Intel® oneAPI Deep Neural Network Developer Guide and Reference

ID 768875
Date 7/13/2023
Public

A newer version of this document is available. Customers should click here to go to the newest version.

Document Table of Contents

struct dnnl::post_ops

Overview

Post-ops. More…

#include <dnnl.hpp>

struct post_ops: public dnnl::handle
{
    // construction

    post_ops();
    post_ops(dnnl_post_ops_t post_ops);

    // methods

    int len() const;
    primitive::kind kind(int index) const;

    void append_sum(
        float scale = 1.f,
        int32_t zero_point = 0,
        memory::data_type data_type = memory::data_type::undef
        );

    void get_params_sum(int index, float& scale) const;
    void get_params_sum(int index, float& scale, memory::data_type& data_type) const;

    void get_params_sum(
        int index,
        float& scale,
        int32_t& zero_point,
        memory::data_type& data_type
        ) const;

    void append_eltwise(algorithm aalgorithm, float alpha, float beta);

    void get_params_eltwise(
        int index,
        algorithm& aalgorithm,
        float& alpha,
        float& beta
        ) const;

    void append_dw(
        memory::data_type weights_data_type,
        memory::data_type bias_data_type,
        memory::data_type dst_data_type,
        memory::dim kernel_size,
        memory::dim stride_size,
        memory::dim padding_l_size
        );

    void get_params_dw(
        int index,
        memory::data_type& weights_data_type,
        memory::data_type& bias_data_type,
        memory::data_type& dst_data_type,
        memory::dim& kernel_size,
        memory::dim& stride_size,
        memory::dim& padding_l_size
        ) const;

    void append_binary(algorithm aalgorithm, const memory::desc& src1_desc);

    void get_params_binary(
        int index,
        algorithm& aalgorithm,
        memory::desc& src1_desc
        ) const;

    void append_prelu(int mask);
    void get_params_prelu(int index, int& mask) const;
};

Inherited Members

public:
    // methods

    handle<T, traits>& operator = (const handle<T, traits>&);
    handle<T, traits>& operator = (handle<T, traits>&&);
    void reset(T t, bool weak = false);
    T get(bool allow_empty = false) const;
    operator T () const;
    operator bool () const;
    bool operator == (const handle<T, traits>& other) const;
    bool operator != (const handle& other) const;

Detailed Documentation

Post-ops.

Post-ops are computations executed after the main primitive computations and are attached to the primitive via primitive attributes.

See also:

Primitive Attributes: Post-ops

Construction

post_ops()

Constructs an empty sequence of post-ops.

post_ops(dnnl_post_ops_t post_ops)

Creates post-ops primitive attribute from a C API dnnl_post_ops_t handle.

The resulting handle is not weak and the C handle will be destroyed during the destruction of the C++ object.

Parameters:

post_ops

The C API post-ops primitive attribute.

Methods

int len() const

Returns the number of post-ops entries.

primitive::kind kind(int index) const

Returns the primitive kind of post-op at entry with a certain index.

Parameters:

index

Index of the post-op to return the kind for.

Returns:

Primitive kind of the post-op at the specified index.

void append_sum(
    float scale = 1.f,
    int32_t zero_point = 0,
    memory::data_type data_type = memory::data_type::undef
    )

Appends an accumulation (sum) post-op.

Prior to accumulating the result, the previous value will be will be reduced by zero point zero_point and multiplied by a scaling factor scale.

The kind of this post-op is dnnl::primitive::kind::sum.

This feature may improve performance for cases like dequantize the asymmetrically quantized sum’s src1 tensor to f32 domain before performing the sum operation by subtracting zero_point before the scaling.

In the simplest case when the accumulation is the only post-op, the computations will be dst[:] := scale * (dst[:] - zero_point) + op(...) instead of dst[:] := op(...).

If data_type is specified, the original dst tensor will be reinterpreted as a tensor with the provided data type. Because it is a reinterpretation, data_type and dst data type should have the same size. As a result, computations will be dst[:] <- scale * (as_data_type(dst[:]) - zero_point) + op(...) instead of dst[:] <- op(...).

NOTE:
This post-op executes in-place and does not change the destination layout.

Parameters:

scale

Scaling factor.

zero_point

Zero point.

data_type

Data type.

void get_params_sum(int index, float& scale) const

Returns the parameters of an accumulation (sum) post-op.

Parameters:

index

Index of the sum post-op.

scale

Scaling factor of the sum post-op.

void get_params_sum(int index, float& scale, memory::data_type& data_type) const

Returns the parameters of an accumulation (sum) post-op.

Parameters:

index

Index of the sum post-op.

scale

Scaling factor of the sum post-op.

data_type

Data type of the sum post-op.

void get_params_sum(
    int index,
    float& scale,
    int32_t& zero_point,
    memory::data_type& data_type
    ) const

Returns the parameters of an accumulation (sum) post-op.

Parameters:

index

Index of the sum post-op.

scale

Scaling factor of the sum post-op.

zero_point

Single scalar int32_t value of zeropoint.

data_type

Data type of the sum post-op.

void append_eltwise(algorithm aalgorithm, float alpha, float beta)

Appends an elementwise post-op.

The kind of this post-op is dnnl::primitive::kind::eltwise.

