Summary: The current syntax for AffineMapAttr and IntegerSetAttr conflict with function types, making it currently impossible to round-trip function types(and e.g. FuncOp) in the IR. This revision changes the syntax for the attributes by wrapping them in a keyword. AffineMapAttr is wrapped with `affine_map<>` and IntegerSetAttr is wrapped with `affine_set<>`.
Reviewed By: nicolasvasilache, ftynse
Differential Revision: https://reviews.llvm.org/D72429
Add one more simplification for floordiv and mod affine expressions.
Examples:
(2*d0 + 1) floordiv 2 is simplified to d0
(8*d0 + 4*d1 + d2) floordiv 4 simplified to 4*d0 + d1 + d2 floordiv 4.
etc.
Similarly, (4*d1 + 1) mod 2 is simplified to 1,
(2*d0 + 8*d1) mod 8 simplified to 2*d0 mod 8.
Change getLargestKnownDivisor to return int64_t to be consistent and
to avoid casting at call sites (since the return value is used in expressions
of int64_t/index type).
Signed-off-by: Uday Bondhugula <uday@polymagelabs.com>
Closestensorflow/mlir#202
COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/mlir/pull/202 from bondhugula:affine b13fcb2f1c00a39ca5434613a02408e085a80e77
PiperOrigin-RevId: 284866710
Now that we have unrolling as a declarative pattern, we can drop a full pass that has gone stale. In the future we may want to add specific unrolling patterns for VectorTransferReadOp.
PiperOrigin-RevId: 283806880
This is step 1/n in refactoring infrastructure along the Vector dialect to make it ready for retargetability and composable progressive lowering.
PiperOrigin-RevId: 280529784
This CL moves VectorOps to Tablegen and cleans up the implementation.
This is almost NFC but 2 changes occur:
1. an interface change occurs in the padding value specification in vector_transfer_read:
the value becomes non-optional. As a shortcut we currently use %f0 for all paddings.
This should become an OpInterface for vectorization in the future.
2. the return type of vector.type_cast is trivial and simplified to `memref<vector<...>>`
Relevant roundtrip and invalid tests that used to sit in core are moved to the vector dialect.
The op documentation is moved to the .td file.
PiperOrigin-RevId: 280430869
This allows mixing linalg operations with vector transfer operations (with additional modifications to affine ops) and is a step towards solving tensorflow/mlir#189.
PiperOrigin-RevId: 275543361
Clipping creates non-affine memory accesses, use std_load and std_store instead of affine_load and affine_store.
In the future we may also want a fill with the neutral element rather than clip, this would make the accesses affine if we wanted more analyses and transformations to happen post lowering to pointwise copies.
PiperOrigin-RevId: 260110503
Standard load and store operations are evolving to be separated from the Affine
constructs. Special affine.load/store have been introduced to uphold the
restrictions of the Affine control flow constructs on their operands.
EDSC-produced loads and stores were originally intended to uphold those
restrictions as well so they should use affine.load/store instead of
std.load/store.
PiperOrigin-RevId: 257443307
Change the AsmPrinter to number values breadth-first so that values in adjacent regions can have the same name. This allows for ModuleOp to contain operations that produce results. This also standardizes the special name of region entry arguments to "arg[0-9+]" now that Functions are also operations.
PiperOrigin-RevId: 257225069
In most places, this is just a name change (with the exception of affine.dma_start swapping the operand positions of its tag memref and num_elements operands).
Significant code changes occur here:
*) Vectorization: LoopAnalysis.cpp, Vectorize.cpp
*) Affine Transforms: Transforms/Utils/Utils.cpp
PiperOrigin-RevId: 256395088
The current syntax separates the name and value with ':', but ':' is already overloaded by several other things(e.g. trailing types). This makes the syntax difficult to parse in some situtations:
Old:
"foo: 10 : i32"
New:
"foo = 10 : i32"
PiperOrigin-RevId: 255097928
This is the standard syntax for types on operations, and is also already used by IntegerAttr and FloatAttr.
Example:
dense<5> : tensor<i32>
dense<[3]> : tensor<1xi32>
PiperOrigin-RevId: 255069157
Trying to activate both LLVM and MLIR passes in mlir-cpu-runner showed name collisions when registering pass names.
One possible way of disambiguating that should also work across dialects is to prepend the dialect name to the passes that specifically operate on that dialect.
With this CL, mlir-cpu-runner tests still run when both LLVM and MLIR passes are registered
--
PiperOrigin-RevId: 246539917
During the pattern rewrite, if the function is changed, i.e. ops created,
deleted or swapped, the pattern rewriter needs to re-scan the function entirely
and apply the patterns again, so the patterns whose root ops have been popped
out from the working list nor an immediate users of the changed ops can be
reconsidered.
