Skip to main content

nchw_bias_add_backward

teeny_kernels::nn::tensor::channel_bias_add

Function nchw_bias_add_backward 

Source 
pub fn nchw_bias_add_backward<T: Triton, D: Float, const BLOCK_HW: i32>(
    dy_ptr: T::Pointer<D>,
    dx_ptr: T::Pointer<D>,
    dbias_ptr: T::Pointer<D>,
    C: i32,
    HW: i32,
)
where
    T::I32Tensor: Tensor<i32, 1> + Comparison<i32, BoolTensor = T::BoolTensor>,
    T::Pointer<D>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<D>>>,
Expand description

NCHW bias add backward: dx = dy, dbias[c] = sum over (B, H, W) of dy.

Grid: [C, B] — one CTA per (channel, batch) pair; single while loop over H*W to avoid nested loops (which ICE the teenyc compiler).