This CL makes some minor changes to the declarative builder Helpers:
1. adds lb, ub, step methods to MemRefView to avoid always having to go through std::get + range;
2. drops MemRefView& from IndexedValue which was just creating ownership concerns. Instead, an IndexedValue only needs to keep track of the ValueHandle from which a MemRefView can be constructed on-demand if necessary.
PiperOrigin-RevId: 237861493
TensorFlow comparison ops like tf.Less supports broadcast behavior but the result
type have different element types as the input types. Extend broadcastable trait
to allow such cases. Added tf.Less to demonstrate it.
PiperOrigin-RevId: 237846127
So that we can use this function to deduce broadcasted shapes elsewhere.
Also added support for unknown dimensions, by following TensorFlow behavior.
PiperOrigin-RevId: 237846065
Below shows the output for an example mlir-opt command line.
mlir-opt foo.mlir -verify-each=false -cse -canonicalize -cse -cse -pass-timing
list view (-pass-timing-display=list):
* In this mode the results are displayed in a list sorted by total time; with each pass/analysis instance aggregated into one unique result. This mode is similar to the output of 'time-passes' in llvm-opt.
===-------------------------------------------------------------------------===
... Pass execution timing report ...
===-------------------------------------------------------------------------===
Total Execution Time: 0.0097 seconds (0.0096 wall clock)
---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name ---
0.0051 ( 58.3%) 0.0001 ( 12.2%) 0.0052 ( 53.8%) 0.0052 ( 53.8%) Canonicalizer
0.0025 ( 29.1%) 0.0005 ( 58.2%) 0.0031 ( 31.9%) 0.0031 ( 32.0%) CSE
0.0011 ( 12.6%) 0.0003 ( 29.7%) 0.0014 ( 14.3%) 0.0014 ( 14.2%) DominanceInfo
0.0087 (100.0%) 0.0009 (100.0%) 0.0097 (100.0%) 0.0096 (100.0%) Total
pipeline view (-pass-timing-display=pipeline):
* In this mode the results are displayed in a nested pipeline view that mirrors the internal pass pipeline that is being executed in the pass manager. This view is useful for understanding specifically which parts of the pipeline are taking the most time, and can also be used to identify when analyses are being invalidated and recomputed.
===-------------------------------------------------------------------------===
... Pass execution timing report ...
===-------------------------------------------------------------------------===
Total Execution Time: 0.0082 seconds (0.0081 wall clock)
---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name ---
0.0042 (100.0%) 0.0039 (100.0%) 0.0082 (100.0%) 0.0081 (100.0%) Function Pipeline
0.0005 ( 11.6%) 0.0008 ( 21.1%) 0.0013 ( 16.1%) 0.0013 ( 16.2%) CSE
0.0002 ( 5.0%) 0.0004 ( 9.3%) 0.0006 ( 7.0%) 0.0006 ( 7.0%) (A) DominanceInfo
0.0026 ( 61.8%) 0.0018 ( 45.6%) 0.0044 ( 54.0%) 0.0044 ( 54.1%) Canonicalizer
0.0005 ( 11.7%) 0.0005 ( 13.0%) 0.0010 ( 12.3%) 0.0010 ( 12.4%) CSE
0.0003 ( 6.1%) 0.0003 ( 8.3%) 0.0006 ( 7.2%) 0.0006 ( 7.1%) (A) DominanceInfo
0.0002 ( 3.8%) 0.0001 ( 2.8%) 0.0003 ( 3.3%) 0.0003 ( 3.3%) CSE
0.0042 (100.0%) 0.0039 (100.0%) 0.0082 (100.0%) 0.0081 (100.0%) Total
PiperOrigin-RevId: 237825367
* before/after pass execution
* after a pass fails
* before/after an analysis is computed
After getting this infrastructure in place, we can start providing common developer utilities like pass timing, IR printing after pass execution, etc.
PiperOrigin-RevId: 237709692
This CL addresses a few post-submit comments:
1. better comments,
2. check number of results before dyn_cast (which is a less common case)
3. test usage for multi-result InstructionHandle
PiperOrigin-RevId: 237549333
This CL adds support for named custom instructions in declarative builders.
To allow this, it introduces a templated `CustomInstruction` class.
This CL also splits ValueHandle which can capture only the **value** in single-valued instructions from InstructionHandle which can capture any instruction but provide no typing and sugaring to extract the potential Value*.
PiperOrigin-RevId: 237543222
This CL added the ability to generate multiple ops using multiple result
patterns, with each of them replacing one result of the matched source op.
Specifically, the syntax is
```
def : Pattern<(SourceOp ...),
[(ResultOp1 ...), (ResultOp2 ...), (ResultOp3 ...)]>;
```
Assuming `SourceOp` has three results.
Currently we require that each result op must generate one result, which
can be lifted later when use cases arise.
To help with cases that certain output is unused and we don't care about it,
this CL also introduces a new directive: `verifyUnusedValue`. Checks will
be emitted in the `match()` method to make sure if the corresponding output
is not unused, `match()` returns with `matchFailure()`.
PiperOrigin-RevId: 237513904
TensorFlow does not allow integers of random bitwidths. It only accepts 8-,
16-, 32-, and 64-bit integer types. Similarly for floating point types, only
half, single, double, and bfloat16 types.
