Skip to content

Commit 5911422

Browse files
HonryHonry
authored
[WebNN EP] Remove NHWC preferred layout (#21570)
Currently WebNN CPU backend has supported NCHW layout in Chromium, we can now drop NHWC preferred layout for CPU backend in WebNN EP to simplify the code.
1 parent bf8855b commit 5911422

10 files changed

+39
-296
lines changed

‎onnxruntime/core/providers/webnn/builders/impl/builder_utils.cc

+9-12
Original file line numberDiff line numberDiff line change
@@ -19,10 +19,9 @@ common::Status ComputeConvPads(const std::vector<int64_t> input_shape,
1919
const std::vector<int64_t>& onnx_strides,
2020
const std::vector<int64_t>& onnx_dilations,
2121
AutoPadType auto_pad_type,
22-
std::vector<int64_t>& pads_out,
23-
bool use_nchw) {
24-
const int64_t input_size_y = use_nchw ? input_shape[2] : input_shape[1];
25-
const int64_t input_size_x = use_nchw ? input_shape[3] : input_shape[2];
22+
std::vector<int64_t>& pads_out) {
23+
const int64_t input_size_y = input_shape[2];
24+
const int64_t input_size_x = input_shape[3];
2625
const int64_t stride_y = onnx_strides[0];
2726
const int64_t stride_x = onnx_strides[1];
2827
const int64_t dilation_y = onnx_dilations[0];
@@ -54,16 +53,15 @@ common::Status HandleAutoPad(const std::vector<int64_t> input_shape,
5453
const std::vector<int64_t>& onnx_strides,
5554
const std::vector<int64_t>& onnx_dilations,
5655
AutoPadType auto_pad_type,
57-
std::vector<int64_t>& pads_out,
58-
bool use_nchw) {
56+
std::vector<int64_t>& pads_out) {
5957
if (AutoPadType::SAME_UPPER == auto_pad_type) {
6058
ORT_RETURN_IF_ERROR(ComputeConvPads(input_shape, weight_size_y, weight_size_x,
6159
onnx_pads, onnx_strides, onnx_dilations,
62-
AutoPadType::SAME_UPPER, pads_out, use_nchw));
60+
AutoPadType::SAME_UPPER, pads_out));
6361
} else {
6462
ORT_RETURN_IF_ERROR(ComputeConvPads(input_shape, weight_size_y, weight_size_x,
6563
onnx_pads, onnx_strides, onnx_dilations,
66-
AutoPadType::SAME_LOWER, pads_out, use_nchw));
64+
AutoPadType::SAME_LOWER, pads_out));
6765
}
6866
return Status::OK();
6967
}
@@ -111,10 +109,9 @@ common::Status ComputeConvTransposePadsAndOutputShape(const std::vector<int64_t>
111109
const std::vector<int64_t>& onnx_output_padding,
112110
AutoPadType auto_pad_type,
113111
std::vector<int64_t>& pads_out,
114-
std::vector<int64_t>& output_shape_out,
115-
bool use_nchw) {
116-
const int64_t input_size_y = use_nchw ? input_shape[2] : input_shape[1];
117-
const int64_t input_size_x = use_nchw ? input_shape[3] : input_shape[2];
112+
std::vector<int64_t>& output_shape_out) {
113+
const int64_t input_size_y = input_shape[2];
114+
const int64_t input_size_x = input_shape[3];
118115
const int64_t stride_y = onnx_strides[0];
119116
const int64_t stride_x = onnx_strides[1];
120117
const int64_t dilation_y = onnx_dilations[0];

‎onnxruntime/core/providers/webnn/builders/impl/builder_utils.h

+2-4
Original file line numberDiff line numberDiff line change
@@ -21,8 +21,7 @@ common::Status HandleAutoPad(const std::vector<int64_t> input_shape,
2121
const std::vector<int64_t>& onnx_strides,
2222
const std::vector<int64_t>& onnx_dilations,
2323
AutoPadType auto_pad_type,
24-
std::vector<int64_t>& pads_out,
25-
bool use_nchw) ORT_MUST_USE_RESULT;
24+
std::vector<int64_t>& pads_out) ORT_MUST_USE_RESULT;
2625

