Skip to content

Add int8/uint8 CPU support for SpaceToDepth and int8 for DepthToSpace #29154

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
tianleiwu merged 3 commits into from
Jun 21, 2026
Merged

Add int8/uint8 CPU support for SpaceToDepth and int8 for DepthToSpace #29154

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1d39f26
Add int8/uint8 CPU support for SpaceToDepth and int8 for DepthToSpace
ArsalanShakil Jun 18, 2026
feaa3d6
Address review feedback for SpaceToDepth/DepthToSpace int8 tests
ArsalanShakil Jun 20, 2026
e4be642
Exclude TensorRT EP from int8 SpaceToDepth/DepthToSpace tests
ArsalanShakil Jun 21, 2026
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
8 changes: 4 additions & 4 deletions docs/OperatorKernels.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -115,8 +115,8 @@ The **OpSet Version** column uses the following notation:
|||[17, 19]|**T1** = tensor(double), tensor(float)<br/> **T2** = tensor(int32), tensor(int64)|
|DeformConv|*in* X:**T**<br> *in* W:**T**<br> *in* offset:**T**<br> *in* B:**T**<br> *in* mask:**T**<br> *out* Y:**T**|22+|**T** = tensor(double), tensor(float)|
|||[19, 21]|**T** = tensor(double), tensor(float)|
|DepthToSpace|*in* input:**T**<br> *out* output:**T**|13+|**T** = tensor(double), tensor(float), tensor(uint8)|
|||[11, 12]|**T** = tensor(double), tensor(float), tensor(uint8)|
|DepthToSpace|*in* input:**T**<br> *out* output:**T**|13+|**T** = tensor(double), tensor(float), tensor(int8), tensor(uint8)|
|||[11, 12]|**T** = tensor(double), tensor(float), tensor(int8), tensor(uint8)|
|||[1, 10]|**T** = tensor(double), tensor(float)|
|DequantizeLinear|*in* x:**T**<br> *in* x_scale:**tensor(float)**<br> *in* x_zero_point:**T**<br> *out* y:**tensor(float)**<br><br>or<br><br>*in* x:**T1**<br> *in* x_scale:**T2**<br> *in* x_zero_point:**T1**<br> *out* y:**T2**<br><br>or<br><br>*in* x:**T1**<br> *in* x_scale:**T2**<br> *in* x_zero_point:**T1**<br> *out* y:**T3**|25+|**T1** = tensor(float8e4m3fn), tensor(float8e4m3fnuz), tensor(float8e5m2), tensor(float8e5m2fnuz), tensor(int16), tensor(int2), tensor(int32), tensor(int4), tensor(int8), tensor(uint16), tensor(uint2), tensor(uint4), tensor(uint8)<br/> **T2** = tensor(float), tensor(float16)|
|||24|**T1** = tensor(float8e4m3fn), tensor(float8e4m3fnuz), tensor(float8e5m2), tensor(float8e5m2fnuz), tensor(int16), tensor(int32), tensor(int4), tensor(int8), tensor(uint16), tensor(uint4), tensor(uint8)<br/> **T2** = tensor(float), tensor(float16)|
Expand Down Expand Up @@ -477,8 +477,8 @@ The **OpSet Version** column uses the following notation:
|||[1, 21]|**T** = tensor(float)|
|Softsign|*in* input:**T**<br> *out* output:**T**|22+|**T** = tensor(float)|
|||[1, 21]|**T** = tensor(float)|
|SpaceToDepth|*in* input:**T**<br> *out* output:**T**|13+|**T** = tensor(double), tensor(float)|
|||[1, 12]|**T** = tensor(double), tensor(float)|
|SpaceToDepth|*in* input:**T**<br> *out* output:**T**|13+|**T** = tensor(double), tensor(float), tensor(int8), tensor(uint8)|
|||[1, 12]|**T** = tensor(double), tensor(float), tensor(int8), tensor(uint8)|
|Split|*in* input:**T**<br> *in* split:**T**<br> *out* outputs...:**T**<br><br>or<br><br>*in* input:**T**<br> *in* split:**tensor(int64)**<br> *out* outputs:**T**<br><br>or<br><br>*in* input:**T**<br> *out* outputs:**T**|18+|**T** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(string), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)|
|||[13, 17]|**T** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(string), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)|
|||[11, 12]|**T** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(string), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)|
Expand Down
55 changes: 40 additions & 15 deletions onnxruntime/core/providers/cpu/tensor/space_depth_ops.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,15 +18,19 @@ ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
12,
KernelDefBuilder()
.TypeConstraint("T", {DataTypeImpl::GetTensorType<float>(),
DataTypeImpl::GetTensorType<double>()}),
DataTypeImpl::GetTensorType<double>(),
DataTypeImpl::GetTensorType<uint8_t>(),
DataTypeImpl::GetTensorType<int8_t>()}),
SpaceToDepth);

