Share threadpool framework between xnnpack and optimized kernels (#11194)

Author: shoumikhin

Package.swift

@@ -19,8 +19,10 @@
 
 import PackageDescription
 
-let debug = "_debug"
-let deliverables = [
+let debug_suffix = "_debug"
+let dependencies_suffix = "_with_dependencies"
+
+let products = [
   "backend_coreml": [
     "frameworks": [
       "Accelerate",
@@ -37,49 +39,64 @@ let deliverables = [
       "MetalPerformanceShadersGraph",
     ],
   ],
-  "backend_xnnpack": [:],
-  "executorch": [:],
+  "backend_xnnpack": [
+    "targets": [
+      "threadpool",
+    ],
+  ],
+  "executorch": [
+    "libraries": [
+      "c++",
+    ],
+  ],
   "kernels_custom": [:],
-  "kernels_optimized": [:],
+  "kernels_optimized": [
+    "frameworks": [
+      "Accelerate",
+    ],
+    "targets": [
+      "threadpool",
+    ],
+  ],
   "kernels_portable": [:],
   "kernels_quantized": [:],
 ].reduce(into: [String: [String: Any]]()) {
   $0[$1.key] = $1.value
-  $0[$1.key + debug] = $1.value
-}.reduce(into: [String: [String: Any]]()) {
-  var newValue = $1.value
-  if $1.key.hasSuffix(debug) {
-    $1.value.forEach { key, value in
-      if key.hasSuffix(debug) {
-        newValue[String(key.dropLast(debug.count))] = value
-      }
-    }
-  }
-  $0[$1.key] = newValue.filter { key, _ in !key.hasSuffix(debug) }
+  $0[$1.key + debug_suffix] = $1.value
 }
 
+let targets = [
+  "threadpool",
+].flatMap { [$0, $0 + debug_suffix] }
+
 let package = Package(
   name: "executorch",
   platforms: [
     .iOS(.v17),
     .macOS(.v10_15),
   ],
-  products: deliverables.keys.map { key in
-    .library(name: key, targets: ["\(key)_dependencies"])
+  products: products.keys.map { key in
+    .library(name: key, targets: ["\(key)\(dependencies_suffix)"])
   }.sorted { $0.name < $1.name },
-  targets: deliverables.flatMap { key, value -> [Target] in
+  targets: targets.map { key in
+    .binaryTarget(
+      name: key,
+      path: "cmake-out/\(key).xcframework"
+    )
+  } + products.flatMap { key, value -> [Target] in
     [
       .binaryTarget(
         name: key,
         path: "cmake-out/\(key).xcframework"
       ),
       .target(
-        name: "\(key)_dependencies",
-        dependencies: [.target(name: key)],
+        name: "\(key)\(dependencies_suffix)",
+        dependencies:([key] +
+          (value["targets"] as? [String] ?? []).map {
+            target in key.hasSuffix(debug_suffix) ? target + debug_suffix : target
+          }).map { .target(name: $0) },
         path: ".Package.swift/\(key)",
-        linkerSettings: [
-          .linkedLibrary("c++")
-        ] +
+        linkerSettings:
           (value["frameworks"] as? [String] ?? []).map { .linkedFramework($0) } +
           (value["libraries"] as? [String] ?? []).map { .linkedLibrary($0) }
       ),
@@ -88,18 +105,17 @@ let package = Package(
     .testTarget(
       name: "tests",
       dependencies: [
-        .target(name: "executorch_debug"),
-        .target(name: "kernels_portable"),
+        .target(name: "executorch\(debug_suffix)"),
+        .target(name: "kernels_optimized\(dependencies_suffix)"),
       ],
       path: "extension/apple/ExecuTorch/__tests__",
       resources: [
-        .copy("resources/add.pte")
+        .copy("resources/add.pte"),
       ],
       linkerSettings: [
-        .linkedLibrary("c++"),
         .unsafeFlags([
           "-Xlinker", "-force_load",
-          "-Xlinker", "cmake-out/kernels_portable.xcframework/macos-arm64/libkernels_portable_macos.a",
+          "-Xlinker", "cmake-out/kernels_optimized.xcframework/macos-arm64/libkernels_optimized_macos.a",
         ])
       ]
     )

