Using the Numba Rewrite Pass for Fun and Optimization
Overview
This section introduces intermediate representation (IR) rewrites, and how they can be used to implement optimizations.
As discussed earlier in “Stage 5a: Rewrite typed IR”, rewriting the Numba IR allows us to perform optimizations that would be much more difficult to perform at the lower LLVM level. Similar to the Numba type and lowering subsystems, the rewrite subsystem is user extensible. This extensibility affords Numba the possibility of supporting a wide variety of domain-specific optimizations (DSO’s).
The remaining subsections detail the mechanics of implementing a rewrite, registering a rewrite with the rewrite registry, and provide examples of adding new rewrites, as well as internals of the array expression optimization pass. We conclude by reviewing some use cases exposed in the examples, as well as reviewing any points where developers should take care.
Rewriting Passes
Rewriting passes have a simple match()
and
apply()
interface. The division between matching and
rewriting follows how one would define a term rewrite in a declarative
domain-specific languages (DSL’s). In such DSL’s, one may write a
rewrite as follows:
<match> => <replacement>
The <match>
and <replacement>
symbols represent IR term
expressions, where the left-hand side presents a pattern to match, and
the right-hand side an IR term constructor to build upon matching.
Whenever the rewrite matches an IR pattern, any free variables in the
left-hand side are bound within a custom environment. When applied,
the rewrite uses the pattern matching environment to bind any free
variables in the right-hand side.
As Python is not commonly used in a declarative capacity, Numba uses object state to handle the transfer of information between the matching and application steps.
The Rewrite
Base Class
- class Rewrite
The
Rewrite
class simply defines an abstract base class for Numba rewrites. Developers should define rewrites as subclasses of this base type, overloading thematch()
andapply()
methods.- pipeline
The pipeline attribute contains the
numba.compiler.Pipeline
instance that is currently compiling the function under consideration for rewriting.
- __init__(self, pipeline, *args, **kws)
The base constructor for rewrites simply stashes its arguments into attributes of the same name. Unless being used in debugging or testing, rewrites should only be constructed by the
RewriteRegistry
in theRewriteRegistry.apply()
method, and the construction interface should remain stable (though the pipeline will commonly contain just about everything there is to know).
- match(self, block, typemap, callmap)
The
match()
method takes four arguments other than self:func_ir: This is an instance of
numba.ir.FunctionIR
for the function being rewritten.block: This is an instance of
numba.ir.Block
. The matching method should iterate over the instructions contained in thenumba.ir.Block.body
member.typemap: This is a Python
dict
instance mapping from symbol names in the IR, represented as strings, to Numba types.callmap: This is another
dict
instance mapping from calls, represented asnumba.ir.Expr
instances, to their corresponding call site type signatures, represented as anumba.typing.templates.Signature
instance.
The
match()
method should return abool
result. ATrue
result should indicate that one or more matches were found, and theapply()
method will return a new replacementnumba.ir.Block
instance. AFalse
result should indicate that no matches were found, and subsequent calls toapply()
will return undefined or invalid results.
- apply(self)
The
apply()
method should only be invoked following a successful call tomatch()
. This method takes no additional parameters other than self, and should return a replacementnumba.ir.Block
instance.As mentioned above, the behavior of calling
apply()
is undefined unlessmatch()
has already been called and returnedTrue
.
Subclassing Rewrite
Before going into the expectations for the overloaded methods any
Rewrite
subclass must have, let’s step back a minute to
review what is taking place here. By providing an extensible
compiler, Numba opens itself to user-defined code generators which may
be incomplete, or worse, incorrect. When a code generator goes awry,
it can cause abnormal program behavior or early termination.
User-defined rewrites add a new level of complexity because they must
not only generate correct code, but the code they generate should
ensure that the compiler does not get stuck in a match/apply loop.
Non-termination by the compiler will directly lead to non-termination
of user function calls.
There are several ways to help ensure that a rewrite terminates:
Typing: A rewrite should generally attempt to decompose composite types, and avoid composing new types. If the rewrite is matching a specific type, changing expression types to a lower-level type will ensure they will no long match after the rewrite is applied.
