xref: /llvm-project/mlir/docs/Bufferization.md (revision ced2fc7819d5ddea616ec330f18e08ff284c1868)

# Bufferization

[TOC]

## Overview

Bufferization in MLIR is the process of converting ops with `tensor` semantics
to ops with `memref` semantics. There are multiple MLIR passes that are related
to bufferization. These passes typically run as one of the last steps in a
pass pipeline, right before lowering to `memref` ops to LLVM. That is because
many transformations are easier or only supported in tensor land; e.g.,
[tile/fuse/… on tensors first](https://llvm.discourse.group/t/rfc-linalg-on-tensors-update-and-comprehensive-bufferization-rfc/3373),
then bufferize the remaining IR.

![bufferization passes](/includes/img/bufferization_passes.svg)

The most important bufferization pass is *One-Shot Bufferize*: This pass
rewrites `tensor` IR to `memref` IR. There are additional helper passes that
preprocess IR (e.g., so that IR can be bufferized more efficiently), perform
buffer-level optimizations such as allocation hoisting, and
[insert buffer deallocation ops](OwnershipBasedBufferDeallocation.md) so that
the resulting `memref` IR has no memory leaks.

## Deprecated Passes

The buffer deallocation pass has been deprecated in favor of the ownership-based
buffer deallocation pipeline. The deprecated pass has some limitations that may
cause memory leaks in the resulting IR.

## What is One-Shot Bufferize?

One-Shot Bufferize is a tensor bufferization pass designed for IR in
[destination-passing style](https://www.microsoft.com/en-us/research/wp-content/uploads/2016/11/dps-fhpc17.pdf),
and with aggressive in-place bufferization.

One-Shot Bufferize is:

*   **Monolithic**: A single MLIR pass does the entire work.

*   **Extensible** via an op interface: All ops that implement
    `BufferizableOpInterface` can be bufferized.

*   A **whole-function at a time analysis**. In-place bufferization decisions
    are made by analyzing SSA use-def chains on tensors. Op interface
    implementations not only provide the rewrite logic from tensor ops to memref
    ops, but also helper methods for One-Shot Bufferize's analysis to query
    information about an op's bufferization/memory semantics.

*   **2-Phase**: Bufferization is internally broken down into 2 steps: First,
    analyze the entire IR and make bufferization decisions. Then, bufferize
    (rewrite) the IR. The analysis has access to exact SSA use-def information.
    It incrementally builds alias and equivalence sets and does not rely on a
    posteriori-alias analysis from preallocated memory.

*   **Greedy**: Operations are analyzed one-by-one and it is decided on the spot
    whether a tensor OpOperand must be copied or not. Heuristics determine the
    order of analysis.

*   **Modular**: The current One-Shot Analysis can be replaced with a different
    analysis. The result of the analysis are queried by the bufferization via
    `AnalysisState`, in particular `AnalysisState::isInPlace`. Any derived class
    of `AnalysisState` that implements a small number virtual functions can
    serve as a custom analysis. It is even possible to run One-Shot Bufferize
    without any analysis (`AlwaysCopyAnalysisState`), in which case One-Shot
    Bufferize copies every buffer before writing to it.

Note that One-Shot Bufferize does not deallocate buffers. That is done by the
[Ownership-based Buffer Deallocation passes](OwnershipBasedBufferDeallocation.md).

## Goals of Bufferization

The high-level goal of every bufferization technique is to:

1. Use as little memory as possible.
2. Copy as little memory as possible.

This implies reusing already allocated buffers when possible, turning
bufferization into an algorithmically complex problem with similarities to
register allocation.

Depending on the concrete use case, there may be additional bufferization
requirements. If the contents of a buffer are expensive to compute, there could
be a tradeoff between *recomputation* and *compute once and copy*. On the
contrary, it may not even be possible to allocate new buffers at runtime on some
architectures.

