Skip to content

Support Half/BFloat16 in cdist #7800

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
swolchok merged 1 commit into from
Jan 23, 2025
Merged

Support Half/BFloat16 in cdist #7800

Show file tree
Hide file tree
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
2 changes: 1 addition & 1 deletion kernels/portable/cpu/op_cdist_forward.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -162,7 +162,7 @@ Tensor& _cdist_forward_out(
ScalarType out_type = out.scalar_type();
constexpr auto name = "_cdist_forward.out";

ET_SWITCH_FLOAT_TYPES(
ET_SWITCH_FLOATHBF16_TYPES(
out_type, ctx, name, CTYPE, [&] { cdist<CTYPE>(x1, x2, out, p); });

return out;
Expand Down
170 changes: 93 additions & 77 deletions kernels/test/op_cdist_forward_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,89 +40,105 @@ class OpCdistForwardOutTest : public ::testing::Test {
// first.
torch::executor::runtime_init();
}
};

TEST_F(OpCdistForwardOutTest, SmokeTest) {
TensorFactory<ScalarType::Float> tfFloat;
template <ScalarType DTYPE>
void test_dtype() {
TensorFactory<DTYPE> tf;

Tensor x1 = tf.make({2, 1, 4, 3}, {0, 1, 2, 3, 5, 4, 3, -3, 7, 1, 6, 2,
-1, 5, 1, 1, -2, 1, 5, 4, 3, 2, -1, 5});
Tensor x2 = tf.make(
{1, 2, 5, 3}, {0, 1, 2, 3, 5, -3, 7, 1, 6, 2, -1, 5, 1, 1, -2,
4, 3, 2, -1, 5, 1, 1, -2, 1, 5, 4, 3, 2, -1, 5});
optional<int64_t> compute_mode = optional<int64_t>();

Tensor x1 =
tfFloat.make({2, 1, 4, 3}, {0, 1, 2, 3, 5, 4, 3, -3, 7, 1, 6, 2,
-1, 5, 1, 1, -2, 1, 5, 4, 3, 2, -1, 5});
Tensor x2 = tfFloat.make(
{1, 2, 5, 3}, {0, 1, 2, 3, 5, -3, 7, 1, 6, 2, -1, 5, 1, 1, -2,
4, 3, 2, -1, 5, 1, 1, -2, 1, 5, 4, 3, 2, -1, 5});
optional<int64_t> compute_mode = optional<int64_t>();
Tensor out = tf.zeros({2, 2, 4, 5});

Tensor out = tfFloat.zeros({2, 2, 4, 5});
Tensor l0 = tf.make(
{2, 2, 4, 5},
{0., 3., 2., 3., 2., 3., 1., 3., 3., 3., 3., 2., 3., 3., 3., 2.,
3., 3., 3., 2., 2., 3., 3., 3., 3., 3., 2., 3., 3., 3., 3., 3.,
3., 3., 3., 2., 3., 2., 3., 3., 3., 2., 3., 3., 3., 3., 3., 3.,
3., 2., 3., 3., 3., 3., 3., 3., 3., 3., 0., 3., 3., 0., 2., 3.,
3., 3., 2., 0., 3., 3., 3., 3., 3., 0., 3., 3., 3., 3., 3., 0.});
op_cdist_forward_out(x1, x2, 0.0, compute_mode, out);
EXPECT_TENSOR_CLOSE(out, l0);

