[mlir] Added support for symbols inside linalg.generic and map concatenation

[limo1996]

Summary

This commit adds functionality needed for implementation of convolutions with
linalg.generic op. Since linalg.generic right now expects indexing maps to be
just permutations, offset indexing needed in convolutions is not possible.
Therefore in this commit we address the issue by adding support for symbols inside
indexing maps which enables more advanced indexing. The upcoming commit will
solve the problem of computing loop bounds from such maps.

This commit is a fold of these commits:

[mlir] Added support for symbols inside linalg.generic indexing maps

---

[mlir] Added support for the shift of the symbols inside AffineExpr

---

[mlir] Added support for symbols in map concatenation

[limo1996]

Fixed clang-tidy warning

[ftynse]

Thanks, a couple of comments

mlir/lib/Dialect/Linalg/Transforms/Loops.cpp

39

Please don't commit commented-out code

173

Please don't use auto unless if it improves readability (e.g. for very long types such as iterators, or when the type is already mentioned in the same statement such as getAttrOfType).

177

Call .reserve to allocate space in the vector before calling push_back in a loop.

178

Can you just do for (unsigned i = 0, e = shapedType.getRank(); i < e; ++i) instead of enumerating the shape and using only the indices?

mlir/test/Dialect/Linalg/loops.mlir

17

Nit: the expression doesn't seem to correspond to a _strided_ convolution

[ftynse]

mlir/test/Dialect/Linalg/loops.mlir

949

Please don't pattern-match on SSA value names (%c0), they are not guaranteed to be stable.

953

Hmm, would it be possible to put this operation before the loops (actually, there is one already, many we can just reuse its value)?

[limo1996]

Most of the comments incorporated

[ftynse]

mlir/lib/IR/AffineMap.cpp

394

This function is still not supposed to work for maps with symbols IIUC

410

Is this change necessary?

422–425

There are more of these, I only commented on the first occurrence.

[limo1996]

mlir/lib/IR/AffineMap.cpp

410

When numSymbols == 0 then numDims == numInputs but yeah no need for the change..

[limo1996]

removed commented code

[limo1996]

mlir/lib/IR/AffineMap.cpp

394

At this point it should. The next diff introduces function that bypasses inversePermutation call

410

Actually at this point there is a need for it. I will revert it in the next commit..

mlir/test/Dialect/Linalg/loops.mlir

949

Hmm I saw them in other tests so I just used them.. When naming convention changes other tests will need changes as well..

953

Hmm I think --cse solves it right?

[ftynse]

mlir/lib/IR/AffineMap.cpp

394

I see two bad practices here:

• trying to commit commented-out code: if the assertion is no longer necessary, remove it completely, add a test that exercises the new behavior, and update the doc;
• commits are not self-contained: if this change is only required for the next commit, it should be included in that commit and removed from this one.

mlir/test/Dialect/Linalg/loops.mlir

949

The fact that other tests currently contradict the testing guide https://mlir.llvm.org/getting_started/TestingGuide/ because of legacy (they likely existed before the current convention was adopted) does not mean you are allowed to commit new code that also contradicts the testing guide. If you notice such tests, the proper hygiene is to update them as well, in a separate commit.

There is _no_ naming convention for SSA names. The only convention is that SSA names can change at any time without any warning, that's why tests should not rely on them. If you disagree with the guide, feel free to start a discussion on the forum.

MLIR currently has ~400 test input files totalling ~70k LoC. Updating matches for SSA names in all of them would be extremely painful, will likely take days or weeks even with automation and require multiple rounds because other commits will be editing tests concurrently. This sounds like extremely poor use of engineering time, compared to several minutes spent writing the proper match conditions for each new commit.

953

If it is simple to do in your code, you should do it instead of expecting the caller to run CSE after the fact. Otherwise, we risk getting into a situation where each pass requires five other cleanup passes to run, each of which also require five other cleanup passes and so on. If doing so feels like you start reimplementing a generic CSE, then you shouldn't do it. It's a trade-off that must be considered rather than shrugged off.

[nicolasvasilache]

mlir/include/mlir/Dialect/Linalg/IR/LinalgStructuredOps.td

560

Can we increase the description and add IR examples ?
In particular, if symbol_source is specified, it seems we need a symbol for every single dim of the corresponding operand?

What happens if the operand is partially static?
Atm it seems we add symbols for everything.
Will partially static operands break something down the line? (If so please at least add a TODO).

Additionally, it seems like symbol_source should be a proper ODS Optional Attribute this way some of the code you write would be auto generated?

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

267

I think most of this could go away with a proper confined integer attr (e.g. https://github.com/llvm/llvm-project/blob/master/mlir/include/mlir/Dialect/StandardOps/IR/Ops.td#L1983)

272

It seems the logic behind this targetRank and its usage should be hidden behind a op.getNumSymbolsPerOperand() that would return a SmallVector<unsigned> where each operand index has the number of symbols.
This would also be somewhat future proof if we decided we wanted to turn the the symbol source into an ArrayAttr

283–285

this seems incorrect as all indexing maps could have the number of symbols of operand X.
If you had the helper above you could check m.getNumSymbols() != 0 && m.getNumSymbols() != op.getNumSymbolsPerOperand()[idx]

mlir/lib/IR/AffineExpr.cpp

101

Drop dims above and make this

return replaceDimsAndSymbols({}, symbols);

?

mlir/lib/IR/AffineMap.cpp

394

I don't see how this is correct.
Bypassing a caller has no incidence on potential other callers and the code would be bugged.

Let's revert the 2 changes to this function please.

410

If it is not necessary let's revert is now, I don't see value in making a change that is reversed in a subsequent commit without a solid justification.

[limo1996]

Changed static addressing of constants in tests. shiftSymbols function modified as Nicolas suggested.

[limo1996]

Documentation for symbol_source attribute extended

[limo1996]

mlir/include/mlir/Dialect/Linalg/IR/LinalgStructuredOps.td

560

Documentation updated..

In particular, if symbol_source is specified, it seems we need a symbol for every single dim of the corresponding operand?

Yes we do.

What happens if the operand is partially static?

It behaves the same as if it was dynamic. Do you think it should behave differently? i.e. to not require a symbol for static dimension?

Will partially static operands break something down the line?

I don't think so

Additionally, it seems like symbol_source should be a proper ODS Optional Attribute this way some of the code you write would be auto generated?

Yes there is a separate WIP revision for that. Once it is approaved I will merge it into this one.

[limo1996]

problem with commented code inside inversePermutation solved

[limo1996]

mlir/test/Dialect/Linalg/loops.mlir

953

I will do it in the follow up commit

[limo1996]

revision D83378 merged: [mlir][WIP] symbol_source attr in linalg.generic proposal

[limo1996]

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

249

@nicolasvasilache @ftynse I found the function name misleading as not only GenericOp and IndexedGenericOp are passed but also LinalgStructuredOps which does not allow me to call functions defined in GenericOpBase. Should I circumvent it somehow or just avoid functions I defined in GenericOpBase?

272

If we turn symbol source into ArrayAttr then we must either pass dimensions from all operands specified in the symbol source into the maps as symbols or we will pass only dimensions of the corresponding operand which is really limited and does not support our case when we need dimensions of the operand in the map that does not correspond to it.

283–285

Yes, all maps can have the number of symbols of operand X. Is it wrong?

If you had the helper above you could check m.getNumSymbols() != 0 && m.getNumSymbols() != op.getNumSymbolsPerOperand()[idx]

This condition basically allows only map at index equal to symbol_source to have symbols which is not correct as in our scenario we want to have symbols in the map corresponding to the input view and not the kernel view.

[ftynse]

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

249

If you can avoid new ops, just do that. Possibly you can have a verifier in the GenericOp that calls this function, and then does some additional things. Otherwise, it's template-parameterized by the actual OpType, so it should be possible to use template specialization to work around missing functions
.

272

A helper function looks reasonable from the code complexity perspective regardless of the further evolution. ArrayAttr discussion is irrelevant for the current patch.

[limo1996]

evolve

[ftynse]

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

298–299

Please don't put usernames after TODO in LLVM codebase.

mlir/lib/Dialect/Linalg/Transforms/Loops.cpp

471

Please use full sentences in comments: start with a capital letter and terminate with a period, https://llvm.org/docs/CodingStandards.html#commenting.

[ftynse]

Works for me after the last comments are addressed.

[nicolasvasilache]

Very sorry for the delay, thanks Jakub!

One more question inline re the dropping of symbols that seems to look a bit fishy.

Thanks for pushing this!

mlir/include/mlir/Dialect/Linalg/IR/LinalgStructuredOps.td

489

thanks!

560

sounds good, thanks!

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

249

This is mostly related to better OpInterface definition.
This is good for now, I'll refactor and cleanup a bit at some later time, thanks!

272

The comment was about code hygiene, not suggesting we should discuss this extension right now.
Basically l 264-272 is an implementation detail that should be hidden away in a helper function today.
If tomorrow the logic grows this new helper function will be the right place to impl this.

283–285

this seems incorrect as all indexing maps could have the number of symbols of operand X.

I was thinking about the case where we have >1 operands that want to propagate their symbols.
This falls into the extension to ArrayAttr that is better disregarded for now.

283–285

This condition basically allows only map at index equal to symbol_source to have symbols which is not correct as in our scenario we want to have symbols in the map corresponding to the input view and not the kernel view.

