diff --git a/cmake/CMakeLists.txt b/cmake/CMakeLists.txt index 158356b67f..3d2c7fa4b7 100644 --- a/cmake/CMakeLists.txt +++ b/cmake/CMakeLists.txt @@ -110,6 +110,13 @@ else() add_definitions(-DUSE_OPENMP) endif() endif() + +if (${CMAKE_SYSTEM_NAME} MATCHES "Darwin") + #For Mac compliance + message("Adding flags for Mac builds") + set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fstack-protector-strong") +endif() + find_package(PNG) set(ENABLE_DATE_TESTING OFF CACHE BOOL "" FORCE) set(USE_SYSTEM_TZ_DB ON CACHE BOOL "" FORCE) diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake index 8acc389302..382be43a9e 100644 --- a/cmake/external/mkldnn.cmake +++ b/cmake/external/mkldnn.cmake @@ -24,7 +24,11 @@ if(WIN32) set(IOMP5MD_SHARED_LIB libiomp5md.dll) endif() else() - set(MKLDNN_SHARED_LIB libmkldnn.so.0) + if (APPLE) + set(MKLDNN_SHARED_LIB libmkldnn.0.dylib) + else() + set(MKLDNN_SHARED_LIB libmkldnn.so.0) + endif() if(onnxruntime_USE_MKLML) set(DOWNLOAD_MKLML ${MKLDNN_SOURCE}/scripts/prepare_mkl.sh) set(MKLML_SHARED_LIB libmklml_intel.so) diff --git a/cmake/onnxruntime_python.cmake b/cmake/onnxruntime_python.cmake index 8f2f55c69c..22db4d45f2 100644 --- a/cmake/onnxruntime_python.cmake +++ b/cmake/onnxruntime_python.cmake @@ -75,13 +75,16 @@ add_dependencies(onnxruntime_pybind11_state ${onnxruntime_pybind11_state_depende if (MSVC) # if MSVC, pybind11 looks for release version of python lib (pybind11/detail/common.h undefs _DEBUG) target_link_libraries(onnxruntime_pybind11_state ${onnxruntime_pybind11_state_libs} ${onnxruntime_EXTERNAL_LIBRARIES} ${PYTHON_LIBRARY_RELEASE} ${ONNXRUNTIME_SO_LINK_FLAG}) +elseif (APPLE) + set_target_properties(onnxruntime_pybind11_state PROPERTIES LINK_FLAGS "-undefined dynamic_lookup") + target_link_libraries(onnxruntime_pybind11_state ${onnxruntime_pybind11_state_libs} ${onnxruntime_EXTERNAL_LIBRARIES} ${ONNXRUNTIME_SO_LINK_FLAG}) + set_target_properties(onnxruntime_pybind11_state PROPERTIES + INSTALL_RPATH "@loader_path" + BUILD_WITH_INSTALL_RPATH TRUE + INSTALL_RPATH_USE_LINK_PATH FALSE) else() target_link_libraries(onnxruntime_pybind11_state ${onnxruntime_pybind11_state_libs} ${onnxruntime_EXTERNAL_LIBRARIES} ${PYTHON_LIBRARY} ${ONNXRUNTIME_SO_LINK_FLAG}) - if (APPLE) - set_target_properties(onnxruntime_pybind11_state PROPERTIES INSTALL_RPATH "@loader_path") - else() - set_target_properties(onnxruntime_pybind11_state PROPERTIES LINK_FLAGS "-Xlinker -rpath=\$ORIGIN") - endif() + set_target_properties(onnxruntime_pybind11_state PROPERTIES LINK_FLAGS "-Xlinker -rpath=\$ORIGIN") endif() set_target_properties(onnxruntime_pybind11_state PROPERTIES PREFIX "") diff --git a/csharp/src/Microsoft.ML.OnnxRuntime/Microsoft.ML.OnnxRuntime.csproj b/csharp/src/Microsoft.ML.OnnxRuntime/Microsoft.ML.OnnxRuntime.csproj index 05ed0811f3..b52213d52b 100644 --- a/csharp/src/Microsoft.ML.OnnxRuntime/Microsoft.ML.OnnxRuntime.csproj +++ b/csharp/src/Microsoft.ML.OnnxRuntime/Microsoft.ML.OnnxRuntime.csproj @@ -12,16 +12,19 @@ Microsoft Corporation This package contains Microsoft's implementation of ONNX runtime, usable in .Net platforms ONNX;ONNX Runtime;Machine Learning - Microsoft Corporation - true + © Microsoft Corporation. All rights reserved. https://github.com/Microsoft/onnxruntime/blob/master/LICENSE - https://github.com/Microsoft/onnxruntime.git - git ..\.. $(OnnxRuntimeCsharpRoot)\..