Make partitioner and preprocess work

1. Make compilation pass (script build.sh)
2. Enable simple function for partitioner
3. Fix errors in preprocess phase.

Co-authored-by: chong-chen <[email protected]>

Author: chong-chen01

backends/samsung/CMakeLists.txt

@@ -11,8 +11,8 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
 get_filename_component(EXECUTORCH_SOURCE_DIR "${CMAKE_CURRENT_LIST_DIR}/../.." ABSOLUTE)
 
-if(NOT DEFINED EXYNOS_AI_LITECORE_PATH)
-	message(FATAL_ERROR "Please define EXYNOS_AI_LIRECORE_PATH by adding cmake parameter -DEXYNOS_AI_LITECORE_PATH=<...>")
+if(NOT DEFINED EXYNOS_AI_LITECORE_ROOT)
+	message(FATAL_ERROR "Please define EXYNOS_AI_LIRECORE_PATH by adding cmake parameter -DEXYNOS_AI_LITECORE_ROOT=<...>")
 endif()
 if(CMAKE_TOOLCHAIN_FILE MATCHES ".*(iOS|ios\.toolchain)\.cmake$")
 	message(FATAL_ERROR "IOS is not supported on Exynos.")
@@ -32,8 +32,12 @@ if(CMAKE_BUILD_TYPE STREQUAL "Release")
     add_definitions(-DNDEBUG)
 endif()
 
-include_directories(${EXECUTORCH_SOURCE_DIR}/..)
-include_directories(${EXYNOS_AI_LITECORE_PATH})
+include_directories(
+    ${EXECUTORCH_SOURCE_DIR}/..
+    ${EXECUTORCH_SOURCE_DIR}/runtime/core/portable_type/c10
+    ${EXYNOS_AI_LITECORE_ROOT}
+)
+add_compile_definitions(C10_USING_CUSTOM_GENERATED_MACROS)
 
 if(${ANDROID})
     find_library(android_log log)
@@ -50,10 +54,10 @@ if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64")
     )
     add_library(PyEnnWrapperAdaptor MODULE)
 
-    find_library(GG_API_LIB NAMES graphgen_api HINTS ${EXYNOS_AI_LITECORE_PATH}/lib64/)
+    find_library(GG_API_LIB NAMES graphgen_api HINTS ${EXYNOS_AI_LITECORE_ROOT}/lib64/)
     add_library(graphgen_api SHARED IMPORTED GLOBAL)
     set_target_properties(graphgen_api PROPERTIES
-        INTERFACE_INCLUDE_DIRECTORIES "${EXYNOS_AI_LITECORE_PATH}/include"
+        INTERFACE_INCLUDE_DIRECTORIES "${EXYNOS_AI_LITECORE_ROOT}/include"
         IMPORTED_LOCATION "${GG_API_LIB}"
     )
 
@@ -96,10 +100,10 @@ if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64")
     # PyGraphWrapperAdaptor
     add_library(PyGraphWrapperAdaptor MODULE)
     #
-    find_library(GRAPH_WRAPPER_LIB NAMES graph_wrapper HINTS ${EXYNOS_AI_LITECORE_PATH}/lib64/)
+    find_library(GRAPH_WRAPPER_LIB NAMES graph_wrapper HINTS ${EXYNOS_AI_LITECORE_ROOT}/lib64/)
     add_library(graph_wrapper SHARED IMPORTED GLOBAL)
     set_target_properties(graph_wrapper PROPERTIES
-            INTERFACE_INCLUDE_DIRECTORIES "${EXYNOS_AI_LITECORE_PATH}/include"
+            INTERFACE_INCLUDE_DIRECTORIES "${EXYNOS_AI_LITECORE_ROOT}/include"
             IMPORTED_LOCATION "${GRAPH_WRAPPER_LIB}"
     )
     set_target_properties(PyGraphWrapperAdaptor PROPERTIES CXX_VISIBILITY_PRESET hidden)
@@ -120,6 +124,7 @@ if(${ANDROID})
     add_library(enn_backend STATIC)
     target_link_libraries(enn_backend PRIVATE enn_logging)
     target_link_options_shared_lib(enn_backend)
+    target_compile_options(enn_backend PRIVATE -Wno-deprecated-declarations)
 
