Reference Lazy Backend (#1045)

* Changed Example MLIR backend to Reference MLIR backend

* Moved reference_ltc_backend into csrc

* Merged sys_utils.h

* Renamed reference_ltc_backend to reference_lazy_backend

* Addressed review comments

* Update docs with new library name

* Removed _REFERENCE_LAZY_BACKEND from .gitignore

* Added reference_lazy_backend to the TorchMLIRPythonModules dependency list

Fixed typo in `ltc_examples.md`

Missed instance where `ltc_backend` was used instead of `lazy_backend`.

.gitignore

@@ -26,6 +26,3 @@ bazel-*
 
 # Autogenerated files
 /python/torch_mlir/csrc/base_lazy_backend/generated
-
-# Example backend
-examples/ltc_backend/ltc_backend/_EXAMPLE_MLIR_BACKEND.cpython-37m-x86_64-linux-gnu.so

CMakeLists.txt

@@ -192,4 +192,3 @@ else()
 endif()
 
 add_subdirectory(test)
-add_subdirectory(examples)

build_tools/autogen_ltc_backend.py

@@ -377,7 +377,7 @@ class GenTorchMlirLTC:
                     // for ops that dont have a corresponding structured kernel or shape definition
 
                     #include "shape_inference.h"
-                    #include "../../utils/exception.h"
+                    #include "../utils/exception.h"
                     namespace torch {{
                     namespace lazy {{
                     {}

docs/ltc_backend.md

@@ -60,12 +60,15 @@ Generated files are created in this directory, which is ignored by version contr
 - `shape_inference.cpp`
   - Implementation of select shape inference functions (most functions are [implemented upstream](https://github.com/pytorch/pytorch/blob/master/torch/csrc/lazy/core/shape_inference.cpp))
 
+### Reference Backend ([`python/torch_mlir/csrc/reference_lazy_backend`](../python/torch_mlir/csrc/reference_lazy_backend))
+
+- `backend_impl.{cpp,h}`
+  - Reference Torch-MLIR LTC backend implementation, which simply stores the MLIR as a string and executes computation on CPU
+- `reference_lazy_backend_pybind.cpp`
+  - pybind for reference Torch-MLIR LTC backend
+
 ### Examples ([`examples`](../examples))
 
-- `examples/ltc_backend/ltc_backend/csrc/backend/backend_impl.{cpp,h}`
-  - Example Torch-MLIR LTC backend implementation, which simply stores the MLIR as a string and executes computation on CPU
-- `examples/ltc_backend/ltc_backend/csrc/example_mlir_backend_pybind.cpp`
-  - pybind for example Torch-MLIR LTC backend
 - `ltc_backend_bert.py`
   - Example HuggingFace BERT model traced by LTC to MLIR
 - `ltc_backend_mnist.py`
@@ -77,7 +80,7 @@ Generated files are created in this directory, which is ignored by version contr
 
 The journey begins with a tensor in PyTorch on the `lazy` device, which may undergo a number of operations during its lifetime.
 ```python
->>> ltc_backend._initialize()
+>>> lazy_backend._initialize()
 >>> x = torch.tensor(..., device='lazy')
 >>> y = torch.tanh(x)
 ...
@@ -116,17 +119,17 @@ Finally, the compiled computation is sent to `TorchMlirBackendImpl::ExecuteCompu
 
 ## Implementing a custom backend
 
-An example implementation of a custom backend is available [here](../examples/ltc_backend/ltc_backend). 
+A reference implementation of a custom backend is available [here](../python/torch_mlir/csrc/reference_lazy_backend/). 
 All the work involved with generating MLIR is handled in the base LTC backend, so vendors only need to worry about implementing `Compile`, `ExecuteComputation`, and some other minor methods to interface with the device.
 
 A pybind is needed to invoke C++ code to register the autogen PyTorch kernels and the custom backend itself.
-Most of the code in the example implementation should be reusable, excluding some debug related function (e.g. `get_latest_computation`).
+Most of the code in the reference implementation should be reusable, excluding some debug related function (e.g. `get_latest_computation`).
 
 ## Future Expansion
 
 There are a number of areas for future improvement:
 - Generate source information in `jit::Graph` so it can be embedded in the MLIR
-- Currently the example backend implementation executes via the `jit::Graph` instead of the MLIR since we currently lack lowerings for many ops, which would make it difficult to run models such as HF BERT
+- Currently the reference backend implementation executes via the `jit::Graph` instead of the MLIR since we currently lack lowerings for many ops, which would make it difficult to run models such as HF BERT
   - In the future, we should change the implementation to lower the MLIR to linalg and execute on a reference backend
 - As new models get tested, we will inevitably run into errors related to unimplemented shape inference functions.
 This problem is simply solved by implementing the missing function, or adding a structured kernel to PyTorch.

docs/ltc_examples.md

@@ -6,10 +6,10 @@ Refer to the main documentation [here](ltc_backend.md).
 ```python
 import torch
 import torch._lazy
-import ltc_backend.ltc_backend._EXAMPLE_MLIR_BACKEND as ltc_backend
+import torch_mlir.reference_lazy_backend._REFERENCE_LAZY_BACKEND as lazy_backend
 
