AMD's CodeXL profiler/debugger has been updated to version 2.3. Updates include new features for CPU, Power, and GPU profiling and graphics debugging, including support for the AMD Ryzen Processor.
CodeXL is available for download at gpuopen.com or directly from the CodeXL 2.3 GitHub repository here.
Major updates:
CPU Profiling
- Support for AMD Ryzen Processor.
- Support Count Mode profiling in CodeXLCpuProfiler CLI. In this mode, the CLI counts the number of occurrences of monitored events.
- CodeXLCpuProfiler CLI supports headless servers to collect and analyze profile data.
- CodeXLCpuProfiler CLI generates Call Graph section if the CallStack sampling is enabled during profile data collection.
- Many GUI fixes and improvements to make it more stable.
Power Profiling
- Support RAPL Energy counters for AMD Ryzen Processor.
- Simplified Summary View for Timeline profiling for AMD Ryzen processor.
- Support AMD Radeon RX 500 Series discrete GPUs.
- Module level attribution (Now available in the Command Line tool on Linux).
- Many GUI fixes and improvements to make it more stable.
GPU Profiling
- Support for the most recent GPUs, including the RX 500 series GPUs.
- Support for the ROCm 1.5 release.
- Support for showing ROCm data transfers on the timeline when performing an Application trace.
- Support for collecting an AQL packet trace for ROCm applications (command line profiler support only).
- Support for specifying profiler start delay and duration (command line profiler support only).
- Command line support for querying the number of passes required by a given set of performance counters.
- Support for “Whole Application Replay”, which allows collecting multi-pass counter sets for applications which don’t support individual kernel replay, like ROCm applications and OpenCL applications that used Shared Virtual Memory (command line profiler support only).
GPU Debugger Back-End
- Remove HSA debugging support
- Disable OpenCL kernel debugging on new OpenCL™ with open source LLVM Lightning compiler path which can’t be supported by the existing AMDIL-based patching debugging technology