\build\Windows $(buildDirectory)\$(Configuration)\$(Configuration) + + + true + + + $(AllowedOutputExtensionsInPackageBuildOutputFolder);.pdb @@ -66,20 +69,20 @@ - + - $(GitCommitHash) @(MajorVersionNumber) $(PackageVersion) - $(PackageVersion)-dev-$(GitCommitHash) + $(PackageVersion)-dev-$(GitCommitHashShort) + diff --git a/docs/CSharp_API.md b/docs/CSharp_API.md index d8339f1fa9..ae333af3d6 100644 --- a/docs/CSharp_API.md +++ b/docs/CSharp_API.md @@ -2,14 +2,14 @@ The ONNX runtime provides a C# .Net binding for running inference on ONNX models in any of the .Net standard platforms. The API is .Net standard 1.1 compliant for maximum portability. This document describes the API. ## NuGet Package -There is a NuGet package Microsoft.ML.OnnxRuntime available for .Net consumers, which includes the prebuilt binaries for ONNX runtime. The API is portable across all platforms and architectures supported by the .Net standard, although currently the NuGet package contains the prebuilt binaries for Windows 10 platform on x64 CPUs only. +The Microsoft.ML.OnnxRuntime Nuget package includes the precompiled binaries for ONNX runtime, and includes libraries for Windows 10 platform and X64 CPUs. The APIs conform to .Net Standard 1.1. ## Getting Started -Here is simple tutorial for getting started with running inference on an existing ONNX model for a given input data (a.k.a query). Say the model is trained using any of the well-known training frameworks and exported as an ONNX model into a file named `model.onnx`. The runtime incarnation of a model is an `InferenceSession` object. You simply construct an `InferenceSession` object using the model file as parameter -- +Here is simple tutorial for getting started with running inference on an existing ONNX model for a given input data. The model is typically trained using any of the well-known training frameworks and exported into the ONNX format. To start scoring using the model, open a session using the `InferenceSession` class, passing in the file path to the model as a parameter. var session = new InferenceSession("model.onnx"); -Once a session is created, you can run queries on the session using your input data, using the `Run` method of the `InferenceSession`. Both input and output of `Run` method are represented as collections of .Net `Tensor` objects (as defined in [System.Numerics.Tensor](https://www.nuget.org/packages/System.Numerics.Tensors)) - +Once a session is created, you can execute queries using the `Run` method of the `InferenceSession` object. Currently, only `Tensor` type of input and outputs are supported. The results of the `Run` method are represented as a collection of .Net `Tensor` objects (as defined in [System.Numerics.Tensor](https://www.nuget.org/packages/System.Numerics.Tensors)). Tensor t1, t2; // let's say data is fed into the Tensor objects var inputs = new List() @@ -19,7 +19,8 @@ Once a session is created, you can run queries on the session using your input d }; IReadOnlyCollection results = session.Run(inputs); -You can load your input data into Tensor objects in several ways. A simple example is to create the Tensor from arrays - +You can load your input data into Tensor objects in several ways. A simple example is to create the Tensor from arrays. + float[] sourceData; // assume your data is loaded into a flat float array int[] dimensions; // and the dimensions of the input is stored here Tensor t1 = new DenseTensor(sourceData, dimensions); @@ -84,7 +85,7 @@ Accessor to the default static option object #### Methods AppendExecutionProvider(ExecutionProvider provider); -Appends execution provider to the session. For any operator in the graph the first execution provider that implements the operator will be user. ExecutionProvider is defined as the following enum - +Appends execution provider to the session. For any operator in the graph the first execution provider that implements the operator will be user. ExecutionProvider is defined as the following enum. enum ExecutionProvider { @@ -112,4 +113,3 @@ The type of Exception that is thrown in most of the error conditions related to - diff --git a/docs/HighLevelDesign.md b/docs/HighLevelDesign.md index 7b9dc27e9a..8e6b955934 100644 --- a/docs/HighLevelDesign.md +++ b/docs/HighLevelDesign.md @@ -1,7 +1,7 @@ # ONNX Runtime High Level Design This document outlines the high level design of -ONNXRuntime - a high performance, cross platform engine. +ONNX Runtime - a high