A command line flag is added to set the max number of iterations rescanning the
function for pattern match. If the rewrite doesn' converge after this number,
this compiling will continue and the result can be sub-optimal.
One unit test is updated because this change fixed the missing optimization opportunities.
--
PiperOrigin-RevId: 244754190
This CL fixes the non-determinism across compilers in an edsc::select expression used in LowerVectorTransfers. This is achieved by factoring the expression out of the function call to ensure a deterministic order of evaluation.
Since the expression is now factored out, fewer IR is generated and the test is updated accordingly.
--
PiperOrigin-RevId: 241679962
This CL removes the reliance of the vectorize pass on the specification of a `fastestVaryingDim` parameter. This parameter is a restriction meant to more easily target a particular loop/memref combination for vectorization and is mainly used for testing.
This also had the side-effect of restricting vectorization patterns to only the ones in which all memrefs were contiguous along the same loop dimension. This simple restriction prevented matmul to vectorize in 2-D.
this CL removes the restriction and adds the matmul test which vectorizes in 2-D along the parallel loops. Support for reduction loops is left for future work.
PiperOrigin-RevId: 240993827
Example:
%call:2 = call @multi_return() : () -> (f32, i32)
use(%calltensorflow/mlir#0, %calltensorflow/mlir#1)
This cl also adds parser support for uniquely named result values. This means that a test writer can now write something like:
%foo, %bar = call @multi_return() : () -> (f32, i32)
use(%foo, %bar)
Note: The printer will still print the collapsed form.
PiperOrigin-RevId: 240860058
This CL revisits the composition of AffineApplyOp for the special case where a symbol
itself comes from an AffineApplyOp.
This is achieved by rewriting such symbols into dims to allow composition to occur mathematically.
The implementation is also refactored to improve readability.
Rationale for locally rewriting symbols as dims:
================================================
The mathematical composition of AffineMap must always concatenate symbols
because it does not have enough information to do otherwise. For example,
composing `(d0)[s0] -> (d0 + s0)` with itself must produce
`(d0)[s0, s1] -> (d0 + s0 + s1)`.
The result is only equivalent to `(d0)[s0] -> (d0 + 2 * s0)` when
applied to the same mlir::Value* for both s0 and s1.
As a consequence mathematical composition of AffineMap always concatenates
symbols.
When AffineMaps are used in AffineApplyOp however, they may specify
composition via symbols, which is ambiguous mathematically. This corner case
is handled by locally rewriting such symbols that come from AffineApplyOp
into dims and composing through dims.
PiperOrigin-RevId: 239791597
This CL removes the dependency of LowerVectorTransfers on the AST version of EDSCs which will be retired.
This exhibited a pretty fundamental staging difference in AST-based vs declarative based emission.
Since the delayed creation with an AST was staged, the loop order came into existence after the clipping expressions were computed.
This now changes as the loops first need to be created declaratively in fixed order and then the clipping expressions are created.
Also, due to lack of staging, coalescing cannot be done on the fly anymore and
needs to be done either as a pre-pass (current implementation) or as a local transformation on the generated IR (future work).
Tests are updated accordingly.
PiperOrigin-RevId: 238971631
- print multiplication by -1 as unary negate; expressions like s0 * -1, d0 * -1
+ d1 will now appear as -s0, -d0 + d1 resp.
- a minor cleanup while on printAffineExprInternal
PiperOrigin-RevId: 230222151
This CL fixes a misunderstanding in how to build DimOp which triggered
execution issues in the CPU path.
The problem is that, given a `memref<?x4x?x8x?xf32>`, the expressions to
construct the dynamic dimensions should be:
`dim %arg, 0 : memref<?x4x?x8x?xf32>`
`dim %arg, 2 : memref<?x4x?x8x?xf32>`
and
`dim %arg, 4 : memref<?x4x?x8x?xf32>`
Before this CL, we wold construct:
`dim %arg, 0 : memref<?x4x?x8x?xf32>`
`dim %arg, 1 : memref<?x4x?x8x?xf32>`
`dim %arg, 2 : memref<?x4x?x8x?xf32>`
and expect the other dimensions to be constants.
This assumption seems consistent at first glance with the syntax of alloc:
```
%tensor = alloc(%M, %N, %O) : memref<?x4x?x8x?xf32>
```
But this was actuallyincorrect.
This CL also makes the relevant functions available to EDSCs and removes
duplication of the incorrect function.