2726
// Compute pads and output shape for ConvTranspose.
2827
common::Status ComputeConvTransposePadsAndOutputShape(const std::vector<int64_t> input_shape,
@@ -34,8 +33,7 @@ common::Status ComputeConvTransposePadsAndOutputShape(const std::vector<int64_t>
3433
const std::vector<int64_t>& onnx_output_padding,
3534
AutoPadType auto_pad_type,
3635
std::vector<int64_t>& pads_out,
37-
std::vector<int64_t>& output_shape_out,
38-
bool use_nchw) ORT_MUST_USE_RESULT;
36+
std::vector<int64_t>& output_shape_out) ORT_MUST_USE_RESULT;
3937

4038
} // namespace webnn
4139
} // namespace onnxruntime

‎onnxruntime/core/providers/webnn/builders/impl/conv_op_builder.cc

+12-159
Original file line numberDiff line numberDiff line change
@@ -18,9 +18,6 @@ namespace webnn {
1818

1919
class ConvOpBuilder : public BaseOpBuilder {
2020
// Add operator related.
21-
public:
22-
void AddInitializersToSkip(ModelBuilder& model_builder, const Node& node) const override;
23-
2421
private:
2522
Status AddToModelBuilderImpl(ModelBuilder& model_builder, const Node& node,
2623
const logging::Logger& logger) const override ORT_MUST_USE_RESULT;
@@ -33,13 +30,6 @@ class ConvOpBuilder : public BaseOpBuilder {
3330
const logging::Logger& logger) const override;
3431
};
3532

36-
void ConvOpBuilder::AddInitializersToSkip(ModelBuilder& model_builder, const Node& node) const {
37-
// skip the weight for conv as we need to transpose for preferred layout NHWC.
38-
if (model_builder.GetPreferredLayout() == DataLayout::NHWC) {
39-
model_builder.AddInitializerToSkip(node.InputDefs()[1]->Name()); // W
40-
}
41-
}
42-
4333
// Helper functions
4434
common::Status SetConvBaseOptions(ModelBuilder& model_builder,
4535
const Node& node, emscripten::val& options,
@@ -48,7 +38,6 @@ common::Status SetConvBaseOptions(ModelBuilder& model_builder,
4838
const std::vector<int64_t>& strides,
4939
const std::vector<int64_t>& dilations,
5040
std::vector<int64_t>& pads,
51-
const bool is_nhwc,
5241
const bool is_conv1d,
5342
const logging::Logger& logger) {
5443
NodeAttrHelper helper(node);
@@ -61,7 +50,7 @@ common::Status SetConvBaseOptions(ModelBuilder& model_builder,
6150
// Calculate explicit padding for autoPad.
6251
if (AutoPadType::SAME_UPPER == auto_pad_type || AutoPadType::SAME_LOWER == auto_pad_type) {
6352
ORT_RETURN_IF_ERROR(HandleAutoPad(input_shape, weight_shape[2], weight_shape[3],
64-
pads, strides, dilations, auto_pad_type, pads_out, !is_nhwc));
53+
pads, strides, dilations, auto_pad_type, pads_out));
6554
pads = pads_out;
6655
}
6756
} else if (node.OpType() == "ConvTranspose") {
@@ -82,7 +71,7 @@ common::Status SetConvBaseOptions(ModelBuilder& model_builder,
8271
// Otherwise compute the output shape, as well as the pads if the auto_pad attribute is SAME_UPPER/SAME_LOWER.
8372
ORT_RETURN_IF_ERROR(ComputeConvTransposePadsAndOutputShape(input_shape, weight_shape[2], weight_shape[3],
8473
pads, strides, dilations, output_padding,
85-
auto_pad_type, pads_out, output_shape, !is_nhwc));
74+
auto_pad_type, pads_out, output_shape));
8675

8776
if (output_shape[0] != -1 && output_shape[1] != -1) {
8877
options.set("outputSizes", emscripten::val::array(GetVecUint32FromVecInt64(output_shape)));
@@ -111,89 +100,6 @@ common::Status SetConvBaseOptions(ModelBuilder& model_builder,
111100
return Status::OK();
112101
}
113102

