Skip to content

dpctl.tensor.floor_divide fixed for signed 0 output #1271

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
Jul 30, 2023
Merged

dpctl.tensor.floor_divide fixed for signed 0 output #1271

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
c31d020
floor_divide fixed for signed 0 output
ndgrigorian Jun 30, 2023
1dd01fa
Reduced number of computations for floor_divide between integers
ndgrigorian Jul 29, 2023
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
89 changes: 40 additions & 49 deletions dpctl/tensor/libtensor/include/kernels/elementwise_functions/floor_divide.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,62 +52,60 @@ namespace tu_ns = dpctl::tensor::type_utils;
template <typename argT1, typename argT2, typename resT>
struct FloorDivideFunctor
{

using supports_sg_loadstore = std::negation<
std::disjunction<tu_ns::is_complex<argT1>, tu_ns::is_complex<argT2>>>;
using supports_vec = std::negation<
std::disjunction<tu_ns::is_complex<argT1>, tu_ns::is_complex<argT2>>>;
using supports_sg_loadstore = std::true_type;
using supports_vec = std::true_type;

resT operator()(const argT1 &in1, const argT2 &in2)
{
auto tmp = in1 / in2;
if constexpr (std::is_integral_v<decltype(tmp)>) {
if constexpr (std::is_unsigned_v<decltype(tmp)>) {
return (in2 == argT2(0)) ? resT(0) : tmp;
if constexpr (std::is_integral_v<argT1> || std::is_integral_v<argT2>) {
if (in2 == argT2(0)) {
return resT(0);
}
if constexpr (std::is_signed_v<argT1> || std::is_signed_v<argT2>) {
auto div = in1 / in2;
auto mod = in1 % in2;
auto corr = (mod != 0 && l_xor(mod < 0, in2 < 0));
return (div - corr);
}
else {
if (in2 == argT2(0)) {
return resT(0);
}
else {
auto rem = in1 % in2;
auto corr = (rem != 0 && ((rem < 0) != (in2 < 0)));
return (tmp - corr);
}
return (in1 / in2);
}
}
else {
return sycl::floor(tmp);
auto div = in1 / in2;
return (div == resT(0)) ? div : resT(std::floor(div));
}
}

template <int vec_sz>
sycl::vec<resT, vec_sz> operator()(const sycl::vec<argT1, vec_sz> &in1,
const sycl::vec<argT2, vec_sz> &in2)
{
auto tmp = in1 / in2;
using tmpT = typename decltype(tmp)::element_type;
if constexpr (std::is_integral_v<tmpT>) {
if constexpr (std::is_signed_v<tmpT>) {
auto rem_tmp = in1 % in2;
if constexpr (std::is_integral_v<resT>) {
sycl::vec<resT, vec_sz> res;
#pragma unroll
for (int i = 0; i < vec_sz; ++i) {
if (in2[i] == argT2(0)) {
tmp[i] = tmpT(0);
}
else {
tmpT corr = (rem_tmp[i] != 0 &&
((rem_tmp[i] < 0) != (in2[i] < 0)));
tmp[i] -= corr;
for (int i = 0; i < vec_sz; ++i) {
if (in2[i] == argT2(0)) {
res[i] = resT(0);
}
else {
res[i] = in1[i] / in2[i];
if constexpr (std::is_signed_v<resT>) {
auto mod = in1[i] % in2[i];
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This performs a second division. How about mod = in1[i] - res[i] * in2[i] instead?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I will try this out as well, I'd like to see how the performance looks.

