[TORCH] Add Kullback-Leibler divergence loss support (#4204)

This PR takes care of #4203.
- e2e support of **aten.kl_div** op supporting all reduction modes
(`mean, sum, batchmean, none`)
- `reduction: batchmean` requires special handling by calling op with
`sum` and then dividing it by input `batch_size`.

Some tests are failing and are marked either in expected failures or
crashing set.
- **config=linalg** | **RuntimeError**: attribute lookup is not defined
on builtin | **LINALG_XFAIL_SET**
- **config=torchdynamo** | **Error**: failed to legalize operation
'`torch.aten.xlogy.Tensor`' | **TORCHDYNAMO_CRASHING_SET**
- **config=onnx** | **RuntimeError**: aten::div() Expected a value of
type 'number' for argument 'other' but instead found type 'Tensor'
Position: 1
        Value: tensor(1)
Declaration: aten::div.Scalar(Tensor self, Scalar other) -> Tensor
Cast error details: Cannot cast tensor(1) to number | **ONNX_XFAIL_SET**
- **config=onnx_tosa** | **Error**: failed to legalize operation
'`torch.aten.size.int`' that was explicitly marked illegal |
**ONNX_TOSA_XFAIL_SET**

---------

Signed-off-by: Zahid Wakeel <zahid.wakeel@multicorewareinc.com>

Author: zahidwx

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

@@ -9506,6 +9506,32 @@ def Torch_AtenPoissonNllLossOp : Torch_Op<"aten.poisson_nll_loss", [
   }];
 }
 
+def Torch_AtenKlDivOp : Torch_Op<"aten.kl_div", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::kl_div : (Tensor, Tensor, int, bool) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self,
+    AnyTorchTensorType:$target,
+    Torch_IntType:$reduction,
+    Torch_BoolType:$log_target
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenKlDivOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 4, 1);
+    }
+    void AtenKlDivOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 4, 1);
+    }
+  }];
+}
+
 def Torch_AtenBincountOp : Torch_Op<"aten.bincount", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -10692,6 +10692,31 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %0 = call @__torch__.torch.jit._shape_functions.unary(%arg0) : (!torch.list<int>) -> !torch.list<int>\n"
 "    return %0 : !torch.list<int>\n"
 "  }\n"
+"  func.func @\"__torch_mlir_shape_fn.aten.kl_div\"(%arg0: !torch.list<int>, %arg1: !torch.list<int>, %arg2: !torch.int, %arg3: !torch.bool) -> !torch.list<int> {\n"
+"    %none = torch.constant.none\n"
+"    %str = torch.constant.str \"AssertionError: Invalid reduction value.\"\n"
+"    %int0 = torch.constant.int 0\n"
+"    %int1 = torch.constant.int 1\n"
+"    %int2 = torch.constant.int 2\n"
+"    %0 = torch.prim.Uninitialized : !torch.list<int>\n"
+"    %1 = torch.aten.eq.int %arg2, %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.unary(%arg0) : (!torch.list<int>) -> !torch.list<int>\n"
+"      torch.prim.If.yield %3 : !torch.list<int>\n"
+"    } else {\n"
+"      %3 = torch.prim.ListConstruct %int1, %int2 : (!torch.int, !torch.int) -> !torch.list<int>\n"
+"      %4 = torch.aten.__contains__.int_list %3, %arg2 : !torch.list<int>, !torch.int -> !torch.bool\n"
+"      %5 = torch.prim.If %4 -> (!torch.list<int>) {\n"
+"        %6 = torch.prim.ListConstruct  : () -> !torch.list<int>\n"
+"        torch.prim.If.yield %6 : !torch.list<int>\n"
+"      } else {\n"
+"        torch.prim.RaiseException %str, %none : !torch.str, !torch.none\n"
+"        torch.prim.If.yield %0 : !torch.list<int>\n"
+"      }\n"
+"      torch.prim.If.yield %5 : !torch.list<int>\n"
+"    }\n"
+"    return %2 : !torch.list<int>\n"
+"  }\n"
 "  func.func @\"__torch_mlir_shape_fn.aten.nll_loss_forward\"(%arg0: !torch.list<int>, %arg1: !torch.list<int>, %arg2: !torch.optional<list<int>>, %arg3: !torch.int, %arg4: !torch.int) -> !torch.tuple<list<int>, list<int>> {\n"
 "    %0 = call @__torch__.torch.jit._shape_functions.nll_loss_forward(%arg0, %arg1, %arg2, %arg3) : (!torch.list<int>, !torch.list<int>, !torch.optional<list<int>>, !torch.int) -> !torch.tuple<list<int>, list<int>>\n"
 "    return %0 : !torch.tuple<list<int>, list<int>>\n"
@@ -14575,6 +14600,14 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    }\n"
 "    return %int3 : !torch.int\n"
 "  }\n"
+"  func.func @\"__torch_mlir_dtype_fn.aten.kl_div\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.tuple<int, int>, %arg2: !torch.int, %arg3: !torch.bool) -> !torch.int {\n"
+"    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
+"    %1:2 = torch.prim.TupleUnpack %arg1 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
+"    %2 = torch.prim.ListConstruct %0#0, %1#0 : (!torch.int, !torch.int) -> !torch.list<optional<int>>\n"
+"    %3 = torch.prim.ListConstruct %0#1, %1#1 : (!torch.int, !torch.int) -> !torch.list<int>\n"
+"    %4 = call @__torch__.torch_mlir.jit_ir_importer.build_tools.library_generator.promote_dtypes(%2, %3) : (!torch.list<optional<int>>, !torch.list<int>) -> !torch.int\n"
+"    return %4 : !torch.int\n"
+"  }\n"
 "  func.func @\"__torch_mlir_dtype_fn.aten.mse_loss\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.tuple<int, int>, %arg2: !torch.int) -> !torch.int {\n"
 "    %none = torch.constant.none\n"
 "    %str = torch.constant.str \"AssertionError: \"\n"

lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp

@@ -10629,6 +10629,84 @@ class DecomposeAtenPoissonNllLossOp
 };
 } // namespace
 
+namespace {
+class DecomposeAtenKlDivOp : public OpRewritePattern<AtenKlDivOp> {
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenKlDivOp op,
+                                PatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    Value self = op.getSelf();
+    Value target = op.getTarget();
+    Value reductionValue = op.getReduction();
+    Value logTargetValue = op.getLogTarget();
+
+    auto selfTy = cast<ValueTensorType>(self.getType());
+    auto targetTy = cast<ValueTensorType>(target.getType());
+    auto outTy = cast<ValueTensorType>(op.getType());
+
+    if (!selfTy.hasSizes() || !targetTy.hasSizes() || !outTy.hasSizes()) {
+      return rewriter.notifyMatchFailure(
+          op, "require self, target and output having sizes!");
+    }
+
+    if (!selfTy.hasDtype() || !targetTy.hasDtype() || !outTy.hasDtype()) {
+      return rewriter.notifyMatchFailure(
+          op, "require self, target and output having dtype!");
+    }
+
+    // Extract boolean value from logTarget argument
+    bool logTargetBool;
+    if (!matchPattern(logTargetValue, m_TorchConstantBool(&logTargetBool)))
+      return rewriter.notifyMatchFailure(
+          op, "Expected a constant boolean value for logTargetBool");
+
+    // Default: target tensor is not in log space
+    Value logOfTarget;
+    if (!logTargetBool) {
+      logOfTarget = rewriter.create<AtenLogOp>(loc, targetTy, target);
+    } else {
+      logOfTarget = target;
+    }
+
+    Value constOne =
+        rewriter.create<ConstantIntOp>(loc, rewriter.getI64IntegerAttr(1));
+    Value subValue = rewriter.create<AtenSubTensorOp>(loc, selfTy, logOfTarget,
+                                                      self, constOne);
+
+    // if target tensor is already in log space
+    if (logTargetBool) {
+      target = rewriter.create<AtenExpOp>(loc, targetTy, target);
+    }
+    Value lossPointwise =
+        rewriter.create<AtenMulTensorOp>(loc, targetTy, target, subValue);
+
+    // Extract reduction int value from reduction argument
+    int64_t reduction;
+    if (!matchPattern(reductionValue, m_TorchConstantInt(&reduction))) {
+      return rewriter.notifyMatchFailure(op,
+                                         "reduction should be a constant int!");
+    }
+
+    Value loss;
+    Value none = rewriter.create<ConstantNoneOp>(loc);
+    // reduction: mean
+    if (reduction == 1) {
+      loss = rewriter.create<AtenMeanOp>(loc, outTy, lossPointwise, none);
+    } else if (reduction == 2) {
+      // reduction: sum
+      loss = rewriter.create<AtenSumOp>(loc, outTy, lossPointwise, none);
+    } else {
+      // reduction: none
+      loss = lossPointwise;
+    }
+
+    rewriter.replaceOp(op, loss);
+
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 class DecomposeAtenBinaryCrossEntropyWithLogitsOp
     : public OpRewritePattern<AtenBinaryCrossEntropyWithLogitsOp> {
@@ -12546,6 +12624,7 @@ class DecomposeComplexOpsPass
     addPatternIfTargetOpIsIllegal<DecomposeAtenPoissonNllLossOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenBinaryCrossEntropyWithLogitsOp>(
         patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenKlDivOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenVarMeanDimOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenTopkOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenArgsortOp>(patterns);

lib/Dialect/Torch/Transforms/LowerToBackendContract.cpp

@@ -587,6 +587,7 @@ static void markDecomposedOpsAsIllegal(MLIRContext *context,
   target.addIllegalOp<AtenFlipudOp>();
   target.addIllegalOp<AtenLogaddexpOp>();
   target.addIllegalOp<AtenLogaddexp2Op>();
+  target.addIllegalOp<AtenKlDivOp>();
 