PiperOrigin-RevId: 237483913
- compute tile sizes based on a simple model that looks at memory footprints
(instead of using the hardcoded default value)
- adjust tile sizes to make them factors of trip counts based on an option
- update loop fusion CL options to allow setting maximal fusion at pass creation
- change an emitError to emitWarning (since it's not a hard error unless the client
treats it that way, in which case, it can emit one)
$ mlir-opt -debug-only=loop-tile -loop-tile test/Transforms/loop-tiling.mlir
test/Transforms/loop-tiling.mlir:81:3: note: using tile sizes [4 4 5 ]
for %i = 0 to 256 {
for %i0 = 0 to 256 step 4 {
for %i1 = 0 to 256 step 4 {
for %i2 = 0 to 250 step 5 {
for %i3 = #map4(%i0) to #map11(%i0) {
for %i4 = #map4(%i1) to #map11(%i1) {
for %i5 = #map4(%i2) to #map12(%i2) {
%0 = load %arg0[%i3, %i5] : memref<8x8xvector<64xf32>>
%1 = load %arg1[%i5, %i4] : memref<8x8xvector<64xf32>>
%2 = load %arg2[%i3, %i4] : memref<8x8xvector<64xf32>>
%3 = mulf %0, %1 : vector<64xf32>
%4 = addf %2, %3 : vector<64xf32>
store %4, %arg2[%i3, %i4] : memref<8x8xvector<64xf32>>
}
}
}
}
}
}
PiperOrigin-RevId: 237461836
Recently, EDSC introduced an eager mode for building IR in different contexts.
Introduce Python bindings support for loop and loop nest contexts of EDSC
builders. The eager mode is built around the notion of ValueHandle, which is
convenience class for delayed initialization and operator overloads. Expose
this class and overloads directly. The model of insertion contexts maps
naturally to Python context manager mechanism, therefore new bindings are
defined bypassing the C APIs. The bindings now provide three new context
manager classes: FunctionContext, LoopContext and LoopNestContext. The last
two can be used with the `with`-construct in Python to create loop (nests) and
obtain handles to the loop induction variables seamlessly:
with LoopContext(lhs, rhs, 1) as i:
lhs + rhs + i
with LoopContext(rhs, rhs + rhs, 2) as j:
x = i + j
Any statement within the Python context will trigger immediate emission of the
corresponding IR constructs into the context owned by the nearest context
manager.
PiperOrigin-RevId: 237447732
These cleanups reflects some recent changes to the LLVM IR Dialect and the
infrastructure that affects it. In particular, add documentation on direct and
indirect function calls as well as remove the `call` and `call0` separation.
Change the prefix of custom types from `!llvm.type` to `!llvm` so that it
matches the IR. Remove the verifier check disallowing conditional branches to
the same block with arguments: identical arguments are now supported, and
different arguments will be caught later.
PiperOrigin-RevId: 237203452
The LLVM IR Dialect strives to be close to the original LLVM IR instructions.
The conversion from the LLVM IR Dialect to LLVM IR proper is mostly mechanical
and can be automated. Implement TableGen support for generating conversions
from a concise pattern form in the TableGen definition of the LLVM IR Dialect
operations. It is used for all operations except calls and branches. These
operations need access to function and block remapping tables and would require
significantly more code to generate the conversions from TableGen definitions
than the current manually written conversions.
This implementation is accompanied by various necessary changes to the TableGen
operation definition infrastructure. In particular, operation definitions now
contain named accessors to results as well as named accessors to the variadic
operand (returning a vector of operands). The base operation support TableGen
file now contains a FunctionAttr definition. The TableGen now allows to query
the names of the operation results.
PiperOrigin-RevId: 237203077
* bool succeeded(Status)
- Return if the status corresponds to a success value.
* bool failed(Status)
- Return if the status corresponds to a failure value.
PiperOrigin-RevId: 237153884
This CL adds the same helper classes that exist in the AST form of EDSCs to support a basic indexing notation and emit the proper load and store operations and capture MemRefViews as function arguments.
This CL also adds a wrapper class LoopNestBuilder to allow generic rank-agnostic loops over indices.
PiperOrigin-RevId: 237113755
When building unstructured control-flow there is a need to construct mlir::Block* before being able to fill them. This invites goto-style programming.
This CL introduces an alternative eager API for BR and COND_BR in which blocks are created eagerly and captured on the fly.
This allows reducing the number of calls to `BlockBuilder` from 4 to 2 in the `builder_blocks_eager` test and from 3 to 2 in the `builder_cond_branch_eager` test.
PiperOrigin-RevId: 237046114
This CL adds support for BranchHandle and BranchBuilder that require a slightly different
abstraction since an mlir::Block is not an mlir::Value.
This CL also adds support for the BR and COND_BR instructions and the relevant tests.
PiperOrigin-RevId: 237034312
This CL reworks the design of EDSCs from first principles.
It introduces a ValueHandle which can hold either:
1. an eagerly typed, delayed Value*
2. a precomputed Value*
A ValueHandle can be manipulated with intrinsic operations a nested within a NestedBuilder. These NestedBuilder are a more idiomatic nested builder abstraction that should feel intuitive to program in C++.
Notably, this abstraction does not require an AST to stage the construction of MLIR snippets from C++. Instead, the abstraction makes use of orderings between definition and declaration of ValueHandles and provides a NestedBuilder and a LoopBuilder helper classes to handle those orderings.
All instruction creations are meant to use the templated ValueHandle::create<> which directly calls mlir::Builder.create<>.
For now the EDSC AST and the builders live side-by-side until the C API is ported.
PiperOrigin-RevId: 237030945
The existing implementation of the Op definition generator assumes and relies
on the fact that native Op Attributes appear after its value-based operands in
the Arguments list. Furthermore, the same order is used in the generated
`build` function for the operation. This is not desirable for some operations
with mandatory attributes that would want the attribute to appear upfront for
better consistency with their textual representation, for example `cmpi` would
prefer the `predicate` attribute to be foremost in the argument list.
Introduce support for using attributes and operands in the Arguments DAG in no
particular order. This is achieved by maintaining a list of Arguments that
point to either the value or the attribute and are used to generate the `build`
method.
