Skip to content

Reduce warpspeed scan to 256 threads/block on NVHPC #7892

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
bernhardmgruber merged 1 commit into from
Mar 9, 2026
Merged

Reduce warpspeed scan to 256 threads/block on NVHPC #7892

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 13 additions & 2 deletions cub/cub/device/dispatch/tuning/tuning_scan.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -573,8 +573,15 @@ struct policy_hub
static constexpr int num_threads_per_warp = 32;

// TODO(bgruber): tune this
static constexpr int num_reduce_warps = 4;
static constexpr int num_scan_stor_warps = 4;
# if _CCCL_COMPILER(NVHPC)
// need to reduce the number of threads to <= 256, so each thread can use up to 255 registers. This avoids an
// error in ptxas, see also: https://github.com/NVIDIA/cccl/issues/7700.
static constexpr int num_reduce_warps = 2;
static constexpr int num_scan_stor_warps = 2;
# else // _CCCL_COMPILER(NVHPC)
static constexpr int num_reduce_warps = 4;
static constexpr int num_scan_stor_warps = 4;
# endif // _CCCL_COMPILER(NVHPC)
static constexpr int num_load_warps = 1;
static constexpr int num_sched_warps = 1;
static constexpr int num_look_ahead_warps = 1;
Expand All @@ -587,6 +594,10 @@ struct policy_hub
num_reduce_warps + num_scan_stor_warps + num_load_warps + num_sched_warps + num_look_ahead_warps;
static constexpr int num_total_threads = num_total_warps * num_threads_per_warp;

# if _CCCL_COMPILER(NVHPC)
static_assert(num_total_threads <= 256);
# endif // _CCCL_COMPILER(NVHPC)

static constexpr int squad_reduce_thread_count = num_reduce_warps * num_threads_per_warp;

// 256 / sizeof(InputValueT) - 1 should minimize bank conflicts (and fits into 48KiB SMEM)
Expand Down
Loading