[mlir][sparse] further implement singleton dimension level type

[wrengr]

Summary

Handle more cases of singleton DLT including direct sparse2sparse conversion. (Followup to D134096)

Depends On D134926

[aartbik]

mlir/include/mlir/ExecutionEngine/SparseTensor/Storage.h

865

you want to be defensive here too with an else assert dense (for the future)

917

period at end

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorConversion.cpp

453

nice refinement!

[wrengr]

rebase

[wrengr]

Adding future-proofing, per nits

[wrengr]

Fixing typo in the new assertions

[stella.stamenova]

This change broke the mlir build on Windows in Debug. I'm actually not entirely sure why it only broke in Debug and not in Release as well because I would expect the issue to be the same in both.

What's happening now is SparseTensorConversion.cpp has calls to a couple of functions defined in Enums.h. However, Enums.h is setup as if it belongs to a shared library (but it doesn't, see comment in the CMakelists file when it was created here: https://reviews.llvm.org/D133462) - so when it is included in SparseTensorConversion.cpp the functions are marked as dllimport but there's no library that exports them, so the build fails to link a number of executables. For example:

cmd.exe /C "cd . && "C:\Program Files\CMake\bin\cmake.exe" -E vs_link_exe --intdir=tools\mlir\tools\mlir-opt\CMakeFiles\mlir-opt.dir --rc=C:\PROGRA~2\WI3CF2~1\10\bin\100203~1.0\x64\rc.exe --mt=C:\PROGRA~2\WI3CF2~1\10\bin\100203~1.0\x64\mt.exe --manifests  -- C:\PROGRA~1\MIB055~1\2022\PROFES~1\VC\Tools\MSVC\1433~1.316\bin\Hostx64\x64\link.exe /nologo @CMakeFiles\mlir-opt.rsp  /out:bin\mlir-opt.exe /implib:lib\mlir-opt.lib /pdb:bin\mlir-opt.pdb /version:0.0 /machine:x64 /STACK:10000000 /debug /INCREMENTAL /subsystem:console  && cd ."
LINK Pass 1: command "C:\PROGRA~1\MIB055~1\2022\PROFES~1\VC\Tools\MSVC\1433~1.316\bin\Hostx64\x64\link.exe /nologo @CMakeFiles\mlir-opt.rsp /out:bin\mlir-opt.exe /implib:lib\mlir-opt.lib /pdb:bin\mlir-opt.pdb /version:0.0 /machine:x64 /STACK:10000000 /debug /INCREMENTAL /subsystem:console /MANIFEST /MANIFESTFILE:tools\mlir\tools\mlir-opt\CMakeFiles\mlir-opt.dir/intermediate.manifest tools\mlir\tools\mlir-opt\CMakeFiles\mlir-opt.dir/manifest.res" failed (exit code 1120) with the following output:
MLIRSparseTensorTransforms.lib(SparseTensorConversion.cpp.obj) : error LNK2019: unresolved external symbol "bool __cdecl mlir::sparse_tensor::isDenseDLT(enum mlir::sparse_tensor::DimLevelType)" (?isDenseDLT@sparse_tensor@mlir@@YA_NW4DimLevelType@12@@Z) referenced in function "bool __cdecl `anonymous namespace'::canUseDirectConversion(class llvm::ArrayRef<enum mlir::sparse_tensor::SparseTensorEncodingAttr::DimLevelType>)" (?canUseDirectConversion@?A0x0125d971@@YA_NV?$ArrayRef@W4DimLevelType@SparseTensorEncodingAttr@sparse_tensor@mlir@@@llvm@@@Z)
MLIRSparseTensorTransforms.lib(SparseTensorConversion.cpp.obj) : error LNK2019: unresolved external symbol "bool __cdecl mlir::sparse_tensor::isCompressedDLT(enum mlir::sparse_tensor::DimLevelType)" (?isCompressedDLT@sparse_tensor@mlir@@YA_NW4DimLevelType@12@@Z) referenced in function "bool __cdecl `anonymous namespace'::canUseDirectConversion(class llvm::ArrayRef<enum mlir::sparse_tensor::SparseTensorEncodingAttr::DimLevelType>)" (?canUseDirectConversion@?A0x0125d971@@YA_NV?$ArrayRef@W4DimLevelType@SparseTensorEncodingAttr@sparse_tensor@mlir@@@llvm@@@Z)
MLIRSparseTensorTransforms.lib(SparseTensorConversion.cpp.obj) : error LNK2019: unresolved external symbol "bool __cdecl mlir::sparse_tensor::isSingletonDLT(enum mlir::sparse_tensor::DimLevelType)" (?isSingletonDLT@sparse_tensor@mlir@@YA_NW4DimLevelType@12@@Z) referenced in function "bool __cdecl `anonymous namespace'::canUseDirectConversion(class llvm::ArrayRef<enum mlir::sparse_tensor::SparseTensorEncodingAttr::DimLevelType>)" (?canUseDirectConversion@?A0x0125d971@@YA_NV?$ArrayRef@W4DimLevelType@SparseTensorEncodingAttr@sparse_tensor@mlir@@@llvm@@@Z)
bin\mlir-opt.exe : fatal error LNK1120: 3 unresolved externals
ninja: build stopped: subcommand failed.

