Use official emsdk bazel toolchain

.bazelrc

@@ -6,22 +6,35 @@
 # editor's search path.
 build --symlink_prefix=dist/
 
-# Use our custom-configured c++ toolchain.
-build:wasm --crosstool_top=//toolchain:emscripten
+# These compile flags are active no matter which build mode we are in
+# (dbg vs opt). For flags specific to build mode, see cc_toolchain_config.bzl.
+build --cxxopt="-std=c++11"
+build --cxxopt="-fno-rtti"
+build --cxxopt="-fno-exceptions"
+build --cxxopt="-fomit-frame-pointer"
 
-# Use --cpu as a differentiator.
-build:wasm --cpu=wasm
+# Remote cache config. Users can add credentials in their .bazelrc.user files.
+build:remote --remote_cache=https://storage.googleapis.com/bazel-remote-cache-tfjs
 
-# Use the default C++ toolchain to build the tools used during the build.
-build:wasm --host_crosstool_top=@bazel_tools//tools/cpp:toolchain
+# Config for Google Cloud continuous integration that uses default credentials.
+build:ci --config=remote
+build:ci --google_default_credentials
 
-# These compile flags are active no matter which build mode we are in
-# (dbg vs opt). For flags specific to build mode, see cc_toolchain_config.bzl.
-build:wasm --cxxopt="-std=c++11"
-build:wasm --cxxopt="-fno-rtti"
-build:wasm --cxxopt="-fno-exceptions"
-build:wasm --cxxopt="-fomit-frame-pointer"
+# This flag is needed to prevent the bazel cache from being invalidated when
+# running bazel via `yarn bazel`.
+# See https://github.com/angular/angular/issues/27514.
+build --incompatible_strict_action_env
+run --incompatible_strict_action_env
+
+# Don't use a sandboxed build since it hurts performance.
+# When sandboxed, the wasm backend takes hours to build on a 32 core machine.
+build --spawn_strategy=worker,local
 
-# Disable sandbox environment because emsdk caches files by writing to
-# home directory.
-build:wasm --spawn_strategy=local
+# Load any settings specific to the current user.
+# .bazelrc.user should appear in .gitignore so that settings are not shared with
+# team members. This needs to be last statement in this config, as the user
+# configuration should be able to overwrite flags from this file.
+# See https://docs.bazel.build/versions/master/best-practices.html#bazelrc
+# (Note that we use .bazelrc.user so the file appears next to .bazelrc in
+# directory listing, rather than user.bazelrc as suggested in the Bazel docs).
+try-import %workspace%/.bazelrc.user

.gitignore

@@ -59,4 +59,7 @@ tfjs-backend-wasm/wasm-out/*.js
 tfjs-backend-wasm/wasm-out/*.wasm
 yalc.lock
 yarn-error.log
-cloudbuild_generated.yml
+cloudbuild_generated.yml
+
+# User-specific .bazelrc
+.bazelrc.user

WORKSPACE

@@ -12,18 +12,28 @@ yarn_install(
     yarn_lock = "//:yarn.lock",
 )
 
-# Make all files under $HOME/emsdk/* visible to the toolchain. The files are
-# available as external/emsdk/emsdk/*
-load("//toolchain:cc_toolchain_config.bzl", "emsdk_configure")
-emsdk_configure(name = "emsdk")
+# Emscripten toolchain
+http_archive(
+    name = "emsdk",
+    strip_prefix = "emsdk-c1589b55641787d55d53e883852035beea9aec3f/bazel",
+    url = "https://github.com/emscripten-core/emsdk/archive/c1589b55641787d55d53e883852035beea9aec3f.tar.gz",
+    sha256 = "7a58a9996b113d3e0675df30b5f17e28aa47de2e684a844f05394fe2f6f12e8e",
+)
+
+load("@emsdk//:deps.bzl", emsdk_deps = "deps")
+emsdk_deps()
+
+load("@emsdk//:emscripten_deps.bzl", emsdk_emscripten_deps = "emscripten_deps")
+emsdk_emscripten_deps()
+
 
 load("@bazel_tools//tools/build_defs/repo:git.bzl", "git_repository")
 # xnnpack used for fast vectorized wasm operations
 git_repository(
     name = "xnnpack",
-    commit = "55d53a4e7079d38e90acd75dd9e4f9e781d2da35",
+    commit = "3bfbdaf00211b313b143af39279bb6bf1f7effc0",
     remote = "https://github.com/google/XNNPACK.git",
-    shallow_since = "1614036677 -0800",
+    shallow_since = "1617056836 -0700",
 )
 