## Destination-Passing Style

Bufferization is an algorithmically complex problem. Given an op with a tensor
result, bufferization has to choose a memref buffer in which the result can be
stored. It is always safe to allocate a brand new buffer, but such a
bufferization strategy would be unacceptable for high-performance codegen. When
choosing an already existing buffer, we must be careful not to accidentally
overwrite data that is still needed later in the program.

To simplify this problem, One-Shot Bufferize was designed to take advantage of
*destination-passing style* (DPS). In MLIR, DPS op should implement the
[`DestinationStyleOpInterface`](https://github.com/llvm/llvm-project/blob/792d437b56adfb3416daf8105942d4899fb82763/mlir/include/mlir/Interfaces/DestinationStyleOpInterface.td).
DPS exists in itself independently of bufferization and is tied to SSA
semantics: many ops are "updating" a part of their input SSA variables. For
example the LLVM instruction
[`insertelement`](https://llvm.org/docs/LangRef.html#insertelement-instruction)
is inserting an element inside a vector. Since SSA values are immutable, the
operation returns a copy of the input vector with the element inserted.
Another example in MLIR is `linalg.generic` on tensors, which always has an
extra `outs` operand for each result, which provides the initial values to
update (for example when the operation is doing a reduction).

`outs` operands are referred to as "destinations" in the following (quotes are
important as this operand isn't modified in place but copied) and comes into
place in the context of bufferization as a possible "anchor" for the
bufferization algorithm. This allows the user to shape the input in a form that
guarantees close to optimal bufferization result when carefully choosing the
SSA value used as "destination".

For every tensor result, a DPS op has a corresponding tensor operand. If there
aren't any other conflicting uses of this tensor, the bufferization can alias
it with the op result and perform the operation "in-place" by reusing the buffer
allocated for this "destination" input.

As an example, consider the following op: `%r = tensor.insert %f into
%t[%idx] : tensor<5xf32>`

![tensor.insert example](/includes/img/bufferization_tensor_insert_dst.svg)

`%t` is the "destination" in this example. When choosing a buffer for the result
`%r`, denoted as `buffer(%r)`, One-Shot Bufferize considers only two options:

1.  `buffer(%r) = buffer(%t)`: store the result in the existing `buffer(%t)`.
    Note that this is not always possible. E.g., if the old contents of
    `buffer(%t)` are still needed. One-Shot Bufferize's main task is to detect
    such cases and fall back to the second option when necessary.
2.  `buffer(%r)` is a newly allocated buffer.

There may be other buffers in the same function that could potentially be used
for `buffer(%r)`, but those are not considered by One-Shot Bufferize to keep the
bufferization simple. One-Shot Bufferize could be extended to consider such
buffers in the future to achieve a better quality of bufferization.

Tensor ops that are not in destination-passing style always bufferized to a
memory allocation. E.g.:

```mlir
%0 = tensor.generate %sz {
^bb0(%i : index):
  %cst = arith.constant 0.0 : f32
  tensor.yield %cst : f32
} : tensor<?xf32>
```

The result of `tensor.generate` does not have a "destination" operand, so
bufferization allocates a new buffer. This could be avoided by instead using an
op such as `linalg.generic`, which can express the same computation with a
"destination" operand, as specified behind outputs (`outs`):

```mlir
#map = affine_map<(i) -> (i)>
%0 = linalg.generic {indexing_maps = [#map], iterator_types = ["parallel"]}
                    outs(%t : tensor<?xf32>) {
  ^bb0(%arg0 : f32):
    %cst = arith.constant 0.0 : f32
    linalg.yield %cst : f32
} -> tensor<?xf32>
```

At first glance, the above `linalg.generic` op may not seem very useful because
the output tensor `%t` is entirely overwritten. Why pass the tensor `%t` as an
operand in the first place? As an example, this can be useful for overwriting a
slice of a tensor:

```mlir
%t = tensor.extract_slice %s [%idx] [%sz] [1] : tensor<?xf32> to tensor<?xf32>
%0 = linalg.generic ... outs(%t) { ... } -> tensor<?xf32>
%1 = tensor.insert_slice %0 into %s [%idx] [%sz] [1]
    : tensor<?xf32> into tensor<?xf32>
```

The above example bufferizes to a `memref.subview`, followed by a
"`linalg.generic` on memrefs" that overwrites the memory of the subview, assuming
that the slice `%t` has no other user. The `tensor.insert_slice` then bufferizes
to a no-op (in the absence of RaW conflicts such as a subsequent read of `%s`).

