Skip to content

fall back gpu_to_gpu floating-point min/max reductions to run_to_run #6462

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
merged 2 commits into from
Nov 4, 2025
Merged

fall back gpu_to_gpu floating-point min/max reductions to run_to_run #6462

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
972d5ed
fall back gpu_to_gpu floating-point min/max reductions to run_to_run
srinivasyadav18 Nov 3, 2025
346f1a7
add test for maximum
srinivasyadav18 Nov 3, 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
9 changes: 7 additions & 2 deletions cub/cub/device/device_reduce.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -469,7 +469,12 @@ public:
constexpr auto float_double_plus =
gpu_gpu_determinism && detail::is_one_of_v<accum_t, float, double> && detail::is_cuda_std_plus_v<ReductionOpT>;

constexpr auto supported = integral_fallback || float_double_plus || !gpu_gpu_determinism;
constexpr auto float_double_min_max_fallback =
gpu_gpu_determinism
&& detail::is_one_of_v<accum_t, float, double> && detail::is_cuda_minimum_maximum_v<ReductionOpT>;

constexpr auto supported =
integral_fallback || float_double_plus || float_double_min_max_fallback || !gpu_gpu_determinism;

// gpu_to_gpu determinism is only supported for integral types with cuda operators, or
// float and double types with ::cuda::std::plus operator
Expand All @@ -496,7 +501,7 @@ public:
// If the conditions for gpu-to-gpu determinism or non-deterministic
// reduction are not met, we fall back to run-to-run determinism.
using determinism_t = ::cuda::std::conditional_t<
(gpu_gpu_determinism && integral_fallback)
(gpu_gpu_determinism && (integral_fallback || float_double_min_max_fallback))
|| (no_determinism && !(is_contiguous_fallback && is_plus_fallback && is_4b_or_greater)),
::cuda::execution::determinism::run_to_run_t,
default_determinism_t>;
Expand Down
95 changes: 59 additions & 36 deletions cub/test/catch2_test_device_reduce_deterministic.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -329,7 +329,9 @@ C2H_TEST("Deterministic Device reduce works with float and double on gpu with di
}

using test_types =
c2h::type_list<int8_t,
c2h::type_list<float,
double,
int8_t,
int32_t,
int64_t
#if TEST_INT128()
Expand All @@ -356,75 +358,96 @@ C2H_TEST("Deterministic Device reduce works with integral types on gpu with diff
c2h::device_vector<type> d_input(num_items);
c2h::gen(C2H_SEED(2), d_input, min_value, max_value);

SECTION("plus")
if constexpr (::cuda::std::is_integral_v<type>)
{
c2h::device_vector<type> d_output(1);
SECTION("plus")
{
c2h::device_vector<type> d_output(1);

auto error =
cub::DeviceReduce::Reduce(d_input.begin(), d_output.begin(), num_items, cuda::std::plus<type>{}, init_t{}, env);
REQUIRE(error == cudaSuccess);
auto error =
cub::DeviceReduce::Reduce(d_input.begin(), d_output.begin(), num_items, cuda::std::plus<type>{}, init_t{}, env);
REQUIRE(error == cudaSuccess);

c2h::host_vector<type> h_input = d_input;
c2h::host_vector<type> h_input = d_input;

c2h::host_vector<type> h_expected(1);
// Requires `std::accumulate` to produce deterministic result which is required for comparison
// with the device RFA result.
// NOTE: `std::reduce` is not equivalent
h_expected[0] = std::accumulate(h_input.begin(), h_input.end(), init_t{}, cuda::std::plus<type>{});
c2h::host_vector<type> h_expected(1);
// Requires `std::accumulate` to produce deterministic result which is required for comparison
// with the device RFA result.
// NOTE: `std::reduce` is not equivalent
h_expected[0] = std::accumulate(h_input.begin(), h_input.end(), init_t{}, cuda::std::plus<type>{});

c2h::host_vector<type> h_output = d_output;
REQUIRE(h_expected == h_output);
}
c2h::host_vector<type> h_output = d_output;
REQUIRE(h_expected == h_output);
}