PiperOrigin-RevId: 237002921
Adds utility to convert slice bounds to a FlatAffineConstraints representation.
Adds utility to FlatAffineConstraints to promote loop IV symbol identifiers to dim identifiers.
PiperOrigin-RevId: 236973261
- change this for consistency - everything else similar takes/returns a
Function pointer - the FuncBuilder ctor,
Block/Value/Instruction::getFunction(), etc.
- saves a whole bunch of &s everywhere
PiperOrigin-RevId: 236928761
This fixes a bug: previously, during conversion function argument
attributes were neither beings passed through nor converted. This fix
extends DialectConversion to allow for simultaneous conversion of the
function type and the argument attributes.
This was important when lowering MLIR to LLVM where attribute
information (e.g. noalias) needs to be preserved in MLIR(LLVMDialect).
Longer run it seems reasonable that we want to convert both the
function attribute and its type and the argument attributes, but that
requires a small refactoring in Function.h to aggregate these three
fields in an inner struct, which will require some discussion.
PiperOrigin-RevId: 236709409
Dialect attributes are defined as:
dialect-namespace `.` attr-name `:` attribute-value
Dialects can override any of the following hooks to verify the validity of a given attribute:
* verifyFunctionAttribute
* verifyFunctionArgAttribute
* verifyInstructionAttribute
PiperOrigin-RevId: 236507970
An analysis can be any class, but it must provide the following:
* A constructor for a given IR unit.
struct MyAnalysis {
// Compute this analysis with the provided module.
MyAnalysis(Module *module);
};
Analyses can be accessed from a Pass by calling either the 'getAnalysisResult<AnalysisT>' or 'getCachedAnalysisResult<AnalysisT>' methods. A FunctionPass may query for a cached analysis on the parent module with 'getCachedModuleAnalysisResult'. Similary, a ModulePass may query an analysis, it doesn't need to be cached, on a child function with 'getFunctionAnalysisResult'.
By default, when running a pass all cached analyses are set to be invalidated. If no transformation was performed, a pass can use the method 'markAllAnalysesPreserved' to preserve all analysis results. As noted above, preserving specific analyses is not yet supported.
PiperOrigin-RevId: 236505642
This CL changes dialect op source files (.h, .cpp, .td) to follow the following
convention:
<full-dialect-name>/<dialect-namespace>Ops.{h|cpp|td}
Builtin and standard dialects are specially treated, though. Both of them do
not have dialect namespace; the former is still named as BuiltinOps.* and the
latter is named as Ops.*.
Purely mechanical. NFC.
PiperOrigin-RevId: 236371358
- detect all parallel loops based on dep information and mark them with a
"parallel" attribute
- add mlir::isLoopParallel(OpPointer<AffineForOp> ...), and refactor an existing method
to use that (reuse some code from @andydavis (cl/236007073) for this)
- a simple/meaningful way to test memref dep test as well
Ex:
$ mlir-opt -detect-parallel test/Transforms/parallelism-detection.mlir
#map1 = ()[s0] -> (s0)
func @foo(%arg0: index) {
%0 = alloc() : memref<1024x1024xvector<64xf32>>
%1 = alloc() : memref<1024x1024xvector<64xf32>>
%2 = alloc() : memref<1024x1024xvector<64xf32>>
for %i0 = 0 to %arg0 {
for %i1 = 0 to %arg0 {
for %i2 = 0 to %arg0 {
%3 = load %0[%i0, %i2] : memref<1024x1024xvector<64xf32>>
%4 = load %1[%i2, %i1] : memref<1024x1024xvector<64xf32>>
%5 = load %2[%i0, %i1] : memref<1024x1024xvector<64xf32>>
%6 = mulf %3, %4 : vector<64xf32>
%7 = addf %5, %6 : vector<64xf32>
store %7, %2[%i0, %i1] : memref<1024x1024xvector<64xf32>>
} {parallel: false}
} {parallel: true}
} {parallel: true}
return
}
PiperOrigin-RevId: 236367368
*) Breaks fusion pass into multiple sub passes over nodes in data dependence graph:
- first pass fuses single-use producers into their unique consumer.
- second pass enables fusing for input-reuse by fusing sibling nodes which read from the same memref, but which do not share dependence edges.
- third pass fuses remaining producers into their consumers (Note that the sibling fusion pass may have transformed a producer with multiple uses into a single-use producer).
*) Fusion for input reuse is enabled by computing a sibling node slice using the load/load accesses to the same memref, and fusion safety is guaranteed by checking that the sibling node memref write region (to a different memref) is preserved.
*) Enables output vector and output matrix computations from KFAC patches-second-moment operation to fuse into a single loop nest and reuse input from the image patches operation.
*) Adds a generic loop utilitiy for finding all sequential loops in a loop nest.
*) Adds and updates unit tests.
PiperOrigin-RevId: 236350987
- add a method to merge and align the spaces (identifiers) of two
FlatAffineConstraints (both get dimension-wise and symbol-wise unique
columns)
- this completes several TODOs, gets rid of previous assumptions/restrictions
in composeMap, unionBoundingBox, and reuses common code
- remove previous workarounds / duplicated funcitonality in
FlatAffineConstraints::composeMap and unionBoundingBox, use mergeAlignIds
from both
PiperOrigin-RevId: 236320581
EDSC Expressions can now be used to build arbitrary MLIR operations identified
by their canonical name, i.e. the name obtained from
`OpClass::getOperationName()` for registered operations. Expose this
functionality to the C API and Python bindings. This exposes builder-level
interface to Python and avoids the need for experimental Python code to
implement EDSC free function calls for constructing each op type.
