Skip to content

[CUB]: use vectorized kernel for triad and add benchmark for dtypes of size 2 #7019

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
NaderAlAwar merged 2 commits into from
Dec 19, 2025
Merged

[CUB]: use vectorized kernel for triad and add benchmark for dtypes of size 2 #7019

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fa9a54c
Use vectorized tuning for triad benchmark for dtypes of size 2
NaderAlAwar Dec 18, 2025
fdf5610
Merge branch 'main' into triad-transform-vectorized-tuning
NaderAlAwar Dec 18, 2025
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
18 changes: 17 additions & 1 deletion cub/cub/device/dispatch/tuning/tuning_transform.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -326,6 +326,15 @@ struct tuning_vec<1200, StoreSize>
static constexpr int items_per_thread = 8;
};

// manually tuned triad on A100
template <int StoreSize, int LoadSize0, int... LoadSizes>
struct tuning_vec<800, StoreSize, LoadSize0, LoadSizes...>
{
static constexpr int block_threads = 128;
static constexpr int vec_size = 4;
static constexpr int items_per_thread = 16;
};

template <bool RequiresStableAddress,
bool DenseOutput,
typename RandomAccessIteratorTupleIn,
Expand Down Expand Up @@ -397,7 +406,14 @@ struct policy_hub<RequiresStableAddress,
block_threads* async_policy::min_items_per_thread,
ldgsts_size_and_align)
> int{max_smem_per_block};
static constexpr bool fallback_to_vectorized = exhaust_smem || no_input_streams || !can_memcpy_all_inputs;

// on Ampere, the vectorized kernel performs better for 1 and 2 byte values
static constexpr bool use_vector_kernel_on_ampere =
((size_of<it_value_t<RandomAccessIteratorsIn>> < 4) && ...) && sizeof...(RandomAccessIteratorsIn) > 1
&& size_of<it_value_t<RandomAccessIteratorOut>> < 4;

static constexpr bool fallback_to_vectorized =
exhaust_smem || no_input_streams || !can_memcpy_all_inputs || use_vector_kernel_on_ampere;

public:
static constexpr auto algorithm =
Expand Down
Loading