nx_pose_cnn_gate.nx
buildroot/runtime/nx_pose_cnn_gate.nx
about
nx_pose_cnn_gate.nx -- executed proof of the sovereign f32 pose-CNN FORWARD PASS end-to-end into the integer
keypoint back-half. A tiny 8x8 single-channel image with impulses runs through a REAL 3x3 f32 conv backbone
(identity weights, so the peak is predictable while the full 9-tap f32 multiply-add path is exercised) + a REAL
1x1 f32 pose head, then the ReLU'd f32 heatmap is fed DIRECTLY (order-isomorphism) into nx_pose_keypoints.
Proves: forward peak + quarter-pixel subpixel through the conv, ReLU zeroing, the f32-bits occlusion threshold,
the ReLU unit behavior, and (C5) that the ReLU is LOAD-BEARING -- integer argmax picks the wrong cell on a raw
negative activation and the right cell only after ReLU. expect_exit: 0
dependencies 4 imports · 0 importers
imports: nx_syscalls.nxnx_f32_cvt.nxnx_pose_cnn.nxnx_pose_keypoints.nx
imported by: nobody (leaf or entry point)
call flow from main pre-order; caps 40 nodes / depth 6 declared; ↻ = already shown
structs
| none |
consts
| none |
functions
| 13 | func gp(s: *u8) -> i64 { var n: i64 = 0; while s[n] != (0 as u8) { n = n + 1 } return sys_write(1, s, n) } |
| 14 | func gn(v: i64) -> i64 |
| 24 | func main(argc: i64, argv: *i64) -> i64 |