RaW conflicts are detected with an analysis of SSA use-def chains (details
later). One-Shot Bufferize works best if there is a single SSA use-def chain,
where the result of a tensor op is the operand of the next tensor ops, e.g.:

```mlir
%0 = "my_dialect.some_op"(%t) : (tensor<?xf32>) -> (tensor<?xf32>)
%1 = "my_dialect.another_op"(%0) : (tensor<?xf32>) -> (tensor<?xf32>)
%2 = "my_dialect.yet_another_op"(%1) : (tensor<?xf32>) -> (tensor<?xf32>)
```

Buffer copies are likely inserted if the SSA use-def chain splits at some point,
e.g.:

```mlir
%0 = "my_dialect.some_op"(%t) : (tensor<?xf32>) -> (tensor<?xf32>)
%1 = "my_dialect.another_op"(%0) : (tensor<?xf32>) -> (tensor<?xf32>)

// "yet_another_op" likely needs to read the data of %0, so "another_op" cannot
// in-place write to buffer(%0).
%2 = "my_dialect.yet_another_op"(%0) : (tensor<?xf32>) -> (tensor<?xf32>)
```

## Tensor / MemRef Boundary

The bufferization dialect provides a few helper ops to connect tensor IR (that
should be bufferized) with existing buffers (that may be allocated/provided by
a different runtime/library/etc.).

`bufferization.to_memref %t` returns the future buffer of a tensor SSA value.
`bufferization.to_tensor %m` returns a tensor SSA value for a given MemRef
buffer. `bufferization.materialize_in_destination` indicates that a tensor value
should materialize in a certain buffer.

Consider the following example, where a TOSA matmul result should materialize in
an existing buffer `%C`:

```mlir
// Batched TOSA matrix multiplication. %A and %B are the
// inputs, %C is the output.
func.func @test_matmul(%A: memref<1x17x19xf32>,
                       %B: memref<1x19x29xf32>,
                       %C: memref<1x17x29xf32>) {

  %A_tensor = bufferization.to_tensor %A restrict : memref<1x17x19xf32> to tensor<1x17x19xf32>
  %B_tensor = bufferization.to_tensor %B restrict : memref<1x19x29xf32> to tensor<1x19x29xf32>

  %0 = tosa.matmul %A_tensor, %B_tensor
      : (tensor<1x17x19xf32>, tensor<1x19x29xf32>) ->
         tensor<1x17x29xf32>

  bufferization.materialize_in_destination
    %0 in restrict writable %C
      : (tensor<1x17x29xf32>, memref<1x17x29xf32>) -> ()

  return
}
```

Note that all bufferization ops in this example have the `restrict` unit
attribute set. This attribute is similar to the C restrict keyword and indicates
that there is no other `to_tensor` or `materialize_in_destination` op with
the same or an aliasing MemRef operand. Only such
`to_tensor`/`materialize_in_destination` ops are supported. The `restrict`
attribute gives strong aliasing guarantees to the bufferization analysis and
allows us to look only at the tensor IR in a program. (Ops that do not operate
on tensors are ignored by the One-Shot Bufferize.)

Also note that `tosa.matmul` cannot be bufferized as is: there is no
`BufferizableOpInterface` implementation for that op. However, the op can be
lowered to a combination of `tensor.empty` and `linalg.matmul`, which can be
bufferized.