Special instructions: A rewrite may synthesize custom operators or use special functions in the target IR. This technique again generates code that is no longer within the domain of the original match, and the rewrite will terminate.
In the “Case study: Array Expressions” subsection, below, we’ll see how the array expression rewriter uses both of these techniques.
Overloading Rewrite.match()
Every rewrite developer should seek to have their implementation of
match()
return a False
value as quickly as
possible. Numba is a just-in-time compiler, and adding compilation
time ultimately adds to the user’s run time. When a rewrite returns
False
for a given block, the registry will no longer process that
block with that rewrite, and the compiler is that much closer to
proceeding to lowering.
This need for timeliness has to be balanced against collecting the necessary information to make a match for a rewrite. Rewrite developers should be comfortable adding dynamic attributes to their subclasses, and then having these new attributes guide construction of the replacement basic block.
Overloading Rewrite.apply()
The apply()
method should return a replacement
numba.ir.Block
instance to replace the basic block that
contained a match for the rewrite. As mentioned above, the IR built
by apply()
methods should preserve the semantics of the
user’s code, but also seek to avoid generating another match for the
same rewrite or set of rewrites.
The Rewrite Registry
When you want to include a rewrite in the rewrite pass, you should
register it with the rewrite registry. The numba.rewrites
module provides both the abstract base class and a class decorator for
hooking into the Numba rewrite subsystem. The following illustrates a
stub definition of a new rewrite:
from numba import rewrites
@rewrites.register_rewrite
class MyRewrite(rewrites.Rewrite):
def match(self, block, typemap, calltypes):
raise NotImplementedError("FIXME")
def apply(self):
raise NotImplementedError("FIXME")
Developers should note that using the class decorator as shown above will register a rewrite at import time. It is the developer’s responsibility to ensure their extensions are loaded before compilation starts.
Case study: Array Expressions
This subsection looks at the array expression rewriter in more depth.
The array expression rewriter, and most of its support functionality,
are found in the numba.npyufunc.array_exprs
module. The
rewriting pass itself is implemented in the RewriteArrayExprs
class. In addition to the rewriter, the
array_exprs
module includes a function for
lowering array expressions,
_lower_array_expr()
. The overall
optimization process is as follows:
RewriteArrayExprs.match()
: The rewrite pass looks for one or more array operations that form an array expression.RewriteArrayExprs.apply()
: Once an array expression is found, the rewriter replaces the individual array operations with a new kind of IR expression, thearrayexpr
.numba.npyufunc.array_exprs._lower_array_expr()
: During lowering, the code generator calls_lower_array_expr()
whenever it finds anarrayexpr
IR expression.
More details on each step of the optimization are given below.
The RewriteArrayExprs.match()
method
The array expression optimization pass starts by looking for array
operations, including calls to supported ufunc
's and
user-defined DUFunc
's. Numba IR follows the
conventions of a static single assignment (SSA) language, meaning that
the search for array operators begins with looking for assignment
instructions.
When the rewriting pass calls the RewriteArrayExprs.match()
method, it first checks to see if it can trivially reject the basic
block. If the method determines the block to be a candidate for
matching, it sets up the following state variables in the rewrite
object:
crnt_block: The current basic block being matched.
typemap: The typemap for the function being matched.
matches: A list of variable names that reference array expressions.
array_assigns: A map from assignment variable names to the actual assignment instructions that define the given variable.
const_assigns: A map from assignment variable names to the constant valued expression that defines the constant variable.
At this point, the match method iterates over the assignment instructions in the input basic block. For each assignment instruction, the matcher looks for one of two things:
Array operations: If the right-hand side of the assignment instruction is an expression, and the result of that expression is an array type, the matcher checks to see if the expression is either a known array operation, or a call to a universal function. If an array operator is found, the matcher stores the left-hand variable name and the whole instruction in the array_assigns member. Finally, the matcher tests to see if any operands of the array operation have also been identified as targets of other array operations. If one or more operands are also targets of array operations, then the matcher will also append the left-hand side variable name to the matches member.
