max_pool2d_forward_f32

void max_pool2d_forward_f32(const float * x, float * y, int * argmax, int B, int C, int H, int W, int KH, int KW, int OH, int OW, int stride_h, int stride_w, int pad_h, int pad_w)

source edit

Single-precision 2-D max-pool forward over an NCHW input.

For every output cell the kernel scans the corresponding KH × KW window and writes both the maximum into y and the flat H*W index of the winning input position into argmax. Pad positions contribute -inf (an all-pad window yields the first pad slot as the argmax).

Shape

$\text{OH} = \lfloor (H + 2\cdot\text{pad\_h} - \text{KH})/\text{stride\_h}\rfloor + 1$ , analogously for $\text{OW}$ .

Parameters

xconst float*

Input tensor of shape (B, C, H, W).

yfloat*

Output tensor of shape (B, C, OH, OW).

argmaxint32_t*

Per-output winning input position, encoded as h*W + w within the 2-D spatial slice; shape (B, C, OH, OW).

B, C, H, Wint

Layout extents of x.

KH, KWint

Window height and width.

OH, OWint

Output spatial extents.

stride_h, stride_wint

Window stride along H and W.

pad_h, pad_wint

Symmetric zero-pad on each spatial axis.