Tensor l0 = tfFloat.make(
{2, 2, 4, 5},
{0., 3., 2., 3., 2., 3., 1., 3., 3., 3., 3., 2., 3., 3., 3., 2.,
3., 3., 3., 2., 2., 3., 3., 3., 3., 3., 2., 3., 3., 3., 3., 3.,
3., 3., 3., 2., 3., 2., 3., 3., 3., 2., 3., 3., 3., 3., 3., 3.,
3., 2., 3., 3., 3., 3., 3., 3., 3., 3., 0., 3., 3., 0., 2., 3.,
3., 3., 2., 0., 3., 3., 3., 3., 3., 0., 3., 3., 3., 3., 3., 0.});
op_cdist_forward_out(x1, x2, 0.0, compute_mode, out);
EXPECT_TENSOR_CLOSE(out, l0);
Tensor l1 = tf.make(
{2, 2, 4, 5},
{0., 12., 11., 7., 5., 9., 7., 10., 8., 12., 12., 18., 9., 5.,
15., 6., 8., 15., 11., 9., 6., 6., 5., 9., 7., 5., 7., 12.,
4., 8., 12., 18., 9., 13., 5., 6., 4., 9., 7., 11., 6., 8.,
17., 13., 9., 5., 13., 14., 6., 6., 9., 9., 8., 10., 12., 7.,
15., 8., 0., 10., 8., 0., 9., 9., 13., 9., 9., 0., 12., 6.,
3., 9., 12., 0., 10., 9., 13., 6., 10., 0.});
op_cdist_forward_out(x1, x2, 1.0, compute_mode, out);
EXPECT_TENSOR_CLOSE(out, l1);

Tensor l1 = tfFloat.make(
{2, 2, 4, 5},
{0., 12., 11., 7., 5., 9., 7., 10., 8., 12., 12., 18., 9., 5.,
15., 6., 8., 15., 11., 9., 6., 6., 5., 9., 7., 5., 7., 12.,
4., 8., 12., 18., 9., 13., 5., 6., 4., 9., 7., 11., 6., 8.,
17., 13., 9., 5., 13., 14., 6., 6., 9., 9., 8., 10., 12., 7.,
15., 8., 0., 10., 8., 0., 9., 9., 13., 9., 9., 0., 12., 6.,
3., 9., 12., 0., 10., 9., 13., 6., 10., 0.});
op_cdist_forward_out(x1, x2, 1.0, compute_mode, out);
EXPECT_TENSOR_CLOSE(out, l1);
Tensor l2 = tf.make(
{2, 2, 4, 5},
{0.00000000, 7.07106781, 8.06225777, 4.12310553, 4.12310553,
5.38516474, 7.00000000, 6.00000000, 6.16441393, 7.48331499,
7.07106781, 12.80624866, 5.74456263, 3.00000000, 10.04987526,
5.09901953, 5.47722578, 8.77496433, 7.68114567, 6.40312433,
4.47213602, 4.24264050, 3.31662488, 5.91608000, 4.12310553,
3.00000000, 5.00000000, 7.87400770, 2.44948983, 6.16441393,
7.87400770, 10.77032948, 6.40312433, 8.30662346, 3.00000000,
4.24264050, 2.44948983, 8.06225777, 4.58257580, 7.68114567,
4.24264050, 5.65685415, 10.24695110, 7.81024981, 5.38516474,
3.31662488, 8.30662346, 8.36660004, 4.24264050, 4.24264050,
5.91608000, 6.40312433, 4.69041586, 6.16441393, 7.07106781,
4.12310553, 10.04987526, 5.47722578, 0.00000000, 7.34846926,
5.47722578, 0.00000000, 7.28010988, 6.40312433, 7.81024981,
5.91608000, 7.28010988, 0.00000000, 7.48331499, 4.24264050,
1.73205078, 6.40312433, 7.48331499, 0.00000000, 6.16441393,
5.38516474, 7.81024981, 4.24264050, 6.16441393, 0.00000000});
op_cdist_forward_out(x1, x2, 2.0, compute_mode, out);
EXPECT_TENSOR_CLOSE(out, l2);

