Add aten::nll_loss_forward op lowering.

The op lowering has been added as a part of `torch-lower-to-linalg`
pass. This takes care of ignore_index but the weight and reduction
operand is still to be accounted for.

Signed-off-by: Prashant Kumar <[email protected]>

Author: Prashant Kumar

e2e_testing/torchscript/main.py

@@ -43,6 +43,7 @@
 from . import scalar
 from . import squeeze
 from . import slice_like
+from . import nll_loss
 
 def _get_argparse():
     config_choices = ['native_torch', 'torchscript', 'refbackend', 'tosa', 'external']

e2e_testing/torchscript/nll_loss.py

@@ -0,0 +1,62 @@
+# Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+# Also available under a BSD-style license. See LICENSE.
+
+import torch
+
+from torch_mlir_e2e_test.torchscript.framework import TestUtils
+from torch_mlir_e2e_test.torchscript.registry import register_test_case
+from torch_mlir_e2e_test.torchscript.annotations import annotate_args, export
+
+# ==============================================================================
+
+
+class NllLossModule(torch.nn.Module):
+
+  def __init__(self):
+    super().__init__()
+
+  @export
+  @annotate_args([
+      None,
+      ([-1, -1], torch.float32, True),
+      ([-1], torch.int64, True),
+  ])
+  # Here the 2nd index is ignored.
+  def forward(self, x, y):
+    return torch.ops.aten.nll_loss_forward(self=x,
+                                           target=y,
+                                           weight=None,
+                                           reduction=0,
+                                           ignore_index=2)[0]
+
+
+@register_test_case(module_factory=lambda: NllLossModule())
+def NllLossModule_basic(module, tu: TestUtils):
+  module.forward(tu.rand(2, 3), torch.tensor([0, 1]))
+
+
+class NllLossModule_ignore_index_out_of_bounds(torch.nn.Module):
+
+  def __init__(self):
+    super().__init__()
+
+  @export
+  @annotate_args([
+      None,
+      ([-1, -1], torch.float32, True),
+      ([-1], torch.int64, True),
+  ])
+  # None of the index is ignored here, since the ignored index is out of bounds.
+  def forward(self, x, y):
+    return torch.ops.aten.nll_loss_forward(self=x,
+                                           target=y,
+                                           weight=None,
+                                           reduction=0,
+                                           ignore_index=10)[0]
+
+
+@register_test_case(module_factory=lambda: NllLossModule_ignore_index_out_of_bounds())
+def NllLossModule_ignore_index(module, tu: TestUtils):
+  module.forward(tu.rand(2, 3), torch.tensor([0, 1]))

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

@@ -1573,6 +1573,25 @@ def Torch_AtenMeanOp : Torch_Op<"aten.mean", [
   let assemblyFormat = "$self `,` $dtype attr-dict `:` type($self) `,` type($dtype) `->` type($result)";
 }
 
+def Torch_AtenNllLossForwardOp : Torch_Op<"aten.nll_loss_forward", [
+    AllowsTypeRefinement,
+    HasValueSemantics
+  ]> {
+  let summary = "Generated op for `aten::nll_loss_forward : (Tensor, Tensor, Tensor?, int, int) -> (Tensor, Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self,
+    AnyTorchTensorType:$target,
+    AnyTorchOptionalTensorType:$weight,
+    Torch_IntType:$reduction,
+    Torch_IntType:$ignore_index
+  );
+  let results = (outs
+    AnyTorchTensorType:$output,
+    AnyTorchTensorType:$total_weight
+  );
+  let assemblyFormat = "$self `,` $target `,` $weight `,` $reduction `,` $ignore_index attr-dict `:` type($self) `,` type($target) `,` type($weight) `,` type($reduction) `,` type($ignore_index) `->` type($output) `,` type($total_weight)";
+}
+
 def Torch_AtenUnsqueezeOp : Torch_Op<"aten.unsqueeze", [
     AllowsTypeRefinement
   ]> {

lib/Conversion/TorchToLinalg/TorchToLinalg.cpp

@@ -1167,6 +1167,102 @@ class ConvertAtenDropoutOp : public OpConversionPattern<AtenDropoutOp> {
 };
 } // namespace
 