Ah I see, thanks for explaining, the symbols of the kernel are indeed used in the input.
Then I'd recommend making things homogeneous from the get go.
Instead of

#conv_1d_accesses = [
  affine_map<(m, n)[s0] -> (m + n - s0 floordiv 2)>, // in
  affine_map<(m, n) -> (n)>, // filter
  affine_map<(m, n) -> (m)> // out
]

let's use:

#conv_1d_accesses = [
  affine_map<(m, n)[s0] -> (m + n - s0 floordiv 2)>, // in
  affine_map<(m, n)[s0] -> (n)>, // filter
  affine_map<(m, n)[s0] -> (m)> // out
]

?

The fact that only the first operand map had [s0] had me think that s0 corresponds to the operand 0 (even though it clearly says 1).

<don't consider for now>
and so in the future, when we want symbols from multiple operands, we can just concatenate them all and still remain homogeneous.
</don't consider for now>

mlir/lib/Dialect/Linalg/Transforms/Loops.cpp

473

This looks quite unsafe to me.
It seems the proper way to drop symbols would be to replace them all with 0.

Is this just a temporary thing that needs to be fixed in the next CL and the state of the codebase is that incorrect code may be generated?

[ftynse]

mlir/lib/Dialect/Linalg/Transforms/Loops.cpp

473

Good point. Can we just return {} to signify failure in presence of symbols? The tests for generated loops can go into the following commit where the loop bound computation is added. Here, we can just keep the "roundtripping" test that makes sure we print back what we parsed.

[limo1996]

All comments resolved

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

249

Ok let me know once it's refactored so I can simplify my code as well. Thanks :)

Code starting at line 264 can then be simplified to:

auto symSrc = op.getSymbolSource();
if (symSrc.hasValue() && symSrc.getValue() >= op.getNumOperands())
  return op.emitOpError("symbol_source index out of range");

272

I think in the future once this function is refactored such that it takes only GenericOp and IndexedGenericOp we can introduce op.getNumSymbols() that would return sum of ranks of operands specified in symbol_source and lines 264-269 can be refactored like this:

auto symSrc = op.getSymbolSource();
if (symSrc.hasValue() && symSrc.getValue() >= op.getNumOperands())
  return op.emitOpError("symbol_source index out of range");

For now I refactored only line 271.

283–285

Yeah exactly, we are on the same page. I added s0 to every map to make it less confusing and also I enforce it now. Thanks for the feedback!

[limo1996]

All comments resolved

[nicolasvasilache]

mlir/include/mlir/Dialect/Linalg/Utils/Utils.h

123 ↗

(On Diff #278770)

you should be able to just do:

linalgOp.indexing_maps().template getAsRange<AffineMapAttr, AffineMap>()

and drop the extra function.

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

272

Ok I see how the fact that it is use on Generic, IndexedGeneric and on the op interface makes it less attractive to refactor now.
I'll fix, thanks!

283–285

Looks great, thanks!

[limo1996]

Last comment of Nicolas resolved + found out one test is not passing which was caused by quite significant bug so fixed that as well

mlir/include/mlir/Dialect/Linalg/IR/LinalgStructuredOps.td

 class GenericOpBase<string mnemonic> : LinalgStructuredBase_Op<mnemonic,
     [SingleBlockImplicitTerminator<"YieldOp">]> {
   let arguments = (ins Variadic<LinalgOperand>:$views,
                    I64Attr:$args_in,
                    I64Attr:$args_out,
                    AffineMapArrayAttr:$indexing_maps,
                    ArrayAttr:$iterator_types,
                    OptionalAttr<StrAttr>:$doc,
-                   OptionalAttr<StrAttr>:$library_call);
+                   OptionalAttr<StrAttr>:$library_call,
+                   Confined<OptionalAttr<I64Attr>,
+                     [IntMinValue<0>]>:$symbol_source);
   let results = (outs Variadic<AnyRankedTensor>:$output_tensors);
   let regions = (region AnyRegion:$region);
   let extraClassDeclaration = [{
     SmallVector<StringRef, 8> linalgTraitAttrNames() {
       return SmallVector<StringRef, 8>{
         getArgsInAttrName(), getArgsOutAttrName(), getDocAttrName(),
         getIndexingMapsAttrName(), getLibraryCallAttrName(),
-        getIteratorTypesAttrName()
+        getIteratorTypesAttrName(), getSymbolSourceAttrName()
       };
     }
 
     unsigned getNumInputs() { return args_in().getSExtValue(); }
 
     unsigned getNumOutputs() { return args_out().getSExtValue(); }
 
     StringRef getLibraryCallName() {
       return library_call().hasValue() ? library_call().getValue() : "";
     }
 
     llvm::Optional<SmallVector<StringRef, 8>> referenceIterators() {
       llvm_unreachable(
         "No such thing as reference iterator types for a generic op.");
-     }
 
     llvm::Optional<SmallVector<AffineMap, 8>> referenceIndexingMaps() {
       llvm_unreachable(
         "No such thing as reference indexing maps for a generic op.");
-     }
+
+    llvm::Optional<int> getSymbolSource() {
+      auto ss = symbol_source();
+      return ss.hasValue() ?
+        llvm::Optional<int>(ss.getValue().getLimitedValue()) : llvm::None;
+    }
   }];
 
   let printer = [{ return ::print(p, *this); }];
   let parser = [{ return ::parseGenericOp(parser, result); }];
 }
 
 /// Index-free GenericOp.
 def GenericOp : GenericOpBase<"generic"> {
Where #trait_attributes is an alias of a dictionary attribute containing:
         external library function that the linalg.generic operation maps to.
         The external library is assumed to be dynamically linked and no strong
         compile-time guarantees are provided. In the absence of such a library
         call, linalg.generic will always lower to loops.
       - iterator_types: an ArrayAttr specifying the type of the enclosing loops.
         Each element of the list represents and iterator of one of the following
         types:
           parallel, reduction, window
+      - symbol_source: index of the operand whose dimensions will be propagated
+        as symbols to the indexing maps. When specified the number of symbols
+        in each of the indexing maps has to be either 0 or the rank of the
+        specified operand.
 
     Example:
     Defining a #matmul_trait attribute in MLIR can be done as follows:
       ```mlir
       #matmul_accesses = [
         (m, n, k) -> (m, k),
         (m, n, k) -> (k, n),
         (m, n, k) -> (m, n)
let description = [{
     transformations can be applied. Such legalization moves tensor return values
     into output buffer operands and updates the region arguments accordingly.
 
     Transformations that create control-flow around linalg.indexed_generic
     operations are not expected to work with tensors because SSA values do not
     escape naturally. Still, transformations and rewrites that take advantage of
     tensor SSA values are expected to be useful and will be added in the near
     future.
+
+    Example of 1D convolution with symbols:
+    ```mlir
+    #conv_1d_accesses = [
+      affine_map<(m, n)[s0] -> (m + n - s0 floordiv 2)>, // in
+      affine_map<(m, n) -> (n)>, // filter
+      affine_map<(m, n) -> (m)> // out
+    ]
+
+    #conv_1d_trait = {
+      doc = "O(m) += I(m + n - size(n) floordiv 2) * K(n)",
+      indexing_maps = #conv_1d_accesses,
+      library_call  = "linalg_conv_1d",
+      iterator_types = ["parallel", "parallel"],
+      symbol_source = 1
+    }
+
+    linalg.generic #conv_1d_trait %in, %filter, %out {
+      ^bb0(%a: f32, %b: f32, %c: f32) :
+        %d = mulf %a, %b : f32
+        %e = addf %c, %d : f32
+        linalg.yield %e : f32
+    } : memref<?xf32>,
+        memref<?xf32>,
+        memref<?xf32>
+    ```
+    where symbol s0 will be substituted with `dim %filter, %c0` i.e. the first
+    and only dimension of the second operand as specified by the symbol_source
+    attribute. Note that maps requiring no symbols does not need to specify
+    them.
   }];
 
   let builders = [
     OpBuilder<
       "OpBuilder &builder, OperationState &result, ArrayRef<Type> resultTypes, "
       "ValueRange args, int64_t inputCount, int64_t outputCount, "
       "ArrayRef<AffineMap> indexingMaps, ArrayRef<StringRef> iteratorTypes, "
       "function_ref<void(OpBuilder &, Location, ValueRange)> = nullptr">

mlir/include/mlir/Dialect/Utils/StructuredOpsUtils.h

inline bool isColumnMajorMatmul(ArrayAttr indexingMaps) {
   bindDims(context, m, n, k);
   auto mapA = AffineMapAttr::get(AffineMap::get(3, 0, {k, n}, context));
   auto mapB = AffineMapAttr::get(AffineMap::get(3, 0, {m, k}, context));
   auto mapC = AffineMapAttr::get(AffineMap::get(3, 0, {n, m}, context));
   auto maps = ArrayAttr::get({mapA, mapB, mapC}, context);
   return indexingMaps == maps;
 }
 
+/// Attribute name for the IntegerAttr which encodes the index of operand
+/// whose dimensions will be propagated as symbols to the indexing maps
+constexpr StringRef getSymbolSourceAttrName() { return "symbol_source"; }
+
 /// Attribute name for the AffineArrayAttr which encodes the relationship
 /// between a structured op iterators' and its operands.
 constexpr StringRef getIndexingMapsAttrName() { return "indexing_maps"; }
 
 /// Attribute name for the StrArrayAttr which encodes the type of a structured
 /// op's iterators.
 constexpr StringRef getIteratorTypesAttrName() { return "iterator_types"; }
 

mlir/include/mlir/IR/AffineExpr.h

public:
   /// Walk all of the AffineExpr's in this expression in postorder.
   void walk(std::function<void(AffineExpr)> callback) const;
 
   /// This method substitutes any uses of dimensions and symbols (e.g.
   /// dim#0 with dimReplacements[0]) and returns the modified expression tree.
   AffineExpr replaceDimsAndSymbols(ArrayRef<AffineExpr> dimReplacements,
                                    ArrayRef<AffineExpr> symReplacements) const;
 