This can be fixed in a couple of ways:

1. Assume that mlir_sparse_tensor_utils is a static library (which it is) and then remove any export/import definitions. This should be fairly straightforward. In this case we might want to consider the library naming as well since mlir_xxx_utils are generally shared libraries.
2. Make mlir_sparse_tensor_utils a shared library and make all the necessary fixes for it to work. One caveat here is that we'd end up with a situation where one of the transform libraries (and as far as I can tell only one of the transform libraries) requires linking to a shared library.

I suppose a third solution would be to move the definitions from Enums.h to the existing MLIRSparseTensorUtils which is also a static library.

Thoughts?

[wrengr]

In D134933#3843372, @stella.stamenova wrote:

What's happening now is SparseTensorConversion.cpp has calls to a couple of functions defined in Enums.h. However, Enums.h is setup as if it belongs to a shared library (but it doesn't, see comment in the CMakelists file when it was created here: https://reviews.llvm.org/D133462) - so when it is included in SparseTensorConversion.cpp the functions are marked as dllimport but there's no library that exports them, so the build fails to link a number of executables.

When factoring out the mlir_sparse_tensor_utils library I was having a lot of issues getting things to work under MSVC, and ultimately I could only resolve the issues by making mlir_sparse_tensor_utils static rather than shared. Fwiw, the chain of reasoning leading to the current state is/was: mlir_c_runner_utils is a shared library so the stuff in ExecutionEngine/SparseTensorUtils.h must be marked with dllimport/dllexport, but SparseTensorUtils.h includes ExecutionEngine/SparseTensor/Enums.h and needs to reexport the enum definitions in order for the SparseTensorUtils.h functions to be usable, so the stuff in Enums.h ends up needing to be dllimport/dllexport as well. (And since Enums.h includes ExecutionEngine/Float16bits.h the same thing happens again, though the mlir_float16_utils library can be made a shared lib without complications.)

As I recall, the only thing mlir_sparse_tensor_utils really needs to export as part of its own API are the enum definitions themselves. So I think the quickest fix might be to just remove the MLIR_SPARSETENSOR_EXPORT attribute from all the functions in Enums.h and only leave it on the enums themselves. That does still leave things in an awkward state, but if it makes things green then that'll buy time to come up with a better fix. Though I'm not sure if leaving the MLIR_SPARSETENSOR_EXPORT attribute on the enum definitions would still cause the sorts of problems you're seeing. This is my first time working with DLLs on Windows, so any enlightenment you can share re best practices would be most welcome.

Also, it seems like the LLVM Windows buildbot is a lot stricter than the MLIR/Phabricator Windows buildbot. Is there any way I can send differentials to be run by the LLVM buildbot before landing them?

This can be fixed in a couple of ways:

1. Assume that mlir_sparse_tensor_utils is a static library (which it is) and then remove any export/import definitions. This should be fairly straightforward. In this case we might want to consider the library naming as well since mlir_xxx_utils are generally shared libraries.
2. Make mlir_sparse_tensor_utils a shared library and make all the necessary fixes for it to work. One caveat here is that we'd end up with a situation where one of the transform libraries (and as far as I can tell only one of the transform libraries) requires linking to a shared library.

I suppose a third solution would be to move the definitions from Enums.h to the existing MLIRSparseTensorUtils which is also a static library.

Thoughts?

Just as a bit of background: The reason we split mlir_sparse_tensor_utils out from mlir_c_runner_utils is because another team at Google wanted to be able to use the C++ library directly rather than going through the C-API exposed by mlir_c_runner_utils. That team doesn't care whether the library is shared vs static nor care about the Windows side, so whatever solution we can come up with should be fine for their needs. Originally I intended to make it a shared library, following the pattern of all the other mlir_xxx_utils libraries under the ExecutionEngine directory, but kept running into MSVC linker errors. Eventually I learned about dllimport/dllexport, which solved most of the problems— but ultimately had to make the library static because std::vector isn't set up to be used/exposed by DLLs.

TBH, I'm not really sure whether mlir_sparse_tensor_utils "should" be static or dynamic (nor what MLIR's stance is on such things). We definitely want to be able to share the enums between both mlir_c_runner_utils and SparseTensorTransforms, since the transforms need to be able to be able to codegen calls into the runtime library, so both the transforms and the runtime need to agree on the enums. And I'd really like to keep the predicate functions side-by-side with the enums, since that helps ensure they remain in sync as we evolve and change the enum definitions (e.g., as in D135004); though the predicate functions only need to be called by SparseTensorTransforms and mlir_sparse_tensor_utils itself, they don't need to be exported from mlir_c_runner_utils. I suppose we could split Enums.h off into a separate library from the rest of mlir_sparse_tensor_utils, but that doesn't help the latter become a sharedlib since there's still the std::vector issue.

Moving things into MLIRSparseTensorUtils is a no-go, since that's a completely unrelated library. (This is an example of why I hate using "utils" rather than more expressive names :) Though I'm totally cool with renaming mlir_sparse_tensor_utils to avoid the name causing any confusion about whether it's a sharedlib or not.

