, the IR becomes:

@T.prim_func
def two_elementwise_transformed_intermediate_buffer(a: T.handle, c: T.handle) -> None:
    A = T.match_buffer(a, (128, 128), "float32")
    B = T.alloc_buffer((8, 8, 16, 16), "float32")
    C = T.match_buffer(c, (128, 128), "float32")
    for i, j in T.grid(128, 128):
        with T.block("B"):
            vi, vj = T.axis.remap("SS", [i, j])
            B[vi // 16, vj // 16, vi % 16, vj % 16] = A[vi, vj] * 2.0
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j])
            C[vi, vj] = B[vi // 16, vj // 16, vi % 16, vj % 16] + 1.0

unannotate(block_or_loop: BlockRV | LoopRV, ann_key: str) → None¶

Unannotate a block/loop’s annotation with key ann_key

Parameters:

block_or_loop (Union[BlockRV, LoopRV]) – The block/loop to be unannotated
ann_key (str) – The annotation key

Examples

Before unannotate, in TensorIR, the IR is:

@T.prim_func
def before_unannotate(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])
            T.block_attr({"ann_key", "ann_value"})
            B[vi, vj] = A[vi, vj] * 2.0

Create the schedule and do annotate:

sch = tir.Schedule(before_unannotate)
sch.unannotate(sch.get_block("B"), "ann_key")
print(sch.mod["main"].script())

After applying unannotate, the IR becomes:

@T.prim_func
def after_unannotate(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

unroll(loop: LoopRV) → None¶

Unroll the input loop. It requires nothing

Parameters:: loop (LoopRV) – The loop to be unrolled

Examples

Before unroll, in TensorIR, the IR is:

@T.prim_func
def before_unroll(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

Create the schedule and do unroll:

sch = tir.Schedule(before_unroll)
i, j = sch.get_loops(sch.get_block("B"))
sch.unroll(i)

After applying unroll, the IR becomes:

@T.prim_func
def after_unroll(a: T.handle, b: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    B = T.match_buffer(b, (128, 128))
    for i in T.unroll(0, 128):
        for j in T.serial(0, 128):
            with T.block("B"):
                vi, vj = T.axis.remap("SS", [i, j])
                B[vi, vj] = A[vi, vj] * 2.0

unsafe_hide_buffer_access(block: BlockRV, buf_type: str, buf_index_array: List[int]) → None¶

Hide some buffer access in a given block. This is an unsafe schedule primitive.

Parameters:

block (BlockRV) – The block where we hide read access.
buf_type (str) – The buffer type: “read”/”write”.
buf_index_array (List[int]) – The array of buffer indices we hide access.

Note

This schedule primitive is unsafe, and may fail dependency analysis. One use case of unsafe_hide_buffer_access is to hide the buffer access to indices buffers (e.g. in sparse computation) so that we can further tensorize the block (the indices buffers appeared in read/write regions may fail the pattern matching in tensorize primitive, and hide the access to these buffers could address the issue).

unsafe_set_dtype(block: BlockRV | str, buffer_index: int, dtype: str) → None¶

Set the data type of a buffer, where the buffer is specified by the a block and write-index.

This schedule primitive is unsafe and may change the correctness of program because of type conversion, please use with caution.

Parameters:

block (Union[BlockRV, str]) – The producer block of the buffer
buffer_index (int) – The index of the buffer in block’s write region
dtype (str) – The data type to be set

Examples

Before unsafe_set_dtype, in TensorIR, the IR is:

@T.prim_func
def before_set_dtype(
    A: T.Buffer((128, 128), "float32"), C: T.Buffer((128, 128), "float32")
) -> None:
    B = T.alloc_buffer((128, 128), dtype="float32")

    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
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j]
            C[vi, vj] = B[vi, vj] + 1.0

Create the schedule and do unsafe_set_dtype:

sch = tir.Schedule(before_set_dtype)
sch.unsafe_set_dtype("B", buffer_index=0, dtype="float16")
print(sch.mod["main"].script())

After applying set_dtype, the IR becomes:

@T.prim_func
def after_set_dtype(
    A: T.Buffer((128, 128), "float32"), C: T.Buffer((128, 128), "float32")
) -> None:
    B = T.alloc_buffer((128, 128), dtype="float16")

    for i, j in T.grid(128, 128):
        with T.block("B"):
            vi, vj = T.axis.remap("SS", [i, j])
            B[vi, vj] = T.cast(A[vi, vj] * 2.0, "float16")
    for i, j in T.grid(128, 128):
        with T.block("C"):
            vi, vj = T.axis.remap("SS", [i, j]
            C[vi, vj] = T.cast(B[vi, vj], "float32") + 1.0

Note

unsafe_set_dtype requires the buffer to be an intermediate buffer defined via alloc_buffer.