+  /// Replace symbols[0 .. numDims - 1] by
+  /// symbols[shift .. shift + numDims - 1].
+  AffineExpr shiftSymbols(unsigned numSymbols, unsigned shift) const;
+
   AffineExpr operator+(int64_t v) const;
   AffineExpr operator+(AffineExpr other) const;
   AffineExpr operator-() const;
   AffineExpr operator-(int64_t v) const;
   AffineExpr operator-(AffineExpr other) const;
   AffineExpr operator*(int64_t v) const;
   AffineExpr operator*(AffineExpr other) const;
   AffineExpr floorDiv(uint64_t v) const;

mlir/lib/Dialect/Linalg/EDSC/Builders.cpp

auto *op =
               edsc::ScopedContext::getLocation(),
               types,
               values,
               IntegerAttr::get(IntegerType::get(64, ctx), nInputs),
               IntegerAttr::get(IntegerType::get(64, ctx), nOutputs),
               builder.getAffineMapArrayAttr(maps),
               builder.getStrArrayAttr(iteratorStrTypes),
               StringAttr() /*doc*/,
-              StringAttr() /*library_call*/
+              StringAttr() /*library_call*/,
+              IntegerAttr() /*symbol_source*/
               /* TODO: other attributes in op */
               )
           .getOperation();
   // clang-format on
 
   using namespace edsc;
   SmallVector<Type, 4> blockTypes;
   blockTypes.reserve(values.size());

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

void GenericOp::build(
     ValueRange args, int64_t inputCount, int64_t outputCount,
     ArrayRef<AffineMap> indexingMaps, ArrayRef<StringRef> iteratorTypes,
     function_ref<void(OpBuilder &, Location, ValueRange)> bodyBuild) {
   build(builder, result, resultTypes, args,
         builder.getI64IntegerAttr(inputCount),
         builder.getI64IntegerAttr(outputCount),
         builder.getAffineMapArrayAttr(indexingMaps),
         builder.getStrArrayAttr(iteratorTypes),
-        /*doc=*/nullptr, /*library_call=*/nullptr);
+        /*doc=*/nullptr, /*library_call=*/nullptr,
+        /*symbol_source=*/nullptr);
   if (!bodyBuild)
     return;
 
   SmallVector<Type, 4> blockArgTypes;
   for (Value arg : args)
     blockArgTypes.push_back(arg.getType().cast<ShapedType>().getElementType());
 
   OpBuilder::InsertionGuard guard(builder);
   auto &region = *result.regions.front();
   Block *bodyBlock = builder.createBlock(&region, region.end(), blockArgTypes);
   bodyBuild(builder, result.location, bodyBlock->getArguments());
 }
 
 void IndexedGenericOp::build(
     OpBuilder &builder, OperationState &result, ArrayRef<Type> resultTypes,
     ValueRange args, int64_t inputCount, int64_t outputCount,
     ArrayRef<AffineMap> indexingMaps, ArrayRef<StringRef> iteratorTypes,
     function_ref<void(OpBuilder &, Location, ValueRange, ValueRange)>
         bodyBuild) {
   build(builder, result, resultTypes, args,
         builder.getI64IntegerAttr(inputCount),
         builder.getI64IntegerAttr(outputCount),
         builder.getAffineMapArrayAttr(indexingMaps),
         builder.getStrArrayAttr(iteratorTypes),
-        /*doc=*/nullptr, /*library_call=*/nullptr);
+        /*doc=*/nullptr, /*library_call=*/nullptr,
+        /*symbol_source=*/nullptr);
   if (!bodyBuild)
     return;
 
   unsigned nLoops = iteratorTypes.size();
   SmallVector<Type, 4> blockArgTypes(nLoops, builder.getIndexType());
   for (Value arg : args)
     blockArgTypes.push_back(arg.getType().cast<ShapedType>().getElementType());
 
if (viewType.getElementType() !=
              << ((i < nInputViews) ? "input " : "output ")
              << "operand: " << viewType;
   }
   return success();
 }
 } // namespace
 
 template <typename GenericOpType>
 static LogicalResult verifyGenericOp(GenericOpType op) {
   auto nInputViews = op.getNumInputs();
   auto nLoops = op.getNumLoops();
   auto nInputsAndOutputBuffers = op.getNumInputsAndOutputBuffers();
   if (nInputsAndOutputBuffers != llvm::size(op.views()))
     return op.emitOpError("expected exactly ")
            << nInputsAndOutputBuffers
            << " inputs (tensor or buffer) and output buffer operands";
 
   auto &region = op.region();
   if (!llvm::hasSingleElement(region))
     return op.emitOpError("expected region with 1 block");
   if (failed(BlockArgsVerifier<GenericOpType>::verify(op, region.front())))
     return failure();
 
+  auto attr = op.template getAttrOfType<IntegerAttr>("symbol_source");
+  int64_t targetRank = 0;
+  if (attr) {
+    unsigned index = attr.getInt();
+    if (index >= op.getNumOperands())
+      return op.emitOpError("symbol_source index out of range");
+    targetRank =
+        op.getOperand(index).getType().template cast<ShapedType>().getRank();
+  }
+
   SmallVector<AffineMap, 4> indexingMaps;
   indexingMaps.reserve(op.indexing_maps().size());
   for (auto en : llvm::enumerate(op.indexing_maps())) {
     auto idx = en.index();
     auto m = en.value().template cast<AffineMapAttr>().getValue();
     indexingMaps.push_back(m); // Save reference to map for further checks.
     auto view = (idx < nInputViews) ? op.getInputShapedType(idx)
                                     : op.getOutputShapedType(idx - nInputViews);
 
-    if (m.getNumSymbols() != 0)
-      return op.emitOpError("expected indexing_map #")
-             << idx << " to have no symbols";
+    if (m.getNumSymbols() != targetRank && m.getNumSymbols() != 0)
+      return op.emitOpError("expected the number of symbols in indexing_map #")
+             << idx << " to either match target rank or to be 0";
 
     if (m.getNumDims() != nLoops)
       return op.emitOpError("expected indexing_map #")
              << idx << " to have " << nLoops
              << " dim(s) to match the number of loops";
 
     if (m.getNumResults() != view.getRank())
       return op.emitOpError("expected indexing_map #")
              << idx << " results to match view rank: " << view;
   }
 
   auto concatMap = concatAffineMaps(indexingMaps);
-  auto aggregateMap = inversePermutation(concatMap);
-  if (!aggregateMap)
+  // TODO(limo): Bound inference for maps with symbols
+  if (!concatMap.getNumSymbols() && !inversePermutation(concatMap))
     return op.emitOpError("expected the concatenation of maps in indexing_map "
                           "to be invertible");
 
   return success();
 }
 
 static LogicalResult verify(GenericOp op) { return verifyGenericOp(op); }
 static LogicalResult verify(IndexedGenericOp op) { return verifyGenericOp(op); }

mlir/lib/Dialect/Linalg/Transforms/DropUnitDims.cpp

LogicalResult matchAndRewrite(GenericOp genericOp,
     for (unsigned i : llvm::seq<unsigned>(0, genericOp.getNumResults()))
       resultTypes.push_back(
           newInputOutputTypes[i + genericOp.getNumOperands()]);
     GenericOp replacementOp = rewriter.create<GenericOp>(
         loc, resultTypes, newOperands, genericOp.args_in(),
         genericOp.args_out(), rewriter.getAffineMapArrayAttr(newIndexingMaps),
         genericOp.iterator_types(),
         /*doc = */ nullptr,
-        /*library_call = */ nullptr);
+        /*library_call = */ nullptr,
+        /*symbol_source = */ nullptr);
     rewriter.inlineRegionBefore(genericOp.region(), replacementOp.region(),
                                 replacementOp.region().begin());
 
     // If any result tensor has a modified shape, then add reshape to recover
     // the original shape.
     SmallVector<Value, 4> resultReplacements;
     for (auto result : llvm::enumerate(replacementOp.getResults())) {
       unsigned index = result.index() + replacementOp.getNumOperands();

mlir/lib/Dialect/Linalg/Transforms/Fusion.cpp

if (isa<GenericOp>(producer.getOperation()) &&
                   rewriter.getUnknownLoc(),
                   consumer.getOperation()->getResultTypes(), fusedOperands,
                   rewriter.getI64IntegerAttr(fusedOperands.size()),
                   rewriter.getI64IntegerAttr(
                       consumer.getOperation()->getNumResults()),
                   rewriter.getArrayAttr(fusedIndexMaps),
                   consumer.iterator_types(),
                   /*doc=*/nullptr,
-                  /*library_call=*/nullptr)
+                  /*library_call=*/nullptr,
+                  /*symbol_source=*/nullptr)
               .getOperation();
     } else {
       fusedOp =
           rewriter
               .create<IndexedGenericOp>(
                   rewriter.getUnknownLoc(),
                   consumer.getOperation()->getResultTypes(), fusedOperands,
                   rewriter.getI64IntegerAttr(fusedOperands.size()),
                   rewriter.getI64IntegerAttr(
                       consumer.getOperation()->getNumResults()),
                   rewriter.getArrayAttr(fusedIndexMaps),
                   consumer.iterator_types(),
                   /*doc=*/nullptr,
-                  /*library_call=*/nullptr)
+                  /*library_call=*/nullptr,
+                  /*symbol_source=*/nullptr)
               .getOperation();
     }
 
