Sigmoid Prototype and Function List

Description

This kernel performs sigmoid (also called as logistic) activation function on input tensor element-wise and stores the result to the output tensor.

\[y_{i} = \frac{1}{1 + e^{{- x}_{i}}}\]

Where:

\(x_{i}\) \(i_{\text{th}}\) value in input tensor

\(y_{i}\) \(i_{\text{th}}\) value in output tensor

This kernel uses a look-up table (LUTs) to perform data transformation. See Look-Up Tables (LUT) Manipulation Prototypes and Function List section and the pseudo-code sample for more details on LUT structure preparation. Use the following functions for the purpose:

  • mli_krn_sigm_get_lut_size

  • mli_krn_sigm_create_lut

Functions

Kernels which implement Sigmoid functions have the following prototype:

mli_status mli_krn_sigm_<data_format>(
   const mli_tensor  *in,
   const mli_lut *lut,
   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:

Sigmoid Function Parameters

Parameter

Type

Description

in

mli_tensor *

[IN] Pointer to constant input tensor.

lut

mli_lut *

[IN] Pointer to a valid LUT table structure prepared for sigmoid activation.

out

mli_tensor *

[IN | OUT] Pointer to output tensor. Result is stored here

List of Available Sigmoid Functions

Function Name

Details

mli_krn_sigm_sa8

All tensors data format: sa8

mli_krn_sigm_fx16

All tensors data format: fx16

Conditions

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

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

  • in tensor 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 tensor must be within [-128, 127] range.

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

Result

These functions modify:

  • Memory pointed by out.data.mem field.

  • el_params field of out tensor.

It is assumed that all the other fields and structures are properly populated to be used in calculations and are not modified by the kernel.

The range of this function is (0, 1). Depending on the data type, quantization parameters of the output tensor are configured in the following way:

  • fx16

    • out.el_params.fx.frac_bits is set to 15. Hence, the maximum representable value of sigmoid is equivalent to 0.999969482421875 (not 1.0).

  • sa8

    • out.el_params.sa.zero_point.mem.i16 is set to -128

    • out.el_params.sa.scale.mem.i16 is set to 1

    • out.el_params.sa.scale_frac_bits.mem.i8 is set to 8

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.