Skip to content

Support Half/BFloat16 in var #7913

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 28, 2025
Merged

Support Half/BFloat16 in var #7913

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
10 changes: 5 additions & 5 deletions kernels/portable/cpu/op_var.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,7 +38,7 @@ void compute_variance(
in,
dim_list,
out_ix);
CTYPE_OUT mean = sum / num;
CTYPE_OUT mean = sum / static_cast<CTYPE_OUT>(num);
CTYPE_OUT sum2 = map_reduce_over_dim_list<CTYPE_IN, CTYPE_OUT>(
[mean](CTYPE_IN v) {
return (
Expand Down Expand Up @@ -90,8 +90,8 @@ Tensor& var_out(

constexpr auto name = "var.out";

ET_SWITCH_FLOAT_TYPES(in.scalar_type(), ctx, name, CTYPE_IN, [&] {
ET_SWITCH_FLOAT_TYPES(out.scalar_type(), ctx, name, CTYPE_OUT, [&] {
ET_SWITCH_FLOATHBF16_TYPES(in.scalar_type(), ctx, name, CTYPE_IN, [&] {
ET_SWITCH_FLOATHBF16_TYPES(out.scalar_type(), ctx, name, CTYPE_OUT, [&] {
compute_variance<CTYPE_IN, CTYPE_OUT>(in, out, dim_list, num, denom);
});
});
Expand Down Expand Up @@ -135,8 +135,8 @@ Tensor& var_correction_out(
const size_t num = get_reduced_dim_product(in, dim_list);
const double denom = num - correction_val;

ET_SWITCH_FLOAT_TYPES(in.scalar_type(), ctx, name, CTYPE_IN, [&] {
ET_SWITCH_FLOAT_TYPES(out.scalar_type(), ctx, name, CTYPE_OUT, [&] {
ET_SWITCH_FLOATHBF16_TYPES(in.scalar_type(), ctx, name, CTYPE_IN, [&] {
ET_SWITCH_FLOATHBF16_TYPES(out.scalar_type(), ctx, name, CTYPE_OUT, [&] {
compute_variance<CTYPE_IN, CTYPE_OUT>(in, out, dim_list, num, denom);
});
});
Expand Down
65 changes: 44 additions & 21 deletions kernels/test/op_var_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,19 @@ using exec_aten::ScalarType;
using exec_aten::Tensor;
using torch::executor::testing::TensorFactory;

namespace {
void expect_tensor_close_with_increased_tol(
const Tensor& actual,
const Tensor& expected) {
if (actual.scalar_type() == ScalarType::BFloat16 ||
actual.scalar_type() == ScalarType::Half) {
EXPECT_TENSOR_CLOSE_WITH_TOL(expected, actual, 1e-2, 1e-2);
} else {
EXPECT_TENSOR_CLOSE(expected, actual);
}
}
} // namespace

class OpVarOutTest : public OperatorTest {
protected:
Tensor& op_var_out(
Expand Down Expand Up @@ -142,7 +155,7 @@ class OpVarOutTest : public OperatorTest {
op_var_out(
self, optional_dim_list, /*unbiased=*/true, /*keepdim=*/true, out);
// clang-format off
EXPECT_TENSOR_CLOSE(out, tf_out.make(
expect_tensor_close_with_increased_tol(out, tf_out.make(
{2, 3, 1},
{
1.666667,
Expand All @@ -160,7 +173,7 @@ class OpVarOutTest : public OperatorTest {
op_var_out(
self, optional_dim_list, /*unbiased=*/true, /*keepdim=*/false, out);
// clang-format off
EXPECT_TENSOR_CLOSE(out, tf_out.make(
expect_tensor_close_with_increased_tol(out, tf_out.make(
{2, 3},
{
1.666667, 1.666667, 1.666667,
Expand All @@ -174,12 +187,14 @@ class OpVarOutTest : public OperatorTest {
optional_dim_list = ArrayRef<int64_t>{dims_2, 2};
op_var_out(
self, optional_dim_list, /*unbiased=*/true, /*keepdim=*/true, out);
EXPECT_TENSOR_CLOSE(out, tf_out.make({1, 1, 4}, {56.0, 56.0, 56.0, 56.0}));
expect_tensor_close_with_increased_tol(
out, tf_out.make({1, 1, 4}, {56.0, 56.0, 56.0, 56.0}));

out = tf_out.zeros({4});
op_var_out(
self, optional_dim_list, /*unbiased=*/true, /*keepdim=*/false, out);
EXPECT_TENSOR_CLOSE(out, tf_out.make({4}, {56.0, 56.0, 56.0, 56.0}));
expect_tensor_close_with_increased_tol(
out, tf_out.make({4}, {56.0, 56.0, 56.0, 56.0}));

