Visible to Intel only — GUID: GUID-9D6CFC01-E0E0-4E55-AAC4-E0F4288C8CBA
Abs
AbsBackward
Add
AvgPool
AvgPoolBackward
BatchNormForwardTraining
BatchNormInference
BatchNormTrainingBackward
BiasAdd
BiasAddBackward
Clamp
ClampBackward
Concat
Convolution
ConvolutionBackwardData
ConvolutionBackwardWeights
ConvTranspose
ConvTransposeBackwardData
ConvTransposeBackwardWeights
Dequantize
Divide
DynamicDequantize
DynamicQuantize
Elu
EluBackward
End
Exp
GELU
GELUBackward
HardSigmoid
HardSigmoidBackward
HardSwish
HardSwishBackward
Interpolate
InterpolateBackward
LayerNorm
LayerNormBackward
LeakyReLU
Log
LogSoftmax
LogSoftmaxBackward
MatMul
Maximum
MaxPool
MaxPoolBackward
Minimum
Mish
MishBackward
Multiply
Pow
PReLU
PReLUBackward
Quantize
Reciprocal
ReduceL1
ReduceL2
ReduceMax
ReduceMean
ReduceMin
ReduceProd
ReduceSum
ReLU
ReLUBackward
Reorder
Round
Select
Sigmoid
SigmoidBackward
SoftMax
SoftMaxBackward
SoftPlus
SoftPlusBackward
Sqrt
SqrtBackward
Square
SquaredDifference
StaticReshape
StaticTranspose
Subtract
Tanh
TanhBackward
TypeCast
Wildcard
enum dnnl_alg_kind_t
enum dnnl_normalization_flags_t
enum dnnl_primitive_kind_t
enum dnnl_prop_kind_t
enum dnnl_query_t
enum dnnl::normalization_flags
enum dnnl::query
struct dnnl_exec_arg_t
struct dnnl_primitive
struct dnnl_primitive_desc
struct dnnl::primitive
struct dnnl::primitive_desc
struct dnnl::primitive_desc_base
enum dnnl_rnn_direction_t
enum dnnl_rnn_flags_t
enum dnnl::rnn_direction
enum dnnl::rnn_flags
struct dnnl::augru_backward
struct dnnl::augru_forward
struct dnnl::gru_backward
struct dnnl::gru_forward
struct dnnl::lbr_augru_backward
struct dnnl::lbr_augru_forward
struct dnnl::lbr_gru_backward
struct dnnl::lbr_gru_forward
struct dnnl::lstm_backward
struct dnnl::lstm_forward
struct dnnl::rnn_primitive_desc_base
struct dnnl::vanilla_rnn_backward
struct dnnl::vanilla_rnn_forward
Visible to Intel only — GUID: GUID-9D6CFC01-E0E0-4E55-AAC4-E0F4288C8CBA
Build Options
oneDNN supports the following build-time options.
CMake Option |
Supported values (defaults in bold) |
Description |
|---|---|---|
ONEDNN_LIBRARY_TYPE |
SHARED , STATIC |
Defines the resulting library type |
ONEDNN_CPU_RUNTIME |
NONE, OMP , TBB, SEQ, THREADPOOL, SYCL |
Defines the threading runtime for CPU engines |
ONEDNN_GPU_RUNTIME |
NONE , OCL, SYCL |
Defines the offload runtime for GPU engines |
ONEDNN_BUILD_EXAMPLES |
ON , OFF |
Controls building the examples |
ONEDNN_BUILD_TESTS |
ON , OFF |
Controls building the tests |
ONEDNN_BUILD_GRAPH |
ON , OFF |
Controls building graph component |
ONEDNN_ENABLE_GRAPH_DUMP |
ON, OFF |
Controls dumping graph artifacts |
ONEDNN_EXPERIMENTAL_GRAPH_COMPILER_BACKEND |
ON, OFF |
Enables the graph compiler backend of the graph component (experimental) |
ONEDNN_ARCH_OPT_FLAGS |
compiler flags |
Specifies compiler optimization flags (see warning note below) |
ONEDNN_ENABLE_CONCURRENT_EXEC |
ON, OFF |
Disables sharing a common scratchpad between primitives in dnnl::scratchpad_mode::library mode |
ONEDNN_ENABLE_JIT_PROFILING |
ON , OFF |
|
ONEDNN_ENABLE_ITT_TASKS |
ON , OFF |
|
ONEDNN_ENABLE_PRIMITIVE_CACHE |
ON , OFF |
Enables primitive cache |
ONEDNN_ENABLE_MAX_CPU_ISA |
ON , OFF |
Enables CPU dispatcher controls |
ONEDNN_ENABLE_CPU_ISA_HINTS |
ON , OFF |
Enables CPU ISA hints |
