Skip to content

Validate ScatterND index depth #32034

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
Akshay Sonawane (apsonawane) merged 4 commits into from
Aug 20, 2026
Merged

Validate ScatterND index depth #32034

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
bc1a512
Validate ScatterND index depth
apsonawane Aug 12, 2026
3a9cf23
Support zero-depth ScatterND indices
apsonawane Aug 13, 2026
6c253a0
Exclude WebGPU from zero-depth ScatterND tests
apsonawane Aug 13, 2026
c86bbe5
Exclude TensorRT from zero-depth ScatterND tests
apsonawane Aug 17, 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
15 changes: 12 additions & 3 deletions onnxruntime/core/providers/cpu/tensor/scatter_nd.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,11 +61,13 @@ struct Prepare {
TData* output_base;
uint64_t element_to_copy;
std::vector<uint64_t> element_offsets;
bool serialize_updates;

Prepare() : input_base(nullptr),
output_base(nullptr),
element_to_copy(0),
element_offsets(0) {}
element_offsets(0),
serialize_updates(false) {}
}; // struct Prepare

template <typename TData>
Expand All @@ -89,7 +91,7 @@ Status PrepareForCompute(OpKernelContext* context, Prepare<TData>& p) {
auto last_indice_dimension = indice_shape[indice_shape.NumDimensions() - 1];

// Re-use input for output. If input/output Tensor* are the same, do not copy.
if (src_base != dst_base) {
if (src_base != dst_base && input_tensor->Shape().Size() > 0) {
if (is_string_type) {
const auto* str_begin = input_tensor->Data<std::string>();
const std::string* str_end = str_begin + input_shape.Size();
Expand All @@ -108,8 +110,9 @@ Status PrepareForCompute(OpKernelContext* context, Prepare<TData>& p) {
}

p.element_to_copy = input_shape.SizeFromDimension(onnxruntime::narrow<size_t>(last_indice_dimension));
p.serialize_updates = last_indice_dimension == 0;
const int64_t* indice_offset = indice_tensor->Data<int64_t>();
auto offset_count = indice_shape.Size() / last_indice_dimension; // Times to copy
auto offset_count = indice_shape.SizeToDimension(indice_shape.NumDimensions() - 1); // Times to copy
p.element_offsets.assign(onnxruntime::narrow<size_t>(offset_count), 0LL);

p.input_base = update_tensor->Data<TData>();
Expand Down Expand Up @@ -302,6 +305,12 @@ struct ScatterNDDispatchTarget {
Status operator()(OpKernelContext* context, concurrency::ThreadPool* tp, ScatterND::Reduction reduction) const {
Prepare<TData> prepare;
ORT_RETURN_IF_ERROR(PrepareForCompute(context, prepare));
if (prepare.element_to_copy == 0 || prepare.element_offsets.empty()) {
return Status::OK();
}
if (prepare.serialize_updates) {
tp = nullptr;
}

auto lambda = [&](ptrdiff_t i) {
switch (reduction) {
Expand Down
30 changes: 18 additions & 12 deletions onnxruntime/core/providers/cuda/tensor/scatter_nd.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -90,18 +90,21 @@ Status ScatterNDDisjointAndNoReduction::ComputeInternal(OpKernelContext* context
const void* input_data = input_tensor->DataRaw();
void* output_data = output_tensor->MutableDataRaw();

if (input_data != output_data) {
if (input_data != output_data && input_tensor->SizeInBytes() > 0) {
// TODO: Run benchmarks to determine if a dedicated kernel doing data copy will be faster than invoking cudaMemcpy ?
CUDA_RETURN_IF_ERROR(
cudaMemcpyAsync(output_data, input_data, input_tensor->SizeInBytes(), cudaMemcpyDeviceToDevice, Stream(context)));
}

const auto num_indices = indices_shape.SizeToDimension(indices_shape.NumDimensions() - 1);
auto last_index_dimension = indices_shape[indices_shape.NumDimensions() - 1];
const auto num_update_elements = input_shape.SizeFromDimension(last_index_dimension);

// Bail out early
if (indices_shape.Size() == 0) {
if (num_indices == 0 || num_update_elements == 0) {
return Status::OK();
}

auto last_index_dimension = indices_shape[indices_shape.NumDimensions() - 1];
size_t element_size = input_tensor->DataType()->Size();

// We need element counts for each dimension and the input dim value for each dimension
Expand All @@ -118,12 +121,12 @@ Status ScatterNDDisjointAndNoReduction::ComputeInternal(OpKernelContext* context
Stream(context),
output_data,
element_size,
indices_shape.Size() / static_cast<size_t>(last_index_dimension),
onnxruntime::narrow<size_t>(num_indices),
indices_tensor->Data<int64_t>(), // only int64_t is supported for indices as per the onnx spec
last_index_dimension,
element_counts_and_input_dims,
updates_tensor->DataRaw(),
input_shape.SizeFromDimension(last_index_dimension)));
onnxruntime::narrow<size_t>(num_update_elements)));

