max_pool3d_forward_f32

void max_pool3d_forward_f32(const float * x, float * y, int * argmax, int B, int C, int D, int H, int W, int KD, int KH, int KW, int OD, int OH, int OW, int sd, int sh, int sw, int pd, int ph, int pw)

source edit

Single-precision 3-D max-pool forward over an NCDHW input.

For every output voxel the kernel scans the corresponding KD × KH × KW window, writes the maximum into y, and stores the flat d*H*W + h*W + w index of the winner into argmax. Pad voxels are treated as -inf.

Shape

$\text{OD} = \lfloor (D + 2\cdot\text{pd} - \text{KD})/\text{sd}\rfloor + 1$ , analogously for $\text{OH}, \text{OW}$ .

Parameters

xconst float*

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

yfloat*

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

argmaxint32_t*

Per-output winning flat 3-D index, shape (B, C, OD, OH, OW).

B, C, D, H, Wint

Layout extents of x.

KD, KH, KWint

Window depth, height, width.

OD, OH, OWint

Output spatial extents.

sd, sh, swint

Window strides along D, H, W.

pd, ph, pwint

Symmetric zero-pads along D, H, W.

void max_pool3d_forward_f32(const float * x, float * y, int * argmax, int B, int C, int D, int H, int W, int KD, int KH, int KW, int OD, int OH, int OW, int sd, int sh, int sw, int pd, int ph, int pw)