NVIDIA/nvbench
1
0
Fork 0
mirror of https://github.com/NVIDIA/nvbench.git synced 2026-03-14 20:27:24 +00:00
Code Issues Packages Projects Releases Wiki Activity

More README cleanup.

Browse Source
This commit is contained in:
Allison Vacanti
2021-03-03 17:25:12 -05:00
parent 21e13f002d
commit d7c34c835d
1 changed files with 7 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
README.md
View File

@@ -286,11 +286,11 @@ NVBENCH_BENCH_TEMPLATE(my_benchmark, NVBENCH_TYPE_AXES(Ts, Us));
# Execution Tags For Special Cases
By default, NVBench assumes that the entire execution time of the
`KernelLauncher` should be measured, and that no syncs are performed (
e.g. `cudaDeviceSynchronize`, `cudaStreamSynchronize`, `cudaEventSynchronize`,
etc).
`KernelLauncher` should be measured, and that no syncs are performed
(e.g. `cudaDeviceSynchronize`, `cudaStreamSynchronize`, `cudaEventSynchronize`,
etc. are not called).
Execution tags may be passed to `state.exec` when this these assumptions are not
Execution tags may be passed to `state.exec` when these assumptions are not
true:
- `nvbench::exec_tag::sync` tells NVBench that the kernel launcher will
@@ -305,13 +305,13 @@ state.exec(nvbench::exec_tag::sync | nvbench::exec_tag::timer,
[](nvbench::launch &launch, auto& timer) { /*...*/ });
```
The following sections provide more detail.
The following sections provide more details on these features.
## Benchmarks that sync: `nvbench::exec_tag::sync`
If a `KernelLauncher` synchronizes the CUDA device internally without passing
this tag, the benchmark will deadlock at runtime. Passing the `sync` tag will
fix this. Note that the `sync` exec tag will disable batch measurements.
this tag, **the benchmark will deadlock at runtime**. Passing the `sync` tag
will fix this issue. Note that this disables batch measurements.
```cpp
void sync_example(nvbench::state& state)