Tensor l2 = tfFloat.make(
{2, 2, 4, 5},
{0.00000000, 7.07106781, 8.06225777, 4.12310553, 4.12310553,
5.38516474, 7.00000000, 6.00000000, 6.16441393, 7.48331499,
7.07106781, 12.80624866, 5.74456263, 3.00000000, 10.04987526,
5.09901953, 5.47722578, 8.77496433, 7.68114567, 6.40312433,
4.47213602, 4.24264050, 3.31662488, 5.91608000, 4.12310553,
3.00000000, 5.00000000, 7.87400770, 2.44948983, 6.16441393,
7.87400770, 10.77032948, 6.40312433, 8.30662346, 3.00000000,
4.24264050, 2.44948983, 8.06225777, 4.58257580, 7.68114567,
4.24264050, 5.65685415, 10.24695110, 7.81024981, 5.38516474,
3.31662488, 8.30662346, 8.36660004, 4.24264050, 4.24264050,
5.91608000, 6.40312433, 4.69041586, 6.16441393, 7.07106781,
4.12310553, 10.04987526, 5.47722578, 0.00000000, 7.34846926,
5.47722578, 0.00000000, 7.28010988, 6.40312433, 7.81024981,
5.91608000, 7.28010988, 0.00000000, 7.48331499, 4.24264050,
1.73205078, 6.40312433, 7.48331499, 0.00000000, 6.16441393,
5.38516474, 7.81024981, 4.24264050, 6.16441393, 0.00000000});
op_cdist_forward_out(x1, x2, 2.0, compute_mode, out);
EXPECT_TENSOR_CLOSE(out, l2);
Tensor l3 = tf.make(
{2, 2, 4, 5},
{0.00000000, 6.00000000, 7.41079521, 3.50339794, 4.02072573,
4.62606478, 7.00000000, 5.14256334, 6.01846170, 6.60385466,
6.00000000, 11.47758675, 5.05277443, 2.57128167, 9.28704357,
5.01329803, 5.11722994, 7.39863634, 7.18551636, 5.73879337,
4.16016769, 4.04124022, 3.07231688, 5.34848118, 3.50339794,
2.57128167, 4.49794149, 7.23042679, 2.15443468, 6.01846170,
6.99319077, 9.25212955, 6.08220196, 7.45903587, 2.57128167,
3.77976322, 2.15443468, 8.00520515, 4.17933941, 7.18551636,
4.04124022, 5.03968430, 8.88326645, 6.74599648, 4.62606478,
3.07231688, 7.45903587, 7.16609573, 4.04124022, 3.77976322,
5.34848118, 6.08220196, 3.95789170, 5.42883539, 6.00000000,
3.50339794, 9.00000000, 5.11722994, 0.00000000, 7.06069660,
5.11722994, 0.00000000, 7.05400419, 6.08220196, 6.74599648,
5.34848118, 7.05400419, 0.00000000, 6.60385466, 4.04124022,
1.44224954, 6.08220196, 6.60385466, 0.00000000, 5.42883539,
4.62606478, 6.74599648, 4.04124022, 5.42883539, 0.00000000});
op_cdist_forward_out(x1, x2, 3.0, compute_mode, out);
if (DTYPE == ScalarType::BFloat16) {
EXPECT_TENSOR_CLOSE_WITH_TOL(
out,
l3,
1e-2,
executorch::runtime::testing::internal::kDefaultBFloat16Atol);
Copy link
Contributor

@manuelcandales manuelcandales Jan 22, 2025

Choose a reason for hiding this comment

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

Why do you need to pass the default atol here?

