From e8b23817a4acc7c2cabc916e732b0e8e7fc8e5bd Mon Sep 17 00:00:00 2001 From: Ranjit Ranjan Date: Mon, 17 Apr 2023 01:26:00 -0500 Subject: [PATCH 2/7] merge changes from tvkaibranch4 to rel-1.14.1 --- .../test/contrib_ops/function_test_util.h | 26 ++++++++- .../test/framework/sparse_kernels_test.cc | 6 ++ .../test/framework/tensorutils_test.cc | 56 +++++++++++++++++++ .../test/framework/test_tensor_loader.cc | 17 +++++- onnxruntime/test/onnx/main.cc | 4 +- onnxruntime/test/onnx/tensorprotoutils.cc | 27 ++++++++- .../optimizer/graph_transform_test_builder.h | 5 ++ .../test/optimizer/initializer_test.cc | 4 +- .../test/optimizer/nchwc_optimizer_test.cc | 4 ++ .../test/providers/provider_test_utils.cc | 3 + .../test/python/onnxruntime_test_python.py | 2 +- .../symbols.exp | 2 + tools/ci_build/gen_def.py | 12 ++-- 13 files changed, 154 insertions(+), 14 deletions(-) create mode 100644 onnxruntime/test/testdata/custom_execution_provider_library/symbols.exp diff --git a/onnxruntime/test/contrib_ops/function_test_util.h b/onnxruntime/test/contrib_ops/function_test_util.h index 9b104eefb..f2b8eed25 100644 --- a/onnxruntime/test/contrib_ops/function_test_util.h +++ b/onnxruntime/test/contrib_ops/function_test_util.h @@ -102,11 +102,33 @@ struct FunctionTestCase { template void AddInput(std::string input_name, std::vector shape) { + std::vector shapeforrandom; + shapeforrandom = shape; + char *bytes1 = (char *)shape.data(); + if (1) { +#ifdef DEBUG_AIX + std::cout<<"Doing byte swapping in AddInput test/contrib_ops/function_test_util.h "<(), shape); input_args.emplace_back(input_name, &arg_type); if (GenData) { - std::vector data = random(shape); + std::vector data = random(shapeforrandom); OrtValue ort_value; CreateMLValue(provider->GetAllocator(0, OrtMemTypeDefault), shape, data, &ort_value); input_values.push_back(std::make_pair(input_name, ort_value)); @@ -149,4 +171,4 @@ struct FunctionTestCase { }; } // namespace test -} // namespace onnxruntime \ No newline at end of file +} // namespace onnxruntime diff --git a/onnxruntime/test/framework/sparse_kernels_test.cc b/onnxruntime/test/framework/sparse_kernels_test.cc index 3b3884325..b38267db4 100644 --- a/onnxruntime/test/framework/sparse_kernels_test.cc +++ b/onnxruntime/test/framework/sparse_kernels_test.cc @@ -705,6 +705,9 @@ struct InsertIndices { // Conversion on the fly to the target data type std::vector indices(indices_data.cbegin(), indices_data.cend()); indices_tp.mutable_raw_data()->assign(reinterpret_cast(indices.data()), indices.size() * sizeof(T)); + if constexpr (endian::native != endian::little) { + utils::ConvertRawDataInTensorProto((ONNX_NAMESPACE::TensorProto*)&indices_tp); + } } } }; @@ -838,6 +841,9 @@ template static void RawDataWriter(const std::vector& values, TensorProto& tp, TensorProto_DataType datatype) { tp.set_data_type(datatype); tp.set_raw_data(values.data(), values.size() * sizeof(T)); + if constexpr (endian::native != endian::little) { + utils::ConvertRawDataInTensorProto((ONNX_NAMESPACE::TensorProto*)&tp); + } } int64_t ActualSize(const TensorProto& actual) { diff --git a/onnxruntime/test/framework/tensorutils_test.cc b/onnxruntime/test/framework/tensorutils_test.cc index e632f0f44..bffcbb048 100644 --- a/onnxruntime/test/framework/tensorutils_test.cc +++ b/onnxruntime/test/framework/tensorutils_test.cc @@ -31,6 +31,9 @@ void TestUnpackFloatTensor(TensorProto_DataType type, const Path& model_path) { memcpy(rawdata + i * sizeof(T), &(f[i]), sizeof(T)); } float_tensor_proto.set_raw_data(rawdata, len); + if constexpr (endian::native != endian::little) { + utils::ConvertRawDataInTensorProto((ONNX_NAMESPACE::TensorProto*)&float_tensor_proto); + } T