[wrengr]

I just uploaded D135502 to see if that fixes things. (Apparently enum class definitions don't need the dllexport/dllimport stuff?) So let's continue the conversation on that differential

mlir/include/mlir/ExecutionEngine/SparseTensor/Storage.h

 class SparseTensorEnumeratorBase;
 
 // These macros ensure consistent error messages, without risk of incuring
 // an additional method call to do so.
 #define ASSERT_VALID_DIM(d)                                                    \
   assert(d < getRank() && "Dimension index is out of bounds");
 #define ASSERT_COMPRESSED_DIM(d)                                               \
   assert(isCompressedDim(d) && "Dimension is not compressed");
+#define ASSERT_COMPRESSED_OR_SINGLETON_DIM(d)                                  \
+  assert((isCompressedDim(d) || isSingletonDim(d)) &&                          \
+         "Dimension is neither compressed nor singleton");
 #define ASSERT_DENSE_DIM(d) assert(isDenseDim(d) && "Dimension is not dense");
 
 /// Abstract base class for `SparseTensorStorage<P,I,V>`.  This class
 /// takes responsibility for all the `<P,I,V>`-independent aspects
 /// of the tensor (e.g., shape, sparsity, permutation).  In addition,
 /// we use function overloading to implement "partial" method
 /// specialization, which the C-API relies on to catch type errors
 /// arising from our use of opaque pointers.
#define DECL_GETPOINTERS(PNAME, P)                                             \
   FOREVERY_FIXED_O(DECL_GETPOINTERS)
 #undef DECL_GETPOINTERS
 
   /// Indices-overhead storage.
 #define DECL_GETINDICES(INAME, I)                                              \
   virtual void getIndices(std::vector<I> **, uint64_t);
   FOREVERY_FIXED_O(DECL_GETINDICES)
 #undef DECL_GETINDICES
+  virtual uint64_t getIndex(uint64_t d, uint64_t pos) const = 0;
 
   /// Primary storage.
 #define DECL_GETVALUES(VNAME, V) virtual void getValues(std::vector<V> **);
   FOREVERY_V(DECL_GETVALUES)
 #undef DECL_GETVALUES
 
   /// Element-wise insertion in lexicographic index order.
 #define DECL_LEXINSERT(VNAME, V) virtual void lexInsert(const uint64_t *, V);
void getPointers(std::vector<P> **out, uint64_t d) final {
     *out = &pointers[d];
   }
   void getIndices(std::vector<I> **out, uint64_t d) final {
     ASSERT_VALID_DIM(d);
     *out = &indices[d];
   }
   void getValues(std::vector<V> **out) final { *out = &values; }
 
+  uint64_t getIndex(uint64_t d, uint64_t pos) const final {
+    ASSERT_COMPRESSED_OR_SINGLETON_DIM(d);
+    assert(pos < indices[d].size() && "Index position is out of bounds");
+    return indices[d][pos]; // Converts the stored `I` into `uint64_t`.
+  }
+
   /// Partially specialize lexicographical insertions based on template types.
   void lexInsert(const uint64_t *cursor, V val) final {
     // First, wrap up pending insertion path.
     uint64_t diff = 0;
     uint64_t top = 0;
     if (!values.empty()) {
       diff = lexDiff(cursor);
       endPath(diff + 1);
void appendIndex(uint64_t d, uint64_t full, uint64_t i) {
     }
   }
 
   /// Writes the given coordinate to `indices[d][pos]`.  This method
   /// checks that `i` is representable in the `I` type; however, it
   /// does not check that `i` is semantically valid (i.e., in bounds
   /// for `dimSizes[d]` and not elsewhere occurring in the same segment).
   void writeIndex(uint64_t d, uint64_t pos, uint64_t i) {
-    ASSERT_COMPRESSED_DIM(d);
+    ASSERT_COMPRESSED_OR_SINGLETON_DIM(d);
     // Subscript assignment to `std::vector` requires that the `pos`-th
     // entry has been initialized; thus we must be sure to check `size()`
     // here, instead of `capacity()` as would be ideal.
     assert(pos < indices[d].size() && "Index position is out of bounds");
     assert(i <= std::numeric_limits<I>::max() &&
            "Index value is too large for the I-type");
     indices[d][pos] = static_cast<I>(i);
   }
 
   /// Computes the assembled-size associated with the `d`-th dimension,
   /// given the assembled-size associated with the `(d-1)`-th dimension.
   /// "Assembled-sizes" correspond to the (nominal) sizes of overhead
   /// storage, as opposed to "dimension-sizes" which are the cardinality
   /// of coordinates for that dimension.
   ///
   /// Precondition: the `pointers[d]` array must be fully initialized
   /// before calling this method.
   uint64_t assembledSize(uint64_t parentSz, uint64_t d) const {
     if (isCompressedDim(d))
       return pointers[d][parentSz];
-    // else if dense:
+    if (isSingletonDim(d))
+      return parentSz; // New size is same as the parent.
+    if (isDenseDim(d))
       return parentSz * getDimSizes()[d];
+    MLIR_SPARSETENSOR_FATAL("unsupported dimension level type");
   }
 
   /// Initializes sparse tensor storage scheme from a memory-resident sparse
   /// tensor in coordinate scheme. This method prepares the pointers and
   /// indices arrays under the given per-dimension dense/sparse annotations.
   ///
   /// Preconditions:
   /// (1) the `elements` must be lexicographically sorted.
private:
 