performance, cross platform engine. ## Key objectives * Maximally and automatically leverage the custom accelerators and runtimes @@ -10,8 +10,8 @@ available on disparate platforms. runtimes. We call this abstraction an [execution provider](../include/onnxruntime/core/framework/execution_provider.h). It defines and exposes a set of its capabilities to ONNXRuntime: a set of single or fused nodes it can -execute, its memory allocator and more. Custom accelerators and runtimes are -instances of execution provider. +execute, its memory allocator, and more. Custom accelerators and runtimes are +instances of execution providers. * We don't expect that an execution provider can always run an ONNX model fully on its device. This means that ONNXRuntime must be able to execute a single model in a heterogeneous environment involving multiple execution providers. @@ -35,46 +35,45 @@ provider using the GetCapability() API. ![ONNXRuntime high level system architecture](https://azurecomcdn.azureedge.net/mediahandler/acomblog/media/Default/blog/228d22d3-6e3e-48b1-811c-1d48353f031c.png) -*Note: TensorRT and nGraph support in the works.* +*Note: TensorRT and nGraph support are in progress* ### More about partitioning -ONNXRuntime partitions a model graph based on the available execution providers -into subgraphs, each for a distinct provider respectively. ONNXRuntime provides -a default execution provider that is used for fallback execution for the +ONNXRuntime partitions a model graph into subgraphs based on the available execution providers, one for each distinct provider. ONNXRuntime provides +a default execution provider that is used as the fallback execution for the operators that cannot be pushed onto the more specialized but more efficient -execution providers. Intuitively we probably want to push computation to the -specialized execution providers as much as possible. +execution providers. Intuitively we want to push computation to more +specialized execution providers whenever possible. We use a simple graph partitioning technique. The available execution providers will be considered in a specific order, and each will be assigned the maximal subgraphs (possibly more than one) that it is able to handle. The -ONNXRuntime-provided default execution provider will be the last one to be +ONNXRuntime-provided default execution provider will be the last one considered, and it ensures completeness. More sophisticated optimizations can be considered in the future (or can even be implemented as a composite execution provider). Conceptually, each partition is reduced to a single fused operator. It is -created by invoking the execution provider's Compile() method and wrap it as a +created by invoking the execution provider's Compile() method and wraps it as a custom operator. Currently we support only synchronous mode of execution. An execution provider exposes its memory allocator, which is used to allocate the input tensors for the execution provider. The rewriting and partitioning transform the -initial model graph into a new graph composed with operators assigned to either +initial model graph into a new graph composed of operators assigned to either the default execution provider or other registered execution -providers. ONNXRuntime execution engine is responsible for running this graph. +providers. The ONNXRuntime execution engine is responsible for running this graph. ## Key design decisions -* Multiple threads should be able to inovke the Run() method on the same +* Multiple threads can invoke the Run() method on the same inference session object. See [API doc](C_API.md) for more details. -* To facilitate the above the Compute() function of all kernels is const +* To facilitate this, the Compute() function of all kernels is const implying the kernels are stateless. -* We call implementations of the operators by execution providers as +* Implementations of