ONEDNN_ENABLE_WORKLOAD |
TRAINING , INFERENCE |
Specifies a set of functionality to be available based on workload |
ONEDNN_ENABLE_PRIMITIVE |
ALL , PRIMITIVE_NAME |
Specifies a set of functionality to be available based on primitives |
ONEDNN_ENABLE_PRIMITIVE_CPU_ISA |
ALL , CPU_ISA_NAME |
Specifies a set of functionality to be available for CPU backend based on CPU ISA |
ONEDNN_ENABLE_PRIMITIVE_GPU_ISA |
ALL , GPU_ISA_NAME |
Specifies a set of functionality to be available for GPU backend based on GPU ISA |
ONEDNN_ENABLE_GEMM_KERNELS_ISA |
ALL , NONE, ISA_NAME |
Specifies a set of functionality to be available for GeMM kernels for CPU backend based on ISA |
ONEDNN_EXPERIMENTAL |
ON, OFF |
Enables experimental features |
ONEDNN_VERBOSE |
ON , OFF |
Enables verbose mode |
ONEDNN_DEV_MODE |
ON, OFF |
Enables internal tracing and debuginfo logging in verbose output (for oneDNN developers) |
ONEDNN_AARCH64_USE_ACL |
ON, OFF |
Enables integration with Arm Compute Library for AArch64 builds |
ONEDNN_BLAS_VENDOR |
NONE , ARMPL |
Defines an external BLAS library to link to for GEMM-like operations |
ONEDNN_GPU_VENDOR |
INTEL , NVIDIA |
Defines GPU vendor for GPU engines |
ONEDNN_DPCPP_HOST_COMPILER |
DEFAULT , GNU C++ compiler executable |
Specifies host compiler executable for SYCL runtime |
ONEDNN_LIBRARY_NAME |
dnnl , library name |
Specifies name of the library |
All building options listed support their counterparts with DNNL prefix instead of ONEDNN. DNNL options would take precedence over ONEDNN versions, if both versions are specified.
All other building options or values that can be found in CMake files are intended for development/debug purposes and are subject to change without notice. Please avoid using them.
Common options
Host compiler
When building oneDNN with oneAPI DPC++/C++ Compiler user can specify a custom host compiler. The host compiler is a compiler that will be used by the main compiler driver to perform host compilation step.
The host compiler can be specified with ONEDNN_DPCPP_HOST_COMPILER CMake option. It should be specified either by name (in this case, the standard system environment variables will be used to discover it) or an absolute path to the compiler executable.
The default value of ONEDNN_DPCPP_HOST_COMPILER is DEFAULT, which is the default host compiler used by the compiler specified with CMAKE_CXX_COMPILER.
The DEFAULT host compiler is the only supported option on Windows. On Linux, user can specify a GNU C++ compiler as the host compiler.
WARNING:
oneAPI DPC++/C++ Compiler requires host compiler to be compatible. The minimum allowed GNU C++ compiler version is 7.4.0. See GCC* Compatibility and Interoperability section in oneAPI DPC++/C++ Compiler Developer Guide.
Configuring functionality
Using ONEDNN_ENABLE_WORKLOAD and ONEDNN_ENABLE_PRIMITIVE it is possible to limit functionality available in the final shared object or statically linked application. This helps to reduce the amount of disk space occupied by an app.
ONEDNN_ENABLE_WORKLOAD
This option supports only two values: TRAINING (the default) and INFERENCE. INFERENCE enables only forward propagation kind part of functionality, removing all backward-related functionality, except those which are dependencies for forward propagation kind part.