 # Register the example LTC backend.
-ltc_backend._initialize()
+lazy_backend._initialize()
 
 device = 'lazy'
 
@@ -22,7 +22,7 @@ torch._lazy.mark_step()
 print('Results:', outputs)
 
 # Optionally dump MLIR graph generated from LTC trace.
-computation = ltc_backend.get_latest_computation()
+computation = lazy_backend.get_latest_computation()
 if computation:
     print(computation.debug_string())
 ```

examples/CMakeLists.txt

@@ -1 +0,0 @@
-add_subdirectory(ltc_backend)

examples/ltc_backend/ltc_backend/csrc/utils/sys_utils.h

@@ -1,26 +0,0 @@
-//===- sys_utils.h --------------------------------------------------------===//
-//
-// 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.
-//
-//===----------------------------------------------------------------------===//
-
-#pragma once
-
-#include <cstdlib>
-#include <string>
-
-namespace sys_util {
-
-template <typename T>
-T GetEnv(const std::string &name, const T &default_value = T(0)) {
-  const char *env = std::getenv(name.c_str());
-  if (!env) {
-    return default_value;
-  }
-  return T(std::atoi(env));
-}
-
-} // namespace sys_util

examples/ltc_backend_bert.py

@@ -113,8 +113,8 @@ def main(device='lazy', full_size=False):
     losses = train(model, num_epochs, num_training_steps, train_dataloader, device)
 
     # Get debug information from LTC
-    if 'ltc_backend' in sys.modules:
-        computation = ltc_backend.get_latest_computation()
+    if 'torch_mlir.reference_lazy_backend._REFERENCE_LAZY_BACKEND' in sys.modules:
+        computation = lazy_backend.get_latest_computation()
         if computation:
             print(computation.debug_string())
 
@@ -148,9 +148,9 @@ if __name__ == "__main__":
             torch._lazy.ts_backend.init()
 
         elif args.device == "MLIR_EXAMPLE":
-            import ltc_backend.ltc_backend._EXAMPLE_MLIR_BACKEND as ltc_backend
+            import torch_mlir.reference_lazy_backend._REFERENCE_LAZY_BACKEND as lazy_backend
 
-            ltc_backend._initialize()
+            lazy_backend._initialize()
 
         device = "lazy"
         print("Initialized backend")

examples/ltc_backend_mnist.py

@@ -65,8 +65,8 @@ def main(device='lazy'):
             torch._lazy.mark_step()
 
     # Get debug information from LTC
-    if 'ltc_backend' in sys.modules:
-        computation = ltc_backend.get_latest_computation()
+    if 'torch_mlir.reference_lazy_backend._REFERENCE_LAZY_BACKEND' in sys.modules:
+        computation = lazy_backend.get_latest_computation()
         if computation:
             print(computation.debug_string())
 
@@ -93,9 +93,9 @@ if __name__ == "__main__":
             torch._lazy.ts_backend.init()
 
         elif args.device == "MLIR_EXAMPLE":
-            import ltc_backend.ltc_backend._EXAMPLE_MLIR_BACKEND as ltc_backend
+            import torch_mlir.reference_lazy_backend._REFERENCE_LAZY_BACKEND as lazy_backend
 
-            ltc_backend._initialize()
+            lazy_backend._initialize()
 
         device = "lazy"
         print("Initialized backend")

python/CMakeLists.txt

@@ -60,7 +60,8 @@ declare_mlir_python_extension(TorchMLIRPythonExtensions.Main
 # Lazy Tensor Core
 ################################################################################
 
-add_subdirectory(torch_mlir/csrc)
+add_subdirectory(torch_mlir/csrc/base_lazy_backend)
+add_subdirectory(torch_mlir/csrc/reference_lazy_backend)
 
 ################################################################################
 # Optionally handle JIT IR importer.
@@ -155,6 +156,6 @@ endif()
 
 # Add Torch-MLIR LTC backend as dependency
 add_dependencies(TorchMLIRPythonModules torch_mlir_ltc_backend)
+add_dependencies(TorchMLIRPythonModules reference_lazy_backend)
 
 add_subdirectory(test)
-

python/torch_mlir/csrc/base_lazy_backend/CMakeLists.txt

@@ -20,15 +20,15 @@ include_directories(BEFORE
 link_directories("${TORCH_INSTALL_PREFIX}/lib")
 
 set(LTC_GENERATED
-  base_lazy_backend/generated/LazyNativeFunctions.cpp
-  base_lazy_backend/generated/RegisterLazy.cpp
-  base_lazy_backend/generated/shape_inference.cpp
+  generated/LazyNativeFunctions.cpp
