Update on "[ET-VK] Replacing use of adaptive_work_group_size function…

… by create_local_wg_size function."

This diff replaces the use of the adaptive_work_group_size function with create_local_wg_size function, which is better tuned for improving shader performance.

Differential Revision: [D66308779](https://our.internmc.facebook.com/intern/diff/D66308779/)

[ghstack-poisoned]

.ci/scripts/gather_test_models.py

@@ -20,16 +20,16 @@
 CUSTOM_RUNNERS = {
     "linux": {
         # This one runs OOM on smaller runner, the root cause is unclear (T163016365)
-        "w2l": "linux.12xlarge",
-        "ic4": "linux.12xlarge",
-        "resnet50": "linux.12xlarge",
-        "llava": "linux.12xlarge",
-        "llama3_2_vision_encoder": "linux.12xlarge",
-        # "llama3_2_text_decoder": "linux.12xlarge",  # TODO: re-enable test when Huy's change is in / model gets smaller.
+        "w2l": "linux.4xlarge.memory",
+        "ic4": "linux.4xlarge.memory",
+        "resnet50": "linux.4xlarge.memory",
+        "llava": "linux.4xlarge.memory",
+        "llama3_2_vision_encoder": "linux.4xlarge.memory",
+        "llama3_2_text_decoder": "linux.4xlarge.memory",
         # This one causes timeout on smaller runner, the root cause is unclear (T161064121)
-        "dl3": "linux.12xlarge",
-        "emformer_join": "linux.12xlarge",
-        "emformer_predict": "linux.12xlarge",
+        "dl3": "linux.4xlarge.memory",
+        "emformer_join": "linux.4xlarge.memory",
+        "emformer_predict": "linux.4xlarge.memory",
     }
 }
 
@@ -39,10 +39,12 @@
     "linux": {
         "mobilebert": 90,
         "emformer_predict": 360,
+        "llama3_2_text_decoder": 360,
     },
     "macos": {
         "mobilebert": 90,
         "emformer_predict": 360,
+        "llama3_2_text_decoder": 360,
     },
 }
 

.ci/scripts/setup-macos.sh

@@ -49,6 +49,9 @@ install_buck() {
 
   rm "${BUCK2}"
   popd
+
+  # Kill all running buck2 daemon for a fresh start
+  buck2 killall || true
 }
 
 function write_sccache_stub() {

.ci/scripts/test_llama.sh

@@ -51,6 +51,9 @@ UPLOAD_DIR="${UPLOAD_DIR:-}"
 # Default PT2E_QUANTIZE to empty string if not set
 PT2E_QUANTIZE="${PT2E_QUANTIZE:-}"
 
+# Default CMake Build Type to release mode
+CMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release}
+
 if [[ $# -lt 4 ]]; then # Assuming 4 mandatory args
     echo "Expecting atleast 4 positional arguments"
     echo "Usage: [...]"
@@ -143,7 +146,7 @@ cmake_install_executorch_libraries() {
     rm -rf cmake-out
     retry cmake \
         -DCMAKE_INSTALL_PREFIX=cmake-out \
-        -DCMAKE_BUILD_TYPE=Debug \
+        -DCMAKE_BUILD_TYPE="$CMAKE_BUILD_TYPE" \
         -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
         -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
         -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
@@ -157,22 +160,22 @@ cmake_install_executorch_libraries() {
         -DQNN_SDK_ROOT="$QNN_SDK_ROOT" \
         -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
         -Bcmake-out .
-    cmake --build cmake-out -j9 --target install --config Debug
+    cmake --build cmake-out -j9 --target install --config "$CMAKE_BUILD_TYPE"
 }
 
 cmake_build_llama_runner() {
     echo "Building llama runner"
     dir="examples/models/llama"
     retry cmake \
         -DCMAKE_INSTALL_PREFIX=cmake-out \
-        -DCMAKE_BUILD_TYPE=Debug \
+        -DCMAKE_BUILD_TYPE="$CMAKE_BUILD_TYPE" \
         -DEXECUTORCH_BUILD_KERNELS_CUSTOM="$CUSTOM" \
         -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
         -DEXECUTORCH_BUILD_XNNPACK="$XNNPACK" \
         -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
         -Bcmake-out/${dir} \
         ${dir}
-    cmake --build cmake-out/${dir} -j9 --config Debug
+    cmake --build cmake-out/${dir} -j9 --config "$CMAKE_BUILD_TYPE"
 
 }
 

.ci/scripts/test_llava.sh

@@ -8,11 +8,11 @@
 set -exu
 # shellcheck source=/dev/null
 
-BUILD_TYPE=${1:-Debug}
 TARGET_OS=${2:-Native}
 BUILD_DIR=${3:-cmake-out}
+CMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release}
 
