Introducing ROCprofiler SDK - The Latest Toolkit for Performance Profiling#
Profiling is the backbone of performance optimization in AI and HPC workloads, enabling developers to extract maximum efficiency from AMD Instinct™ GPUs. With ROCm’s rapid evolution, the need for a unified, scalable, and extensible profiling framework has never been more critical. The new ROCprofiler-SDK framework represents a significant step forward in profiling capabilities, offering enhanced features, streamlined integration, and a better user experience while also solving past limitations with former profiler interface versions. This guide aims to help users seamlessly transition from legacy profiling tools to the ROCprofiler-SDK infrastructure. We will explore new features, highlight key differences from previous tools, and provide actionable steps for a smooth migration.
Why ROCprofiler-SDK#
Efficient & Reliable Initialization: Improved tool initialization, reducing setup complexities and overhead.
Multi-Tool Support – Enables multiple tools to simultaneously use the same profiling services without interference.
Simplified Data Collection Control – Manage multiple profiling services easily with unified API control.
Enhanced Error Handling & Logging – Better diagnostics with improved error checking and logging mechanisms.
PC Sampling (Beta) – Advanced sampling capabilities for deeper insights into where time is spent in GPU kernels.
Backward ABI Compatibility – Ensures seamless migration for existing tools using former ROCm profiling interfaces.
Key value proposition#
Let’s look at the key highlights in the ROCprofiler SDK and why you should migrate:
Collect GPU Performance Metrics
Access GPU hardware counters to monitor execution time, memory usage, cache hits/misses, and more.
Identify GPU bottlenecks and optimize compute, memory, and bandwidth utilization.
Trace API Calls & GPU Workload Execution
HIP API Tracing - Capture HIP runtime API calls, their execution order, and overhead for application debugging and performance tuning.
Kernel Tracing - Track kernel launches and execution to debug and optimize your application.
HSA API Tracing - Monitor low-level HSA API calls for a deeper understanding of how workloads interact with the ROCm runtime.
Memory Copy Tracing - Trace device to host (D2H) or Host to Device (H2D) copies.
Marker Tracing: Tracing user-instrumented
roctxmarkers in application code.and many more expanded capabilities like tracking Page Migration events, tracing Scratch Memory operations etc.
With ROCprofiler SDK, developers get a stable, future-proof profiling interface that already integrates with ROCm performance profiling tools like ROCm Systems Profiler & ROCm Compute Profiler.
Evolving ROCprofiler: A more efficient and flexible profiling SDK#
Earlier versions of ROCm profiling interfaces included:
ROCprofiler & ROCTracer (v1) – Separate libraries for hardware counter collection and device activity tracing.
ROCprofiler v2 – A bundled version combining the above libraries, but still inheriting limitations as described below.
The major pain points in these implementations were:
Unnecessary Overhead & Lack of Service Awareness
In previous versions, tools accessed profiling services through
roctracer_init(), which initialized all available services by default, regardless of their need.This led to unnecessary runtime overhead since every tool had to prepare for unused services.
Thread-Safety Challenges & Indirection Overhead
Because services were always enabled, runtime API calls had to go through wrappers that checked service configurations dynamically. This introduced synchronization complexity, requiring multiple threads to coordinate access to profiling data leading to race conditions and increased contention.
This increased latency and made managing data across multiple threads difficult.
Limited Multi-Tool Support
There was no structured way for multiple profiling tools to use the same services at the same time.
This restricted developers from running multiple profiling tools in parallel.
Poor Testing Coverage & Lack of Validation
Code coverage was extremely low, making profiling results less reliable.
Many tests only checked the exit code of
rocprofexecution and missed result verification tests.
ROCprofiler SDK introduces contexts, a powerful mechanism that resolves the inefficiencies of previous profiling approaches.
Context-Based Service Configuration
Tools now can declare which profiling services they need in advance by creating contexts.
Only the requested services are initialized, reducing unnecessary setup overhead.
More Efficient & Thread-Safe Profiling
Since services are only initialized when explicitly requested, ROCprofiler SDK removes unnecessary API wrappers, eliminating unnecessary overhead.
True Multi-Tool Support
Multiple profiling tools can now use the same services simultaneously without interference.
This makes it easier to run different profiling tools in parallel.
Cross-Architecture Testing & Support
ROCprofiler SDK undergoes rigorous testing with comprehensive test suites ensuring compatibility across multiple GPU architectures and robust performance in diverse environments.
Enhanced Adaptability and API improvements
Frequent API improvements make it easier to identify newly added features and adapt tools accordingly.
Consistent naming conventions in the API provide clarity and improve usability.
Extensive documentation ensures that every feature is well-documented, making integration smoother for developers.
What is in ROCprofiler-SDK?#
ROCprofiler API (referred to as
rocprofiler-sdkon Github)The core API is used for tracing and profiling workloads running on AMD GPUs.
Provides direct access to GPU performance counters, kernel execution tracing, and communication profiling.
Used by third-party profiling tools and ROCm-based profilers like ROCm Systems Profiler, ROCm Compute Profiler, and rocprofv3 CLI tool.
rocprofv3ToolA command-line interface (CLI) for executing profiling workflows.
Allows users to collect, analyze, and export profiling data for performance tuning.
Simplifies profiling tasks for developers who need quick insights into GPU execution.
roctxA marker library for code annotations, enabling developers to insert markers in their code for performance analysis on specific code sections.
One of the most noticeable improvements in ROCprofiler SDK is the
refined and more intuitive command-line interface. The transition from
rocprof v1/v2 to rocprof v3 brings a cleaner, more structured
command-line set, along with powerful new features that enhance
usability, flexibility, and profiling efficiency.
