Skip to content

Add complex dtype support to mul #10560

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
facebook-github-bot merged 1 commit into from
May 1, 2025
Merged

Add complex dtype support to mul #10560

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
106 changes: 74 additions & 32 deletions kernels/optimized/cpu/op_mul.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -120,21 +120,47 @@ Tensor& opt_mul_out(
out,
"Failed to resize output tensor.");

ET_SWITCH_REALB_TYPES(out_type, ctx, "mul.out", CTYPE, [&]() {
using Vec = executorch::vec::Vectorized<CTYPE>;
executorch::vec::map2<CTYPE>(
[](Vec x, Vec y) { return x * y; },
out.mutable_data_ptr<CTYPE>(),
a.const_data_ptr<CTYPE>(),
b.const_data_ptr<CTYPE>(),
out.numel());
});
if (executorch::runtime::isComplexType(out_type)) {
ET_KERNEL_CHECK(
ctx, a_type == b_type && a_type == out_type, InvalidArgument, out);

ET_SWITCH_COMPLEXH_TYPES(out_type, ctx, "mul.out", CTYPE, [&]() {
using Vec = executorch::vec::Vectorized<CTYPE>;
executorch::vec::map2<CTYPE>(
[](Vec x, Vec y) { return x * y; },
out.mutable_data_ptr<CTYPE>(),
a.const_data_ptr<CTYPE>(),
b.const_data_ptr<CTYPE>(),
out.numel());
});
} else {
ET_SWITCH_REALB_TYPES(out_type, ctx, "mul.out", CTYPE, [&]() {
using Vec = executorch::vec::Vectorized<CTYPE>;
executorch::vec::map2<CTYPE>(
[](Vec x, Vec y) { return x * y; },
out.mutable_data_ptr<CTYPE>(),
a.const_data_ptr<CTYPE>(),
b.const_data_ptr<CTYPE>(),
out.numel());
});
}
} else if (selected_optimized_path != ElementwiseOptimizedPath::kNone) {
ET_SWITCH_REALB_TYPES(out_type, ctx, "mul.out", CTYPE, [&]() {
auto mul_lambda = [](auto x, auto y) { return x * y; };
return torch::executor::handle_broadcast_elementwise<CTYPE>(
ctx, mul_lambda, a, b, out, selected_optimized_path);
});
if (executorch::runtime::isComplexType(out_type)) {
ET_KERNEL_CHECK(
ctx, a_type == b_type && a_type == out_type, InvalidArgument, out);

ET_SWITCH_COMPLEXH_TYPES(out_type, ctx, "mul.out", CTYPE, [&]() {
auto mul_lambda = [](auto x, auto y) { return x * y; };
return torch::executor::handle_broadcast_elementwise<CTYPE>(
ctx, mul_lambda, a, b, out, selected_optimized_path);
});
} else {
ET_SWITCH_REALB_TYPES(out_type, ctx, "mul.out", CTYPE, [&]() {
auto mul_lambda = [](auto x, auto y) { return x * y; };
return torch::executor::handle_broadcast_elementwise<CTYPE>(
ctx, mul_lambda, a, b, out, selected_optimized_path);
});
}
} else {
ScalarType common_type =
promoteTypes(a_type, b_type, /*half_to_float*/ true);
Expand All @@ -146,26 +172,42 @@ Tensor& opt_mul_out(
InvalidArgument,
out);

ET_SWITCH_REALHBBF16_TYPES(a_type, ctx, "mul.out", CTYPE_A, [&]() {
ET_SWITCH_REALHBBF16_TYPES(b_type, ctx, "mul.out", CTYPE_B, [&]() {
using CTYPE_IN = typename torch::executor::
promote_types<CTYPE_A, CTYPE_B, /*half_to_float*/ true>::type;
ET_DCHECK(CppTypeToScalarType<CTYPE_IN>::value == common_type);
ET_SWITCH_REALHBBF16_TYPES(out_type, ctx, "mul.out", CTYPE_OUT, [&]() {
apply_binary_elementwise_fn<CTYPE_A, CTYPE_B, CTYPE_OUT>(
[](const CTYPE_A val_a, const CTYPE_B val_b) {
CTYPE_IN a_casted = static_cast<CTYPE_IN>(val_a);
CTYPE_IN b_casted = static_cast<CTYPE_IN>(val_b);
CTYPE_IN value = a_casted * b_casted;

return static_cast<CTYPE_OUT>(value);
},
a,
b,
out);
if (executorch::runtime::isComplexType(a_type) ||
executorch::runtime::isComplexType(b_type) ||
executorch::runtime::isComplexType(out_type)) {
ET_KERNEL_CHECK(
ctx, a_type == b_type && a_type == out_type, InvalidArgument, out);

ET_SWITCH_COMPLEXH_TYPES(out_type, ctx, "mul.out", CTYPE, [&]() {
apply_binary_elementwise_fn<CTYPE, CTYPE, CTYPE>(
[](const CTYPE val_a, const CTYPE val_b) { return val_a * val_b; },
a,
b,
out);
});
} else {
ET_SWITCH_REALHBBF16_TYPES(a_type, ctx, "mul.out", CTYPE_A, [&]() {
ET_SWITCH_REALHBBF16_TYPES(b_type, ctx, "mul.out", CTYPE_B, [&]() {
using CTYPE_IN = typename torch::executor::
promote_types<CTYPE_A, CTYPE_B, /*half_to_float*/ true>::type;
ET_DCHECK(CppTypeToScalarType<CTYPE_IN>::value == common_type);
ET_SWITCH_REALHBBF16_TYPES(
out_type, ctx, "mul.out", CTYPE_OUT, [&]() {
apply_binary_elementwise_fn<CTYPE_A, CTYPE_B, CTYPE_OUT>(
[](const CTYPE_A val_a, const CTYPE_B val_b) {
CTYPE_IN a_casted = static_cast<CTYPE_IN>(val_a);
CTYPE_IN b_casted = static_cast<CTYPE_IN>(val_b);
CTYPE_IN value = a_casted * b_casted;

