device-code-proof

Prove an operator refactor changed nothing in the device code, without a build tree and without a GPU, by compiling the changed translation unit before and after and classifying the SASS/PTX difference. Use when asked to show a refactor is bit-exact, to satisfy VER-6 / NEEDS-LOCAL-PROOF from `tools/refactor_op.py --phase verify`, or to explain why codegen changed. Produces a classified per-kernel report; a non-empty diff is not a failure and an unexplained one is not a pass.

Jump to install Review files

Repository: CVCUDA/CV-CUDA
Last source activity: October 20, 2026 at 16:14
Detected SKILL.md language: English
Stars: 2,731
Forks: 267

Install options

The review-first prompt is selected by default. You can switch to a direct command or download a local copy.

Install with an AI assistant

Copy this prompt into the AI assistant you're using.

Follow https://skillsmp.com/skill-install/prompt.md to install only the Agent Skill at https://github.com/CVCUDA/CV-CUDA/tree/main/.agents/skills/op-tools/skills/device-code-proof.

A direct command skips the review prompt. Inspect the source before running it.

npx skills add https://github.com/CVCUDA/CV-CUDA --skill device-code-proof

The command stays on one line. Scroll horizontally to inspect it before copying.

Prefer a local copy? Download the files currently available to SkillsMP.

Review the source files

Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.

SKILL.md

Source instructions · Read-only preview

name: device-code-proof
description: Prove an operator refactor changed nothing in the device code, without a build tree and without a GPU, by compiling the changed translation unit before and after and classifying the SASS/PTX difference. Use when asked to show a refactor is bit-exact, to satisfy VER-6 / NEEDS-LOCAL-PROOF from `tools/refactor_op.py --phase verify`, or to explain why codegen changed. Produces a classified per-kernel report; a non-empty diff is not a failure and an unexplained one is not a pass.

[//]: # "SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved." [//]: # "SPDX-License-Identifier: Apache-2.0" # Device Code Proof Nested skill of the **op-tools** bundle (`.agents/skills/op-tools/`). Follow the shared operating contract in `.agents/skills/op-tools/reference/conventions.md`; the full procedure for this command is `.agents/skills/op-tools/reference/device-code-proof.md`. - **Run:** `python3 tools/device_code_proof.py <Operator> --base <ref>` (or `--file <TU>`; `--stage sass` for the strongest nvcc-only tier). - **Interpret per:** the VER-6 leg of `.agents/guidance/REFACTOR_OP_GUIDELINES.md` — evidence for the leg, never a replacement for the frozen test run. **Delegation rule.** At top level, delegate the full procedure to the `device-code-proof-worker` sub-agent (`.agents/agents/device-code-proof-worker.md`) and read back its verdict; when you are already inside a delegated sub-agent run, or the task is a single checker invocation, run inline. The worker is an optional accelerator — `.agents/agents/` is internal-only and the release trim removes it, so the reference doc above remains the complete procedure.

View on GitHub

Back to search Back to installation

op-tools

Hub of the CV-CUDA op-tools bundle — the operator-lifecycle command families (make-op, review-op, optimize-op, refactor-op, device-code-proof) and their shared operating contract. Use when asked which op command applies to a task or how the op tooling fits together; the per-command skills remain the entry points for the work itself.

2026-10-20

make-op-scaffold

Scaffold a new CV-CUDA operator — a complete, wired, building skeleton — and delegate the implementation to a human or another AI. Use when asked to set up / stub out a new operator without implementing it, or to bootstrap one for someone else to finish. Supports --bare (skip the spec when the definition itself is delegated).

2026-10-20

make-op

Add a new CV-CUDA operator end-to-end per .agents/guidance/MAKE_OP_GUIDELINES.md, with a deterministically-enforced definition-of-done. Use when asked to create/add a new operator, scaffold one, or verify that a new operator is complete (approved spec, wired scaffold, bit-exact regression tests across the declared support matrix, required layout parity, complement negatives, docs + relnote, benched dtypes, tests that run and pass, optimization-ready).

2026-10-20

View all from this repository

device-code-proof

Source facts

Install options

Install with an AI assistant

Review the source files

More from this repository

Install with an AI assistant

More from this repository