 # The libraries below are transitive dependencies of XNNPACK that we need to

tfjs-backend-wasm/scripts/build-ci.sh

@@ -17,5 +17,4 @@
 set -e
 
 yarn tsc
-
-./scripts/build-wasm.sh
+BAZEL_REMOTE="--config=ci" ./scripts/build-wasm.sh

tfjs-backend-wasm/scripts/build-wasm.sh

@@ -24,25 +24,28 @@ set -e
 set -x
 
 # Default build.
-yarn bazel build -c opt //tfjs-backend-wasm/src/cc:tfjs-backend-wasm.js --config=wasm
+yarn bazel build $BAZEL_REMOTE -c opt //tfjs-backend-wasm/src/cc:tfjs-backend-wasm
 # The typescript code and karma config expect the output of emscripten to be in
 # wasm-out/ so we copy the bazel output there.
-cp -f ../../dist/bin/tfjs-backend-wasm/src/cc/tfjs-backend-wasm.js \
-      ../../dist/bin/tfjs-backend-wasm/src/cc/tfjs-backend-wasm.wasm \
+cp -f ../../dist/bin/tfjs-backend-wasm/src/cc/tfjs-backend-wasm/tfjs-backend-wasm.js \
+      ../../dist/bin/tfjs-backend-wasm/src/cc/tfjs-backend-wasm/tfjs-backend-wasm.wasm \
       ../wasm-out/
 
 if [[ "$1" != "--dev" ]]; then
-  # SIMD build.
-  yarn bazel build -c opt //tfjs-backend-wasm/src/cc:tfjs-backend-wasm-simd.js --config=wasm --copt="-msimd128"
-  cp -f ../../dist/bin/tfjs-backend-wasm/src/cc/tfjs-backend-wasm-simd.wasm \
-        ../wasm-out/
-
-  # Threaded + SIMD build.
-  yarn bazel build -c opt //tfjs-backend-wasm/src/cc:tfjs-backend-wasm-threaded-simd.js --config=wasm --copt="-pthread" --copt="-msimd128"
-  cp -f ../../dist/bin/tfjs-backend-wasm/src/cc/tfjs-backend-wasm-threaded-simd.js \
-        ../../dist/bin/tfjs-backend-wasm/src/cc/tfjs-backend-wasm-threaded-simd.worker.js \
-        ../../dist/bin/tfjs-backend-wasm/src/cc/tfjs-backend-wasm-threaded-simd.wasm \
-        ../wasm-out/
+  # SIMD and threaded + SIMD builds.
+  yarn bazel build $BAZEL_REMOTE -c opt --copt="-msimd128" //tfjs-backend-wasm/src/cc:tfjs-backend-wasm-simd \
+    //tfjs-backend-wasm/src/cc:tfjs-backend-wasm-threaded-simd
+  # Copy SIMD
+  cp -f ../../dist/bin/tfjs-backend-wasm/src/cc/tfjs-backend-wasm-simd/tfjs-backend-wasm.wasm \
+    ../wasm-out/tfjs-backend-wasm-simd.wasm
+
+  # Copy threaded
+  cp -f ../../dist/bin/tfjs-backend-wasm/src/cc/tfjs-backend-wasm-threaded-simd/tfjs-backend-wasm.js \
+    ../wasm-out/tfjs-backend-wasm-threaded-simd.js
+  cp -f ../../dist/bin/tfjs-backend-wasm/src/cc/tfjs-backend-wasm-threaded-simd/tfjs-backend-wasm.worker.js \
+    ../wasm-out/tfjs-backend-wasm-threaded-simd.worker.js
+  cp -f ../../dist/bin/tfjs-backend-wasm/src/cc/tfjs-backend-wasm-threaded-simd/tfjs-backend-wasm.wasm \
+    ../wasm-out/tfjs-backend-wasm-threaded-simd.wasm
 
