Amd smi

[djwoun]

Pull Request Description

AMD SMI PAPI Component

Overview

This component adds AMD SMI to PAPI. It discovers GPUs at runtime and exposes AMD SMI-reported metrics (temperatures, power, fans, PCIe, RAS/ECC, cache/VRAM info, utilization, counters, etc.) as PAPI native events.

Targeted for ROCm 6.4.0 (AMD SMI ~25.3) and expected to work with ROCm(6.3.4 ~ 6.4.3) releases as well.
Tested on a combination of MI210 & MI300 for ROCm:

• 6.4.0
• 6.3.4

---

Usage

# point to your ROCm install (tested with 6.4.0)
export PAPI_AMDSMI_ROOT=/opt/rocm-6.4.0

# configure PAPI
./configure --prefix=${INSTDIR} --with-components="amd_smi"

---

How it fits together (high-level)

• linux-amd-smi.c: implements the PAPI vector (init/start/stop/read/reset/etc.) and delegates to the internal amds_* APIs.
• amds.c: dlopen AMD SMI, discover devices, build the native event table, and wire up accessors.
• amds_accessors.c: one function per metric — actually calls AMD SMI library to read and write values.
• amds_ctx.c: per-eventset lifecycle — open/close/start/stop/read/write/reset groups of events and enforce device usage.
• amds_evtapi.c: native event enumeration — code↔name↔description helpers for PAPI.
• amds_priv.h: internal types (e.g., native_event_t), globals, and AMD SMI function-pointer declarations.
• amds_funcs.h: list of AMD SMI API calls used (generates the function-pointer declarations/definitions).
• htable.h: lightweight string→event lookup (for fast name→code mapping).
• amds.h: public “component-internal” API used across the above files.
• Rules.amd_smi: build glue to include this component in PAPI.

---

File-by-file (concise)

• linux-amd-smi.c
  Declares the papi_vector_t for this component; initializes on first use; hands off work to amds_* for device/event management; implements PAPI hooks (init_component, update_control_state, start, read, stop, reset, shutdown, and native-event queries).

• amds.c
  Dynamically loads libamd_smi.so, resolves AMD SMI symbols, discovers sockets/devices, and builds the native event table. Defines helpers to add simple and counter-based events. Manages global teardown (destroy event table, close library).

• amds_accessors.c
  Implements the accessors that read/write individual metrics (e.g., temperatures, fans, PCIe, energy, power caps, RAS/ECC, clocks, VRAM, link topology, XGMI/PCIe metrics, firmware/board info, etc.). Each accessor maps an event’s (variant, subvariant) to the right SMI call and returns the value.

• amds_ctx.c
  Provides the per-eventset context:

  • amds_ctx_open/close — acquire/release devices, run per-event open/close hooks.
  • amds_ctx_start/stop — start/stop counters where needed.
  • amds_ctx_read/write/reset — read current values, optionally write supported controls (e.g., power cap), zero software view.

• amds_evtapi.c
  Implements native-event enumeration for PAPI (enum, code_to_name, name_to_code, code_to_descr) using the in-memory event table and a small hash map for fast lookups.

• amds_priv.h
  Internal definitions: native_event_t (name/descr/device/mode/value + open/close/start/stop/access callbacks), global getters, and the AMD SMI function-pointer declarations (via amds_funcs.h).

• amds_funcs.h
  Centralized macro list of AMD SMI APIs used by the component; generates function-pointer declarations/definitions so amds.c can dlsym() them at runtime. Conditional entries handle newer SMI features.

• htable.h
  Minimal chained hash table for name→event mapping; used by amds_evtapi.c to resolve native event names quickly.

• amds.h
  Public, component-internal API across files: init/shutdown, native-event queries, context ops, and error-string retrieval.

• Rules.amd_smi
  Build integration for PAPI’s make system; compiles this component and sets include/library paths for AMD SMI.

---

Author Checklist

• Description
  Why this PR exists. Reference all relevant information, including background, issues, test failures, etc
• Commits
  Commits are self contained and only do one thing
  Commits have a header of the form: module: short description
  Commits have a body (whenever relevant) containing a detailed description of the addressed problem and its solution
• Tests
  The PR needs to pass all the tests