vectorize(loop: LoopRV) → None¶

Vectorize the input loop. It requires: 1) The scope block that the loop is in should have stage-pipeline property 2) All the blocks under the loop are complete blocks or reduction blocks, and have affine bindings 3) For each block under the loop, the loop can only be contained in data-parallel block iters’ bindings

Parameters:: loop (LoopRV) – The loop to be vectorized

Examples

Before vectorize, in TensorIR, the IR is:

@T.prim_func
def before_vectorize(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

Create the schedule and do vectorize:

sch = tir.Schedule(before_vectorize)
i, j = sch.get_loops(sch.get_block("B"))
sch.vectorize(j)

After applying vectorize, the IR becomes:

@T.prim_func
def after_vectorize(a: T.handle, b: T.handle) -> None:
    A = T.match_buffer(a, (128, 128))
    B = T.match_buffer(b, (128, 128))
    for i in T.serial(0, 128):
        for j in T.vectorized(0, 128):
            with T.block("B"):
                vi, vj = T.axis.remap("SS", [i, j])
                B[vi, vj] = A[vi, vj] * 2.0

work_on(func_name: str) → None¶

Instruct the schedule to work on a function in the IRModule.

By default, the schedule works on the function with the name “main”, or the only function in the IRModule if there is only one. If there is multiple functions in the IRModule, and none of their names are “main”, users will have to call this method to explicitly specify which function to work on.

This sugar function will guide the GetBlock method if its func_name is not specified.

Parameters:: func_name (str) – The name of the function to work on.

class tvm.tir.schedule.ScheduleDebugMask(value)¶

The bitmask of the debug_mask flag in the ScheduleState class.

If the debug_mask flag has a certain bit on, then the correpsonding verification pass will be conducted. For example, if (debug_mask & VERIFY_SREF_TREE) != 0, then the correctness of the sref tree will be verified after each schedule instruction.

VERIFY_SREF_TREE¶

Verify the correctness of the sref tree

Type:: int = 1

VERIFY_CACHED_FLAGS¶

Verify the correctness of affine_binding, region_cover and stage_pipeline

Type:: int = 2

exception tvm.tir.schedule.ScheduleError¶: Error that happens during TensorIR scheduling.

class tvm.tir.schedule.ScheduleState(mod: PrimFunc | IRModule, *, debug_mask: str | int = 'none', enable_check: bool = True)¶

The state of scheduling, which exposes a Replace method as the primary resort for all the scheduling primitives to manipulate the TensorIR.

The data structure contains the following information 1) The AST being scheduled (mod) 2) The sref tree of schedulable statements (indicated by the srefs) 3) The dependency information of each block scope (block_info) 4) A reverse mapping from the AST nodes to that in the sref tree (get_sref) 5) A debug flag, if set, extra checking is enabled (debug_mask) 6) A enable check flag, if False, some prerequisite checks are disabled.

Parameters:

mod (IRModule) – The AST of the module being scheduled
debug_mask (int) – Do extra correctness checking after the object construction and each time after calling the Replace method.
enable_check (bool) – Indicates whether we enable prerequisite checks for some schedule primitives or not, defaults to True.

get_block_scope(block_sref: StmtSRef) → BlockScope¶

Get the BlockScope correpsonding to the block sref

Parameters:: block_sref (StmtSRef) – The block sref to be retrieved
Returns:: sref – The corresponding sref
Return type:: StmtSRef

get_sref(stmt: Block | For) → StmtSRef | None¶

Return the corresponding sref that points to the stmt

Parameters:: stmt (Union[Block, For]) – The schedulable statement in the TensorIR to be retrieved for its sref
Returns:: sref – The corresponding sref
Return type:: StmtSRef

replace(src_sref: StmtSRef, tgt_stmt: Block | For | BlockRealize, block_sref_reuse: Dict[Block, Block] | None = None) → None¶

Replace the part of the AST, as being pointed to by src_sref, with a specific statement tgt_stmt, and maintain the sref tree accordingly. Replace will try to perform copy on write as much as possible when the ScheduleState holds the only copy to the IRModule and IR nodes.

Only 3 types of replacements are allowed: from src_sref->stmt to tgt_stmt. 1) Block -> Block 2) Loop -> Loop 3) Loop -> BlockRealize

Parameters:

src_sref (StmtSRef) – The sref to the statement to be replaced in the TensorIR AST
tgt_stmt (Union[Block, For, BlockRealize]) – The statement to be replaced to
block_sref_reuse (Optional[Dict[Block, Block]] = None) – Maps an old block (to be replaced in the subtree under src_sref->stmt) to a new block (replaced to, in the subtree under tgt_stmt), and enforces reuse of srefs between them (rather than create new srefs) i.e. after being replaced, the sref that points to the old block will point to the new one

Note

The reuse of loop srefs are detected automatically according to the reuse of loop vars.

class tvm.tir.schedule.StmtSRef¶

An object that refers to schedulable elements in the TensorIR, aka “sref”.