   node ./create-worker-module.js
   node ./patch-threaded-simd-module.js

tfjs-backend-wasm/scripts/create-worker-module.js

@@ -26,5 +26,6 @@ const BASE_PATH = '../wasm-out/';
 const WORKER_PATH = `${BASE_PATH}tfjs-backend-wasm-threaded-simd.worker.js`;
 
 const workerContents = fs.readFileSync(WORKER_PATH, "utf8");
+fs.chmodSync(WORKER_PATH, 0o644);
 fs.writeFileSync(`${WORKER_PATH}`,
   `export const wasmWorkerContents = '${workerContents.trim()}';`);

tfjs-backend-wasm/src/cc/BUILD

@@ -1,3 +1,4 @@
+load("@emsdk//emscripten_toolchain:wasm_rules.bzl", "wasm_cc_binary")
 load(":build_defs.bzl", "tfjs_cc_library", "tfjs_unit_test")
 
 # Emcripten produces a much larger wasm bundle unless the cc_binary has srcs
@@ -36,54 +37,22 @@ cc_binary(
     ],
 )
 
-# This build rule generates tfjs-backend-wasm-simd.{js,wasm}.
-#
-# We only need the .wasm file, not the .js file, because it will be loadsed by
-# the tfjs-backend-wasm.js file (generated from the previous build rule).
-#
-# See scripts/build-wasm.sh where we only copy the .wasm file to wasm-out.
-cc_binary(
-    name = "tfjs-backend-wasm-simd.js",
-    srcs = ["backend.cc"] + KERNELS_WITH_KEEPALIVE,
-    linkopts = [
-        "-s ALLOW_MEMORY_GROWTH=1",
-        "-s DEFAULT_LIBRARY_FUNCS_TO_INCLUDE=[]",
-        "-s DISABLE_EXCEPTION_CATCHING=1",
-        "-s FILESYSTEM=0",
-        "-s EXIT_RUNTIME=0",
-        "-s EXPORTED_FUNCTIONS='[\"_malloc\", \"_free\"]'",
-        "-s EXTRA_EXPORTED_RUNTIME_METHODS='[\"cwrap\"]'",
-        "-s MALLOC=emmalloc",
-    ],
-    deps = [
-        ":all_kernels",
-        ":backend",
-    ],
+wasm_cc_binary(
+    name = "tfjs-backend-wasm",
+    cc_target = ":tfjs-backend-wasm.js",
 )
 
-# This build rule generates tfjs-backend-wasm-threaded-simd.{js,wasm} and
-# tfjs-backend-wasm-threaded-simd.worker.js.
-cc_binary(
-    name = "tfjs-backend-wasm-threaded-simd.js",
-    srcs = ["backend.cc"] + KERNELS_WITH_KEEPALIVE,
-    linkopts = [
-        "-s ALLOW_MEMORY_GROWTH=1",
-        "-s DEFAULT_LIBRARY_FUNCS_TO_INCLUDE=[]",
-        "-s DISABLE_EXCEPTION_CATCHING=1",
-        "-s FILESYSTEM=0",
-        "-s EXIT_RUNTIME=0",
-        "-s EXPORTED_FUNCTIONS='[\"_malloc\", \"_free\"]'",
-        "-s EXTRA_EXPORTED_RUNTIME_METHODS='[\"cwrap\"]'",
-        "-s MODULARIZE=1",
-        "-s EXPORT_NAME=WasmBackendModuleThreadedSimd",
-        "-s MALLOC=emmalloc",
-        "-s USE_PTHREADS=1",
-        "-s PROXY_TO_PTHREAD=1",
-    ],
-    deps = [
-        ":all_kernels",
-        ":backend",
-    ],
+wasm_cc_binary(
+    name = "tfjs-backend-wasm-simd",
+    cc_target = ":tfjs-backend-wasm.js",
+    simd = True,
+)
+
+wasm_cc_binary(
+    name = "tfjs-backend-wasm-threaded-simd",
+    cc_target = ":tfjs-backend-wasm.js",
+    simd = True,
+    threads = "emscripten",
 )
 
