[TORCH][MLIR] Add and templatize lowering of [aten.zeros|aten.ones|aten.empty] ops

- Templatize `aten.zeros` and `aten.ones` ops lowering.
- Add E2E support for `aten.empty` op.
- Add Integer type support in `aten.mul.Scalar` op lowering.

Signed-Off-by: Gaurav Shukla <[email protected]>

Author: Gaurav Shukla

e2e_testing/torchscript/basic.py

@@ -637,6 +637,57 @@ def OnesModuleFalsePinMemory_basic(module, tu: TestUtils):
 
 # ==============================================================================
 
+class EmptyIntModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+    ])
+    def forward(self):
+        return 0 * torch.empty((3, 4), dtype=torch.int64)
+
+@register_test_case(module_factory=lambda: EmptyIntModule())
+def EmptyModule_int(module, tu: TestUtils):
+    module.forward()
+
+# ==============================================================================
+
+class EmptyFloatModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+    ])
+    def forward(self):
+        return torch.abs(torch.empty((3, 4), dtype=torch.float32)) > -1.0
+
+@register_test_case(module_factory=lambda: EmptyFloatModule())
+def EmptyModule_float(module, tu: TestUtils):
+    module.forward()
+
+
+class EmptyFalsePinMemoryModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+    ])
+    def forward(self):
+        return torch.abs(torch.empty((3, 4), dtype=torch.float32, 
+                                     pin_memory=False)) > -1.0 
+
+@register_test_case(module_factory=lambda: EmptyFalsePinMemoryModule())
+def EmptyModule_falsePinMemory(module, tu: TestUtils):
+    module.forward()
+
+# ==============================================================================
+
 class ContiguousModule(torch.nn.Module):
     def __init__(self):
         super().__init__()

e2e_testing/torchscript/elementwise.py

@@ -443,8 +443,24 @@ def RsubModule_noalpha_basic(module, tu: TestUtils):
 
 # ==============================================================================
 
+class ElementwiseMulScalarIntModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
 
-class ElementwiseMulScalarModule(torch.nn.Module):
+    @export
+    @annotate_args([
+        None,
+        ([-1, -1], torch.int64, True),
+    ])
+    def forward(self, x):
+        return torch.mul(x, 4)
+
+@register_test_case(module_factory=lambda: ElementwiseMulScalarIntModule())
+def ElementwiseMulScalarModule_int(module, tu: TestUtils):
+    module.forward(torch.randint(10, (3, 4)))
+
+
+class ElementwiseMulScalarFloatModule(torch.nn.Module):
     def __init__(self):
         super().__init__()
 
@@ -456,11 +472,27 @@ def __init__(self):
     def forward(self, x):
         return torch.mul(x, 100.0)
 
-@register_test_case(module_factory=lambda: ElementwiseMulScalarModule())
-def ElementwiseMulScalarModule_basic(module, tu: TestUtils):
+@register_test_case(module_factory=lambda: ElementwiseMulScalarFloatModule())
+def ElementwiseMulScalarModule_float(module, tu: TestUtils):
     module.forward(tu.rand(3, 4))
 
 
+class ElementwiseMulScalarModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+        ([-1, -1], torch.int64, True),
+    ])
+    def forward(self, x):
+        return torch.mul(x, 8.0)
+
+@register_test_case(module_factory=lambda: ElementwiseMulScalarModule())
+def ElementwiseMulScalarModule_basic(module, tu: TestUtils):
+    module.forward(torch.randint(10, (3, 4)))
+
 
 class ElementwiseMulTensorFloatModule(torch.nn.Module):
     def __init__(self):

lib/Conversion/TorchToLinalg/TorchToLinalg.cpp

@@ -189,6 +189,21 @@ static Value createInitTensor(OpBuilder &b, Location loc, ValueRange sizes,
   Value initTensor = b.create<linalg::InitTensorOp>(loc, sizes, elemTy);
   return b.create<linalg::FillOp>(loc, initElem, initTensor).getResult(0);
 }
+// Creates a constant of type `elemType` with value `val`.
+static Value getConstant(OpBuilder &b, Location loc, int64_t val,
+                         Type elemType) {
+  Attribute attr = {};
+  if (elemType.isa<mlir::FloatType>())
+    attr = b.getFloatAttr(elemType, val);
+  if (elemType.isa<mlir::IndexType>())
+    attr = b.getIndexAttr(val);
+  if (elemType.isa<mlir::IntegerType>())
+    attr = b.getIntegerAttr(
+        elemType, APInt(elemType.cast<IntegerType>().getWidth(), val));
+  if (!attr)
+    return nullptr;
+  return b.create<arith::ConstantOp>(loc, elemType, attr);
+}
 
