[torch-dialect] emit aten.as_strided op

e2e_testing/xfail_sets.py

@@ -312,6 +312,7 @@
 }
 
 STABLEHLO_PASS_SET = {
+    "AsStridedStaticModule_basic",
     "AliasModule_basic",
     "AllBoolFalseModule_basic",
     "AllBoolTrueModule_basic",
@@ -810,6 +811,7 @@
 # Write the TOSA set as a "passing" set as it is very early in development
 # and very few tests work yet.
 TOSA_PASS_SET = {
+    "AsStridedStaticModule_basic",
     "AliasModule_basic",
     "MaxPool2dEmptyStrideStaticModule_basic",
     "ConstantBoolParameterModule_basic",

include/torch-mlir/Dialect/Torch/IR/GeneratedTorchOps.td

@@ -8028,6 +8028,32 @@ def Torch_AtenViewOp : Torch_Op<"aten.view", [
   let hasFolder = 1;
 }
 
+def Torch_AtenAsStridedOp : Torch_Op<"aten.as_strided", [
+    AllowsTypeRefinement,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::as_strided : (Tensor, int[], int[], int?) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self,
+    AnyTorchListOfTorchIntType:$size,
+    AnyTorchListOfTorchIntType:$stride,
+    AnyTorchOptionalIntType:$storage_offset
+  );
+  let results = (outs
+    AnyTorchTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenAsStridedOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 4, 1);
+    }
+    void AtenAsStridedOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 4, 1);
+    }
+  }];
+  let hasFolder = 1;
+}
+
 def Torch_Aten_UnsafeViewOp : Torch_Op<"aten._unsafe_view", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Dialect/Torch/IR/TorchOps.cpp

@@ -895,6 +895,52 @@ OpFoldResult AtenViewOp::fold(FoldAdaptor adaptor) {
   return getOperand(0);
 }
 
+//===----------------------------------------------------------------------===//
+// AtenAsStrideOp
+//===----------------------------------------------------------------------===//
+
+OpFoldResult AtenAsStridedOp::fold(FoldAdaptor adaptor) {
+  auto inputType = getOperand(0).getType().dyn_cast<BaseTensorType>();
+  if (!inputType || !inputType.hasSizes() || !inputType.areAllSizesKnown())
+    return nullptr;
+
+  auto outType = getType().dyn_cast<BaseTensorType>();
+  if (!outType || !outType.hasSizes() || !outType.areAllSizesKnown())
+    return nullptr;
+
+  int64_t storageOffset;
+  if (!getStorageOffset().getType().isa<Torch::NoneType>()) {
+    if (!matchPattern(getStorageOffset(), m_TorchConstantInt(&storageOffset)) ||
+        storageOffset != 0)
+      return nullptr;
+  }
+
+  // Check if the shapes of input tensor and output tensor are totally same.
+  ArrayRef<int64_t> inputSizes = inputType.getSizes();
+  ArrayRef<int64_t> outSizes = outType.getSizes();
+  if (inputSizes.size() != outSizes.size())
+    return nullptr;
+  for (int i = 0, e = inputSizes.size(); i < e; ++i) {
+    if (inputSizes[i] != outSizes[i])
+      return nullptr;
+  }
+
+  // Check if the elements of output tensor are fetched sequentially from input
+  // tensor's storage.
+  SmallVector<int64_t> strides;
+  if (!matchPattern(getStride(), m_TorchListOfConstantInts(strides)))
+    return nullptr;
+
+  if (strides.size() != inputSizes.size() || strides[strides.size() - 1] != 1)
+    return nullptr;
+  for (int i = inputSizes.size() - 2; i >= 0; --i) {
+    if (strides[i] != inputSizes[i + 1] * strides[i + 1])
+      return nullptr;
+  }
+
+  return getOperand(0);
+}
+
 //===----------------------------------------------------------------------===//
 // PrimsViewOfOp
 //===----------------------------------------------------------------------===//

