engine: refuse any blob the kernels cannot run safely

[widgetii]

From the review of majestic's copy of this engine (widgetii/majestic#1030): the model is a file on the camera, so a malformed one must be refused when it is opened, not trusted into a kernel.

parse() now requires:

• an even input size, because NV21 carries chroma in pairs;
• int8 zero points throughout;
• a 1×1 convolution with stride 1 and no padding, as its kernel assumes, and padding below the kernel size for the others;
• requantization shifts in range for the add as well as the convolution;
• every op to read a tensor an earlier op produced as int8, and to write a new one;
• float detection heads on matching grids;
• all activations together within 8 MiB, counted in 64 bits.

The decode bounds each box distance before it becomes an int. The lab-only profiling code is removed.

Unchanged outputs: a full scripts/build.sh produces the same person.pdnet and person.json (identical SHA256SUMS) and the same points.md. The host and ARMv7 NEON builds are still bit-identical to the reference, and the ARM build finds exactly the host's boxes on all 3000 crops. majestic's tests/detect/pdnet.c exercises each rejection.

[qodo-free-for-open-source-projects]

PR Summary by Qodo

Reject unsafe model blobs before running detector kernels

🐞 Bug fix 🕐 20-40 Minutes

AI Description

• Reject malformed models at open time when tensor types, operation parameters, or activation sizes
 are unsafe.
• Bound decoded box distances before integer conversion and remove lab-only profiling code.
• Reported build checks show unchanged release artifacts and matching host and ARM detections.

Diagram

graph TD
  Blob["Model blob"] --> Parse{"Parse valid?"} --> Buffers["Typed buffers"] --> Kernels["Int8 kernels"] --> Heads["Float heads"] --> Decode["Bounded decode"] --> Boxes["Person boxes"]
  Parse --> Reject["Reject blob"]

Loading

High-Level Assessment

Keep validation in parse(): it rejects unsafe files before allocation or kernel execution, while tracking tensor kinds in operation order matches how the engine runs. Runtime checks would defer failures until detection; a buffer-liveness planner would add complexity without a clear need under the 8 MiB activation cap.

Files changed (1) +67 / -61

Bug fix (1) +67 / -61

pdnet.cValidate model execution safety at open time +67/-61

Validate model execution safety at open time

• The parser now checks even NV21 dimensions, int8 zero points, convolution constraints, requantization shifts, tensor production order and types, matching float detection heads, and an 8 MiB aggregate activation limit. Allocation uses the validated tensor type; decoding bounds box distances before integer conversion. Lab-only per-operation profiling code is removed.

engine/pdnet.c

[qodo-free-for-open-source-projects]

Code Review by Qodo

🐞 Bugs (0) 📘 Rule violations (0) 📎 Requirement gaps (0) 🎨 UX issues (0) 🔗 Cross-repo conflicts (0) 📜 Skill insights (0)

1. Large box distances change detections ✓ Resolved

Description

pdnet_detect() caps every regression distance at 1024 stride units before calculating box edges,
even when the original distance would produce valid integer coordinates. A finite head output of
1025 at stride 8, for example, now yields a smaller box, which can change overlap matching and the
detections used to score presets.

Code

engine/pdnet.c[811]

+                    r[s4] = fminf(fmaxf(0.0f, tr->f[(size_t)i * 4 + s4]), 1024.0f);

Evidence

The new clamp feeds all four box-edge calculations; the previous calculations used each nonnegative
distance without an upper cap. The preset evaluation matches emitted boxes by
intersection-over-union, so a changed box can change its classification.

engine/pdnet.c[803-819]
tools/points.py[27-52]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
The fixed 1024-unit cap changes safely representable decoded boxes, rather than only preventing unsafe integer conversions.
## Fix Focus Areas
- engine/pdnet.c[806-819]
## Recommended Fix
Bound decoded coordinates according to the integer range and the subsequent box-width calculations, preserving the existing result whenever those calculations are safe.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---

Tip of the day

💡 Did you know, you can tweak Display settings with a live preview to see your comment before it ships

More tips ↗ | Customize Qodo ↗ | Qodo docs ↗