[mlir][sparse] support sparse dilated convolution.

mlir/lib/Dialect/SparseTensor/Transforms/Utils/LoopEmitter.cpp

@@ -313,16 +313,16 @@ void LoopEmitter::initSubSectIterator(OpBuilder &builder, Location loc) {
         // Compute the subsection size.
         Value size = c0;
         for (auto [loop, stride] : remDepStack[t][lvl]) {
-          Value loopHi = loopHighs[loop];
-          size = ADDI(size, MULI(loopHi, C_IDX(stride)));
+          Value idxMax = SUBI(loopHighs[loop], C_IDX(1));
+          size = ADDI(size, ADDI(MULI(idxMax, C_IDX(stride)), C_IDX(1)));
         }
         it = makeNonEmptySubSectIterator(builder, loc, parent, loopHighs[loop],
                                          std::move(lvlIt), size, curDep.second);
       } else {
-        Value size = loopHighs[loop];
         const SparseIterator &subSectIter = *iters[t][lvl].back();
-        it = makeTraverseSubSectIterator(subSectIter, *parent, std::move(lvlIt),
-                                         size, curDep.second);
+        it = makeTraverseSubSectIterator(builder, loc, subSectIter, *parent,
+                                         std::move(lvlIt), loopHighs[loop],
+                                         curDep.second);
       }
       lastIter[t] = it.get();
       iters[t][lvl].emplace_back(std::move(it));

mlir/lib/Dialect/SparseTensor/Transforms/Utils/SparseTensorLevel.cpp

@@ -665,13 +665,10 @@ class SubSectIterator : public SparseIterator {
 public:
   SubSectIterator(const NonEmptySubSectIterator &subSect,
                   const SparseIterator &parent,
-                  std::unique_ptr<SparseIterator> &&wrap, Value size,
-                  unsigned stride)
+                  std::unique_ptr<SparseIterator> &&wrap)
       : SparseIterator(IterKind::kSubSect, *wrap,
                        /*extraCursorCnt=*/wrap->randomAccessible() ? 0 : 1),
-        subSect(subSect), wrap(std::move(wrap)), parent(parent), size(size),
-        stride(stride), helper(*this) {
-    assert(stride == 1 && "Not implemented.");
+        subSect(subSect), wrap(std::move(wrap)), parent(parent), helper(*this) {
     assert(subSect.tid == tid && subSect.lvl == lvl);
     assert(parent.kind != IterKind::kSubSect || parent.lvl + 1 == lvl);
   };
@@ -693,7 +690,9 @@ class SubSectIterator : public SparseIterator {
 
   bool randomAccessible() const override { return wrap->randomAccessible(); };
   bool iteratableByFor() const override { return randomAccessible(); };
-  Value upperBound(OpBuilder &b, Location l) const override { return size; }
+  Value upperBound(OpBuilder &b, Location l) const override {
+    return subSect.subSectSz;
+  }
   std::pair<Value, Value> getCurPosition() const override {
     return wrap->getCurPosition();
   };
@@ -711,7 +710,7 @@ class SubSectIterator : public SparseIterator {
         assert(p->lvl + 1 == lvl);
         wrap->genInit(b, l, p);
         // Linearize the dense subsection index.
-        nxLvlTupleStart = MULI(size, p->getNxLvlTupleId(b, l));
+        nxLvlTupleStart = MULI(subSect.subSectSz, p->getNxLvlTupleId(b, l));
       } else {
         assert(subSect.lvl == lvl && subSect.isSubSectRoot());
         wrap->deserialize(subSect.delegate->serialize());
@@ -765,9 +764,6 @@ class SubSectIterator : public SparseIterator {
   std::unique_ptr<SparseIterator> wrap;
   const SparseIterator &parent;
 
-  Value size;
-  unsigned stride;
-
   SubSectIterHelper helper;
 };
 
@@ -1330,29 +1326,19 @@ sparse_tensor::makeSlicedLevelIterator(std::unique_ptr<SparseIterator> &&sit,
   return std::make_unique<FilterIterator>(std::move(sit), offset, stride, size);
 }
 
-template <typename IterType>
 static const SparseIterator *tryUnwrapFilter(const SparseIterator *it) {
   auto *filter = llvm::dyn_cast_or_null<FilterIterator>(it);
-  if (filter && llvm::isa<IterType>(filter->wrap.get())) {
+  if (filter)
     return filter->wrap.get();
-  }
   return it;
 }
-template <typename IterType>
-static const IterType *unwrapFilter(const SparseIterator *it) {
-  auto *filter = llvm::dyn_cast_or_null<FilterIterator>(it);
-  if (filter) {
-    return llvm::cast<IterType>(filter->wrap.get());
-  }
-  return llvm::cast<IterType>(it);
-}
 
 std::unique_ptr<SparseIterator> sparse_tensor::makeNonEmptySubSectIterator(
     OpBuilder &b, Location l, const SparseIterator *parent, Value loopBound,
     std::unique_ptr<SparseIterator> &&delegate, Value size, unsigned stride) {
 
