Scarfski/tracy
1
0
Fork 0
mirror of https://github.com/wolfpld/tracy.git synced 2026-09-01 07:58:31 +00:00
Code Issues Packages Projects Releases Wiki Activity

Refine tracy cuda documentation

Browse Source
This commit is contained in:
Fan Chen
2025-05-22 19:26:57 -04:00
parent 467f39b222
commit feda21b047
1 changed files with 8 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
manual/tracy.tex
View File

@@ -1694,13 +1694,17 @@ Similar to Vulkan and OpenGL, you also need to periodically collect the OpenCL e
\subsubsection{CUDA}
CUDA support is enabled by including the \texttt{public/tracy/TracyCUDA.hpp} header file. Tracing CUDA requires the creation of a Tracy CUDA context using the macro \texttt{TracyCUDAContext()}, which returns an instance of a \texttt{TracyCUDACtx} object. This context must be used when profiling CUDA events. Cleanup is handled using the \texttt{TracyCUDAContextDestroy(ctx)} macro. To assign a custom name to the context, use the \texttt{TracyCUDAContextName(ctx, name, size)} macro.
CUDA support is enabled by including the \texttt{public/tracy/TracyCUDA.hpp} header file. To use it, the NVIDIA CUPTI library is required. This library comes with the NVIDIA CUDA Toolkit and is located at \texttt{CUDA\_INSTALLATION\_PATH/extras/CUPTI}.
To begin profiling CUDA events, call the \texttt{TracyCUDAStartProfiling(ctx)} macro. This initiates the profiling of CUDA events, and GPU zones for key CUDA activities will be created automatically.
Tracing CUDA requires the creation of a Tracy CUDA context using the macro \texttt{TracyCUDAContext()}, which returns an instance of a \texttt{TracyCUDACtx} object. TracyCUDA allows only a single \texttt{TracyCUDACtx} object at any given time. Subsequent calls to \texttt{TracyCUDAContext()} will return the same reference-counted object. There is no need for clients to instantiate multiple \texttt{TracyCUDACtx} objects, as a single context is capable of instrumenting all CUDA contexts and streams. This context must be used when profiling CUDA events.
The \texttt{TracyCUDACollect(ctx)} macro flushes CUDA activities and collects the associated events, resulting in the creation of GPU zones. This step is optional, as the collector thread is enabled by default.
Cleanup is handled using the \texttt{TracyCUDAContextDestroy(ctx)} macro. To assign a custom name to the context, use the \texttt{TracyCUDAContextName(ctx, name, size)} macro.
To stop profiling, call the \texttt{TracyCUDAStopProfiling(ctx)} macro. This terminates the profiling of CUDA events.
To begin instrumentation of all CUDA API calls, use the \texttt{TracyCUDAStartProfiling(ctx)} macro. This initiates the profiling of CUDA events, including relevant GPU activity such as kernel execution, memory transfers, and synchronization. This instrumentation is automatic and requires no code annotation\footnote{CUDA does not provide an API to retrieve timestamps associated with events. Therefore, the typical GPU instrumentation design of Tracy cannot be applied.}.
Unlike other GPU backends in Tracy, there is no need to call \texttt{TracyCUDACollect(ctx)} periodically, since a background collector thread is enabled by default. This behavior can be disabled by defining \texttt{TRACY\_CUDA\_ENABLE\_COLLECTOR\_THREAD} as \texttt{0} prior to including \texttt{TracyCUDA.hpp}.
To stop profiling, call the \texttt{TracyCUDAStopProfiling(ctx)} macro.
\subsubsection{Multiple zones in one scope}