 test_suite(

tfjs-backend-wasm/src/cc/binary.cc

@@ -61,7 +61,6 @@ void binary_xnn_f32(const size_t a_id, const size_t* a_shape_ptr,
   } else {
     binary_op = cache_result->second;
   }
-  const size_t batch_size = out_info.size;
   xnn_status status =
       setup_op(binary_op, a_shape_len, a_shape_ptr, b_shape_len, b_shape_ptr,
                a_buf, b_buf, out_buf, tfjs::backend::threadpool);

tfjs-backend-wasm/src/cc/kernels/All.cc

@@ -33,7 +33,6 @@ void All(const size_t x_id, const size_t reduce_size, const size_t out_id) {
   auto& out_info = backend::get_tensor_info_out(out_id);
 
   const bool* x_buf = x_info.b();
-  const size_t x_size = x_info.size;
 
   bool* out_buf = out_info.b_write();
   const size_t out_size = out_info.size;

tfjs-backend-wasm/src/cc/kernels/Any.cc

@@ -33,7 +33,6 @@ void Any(const size_t x_id, const size_t reduce_size, const size_t out_id) {
   auto& out_info = backend::get_tensor_info_out(out_id);
 
   const bool* x_buf = x_info.b();
-  const size_t x_size = x_info.size;
 
   bool* out_buf = out_info.b_write();
   const size_t out_size = out_info.size;

tfjs-backend-wasm/src/cc/kernels/CropAndResize.cc

@@ -91,16 +91,12 @@ void CropAndResize(size_t images_id, size_t boxes_id, size_t box_ind_id,
   auto& out_info = backend::get_tensor_info_out(out_id);
 
   const float* images_buf = images_info.f32();
-  const size_t images_size = images_info.size;
 
   const float* boxes_buf = boxes_info.f32();
-  const size_t boxes_size = boxes_info.size;
 
   const int* box_ind_buf = box_ind_info.i32();
-  const size_t box_ind_size = box_ind_info.size;
 
   float* out_buf = out_info.f32_write();
-  const size_t out_size = out_info.size;
 
   const size_t batch = images_shape[0];
   const size_t image_height = images_shape[1];

tfjs-backend-wasm/src/cc/kernels/FloorDiv.cc

@@ -35,7 +35,6 @@ void FloorDiv(const size_t a_id, const size_t* a_shape_ptr,
               const size_t a_shape_len, const size_t b_id,
               const size_t* b_shape_ptr, const size_t b_shape_len,
               const DType dtype, const size_t out_id) {
-  auto& a_info = backend::get_tensor_info(a_id);
   switch (dtype) {
     case DType::float32:
       binary_f32(a_id, b_id, out_id,

tfjs-backend-wasm/src/cc/kernels/Max.cc

@@ -33,7 +33,6 @@ void Max(const size_t x_id, const size_t reduce_size, const size_t out_id) {
   auto& out_info = backend::get_tensor_info_out(out_id);
 
   const float* x_buf = x_info.f32();
-  const size_t x_size = x_info.size;
 
   float* out_buf = out_info.f32_write();
   const size_t out_size = out_info.size;

tfjs-backend-wasm/src/cc/kernels/Mean.cc

@@ -34,15 +34,13 @@ void Mean(const size_t x_id, const size_t reduce_size, const size_t out_id) {
   auto& out_info = backend::get_tensor_info_out(out_id);
 
   const float* x_buf = x_info.f32();
-  const size_t x_size = x_info.size;
 
   float* out_buf = out_info.f32_write();
   const size_t out_size = out_info.size;
 
   const float* x_offset = x_buf;
 
   for (size_t i = 0; i < out_size; ++i) {
-    const size_t offset = i * reduce_size;
     float sum = 0;
 
     const float* x_iter_end = x_offset + reduce_size;

