Jacque PR

include/torch-mlir/Conversion/Passes.td

@@ -26,6 +26,11 @@ def ConvertTorchToSCF: Pass<"convert-torch-to-scf", "func::FuncOp"> {
   let constructor = "mlir::torch::createConvertTorchToSCFPass()";
 }
 
+def ConvertTorchToMLProgram : Pass<"convert-torch-to-mlprogram", "ModuleOp"> {
+  let summary = "Convert Torch ops to MLProgram ops";
+  let constructor = "mlir::torch::createConvertTorchToMLProgramPass()";
+}
+
 def ConvertTorchToLinalg : Pass<"convert-torch-to-linalg", "func::FuncOp"> {
   let summary = "Convert recognized Torch ops to Linalg ops";
   let description = [{

include/torch-mlir/Conversion/TorchToMLProgram/TorchToMLProgram.h

@@ -0,0 +1,23 @@
+//===------------------------------------------------------------*- C++ -*-===//
+//
+// 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.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef TORCHMLIR_CONVERSION_TORCHTOMLPROGRAM_TORCHTOMLPROGRAM_H
+#define TORCHMLIR_CONVERSION_TORCHTOMLPROGRAM_TORCHTOMLPROGRAM_H
+
+#include "mlir/IR/BuiltinOps.h"
+#include "mlir/Pass/Pass.h"
+#include <memory>
+
+namespace mlir {
+namespace torch {
+std::unique_ptr<OperationPass<ModuleOp>> createConvertTorchToMLProgramPass();
+}
+} // namespace mlir
+
+#endif // TORCHMLIR_CONVERSION_TORCHTOMLPROGRAM_TORCHTOMLPROGRAM_H

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

@@ -302,6 +302,41 @@ def Torch_InitializeGlobalSlotsOp : Torch_Op<"initialize.global_slots", [
   let hasVerifier = 1;
 }
 
+def Torch_ResourceNonValueTensorLiteralOp : Torch_Op<"tensor.resource.literal", [
+    AllowsTypeRefinement,
+    AllowedInModuleInitializer
+  ]> {
+  let summary = "Create a value of !torch.tensor type from an external literal";
+  let description = [{
+  }];
+  let arguments = (ins
+    FlatSymbolRefAttr:$sym_name,
+    ElementsAttr:$value
+  );
+  let results = (outs Torch_NonValueTensorType:$result);
+
+  let assemblyFormat = [{
+    `(` $sym_name `=` $value `)` attr-dict `:` qualified(type($result))
+  }];
+}
+
+def Torch_ResourceValueTensorLiteralOp : Torch_Op<"vtensor.resource.literal", [
+    AllowsTypeRefinement,
+  ]> {
+  let summary = "Create a value of !torch.vtensor type from an external literal";
+  let description = [{
+  }];
+  let arguments = (ins
+    FlatSymbolRefAttr:$sym_name,
+    ElementsAttr:$value
+  );
+  let results = (outs Torch_ValueTensorType:$result);
+
+  let assemblyFormat = [{
+    `(` $sym_name `=` $value `)` attr-dict `:` qualified(type($result))
+  }];
+}
+
 def Torch_GlobalSlotInitOp : Torch_Op<"global_slot.init", [
     Terminator,
     HasParent<"::mlir::torch::Torch::GlobalSlotOp">]> {

lib/Conversion/CMakeLists.txt

@@ -1,6 +1,7 @@
 add_subdirectory(TorchToLinalg)
 add_subdirectory(TorchToSCF)
 add_subdirectory(TorchToArith)
+add_subdirectory(TorchToMLProgram)
 add_subdirectory(TorchToTosa)
 if(TORCH_MLIR_ENABLE_MHLO)
   add_subdirectory(TorchToMhlo)
@@ -15,6 +16,7 @@ set(linked_libs TorchMLIRTorchToLinalg
                 TorchMLIRTorchToArith
                 TorchMLIRTorchToTosa
                 TorchMLIRTorchToTMTensor
+                TorchMLIRTorchToMLProgram
                 TorchMLIRTorchConversionToMLProgram
                 TorchMLIRConversionUtils)
 if(TORCH_MLIR_ENABLE_MHLO)

lib/Conversion/PassDetail.h

@@ -12,6 +12,7 @@
 
 #include "mlir/Dialect/Func/IR/FuncOps.h"
 #include "mlir/Pass/Pass.h"
+#include "mlir/IR/BuiltinOps.h"
 
