saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-23 19:32:23 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

[WebNN] QDQ's axis should be used for broadcasting (#22721)

Browse source 
For per-axis quantization/dequantization, WebNN requires the scale and
zero_point inputs to be broadcastable. Axis should be used for reshape
these two inputs.
This commit is contained in:
Wanming Lin 2024-11-10 10:19:46 +08:00 committed by GitHub
parent d3ad76b2cf
commit b9b1a0353a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 47 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

49
onnxruntime/core/providers/webnn/builders/impl/qdq_op_builder.cc
View file

@ -35,11 +35,13 @@ Status QDQOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder,
std::vector<int64_t> input_shape;
std::vector<int64_t> scale_shape;
std::vector<uint32_t> zero_point_shape;
ORT_RETURN_IF_NOT(GetShape(*input_defs[0], input_shape, logger), "Cannot get input shape");
ORT_RETURN_IF_NOT(GetShape(*input_defs[1], scale_shape, logger), "Cannot get scale shape");
int32_t input_type = 0;
int32_t output_type = 0;
int32_t zero_point_type = 0;
bool has_zero_point = false;
ORT_RETURN_IF_NOT(GetType(*input_defs[0], input_type, logger), "Cannot get input data type");
ORT_RETURN_IF_NOT(GetType(*output_defs[0], output_type, logger), "Cannot get output data type");
@ -49,12 +51,55 @@ Status QDQOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder,
if (input_defs.size() == 3 && input_defs[2]->Exists()) {
zero_point = model_builder.GetOperand(node.InputDefs()[2]->Name());
has_zero_point = true;
} else {
// DequantizeLinear: x_zero_point's data type equals to input data type
// QuantizeLinear: x_zero_point's data type equals to output data type
// WebNN requires the zero_point to have the same shape as the scale
zero_point_type = op_type == "DequantizeLinear" ? input_type : output_type;
const auto zero_point_shape = GetVecUint32FromVecInt64(scale_shape);
}
const auto input_rank = input_shape.size();
NodeAttrHelper helper(node);
int32_t block_size = helper.Get("block_size", 0);
int32_t axis = helper.Get("axis", 1);
if (axis < 0) {
axis = SafeInt<int32_t>(HandleNegativeAxis(axis, input_rank));
}
// For per-axis quantization/dequantization and axis is not equal to input_rank - 1,
// we need to reshape the scale and zero_point tensors to make them broadcastable with the input tensor.
if (scale_shape.size() == 1 && input_rank > 1 &&
block_size == 0 && axis != static_cast<int32_t>(input_rank - 1)) {
// Insert ones before and after the axis dimension for broadcasting of scale tensor.
std::vector<uint32_t> target_shape{SafeInt<uint32_t>(input_shape[axis])};
target_shape.insert(target_shape.begin(), axis, 1);
target_shape.insert(target_shape.end(), input_rank - axis - 1, 1);
// zero_point has the same shape as the scale tensor.
zero_point_shape = target_shape;
emscripten::val reshape_scale_options = emscripten::val::object();
reshape_scale_options.set("label", node.Name() + "_reshape_scale");
scale = model_builder.GetBuilder().call<emscripten::val>("reshape",
scale,
emscripten::val::array(target_shape),
reshape_scale_options);
if (has_zero_point) {
// Reshape the zero_point tensor too.
emscripten::val reshape_zero_point_options = emscripten::val::object();
reshape_zero_point_options.set("label", node.Name() + "_reshape_zero_point");
zero_point = model_builder.GetBuilder().call<emscripten::val>("reshape",
zero_point,
emscripten::val::array(target_shape),
reshape_zero_point_options);
}
}
// If zero_point is not provided, create a zero constant with the same shape as the scale tensor.
if (!has_zero_point) {
if (zero_point_shape.empty()) {
// zero_point has the same shape as the scale tensor.
zero_point_shape = GetVecUint32FromVecInt64(scale_shape);
}
zero_point = model_builder.GetZeroConstant(zero_point_type, zero_point_shape);
}