tfjs-backend-wasm/src/cc/kernels/Min.cc

@@ -33,7 +33,6 @@ void Min(const size_t x_id, const size_t reduce_size, const size_t out_id) {
   auto& out_info = backend::get_tensor_info_out(out_id);
 
   const float* x_buf = x_info.f32();
-  const size_t x_size = x_info.size;
 
   float* out_buf = out_info.f32_write();
   const size_t out_size = out_info.size;

tfjs-backend-wasm/src/cc/kernels/Pow.cc

@@ -27,8 +27,13 @@ template <class T>
 inline T power(T a, T b) {
   return pow(a, b);
 }
+
+inline bool power_bool(bool a, bool b) {
+  return static_cast<bool>(pow(static_cast<float>(a), static_cast<float>(b)));
+}
 }  // namespace
 
+
 namespace tfjs {
 namespace wasm {
 // We use C-style API to interface with Javascript.
@@ -48,7 +53,7 @@ void Pow(const size_t a_id, const size_t* a_shape_ptr, const size_t a_shape_len,
       binary_i32(a_id, b_id, out_id, power<int32_t>);
       break;
     case DType::boolean:
-      binary_bool(a_id, b_id, out_id, power<bool>);
+      binary_bool(a_id, b_id, out_id, power_bool);
       break;
     default:
       util::warn("Pow for tensor ids %d and %d failed. Unknown dtype %d", a_id,

tfjs-backend-wasm/src/cc/kernels/Prod.cc

@@ -29,15 +29,13 @@ void prod(const size_t x_id, const size_t reduce_size, const size_t out_id) {
   auto& out_info = tfjs::backend::get_tensor_info_out(out_id);
 
   const T* x_buf = reinterpret_cast<const T *>(x_info.memory_offset);
-  const size_t x_size = x_info.size;
 
   T* out_buf = reinterpret_cast<T *>(out_info.memory_offset);
   const size_t out_size = out_info.size;
 
   const T* x_offset = x_buf;
 
   for (size_t i = 0; i < out_size; ++i) {
-    const size_t offset = i * reduce_size;
     T product = 1;
 
     const T* x_iter_end = x_offset + reduce_size;

tfjs-backend-wasm/src/cc/kernels/Sum.cc

@@ -34,15 +34,13 @@ void Sum(const size_t x_id, const size_t reduce_size, const size_t out_id) {
   auto& out_info = backend::get_tensor_info_out(out_id);
 
   const float* x_buf = x_info.f32();
-  const size_t x_size = x_info.size;
 
   float* out_buf = out_info.f32_write();
   const size_t out_size = out_info.size;
 
   const float* x_offset = x_buf;
 
   for (size_t i = 0; i < out_size; ++i) {
-    const size_t offset = i * reduce_size;
     float sum = 0;
 
     const float* x_iter_end = x_offset + reduce_size;

tfjs-backend-wasm/src/cc/non_max_suppression_impl.cc

@@ -116,7 +116,7 @@ const NonMaxSuppressionResult* non_max_suppression_impl(
   std::vector<int32_t> selected_indices;
   std::vector<float> selected_scores;
   Candidate candidate;
-  float iou, original_score;
+  float original_score;
 
   while (selected_indices.size() < max_out_size &&
          !candidate_priority_queue.empty()) {

tfjs-backend-wasm/yarn.lock

@@ -984,18 +984,11 @@
 
 "@tensorflow/tfjs-backend-cpu@link:../tfjs-backend-cpu":
   version "0.0.0"
-  dependencies:
-    "@types/seedrandom" "2.4.27"
-    seedrandom "2.4.3"
+  uid ""
 
 "@tensorflow/tfjs-core@link:../tfjs-core":
   version "0.0.0"
-  dependencies:
-    "@types/offscreencanvas" "~2019.3.0"
-    "@types/seedrandom" "2.4.27"
-    "@types/webgl-ext" "0.0.30"
-    node-fetch "~2.6.1"
-    seedrandom "2.4.3"
+  uid ""
 
 "@types/emscripten@~0.0.34":
   version "0.0.34"