 namespace mlir {
 namespace torch {

lib/Conversion/Passes.cpp

@@ -20,7 +20,7 @@
 #include "torch-mlir/Conversion/TorchToMhlo/TorchToMhlo.h"
 #include "torch-mlir/Conversion/TorchToTMTensor/TorchToTMTensor.h"
 #include "torch-mlir/Conversion/TorchConversionToMLProgram/TorchConversionToMLProgram.h"
-
+#include "torch-mlir/Conversion/TorchToMLProgram/TorchToMLProgram.h"
 //===----------------------------------------------------------------------===//
 // Pass registration
 //===----------------------------------------------------------------------===//

lib/Conversion/TorchToMLProgram/CMakeLists.txt

@@ -0,0 +1,21 @@
+add_mlir_conversion_library(TorchMLIRTorchToMLProgram
+  TorchToMLProgram.cpp
+
+  ADDITIONAL_HEADER_DIRS
+  ${PROJECT_SOURCE_DIR}/include/torch-mlir/Conversion/TorchToMLProgram
+
+  DEPENDS
+  TorchMLIRConversionPassIncGen
+
+  LINK_COMPONENTS
+  Core
+
+  LINK_LIBS PUBLIC
+  MLIRIR
+  MLIRPass
+  MLIRFuncDialect
+  MLIRMLProgramDialect
+  TorchMLIRTorchDialect
+)
+
+torch_mlir_target_includes(TorchMLIRTorchToMLProgram)

lib/Conversion/TorchToMLProgram/TorchToMLProgram.cpp

@@ -0,0 +1,64 @@
+//===----------------------------------------------------------------------===//
+//
+// 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.
+//
+//===----------------------------------------------------------------------===//
+
+#include "torch-mlir/Conversion/TorchToMLProgram/TorchToMLProgram.h"
+
+#include "../PassDetail.h"
+#include "mlir/Dialect/MLProgram/IR/MLProgram.h"
+#include "mlir/Dialect/Traits.h"
+#include "mlir/Transforms/DialectConversion.h"
+#include "torch-mlir/Dialect/Torch/IR/TorchOps.h"
+#include "torch-mlir/Dialect/TorchConversion/IR/TorchConversionOps.h"
+#include "torch-mlir/Dialect/TorchConversion/Transforms/BackendTypeConversion.h"
+
+using namespace mlir;
+using namespace mlir::torch;
+
+namespace {
+class ConvertTorchToMLProgram
+    : public ConvertTorchToMLProgramBase<ConvertTorchToMLProgram> {
+public:
+  void getDependentDialects(DialectRegistry &registry) const override {
+    registry.insert<ml_program::MLProgramDialect>();
+    TorchConversion::getBackendTypeConversionDependentDialects(registry);
+  }
+
+  void runOnOperation() override {
+    auto module = getOperation();
+    MLIRContext *context = &getContext();
+    ConversionTarget dummyTarget(*context);
+    TypeConverter typeConverter;
+    typeConverter.addConversion([](Type type) { return type; });
+    TorchConversion::setupBackendTypeConversion(dummyTarget, typeConverter);
+
+    auto globalBuilder =
+        OpBuilder::atBlockBegin(&*module.getBodyRegion().begin());
+    module.walk([&](Torch::ResourceValueTensorLiteralOp op) {
+      auto type = typeConverter.convertType(op.getType());
+      globalBuilder.create<ml_program::GlobalOp>(
+          op.getLoc(), op.getSymNameAttr().getAttr(), type,
+          /*is_mutable=*/true, // Just to enable generator
+          /*value=*/op.getValue(),
+          globalBuilder.getStringAttr("public"));
+      OpBuilder builder(op);
+      auto loadConst = builder.create<ml_program::GlobalLoadOp>(
+          op.getLoc(), type, op.getSymNameAttr());
+      Value torchTensor = builder.create<TorchConversion::FromBuiltinTensorOp>(
+          op.getLoc(), op.getType(), loadConst);
+      op.replaceAllUsesWith(torchTensor);
+      op.erase();
+    });
+  }
+};
+} // namespace
+
+std::unique_ptr<OperationPass<ModuleOp>>
+mlir::torch::createConvertTorchToMLProgramPass() {
+  return std::make_unique<ConvertTorchToMLProgram>();
+}

lib/Dialect/Torch/Transforms/ReduceOpVariants.cpp

@@ -240,6 +240,19 @@ reduceNonValueTensorLiteralOpToValueTensorLiteralOp(NonValueTensorLiteralOp op,
   return success();
 }
 