+  generated/RegisterLazy.cpp
+  generated/shape_inference.cpp
 )
 set(LTC_BACKEND_DEPENDS
-  base_lazy_backend/mlir_lowering_context.cpp
-  base_lazy_backend/mlir_native_functions.cpp
-  base_lazy_backend/mlir_node_lowering.cpp
-  base_lazy_backend/shape_inference.cpp
+  mlir_lowering_context.cpp
+  mlir_native_functions.cpp
+  mlir_node_lowering.cpp
+  shape_inference.cpp
 )
 
 # Generate Lazy IR Nodes
@@ -57,10 +57,10 @@ add_custom_target(
 add_library(torch_mlir_ltc_backend SHARED
   ${LTC_GENERATED}
   ${LTC_BACKEND_DEPENDS}
-  base_lazy_backend/backend_impl.cpp
-  base_lazy_backend/mlir_node.cpp
-  base_lazy_backend/ops/device_data.cpp
-  base_lazy_backend/ops/generic.cpp
+  backend_impl.cpp
+  mlir_node.cpp
+  ops/device_data.cpp
+  ops/generic.cpp
 )
 target_compile_features(torch_mlir_ltc_backend PRIVATE cxx_std_17)
 

python/torch_mlir/csrc/base_lazy_backend/backend_impl.cpp

@@ -15,12 +15,12 @@
 #include <torch/csrc/lazy/backend/lowering_context.h>
 #include <torch/csrc/lazy/core/shape.h>
 
-#include "../utils/debug.h"
-#include "../utils/exception.h"
 #include "backend_impl.h"
 #include "ir_builder.h"
 #include "mlir_lowering_context.h"
 #include "ops/device_data.h"
+#include "utils/debug.h"
+#include "utils/exception.h"
 
 namespace torch {
 namespace lazy {

python/torch_mlir/csrc/base_lazy_backend/ir_builder.h

@@ -22,7 +22,7 @@
 #include "mlir_node.h"
 #include "ops/device_data.h"
 #include "ops/generic.h"
-#include "../utils/exception.h"
+#include "utils/exception.h"
 
 // This file contains the TorchMlir IrBuilder
 

python/torch_mlir/csrc/base_lazy_backend/mlir_lowering_context.cpp

@@ -17,13 +17,13 @@
 #include <torch/csrc/lazy/core/lazy_graph_executor.h>
 
 #include "../../dialects/torch/importer/jit_ir/csrc/function_importer.h"
-#include "../utils/debug.h"
-#include "../utils/exception.h"
 #include "backend_impl.h"
 #include "mlir-c/Registration.h"
 #include "mlir_lowering_context.h"
 #include "mlir_node.h"
 #include "torch-mlir-c/Registration.h"
+#include "utils/debug.h"
+#include "utils/exception.h"
 
 namespace torch {
 namespace lazy {

python/torch_mlir/csrc/base_lazy_backend/mlir_native_functions.cpp

@@ -10,11 +10,11 @@
 // https://github.com/pytorch/pytorch/blob/master/torch/csrc/lazy/ts_backend/ts_native_functions.cpp
 //===----------------------------------------------------------------------===//
 
-#include <ATen/InferSize.h>
-#include <ATen/Operators.h>
 #include <ATen/FunctionalTensorWrapper.h>
+#include <ATen/InferSize.h>
 #include <ATen/MetaFunctions.h>
 #include <ATen/NativeFunctions.h>
+#include <ATen/Operators.h>
 #include <ATen/native/BinaryOps.h>
 #include <ATen/native/CPUFallback.h>
 #include <ATen/ops/empty.h>
@@ -28,12 +28,11 @@
 #include <torch/csrc/lazy/core/tensor_util.h>
 #include <torch/library.h>
 
-
-#include "../utils/exception.h"
-#include "../utils/sys_utils.h"
-#include "generated/shape_inference.h"
 #include "generated/LazyNativeFunctions.h"
+#include "generated/shape_inference.h"
 #include "ops/to_copy.h"
+#include "utils/exception.h"
+#include "utils/sys_utils.h"
 
 namespace torch {
 namespace lazy {
@@ -174,7 +173,6 @@ at::Tensor LazyNativeFunctions::cat(at::TensorList tensors, int64_t dim) {
   // return result;
 }
 
-
 // clone is special in LT because we make it a no-op.
 // This should be safe to do, because every operator in the LT is functional.
 at::Tensor LazyNativeFunctions::clone(
@@ -294,8 +292,12 @@ at::Tensor LazyNativeFunctions::_to_copy(
     // See Note [Lazy Tensor Functionalization]
     // Invariant: if the functionalization key is in the exclude set, then we're expected
     // to return an ordinary tensor, which will be "lifted" into a functional wrapper later.
-      bool functionalize_output = !c10::impl::tls_local_dispatch_key_set().excluded_.has(c10::DispatchKey::Functionalize);
-      return torch::lazy::to_lazy_tensor(self, options, *device, /*non_blocking=*/non_blocking, /*functionalize_output=*/functionalize_output);
+    bool functionalize_output =
+        !c10::impl::tls_local_dispatch_key_set().excluded_.has(