     // Construct an AffineMap from consumer loops to producer loops.
     // consumer loop -> tensor index
     AffineMap consumerResultIndexMap =
         consumer.getInputIndexingMap(consumerIdx);
     // producer loop -> tensor index
SmallVector<Attribute, 4> indexMapAttrs = llvm::to_vector<4>(
           return AffineMapAttr::get(map);
         }));
     auto fusedOp = rewriter.create<LinalgOpTy>(
         rewriter.getUnknownLoc(), consumer.getResultTypes(), fusedOperands,
         rewriter.getI64IntegerAttr(fusedOperands.size()),
         rewriter.getI64IntegerAttr(consumer.getNumResults()),
         rewriter.getArrayAttr(indexMapAttrs), consumer.iterator_types(),
         /*doc=*/nullptr,
-        /*library_call=*/nullptr);
+        /*library_call=*/nullptr,
+        /*symbol_source=*/nullptr);
     auto &fusedRegion = fusedOp.region();
     rewriter.cloneRegionBefore(consumer.region(), fusedRegion,
                                fusedRegion.begin());
     return fusedOp;
   }
 };
 
 /// Implementation of fusion on tensor ops when consumer is a TensorReshapeOp.
static Operation *fuse(LinalgOpTy producer, TensorReshapeOp consumer,
 
     auto fusedOp = rewriter.create<LinalgOpTy>(
         rewriter.getUnknownLoc(), consumer.getResultType(),
         producer.getOperands(),
         rewriter.getI64IntegerAttr(producer.getNumOperands()),
         rewriter.getI64IntegerAttr(1), rewriter.getArrayAttr(indexMapAttrs),
         producer.iterator_types(),
         /*doc=*/nullptr,
-        /*library_call=*/nullptr);
+        /*library_call=*/nullptr,
+        /*symbol_source=*/nullptr);
     auto &fusedRegion = fusedOp.region();
     rewriter.cloneRegionBefore(producer.region(), fusedRegion,
                                fusedRegion.begin());
     return fusedOp;
   }
 };
 
 /// Implementation of fusion on tensor ops when producer is a splat constant.
static Operation *fuse(ConstantOp producer, LinalgOpTy consumer,
 
     auto fusedOp = rewriter.create<LinalgOpTy>(
         rewriter.getUnknownLoc(), consumer.getResultTypes(), fusedOperands,
         rewriter.getI64IntegerAttr(consumer.getNumOperands() - 1),
         rewriter.getI64IntegerAttr(consumer.getNumResults()),
         rewriter.getAffineMapArrayAttr(fusedIndexMaps),
         consumer.iterator_types(),
         /*doc=*/nullptr,
-        /*library_call=*/nullptr);
+        /*library_call=*/nullptr,
+        /*symbol_source=*/nullptr);
 
     // Map the block argument corresponding to the replaced argument with the
     // scalar constant.
     Region &consumerRegion = consumer.region();
     Block &entryBlock = *consumerRegion.begin();
     unsigned argIndex =
         entryBlock.getNumArguments() - consumer.getNumOperands() + consumerIdx;
     BlockAndValueMapping mapping;

mlir/lib/Dialect/Linalg/Transforms/Loops.cpp

 using edsc::op::operator+;
 
 static SmallVector<Value, 8> makeCanonicalAffineApplies(OpBuilder &b,
                                                         Location loc,
                                                         AffineMap map,
                                                         ArrayRef<Value> vals) {
   if (map.isEmpty())
     return {};
-  assert(map.getNumSymbols() == 0);
+
   assert(map.getNumInputs() == vals.size());
   SmallVector<Value, 8> res;
   res.reserve(map.getNumResults());
   auto dims = map.getNumDims();
   for (auto e : map.getResults()) {
-    auto exprMap = AffineMap::get(dims, 0, e);
+    auto exprMap = AffineMap::get(dims, map.getNumSymbols(), e);
     SmallVector<Value, 4> operands(vals.begin(), vals.end());
     canonicalizeMapAndOperands(&exprMap, &operands);
     res.push_back(affine_apply(exprMap, operands));
   }
   return res;
 }
 
 static SmallVector<Value, 4> permuteIvs(ArrayRef<Value> ivs,
void emitScalarImplementation(ArrayRef<Value> allIvs,
   unsigned nOutputs = linalgOp.getNumOutputs();
   SmallVector<Value, 4> indexedValues;
   indexedValues.reserve(nInputs + nOutputs);
 
   // TODO: Avoid the loads if the corresponding argument of the
   // region has no uses.
   // 1.a. Emit load from input views.
   for (unsigned i = 0; i < nInputs; ++i) {
+    auto attr = linalgOp.template getAttrOfType<IntegerAttr>("symbol_source");
+    AffineMap map = linalgOp.getInputIndexingMap(i);
+    auto allIvsPlusDims = SmallVector<Value, 4>(allIvs.begin(), allIvs.end());
+    if (attr && map.getNumSymbols() > 0) {
+      auto operand = linalgOp.getOperand(attr.getInt());
+      auto shapedType = operand.getType().template cast<ShapedType>();
+      allIvsPlusDims.reserve(allIvs.size() + shapedType.getRank());
+      for (unsigned idx = 0, e = shapedType.getRank(); idx < e; ++idx)
+        allIvsPlusDims.push_back(b.create<DimOp>(loc, operand, idx));
+    }
     auto indexing = makeCanonicalAffineApplies(
-        b, loc, linalgOp.getInputIndexingMap(i), allIvs);
+        b, loc, linalgOp.getInputIndexingMap(i), allIvsPlusDims);
     // Passing through IndexedValueType emits the proper load operation.
     indexedValues.push_back(IndexedValueType(linalgOp.getInput(i))(indexing));
   }
   // 1.b. Emit load from output views.
   for (unsigned i = 0; i < nOutputs; ++i) {
     auto indexing = makeCanonicalAffineApplies(
         b, loc, linalgOp.getOutputIndexingMap(i), allIvs);
     // Passing through IndexedValueType emits the proper load operation.
Optional<LinalgLoops> linalgOpToLoopsImpl(Operation *op, OpBuilder &builder) {
   // permutation map (which is asserted in the inverse calculation).
   auto linalgOp = cast<ConcreteOpTy>(op);
   assert(linalgOp.hasBufferSemantics() &&
          "expected linalg op with buffer semantics");
   auto mapsRange =
       linalgOp.indexing_maps().template getAsRange<AffineMapAttr>();
   auto maps = llvm::to_vector<8>(
       llvm::map_range(mapsRange, [](AffineMapAttr a) { return a.getValue(); }));
-  AffineMap invertedMap = inversePermutation(concatAffineMaps(maps));
+  AffineMap map = concatAffineMaps(maps);
+  // ignore symbols for now as they are not supported
+  AffineMap invertedMap = inversePermutation(
+      AffineMap::get(map.getNumDims(), 0, map.getResults(), map.getContext()));
   if (!invertedMap)
     return {};
   if (invertedMap.isEmpty()) {
     emitScalarImplementation<IndexedValueTy>({}, linalgOp);
     return LinalgLoops();
   }
 
   SmallVector<Value, 4> allIvs;

mlir/lib/Dialect/Linalg/Transforms/TensorsToBuffers.cpp

for (auto result : results) {
       newArgs.push_back(alloc);
       newResults.push_back(alloc);
     }
 
     // Generate a new linalg operation that works on buffers.
     auto linalgOp = rewriter.create<linalg::GenericOp>(
         loc, llvm::None, newArgs, rewriter.getI64IntegerAttr(operands.size()),
         rewriter.getI64IntegerAttr(results.size()), op.indexing_maps(),
-        op.iterator_types(), op.docAttr(), op.library_callAttr());
+        op.iterator_types(), op.docAttr(), op.library_callAttr(),
+        op.symbol_sourceAttr());
 
     // Create a new block in the region of the new Generic Op.
     Block &oldBlock = op.getRegion().front();
     Region &newRegion = linalgOp.region();
     Block *newBlock = rewriter.createBlock(&newRegion, newRegion.begin(),
                                            oldBlock.getArgumentTypes());
 
     // Add the result arguments to the new block.

mlir/lib/IR/AffineExpr.cpp

case AffineExprKind::Mod:
     auto newRHS = rhs.replaceDimsAndSymbols(dimReplacements, symReplacements);
     if (newLHS == lhs && newRHS == rhs)
       return *this;
     return getAffineBinaryOpExpr(getKind(), newLHS, newRHS);
   }
   llvm_unreachable("Unknown AffineExpr");
 }
 
+/// Replace symbols[0 .. numDims - 1] by symbols[shift .. shift + numDims - 1].
+AffineExpr AffineExpr::shiftSymbols(unsigned numSymbols, unsigned shift) const {
+  SmallVector<AffineExpr, 4> symbols;
+  for (unsigned idx = 0; idx < numSymbols; ++idx)
+    symbols.push_back(getAffineSymbolExpr(idx + shift, getContext()));
+  return replaceDimsAndSymbols({}, symbols);
+}
+
 /// Returns true if this expression is made out of only symbols and
 /// constants (no dimensional identifiers).
 bool AffineExpr::isSymbolicOrConstant() const {
   switch (getKind()) {
   case AffineExprKind::Constant:
     return true;
   case AffineExprKind::DimId:
     return false;

mlir/lib/IR/AffineMap.cpp

uniqueExprs.erase(std::unique(uniqueExprs.begin(), uniqueExprs.end()),
                     uniqueExprs.end());
   return AffineMap::get(map.getNumDims(), map.getNumSymbols(), uniqueExprs,
                         map.getContext());
 }
 
