[MLIR][TORCH] Add E2E support for aten.zeros op

This commit adds lowering of `aten.zeros` op.

Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>

e2e_testing/torchscript/basic.py

@@ -900,7 +900,6 @@ class NumelModule(torch.nn.Module):
         None,
         ([-1, -1, -1], torch.float32, True),
     ])
-
     def forward(self, input):
         return torch.numel(input)
 
@@ -918,10 +917,89 @@ class NumelZeroRankModule(torch.nn.Module):
         None,
         ([], torch.int64, True),
     ])
-
     def forward(self, input):
         return torch.numel(input)
 
 @register_test_case(module_factory=lambda: NumelZeroRankModule())
 def NumelZeroRankModule_basic(module, tu: TestUtils):
     module.forward(torch.randint(10,[]))
+
+
+class ZerosModuleInt2D(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+    ])
+    def forward(self):
+        return torch.zeros(3, 4, dtype=torch.int64)
+
+@register_test_case(module_factory=lambda: ZerosModuleInt2D())
+def ZerosModuleInt2D_basic(module, tu: TestUtils):
+    module.forward()
+
+
+class ZerosModuleInt3D(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+    ])
+    def forward(self):
+        return torch.zeros(3, 4, 5, dtype=torch.int64)
+
+@register_test_case(module_factory=lambda: ZerosModuleInt3D())
+def ZerosModuleInt3D_basic(module, tu: TestUtils):
+    module.forward()
+
+
+class ZerosModuleFloat2D(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+    ])
+    def forward(self):
+        return torch.zeros(3, 4, dtype=torch.float32)
+
+@register_test_case(module_factory=lambda: ZerosModuleFloat2D())
+def ZerosModuleFloat2D_basic(module, tu: TestUtils):
+    module.forward()
+
+
+class ZerosModuleFloat3D(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+    ])
+    def forward(self):
+        return torch.zeros(3, 4, 5, dtype=torch.float32)
+
+@register_test_case(module_factory=lambda: ZerosModuleFloat3D())
+def ZerosModuleFloat3D_basic(module, tu: TestUtils):
+    module.forward()
+
+
+class ZerosModuleFalsePinMemory(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+    ])
+    def forward(self):
+        return torch.zeros(3, 4, dtype=torch.float32, pin_memory=False)
+
+@register_test_case(module_factory=lambda: ZerosModuleFalsePinMemory())
+def ZerosModuleFalsePinMemory_basic(module, tu: TestUtils):
+    module.forward()

lib/Conversion/TorchToLinalg/TorchToLinalg.cpp

@@ -3269,24 +3269,44 @@ public:
 } // namespace
 
 namespace {
-class ConvertAtenOnesOp : public OpConversionPattern<AtenOnesOp> {
+// Converts AtenOnesOp and AtenZerosOp.
-public:
+struct ConvertAtenOnesZerosOp : ConversionPattern {
-  using OpConversionPattern::OpConversionPattern;
+  ConvertAtenOnesZerosOp(TypeConverter &typeConverter, MLIRContext *context)
+      : ConversionPattern(typeConverter, MatchAnyOpTypeTag(), /*benefit=*/1,
+                          context) {}
+
   LogicalResult
-  matchAndRewrite(AtenOnesOp op, OpAdaptor adaptor,
+  matchAndRewrite(Operation *op, ArrayRef<Value> operands,
                   ConversionPatternRewriter &rewriter) const override {
+    if (!isa<AtenOnesOp, AtenZerosOp>(op))
+      return rewriter.notifyMatchFailure(op,
+                                         "not a supported ones or zeros op");
+
     if (failed(verifyLinalgCompatibleTypes(op, rewriter)))
       return failure();
-    Location loc = op.getLoc();
+    Location loc = op->getLoc();
+
+    SmallVector<Value, 3> opArguments;
+    int64_t elementValue;
+
+    if (AtenOnesOp onesOp = dyn_cast<AtenOnesOp>(op)) {
+      opArguments.insert(opArguments.end(),
+                         {onesOp.size(), onesOp.layout(), onesOp.pin_memory()});
+      elementValue = 1;
+    } else if (AtenZerosOp zerosOp = dyn_cast<AtenZerosOp>(op)) {
+      opArguments.insert(opArguments.end(), {zerosOp.size(), zerosOp.layout(),
+                                             zerosOp.pin_memory()});
+      elementValue = 0;
+    }
 