Constants: Constants (even scalars) can be operands to array operations. Without worrying about the constant being apart of an array expression, the matcher stores constant names and values in the const_assigns member.
The end of the matching method simply checks for a non-empty matches
list, returning True
if there were one or more matches, and
False
when matches is empty.
The RewriteArrayExprs.apply()
method
When one or matching array expressions are found by
RewriteArrayExprs.match()
, the rewriting pass will call
RewriteArrayExprs.apply()
. The apply method works in two
passes. The first pass iterates over the matches found, and builds a
map from instructions in the old basic block to new instructions in
the new basic block. The second pass iterates over the instructions
in the old basic block, copying instructions that are not changed by
the rewrite, and replacing or deleting instructions that were
identified by the first pass.
The RewriteArrayExprs._handle_matches()
implements the first
pass of the code generation portion of the rewrite. For each match,
this method builds a special IR expression that contains an expression
tree for the array expression. To compute the leaves of the
expression tree, the _handle_matches()
method
iterates over the operands of the identified root operation. If the
operand is another array operation, it is translated into an
expression sub-tree. If the operand is a constant,
_handle_matches()
copies the constant value.
Otherwise, the operand is marked as being used by an array expression.
As the method builds array expression nodes, it builds a map from old
instructions to new instructions (replace_map), as well as sets of
variables that may have moved (used_vars), and variables that should
be removed altogether (dead_vars). These three data structures are
returned back to the calling RewriteArrayExprs.apply()
method.
The remaining part of the RewriteArrayExprs.apply()
method
iterates over the instructions in the old basic block. For each
instruction, this method either replaces, deletes, or duplicates that
instruction based on the results of
RewriteArrayExprs._handle_matches()
. The following list
describes how the optimization handles individual instructions:
When an instruction is an assignment,
apply()
checks to see if it is in the replacement instruction map. When an assignment instruction is found in the instruction map,apply()
must then check to see if the replacement instruction is also in the replacement map. The optimizer continues this check until it either arrives at aNone
value or an instruction that isn’t in the replacement map. Instructions that have a replacement that isNone
are deleted. Instructions that have a non-None
replacement are replaced. Assignment instructions not in the replacement map are appended to the new basic block with no changes made.When the instruction is a delete instruction, the rewrite checks to see if it deletes a variable that may still be used by a later array expression, or if it deletes a dead variable. Delete instructions for used variables are added to a map of deferred delete instructions that
apply()
uses to move them past any uses of that variable. The loop copies delete instructions for non-dead variables, and ignores delete instructions for dead variables (effectively removing them from the basic block).All other instructions are appended to the new basic block.
Finally, the apply()
method returns the new
basic block for lowering.
The _lower_array_expr()
function
If we left things at just the rewrite, then the lowering stage of the
compiler would fail, complaining it doesn’t know how to lower
arrayexpr
operations. We start by hooking a lowering function
into the target context whenever the RewriteArrayExprs
class
is instantiated by the compiler. This hook causes the lowering pass to
call _lower_array_expr()
whenever it
encounters an arrayexr
operator.
This function has two steps:
Synthesize a Python function that implements the array expression: This new Python function essentially behaves like a Numpy
ufunc
, returning the result of the expression on scalar values in the broadcasted array arguments. The lowering function accomplishes this by translating from the array expression tree into a Python AST.Compile the synthetic Python function into a kernel: At this point, the lowering function relies on existing code for lowering ufunc and DUFunc kernels, calling
numba.targets.numpyimpl.numpy_ufunc_kernel()
after defining how to lower calls to the synthetic function.
The end result is similar to loop lifting in Numba’s object mode.
Conclusions and Caveats
We have seen how to implement rewrites in Numba, starting with the interface, and ending with an actual optimization. The key points of this section are:
When writing a good plug-in, the matcher should try to get a go/no-go result as soon as possible.
The rewrite application portion can be more computationally expensive, but should still generate code that won’t cause infinite loops in the compiler.
We use object state to communicate any results of matching to the rewrite application pass.