Skip to content

Refactor op_div: fix bug + enable Half/Bfloat16 #13740

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 1 commit into from
Aug 28, 2025
Merged

Refactor op_div: fix bug + enable Half/Bfloat16 #13740

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
179 changes: 80 additions & 99 deletions kernels/optimized/cpu/op_div.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,7 @@
#include <ATen/cpu/vec/vec.h>
#include <executorch/kernels/optimized/cpu/binary_ops.h>
#include <executorch/kernels/portable/cpu/scalar_utils.h>
#include <executorch/kernels/portable/cpu/util/broadcast_util.h>
#include <executorch/kernels/portable/cpu/util/elementwise_util.h>
#include <executorch/runtime/kernel/kernel_includes.h>
#include <executorch/runtime/platform/assert.h>

Expand All @@ -20,7 +20,7 @@ namespace native {

namespace {

ScalarType get_compute_type(ScalarType a_type, ScalarType b_type) {
ScalarType get_common_type(ScalarType a_type, ScalarType b_type) {
ET_CHECK(
!isComplexType(a_type) && !isQIntType(a_type) && !isBitsType(a_type));
ET_CHECK(
Expand All @@ -43,14 +43,27 @@ Tensor& opt_div_out(
const Tensor& a,
const Tensor& b,
Tensor& out) {
(void)ctx;
// Check Dim Order
ET_KERNEL_CHECK(
ctx, tensors_have_same_dim_order(a, b, out), InvalidArgument, out);

// Resize
ET_KERNEL_CHECK(
ctx,
resize_to_broadcast_target_size(a, b, out) == Error::Ok,
InvalidArgument,
out);

// @lint-ignore CLANGTIDY facebook-hte-CArray
static constexpr const char op_name[] = "div.out";

ScalarType a_type = a.scalar_type();
ScalarType b_type = b.scalar_type();
ScalarType out_type = out.scalar_type();

if (a.numel() == 1 || b.numel() == 1) {
if (a_type == b_type && a_type == out_type && a_type != ScalarType::Half) {
if (a_type == b_type && a_type == out_type && a_type != ScalarType::Half &&
a_type != ScalarType::BFloat16) {
const Tensor* tensor;
const Tensor* scalar;
ScalarType tensor_type;
Expand All @@ -66,13 +79,8 @@ Tensor& opt_div_out(
scalar = &b;
scalar_type = b_type;
}
ET_KERNEL_CHECK(
ctx,
resize_to_broadcast_target_size(a, b, out) == Error::Ok,
InvalidArgument,
out);
ET_SWITCH_REALB_TYPES(tensor_type, ctx, "div.out", CTYPE, [&]() {
ET_SWITCH_REALB_TYPES(scalar_type, ctx, "div.out", CTYPE_SCALAR, [&]() {
ET_SWITCH_REALB_TYPES(tensor_type, ctx, op_name, CTYPE, [&]() {
ET_SWITCH_REALB_TYPES(scalar_type, ctx, op_name, CTYPE_SCALAR, [&]() {
CTYPE_SCALAR scalar_val = *scalar->const_data_ptr<CTYPE_SCALAR>();
CTYPE scalar_casted = static_cast<CTYPE>(scalar_val);

Expand Down Expand Up @@ -101,16 +109,7 @@ Tensor& opt_div_out(

auto selected_optimized_path = select_optimized_path(a, b, out);
if (selected_optimized_path == ElementwiseOptimizedPath::kTreatAs1d) {
// Resize for dynamic shape
auto error = resize_tensor(out, a.sizes());
ET_KERNEL_CHECK_MSG(
ctx,
error == Error::Ok,
InvalidArgument,
out,
"Failed to resize output tensor.");

ET_SWITCH_REAL_TYPES_AND(Bool, out_type, ctx, "div.out", CTYPE, [&]() {
ET_SWITCH_REALB_TYPES(out_type, ctx, op_name, CTYPE, [&]() {
using Vec = at::vec::Vectorized<CTYPE>;
at::vec::map2<CTYPE>(
[](Vec x, Vec y) { return x / y; },
Expand All @@ -122,7 +121,7 @@ Tensor& opt_div_out(
} else if (selected_optimized_path != ElementwiseOptimizedPath::kNone) {
// Reason for using alpha is becasuse handle_broadcast_elementwise
// is used for add and sub as well:
ET_SWITCH_REALB_TYPES(out_type, ctx, "div.out", CTYPE, [&]() {
ET_SWITCH_REALB_TYPES(out_type, ctx, op_name, CTYPE, [&]() {
if (selected_optimized_path ==
ElementwiseOptimizedPath::kBroadcast2dBy1dReverseArguments ||
selected_optimized_path ==
Expand All @@ -139,33 +138,21 @@ Tensor& opt_div_out(
}
});
} else {
ScalarType common_type = get_compute_type(a_type, b_type);
ET_KERNEL_CHECK(ctx, canCast(common_type, out_type), InvalidArgument, out);