   /// Finalize the sparse pointer structure at this dimension.
   void finalizeSegment(uint64_t d, uint64_t full = 0, uint64_t count = 1) {
     if (count == 0)
       return; // Short-circuit, since it'll be a nop.
     if (isCompressedDim(d)) {
       appendPointer(d, indices[d].size(), count);
     } else if (isSingletonDim(d)) {
-      return;
+      return; // Nothing to finalize.
     } else { // Dense dimension.
       ASSERT_DENSE_DIM(d);
       const uint64_t sz = getDimSizes()[d];
       assert(sz >= full && "Segment is overfull");
       count = detail::checkedMul(count, sz - full);
       // For dense storage we must enumerate all the remaining coordinates
       // in this dimension (i.e., coordinates after the last non-zero
       // element), and either fill in their zero values or else recurse
void endPath(uint64_t diff) {
     for (uint64_t i = 0; i < rank - diff; ++i) {
       const uint64_t d = rank - i - 1;
       finalizeSegment(d, idx[d] + 1);
     }
   }
 
   /// Continues a single insertion path, outer to inner.
   void insPath(const uint64_t *cursor, uint64_t diff, uint64_t top, V val) {
-    ASSERT_VALID_DIM(diff);
     const uint64_t rank = getRank();
+    assert(diff <= rank && "Dimension-diff is out of bounds");
     for (uint64_t d = diff; d < rank; ++d) {
       const uint64_t i = cursor[d];
       appendIndex(d, top, i);
       top = 0;
       idx[d] = i;
     }
     values.push_back(val);
   }
private:
   friend class SparseTensorEnumerator<P, I, V>;
 
   std::vector<std::vector<P>> pointers;
   std::vector<std::vector<I>> indices;
   std::vector<V> values;
   std::vector<uint64_t> idx; // index cursor for lexicographic insertion.
 };
 
+#undef ASSERT_COMPRESSED_OR_SINGLETON_DIM
 #undef ASSERT_COMPRESSED_DIM
 #undef ASSERT_VALID_DIM
 
 //===----------------------------------------------------------------------===//
 /// A (higher-order) function object for enumerating the elements of some
 /// `SparseTensorStorage` under a permutation.  That is, the `forallElements`
 /// method encapsulates the loop-nest for enumerating the elements of
 /// the source tensor (in whatever order is best for the source tensor),
if (d == Base::getRank()) {
       const std::vector<I> &indicesD = src.indices[d];
       assert(pstop <= indicesD.size() && "Index position is out of bounds");
       uint64_t &cursorReordD = this->cursor[this->reord[d]];
       for (uint64_t pos = pstart; pos < pstop; ++pos) {
         cursorReordD = static_cast<uint64_t>(indicesD[pos]);
         forallElements(yield, pos, d + 1);
       }
     } else if (src.isSingletonDim(d)) {
-      MLIR_SPARSETENSOR_FATAL("unsupported dimension level type");
+      this->cursor[this->reord[d]] = src.getIndex(d, parentPos);
+      forallElements(yield, parentPos, d + 1);
     } else { // Dense dimension.
       assert(src.isDenseDim(d)); // TODO: reuse the ASSERT_DENSE_DIM message
       const uint64_t sz = src.getDimSizes()[d];
       const uint64_t pstart = parentPos * sz;
       uint64_t &cursorReordD = this->cursor[this->reord[d]];
       for (uint64_t i = 0; i < sz; ++i) {
         cursorReordD = i;
         forallElements(yield, pstart + i, d + 1);
SparseTensorStorage<P, I, V> *SparseTensorStorage<P, I, V>::newSparseTensor(
     const DimLevelType *sparsity, SparseTensorCOO<V> *coo) {
   if (coo) {
     const auto &coosz = coo->getDimSizes();
 #ifndef NDEBUG
     detail::assertPermutedSizesMatchShape(coosz, rank, perm, shape);
 #endif
     return new SparseTensorStorage<P, I, V>(coosz, perm, sparsity, coo);
   }
-  // else
   std::vector<uint64_t> permsz(rank);
   for (uint64_t r = 0; r < rank; ++r) {
     assert(shape[r] > 0 && "Dimension size zero has trivial storage");
     permsz[perm[r]] = shape[r];
   }
   // We pass the null `coo` to ensure we select the intended constructor.
   return new SparseTensorStorage<P, I, V>(permsz, perm, sparsity, coo);
 }
for (uint64_t rank = getRank(), r = 0; r < rank; ++r) {
       }
       // Update assembled-size for the next iteration.
       parentSz = assembledSize(parentSz, r);
       // Ideally we need only `indices[r].reserve(parentSz)`, however
       // the `std::vector` implementation forces us to initialize it too.
       // That is, in the yieldPos loop we need random-access assignment
       // to `indices[r]`; however, `std::vector`'s subscript-assignment
       // only allows assigning to already-initialized positions.
-      if (isCompressedDim(r))
+      if (isCompressedDim(r) || isSingletonDim(r))
         indices[r].resize(parentSz, 0);
     }
     values.resize(parentSz, 0); // Both allocate and zero-initialize.
   }
   // The yieldPos loop
   enumerator->forallElements([this](const std::vector<uint64_t> &ind, V val) {
     uint64_t parentSz = 1, parentPos = 0;
     for (uint64_t rank = getRank(), r = 0; r < rank; ++r) {
       if (isCompressedDim(r)) {
         // If `parentPos == parentSz` then it's valid as an array-lookup;
         // however, it's semantically invalid here since that entry
         // does not represent a segment of `indices[r]`.  Moreover, that
         // entry must be immutable for `assembledSize` to remain valid.
         assert(parentPos < parentSz && "Pointers position is out of bounds");
         const uint64_t currentPos = pointers[r][parentPos];
         // This increment won't overflow the `P` type, since it can't
         // exceed the original value of `pointers[r][parentPos+1]`
         // which was already verified to be within bounds for `P`
         // when it was written to the array.
         pointers[r][parentPos]++;
         writeIndex(r, currentPos, ind[r]);
         parentPos = currentPos;
       } else if (isSingletonDim(r)) {
+        writeIndex(r, parentPos, ind[r]);
         // the new parentPos equals the old parentPos.
       } else { // Dense dimension.
         ASSERT_DENSE_DIM(r);
         parentPos = parentPos * getDimSizes()[r] + ind[r];
       }
       parentSz = assembledSize(parentSz, r);
     }
     assert(parentPos < values.size() && "Value position is out of bounds");
if (isCompressedDim(r)) {
              "Pointers got corrupted");
       // TODO: optimize this by using `memmove` or similar.
       for (uint64_t n = 0; n < parentSz; ++n) {
         const uint64_t parentPos = parentSz - n;
         pointers[r][parentPos] = pointers[r][parentPos - 1];
       }
       pointers[r][0] = 0;
     }
+    // Both dense and singleton are no-ops for the finalizeYieldPos loop
     parentSz = assembledSize(parentSz, r);
   }
 }
 
