Update on "[ET-VK] Removed shared memory usage and simplied conv2d dw op shader to improve performance."

This diff removes shared memory usage in `conv2d_dw_output_tile.glsl` shader to improve performance.
Makes sum a one dimensional array, and moves bias application before storing texel.

Differential Revision: [D75499165](https://our.internmc.facebook.com/intern/diff/D75499165/)

[ghstack-poisoned]

Author: trivedivivek

.ci/scripts/wheel/test_linux.py

@@ -14,6 +14,10 @@
             test_base.ModelTest(
                 model=Model.Mv3,
                 backend=Backend.XnnpackQuantizationDelegation,
-            )
+            ),
+            test_base.ModelTest(
+                model=Model.Mv3,
+                backend=Backend.CoreMlExportOnly,
+            ),
         ]
     )

.ci/scripts/wheel/test_macos.py

@@ -17,7 +17,7 @@
             ),
             test_base.ModelTest(
                 model=Model.Mv3,
-                backend=Backend.CoreMlTest,
+                backend=Backend.CoreMlExportAndTest,
             ),
         ]
     )

.github/workflows/apple-perf.yml

@@ -386,8 +386,7 @@ jobs:
         echo "::endgroup::"
 
         echo "::group::Build ExecuTorch iOS frameworks"
-        PYTHON_EXECUTABLE=python ${CONDA_RUN} --no-capture-output \
-          scripts/build_apple_frameworks.sh --Release --Debug --coreml --custom --mps --optimized --portable --quantized --xnnpack
+        PYTHON_EXECUTABLE=python ${CONDA_RUN} --no-capture-output scripts/build_apple_frameworks.sh
         echo "::endgroup::"
 
         # NB: Although exported models can be copied to this directory and bundled together with the

.github/workflows/apple.yml

@@ -156,6 +156,7 @@ jobs:
           "kernels_optimized"
           "kernels_portable"
           "kernels_quantized"
+          "threadpool"
         )
 
         .ci/scripts/setup-conda.sh
@@ -173,8 +174,7 @@ jobs:
         backends/apple/mps/install_requirements.sh
 
         # Build iOS Frameworks
-        PYTHON_EXECUTABLE=python ${CONDA_RUN} --no-capture-output \
-        scripts/build_apple_frameworks.sh --Release --Debug --coreml --custom --mps --optimized --portable --quantized --xnnpack
+        PYTHON_EXECUTABLE=python ${CONDA_RUN} --no-capture-output scripts/build_apple_frameworks.sh
 
         # Bundle iOS Frameworks
         for FRAMEWORK in "${FRAMEWORKS[@]}"; do (
@@ -314,8 +314,7 @@ jobs:
         echo "::endgroup::"
 
         echo "::group::Build ExecuTorch iOS frameworks"
-        PYTHON_EXECUTABLE=python ${CONDA_RUN} --no-capture-output \
-          scripts/build_apple_frameworks.sh --Release --Debug --coreml --custom --mps --optimized --portable --quantized --xnnpack
+        PYTHON_EXECUTABLE=python ${CONDA_RUN} --no-capture-output scripts/build_apple_frameworks.sh
         echo "::endgroup::"
 
         echo "::group::Build ExecuTorch benchmark app"

.lintrunner.toml

@@ -386,19 +386,14 @@ exclude_patterns = [
     "third-party/**",
     # TODO: remove exceptions as we migrate
     # backends
-    "backends/arm/quantizer/**",
-    "backends/arm/test/ops/**",
     "backends/vulkan/quantizer/**",
     "backends/vulkan/test/**",
-    "backends/qualcomm/quantizer/**",
-    "examples/qualcomm/**",
     "backends/xnnpack/quantizer/**",
     "backends/xnnpack/test/**",
     "exir/tests/test_passes.py",
     "extension/llm/export/builder.py",
     "extension/llm/export/quantizer_lib.py",
     "exir/tests/test_memory_planning.py",
-    "backends/transforms/duplicate_dynamic_quant_chain.py",
     "exir/backend/test/demos/test_xnnpack_qnnpack.py",
 ]
 

CMakeLists.txt

@@ -565,8 +565,6 @@ if(EXECUTORCH_BUILD_PTHREADPOOL AND EXECUTORCH_BUILD_CPUINFO)
 endif()
 
 if(EXECUTORCH_BUILD_PYBIND)
-  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/third-party/pybind11)
-
   if(NOT EXECUTORCH_BUILD_EXTENSION_DATA_LOADER)
     add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/data_loader)
   endif()
@@ -602,7 +600,7 @@ if(EXECUTORCH_BUILD_PYBIND)
     list(APPEND _dep_libs portable_ops_lib)
   endif()
 
-  if(EXECUTORCH_BUILD_COREML)
+  if(EXECUTORCH_BUILD_COREML AND APPLE)
     list(APPEND _dep_libs coremldelegate)
   endif()
 
@@ -701,7 +699,7 @@ if(EXECUTORCH_BUILD_EXECUTOR_RUNNER)
     list(APPEND _executor_runner_libs etdump flatccrt)
   endif()
 
-  if(EXECUTORCH_BUILD_COREML)
+  if(EXECUTORCH_BUILD_COREML AND APPLE)
     list(APPEND _executor_runner_libs coremldelegate)
   endif()
 

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",
         ])
       ]
     )