   // Try unwrap the NonEmptySubSectIterator from a filter parent.
-  parent = tryUnwrapFilter<NonEmptySubSectIterator>(parent);
+  parent = tryUnwrapFilter(parent);
   auto it = std::make_unique<NonEmptySubSectIterator>(
       b, l, parent, std::move(delegate), size);
 
@@ -1366,12 +1352,22 @@ std::unique_ptr<SparseIterator> sparse_tensor::makeNonEmptySubSectIterator(
 }
 
 std::unique_ptr<SparseIterator> sparse_tensor::makeTraverseSubSectIterator(
-    const SparseIterator &subSectIter, const SparseIterator &parent,
-    std::unique_ptr<SparseIterator> &&wrap, Value size, unsigned stride) {
+    OpBuilder &b, Location l, const SparseIterator &subSectIter,
+    const SparseIterator &parent, std::unique_ptr<SparseIterator> &&wrap,
+    Value loopBound, unsigned stride) {
+
   // This must be a subsection iterator or a filtered subsection iterator.
-  auto &subSect = *unwrapFilter<NonEmptySubSectIterator>(&subSectIter);
-  return std::make_unique<SubSectIterator>(subSect, parent, std::move(wrap),
-                                           size, stride);
+  auto &subSect =
+      llvm::cast<NonEmptySubSectIterator>(*tryUnwrapFilter(&subSectIter));
+
+  auto it = std::make_unique<SubSectIterator>(
+      subSect, *tryUnwrapFilter(&parent), std::move(wrap));
+
+  if (stride != 1) {
+    return std::make_unique<FilterIterator>(std::move(it), /*offset=*/C_IDX(0),
+                                            C_IDX(stride), /*size=*/loopBound);
+  }
+  return it;
 }
 
 #undef CMPI

mlir/lib/Dialect/SparseTensor/Transforms/Utils/SparseTensorLevel.h

@@ -300,8 +300,9 @@ std::unique_ptr<SparseIterator> makeNonEmptySubSectIterator(
 /// Helper function to create a SparseIterator object that iterate over a
 /// non-empty subsection created by NonEmptySubSectIterator.
 std::unique_ptr<SparseIterator> makeTraverseSubSectIterator(
-    const SparseIterator &subsectIter, const SparseIterator &parent,
-    std::unique_ptr<SparseIterator> &&delegate, Value size, unsigned stride);
+    OpBuilder &b, Location l, const SparseIterator &subsectIter,
+    const SparseIterator &parent, std::unique_ptr<SparseIterator> &&wrap,
+    Value loopBound, unsigned stride);
 
 } // namespace sparse_tensor
 } // namespace mlir

mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_dilated_conv_2d_nhwc_hwcf.mlir

