Add implicit stream benchmarking support #76
Conversation
|
Thanks for the PR! I'll take a look at this once I get the Thrust/CUB 1.16 release candidate prepped. Should be later this week if all goes well. |
alliepiper
added
type: enhancement
|
@PointKernel Could you add an example for this feature in |
Also update to use `nvbench::make_cuda_stream_view`.
There was a problem hiding this comment.
I pushed a couple of commits that update the documentation and add a nvbench::make_cuda_stream_view method. I think that's a clean solution to the non-owning API issue.
This looks good to me -- thanks for working on this @PointKernel and @jrhemstad!
docs/benchmarks.md
Outdated
| NVBench records the elapsed time of work on a CUDA stream for each iteration of a benchmark. | ||
| By default, NVBench creates and provides an explicit stream via `launch::get_stream()` | ||
| to pass to every stream-ordered operation. |
There was a problem hiding this comment.
This isn't quite true -- for the isolated/cold measurements, each iteration is recorded, but for the batch/hot measurements, several iterations are lumped together in a single timer.
I'd also move this down into it's own section -- this section is meant to give an extremely brief overview of a minimal benchmark specification and introduce key concepts. Using an explicit stream is an advanced usecase that should have it's own section.
I'll push a commit to this branch that restructures this a bit, since I'm pretty picky about these docs 😅
nvbench/detail/measure_cupti.cu
Outdated
| try : m_state | ||
| { | ||
| exec_state | ||
| } | ||
| , m_launch{m_state.get_cuda_stream()}, | ||
| m_cupti{*m_state.get_device(), add_metrics(m_state)} |
There was a problem hiding this comment.
Heh. Understandably, clang-format is not a fan of initializer-scope try statements. I'll clean this up a bit in my follow up patch.
Closes #13
This PR adds support to benchmark functions that do not expose stream parameters.