the operators by execution providers are called kernels. Each execution provider supports a subset of the (ONNX) operators/kernels. -* ONNXRuntime runtime guarantees that all operators are supported by the default +* The ONNXRuntime runtime guarantees that all operators are supported by the default execution provider. * Tensor representation: ONNXRuntime will utilize a standard representation for the tensor runtime values. The execution providers can internally use a -different representation, if they choose to, but it is their responsibility to +different representation if they choose to, but it is their responsibility to convert the values from/to the standard representation at the boundaries of their subgraph. diff --git a/onnxruntime/contrib_ops/contrib_ops.cc b/onnxruntime/contrib_ops/contrib_ops.cc index f6151690e7..3b3d948ee9 100644 --- a/onnxruntime/contrib_ops/contrib_ops.cc +++ b/onnxruntime/contrib_ops/contrib_ops.cc @@ -147,7 +147,7 @@ The dequantization formula is y = (x - x_zero_point) * x_scale. .SinceVersion(1) .SetDoc(R"DOC( The convolution operator consumes a quantized input tensor, its scale and zero point, -a quantized filter, its scale and zero point, and output’s scale and zero point, +a quantized filter, its scale and zero point, and output's scale and zero point, and computes the quantized output. Each scale and zero point pair must have same shape. It means they must be either scalars (per tensor) or 1-D tensors (per channel).)DOC") .Input( @@ -163,8 +163,8 @@ It means they must be either scalars (per tensor) or 1-D tensors (per channel).) "to arrive with the dimension denotation of [DATA_BATCH, " "DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE ...].", "T1") - .Input(1, "x_scale", "Scale tensor for input ‘x’. It could be a scalar or a 1-D tensor, which means a per-tensor or per-channel quantization. If it’s a 1-D tensor, its number of elements should be equal to the number of channels of input ‘x’.", "T3") - .Input(2, "x_zero_point", "Zero point tensor for input ‘x’. It could be a scalar or a 1-D tensor, which means a per-tensor or per-channel quantization. If it’s a 1-D tensor, its number of elements should be equal to the number of channels of input ‘x’.", "T1") + .Input(1, "x_scale", "Scale tensor for input 'x'. It could be a scalar or a 1-D tensor, which means a per-tensor or per-channel quantization. If it's a 1-D tensor, its number of elements should be equal to the number of channels of input 'x'.", "T3") + .Input(2, "x_zero_point", "Zero point tensor for input 'x'. It could be a scalar or a 1-D tensor, which means a per-tensor or per-channel quantization. If it's a 1-D tensor, its number of elements should be equal to the number of channels of input 'x'.", "T1") .Input( 3, "w", @@ -183,10 +183,10 @@ It means they must be either scalars (per tensor) or 1-D tensors (per channel).) "(assuming zero based indices for the shape array). " "Or in other words FILTER_IN_CHANNEL should be equal to DATA_CHANNEL. ", "T1") - .Input(4, "w_scale", "Scale tensor for input ‘w’. It could be a scalar or a 1-D tensor, which means a per-tensor or per-channel quantization. If it’s a 1-D tensor, its number of elements should be equal to the number of channels of input ‘w’.", "T3") - .Input(5, "w_zero_point", "Scale tensor for input ‘w’. It could be a scalar or a 1-D tensor, which means a per-tensor or per-channel quantization. If it’s a 1-D tensor, its number of elements should be equal to the number of channels of input ‘w’.", "T1") - .Input(6, "y_scale", "Scale tensor for output ‘y’. It could be a scalar or a 1-D tensor, which means a per-tensor or per-channel quantization. If it’s a 1-D tensor, its number of elements should be equal to the number of channels of input ‘y’.", "T3") - .Input(7, "y_zero_point", "Scale tensor for output ‘y’. It could be a scalar or a 1-D tensor, which means a per-tensor or per-channel quantization. If