 AffineMap mlir::inversePermutation(AffineMap map) {
   if (map.isEmpty())
     return map;
   assert(map.getNumSymbols() == 0 && "expected map without symbols");
   SmallVector<AffineExpr, 4> exprs(map.getNumDims());
   for (auto en : llvm::enumerate(map.getResults())) {
     auto expr = en.value();
     // Skip non-permutations.
     if (auto d = expr.dyn_cast<AffineDimExpr>()) {
       if (exprs[d.getPosition()])
         continue;
       exprs[d.getPosition()] = getAffineDimExpr(en.index(), d.getContext());
     }
   }
   SmallVector<AffineExpr, 4> seenExprs;
   seenExprs.reserve(map.getNumDims());
   for (auto expr : exprs)
     if (expr)
       seenExprs.push_back(expr);
   if (seenExprs.size() != map.getNumInputs())
     return AffineMap();
   return AffineMap::get(map.getNumResults(), 0, seenExprs, map.getContext());
 }
 
 AffineMap mlir::concatAffineMaps(ArrayRef<AffineMap> maps) {
-  unsigned numResults = 0;
+  unsigned numResults = 0, numDims = 0, numSymbols = 0;
   for (auto m : maps)
     numResults += m.getNumResults();
-  unsigned numDims = 0;
   SmallVector<AffineExpr, 8> results;
   results.reserve(numResults);
   for (auto m : maps) {
-    assert(m.getNumSymbols() == 0 && "expected map without symbols");
-    results.append(m.getResults().begin(), m.getResults().end());
+    for (auto res : m.getResults())
+      results.push_back(res.shiftSymbols(m.getNumSymbols(), numSymbols));
+
+    numSymbols += m.getNumSymbols();
     numDims = std::max(m.getNumDims(), numDims);
   }
-  return AffineMap::get(numDims, /*numSymbols=*/0, results,
+  return AffineMap::get(numDims, numSymbols, results,
                         maps.front().getContext());
 }
 
 //===----------------------------------------------------------------------===//
 // MutableAffineMap.
 //===----------------------------------------------------------------------===//
 
 MutableAffineMap::MutableAffineMap(AffineMap map)

mlir/test/Dialect/Linalg/invalid.mlir

linalg.generic {
     ^bb(%0: f32):
       linalg.yield
   }: memref<f32>
 }
 
 // -----
 
 func @generic_symbol_in_map(%arg0: memref<i32>) {
-  // expected-error @+1 {{op expected indexing_map #0 to have no symbols}}
+  // expected-error @+1 {{expected the number of symbols in indexing_map #0 to either match target rank or to be 0}}
   linalg.generic {
     args_in = 0,
     args_out = 1,
     indexing_maps =  [ affine_map<()[N] -> (0)> ],
     iterator_types = ["parallel"]
   } %arg0 {
     ^bb(%i : i32):
     linalg.yield %i : i32
   }: memref<i32>
 }
 
 // -----
 
+func @generic_symbol_source_out_of_range(%arg0: memref<i32>) {
+  // expected-error @+1 {{symbol_source index out of range}}
+  linalg.generic {
+    args_in = 0,
+    args_out = 1,
+    indexing_maps =  [ affine_map<()[N] -> (0)> ],
+    iterator_types = ["parallel"],
+    symbol_source = 1
+  } %arg0 {
+    ^bb(%i : i32):
+    linalg.yield %i : i32
+  }: memref<i32>
+}
+
+// -----
+
 func @generic_wrong_dim_in_map(%arg0: memref<1xi32>) {
   // expected-error @+1 {{op expected indexing_map #0 to have 1 dim(s) to match the number of loops}}
   linalg.generic {
     args_in = 0,
     args_out = 1,
     indexing_maps =  [ affine_map<() -> (0)> ],
     iterator_types = ["parallel"]
   } %arg0 {

mlir/test/Dialect/Linalg/loops.mlir

 // RUN: mlir-opt %s -convert-linalg-to-loops | FileCheck --check-prefix=CHECKLOOP %s
 // RUN: mlir-opt %s -convert-linalg-to-parallel-loops | FileCheck --check-prefix=CHECKPARALLEL %s
 
 // Test that we can lower all the way to LLVM without crashing, don't check results here.
 // RUN: mlir-opt %s -convert-linalg-to-loops -convert-linalg-to-llvm -o=/dev/null 2>&1
 
 // CHECKLOOP-DAG: #[[$strided1D:.*]] = affine_map<(d0)[s0] -> (d0 + s0)>
 // CHECKLOOP-DAG: #[[$strided2D:.*]] = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1)>
 // CHECKLOOP-DAG: #[[$strided3D:.*]] = affine_map<(d0, d1, d2)[s0, s1, s2] -> (d0 * s1 + s0 + d1 * s2 + d2)>
 // CHECKLOOP-DAG: #[[$strided4D:.*]] = affine_map<(d0, d1, d2, d3)[s0, s1, s2, s3] -> (d0 * s1 + s0 + d1 * s2 + d2 * s3 + d3)>
 // CHECKLOOP-DAG: #[[$clampMinMap:.*]] = affine_map<(d0) -> (d0, 0)>
 
 // CHECKLOOP-DAG: #[[$stride1Dilation1:.*]] = affine_map<(d0, d1) -> (d0  + d1)>
 // CHECKLOOP-DAG: #[[$stride2Dilation1:.*]] = affine_map<(d0, d1) -> (d0 * 2 + d1)>
 // CHECKLOOP-DAG: #[[$stride2Dilation4:.*]] = affine_map<(d0, d1) -> (d0 * 2 + d1 * 4)>
 // CHECKLOOP-DAG: #[[$stride3Dilation5:.*]] = affine_map<(d0, d1) -> (d0 * 3 + d1 * 5)>
+// CHECKLOOP-DAG: #[[$convMap:.*]] = affine_map<(d0, d1)[s0] -> (d0 + d1 - s0 floordiv 2)>
 
 // CHECKPARALLEL-DAG: #[[$strided1D:.*]] = affine_map<(d0)[s0] -> (d0 + s0)>
 // CHECKPARALLEL-DAG: #[[$strided2D:.*]] = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1)>
 // CHECKPARALLEL-DAG: #[[$strided3D:.*]] = affine_map<(d0, d1, d2)[s0, s1, s2] -> (d0 * s1 + s0 + d1 * s2 + d2)>
 // CHECKPARALLEL-DAG: #[[$strided4D:.*]] = affine_map<(d0, d1, d2, d3)[s0, s1, s2, s3] -> (d0 * s1 + s0 + d1 * s2 + d2 * s3 + d3)>
 // CHECKPARALLEL-DAG: #[[$clampMinMap:.*]] = affine_map<(d0) -> (d0, 0)>
 
 // CHECKPARALLEL-DAG: #[[$stride1Dilation1:.*]] = affine_map<(d0, d1) -> (d0  + d1)>
 // CHECKPARALLEL-DAG: #[[$stride2Dilation1:.*]] = affine_map<(d0, d1) -> (d0 * 2 + d1)>
 // CHECKPARALLEL-DAG: #[[$stride2Dilation4:.*]] = affine_map<(d0, d1) -> (d0 * 2 + d1 * 4)>
 // CHECKPARALLEL-DAG: #[[$stride3Dilation5:.*]] = affine_map<(d0, d1) -> (d0 * 3 + d1 * 5)>
+// CHECKPARALLEL-DAG: #[[$convMap:.*]] = affine_map<(d0, d1)[s0] -> (d0 + d1 - s0 floordiv 2)>
 
 
 func @matmul(%arg0: memref<?xi8>, %M: index, %N: index, %K: index) {
   %c0 = constant 0 : index
   %c1 = constant 1 : index
   %A = view %arg0[%c0][%M, %K] : memref<?xi8> to memref<?x?xf32>
   %B = view %arg0[%c0][%K, %N] : memref<?xi8> to memref<?x?xf32>
   %C = view %arg0[%c0][%M, %N] : memref<?xi8> to memref<?x?xf32>