ET_KERNEL_CHECK(
ctx,
resize_to_broadcast_target_size(a, b, out) == Error::Ok,
InvalidArgument,
out);

ET_SWITCH_REALB_TYPES(a_type, ctx, "div.out", CTYPE_A, [&]() {
ET_SWITCH_REALB_TYPES(b_type, ctx, "div.out", CTYPE_B, [&]() {
ET_SWITCH_REALB_TYPES(common_type, ctx, "div.out", CTYPE_IN, [&]() {
ET_SWITCH_REALB_TYPES(out_type, ctx, "div.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);
});
});
});
ScalarType common_type = get_common_type(a.scalar_type(), b.scalar_type());
ScalarType compute_type = utils::get_compute_type(common_type);

ET_SWITCH_FLOAT_TYPES(compute_type, ctx, op_name, CTYPE_COMPUTE, [&]() {
utils::apply_bitensor_elementwise_fn<
CTYPE_COMPUTE,
op_name,
utils::SupportedTensorDtypes::FLOATHBF16>(
[](const auto val_a, const auto val_b) { return val_a / val_b; },
ctx,
a,
utils::SupportedTensorDtypes::REALHBBF16,
b,
utils::SupportedTensorDtypes::REALHBBF16,
out);
});
}

Expand All @@ -177,63 +164,57 @@ Tensor& opt_div_scalar_out(
const Tensor& a,
const Scalar& b,
Tensor& out) {
(void)ctx;

ScalarType a_type = a.scalar_type();
ScalarType b_type = utils::get_scalar_dtype(b);
ScalarType common_type = isFloatingType(a_type) ? a_type : ScalarType::Float;
ScalarType out_type = out.scalar_type();

ET_CHECK(common_type == out_type);

// Resize for dynamic shape
auto error = resize_tensor(out, a.sizes());
ET_CHECK_MSG(error == Error::Ok, "Failed to resize output tensor.");

if (a_type == common_type && a_type == out_type) {
ET_SWITCH_REAL_TYPES(a_type, ctx, "div.Scalar_out", CTYPE, [&]() {
ET_SWITCH_REAL_TYPES_AND(
Bool, b_type, ctx, "div.Scalar_out", CTYPE_B, [&]() {
CTYPE_B b_val;
ET_EXTRACT_SCALAR(b, b_val);
CTYPE b_casted = static_cast<CTYPE>(b_val);

using Vec = at::vec::Vectorized<CTYPE>;
Vec inv_b_casted_vec(CTYPE(1) / b_casted);
at::vec::map<CTYPE>(
[inv_b_casted_vec](Vec x) { return x * inv_b_casted_vec; },
out.mutable_data_ptr<CTYPE>(),
a.const_data_ptr<CTYPE>(),
out.numel());
});
// Check Common Dtype
ET_KERNEL_CHECK(ctx, common_type == out_type, InvalidArgument, out);

// Check Dim Order
ET_KERNEL_CHECK(
ctx, tensors_have_same_dim_order(a, out), InvalidArgument, out);

// Resize
ET_KERNEL_CHECK(
ctx, resize_tensor(out, a.sizes()) == Error::Ok, InvalidArgument, out);

// @lint-ignore CLANGTIDY facebook-hte-CArray
static constexpr const char op_name[] = "div.Scalar_out";

if (a_type == common_type && a_type == out_type &&
a_type != ScalarType::Half && a_type != ScalarType::BFloat16) {
ET_SWITCH_REAL_TYPES(a_type, ctx, op_name, CTYPE, [&]() {
ET_SWITCH_REALB_TYPES(b_type, ctx, op_name, CTYPE_B, [&]() {
CTYPE_B b_val;
ET_EXTRACT_SCALAR(b, b_val);
CTYPE b_casted = static_cast<CTYPE>(b_val);

using Vec = at::vec::Vectorized<CTYPE>;
Vec inv_b_casted_vec(CTYPE(1) / b_casted);
at::vec::map<CTYPE>(
[inv_b_casted_vec](Vec x) { return x * inv_b_casted_vec; },
out.mutable_data_ptr<CTYPE>(),
a.const_data_ptr<CTYPE>(),
out.numel());
});
});
} else {
ET_SWITCH_REAL_TYPES_AND(
Bool, a_type, ctx, "div.Scalar_out", CTYPE_A, [&]() {
ET_SWITCH_REAL_TYPES_AND(
Bool, b_type, ctx, "div.Scalar_out", CTYPE_B, [&]() {
ET_SWITCH_REAL_TYPES(
common_type, ctx, "div.Scalar_out", CTYPE_IN, [&]() {
ET_SWITCH_REAL_TYPES(
out_type, ctx, "div.Scalar_out", CTYPE_OUT, [&]() {
CTYPE_B b_val;
ET_EXTRACT_SCALAR(b, b_val);
CTYPE_IN b_casted = static_cast<CTYPE_IN>(b_val);
CTYPE_IN inv_b_casted = CTYPE_IN(1) / b_casted;