   for (auto &opName : backendLegalOpsSet) {
     target.addLegalOp(

projects/pt1/e2e_testing/xfail_sets.py

@@ -39,6 +39,8 @@
     "AtenSymConstrainRange_basic",
     "AtenSymConstrainRangeForSize_basic",
     "Aten_AssertScalar_basic",
+    # RuntimeError: attribute lookup is not defined on builtin:
+    "KlDivLossModule_batchmean_reduction_basic",
 }
 
 if torch_version_for_comparison() < version.parse("2.5.0.dev"):
@@ -386,6 +388,12 @@
     "MaxPool3dStaticModule_basic",
     # Looks like incorrect fx graph conversion
     "ElementwiseAddScalar_TensorLiteralInt32_Module_basic",
+    # error: failed to legalize operation 'torch.aten.xlogy.Tensor'
+    "KlDivLossModule_default_basic",
+    "KlDivLossModule_reduction_is_none_basic",
+    "KlDivLossModule_mean_reduction_basic",
+    "KlDivLossModule_sum_reduction_basic",
+    "KlDivLossModule_batchmean_reduction_basic",
 }
 
 FX_IMPORTER_XFAIL_SET = {
@@ -3087,6 +3095,7 @@
     "PoissonNLLLossMeanReductionModule_basic",
     "PoissonNLLLossSumReductionModule_basic",
     "PoissonNLLLossNonDefaultEpsModule_basic",
+    "KlDivLossModule_batchmean_reduction_basic",
     "NormScalarComplexModule_basic",
     "NormScalarModule_basic",
     "NormScalarOptDimKeepDimComplexModule_basic",
@@ -3982,6 +3991,12 @@
     "NllLossStaticModule_mean_basic",
     "NllLossStaticModule_sum_basic",
     "NllLossStaticModule_weight_basic",
+    "KlDivLossModule_default_basic",
+    "KlDivLossModule_reduction_is_none_basic",
+    "KlDivLossModule_reduction_is_none_log_target_is_true_basic",
+    "KlDivLossModule_mean_reduction_basic",
+    "KlDivLossModule_sum_reduction_basic",
+    "KlDivLossModule_batchmean_reduction_basic",
     "Exp2StaticModule_basic",
     "ElementwiseRreluWithNoiseEvalModule_basic",
     "ElementwiseRreluWithNoiseEvalStaticModule_basic",

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

@@ -2174,6 +2174,14 @@ def aten〇tril_indices〡shape(row: int, col: int, offset: int = 0, dtype: Opti
 def aten〇deg2rad〡shape(self: List[int]) -> List[int]:
     return upstream_shape_functions.unary(self)
 
+def aten〇kl_div〡shape(self: List[int], target: List[int], reduction: int = 1, log_target: bool = False) -> List[int]:
+    if reduction == 0:
+        return upstream_shape_functions.unary(self)
+    elif reduction in [1, 2]:
+        return []
+    else:
+        assert False, "Invalid reduction value."
+
 @check_shape_function([
     Invocation(TensorOfShape(2, 3), LongTensorOfShape(2), None, 1, -100), # Basic case.
     Invocation(TensorOfShape(3), LongTensorOfShape(), None, 1, -100), # No batch dim.
@@ -4552,6 +4560,14 @@ def aten〇_int_mm〡dtype(self_rank_dtype: Tuple[int, int], mat2_rank_dtype: Tu
     assert mat2_dtype == torch.int8
     return torch.int32
 
+def aten〇kl_div〡dtype(self_rank_dtype: Tuple[int, int], target_rank_dtype: Tuple[int, int], reduction: int = 1, log_target: bool = False) -> int:
+    self_rank, self_dtype = self_rank_dtype
+    target_rank, target_dtype = target_rank_dtype
+    ranks: List[Optional[int]] = [self_rank, target_rank]
+    dtypes = [self_dtype, target_dtype]
+    promoted_dtype = promote_dtypes(ranks, dtypes)
+    return promoted_dtype
+
 @check_dtype_function(_check_two_tensor_op(
     output_error_types={torch.bool, torch.int8, torch.uint8, torch.int16, torch.int32, torch.int64}))
 def aten〇mse_loss〡dtype(self_rank_dtype: Tuple[int, int], target_rank_dtype: Tuple[int, int], reduction: int = 1) -> int:

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

@@ -764,6 +764,7 @@ def emit_with_mutating_variants(key, **kwargs):
     emit(
         "aten::poisson_nll_loss : (Tensor, Tensor, bool, bool, float, int) -> (Tensor)"
     )
+    emit("aten::kl_div : (Tensor, Tensor, int, bool) -> (Tensor)")
     emit("aten::bincount : (Tensor, Tensor?, int) -> (Tensor)")
     emit("aten::linalg_vector_norm : (Tensor, Scalar, int[]?, bool, int?) -> (Tensor)")
     emit("aten::linalg_norm : (Tensor, Scalar?, int[]?, bool, int?) -> (Tensor)")

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

@@ -62,3 +62,4 @@ def register_all_tests():
     from . import gridsampler
     from . import meshgrid
     from . import timeout
+    from . import kl_div_loss