ONEDNN_ENABLE_PRIMITIVE
This option supports several values: ALL (the default) which enables all primitives implementations or a set of BATCH_NORMALIZATION, BINARY, CONCAT, CONVOLUTION, DECONVOLUTION, ELTWISE, INNER_PRODUCT, LAYER_NORMALIZATION, LRN, MATMUL, POOLING, PRELU, REDUCTION, REORDER, RESAMPLING, RNN, SHUFFLE, SOFTMAX, SUM. When a set is used, only those selected primitives implementations will be available. Attempting to use other primitive implementations will end up returning an unimplemented status when creating primitive descriptor. In order to specify a set, a CMake-style string should be used, with semicolon delimiters, as in this example:
-DONEDNN_ENABLE_PRIMITIVE=CONVOLUTION;MATMUL;REORDER
ONEDNN_ENABLE_PRIMITIVE_CPU_ISA
This option supports several values: ALL (the default) which enables all ISA implementations or one of SSE41, AVX2, AVX512, and AMX. Values are linearly ordered as SSE41 <AVX2 <AVX512 <AMX. When specified, selected ISA and all ISA that are “smaller” will be available. Example that enables SSE41 and AVX2 sets:
-DONEDNN_ENABLE_PRIMITIVE_CPU_ISA=AVX2
ONEDNN_ENABLE_PRIMITIVE_GPU_ISA
This option supports several values: ALL (the default) which enables all ISA implementations or any set of GEN9, GEN11, XELP, XEHP, XEHPG and XEHPC. Selected ISA will enable correspondent parts in just-in-time kernel generation based implementations. OpenCL based kernels and implementations will always be available. Example that enables XeLP and XeHP set:
-DONEDNN_ENABLE_PRIMITIVE_GPU_ISA=XELP;XEHP
ONEDNN_ENABLE_GEMM_KERNELS_ISA
This option supports several values: ALL (the default) which enables all ISA kernels from x64/gemm folder, NONE which disables all kernels and removes correspondent interfaces, or one of SSE41, AVX2, and AVX512. Values are linearly ordered as SSE41 <AVX2 <AVX512. When specified, selected ISA and all ISA that are “smaller” will be available. Example that leaves SSE41 and AVX2 sets, but removes AVX512 and AMX kernels:
-DONEDNN_ENABLE_GEMM_KERNELS_ISA=AVX2
CPU Options
Intel Architecture Processors and compatible devices are supported by oneDNN CPU engine. The CPU engine is built by default but can be disabled at build time by setting ONEDNN_CPU_RUNTIME to NONE. In this case, GPU engine must be enabled.
Targeting Specific Architecture
oneDNN uses JIT code generation to implement most of its functionality and will choose the best code based on detected processor features. However, some oneDNN functionality will still benefit from targeting a specific processor architecture at build time. You can use ONEDNN_ARCH_OPT_FLAGS CMake option for this.
For Intel(R) C++ Compilers, the default option is -xSSE4.1, which instructs the compiler to generate the code for the processors that support SSE4.1 instructions. This option would not allow you to run the library on older processor architectures.
For GNU* Compilers and Clang, the default option is -msse4.1.
WARNING:
While use of ONEDNN_ARCH_OPT_FLAGS option gives better performance, the resulting library can be run only on systems that have instruction set compatible with the target instruction set. Therefore, ARCH_OPT_FLAGS should be set to an empty string ("") if the resulting library needs to be portable.
Runtime CPU dispatcher control
oneDNN JIT relies on ISA features obtained from the processor it is being run on. There are situations when it is necessary to control this behavior at run-time to, for example, test SSE4.1 code on an AVX2-capable processor. The ONEDNN_ENABLE_MAX_CPU_ISA build option controls the availability of this feature. See CPU Dispatcher Control for more information.
Runtime CPU ISA hints
For performance reasons, sometimes oneDNN JIT needs to be provided with extra hints so as to prefer or avoid particular CPU ISA feature. For example, one might want to disable Zmm registers usage in order to take advantage of higher clock speed. The ONEDNN_ENABLE_CPU_ISA_HINTS build option makes this feature available at runtime. See CPU ISA Hints for more information.
Runtimes
CPU engine can use OpenMP, Threading Building Blocks (TBB) or sequential threading runtimes. OpenMP threading is the default build mode. This behavior is controlled by the ONEDNN_CPU_RUNTIME CMake option.
OpenMP
oneDNN uses OpenMP runtime library provided by the compiler.
When building oneDNN with oneAPI DPC++/C++ Compiler the library will link to Intel OpenMP runtime. This behavior can be changed by changing the host compiler with ONEDNN_DPCPP_HOST_COMPILER option.
WARNING:
Because different OpenMP runtimes may not be binary-compatible, it’s important to ensure that only one OpenMP runtime is used throughout the application. Having more than one OpenMP runtime linked to an executable may lead to undefined behavior including incorrect results or crashes. However as long as both the library and the application use the same or compatible compilers there would be no conflicts.
Threading Building Blocks (TBB)
To build oneDNN with TBB support, set ONEDNN_CPU_RUNTIME to TBB :
$ cmake -DONEDNN_CPU_RUNTIME=TBB ..
Optionally, set the TBBROOT environmental variable to point to the TBB installation path or pass the path directly to CMake:
$ cmake -DONEDNN_CPU_RUNTIME=TBB -DTBBROOT=/opt/intel/path/tbb ..
oneDNN has functional limitations if built with TBB:
Winograd convolution algorithm is not supported for fp32 backward by data and backward by weights propagation.