-echo "Building with BUILD_TYPE: $BUILD_TYPE, TARGET_OS: $TARGET_OS, BUILD_DIR: $BUILD_DIR"
+echo "Building with CMAKE_BUILD_TYPE: $CMAKE_BUILD_TYPE, TARGET_OS: $TARGET_OS, BUILD_DIR: $BUILD_DIR"
 
 if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
     PYTHON_EXECUTABLE=python3
@@ -32,7 +32,7 @@ if hash nproc &> /dev/null; then NPROC=$(nproc); fi
 
 EXECUTORCH_COMMON_CMAKE_ARGS="                      \
         -DCMAKE_INSTALL_PREFIX=${BUILD_DIR}         \
-        -DCMAKE_BUILD_TYPE=${BUILD_TYPE}            \
+        -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}            \
         -DEXECUTORCH_ENABLE_LOGGING=ON              \
         -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON      \
         -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
@@ -49,7 +49,7 @@ cmake_install_executorch_libraries() {
         ${EXECUTORCH_COMMON_CMAKE_ARGS} \
         -B${BUILD_DIR} .
 
-    cmake --build ${BUILD_DIR} -j${NPROC} --target install --config ${BUILD_TYPE}
+    cmake --build ${BUILD_DIR} -j${NPROC} --target install --config ${CMAKE_BUILD_TYPE}
 }
 
 cmake_install_executorch_libraries_for_android() {
@@ -59,14 +59,14 @@ cmake_install_executorch_libraries_for_android() {
         ${EXECUTORCH_COMMON_CMAKE_ARGS}                                         \
         -B${BUILD_DIR} .
 
-    cmake --build ${BUILD_DIR} -j${NPROC} --target install --config ${BUILD_TYPE}
+    cmake --build ${BUILD_DIR} -j${NPROC} --target install --config ${CMAKE_BUILD_TYPE}
 }
 
 
 LLAVA_COMMON_CMAKE_ARGS="                        \
         -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
         -DCMAKE_INSTALL_PREFIX=${BUILD_DIR}      \
-        -DCMAKE_BUILD_TYPE=${BUILD_TYPE}         \
+        -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}         \
         -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON     \
         -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON  \
         -DEXECUTORCH_BUILD_XNNPACK=ON"
@@ -81,7 +81,7 @@ cmake_build_llava_runner() {
         -B${BUILD_DIR}/${dir}             \
         ${dir}
 
-    cmake --build ${BUILD_DIR}/${dir} -j${NPROC} --config ${BUILD_TYPE}
+    cmake --build ${BUILD_DIR}/${dir} -j${NPROC} --config ${CMAKE_BUILD_TYPE}
 }
 
 
@@ -98,7 +98,7 @@ cmake_build_llava_runner_for_android() {
         -B${BUILD_DIR}/${dir}                                                   \
         ${dir}
 
-    cmake --build ${BUILD_DIR}/${dir} -j${NPROC} --config ${BUILD_TYPE}
+    cmake --build ${BUILD_DIR}/${dir} -j${NPROC} --config ${CMAKE_BUILD_TYPE}
 }
 
 # only export the one without custom op for now since it's

.github/workflows/apple.yml

@@ -42,6 +42,8 @@ jobs:
 
   build-demo-ios:
     name: build-demo-ios
+    # NB: Don't run this on fork PRs because they won't have access to the secret and would fail anyway
+    if: ${{ !github.event.pull_request.head.repo.fork }}
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     secrets: inherit
     with:
@@ -190,6 +192,8 @@ jobs:
         ) done
 
   upload-frameworks-ios:
+    # NB: Don't run this on fork PRs because they won't have access to the secret and would fail anyway
+    if: ${{ !github.event.pull_request.head.repo.fork }}
     runs-on: ubuntu-22.04
     needs: [build-frameworks-ios, set-version]
     timeout-minutes: 30
@@ -278,6 +282,8 @@ jobs:
 
   build-benchmark-app:
     name: build-benchmark-app
+    # NB: Don't run this on fork PRs because they won't have access to the secret and would fail anyway
+    if: ${{ !github.event.pull_request.head.repo.fork }}
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     secrets: inherit
     with:

.github/workflows/ghstack_land.yml

@@ -3,21 +3,7 @@ on:
   pull_request:
     types: [closed]
     branches:
-      - 'gh/cccclai/[0-9]+/base'
-      - 'gh/dbort/[0-9]+/base'
-      - 'gh/dvorjackz/[0-9]+/base'
-      - 'gh/guangy10/[0-9]+/base'
-      - 'gh/helunwencser/[0-9]+/base'
-      - 'gh/jorgep31415/[0-9]+/base'
-      - 'gh/kimishpatel/[0-9]+/base'
-      - 'gh/kirklandsign/[0-9]+/base'
-      - 'gh/larryliu0820/[0-9]+/base'
-      - 'gh/lucylq/[0-9]+/base'
-      - 'gh/manuelcandales/[0-9]+/base'
-      - 'gh/mcr229/[0-9]+/base'
-      - 'gh/swolchok/[0-9]+/base'
-      - 'gh/SS-JIA/[0-9]+/base'
-      - 'gh/trivedivivek/[0-9]+/base'
+      - 'gh/*/[0-9]+/base'
 
 jobs:
   ghstack_merge_to_main:

