3 Shape inference as discussed here is considered a specific instance of type
4 inference for [ShapedType][ShapedType]. Type constraints are along (at least)
5 three axis: 1) elemental type, 2) rank (including static or dynamic), 3)
6 dimensions. While some operations have no compile time fixed shape (e.g., output
7 shape is dictated by data) we could still have some knowledge of
8 constraints/bounds in the system for that operation (e.g., the output of a
9 `tf.where` is at most the size of the input data). That is, there are additional
10 valuable constraints that could be captured even without full knowledge of the
13 Type inference is currently modelled executionally for operation creation using the
14 [`InferTypeOpInterface`][InferTypeOpInterface], while
15 `InferShapedTypeOpInterface` is used to implement the shape and element type
16 inference. The return type can often be deduced from the deduced return shape
17 and elemental type (queryable from `InferShapedTypeOpInterface`) and so type
18 inference for tensor types can be implemented with `InferShapedTypeOpInterface`.
24 The C++ interfaces are the base mechanism whereby shape inference is queried and
25 executed, but not the intended way to specify shape constraints in general.
27 Initially the shape inference will be declaratively specified using:
29 * Constraints on the operands of an operation directly. For example
30 constraining the input type to be tensor/vector elements or that the
31 elemental type be of a specific type (e.g., output of computing the size
32 of a value is of elemental type `i1`) or class (e.g., float-like).
33 * Constraints across operands and results of an operation.
35 - For example, specifying equality constraints on type/constituents of a
36 type (shape and elemental type) between operands and results (e.g., the
37 output type of an add is the same as those of the input operands).
39 NOTE: The C++ shape functions are an intermediate step until the shape dialect
40 is more full-fledged, at which point the C++ functions should become the
45 Shape inference is currently tested alongside type inference by
46 `TestReturnTypeDriver` in the test dialect. This driver performs two checks:
48 1. Verification that the return types specified matches the inferred types. This
49 explicit check will be removed and made part of Op verification instead.
50 2. Test the creation of Ops without specifying the return type explicitly in
51 function `testCreateFunctions` by creating new binary Ops (Op classes
52 specified in `TestReturnTypeDriver`) using 1) all operands to
53 `testCreateFunctions` as both operands, and 2) using combinations of input
54 operands of the function.
58 This section details the shape type inference dialect (`shape`). The initial
59 focus will be on shape functions that describe shape functions could be used in
60 runtime and compiler (for constructions of ops/refinement of shapes, reification
61 of dynamic allocations for dialect including TF, TFLite, XLA & tensor compute
62 dialect under discussion).
64 This will focus on the shape functions (e.g., determine the rank and dimensions
65 of the output shape). As shown in the shaped container type, shape will be one
66 of 3 components, the others being elemental type and attribute (which is
67 currently left open with the intention of supporting extensions such as layouts
68 or bounded shapes at a later point). This allows for decoupling of these:
70 * Not all the information is needed for all analysis;
71 * Not all shape functions need to provide all the information (e.g., one could
72 define a base class function that only populates element type but composes
74 * It allows reusing the constraints between, say, Tensor and Memref
75 representation of an operation;
77 An argument could be made that these are metadata function instead of shape
78 functions, with some considering shape and elemental types different and some considering them both as
79 part of shape. But `shape function` is IMHO descriptive and metadata can span
80 too large a range of potential uses/values.
84 The requirements for the shape inference functions are determined by the
85 requirements of shape inference, but we believe the requirements below still
86 allow freedom to consider different shape inference approaches and so we do not
87 impose a particular shape inference approach here.
89 #### Shape inference functions
91 * **Expressiveness** shape functions need to support programs where tensors
92 have shapes that are not known statically (for example, `tensor<16x?xf32>`
94 * **Shape error detection** Many operations will have constraints on their
95 operands. If the constraints are not satisfied or cannot be determined if
96 satisfied statically, then a runtime check/assertion could be generated.
98 * This also aligns with the requirement that the shape function description
99 should be usable by both the compiler and runtime.