     set(__enn_executor_runner_srcs ${EXECUTORCH_SOURCE_DIR}/examples/samsung/executor_runner/enn_executor_runner.cpp)
     add_executable(enn_executor_runner ${__enn_executor_runner_srcs})

backends/samsung/aot/PyEnnWrapperAdaptor.h

@@ -13,6 +13,9 @@
 #include <pybind11/numpy.h>
 #include <pybind11/pybind11.h>
 
+#include <executorch/backends/samsung/compile_options_def_generated.h>
+#include <executorch/backends/samsung/runtime/logging.h>
+
 #include <iostream>
 #include <memory>
 #include <vector>
@@ -29,18 +32,26 @@ class PyEnnWrapper {
 
   void Init(const py::bytes& compile_opts) {
     graphgen_instance_ = graphgen_create();
+    option_buf_ = enn_option::GetEnnExecuTorchOptions(
+        compile_opts.cast<std::string_view>().data());
   }
 
   bool IsNodeSupportedByBackend() {
-    return False;
+    return false;
   }
 
   py::array_t<char> Compile(const py::array_t<char>& model_buffer) {
     if (graphgen_instance_ == nullptr) {
       ENN_LOG_ERROR("Please call `Init()` first before compile.");
       return py::array_t<char>();
     }
-
+    auto soc_name = option_buf_->chipset();
+    if (graphgen_initialize_context(graphgen_instance_, soc_name) !=
+        GraphGenResult::SUCCESS) {
+      ENN_LOG_ERROR(
+          "Unsupported Soc (%d), please check your chipset version.", soc_name);
+      return py::array_t<char>();
+    }
 
     auto m_buf_info = model_buffer.request();
     auto* model_buf_ptr = reinterpret_cast<uint8_t*>(m_buf_info.ptr);
@@ -52,7 +63,7 @@ class PyEnnWrapper {
       return py::array_t<char>();
     }
 
-   auto result = py::array_t<char>({nnc_buffer->size}, {sizeof(char)});
+    auto result = py::array_t<char>({nnc_buffer->size}, {sizeof(char)});
     auto result_buf = result.request();
     memcpy(result_buf.ptr, nnc_buffer->addr, nnc_buffer->size);
 
@@ -73,6 +84,8 @@ class PyEnnWrapper {
  private:
   // pointer to enn software entry
   void* graphgen_instance_ = nullptr;
+  // enn compilation option buf
+  const enn_option::EnnExecuTorchOptions* option_buf_ = nullptr;
 };
 } // namespace enn
 } // namespace executor

backends/samsung/build.sh

@@ -18,7 +18,7 @@ function usage() {
     <command>           <argument>                 <description>
 
     --sdk                                      The path of downloaded ENN SDK, which is required for building.
-                                               Or export EXYNOS_AI_LITECORE_ROOT=/path/to/enn_sdk_xxx
+                                               Or export EXYNOS_AI_LITECORE_ROOT=/path/to/xxx
     --ndk                                      The path of Android NDK, or export ANDROID_NDK_ROOT=/path/to/ndk.
 
     --build, -b     [x86_64, android, all]     Default is all, x86_64 target to offline compilation,
@@ -34,14 +34,16 @@ function build_x86_64() {
     exit 1
   fi
 
-  echo EXYNOS_AI_LITECORE_ROOT: ${EXYNOS_AI_LITECORE_ROOT}
-  echo ANDROID_NDK_ROOT: ${ANDROID_NDK_ROOT}
+  echo "EXYNOS_AI_LITECORE_ROOT: ${EXYNOS_AI_LITECORE_ROOT}"
+  echo "ANDROID_NDK_ROOT: ${ANDROID_NDK_ROOT}"
 