## Using One-Shot Bufferize

MLIR provides a pass
[`-one-shot-bufferize`](https://mlir.llvm.org/docs/Passes/#-one-shot-bufferize-one-shot-bufferize)
that performs an analysis and bufferizes all ops with tensor semantics that
implement `BufferizableOpInterface`. For modularity reasons, these op interface
implementations are typically external models that live in a dialect's
"Transforms" build unit. (External models are a mechanism for implementing an op
interface in a different build unit.) It is the user's responsibility to ensure
that all needed external models are registered before running One-Shot
Bufferize.

By default, One-Shot Bufferize fails when it encounters an op with tensor
semantics (i.e., tensor result or tensor operand) that is not bufferizable
(i.e., does not implement `BufferizableOpInterface`). This can be avoided with
`allow-unknown-ops`. In that case, One-Shot Bufferize inserts
`to_memref`/`to_tensor` ops around the bufferization boundary.

One-Shot Bufferize can be configured to bufferize only ops from a set of
dialects with `dialect-filter`.

One-Shot Bufferize can also be called programmatically with
[`bufferization::runOneShotBufferize`](https://github.com/llvm/llvm-project/blob/ae2764e835a26bad9774803eca0a6530df2a3e2d/mlir/include/mlir/Dialect/Bufferization/Transforms/OneShotAnalysis.h#L167).
Alternatively,
[`bufferization::bufferizeOp`](https://github.com/llvm/llvm-project/blob/ae2764e835a26bad9774803eca0a6530df2a3e2d/mlir/include/mlir/Dialect/Bufferization/Transforms/Bufferize.h#L78)
skips the analysis and inserts a copy on every buffer write.

By default, function boundaries are not bufferized. This is because there are
currently limitations around function graph bufferization: recursive
calls are not supported. As long as there are no recursive calls, function
boundary bufferization can be enabled with `bufferize-function-boundaries`. Each
tensor function argument and tensor function result is then turned into a
memref. The layout map of the memref type can be controlled with
`function-boundary-type-conversion`.

## Memory Layouts

One-Shot Bufferize bufferizes ops from top to bottom. This works well when all
ops are bufferizable. However, when encountering a non-bufferizable tensor with
`allow-unknown-ops`, One-Shot Bufferize must insert `to_memref` ops at the
bufferization boundary and decide on a memref type. By default, One-Shot
Bufferize choose the most dynamic memref type wrt. layout maps. E.g.:

```mlir
%0 = "my_dialect.unbufferizable_op(%t) : (tensor<?x?xf32>) -> (tensor<?x?xf32>)
%1 = tensor.extract %0[%idx1, %idx2] : tensor<?xf32>
```

When bufferizing the above IR, One-Shot Bufferize inserts a `to_memref` ops with
dynamic offset and strides:

```mlir
%0 = "my_dialect.unbufferizable_op(%t) : (tensor<?x?xf32>) -> (tensor<?x?xf32>)
%0_m = bufferization.to_memref %0 : memref<?x?xf32, strided<[?, ?], offset: ?>>
%1 = memref.load %0_m[%idx1, %idx2] : memref<?x?xf32, strided<[?, ?], offset: ?>>
```

All users of `%0` have fully dynamic layout maps. This ensures that the
bufferized IR composes well with future bufferizations of `unbufferizable_op`
(maybe bufferized by another pass), regardless of the exact memref type of the
future bufferization. If the op turns out to be bufferized to an op with a
simpler memref type (e.g., identity layout map), we expect that canonicalization
patterns would clean up unnecessarily dynamic layout maps. (Some of these
canonicalization patterns may not be implemented yet.)

One-Shot Bufferize tries to infer the most precise memref type when bufferizing
an op. If the entire IR is bufferizable, we do not have to resort to
conservatively use fully dynamic layout maps. In that case, we also do not have
to rely on canonicalization patterns to clean up the bufferized IR.