// dim list with negative dimensions should work
out = tf_out.zeros({2, 1, 4});
Expand All @@ -188,7 +203,7 @@ class OpVarOutTest : public OperatorTest {
op_var_out(
self, optional_dim_list, /*unbiased=*/false, /*keepdim=*/true, out);
// clang-format off
EXPECT_TENSOR_CLOSE(out, tf_out.make(
expect_tensor_close_with_increased_tol(out, tf_out.make(
{2, 1, 4},
{
10.666667, 10.666667, 10.666667, 10.666667,
Expand All @@ -201,18 +216,19 @@ class OpVarOutTest : public OperatorTest {
out = tf_out.zeros({1, 1, 1});
optional<ArrayRef<int64_t>> null_dim_list;
op_var_out(self, null_dim_list, /*unbiased=*/true, /*keepdim=*/true, out);
EXPECT_TENSOR_CLOSE(out, tf_out.make({1, 1, 1}, {50.0}));
expect_tensor_close_with_increased_tol(out, tf_out.make({1, 1, 1}, {50.0}));

optional<ArrayRef<int64_t>> empty_dim_list{ArrayRef<int64_t>{}};
op_var_out(self, empty_dim_list, /*unbiased=*/false, /*keepdim=*/true, out);
EXPECT_TENSOR_CLOSE(out, tf_out.make({1, 1, 1}, {47.916668}));
expect_tensor_close_with_increased_tol(
out, tf_out.make({1, 1, 1}, {47.916668}));

out = tf_out.zeros({});
op_var_out(self, null_dim_list, /*unbiased=*/false, /*keepdim=*/false, out);
EXPECT_TENSOR_CLOSE(out, tf_out.make({}, {47.916668}));
expect_tensor_close_with_increased_tol(out, tf_out.make({}, {47.916668}));

op_var_out(self, empty_dim_list, /*unbiased=*/true, /*keepdim=*/false, out);
EXPECT_TENSOR_CLOSE(out, tf_out.make({}, {50.0}));
expect_tensor_close_with_increased_tol(out, tf_out.make({}, {50.0}));
}
};

Expand All @@ -227,6 +243,20 @@ class OpVarCorrectionOutTest : public OperatorTest {
return torch::executor::aten::var_outf(
context_, self, dim, correction, keepdim, out);
}

template <ScalarType DTYPE>
void test_dtype() {
TensorFactory<DTYPE> tf;

Tensor x = tf.make({2, 3}, {4.9, 4.0, 5.6, 3.8, 4.9, 5.6});
Tensor expected = tf.make({2}, {0.72693, 0.93032});
optional<Scalar> correction(1.23);
Tensor out = tf.zeros({2});

op_var_correction_out(
x, ArrayRef<int64_t>{1}, correction, /*keepdim=*/false, out);
expect_tensor_close_with_increased_tol(out, expected);
}
};

TEST_F(OpVarOutTest, InvalidDimensionListDies) {
Expand Down Expand Up @@ -303,9 +333,9 @@ TEST_F(OpVarOutTest, AllFloatInputFloatOutputPasses) {
test_var_out_dtype<ScalarType::INPUT_DTYPE, ScalarType::OUTPUT_DTYPE>();

#define TEST_ENTRY(INPUT_CTYPE, INPUT_DTYPE) \
ET_FORALL_FLOAT_TYPES_WITH2(INPUT_CTYPE, INPUT_DTYPE, TEST_KERNEL);
ET_FORALL_FLOATHBF16_TYPES_WITH2(INPUT_CTYPE, INPUT_DTYPE, TEST_KERNEL);

ET_FORALL_FLOAT_TYPES(TEST_ENTRY);
ET_FORALL_FLOATHBF16_TYPES(TEST_ENTRY);
#undef TEST_ENTRY
#undef TEST_KERNEL
}
Expand Down Expand Up @@ -387,14 +417,7 @@ TEST_F(OpVarOutTest, DynamicShapeUnbound) {
}

TEST_F(OpVarCorrectionOutTest, SmokeTest) {
TensorFactory<ScalarType::Float> tf;

Tensor x = tf.make({2, 3}, {4.9, 4.0, 5.6, 3.8, 4.9, 5.6});
Tensor expected = tf.make({2}, {0.72693, 0.93032});
optional<Scalar> correction(1.23);
Tensor out = tf.zeros({2});

op_var_correction_out(
x, ArrayRef<int64_t>{1}, correction, /*keepdim=*/false, out);
EXPECT_TENSOR_CLOSE(out, expected);
#define TEST_ENTRY(ctype, dtype) test_dtype<ScalarType::dtype>();
ET_FORALL_FLOATHBF16_TYPES(TEST_ENTRY);
#undef TEST_ENTRY
}
Loading