ONNX_CPU_OPERATOR_KERNEL(
SpaceToDepth,
13,
KernelDefBuilder()
.TypeConstraint("T", {DataTypeImpl::GetTensorType<float>(),
DataTypeImpl::GetTensorType<double>()}),
DataTypeImpl::GetTensorType<double>(),
DataTypeImpl::GetTensorType<uint8_t>(),
DataTypeImpl::GetTensorType<int8_t>()}),
SpaceToDepth);

ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
Expand All @@ -44,7 +48,8 @@ ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
KernelDefBuilder()
.TypeConstraint("T", {DataTypeImpl::GetTensorType<float>(),
DataTypeImpl::GetTensorType<double>(),
DataTypeImpl::GetTensorType<uint8_t>()}),
DataTypeImpl::GetTensorType<uint8_t>(),
DataTypeImpl::GetTensorType<int8_t>()}),
DepthToSpace);

ONNX_CPU_OPERATOR_KERNEL(
Expand All @@ -53,7 +58,8 @@ ONNX_CPU_OPERATOR_KERNEL(
KernelDefBuilder()
.TypeConstraint("T", {DataTypeImpl::GetTensorType<float>(),
DataTypeImpl::GetTensorType<double>(),
DataTypeImpl::GetTensorType<uint8_t>()}),
DataTypeImpl::GetTensorType<uint8_t>(),
DataTypeImpl::GetTensorType<int8_t>()}),
DepthToSpace);

// intermediate tensor shapes are:
Expand All @@ -70,15 +76,15 @@ using ConstEigenTensorMap = Eigen::TensorMap<Eigen::Tensor<const T, Intermediate
// helper method to fill in output buffer
// only this portion is templated to minimize binary size
template <typename T>
static void SpaceDepthOpCpuImpl(const Tensor& input, Tensor& output,
static void SpaceDepthOpCpuImpl(const T* input_data, T* output_data,
const std::array<Eigen::DenseIndex, IntermediateTensorRank>& permutation,
const Eigen::DenseIndex batch_size, // dim0 in both input and output
const Eigen::DenseIndex in_dim1, const Eigen::DenseIndex in_dim2, const Eigen::DenseIndex in_dim3,
const Eigen::DenseIndex in_dim4, const Eigen::DenseIndex in_dim5,
const Eigen::DenseIndex out_dim1, const Eigen::DenseIndex out_dim2, const Eigen::DenseIndex out_dim3,
const Eigen::DenseIndex out_dim4, const Eigen::DenseIndex out_dim5) {
EigenTensorMap<T>(output.MutableData<T>(), batch_size, out_dim1, out_dim2, out_dim3, out_dim4, out_dim5) =
ConstEigenTensorMap<T>(input.Data<T>(), batch_size,
EigenTensorMap<T>(output_data, batch_size, out_dim1, out_dim2, out_dim3, out_dim4, out_dim5) =
ConstEigenTensorMap<T>(input_data, batch_size,
in_dim1, in_dim2, in_dim3, in_dim4, in_dim5)
.shuffle(permutation);
}
Expand Down Expand Up @@ -109,7 +115,7 @@ Status SpaceToDepth::Compute(OpKernelContext* context) const {
std::array<Eigen::DenseIndex, IntermediateTensorRank> permutation{{0, 3, 5, 1, 2, 4}};

