AMDGPU Inspect

Structured AMDGPU code-object and ISA inspection for DeepSeek Harness.

The plugin turns an existing HSACO or AMDGPU ELF file into bounded JSON that an agent can query. It reports compiler artifacts without compiling source, executing a kernel, predicting performance, or prescribing an optimization.

AMDGPU ELF -> llvm-readelf + llvm-objdump -> normalized object model -> DSH tools

Tools

amdgpu_object_inspect

List kernels and report public code-object metadata plus instruction-category counts:

{ "path": "/tmp/vector-add.hsaco" }

amdgpu_isa_query

Query one kernel's normalized instruction stream without placing the complete disassembly in model context:

{
  "path": "/tmp/vector-add.hsaco",
  "kernel": "vector_add",
  "opcodes": ["scratch_*", "s_waitcnt"],
  "cursor": 0,
  "limit": 40
}

The result contains addresses, opcodes, operands, categories, and a nextCursor when more matches remain.

Install

Requirements:

• Linux
• DeepSeek Harness developer preview
• ROCm LLVM tools in /opt/rocm/llvm/bin or on PATH

Install the bundle from GitHub:

dsh plugin --profile web add github:yujiaoliang/dsh-plugin-amdgpu-inspect
dsh --profile web

The package ships plain ESM JavaScript and needs no install-time build script.

Kernel X-Ray Skill

skills/kernel-xray/SKILL.md is the opinionated layer: it tells an agent how to compile before/after artifacts, query this plugin, form hypotheses, and require a workload measurement before claiming a speedup. The plugin itself remains a low-level facts API.

Design

• llvm-readelf --notes supplies public AMDGPU metadata.
• llvm-objdump --disassemble --demangle supplies ISA.
• A one-artifact cache is keyed by absolute path, size, and modification time.
• ISA queries are capped at 200 records and support deterministic cursors.
• Child processes run without a shell and honor tool-call cancellation.
• Inputs are limited to regular files no larger than 256 MiB.

Security boundary

This version invokes LLVM and reads the artifact directly on the DSH host. It does not inherit a deployment's virtual filesystem or subprocess provider, so treat access as host-local rather than sandboxed. It never loads the artifact into an HSA runtime or executes GPU code. LLVM diagnostics can expose local paths.