const size_t n = a.numel();
const CTYPE_A* a_data = a.const_data_ptr<CTYPE_A>();
CTYPE_OUT* out_data =
out.mutable_data_ptr<CTYPE_OUT>();
for (auto i = 0; i < n; ++i) {
out_data[i] = static_cast<CTYPE_OUT>(
static_cast<CTYPE_IN>(a_data[i]) *
inv_b_casted);
}
});
});
});
});
ScalarType compute_type = utils::get_compute_type(common_type);

ET_SWITCH_FLOAT_TYPES(compute_type, ctx, op_name, CTYPE_COMPUTE, [&]() {
const CTYPE_COMPUTE val_b = utils::scalar_to<CTYPE_COMPUTE>(b);
utils::apply_unitensor_elementwise_fn<
CTYPE_COMPUTE,
op_name,
utils::SupportedTensorDtypes::SAME_AS_COMMON>(
[val_b](const auto val_a) { return val_a / val_b; },
ctx,
a,
utils::SupportedTensorDtypes::REALHBBF16,
out);
});
}

return out;
Expand Down
22 changes: 19 additions & 3 deletions kernels/test/op_div_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,7 +54,7 @@ class OpDivOutTest : public OperatorTest {
#define ENUMERATE_TEST_ENTRY(ctype, dtype) \
test_div<DTYPE_A, DTYPE_B, ScalarType::dtype>();

ET_FORALL_FLOAT_TYPES(ENUMERATE_TEST_ENTRY)
ET_FORALL_FLOATHBF16_TYPES(ENUMERATE_TEST_ENTRY)

#undef ENUMERATE_TEST_ENTRY
}
Expand All @@ -64,7 +64,7 @@ class OpDivOutTest : public OperatorTest {
#define ENUMERATE_TEST_ENTRY(ctype, dtype) \
test_div_enumerate_out_types<DTYPE_A, ScalarType::dtype>();

ET_FORALL_REAL_TYPES(ENUMERATE_TEST_ENTRY)
ET_FORALL_REALHBF16_TYPES(ENUMERATE_TEST_ENTRY)

#undef ENUMERATE_TEST_ENTRY
}
Expand Down Expand Up @@ -183,7 +183,7 @@ void OpDivOutTest::test_div_enumerate_a_types() {
#define ENUMERATE_TEST_ENTRY(ctype, dtype) \
test_div_enumerate_b_types<ScalarType::dtype>();

ET_FORALL_REAL_TYPES(ENUMERATE_TEST_ENTRY)
ET_FORALL_REALHBF16_TYPES(ENUMERATE_TEST_ENTRY)

test_div<ScalarType::Bool, ScalarType::Float, ScalarType::Float>();

Expand Down Expand Up @@ -283,6 +283,22 @@ TEST_F(OpDivOutTest, BroadcastScalarSupported2) {
EXPECT_TENSOR_EQ(out, ret);
}

TEST_F(OpDivOutTest, BroadcastSupported3) {
TensorFactory<ScalarType::Float> tf;

Tensor a = tf.make({5}, {2, 3, 4, 5, 6});
Tensor b = tf.make({1, 5}, {2, 1, 2, 2, 3});

// Destination for the broadcasting div. Follow the broadcasting rules in
// https://fburl.com/n9wl4d0o
Tensor out = tf.zeros({1, 5});

op_div_out(a, b, out);

Tensor ret = tf.make({1, 5}, {1, 3, 2, 2.5, 2});
EXPECT_TENSOR_EQ(out, ret);
}

TEST_F(OpDivOutTest, BroadcastScalarRank0Supported) {
TensorFactory<ScalarType::Float> tf;

Expand Down
2 changes: 2 additions & 0 deletions shim_et/xplat/executorch/kernels/optimized/op_registration_util.bzl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -180,6 +180,8 @@ OPTIMIZED_ATEN_OPS = (
":binary_ops",
"//executorch/kernels/portable/cpu:scalar_utils",
"//executorch/kernels/portable/cpu/util:broadcast_util",
"//executorch/kernels/portable/cpu/util:dtype_util",
"//executorch/kernels/portable/cpu/util:elementwise_util",
"//executorch/runtime/core/portable_type/c10/c10:aten_headers_for_executorch",
],
),
Expand Down
Loading