SECTION("minimum")
{
c2h::device_vector<type> d_output(1);
SECTION("bitwise xor")
{
c2h::device_vector<type> d_output(1);

init_t init_value{cuda::std::numeric_limits<init_t>::max()};
init_t init_value{};

auto error =
cub::DeviceReduce::Reduce(d_input.begin(), d_output.begin(), num_items, cuda::minimum<init_t>{}, init_value, env);
REQUIRE(error == cudaSuccess);
auto error = cub::DeviceReduce::Reduce(
d_input.begin(), d_output.begin(), num_items, cuda::std::bit_xor<>{}, init_value, env);
REQUIRE(error == cudaSuccess);

c2h::host_vector<type> h_input = d_input;
c2h::host_vector<type> h_expected(1);
h_expected[0] = std::accumulate(h_input.begin(), h_input.end(), type{init_value}, cuda::minimum<>{});
c2h::host_vector<type> h_input = d_input;
c2h::host_vector<type> h_expected(1);
h_expected[0] = std::accumulate(h_input.begin(), h_input.end(), type{init_value}, cuda::std::bit_xor<type>{});

c2h::host_vector<type> h_output = d_output;
REQUIRE(h_expected == h_output);
c2h::host_vector<type> h_output = d_output;
REQUIRE(h_expected == h_output);
}

SECTION("logical or")
{
c2h::device_vector<type> d_output(1);

init_t init_value{};

auto error = cub::DeviceReduce::Reduce(
d_input.begin(), d_output.begin(), num_items, cuda::std::logical_or<>{}, init_value, env);
REQUIRE(error == cudaSuccess);

c2h::host_vector<type> h_input = d_input;
c2h::host_vector<type> h_expected(1);
h_expected[0] = std::accumulate(h_input.begin(), h_input.end(), type{init_value}, cuda::std::logical_or<>{});

c2h::host_vector<type> h_output = d_output;
REQUIRE(h_expected == h_output);
}
}

SECTION("bitwise xor")
SECTION("minimum")
{
c2h::device_vector<type> d_output(1);

init_t init_value{};
init_t init_value{cuda::std::numeric_limits<init_t>::max()};

auto error =
cub::DeviceReduce::Reduce(d_input.begin(), d_output.begin(), num_items, cuda::std::bit_xor<>{}, init_value, env);
cub::DeviceReduce::Reduce(d_input.begin(), d_output.begin(), num_items, cuda::minimum<init_t>{}, init_value, env);
REQUIRE(error == cudaSuccess);

c2h::host_vector<type> h_input = d_input;
c2h::host_vector<type> h_expected(1);
h_expected[0] = std::accumulate(h_input.begin(), h_input.end(), type{init_value}, cuda::std::bit_xor<type>{});
h_expected[0] = std::accumulate(h_input.begin(), h_input.end(), type{init_value}, cuda::minimum<>{});

c2h::host_vector<type> h_output = d_output;
REQUIRE(h_expected == h_output);
}

SECTION("logical or")
SECTION("maximum")
{
c2h::device_vector<type> d_output(1);

init_t init_value{};
init_t init_value{cuda::std::numeric_limits<init_t>::min()};

auto error = cub::DeviceReduce::Reduce(
d_input.begin(), d_output.begin(), num_items, cuda::std::logical_or<>{}, init_value, env);
auto error =
cub::DeviceReduce::Reduce(d_input.begin(), d_output.begin(), num_items, cuda::maximum<init_t>{}, init_value, env);
REQUIRE(error == cudaSuccess);

c2h::host_vector<type> h_input = d_input;
c2h::host_vector<type> h_expected(1);
h_expected[0] = std::accumulate(h_input.begin(), h_input.end(), type{init_value}, cuda::std::logical_or<>{});
h_expected[0] = std::accumulate(h_input.begin(), h_input.end(), type{init_value}, cuda::maximum<>{});

c2h::host_vector<type> h_output = d_output;
REQUIRE(h_expected == h_output);
Expand Down