NVIDIA CUDA Profiling Tools Interface (CUPTI) - CUDA Toolkit 12.4
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 12.4 Update 1
- Fixed a crash when API
cuptiFinalize
is used for applications using CUDA Graph. - Fixed a crash which can occur while tracing memcpy and memset nodes in a graph when using graph level tracing..
- Skip delivering worker thread buffers on internal flush if the worker thread buffer is not full.
Resolved Issues
Updates in CUDA Toolkit 12.4
- Added tracing support for applications using Green contexts. Added two new fields
isGreenContext
andparentContextId
in the context activity record. The activity record CUpti_ActivityContext is deprecated and it is replaced by a new activity recordCUpti_ActivityContext2
. - CUDA API calls are completed asynchronously from the perspective of the host CPU. This is accomplished by queuing the work slated for the GPU into a structure known as a command buffer. If there is insufficient space available in the command buffer when attempting to call a CUDA API, the host call will block until space becomes available. The user should be able to identify when this situation occurs. This is indicated using the new attribute
CUPTI_ACTIVITY_OVERHEAD_COMMAND_BUFFER_FULL
added in the activity overhead enumCUpti_ActivityOverheadKind
. To provide additional details about the overhead, a new fieldoverheadData
is added in the overhead activity record. Activity recordCUpti_ActivityOverhead2
is deprecated and it is replaced by the new activity recordCUpti_ActivityOverhead3
. - Added process ID and thread ID in the JIT activity record. To accomodate this change, activity record
CUpti_ActivityJit
is deprecated and it is replaced by a new activity recordCUpti_ActivityJit2
. - To correlate the sampling data for a kernel with the launch API in the serial mode of the PC Sampling APIs, a new field
correlationId
is added in the structCUpti_PCSamplingPCData
. - For PC Sampling APIs, total (
smsp__pcsamp_sample_count
) and dropped (smsp__pcsamp_samples_data_dropped
) sample counts are collected by default. - Fixed the issue for overhead records showing the default thread ID than the one requested using the API
cuptiSetThreadIdType()
. - Fixed instruction level SASS metrics profiling for CUDA Graph applications.
- When a device graph is first launched from the device and it is not launched from the host earlier, end timestamp could be 0 for graph-level tracing on Ampere and later GPU architectures. This issue is fixed.
New Features
Resolved Issues
Requirements
- Linux x86_64[1]
- Windows x86_64[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 12.4 Update 1 production release
- 551.78 (Windows)
- 550.54.15 (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 12.4 Update 1 and specifying options, be sure to select CUDA > Development > Tools > CUPTI.