+// Given `input`, `target`, `nll_loss_forward` is given by:
+//   for i in range(0, len(target)):
+//     indi = target[i];
+//     nll_loss_forward[i] = -(input[i][indi]);
+// TODO: `weight` and `reduction` operands are still to be taken care of.
+namespace {
+class ConvertAtenNllLossForwardOp
+    : public OpConversionPattern<AtenNllLossForwardOp> {
+public:
+  using OpConversionPattern::OpConversionPattern;
+  LogicalResult
+  matchAndRewrite(AtenNllLossForwardOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    if (failed(verifyLinalgCompatibleTypes(op, rewriter)))
+      return failure();
+    Location loc = op->getLoc();
+    Value input = adaptor.self();
+    Value target = adaptor.target();
+    Value weight = adaptor.weight();
+
+    int64_t reduce_dim;
+    if (!matchPattern(op.reduction(), m_TorchConstantInt(&reduce_dim)))
+      return rewriter.notifyMatchFailure(op, "dim must be constant");
+
+    // TODO: Handle reduction.
+    if (reduce_dim != 0)
+      return rewriter.notifyMatchFailure(
+          op, "reduction along dimensions is not supported.");
+
+    // TODO: Incorporate the weight argument.
+    if (!weight.getType().isa<mlir::torch::Torch::NoneType>())
+      return rewriter.notifyMatchFailure(
+          op, "Unimplemented, the weight operand is not incorporated.");
+
+    Value ignoreIndex = adaptor.ignore_index();
+    Value ignoreIndexVal = castIntToIndex(rewriter, loc, ignoreIndex);
+
+    unsigned inputRank = input.getType().cast<RankedTensorType>().getRank();
+    unsigned targetRank = target.getType().cast<RankedTensorType>().getRank();
+
+    // TODO: Cases with targetRank != 1 where `Mean` reduction is required.
+    if (inputRank != 2 || targetRank != 1) {
+      return rewriter.notifyMatchFailure(
+          op, "expected  input and target to be rank 2 and 1 respectively");
+    }
+    RankedTensorType resultType = getTypeConverter()
+                          ->convertType(op->getResult(0).getType())
+                          .cast<RankedTensorType>();
+
+    Type elementType = resultType.getElementType();
+
+    Value targetDim = getDimOp(rewriter, loc, target, 0);
+    Value initTensor0 =
+        createZeroInitTensor(rewriter, loc, {targetDim}, elementType);
+    Value zeroVal = rewriter.create<arith::ConstantOp>(
+        loc, rewriter.getZeroAttr(elementType));
+
+    SmallVector<AffineExpr> targetExpr;
+    targetExpr.push_back(rewriter.getAffineDimExpr(0));
+    SmallVector<StringRef> iteratorTypes{getParallelIteratorTypeName()};
+    auto indexingMaps = AffineMap::inferFromExprList({targetExpr, targetExpr});
+    Value finalRes =
+        rewriter
+            .create<linalg::GenericOp>(
+                loc, resultType, ValueRange{target}, initTensor0,
+                /*indexingMaps=*/indexingMaps,
+                /*iteratorTypes=*/iteratorTypes,
+                [&](OpBuilder &b, Location loc, ValueRange args) {
+                  Value indTarget = rewriter.create<arith::IndexCastOp>(
+                      loc, rewriter.getIndexType(), args[0]);
+                  Value indI = rewriter.create<linalg::IndexOp>(loc, 0);
+
+                  // The final result is given by:
+                  // final_res = (indI == ignoreIndexVal) ? 0 :
+                  // input[indI][IndTarget]
+                  Value cmpEq = rewriter.create<arith::CmpIOp>(
+                      loc, arith::CmpIPredicate::eq, indI, ignoreIndexVal);
+                  Value result = rewriter.create<tensor::ExtractOp>(
+                      loc, input, ValueRange{indI, indTarget});
+                  Value negate =
+                      rewriter.create<arith::NegFOp>(loc, elementType, result);
+                  Value selectFinal = rewriter.create<mlir::SelectOp>(
+                      loc, cmpEq, zeroVal, negate);
+                  b.create<linalg::YieldOp>(loc, selectFinal);
+                })
+            .getResult(0);
+
+    // TODO: Update the second result tensor.
+    Value weightUpdated =
+        createZeroInitTensor(rewriter, loc, {}, elementType);
+    rewriter.replaceOp(op, {finalRes, weightUpdated});
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 // See comments at in convertMmOp and the heading for this section for general
 // considerations. This function needs to be auto-generated.
@@ -3372,6 +3468,8 @@ class ConvertTorchToLinalg
     patterns.add<ConvertAtenNumelOp>(typeConverter, context);
     target.addIllegalOp<AtenSliceTensorOp>();
     patterns.add<ConvertAtenSliceTensorOp>(typeConverter, context);
+    target.addIllegalOp<AtenNllLossForwardOp>();
+    patterns.add<ConvertAtenNllLossForwardOp>(typeConverter, context);
 
