ROCm profiling tools overview

This topic provides an overview of the ROCm profiling tool suite, covering the six tools, their relationships, and their dependencies. The tools are complementary and designed to be used together in sequence, moving from coarse system-level characterization down to fine-grained instruction-level analysis.

For guidance on choosing the right tool for your performance investigation, see Choosing the right ROCm profiling tool.

The suite includes the following tools:

Tool	Description
ROCprofiler-SDK	A C++ library that provides the low-level profiling and tracing infrastructure.
rocprofv3	A command-line interface that exposes ROCprofiler-SDK capabilities without requiring any code changes, including advanced features such as ATT, PC Sampling, and SPM.
ROCm Systems Profiler	A system-wide profiler that captures CPU, GPU, MPI, and OS activity in a unified timeline.
ROCm Compute Profiler	A kernel-level analysis tool that collects hundreds of hardware counters and produces roofline, Speed-of-Light, and memory bandwidth analysis.
ROCprof Compute Viewer	A desktop GUI for visualizing instruction-level ATT data, showing wavefront execution timelines, ISA hotspots, and stall attribution at cycle granularity.
ROCm Optiq (ROCm Extras)	A unified visualization application for exploring system- and kernel-level profiling data from rocprofiler-systems and rocprofiler-compute, with no ROCm installation required on the analysis machine.

The ROCm profiling stack

All user-facing tools are built on rocprofiler-sdk, which provides the interception points into the ROCm runtime. The SDK itself depends on the ROCm runtime layer (HIP, HSA, KFD). Tools are independent of each other — none of the user-facing tools routes through another user-facing tool.

The three levels are:

1. ROCm runtime layer — HIP runtime, HSA runtime, and the KFD kernel driver. This is where GPU work is submitted and hardware access is gated.

2. ROCprofiler-SDK — The profiling infrastructure library. It intercepts API calls, reads hardware performance counters, streams PC samples, and captures Advanced Thread Trace data. All higher-level tools use this library.

3. User-facing tools — rocprofv3, rocprofiler-systems, and rocprofiler-compute each build on rocprofiler-sdk. They differ in scope (instruction vs. kernel vs. system) and in how they present results.

Tools

The following sections describe each tool in the ROCm profiling suite, covering what it does, who it’s for, and how it fits in the stack.

ROCprofiler-SDK

ROCprofiler-SDK is a C++ library that provides the profiling and tracing infrastructure on which the ROCm profiling tools are built. It isn’t a user-facing tool; it’s an API consumed programmatically by tool developers, framework integrators, and advanced users. The library provides a unified, thread-safe interface for:

• API tracing: interception and timing of HIP, HSA, ROCTx, and RCCL function calls

• Hardware counter collection: per-kernel PMU performance counter reads from the GPU

• Streaming Performance Monitoring (SPM): continuous hardware counter samples streamed into a ring buffer at a programmable interval, independent of kernel dispatch boundaries

• PC sampling: periodic capture of the program counter and execution state of running GPU wavefronts

• Advanced Thread Trace (ATT): instruction-level wavefront execution history

• Correlation IDs: linking of profiling records across API boundaries and CPU/GPU timelines

• Code object tracking: monitoring of GPU binary load and unload events

Most users interact with ROCprofiler-SDK indirectly through the higher-level tools. For guidance on when to use ROCprofiler-SDK directly, see When to choose ROCprofiler-SDK. For a quick overview, see What is ROCprofiler-SDK?.

rocprofv3

rocprofv3 is the official command-line interface for rocprofiler-sdk. It requires no source-code changes, no recompilation, and no tool library development. Run your application under rocprofv3 and it writes structured output files containing trace data, counter values, or samples. For guidance on when to use rocprofv3, see When to choose rocprofv3. For a quick overview, see rocprofv3 — command-line profiling interface.

ROCm Systems Profiler

ROCm Systems Profiler (rocprofiler-systems) captures the entire application stack in a single unified timeline: CPU function calls, GPU kernel dispatches, MPI collectives, OpenMP regions, memory allocations, OS scheduling events, and GPU telemetry (temperature, power, utilization, bandwidth).

