Skip to main content

conv2d_forward

Function conv2d_forward 

Source
pub fn conv2d_forward<T: Triton, D: Num, const KH: i32, const KW: i32, const STRIDE_H: i32, const STRIDE_W: i32, const PAD_H: i32, const PAD_W: i32, const G: i32, const BLOCK_OW: i32>(
    x_ptr: T::Pointer<D>,
    w_ptr: T::Pointer<D>,
    y_ptr: T::Pointer<D>,
    _B: i32,
    C_IN: i32,
    C_OUT: i32,
    H: i32,
    W: i32,
    OH: i32,
    OW: i32,
)
where T::I32Tensor: Tensor<i32, 1> + Comparison<i32, BoolTensor = T::BoolTensor>, T::BoolTensor: BitAnd<Output = T::BoolTensor>, T::Pointer<D>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<D>>>,
Expand description

2-D convolution forward pass (supports grouped and depthwise conv).

Grid: one CTA per (b, c_out, oh, ow-tile): pid = ((b * C_OUT + c_out) * OH + oh) * num_ow_tiles + ow_tile

Each CTA computes a BLOCK_OW-wide strip of output columns by iterating over (C_IN/G) * KH * KW combinations for its assigned group. The weight for each (c_in_local, kh, kw) is loaded as a [1] tensor and broadcast to [BLOCK_OW].

G is the number of groups (1 = standard conv, G = C_IN = C_OUT for depthwise). Weight layout: [C_OUT, C_IN/G, KH, KW].

Zero-padding of PAD_H / PAD_W elements is applied on each spatial side. OH = (H + 2*PAD_H - KH) / STRIDE_H + 1, OW = (W + 2*PAD_W - KW) / STRIDE_W + 1.