114-
// Both depthwise Conv and ConvTranspose share the same logic to add the layout.
115-
Status AddInitializerInNewLayout(ModelBuilder& model_builder,
116-
const std::string& name,
117-
bool is_conv,
118-
bool is_conv1d) {
119-
const auto& tensor = *model_builder.GetInitializerTensors().at(name);
120-
auto data_type = tensor.data_type();
121-
122-
const auto& shape = tensor.dims();
123-
std::vector<uint32_t> dims = GetVecUint32FromVecInt64(std::vector<int64_t>(std::begin(shape), std::end(shape)));
124-
125-
if (is_conv1d) {
126-
// Support conv1d by prepending a 1 size dimension.
127-
dims.push_back(1);
128-
}
129-
130-
const uint8_t* src = nullptr;
131-
Initializer unpacked_tensor(tensor, model_builder.GetGraphViewer().ModelPath());
132-
src = unpacked_tensor.DataAsByteSpan().data();
133-
const auto out_t = dims[0], in_t = dims[1],
134-
h_t = dims[2], w_t = dims[3];
135-
std::vector<uint32_t> dest_shape;
136-
if (is_conv == 1)
137-
dest_shape = {out_t, h_t, w_t, in_t}; // L_0231
138-
else
139-
dest_shape = {in_t, h_t, w_t, out_t}; // L_1230 for depthwise conv and convTranspose weight
140-
141-
SafeInt<size_t> num_elements = SafeInt<size_t>(Product(dest_shape));
142-
143-
size_t element_size{0};
144-
switch (data_type) {
145-
case ONNX_NAMESPACE::TensorProto_DataType_UINT8:
146-
element_size = sizeof(uint8_t);
147-
break;
148-
case ONNX_NAMESPACE::TensorProto_DataType_INT8:
149-
element_size = sizeof(int8_t);
150-
break;
151-
case ONNX_NAMESPACE::TensorProto_DataType_FLOAT16:
152-
element_size = sizeof(uint16_t);
153-
break;
154-
case ONNX_NAMESPACE::TensorProto_DataType_FLOAT:
155-
element_size = sizeof(float);
156-
break;
157-
default:
158-
break;
159-
}
160-
std::unique_ptr<uint8_t[]> buffer_holder(new uint8_t[element_size * num_elements]);
161-
uint8_t* buffer = buffer_holder.get();
162-
163-
for (uint32_t out = 0; out < out_t; out++) {
164-
for (uint32_t in = 0; in < in_t; in++) {
165-
for (uint32_t h = 0; h < h_t; h++) {
166-
for (uint32_t w = 0; w < w_t; w++) {
167-
auto onnx_idx = out * in_t * h_t * w_t +
168-
in * h_t * w_t +
169-
h * w_t +
170-
w;
171-
172-
uint32_t nnapi_idx;
173-
if (is_conv == 1) { // L_0231
174-
nnapi_idx = out * h_t * w_t * in_t +
175-
h * w_t * in_t +
176-
w * in_t +
177-
in;
178-
} else { // L_1230 for depthwise conv weight
179-
nnapi_idx = in * h_t * w_t * out_t +
180-
h * w_t * out_t +
181-
w * out_t +
182-
out;
183-
}
184-
185-
for (size_t i = 0; i < element_size; i++) {
186-
buffer[element_size * nnapi_idx + i] = src[element_size * onnx_idx + i];
187-
}
188-
}
189-
}
190-
}
191-
}
192-
ORT_RETURN_IF_ERROR(model_builder.AddOperandFromPersistMemoryBuffer(name, buffer, num_elements * element_size,
193-
dest_shape, data_type));
194-
return Status::OK();
195-
}
196-
197103
// Add operator related.
198104

199105
Status ConvOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder, const Node& node,
@@ -203,7 +109,6 @@ Status ConvOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder, const N
203109
const auto& op_type = node.OpType();
204110
emscripten::val input = model_builder.GetOperand(input_defs[0]->Name());
205111
emscripten::val output = emscripten::val::object();
206-
const auto& initializers(model_builder.GetInitializerTensors());
207112

208113
std::vector<int64_t> input_shape;
209114
ORT_RETURN_IF_NOT(GetShape(*input_defs[0], input_shape, logger), "Cannot get input shape");
@@ -216,19 +121,11 @@ Status ConvOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder, const N
216121
auto dilations = helper.Get("dilations", std::vector<int64_t>{1, 1});
217122
auto pads = helper.Get("pads", std::vector<int64_t>{0, 0, 0, 0});
218123

