[mlir][sparse] external entry method wrapper for sparse tensors (#80326)

Similar to the emit_c_interface, this pull request adds a pass that
converts public entry methods that use sparse tensors as input
parameters and/or output return values into wrapper functions that
[dis]assemble the individual tensors that constitute the actual storage
used externally into MLIR sparse tensors. This pass can be used to
prepare the public entry methods of a program that is compiled by the
MLIR sparsifier to interface with an external runtime, e.g., when
passing sparse tensors as numpy arrays from and to Python. Note that
eventual bufferization decisions (e.g. who [de]allocates the underlying
memory) should be resolved in agreement with the external runtime
(Python, PyTorch, JAX, etc.)

mlir/include/mlir/Dialect/SparseTensor/Transforms/Passes.h

@@ -56,6 +56,14 @@ enum class SparseEmitStrategy {
 #define GEN_PASS_DECL
 #include "mlir/Dialect/SparseTensor/Transforms/Passes.h.inc"
 
+//===----------------------------------------------------------------------===//
+// The SparseAssembler pass.
+//===----------------------------------------------------------------------===//
+
+void populateSparseAssembler(RewritePatternSet &patterns);
+
+std::unique_ptr<Pass> createSparseAssembler();
+
 //===----------------------------------------------------------------------===//
 // The SparseReinterpretMap pass.
 //===----------------------------------------------------------------------===//

mlir/include/mlir/Dialect/SparseTensor/Transforms/Passes.td

@@ -11,6 +11,26 @@
 
 include "mlir/Pass/PassBase.td"
 
+def SparseAssembler : Pass<"sparse-assembler", "ModuleOp"> {
+  let summary = "Add [dis]assemble operations on external sparse tensors";
+  let description = [{
+    A pass that converts public entry methods that use sparse tensors as
+    input parameters and/or output return values into wrapper functions
+    that [dis]assemble the individual tensors that constitute the actual
+    storage used externally into MLIR sparse tensors. This pass can be used
+    to prepare the public entry methods of a program that is compiled by the
+    MLIR sparsifier to interface with an external runtime, e.g., when passing
+    sparse tensors as numpy arrays from and to Python. Note that eventual
+    bufferization decisions (e.g. who [de]allocates the underlying memory)
+    should be resolved in agreement with the external runtime.
+  }];
+  let constructor = "mlir::createSparseAssembler()";
+  let dependentDialects = [
+    "sparse_tensor::SparseTensorDialect",
+    "tensor::TensorDialect",
+  ];
+}
+
 def SparseReinterpretMap : Pass<"sparse-reinterpret-map", "ModuleOp"> {
   let summary = "Reinterprets sparse tensor type mappings";
   let description = [{
@@ -190,7 +210,6 @@ def LowerForeachToSCF : Pass<"lower-sparse-foreach-to-scf", "func::FuncOp"> {
   ];
 }
 
-
 def SparseTensorConversionPass : Pass<"sparse-tensor-conversion", "ModuleOp"> {
   let summary = "Convert sparse tensors and primitives to library calls";
   let description = [{

mlir/lib/Dialect/SparseTensor/Transforms/CMakeLists.txt

@@ -1,6 +1,7 @@
 add_mlir_dialect_library(MLIRSparseTensorTransforms
   # Rewriting.
   BufferizableOpInterfaceImpl.cpp
+  SparseAssembler.cpp
   SparseBufferRewriting.cpp
   SparseGPUCodegen.cpp
   SparseReinterpretMap.cpp

