Since the goal of the LLVM IR dialect is to reflect LLVM IR in MLIR, the
dialect and the conversion procedure must account for the differences betweeen
block arguments and LLVM IR PHI nodes. In particular, LLVM IR disallows PHI
nodes with different values coming from the same source. Therefore, the LLVM IR
dialect now disallows `cond_br` operations that have identical successors
accepting arguments, which would lead to invalid PHI nodes. The conversion
process resolves the potential PHI source ambiguity by injecting dummy blocks
if the same block is used more than once as a successor in an instruction.
These dummy blocks branch unconditionally to the original successors, pass them
the original operands (available in the dummy block because it is dominated by
the original block) and are used instead of them in the original terminator
operation.
PiperOrigin-RevId: 235682798
Addressing post-submit comments. The `getelementptr` operation now supports
non-constant indexes, similarly to LLVM, and this functionality is exercised by
the lowering to the dialect. Update the documentation accordingly.
List the values of integer comparison predicates, which currently correspond to
those of CmpIOp in MLIR. Ideally, we would use strings instead, but it
requires additional support for argument conversion in both the dialect
lowering pass and the LLVM translator.
PiperOrigin-RevId: 235678877
This CL adds a primitive to perform stripmining of a loop by a given factor and
sinking it under multiple target loops.
In turn this is used to implement imperfectly nested loop tiling (with interchange) by repeatedly calling the stripmineSink primitive.
The API returns the point loops and allows repeated invocations of tiling to achieve declarative, multi-level, imperfectly-nested tiling.
Note that this CL is only concerned with the mechanical aspects and does not worry about analysis and legality.
The API is demonstrated in an example which creates an EDSC block, emits the corresponding MLIR and applies imperfectly-nested tiling:
```cpp
auto block = edsc::block({
For(ArrayRef<edsc::Expr>{i, j}, {zero, zero}, {M, N}, {one, one}, {
For(k1, zero, O, one, {
C({i, j, k1}) = A({i, j, k1}) + B({i, j, k1})
}),
For(k2, zero, O, one, {
C({i, j, k2}) = A({i, j, k2}) + B({i, j, k2})
}),
}),
});
// clang-format on
emitter.emitStmts(block.getBody());
auto l_i = emitter.getAffineForOp(i), l_j = emitter.getAffineForOp(j),
l_k1 = emitter.getAffineForOp(k1), l_k2 = emitter.getAffineForOp(k2);
auto indicesL1 = mlir::tile({l_i, l_j}, {512, 1024}, {l_k1, l_k2});
auto l_ii1 = indicesL1[0][0], l_jj1 = indicesL1[1][0];
mlir::tile({l_jj1, l_ii1}, {32, 16}, l_jj1);
```
The edsc::Expr for the induction variables (i, j, k_1, k_2) provide the programmatic hooks from which tiling can be applied declaratively.
PiperOrigin-RevId: 235548228
Leverage the recently introduced support for multiple argument groups and
multiple destination blocks in EDSC Expressions to implement conditional
branches in EDSC. Conditional branches have two successors and three argument
groups. The first group contains a single expression of i1 type that
corresponds to the condition of the branch. The two following groups contain
arguments of the two successors of the conditional branch instruction, in the
same order as the successors. Expose this instruction to the C API and Python
bindings.
PiperOrigin-RevId: 235542768
The new implementation of blocks was designed to support blocks with arguments.
More specifically, StmtBlock can be constructed with a list of Bindables that
will be bound to block aguments upon construction. Leverage this functionality
to implement branch instructions with arguments.
This additionally requires the statement storage to have a list of successors,
similarly to core IR operations.