I have also seen sources suggest that the remainder is, in some cases, a byproduct of division and the compiler can optimize these operations when nearby.

i.e., here under notes

auto corr = (mod != 0 && l_xor(mod < 0, in2[i] < 0));
res[i] -= corr;
}
}
}
else {
return res;
}
else {
auto tmp = in1 / in2;
using tmpT = typename decltype(tmp)::element_type;
#pragma unroll
for (int i = 0; i < vec_sz; ++i) {
if (in2[i] == argT2(0)) {
tmp[i] = tmpT(0);
}
for (int i = 0; i < vec_sz; ++i) {
if (in2[i] != argT2(0)) {
tmp[i] = std::floor(tmp[i]);
}
}
if constexpr (std::is_same_v<resT, tmpT>) {
Expand All @@ -118,19 +116,12 @@ struct FloorDivideFunctor
return vec_cast<resT, tmpT, vec_sz>(tmp);
}
}
else {
sycl::vec<resT, vec_sz> res = sycl::floor(tmp);
if constexpr (std::is_same_v<resT,
typename decltype(res)::element_type>)
{
return res;
}
else {
using dpctl::tensor::type_utils::vec_cast;
return vec_cast<resT, typename decltype(res)::element_type,
vec_sz>(res);
}
}
}

private:
bool l_xor(bool b1, bool b2) const
{
return (b1 != b2);
}
};

Expand Down
50 changes: 36 additions & 14 deletions dpctl/tests/elementwise/test_floor_divide.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -203,16 +203,6 @@ def test_floor_divide_gh_1247():
dpt.asnumpy(res), np.full(res.shape, -1, dtype=res.dtype)
)

# attempt to invoke sycl::vec overload using a larger array
x = dpt.arange(-64, 65, 1, dtype="i4")
np.testing.assert_array_equal(
dpt.asnumpy(dpt.floor_divide(x, 3)), np.floor_divide(dpt.asnumpy(x), 3)
)
np.testing.assert_array_equal(
dpt.asnumpy(dpt.floor_divide(x, -3)),
np.floor_divide(dpt.asnumpy(x), -3),
)


@pytest.mark.parametrize("dtype", _no_complex_dtypes[1:9])
def test_floor_divide_integer_zero(dtype):
Expand All @@ -226,10 +216,42 @@ def test_floor_divide_integer_zero(dtype):
dpt.asnumpy(res), np.zeros(x.shape, dtype=res.dtype)
)

# attempt to invoke sycl::vec overload using a larger array
x = dpt.arange(129, dtype=dtype, sycl_queue=q)
y = dpt.zeros_like(x, sycl_queue=q)

def test_floor_divide_special_cases():
q = get_queue_or_skip()

x = dpt.empty(1, dtype="f4", sycl_queue=q)
y = dpt.empty_like(x)
x[0], y[0] = dpt.inf, dpt.inf
res = dpt.floor_divide(x, y)
with np.errstate(all="ignore"):
res_np = np.floor_divide(dpt.asnumpy(x), dpt.asnumpy(y))
np.testing.assert_array_equal(dpt.asnumpy(res), res_np)

x[0], y[0] = 0.0, -1.0
res = dpt.floor_divide(x, y)
x_np = dpt.asnumpy(x)
y_np = dpt.asnumpy(y)
res_np = np.floor_divide(x_np, y_np)
np.testing.assert_array_equal(dpt.asnumpy(res), res_np)

res = dpt.floor_divide(y, x)
with np.errstate(all="ignore"):
res_np = np.floor_divide(y_np, x_np)
np.testing.assert_array_equal(dpt.asnumpy(res), res_np)

x[0], y[0] = -1.0, dpt.inf
res = dpt.floor_divide(x, y)
np.testing.assert_array_equal(
dpt.asnumpy(res), np.zeros(x.shape, dtype=res.dtype)
dpt.asnumpy(res), np.asarray([-0.0], dtype="f4")
)

res = dpt.floor_divide(y, x)
np.testing.assert_array_equal(
dpt.asnumpy(res), np.asarray([-dpt.inf], dtype="f4")
)

x[0], y[0] = 1.0, dpt.nan
res = dpt.floor_divide(x, y)
res_np = np.floor_divide(dpt.asnumpy(x), dpt.asnumpy(y))
np.testing.assert_array_equal(dpt.asnumpy(res), res_np)