Skip to content

Use WebNN async API #51

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
Honry wants to merge 1 commit into
base: webnn_delegate_side_module
Choose a base branch
from
Open

Use WebNN async API #51

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
58 changes: 46 additions & 12 deletions tensorflow/lite/delegates/webnn/webnn_delegate.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -315,22 +315,35 @@ class Subgraph {
}
}


emscripten::val graph_inputs = emscripten::val::object();
for (auto& i : compute_inputs) {
std::string name = std::to_string(i);
auto input_size = context->tensors[i].bytes / 4;
graph_inputs.set(name, emscripten::val::global("Float32Array").new_(input_size));
}

emscripten::val graph_outputs = emscripten::val::object();
emscripten::val named_operands = emscripten::val::object();
for (auto o : outputs) {
std::string name = std::to_string(o);
if (!webnn_operands.at(o).as<bool>()) {
TF_LITE_KERNEL_LOG(context, "Invalid operand");
return nullptr;
}
auto output_size = context->tensors[o].bytes / 4;
graph_outputs.set(name, emscripten::val::global("Float32Array").new_(output_size));
named_operands.set(name, webnn_operands.at(o));
}

emscripten::val wnn_graph = wnn_builder.call<emscripten::val>("buildSync", named_operands);
emscripten::val wnn_graph = wnn_builder.call<emscripten::val>("build", named_operands).await();
if (!wnn_graph.as<bool>()) {
TF_LITE_KERNEL_LOG(context, "failed to build WebNN graph");
return nullptr;
}
return new Subgraph(delegate->wnn_context_, wnn_graph, std::move(compute_inputs), std::move(outputs));
return new Subgraph(delegate->wnn_context_, wnn_graph,
std::move(compute_inputs), std::move(outputs),
graph_inputs, graph_outputs);
}

TfLiteStatus Prepare(TfLiteContext* context) { return kTfLiteOk; }
Expand All @@ -356,27 +369,43 @@ class Subgraph {
}
}

std::unordered_map<std::string, emscripten::val> output_views;
if (any_pointers_changed) {
graph_inputs_ = emscripten::val::object();
for (int t : inputs_) {
std::string name = std::to_string(t);
auto input_size = context->tensors[t].bytes / 4;
auto input_data = context->tensors[t].data.f;
emscripten::val view{ emscripten::typed_memory_view(input_size, input_data) };
graph_inputs_.set(name, view);
graph_inputs_[name].call<void>("set", view);
}

graph_outputs_ = emscripten::val::object();
for (int t : outputs_) {
std::string name = std::to_string(t);
auto output_size = context->tensors[t].bytes / 4;
auto output_data = context->tensors[t].data.f;
emscripten::val view{emscripten::typed_memory_view(output_size, output_data)};
graph_outputs_.set(name, view);
output_views.insert({name, view});
}
}

const emscripten::val results =
wnn_context_
.call<emscripten::val>("compute", wnn_graph_, graph_inputs_, graph_outputs_)
.await();

if (any_pointers_changed) {
// Copy the outputs from pre-allocated ArrayBuffers back to the Wasm ArrayBuffer.
for (int t : outputs_) {
std::string name = std::to_string(t);
emscripten::val view = output_views.at(name);
view.call<void>("set", results["outputs"][name]);
}
}

wnn_context_.call<void>("computeSync", wnn_graph_, graph_inputs_, graph_outputs_);
// WebNN compute() method would return the input and output buffers via the
// promise resolution. Reuse the buffers to avoid additional allocation.
graph_inputs_ = results["inputs"];
graph_outputs_ = results["outputs"];

return kTfLiteOk;
}
Expand Down Expand Up @@ -3692,16 +3721,21 @@ class Subgraph {
}

private:
Subgraph(emscripten::val context, emscripten::val graph, std::unordered_set<int>&& inputs, std::unordered_set<int>&& outputs)
: wnn_context_(context), wnn_graph_(graph), inputs_(inputs), outputs_(outputs) {
Subgraph(emscripten::val context, emscripten::val graph,
std::unordered_set<int>&& inputs, std::unordered_set<int>&& outputs,
emscripten::val graph_inputs, emscripten::val graph_outputs)
: wnn_context_(context),
wnn_graph_(graph),
inputs_(inputs),
outputs_(outputs),
graph_inputs_(graph_inputs),
graph_outputs_(graph_outputs) {
for (auto& i : inputs_) {
externals_[i] = nullptr;
}
for (auto& o : outputs_) {
externals_[o] = nullptr;
}
graph_inputs_ = emscripten::val::object();
graph_outputs_ = emscripten::val::object();
}

emscripten::val wnn_context_ = emscripten::val::object();
Expand Down Expand Up @@ -4147,7 +4181,7 @@ TfLiteDelegate* TfLiteWebNNDelegateCreate(
emscripten::val(power_preference_name));
context_options.set("numThreads", options->numThreads);
emscripten::val wnn_context =
ml.call<emscripten::val>("createContextSync", context_options);
ml.call<emscripten::val>("createContext", context_options).await();

if (!wnn_context.as<bool>()) {
TFLITE_LOG_PROD_ONCE(tflite::TFLITE_LOG_ERROR,
Expand Down