+#undef ASSERT_DENSE_DIM
+
 } // namespace sparse_tensor
 } // namespace mlir
 
 #undef ASSERT_DENSE_DIM
 
 #endif // MLIR_EXECUTIONENGINE_SPARSETENSOR_STORAGE_H

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorConversion.cpp

 }
 
 /// Determine if the runtime library supports direct conversion to the
 /// given target `dimTypes`.
 static bool canUseDirectConversion(
     ArrayRef<SparseTensorEncodingAttr::DimLevelType> dimTypes) {
   bool alreadyCompressed = false;
   for (uint64_t rank = dimTypes.size(), r = 0; r < rank; r++) {
-    switch (dimTypes[r]) {
-    case SparseTensorEncodingAttr::DimLevelType::Compressed:
+    const DimLevelType dlt = dimLevelTypeEncoding(dimTypes[r]);
+    if (isCompressedDLT(dlt)) {
       if (alreadyCompressed)
         return false; // Multiple compressed dimensions not yet supported.
       alreadyCompressed = true;
-      break;
-    case SparseTensorEncodingAttr::DimLevelType::Dense:
+    } else if (isDenseDLT(dlt)) {
       if (alreadyCompressed)
         return false; // Dense after Compressed not yet supported.
-      break;
-    default: // TODO: investigate
+    } else if (isSingletonDLT(dlt)) {
+      // Direct conversion doesn't have any particular problems with
+      // singleton after compressed.
+    } else { // TODO: investigate
       return false;
     }
   }
   return true;
 }
 
 /// Helper method to translate indices during a reshaping operation.
 /// TODO: provide as general utility to MLIR at large?

mlir/lib/ExecutionEngine/SparseTensor/NNZ.cpp

 /// does not actually populate the statistics, however; for that see
 /// `initialize`.
 ///
 /// Precondition: `dimSizes` must not contain zeros.
 SparseTensorNNZ::SparseTensorNNZ(const std::vector<uint64_t> &dimSizes,
                                  const std::vector<DimLevelType> &sparsity)
     : dimSizes(dimSizes), dimTypes(sparsity), nnz(getRank()) {
   assert(dimSizes.size() == dimTypes.size() && "Rank mismatch");
-  bool uncompressed = true;
-  (void)uncompressed;
+  bool alreadyCompressed = false;
+  (void)alreadyCompressed;
   uint64_t sz = 1; // the product of all `dimSizes` strictly less than `r`.
   for (uint64_t rank = getRank(), r = 0; r < rank; r++) {
-    switch (dimTypes[r]) {
-    case DimLevelType::kCompressed:
-      assert(uncompressed &&
-             "Multiple compressed layers not currently supported");
-      uncompressed = false;
+    const DimLevelType dlt = sparsity[r];
+    if (isCompressedDLT(dlt)) {
+      if (alreadyCompressed)
+        MLIR_SPARSETENSOR_FATAL(
+      alreadyCompressed = true;
       nnz[r].resize(sz, 0); // Both allocate and zero-initialize.
-      break;
-    case DimLevelType::kDense:
-      assert(uncompressed && "Dense after compressed not currently supported");
-      break;
-    case DimLevelType::kSingleton:
+    } else if (isDenseDLT(dlt)) {
+      if (alreadyCompressed)
+        MLIR_SPARSETENSOR_FATAL(
+            "Dense after compressed not currently supported");
+    } else if (isSingletonDLT(dlt)) {
       // Singleton after Compressed causes no problems for allocating
       // `nnz` nor for the yieldPos loop.  This remains true even
       // when adding support for multiple compressed dimensions or
       // for dense-after-compressed.
-      break;
-    default:
-      MLIR_SPARSETENSOR_FATAL("unsupported dimension level type");
+    } else {
+      MLIR_SPARSETENSOR_FATAL("unsupported dimension level type: %d\n",
+                              static_cast<uint8_t>(dlt));
     }
     sz = detail::checkedMul(sz, dimSizes[r]);
   }
 }
 