float_data2[4]; auto status = UnpackTensor(float_tensor_proto, model_path, float_data2, 4); EXPECT_TRUE(status.IsOK()) << status.ErrorMessage(); @@ -104,6 +107,23 @@ std::vector CreateValues() { template void WriteDataToFile(FILE* fp, const std::vector& test_data) { + char *bytes1 = (char *)test_data.data(); + if constexpr (endian::native != endian::little) { + const size_t element_size = sizeof(T); + const size_t num_elements = test_data.size(); + for (size_t i = 0; i < num_elements; ++i) { + char* start_byte = bytes1 + i * element_size; + char* end_byte = start_byte + element_size - 1; + /* keep swapping */ + for (size_t count = 0; count < element_size / 2; ++count) { + char temp = *start_byte; + *start_byte = *end_byte; + *end_byte = temp; + ++start_byte; + --end_byte; + } + } + } size_t size_in_bytes = test_data.size() * sizeof(T); ASSERT_EQ(size_in_bytes, fwrite(test_data.data(), 1, size_in_bytes, fp)); } @@ -148,6 +168,24 @@ void UnpackAndValidate(const TensorProto& tensor_proto, const Path& model_path, auto st = utils::UnpackTensor(tensor_proto, model_path, val.data(), test_data.size()); ASSERT_TRUE(st.IsOK()) << st.ErrorMessage(); + char *bytes1 = (char *)val.data(); + if constexpr (endian::native != endian::little) { + const size_t element_size = sizeof(T); + const size_t num_elements = test_data.size(); + for (size_t i = 0; i < num_elements; ++i) { + char* start_byte = bytes1 + i * element_size; + char* end_byte = start_byte + element_size - 1; + /* keep swapping */ + for (size_t count = 0; count < element_size / 2; ++count) { + char temp = *start_byte; + *start_byte = *end_byte; + *end_byte = temp; + ++start_byte; + --end_byte; + } + } + } + // Validate data for (size_t i = 0; i < test_data.size(); i++) { ASSERT_TRUE(val[i] == test_data[i]); // need to use ASSERT_TRUE with '==' to handle MFLoat16 and BFloat16 @@ -325,6 +363,24 @@ static void TestConstantNodeConversionWithExternalData(TensorProto_DataType type std::vector val(test_data.size()); auto st = utils::UnpackTensor(tp, model_path, val.data(), test_data.size()); ASSERT_TRUE(st.IsOK()) << st.ErrorMessage(); + char *bytes1 = (char *)val.data(); + if constexpr (endian::native != endian::little) { + const size_t element_size = sizeof(T); + const size_t num_elements = test_data.size(); + for (size_t i = 0; i < num_elements; ++i) { + char* start_byte = bytes1 + i * element_size; + char* end_byte = start_byte + element_size - 1; + /* keep swapping */ + for (size_t count = 0; count < element_size / 2; ++count) { + char temp = *start_byte; + *start_byte = *end_byte; + *end_byte = temp; + ++start_byte; + --end_byte; + } + } + } + for (size_t i = 0; i < test_data.size(); i++) { ASSERT_EQ(val[i], test_data[i]); } diff --git a/onnxruntime/test/framework/test_tensor_loader.cc b/onnxruntime/test/framework/test_tensor_loader.cc index f627409bb..29f6e7e89 100644 --- a/onnxruntime/test/framework/test_tensor_loader.cc +++ b/onnxruntime/test/framework/test_tensor_loader.cc @@ -72,6 +72,21 @@ static void run_external_data_test() { std::unique_ptr file_deleter(const_cast(filename.c_str()), DeleteFileFromDisk); float test_data[] = {1.0f, 2.2f, 3.5f}; + int element_size = sizeof(float); + char *bytes = (char *)&test_data; + const size_t num_elements = sizeof(test_data); + for (size_t i = 0; i < num_elements; ++i) { + char* start_byte = bytes + i * element_size; + char* end_byte = start_byte + element_size - 1; + /* keep swapping */ + for (size_t count = 0; count < element_size / 2; ++count) { + char temp = *start_byte; + *start_byte = *end_byte; + *end_byte = temp; + ++start_byte; + --end_byte; + } + } ASSERT_EQ(sizeof(test_data), fwrite(test_data, 1, sizeof(test_data), fp)); ASSERT_EQ(0, fclose(fp)); // construct a tensor proto @@ -97,7 +112,7 @@ static void run_external_data_test() { ASSERT_NE(len, (DWORD)0); cwd.append(ORT_TSTR("\\fake.onnx")); #else - char* p = getcwd(nullptr, 0); + char* p = getcwd(nullptr, 512); ASSERT_NE(p, nullptr); cwd = p; free(p); diff --git a/onnxruntime/test/onnx/main.cc b/onnxruntime/test/onnx/main.cc index 11d635519..1fc4c9bda 100644 --- a/onnxruntime/test/onnx/main.cc +++ b/onnxruntime/test/onnx/main.cc @@ -6,8 +6,10 @@ #include #include #include -#ifdef _WIN32 +#if defined(_WIN32) #include "getopt.h" +#elif defined(_AIX) +#include #else #include #include diff --git a/onnxruntime/test/onnx/tensorprotoutils.cc b/onnxruntime/test/onnx/tensorprotoutils.cc index afafd0c6b..22929857c 100644 --- a/onnxruntime/test/onnx/tensorprotoutils.cc +++ b/onnxruntime/test/onnx/tensorprotoutils.cc @@ -69,10 +69,31 @@ static void UnpackTensorWithRawData(const void* raw_data, size_t raw_data_length expected_size_in_bytes, ", got ", raw_data_length), OrtErrorCode::ORT_FAIL); if constexpr (endian::native != endian::little) { - ORT_CXX_API_THROW("UnpackTensorWithRawData only handles little-endian native byte order for now.", - OrtErrorCode::ORT_NOT_IMPLEMENTED); + /* Convert Endianness and copy */ + char *nraw_data = (char*)malloc(raw_data_length); + memcpy(nraw_data, raw_data, raw_data_length); + char* bytes = nraw_data; + size_t element_size = sizeof(T); + size_t num_elements = raw_data_length / element_size; + + for (size_t i = 0; i < num_elements; ++i) { + char* start_byte = bytes + i * element_size; + char* end_byte = start_byte + element_size - 1; + /* keep swapping */ + for (size_t count = 0; count < element_size / 2; ++count) { + char temp = *start_byte; + *start_byte = *end_byte; + *end_byte = temp; + ++start_byte; + --end_byte; + } + } + memcpy(p_data, nraw_data, raw_data_length); + free(nraw_data); } - memcpy(p_data, raw_data, raw_data_length); + else{ + memcpy(p_data, raw_data, raw_data_length); + } } // This macro doesn't work for Float16/bool/string tensors diff --git a/onnxruntime/test/optimizer/graph_transform_test_builder.h b/onnxruntime/test/optimizer/graph_transform_test_builder.h index 14c73b2b5..3678fa4d2 100644 --- a/onnxruntime/test/optimizer/graph_transform_test_builder.h +++ b/onnxruntime/test/optimizer/graph_transform_test_builder.h @@ -11,6 +11,7 @@ #include "core/framework/framework_common.h" #include "core/optimizer/graph_transformer_level.h" #include "core/graph/onnx_protobuf.h" +#include "core/framework/tensorprotoutils.h" #include "test/framework/test_utils.h" #include "test/common/tensor_op_test_utils.h" #include "test/framework/test_utils.h" @@ -143,6 +144,10 @@ class ModelTestBuilder { tensor_proto.set_data_type(utils::ToTensorProtoElementType()); tensor_proto.set_raw_data(data.data(), data.size() * sizeof(T)); + if (endian::native != endian::little){ + ::onnxruntime::utils::ConvertRawDataInTensorProto(&tensor_proto); + } + for (auto& dim : shape) { tensor_proto.add_dims(dim); } diff --git a/onnxruntime/test/optimizer/initializer_test.cc b/onnxruntime/test/optimizer/initializer_test.cc index e73950bbd..b86d8f65e 100644 --- a/onnxruntime/test/optimizer/initializer_test.cc +++ b/onnxruntime/test/optimizer/initializer_test.cc @@ -163,7 +163,9 @@ void TestInitializerRawData() { tensor_proto.add_dims(3); tensor_proto.add_dims(4); tensor_proto.set_raw_data(data.data(), data.size() * sizeof(T)); - + if constexpr (endian::native != endian::little) { + utils::ConvertRawDataInTensorProto((ONNX_NAMESPACE::TensorProto *) &tensor_proto); + } const