+            c10::DispatchKey::Functionalize);
+    return torch::lazy::to_lazy_tensor(
+        self, options, *device, /*non_blocking=*/non_blocking,
+        /*functionalize_output=*/functionalize_output);
   } else if (device && device->type() != c10::kLazy) {
     // Case 2: lazy->eager (forces a graph break since we are materializing a tensor)
 
@@ -368,7 +370,8 @@ at::Tensor LazyNativeFunctions::empty(
   auto x_result = at::empty(size, options, memory_format);
   auto tensor = CreateLtcTensor(x_result, GetLtcDevice(device));
   // See Note [Lazy Tensor Functionalization]
-  if (c10::impl::tls_local_dispatch_key_set().excluded_.has(c10::DispatchKey::Functionalize)) {
+  if (c10::impl::tls_local_dispatch_key_set().excluded_.has(
+          c10::DispatchKey::Functionalize)) {
     // Invariant: if the functionalization key is in the exclude set, then we're expected
     // to return an ordinary tensor, which will be "lifted" into a functional wrapper later.
     return tensor;
@@ -409,7 +412,8 @@ at::Tensor LazyNativeFunctions::_unsafe_view(
 // LazyTensor always opts into functionalization.
 // "lifting" a tensor for functionalization means wrapping it in a FunctionalTensorWrapper object.
 at::Tensor LazyNativeFunctions::lift(const at::Tensor& tensor) {
-  TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(tensor));
+  TORCH_INTERNAL_ASSERT(
+      !at::functionalization::impl::isFunctionalTensor(tensor));
   return at::functionalization::impl::to_functional_tensor(tensor);
 }
 
@@ -418,43 +422,75 @@ at::Tensor LazyNativeFunctions::lift(const at::Tensor& tensor) {
 // These are all composite ops that LTC can technically re-use / get for free,
 // but we need to "functionalize" them to remove the view ops before we can use them.
 at::Tensor LazyNativeFunctions::block_diag(at::TensorList tensors) {
-  return at::functionalization::functionalize_aten_op<ATEN_OP(block_diag)>::call(tensors);
+  return at::functionalization::functionalize_aten_op<ATEN_OP(
+      block_diag)>::call(tensors);
 }
-at::Tensor LazyNativeFunctions::new_empty_strided(const at::Tensor& self, at::IntArrayRef size, at::IntArrayRef stride, c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout, c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
-  return at::functionalization::functionalize_aten_op<ATEN_OP(new_empty_strided)>::call(self, size, stride, dtype, layout, device, pin_memory);
+at::Tensor LazyNativeFunctions::new_empty_strided(
+    const at::Tensor& self, at::IntArrayRef size, at::IntArrayRef stride,
+    c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout,
+    c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
+  return at::functionalization::
+      functionalize_aten_op<ATEN_OP(new_empty_strided)>::call(
+          self, size, stride, dtype, layout, device, pin_memory);
 }
 
-at::Tensor LazyNativeFunctions::narrow_copy(const at::Tensor& self, int64_t dim, int64_t start, int64_t length) {
-  return at::functionalization::functionalize_aten_op<ATEN_OP(narrow_copy)>::call(self, dim, start, length);
+at::Tensor LazyNativeFunctions::narrow_copy(
+    const at::Tensor& self, int64_t dim, int64_t start, int64_t length) {
+  return at::functionalization::functionalize_aten_op<ATEN_OP(
+      narrow_copy)>::call(self, dim, start, length);
 }
-at::Tensor LazyNativeFunctions::pixel_shuffle(const at::Tensor & self, int64_t upscale_factor) {
-  return at::functionalization::functionalize_aten_op<ATEN_OP(pixel_shuffle)>::call(self, upscale_factor);
+at::Tensor LazyNativeFunctions::pixel_shuffle(
+    const at::Tensor& self, int64_t upscale_factor) {
+  return at::functionalization::functionalize_aten_op<ATEN_OP(
+      pixel_shuffle)>::call(self, upscale_factor);
 }
-at::Tensor LazyNativeFunctions::pixel_unshuffle(const at::Tensor & self, int64_t downscale_factor) {
-  return at::functionalization::functionalize_aten_op<ATEN_OP(pixel_unshuffle)>::call(self, downscale_factor);
+at::Tensor LazyNativeFunctions::pixel_unshuffle(
+    const at::Tensor& self, int64_t downscale_factor) {
+  return at::functionalization::functionalize_aten_op<ATEN_OP(