   cmake \
         -DCMAKE_INSTALL_PREFIX=${X86_64_BUILD_DIR} \
         -DEXYNOS_AI_LITECORE_ROOT=${EXYNOS_AI_LITECORE_ROOT} \
-        -DEXECUTORCH_BUILD_ENN=ON \
+        -DEXECUTORCH_BUILD_SAMSUNG=ON \
         -DEXECUTORCH_BUILD_DEVTOOLS=ON \
+        -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
+	      -DEXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR=ON \
         -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
         -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
         -S ${PROJECT_DIR} \
@@ -70,12 +72,15 @@ function build_android() {
         -DCMAKE_TOOLCHAIN_FILE="${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake" \
         -DANDROID_ABI="${ANDROID_ABI}" \
         -DCMAKE_BUILD_TYPE=Release \
-        -DEXECUTORCH_BUILD_ENN=ON \
+        -DEXECUTORCH_BUILD_SAMSUNG=ON \
         -DEXYNOS_AI_LITECORE_ROOT=${EXYNOS_AI_LITECORE_ROOT} \
-        -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
         -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
+        -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
+	      -DEXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR=ON \
         -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
         -DEXECUTORCH_ENABLE_LOGGING=1 \
+        -DEXECUTORCH_BUILD_DEVTOOLS=ON \
+        -DEXECUTORCH_ENABLE_EVENT_TRACER=ON \
         -S ${PROJECT_DIR} \
         -B ${ANDROID_BUILD_DIR}
 

backends/samsung/enn_preprocess.py

@@ -15,6 +15,12 @@
 )
 from executorch.backends.samsung.serialization.enn_graph_schema import EnnGraph
 from executorch.backends.samsung.utils.utils import get_compile_spec
+from executorch.backends.transforms.addmm_mm_to_linear import AddmmToLinearTransform
+from executorch.backends.transforms.fuse_batch_norm_with_conv import (
+    FuseBatchNormWithConvPass,
+)
+
+from executorch.backends.transforms.remove_getitem_op import RemoveGetItemPass
 
 from executorch.exir.backend.backend_details import (
     BackendDetails,
@@ -40,12 +46,17 @@ def preprocess(
         )
         enn_wrapper.Init(option_spec.value)
 
-        enn_preprocess_passes = PassManager(passes=[])
+        enn_preprocess_passes = PassManager(
+            passes=[
+                FuseBatchNormWithConvPass(edge_program),
+                AddmmToLinearTransform(),
+                RemoveGetItemPass(),
+            ]
+        )
         pass_result = enn_preprocess_passes(edge_program.graph_module)
         assert pass_result is not None
 
         enn_graph = EnnGraph()
-        enn_graph.init("UnknownName", "")
         # node visitors
         node_visitors = get_node_visitors(edge_program)
 
@@ -71,7 +82,6 @@ def preprocess(
                 raise RuntimeError(f"{node.op}" " is not supported in ENN Delegate")
 
         # Compile Graph
-        enn_wrapper.Destroy()
         enn_graph.finish()
         ser_buf = enn_graph.serialize()
         enn_context_binary = enn_wrapper.Compile(ser_buf)

backends/samsung/partition/enn_partitioner.py

@@ -3,15 +3,18 @@
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
-
 import logging
-from typing import Any, Callable, Dict, List, Optional, Tuple
+from typing import Any, Dict, List
+
+import executorch.backends.samsung.builders.node_visitor as node_visitor
 
 import executorch.backends.samsung.python.PyEnnWrapperAdaptor as PyEnnWrapper
 
 import torch
 from executorch.backends.samsung.enn_preprocess import EnnBackend
-
+from executorch.backends.samsung.serialization.compile_options import (
+    ENN_COMPILE_OPTION_TITLE,
+)
 from executorch.backends.samsung.utils.utils import get_compile_spec
 from executorch.exir.backend.backend_details import CompileSpec
 from executorch.exir.backend.canonical_partitioners.pattern_op_partitioner import (
@@ -23,9 +26,15 @@
     PartitionResult,
 )
 from executorch.exir.backend.utils import tag_constant_data
+
+from executorch.exir.dialects._ops import ops as exir_ops
 from torch.fx.passes.infra.partitioner import Partition
 from torch.fx.passes.operator_support import OperatorSupportBase
 
+SUPPORTED_OPS = [
+    exir_ops.edge.aten.addmm.default,
+]
+
 
 class EnnOperatorSupport(OperatorSupportBase):
 