This modification required exposing mlir::Attribute to the C API and Python
bindings, which only supports integer attributes for now.
This is step 4/n to making EDSCs more generalizable.
PiperOrigin-RevId: 236306776
When the LLVM IR dialect was implemented, TableGen operation definition scheme
did not support operations with variadic results. Therefore, the `call`
instruction was split into `call` and `call0` for the single- and zero-result
calls (LLVM does not support multi-result operations). Unify `call` and
`call0` using the recently added TableGen support for operations with Variadic
results. Explicitly verify that the new operation has 0 or 1 results. As a
side effect, this change enables clean-ups in the conversion to the LLVM IR
dialect that no longer needs to rely on wrapped LLVM IR void types when
constructing zero-result calls.
PiperOrigin-RevId: 236119197
- detect more trivially redundant constraints in
FlatAffineConstraints::removeTrivialRedundantConstraints. Redundancy due to
constraints that only differ in the constant part (eg., 32i + 64j - 3 >= 0, 32 +
64j - 8 >= 0) is now detected. The method is still linear-time and does
a single scan over the FlatAffineConstraints buffer. This detection is useful
and needed to eliminate redundant constraints generated after FM elimination.
- update GCDTightenInequalities so that we also normalize by the GCD while at
it. This way more constraints will show up as redundant (232i - 203 >= 0
becomes i - 1 >= 0 instead of 232i - 232 >= 0) without having to call
normalizeConstraintsByGCD.
- In FourierMotzkinEliminate, call GCDTightenInequalities and
normalizeConstraintsByGCD before calling removeTrivialRedundantConstraints()
- so that more redundant constraints are detected. As a result, redundancy
due to constraints like i - 5 >= 0, i - 7 >= 0, 2i - 5 >= 0, 232i - 203 >=
0 is now detected (here only i >= 7 is non-redundant).
As a result of these, a -memref-bound-check on the added test case runs in 16ms
instead of 1.35s (opt build) and no longer returns a conservative result.
PiperOrigin-RevId: 235983550
The definitions of derived passes have now changed and passes must adhere to the following:
* Inherit from a CRTP base class FunctionPass/ModulePass.
- This class provides several necessary utilities for the transformation:
. Access to the IR unit being transformed (getFunction/getModule)
. Various utilities for pass identification and registration.
* Provide a 'PassResult runOn(Function|Module)()' method to transform the IR.
- This replaces the runOn* functions from before.
This patch also introduces the notion of the PassManager. This allows for simplified construction of pass pipelines and acts as the sole interface for executing passes. This is important as FunctionPass will no longer have a 'runOnModule' method.
PiperOrigin-RevId: 235952008
- clean up loop fusion CL options for promoting local buffers to fast memory
space
- add parameters to loop fusion pass instantiation
PiperOrigin-RevId: 235813419
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
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
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
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
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
- 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
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
A recent change introduced a possibility to run LLVM IR transformation during
JIT-compilation in the ExecutionEngine. Provide helper functions that
construct IR transformers given either clang-style optimization levels or a
list passes to run. The latter wraps the LLVM command line option parser to
parse strings rather than actual command line arguments. As a result, we can
run either of
mlir-cpu-runner -O3 input.mlir
mlir-cpu-runner -some-mlir-pass -llvm-opts="-llvm-pass -other-llvm-pass"
to combine different transformations. The transformer builder functions are
provided as a separate library that depends on LLVM pass libraries unlike the
main execution engine library. The library can be used for integrating MLIR
execution engine into external frameworks.
PiperOrigin-RevId: 234173493
*) Adds utility to LoopUtils to perform loop interchange of two AffineForOps.
*) Adds utility to LoopUtils to sink a loop to a specified depth within a loop nest, using a series of loop interchanges.
*) Computes dependences between all loads and stores in the loop nest, and classifies each loop as parallel or sequential.
*) Computes loop interchange permutation required to sink sequential loops (and raise parallel loop nests) while preserving relative order among them.
*) Checks each dependence against the permutation to make sure that dependences would not be violated by the loop interchange transformation.
*) Calls loop interchange in LoopFusion pass on consumer loop nests before fusing in producers, sinking loops with loop carried dependences deeper into the consumer loop nest.
*) Adds and updates related unit tests.
PiperOrigin-RevId: 234158370
We specify op operands and results in TableGen op definition using the same syntax.
They should be modelled similarly in TableGen driver wrapper classes.
PiperOrigin-RevId: 234153332
Function types are built-in in MLIR and affect the validity of the IR itself.
However, advanced target dialects such as the LLVM IR dialect may include
custom function types. Until now, dialect conversion was expecting function
types not to be converted to the custom type: although the signatures was
allowed to change, the outer type must have been an mlir::FunctionType. This
effectively prevented dialect conversion from creating instructions that
operate on values of the custom function type.
Dissociate function signature conversion from general type conversion.
Function signature conversion must still produce an mlir::FunctionType and is
used in places where built-in types are required to make IR valid. General
type conversion is used for SSA values, including function and block arguments
and function results.
Exercise this behavior in the LLVM IR dialect conversion by converting function
types to LLVM IR function pointer types. The pointer to a function is chosen
to provide consistent lowering of higher-order functions: while it is possible
to have a value of function type, it is not possible to create a function type
accepting a returning another function type.
PiperOrigin-RevId: 234124494
Update FlatAffineConstraints::getLower/UpperBounds to project to the identifier for which bounds are being computed. This change enables computing bounds on an identifier which were previously dependent on the bounds of another identifier.
PiperOrigin-RevId: 234017514
If we see an add op adding a constant value to a convolution op with constant
bias, we can fuse the add into the convolution op by constant folding the
bias and the add op's constant operand.
This CL also removes dangling RewriterGen check that prevents us from using
nested DAG nodes in result patterns, which is already supported.