 /// Lexicographically enumerates all indicies for dimensions strictly
 /// less than `stopDim`, and passes their nnz statistic to the callback.
 /// Since our use-case only requires the statistic not the coordinates
 /// themselves, we do not bother to construct those coordinates.
 void SparseTensorNNZ::forallIndices(uint64_t stopDim,
                                     SparseTensorNNZ::NNZConsumer yield) const {
   assert(stopDim < getRank() && "Dimension out of bounds");
-  assert(dimTypes[stopDim] == DimLevelType::kCompressed &&
+  assert(isCompressedDLT(dimTypes[stopDim]) &&
          "Cannot look up non-compressed dimensions");
   forallIndices(yield, stopDim, 0, 0);
 }
 
 /// Adds a new element (i.e., increment its statistics).  We use
 /// a method rather than inlining into the lambda in `initialize`,
 /// to avoid spurious templating over `V`.  And this method is private
 /// to avoid needing to re-assert validity of `ind` (which is guaranteed
 /// by `forallElements`).
 void SparseTensorNNZ::add(const std::vector<uint64_t> &ind) {
   uint64_t parentPos = 0;
   for (uint64_t rank = getRank(), r = 0; r < rank; ++r) {
-    if (dimTypes[r] == DimLevelType::kCompressed)
+    if (isCompressedDLT(dimTypes[r]))
       nnz[r][parentPos]++;
     parentPos = parentPos * dimSizes[r] + ind[r];
   }
 }
 
 /// Recursive component of the public `forallIndices`.
 void SparseTensorNNZ::forallIndices(SparseTensorNNZ::NNZConsumer yield,
                                     uint64_t stopDim, uint64_t parentPos,

mlir/test/Dialect/SparseTensor/conversion_sparse2sparse.mlir

 //   CHECK-AUTO-DAG: %[[Y:.*]] = memref.cast %[[Q]] : memref<1xindex> to memref<?xindex>
 //   CHECK-AUTO-DAG: %[[Z:.*]] = memref.cast %[[R]] : memref<1xindex> to memref<?xindex>
 //       CHECK-AUTO: %[[T:.*]] = call @newSparseTensor(%[[X]], %[[Y]], %[[Z]], %{{.*}}, %{{.*}}, %{{.*}}, %[[SparseToSparse]], %[[A]])
 //       CHECK-AUTO: return %[[T]] : !llvm.ptr<i8>
 func.func @sparse_convert(%arg0: tensor<?xf32, #SparseVector64>) -> tensor<?xf32, #SparseVector32> {
   %0 = sparse_tensor.convert %arg0 : tensor<?xf32, #SparseVector64> to tensor<?xf32, #SparseVector32>
   return %0 : tensor<?xf32, #SparseVector32>
 }
+
+#SparseSingleton64 = #sparse_tensor.encoding<{
+  dimLevelType = ["singleton"],
+  pointerBitWidth = 64,
+  indexBitWidth = 64
+}>
+
+#SparseSingleton32 = #sparse_tensor.encoding<{
+  dimLevelType = ["singleton"],
+  pointerBitWidth = 32,
+  indexBitWidth = 32
+}>
+
+// CHECK-COO-LABEL: func @sparse_convert_singleton(
+//  CHECK-COO-SAME: %[[A:.*]]: !llvm.ptr<i8>)
+//  CHECK-COO-DAG:  %[[ToCOO:.*]] = arith.constant 5 : i32
+//  CHECK-COO-DAG:  %[[FromCOO:.*]] = arith.constant 2 : i32
+//   CHECK-COO-DAG: %[[P:.*]] = memref.alloca() : memref<1xi8>
+//   CHECK-COO-DAG: %[[Q:.*]] = memref.alloca() : memref<1xindex>
+//   CHECK-COO-DAG: %[[R:.*]] = memref.alloca() : memref<1xindex>
+//   CHECK-COO-DAG: %[[X:.*]] = memref.cast %[[P]] : memref<1xi8> to memref<?xi8>
+//   CHECK-COO-DAG: %[[Y:.*]] = memref.cast %[[Q]] : memref<1xindex> to memref<?xindex>
+//   CHECK-COO-DAG: %[[Z:.*]] = memref.cast %[[R]] : memref<1xindex> to memref<?xindex>
+//       CHECK-COO: %[[C:.*]] = call @newSparseTensor(%[[X]], %[[Y]], %[[Z]], %{{.*}}, %{{.*}}, %{{.*}}, %[[ToCOO]], %[[A]])
+//       CHECK-COO: %[[T:.*]] = call @newSparseTensor(%[[X]], %[[Y]], %[[Z]], %{{.*}}, %{{.*}}, %{{.*}}, %[[FromCOO]], %[[C]])
+//       CHECK-COO: call @delSparseTensorCOOF32(%[[C]])
+//       CHECK-COO: return %[[T]] : !llvm.ptr<i8>
+// CHECK-AUTO-LABEL: func @sparse_convert_singleton(
+//  CHECK-AUTO-SAME: %[[A:.*]]: !llvm.ptr<i8>)
+//   CHECK-AUTO-DAG: %[[SparseToSparse:.*]] = arith.constant 3 : i32
+//   CHECK-AUTO-DAG: %[[P:.*]] = memref.alloca() : memref<1xi8>
+//   CHECK-AUTO-DAG: %[[Q:.*]] = memref.alloca() : memref<1xindex>
+//   CHECK-AUTO-DAG: %[[R:.*]] = memref.alloca() : memref<1xindex>
+//   CHECK-AUTO-DAG: %[[X:.*]] = memref.cast %[[P]] : memref<1xi8> to memref<?xi8>
+//   CHECK-AUTO-DAG: %[[Y:.*]] = memref.cast %[[Q]] : memref<1xindex> to memref<?xindex>
+//   CHECK-AUTO-DAG: %[[Z:.*]] = memref.cast %[[R]] : memref<1xindex> to memref<?xindex>
+//       CHECK-AUTO: %[[T:.*]] = call @newSparseTensor(%[[X]], %[[Y]], %[[Z]], %{{.*}}, %{{.*}}, %{{.*}}, %[[SparseToSparse]], %[[A]])
+//       CHECK-AUTO: return %[[T]] : !llvm.ptr<i8>
+func.func @sparse_convert_singleton(%arg0: tensor<?xf32, #SparseSingleton64>) -> tensor<?xf32, #SparseSingleton32> {
+  %0 = sparse_tensor.convert %arg0 : tensor<?xf32, #SparseSingleton64> to tensor<?xf32, #SparseSingleton32>
+  return %0 : tensor<?xf32, #SparseSingleton32>
+}

mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_conversion_sparse2sparse.mlir

#Tensor2 = #sparse_tensor.encoding<{
   dimLevelType = [ "dense", "compressed", "dense" ]
 }>
 
 #Tensor3 = #sparse_tensor.encoding<{
   dimLevelType = [ "dense", "dense", "compressed" ],
   dimOrdering = affine_map<(i,j,k) -> (i,k,j)>
 }>
 
+#SingletonTensor1 = #sparse_tensor.encoding<{
+  dimLevelType = [ "dense", "compressed", "singleton" ]
+}>
+
+// This also checks the compressed->dense conversion (when there are zeros).
+#SingletonTensor2 = #sparse_tensor.encoding<{
+  dimLevelType = [ "dense", "dense", "singleton" ]
+}>
+
+// This also checks the singleton->compressed conversion.
+#SingletonTensor3 = #sparse_tensor.encoding<{
+  dimLevelType = [ "dense", "dense", "compressed" ]
+}>
+
 module {
   //
   // Utilities for output and releasing memory.
   //
   func.func @dump(%arg0: tensor<2x3x4xf64>) {
     %c0 = arith.constant 0 : index
     %d0 = arith.constant -1.0 : f64
     %0 = vector.transfer_read %arg0[%c0, %c0, %c0], %d0: tensor<2x3x4xf64>, vector<2x3x4xf64>
     vector.print %0 : vector<2x3x4xf64>
     return
   }
   func.func @dumpAndRelease_234(%arg0: tensor<2x3x4xf64>) {
     call @dump(%arg0) : (tensor<2x3x4xf64>) -> ()
     return
   }
 
   //
-  // Main driver.
+  // The first test suite (for non-singleton DimLevelTypes).
   //
-  func.func @entry() {
+  func.func @testNonSingleton() {
     //
     // Initialize a 3-dim dense tensor.
     //
     %src = arith.constant dense<[
        [  [  1.0,  2.0,  3.0,  4.0 ],
           [  5.0,  6.0,  7.0,  8.0 ],
           [  9.0, 10.0, 11.0, 12.0 ] ],
        [  [ 13.0, 14.0, 15.0, 16.0 ],
func.func @testNonSingleton() {
     bufferization.dealloc_tensor %t23 : tensor<2x3x4xf64, #Tensor3>
     bufferization.dealloc_tensor %t31 : tensor<2x3x4xf64, #Tensor1>
     bufferization.dealloc_tensor %s1 : tensor<2x3x4xf64, #Tensor1>
     bufferization.dealloc_tensor %s2 : tensor<2x3x4xf64, #Tensor2>
     bufferization.dealloc_tensor %s3 : tensor<2x3x4xf64, #Tensor3>
 