if (input.IsDataType<float>()) {
SpaceDepthOpCpuImpl<float>(input, output, permutation,
SpaceDepthOpCpuImpl<float>(input.Data<float>(), output.MutableData<float>(), permutation,
onnxruntime::narrow<ptrdiff_t>(batch),
onnxruntime::narrow<std::ptrdiff_t>(input_depth),
onnxruntime::narrow<std::ptrdiff_t>(input_height / blocksize_),
Expand All @@ -122,7 +128,7 @@ Status SpaceToDepth::Compute(OpKernelContext* context) const {
onnxruntime::narrow<std::ptrdiff_t>(input_height / blocksize_),
onnxruntime::narrow<std::ptrdiff_t>(input_width / blocksize_));
} else if (input.IsDataType<double>()) {
SpaceDepthOpCpuImpl<double>(input, output, permutation,
SpaceDepthOpCpuImpl<double>(input.Data<double>(), output.MutableData<double>(), permutation,
onnxruntime::narrow<ptrdiff_t>(batch),
onnxruntime::narrow<std::ptrdiff_t>(input_depth),
onnxruntime::narrow<std::ptrdiff_t>(input_height / blocksize_),
Expand All @@ -134,8 +140,24 @@ Status SpaceToDepth::Compute(OpKernelContext* context) const {
onnxruntime::narrow<std::ptrdiff_t>(input_depth),
onnxruntime::narrow<std::ptrdiff_t>(input_height / blocksize_),
onnxruntime::narrow<std::ptrdiff_t>(input_width / blocksize_));
} else if (input.IsDataType<uint8_t>() || input.IsDataType<int8_t>()) {
// uint8_t and int8_t share a single implementation: the op only moves 8-bit
// elements around, so the signedness of the data does not matter.
SpaceDepthOpCpuImpl<uint8_t>(static_cast<const uint8_t*>(input.DataRaw()),
static_cast<uint8_t*>(output.MutableDataRaw()), permutation,
onnxruntime::narrow<ptrdiff_t>(batch),
onnxruntime::narrow<std::ptrdiff_t>(input_depth),
onnxruntime::narrow<std::ptrdiff_t>(input_height / blocksize_),
onnxruntime::narrow<std::ptrdiff_t>(blocksize_),
onnxruntime::narrow<std::ptrdiff_t>(input_width / blocksize_),
onnxruntime::narrow<std::ptrdiff_t>(blocksize_),
onnxruntime::narrow<ptrdiff_t>(blocksize_),
onnxruntime::narrow<ptrdiff_t>(blocksize_),
onnxruntime::narrow<std::ptrdiff_t>(input_depth),
onnxruntime::narrow<std::ptrdiff_t>(input_height / blocksize_),
onnxruntime::narrow<std::ptrdiff_t>(input_width / blocksize_));
} else {
// user will not see this as the kernel doesn't claim support for types other than float and double
// user will not see this as the kernel doesn't claim support for types other than float, double, uint8_t and int8_t
return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "Unsupported input type in SpaceToDepth op: ", input.DataType());
}