it’s a 1-D tensor, its number of elements should be equal to the number of channels of input ‘y’.", "T1") + .Input(4, "w_scale", "Scale tensor for input 'w'. It could be a scalar or a 1-D tensor, which means a per-tensor or per-channel quantization. If it's a 1-D tensor, its number of elements should be equal to the number of channels of input 'w'.", "T3") + .Input(5, "w_zero_point", "Scale tensor for input 'w'. It could be a scalar or a 1-D tensor, which means a per-tensor or per-channel quantization. If it's a 1-D tensor, its number of elements should be equal to the number of channels of input 'w'.", "T1") + .Input(6, "y_scale", "Scale tensor for output 'y'. It could be a scalar or a 1-D tensor, which means a per-tensor or per-channel quantization. If it's a 1-D tensor, its number of elements should be equal to the number of channels of input 'y'.", "T3") + .Input(7, "y_zero_point", "Scale tensor for output 'y'. It could be a scalar or a 1-D tensor, which means a per-tensor or per-channel quantization. If it's a 1-D tensor, its number of elements should be equal to the number of channels of input 'y'.", "T1") .Input(8, "B", "Optional 1D bias to be added to the convolution, has size of M.", "T2", OpSchema::Optional) .Output( 0, diff --git a/onnxruntime/core/providers/cpu/reduction/reduction_ops.cc b/onnxruntime/core/providers/cpu/reduction/reduction_ops.cc index ae17d9c563..a332aa2c33 100644 --- a/onnxruntime/core/providers/cpu/reduction/reduction_ops.cc +++ b/onnxruntime/core/providers/cpu/reduction/reduction_ops.cc @@ -34,14 +34,20 @@ REGISTER_UNARY_ELEMENTWISE_KERNEL(ReduceSumSquare, 1); REGISTER_UNARY_ELEMENTWISE_KERNEL(ArgMax, 1); REGISTER_UNARY_ELEMENTWISE_KERNEL(ArgMin, 1); +// When all reduce axises located at the tail of the dims, quite general cases, transpose and extra +// copy could be skiped to improve performance, if required by check_no_transpose = true; +// return value: true means transposedInputData is not created/copied, input tensor data could +// be direct use as row major matrix [block_size, blocks], where blocks is the +// size of each reduce. template -void PrepareForReduce(OpKernelContext* ctx, +bool PrepareForReduce(OpKernelContext* ctx, std::vector& transposedInputData, Tensor** reducedTensor, int64_t& block_size, int64_t& blocks, const std::vector& axes_, - bool keepdims_) { + bool keepdims_, + bool check_no_transpose = false) { const Tensor* input_tensor_ptr = ctx->Input(0); ONNXRUNTIME_ENFORCE(input_tensor_ptr != nullptr); const Tensor& input = *input_tensor_ptr; @@ -51,8 +57,6 @@ void PrepareForReduce(OpKernelContext* ctx, ONNXRUNTIME_ENFORCE(axe >= 0 && axe < (int64_t)ndim, "Axis attribute out of range"); } - transposedInputData.resize(input.Shape().Size(), 0); - std::vector axes = axes_; if (axes.empty()) { // This is the default case for non-arg kind reductions. Reduce on all dimensions. @@ -62,6 +66,13 @@ void PrepareForReduce(OpKernelContext* ctx, std::sort(axes.begin(), axes.end()); + // If all reduced axes are located at the tail of the input shape, then copy could be skipped is required + bool need_copy = true; + if (axes.size() <= ndim && axes.front() == static_cast(ndim - axes.size()) + && axes.back() == static_cast(ndim) - 1) { + need_copy = false; + } + vector keep_axis(ndim, true); for (auto i : axes) { keep_axis[i] = false; @@ -96,7 +107,6 @@ void PrepareForReduce(OpKernelContext* ctx, } const T* from_data = input.template Data(); - T* to_data = &transposedInputData[0]; size_t count = input.Shape().Size(); //set to-be-reduced axes to one. squeeze is keepdims_ is false @@ -117,9 +127,15 @@ void PrepareForReduce(OpKernelContext* ctx, block_size = input.Shape().Size() / first_dim; blocks = first_dim; + if (!need_copy && check_no_transpose) { + return true; + } + + transposedInputData.resize(input.Shape().Size(), 0); + T* to_data = &transposedInputData[0]; if (num_axes < 2 || n_shared_idxs == num_axes) { memcpy(to_data, from_data, count * sizeof(T)); - return; + return false; } int itr_axes = num_axes - n_shared_idxs; @@ -178,6 +194,7 @@ void PrepareForReduce(OpKernelContext* ctx, } } } + return false; } template @@ -272,12 +289,22 @@ Status ReduceMean::Compute(OpKernelContext* ctx) const { std::vector transposedInputData; int64_t block_size, blocks; Tensor* reduced; - PrepareForReduce(ctx, transposedInputData, &reduced, block_size, blocks, axes_, keepdims_); + bool no_transpose = PrepareForReduce(ctx, transposedInputData, &reduced, block_size, blocks, axes_, keepdims_, true); T* output_data = reduced->template MutableData(); - EigenVectorMap out_vec(output_data, block_size); - out_vec = ConstEigenMatrixMap(&transposedInputData[0], block_size, blocks).rowwise().mean(); + if (no_transpose) { + const T* input_data = ctx->Input(0)->template Data(); + + #pragma omp parallel for + for (int64_t i = 0; i < block_size; ++i) { + output_data[i] = ConstEigenVectorMap(input_data + (i * blocks), blocks).mean(); + } + } + else { + EigenVectorMap out_vec(output_data, block_size); + out_vec = ConstEigenMatrixMap(&transposedInputData[0], block_size, blocks).rowwise().mean(); + } return Status::OK(); } @@ -317,12 +344,22 @@ Status ReduceSum::Compute(OpKernelContext* ctx) const { std::vector transposedInputData; int64_t block_size, blocks; Tensor* reduced; - PrepareForReduce(ctx, transposedInputData, &reduced, block_size, blocks, axes_, keepdims_); + bool no_transpose = PrepareForReduce(ctx, transposedInputData, &reduced, block_size, blocks, axes_, keepdims_, true); T* output_data = reduced->template MutableData(); - EigenVectorMap out_vec(output_data, block_size); - out_vec = ConstEigenMatrixMap(&transposedInputData[0], block_size, blocks).rowwise().sum(); + if (no_transpose) { + const T* input_data = ctx->Input(0)->template Data(); + + #pragma omp parallel for + for (int64_t i = 0; i < block_size; ++i) { + output_data[i] = ConstEigenVectorMap(input_data + (i * blocks), blocks).sum(); + } + } + else { + EigenVectorMap out_vec(output_data, block_size); + out_vec = ConstEigenMatrixMap(&transposedInputData[0], block_size, blocks).rowwise().sum(); + } return Status::OK(); } diff --git a/setup.py b/setup.py index ea32fa84bf..812886f0bd 100644 --- a/setup.py +++ b/setup.py @@ -25,6 +25,8 @@ except ImportError: # Additional binaries if platform.system() == 'Linux': libs = ['onnxruntime_pybind11_state.so', 'libmkldnn.so.0', 'libmklml_intel.so', 'libiomp5.so'] +elif platform.system() == "Darwin": + libs = ['onnxruntime_pybind11_state.so', 'libmkldnn.0.dylib'] # TODO add libmklml and libiomp5 later. else: libs = ['onnxruntime_pybind11_state.pyd', 'mkldnn.dll', 'mklml.dll', 'libiomp5md.dll'] diff --git a/tools/ci_build/github/azure-pipelines/azure-pipelines.yml b/tools/ci_build/github/azure-pipelines/azure-pipelines.yml index 1ebfce85de..047e0cda91 100644 --- a/tools/ci_build/github/azure-pipelines/azure-pipelines.yml +++ b/tools/ci_build/github/azure-pipelines/azure-pipelines.yml @@ -5,7 +5,7 @@ jobs: pool: Linux-CPU steps: - - script: 'tools/ci_build/github/linux/run_dockerbuild.sh -o ubuntu16.04 -d cpu -r $(Build.BinariesDirectory)' + - script: 'tools/ci_build/github/linux/run_dockerbuild.sh -o ubuntu16.04 -d cpu -r $(Build.BinariesDirectory) -x "--use_mklml"' displayName: 'Command Line Script' env: AZURE_BLOB_KEY: $(onnxruntime-storage-key) diff --git a/tools/ci_build/github/linux/run_build.sh b/tools/ci_build/github/linux/run_build.sh index 083204a6a6..83110363ce 100755 --- a/tools/ci_build/github/linux/run_build.sh +++ b/tools/ci_build/github/linux/run_build.sh @@ -38,6 +38,6 @@ else --config Debug Release --build_shared_lib \ --skip_submodule_sync \ --enable_pybind \ - --parallel --use_mkldnn --use_mklml --build_shared_lib $BUILD_EXTR_PAR + --parallel --use_mkldnn --build_shared_lib $BUILD_EXTR_PAR /home/onnxruntimedev/Release/onnx_test_runner /data/onnx fi