extension/benchmark/apple/Benchmark/Benchmark.xcodeproj/project.pbxproj

@@ -25,7 +25,6 @@
 		03F181542D7262FC0058BDF9 /* executorch in Frameworks */ = {isa = PBXBuildFile; productRef = 03F181532D7262FC0058BDF9 /* executorch */; };
 		03F181562D7262FC0058BDF9 /* kernels_custom in Frameworks */ = {isa = PBXBuildFile; productRef = 03F181552D7262FC0058BDF9 /* kernels_custom */; };
 		03F181582D7262FC0058BDF9 /* kernels_optimized in Frameworks */ = {isa = PBXBuildFile; productRef = 03F181572D7262FC0058BDF9 /* kernels_optimized */; };
-		03F1815A2D7262FC0058BDF9 /* kernels_portable in Frameworks */ = {isa = PBXBuildFile; productRef = 03F181592D7262FC0058BDF9 /* kernels_portable */; };
 		03F1815C2D7262FC0058BDF9 /* kernels_quantized in Frameworks */ = {isa = PBXBuildFile; productRef = 03F1815B2D7262FC0058BDF9 /* kernels_quantized */; };
 		30AA4B602DC0766800B1BE50 /* pcre2_regex.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 30AA4B5A2DC0766800B1BE50 /* pcre2_regex.cpp */; };
 		30AA4B612DC0766800B1BE50 /* regex.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 30AA4B5D2DC0766800B1BE50 /* regex.cpp */; };
@@ -126,7 +125,6 @@
 				03F1814E2D7262FC0058BDF9 /* backend_coreml in Frameworks */,
 				03F181522D7262FC0058BDF9 /* backend_xnnpack in Frameworks */,
 				03F181562D7262FC0058BDF9 /* kernels_custom in Frameworks */,
-				03F1815A2D7262FC0058BDF9 /* kernels_portable in Frameworks */,
 				03F181582D7262FC0058BDF9 /* kernels_optimized in Frameworks */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
@@ -741,10 +739,6 @@
 			isa = XCSwiftPackageProductDependency;
 			productName = kernels_optimized;
 		};
-		03F181592D7262FC0058BDF9 /* kernels_portable */ = {
-			isa = XCSwiftPackageProductDependency;
-			productName = kernels_portable;
-		};
 		03F1815B2D7262FC0058BDF9 /* kernels_quantized */ = {
 			isa = XCSwiftPackageProductDependency;
 			productName = kernels_quantized;

scripts/build_apple_frameworks.sh

@@ -32,6 +32,12 @@ libextension_module.a,\
 libextension_tensor.a,\
 :$HEADERS_RELATIVE_PATH:ExecuTorch"
 
+FRAMEWORK_THREADPOOL="threadpool:\
+libcpuinfo.a,\
+libextension_threadpool.a,\
+libpthreadpool.a,\
+:"
+
 FRAMEWORK_BACKEND_COREML="backend_coreml:\
 libcoreml_util.a,\
 libcoreml_inmemoryfs.a,\
@@ -44,9 +50,6 @@ libmpsdelegate.a,\
 
 FRAMEWORK_BACKEND_XNNPACK="backend_xnnpack:\
 libXNNPACK.a,\
-libcpuinfo.a,\
-libextension_threadpool.a,\
-libpthreadpool.a,\
 libxnnpack_backend.a,\
 libmicrokernels-prod.a,\
 :"
@@ -59,6 +62,7 @@ FRAMEWORK_KERNELS_OPTIMIZED="kernels_optimized:\
 libcpublas.a,\
 liboptimized_kernels.a,\
 liboptimized_native_cpu_ops_lib.a,\
+libportable_kernels.a,\
 :"
 
 FRAMEWORK_KERNELS_PORTABLE="kernels_portable:\
@@ -230,6 +234,7 @@ for mode in "${MODES[@]}"; do
   done
 
   append_framework_flag "" "$FRAMEWORK_EXECUTORCH" "$mode"
+  append_framework_flag "" "$FRAMEWORK_THREADPOOL" "$mode"
   append_framework_flag "EXECUTORCH_BUILD_COREML" "$FRAMEWORK_BACKEND_COREML" "$mode"
   append_framework_flag "EXECUTORCH_BUILD_MPS" "$FRAMEWORK_BACKEND_MPS" "$mode"
   append_framework_flag "EXECUTORCH_BUILD_XNNPACK" "$FRAMEWORK_BACKEND_XNNPACK" "$mode"