Note: There are some bufferizable ops for which a percise layout map cannot be
inferred. E.g., a `tensor.cast` from a `tensor<*xf32>` to a `tensor<?x?xf32>`
must be bufferized to a `memref.cast` with a memref type that has a fully
dynamic layout map.

One-Shot Bufferize has an option `unknown-type-conversion` to control the
generation of layout maps when no precise layout can be inferred:

*   `fully-dynamic-layout-map` uses fully dynamic layout maps and is the default
    behavior. This composes well when IR is partially bufferized.
*   `identity-layout-map` uses static identity layout maps. This option can be
    useful for legacy code that cannot handle memref types with layout maps.
    Note that this setting can lead to additional buffer copies when folding a
    `to_tensor`/`to_memref` pair with memref types that are not cast-compatible.

Note: The `unknown-type-conversion` option does not affect layout maps of
function signatures. There is a separate `function-signature-type-conversion`
option that controls layout maps of function parameters and function results.

## Extending One-Shot Bufferize

Custom ops can be bufferized if they implement `BufferizableOpInterface`. Users
must at least implement the following interface methods.

*   `bufferizesToMemoryRead`: Return `true` if the buffer of the given tensor
    OpOperand is read.
*   `bufferizesToMemoryWrite`: Return `true` if the buffer of the given tensor
    OpOperand is written (if bufferizing in-place).
*   `getAliasingOpResult`: Return the OpResults that may share the same buffer
    as the given OpOperand. This interface method describes to
    OpOperand-to-OpResult mapping wrt. destination-passing style.
*   `bufferRelation`: Return `BufferRelation::Equivalent` if the given OpResult
    is the exact same memref as the aliasing OpOperand after bufferization (in
    case of in-place bufferization). Otherwise, (e.g., they overlap but are not
    necessarily the exact same memrefs), `BufferRelation::Unknown` should be
    returned. Additional buffer relations will be added in the future, but
    `BufferRelation::Unknown` is always safe.
*   `bufferize`: Rewrite the op with the given rewriter. Ops should be replaced
    with `bufferization::replaceOpWithBufferizedValues`.

To get a better intuition of the interface methods, we invite users to take a
look at existing implementations in MLIR, e.g., the implementation of
`tensor.insert` or `tensor.extract`.

Interface implementations of DPS ops (that implement
`DestinationStyleOpInterface`) can derive from
`DstBufferizableOpInterfaceExternalModel`, which provides all necessary
method implementations except for `bufferize`.

## Debugging Buffer Copies

To get a better understanding of why One-Shot Bufferize introduced a buffer
copy, users can run the pass with `test-analysis-only print-conflicts`. Every
tensor op is then annotated with an attribute that has a boolean value for each
tensor OpOperand. `true` means that the OpOperand bufferizes in-place. `false`
means that the OpOperand bufferizes out-of-place and a buffer copy will be
inserted.

There are two reasons why a buffer copy may be inserted.

1.  Due to a RaW conflict, it is not safe to bufferize in-place. I.e., the
    overwritten data is still needed.
2.  The buffer is not writable. E.g., `memref.global` buffers that are the
    result of `arith.constant` ops are never modified.

In the first case, `print-conflicts` illustrates the conflict in the form of a
("read", "conflicting write", "last write") tuple.

A RaW conflict consists of three parts, in the following order according to
op dominance:

1. **Definition:** A tensor `%t` is defined.
2. **Conflicting Write:** An operation writes to `buffer(%t)`.
3. **Read:** An operation reads `%t`.

When such a RaW conflict is detected during the analysis phase, One-Shot
Bufferize will insert a buffer copy for the conflicting write.

