Reference 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

@@ -165,4 +165,3 @@ else()
 endif()
 
 add_subdirectory(test)
-add_subdirectory(examples)

build_tools/autogen_ltc_backend.py

@@ -377,7 +377,7 @@ def extract_signatures(text):
                     // 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'
 

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 @@ def main(device='lazy', full_size=False):
             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 forward(self, x):
             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 @@ def forward(self, x):
             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.
@@ -162,6 +163,6 @@ 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/CMakeLists.txt → 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 {