Visible to Intel only — GUID: GUID-12777888-CFAF-4C70-B7EA-B3964A30B754
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
HardSwish
HardSwishBackward
Interpolate
InterpolateBackward
LayerNorm
LayerNormBackward
LeakyReLU
Log
LogSoftmax
LogSoftmaxBackward
MatMul
Maximum
MaxPool
MaxPoolBackward
Minimum
Mish
MishBackward
Multiply
PReLU
PReLUBackward
Quantize
Reciprocal
ReduceL1
ReduceL2
ReduceMax
ReduceMean
ReduceMin
ReduceProd
ReduceSum
ReLU
ReLUBackward
Reorder
Round
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-12777888-CFAF-4C70-B7EA-B3964A30B754
DPC++ Interoperability
Overview
oneDNN may use the DPC++ runtime for CPU and GPU engines to interact with the hardware. Users may need to use oneDNN with other code that uses DPC++. For that purpose, the library provides API extensions to interoperate with underlying SYCL objects. This interoperability API is defined in the dnnl_sycl.hpp header.
One of the possible scenarios is executing a SYCL kernel for a custom operation not provided by oneDNN. In this case, the library provides all the necessary API to “seamlessly” submit a kernel, sharing the execution context with oneDNN: using the same device and queue.
The interoperability API is provided for two scenarios:
Construction of oneDNN objects based on existing SYCL objects
Accessing SYCL objects for existing oneDNN objects
The mapping between oneDNN and SYCL objects is provided in the following table:
oneDNN object |
SYCL object(s) |
|---|---|
Engine |
sycl::device and sycl::context |
Stream |
sycl::queue |
Memory (Buffer-based) |
sycl::buffer<uint8_t, 1> |
Memory (USM-based) |
Unified Shared Memory (USM) pointer |
The table below summarizes how to construct oneDNN objects based on SYCL objects and how to query underlying SYCL objects for existing oneDNN objects.
oneDNN object |
API to construct oneDNN object |
API to access SYCL object(s) |
|---|---|---|
Engine |
dnnl::sycl_interop::make_engine(const sycl::device &, const sycl::context &) |
dnnl::sycl_interop::get_device(const engine &)dnnl::sycl_interop::get_context(const engine &) |
Stream |
dnnl::sycl_interop::make_stream(const engine &, sycl::queue &) |
dnnl::sycl_interop::get_queue(const stream &) |
Memory (Buffer-based) |
dnnl::sycl_interop::make_memory(const memory::desc &, const engine &, sycl::buffer<T, ndims> &) |
dnnl::sycl_interop::get_buffer<T, ndims>(const memory &) |
Memory (USM-based) |
NOTE:
Internally, library buffer-based memory objects use 1D uint8_t SYCL buffers; however, the user may initialize and access memory using SYCL buffers of a different type. In this case, buffers will be reinterpreted to the underlying type sycl::buffer<uint8_t, 1>.
SYCL Buffers and DPC++ USM Interfaces for Memory Objects
The memory model in SYCL 1.2.1 is based on SYCL buffers. DPC++ further extends the programming model with a Unified Shared Memory (USM) alternative, which provides the ability to allocate and use memory in a uniform way on host and DPC++ devices.
oneDNN supports both programming models. USM is the default and can be used with the usual oneDNN memory API. The buffer-based programming model requires using the interoperability API.
To construct a oneDNN memory object, use one of the following interfaces:
dnnl::sycl_interop::make_memory(const memory::desc &, const engine &, sycl_interop::memory_kind kind, void *handle)
Constructs a USM-based or buffer-based memory object depending on memory allocation kind kind. The handle could be one of special values DNNL_MEMORY_ALLOCATE or DNNL_MEMORY_NONE, or it could be a user-provided USM pointer. The latter works only when kind is dnnl::sycl_interop::memory_kind::usm.
dnnl::memory(const memory::desc &, const engine &, void *)
Constructs a USM-based memory object. The call is equivalent to calling the function above with with kind equal to dnnl::sycl_interop::memory_kind::usm.
dnnl::sycl_interop::make_memory(const memory::desc &, const engine &, sycl::buffer<T, ndims> &)
Constructs a buffer-based memory object based on a user-provided SYCL buffer.
To identify whether a memory object is USM-based or buffer-based, dnnl::sycl_interop::get_memory_kind() query can be used.
Handling Dependencies with USM
SYCL queues could be in-order or out-of-order. For out-of-order queues, the order of execution is defined by the dependencies between SYCL tasks. The runtime tracks dependencies based on accessors created for SYCL buffers. USM pointers cannot be used to create accessors and users must handle dependencies on their own using SYCL events.
oneDNN provides two mechanisms to handle dependencies when USM memory is used:
dnnl::sycl_interop::execute() interface
This interface enables you to pass dependencies between primitives using SYCL events. In this case, the user is responsible for passing proper dependencies for every primitive execution.
In-order oneDNN stream
oneDNN enables you to create in-order streams when submitted primitives are executed in the order they were submitted. Using in-order streams prevents possible read-before-write or concurrent read/write issues.