Struct ElementwiseTemplate 

pub struct ElementwiseTemplate {
    pub op: ElementwiseOp,
    pub precision: PtxType,
    pub target: SmVersion,
}

Template for generating elementwise PTX kernels.

Combines an ElementwiseOp, a precision (PtxType), and a target architecture (SmVersion) to produce a complete PTX module string.

The generated kernel handles global thread indexing and bounds checking. For complex activations (GELU, sigmoid, SiLU), the template emits approximate PTX instruction sequences using ex2.approx and rcp.approx.

Fields

op: ElementwiseOp

The elementwise operation to generate.

precision: PtxType

The data precision for computation (e.g., PtxType::F32).

target: SmVersion

The target GPU architecture.

Implementations

impl ElementwiseTemplate

pub const fn new( op: ElementwiseOp, precision: PtxType, target: SmVersion, ) -> Self

Creates a new elementwise template with the given parameters.

pub fn kernel_name(&self) -> String

Returns the kernel function name derived from the operation and precision.

The name follows the pattern elementwise_{op}_{type}, for example elementwise_add_f32 or elementwise_relu_f16.

pub fn generate(&self) -> Result<String, PtxGenError>

Generates the complete PTX module text for this elementwise operation.

Errors

Returns PtxGenError if the precision type is unsupported for the requested operation or if PTX text generation fails.