[Torch] support binary_cross_entropy_with_logits decomposition (#3741)

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

@@ -9224,6 +9224,33 @@ def Torch_AtenBinaryCrossEntropyBackwardOp : Torch_Op<"aten.binary_cross_entropy
   }];
 }
 
+def Torch_AtenBinaryCrossEntropyWithLogitsOp : Torch_Op<"aten.binary_cross_entropy_with_logits", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::binary_cross_entropy_with_logits : (Tensor, Tensor, Tensor?, Tensor?, int) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self,
+    AnyTorchTensorType:$target,
+    AnyTorchOptionalTensorType:$weight,
+    AnyTorchOptionalTensorType:$pos_weight,
+    Torch_IntType:$reduction
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenBinaryCrossEntropyWithLogitsOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 5, 1);
+    }
+    void AtenBinaryCrossEntropyWithLogitsOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 5, 1);
+    }
+  }];
+}
+
 def Torch_AtenLogSigmoidForwardOp : Torch_Op<"aten.log_sigmoid_forward", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -10289,6 +10289,18 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %0 = call @__torch__.torch.jit._shape_functions.cross_entropy_loss(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5) : (!torch.list<int>, !torch.list<int>, !torch.optional<list<int>>, !torch.int, !torch.int, !torch.float) -> !torch.list<int>\n"
 "    return %0 : !torch.list<int>\n"
 "  }\n"
+"  func.func @\"__torch_mlir_shape_fn.aten.binary_cross_entropy_with_logits\"(%arg0: !torch.list<int>, %arg1: !torch.list<int>, %arg2: !torch.optional<list<int>>, %arg3: !torch.optional<list<int>>, %arg4: !torch.int) -> !torch.list<int> {\n"
+"    %int0 = torch.constant.int 0\n"
+"    %0 = torch.prim.ListConstruct  : () -> !torch.list<int>\n"
+"    %1 = torch.aten.eq.int %arg4, %int0 : !torch.int, !torch.int -> !torch.bool\n"
+"    %2 = torch.prim.If %1 -> (!torch.list<int>) {\n"
+"      %3 = func.call @__torch__.torch.jit._shape_functions._copy(%arg0) : (!torch.list<int>) -> !torch.list<int>\n"
+"      torch.prim.If.yield %3 : !torch.list<int>\n"
+"    } else {\n"
+"      torch.prim.If.yield %0 : !torch.list<int>\n"
+"    }\n"
+"    return %2 : !torch.list<int>\n"
+"  }\n"
 "  func.func @\"__torch_mlir_shape_fn.aten.native_layer_norm\"(%arg0: !torch.list<int>, %arg1: !torch.list<int>, %arg2: !torch.optional<list<int>>, %arg3: !torch.optional<list<int>>, %arg4: !torch.float) -> !torch.tuple<list<int>, list<int>, list<int>> {\n"
 "    %0 = call @__torch__.torch.jit._shape_functions.native_layer_norm(%arg0, %arg1) : (!torch.list<int>, !torch.list<int>) -> !torch.tuple<list<int>, list<int>, list<int>>\n"
 "    return %0 : !torch.tuple<list<int>, list<int>, list<int>>\n"
@@ -14634,6 +14646,10 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    }\n"
 "    return %4 : !torch.int\n"
 "  }\n"
+"  func.func @\"__torch_mlir_dtype_fn.aten.binary_cross_entropy_with_logits\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.tuple<int, int>, %arg2: !torch.optional<tuple<int, int>>, %arg3: !torch.optional<tuple<int, int>>, %arg4: !torch.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.renorm\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.number, %arg2: !torch.int, %arg3: !torch.number) -> !torch.int {\n"
 "    %none = torch.constant.none\n"
 "    %str = torch.constant.str \"AssertionError: \"\n"

lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp

@@ -8799,6 +8799,77 @@ public:
 };
 } // namespace
 
+namespace {
+class DecomposeAtenBinaryCrossEntropyWithLogitsOp
+    : public OpRewritePattern<AtenBinaryCrossEntropyWithLogitsOp> {
+  using OpRewritePattern<AtenBinaryCrossEntropyWithLogitsOp>::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenBinaryCrossEntropyWithLogitsOp op,
+                                PatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    auto self = op.getSelf();
+    auto target = op.getTarget();
+    auto posWeight = op.getPosWeight();
+    auto weight = op.getWeight();
+    auto reduction = op.getReduction();
+
+    Value loss;
+    auto one =
+        rewriter.create<ConstantIntOp>(loc, rewriter.getI64IntegerAttr(1));
+    auto _one =
+        rewriter.create<ConstantIntOp>(loc, rewriter.getI64IntegerAttr(-1));
+
+    auto _target =
+        rewriter.create<AtenMulScalarOp>(loc, target.getType(), target, _one);
+    auto _target_1 = rewriter.create<AtenAddScalarOp>(loc, _target.getType(),
+                                                      _target, one, one);
+    Value mm =
+        rewriter.create<AtenMulTensorOp>(loc, self.getType(), _target_1, self);
+    Value logSigm =
+        rewriter.create<AtenLogSigmoidOp>(loc, self.getType(), self);
+
+    if (!isa<Torch::NoneType>(posWeight.getType())) {
+      auto logWeight = rewriter.create<AtenAddScalarOp>(
+          loc, posWeight.getType(),
+          rewriter.create<AtenSubScalarOp>(loc, posWeight.getType(), posWeight,
+                                           one, one),
+          one, one);
+      loss = rewriter.create<AtenSubTensorOp>(
+          loc, mm.getType(), mm,
+          rewriter.create<AtenMulTensorOp>(loc, logWeight.getType(), logWeight,
+                                           logSigm),
+          one);
+    } else {
+      loss =
+          rewriter.create<AtenSubTensorOp>(loc, mm.getType(), mm, logSigm, one);
+    }
+
+    if (!isa<Torch::NoneType>(weight.getType())) {
+      loss =
+          rewriter.create<AtenMulTensorOp>(loc, loss.getType(), loss, weight);
+    }
+
+    // apply loss reduction.
+    int64_t reductionInt;
+    if (!matchPattern(reduction, m_TorchConstantInt(&reductionInt))) {
+      return rewriter.notifyMatchFailure(op, "no reduction type is appointed!");
+    }
+
+    auto none = rewriter.create<ConstantNoneOp>(loc);
+    Value res;
+    if (reductionInt == 1) {
+      res = rewriter.create<AtenMeanOp>(loc, op.getType(), loss, none);
+    } else if (reductionInt == 2) {
+      res = rewriter.create<AtenSumOp>(loc, op.getType(), loss, none);
+    } else {
+      res = loss;
+    }
+
+    rewriter.replaceOp(op, res);
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 class DecomposeAtenOneHotOp : public OpRewritePattern<AtenOneHotOp> {
   using OpRewritePattern<AtenOneHotOp>::OpRewritePattern;
@@ -9936,6 +10007,8 @@ public:
     addPatternIfTargetOpIsIllegal<DecomposeAtenMovedimIntOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenOneHotOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenCrossEntropyLossOp>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenBinaryCrossEntropyWithLogitsOp>(
+        patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenVarMeanDimOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenTopkOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenHannWindowPeriodicOp>(patterns);

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

@@ -1993,6 +1993,14 @@ def aten〇mse_loss〡shape(self: List[int], target: List[int], reduction: int =
 def aten〇cross_entropy_loss〡shape(self: List[int], target: List[int], weight: Optional[List[int]] = None, reduction: int = 1, ignore_index: int = -100, label_smoothing: float = 0.) -> List[int]:
     return upstream_shape_functions.cross_entropy_loss(self, target, weight, reduction, ignore_index, label_smoothing)
 
+def aten〇binary_cross_entropy_with_logits〡shape(self: List[int], target: List[int], weight: Optional[List[int]] = None, pos_weight: Optional[List[int]] = None, reduction: int = 1) -> List[int]:
+    scalar_shape: List[int] = []
+    if reduction == 0:
+        result_shape = upstream_shape_functions._copy(self)
+    else:
+        result_shape = scalar_shape
+    return result_shape
+
 @check_shape_function([
     Invocation(TensorOfShape(2, 5, 2, 2, 3), [2, 2, 3], None, None, 1e-6), # Basic case.
 ])
@@ -4958,6 +4966,10 @@ def aten〇linalg_norm〡dtype(self_rank_dtype: Tuple[int, int], ord: Optional[U
         return dtype
     return aten〇std〡dtype(self_rank_dtype)
 
+def aten〇binary_cross_entropy_with_logits〡dtype(self_rank_dtype: Tuple[int, int], target_rank_dtype: Tuple[int, int], weight_rank_dtype: Optional[Tuple[int, int]] = None, pos_weight_rank_dtype: Optional[Tuple[int, int]] = None, reduction: int = 1) -> int:
+    self_rank, self_dtype = self_rank_dtype
+    return self_dtype
+
 @check_dtype_function(
     _check_tensors_with_the_same_dtype(
         tensor_shapes=[(3,3)],

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

@@ -743,6 +743,9 @@ def emit_ops(emitter_td: TextEmitter, registry: Registry):
     emit(
         "aten::binary_cross_entropy_backward : (Tensor, Tensor, Tensor, Tensor?, int) -> (Tensor)"
     )
+    emit(
+        "aten::binary_cross_entropy_with_logits : (Tensor, Tensor, Tensor?, Tensor?, int) -> (Tensor)"
+    )
     emit("aten::log_sigmoid_forward : (Tensor) -> (Tensor, Tensor)")
     emit("aten::log_sigmoid_backward : (Tensor, Tensor, Tensor) -> (Tensor)")
     emit("aten::sigmoid_backward : (Tensor, Tensor) -> (Tensor)")

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

@@ -2294,6 +2294,29 @@ def CrossEntropyLossNoReductionModule_basic(module, tu: TestUtils):
     module.forward(tu.rand(8, 2), tu.randint(8, high=2))
 
 
+class BinaryCrossEntropyWithLogitsStaticModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([8, 2], torch.float32, True),
+            ([8, 2], torch.float32, True),
+        ]
+    )
+    def forward(self, input, target):
+        return torch.ops.aten.binary_cross_entropy_with_logits(
+            input, target, reduction=0
+        )
+
+
+@register_test_case(module_factory=lambda: BinaryCrossEntropyWithLogitsStaticModule())
+def BinaryCrossEntropyWithLogitsStaticModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(8, 2), tu.rand(8, 2))
+
+
 # ==============================================================================