+static LogicalResult
+reduceResourceNonValueTensorLiteralOpToResourceValueTensorLiteralOp(
+    ResourceNonValueTensorLiteralOp op, PatternRewriter &rewriter) {
+  Value valueTensor = rewriter.create<ResourceValueTensorLiteralOp>(
+      op->getLoc(),
+      op.getType().cast<NonValueTensorType>().getWithValueSemantics(),
+      op.getSymName(), op.getValue());
+  Value tensor =
+      copyTensorToType(rewriter, op->getLoc(), op.getType(), valueTensor);
+  rewriter.replaceOp(op, {tensor});
+  return success();
+}
+
 namespace {
 class ReduceOpVariantsPass : public ReduceOpVariantsBase<ReduceOpVariantsPass> {
   void runOnOperation() override {
@@ -248,10 +261,13 @@ class ReduceOpVariantsPass : public ReduceOpVariantsBase<ReduceOpVariantsPass> {
     patterns.add<ConvertHasValueSemanticsOpsToValueTensors>(context);
     patterns.add<ReduceTrailingUnderscoreInplaceVariant>(context);
     patterns.add(reduceNonValueTensorLiteralOpToValueTensorLiteralOp);
+    patterns.add(
+        reduceResourceNonValueTensorLiteralOpToResourceValueTensorLiteralOp);
     patterns.add<ReduceNonValueSemanticOps>(context);
 
     ConversionTarget target(*context);
     target.addIllegalOp<NonValueTensorLiteralOp>();
+    target.addIllegalOp<ResourceNonValueTensorLiteralOp>();
     target.addIllegalOp<AtenBernoulli_FloatOp>();
     target.markUnknownOpDynamicallyLegal([](Operation *op) {
       if (op->hasTrait<Torch::OpTrait::HasValueSemantics>()) {

lib/Dialect/TorchConversion/Transforms/Passes.cpp

@@ -26,6 +26,7 @@
 #include "torch-mlir/Conversion/TorchToMhlo/TorchToMhlo.h"
 #endif
 #include "torch-mlir/Dialect/Torch/Transforms/Passes.h"
+#include "torch-mlir/Conversion/TorchToMLProgram/TorchToMLProgram.h"
 
 using namespace mlir;
 using namespace mlir::torch;
@@ -64,6 +65,7 @@ void mlir::torch::registerTorchConversionPasses() {
 
 void TorchConversion::createTorchBackendToLinalgOnTensorsBackendPipeline(
     OpPassManager &pm) {
+  pm.addPass(createConvertTorchToMLProgramPass());
   // Lower to linalg + guards which is the input to codegen backends.
   // We do this first as it tends to involve pattern-matching against constants,
   // (e.g. dimensions which must be constant in a ranked programming model)

lib/Dialect/TorchConversion/Transforms/VerifyLinalgOnTensorsBackendContract.cpp

@@ -17,6 +17,7 @@
 #include "mlir/Dialect/MLProgram/IR/MLProgram.h"
 #include "mlir/Dialect/Math/IR/Math.h"
 #include "mlir/Dialect/SCF/IR/SCF.h"
+#include "mlir/Dialect/MLProgram/IR/MLProgram.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/IR/OpDefinition.h"
 #include "mlir/Transforms/DialectConversion.h"
@@ -83,6 +84,8 @@ class VerifyLinalgOnTensorsBackendContractPass
     target.addDynamicallyLegalDialect<scf::SCFDialect>(opHasLegalTypes);
     target.addDynamicallyLegalDialect<ml_program::MLProgramDialect>(
         opHasLegalTypes);
+    target.addDynamicallyLegalDialect<ml_program::MLProgramDialect>(
+        opHasLegalTypes);
 
     // ConstantOp is used for tensors and for scalars.
     target.addDynamicallyLegalOp<arith::ConstantOp>(opHasLegalTypes);

python/torch_mlir/__init__.py

@@ -252,6 +252,7 @@ def compile(model: torch.nn.Module,
             use_tracing: bool = False,
             ignore_traced_shapes=False,
             backend_legal_ops: Optional[Sequence[str]] = None,
+            use_external_references_if_numel_exceeds: Optional[int] = None,
             verbose: bool = False):
     """Convert a PyTorch model to MLIR.
 