In the simplest case when the elementwise is the only post-op, the computations would be dst[:] := eltwise_op (op(...)) instead of dst[:] <- op(...), where eltwise_op is configured with the given parameters.

Parameters:

aalgorithm

Elementwise algorithm.

alpha

Alpha parameter for the elementwise algorithm.

beta

Beta parameter for the elementwise algorithm.

void get_params_eltwise(
    int index,
    algorithm& aalgorithm,
    float& alpha,
    float& beta
    ) const

Returns parameters of an elementwise post-op.

Parameters:

index

Index of the post-op.

aalgorithm

Output elementwise algorithm kind.

alpha

Output alpha parameter for the elementwise algorithm.

beta

Output beta parameter for the elementwise algorithm.

void append_dw(
    memory::data_type weights_data_type,
    memory::data_type bias_data_type,
    memory::data_type dst_data_type,
    memory::dim kernel_size,
    memory::dim stride_size,
    memory::dim padding_l_size
    )

Appends a depthwise post-op convolution.

This post-op can only be fused with a 2D 1x1 convolution (convolution with weights spatial dimension equal to 1 i.e., kh=kw=1).

The kind of this post-op is dnnl_convolution.

The number of outputs for primitive remain same as before. The output spatial size can be derived as below:

output_height = ceil(output_height_1x1_convolution, stride) output_width = ceil(output_width_1x1_convolution, stride)

See dev_guide_attributes_post_ops_depthwise and dev_guide_attributes_post_ops_depthwise_fusion for more info.

Parameters:

weights_data_type

Weights data type of depthwise post-op

bias_data_type

Bias data type of depthwise post-op

dst_data_type

Output data type of depthwise post-op

kernel_size

Size of kernel of depthwise post-op

stride_size

Size of stride of depthwise post-op

padding_l_size

Size of left and top paddings of depthwise post-op

void get_params_dw(
    int index,
    memory::data_type& weights_data_type,
    memory::data_type& bias_data_type,
    memory::data_type& dst_data_type,
    memory::dim& kernel_size,
    memory::dim& stride_size,
    memory::dim& padding_l_size
    ) const

Returns the parameters of an depthwise post-op.

Parameters:

index

Index of the elementwise post-op.

weights_data_type

Weights data type of depthwise post-op

bias_data_type

Bias data type of depthwise post-op

dst_data_type

Output data type of depthwise post-op

kernel_size

Size of kernel of depthwise post-op

stride_size

Size of stride of depthwise post-op

padding_l_size

Size of left and top paddings of depthwise post-op

void append_binary(algorithm aalgorithm, const memory::desc& src1_desc)

Appends a binary post-op.

The kind of this post operation is dnnl_binary.

In the simplest case when the binary is the only post operation, the computations would be:

dst[:] <- binary_op (dst[:], another_input[:])

where binary_op is configured with the given parameters. binary_op supports broadcast semantics for a second operand.

Parameters:

aalgorithm

Binary algorithm for the post-op.

src1_desc

Memory descriptor of a second operand.

void get_params_binary(
    int index,
    algorithm& aalgorithm,
    memory::desc& src1_desc
    ) const

Returns the parameters of a binary post-op.

Parameters:

index

Index of the binary post-op.

aalgorithm

Output binary algorithm kind.

src1_desc

Output memory descriptor of a second operand.

void append_prelu(int mask)

Appends a prelu forward post-op.

The kind of this post-op is dnnl::primitive::kind::prelu.

The post-op can be defined as:

dst[:] <- prelu(dst[:], weights[:])
prelu:
dst[:] <- dst[:] if dst[:] > 0
dst[:] <- dst[:] * weights[:] if dst[:] <= 0

Example usage:

int mb = 32, oc = 32,
    oh = 14, ow = 14; // convolution output params
// unique weights per output channel
vector<float> weights = { ... };
int oc_dim = 1; // mb_dim = 0, channel_dim = 1, height_dim = 2, ...

// construct a convolution descriptor
dnnl::convolution::desc conv_d;

dnnl::primitive_attr attr;
attr.append_prelu(1 << oc_dim);

dnnl::primitive_desc conv_pd(conv_d, attr, engine);
memory prelu_weights({{1}, dt::f32, {1}}, eng, weights.data());

std::unordered_map<int, memory> conv_args;

conv_args.insert(
 {DNNL_ARG_ATTR_MULTIPLE_POST_OP(0) | DNNL_ARG_WEIGHTS, prelu_weights})
NOTE:
The order of dimensions does not depend on how elements are laid out in memory. For example:
  • for a 2D CNN activations tensor the order is always (n, c)

  • for a 4D CNN activations tensor the order is always (n, c, h, w)

  • for a 5D CNN weights tensor the order is always (g, oc, ic, kh, kw)

Prelu weights tensor is passed in runtime execution phase. Prelu weights tensor data type is implicitly assumed as f32 using plain layout (a, ab, acb, acdb, acdeb).

Parameters:

mask

Defines the correspondence between the output tensor dimensions and the prelu weights tensor. The set i-th bit indicates that a dedicated weights value is used for each index along that dimension. Set the mask to 0 to use a common weights value for the whole output tensor.

void get_params_prelu(int index, int& mask) const

Returns the parameters of a prelu post-op.

Parameters:

index

Index of the prelu post-op.

mask

Weights mask of prelu post-op.