Move SASS diff description to a SKILL by bernhardmgruber · Pull Request #9767 · NVIDIA/cccl

bernhardmgruber · 2026-07-09T12:10:56Z

Fixes: #7829

coderabbitai · 2026-07-09T12:17:50Z

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 2636a2b9-64d8-490f-849b-d62a98bcba97

📥 Commits

Reviewing files that changed from the base of the PR and between f956a5f and 9a3165a.

📒 Files selected for processing (3)

.agent/skills/sass-diff/SKILL.md
.claude/skills/sass-diff/SKILL.md
AGENTS.md

💤 Files with no reviewable changes (1)

AGENTS.md

✅ Files skipped from review due to trivial changes (2)

.claude/skills/sass-diff/SKILL.md
.agent/skills/sass-diff/SKILL.md

📝 Walkthrough

Summary by CodeRabbit

Documentation
- Added a new guide for comparing CUDA SASS/PTX outputs, including recommended tooling, required inputs, and rules to ignore common noise (e.g., addresses/build metadata/timestamps).
- Updated the skills documentation to reference the shared guide for SASS diffs.
- Removed outdated SASS-diff instructions from the main agent documentation so guidance is centralized.

Walkthrough

Adds a new skill file for SASS/PTX diff checks, links it from the Claude skills path, and removes the old SASS Diffs section from AGENTS.md.

Changes

SASS diff skill relocation

Layer / File(s)	Summary
New sass-diff skill spec and symlink, AGENTS.md cleanup `.agent/skills/sass-diff/SKILL.md`, `.claude/skills/sass-diff/SKILL.md`, `AGENTS.md`	Adds a new skill spec covering inputs, normalization, comparison, and reporting rules for SASS/PTX diffs; creates a symlink in the Claude skills directory pointing to the new file; removes the equivalent SASS Diffs section from AGENTS.md.

Assessment against linked issues

Objective	Addressed	Explanation
Move SASS diff instructions from AGENTS.md into a SKILLS.md-style skill file [`#7829`]	✅
Expose the instructions as a separately addressable skill [`#7829`]	✅

_{Comment @coderabbitai help to get the list of available commands.}

coderabbitai

Actionable comments posted: 2

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 80db501d-a267-462f-ad53-e71ae893129b

📥 Commits

Reviewing files that changed from the base of the PR and between f51b10a and f956a5f.

📒 Files selected for processing (3)

.agent/skills/sass-diff/SKILL.md
.claude/skills/sass-diff/SKILL.md
AGENTS.md

💤 Files with no reviewable changes (1)

AGENTS.md

coderabbitai · 2026-07-09T12:17:53Z

+## Inputs to establish
+
+* Compiled binary under test
+* The CUDA SM architectures to compile for. Try to detect this from the code and offer the user a list of suggestions.
+  The user must conform or provide this list.
+* Baseline disassembly (from the previous commit/branch or the current commit without the changes in the working copy).
+* Comparison disassembly (form the current commit/branch or the current commit with the changes in the working copy).
+* By default, prefer `cuobjdump -sass` to inspect SASS changes.
+  Use `cuobjdump -ptx` if the request is to check for PTX changes instead.
+


🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

important: This still starts from prebuilt disassemblies. The issue requires the skill to take baseline and modified CUDA sources plus a target SM, compile both with the same nvcc flags, and then diff the cuobjdump output; otherwise agents can compare non-equivalent artifacts.

🧰 Tools

🪛 LanguageTool

[grammar] ~21-~21: Ensure spelling is correct
Context: ...orking copy). * Comparison disassembly (form the current commit/branch or the curren...

(QB_NEW_EN_ORTHOGRAPHY_ERROR_IDS_1)

coderabbitai · 2026-07-09T12:17:53Z

+## Reporting
+
+* If any non-trivial change was detected, the top 5 regions where a non-trivial change was detected,
+  including the name of the kernel they appeared in.
+* A short summary of the diff type (opcode change, memory access size change, size delta, control-flow, etc.).
+* Explicitly state if only noise was detected after normalization.
+* If you are not sure if the differences are impactful, show it and ask the user for guidance.
+* Keep the disassembly dumps available for reference and show the command to the user to generate a diff.


🚀 Performance & Scalability | 🟠 Major | ⚡ Quick win

important: The reporting section is missing the required performance regressions. Add explicit register-count and occupancy deltas to the summary so those regressions are surfaced instead of normalized away.

Fixes: NVIDIA#7829

Jacobfaib · 2026-07-09T12:49:11Z

Note #9750 symlinks the whole folder (and is just waiting on fixes in #9754 to merge)

Jacobfaib · 2026-07-09T12:51:26Z

+Apply these transforms to both baseline and candidate listings before diffing.
+Write the normalized listings to separate files.
+
+* Remove addresses/offsets/hex location prefixes.
+* Remove build IDs, timestamps, absolute paths, temp directories, and compiler banners.
+* Normalize whitespace and alignment to single spaces.
+* Remove empty lines and purely comment lines.