@@ -349,6 +350,7 @@ def compile(model: torch.nn.Module,
     mb = ModuleBuilder()
     import_options = ImportOptions()
     import_options.ignoreExistingTensorShapesAndDtypes = ignore_traced_shapes
+    import_options.useExternalReferencesIfNumelExceeds = use_external_references_if_numel_exceeds
     try:
         original_stderr = sys.stderr
         sys.stderr = StringIO()

python/torch_mlir/dialects/torch/importer/jit_ir/csrc/import_options.h

@@ -10,6 +10,8 @@
 #ifndef TORCHMLIRJITIRIMPORTER_CSRC_IMPORT_OPTIONS_H
 #define TORCHMLIRJITIRIMPORTER_CSRC_IMPORT_OPTIONS_H
 
+#include "c10/util/Optional.h"
+
 namespace torch_mlir {
 // Common import options across importers. We define this as a struct to avoid
 // an unstructured proliferation of different kinds of ways to control different
@@ -33,6 +35,9 @@ struct ImportOptions {
   // In that case, the appropriate shape information is provided via the type
   // bound annotations on the function arguments instead.
   bool ignoreExistingTensorShapesAndDtypes = false;
+  // If this is set, then external constant references will be used when
+  // importing tensors with numel exceeding the given threshold.
+  c10::optional<unsigned> useExternalReferencesIfNumelExceeds = c10::nullopt;
 };
 } // namespace torch_mlir
 

python/torch_mlir/dialects/torch/importer/jit_ir/csrc/import_options_pybind.cpp

@@ -20,5 +20,7 @@ void torch_mlir::initImportOptionsBindings(py::module &m) {
       .def_readwrite("assumeTensorsHaveValueSemantics",
                      &ImportOptions::assumeTensorsHaveValueSemantics)
       .def_readwrite("ignoreExistingTensorShapesAndDtypes",
-                     &ImportOptions::ignoreExistingTensorShapesAndDtypes);
+                     &ImportOptions::ignoreExistingTensorShapesAndDtypes)
+      .def_readwrite("useExternalReferencesIfNumelExceeds",
+                     &ImportOptions::useExternalReferencesIfNumelExceeds);
 }

python/torch_mlir/dialects/torch/importer/jit_ir/csrc/ivalue_importer.cpp

@@ -372,16 +372,31 @@ MlirValue IValueImporter::rawImportIValue(c10::IValue ivalue) {
 
 MlirValue IValueImporter::importTensor(c10::IValue ivalue) {
   assert(ivalue.isTensor() && "expected a tensor!");
+  at::Tensor tensor = ivalue.toTensor();
 
   // TODO: Can we do better?
   MlirLocation loc = mlirLocationUnknownGet(context);
 
   // Import the bulk tensor representation.
-  at::Tensor tensor = ivalue.toTensor().contiguous();
   MlirAttribute denseElements = convertTensorToMlirElementsAttr(tensor, loc);
 
+  MlirValue tensorReprValue;
+  if (importOptions.useExternalReferencesIfNumelExceeds.has_value() &&
+      tensor.numel() >
+          importOptions.useExternalReferencesIfNumelExceeds.value()) {
+    MlirType type = torchMlirTorchNonValueTensorTypeGet(
+        context, tensor.sizes().size(), tensor.sizes().data(),
+        getMlirTypeForTorchScalarType(loc, tensor.scalar_type()));
+    MlirAttribute symName = mlirFlatSymbolRefAttrGet(
+        context, toMlirStringRef(
+                      c10::QualifiedName(attributeNameStack).qualifiedName()));
+    MlirOperation tensorOp = createMlirOperationAtEnd(
+        importBlock, "torch.tensor.resource.literal", loc, type,
+        toMlirNamedAttribute("sym_name", symName),
+        toMlirNamedAttribute("value", denseElements));
+    tensorReprValue = mlirOperationGetResult(tensorOp, 0);
+  } else {
     MlirOperation tensorOp;
-
     if (importOptions.assumeTensorsHaveValueSemantics) {
       tensorOp = createMlirOperationAtEnd(
           importBlock, "torch.vtensor.literal", loc,
@@ -393,9 +408,8 @@ MlirValue IValueImporter::importTensor(c10::IValue ivalue) {
           torchMlirTorchNonValueTensorTypeGetFromAttribute(denseElements),
           toMlirNamedAttribute("value", denseElements));
     }
-
-  MlirValue tensorReprValue = mlirOperationGetResult(tensorOp, 0);
-
+    tensorReprValue = mlirOperationGetResult(tensorOp, 0);
+  }
   // Construct the complete tensor value. This is trivial for most tensors, but
   // for quantized tensors (and probably sparse too, TBD) there is more for us
   // to do.