 //       CHECKPARALLEL: scf.parallel (%[[b:.*]], %[[m:.*]], %[[n:.*]]) = ({{.*}}) to (%[[B]], %[[M]], %[[N]]) step ({{.*}}) {
 //       CHECKPARALLEL:   scf.for %[[k:.*]] = %{{.*}} to %[[K]] step %{{.*}} {
 //       CHECKPARALLEL:       %[[va:.*]] = load %[[mA]][%[[b]], %[[m]], %[[k]]] : memref<?x?x?xf32>
 //       CHECKPARALLEL:       %[[vb:.*]] = load %[[mB]][%[[b]], %[[k]], %[[n]]] : memref<?x?x?xf32>
 //       CHECKPARALLEL:       %[[vc:.*]] = load %[[mC]][%[[b]], %[[m]], %[[n]]] : memref<?x?x?xf32>
 //       CHECKPARALLEL:       %[[inc:.*]] = mulf %[[va]], %[[vb]] : f32
 //       CHECKPARALLEL:       %[[res:.*]] = addf %[[vc]], %[[inc]] : f32
 //       CHECKPARALLEL:       store %[[res]], %[[mC]][%[[b]], %[[m]], %[[n]]] : memref<?x?x?xf32>
+
+#conv_1d_accesses = [
+  affine_map<(m, n)[s0] -> (m + n - s0 floordiv 2)>, // in
+  affine_map<(m, n) -> (n)>, // filter
+  affine_map<(m, n) -> (m)> // out
+]
+
+#conv_1d_trait = {
+  args_in = 2,
+  args_out = 1,
+  doc = "C(m) += A(m) * B(n)",
+  indexing_maps = #conv_1d_accesses,
+  library_call  = "linalg_conv_1d",
+  n_views = [2, 1],
+  iterator_types = ["parallel", "parallel"],
+  symbol_source = 1
+}
+
+func @conv1d(%filter : memref<?xf32>, %in : memref<?xf32>, %out :  memref<?xf32>) -> () {
+  linalg.generic #conv_1d_trait %in, %filter, %out {
+    ^bb0(%a: f32, %b: f32, %c: f32) :
+      %d = mulf %a, %b : f32
+      %e = addf %c, %d : f32
+      linalg.yield %e : f32
+  } : memref<?xf32>,
+      memref<?xf32>,
+      memref<?xf32>
+  return
+}
+
+// CHECKLOOP-LABEL: @conv1d
+//  CHECKLOOP-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?xf32>
+//  CHECKLOOP-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?xf32>
+//  CHECKLOOP-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?xf32>
+//       CHECKLOOP: %[[c0:.*]] = constant 0 : index
+//       CHECKLOOP: %[[dim0:.*]] = dim %[[arg0]], %[[c0]] : memref<?xf32>
+//       CHECKLOOP: %[[dim1:.*]] = dim %[[arg2]], %[[c0]] : memref<?xf32>
+//       CHECKLOOP: scf.for %[[b:.*]] = %{{.*}} to %[[dim1]] step %{{.*}} {
+//       CHECKLOOP:   scf.for %[[m:.*]] = %{{.*}} to %[[dim0]] step %{{.*}} {
+//       CHECKLOOP:     %[[dim2:.*]] = dim %[[arg0]], %[[c0]] : memref<?xf32>
+//       CHECKLOOP:     %[[aff:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim2]]]
+//       CHECKLOOP:     %[[va:.*]] = load %[[arg1]][%[[aff]]] : memref<?xf32>
+//       CHECKLOOP:     %[[vb:.*]] = load %[[arg0]][%[[m]]] : memref<?xf32>
+//       CHECKLOOP:     %[[vc:.*]] = load %[[arg2]][%[[b]]] : memref<?xf32>
+//       CHECKLOOP:     %[[inc:.*]] = mulf %[[va]], %[[vb]] : f32
+//       CHECKLOOP:     %[[res:.*]] = addf %[[vc]], %[[inc]] : f32
+//       CHECKLOOP:     store %[[res]], %[[arg2]][%[[b]]] : memref<?xf32>
+
+// CHECKPARALLEL-LABEL: @conv1d
+//  CHECKPARALLEL-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?xf32>
+//  CHECKPARALLEL-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?xf32>
+//  CHECKPARALLEL-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?xf32>
+//       CHECKPARALLEL: %[[c0:.*]] = constant 0 : index
+//       CHECKPARALLEL: %[[dim0:.*]] = dim %[[arg0]], %[[c0]] : memref<?xf32>
+//       CHECKPARALLEL: %[[dim1:.*]] = dim %[[arg2]], %[[c0]] : memref<?xf32>
+//       CHECKPARALLEL: scf.parallel (%[[b:.*]], %[[m:.*]]) = (%{{.*}}, %{{.*}}) to (%[[dim1]], %[[dim0]]) step ({{.*}}) {
+//       CHECKPARALLEL:   %[[dim2:.*]] = dim %[[arg0]], %[[c0]] : memref<?xf32>
+//       CHECKPARALLEL:   %[[aff:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim2]]]
+//       CHECKPARALLEL:   %[[va:.*]] = load %[[arg1]][%[[aff]]] : memref<?xf32>
+//       CHECKPARALLEL:   %[[vb:.*]] = load %[[arg0]][%[[m]]] : memref<?xf32>
+//       CHECKPARALLEL:   %[[vc:.*]] = load %[[arg2]][%[[b]]] : memref<?xf32>
+//       CHECKPARALLEL:   %[[inc:.*]] = mulf %[[va]], %[[vb]] : f32
+//       CHECKPARALLEL:   %[[res:.*]] = addf %[[vc]], %[[inc]] : f32
+//       CHECKPARALLEL:   store %[[res]], %[[arg2]][%[[b]]] : memref<?xf32>
+
+#conv_2d_accesses = [
+  affine_map<(m, n, m1, n1)[s0, s1] -> (m + m1 - s0 floordiv 2, n + n1 - s1 floordiv 2)>, // in
+  affine_map<(m, n, m1, n1) -> (m1, n1)>, // filter
+  affine_map<(m, n, m1, n1) -> (m, n)> // out
+]
+
+#conv_2d_trait = {
+  args_in = 2,
+  args_out = 1,
+  doc = "C(m,n) += A(m,n) * B(m1,n1)",
+  indexing_maps = #conv_2d_accesses,
+  library_call  = "linalg_conv_2d",
+  n_views = [2, 1],
+  iterator_types = ["parallel", "parallel", "parallel", "parallel"],
+  symbol_source = 1
+}
+
+func @conv2d(%filter : memref<?x?xf32>, %in : memref<?x?xf32>, %out :  memref<?x?xf32>) -> () {
+  linalg.generic #conv_2d_trait %in, %filter, %out {
+    ^bb0(%a: f32, %b: f32, %c: f32) :
+      %d = mulf %a, %b : f32
+      %e = addf %c, %d : f32
+      linalg.yield %e : f32
+  } : memref<?x?xf32>,
+      memref<?x?xf32>,
+      memref<?x?xf32>
+  return
+}
+
+// CHECKLOOP-LABEL: @conv2d
+//  CHECKLOOP-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?x?xf32>
+//  CHECKLOOP-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?x?xf32>
+//  CHECKLOOP-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?x?xf32>
+//       CHECKLOOP: %[[c0:.*]] = constant 0 : index
+//       CHECKLOOP: %[[c1:.*]] = constant 1 : index
+//       CHECKLOOP: %[[dim0:.*]] = dim %[[arg0]], %[[c0]] : memref<?x?xf32>
+//       CHECKLOOP: %[[dim1:.*]] = dim %[[arg0]], %[[c1]] : memref<?x?xf32>
+//       CHECKLOOP: %[[dim2:.*]] = dim %[[arg2]], %[[c0]] : memref<?x?xf32>
+//       CHECKLOOP: %[[dim3:.*]] = dim %[[arg2]], %[[c1]] : memref<?x?xf32>
+//       CHECKLOOP: scf.for %[[i0:.*]] = %{{.*}} to %[[dim2]] step %{{.*}} {
+//       CHECKLOOP:   scf.for %[[i1:.*]] = %{{.*}} to %[[dim3]] step %{{.*}} {
+//       CHECKLOOP:     scf.for %[[i2:.*]] = %{{.*}} to %[[dim0]] step %{{.*}} {
+//       CHECKLOOP:       scf.for %[[i3:.*]] = %{{.*}} to %[[dim1]] step %{{.*}} {
+//       CHECKLOOP:         %[[dim4:.*]] = dim %[[arg0]], %[[c0]] : memref<?x?xf32>
+//       CHECKLOOP:         %[[dim5:.*]] = dim %[[arg0]], %[[c1]] : memref<?x?xf32>
+//       CHECKLOOP:         %[[aff1:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim4]]]
+//       CHECKLOOP:         %[[aff2:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim5]]]
+//       CHECKLOOP:         %[[va:.*]] = load %[[arg1]][%[[aff1]], %[[aff2]]] : memref<?x?xf32>
+//       CHECKLOOP:         %[[vb:.*]] = load %[[arg0]][%[[i2]], %[[i3]]] : memref<?x?xf32>
+//       CHECKLOOP:         %[[vc:.*]] = load %[[arg2]][%[[i0]], %[[i1]]] : memref<?x?xf32>
+//       CHECKLOOP:         %[[inc:.*]] = mulf %[[va]], %[[vb]] : f32
+//       CHECKLOOP:         %[[res:.*]] = addf %[[vc]], %[[inc]] : f32
+//       CHECKLOOP:         store %[[res]], %[[arg2]][%[[i0]], %[[i1]]] : memref<?x?xf32>
+
+// CHECKPARALLEL-LABEL: @conv2d
+//  CHECKPARALLEL-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?x?xf32>
+//  CHECKPARALLEL-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?x?xf32>
+//  CHECKPARALLEL-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?x?xf32>
+//       CHECKPARALLEL: %[[c0:.*]] = constant 0 : index
+//       CHECKPARALLEL: %[[c1:.*]] = constant 1 : index
+//       CHECKPARALLEL: %[[dim0:.*]] = dim %[[arg0]], %[[c0]] : memref<?x?xf32>
+//       CHECKPARALLEL: %[[dim1:.*]] = dim %[[arg0]], %[[c1]] : memref<?x?xf32>
+//       CHECKPARALLEL: %[[dim2:.*]] = dim %[[arg2]], %[[c0]] : memref<?x?xf32>
+//       CHECKPARALLEL: %[[dim3:.*]] = dim %[[arg2]], %[[c1]] : memref<?x?xf32>