Becauase successor chains can form loops, we need a possibility to decouple
block declaration, after which it becomes usable by branch instructions, from
block body definition. This is achieved by creating an empty block and by
resetting its body with a new list of instructions. Note that assigning a
block from another block will not affect any instructions that may have
designated this block as their successor (this behavior is necessary to make
value-type semantics of EDSC types consistent). Combined, one can now write
generators like
EDSCContext context;
Type indexType = ...;
Bindable i(indexType), ii(indexType), zero(indexType), one(indexType);
StmtBlock loopBlock({i}, {});
loopBlock.set({ii = i + one,
Branch(loopBlock, {ii})});
MLIREmitter(&builder)
.bindConstant<ConstantIndexOp>(zero, 0)
.bindConstant<ConstantIndexOp>(one, 1)
.emitStmt(Branch(loopBlock, {zero}));
where the emitter will emit the statement and its successors, if present.
PiperOrigin-RevId: 235541892
This came up in post-submit review. Use LLVM's support for outputting APInt
values directly instead of obtaining a 64-bit integer value from APInt, which
will not work for wider integers.
PiperOrigin-RevId: 235531574
Previously we have `auto pos = std::string::find(...) != std::string::npos` as
if condition to control substring substitution. Instead of the position for the
found substring, `pos` will be a boolean value indicating found nor not. Then
used as the replace start position, we were always replacing starting from 0 or
1. If the replaced substring also has the pattern to be matched, we'll see
an infinite loop.
PiperOrigin-RevId: 235504681
Analysis - NFC
- refactor AffineExprFlattener (-> SimpleAffineExprFlattener) so that it
doesn't depend on FlatAffineConstraints, and so that FlatAffineConstraints
could be moved out of IR/; the simplification that the IR needs for
AffineExpr's doesn't depend on FlatAffineConstraints
- have AffineExprFlattener derive from SimpleAffineExprFlattener to use for
all Analysis/Transforms purposes; override addLocalFloorDivId in the derived
class
- turn addAffineForOpDomain into a method on FlatAffineConstraints
- turn AffineForOp::getAsValueMap into an AffineValueMap ctor
PiperOrigin-RevId: 235283610
The only reason in starting with a fixedpoint add is that it is the absolute simplest variant and illustrates the level of abstraction I'm aiming for.
The overall flow would be:
1. Determine quantization parameters (out of scope of this cl).
2. Source dialect rules to lower supported math ops to the quantization dialect (out of scope of this cl).
3. Quantization passes: [-quant-convert-const, -quant-lower-uniform-real-math, -quant-lower-unsupported-to-float] (the last one not implemented yet)
4. Target specific lowering of the integral arithmetic ops (roughly at the level of gemmlowp) to more fundamental operations (i.e. calls to gemmlowp, simd instructions, DSP instructions, etc).
How I'm doing this should facilitate implementation of just about any kind of backend except TFLite, which has a very course, adhoc surface area for its quantized kernels. Options there include (I'm not taking an opinion on this - just trying to provide options):
a) Not using any of this: just match q/dbarrier + tf math ops to the supported TFLite quantized op set.
b) Implement the more fundamental integer math ops on TFLite and convert to those instead of the current op set.
Note that I've hand-waved over the process of choosing appropriate quantization parameters. Getting to that next. As you can see, different implementations will likely have different magic combinations of specific math support, and we will need the target system that has been discussed for some of the esoteric cases (i.e. many DSPs only support POT fixedpoint).
Two unrelated changes to the overall goal of this CL and can be broken out of desired:
- Adding optional attribute support to TabelGen
- Allowing TableGen native rewrite hooks to return nullptr, signalling that no rewrite has been done.
PiperOrigin-RevId: 235267229
Add a documentation page on the key points of the conversion to LLVM IR. This
focuses on the aspects of conversion that are relevant for integration of the
LLVM IR dialect (and produced LLVM IR that is mostly a one-to-one translation)
into other projects. In particular, it describes the type conversion rules and
the memref model supporting dynamic sizes.