219-
const bool is_nhwc = model_builder.GetPreferredLayout() == DataLayout::NHWC;
220124
const bool is_conv1d = input_shape.size() == 3 && weight_shape.size() == 3;
221-
const bool is_constant_weight = Contains(initializers, weight_name);
222125
// Support conv1d by prepending a 1 or 2 size dimensions.
223126
if (is_conv1d) {
224127
// Reshape input.
225-
if (is_nhwc) {
226-
// For NHWC preferred layout, the input has been transposed.
227-
// For conv1d it is NCD1 -> ND1C, so we need to prepend 1 to the index 2.
228-
input_shape.insert(input_shape.begin() + 2, 1);
229-
} else {
230-
input_shape.push_back(1);
231-
}
128+
input_shape.push_back(1);
232129
std::vector<uint32_t> new_shape = GetVecUint32FromVecInt64(input_shape);
233130
input = model_builder.GetBuilder().call<emscripten::val>("reshape", input, emscripten::val::array(new_shape));
234131

@@ -244,63 +141,19 @@ Status ConvOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder, const N
244141
emscripten::val options = emscripten::val::object();
245142
options.set("label", node.Name());
246143
ORT_RETURN_IF_ERROR(SetConvBaseOptions(
247-
model_builder, node, options, input_shape, weight_shape, strides, dilations, pads, is_nhwc, is_conv1d, logger));
248-
bool depthwise = false;
249-
if (op_type == "Conv" || op_type == "ConvInteger") {
250-
int groups = options["groups"].as<int>();
251-
if (is_nhwc) {
252-
depthwise = (groups == input_shape[3] && groups != 1);
253-
options.set("inputLayout", emscripten::val("nhwc"));
254-
if (is_constant_weight) {
255-
ORT_RETURN_IF_ERROR(AddInitializerInNewLayout(model_builder, weight_name, !depthwise, is_conv1d));
256-
}
257-
if (!depthwise) {
258-
options.set("filterLayout", emscripten::val("ohwi"));
259-
} else {
260-
options.set("filterLayout", emscripten::val("ihwo"));
261-
}
262-
}
263-
} else { // ConvTranspose
264-
if (is_nhwc) {
265-
options.set("inputLayout", emscripten::val("nhwc"));
266-
options.set("filterLayout", emscripten::val("ohwi"));
267-
if (is_constant_weight) {
268-
ORT_RETURN_IF_ERROR(AddInitializerInNewLayout(model_builder, weight_name, true, is_conv1d));
269-
}
270-
}
271-
}
272-
144+
model_builder, node, options, input_shape, weight_shape, strides, dilations, pads, is_conv1d, logger));
273145
emscripten::val filter = model_builder.GetOperand(weight_name);
274146

275147
if (is_conv1d) {
276148
// Reshape weight to 4D for conv1d.
277-
if (!is_nhwc || !is_constant_weight) {
278-
// The weight_shape has been appended 1's, reshape weight operand.
279-
std::vector<uint32_t> new_shape = GetVecUint32FromVecInt64(weight_shape);
280-
emscripten::val reshape_options = emscripten::val::object();
281-
reshape_options.set("label", node.Name() + "_reshape_filter");
282-
filter = model_builder.GetBuilder().call<emscripten::val>("reshape",
283-
filter,
284-
emscripten::val::array(new_shape),
285-
reshape_options);
286-
}
287-
}
288-
289-
emscripten::val transpose_options = emscripten::val::object();
290-
if (is_nhwc && !is_constant_weight) {
291-
// For NHWC preferred layout, if the weight is input:
292-
// - Transpose it from iohw -> ohwi for convTranspose.
293-
// - Transpose it from oihw -> ihwo for depthwise conv.
294-
// - Transpose it from oihw -> ohwi for conv.
295-
std::vector<uint32_t> perm(4);
296-
if (op_type == "ConvTranspose" || depthwise) {
297-
perm = {1, 2, 3, 0}; // L_1230 for depthwise conv and convTranspose weight
298-
} else {
299-
perm = {0, 2, 3, 1}; // L_0231
300-
}
301-
transpose_options.set("permutation", emscripten::val::array(perm));
302-
transpose_options.set("label", node.Name() + "_transpose_filter");
303-
filter = model_builder.GetBuilder().call<emscripten::val>("transpose", filter, transpose_options);
149+
// The weight_shape has been appended 1's, reshape weight operand.
150+
std::vector<uint32_t> new_shape = GetVecUint32FromVecInt64(weight_shape);
151+
emscripten::val reshape_options = emscripten::val::object();
152+
reshape_options.set("label", node.Name() + "_reshape_filter");
153+
filter = model_builder.GetBuilder().call<emscripten::val>("reshape",
154+
filter,
155+
emscripten::val::array(new_shape),
156+
reshape_options);
304157
}
305158