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

@@ -0,0 +1,239 @@
+//===- SparseAssembler.cpp - adds wrapper method around sparse types ------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "Utils/CodegenUtils.h"
+
+#include "mlir/Dialect/SparseTensor/IR/SparseTensor.h"
+#include "mlir/Dialect/SparseTensor/IR/SparseTensorStorageLayout.h"
+#include "mlir/Dialect/SparseTensor/IR/SparseTensorType.h"
+#include "mlir/Dialect/SparseTensor/Transforms/Passes.h"
+#include "mlir/Dialect/Tensor/IR/Tensor.h"
+#include "llvm/Support/FormatVariadic.h"
+
+using namespace mlir;
+using namespace sparse_tensor;
+
+//===----------------------------------------------------------------------===//
+// Helper methods.
+//===----------------------------------------------------------------------===//
+
+// TODO: reuse StorageLayout::foreachField?
+
+// TODO: we need COO AoS and SoA
+
+// Convert type range to new types range, with sparse tensors externalized.
+void convTypes(TypeRange types, SmallVectorImpl<Type> &convTypes,
+               SmallVectorImpl<Type> *extraTypes = nullptr) {
+  for (auto type : types) {
+    // All "dense" data passes through unmodified.
+    if (!getSparseTensorEncoding(type)) {
+      convTypes.push_back(type);
+      continue;
+    }
+    // Convert the external representation of the values array.
+    const SparseTensorType stt(cast<RankedTensorType>(type));
+    auto shape = {ShapedType::kDynamic};
+    auto vtp = RankedTensorType::get(shape, stt.getElementType());
+    convTypes.push_back(vtp);
+    if (extraTypes)
+      extraTypes->push_back(vtp);
+    // Convert the external representations of the pos/crd arrays.
+    for (Level lvl = 0, lvlRank = stt.getLvlRank(); lvl < lvlRank; lvl++) {
+      const auto lt = stt.getLvlType(lvl);
+      if (isCompressedLT(lt) || isLooseCompressedLT(lt)) {
+        auto ptp = RankedTensorType::get(shape, stt.getPosType());
+        auto ctp = RankedTensorType::get(shape, stt.getCrdType());
+        convTypes.push_back(ptp);
+        convTypes.push_back(ctp);
+        if (extraTypes) {
+          extraTypes->push_back(ptp);
+          extraTypes->push_back(ctp);
+        }
+      } else {
+        assert(isDenseLT(lt)); // TODO: handle other cases
+      }
+    }
+  }
+}
+
+// Convert input and output values to [dis[assemble ops for sparse tensors.
+void convVals(OpBuilder &builder, Location loc, TypeRange types,
+              ValueRange fromVals, ValueRange extraVals,
+              SmallVectorImpl<Value> &toVals, unsigned extra, bool isIn) {
+  unsigned idx = 0;
+  for (auto type : types) {
+    // All "dense" data passes through unmodified.
+    if (!getSparseTensorEncoding(type)) {
+      toVals.push_back(fromVals[idx++]);
+      continue;
+    }
+    // Convert the external representation of the values array.
+    auto rtp = cast<RankedTensorType>(type);
+    const SparseTensorType stt(rtp);
+    auto shape = {ShapedType::kDynamic};
+    SmallVector<Value> inputs;
+    SmallVector<Type> retTypes;
+    SmallVector<Type> cntTypes;
+    // Collect the external representation of the values array for
+    // input or the outgoing sparse tensor for output.
+    inputs.push_back(fromVals[idx++]);
+    if (!isIn) {
+      inputs.push_back(extraVals[extra++]);
+      retTypes.push_back(RankedTensorType::get(shape, stt.getElementType()));
+      cntTypes.push_back(builder.getIndexType());
+    }
+    // Collect the external representations of the pos/crd arrays.
+    for (Level lvl = 0, lvlRank = stt.getLvlRank(); lvl < lvlRank; lvl++) {
+      const auto lt = stt.getLvlType(lvl);
+      if (isCompressedLT(lt) || isLooseCompressedLT(lt)) {
+        if (isIn) {
+          inputs.push_back(fromVals[idx++]);
+          inputs.push_back(fromVals[idx++]);
+        } else {
+          Type pTp = stt.getPosType();
+          Type cTp = stt.getCrdType();
+          inputs.push_back(extraVals[extra++]);
+          inputs.push_back(extraVals[extra++]);
+          retTypes.push_back(RankedTensorType::get(shape, pTp));
+          retTypes.push_back(RankedTensorType::get(shape, cTp));
+          cntTypes.push_back(pTp);
+          cntTypes.push_back(cTp);
+        }
+      } else {
+        assert(isDenseLT(lt)); // TODO: handle other cases
+      }
+    }
+    if (isIn) {
+      // Assemble multiple inputs into a single sparse tensor.
+      auto a = builder.create<sparse_tensor::AssembleOp>(loc, rtp, inputs);
+      toVals.push_back(a.getResult());
+    } else {
+      // Disassemble a single sparse input into multiple outputs.
+      // Note that this includes the counters, which are dropped.
+      unsigned len = retTypes.size();
+      retTypes.append(cntTypes);
+      auto d =
+          builder.create<sparse_tensor::DisassembleOp>(loc, retTypes, inputs);
+      for (unsigned i = 0; i < len; i++)
+        toVals.push_back(d.getResult(i));
+    }
+  }
+}
+
+//===----------------------------------------------------------------------===//
+// Rewriting rules.
+//===----------------------------------------------------------------------===//
+
+namespace {
+
+// A rewriting rules that converts public entry methods that use sparse tensors
+// as input parameters and/or output return values into wrapper functions
+// that [dis]assemble the individual tensors that constitute the actual
+// storage used externally into MLIR sparse tensors.
+//
+// In particular, each sparse tensor input
+//
+// void foo(..., t, ...) { }
+//
+// adds the following strucuture in a wrapper
+//
+// void spiface_foo(..., t1..tn, ...) {
+//   t = assemble t1..tn
+//   foo(..., t, ...)
+// }
+//
+// and likewise, each output tensor
+//
+// ... T ... bar(...) { return ..., t, ...; }
+//
+// adds the following structure in a wrapper
+//
+// ... T1..TN ... spiface_bar(..., t1'..tn') {
+//   ..., t, ... = bar(...)
+//   t1..tn = disassemble t, t1'..tn'
+//   return ..., t1..tn, ...
+// }
+//
+// TODO: refine output sparse tensors to work well with external framework
+//
+// TODO: use "inlining" instead of a wrapper?
+//
+struct SparseFuncAssembler : public OpRewritePattern<func::FuncOp> {
+  using OpRewritePattern::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(func::FuncOp funcOp,
+                                PatternRewriter &rewriter) const override {
+    // Only a rewrite an entry with the c-interface requested.
+    if (!funcOp->getAttrOfType<UnitAttr>(
+            LLVM::LLVMDialect::getEmitCWrapperAttrName()))
+      return failure();
+
+    // Translate sparse tensor types to external types.
+    SmallVector<Type> inputTypes;
+    SmallVector<Type> outputTypes;
+    SmallVector<Type> extraTypes;
+    convTypes(funcOp.getArgumentTypes(), inputTypes);
+    convTypes(funcOp.getResultTypes(), outputTypes, &extraTypes);
+
+    // Only sparse inputs or outputs need a wrapper function.
+    if (inputTypes.size() == funcOp.getArgumentTypes().size() &&
+        outputTypes.size() == funcOp.getResultTypes().size())
+      return failure();
+
+    // Start the new wrapper function. Together with the c-interface mangling,
+    // a sparse external entry point eventually will have a name like:
+    //    _mlir_ciface_spiface_XXX(...)
+    Location loc = funcOp.getLoc();
+    ModuleOp modOp = funcOp->getParentOfType<ModuleOp>();
+    MLIRContext *context = modOp.getContext();
+    OpBuilder moduleBuilder(modOp.getBodyRegion());
+    std::string wrapper = llvm::formatv("spiface_{0}", funcOp.getName()).str();
+    unsigned extra = inputTypes.size();
+    inputTypes.append(extraTypes);
+    auto func = moduleBuilder.create<func::FuncOp>(
+        loc, wrapper, FunctionType::get(context, inputTypes, outputTypes));
+    func.setPublic();
+    func->setAttr(LLVM::LLVMDialect::getEmitCWrapperAttrName(),
+                  UnitAttr::get(context));
+
+    // Construct new wrapper function body.
+    auto org = SymbolRefAttr::get(context, funcOp.getName());
+    OpBuilder::InsertionGuard insertionGuard(rewriter);
+    Block *body = func.addEntryBlock();
+    rewriter.setInsertionPointToStart(body);
+
+    // Convert inputs.
+    SmallVector<Value> inputs;
+    convVals(rewriter, loc, funcOp.getArgumentTypes(), body->getArguments(),
+             ValueRange(), inputs, 0, /*isIn=*/true);
+
+    // Call original function.
+    auto call = rewriter.create<func::CallOp>(loc, funcOp.getResultTypes(), org,
+                                              inputs);
+
+    // Convert outputs and return.
+    SmallVector<Value> outputs;
+    convVals(rewriter, loc, funcOp.getResultTypes(), call.getResults(),
+             body->getArguments(), outputs, extra, /*isIn=*/false);
+    rewriter.create<func::ReturnOp>(loc, outputs);
+
+    // Strip the c-interface attribute from the original function.
+    funcOp->removeAttr(LLVM::LLVMDialect::getEmitCWrapperAttrName());
+    return success();
+  }
+};
+
+} // namespace
+
+//===----------------------------------------------------------------------===//
+// Public method for populating conversion rules.
+//===----------------------------------------------------------------------===//
+
+void mlir::populateSparseAssembler(RewritePatternSet &patterns) {
+  patterns.add<SparseFuncAssembler>(patterns.getContext());
+}