Expand Down Expand Up @@ -173,7 +195,7 @@ Status DepthToSpace::Compute(OpKernelContext* context) const {
: std::array<Eigen::DenseIndex, IntermediateTensorRank>{{0, 1, 4, 2, 5, 3}};

if (input.IsDataType<float>()) {
SpaceDepthOpCpuImpl<float>(input, output, permutation,
SpaceDepthOpCpuImpl<float>(input.Data<float>(), output.MutableData<float>(), permutation,
onnxruntime::narrow<std::ptrdiff_t>(batch),
onnxruntime::narrow<std::ptrdiff_t>(dim1),
onnxruntime::narrow<std::ptrdiff_t>(blocksize_),
Expand All @@ -186,7 +208,7 @@ Status DepthToSpace::Compute(OpKernelContext* context) const {
onnxruntime::narrow<std::ptrdiff_t>(input_width),
onnxruntime::narrow<std::ptrdiff_t>(blocksize_));
} else if (input.IsDataType<double>()) {
SpaceDepthOpCpuImpl<double>(input, output, permutation,
SpaceDepthOpCpuImpl<double>(input.Data<double>(), output.MutableData<double>(), permutation,
onnxruntime::narrow<std::ptrdiff_t>(batch),
onnxruntime::narrow<std::ptrdiff_t>(dim1),
onnxruntime::narrow<std::ptrdiff_t>(blocksize_),
Expand All @@ -198,8 +220,11 @@ Status DepthToSpace::Compute(OpKernelContext* context) const {
onnxruntime::narrow<std::ptrdiff_t>(blocksize_),
onnxruntime::narrow<std::ptrdiff_t>(input_width),
onnxruntime::narrow<std::ptrdiff_t>(blocksize_));
} else if (input.IsDataType<uint8_t>()) {
SpaceDepthOpCpuImpl<uint8_t>(input, output, permutation,
} else if (input.IsDataType<uint8_t>() || input.IsDataType<int8_t>()) {
// uint8_t and int8_t share a single implementation: the op only moves 8-bit
// elements around, so the signedness of the data does not matter.
SpaceDepthOpCpuImpl<uint8_t>(static_cast<const uint8_t*>(input.DataRaw()),
static_cast<uint8_t*>(output.MutableDataRaw()), permutation,
onnxruntime::narrow<std::ptrdiff_t>(batch),
onnxruntime::narrow<std::ptrdiff_t>(dim1),
onnxruntime::narrow<std::ptrdiff_t>(blocksize_),
Expand All @@ -212,7 +237,7 @@ Status DepthToSpace::Compute(OpKernelContext* context) const {
onnxruntime::narrow<std::ptrdiff_t>(input_width),
onnxruntime::narrow<std::ptrdiff_t>(blocksize_));
} else {
// user will not see this as the kernel doesn't claim support for types other than float and double
// user will not see this as the kernel doesn't claim support for types other than float, double, uint8_t and int8_t
return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "Unsupported input type in DepthToSpace op: ", input.DataType());
}

Expand Down
88 changes: 87 additions & 1 deletion onnxruntime/test/providers/cpu/tensor/space_depth_ops_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@ template <typename T>
class TensorOpTest : public ::testing::Test {
};

using TensorOpTestTypes = ::testing::Types<float, MLFloat16, uint8_t>;
using TensorOpTestTypes = ::testing::Types<float, MLFloat16, uint8_t, int8_t>;
TYPED_TEST_SUITE(TensorOpTest, TensorOpTestTypes);

TEST(TensorOpTest, SpaceToDepthTest_1) {
Expand Down Expand Up @@ -159,6 +159,71 @@ TEST(TensorOpTest, SpaceToDepthTest_3) {
test.Run();
}

TYPED_TEST(TensorOpTest, SpaceToDepthTest_int) {
// Same data as SpaceToDepthTest_2 (values 0..107 fit in both int8_t and uint8_t),
// exercised across the supported element types.
OpTester test("SpaceToDepth");
Comment thread
ArsalanShakil marked this conversation as resolved.
Outdated
constexpr int64_t blocksize = 3;
test.AddAttribute("blocksize", blocksize);
constexpr int64_t N = 2, C = 3, H = 3, W = 6;
Comment thread
ArsalanShakil marked this conversation as resolved.
const std::vector<float> X = {
0., 1., 2., 3., 4., 5., 6., 7., 8., 9., 10.,
11., 12., 13., 14., 15., 16., 17., 18., 19., 20., 21.,
22., 23., 24., 25., 26., 27., 28., 29., 30., 31., 32.,
33., 34., 35., 36., 37., 38., 39., 40., 41., 42., 43.,
44., 45., 46., 47., 48., 49., 50., 51., 52., 53., 54.,
55., 56., 57., 58., 59., 60., 61., 62., 63., 64., 65.,
66., 67., 68., 69., 70., 71., 72., 73., 74., 75., 76.,
77., 78., 79., 80., 81., 82., 83., 84., 85., 86., 87.,
88., 89., 90., 91., 92., 93., 94., 95., 96., 97., 98.,
99., 100., 101., 102., 103., 104., 105., 106., 107.};