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -6824,6 +6824,9 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "  func.func @\"__torch_mlir_shape_fn.aten.resize_\"(%arg0: !torch.list<int>, %arg1: !torch.list<int>, %arg2: !torch.optional<int>) -> !torch.list<int> {\n"
 "    return %arg1 : !torch.list<int>\n"
 "  }\n"
+"  func.func @\"__torch_mlir_shape_fn.aten.as_strided\"(%arg0: !torch.list<int>, %arg1: !torch.list<int>, %arg2: !torch.list<int>, %arg3: !torch.optional<int>) -> !torch.list<int> {\n"
+"    return %arg1 : !torch.list<int>\n"
+"  }\n"
 "  func.func @\"__torch_mlir_shape_fn.aten.max_pool2d\"(%arg0: !torch.list<int>, %arg1: !torch.list<int>, %arg2: !torch.list<int>, %arg3: !torch.list<int>, %arg4: !torch.list<int>, %arg5: !torch.bool) -> !torch.list<int> {\n"
 "    %0 = call @__torch__.torch.jit._shape_functions.max_pool2d(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5) : (!torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.bool) -> !torch.list<int>\n"
 "    return %0 : !torch.list<int>\n"
@@ -8520,6 +8523,10 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
 "    return %0#1 : !torch.int\n"
 "  }\n"
+"  func.func @\"__torch_mlir_dtype_fn.aten.as_strided\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.list<int>, %arg2: !torch.list<int>, %arg3: !torch.optional<int>) -> !torch.int {\n"
+"    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
+"    return %0#1 : !torch.int\n"
+"  }\n"
 "  func.func @\"__torch_mlir_dtype_fn.aten.roll\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.list<int>, %arg2: !torch.list<int>) -> !torch.int {\n"
 "    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
 "    return %0#1 : !torch.int\n"

lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp

@@ -4585,6 +4585,123 @@ class DecomposeAtenSignOp : public OpRewritePattern<AtenSignOp> {
 };
 } // namespace
 
+namespace {
+// Decompose `aten.as_strided` into `aten.flatten.using_ints`, `aten.gather` and
+// `aten.view`.
+// This decomposition is a little hacky. Since aten.as_strided is a
+// view-like op, while aten.gather is not. But it might be okay
+// in torch-mlir, since we've already assumed view-like ops to be of value
+// semantics.
+class DecomposeAtenAsStridedOp : public OpRewritePattern<AtenAsStridedOp> {
+public:
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenAsStridedOp op,
+                                PatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    auto inputType = op.getSelf().getType().dyn_cast<BaseTensorType>();
+    if (!inputType || !inputType.hasSizes() || !inputType.areAllSizesKnown()) {
+      return rewriter.notifyMatchFailure(
+          op, "only handle input tensor with static shape information");
+    }
+    ArrayRef<int64_t> inputSizes = inputType.getSizes();
+
+    SmallVector<int64_t> outSizes;
+    if (!matchPattern(op.getSize(), m_TorchListOfConstantInts(outSizes))) {
+      return rewriter.notifyMatchFailure(
+          op, "out size must be a list of constant integers");
+    }
+    SmallVector<int64_t> strides;
+    if (!matchPattern(op.getStride(), m_TorchListOfConstantInts(strides))) {
+      return rewriter.notifyMatchFailure(
+          op, "strides must be a list of constant integers");
+    }
+    if (strides.size() != outSizes.size()) {
+      return rewriter.notifyMatchFailure(op,
+                                         "the stride size is expected to be "
+                                         "the same as that of output tensor");
+    }
+    int64_t storageOffset = 0;
+    if (!op.getStorageOffset().getType().isa<Torch::NoneType>() &&
+        !matchPattern(op.getStorageOffset(),
+                      m_TorchConstantInt(&storageOffset))) {
+      return rewriter.notifyMatchFailure(
+          op, "storage offset must be a constant integer");
+    }
+
+    int64_t inputTotalSize = 1;
+    for (auto inputDimSize : inputSizes) {
+      inputTotalSize *= inputDimSize;
+    }
+
+    auto flattenInputType = inputType.getWithSizesAndDtype(
+        SmallVector<int64_t>(1, inputTotalSize), inputType.getOptionalDtype());
+    Value startDim = rewriter.create<Torch::ConstantIntOp>(
+        loc, rewriter.getI64IntegerAttr(0));
+    Value endDim = rewriter.create<Torch::ConstantIntOp>(
+        loc, rewriter.getI64IntegerAttr(inputSizes.size() - 1));
+    Value flattenInput = rewriter.create<Torch::AtenFlattenUsingIntsOp>(
+        loc, flattenInputType, op.getSelf(), startDim, endDim);
+
+    int64_t outTotalSize = 1;
+    for (auto outDimSize : outSizes) {
+      outTotalSize *= outDimSize;
+    }
+
+    // Generate index for gather op;
+    DenseSet<int64_t> visitedStoragePlaces;
+    SmallVector<int64_t> gatherIndicies;
+    for (int64_t i = 0; i < outTotalSize; ++i) {
+      int64_t newI = i;
+      int64_t index = 0;
+      for (int64_t d = outSizes.size() - 1; d >= 0; --d) {
+        index += newI % outSizes[d] * strides[d];
+        newI /= outSizes[d];
+      }
+
+      // We can not handle the situation when the view is "overlapped"
+      // Ref: https://pytorch.org/docs/stable/generated/torch.as_strided.html
+      if (visitedStoragePlaces.find(index) != visitedStoragePlaces.end()) {
+        return rewriter.notifyMatchFailure(
+            op, "multiple indices of new tensor are mapped to the same storage "
+                "location");
+      }
+
+      visitedStoragePlaces.insert(index);
+      gatherIndicies.push_back(index + storageOffset);
+    }
+    auto gatherOpType = inputType.getWithSizesAndDtype(
+        {outTotalSize}, inputType.getOptionalDtype());
+    Value gatherDim = rewriter.create<Torch::ConstantIntOp>(
+        loc, rewriter.getI64IntegerAttr(0));
+
+    Value gatherIndex = rewriter.create<Torch::ValueTensorLiteralOp>(
+        loc,
+        inputType.getWithSizesAndDtype({outTotalSize},
+                                       rewriter.getIntegerType(64, true)),
+        DenseIntElementsAttr::get(
+            RankedTensorType::get({static_cast<int64_t>(outTotalSize)},
+                                  rewriter.getIntegerType(64, true)),
+            gatherIndicies));
+    Value sparseGrad = rewriter.create<Torch::ConstantBoolOp>(loc, false);
+    auto gatherOp = rewriter.create<Torch::AtenGatherOp>(
+        loc, gatherOpType, flattenInput, gatherDim, gatherIndex, sparseGrad);
+
+    SmallVector<Value> viewSizes;
+    for (auto outDimSize : outSizes) {
+      viewSizes.push_back(rewriter.create<Torch::ConstantIntOp>(
+          loc, rewriter.getI64IntegerAttr(outDimSize)));
+    }
+    auto viewSizesListConstructOp = rewriter.create<Torch::PrimListConstructOp>(
+        loc, Torch::ListType::get(rewriter.getType<Torch::IntType>()),
+        ValueRange(viewSizes));
+
+    rewriter.replaceOpWithNewOp<AtenViewOp>(op, op.getType(), gatherOp,
+                                            viewSizesListConstructOp);
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 class DecomposeComplexOpsPass
     : public DecomposeComplexOpsBase<DecomposeComplexOpsPass> {
@@ -4754,6 +4871,7 @@ class DecomposeComplexOpsPass
     addPatternIfTargetOpIsIllegal<DecomposeAtenScalarTensor>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenScatterValueOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenSignOp>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenAsStridedOp>(patterns);
 