PiperOrigin-RevId: 233989654
For ops with the SameOperandsAndResultType trait, we know that all result types
should be the same as the first operand's type. So we can generate a build()
method without requiring result types as parameters and also invoke this method
when constructing such ops during expanding rewrite patterns.
Similarly for ops have broadcast behavior, we can define build() method to use
the deduced type as the result type. So we can also calling into this build()
method when constructing ops in RewriterGen.
PiperOrigin-RevId: 233988307
EDSC expressions evolved to have different types of underlying storage.
Separate classes are used for unary, binary, ternary and variadic expressions.
The latter covers all the needs of the three special cases. Remove these
special cases and use a single ExprStorage class everywhere while maintaining
the same APIs at the Expr level (ExprStorage is an internal implementation
class).
This is step 1/n to converging EDSC expressions and Ops and making EDSCs
support custom operations.
PiperOrigin-RevId: 233704912
In the current state, edsc::Expr and edsc::Stmt overload operators to construct
other Exprs and Stmts. This includes some unconventional overloads of the
`operator==` to create a comparison expression and of the `operator!` to create
a negation expression. This situation could lead to unpleasant surprises where
the code does not behave like expected. Make all Expr and Stmt construction
operators free functions and move them to the `edsc::op` namespace. Callers
willing to use these operators must explicitly include them with the `using`
declaration. This can be done in some local scope.
Additionally, we currently emit signed comparisons for order-comparison
operators. With namespaces, we can later introduce two sets of operators in
different namespace, e.g. `edsc::op::sign` and `edsc::op::unsign` to clearly
state which kind of comparison is implied.
PiperOrigin-RevId: 233578674
Associates opaque constants with a particular dialect. Adds general mechanism to register dialect-specific hooks defined in external components. Adds hooks to decode opaque tensor constant and extract an element of an opaque tensor constant.
This CL does not change the existing mechanism for registering constant folding hook yet. One thing at a time.
PiperOrigin-RevId: 233544757
- for the DMA buffers being allocated (and their tags), generate corresponding deallocs
- minor related update to replaceAllMemRefUsesWith and PipelineDataTransfer pass
Code generation for DMA transfers was being done with the initial simplifying
assumption that the alloc's would map to scoped allocations, and so no
deallocations would be necessary. Drop this assumption to generalize. Note that
even with scoped allocations, unrolling loops that have scoped allocations
could create a series of allocations and exhaustion of fast memory. Having a
end of lifetime marker like a dealloc in fact allows creating new scopes if
necessary when lowering to a backend and still utilize scoped allocation.
DMA buffers created by -dma-generate are guaranteed to have either
non-overlapping lifetimes or nested lifetimes.
PiperOrigin-RevId: 233502632
Previously we were using PatternRewrite::replaceOpWithNewOp() to both create the new op
inline and rewrite the matched op. That does not work well if we want to generate multiple
ops in a sequence. To support that, this CL changed to assign each newly created op to a
separate variable.
This CL also refactors how PatternEmitter performs the directive dispatch logic.
PiperOrigin-RevId: 233206819
That allows TensorFlow Add and Div ops to use Broadcastable op trait instead of
more restrictive SameValueType op trait.
That in turn allows TensorFlow ops to be registered by defining GET_OP_LIST and
including the generated ops file. Currently, tf-raise-control-flow pass tests
are using dynamic shapes in tf.Add op and AddOp can't be registered without
supporting the dynamic shapes.
TESTED with unit tests
PiperOrigin-RevId: 232927998
* Add tf.LeakyRelu op definition + folders (well one is really canonicalizer)
* Change generated error message to use attribute description instead;
* Change the return type of F32Attr to be APFloat - internally it is already
stored as APFloat so let the caller decides if they want to convert it or
not. I could see varying opinions here though :) (did not change i32attr
similarly)
PiperOrigin-RevId: 232923358
Aggregate types where at least one dimension is zero do not fully make sense as
they cannot contain any values (their total size is zero). However, TensorFlow
and XLA support tensors with zero sizes, so we must support those too. This is
relatively safe since, unlike vectors and memrefs, we don't have first-class
element accessors for MLIR tensors.
To support sparse element attributes of vector types that have no non-zero
elements, make sure that index and value element attributes have tensor type so
that we never need to create a zero vector type internally. Note that this is
already consistent with the inline documentation of the sparse elements
attribute. Users of the sparse elements attribute should not rely on the
storage schema anyway.
PiperOrigin-RevId: 232896707
The current ExecutionEngine flow generates the LLVM IR from MLIR and
JIT-compiles it as is without any transformation. It thus misses the
opportunity to perform optimizations supported by LLVM or collect statistics
about the module. Modify the Orc JITter to perform transformations on the LLVM
IR. Accept an optional LLVM module transformation function when constructing
the ExecutionEngine and use it while JIT-compiling. This prevents MLIR
ExecutionEngine from depending on LLVM passes; its clients should depend on the
passes they require.
PiperOrigin-RevId: 232877060
Instead, we deduce the result type from the given attribute.
This is in preparation for generating constant ops with TableGen.
PiperOrigin-RevId: 232723467
*) Adds parameter to public API of MemRefRegion::compute for passing in the slice loop bounds to compute the memref region of the loop nest slice.
*) Exposes public method MemRefRegion::getRegionSize for computing the size of the memref region in bytes.
PiperOrigin-RevId: 232706165
Previously, we were using the trait mechanism to specify that an op has variadic operands.
That led a discrepancy between how we handle ops with deterministic number of operands.
Besides, we have no way to specify the constraints and match against the variadic operands.
This CL introduced Variadic<Type> as a way to solve the above issues.