Initializer init(tensor_proto, Path()); for (size_t idx = 0; idx < data.size(); idx++) { diff --git a/onnxruntime/test/optimizer/nchwc_optimizer_test.cc b/onnxruntime/test/optimizer/nchwc_optimizer_test.cc index 06beedbd5..3038bea52 100644 --- a/onnxruntime/test/optimizer/nchwc_optimizer_test.cc +++ b/onnxruntime/test/optimizer/nchwc_optimizer_test.cc @@ -6,6 +6,7 @@ #include "core/mlas/inc/mlas.h" #include "core/session/environment.h" #include "core/session/inference_session.h" +#include "core/framework/tensorprotoutils.h" #include "test/compare_ortvalue.h" #include "test/test_environment.h" #include "test/framework/test_utils.h" @@ -64,6 +65,9 @@ struct NchwcTestHelper { tensor_proto.set_data_type(utils::ToTensorProtoElementType()); tensor_proto.set_raw_data(data.data(), data.size() * sizeof(T)); + if constexpr (endian::native != endian::little) { + ::onnxruntime::utils::ConvertRawDataInTensorProto((ONNX_NAMESPACE::TensorProto*)&tensor_proto); + } for (auto& dim : shape) { tensor_proto.add_dims(dim); } diff --git a/onnxruntime/test/providers/provider_test_utils.cc b/onnxruntime/test/providers/provider_test_utils.cc index fa8ef69f9..82823b9bc 100644 --- a/onnxruntime/test/providers/provider_test_utils.cc +++ b/onnxruntime/test/providers/provider_test_utils.cc @@ -776,6 +776,9 @@ void OpTester::AddInitializers(onnxruntime::Graph& graph) { } else { auto buffer_size = tensor.DataType()->Size() * shape.Size(); tensor_proto.set_raw_data(tensor.DataRaw(), buffer_size); + if constexpr (endian::native != endian::little) { + ::onnxruntime::utils::ConvertRawDataInTensorProto((ONNX_NAMESPACE::TensorProto*)&tensor_proto); + } } // 4. name tensor_proto.set_name(data.def_.Name()); diff --git a/onnxruntime/test/python/onnxruntime_test_python.py b/onnxruntime/test/python/onnxruntime_test_python.py index 89fd90ad3..d6e54bde9 100644 --- a/onnxruntime/test/python/onnxruntime_test_python.py +++ b/onnxruntime/test/python/onnxruntime_test_python.py @@ -1213,7 +1213,7 @@ class TestInferenceSession(unittest.TestCase): ) def testMemoryArenaShrinkage(self): - if platform.architecture()[0] == "32bit" or "ppc" in platform.machine() or "powerpc" in platform.machine(): + if platform.architecture()[0] == "32bit" or "ppc" in platform.machine() or "powerpc" in platform.machine() or "powerpc" in platform.processor(): # on x86 or ppc builds, the CPU allocator does not use an arena print("Skipping testMemoryArenaShrinkage in 32bit or powerpc platform.") else: diff --git a/onnxruntime/test/testdata/custom_execution_provider_library/symbols.exp b/onnxruntime/test/testdata/custom_execution_provider_library/symbols.exp new file mode 100644 index 000000000..c9a5232be --- /dev/null +++ b/onnxruntime/test/testdata/custom_execution_provider_library/symbols.exp @@ -0,0 +1,2 @@ +GetProvider +ProviderHashFunc diff --git a/tools/ci_build/gen_def.py b/tools/ci_build/gen_def.py index b020fd2f9..5fc8a5fb3 100755 --- a/tools/ci_build/gen_def.py +++ b/tools/ci_build/gen_def.py @@ -41,8 +41,9 @@ with open(args.output, "w") as file: elif args.style == "xcode": pass # xcode compile don't has any header. else: - file.write("VERS_%s {\n" % VERSION_STRING) - file.write(" global:\n") + pass + #file.write("VERS_%s {\n" % VERSION_STRING) + #file.write(" global:\n") for symbol in symbols: if args.style == "vc": @@ -54,9 +55,10 @@ with open(args.output, "w") as file: symbol_index += 1 if args.style == "gcc": - file.write(" local:\n") - file.write(" *;\n") - file.write("}; \n") + pass + #file.write(" local:\n") + #file.write(" *;\n") + #file.write("}; \n") with open(args.output_source, "w") as file: file.write("#include \n") -- 2.38.3