Support prims.sum (#4052)

This is supported by decomposing into `aten.sum`.

Author: mgehre-amd

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

@@ -18369,6 +18369,31 @@ def Torch_PrimsSqueezeOp : Torch_Op<"prims.squeeze", [
   }];
 }
 
+def Torch_PrimsSumOp : Torch_Op<"prims.sum", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `prims::sum : (Tensor, int[]?, int?) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$inp,
+    AnyTorchOptionalListOfTorchIntType:$dims,
+    AnyTorchOptionalIntType:$output_dtype
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult PrimsSumOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 3, 1);
+    }
+    void PrimsSumOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 3, 1);
+    }
+  }];
+}
+
 def Torch_PrimsViewOfOp : Torch_Op<"prims.view_of", [
     AllowsTypeRefinement,
     ReadOnly

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -7542,6 +7542,12 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %1 = call @__torch__.torch.jit._shape_functions.sum_mean_dim(%arg0, %arg1, %arg2, %0) : (!torch.list<int>, !torch.optional<list<int>>, !torch.bool, !torch.any) -> !torch.list<int>\n"
 "    return %1 : !torch.list<int>\n"
 "  }\n"
+"  func.func @\"__torch_mlir_shape_fn.prims.sum\"(%arg0: !torch.list<int>, %arg1: !torch.optional<list<int>>, %arg2: !torch.optional<int>) -> !torch.list<int> {\n"
+"    %false = torch.constant.bool false\n"
+"    %0 = torch.derefine %arg2 : !torch.optional<int> to !torch.any\n"
+"    %1 = call @__torch__.torch.jit._shape_functions.sum_mean_dim(%arg0, %arg1, %false, %0) : (!torch.list<int>, !torch.optional<list<int>>, !torch.bool, !torch.any) -> !torch.list<int>\n"
+"    return %1 : !torch.list<int>\n"
+"  }\n"
 "  func.func @\"__torch_mlir_shape_fn.aten.prod.dim_int\"(%arg0: !torch.list<int>, %arg1: !torch.int, %arg2: !torch.bool, %arg3: !torch.optional<int>) -> !torch.list<int> {\n"
 "    %0 = torch.prim.ListConstruct %arg1 : (!torch.int) -> !torch.list<int>\n"
 "    %1 = torch.derefine %0 : !torch.list<int> to !torch.optional<list<int>>\n"
@@ -11889,6 +11895,19 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    }\n"
 "    return %2 : !torch.int\n"
 "  }\n"
+"  func.func @\"__torch_mlir_dtype_fn.prims.sum\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.optional<list<int>>, %arg2: !torch.optional<int>) -> !torch.int {\n"
+"    %str = torch.constant.str \"AssertionError: \"\n"
+"    %none = torch.constant.none\n"
+"    %0 = torch.aten.__is__ %arg2, %none : !torch.optional<int>, !torch.none -> !torch.bool\n"
+"    torch.prim.If %0 -> () {\n"
+"      torch.prim.If.yield\n"
+"    } else {\n"
+"      torch.prim.RaiseException %str, %none : !torch.str, !torch.none\n"
+"      torch.prim.If.yield\n"
+"    }\n"
+"    %1:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
+"    return %1#1 : !torch.int\n"
+"  }\n"
 "  func.func @\"__torch_mlir_dtype_fn.aten.abs\"(%arg0: !torch.tuple<int, int>) -> !torch.int {\n"
 "    %int6 = torch.constant.int 6\n"
 "    %int9 = torch.constant.int 9\n"

lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp

@@ -10495,6 +10495,23 @@ class DecomposeAtenScatterValueOp
 };
 } // namespace
 
+namespace {
+// Decompose prims.sum into aten.sum
+class DecomposePrimsSumOp : public OpRewritePattern<PrimsSumOp> {
+public:
+  using OpRewritePattern<PrimsSumOp>::OpRewritePattern;
+  LogicalResult matchAndRewrite(PrimsSumOp op,
+                                PatternRewriter &rewriter) const override {
+    Value cstFalse = rewriter.create<Torch::ConstantBoolOp>(op.getLoc(), false);
+
+    rewriter.replaceOpWithNewOp<AtenSumDimIntListOp>(
+        op, op.getType(), op.getInp(), op.getDims(), /*keepdim=*/cstFalse,
+        op.getOutputDtype());
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 // Decompose `aten.sgn` op into comparisons and aten.where.
 class DecomposeAtenSgnOp : public OpRewritePattern<AtenSgnOp> {
@@ -11812,6 +11829,7 @@ class DecomposeComplexOpsPass
     addPatternIfTargetOpIsIllegal<DecomposeAtenScalarTensor>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenScatterValueOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenSgnOp>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposePrimsSumOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenTypeAsOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenTileOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenReshapeAsOp>(patterns);