     GreedyRewriteConfig config;
     config.useTopDownTraversal = true;

lib/Dialect/Torch/Transforms/LowerToBackendContract.cpp

@@ -483,6 +483,7 @@ static void markDecomposedOpsAsIllegal(MLIRContext *context,
   target.addIllegalOp<AtenTopkOp>();
   target.addIllegalOp<AtenScalarTensorOp>();
   target.addIllegalOp<AtenScatterValueOp>();
+  target.addIllegalOp<AtenAsStridedOp>();
   for (auto &opName : backendLegalOpsSet) {
     target.addLegalOp(
         OperationName(kTorchOpPrefix + opName.first().str(), context));

lib/Dialect/Torch/Utils/Utils.cpp

@@ -205,7 +205,8 @@ bool Torch::isViewLikeOp(Operation *op) {
              AtenTransposeIntOp, AtenUnsqueezeOp, AtenViewOp,
              TensorStaticInfoCastOp, AtenToDtypeLayoutOp, AtenNumpyTOp,
              AtenNarrowOp, AtenToDeviceOp, PrimsSqueezeOp, AtenMovedimIntOp,
-             PrimsViewOfOp, AtenRealOp, AtenImagOp, AtenViewAsComplexOp>(op);
+             PrimsViewOfOp, AtenRealOp, AtenImagOp, AtenViewAsComplexOp,
+             AtenAsStridedOp>(op);
 }
 
 Value Torch::getConstantWithGivenDtypeAndValue(PatternRewriter &rewriter,

python/torch_mlir/dialects/torch/importer/jit_ir/build_tools/abstract_interp_lib_gen.py

@@ -469,6 +469,9 @@ def aten〇_unsafe_view〡shape(self: List[int], size: List[int]) -> List[int]:
 def aten〇resize_〡shape(self: List[int], size: List[int], memory_format: Optional[int] = None) -> List[int]:
     return size
 
+def aten〇as_strided〡shape(self: List[int], size: List[int], stride: List[int], storage_offset: Optional[int] = None) -> List[int]:
+    return size
+
 def aten〇max_pool2d〡shape(self: List[int], kernel_size: List[int], stride: List[int] = (), padding: List[int] = (0, 0,), dilation: List[int] = (1, 1,), ceil_mode: bool = False) -> List[int]:
     return upstream_shape_functions.max_pool2d(self, kernel_size, stride, padding, dilation, ceil_mode)
 
@@ -1732,6 +1735,11 @@ def aten〇resize_〡dtype(self_rank_dtype: Tuple[int, int], size: List[int], me
     self_rank, self_dtype = self_rank_dtype
     return self_dtype
 