@@ -36,7 +45,10 @@ def __init__(
     ):
         self.edge_program = edge_program
         self.enn_wrapper = PyEnnWrapper.EnnWrapper()
-        option_spec = get_compile_spec(compile_specs, "Exynos compile", required=True)
+        self.node_visitors = node_visitor.get_node_visitors(edge_program)
+        option_spec = get_compile_spec(
+            compile_specs, ENN_COMPILE_OPTION_TITLE, required=True
+        )
         self.enn_wrapper.Init(option_spec.value)
 
     def is_node_supported(self, _, node: torch.fx.Node) -> bool:
@@ -50,6 +62,9 @@ def is_node_supported(self, _, node: torch.fx.Node) -> bool:
         ]:
             return False
 
+        if node.target in SUPPORTED_OPS or node.target.__name__ in self.node_visitors:
+            return True
+
         supported = self.enn_wrapper.IsNodeSupportedByBackend()
         return supported
 
@@ -67,7 +82,7 @@ def __init__(self, compile_specs: List[CompileSpec]):
     def generate_partitions(
         self, edge_program: torch.export.ExportedProgram
     ) -> List[Any]:
-
+        self.op_support_checker = EnnOperatorSupport(edge_program, self.compile_specs)
         return generate_partitions_from_list_of_nodes(
             edge_program.graph_module,
             op_support=self.op_support_checker,

backends/samsung/python/.gitignore

@@ -0,0 +1,5 @@
+# exclude the pybind libraries
+*
+
+# keep folder for ease of use
+!.gitignore

backends/samsung/utils/export_utils.py

@@ -4,26 +4,31 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-from typing import List, Optional, Tuple
+from typing import Optional, Tuple
 
 import executorch.exir as exir
 import torch
-from executorch.exir import EdgeCompileConfig, ExportedProgram
+from executorch.backends.samsung.partition.enn_partitioner import EnnPartitioner
+from executorch.backends.transforms.remove_clone_ops import RemoveCloneOpsTransform
+from executorch.exir import EdgeCompileConfig
 from executorch.exir.backend.backend_details import CompileSpec
-from executorch.exir.program._program import (
-    to_edge_transform_and_lower,
-)
+from executorch.exir.program._program import to_edge_transform_and_lower
 
 
 def to_edge_transform_and_lower_to_enn(
     module: torch.nn.Module,
     inputs: Tuple[torch.Tensor],
-    custom_pass_config: List[PassType] = None,
     compile_specs: Optional[CompileSpec] = None,
 ) -> exir.ExecutorchProgramManager:
-    assert compile_specs is not None, "For now, we must deliver complile specs"
+    assert (
+        compile_specs is not None
+    ), "Please provide compile specifications for enn backend"
     prog = torch.export.export(module, inputs)
+
+    ahead_pass_list = [RemoveCloneOpsTransform()]
     return to_edge_transform_and_lower(
         prog,
-        compile_config=[],
+        ahead_pass_list,
+        {"forward": [EnnPartitioner(compile_specs)]},
+        compile_config=EdgeCompileConfig(_skip_dim_order=True),
     )

examples/samsung/executor_runner/enn_executor_runner.cpp

@@ -1,6 +1,7 @@
 /*
- *  Copyright (c) 2025 Samsung Electronics Co. LTD
- *  All rights reserved
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * Copyright (c) 2025 Samsung Electronics Co. LTD
+ * All rights reserved
  *
  * This source code is licensed under the BSD-style license found in the
  * LICENSE file in the root directory of this source tree.
@@ -265,9 +266,11 @@ int main(int argc, char** argv) {
   status = method->get_outputs(outputs.data(), outputs.size());
   ET_CHECK(status == Error::Ok);
 
-  for (size_t output_index = 0; output_index < method->outputs_size(); ++output_index) {
-	  // Save the results to given directory in order.
-       saveOutput(output_tensor, output_index);
+  for (size_t output_index = 0; output_index < method->outputs_size();
+       ++output_index) {
+    auto output_tensor = outputs[output_index].toTensor();
+    // Save the results to given directory in order.
+    saveOutput(output_tensor, output_index);
   }
 
   return 0;