Threadpool
To build oneDNN with support for threadpool threading, set ONEDNN_CPU_RUNTIME to THREADPOOL
$ cmake -DONEDNN_CPU_RUNTIME=THREADPOOL ..
The _ONEDNN_TEST_THREADPOOL_IMPL CMake variable controls which of the three threadpool implementations would be used for testing: STANDALONE, TBB, or EIGEN. The latter two require also passing TBBROOT or Eigen3_DIR paths to CMake. For example:
$ cmake -DONEDNN_CPU_RUNTIME=THREADPOOL -D_ONEDNN_TEST_THREADPOOL_IMPL=EIGEN -DEigen3_DIR=/path/to/eigen/share/eigen3/cmake ..
Threadpool threading support is experimental and has the same limitations as TBB plus more:
As threadpools are attached to streams which are only passed during primitive execution, work decomposition is performed statically at the primitive creation time. At the primitive execution time, the threadpool is responsible for balancing the static decomposition from the previous item across available worker threads.
AArch64 Options
oneDNN includes experimental support for Arm 64-bit Architecture (AArch64). By default, AArch64 builds will use the reference implementations throughout. The following options enable the use of AArch64 optimised implementations for a limited number of operations, provided by AArch64 libraries.
AArch64 build configuration |
CMake Option |
Environment variables |
Dependencies |
|---|---|---|---|
Arm Compute Library based primitives |
ONEDNN_AARCH64_USE_ACL=ON |
ACL_ROOT_DIR=*</path/to/ComputeLibrary>* |
|
Vendor BLAS library support |
ONEDNN_BLAS_VENDOR=ARMPL |
None |
Arm Compute Library
Arm Compute Library is an open-source library for machine learning applications. The development repository is available from mlplatform.org, and releases are also available on GitHub. The ONEDNN_AARCH64_USE_ACL CMake option is used to enable Compute Library integration:
$ cmake -DONEDNN_AARCH64_USE_ACL=ON ..
This assumes that the environment variable ACL_ROOT_DIR is set to the location of Arm Compute Library, which must be downloaded and built independently of oneDNN.
WARNING:
For a debug build of oneDNN it is advisable to specify a Compute Library build which has also been built with debug enabled.
WARNING:
oneDNN only supports builds with Compute Library v23.02.1 or later.
Vendor BLAS libraries
oneDNN can use a standard BLAS library for GEMM operations. The ONEDNN_BLAS_VENDOR build option controls BLAS library selection, and defaults to NONE. For AArch64 builds with GCC, use the Arm Performance Libraries :
$ cmake -DONEDNN_BLAS_VENDOR=ARMPL ..
Additional options available for development/debug purposes. These options are subject to change without notice, see https://github.com/oneapi-src/oneDNN/blob/master/cmake/options.cmake for details.
GPU Options
Intel Processor Graphics is supported by oneDNN GPU engine. GPU engine is disabled in the default build configuration.
Runtimes
To enable GPU support you need to specify the GPU runtime by setting ONEDNN_GPU_RUNTIME CMake option. The default value is "NONE" which corresponds to no GPU support in the library.
OpenCL*
OpenCL runtime requires Intel(R) SDK for OpenCL* applications. You can explicitly specify the path to the SDK using -DOPENCLROOT CMake option.
$ cmake -DONEDNN_GPU_RUNTIME=OCL -DOPENCLROOT=/path/to/opencl/sdk ..
Graph component limitations
The graph component can be enabled via the build option ONEDNN_BUILD_GRAPH. But the build option does not work with some values of other build options. Specifying the options and values simultaneously in one build will lead to a CMake error.
CMake Option |
Unsupported Values |
|---|---|
ONEDNN_GPU_RUNTIME |
OCL |
ONEDNN_GPU_VENDOR |
NVIDIA |
ONEDNN_ENABLE_PRIMITIVE |
PRIMITIVE_NAME |
ONEDNN_ENABLE_WORKLOAD |
INFERENCE |
Graph Compiler Backend Limitations
As a backend of the graph component, besides the options described in Graph component limitations, graph compiler backend has some extra limitations. Specifying unsupported build options will lead to a CMake error.
CMake Option |
Unsupported Values |
|---|---|
ONEDNN_CPU_RUNTIME |
THREADPOOL, SYCL |
ONEDNN_GPU_RUNTIME |
OCL, SYCL |
Besides, the instructions contained in the kernels generated by the graph compiler backend are AVX512_CORE or above, so these kernels will not be dispatched on systems that do not have corresponding instruction sets support.