**Example**

```mlir
// RUN: mlir-opt %s -one-shot-bufferize="bufferize-function-boundaries test-analysis-only print-conflicts"
func.func @test(%arg0: f32, %arg1: f32, %arg2: index, %arg3: index) -> (f32, tensor<3xf32>) {
  // Create a new tensor with [%arg0, %arg0, %arg0].
  %0 = tensor.from_elements %arg0, %arg0, %arg0 : tensor<3xf32>

  // Insert something into the new tensor.
  %1 = tensor.insert %arg1 into %0[%arg2] : tensor<3xf32>

  // Read from the old tensor.
  %r = tensor.extract %0[%arg3] : tensor<3xf32>

  // Return the extracted value and the result of the insertion.
  func.return %r, %1 : f32, tensor<3xf32>
}
```

The output IR is as follows:

```mlir
func.func @test(%arg0: f32, %arg1: f32, %arg2: index, %arg3: index) -> (f32, tensor<3xf32>) {
  %from_elements = tensor.from_elements %arg0, %arg0, %arg0 {"C_0[DEF: result 0]"} : tensor<3xf32>
  %inserted = tensor.insert %arg1 into %from_elements[%arg2] {"C_0[CONFL-WRITE: 1]", __inplace_operands_attr__ = ["none", "false", "none"]} : tensor<3xf32>
  %extracted = tensor.extract %from_elements[%arg3] {"C_0[READ: 0]", __inplace_operands_attr__ = ["true", "none"]} : tensor<3xf32>
  return {__inplace_operands_attr__ = ["none", "true"]} %extracted, %inserted : f32, tensor<3xf32>
}
```

Note that the IR was not bufferized. It was merely annotated with the results
of the bufferization analysis. Every operation with tensor semantics has a
`__inplace_operands_attr__` attribute with one value per operand. If an operand
is not a tensor, the respective value is `none`. Otherwise, if the operand was
decided to be bufferized in-place, the value is `true`. A value of `false`
indicates a buffer copy. In the above example, a buffer copy would be inserted
for `tensor.insert`, so that it does not overwrite `buffer(%from_elements)`,
which is still needed for `tensor.extract`.

For each RaW (there is only one in the example), three `C_i` attributes were
added:

* `C_0[DEF: result 0]`: A tensor is defined: 0-th result of
  `tensor.from_elements`.
* `C_0[CONFL-WRITE: 1]`: An operation (if bufferized in-place) would write into
  the future buffer of the defined tensor: 1-st operand of `tensor.insert`.
* `C_0[READ: 0]`: An operation reads the tensor definition: 0-th operand of
  `tensor.extract`.

The fully bufferized IR (with the inserted buffer copy) is as follows:

```mlir
func.func @test(%arg0: f32, %arg1: f32, %arg2: index, %arg3: index) -> (f32, memref<3xf32>) {
  %c2 = arith.constant 2 : index
  %c1 = arith.constant 1 : index
  %c0 = arith.constant 0 : index
  %alloc = memref.alloc() {alignment = 64 : i64} : memref<3xf32>
  memref.store %arg0, %alloc[%c0] : memref<3xf32>
  memref.store %arg0, %alloc[%c1] : memref<3xf32>
  memref.store %arg0, %alloc[%c2] : memref<3xf32>
  %alloc_0 = memref.alloc() {alignment = 64 : i64} : memref<3xf32>
  memref.copy %alloc, %alloc_0 : memref<3xf32> to memref<3xf32>
  memref.store %arg1, %alloc_0[%arg2] : memref<3xf32>
  %0 = memref.load %alloc[%arg3] : memref<3xf32>
  return %0, %alloc_0 : f32, memref<3xf32>
}
```

To get a better understanding of the SSA Use-Def Chain Analysis and the RaW
conflict detection algorithm, interested users may want to refer to:

* [Original design document](https://discourse.llvm.org/uploads/short-url/5kckJ3DftYwQokG252teFgw3sYa.pdf)
* [ODM talk](https://youtu.be/TXEo59CYS9A), ([slides](https://mlir.llvm.org/OpenMeetings/2022-01-13-One-Shot-Bufferization.pdf)).
* [LLVM Dev Meeting 2023 tutorial slides](https://m-sp.org/downloads/llvm_dev_2023.pdf)