@@ -0,0 +1,148 @@
+//--------------------------------------------------------------------------------------------------
+// WHEN CREATING A NEW TEST, PLEASE JUST COPY & PASTE WITHOUT EDITS.
+//
+// Set-up that's shared across all tests in this directory. In principle, this
+// config could be moved to lit.local.cfg. However, there are downstream users that
+//  do not use these LIT config files. Hence why this is kept inline.
+//
+// DEFINE: %{sparsifier_opts} = enable-runtime-library=true
+// DEFINE: %{sparsifier_opts_sve} = enable-arm-sve=true %{sparsifier_opts}
+// DEFINE: %{compile} = mlir-opt %s --sparsifier="%{sparsifier_opts}"
+// DEFINE: %{compile_sve} = mlir-opt %s --sparsifier="%{sparsifier_opts_sve}"
+// DEFINE: %{run_libs} = -shared-libs=%mlir_c_runner_utils,%mlir_runner_utils
+// DEFINE: %{run_opts} = -e entry -entry-point-result=void
+// DEFINE: %{run} = mlir-cpu-runner %{run_opts} %{run_libs}
+// DEFINE: %{run_sve} = %mcr_aarch64_cmd --march=aarch64 --mattr="+sve" %{run_opts} %{run_libs}
+//
+// DEFINE: %{env} =
+//--------------------------------------------------------------------------------------------------
+
+// RUN: %{compile} | %{run} | FileCheck %s
+//
+// Do the same run, but now with direct IR generation.
+// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false enable-buffer-initialization=true
+// RUN: %{compile} | %{run} | FileCheck %s
+//
+// Do the same run, but now with direct IR generation and vectorization.
+// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false enable-buffer-initialization=true vl=2 reassociate-fp-reductions=true enable-index-optimizations=true
+// RUN: %{compile} | %{run} | FileCheck %s
+//
+// Do the same run, but now with direct IR generation and VLA vectorization.
+// RUN: %if mlir_arm_sve_tests %{ %{compile_sve} | %{run_sve} | FileCheck %s %}
+
+#CCCC = #sparse_tensor.encoding<{
+  map = (d0, d1, d2, d3) -> (d0 : compressed, d1 : compressed, d2 : compressed, d3 : compressed)
+}>
+
+#CDCC = #sparse_tensor.encoding<{
+  map = (d0, d1, d2, d3) -> (d0 : compressed, d1 : dense, d2 : compressed, d3 : compressed)
+}>
+
+// Creates and returns 4-D buffer of size (%s1, %s2, %s3, %s4) filled with the value %f
+func.func @alloc_4d_filled_f32(%s1 : index, %s2 : index, %s3 : index, %s4 : index, %f : f32) -> tensor<?x?x?x?xf32> {
+  %buf = tensor.empty(%s1, %s2, %s3, %s4) : tensor<?x?x?x?xf32>
+  %ret = linalg.fill ins(%f : f32) outs(%buf : tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32>
+  return %ret : tensor<?x?x?x?xf32>
+}
+
+func.func @conv_2d_nhwc_hwcf(%arg0: tensor<?x?x?x?xf32>, %arg1: tensor<?x?x?x?xf32>, %arg2: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32> {
+  %ret = linalg.conv_2d_nhwc_hwcf {dilations = dense<2> : tensor<2xi64>,
+                                     strides = dense<1> : tensor<2xi64>}
+     ins (%arg0, %arg1: tensor<?x?x?x?xf32>, tensor<?x?x?x?xf32>)
+    outs (%arg2: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32>
+  return %ret : tensor<?x?x?x?xf32>
+}
+
+func.func @conv_2d_nhwc_hwcf_CCCC(%arg0: tensor<?x?x?x?xf32, #CCCC>, %arg1: tensor<?x?x?x?xf32>, %arg2: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32> {
+  %ret = linalg.conv_2d_nhwc_hwcf {dilations = dense<2> : tensor<2xi64>,
+                                     strides = dense<1> : tensor<2xi64>}
+     ins (%arg0, %arg1: tensor<?x?x?x?xf32, #CCCC>, tensor<?x?x?x?xf32>)
+    outs (%arg2: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32>
+  return %ret : tensor<?x?x?x?xf32>
+}
+
+func.func @conv_2d_nhwc_hwcf_CDCC(%arg0: tensor<?x?x?x?xf32, #CDCC>, %arg1: tensor<?x?x?x?xf32>, %arg2: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32> {
+  %ret = linalg.conv_2d_nhwc_hwcf {dilations = dense<2> : tensor<2xi64>,
+                                     strides = dense<1> : tensor<2xi64>}
+     ins (%arg0, %arg1: tensor<?x?x?x?xf32, #CDCC>, tensor<?x?x?x?xf32>)
+    outs (%arg2: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32>
+  return %ret : tensor<?x?x?x?xf32>
+}
+
+func.func @conv_2d_nhwc_hwcf_dual_CDCC(%arg0: tensor<?x?x?x?xf32, #CDCC>, %arg1: tensor<?x?x?x?xf32, #CDCC>, %arg2: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32> {
+  %ret = linalg.conv_2d_nhwc_hwcf {dilations = dense<2> : tensor<2xi64>,
+                                     strides = dense<1> : tensor<2xi64>}
+     ins (%arg0, %arg1: tensor<?x?x?x?xf32, #CDCC>, tensor<?x?x?x?xf32, #CDCC>)