 // Helper function to caculate the output tensor dims for convolution-like ops.
 // Along each dim:
@@ -1828,13 +1843,14 @@ static Value createLinalgPayloadCalculationForElementwiseOp(
     Type dtype = converter->convertType(mulScalar.getType())
                      .cast<RankedTensorType>()
                      .getElementType();
-    if (!dtype.isa<mlir::FloatType>()) {
-      mulScalar.emitError("unimplemented: non-floating point dtype");
-      return nullptr;
-    }
-    Value self = payloadArgs[0];
-    Value other = convertScalarToDtype(b, loc, operands[1], dtype);
-    return b.create<arith::MulFOp>(loc, self, other);
+    Value lhs = convertScalarToDtype(b, loc, payloadArgs[0], dtype);
+    Value rhs = convertScalarToDtype(b, loc, operands[1], dtype);
+    if (dtype.isa<mlir::FloatType>())
+      return b.create<arith::MulFOp>(loc, lhs, rhs);
+    if (dtype.isa<mlir::IntegerType>())
+      return b.create<arith::MulIOp>(loc, lhs, rhs);
+    mulScalar.emitError("unimplemented: Only integer/float dtype supported");
+    return nullptr;
   }
   if (auto atenToDtype = dyn_cast<AtenToDtypeOp>(op)) {
     Value input = payloadArgs[0];
@@ -3417,83 +3433,68 @@ class ConvertAtenContiguousOp : public OpConversionPattern<AtenContiguousOp> {
 } // namespace
 
 namespace {
-// Converts AtenOnesOp and AtenZerosOp.
-struct ConvertAtenOnesZerosOp : ConversionPattern {
-  ConvertAtenOnesZerosOp(TypeConverter &typeConverter, MLIRContext *context)
-      : ConversionPattern(typeConverter, MatchAnyOpTypeTag(), /*benefit=*/1,
-                          context) {}
-
+// Converts constant tensor allocation like ops.
+template <typename OpTy>
+class ConvertConstantTensorAllocOp : public OpConversionPattern<OpTy> {
+public:
+  using OpConversionPattern<OpTy>::OpConversionPattern;
   LogicalResult
-  matchAndRewrite(Operation *op, ArrayRef<Value> operands,
+  matchAndRewrite(OpTy op, typename OpTy::Adaptor adaptor,
                   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();
-
-    Value size, layout, pin_memory;
-    int64_t elementValue;
 
-    if (AtenOnesOp onesOp = dyn_cast<AtenOnesOp>(op)) {
-      size = onesOp.size();
-      layout = onesOp.layout();
-      pin_memory = onesOp.pin_memory();
-      elementValue = 1;
-    } else if (AtenZerosOp zerosOp = dyn_cast<AtenZerosOp>(op)) {
-      size = zerosOp.size();
-      layout = zerosOp.layout();
-      pin_memory = zerosOp.pin_memory();
-      elementValue = 0;
+    // Currently memory pinning and layout features are not supported.
+    if (!op.layout().getType().template isa<Torch::NoneType>())
+      return rewriter.notifyMatchFailure(
+          op, "unimplemented: only default layout is supported");
+    bool pinMemory;
+    if (!op.pin_memory().getType().template isa<Torch::NoneType>() &&
+        (!matchPattern(op.pin_memory(), m_TorchConstantBool(&pinMemory)) ||
+         pinMemory)) {
+      return rewriter.notifyMatchFailure(
+          op, "unimplemented: pin_memory must be either None or false");
     }
 
-    // We ignore device, but add simple asserts for unimplemented kwargs
-    if (!layout.getType().isa<Torch::NoneType>())
-      return rewriter.notifyMatchFailure(op,
-                                         "only default layout is supported");
-
-    bool pinMemory = false;
-    if (!pin_memory.getType().isa<Torch::NoneType>() &&
-        !matchPattern(pin_memory, m_TorchConstantBool(&pinMemory))) {
-      return rewriter.notifyMatchFailure(
-          op, "pin_memory must be constant bool or None");
+    // Memory formats are not supported in the case of `AtenEmptyMemoryFormat`.
+    if constexpr (std::is_same<OpTy, AtenEmptyMemoryFormatOp>::value) {
+      if (!op.memory_format().getType().template isa<Torch::NoneType>())
+        return rewriter.notifyMatchFailure(
+            op, "unimplemented: only default memory format is supported");
     }
-    if (pinMemory)
-      return rewriter.notifyMatchFailure(op, "memory pinning not supported");
 