PiperOrigin-RevId: 235190772
The LLVM IR pass was bootstrapped without user documentation, following LLVM's
language reference and existing conversions between MLIR standard operations
and LLVM IR instructions. Provide concise documentation of the LLVM IR dialect
operations. This documentation does not describe the semantics of the
operations, which should match that of LLVM IR, but highlights the structural
differences in operation definitions, in particular using attributes instead of
constant-only values. It also describes pseudo-operations that exist only to
make the LLVM IR dialect self-contained within MLIR.
While it could have been possible to generate operation description from
TableGen, this opts for a more concise format where groups of related
operations are described together.
PiperOrigin-RevId: 235149136
Add support for lowering DivF and RemF to LLVM::FDiv and LLMV::FRem
respectively. The lowering is a trivial one-to-one transformation.
The corresponding operations already existed in the LLVM IR dialect and can be
lowered to the LLVM IR proper. Add the necessary tests for scalar and vector
forms.
PiperOrigin-RevId: 234984608
This change introduces three new operators in EDSC: Div (also exposed
via Expr.__div__ aka /) -- floating-point division, FloorDiv and CeilDiv
for flooring/ceiling index division.
The lowering to LLVM will be implemented in b/124872679.
PiperOrigin-RevId: 234963217
Introduce support for binding MLIR functions as constant expressions. Standard
constant operation supports functions as possible constant values.
Provide C APIs to look up existing named functions in an MLIR module and expose
them to the Python bindings. Provide Python bindings to declare a function in
an MLIR module without defining it and to add a definition given a function
declaration. These declarations are useful when attempting to link MLIR
modules with, e.g., the standard library.
Introduce EDSC support for direct and indirect calls to other MLIR functions.
Internally, an indirect call is always emitted to leverage existing support for
delayed construction of MLIR Values using EDSC Exprs. If the expression is
bound to a constant function (looked up or declared beforehand), MLIR constant
folding will be able to replace an indirect call by a direct call. Currently,
only zero- and one-result functions are supported since we don't have support
for multi-valued expressions in EDSC.
Expose function calling interface to Python bindings on expressions by defining
a `__call__` function accepting a variable number of arguments.
PiperOrigin-RevId: 234959444
* Introduce a OpTrait class in C++ to wrap the TableGen definition;
* Introduce PredOpTrait and rename previous usage of OpTrait to NativeOpTrait;
* PredOpTrait allows specifying a trait of the operation by way of predicate on the operation. This will be used in future to create reusable set of trait building blocks in the definition of operations. E.g., indicating whether to operands have the same type and allowing locally documenting op requirements by trait composition.
- Some of these building blocks could later evolve into known fixed set as LLVMs backends do, but that can be considered with more data.
* Use the modelling to address one verify TODO in a very local manner.
This subsumes the current custom verify specification which will be removed in a separate mechanical CL.
PiperOrigin-RevId: 234827169
A recent change made ConstantOp::build accept a NumericAttr or assert that a
generic Attribute is in fact a NumericAttr. The rationale behind the change
was that NumericAttrs have a type that can be used as the result type of the
constant operation. FunctionAttr also has a type, and it is valid to construct
function-typed constants as exercised by the parser.mlir test. Relax
ConstantOp::build back to take a generic Attribute. In the overload that only
takes an attribute, assert that the Attribute is either a NumericAttr or a
FunctionAttr, because it is necessary to extract the type. In the overload
that takes both type type and the attribute, delegate the attribute type
checking to ConstantOp::verify to prevent non-Builder-based Op construction
mechanisms from creating invalid IR.
PiperOrigin-RevId: 234798569
The recent rework of MLIREmitter switched to using the generic call to
`builder.createOperation` from OperationState instead of individual customized
calls to `builder.create<>`. As a result, regular non-composed affine apply
operations where emitted. Introduce a special case in Expr::build to always
create composed affine maps instead, as it used to be the case before the
rework.
Such special-casing goes against the idea of EDSC generality and extensibility.
Instead, we should consider declaring the composed form canonical for
affine.apply operations and using the builder support for creating operations
and canonicalizing them immediately (ongoing effort).