PiperOrigin-RevId: 232656104
* AffineStructures has moved to IR.
* simplifyAffineExpr/simplifyAffineMap/getFlattenedAffineExpr have moved to IR.
* makeComposedAffineApply/fullyComposeAffineMapAndOperands have moved to AffineOps.
* ComposeAffineMaps is replaced by AffineApplyOp::canonicalize and deleted.
PiperOrigin-RevId: 232586468
Motivation for this change is to remove redundant TF type attributes for
TensorFlow ops. For example, tf$T: "tfdtype$DT_FLOAT". Type attributes can be derived using the MLIR operand or result MLIR types, attribute names and their mapping. This will also allow constant folding of instructions generated within MLIR (and not imported from TensorFlow) without adding type attributes for the instruction.
Derived attributes are populated while exporting MLIR to TF GraphDef using
auto-generated populators. Populators are only available for the ops that are generated by the TableGen.
Also, fixed Operator::getNumArgs method to exclude derived attributes as they are not
part of the arguments.
TESTED with unit test
PiperOrigin-RevId: 232531561
In optional attribute dictionary used, among others, in the generic form of the
ops, attribute types for integers and floats are omitted. This could lead to
inconsistencies when round-tripping the IR, in particular the attributes are
created with incorrect types after parsing (integers default to i64, floats
default to f64). Provide API to emit a trailing type after the attribute for
integers and floats. Use it while printing the optional attribute dictionary.
Omitting types for i64 and f64 is a pragmatic decision that minimizes changes
in tests. We may want to reconsider in the future and always print types of
attributes in the generic form.
PiperOrigin-RevId: 232480116
- use getAccessMap() instead of repeating it
- fold getMemRefRegion into MemRefRegion ctor (more natural, avoid heap
allocation and unique_ptr where possible)
- change extractForInductionVars - MutableArrayRef -> ArrayRef for the
arguments. Since the method is just returning copies of 'Value *', the client
can't mutate the pointers themselves; it's fine to mutate the 'Value''s
themselves, but that doesn't mutate the pointers to those.
- change the way extractForInductionVars returns (see b/123437690)
PiperOrigin-RevId: 232359277
loops), (2) take into account fast memory space capacity and lower 'dmaDepth'
to fit, (3) add location information for debug info / errors
- change dma-generate pass to work on blocks of instructions (start/end
iterators) instead of 'for' loops; complete TODOs - allows DMA generation for
straightline blocks of operation instructions interspersed b/w loops
- take into account fast memory capacity: check whether memory footprint fits
in fastMemoryCapacity parameter, and recurse/lower the depth at which DMA
generation is performed until it does fit in the provided memory
- add location information to MemRefRegion; any insufficient fast memory
capacity errors or debug info w.r.t dma generation shows location information
- allow DMA generation pass to be instantiated with a fast memory capacity
option (besides command line flag)
- change getMemRefRegion to return unique_ptr's
- change getMemRefFootprintBytes to work on a 'Block' instead of 'ForInst'
- other helper methods; add postDomInstFilter option for
replaceAllMemRefUsesWith; drop forInst->walkOps, add Block::walkOps methods
Eg. output
$ mlir-opt -dma-generate -dma-fast-mem-capacity=1 /tmp/single.mlir
/tmp/single.mlir:9:13: error: Total size of all DMA buffers' for this block exceeds fast memory capacity
for %i3 = (d0) -> (d0)(%i1) to (d0) -> (d0 + 32)(%i1) {
^
$ mlir-opt -debug-only=dma-generate -dma-generate -dma-fast-mem-capacity=400 /tmp/single.mlir
/tmp/single.mlir:9:13: note: 8 KiB of DMA buffers in fast memory space for this block
for %i3 = (d0) -> (d0)(%i1) to (d0) -> (d0 + 32)(%i1) {
PiperOrigin-RevId: 232297044
They are essentially both modelling MLIR OpTrait; the former achieves the
purpose via introducing corresponding symbols in TableGen, while the latter
just uses plain strings.
Unify them to provide a single mechanism to avoid confusion and to better
reflect the definitions on MLIR C++ side.
Ideally we should be able to deduce lots of these traits automatically via
other bits of op definitions instead of manually specifying them; but not
for now though.
PiperOrigin-RevId: 232191401
These attribute kinds are different from the rest in the sense that their types are defined
in MLIR's type hierarchy and we can build constant op out of them.
By defining this middle-level base class, we have a unified way to test and query the type
of these attributes, which will be useful when constructing constant ops of various dialects.
This CL also added asserts to reject non-NumericAttr in constant op's build() method.
PiperOrigin-RevId: 232188178
This CL added a tblgen::DagLeaf wrapper class with several helper methods for handling
DAG arguments. It helps to refactor the rewriter generation logic to be more higher
level.
This CL also added a tblgen::ConstantAttr wrapper class for constant attributes.
PiperOrigin-RevId: 232050683
This CL applies the following simplifications to EDSCs:
1. Rename Block to StmtList because an MLIR Block is a different, not yet
supported, notion;
2. Rework Bindable to drop specific storage and just use it as a simple wrapper
around Expr. The only value of Bindable is to force a static cast when used by
the user to bind into the emitter. For all intended purposes, Bindable is just
a lightweight check that an Expr is Unbound. This simplifies usage and reduces
the API footprint. After playing with it for some time, it wasn't worth the API
cognition overhead;
3. Replace makeExprs and makeBindables by makeNewExprs and copyExprs which is
more explicit and less easy to misuse;
4. Add generally useful functionality to MLIREmitter:
a. expose zero and one for the ubiquitous common lower bounds and step;
b. add support to create already bound Exprs for all function arguments as
well as shapes and views for Exprs bound to memrefs.