+@check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1, size=[1], stride=[1]))
+def aten〇as_strided〡dtype(self_rank_dtype: Tuple[int, int], size: List[int], stride: List[int], storage_offset: Optional[int] = None) -> int:
+    self_rank, self_dtype = self_rank_dtype
+    return self_dtype
+
 @check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1, shifts=[0], dims=[0]))
 def aten〇roll〡dtype(self_rank_dtype: Tuple[int, int], shifts: List[int], dims: List[int] = ()) -> int:
     self_rank, self_dtype = self_rank_dtype

python/torch_mlir/dialects/torch/importer/jit_ir/build_tools/torch_ods_gen.py

@@ -523,6 +523,7 @@ def emit_with_mutating_variants(key, **kwargs):
     emit("aten::to.device : (Tensor, Device, int, bool, bool, int?) -> (Tensor)")
     emit("aten::type_as : (Tensor, Tensor) -> (Tensor)", has_folder=True)
     emit("aten::view : (Tensor, int[]) -> (Tensor)", has_folder=True)
+    emit("aten::as_strided : (Tensor, int[], int[], int?) -> (Tensor)", has_folder=True)
     emit("aten::_unsafe_view : (Tensor, int[]) -> (Tensor)")
     emit("aten::where.self : (Tensor, Tensor, Tensor) -> (Tensor)")
     emit("aten::where.Scalar : (Tensor, Scalar, Scalar) -> (Tensor)")

python/torch_mlir_e2e_test/test_suite/reshape_like.py

@@ -710,4 +710,23 @@ def forward(self, a):
 
 @register_test_case(module_factory=lambda: ReshapeAliasCollapseModule())
 def ReshapeAliasCollapseModule_basic(module, tu: TestUtils):
-    module.forward(tu.rand(2, 4))
+    module.forward(tu.rand(2, 4))
+
+# ==============================================================================
+
+class AsStridedStaticModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+        ([3, 3], torch.float32, True),
+    ])
+
+    def forward(self, x):
+        return torch.ops.aten.as_strided(x, (2, 2), (1, 2))
+
+@register_test_case(module_factory=lambda: AsStridedStaticModule())
+def AsStridedStaticModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(3, 3))

test/Dialect/Torch/canonicalize.mlir

@@ -2065,3 +2065,29 @@ func.func @torch.aten.add$fold() -> !torch.float {
     %0 = torch.aten.add %float1, %float2 : !torch.float, !torch.float -> !torch.float
     return %0 : !torch.float
 }
+
+// CHECK-LABEL:  func.func @torch.aten.as_strided$none_storage_offset(
+// CHECK-SAME:                                    %[[ARG:.*]]: !torch.vtensor<[3,3],f32>) -> !torch.vtensor<[3,3],f32> {
+// CHECK:          return %[[ARG]] : !torch.vtensor<[3,3],f32>
+func.func @torch.aten.as_strided$none_storage_offset(%arg0: !torch.vtensor<[3,3],f32>) -> !torch.vtensor<[3,3],f32> {
+  %int3 = torch.constant.int 3
+  %int1 = torch.constant.int 1
+  %none = torch.constant.none
+  %0 = torch.prim.ListConstruct %int3, %int3 : (!torch.int, !torch.int) -> !torch.list<int>
+  %1 = torch.prim.ListConstruct %int3, %int1 : (!torch.int, !torch.int) -> !torch.list<int>
+  %2 = torch.aten.as_strided %arg0, %0, %1, %none : !torch.vtensor<[3,3],f32>, !torch.list<int>, !torch.list<int>, !torch.none -> !torch.vtensor<[3,3],f32>
+  return %2 : !torch.vtensor<[3,3],f32>
+}
+
+// CHECK-LABEL:  func.func @torch.aten.as_strided$zero_storage_offset(
+// CHECK-SAME:                                    %[[ARG:.*]]: !torch.vtensor<[3,3],f32>) -> !torch.vtensor<[3,3],f32> {
+// CHECK:          return %[[ARG]] : !torch.vtensor<[3,3],f32>
+func.func @torch.aten.as_strided$zero_storage_offset(%arg0: !torch.vtensor<[3,3],f32>) -> !torch.vtensor<[3,3],f32> {
+  %int3 = torch.constant.int 3
+  %int1 = torch.constant.int 1
+  %int0 = torch.constant.int 0
+  %0 = torch.prim.ListConstruct %int3, %int3 : (!torch.int, !torch.int) -> !torch.list<int>
+  %1 = torch.prim.ListConstruct %int3, %int1 : (!torch.int, !torch.int) -> !torch.list<int>
+  %2 = torch.aten.as_strided %arg0, %0, %1, %int0 : !torch.vtensor<[3,3],f32>, !torch.list<int>, !torch.list<int>, !torch.int -> !torch.vtensor<[3,3],f32>
+  return %2 : !torch.vtensor<[3,3],f32>
+}