Skip to main content

psa_extract_v_nchw

vision_rs::models::yolo::kernels::attention::psa

Function psa_extract_v_nchw 

Source 
pub fn psa_extract_v_nchw<T: Triton, D: Float, const KEY_DIM: i32>(
    qkv_ptr: T::Pointer<D>,
    v_ptr: T::Pointer<D>,
    qkv_h: i32,
    c: i32,
    H: i32,
    W: i32,
    num_heads: 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

Extracts V channels from QKV NCHW into V NCHW.

Input: qkv_ptr — [B, qkv_h, H, W] NCHW Output: v_ptr — [B, c, H, W] NCHW (c = num_heads * 2 * KEY_DIM)

Grid: [BH * N, 1, 1]. Block: [KEY_DIM, 1, 1].