const std::vector<float> result = {
0., 3., 18., 21., 36., 39., 1., 4., 19., 22., 37.,
40., 2., 5., 20., 23., 38., 41., 6., 9., 24., 27.,
42., 45., 7., 10., 25., 28., 43., 46., 8., 11., 26.,
29., 44., 47., 12., 15., 30., 33., 48., 51., 13., 16.,
31., 34., 49., 52., 14., 17., 32., 35., 50., 53., 54.,
57., 72., 75., 90., 93., 55., 58., 73., 76., 91., 94.,
56., 59., 74., 77., 92., 95., 60., 63., 78., 81., 96.,
99., 61., 64., 79., 82., 97., 100., 62., 65., 80., 83.,
98., 101., 66., 69., 84., 87., 102., 105., 67., 70., 85.,
88., 103., 106., 68., 71., 86., 89., 104., 107.};

const std::vector<int64_t> output_shape = {2, 27, 1, 2};

if constexpr (std::is_same<TypeParam, float>::value) {
test.AddInput<float>("input", {N, C, H, W}, X);
test.AddOutput<float>("output", output_shape, result);
} else if constexpr (std::is_same<TypeParam, MLFloat16>::value) {
std::vector<TypeParam> X_fp16(X.size());
std::vector<TypeParam> result_fp16(result.size());
ConvertFloatToMLFloat16(X.data(), X_fp16.data(), X.size());
ConvertFloatToMLFloat16(result.data(), result_fp16.data(), result.size());
test.AddInput<TypeParam>("input", {N, C, H, W}, X_fp16);
test.AddOutput<TypeParam>("output", output_shape, result_fp16);
} else if constexpr (std::is_same<TypeParam, uint8_t>::value) {
std::vector<uint8_t> X_u8(X.size());
std::vector<uint8_t> result_u8(result.size());
ConvertFloatToUint8_t(X.data(), X_u8.data(), X.size());
ConvertFloatToUint8_t(result.data(), result_u8.data(), result.size());
test.AddInput<uint8_t>("input", {N, C, H, W}, X_u8);
test.AddOutput<uint8_t>("output", output_shape, result_u8);
} else if constexpr (std::is_same<TypeParam, int8_t>::value) {
std::vector<int8_t> X_i8(X.size());
std::vector<int8_t> result_i8(result.size());
ConvertFloatToInt8_t(X.data(), X_i8.data(), X.size());
ConvertFloatToInt8_t(result.data(), result_i8.data(), result.size());
test.AddInput<int8_t>("input", {N, C, H, W}, X_i8);
test.AddOutput<int8_t>("output", output_shape, result_i8);
} else {
ORT_THROW("Type not supported");
}

// type not supported by QNN EP: MLFloat16 and unsigned char
test.Run(OpTester::ExpectResult::kExpectSuccess, "", {kQnnExecutionProvider});
Comment thread
ArsalanShakil marked this conversation as resolved.
Outdated
}