Copy link
Contributor Author

@swolchok swolchok Jan 22, 2025

Choose a reason for hiding this comment

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

macros don't have default arguments

} else {
EXPECT_TENSOR_CLOSE(out, l3);
}

Tensor l3 = tfFloat.make(
{2, 2, 4, 5},
{0.00000000, 6.00000000, 7.41079521, 3.50339794, 4.02072573, 4.62606478,
7.00000000, 5.14256334, 6.01846170, 6.60385466, 6.00000000, 11.47758675,
5.05277443, 2.57128167, 9.28704357, 5.01329803, 5.11722994, 7.39863634,
7.18551636, 5.73879337, 4.16016769, 4.04124022, 3.07231688, 5.34848118,
3.50339794, 2.57128167, 4.49794149, 7.23042679, 2.15443468, 6.01846170,
6.99319077, 9.25212955, 6.08220196, 7.45903587, 2.57128167, 3.77976322,
2.15443468, 8.00520515, 4.17933941, 7.18551636, 4.04124022, 5.03968430,
8.88326645, 6.74599648, 4.62606478, 3.07231688, 7.45903587, 7.16609573,
4.04124022, 3.77976322, 5.34848118, 6.08220196, 3.95789170, 5.42883539,
6.00000000, 3.50339794, 9.00000000, 5.11722994, 0.00000000, 7.06069660,
5.11722994, 0.00000000, 7.05400419, 6.08220196, 6.74599648, 5.34848118,
7.05400419, 0.00000000, 6.60385466, 4.04124022, 1.44224954, 6.08220196,
6.60385466, 0.00000000, 5.42883539, 4.62606478, 6.74599648, 4.04124022,
5.42883539, 0.00000000});
op_cdist_forward_out(x1, x2, 3.0, compute_mode, out);
EXPECT_TENSOR_CLOSE(out, l3);
Tensor linf = tf.make(
{2, 2, 4, 5},
{0., 5., 7., 3., 4., 4., 7., 4., 6., 6., 5., 10., 4., 2., 9., 5.,
5., 6., 7., 5., 4., 4., 3., 5., 3., 2., 4., 7., 2., 6., 6., 8.,
6., 7., 2., 3., 2., 8., 4., 7., 4., 4., 8., 6., 4., 3., 7., 6.,
4., 3., 5., 6., 3., 5., 5., 3., 8., 5., 0., 7., 5., 0., 7., 6.,
6., 5., 7., 0., 6., 4., 1., 6., 6., 0., 5., 4., 6., 4., 5., 0.});
op_cdist_forward_out(x1, x2, INFINITY, compute_mode, out);
EXPECT_TENSOR_CLOSE(out, linf);
}
};

Tensor linf = tfFloat.make(
{2, 2, 4, 5},
{0., 5., 7., 3., 4., 4., 7., 4., 6., 6., 5., 10., 4., 2., 9., 5.,
5., 6., 7., 5., 4., 4., 3., 5., 3., 2., 4., 7., 2., 6., 6., 8.,
6., 7., 2., 3., 2., 8., 4., 7., 4., 4., 8., 6., 4., 3., 7., 6.,
4., 3., 5., 6., 3., 5., 5., 3., 8., 5., 0., 7., 5., 0., 7., 6.,
6., 5., 7., 0., 6., 4., 1., 6., 6., 0., 5., 4., 6., 4., 5., 0.});
op_cdist_forward_out(x1, x2, INFINITY, compute_mode, out);
EXPECT_TENSOR_CLOSE(out, linf);
TEST_F(OpCdistForwardOutTest, SmokeTest) {
#define TEST_ENTRY(ctype, dtype) test_dtype<ScalarType::dtype>();
ET_FORALL_FLOATHBF16_TYPES(TEST_ENTRY);
#undef TEST_ENTRY
}
3 changes: 3 additions & 0 deletions runtime/core/exec_aten/testing_util/tensor_util.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -80,6 +80,9 @@ double default_atol_for_type(ScalarType t) {
if (t == ScalarType::Half) {
return internal::kDefaultHalfAtol;
}
if (t == ScalarType::BFloat16) {
return internal::kDefaultBFloat16Atol;
Copy link
Contributor

@manuelcandales manuelcandales Jan 22, 2025

Choose a reason for hiding this comment

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

This is great, I wasn't aware of these defaults. Do these play well with the EXPECT_TENSOR_CLOSE_WITH_TOL macro?

Copy link
Contributor Author

@swolchok swolchok Jan 22, 2025

Choose a reason for hiding this comment

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

yes, though atol matters a lot less to me than rtol in all these diffs

Copy link
Contributor

@manuelcandales manuelcandales Jan 23, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

see my comment on #7807 , I think you should also introduce kDefaultHalfRtol and kDefaultBFloat16Rtol.

Copy link
Contributor Author

@swolchok swolchok Jan 23, 2025

Choose a reason for hiding this comment

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

replied; 1e-2 is not needed for all ops

}
return internal::kDefaultAtol;
}
} // namespace
Expand Down
Loading