100 * Shape error functions should be easy to understand, at least what
101 constraint of the operation is violated. This also requires that shape
102 function error messages should be configurable by the author of the
103 shape function (e.g., the author would be able to give the semantic
104 constraint invalidated rather the low-level check that failed).
105 * The static analysis may be used to eliminate run-time checks that are
107 * Ideally all would eventually (see section
108 [Inlining shape checking](#inline)) be elided.
109 * Only reporting errors which are guaranteed to occur at runtime. If an error is only
110 possible (rather than guaranteed) then we use a runtime assertion to fail and produce an error
111 message with the invariant violated.
113 * Shape functions usable by compiler and runtime.
115 * This does not mean the exact same C++ function, but rather the
116 description should be consumable by either.
117 * Shape function description should not be constrained by either runtime
118 or compiler's type system to handle types only used for analysis. That
119 is, these two type systems differ and both should be supported, but the
120 intersection of the two should not be required. As a particular example,
121 if a compiler only wants to differentiate exact shapes vs dynamic
122 shapes, then it need not consider a more generic shape lattice even
123 though the shape description supports it.
125 * Declarative (e.g., analyzable at compile time, possible to generate
126 different versions for different use cases)
128 * This may not strictly be a requirement, but a way to handle the former:
129 a declarative specification could be reused by both while avoiding a
130 need to map to or from a 3rd representation given these two systems
131 have/and will have different types.
133 * Shape inference functions are expressible at runtime
135 * User can define a shape function for a new operation dynamically at runtime,
136 this allows for vendors to describe an operation and shape function
139 This requirement is on the wishlist.
141 * Doesn't require graph-wide shape information (e.g., only require local
144 * Shape functions should be cheap to invoke on each kernel launch.
145 * Shape function can be dictated by arguments (operands, attributes and regions)
146 only (e.g., same operands as the corresponding operation could be
147 constructed & invoked with).
148 * Shape information that needs higher-level/graph information should use
149 richer types (e.g., `TensorList<F32>`);
150 * The function should be invocable before/while constructing an op (e.g.,
151 can't rely on the op being constructed).
153 * Shape functions should be pure functions.
155 * Should support functions whose type is only known dynamically (e.g.,
158 * Without needing to invoke the op (e.g., reading a file once for
159 determining the shape & then post to be able to actually consume the
162 * The shape function operation dialect should be interoperable with non-shape function dialect operations.
164 * There may be a common set of operations that satisfy most uses (e.g., merge,
165 equal_type, arithmetic expressions, slice, concat, pattern matching on
166 attributes such as padding etc.) that will be discovered and could cover
167 a large percentage of the use cases. Among these there will be some
168 which carry extra semantic info that could be used for symbolic
169 constraints (e.g., checking equality of two dimensions resulting in
170 setting an equality constraint) and higher-order interpretation for
173 It is therefore beneficial (but not required) to reuse operations,
174 especially as for statically known shapes, arbitrary arithmetic
175 computations could still be performed. This means that the computations
176 performed statically may or may not be supported by an arbitrary solver,
177 but would still be allowed.
179 * The shape function should be expandable such that symbolic equality and
180 upper bound constraints (say) could be represented and may be propagated by
183 * E.g., the shape functions may contain more information that is only
184 useful when used from shape inference;
186 * Shape functions are allowed to fail and report an error. The error reporting
187 should report the location of the operation that failed with, where
188 possible, a user actionable error message.
190 * These failures could become inlined and become runtime failures with
191 runtime values and error messages.
192 * Reporting errors should be optional. E.g., The same function
193 may be used as to query validity without reporting an error.
197 1. The shape dialect is an IR representations and not a programming language;
198 * While the functions should be readable, it doesn't carry the
199 conveniences of a programming language. Deciding how people write these
200 things, e.g. a mini dsl, a C++ API that generates them, extracting them
201 programmatically from `SetShapeFn` calls, etc., is still TBD.