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

@@ -22,6 +22,7 @@
 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
 
 namespace mlir {
+#define GEN_PASS_DEF_SPARSEASSEMBLER
 #define GEN_PASS_DEF_SPARSEREINTERPRETMAP
 #define GEN_PASS_DEF_PRESPARSIFICATIONREWRITE
 #define GEN_PASS_DEF_SPARSIFICATIONPASS
@@ -46,6 +47,18 @@ namespace {
 // Passes implementation.
 //===----------------------------------------------------------------------===//
 
+struct SparseAssembler : public impl::SparseAssemblerBase<SparseAssembler> {
+  SparseAssembler() = default;
+  SparseAssembler(const SparseAssembler &pass) = default;
+
+  void runOnOperation() override {
+    auto *ctx = &getContext();
+    RewritePatternSet patterns(ctx);
+    populateSparseAssembler(patterns);
+    (void)applyPatternsAndFoldGreedily(getOperation(), std::move(patterns));
+  }
+};
+
 struct SparseReinterpretMap
     : public impl::SparseReinterpretMapBase<SparseReinterpretMap> {
   SparseReinterpretMap() = default;
@@ -378,6 +391,10 @@ struct StorageSpecifierToLLVMPass
 // Pass creation methods.
 //===----------------------------------------------------------------------===//
 
+std::unique_ptr<Pass> mlir::createSparseAssembler() {
+  return std::make_unique<SparseAssembler>();
+}
+
 std::unique_ptr<Pass> mlir::createSparseReinterpretMapPass() {
   return std::make_unique<SparseReinterpretMap>();
 }