The ROCprofiler-SDK also consists of a suite of API services
designed to help developers profile and optimize GPU applications on
the ROCm platform. These services provide deep insights into
application performance and execution behavior, enable
precise profiling, and allow developers to fine-tune their GPU workloads.
Buffered Services – Efficiently manage and store profiling data to prevent system overload.
Callback Tracing – Monitor and log runtime API calls for in-depth execution analysis.
Counter Collection – Gather hardware performance metrics such as memory usage and compute unit activity.
Runtime Intercept Tables – Capture function calls and their parameters to analyze application interactions.
PC Sampling – Periodically record program counters to identify performance hotspots.
For hands-on examples of these APIs in action, check out
the ROCprofiler-SDK Samples Repository on GitHub: Explore the
Samples
Here.
Instructions for building and running your samples are given here:
Instruction to Build & Run ROCprofiler SDK
Samples
State of Command-Line Profiling Tools#
Feature |
rocprof (v1) |
rocprofv2 |
rocprofv3 (ROCprofiler SDK) |
|---|---|---|---|
Brief Description |
Legacy tool for tracing and performance counter collection |
Added functionalities, AMDGPU support, and more output formats |
A more efficient and flexible tool with advanced features, emphasizing stability and robustness |
Underlying Libraries |
ROCprofiler, ROCTracer |
ROCprofilerV2 API |
rocprofiler-sdk |
Output Formats |
CSV, JSON |
CSV, JSON, Ptrace |
CSV, JSON, Ptrace, OTF2 |
Visualization Formats |
JSON (Perfetto) |
Ptrace (Perfetto) |
Ptrace (Perfetto), OTF2 (Vampir) |
Documentation |
|||
Status |
Only critical bug fixes |
Not maintained anymore |
Under active development |
For a detailed comparison of command-line changes, refer to: 🔗 Comparing Command-Line Tool Options: rocprof v1/v2 vs. rocprof v3.
Third-Party Tools Adopting ROCprofiler SDK#
The adoption of ROCprofiler SDK by industry-leading third-party profiling tools underscores its robust capabilities and improved performance with refined design.
Some leading HPC performance analysis tools are actively participating in hackathons and have begun integrating ROCprofiler SDK into their development:
✅ TAU – Comprehensive profiling tool suite for multi-threaded applications.
✅ PAPI – Universal Interface for accessing hardware performance counters.
✅ HPCToolkit – Integrated suite of tools for measurement and analysis of program performance.
✅ Score-P – Highly scalable suite of tools for profiling and event tracing of HPC applications.
✅ Linaro MAP and Performance Reports – Comprehensive tool enabling joint CPU-GPU profiling and providing comprehensive performance reports.
These tools provide critical profiling, tracing, and performance monitoring for developers optimizing AI and HPC workloads at scale. Their migration to ROCprofiler SDK ensures improved profiling accuracy, enhanced thread safety, and better API extensibility for the next generation of ROCm-powered applications.
Join the Future of ROCm Profiling#
With industry adoption gaining momentum, now is the time for other developers and tool vendors to migrate to ROCprofiler SDK. Experience a better and an efficient way to profile workloads on AMD GPUs.
Explore the ROCprofiler SDK GitHub Repository:ROCm/rocprofiler-sdk
Access Documentation here: ROCprofiler SDK Documentation
To conduct a hackathon or assist with your tool migration reach out to us: 📧 dl.ROCm-Profiler.support@amd.com
End of Life Notice: Transition to ROCprofiler SDK#
We are phasing out development and support for ROCTracer, ROCprofiler, rocprof, and rocprofv2 in favor of ROCprofiler-SDK in upcoming ROCm releases. Starting with the ROCm 6.4 release, only critical defect fixes will be addressed for older versions of the profiling tools and libraries. We encourage all users to upgrade to the latest version of the ROCprofiler-SDK library and the rocprofv3 tool to ensure continued support and access to new features. ROCprofiler-SDK is still in beta today and will be production-ready by ROCm 6.5.
We anticipate end of life for ROCprofiler V1/V2 and ROCTracer within nine months after ROCm 6.5 release, aligning with Q1 2026.
Summary#
In this blog, we explored how ROCprofiler SDK is revolutionizing ROCm profiling by replacing legacy tools like ROCprofiler v1/v2 and ROCTracer with a unified, scalable, and efficient framework. We discussed the challenges of older profiling interfaces and how ROCprofiler SDK solves those challenges. The blog highlighted key capabilities, including GPU performance metrics collection, API/kernel tracing, memory analysis, and communication profiling, and showcased the migration of leading third-party tools to ROCprofiler SDK. Additionally, we provided a detailed comparison of command-line changes from rocprof v1/v2 to rocprof v3, outlined the deprecation timeline for older tools, and emphasized the need for developers to transition to ROCprofiler SDK. As ROCprofiler SDK continues to evolve, AMD is committed to expanding its capabilities to meet the growing demands of AI and HPC workloads. Here’s a glimpse into some exciting upcoming features that will make profiling even more powerful and insightful:
Virtualization Support – Enables profiling in virtualized environments, making it easier to analyze performance in cloud workloads.
Support for Interconnect Metrics – Adds xGMI and PCIe monitoring, allowing developers to track bandwidth utilization and data transfer efficiency.
Post-Processing Capabilities – Introduces data processing features that let users analyze and manipulate profiling data after collection for deeper insights.
More Granular Performance Insights – Future updates will bring fine-grained profiling capabilities, providing detailed visibility into memory operations and execution behavior.
Upgrade today and future-proof your performance analysis workflows with ROCprofiler SDK!