     return
   }
+
+  //
+  // The second test suite (for singleton DimLevelTypes).
+  //
+  func.func @testSingleton() {
+    //
+    // Initialize a 3-dim dense tensor with the 3rd dim being singleton.
+    //
+    %src = arith.constant dense<[
+       [  [  1.0,  0.0,  0.0,  0.0 ],
+          [  0.0,  6.0,  0.0,  0.0 ],
+          [  0.0,  0.0, 11.0,  0.0 ] ],
+       [  [  0.0, 14.0,  0.0,  0.0 ],
+          [  0.0,  0.0,  0.0, 20.0 ],
+          [ 21.0,  0.0,  0.0,  0.0 ] ]
+    ]> : tensor<2x3x4xf64>
+
+    //
+    // Convert dense tensor directly to various sparse tensors.
+    //
+    %s1 = sparse_tensor.convert %src : tensor<2x3x4xf64> to tensor<2x3x4xf64, #SingletonTensor1>
+    %s2 = sparse_tensor.convert %src : tensor<2x3x4xf64> to tensor<2x3x4xf64, #SingletonTensor2>
+    %s3 = sparse_tensor.convert %src : tensor<2x3x4xf64> to tensor<2x3x4xf64, #SingletonTensor3>
+
+    //
+    // Convert sparse tensor directly to another sparse format.
+    //
+    %t12 = sparse_tensor.convert %s1 : tensor<2x3x4xf64, #SingletonTensor1> to tensor<2x3x4xf64, #SingletonTensor2>
+    %t13 = sparse_tensor.convert %s1 : tensor<2x3x4xf64, #SingletonTensor1> to tensor<2x3x4xf64, #SingletonTensor3>
+    %t21 = sparse_tensor.convert %s2 : tensor<2x3x4xf64, #SingletonTensor2> to tensor<2x3x4xf64, #SingletonTensor1>
+    %t23 = sparse_tensor.convert %s2 : tensor<2x3x4xf64, #SingletonTensor2> to tensor<2x3x4xf64, #SingletonTensor3>
+    %t31 = sparse_tensor.convert %s3 : tensor<2x3x4xf64, #SingletonTensor3> to tensor<2x3x4xf64, #SingletonTensor1>
+    %t32 = sparse_tensor.convert %s3 : tensor<2x3x4xf64, #SingletonTensor3> to tensor<2x3x4xf64, #SingletonTensor2>
+
+    //
+    // Convert sparse tensor back to dense.
+    //
+    %d12 = sparse_tensor.convert %t12 : tensor<2x3x4xf64, #SingletonTensor2> to tensor<2x3x4xf64>
+    %d13 = sparse_tensor.convert %t13 : tensor<2x3x4xf64, #SingletonTensor3> to tensor<2x3x4xf64>
+    %d21 = sparse_tensor.convert %t21 : tensor<2x3x4xf64, #SingletonTensor1> to tensor<2x3x4xf64>
+    %d23 = sparse_tensor.convert %t23 : tensor<2x3x4xf64, #SingletonTensor3> to tensor<2x3x4xf64>
+    %d31 = sparse_tensor.convert %t31 : tensor<2x3x4xf64, #SingletonTensor1> to tensor<2x3x4xf64>
+    %d32 = sparse_tensor.convert %t32 : tensor<2x3x4xf64, #SingletonTensor2> to tensor<2x3x4xf64>
+
+    //
+    // Check round-trip equality.  And release dense tensors.
+    //
+    // CHECK-COUNT-7: ( ( ( 1, 0, 0, 0 ), ( 0, 6, 0, 0 ), ( 0, 0, 11, 0 ) ), ( ( 0, 14, 0, 0 ), ( 0, 0, 0, 20 ), ( 21, 0, 0, 0 ) ) )
+    call @dump(%src) : (tensor<2x3x4xf64>) -> ()
+    call @dumpAndRelease_234(%d12) : (tensor<2x3x4xf64>) -> ()
+    call @dumpAndRelease_234(%d13) : (tensor<2x3x4xf64>) -> ()
+    call @dumpAndRelease_234(%d21) : (tensor<2x3x4xf64>) -> ()
+    call @dumpAndRelease_234(%d23) : (tensor<2x3x4xf64>) -> ()
+    call @dumpAndRelease_234(%d31) : (tensor<2x3x4xf64>) -> ()
+    call @dumpAndRelease_234(%d32) : (tensor<2x3x4xf64>) -> ()
+
+    //
+    // Release sparse tensors.
+    //
+    bufferization.dealloc_tensor %t12 : tensor<2x3x4xf64, #SingletonTensor2>
+    bufferization.dealloc_tensor %t13 : tensor<2x3x4xf64, #SingletonTensor3>
+    bufferization.dealloc_tensor %t21 : tensor<2x3x4xf64, #SingletonTensor1>
+    bufferization.dealloc_tensor %t23 : tensor<2x3x4xf64, #SingletonTensor3>
+    bufferization.dealloc_tensor %t31 : tensor<2x3x4xf64, #SingletonTensor1>
+    bufferization.dealloc_tensor %t32 : tensor<2x3x4xf64, #SingletonTensor2>
+
+    return
+  }
+
+  //
+  // Main driver.
+  //
+  func.func @entry() {
+    call @testNonSingleton() : () -> ()
+    call @testSingleton() : () -> ()
+    return
+  }
 }

mlir/test/Integration/Dialect/SparseTensor/python/test_stress.py

def main():
   print("\nTEST: test_stress")
   with ir.Context() as ctx, ir.Location.unknown():
     vl = 1
     e = False
     # Disable direct sparse2sparse conversion, because it doubles the time!
     # TODO: While direct s2s is far too slow for per-commit testing,
     # we should have some framework ensure that we run this test with
     # `s2s=0` on a regular basis, to ensure that it does continue to work.
+    # TODO: be sure to test s2s=0 together with singletons.
     s2s = 1
     sparsification_options = (
         f'parallelization-strategy=none '
         f'vectorization-strategy=none '
         f'vl={vl} '
         f'enable-simd-index32={e} '
         f's2s-strategy={s2s}')
     compiler = sparse_compiler.SparseCompiler(
         options=sparsification_options, opt_level=0, shared_libs=[support_lib])
     f64 = ir.F64Type.get()
     # Be careful about increasing this because
-    #     len(types) = 1 + 2^rank * rank! * len(bitwidths)^2
+    #     len(types) = 1 + len(level_choices)^rank * rank! * len(bitwidths)^2
     shape = range(2, 6)
     rank = len(shape)
     # All combinations.
+    # TODO: add singleton here too; which requires updating how `np_arg0`
+    # is initialized below.
     levels = list(itertools.product(*itertools.repeat(
       [st.DimLevelType.dense, st.DimLevelType.compressed], rank)))
     # All permutations.
     orderings = list(map(ir.AffineMap.get_permutation,
       itertools.permutations(range(rank))))
     bitwidths = [0]
     # The first type must be a dense tensor for numpy conversion to work.
     types = [ir.RankedTensorType.get(shape, f64)]