.github/workflows/pull.yml

@@ -332,7 +332,7 @@ jobs:
       docker-image: executorch-ubuntu-22.04-clang12
 
   unittest-arm:
-    uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     with:
       runner: linux.2xlarge
       docker-image: executorch-ubuntu-22.04-arm-sdk

.github/workflows/trunk.yml

@@ -131,7 +131,7 @@ jobs:
 
   test-arm-backend-delegation:
     name: test-arm-backend-delegation
-    uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     with:
       runner: linux.2xlarge
       docker-image: executorch-ubuntu-22.04-arm-sdk
@@ -157,7 +157,7 @@ jobs:
 
   test-arm-reference-delegation:
     name: test-arm-reference-delegation
-    uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     with:
       runner: linux.2xlarge
       docker-image: executorch-ubuntu-22.04-arm-sdk
@@ -290,7 +290,7 @@ jobs:
   #       ${CONDA_RUN} python -m unittest examples.models.llava.test.test_llava
 
   #       # run e2e (export, tokenizer and runner)
-  #       PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/test_llava.sh Release
+  #       PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/test_llava.sh
 
   test-qnn-model:
     name: test-qnn-model
@@ -351,6 +351,8 @@ jobs:
         done
 
   test-huggingface-transformers:
+    # NB: Don't run this on fork PRs because they won't have access to the secret and would fail anyway
+    if: ${{ !github.event.pull_request.head.repo.fork }}
     name: test-huggingface-transformers
     uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
     secrets: inherit

CMakeLists.txt

@@ -682,6 +682,22 @@ if(EXECUTORCH_BUILD_PTHREADPOOL
 endif()
 
 if(EXECUTORCH_BUILD_PYBIND)
+  # Setup RPATH.
+  # See https://gitlab.kitware.com/cmake/community/-/wikis/doc/cmake/RPATH-handling
+  if(APPLE)
+    set(CMAKE_MACOSX_RPATH ON)
+    set(_rpath_portable_origin "@loader_path")
+  else()
+    set(_rpath_portable_origin $ORIGIN)
+  endif(APPLE)
+  # Use separate rpaths during build and install phases
+  set(CMAKE_SKIP_BUILD_RPATH  FALSE)
+  # Don't use the install-rpath during the build phase
+  set(CMAKE_BUILD_WITH_INSTALL_RPATH FALSE)
+  set(CMAKE_INSTALL_RPATH "${_rpath_portable_origin}")
+  # Automatically add all linked folders that are NOT in the build directory to
+  # the rpath (per library?)
+  set(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/third-party/pybind11)
 
   if(NOT EXECUTORCH_BUILD_EXTENSION_DATA_LOADER)
@@ -765,46 +781,6 @@ if(EXECUTORCH_BUILD_PYBIND)
   target_include_directories(portable_lib PRIVATE ${TORCH_INCLUDE_DIRS})
   target_compile_options(portable_lib PUBLIC ${_pybind_compile_options})
   target_link_libraries(portable_lib PRIVATE ${_dep_libs})
-  if(APPLE)
-    # pip wheels will need to be able to find the torch libraries. On Linux, the
-    # .so has non-absolute dependencies on libs like "libtorch.so" without
-    # paths; as long as we `import torch` first, those dependencies will work.
-    # But Apple dylibs do not support non-absolute dependencies, so we need to
-    # tell the loader where to look for its libraries. The LC_LOAD_DYLIB entries
-    # for the torch libraries will look like "@rpath/libtorch.dylib", so we can
-    # add an LC_RPATH entry to look in a directory relative to the installed
-    # location of our _portable_lib.so file. To see these LC_* values, run
-    # `otool -l _portable_lib*.so`.
-    set_target_properties(
-      portable_lib
-      PROPERTIES # Assume that this library will be installed in
-                 # `site-packages/executorch/extension/pybindings`, and that
-                 # the torch libs are in `site-packages/torch/lib`.
-                 BUILD_RPATH "@loader_path/../../../torch/lib"
-                 INSTALL_RPATH "@loader_path/../../../torch/lib"
-                 # Assume <executorch> is the root `site-packages/executorch`
-                 # Need to add <executorch>/extension/llm/custom_ops for
-                 # libcustom_ops_aot_lib.dylib
-                 BUILD_RPATH "@loader_path/../../extension/llm/custom_ops"
-                 INSTALL_RPATH "@loader_path/../../extension/llm/custom_ops"
-                 # Need to add <executorch>/kernels/quantized for
-                 # libquantized_ops_aot_lib.dylib
-                 BUILD_RPATH "@loader_path/../../kernels/quantized"
-                 INSTALL_RPATH "@loader_path/../../kernels/quantized"
-    )
-  else()
-    set_target_properties(
-      portable_lib
-      PROPERTIES
-        # Assume <executorch> is the root `site-packages/executorch`
-        # Need to add <executorch>/extension/llm/custom_ops for
-        # libcustom_ops_aot_lib
-        # Need to add <executorch>/kernels/quantized for
-        # libquantized_ops_aot_lib
-        BUILD_RPATH
-        "$ORIGIN:$ORIGIN/../../extension/llm/custom_ops:$ORIGIN/../../kernels/quantized"
-    )
-  endif()
 
