Element-wise Kernels Group¶

Element-wise Operations Prototype and Function List¶

Description¶

The Element-wise Group describes operations that are applied element-by-element on two tensors of the same shape and return a tensor of the same shape. These kernels can also be used for broadcasting a scalar value. One of the input tensors can be a scalar tensor. In that case the operation is applied to the scalar value and each element of the other tensor.

\(\text{out}_{i} = operation(\text{in}_{i}^{1},\ \text{in}_{i}^{2}\))

\(\text{out}_{i} = operation(\text{in}_{\text{scalar}}^{1},\ \text{in}_{i}^{2}\))

Functions¶

Kernels which implement Element-wise functions have the following prototype:

mli_status mli_krn_eltwise_<operation>_<datatype> (
   const mli_tensor *in1,
   const mli_tensor *in2,
   mli_tensor *out);

Element-wise Group Function Parameters¶
Parameter	Type	Description
`in1`	`mli_tensor *`	[IN] Pointer to constant input tensor.
`in2`	`mli_tensor *`	[IN] Pointer to constant input tensor.
`out`	`mli_tensor *`	[IN \| OUT] Pointer to output tensor. Result is stored here.

List of Available Element-Wise Functions¶
Function Name	Operation	in1 / in2 / out data format
`mli_krn_eltwise_add_sa8`	Addition	sa8
`mli_krn_eltwise_add_fx16`	Addition	fx16
`mli_krn_eltwise_sub_sa8`	Subtract	sa8
`mli_krn_eltwise_sub_fx16`	Subtract	fx16
`mli_krn_eltwise_min_sa8`	Minimum	sa8
`mli_krn_eltwise_min_fx16`	Minimum	fx16
`mli_krn_eltwise_max_sa8`	Maximum	sa8
`mli_krn_eltwise_max_fx16`	Maximum	fx16
`mli_krn_eltwise_mul_sa8`	Multiply	sa8
`mli_krn_eltwise_mul_fx16`	Multiply	fx16

Conditions¶

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

in1, in2 and out tensors must be valid (see mli_tensor Structure Field Descriptions) and satisfy data requirements of the specific version of the kernel.

Shapes of in1, in2 and out tensors must be compatible, which implies the following requirements:

in1 and in2 tensors must be of the same shape, or one of them can be a tensor-scalar (see data field description in the Table mli_tensor Structure Field Descriptions)

out tensors must be of the same shape as a non-scalar input tensor.

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

For mli_krn_eltwise_min_* and mli_krn_eltwise_max_* functions, the following additional restriction apply

in1, in2 tensors must have the same quantization parameters. It means that el_params union of tensors must be the same. For other elementwise functions this restriction is not applicable.

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

in1, in2 and out tensors must be quantized on the tensor level. This implies that each tensor contains a single scale factor and a single zero offset.

Zero offset of in1, in2 and out tensors 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 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.

If the result of an operation is out of container’s range, it is saturated to the container’s limit.

The kernel supports in-place computation. It means that output and input 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 input and output 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.