Skip to content

[ET] enabling half dtype output for dequantization and making logic consistent #11669

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

Open
wants to merge 2 commits into
base: gh/ahmtox/12/orig
Choose a base branch
from
Open

[ET] enabling half dtype output for dequantization and making logic consistent #11669

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
6cb1fac
[ET] enabling half dtype output for dequantization and making logic c…
Jun 13, 2025
a6ffb06
Merge branch 'gh/ahmtox/12/orig' into gh/ahmtox/17/orig
ahmtox Jun 14, 2025
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
46 changes: 24 additions & 22 deletions kernels/quantized/cpu/op_dequantize.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -288,16 +288,16 @@ Tensor& dequantize_per_tensor_out(
static_cast<float>(scale)); \
} \
} break;
#define CALCULATE_INT_TYPE(IN_CTYPE, in_dtype) \
case ScalarType::in_dtype: \
switch (out.scalar_type()) { \
ET_FORALL_FLOAT_TYPES_WITH(IN_CTYPE, DEQUANTIZE_IMPL); \
default: \
ET_CHECK_MSG( \
false, \
"Unhandled output dtype %" PRId8, \
static_cast<int8_t>(out.scalar_type())); \
} \
#define CALCULATE_INT_TYPE(IN_CTYPE, in_dtype) \
case ScalarType::in_dtype: \
switch (out.scalar_type()) { \
ET_FORALL_FLOATH_TYPES_WITH(IN_CTYPE, DEQUANTIZE_IMPL); \
default: \
ET_CHECK_MSG( \
false, \
"Unhandled output dtype %" PRId8, \
static_cast<int8_t>(out.scalar_type())); \
} \
break;

switch (input.scalar_type()) {
Expand Down Expand Up @@ -459,7 +459,8 @@ Tensor& dequantize_per_channel_out(
} \
out_data_ptr[current_ix] = \
static_cast<CTYPE_OUT>( \
input_data_ptr[current_ix] - zero_point) * \
input_data_ptr[current_ix] - \
static_cast<int32_t>(zero_point)) * \
_scale; \
} \
}, \
Expand All @@ -478,23 +479,24 @@ Tensor& dequantize_per_channel_out(
apply_over_dim_list( \
[input_data_ptr, out_data_ptr, _scale, _zero_point](size_t in_ix) { \
out_data_ptr[in_ix] = static_cast<CTYPE_OUT>( \
(input_data_ptr[in_ix] - _zero_point) * _scale); \
(input_data_ptr[in_ix] - static_cast<int32_t>(_zero_point)) * \
_scale); \
}, \
input, \
optional_dim_list, \
channel_ix); \
} \
break;
#define CALCULATE_FLOAT_TYPE(CTYPE_IN, in_dtype) \
case ScalarType::in_dtype: \
switch (out.scalar_type()) { \
ET_FORALL_FLOAT_TYPES_WITH(CTYPE_IN, DEQUANTIZE_IMPL); \
default: \
ET_CHECK_MSG( \
false, \
"Unhandled output dtype %" PRId8, \
static_cast<int8_t>(out.scalar_type())); \
} \
#define CALCULATE_FLOAT_TYPE(CTYPE_IN, in_dtype) \
case ScalarType::in_dtype: \
switch (out.scalar_type()) { \
ET_FORALL_FLOATH_TYPES_WITH(CTYPE_IN, DEQUANTIZE_IMPL); \
default: \
ET_CHECK_MSG( \
false, \
"Unhandled output dtype %" PRId8, \
static_cast<int8_t>(out.scalar_type())); \
} \
break;

switch (input.scalar_type()) {
Expand Down
90 changes: 90 additions & 0 deletions kernels/quantized/test/op_dequantize_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,6 +67,96 @@ TEST(OpDequantizeOutTest, AllDtypesSupported) {
test_dtype<ScalarType::Int>();
}

/// Test all supported output dtypes for dequantization
template <ScalarType OUT_DTYPE>
void test_output_dtype() {
TensorFactory<ScalarType::Byte> tf;

Tensor input = tf.full({3, 5}, 100);
double scale = 0.5;
int64_t zero_point = 30;
int64_t quant_min = 0;
int64_t quant_max = 255;

TensorFactory<OUT_DTYPE> tfo;
Tensor out = tfo.zeros({3, 5});
// (100 - 30) * 0.5 = 35
Tensor expected = tfo.full({3, 5}, 35);
dequantize_per_tensor_out(
input,
scale,
zero_point,
quant_min,
quant_max,
ScalarType::Byte,
optional<ScalarType>(OUT_DTYPE),
out);

EXPECT_TENSOR_EQ(out, expected);
}

TEST(OpDequantizeOutTest, AllOutputDtypesSupported) {
et_pal_init();
test_output_dtype<ScalarType::Float>();
test_output_dtype<ScalarType::Double>();
test_output_dtype<ScalarType::Half>();
}

TEST(OpDequantizeOutTest, HalfOutput) {
et_pal_init();
TensorFactory<ScalarType::Byte> tf;

Tensor input = tf.full({3, 5}, 10);
double scale = 0.5;
int64_t zero_point = 100000;
int64_t quant_min = 0;
int64_t quant_max = 255;

TensorFactory<ScalarType::Half> tfo;
Tensor out = tfo.zeros({3, 5});
// (10 - 100000) * 0.5 = -49995
dequantize_per_tensor_out(
input,
scale,
zero_point,
quant_min,
quant_max,
ScalarType::Byte,
optional<ScalarType>(ScalarType::Half),
out);

// The expected result should be (10 - 100000) * 0.5 = -49995
Tensor expected = tfo.full({3, 5}, -49995);
EXPECT_TENSOR_EQ(out, expected);
}

TEST(OpDequantizeOutTest, DoubleOutput) {
et_pal_init();
TensorFactory<ScalarType::Byte> tf;

Tensor input = tf.full({3, 5}, 10);
double scale = 0.5;
int64_t zero_point = 100000;
int64_t quant_min = 0;
int64_t quant_max = 255;

TensorFactory<ScalarType::Double> tfo;
Tensor out = tfo.zeros({3, 5});
dequantize_per_tensor_out(
input,
scale,
zero_point,
quant_min,
quant_max,
ScalarType::Byte,
optional<ScalarType>(ScalarType::Double),
out);

// The expected result should be (10 - 100000) * 0.5 = -49995
Tensor expected = tfo.full({3, 5}, -49995);
EXPECT_TENSOR_EQ(out, expected);
}

TEST(OpDequantizeOutTest, NonWholeNumbers) {
et_pal_init();
TensorFactory<ScalarType::Byte> tf;
Expand Down
5 changes: 5 additions & 0 deletions runtime/core/exec_aten/util/scalar_type_util.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -199,6 +199,11 @@ ET_FORALL_SCALAR_TYPES(SPECIALIZE_CppTypeToScalarType)
_(ANOTHER_INPUT, float, Float) \
_(ANOTHER_INPUT, double, Double)

#define ET_FORALL_FLOATH_TYPES_WITH(ANOTHER_INPUT, _) \
_(ANOTHER_INPUT, float, Float) \
_(ANOTHER_INPUT, double, Double) \
_(ANOTHER_INPUT, ::executorch::aten::Half, Half)

#define ET_FORALL_FLOAT_TYPES_WITH2(ANOTHER_INPUT1, ANOTHER_INPUT2, _) \
_(ANOTHER_INPUT1, ANOTHER_INPUT2, float, Float) \
_(ANOTHER_INPUT1, ANOTHER_INPUT2, double, Double)
Expand Down
Loading