   install(TARGETS portable_lib
           LIBRARY DESTINATION executorch/extension/pybindings

backends/apple/coreml/runtime/test/ETCoreMLModelDebuggerTests.mm

@@ -151,7 +151,6 @@ - (void)testMV3ProgramDebugging {
     XCTAssertNotNil(debuggingResults[make_path_with_output_name("aten__native_batch_norm_legit_no_training_default_13_cast_fp16")]);
     XCTAssertNotNil(debuggingResults[make_path_with_output_name("_inversed_aten_div_tensor_24_cast_fp16")]);
     XCTAssertNotNil(debuggingResults[make_path_with_output_name("aten_mean_dim_7_cast_fp16")]);
-    XCTAssertNotNil(debuggingResults[make_path_with_output_name("aten_clamp_default_54_cast_fp16")]);
     XCTAssertNotNil(debuggingResults[make_path_with_output_name("aten__native_batch_norm_legit_no_training_default_22_cast_fp16")]);
     XCTAssertNotNil(debuggingResults[make_path_with_output_name("aten_mul_tensor_27_cast_fp16")]);
 }

backends/apple/coreml/runtime/test/ETCoreMLModelProfilerTests.mm

@@ -146,7 +146,6 @@ - (void)testMV3ProgramProfiling {
             XCTAssertNotNil(profilingResult[make_path_with_output_name("aten__native_batch_norm_legit_no_training_default_13_cast_fp16")]);
             XCTAssertNotNil(profilingResult[make_path_with_output_name("_inversed_aten_div_tensor_24_cast_fp16")]);
             XCTAssertNotNil(profilingResult[make_path_with_output_name("aten_mean_dim_7_cast_fp16")]);
-            XCTAssertNotNil(profilingResult[make_path_with_output_name("aten_clamp_default_54_cast_fp16")]);
             XCTAssertNotNil(profilingResult[make_path_with_output_name("aten__native_batch_norm_legit_no_training_default_22_cast_fp16")]);
             XCTAssertNotNil(profilingResult[make_path_with_output_name("aten_mul_tensor_27_cast_fp16")]);
         };

backends/arm/TARGETS

@@ -110,3 +110,14 @@ python_library(
         "//executorch/backends/arm/operators:node_visitor",
     ],
 )
+
+python_library(
+    name = "arm_model_evaluator",
+    src = [
+        "util/arm_model_evaluator.py",
+    ],
+    typing = True,
+    deps = [
+        "//caffe2:torch",
+    ]
+)

backends/arm/_passes/arm_pass_manager.py

@@ -29,8 +29,8 @@
     DecomposeSoftmaxesPass,
 )
 from executorch.backends.arm._passes.decompose_var_pass import DecomposeVarPass
-from executorch.backends.arm._passes.insert_squeeze_after_sum_pass import (
-    InsertSqueezeAfterSumPass,
+from executorch.backends.arm._passes.keep_dims_false_to_squeeze_pass import (
+    KeepDimsFalseToSqueezePass,
 )
 from executorch.backends.arm._passes.match_arg_ranks_pass import MatchArgRanksPass
 from executorch.backends.arm._passes.meandim_to_averagepool_pass import (
@@ -71,7 +71,7 @@ def transform_to_backend_pipeline(
         self.add_pass(DecomposeMeanDimPass())
         self.add_pass(MatchArgRanksPass(exported_program))
         self.add_pass(DecomposeDivPass())
-        self.add_pass(InsertSqueezeAfterSumPass())
+        self.add_pass(KeepDimsFalseToSqueezePass())
         self.add_pass(ConvertSplitToSlicePass())
         self.add_pass(Conv1dUnsqueezePass(exported_program))
         self.add_pass(DecomposeSoftmaxesPass())