return static_cast<CTYPE_OUT>(value);
},
a,
b,
out);
});
});
});
});
}
}

return out;
Expand Down
43 changes: 30 additions & 13 deletions kernels/portable/cpu/op_mul.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,25 +47,42 @@ Tensor& mul_out(
ET_KERNEL_CHECK(
ctx,
(executorch::runtime::isRealType(compute_type) ||
executorch::runtime::isComplexType(compute_type) ||
compute_type == ScalarType::Bool),
InvalidArgument,
out);

ET_SWITCH_REALB_TYPES(compute_type, ctx, op_name, CTYPE_COMPUTE, [&]() {
utils::apply_bitensor_elementwise_fn<
CTYPE_COMPUTE,
op_name,
utils::SupportedTensorDtypes::REALHBBF16>(
[](const CTYPE_COMPUTE val_a, const CTYPE_COMPUTE val_b) {
return val_a * val_b;
},
if (executorch::runtime::isComplexType(compute_type)) {
ET_KERNEL_CHECK(
ctx,
a,
utils::SupportedTensorDtypes::REALHBBF16,
b,
utils::SupportedTensorDtypes::REALHBBF16,
a.scalar_type() == b.scalar_type() &&
a.scalar_type() == out.scalar_type(),
InvalidArgument,
out);
});
ET_SWITCH_COMPLEXH_TYPES(out.scalar_type(), ctx, "mul.out", CTYPE, [&]() {
apply_binary_elementwise_fn<CTYPE, CTYPE, CTYPE>(
[](const CTYPE val_a, const CTYPE val_b) { return val_a * val_b; },
a,
b,
out);
});
} else {
ET_SWITCH_REALB_TYPES(compute_type, ctx, op_name, CTYPE_COMPUTE, [&]() {
utils::apply_bitensor_elementwise_fn<
CTYPE_COMPUTE,
op_name,
utils::SupportedTensorDtypes::REALHBBF16>(
[](const CTYPE_COMPUTE val_a, const CTYPE_COMPUTE val_b) {
return val_a * val_b;
},
ctx,
a,
utils::SupportedTensorDtypes::REALHBBF16,
b,
utils::SupportedTensorDtypes::REALHBBF16,
out);
});
}

return out;
}
Expand Down
42 changes: 42 additions & 0 deletions kernels/test/op_mul_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -322,6 +322,38 @@ class OpMulOutTest : public OperatorTest {
EXPECT_TENSOR_CLOSE(op_mul_out(a, b, out), expected);
EXPECT_TENSOR_CLOSE(op_mul_out(b, a, out), expected);
}

template <typename CTYPE, ScalarType DTYPE>
void test_complex_dtype() {
TensorFactory<DTYPE> tf;
const std::vector<int32_t> sizes = {2, 2};

// Create complex tensors with real and imaginary parts
Tensor x =
tf.make(sizes, {CTYPE(1, 2), CTYPE(3, 4), CTYPE(5, 6), CTYPE(7, 8)});

Tensor y =
tf.make(sizes, {CTYPE(2, 3), CTYPE(4, 5), CTYPE(6, 7), CTYPE(8, 9)});

// Expected result: (a+bi) * (c+di) = (ac-bd) + (ad+bc)i
// (1+2i) * (2+3i) = (1*2-2*3) + (1*3+2*2)i = -4 + 7i
// (3+4i) * (4+5i) = (3*4-4*5) + (3*5+4*4)i = -8 + 31i
// (5+6i) * (6+7i) = (5*6-6*7) + (5*7+6*6)i = -12 + 71i
// (7+8i) * (8+9i) = (7*8-8*9) + (7*9+8*8)i = -16 + 127i
Tensor expected = tf.make(
sizes, {CTYPE(-4, 7), CTYPE(-8, 31), CTYPE(-12, 71), CTYPE(-16, 127)});

Tensor out = tf.make(
{2, 2},
{
CTYPE(0, 0),
CTYPE(0, 0),
CTYPE(0, 0),
CTYPE(0, 0),
});
op_mul_out(x, y, out);
EXPECT_TENSOR_CLOSE(out, expected);
}
};

class OpMulScalarOutTest : public OperatorTest {
Expand Down Expand Up @@ -472,6 +504,16 @@ TEST_F(OpMulOutTest, BothScalarInputBroadcastTest) {
test_both_scalar_input_broadcast<ScalarType::BFloat16>();
}

TEST_F(OpMulOutTest, AllComplexDtypesSupported) {
#define TEST_ENTRY(ctype, dtype) test_complex_dtype<ctype, ScalarType::dtype>();
if (torch::executor::testing::SupportedFeatures::get()->is_aten) {
ET_FORALL_COMPLEX_TYPES(TEST_ENTRY);
} else {
ET_FORALL_COMPLEXH_TYPES(TEST_ENTRY);
}
#undef TEST_ENTRY
}

TEST_F(OpMulOutTest, MismatchedOutputShapesDies) {
if (SupportedFeatures::get()->is_aten) {
GTEST_SKIP() << "ATen currently supports mismatched shapes";
Expand Down
Loading