202 1. Describe the shape inference approach that will use the shape functions;
203 * The goal is that the shape functions and the constraints one could
204 obtain from them are general enough that they would be useful for
205 various analysis. But whether we follow very simple (e.g., only fully
206 static information is used for shape output, unranked for everything
207 else) to very advance (e.g., expression trees of symbolic constants) can
208 be evaluated independently of this proposal and with concrete benefit
210 1. Describe the approach whereby error messages will be generated;
211 * While the shape functions will be able to emit errors optionally, it
212 will be possible to dictate when they emit an error. This enables
213 deciding whether or which error to emit: there have been proposals in
214 the literature that the iteration order for shape inference affect the
215 quality of the error message produced, and the shape functions do not
217 1. Flow sensitive shape functions;
218 * To enable scalable/cheap shape inference, the shape functions do not
219 intend to provide flow sensitive information. This facility could
220 potentially be built as part of some higher order analysis that reuse
221 the shape functions/constraints due to the shape functions.
222 1. All static functions are usable for dynamic/unknown shapes;
223 * More involved computations can be performed with statically known shapes
224 than what can be sensibly analyzed with unknown/symbolic variables.
228 #### Inline shape inference checks {#inline}
230 Shape functions should be lowerable to runtime checks for validity. E.g. verify
231 as much as possible statically, but enable generating instructions to compute the
232 shape dynamically and or falling back to runtime checks for attributes not
233 verifiable at compile time. These checks inserted should ideally only check that
234 which could not have been verified statically.
236 These inlined calls could interfere with optimization patterns/passes (e.g.,
237 shape inference should not insert constructs that interfere with optimization
238 patterns) and so could be delayed until later (with another round of
239 optimizations, constant folding, CSE, etc., that should remove redundant runtime
242 ### Possibly Asked Questions
244 #### What about ODS specifications of operations?
246 In ODS we have been recording the constraints for the operands & attributes of
247 an operation. Where these are sufficient to constrain the output shape (e.g.,
248 `SameOperandAndResultType` or broadcastable) we should generate the shape
249 function from those. Where not, an explicit shape function should be specified
250 (spelling TBD but currently considering using the MLIR textual form as
251 serialization approach).
253 #### Why not extract the shape function from reference implementation?
255 This could be done in future! The extracted shape function would use the shape
256 inference dialect, so we are starting there. Especially for operations described in a
257 structured way, one could autogenerate the shape function.
259 #### How/in what language will the shape functions be authored?
261 TBD. open to many approaches and suggestions, starting on the IR produced by
262 whatever language is the priority of this proposal.
264 #### What shape inference approach is being suggested here?
266 None. There are multiple different shape inference approaches that we could
267 layer on top of these. From the most basic (always return unranked), to more
268 useful (return fixed shape for constant inputs/arguments) to the more advanced
269 (create logical conjunctions of algebraic statements between symbolic named
274 1. Should shape functions that produce dynamic outputs given all statically
275 shaped inputs be marked specially? E.g., read from file.
277 TODO: Add examples here.
279 ## WIP/Future considerations
281 Shape functions are determined by attributes and could be arbitrarily
282 complicated with a wide-range of specification possibilities. Equality
283 relationships are common (e.g., the elemental type of the output matches the
284 primitive type of the inputs, both inputs have exactly the same type [primitive
285 type and shape]) and so these should be easy to specify. Algebraic relationships
286 would also be common (e.g., a concat of `[n,m]` and `[n,m]` matrix along axis 0
287 is `[n+n, m]` matrix), while some ops only have defined shapes under certain
288 cases (e.g., matrix multiplication of `[a,b]` and `[c,d]` is only defined if `b
291 Instead of specifying an additional mechanism to specify a shape transfer
292 function, the reference implementation of the operation will be used to derive
293 the shape function. The reference implementation is general and can support the
294 arbitrary computations needed to specify output shapes.
296 [InferTypeOpInterface]: https://github.com/llvm/llvm-project/tree/main/mlir/include/mlir/Interfaces/InferTypeOpInterface.td
297 [ShapedType]: https://github.com/llvm/llvm-project/tree/main/mlir/include/mlir/IR/BuiltinTypes.h