+      pixel_unshuffle)>::call(self, downscale_factor);
 }
-at::Tensor LazyNativeFunctions::select_backward(const at::Tensor & grad_output, at::IntArrayRef input_sizes, int64_t dim, int64_t index) {
-  return at::functionalization::functionalize_aten_op<ATEN_OP(select_backward)>::call(grad_output, input_sizes, dim, index);
+at::Tensor LazyNativeFunctions::select_backward(
+    const at::Tensor& grad_output, at::IntArrayRef input_sizes, int64_t dim,
+    int64_t index) {
+  return at::functionalization::functionalize_aten_op<ATEN_OP(
+      select_backward)>::call(grad_output, input_sizes, dim, index);
 }
-at::Tensor LazyNativeFunctions::slice_backward(const at::Tensor & grad_output, at::IntArrayRef input_sizes, int64_t dim, int64_t start, int64_t end, int64_t step) {
-  return at::functionalization::functionalize_aten_op<ATEN_OP(slice_backward)>::call(grad_output, input_sizes, dim, start, end, step);
+at::Tensor LazyNativeFunctions::slice_backward(
+    const at::Tensor& grad_output, at::IntArrayRef input_sizes, int64_t dim,
+    int64_t start, int64_t end, int64_t step) {
+  return at::functionalization::functionalize_aten_op<ATEN_OP(
+      slice_backward)>::call(grad_output, input_sizes, dim, start, end, step);
 }
-at::Tensor LazyNativeFunctions::diagonal_backward(const at::Tensor & grad_output, at::IntArrayRef input_sizes, int64_t offset, int64_t dim1, int64_t dim2) {
-  return at::functionalization::functionalize_aten_op<ATEN_OP(diagonal_backward)>::call(grad_output, input_sizes, offset, dim1, dim2);
+at::Tensor LazyNativeFunctions::diagonal_backward(
+    const at::Tensor& grad_output, at::IntArrayRef input_sizes, int64_t offset,
+    int64_t dim1, int64_t dim2) {
+  return at::functionalization::functionalize_aten_op<ATEN_OP(
+      diagonal_backward)>::call(grad_output, input_sizes, offset, dim1, dim2);
 }
-at::Tensor LazyNativeFunctions::_trilinear(const at::Tensor & i1, const at::Tensor & i2, const at::Tensor & i3, at::IntArrayRef expand1, at::IntArrayRef expand2, at::IntArrayRef expand3, at::IntArrayRef sumdim, int64_t unroll_dim) {
-  return at::functionalization::functionalize_aten_op<ATEN_OP(_trilinear)>::call(i1, i2, i3, expand1, expand2, expand3, sumdim, unroll_dim);
+at::Tensor LazyNativeFunctions::_trilinear(
+    const at::Tensor& i1, const at::Tensor& i2, const at::Tensor& i3,
+    at::IntArrayRef expand1, at::IntArrayRef expand2, at::IntArrayRef expand3,
+    at::IntArrayRef sumdim, int64_t unroll_dim) {
+  return at::functionalization::functionalize_aten_op<ATEN_OP(_trilinear)>::
+      call(i1, i2, i3, expand1, expand2, expand3, sumdim, unroll_dim);
 }
-::std::tuple<at::Tensor,at::Tensor> LazyNativeFunctions::linalg_inv_ex(const at::Tensor & self, bool check_errors) {
-  return at::functionalization::functionalize_aten_op<ATEN_OP(linalg_inv_ex)>::call(self, check_errors);
+::std::tuple<at::Tensor, at::Tensor>
+LazyNativeFunctions::linalg_inv_ex(const at::Tensor& self, bool check_errors) {
+  return at::functionalization::functionalize_aten_op<ATEN_OP(
+      linalg_inv_ex)>::call(self, check_errors);
 }
-at::Tensor LazyNativeFunctions::linalg_pinv(const at::Tensor & self, const c10::optional<at::Tensor> & atol, const c10::optional<at::Tensor> & rtol, bool hermitian) {
-  return at::functionalization::functionalize_aten_op<ATEN_OP2(linalg_pinv, atol_rtol_tensor)>::call(self, atol, rtol, hermitian);
+at::Tensor LazyNativeFunctions::linalg_pinv(
+    const at::Tensor& self, const c10::optional<at::Tensor>& atol,
+    const c10::optional<at::Tensor>& rtol, bool hermitian) {
+  return at::functionalization::functionalize_aten_op<ATEN_OP2(
+      linalg_pinv, atol_rtol_tensor)>::call(self, atol, rtol, hermitian);
 }
 
 // functionalize_aten_op can't handle out= ops directly.
 // Instead, we can call the composite kernel from core, and copy and mutations back to the inputs.
-at::Tensor & LazyNativeFunctions::logsumexp_out(const at::Tensor & self, at::IntArrayRef dim, bool keepdim, at::Tensor& out) {
+at::Tensor& LazyNativeFunctions::logsumexp_out(
+    const at::Tensor& self, at::IntArrayRef dim, bool keepdim,