+//       CHECKPARALLEL: scf.parallel (%[[i0:.*]], %[[i1:.*]], %[[i2:.*]], %[[i3:.*]]) = (%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) to (%[[dim2]], %[[dim3]], %[[dim0]], %[[dim1]]) step ({{.*}}) {
+//       CHECKPARALLEL:   %[[dim4:.*]] = dim %[[arg0]], %[[c0]] : memref<?x?xf32>
+//       CHECKPARALLEL:   %[[dim5:.*]] = dim %[[arg0]], %[[c1]] : memref<?x?xf32>
+//       CHECKPARALLEL:   %[[aff1:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim4]]]
+//       CHECKPARALLEL:   %[[aff2:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim5]]]
+//       CHECKPARALLEL:   %[[va:.*]] = load %[[arg1]][%[[aff1]], %[[aff2]]] : memref<?x?xf32>
+//       CHECKPARALLEL:   %[[vb:.*]] = load %[[arg0]][%[[i2]], %[[i3]]] : memref<?x?xf32>
+//       CHECKPARALLEL:   %[[vc:.*]] = load %[[arg2]][%[[i0]], %[[i1]]] : memref<?x?xf32>
+//       CHECKPARALLEL:   %[[inc:.*]] = mulf %[[va]], %[[vb]] : f32
+//       CHECKPARALLEL:   %[[res:.*]] = addf %[[vc]], %[[inc]] : f32
+//       CHECKPARALLEL:   store %[[res]], %[[arg2]][%[[i0]], %[[i1]]] : memref<?x?xf32>
+
+#conv_3d_accesses = [
+  affine_map<(m, n, k, m1, n1, k1)[s0, s1, s2] -> (m + m1 - s0 floordiv 2, n + n1 - s1 floordiv 2, k + k1 - s2 floordiv 2)>, // in
+  affine_map<(m, n, k, m1, n1, k1) -> (m1, n1, k1)>, // filter
+  affine_map<(m, n, k, m1, n1, k1) -> (m, n, k)> // out
+]
+
+#conv_3d_trait = {
+  args_in = 2,
+  args_out = 1,
+  doc = "C(m,n,k) += A(m,n,k) * B(m1,n1,k1)",
+  indexing_maps = #conv_3d_accesses,
+  library_call  = "linalg_conv_3d",
+  n_views = [2, 1],
+  iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel", "parallel"],
+  symbol_source = 1
+}
+
+func @conv3d(%filter : memref<?x?x?xf32>, %in : memref<?x?x?xf32>, %out :  memref<?x?x?xf32>) -> () {
+  linalg.generic #conv_3d_trait %in, %filter, %out {
+    ^bb0(%a: f32, %b: f32, %c: f32) :
+      %d = mulf %a, %b : f32
+      %e = addf %c, %d : f32
+      linalg.yield %e : f32
+  } : memref<?x?x?xf32>,
+      memref<?x?x?xf32>,
+      memref<?x?x?xf32>
+  return
+}
+
+// CHECKLOOP-LABEL: @conv3d
+//  CHECKLOOP-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?x?x?xf32>
+//  CHECKLOOP-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?x?x?xf32>
+//  CHECKLOOP-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?x?x?xf32>
+//       CHECKLOOP: %[[c0:.*]] = constant 0 : index
+//       CHECKLOOP: %[[c1:.*]] = constant 1 : index
+//       CHECKLOOP: %[[c2:.*]] = constant 2 : index
+//       CHECKLOOP: %[[dim0:.*]] = dim %[[arg0]], %[[c0]] : memref<?x?x?xf32>
+//       CHECKLOOP: %[[dim1:.*]] = dim %[[arg0]], %[[c1]] : memref<?x?x?xf32>
+//       CHECKLOOP: %[[dim2:.*]] = dim %[[arg0]], %[[c2]] : memref<?x?x?xf32>
+//       CHECKLOOP: %[[dim3:.*]] = dim %[[arg2]], %[[c0]] : memref<?x?x?xf32>
+//       CHECKLOOP: %[[dim4:.*]] = dim %[[arg2]], %[[c1]] : memref<?x?x?xf32>
+//       CHECKLOOP: %[[dim5:.*]] = dim %[[arg2]], %[[c2]] : memref<?x?x?xf32>
+//       CHECKLOOP: scf.for %[[i0:.*]] = %{{.*}} to %[[dim3]] step %{{.*}} {
+//       CHECKLOOP:   scf.for %[[i1:.*]] = %{{.*}} to %[[dim4]] step %{{.*}} {
+//       CHECKLOOP:     scf.for %[[i2:.*]] = %{{.*}} to %[[dim5]] step %{{.*}} {
+//       CHECKLOOP:       scf.for %[[i3:.*]] = %{{.*}} to %[[dim0]] step %{{.*}} {
+//       CHECKLOOP:         scf.for %[[i4:.*]] = %{{.*}} to %[[dim1]] step %{{.*}} {
+//       CHECKLOOP:           scf.for %[[i5:.*]] = %{{.*}} to %[[dim2]] step %{{.*}} {
+//       CHECKLOOP:             %[[dim6:.*]] = dim %[[arg0]], %[[c0]] : memref<?x?x?xf32>
+//       CHECKLOOP:             %[[dim7:.*]] = dim %[[arg0]], %[[c1]] : memref<?x?x?xf32>
+//       CHECKLOOP:             %[[dim8:.*]] = dim %[[arg0]], %[[c2]] : memref<?x?x?xf32>
+//       CHECKLOOP:             %[[aff1:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim6]]]
+//       CHECKLOOP:             %[[aff2:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim7]]]
+//       CHECKLOOP:             %[[aff3:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim8]]]
+//       CHECKLOOP:             %[[va:.*]] = load %[[arg1]][%[[aff1]], %[[aff2]], %[[aff3]]] : memref<?x?x?xf32>
+//       CHECKLOOP:             %[[vb:.*]] = load %[[arg0]][%[[i3]], %[[i4]], %[[i5]]] : memref<?x?x?xf32>
+//       CHECKLOOP:             %[[vc:.*]] = load %[[arg2]][%[[i0]], %[[i1]], %[[i2]]] : memref<?x?x?xf32>
+//       CHECKLOOP:             %[[inc:.*]] = mulf %[[va]], %[[vb]] : f32
+//       CHECKLOOP:             %[[res:.*]] = addf %[[vc]], %[[inc]] : f32
+//       CHECKLOOP:             store %[[res]], %[[arg2]][%[[i0]], %[[i1]], %[[i2]]] : memref<?x?x?xf32>
+
+// CHECKPARALLEL-LABEL: @conv3d
+//  CHECKPARALLEL-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?x?x?xf32>
+//  CHECKPARALLEL-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?x?x?xf32>
+//  CHECKPARALLEL-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?x?x?xf32>
+//       CHECKPARALLEL: %[[c0:.*]] = constant 0 : index
+//       CHECKPARALLEL: %[[c1:.*]] = constant 1 : index
+//       CHECKPARALLEL: %[[c2:.*]] = constant 2 : index
+//       CHECKPARALLEL: %[[dim0:.*]] = dim %[[arg0]], %[[c0]] : memref<?x?x?xf32>
+//       CHECKPARALLEL: %[[dim1:.*]] = dim %[[arg0]], %[[c1]] : memref<?x?x?xf32>
+//       CHECKPARALLEL: %[[dim2:.*]] = dim %[[arg0]], %[[c2]] : memref<?x?x?xf32>
+//       CHECKPARALLEL: %[[dim3:.*]] = dim %[[arg2]], %[[c0]] : memref<?x?x?xf32>
+//       CHECKPARALLEL: %[[dim4:.*]] = dim %[[arg2]], %[[c1]] : memref<?x?x?xf32>
+//       CHECKPARALLEL: %[[dim5:.*]] = dim %[[arg2]], %[[c2]] : memref<?x?x?xf32>
+//       CHECKPARALLEL: scf.parallel (%[[i0:.*]], %[[i1:.*]], %[[i2:.*]], %[[i3:.*]], %[[i4:.*]], %[[i5:.*]]) = (%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) to (%[[dim3]], %[[dim4]], %[[dim5]], %[[dim0]], %[[dim1]], %[[dim2]]) step ({{.*}}) {
+//       CHECKPARALLEL:   %[[dim6:.*]] = dim %[[arg0]], %[[c0]] : memref<?x?x?xf32>
+//       CHECKPARALLEL:   %[[dim7:.*]] = dim %[[arg0]], %[[c1]] : memref<?x?x?xf32>
+//       CHECKPARALLEL:   %[[dim8:.*]] = dim %[[arg0]], %[[c2]] : memref<?x?x?xf32>
+//       CHECKPARALLEL:   %[[aff1:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim6]]]
+//       CHECKPARALLEL:   %[[aff2:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim7]]]
+//       CHECKPARALLEL:   %[[aff3:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim8]]]
+//       CHECKPARALLEL:   %[[va:.*]] = load %[[arg1]][%[[aff1]], %[[aff2]], %[[aff3]]] : memref<?x?x?xf32>
+//       CHECKPARALLEL:   %[[vb:.*]] = load %[[arg0]][%[[i3]], %[[i4]], %[[i5]]] : memref<?x?x?xf32>
+//       CHECKPARALLEL:   %[[vc:.*]] = load %[[arg2]][%[[i0]], %[[i1]], %[[i2]]] : memref<?x?x?xf32>
+//       CHECKPARALLEL:   %[[inc:.*]] = mulf %[[va]], %[[vb]] : f32
+//       CHECKPARALLEL:   %[[res:.*]] = addf %[[vc]], %[[inc]] : f32
+//       CHECKPARALLEL:   store %[[res]], %[[arg2]][%[[i0]], %[[i1]], %[[i2]]] : memref<?x?x?xf32>
+
+#conv_4d_accesses = [
+  affine_map<(m, n, k, l, m1, n1, k1, l1)[s0, s1, s2, s3] -> (m + m1 - s0 floordiv 2, n + n1 - s1 floordiv 2, k + k1 - s2 floordiv 2, l + l1 - s3 floordiv 2)>, // in
+  affine_map<(m, n, k, l, m1, n1, k1, l1) -> (m1, n1, k1, l1)>, // filter
+  affine_map<(m, n, k, l, m1, n1, k1, l1) -> (m, n, k, l)> // out
+]
+
+#conv_4d_trait = {
+  args_in = 2,
+  args_out = 1,
+  doc = "C(m,n,k,l) += A(m,n,k,l) * B(m1,n1,k1,l1)",
+  indexing_maps = #conv_4d_accesses,
+  library_call  = "linalg_conv_4d",
+  n_views = [2, 1],
+  iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel"],
+  symbol_source = 1
+}
+
+func @conv4d(%filter : memref<?x?x?x?xf32>, %in : memref<?x?x?x?xf32>, %out :  memref<?x?x?x?xf32>) -> () {
