Skip to main content

default_compute_for

atomr_accel_cuda::kernel::tensor::contract

Function default_compute_for 

Source 
pub fn default_compute_for<T: TensorSupported>() -> ComputeDesc
Expand description

Pick the canonical compute descriptor for T. Mirrors NVIDIA’s guidance: f32 inputs default to MIN_32F, f64 to MIN_64F, half/bf16 accumulate in f32 (MIN_32F).