return Status::OK();
}
Expand All @@ -145,20 +148,23 @@ Status ScatterNDWithAtomicReduction::ComputeInternal(OpKernelContext* context) c
const void* input_data = input_tensor->DataRaw();
void* output_data = output_tensor->MutableDataRaw();

if (input_data != output_data) {
if (input_data != output_data && input_tensor->SizeInBytes() > 0) {
// TODO: Run benchmarks to determine if a dedicated kernel doing data copy will
// be faster than invoking cudaMemcpy ?
CUDA_RETURN_IF_ERROR(
cudaMemcpyAsync(output_data, input_data, input_tensor->SizeInBytes(),
cudaMemcpyDeviceToDevice, Stream(context)));
}

const auto num_indices = indices_shape.SizeToDimension(indices_shape.NumDimensions() - 1);
auto last_index_dimension = indices_shape[indices_shape.NumDimensions() - 1];
const auto num_update_elements = input_shape.SizeFromDimension(last_index_dimension);

// Bail out early
if (indices_shape.Size() == 0) {
if (num_indices == 0 || num_update_elements == 0) {
return Status::OK();
}

auto last_index_dimension = indices_shape[indices_shape.NumDimensions() - 1];
ElementCountsAndInputDimsSpanOrGpu element_counts_and_input_dims;
CudaAsyncBuffer<int64_t> element_counts_and_input_dims_gpu(this);
ORT_RETURN_IF_ERROR(InitializeElementCountsAndInputDimsSpanOrGpu(last_index_dimension, input_shape,
Expand All @@ -173,12 +179,12 @@ Status ScatterNDWithAtomicReduction::ComputeInternal(OpKernelContext* context) c
Stream(context),
output_data,
element_size,
indices_shape.Size() / static_cast<size_t>(last_index_dimension),
onnxruntime::narrow<size_t>(num_indices),
indices_tensor->Data<int64_t>(), // only int64_t is supported for indices as per the onnx spec
last_index_dimension,
element_counts_and_input_dims,
updates_tensor->DataRaw(),
input_shape.SizeFromDimension(last_index_dimension)));
onnxruntime::narrow<size_t>(num_update_elements)));
} break;
case ScatterNDReduction::Add:
case ScatterNDReduction::Min:
Expand All @@ -189,12 +195,12 @@ Status ScatterNDWithAtomicReduction::ComputeInternal(OpKernelContext* context) c
Stream(context),
output_data,
element_type,
indices_shape.Size() / static_cast<size_t>(last_index_dimension),
onnxruntime::narrow<size_t>(num_indices),
indices_tensor->Data<int64_t>(), // only int64_t is supported for indices as per the onnx spec
last_index_dimension,
element_counts_and_input_dims,
updates_tensor->DataRaw(),
input_shape.SizeFromDimension(last_index_dimension),
onnxruntime::narrow<size_t>(num_update_elements),
reduction_));
} break;
default:
Expand Down
31 changes: 31 additions & 0 deletions onnxruntime/test/providers/cpu/tensor/scatter_nd_op_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -246,5 +246,36 @@ TEST(ScatterNDOpTest, ScatterND_empty_indices) {
test1.Run(OpTester::ExpectResult::kExpectSuccess, "", {kDmlExecutionProvider});
}

TEST(ScatterNDOpTest, ScatterND_zero_index_depth_updates_entire_tensor) {
OpTester test("ScatterND", 18);
test.AddInput<float>("data", {2, 3}, {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f});
test.AddInput<int64_t>("indices", {1, 0}, {});
test.AddInput<float>("updates", {1, 2, 3}, {10.0f, 20.0f, 30.0f, 40.0f, 50.0f, 60.0f});
test.AddOutput<float>("output", {2, 3}, {10.0f, 20.0f, 30.0f, 40.0f, 50.0f, 60.0f});
test.Run(OpTester::ExpectResult::kExpectSuccess, "",
{kTensorrtExecutionProvider, kWebGpuExecutionProvider});
}

TEST(ScatterNDOpTest, ScatterND_zero_index_depth_adds_multiple_updates) {
OpTester test("ScatterND", 18);
test.AddAttribute("reduction", "add");
test.AddInput<float>("data", {2}, {1.0f, 2.0f});
test.AddInput<int64_t>("indices", {2, 0}, {});
test.AddInput<float>("updates", {2, 2}, {10.0f, 20.0f, 100.0f, 200.0f});
test.AddOutput<float>("output", {2}, {111.0f, 222.0f});
test.Run(OpTester::ExpectResult::kExpectSuccess, "",
{kTensorrtExecutionProvider, kWebGpuExecutionProvider});
}

TEST(ScatterNDOpTest, ScatterND_zero_index_depth_empty_data) {
OpTester test("ScatterND", 18);
test.AddInput<float>("data", {0, 3}, {});
test.AddInput<int64_t>("indices", {1, 0}, {});
test.AddInput<float>("updates", {1, 0, 3}, {});
test.AddOutput<float>("output", {0, 3}, {});
test.Run(OpTester::ExpectResult::kExpectSuccess, "",
{kTensorrtExecutionProvider, kWebGpuExecutionProvider});
}

} // namespace test
} // namespace onnxruntime
Loading