306159
if (op_type == "Conv") {

‎onnxruntime/core/providers/webnn/builders/impl/normalization_op_builder.cc

+1-8
Original file line numberDiff line numberDiff line change
@@ -79,9 +79,6 @@ Status NormalizationOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder
7979
ORT_RETURN_IF_NOT(input_defs.size() == 5, "BatchNormalization requires five inputs.");
8080
emscripten::val mean = model_builder.GetOperand(input_defs[3]->Name());
8181
emscripten::val variance = model_builder.GetOperand(input_defs[4]->Name());
82-
if (model_builder.GetPreferredLayout() == DataLayout::NHWC) {
83-
options.set("axis", rank - 1);
84-
}
8582

8683
output = model_builder.GetBuilder().call<emscripten::val>("batchNormalization", input, mean, variance, options);
8784
} else if (op_type == "LayerNormalization") {
@@ -104,9 +101,8 @@ Status NormalizationOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder
104101
std::back_inserter(new_shape),
105102
[](int64_t dim) -> uint32_t { return SafeInt<uint32_t>(dim); });
106103

107-
size_t insertion_offset = (model_builder.GetPreferredLayout() == DataLayout::NHWC) ? 2 : 3;
108104
ptrdiff_t excess_rank = new_shape.size() - webnn_shape_rank;
109-
auto insertion_point = new_shape.begin() + insertion_offset;
105+
auto insertion_point = new_shape.begin() + 3;
110106
if (input_shape.size() < webnn_shape_rank) {
111107
// Pad the shape with extra 1's to satisfy WebNN v1's rank requirements.
112108
new_shape.insert(insertion_point, -excess_rank, 1);
@@ -125,9 +121,6 @@ Status NormalizationOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder
125121
reshape_input_options);
126122
}
127123

128-
if (model_builder.GetPreferredLayout() == DataLayout::NHWC) {
129-
options.set("layout", emscripten::val("nhwc"));
130-
}
131124
output = model_builder.GetBuilder().call<emscripten::val>("instanceNormalization", input, options);
132125
// Reshape back to the original output shape for 3D input.
133126
if (input_shape.size() != 4) {

‎onnxruntime/core/providers/webnn/builders/impl/pool_op_builder.cc

+2-7
Original file line numberDiff line numberDiff line change
@@ -70,11 +70,7 @@ Status PoolOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder,
7070
options.set("strides", emscripten::val::array(strides));
7171
const auto dilations = helper.Get("dilations", std::vector<int32_t>{1, 1});
7272
options.set("dilations", emscripten::val::array(dilations));
73-
if (model_builder.GetPreferredLayout() == DataLayout::NHWC) {
74-
options.set("layout", emscripten::val("nhwc"));
75-
} else {
76-
options.set("layout", emscripten::val("nchw"));
77-
}
73+
options.set("layout", emscripten::val("nchw"));
7874

7975
// Add Padding.
8076
// Usually using autopadding is more efficient than using explicit padding.
@@ -93,8 +89,7 @@ Status PoolOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder,
9389
helper.Get("strides", std::vector<int64_t>{1, 1}),
9490
helper.Get("dilations", std::vector<int64_t>{1, 1}),
9591
auto_pad_type,
96-
pads_out,
97-
model_builder.GetPreferredLayout() == DataLayout::NCHW));
92+
pads_out));
9893
pads = GetVecUint32FromVecInt64(pads_out);
9994
}
10095
// Permute the ONNX's pads, which is [beginning_height, beginning_width, ending_height, ending_width],

0 commit comments

Comments
 (0)