TEST(TensorOpTest, DepthToSpaceTest_1) {
OpTester test("DepthToSpace", 7); // create an opset 7 model
constexpr int64_t blocksize = 2;
Expand Down Expand Up @@ -319,6 +384,13 @@ TYPED_TEST(TensorOpTest, DepthToSpaceTest_3) {
ConvertFloatToUint8_t(result.data(), result_u8.data(), result.size());
test.AddInput<uint8_t>("input", {N, C, H, W}, X_u8);
test.AddOutput<uint8_t>("output", {2, 3, 6, 4}, result_u8);
} else if constexpr (std::is_same<TypeParam, int8_t>::value) {
std::vector<int8_t> X_i8(X.size());
std::vector<int8_t> result_i8(result.size());
ConvertFloatToInt8_t(X.data(), X_i8.data(), X.size());
Comment thread
tianleiwu marked this conversation as resolved.
ConvertFloatToInt8_t(result.data(), result_i8.data(), result.size());
test.AddInput<int8_t>("input", {N, C, H, W}, X_i8);
test.AddOutput<int8_t>("output", {2, 3, 6, 4}, result_i8);
Comment thread
ArsalanShakil marked this conversation as resolved.
} else {
ORT_THROW("Type not supported");
}
Expand Down Expand Up @@ -383,6 +455,13 @@ TYPED_TEST(TensorOpTest, DepthToSpaceTest_4) {
ConvertFloatToUint8_t(result.data(), result_u8.data(), result.size());
test.AddInput<uint8_t>("input", {N, C, H, W}, X_u8);
test.AddOutput<uint8_t>("output", {2, 3, 6, 4}, result_u8);
} else if constexpr (std::is_same<TypeParam, int8_t>::value) {
std::vector<int8_t> X_i8(X.size());
std::vector<int8_t> result_i8(result.size());
ConvertFloatToInt8_t(X.data(), X_i8.data(), X.size());
ConvertFloatToInt8_t(result.data(), result_i8.data(), result.size());
test.AddInput<int8_t>("input", {N, C, H, W}, X_i8);
test.AddOutput<int8_t>("output", {2, 3, 6, 4}, result_i8);
Comment thread
ArsalanShakil marked this conversation as resolved.
} else {
ORT_THROW("Type not supported");
}
Expand Down Expand Up @@ -429,6 +508,13 @@ TYPED_TEST(TensorOpTest, DepthToSpaceTest_5) {
ConvertFloatToUint8_t(result.data(), result_u8.data(), result.size());
test.AddInput<uint8_t>("input", {N, C, H, W}, X_u8);
test.AddOutput<uint8_t>("output", {1, 1, 4, 6}, result_u8);
} else if constexpr (std::is_same<TypeParam, int8_t>::value) {
std::vector<int8_t> X_i8(X.size());
std::vector<int8_t> result_i8(result.size());
ConvertFloatToInt8_t(X.data(), X_i8.data(), X.size());
ConvertFloatToInt8_t(result.data(), result_i8.data(), result.size());
test.AddInput<int8_t>("input", {N, C, H, W}, X_i8);
test.AddOutput<int8_t>("output", {1, 1, 4, 6}, result_i8);
} else {
ORT_THROW("Type not supported");
}
Expand Down
6 changes: 6 additions & 0 deletions onnxruntime/test/unittest_util/conversion.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,12 @@ inline void ConvertFloatToUint8_t(const float* f_datat, uint8_t* u8_data, size_t
output_vector = in_vector.template cast<uint8_t>();
}

inline void ConvertFloatToInt8_t(const float* f_datat, int8_t* i8_data, size_t input_size) {
auto in_vector = ConstEigenVectorMap<float>(f_datat, input_size);
auto output_vector = EigenVectorMap<int8_t>(static_cast<int8_t*>(static_cast<void*>(i8_data)), input_size);
output_vector = in_vector.template cast<int8_t>();
}

inline void ConvertMLFloat16ToFloat(const MLFloat16* h_data, float* f_data, size_t input_size) {
auto in_vector =
ConstEigenVectorMap<Eigen::half>(static_cast<const Eigen::half*>(static_cast<const void*>(h_data)), input_size);
Expand Down
Loading