projects/pt1/e2e_testing/xfail_sets.py

@@ -3017,6 +3017,7 @@
     "PrimsConvertElementTypeModule_basic",
     "PrimsSqueezeEmptyDimensionsModule_basic",
     "PrimsSqueezeModule_basic",
+    "PrimsSumFloatModule_basic",
     "PrimsViewOfModule_basic",
     "PrimsViewOfZeroRankModule_basic",
     "QuantizedReluInt8_basic",

projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/abstract_interp_lib_gen.py

@@ -809,6 +809,9 @@ def aten〇mean〇dim〡shape(self: List[int], dim: Optional[List[int]], keepdim
 def aten〇sum〇dim_IntList〡shape(self: List[int], dim: Optional[List[int]], keepdim: bool = False, dtype: Optional[int] = None) -> List[int]:
     return upstream_shape_functions.sum_mean_dim(self, dim, keepdim, dtype)
 
+def prims〇sum〡shape(inp: List[int], dims: Optional[List[int]], output_dtype: Optional[int] = None) -> List[int]:
+    return upstream_shape_functions.sum_mean_dim(inp, dims, False, output_dtype)
+
 def aten〇prod〇dim_int〡shape(self: List[int], dim: int, keepdim: bool = False, dtype: Optional[int] = None) -> List[int]:
     return upstream_shape_functions.sum_mean_dim(self, [dim], keepdim, dtype)
 
@@ -2892,6 +2895,15 @@ def prims〇sqrt〡dtype(self_rank_dtype: Tuple[int, int]) -> int:
         return self_dtype
     return _get_dtype_of_floating_point_op(self_dtype)
 
+@check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1, dims=[0]))
+def prims〇sum〡dtype(inp_rank_dtype: Tuple[int, int], dims: Optional[List[int]], output_dtype: Optional[int] = None) -> int:
+    # When invoking prims.sum() with the output_dtype argument, pytorch
+    # complains that the argument is not known.
+    # See https://github.com/pytorch/pytorch/issues/102610
+    assert output_dtype is None
+    inp_rank, inp_dtype = inp_rank_dtype
+    return inp_dtype
+
 @check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1))
 def aten〇abs〡dtype(self_rank_dtype: Tuple[int, int]) -> int:
     self_rank, self_dtype = self_rank_dtype

projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/torch_ods_gen.py

@@ -1277,6 +1277,7 @@ def emit_with_mutating_variants(key, **kwargs):
     emit("prims::collapse : (Tensor, int, int) -> (Tensor)")
     emit("prims::split_dim : (Tensor, int, int) -> (Tensor)")
     emit("prims::squeeze : (Tensor, int[]) -> (Tensor)")
+    emit("prims::sum : (Tensor, int[]?, int?) -> (Tensor)")
     emit("prims::view_of : (Tensor) -> (Tensor)", has_folder=True)
     emit("prims::iota : (int, int, int, int, Device, bool) -> (Tensor)")
 

projects/pt1/python/torch_mlir_e2e_test/test_suite/reduction.py

@@ -81,6 +81,29 @@ def ReduceSumElementTypeBoolModule_basic(module, tu: TestUtils):
 # ==============================================================================
 
 
+class PrimsSumFloatModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([-1, -1, -1], torch.float32, True),
+        ]
+    )
+    def forward(self, a):
+        return torch.ops.prims.sum(a, (0, 1))
+
+
+@register_test_case(module_factory=lambda: PrimsSumFloatModule())
+def PrimsSumFloatModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(3, 4, 5))
+
+
+# ==============================================================================
+
+
 class ReduceProdFloatModule(torch.nn.Module):
     def __init__(self):
         super().__init__()