+    at::Tensor& out) {
   auto self_wrapped = at::functionalization::impl::to_functional_tensor(self);
   auto out_wrapped = at::functionalization::impl::to_functional_tensor(out);
   // directly call the composite kernel from core.

python/torch_mlir/csrc/base_lazy_backend/mlir_node.cpp

@@ -11,7 +11,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "mlir_node.h"
-#include "../utils/exception.h"
+#include "utils/exception.h"
 
 namespace torch {
 namespace lazy {
@@ -74,7 +74,8 @@ hash_t TorchMlirNode::shapeHash() const { return shape_hash_; }
 OpKind TorchMlirTensorList::ClassOpKind() {
   // Note: this OpKind is separate from ltc_ops.h since it would be a circular
   // import otherwise
-  static const OpKind tensor_list_opkind = OpKind::Get("lazy_tensors::tensor_list");
+  static const OpKind tensor_list_opkind =
+      OpKind::Get("lazy_tensors::tensor_list");
   return tensor_list_opkind;
 }
 

python/torch_mlir/csrc/base_lazy_backend/mlir_node.h

@@ -18,9 +18,9 @@
 #include <torch/csrc/lazy/core/ir.h>
 #include <torch/csrc/lazy/core/shape.h>
 
-#include "../utils/debug.h"
-#include "../utils/exception.h"
 #include "mlir_lowering_context.h"
+#include "utils/debug.h"
+#include "utils/exception.h"
 
 namespace torch {
 namespace lazy {
@@ -60,7 +60,6 @@ private:
   hash_t dag_hash_;
 };
 
-
 // TensorList represents an at::TensorList which is a vector[Tensor] but is also
 // a first-class IValue and can be fed as a single input to a TS program.  It is
 // much easier to handle TensorLists in Lazy Tensor code if they are represented

python/torch_mlir/csrc/base_lazy_backend/mlir_node_lowering.cpp

@@ -209,17 +209,17 @@ GenerateClone(torch::jit::Value* val, TorchMlirFunction function) {
   return cloned.front();
 }
 
-
-void GenerateCopy(torch::jit::Value* destination, torch::jit::Value* source, TorchMlirFunction function) {
+void GenerateCopy(
+    torch::jit::Value* destination, torch::jit::Value* source,
+    TorchMlirFunction function) {
   std::vector<torch::jit::NamedValue> arguments;
   arguments.emplace_back(destination);
   arguments.emplace_back(source);
   LowerBuiltin(
-        at::aten::copy_,
-        c10::ArrayRef<Shape>(compute_shape_copy(source->type())), function, arguments);
+      at::aten::copy_, c10::ArrayRef<Shape>(compute_shape_copy(source->type())),
+      function, arguments);
 }
 
-
 torch::jit::Value* GenerateSlice(
     torch::jit::Value* base, int64_t dim, int64_t start, int64_t end,
     int64_t step, TorchMlirFunction function) {
@@ -234,8 +234,7 @@ torch::jit::Value* GenerateSlice(
       at::aten::slice,
       c10::ArrayRef<Shape>(
           compute_shape_slice(base->type(), dim, start, end, step)),
-      function,
-      arguments);
+      function, arguments);
   CHECK_EQ(selected.size(), 1);
   return selected.front();
 }

python/torch_mlir/csrc/base_lazy_backend/shape_inference.cpp

@@ -11,8 +11,8 @@
 #include <c10/util/Optional.h>
 #include <cmath>
 
-#include "../utils/exception.h"
 #include "generated/shape_inference.h"
+#include "utils/exception.h"
 
 namespace torch {
 namespace lazy {

python/torch_mlir/csrc/base_lazy_backend/utils/sys_utils.h

@@ -5,6 +5,15 @@
 
 namespace sys_util {
 
+template <typename T>
+static T GetEnv(const std::string& name, const T& default_value = T(0)) {
+  const char* env = std::getenv(name.c_str());
+  if (!env) {
+    return default_value;
+  }
+  return T(std::atoi(env));
+}
+
 static bool GetEnvBool(const char* name, bool defval) {
   const char* env = std::getenv(name);
   if (env == nullptr) {

python/torch_mlir/csrc/reference_lazy_backend/CMakeLists.txt

@@ -30,22 +30,18 @@ include_directories(BEFORE
         ${PROJECT_SOURCE_DIR}/python
         )
 link_directories("${TORCH_INSTALL_PREFIX}/lib")
-link_directories(${CMAKE_CURRENT_SOURCE_DIR}/ltc_backend/lib)
-add_link_options(-Wl,-rpath,$ORIGIN/ltc_backend/lib)
+link_directories(${CMAKE_CURRENT_SOURCE_DIR}/lib)
+add_link_options(-Wl,-rpath,$ORIGIN/lib)
 
-file(GLOB LTC_BACKEND_CSRC CONFIGURE_DEPENDS
-        "ltc_backend/csrc/*.h"
-        "ltc_backend/csrc/*.cc"
