NVIDIA CUDA Profiling Tools Interface (CUPTI) - 1.0
The NVIDIA? CUDA Profiling Tools Interface (CUPTI) is a dynamic library that enables the creation of profiling and tracing tools that target CUDA applications. CUPTI provides a set of four APIs targeted at ISVs creating profilers and other performance optimization tools:
- the Activity API
- the Callback API
- the Event API, and
- the Metric API.
Using these CUPTI APIs, independent software developers can create profiling tools that give insight into the CPU and GPU behavior of CUDA applications. Normally packaged with the CUDA Toolkit, this stand-alone version of CUPTI includes a new set of metric APIs for compute capability 7.0 and above. These provide low overhead and deterministic profiling on the target system.
- Trace CUDA API usage by registering callbacks for API calls of interest
- Full support for entry and exit points in the CUDA C Runtime (CUDART) and CUDA Driver
- Sample hardware and software event counters, including:
- Instruction count and throughput
- Memory load/store events and throughput
- Cache hits/misses
- Branches and divergent branches
- Many more
- Enables automated bottleneck identification based on metrics such as instruction throughput, memory throughput, and more
- Normalized timestamps for CPU and GPU events
- Independent CUPTI release
CUPTI version 1.0 is the first version to be released outside of the CUDA Tool Kit, and includes the features and functionality of the CUPTI found in CUDA 10.0, plus more!
- New metric APIs for Compute Capability 7.0 and 7.5
This release includes a new set of metric APIs for
- compute capability 7.0 (Volta GPU architectures) and
- compute capability 7.5 (Turing GPU architectures).
- Support for Legacy metric APIs
- Both the old and new metric APIs are supported for compute capability 7.0. This is to enable transition of code to the new metric APIs, but one cannot mix the usage of the old and new metric APIs.
- The earlier event and old metric APIs are not supported for the new Turing GPU architecture (compute capability 7.5).
Variations from the CUPTI found in the CUDA Toolkit 10.0
- The CUPTI version bundled with the CUDA 10.0 Toolkit does not include the new metric APIs.
- The CUPTI version bundled with the CUDA 10.0 Toolkit does not support compute capability 7.5 (Turing GPU architecture).
Variations from the CUPTI found in the version 1.0.6
- Fixed PC Sampling issue with values getting accumulated across launches for the same kernel.
- Fixed PC Sampling cuptiActivityConfigurePCSampling() API returning CUPTI_ERROR_INVALID_PARAMETER for a sampling period value of CUPTI_ACTIVITY_PC_SAMPLING_PERIOD_MAX.
- Linux x86_64
- Windows x86_64
- Volta: GV100
- Turing: TU102, TU104, TU106
- CUPTI 1.0 should be used along with the CUDA Toolkit 10.0 production release
- Please use the drivers provided with CUDA Toolkit 10.0 production release or a more recent version.
- CUPTI 1.0 Product Documentation
- CUDA Toolkit 10.0 Documentation
- doc/html/index.html - CUPTI HTML document
- doc/pdf/Cupti.pdf - CUPTI PDF document
- include/ - folder containing the header files
- lib/x64 - folder containing CUPTI host and target libraries and the Perfworks host library
- samples/ - folder containing samples to use the CUPTI APIs
To provide feedback, request additional features, or report issues, please use the Developer Forums.
Files included in this package:
How to extract, build and run profiler samples on Linux/Windows:
- Extract CUPTI package.
tar -zxvf cupti-linux-cuda10-1.0.7.tar.gz
- Build profiler host utility.
- cd Cupti_CUDA10_1.0.7-25030996/samples/extensions/src/profilerhost_util
- Export LD_LIBRARY_PATH = `pwd`:../../lib/x64
- Change directory to sample directory
- make run
- Repeat step 4 and 5 for other samples