ExecutionSession constructor without entry function name (llvm#1163)

* ExecutionSession constructor without entry function name

Signed-off-by: Tung D. Le <[email protected]>

* clang-format

Signed-off-by: Tung D. Le <[email protected]>

* Do not use run_main_graph in ExecutionSession

Signed-off-by: Tung D. Le <[email protected]>

* Revise

Signed-off-by: Tung D. Le <[email protected]>

* Add omEntryPointName to pass Windows CI

Signed-off-by: Tung D. Le <[email protected]>

* Address review comments

Signed-off-by: Tung D. Le <[email protected]>

* Use run_main_graph as the default entry point

Signed-off-by: Tung D. Le <[email protected]>

* undo numerical.def

Signed-off-by: Tung D. Le <[email protected]>

* Change messages

Signed-off-by: Tung D. Le <[email protected]>

Co-authored-by: Alexandre Eichenberger <[email protected]>

Author: tungld

docs/UsingPyRuntime.md

@@ -34,12 +34,10 @@ The complete interface to ExecutionSession can be seen in the sources mentioned
 using the constructor and run method is enough to perform inferences.
 
 ```python
-def __init__(self, path: str, entry_point: str):
+def __init__(self, path: str):
     """
     Args:
         path: relative or absolute path to your .so model.
-        entry_point: function generated by onnx-mlir to call inferences.
-            Use 'run_main_graph'.
     """
 
 def run(self, input: List[ndarray]) -> List[ndarray]:
@@ -72,7 +70,7 @@ import numpy as np
 from PyRuntime import ExecutionSession
 
 model = 'model.so' # LeNet from ONNX Zoo compiled with onnx-mlir
-session = ExecutionSession(model, "run_main_graph")
+session = ExecutionSession(model)
 print("input signature in json", session.input_signature())
 print("output signature in json",session.output_signature())
 input = np.full((1, 1, 28, 28), 1, np.dtype(np.float32))

docs/mnist_example/README.md

@@ -198,7 +198,7 @@ The runtime use an `ExecutionSession` object to hold a specific model and entry
 ``` Python
 # Load the model mnist.so compiled with onnx-mlir.
 model = 'mnist.so'
-session = ExecutionSession(model, "run_main_graph")
+session = ExecutionSession(model)
 # Print the models input/output signature, for display.
 # If there are problems with the signature functions, they can be simply commented out.
 print("input signature in json", session.input_signature())

docs/mnist_example/mnist.py

@@ -3,7 +3,7 @@
 
 # Load the model mnist.so compiled with onnx-mlir.
 model = './mnist.so'
-session = ExecutionSession(model, "run_main_graph")
+session = ExecutionSession(model)
 # Print the models input/output signature, for display.
 # Signature functions for info only, commented out if they cause problems.
 print("input signature in json", session.input_signature())

src/Conversion/KrnlToLLVM/KrnlToLLVM.cpp

@@ -1194,14 +1194,19 @@ class KrnlEntryPointOpLowering : public OpRewritePattern<KrnlEntryPointOp> {
               KrnlEntryPointOp::getEntryPointFuncAttrName())
             .getLeafReference()
             .getValue();
-    auto dynEntryPointName = "run_" + staticEntryPointFuncName;
-    assert(module.lookupSymbol(dynEntryPointName.str()) == nullptr &&
-           "dynamic entry point name is not unique");
+
+    // When there is only a single entry point function in a model, use
+    // "run_main_graph" as the default name.
+    // TODO(tung): support multiple entry point functions.
+    std::string entryPointName = "run_main_graph";
+    assert(module.lookupSymbol(entryPointName) == nullptr &&
+           "Only support a single entry point function.");
+
     rewriter.eraseOp(op);
     auto dynEntryPointFuncTy =
         LLVM::LLVMFunctionType::get(opaquePtrTy, {opaquePtrTy}, false);
     auto dynamicEntryPointFunc = rewriter.create<LLVM::LLVMFuncOp>(
-        loc, dynEntryPointName.str(), dynEntryPointFuncTy);
+        loc, entryPointName, dynEntryPointFuncTy);
     auto &entryPointEntryBlock =
         createEntryBlock(dynEntryPointFuncTy, dynamicEntryPointFunc, loc);
     rewriter.setInsertionPointToStart(&entryPointEntryBlock);

src/Runtime/ExecutionSession.cpp

@@ -25,6 +25,9 @@ namespace onnx_mlir {
 const std::string ExecutionSession::_inputSignatureName = "omInputSignature";
 const std::string ExecutionSession::_outputSignatureName = "omOutputSignature";
 
