NVIDIA CUDA Profiling Tools Interface (CUPTI) - CUDA Toolkit 12.3
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 APIs targeted at ISVs creating profilers and other performance optimization tools:
- the Activity API,
- the Callback API,
- the Event API,
- the Metric API,
- the Profiler API,
- the PC Sampling API,
- the SASS Metric API, and
- the Checkpoint API
Using these CUPTI APIs, independent software developers can create profiling tools that provide low and deterministic profiling overhead on the target system, while giving insight into the CPU and GPU behavior of CUDA applications. Normally packaged with the CUDA Toolkit, NVIDIA occasionally uses this page to provide CUPTI improvements and bug fixes between toolkit releases.
Revision History
Key Features
- Trace CUDA API by registering callbacks for API calls of interest
- Full support for entry and exit points in the CUDA C Runtime (CUDART) and CUDA Driver
- GPU workload trace for the activities happening on the GPU, which includes kernel executions, memory operations (e.g., Host-to-Device memory copies) and memset operations.
- CUDA Unified Memory trace for transfers from host to device, device to host, device to device and page faults on CPU and GPU etc.
- Normalized timestamps for CPU and GPU trace
- Profile hardware and software event counters, including:
- Utilization metrics for various hardware units
- 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
- Range profiling to enable metric collection over concurrent kernel launches within a range
- Metrics attribution at the high-level source code and the executed assembly instructions.
- Device-wide sampling of the program counter (PC). The PC Sampling gives the number of samples for each source and assembly line with various stall reasons.
Updates in CUDA Toolkit 12.3 Update 1
- To provide normalized timestamps for all activities, CUPTI uses linear interpolation for conversion from GPU timestamps to CPU timestamps. This was broken with CUDA 12.3 causing spurious gaps or overlap on Tegra platforms. Fixed the issue.
Resolved Issues
Updates in CUDA Toolkit 12.3
- New attributes:
CUPTI_ACTIVITY_OVERHEAD_RUNTIME_TRIGGERED_MODULE_LOADING
andCUPTI_ACTIVITY_OVERHEAD_LAZY_FUNCTION_LOADING
are added in the activity overhead enumCUpti_ActivityOverheadKind
to provide the overhead information for CUDA runtime triggered module loading and lazy function loading respectively. - New API
cuptiGetGraphExecId
provides the unique ID of the executable graph. - Added support for collecting graph level trace for device launched graphs. A new
API
cuptiActivityEnableDeviceGraph
is added to enable the collection of records for device launched graphs. - CUDA Graphs can be executed on multiple devices i.e. the root node could
be launched on one device and the leaf node could be launched on the
another device. New fields
endDeviceId
andendContextId
are added to identify the ids of device and context respectively which are used to execute the last node of the graph. To accomodate this change, activity recordCUpti_ActivityGraphTrace
is deprecated and it is replaced by a new activity recordCUpti_ActivityGraphTrace2
. - Added WSL profiling support on Windows 10 WSL with OS build version 19044 and greater. WSL profiling is not supported on Windows 10 WSL for systems that exceed 1 TB of system memory.
- Several performance improvements are done in the tracing path. One of the key
improvements is to allow clients to request CUPTI to maintain the activity buffers
at the thread level instead of global buffers. This can be achieved by setting the
option
CUPTI_ACTIVITY_ATTR_PER_THREAD_ACTIVITY_BUFFER
of the enumCUpti_ActivityAttribute
. This can help in reducing the collection overhead for applications which launch CUDA activities from multiple host threads. - Frame pointers are enabled for Linux x86_64 libraries.
- The deprecated Activity APIs and structures have been moved to a new header cupti_activity_deprecated.h, which is included in the header cupti_activity.h. Header cupti_activity.h contains only the latest version of APIs and structures.
- CUPTI no longer uses profiling semaphore pool to
store the profiling data. Coresponding attributes
CUPTI_ACTIVITY_ATTR_PROFILING_SEMAPHORE_POOL_SIZE
,CUPTI_ACTIVITY_ATTR_PROFILING_SEMAPHORE_POOL_LIMIT
andCUPTI_ACTIVITY_ATTR_PROFILING_SEMAPHORE_PRE_ALLOCATE_VALUE
have been deprecated. - Fixed SASS metric profiling for cuda graph.
- Fixed race condition in the API
cuptiSetThreadIdType
for late subscription
New Features
Resolved Issues
Requirements
- Linux x86_64[1]
- Windows x86_64[1]
- Linux PowerPC[1]
- Linux aarch64 SBSA[1]
- DRIVE OS QNX aarch64[2]
- DRIVE OS Linux aarch64[2]
- Activity and Callback APIs
- All architectures supported by CUDA Toolkit
- Event and Metric APIs
- Volta
- Pascal
- Maxwell
- Profiling and PC Sampling APIs
- Hopper: GH100
- Ada: AD10x
- Ampere: A100 with Multi-Instance GPU, GA10x
- Turing
- Volta: GV100, GV10B
- CUPTI can be found in the CUDA Toolkit 12.3 Update 1 production release
- 546.12 (Windows)
- 545.23.08 (Linux)
Supported platforms
[2] available in the Embedded or Drive toolkits only
Supported NVIDIA GPU architectures
CUDA Toolkit
Drivers
-
Please use the following drivers
Documentation
Support
To provide feedback, request additional features, or report issues, please use the Developer Forums.
Installation Overview
When installing CUDA Toolkit 12.3 Update 1 and specifying options, be sure to select CUDA > Development > Tools > CUPTI.