-    SmallVector<Value> sizes, sizeIndex;
-    if (!getListConstructElements(size, sizes)) {
+    Location loc = op.getLoc();
+    TypeConverter *typeConverter = this->getTypeConverter();
+    SmallVector<Value> resultSizeTorchInt, resultSize, resultSizeIndex;
+    if (!getListConstructElements(op.size(), resultSizeTorchInt)) {
       return rewriter.notifyMatchFailure(
-          op, "size must be created by ListConstruct");
+          op, "unimplemented: size must be constructed using ListConstruct");
     }
-    sizes = getTypeConvertedValues(rewriter, loc, getTypeConverter(), sizes);
-    for (size_t i = 0; i < sizes.size(); i++)
-      sizeIndex.push_back(castIntToIndex(rewriter, loc, sizes[i]));
-
-    RankedTensorType newResultType =
-        getTypeConverter()
-            ->convertType(op->getResult(0).getType())
-            .cast<RankedTensorType>();
-    Type outElementType = newResultType.getElementType();
-
-    Value constantOp = rewriter.create<arith::ConstantOp>(
-        loc, outElementType,
-        (outElementType.isa<mlir::FloatType>()
-             ? rewriter.getFloatAttr(outElementType, elementValue)
-                   .cast<mlir::Attribute>()
-             : rewriter.getIntegerAttr(outElementType, elementValue)
-                   .cast<mlir::Attribute>()));
-    Value outTensor = rewriter
-                          .create<linalg::InitTensorOp>(
-                              loc, sizeIndex, newResultType.getElementType())
-                          .getResult();
-    Value fillOp = rewriter.create<linalg::FillOp>(loc, constantOp, outTensor)
-                       .getResult(0);
+    resultSize = getTypeConvertedValues(rewriter, loc, typeConverter,
+                                        resultSizeTorchInt);
+    for (auto size : resultSize)
+      resultSizeIndex.push_back(castIntToIndex(rewriter, loc, size));
 
-    rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, fillOp);
+    auto resultType =
+        typeConverter->convertType(op.getType()).template cast<RankedTensorType>();
+    Type outElemType = resultType.getElementType();
 
+    // Create an uninitialized tensor of `resultSize` shape. It will be returned
+    // without initialization/filling in the case of `AtenEmptyMemoryFormatOp`.
+    Value outputTensor = rewriter.create<linalg::InitTensorOp>(
+        loc, resultSizeIndex, outElemType);
+
+    // `AtenZeros` and `AtenOnes` ops will be filled with corresponding values.
+    if (std::is_same<OpTy, AtenZerosOp>::value) {
+      Value zero = getConstant(rewriter, loc, 0, outElemType);
+      outputTensor =
+          rewriter.create<linalg::FillOp>(loc, zero, outputTensor).getResult(0);
+    } else if (std::is_same<OpTy, AtenOnesOp>::value) {
+      Value one = getConstant(rewriter, loc, 1, outElemType);
+      outputTensor =
+          rewriter.create<linalg::FillOp>(loc, one, outputTensor).getResult(0);
+    }
+    rewriter.replaceOpWithNewOp<tensor::CastOp>(op, resultType, outputTensor);
     return success();
   }
 };
@@ -3704,8 +3705,15 @@ class ConvertTorchToLinalg
     patterns.add<ConvertAtenSizeIntOp>(typeConverter, context);
     target.addIllegalOp<AtenEmbeddingOp>();
     patterns.add<ConvertAtenEmbeddingOp>(typeConverter, context);
-    target.addIllegalOp<AtenOnesOp, AtenZerosOp>();
-    patterns.add<ConvertAtenOnesZerosOp>(typeConverter, context);
+    target.addIllegalOp<AtenEmptyMemoryFormatOp>();
+    patterns.add<ConvertConstantTensorAllocOp<AtenEmptyMemoryFormatOp>>(
+        typeConverter, context);
+    target.addIllegalOp<AtenZerosOp>();
+    patterns.add<ConvertConstantTensorAllocOp<AtenZerosOp>>(typeConverter,
+                                                            context);
+    target.addIllegalOp<AtenOnesOp>();
+    patterns.add<ConvertConstantTensorAllocOp<AtenOnesOp>>(typeConverter,
+                                                           context);
     target.addIllegalOp<AtenContiguousOp>();
     patterns.add<ConvertAtenContiguousOp>(typeConverter, context);
     target.addIllegalOp<AtenIntTensorOp>();