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