PiperOrigin-RevId: 229622766
This CL adds a short term remedy to an issue that was found during execution
tests.
Lowering of vector transfer ops uses the permutation map to determine which
ForInst have been super-vectorized. During materialization to HW vector sizes
however, some of those dimensions may be fully unrolled and do not appear in
the permutation map.
Such dimensions were then not clipped and may have accessed out of bounds.
This CL conservatively clips all dimensions to ensure no out of bounds access.
The longer term solution is still up for debate but will probably require
either passing more information between Materialization and lowering, or just
merging the 2 passes.
PiperOrigin-RevId: 228980787
This CL is the 5th on the path to simplifying AffineMap composition.
This removes the distinction between normalized single-result AffineMap and
more general composed multi-result map.
One nice byproduct of making the implementation driven by single-result is
that the multi-result extension is a trivial change: the implementation is
still single-result and we just use:
```
unsigned idx = getIndexOf(...);
map.getResult(idx);
```
This CL also fixes an AffineNormalizer implementation issue related to symbols.
Namely it stops performing substitutions on symbols in AffineNormalizer and
instead concatenates them all to be consistent with the call to
`AffineMap::compose(AffineMap)`. This latter call to `compose` cannot perform
simplifications of symbols coming from different maps based on positions only:
i.e. dims are applied and renumbered but symbols must be concatenated.
The only way to determine whether symbols from different AffineApply are the
same is to look at the concrete values. The canonicalizeMapAndOperands is thus
extended with behavior to support replacing operands that appear multiple
times.
Lastly, this CL demonstrates that the implementation is correct by rewriting
ComposeAffineMaps using only `makeComposedAffineApply`. The implementation
uses a matcher because AffineApplyOp are introduced as composed operations on
the fly instead of iteratively forwardSubstituting. For this purpose, a walker
would revisit freshly introduced AffineApplyOp. Regardless, ComposeAffineMaps
is scheduled to disappear, this CL replaces the implementation based on
iterative `forwardSubstitute` by a composed-by-construction
`makeComposedAffineApply`.
Remaining calls to `forwardSubstitute` will be removed in the next CL.
PiperOrigin-RevId: 228830443
Supervectorization does not plan on handling multi-result AffineMaps and
non-canonical chains of > 1 AffineApplyOp.
This CL uses the simpler single-result unbounded AffineApplyOp in the
MaterializeVectors pass.
PiperOrigin-RevId: 228469085
This CL is the 2nd on the path to simplifying AffineMap composition.
This CL uses the now accepted `AffineExpr::compose(AffineMap)` to
implement `AffineMap::compose(AffineMap)`.
Implications of keeping the simplification function in
Analysis are documented where relevant.
PiperOrigin-RevId: 228276646
This CL is the 1st on the path to simplifying AffineMap composition.
This CL uses the now accepted AffineExpr.replaceDimsAndSymbols to
implement `AffineExpr::compose(AffineMap)`.
Arguably, `simplifyAffineExpr` should be part of IR and not Analysis but
this CL does not yet pull the trigger on that.
PiperOrigin-RevId: 228265845
Supervectorization does not plan on handling multi-result AffineMaps and
non-canonical chains of > 1 AffineApplyOp.
This CL introduces a simpler abstraction and composition of single-result
unbounded AffineApplyOp by using the existing unbound AffineMap composition.
This CL adds a simple API call and relevant tests:
```c++
OpPointer<AffineApplyOp> makeNormalizedAffineApply(
FuncBuilder *b, Location loc, AffineMap map, ArrayRef<Value*> operands);
```
which creates a single-result unbounded AffineApplyOp.
The operands of AffineApplyOp are not themselves results of AffineApplyOp by
consrtuction.
This represent the simplest possible interface to complement the composition
of (mathematical) AffineMap, for the cases when we are interested in applying
it to Value*.
In this CL the composed AffineMap is not compressed (i.e. there exist operands
that are not part of the result). A followup commit will compress to normal
form.
The single-result unbounded AffineApplyOp abstraction will be used in a
followup CL to support the MaterializeVectors pass.
PiperOrigin-RevId: 227879021
The entire compiler now looks at structural properties of the function (e.g.
does it have one block, does it contain an if/for stmt, etc) so the only thing
holding up this difference is round tripping through the parser/printer syntax.
Removing this shrinks the compile by ~140LOC.
This is step 31/n towards merging instructions and statements. The last step
is updating the docs, which I will do as a separate patch in order to split it
from this mostly mechanical patch.