+    outs (%arg2: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32>
+  return %ret : tensor<?x?x?x?xf32>
+}
+
+
+func.func @entry() {
+  %c0 = arith.constant 0 : index
+  %c1 = arith.constant 1 : index
+  %c3 = arith.constant 3 : index
+  %c5 = arith.constant 5 : index
+  %c6 = arith.constant 6 : index
+  %c7 = arith.constant 7 : index
+  %f10 = arith.constant 10.00000e+00 : f32
+  %val = arith.constant 2.00000e+00 : f32
+  %zero = arith.constant 0.00000e+00 : f32
+
+  %filter2D_nhwc = call @alloc_4d_filled_f32(%c3, %c3, %c3, %c1, %val) :(index, index, index, index, f32) -> (tensor<?x?x?x?xf32>)
+  %in2D_tmp = call @alloc_4d_filled_f32(%c3, %c7, %c7, %c3, %f10) : (index, index, index, index, f32) -> (tensor<?x?x?x?xf32>)
+  %in2D_nhwc = tensor.insert %zero into %in2D_tmp[%c0, %c1, %c1, %c0] : tensor<?x?x?x?xf32>
+  %out2D_nhwc = call @alloc_4d_filled_f32(%c3, %c3, %c3, %c1, %zero) : (index, index, index, index, f32) -> (tensor<?x?x?x?xf32>)
+
+  %in2D_nhwc_CCCC = sparse_tensor.convert %in2D_nhwc
+    : tensor<?x?x?x?xf32> to tensor<?x?x?x?xf32, #CCCC>
+  %filter2D_nhwc_CDCC = sparse_tensor.convert %filter2D_nhwc
+    : tensor<?x?x?x?xf32> to tensor<?x?x?x?xf32, #CDCC>
+  %in2D_nhwc_CDCC = sparse_tensor.convert %in2D_nhwc
+    : tensor<?x?x?x?xf32> to tensor<?x?x?x?xf32, #CDCC>
+
+  %dense_ret = call @conv_2d_nhwc_hwcf(%in2D_nhwc, %filter2D_nhwc, %out2D_nhwc) : (tensor<?x?x?x?xf32>, tensor<?x?x?x?xf32>, tensor<?x?x?x?xf32>) -> (tensor<?x?x?x?xf32>)
+  %CCCC_ret = call @conv_2d_nhwc_hwcf_CCCC(%in2D_nhwc_CCCC, %filter2D_nhwc, %out2D_nhwc) : (tensor<?x?x?x?xf32, #CCCC>, tensor<?x?x?x?xf32>, tensor<?x?x?x?xf32>) -> (tensor<?x?x?x?xf32>)
+  %CDCC_ret = call @conv_2d_nhwc_hwcf_CDCC(%in2D_nhwc_CDCC, %filter2D_nhwc, %out2D_nhwc) : (tensor<?x?x?x?xf32, #CDCC>, tensor<?x?x?x?xf32>, tensor<?x?x?x?xf32>) -> (tensor<?x?x?x?xf32>)
+
+  %dual_CDCC_ret = call @conv_2d_nhwc_hwcf_dual_CDCC(%in2D_nhwc_CDCC, %filter2D_nhwc_CDCC, %out2D_nhwc)
+    : (tensor<?x?x?x?xf32, #CDCC>, tensor<?x?x?x?xf32, #CDCC>, tensor<?x?x?x?xf32>) -> (tensor<?x?x?x?xf32>)
+
+  // CHECK:      ( ( ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 520 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ) ),
+  // CHECK-SAME:   ( ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ) ),
+  // CHECK-SAME:   ( ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ) ) )
+  %dense_v = vector.transfer_read %dense_ret[%c0, %c0, %c0, %c0], %zero
+      : tensor<?x?x?x?xf32>, vector<3x3x3x1xf32>
+  vector.print %dense_v : vector<3x3x3x1xf32>
+
+  // CHECK-NEXT: ( ( ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 520 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ) ),
+  // CHECK-SAME:   ( ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ) ),
+  // CHECK-SAME:   ( ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ) ) )
+  %v_dual = vector.transfer_read %dual_CDCC_ret[%c0, %c0, %c0, %c0], %zero
+      : tensor<?x?x?x?xf32>, vector<3x3x3x1xf32>
+  vector.print %v_dual : vector<3x3x3x1xf32>
+
+  // CHECK-NEXT: ( ( ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 520 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ) ),
+  // CHECK-SAME:   ( ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ) ),
+  // CHECK-SAME:   ( ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ) ) )
+  %v1 = vector.transfer_read %CCCC_ret[%c0, %c0, %c0, %c0], %zero
+      : tensor<?x?x?x?xf32>, vector<3x3x3x1xf32>
+  vector.print %v1 : vector<3x3x3x1xf32>
+
+  // CHECK-NEXT: ( ( ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 520 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ) ),
+  // CHECK-SAME:   ( ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ) ),
+  // CHECK-SAME:   ( ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ), ( ( 540 ), ( 540 ), ( 540 ) ) ) )
+  %v2 = vector.transfer_read %CDCC_ret[%c0, %c0, %c0, %c0], %zero
+      : tensor<?x?x?x?xf32>, vector<3x3x3x1xf32>
+  vector.print %v1 : vector<3x3x3x1xf32>
+
+  // Free the resources.
+  bufferization.dealloc_tensor %in2D_nhwc : tensor<?x?x?x?xf32>
+  bufferization.dealloc_tensor %filter2D_nhwc : tensor<?x?x?x?xf32>
+  bufferization.dealloc_tensor %out2D_nhwc : tensor<?x?x?x?xf32>
+
+  bufferization.dealloc_tensor %filter2D_nhwc_CDCC : tensor<?x?x?x?xf32, #CDCC>
+  bufferization.dealloc_tensor %in2D_nhwc_CCCC : tensor<?x?x?x?xf32, #CCCC>
+  bufferization.dealloc_tensor %in2D_nhwc_CDCC : tensor<?x?x?x?xf32, #CDCC>
+  return
+}