5. Delete Stmt::operator= and replace by a `Stmt::set` method which is more
explicit.
6. Make Stmt::operator Expr() explicit.
7. Indexed.indices assertions are removed to pave the way for expressing slices
and views as well as to work with 0-D memrefs.
The CL plugs those simplifications with TableGen and allows emitting a full MLIR function for
pointwise add.
This "x.add" op is both type and rank-agnostic (by allowing ArrayRef of Expr
passed to For loops) and opens the door to spinning up a composable library of
existing and custom ops that should automate a lot of the tedious work in
TF/XLA -> MLIR.
Testing needs to be significantly improved but can be done in a separate CL.
PiperOrigin-RevId: 231982325
This allow for arbitrarily complex builder patterns which is meant to cover initial cases while the modelling is improved and long tail cases/cases for which expanding the DSL would result in worst overall system.
NFC just sorting the emit replace methods alphabetical in the class and file body.
PiperOrigin-RevId: 231890352
A performance issue was reported due to the usage of NestedMatcher in
ComposeAffineMaps. The main culprit was the ubiquitous copies that were
occuring when appending even a single element in `matchOne`.
This CL generally simplifies the implementation and removes one level of indirection by getting rid of
auxiliary storage as well as simplifying the API.
The users of the API are updated accordingly.
The implementation was tested on a heavily unrolled example with
ComposeAffineMaps and is now close in performance with an implementation based
on stateless InstWalker.
As a reminder, the whole ComposeAffineMaps pass is slated to disappear but the bug report was very useful as a stress test for NestedMatchers.
Lastly, the following cleanups reported by @aminim were addressed:
1. make NestedPatternContext scoped within runFunction rather than at the Pass level. This was caused by a previous misunderstanding of Pass lifetime;
2. use defensive assertions in the constructor of NestedPatternContext to make it clear a unique such locally scoped context is allowed to exist.
PiperOrigin-RevId: 231781279
This CL mandated TypeConstraint and Type to provide descriptions and fixed
various subclasses and definitions to provide so. The purpose is to enforce
good documentation; using empty string as the default just invites oversight.
PiperOrigin-RevId: 231579629
* Emitted result lists for ops.
* Changed to allow empty summary and description for ops.
* Avoided indenting description to allow proper MarkDown rendering of
formatting markers inside description content.
* Used fixed width font for operand/attribute names.
* Massaged TensorFlow op docs and generated dialect op doc.
PiperOrigin-RevId: 231427574
Similar to op operands and attributes, use DAG to specify operation's results.
This will allow us to provide names and matchers for outputs.
Also Defined `outs` as a marker to indicate the start of op result list.
PiperOrigin-RevId: 231422455
Python modules cannot be defined under a directory that has a `-` character in its name inside of Google code.
Rename to `google_mlir` which circumvents this limitation.
PiperOrigin-RevId: 231329321
Update to allow constant attribute values to be used to match or as result in rewrite rule. Define variable ctx in the matcher to allow matchers to refer to the context of the operation being matched.
PiperOrigin-RevId: 231322019
This CL adds support for calling EDSCs from other languages than C++.
Following the LLVM convention this CL:
1. declares simple opaque types and a C API in mlir-c/Core.h;
2. defines the implementation directly in lib/EDSC/Types.cpp and
lib/EDSC/MLIREmitter.cpp.
Unlike LLVM however the nomenclature for these types and API functions is not
well-defined, naming suggestions are most welcome.
To avoid the need for conversion functions, Types.h and MLIREmitter.h include
mlir-c/Core.h and provide constructors and conversion operators between the
mlir::edsc type and the corresponding C type.
In this first commit, mlir-c/Core.h only contains the types for the C API
to allow EDSCs to work from Python. This includes both a minimal set of core
MLIR
types (mlir_context_t, mlir_type_t, mlir_func_t) as well as the EDSC types
(edsc_mlir_emitter_t, edsc_expr_t, edsc_stmt_t, edsc_indexed_t). This can be
restructured in the future as concrete needs arise.
For now, the API only supports:
1. scalar types;
2. memrefs of scalar types with static or symbolic shapes;
3. functions with input and output of these types.
The C API is not complete wrt ownership semantics. This is in large part due
to the fact that python bindings are written with Pybind11 which allows very
idiomatic C++ bindings. An effort is made to write a large chunk of these
bindings using the C API but some C++isms are used where the design benefits
from this simplication. A fully isolated C API will make more sense once we
also integrate with another language like Swift and have enough use cases to
drive the design.
Lastly, this CL also fixes a bug in mlir::ExecutionEngine were the order of
declaration of llvmContext and the JIT result in an improper order of
destructors (which used to crash before the fix).
PiperOrigin-RevId: 231290250
Similar to other tblgen:: abstractions, tblgen::Pattern hides the native TableGen
API and provides a nicer API that is more coherent with the TableGen definitions.
PiperOrigin-RevId: 231285143
* Matching an attribute and specifying a attribute constraint is the same thing executionally, so represent it such.
* Extract AttrConstraint helper to match TypeConstraint and use that where mAttr was previously used in RewriterGen.
PiperOrigin-RevId: 231213580
Addresses b/122486036
This CL addresses some leftover crumbs in AffineMap and IntegerSet by removing
the Null method and cleaning up the constructors.
As the ::Null uses were tracked down, opportunities appeared to untangle some
of the Parsing logic and make it explicit where AffineMap/IntegerSet have
ambiguous syntax. Previously, ambiguous cases were hidden behind the implicit
pointer values of AffineMap* and IntegerSet* that were passed as function
parameters. Depending the values of those pointers one of 3 behaviors could
occur.
This parsing logic convolution is one of the rare cases where I would advocate
for code duplication. The more proper fix would be to make the syntax
unambiguous or to allow some lookahead.