PiperOrigin-RevId: 227540453
This CL introduces a simple set of Embedded Domain-Specific Components (EDSCs)
in MLIR components:
1. a `Type` system of shell classes that closely matches the MLIR type system. These
types are subdivided into `Bindable` leaf expressions and non-bindable `Expr`
expressions;
2. an `MLIREmitter` class whose purpose is to:
a. maintain a map of `Bindable` leaf expressions to concrete SSAValue*;
b. provide helper functionality to specify bindings of `Bindable` classes to
SSAValue* while verifying comformable types;
c. traverse the `Expr` and emit the MLIR.
This is used on a concrete example to implement MemRef load/store with clipping in the
LowerVectorTransfer pass. More specifically, the following pseudo-C++ code:
```c++
MLFuncBuilder *b = ...;
Location location = ...;
Bindable zero, one, expr, size;
// EDSL expression
auto access = select(expr < zero, zero, select(expr < size, expr, size - one));
auto ssaValue = MLIREmitter(b)
.bind(zero, ...)
.bind(one, ...)
.bind(expr, ...)
.bind(size, ...)
.emit(location, access);
```
is used to emit all the MLIR for a clipped MemRef access.
This simple EDSL can easily be extended to more powerful patterns and should
serve as the counterpart to pattern matchers (and could potentially be unified
once we get enough experience).
In the future, most of this code should be TableGen'd but for now it has
concrete valuable uses: make MLIR programmable in a declarative fashion.
This CL also adds Stmt, proper supporting free functions and rewrites
VectorTransferLowering fully using EDSCs.
The code for creating the EDSCs emitting a VectorTransferReadOp as loops
with clipped loads is:
```c++
Stmt block = Block({
tmpAlloc = alloc(tmpMemRefType),
vectorView = vector_type_cast(tmpAlloc, vectorMemRefType),
ForNest(ivs, lbs, ubs, steps, {
scalarValue = load(scalarMemRef, accessInfo.clippedScalarAccessExprs),
store(scalarValue, tmpAlloc, accessInfo.tmpAccessExprs),
}),
vectorValue = load(vectorView, zero),
tmpDealloc = dealloc(tmpAlloc.getLHS())});
emitter.emitStmt(block);
```
where `accessInfo.clippedScalarAccessExprs)` is created with:
```c++
select(i + ii < zero, zero, select(i + ii < N, i + ii, N - one));
```
The generated MLIR resembles:
```mlir
%1 = dim %0, 0 : memref<?x?x?x?xf32>
%2 = dim %0, 1 : memref<?x?x?x?xf32>
%3 = dim %0, 2 : memref<?x?x?x?xf32>
%4 = dim %0, 3 : memref<?x?x?x?xf32>
%5 = alloc() : memref<5x4x3xf32>
%6 = vector_type_cast %5 : memref<5x4x3xf32>, memref<1xvector<5x4x3xf32>>
for %i4 = 0 to 3 {
for %i5 = 0 to 4 {
for %i6 = 0 to 5 {
%7 = affine_apply #map0(%i0, %i4)
%8 = cmpi "slt", %7, %c0 : index
%9 = affine_apply #map0(%i0, %i4)
%10 = cmpi "slt", %9, %1 : index
%11 = affine_apply #map0(%i0, %i4)
%12 = affine_apply #map1(%1, %c1)
%13 = select %10, %11, %12 : index
%14 = select %8, %c0, %13 : index
%15 = affine_apply #map0(%i3, %i6)
%16 = cmpi "slt", %15, %c0 : index
%17 = affine_apply #map0(%i3, %i6)
%18 = cmpi "slt", %17, %4 : index
%19 = affine_apply #map0(%i3, %i6)
%20 = affine_apply #map1(%4, %c1)
%21 = select %18, %19, %20 : index
%22 = select %16, %c0, %21 : index
%23 = load %0[%14, %i1, %i2, %22] : memref<?x?x?x?xf32>
store %23, %5[%i6, %i5, %i4] : memref<5x4x3xf32>
}
}
}
%24 = load %6[%c0] : memref<1xvector<5x4x3xf32>>
dealloc %5 : memref<5x4x3xf32>
```
In particular notice that only 3 out of the 4-d accesses are clipped: this
corresponds indeed to the number of dimensions in the super-vector.
This CL also addresses the cleanups resulting from the review of the prevous
CL and performs some refactoring to simplify the abstraction.
PiperOrigin-RevId: 227367414
printing the entry block in a CFG function's argument line. Since I'm touching
all of the testcases anyway, change the argument list from printing as
"%arg : type" to "%arg: type" which is more consistent with bb arguments.
In addition to being more consistent, this is a much nicer look for cfg functions.
PiperOrigin-RevId: 227240069