     if (failed(applyPartialConversion(getOperation(), target,
                                       std::move(patterns))))

lib/Dialect/Torch/Transforms/RefineTypes.cpp

@@ -454,8 +454,11 @@ class TypeAnalyzer : public ForwardDataFlowAnalysis<ValueKnowledge> {
       return visitAtenAddCLikeOp(op, operands);
     } else if (auto scalarOp = dyn_cast<AtenAddIntOp>(op)) {
       return visitBinaryScalarOp(scalarOp);
+    }else if (auto nllForwardOp = dyn_cast<AtenNllLossForwardOp>(op)) {
+      return visitAtenNllLossForwardOp(nllForwardOp, operands);
     }
 
+
     // Otherwise, this is an unknown operation. Just mark all results as
     // having reached a pessimistic fixpoint.
     return markAllPessimisticFixpoint(op->getResults());
@@ -580,6 +583,10 @@ class TypeAnalyzer : public ForwardDataFlowAnalysis<ValueKnowledge> {
   ChangeResult
   visitAten_SoftmaxOp(Aten_SoftmaxOp op,
                       ArrayRef<LatticeElement<ValueKnowledge> *> operands);
+
+  ChangeResult
+  visitAtenNllLossForwardOp(AtenNllLossForwardOp op,
+                      ArrayRef<LatticeElement<ValueKnowledge> *> operands);
 };
 } // namespace
 
@@ -927,6 +934,38 @@ ChangeResult TypeAnalyzer::visitAtenSqueezeOp(
   return getLatticeElement(op.getResult()).join(knowledge);
 }
 
+ChangeResult TypeAnalyzer::visitAtenNllLossForwardOp(
+    AtenNllLossForwardOp op,
+    ArrayRef<LatticeElement<ValueKnowledge> *> operands) {
+  auto self = operands[0]->getValue();
+  auto outputKnowledge =
+      ValueKnowledge::getNotNonePessimisticValueState(op.getContext());
+
+  // Contains Knowledge of shape and dtype for the 1st result.
+  outputKnowledge.dtype = self.dtype;
+  int64_t reduction;
+  unsigned resultRank = self.sizes.size();
+
+  // Contains Knowledge of shape and dtype for the 2nd result.
+  auto totalWeightKnowledge =
+      ValueKnowledge::getNotNonePessimisticValueState(op.getContext());
+  totalWeightKnowledge.dtype = self.dtype;
+  totalWeightKnowledge.sizes.resize(0, kUnknownSize);
+  totalWeightKnowledge.hasSizes = true;
+
+  if (self.hasSizes &&
+      matchPattern(op.reduction(), m_TorchConstantInt(&reduction))) {
+    // reduction == 1 means reduce 1st dim.
+    resultRank = reduction == 1 ? resultRank - 1 : resultRank;
+  }
+  outputKnowledge.sizes.resize(resultRank - 1, kUnknownSize);
+  outputKnowledge.hasSizes = true;
+  auto resultLattice = getLatticeElement(op.getResult(0)).join(outputKnowledge);
+  resultLattice |=
+      getLatticeElement(op.getResult(1)).join(totalWeightKnowledge);
+  return resultLattice;
+}
+
 ChangeResult TypeAnalyzer::visitAtenUnsqueezeOp(
     AtenUnsqueezeOp op, ArrayRef<LatticeElement<ValueKnowledge> *> operands) {
   auto operand = operands[0]->getValue();

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::sqrt : (Tensor) -> (Tensor)")
         emit("aten::_softmax : (Tensor, int, bool) -> (Tensor)")
         emit("aten::mean : (Tensor, int?) -> (Tensor)")
+        emit("aten::nll_loss_forward : (Tensor, Tensor, Tensor?, int, int) -> (Tensor, Tensor)")
 
         # Misc tensor ops.
         emit("aten::unsqueeze : (Tensor, int) -> (Tensor)")