Parametric ReLU (PReLU) Prototype and Function List¶

Description¶

This kernel performs Parametric Rectified Linear Unit (PReLU) with a negative slope activation function. It transforms each element of input tensor according to the following formula:

\[\begin{split}y_{i} = \Big\{ { \begin{matrix} x_{i}\text{ if }x_{i} \geq 0 \\ {\alpha}*x_{i}\text{ if }x_{i} < 0 \\ \end{matrix}}\end{split}\]

Where:

\(x_{i}\) - \(i_{\text{th}}\) value in input data subset

\(y_{i}\) - \(i_{\text{th}}\) value in output data subset

\(\alpha\) - coefficient of the negative slope for the specific data subset

While for Leaky ReLU, the whole tensor shares only the \(\alpha\) coefficient, for PRelu an array of slope coefficients is shared across an axis. Hence, for each slice along the specified axis an individual \(\alpha\) slope coefficient is used.

The “shared axis” feature found in some frameworks is not supported in MLI. This functionality can instead be achieved in several iterations using the PReLU kernel and the mem_strides feature. One iteration implies creating subtensors from in and slope_coeff tensors using memstrides and applying the PReLU kernel on them.

Functions¶

Kernels which implement Parametric ReLU functions have the following prototype:

mli_status mli_krn_prelu_<data_format>(
   const mli_tensor  *in,
   const mli_tensor  *slope_coeff,
   const mli_prelu_cfg  *cfg,
   mli_tensor  *out);

where data_format is one of the data formats listed in Table MLI Data Formats and the function parameters are shown in the following table:

Parametric ReLU Function Parameters¶
Parameter	Type	Description
`in`	`mli_tensor *`	[IN] Pointer to constant input tensor.
`slope_coeff`	`mli_tensor *`	[IN] Pointer to tensor with negative slope coefficients.
`cfg`	`mli_prelu_cfg *`	[IN] Pointer to PReLU parameters structure.
`out`	`mli_tensor *`	[IN \| OUT] Pointer to output tensor. Result is stored here

mli_prelu_cfg is defined as:

typedef struct {
    int32_t axis;
} mli_prelu_cfg;

mli_prelu_cfg Structure Field Description¶
Field Name	Type	Description
`axis`	`int32_t`	An axis along which the function is computed. Axis corresponds to index of tensor’s dimension starting from 0. For instance, having feature map in HWC layout, axis == 0 corresponds to H dimension. If axis < 0, the function is applied to the whole tensor.

List of Available PReLU Functions¶
Function Name	Details
`mli_krn_prelu_sa8`	All tensors data format: sa8
`mli_krn_prelu_fx16`	All tensors data format: fx16

Conditions¶

Ensure that you satisfy the following general conditions before calling the function:

in, out and slope_coeff tensors must be valid (see mli_tensor Structure Field Descriptions).

in and out tensors must be of the same shape.

slope_coeff tensor must satisfy the following shape requirements depending on the axis parameter of cfg structure:

axis < 0 : slope_coeff tensor must be a valid tensor-scalar (see data field description in the Table mli_tensor Structure Field Descriptions).

axis >= 0 : slope_coeff is a one-dimensional tensor (rank==1). Its length must be equal to axis dimension of in tensor (e.g. in.shape[cfg.axis]).

axis parameter of cfg structure can be negative and must be less than in tensor rank.

mem_stride of the innermost dimension must be equal to 1 for all the tensors.

For fx16 versions of kernel, in addition to general conditions, ensure that you satisfy the following quantization conditions before calling the function:

The number of frac_bits in the in and out tensors must be equal.

For sa8 versions of kernel, in addition to general conditions, ensure that you satisfy the following quantization conditions before calling the function:

in, out and slope_coeff tensors must be quantized on the tensor level. This implies that the tensor contains a single scale factor and a single zero offset.

Zero offset of in and out tensors must be within [-128, 127] range.

Zero offset of slope_coeffs tensor must be within [-16384, 16383] range.

Ensure that you satisfy the platform-specific conditions in addition to those listed above (see the Platform Specific Details chapter).

Result¶

These functions only modify the memory pointed by out.data.mem field. It is assumed that all the other fields of out tensor are properly populated to be used in calculations and are not modified by the kernel.

The kernel supports in-place computation. It means that out and in tensor structures can point to the same memory with the same memory strides but without shift. It can affect performance for some platforms.

Warning

Only an exact overlap of starting address and memory stride of the in and out tensors is acceptable. Partial overlaps result in undefined behavior.

Depending on the debug level (see section Error Codes) this function performs a parameter check and returns the result as an mli_status code as described in section Kernel Specific Configuration Structures.