IME these skills work much better if the agent has a set of scripts they can run that do these highly deterministic things for them. You should ask the bot to generate such a script for you and check it in with the skill.

+1. Otherwise, the agent needs to write the script every time

Jacobfaib · 2026-07-09T12:52:28Z

+* The CUDA SM architectures to compile for. Try to detect this from the code and offer the user a list of suggestions.
+  The user must confirm or provide this list.


I would note that the bot should also check the compile_commands.json to get a definitive answer for this. The architectures will be listed explicitly in the compiler flags.

sleeepyjack · 2026-07-09T14:17:10Z

+## Normalization rules (strip known noise)
+
+Apply these transforms to both baseline and candidate listings before diffing.
+Write the normalized listings to separate files.
+
+* Remove addresses/offsets/hex location prefixes.
+* Remove build IDs, timestamps, absolute paths, temp directories, and compiler banners.
+* Normalize whitespace and alignment to single spaces.
+* Remove empty lines and purely comment lines.
+
+## Comparison rules (what matters)
+
+Ignore as trivial:
+
+* Register renaming with identical instruction sequence and operands.
+* Pure label renumbering or reordering of identical basic blocks.
+* Formatting-only differences or reordered symbol tables.


Can we add a checked-in normalization script so agents do not rewrite this logic each time?

I use something like this (also agent-generated but it works for me so far).

sleeepyjack · 2026-07-09T14:18:32Z

+* The CUDA SM architectures to compile for. Try to detect this from the code and offer the user a list of suggestions.
+  The user must confirm or provide this list.


Suggested change

* The CUDA SM architectures to compile for. Try to detect this from the code and offer the user a list of suggestions.

The user must confirm or provide this list.

* The CUDA SM architectures to compile for. Try to detect this from the code and offer the user a list of suggestions.

The user must confirm or provide this list.

* Prefer `compile_commands.json` when available to recover the actual `nvcc` flags and `--generate-code` / `-arch` settings.

sleeepyjack · 2026-07-09T14:19:33Z

+
+* If any non-trivial change was detected, the top 5 regions where a non-trivial change was detected,
+  including the name of the kernel they appeared in.
+* A short summary of the diff type (opcode change, memory access size change, size delta, control-flow, etc.).


Suggested change

* A short summary of the diff type (opcode change, memory access size change, size delta, control-flow, etc.).

* A short summary of the diff type, including opcode changes, memory access size/cache policy changes, control-flow changes, register-count changes, spills/local memory, shared memory, and occupancy-relevant resource deltas.

github-project-automation Bot added this to CCCL Jul 9, 2026

github-project-automation Bot moved this to Todo in CCCL Jul 9, 2026

cccl-authenticator-app Bot moved this from Todo to In Review in CCCL Jul 9, 2026

coderabbitai Bot reviewed Jul 9, 2026

View reviewed changes

bernhardmgruber added 2 commits July 9, 2026 14:33

Move SASS diff description to a SKILL

f7cd822

Fixes: NVIDIA#7829

Small fixes

9a3165a

bernhardmgruber force-pushed the sass_diff_skill branch from f956a5f to 9a3165a Compare July 9, 2026 12:33

Jacobfaib reviewed Jul 9, 2026

View reviewed changes

sleeepyjack reviewed Jul 9, 2026

View reviewed changes

		* The CUDA SM architectures to compile for. Try to detect this from the code and offer the user a list of suggestions.
		The user must confirm or provide this list.

	* A short summary of the diff type (opcode change, memory access size change, size delta, control-flow, etc.).
	* A short summary of the diff type, including opcode changes, memory access size/cache policy changes, control-flow changes, register-count changes, spills/local memory, shared memory, and occupancy-relevant resource deltas.

Uh oh!

Conversation

bernhardmgruber commented Jul 9, 2026

Uh oh!

coderabbitai Bot commented Jul 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Walkthrough

Changes

Assessment against linked issues

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

coderabbitai Bot Jul 9, 2026

Choose a reason for hiding this comment

Uh oh!

coderabbitai Bot Jul 9, 2026

Choose a reason for hiding this comment

Uh oh!

Jacobfaib Jul 9, 2026

Choose a reason for hiding this comment

Uh oh!

Jacobfaib Jul 9, 2026

Choose a reason for hiding this comment

Uh oh!

fbusato Jul 9, 2026

Choose a reason for hiding this comment

Uh oh!

Jacobfaib Jul 9, 2026

Choose a reason for hiding this comment

Uh oh!

sleeepyjack Jul 9, 2026

Choose a reason for hiding this comment

Uh oh!

sleeepyjack Jul 9, 2026

Choose a reason for hiding this comment

Uh oh!

sleeepyjack Jul 9, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

coderabbitai Bot commented Jul 9, 2026 •

edited

Loading