PiperOrigin-RevId: 231058512
This CL follows up on a memory leak issue related to SmallVector growth that
escapes the BumpPtrAllocator.
The fix is to properly use ArrayRef and placement new to define away the
issue.
The following renaming is also applied:
1. MLFunctionMatcher -> NestedPattern
2. MLFunctionMatches -> NestedMatch
As a consequence all allocations are now guaranteed to live on the BumpPtrAllocator.
PiperOrigin-RevId: 231047766
- Update createAffineComputationSlice to generate a sequence of single result
affine apply ops instead of one multi-result affine apply
- update pipeline-data-transfer test case; while on this, also update the test
case to use only single result affine maps, and make it more robust to
change.
PiperOrigin-RevId: 230965478
This commit introduces a generic dialect conversion/lowering/legalization pass
and illustrates it on StandardOps->LLVMIR conversion.
It partially reuses the PatternRewriter infrastructure and adds the following
functionality:
- an actual pass;
- non-default pattern constructors;
- one-to-many rewrites;
- rewriting terminators with successors;
- not applying patterns iteratively (unlike the existing greedy rewrite driver);
- ability to change function signature;
- ability to change basic block argument types.
The latter two things required, given the existing API, to create new functions
in the same module. Eventually, this should converge with the rest of
PatternRewriter. However, we may want to keep two pass versions: "heavy" with
function/block argument conversion and "light" that only touches operations.
This pass creates new functions within a module as a means to change function
signature, then creates new blocks with converted argument types in the new
function. Then, it traverses the CFG in DFS-preorder to make sure defs are
converted before uses in the dominated blocks. The generic pass has a minimal
interface with two hooks: one to fill in the set of patterns, and another one
to convert types for functions and blocks. The patterns are defined as
separate classes that can be table-generated in the future.
The LLVM IR lowering pass partially inherits from the existing LLVM IR
translator, in particular for type conversion. It defines a conversion pattern
template, instantiated for different operations, and is a good candidate for
tablegen. The lowering does not yet support loads and stores and is not
connected to the translator as it would have broken the existing flows. Future
patches will add missing support before switching the translator in a single
patch.
PiperOrigin-RevId: 230951202
This CL adds a new marker, replaceWithValue, to indicate that no new result
op is generated by applying a pattern. Instead, the matched DAG is replaced
by an existing SSA value.
Converted the tf.Identity converter to use the pattern.
PiperOrigin-RevId: 230922323
This implements a simple CPU runner based on LLVM Orc JIT. The base
functionality is provided by the ExecutionEngine class that compiles and links
the module, and provides an interface for obtaining function pointers to the
JIT-compiled MLIR functions and for invoking those functions directly. Since
function pointers need to be casted to the correct pointer type, the
ExecutionEngine wraps LLVM IR functions obtained from MLIR into a helper
function with the common signature `void (void **)` where the single argument
is interpreted as a list of pointers to the actual arguments passed to the
function, eventually followed by a pointer to the result of the function.
Additionally, the ExecutionEngine is set up to resolve library functions to
those available in the current process, enabling support for, e.g., simple C
library calls.
For integration purposes, this also provides a simplistic runtime for memref
descriptors as expected by the LLVM IR code produced by MLIR translation. In
particular, memrefs are transformed into LLVM structs (can be mapped to C
structs) with a pointer to the data, followed by dynamic sizes. This
implementation only supports statically-shaped memrefs of type float, but can
be extened if necessary.
Provide a binary for the runner and a test that exercises it.
PiperOrigin-RevId: 230876363
- introduce a way to compute union using symbolic rectangular bounding boxes
- handle multiple load/store op's to the same memref by taking a union of the regions
- command-line argument to provide capacity of the fast memory space
- minor change to replaceAllMemRefUsesWith to not generate affine_apply if the
supplied index remap was identity
PiperOrigin-RevId: 230848185
canonicalizations of operations. The ultimate important user of this is
going to be a funcBuilder->foldOrCreate<YourOp>(...) API, but for now it
is just a more convenient way to write certain classes of canonicalizations
(see the change in StandardOps.cpp).
NFC.
PiperOrigin-RevId: 230770021
Example inline notation:
trailing-location ::= 'loc' '(' location ')'
// FileLineCol Location.
%1 = "foo"() : () -> i1 loc("mysource.cc":10:8)
// Name Location
return loc("foo")
// CallSite Location
return loc(callsite("foo" at "mysource.cc":19:9))
// Fused Location
/// Without metadata
func @inline_notation() loc(fused["foo", "mysource.cc":10:8])
/// With metadata
return loc(fused<"myPass">["foo", "foo2"])
// Unknown location.
return loc(unknown)
Locations are currently only printed with inline notation at the line of each instruction. Further work is needed to allow for reference notation, e.g:
...
return loc 1
}
...
loc 1 = "source.cc":10:1
PiperOrigin-RevId: 230587621
This CL just changes various docs and comments to use the term "generic" and
"custom" when mentioning assembly forms. To be consist, several methods are
also renamed:
* FunctionParser::parseVerboseOperation() -> parseGenericOperation()
* ModuleState::hasShorthandForm() -> hasCustomForm()
* OpAsmPrinter::printDefaultOp() -> printGenericOp()
PiperOrigin-RevId: 230568819
- update fusion cost model to fuse while tolerating a certain amount of redundant
computation; add cl option -fusion-compute-tolerance
evaluate memory footprint and intermediate memory reduction
- emit debug info from -loop-fusion showing what was fused and why
- introduce function to compute memory footprint for a loop nest
- getMemRefRegion readability update - NFC
PiperOrigin-RevId: 230541857
Nicolas Vasilache