     // We ignore device, but add simple asserts for unimplemented kwargs
-    if (!op.layout().getType().isa<Torch::NoneType>())
+    if (!opArguments[1].getType().isa<Torch::NoneType>())
       return rewriter.notifyMatchFailure(op,
                                          "only default layout is supported");
 
     bool pinMemory = false;
-    if (!op.pin_memory().getType().isa<Torch::NoneType>() &&
+    if (!opArguments[2].getType().isa<Torch::NoneType>() &&
-        !matchPattern(op.pin_memory(), m_TorchConstantBool(&pinMemory))) {
+        !matchPattern(opArguments[2], m_TorchConstantBool(&pinMemory))) {
       return rewriter.notifyMatchFailure(
           op, "pin_memory must be constant bool or None");
     }
@@ -3294,7 +3314,7 @@ public:
       return rewriter.notifyMatchFailure(op, "memory pinning not supported");
 
     SmallVector<Value> size, sizeIndex;
-    if (!getListConstructElements(op.size(), size)) {
+    if (!getListConstructElements(opArguments[0], size)) {
       return rewriter.notifyMatchFailure(
           op, "size must be created by ListConstruct");
     }
@@ -3303,21 +3323,24 @@ public:
       sizeIndex.push_back(castIntToIndex(rewriter, loc, size[i]));
 
     RankedTensorType newResultType =
-        getTypeConverter()->convertType(op.getType()).cast<RankedTensorType>();
+        getTypeConverter()
+            ->convertType(op->getResult(0).getType())
+            .cast<RankedTensorType>();
     Type outElementType = newResultType.getElementType();
 
-    Value one = rewriter.create<arith::ConstantOp>(
+    Value constantOp = rewriter.create<arith::ConstantOp>(
         loc, outElementType,
         (outElementType.isa<mlir::FloatType>()
-             ? rewriter.getFloatAttr(outElementType, 1).cast<mlir::Attribute>()
+             ? rewriter.getFloatAttr(outElementType, elementValue)
-             : rewriter.getIntegerAttr(outElementType, 1)
+                   .cast<mlir::Attribute>()
+             : rewriter.getIntegerAttr(outElementType, elementValue)
                    .cast<mlir::Attribute>()));
     Value outTensor = rewriter
                           .create<linalg::InitTensorOp>(
                               loc, sizeIndex, newResultType.getElementType())
                           .getResult();
-    Value fillOp =
+    Value fillOp = rewriter.create<linalg::FillOp>(loc, constantOp, outTensor)
-        rewriter.create<linalg::FillOp>(loc, one, outTensor).getResult(0);
+                       .getResult(0);
 
     rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, fillOp);
 
@@ -3452,8 +3475,8 @@ public:
     patterns.add<ConvertAtenSizeIntOp>(typeConverter, context);
     target.addIllegalOp<AtenEmbeddingOp>();
     patterns.add<ConvertAtenEmbeddingOp>(typeConverter, context);
-    target.addIllegalOp<AtenOnesOp>();
+    target.addIllegalOp<AtenOnesOp, AtenZerosOp>();
-    patterns.add<ConvertAtenOnesOp>(typeConverter, context);
+    patterns.add<ConvertAtenOnesZerosOp>(typeConverter, context);
     target.addIllegalOp<AtenContiguousOp>();
     patterns.add<ConvertAtenContiguousOp>(typeConverter, context);
     target.addIllegalOp<AtenIntTensorOp>();

lib/Dialect/Torch/Transforms/RefineTypes.cpp

@@ -458,7 +458,6 @@ public:
       return visitAtenNllLossForwardOp(nllForwardOp, operands);
     }
 
-
     // Otherwise, this is an unknown operation. Just mark all results as
     // having reached a pessimistic fixpoint.
     return markAllPessimisticFixpoint(op->getResults());