+ExecutionSession::ExecutionSession(std::string sharedLibPath)
+    : ExecutionSession::ExecutionSession(sharedLibPath, "") {}
+
 ExecutionSession::ExecutionSession(
     std::string sharedLibPath, std::string entryPointName) {
 
@@ -36,6 +39,11 @@ ExecutionSession::ExecutionSession(
     throw std::runtime_error(errStr.str());
   }
 
+  // When entry point name is not given, use the default "run_main_graph".
+  // TODO(tung): support multiple entry point functions.
+  if (entryPointName.empty())
+    entryPointName = "run_main_graph";
+
   _entryPointFunc = reinterpret_cast<entryPointFuncType>(
       _sharedLibraryHandle.getAddressOfSymbol(entryPointName.c_str()));
   if (!_entryPointFunc) {

src/Runtime/ExecutionSession.hpp

@@ -31,6 +31,7 @@ using OMTensorUniquePtr = std::unique_ptr<OMTensor, decltype(&omTensorDestroy)>;
 
 class ExecutionSession {
 public:
+  ExecutionSession(std::string sharedLibPath);
   ExecutionSession(std::string sharedLibPath, std::string entryPointName);
 
   // Use custom deleter since forward declared OMTensor hides destructor

src/Runtime/PyExecutionSession.hpp

@@ -25,8 +25,10 @@ namespace onnx_mlir {
 
 class PyExecutionSession : public onnx_mlir::ExecutionSession {
 public:
+  PyExecutionSession(std::string sharedLibPath)
+      : onnx_mlir::ExecutionSession(sharedLibPath) {}
   PyExecutionSession(std::string sharedLibPath, std::string entryPointName)
-      : onnx_mlir::ExecutionSession(sharedLibPath, entryPointName){};
+      : onnx_mlir::ExecutionSession(sharedLibPath, entryPointName) {}
 
   std::vector<py::array> pyRun(const std::vector<py::array> &inputsPyArray);
 
@@ -37,6 +39,7 @@ class PyExecutionSession : public onnx_mlir::ExecutionSession {
 
 PYBIND11_MODULE(PyRuntime, m) {
   py::class_<onnx_mlir::PyExecutionSession>(m, "ExecutionSession")
+      .def(py::init<const std::string &>())
       .def(py::init<const std::string &, const std::string &>())
       .def("run", &onnx_mlir::PyExecutionSession::pyRun)
       .def("input_signature", &onnx_mlir::PyExecutionSession::pyInputSignature)

test/backend/inference_backend.py

@@ -972,7 +972,6 @@ def JniExecutionSession(jar_name, inputs):
 class EndiannessAwareExecutionSession(object):
     def __init__(self, model):
         self.model = model
-        self.entry_point = "run_main_graph"
         self.exec_name = None
         # Compiling the model in advance if not testing constants, so that
         # the model is compiled once and used multiple times.
@@ -1060,7 +1059,7 @@ def run(self, inputs, **kwargs):
                 inputs = self.turn_model_input_to_constant(inputs)
                 self.exec_name = compile_model(self.model, args.emit)
             if args.emit == "lib":
-                session = ExecutionSession(self.exec_name, self.entry_point)
+                session = ExecutionSession(self.exec_name)
                 outputs = session.run(inputs)
                 # print('input='+str(inputs), file=sys.stderr)
                 # print('output='+str(outputs), file=sys.stderr)
@@ -1079,7 +1078,7 @@ def run(self, inputs, **kwargs):
                 "Cannot deduce desired output endianness, using native endianness by default."
             )
             if args.emit == "lib":
-                session = ExecutionSession(self.exec_name, self.entry_point)
+                session = ExecutionSession(self.exec_name)
                 outputs = session.run(inputs)
             elif args.emit == "jni":
                 outputs = JniExecutionSession(self.exec_name, inputs)

test/backend/signature_backend.py

@@ -166,14 +166,13 @@ def run(test_self, device):  # type: (Any, Text) -> None
 class SignatureExecutionSession(object):
     def __init__(self, model):
         self.model = model
-        self.entry_point = "run_main_graph"
         self.exec_name = compile_model(self.model, args.emit)
 
     def run(self, **kwargs):
         sys.path.append(RUNTIME_DIR)
         from PyRuntime import ExecutionSession
 
-        session = ExecutionSession(self.exec_name, self.entry_point)
+        session = ExecutionSession(self.exec_name)
         output = session.input_signature()
         return output
 

test/numerical/TestConv.cpp

@@ -141,8 +141,7 @@ bool isOMConvTheSameAsNaiveImplFor(const int N, const int C, const int H,
           /*output conv param*/ NOut, COut, HOut, WOut))
     return false;
 