PiperOrigin-RevId: 234790129
Introduce a type-safe way of building a 'for' loop with max/min bounds in EDSC.
Define new types MaxExpr and MinExpr in C++ EDSC API and expose them to Python
bindings. Use values of these type to construct 'for' loops with max/min in
newly introduced overloads of the `edsc::For` factory function. Note that in C
APIs, we still must expose MaxMinFor as a different function because C has no
overloads. Also note that MaxExpr and MinExpr do _not_ derive from Expr
because they are not allowed to be used in a regular Expr context (which may
produce `affine.apply` instructions not expecting `min` or `max`).
Factory functions `Min` and `Max` in Python can be further overloaded to
produce chains of comparisons and selects on non-index types. This is not
trivial in C++ since overloaded functions cannot differ by the return type only
(`MaxExpr` or `Expr`) and making `MaxExpr` derive from `Expr` defies the
purpose of type-safe construction.
PiperOrigin-RevId: 234786131
MLIR supports 'for' loops with lower(upper) bound defined by taking a
maximum(minimum) of a list of expressions, but does not have first-class affine
constructs for the maximum(minimum). All these expressions must have affine
provenance, similarly to a single-expression bound. Add support for
constructing such loops using EDSC. The expression factory function is called
`edsc::MaxMinFor` to (1) highlight that the maximum(minimum) operation is
applied to the lower(upper) bound expressions and (2) differentiate it from a
`edsc::For` that creates multiple perfectly nested loops (and should arguably
be called `edsc::ForNest`).
PiperOrigin-RevId: 234785996
Introduce a functionality to create EDSC expressions from typed constants.
This complements the current functionality that uses "unbound" expressions and
binds them to a specific constant before emission. It comes in handy in cases
where we want to check if something is a constant early during construciton
rather than late during emission, for example multiplications and divisions in
affine expressions. This is also consistent with MLIR vision of constants
being defined by an operation (rather than being special kinds of values in the
IR) by exposing this operation as EDSC expression.
PiperOrigin-RevId: 234758020
This CL fixes 2 recent issues with EDSCs:
1. the type of the LHS in Stmt::operator=(Expr rhs) should be the same as the (asserted unique) return type;
2. symbols coming from DimOp should be admissible as lower / upper bounds in For
The relevant tests are added.
PiperOrigin-RevId: 234750249
- compute slices precisely where the destination iteration depends on multiple source
iterations (instead of over-approximating to the whole source loop extent)
- update unionBoundingBox to deal with input with non-matching symbols
- reenable disabled backend test case
PiperOrigin-RevId: 234714069
- hoist DMAs past all loops immediately surrounding the region that the latter
is invariant on - do this at DMA generation time itself
PiperOrigin-RevId: 234628447
Originally, edsc::Expr had a long enum edsc::ExprKind with all supported types
of operations. Recent Expr extensibility support removed the need to specify
supported types in advance. Replace the no-longer-used blocks of enum values
reserved for unary/binary/ternary/variadic expressions with simple values (it
is still useful to know if an expression is, e.g., binary to access it through
a simpler API).
Furthermore, wrap string-comparison now used to identify specific ops into an
`Expr::is_op<>` function template, that acts similarly to `Instruction::isa<>`.
Introduce `{Unary,Binary,Ternary,Variadic}Expr::make<> ` function template that
creates a Expression emitting the MLIR Op specified as template argument.
PiperOrigin-RevId: 234612916
Expose the result types of edsc::Expr, which are now stored for all types of
Exprs and not only for the variadic ones. Require return types when an Expr is
constructed, if it will ever have some. An empty return type list is
interpreted as an Expr that does not create a value (e.g. `return` or `store`).