+  linalg.generic #conv_4d_trait %in, %filter, %out {
+    ^bb0(%a: f32, %b: f32, %c: f32) :
+      %d = mulf %a, %b : f32
+      %e = addf %c, %d : f32
+      linalg.yield %e : f32
+  } : memref<?x?x?x?xf32>,
+      memref<?x?x?x?xf32>,
+      memref<?x?x?x?xf32>
+  return
+}
+
+// CHECKLOOP-LABEL: @conv4d
+//  CHECKLOOP-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?x?x?x?xf32>
+//  CHECKLOOP-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?x?x?x?xf32>
+//  CHECKLOOP-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?x?x?x?xf32>
+//       CHECKLOOP: %[[c0:.*]] = constant 0 : index
+//       CHECKLOOP: %[[c1:.*]] = constant 1 : index
+//       CHECKLOOP: %[[c2:.*]] = constant 2 : index
+//       CHECKLOOP: %[[c3:.*]] = constant 3 : index
+//       CHECKLOOP: %[[dim0:.*]] = dim %[[arg0]], %[[c0]] : memref<?x?x?x?xf32>
+//       CHECKLOOP: %[[dim1:.*]] = dim %[[arg0]], %[[c1]] : memref<?x?x?x?xf32>
+//       CHECKLOOP: %[[dim2:.*]] = dim %[[arg0]], %[[c2]] : memref<?x?x?x?xf32>
+//       CHECKLOOP: %[[dim3:.*]] = dim %[[arg0]], %[[c3]] : memref<?x?x?x?xf32>
+//       CHECKLOOP: %[[dim4:.*]] = dim %[[arg2]], %[[c0]] : memref<?x?x?x?xf32>
+//       CHECKLOOP: %[[dim5:.*]] = dim %[[arg2]], %[[c1]] : memref<?x?x?x?xf32>
+//       CHECKLOOP: %[[dim6:.*]] = dim %[[arg2]], %[[c2]] : memref<?x?x?x?xf32>
+//       CHECKLOOP: %[[dim7:.*]] = dim %[[arg2]], %[[c3]] : memref<?x?x?x?xf32>
+//       CHECKLOOP: scf.for %[[i0:.*]] = %{{.*}} to %[[dim4]] step %{{.*}} {
+//       CHECKLOOP:   scf.for %[[i1:.*]] = %{{.*}} to %[[dim5]] step %{{.*}} {
+//       CHECKLOOP:     scf.for %[[i2:.*]] = %{{.*}} to %[[dim6]] step %{{.*}} {
+//       CHECKLOOP:       scf.for %[[i3:.*]] = %{{.*}} to %[[dim7]] step %{{.*}} {
+//       CHECKLOOP:         scf.for %[[i4:.*]] = %{{.*}} to %[[dim0]] step %{{.*}} {
+//       CHECKLOOP:           scf.for %[[i5:.*]] = %{{.*}} to %[[dim1]] step %{{.*}} {
+//       CHECKLOOP:             scf.for %[[i6:.*]] = %{{.*}} to %[[dim2]] step %{{.*}} {
+//       CHECKLOOP:               scf.for %[[i7:.*]] = %{{.*}} to %[[dim3]] step %{{.*}} {
+//       CHECKLOOP:                 %[[dim8:.*]] = dim %[[arg0]], %[[c0]] : memref<?x?x?x?xf32>
+//       CHECKLOOP:                 %[[dim9:.*]] = dim %[[arg0]], %[[c1]] : memref<?x?x?x?xf32>
+//       CHECKLOOP:                 %[[dim10:.*]] = dim %[[arg0]], %[[c2]] : memref<?x?x?x?xf32>
+//       CHECKLOOP:                 %[[dim11:.*]] = dim %[[arg0]], %[[c3]] : memref<?x?x?x?xf32>
+//       CHECKLOOP:                 %[[aff1:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim8]]]
+//       CHECKLOOP:                 %[[aff2:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim9]]]
+//       CHECKLOOP:                 %[[aff3:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim10]]]
+//       CHECKLOOP:                 %[[aff4:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim11]]]
+//       CHECKLOOP:                 %[[va:.*]] = load %[[arg1]][%[[aff1]], %[[aff2]], %[[aff3]], %[[aff4]]] : memref<?x?x?x?xf32>
+//       CHECKLOOP:                 %[[vb:.*]] = load %[[arg0]][%[[i4]], %[[i5]], %[[i6]], %[[i7]]] : memref<?x?x?x?xf32>
+//       CHECKLOOP:                 %[[vc:.*]] = load %[[arg2]][%[[i0]], %[[i1]], %[[i2]], %[[i3]]] : memref<?x?x?x?xf32>
+//       CHECKLOOP:                 %[[inc:.*]] = mulf %[[va]], %[[vb]] : f32
+//       CHECKLOOP:                 %[[res:.*]] = addf %[[vc]], %[[inc]] : f32
+//       CHECKLOOP:                 store %[[res]], %[[arg2]][%[[i0]], %[[i1]], %[[i2]], %[[i3]]] : memref<?x?x?x?xf32>
+
+// CHECKPARALLEL-LABEL: @conv4d
+//  CHECKPARALLEL-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?x?x?x?xf32>
+//  CHECKPARALLEL-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?x?x?x?xf32>
+//  CHECKPARALLEL-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?x?x?x?xf32>
+//       CHECKPARALLEL: %[[c0:.*]] = constant 0 : index
+//       CHECKPARALLEL: %[[c1:.*]] = constant 1 : index
+//       CHECKPARALLEL: %[[c2:.*]] = constant 2 : index
+//       CHECKPARALLEL: %[[c3:.*]] = constant 3 : index
+//       CHECKPARALLEL: %[[dim0:.*]] = dim %[[arg0]], %[[c0]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL: %[[dim1:.*]] = dim %[[arg0]], %[[c1]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL: %[[dim2:.*]] = dim %[[arg0]], %[[c2]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL: %[[dim3:.*]] = dim %[[arg0]], %[[c3]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL: %[[dim4:.*]] = dim %[[arg2]], %[[c0]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL: %[[dim5:.*]] = dim %[[arg2]], %[[c1]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL: %[[dim6:.*]] = dim %[[arg2]], %[[c2]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL: %[[dim7:.*]] = dim %[[arg2]], %[[c3]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL: scf.parallel (%[[i0:.*]], %[[i1:.*]], %[[i2:.*]], %[[i3:.*]], %[[i4:.*]], %[[i5:.*]], %[[i6:.*]], %[[i7:.*]]) = (%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) to (%[[dim4]], %[[dim5]], %[[dim6]], %[[dim7]], %[[dim0]], %[[dim1]], %[[dim2]], %[[dim3]]) step ({{.*}}) {
+//       CHECKPARALLEL:   %[[dim8:.*]] = dim %[[arg0]], %[[c0]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL:   %[[dim9:.*]] = dim %[[arg0]], %[[c1]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL:   %[[dim10:.*]] = dim %[[arg0]], %[[c2]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL:   %[[dim11:.*]] = dim %[[arg0]], %[[c3]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL:   %[[aff1:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim8]]]
+//       CHECKPARALLEL:   %[[aff2:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim9]]]
+//       CHECKPARALLEL:   %[[aff3:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim10]]]
+//       CHECKPARALLEL:   %[[aff4:.*]] = affine.apply #[[$convMap]](%{{.*}}, %{{.*}})[%[[dim11]]]
+//       CHECKPARALLEL:   %[[va:.*]] = load %[[arg1]][%[[aff1]], %[[aff2]], %[[aff3]], %[[aff4]]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL:   %[[vb:.*]] = load %[[arg0]][%[[i4]], %[[i5]], %[[i6]], %[[i7]]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL:   %[[vc:.*]] = load %[[arg2]][%[[i0]], %[[i1]], %[[i2]], %[[i3]]] : memref<?x?x?x?xf32>
+//       CHECKPARALLEL:   %[[inc:.*]] = mulf %[[va]], %[[vb]] : f32
+//       CHECKPARALLEL:   %[[res:.*]] = addf %[[vc]], %[[inc]] : f32
+//       CHECKPARALLEL:   store %[[res]], %[[arg2]][%[[i0]], %[[i1]], %[[i2]], %[[i3]]] : memref<?x?x?x?xf32>

mlir/test/lib/Transforms/TestBufferPlacement.cpp

matchAndRewrite(linalg::GenericOp op, ArrayRef<Value> operands,
         newArgs.push_back(alloc);
         newResults.push_back(alloc);
       }
 
       // Generate a new linalg operation that works on buffers.
       auto linalgOp = rewriter.create<linalg::GenericOp>(
           loc, llvm::None, newArgs, rewriter.getI64IntegerAttr(operands.size()),
           rewriter.getI64IntegerAttr(results.size()), op.indexing_maps(),
-          op.iterator_types(), op.docAttr(), op.library_callAttr());
+          op.iterator_types(), op.docAttr(), op.library_callAttr(),
+          op.symbol_sourceAttr());
 
       // Create a new block in the region of the new Generic Op.
       Block &oldBlock = op.getRegion().front();
       Region &newRegion = linalgOp.region();
       Block *newBlock = rewriter.createBlock(&newRegion, newRegion.begin(),
                                              oldBlock.getArgumentTypes());
 
       // Map the old block arguments to the new ones.