This system-wide view is essential for workloads where the GPU is one component among many. In a distributed training job, for example, the GPU may appear idle not because the kernels are slow but because a data loader is starved, an MPI collective is blocking, or Python GIL contention is stalling the dispatch queue. For guidance on when to use rocprofiler-systems, see When to choose ROCm Systems Profiler. For an overview, see What is ROCm Systems Profiler?.

Two instrumentation modes are available:

• Dynamic binary instrumentation (rocprof-sys-instrument) — inserts timing probes at every function entry and exit without source changes. Deterministic and complete, but carries higher overhead.

• Call-stack sampling (rocprof-sys-sample) — periodically captures the CPU call stack at a configurable frequency. Lower overhead, statistical coverage; suitable for long-running production workloads.

ROCm Compute Profiler

ROCm Compute Profiler (rocprofiler-compute) performs deep per-kernel hardware analysis on AMD Instinct GPUs. Where rocprofiler-systems shows which kernels are slow, rocprofiler-compute explains why at the hardware level. For guidance on when to use rocprofiler-compute, see When to choose ROCm Compute Profiler. For an overview, see What is ROCm Compute Profiler?.

AMD Instinct GPUs expose hundreds of PMU counters across dozens of hardware blocks (compute units, L1 cache, L2 cache, HBM controllers, shader processor input, and more). Because the hardware multiplexers can’t read all counters simultaneously, rocprofiler-compute replays the target application across multiple passes, collecting a different counter subset on each pass, then derives meaningful metrics from the aggregated values.

Visualization tools

The ROCm profiling stack includes two dedicated visualization tools that operate on the output produced by rocprofv3, rocprofiler-systems, and rocprofiler-compute. They serve distinct purposes at different levels of the analysis hierarchy.

Tool	Input	Level	Primary use
ROCprof Compute Viewer (RCV)	rocprofv3 ATT trace files (JSON)	Instruction-level	Wavefront execution, ISA hotspots, stall analysis
ROCm Optiq	.db / .rpd from rocprofiler-compute or rocprofiler-systems	System- and kernel-level	Timeline visualization, kernel metrics, CPU-GPU interaction

ROCprof Compute Viewer

ROCprof Compute Viewer (RCV) is a Qt desktop GUI for visualizing ATT data collected by rocprofv3 --att. It operates at the instruction level, below the kernel level, showing the execution timeline of individual wavefronts across compute units.

RCV answers questions that counter-based analysis can’t: which specific ISA instruction is stalling, how wavefronts occupy execution pipelines over time, and where LDS bank conflicts or memory latency consumes cycles. For guidance on when to use RCV, see When to choose RCV. For a quick overview, see What is RCV?.

ROCm Optiq

Note

ROCm Optiq is currently in beta and under active development.

ROCm Optiq is a unified desktop visualization application for .db and .rpd profiling databases produced by rocprofiler-systems and rocprofiler-compute. A key design goal is portability: Optiq has no dependency on the ROCm stack, so trace files collected on a GPU cluster can be analyzed on any Windows, Linux, or macOS machine, including developer workstations without AMD hardware. For guidance on when to use Optiq, see When to choose ROCm Optiq. For a quick overview, see What is ROCm Optiq?.

Tool relationships and dependencies

This section covers the runtime dependencies of each tool and how their outputs connect across the investigation sequence.

Dependencies

All user-facing tools depend directly on rocprofiler-sdk. rocprofiler-systems additionally depends on OS-level and runtime-level data sources for its cross-domain instrumentation capabilities.

The following table lists the direct dependencies and additional runtime requirements for each tool.

Tool	Direct dependencies	Additional runtime requirements
rocprofiler-sdk	aqlprofile (bundled), rocprofiler-register	ROCm runtime (HIP + HSA), KFD kernel driver
rocprofv3	rocprofiler-sdk	—
rocprofiler-systems	rocprofiler-sdk, Linux perf/libpfm, amd-smi	PMPI (optional, for MPI tracing), OMPT (optional, for OpenMP tracing)
rocprofiler-compute	rocprofiler-sdk, amd-smi, rocminfo	Python 3.8+

Data flow between tools

The tools are designed to be used in sequence. Each tool’s output directs what to investigate with the next tool.

Recommended profiling workflow

For most performance investigations, work top-down through the stack. The following workflow demonstrates a complete investigation cycle using all four levels of the tool stack: