| Copyright 2021 NVIDIA Corporation. All rights reserved |
|
|
| Profiling API injection sample code |
|
|
| Build this sample with |
|
|
| make CUDA_INSTALL_PATH=/path/to/cuda |
|
|
| This x86 linux-only sample contains 4 build targets: |
|
|
| libinjection_1.so |
| * Minimal injection sample code showing how to write an injection library using |
| either LD_PRELOAD to perform dlsym() interception, or CUDA's injection support with |
| CUDA_INJECTION64_PATH. |
| * When CUDA_INJECTION64_PATH is set to a shared library, at initialization, CUDA |
| will load the shared object and call the function named 'InitializeInjection'. |
| * When LD_PRELOAD is set to a shared library, its symbols will be preferrentially |
| used to resolve dynamic linking. When an application dynamically links in the |
| dlsym() call, this version of dlsym() is provided instead of the default system |
| version. In this case, dlsym() is used to call CUPTI initialization code, then |
| call an internal name for the system dlsym(), ensuring that the original functionality |
| of dlsym() is preserved. |
| *** While this sample shows potential use of LD_PRELOAD, CUPTI does not currently |
| recommend using this means of injecting a tool into a process - CUPTI's initialization |
| may run before other objects are constructed, causing potential undefined behavior. |
| For this reason we only recommend using CUDA_INJECTION64_PATH to guarantee |
| correct behavior. *** |
|
|
| libinjection_2.so |
| * Expands on the injection_1 sample to add CUPTI Callback and Profiler API calls |
| * Registers callbacks for cuLaunchKernel and context creation. This will be |
| sufficient for many target applications, but others may require other launches |
| to be matched, eg cuLaunchCoooperativeKernel or cuLaunchGrid. See the Callback |
| API for all possible kernel launch callbacks. |
| * Creates a Profiler API configuration for each context in the target (using the |
| context creation callback). The Profiler API is configured using Kernel Replay |
| and Auto Range modes with a configurable number of kernel launches within a pass. |
| * The kernel launch callback is used to track how many kernels have launched in |
| a given context's current pass, and if the pass reached its maximum count, it |
| prints the metrics and starts a new pass. |
| * At exit, any context with an unprocessed metrics (any which had partially |
| completed a pass) print their data. |
| * This library links in the profilerHostUtils library which may be built from the |
| cuda/extras/CUPTI/samples/extensions/src/profilerhost_util/ directory |
|
|
| simple_target |
| * Very simple executable which calls a kernel several times with increasing amount |
| of work per call. |
|
|
| complex_target |
| * More complicated example (similar to the concurrent_profiling sample) which |
| launches several patterns of kernels - using default stream, multiple streams, |
| and multiple devices if there are more than one device. |
|
|
| To use the injection library, set CUDA_INJECTION64_PATH to point to that library |
| when you launch the target application: |
|
|
| env CUDA_INJECTION64_PATH=./libinjection_2.so ./simple_target |
|
|