Glossary - Block sref: An StmtSref that points to a TensorIR block. - Loop sref: An StmtSRef that points to a TensorIR for loop. - Parent sref: The parent sref of an sref is the block/loop sref that points to its closest schedulable statement of its ancestors on the TensorIR AST. - Root sref: Sref to the root block. Every sref has exactly one parent sref except for root sref. - Sref tree: The parent-children-relationship of srefs that forms a tree, uniquely determined by the TensorIR AST.

static inline_mark() → StmtSRef¶: A special StmtSRef, which doesn’t point to any stmt in the AST, only serving as a “mark” to hint compute-at to do the work of compute-inline

property parent: StmtSRef | None¶: The parent sref

static root_mark() → StmtSRef¶: A special StmtSRef, which doesn’t point to any stmt in the AST, only serving as a “mark” to hint compute-at to do nothing

property stmt: Block | For | None¶: The block/for stmt the object refers to

class tvm.tir.schedule.Trace(insts: List[Instruction], decisions: Dict[Instruction, Any])¶

An execution trace of a scheduling program.

A trace has two parts: 1) The instructions invoked so far 2) The random decisions made upon those instructions, if any

A trace can be serialized to: 1) Roundtrippable JSON format: can be saved to file and loaded back 2) Python syntax: allows users to copy-paste the trace to reproduce the scheduling process

A trace can be applied to a TensorIR schedule by re-applying all its instructions possibly with their decisions accordingly. Re-sampling is invoked if a sampling instruction doesn’t have its corresponding decision; Otherwise the existing decision will be reused accordingly.

insts¶

The instructions invoked so far in the program execution

Type:: List[Instruction]

decisions¶

The random decisions made upon those instructions

Type:: Dict[Instruction, DECISION_TYPE]

append(inst: Instruction, decision: Any | None = None) → None¶

Append a new instruction to the trace

Parameters:

insts (Instruction) – The new instruction to be appended
decision (Optional[DECISION_TYPE] = None) – The random decision made on this instruction

static apply_json_to_schedule(json_obj: Any, sch: Schedule) → None¶

Apply a JSON-serialized trace to a TensorIR schedule

Parameters:

json_obj (JSON_TYPE) – The JSON-serialized trace
sch (Schedule) – The TensorIR schedule

apply_to_schedule(sch: Schedule, remove_postproc: bool, decision_provider: Callable[[Instruction, List[Any], List[Any], Any], Any] | None = None) → None¶

Apply the trace to a TensorIR schedule

Parameters:

sch (Schedule) – The schedule to be applied onto
remove_postproc (bool) – If postprocessing instructions are removed
decision_provider (Optional[Callable] = None) – A callback that allows users to mutate decisions on the fly when applying instructions. The signature of the callback is: - The 1st argument: The instruction - The 2nd argument: The input random variables - The 3rd argument: The attributes - The 4th argument: The decision - Return: A new decision

as_json(remove_postproc: bool = False) → Any¶

Serialize the trace as a JSON-style object

Parameters:: remove_postproc (bool = False) – If postprocessing instructions are removed
Returns:: json – The JSON-style object
Return type:: JSON_TYPE

as_python(remove_postproc: bool = False) → List[str]¶

Serialize the trace as a sequence of python statements

Parameters:: remove_postproc (bool = False) – If postprocessing instructions are removed
Returns:: py_stmts – A sequence of python statements
Return type:: List[str]

get_decision(inst: Instruction) → Any | None¶

Retrieve the decision made on a specific instruction

Parameters:: insts (Instruction) – The instruction whose decision is to be retrieved
Returns:: decision – The corresponding decision; None if there is no decision made on the instruction
Return type:: Optional[DECISION_TYPE]

pop() → Instruction | None¶

Remove the last instruction, along with the decision made on that instruction, if any

Returns:: popped_inst – Returns the instruction removed; NullOpt if the trace is empty
Return type:: Instruction

show(style: str | None = None, black_format: bool = False) → None¶

A sugar for print highlighted TVM script.

Parameters:

style (str, optional) – Pygmentize printing style, auto-detected if None. See tvm.script.highlight.cprint for more details.
black_format (bool) – If true, use the formatter Black to format the TVMScript. If None, determine based on the “TVM_BLACK_FORMAT” environment variable.

simplified(remove_postproc: bool) → Trace¶

Simplify the trace with dead-code elimination

Parameters:: remove_postproc (bool) – If postprocessing instructions are removed
Returns:: trace – A simplified trace
Return type:: Trace

with_decision(inst: Instruction, decision: Any, remove_postproc: bool) → Trace¶

Create a new trace with an instruction whose decision is changed, assuming this instruction exists in the resulting trace

Parameters:

inst (Instruction) – The instruction whose decision is to be changed
decision (DECISION_TYPE) – The decision to be changed to
remove_postproc (bool) – If postprocessing instructions are removed

Returns:

trace – The new trace with the decision changed

Return type:

Trace