Conceptually, all edss::Exprs are now typed, with the type being a (potentially
empty) tuple of return types. Unbound expressions and Bindables must now be
constructed with a specific type they will take. This makes EDSC less
evidently type-polymorphic, but we can still write generic code such as
Expr sumOfSquares(Expr lhs, Expr rhs) { return lhs * lhs + rhs * rhs; }
and use it to construct different typed expressions as
sumOfSquares(Bindable(IndexType::get(ctx)), Bindable(IndexType::get(ctx)));
sumOfSquares(Bindable(FloatType::getF32(ctx)),
Bindable(FloatType::getF32(ctx)));
On the positive side, we get the following.
1. We can now perform type checking when constructing Exprs rather than during
MLIR emission. Nevertheless, this is still duplicates the Op::verify()
until we can factor out type checking from that.
2. MLIREmitter is significantly simplified.
3. ExprKind enum is only used for actual kinds of expressions. Data structures
are converging with AbstractOperation, and the users can now create a
VariadicExpr("canonical_op_name", {types}, {exprs}) for any operation, even
an unregistered one without having to extend the enum and make pervasive
changes to EDSCs.
On the negative side, we get the following.
1. Typed bindables are more verbose, even in Python.
2. We lose the ability to do print debugging for higher-level EDSC abstractions
that are implemented as multiple MLIR Ops, for example logical disjunction.
This is the step 2/n towards making EDSC extensible.
***
Move MLIR Op construction from MLIREmitter::emitExpr to Expr::build since Expr
now has sufficient information to build itself.
This is the step 3/n towards making EDSC extensible.
Both of these strive to minimize the amount of irrelevant changes. In
particular, this introduces more complex pretty-printing for affine and binary
expression to make sure tests continue to pass. It also relies on string
comparison to identify specific operations that an Expr produces.
PiperOrigin-RevId: 234609882
This CL extended TableGen Operator class to provide accessors for information on op
results.
In OpDefinitionGen, added checks to make sure only the last result can be variadic,
and adjusted traits and builders generation to consider variadic results.
PiperOrigin-RevId: 234596124
EDSC currently implement a block as a statement that is itself a list of
statements. This suffers from two modeling problems: (1) these blocks are not
addressable, i.e. one cannot create an instruction where thus constructed block
is a successor; (2) they support block nesting, which is not supported by MLIR
blocks. Furthermore, emitting such "compound statement" (misleadingly named
`Block` in Python bindings) does not actually produce a new Block in the IR.
Implement support for creating actual IR Blocks in EDSC. In particular, define
a new StmtBlock EDSC class that is neither an Expr nor a Stmt but contains a
list of Stmts. Additionally, StmtBlock may have (early-) typed arguments.
These arguments are Bindable expressions that can be used inside the block.
Provide two calls in the MLIREmitter, `emitBlock` that actually emits a new
block and `emitBlockBody` that only emits the instructions contained in the
block without creating a new block. In the latter case, the instructions must
not use block arguments.
Update Python bindings to make it clear when instruction emission happens
without creating a new block.
PiperOrigin-RevId: 234556474
The parameter to emitStandaloneParamBuilder() was renamed from hasResultType to
isAllSameType, which is the opposite boolean value. The logic should be changed
to make them consistent.
Also re-ordered some methods in Operator. And few other tiny improvements.
PiperOrigin-RevId: 234478316
generation pass to make it drop certain assumptions, complete TODOs.
- multiple fixes for getMemoryFootprintBytes
- pass loopDepth correctly from getMemoryFootprintBytes()
- use union while computing memory footprints
- bug fixes for addAffineForOpDomain
- take into account loop step
- add domains of other loop IVs in turn that might have been used in the bounds
- dma-generate: drop assumption of "non-unit stride loops being tile space loops
and skipping those and recursing to inner depths"; DMA generation is now purely
based on available fast mem capacity and memory footprint's calculated
- handle memory region compute failures/bailouts correctly from dma-generate
- loop tiling cleanup/NFC
- update some debug and error messages to use emitNote/emitError in
pipeline-data-transfer pass - NFC
PiperOrigin-RevId: 234245969
Alex Zinenko