-        "ltc_backend/csrc/*.cpp"
-        "ltc_backend/csrc/*/*.h"
-        "ltc_backend/csrc/*/*.cc"
-        "ltc_backend/csrc/*/*.cpp"
+set(REFERENCE_LAZY_BACKEND_CSRC
+        backend_impl.cpp
+        reference_lazy_backend_pybind.cpp
         )
-add_library(example_mlir_ltc_backend SHARED ${LTC_BACKEND_CSRC})
-add_dependencies(example_mlir_ltc_backend
+add_library(reference_lazy_backend SHARED ${REFERENCE_LAZY_BACKEND_CSRC})
+add_dependencies(reference_lazy_backend
         torch_mlir_ltc_backend
         )
-target_link_libraries(example_mlir_ltc_backend
+target_link_libraries(reference_lazy_backend
         ${TORCH_LIBRARIES}
         ${Python3_LIBRARIES}
         torch_python
@@ -53,9 +49,9 @@ target_link_libraries(example_mlir_ltc_backend
         )
 
 message(STATUS "TORCH_CXXFLAGS=${TORCH_CXXFLAGS} -Wno-pedantic")
-set_target_properties(example_mlir_ltc_backend PROPERTIES
-        LIBRARY_OUTPUT_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/ltc_backend/"
-        OUTPUT_NAME _EXAMPLE_MLIR_BACKEND
+set_target_properties(reference_lazy_backend PROPERTIES
+        LIBRARY_OUTPUT_DIRECTORY "${TORCH_MLIR_PYTHON_PACKAGES_DIR}/torch_mlir/torch_mlir/reference_lazy_backend"
+        OUTPUT_NAME _REFERENCE_LAZY_BACKEND
         PREFIX "${PYTHON_MODULE_PREFIX}"
         SUFFIX "${PYTHON_MODULE_EXTENSION}"
         CXX_VISIBILITY_PRESET "hidden"

python/torch_mlir/csrc/reference_lazy_backend/backend_impl.cpp

@@ -15,8 +15,8 @@
 #include <torch_mlir/csrc/base_lazy_backend/backend_impl.h>
 #include <torch_mlir/csrc/base_lazy_backend/generated/LazyNativeFunctions.h>
 #include <torch_mlir/csrc/base_lazy_backend/mlir_lowering_context.h>
-#include <torch_mlir/csrc/utils/debug.h>
-#include <torch_mlir/csrc/utils/exception.h>
+#include <torch_mlir/csrc/base_lazy_backend/utils/debug.h>
+#include <torch_mlir/csrc/base_lazy_backend/utils/exception.h>
 
 #include "backend_impl.h"
 
@@ -25,8 +25,8 @@ using namespace torch::lazy;
 namespace torch {
 namespace lazy {
 
-struct ExampleMlirBackendDeviceType : public BackendDeviceType {
-  ExampleMlirBackendDeviceType(std::string device_type)
+struct ReferenceLazyBackendDeviceType : public BackendDeviceType {
+  ReferenceLazyBackendDeviceType(std::string device_type)
       : device_type_(device_type) {}
 
   std::string toString() const override { return device_type_; }
@@ -34,9 +34,9 @@ struct ExampleMlirBackendDeviceType : public BackendDeviceType {
   std::string device_type_;
 };
 
-class ExampleMlirBackendImpl : public torch::lazy::TorchMlirBackendImpl {
+class ReferenceLazyBackendImpl : public torch::lazy::TorchMlirBackendImpl {
 public:
-  ExampleMlirBackendImpl() : default_device_type_("Magic") {}
+  ReferenceLazyBackendImpl() : default_device_type_("Magic") {}
 
   /**
    * Configuration
@@ -48,8 +48,8 @@ public:
   /**
    * Lowering, Compilation, Execution
    * */
-  std::vector<std::string>
-  GetCompilationDevices(const std::string &device,
+  std::vector<std::string> GetCompilationDevices(
+      const std::string& device,
       c10::ArrayRef<std::string> devices) const override {
     return std::vector<std::string>(devices.begin(), devices.end());
   };
@@ -70,8 +70,8 @@ public:
     return instances;
   }
 
-  std::vector<BackendDataPtr>
-  ExecuteComputation(torch::lazy::ComputationPtr computation,
+  std::vector<BackendDataPtr> ExecuteComputation(
+      torch::lazy::ComputationPtr computation,
       c10::ArrayRef<BackendDataPtr> arguments,
       const BackendDevice& device) const override {
     PRINT_FUNCTION();
@@ -80,7 +80,8 @@ public:
     // function, we will generate a list of BackendData that corresponds to the
     // return values in the MLIR.
 
-    auto mlir_computation = static_cast<TorchMlirComputation *>(computation.get());
+    auto mlir_computation =
+        static_cast<TorchMlirComputation*>(computation.get());
 
     // Vendor backend specific execution can be inserted here.
     //
@@ -128,7 +129,7 @@ public:
   }
 
   void SetDefaultDeviceType(std::string device_type) {
-    default_device_type_ = ExampleMlirBackendDeviceType(device_type);