-  onnx_mlir::ExecutionSession sess(
-      getSharedLibName(SHARED_LIB_BASE.str()), "run_main_graph");
+  onnx_mlir::ExecutionSession sess(getSharedLibName(SHARED_LIB_BASE.str()));
 
   std::vector<OMTensorUniquePtr> inputs;
   auto xOmt = OMTensorUniquePtr(

test/numerical/TestGRU.cpp

@@ -46,8 +46,7 @@ bool isOMGRUTheSameAsNaiveImplFor(const int direction, const int S, const int B,
           /* GRU param out*/
           D, xShape, hShape, wOmt, rOmt, bOmt))
     return false;
-  onnx_mlir::ExecutionSession sess(
-      getSharedLibName(SHARED_LIB_BASE.str()), "run_main_graph");
+  onnx_mlir::ExecutionSession sess(getSharedLibName(SHARED_LIB_BASE.str()));
 
   std::vector<OMTensorUniquePtr> inputs;
   auto xOmt = OMTensorUniquePtr(

test/numerical/TestGemm.cpp

@@ -76,8 +76,7 @@ bool isOMGemmTheSameAsNaiveImplFor(const int I, const int J, const int K,
           aShape, bShape, cShape))
     return false;
 
-  onnx_mlir::ExecutionSession sess(
-      getSharedLibName(SHARED_LIB_BASE.str()), "run_main_graph");
+  onnx_mlir::ExecutionSession sess(getSharedLibName(SHARED_LIB_BASE.str()));
 
   std::vector<OMTensorUniquePtr> inputs;
   auto aOmt = OMTensorUniquePtr(

test/numerical/TestLSTM.cpp

@@ -48,8 +48,7 @@ bool isOMLSTMTheSameAsNaiveImplFor(const int direction, const int S,
           D, xShape, hShape, cShape, wOmt, rOmt, bOmt, pOmt))
     return false;
 
-  onnx_mlir::ExecutionSession sess(
-      getSharedLibName(SHARED_LIB_BASE.str()), "run_main_graph");
+  onnx_mlir::ExecutionSession sess(getSharedLibName(SHARED_LIB_BASE.str()));
 
   std::vector<OMTensorUniquePtr> inputs;
   auto xOmt = OMTensorUniquePtr(

test/numerical/TestLoop.cpp

@@ -87,8 +87,7 @@ bool isOMLoopTheSameAsNaiveImplFor(std::string moduleIR,
   auto module = mlir::parseSourceString(moduleIR, &ctx);
   OwningModuleRef moduleRef(std::move(module));
   compileModule(moduleRef, ctx, SHARED_LIB_BASE.str(), onnx_mlir::EmitLib);
-  onnx_mlir::ExecutionSession sess(
-      getSharedLibName(SHARED_LIB_BASE.str()), "run_main_graph");
+  onnx_mlir::ExecutionSession sess(getSharedLibName(SHARED_LIB_BASE.str()));
 
   std::vector<OMTensorUniquePtr> inputs;
   auto tripCountTensor = OMTensorUniquePtr(

test/numerical/TestMatMul2D.cpp

@@ -39,8 +39,7 @@ bool isOMMatmulTheSameAsNaiveImplFor(const int I, const int J, const int K) {
           I, J, K))
     return false;
 
-  onnx_mlir::ExecutionSession sess(
-      getSharedLibName(SHARED_LIB_BASE.str()), "run_main_graph");
+  onnx_mlir::ExecutionSession sess(getSharedLibName(SHARED_LIB_BASE.str()));
 
   std::vector<OMTensorUniquePtr> inputs;
   auto aOmt = OMTensorUniquePtr(

test/numerical/TestRNN.cpp

@@ -47,8 +47,7 @@ bool isOMRNNTheSameAsNaiveImplFor(const int direction, const int S, const int B,
           D, xShape, hShape, wOmt, rOmt, bOmt))
     return false;
 
-  onnx_mlir::ExecutionSession sess(
-      getSharedLibName(SHARED_LIB_BASE.str()), "run_main_graph");
+  onnx_mlir::ExecutionSession sess(getSharedLibName(SHARED_LIB_BASE.str()));
 
   std::vector<OMTensorUniquePtr> inputs;
   auto xOmt = OMTensorUniquePtr(

utils/RunONNXModel.py

@@ -306,7 +306,7 @@ def main():
         # Use the generated shared library to create an execution session.
         print("Loading the compiled model ...")
         start = time.perf_counter()
-        sess = ExecutionSession(shared_lib_path, "run_main_graph")
+        sess = ExecutionSession(shared_lib_path)
         end = time.perf_counter()
         print("  took ", end - start, " seconds.\n")