+    default_device_type_ = ReferenceLazyBackendDeviceType(device_type);
   }
 
   /**
@@ -146,19 +147,19 @@ public:
   }
 
 private:
-  ExampleMlirBackendDeviceType default_device_type_;
+  ReferenceLazyBackendDeviceType default_device_type_;
 };
 
-BackendImplInterface *GetExampleMlirBackendImpl() {
-  static ExampleMlirBackendImpl *example_mlir_backend_impl =
-      new ExampleMlirBackendImpl();
-  return example_mlir_backend_impl;
+BackendImplInterface* GetReferenceLazyBackendImpl() {
+  static ReferenceLazyBackendImpl* reference_lazy_backend_impl =
+      new ReferenceLazyBackendImpl();
+  return reference_lazy_backend_impl;
 }
 
-void InitExampleMlirBackend() {
+void InitReferenceLazyBackend() {
   at::RegisterTorchMlirLazyNativeFunctions();
   static std::unique_ptr<BackendRegistrar> g_registrar;
-  g_registrar.reset(new BackendRegistrar(GetExampleMlirBackendImpl()));
+  g_registrar.reset(new BackendRegistrar(GetReferenceLazyBackendImpl()));
 }
 
 ComputationPtr& GetLatestComputation() {

python/torch_mlir/csrc/reference_lazy_backend/backend_impl.h

@@ -19,9 +19,9 @@ TORCH_API void RegisterTorchMlirLazyNativeFunctions();
 namespace torch {
 namespace lazy {
 
-torch::lazy::BackendImplInterface *GetExampleMlirBackendImpl();
+torch::lazy::BackendImplInterface* GetReferenceLazyBackendImpl();
 
-void InitExampleMlirBackend();
+void InitReferenceLazyBackend();
 
 ComputationPtr& GetLatestComputation();
 

python/torch_mlir/csrc/reference_lazy_backend/reference_lazy_backend_pybind.cpp

@@ -1,4 +1,4 @@
-//===- example_mlir_backend_pybind.cpp ------------------------------------===//
+//===- reference_lazy_backend_pybind.cpp ----------------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
@@ -11,13 +11,13 @@
 #include "torch/csrc/lazy/backend/backend_interface.h"
 
 #include <torch_mlir/csrc/base_lazy_backend/mlir_lowering_context.h>
+#include <torch_mlir/csrc/base_lazy_backend/utils/sys_utils.h>
 
 #include <exception>
 #include <iostream>
 #include <string>
 
-#include "backend/backend_impl.h"
-#include "utils/sys_utils.h"
+#include "backend_impl.h"
 
 namespace py = pybind11;
 
@@ -34,12 +34,12 @@ struct NoGilSection {
  * @brief Install the plugin
  */
 void Initialize() {
-  // Initialize the Example MLIR LTC Backend
-  torch::lazy::InitExampleMlirBackend();
+  // Initialize the Reference Lazy Backend
+  torch::lazy::InitReferenceLazyBackend();
 
   // sanity check
   const torch::lazy::BackendImplInterface* mlir_backend =
-      torch::lazy::GetExampleMlirBackendImpl();
+      torch::lazy::GetReferenceLazyBackendImpl();
   const torch::lazy::BackendImplInterface* lazy_backend =
       torch::lazy::getBackend();
   if (lazy_backend != mlir_backend) {
@@ -62,12 +62,12 @@ void Shutdown() {
 }
 } // anonymous namespace
 
-PYBIND11_MODULE(_EXAMPLE_MLIR_BACKEND, m) {
+PYBIND11_MODULE(_REFERENCE_LAZY_BACKEND, m) {
   py::class_<torch::lazy::TorchMlirComputation>(m, "TorchMlirComputation")
       .def("to_string", &torch::lazy::TorchMlirComputation::to_string)
       .def("debug_string", &torch::lazy::TorchMlirComputation::debug_string);
 
-  m.doc() = ("pybind11 for example MLIR LTC backend.");
+  m.doc() = ("pybind11 for the Reference Lazy backend.");
   m.def("get_latest_computation", []() {
     auto computation = static_cast<torch::lazy::TorchMlirComputation*>(
         torch::lazy::GetLatestComputation().get());

python/torch_mlir_e2e_test/torchscript/configs/lazy_tensor_core.py

@@ -3,7 +3,7 @@
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 # Also available under a BSD-style license. See LICENSE.
 
-import ltc_backend.ltc_backend._EXAMPLE_MLIR_BACKEND as ltc_backend
+import torch_mlir.reference_lazy_backend._REFERENCE_LAZY_BACKEND as lazy_backend
 import torch
 from torch.utils._pytree import tree_map
 
@@ -20,7 +20,7 @@ class LazyTensorCoreTestConfig(TestConfig):
 
     def __init__(self):
         super().__init__()
-        ltc_backend._initialize()
+        lazy_backend._initialize()
 
     def compile(self, program: torch.nn.Module) -> torch.nn.Module:
         return program.to('lazy')