Reuse GELU implementation from PyTorch core

Pull Request resolved: #7041

kernels/optimized doesn't need to support embedded systems, so it can just take a header-only dep on PyTorch.

Note that, because we will pick up Sleef internally and ignore it
externally thanks to ATen vec, this PR gets to enable optimized GELU in OSS.

Testing: CI to make sure this doesn't break mobile build modes; happy to take advice on anything not currently covered that might break.
ghstack-source-id: 263918321
@exported-using-ghexport

Differential Revision: [D66335522](https://our.internmc.facebook.com/intern/diff/D66335522/)

Author: swolchok

.ci/scripts/build_llama_android.sh

@@ -10,6 +10,12 @@ set -exu
 # shellcheck source=/dev/null
 source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 
+if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
+  PYTHON_EXECUTABLE=python3
+fi
+which "${PYTHON_EXECUTABLE}"
+CMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')"
+
 install_executorch_and_backend_lib() {
   echo "Installing executorch and xnnpack backend"
   clean_executorch_install_folders
@@ -22,6 +28,7 @@ install_executorch_and_backend_lib() {
     -DANDROID_ABI="${ANDROID_ABI}" \
     -DCMAKE_INSTALL_PREFIX=cmake-android-out \
     -DCMAKE_BUILD_TYPE=Release \
+    -DCMAKE_PREFIX_PATH="${CMAKE_PREFIX_PATH}" \
     -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
     -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
     -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
@@ -47,6 +54,7 @@ build_llama_runner() {
     -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
     -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
     -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
+    -DCMAKE_PREFIX_PATH="${CMAKE_PREFIX_PATH}" \
     -Bcmake-android-out/examples/models/llama examples/models/llama
 
     cmake --build cmake-android-out/examples/models/llama -j4 --config Release

.ci/scripts/test_llama.sh

@@ -154,6 +154,7 @@ cmake_install_executorch_libraries() {
     rm -rf cmake-out
     retry cmake \
         -DCMAKE_INSTALL_PREFIX=cmake-out \
+        -DCMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')" \
         -DCMAKE_BUILD_TYPE="$CMAKE_BUILD_TYPE" \
         -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
         -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \

.ci/scripts/test_model.sh

@@ -50,10 +50,12 @@ prepare_artifacts_upload() {
 
 build_cmake_executor_runner() {
   echo "Building executor_runner"
+  CMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')"
   rm -rf ${CMAKE_OUTPUT_DIR}
   cmake -DCMAKE_BUILD_TYPE=Debug \
       -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
       -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
+      -DCMAKE_PREFIX_PATH="$CMAKE_PREFIX_PATH" \
       -B${CMAKE_OUTPUT_DIR} .
 
   cmake --build ${CMAKE_OUTPUT_DIR} -j4 --config Debug
@@ -98,8 +100,7 @@ test_model() {
 
 build_cmake_xnn_executor_runner() {
   echo "Building xnn_executor_runner"
-  SITE_PACKAGES="$(${PYTHON_EXECUTABLE} -c 'from distutils.sysconfig import get_python_lib; print(get_python_lib())')"
-  CMAKE_PREFIX_PATH="${SITE_PACKAGES}/torch"
+  CMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')"
 
   (rm -rf ${CMAKE_OUTPUT_DIR} \
     && mkdir ${CMAKE_OUTPUT_DIR} \

.ci/scripts/test_phi_3_mini.sh

@@ -22,8 +22,10 @@ NPROC=8
 if hash nproc &> /dev/null; then NPROC=$(nproc); fi
 
 cmake_install_executorch_libraries() {
+  CMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')"
   cmake -DPYTHON_EXECUTABLE=python \
       -DCMAKE_INSTALL_PREFIX=${BUILD_DIR} \
+      -DCMAKE_PREFIX_PATH="${CMAKE_PREFIX_PATH}" \
       -DEXECUTORCH_ENABLE_LOGGING=1 \
       -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
       -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
@@ -39,8 +41,10 @@ cmake_install_executorch_libraries() {
 }
 
 cmake_build_phi_3_mini() {
+  CMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')"
   cmake -DPYTHON_EXECUTABLE=$PYTHON_EXECUTABLE \
       -DCMAKE_INSTALL_PREFIX=${BUILD_DIR} \
+      -DCMAKE_PREFIX_PATH="${CMAKE_PREFIX_PATH}" \
       -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
       -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
       -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \

.ci/scripts/utils.sh

@@ -136,6 +136,7 @@ cmake_install_executorch_lib() {
   clean_executorch_install_folders
   retry cmake -DBUCK2="$BUCK" \
           -DCMAKE_INSTALL_PREFIX=cmake-out \
+          -DCMAKE_PREFIX_PATH="$($PYTHON_EXECUTABLE -c 'import torch as _; print(_.__path__[0])')" \
           -DCMAKE_BUILD_TYPE=Release \
           -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
           -Bcmake-out .

.github/workflows/pull.yml

@@ -147,6 +147,8 @@ jobs:
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
         conda activate "${CONDA_ENV}"
 
+        source .ci/scripts/utils.sh
+        install_executorch "use-pt-pinned-commit"
         BUILD_TOOL="cmake"
         PYTHON_EXECUTABLE=python \
         bash .ci/scripts/build_llama_android.sh  "${BUILD_TOOL}"

.github/workflows/trunk.yml

@@ -394,6 +394,7 @@ jobs:
         rm -rf cmake-out
         cmake \
             -DCMAKE_INSTALL_PREFIX=cmake-out \
+            -DCMAKE_PREFIX_PATH="$(python -c 'import torch as _; print(_.__path__[0])')" \
             -DCMAKE_BUILD_TYPE=Release \
             -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
             -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
@@ -411,6 +412,7 @@ jobs:
         cmake \
             -DCMAKE_INSTALL_PREFIX=cmake-out \
             -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_PREFIX_PATH="$(python -c 'import torch as _; print(_.__path__[0])')" \
             -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
             -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
             -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \

build/Utils.cmake

@@ -321,3 +321,20 @@ function(resolve_python_executable)
     )
   endif()
 endfunction()
+
+# find_package(Torch CONFIG REQUIRED) replacement for targets that
+# have a header-only Torch dependency. Because find_package sets
+# variables in the parent scope, we use a macro to preserve this
+# rather than maintaining our own list of those variables.
+macro(find_package_torch_headers)
+  # We cannot simply use CMAKE_FIND_ROOT_PATH_BOTH, because that does
+  # not propagate into TorchConfig.cmake.
+  foreach(mode_kind IN ITEMS PACKAGE LIBRARY INCLUDE)
+    set(OLD_CMAKE_FIND_ROOT_PATH_MODE_${mode_kind} ${CMAKE_FIND_ROOT_PATH_MODE_${mode_kind}})
+    set(CMAKE_FIND_ROOT_PATH_MODE_${mode_kind} BOTH)
+  endforeach()
+  find_package(Torch CONFIG REQUIRED)
+  foreach(mode_kind IN ITEMS PACKAGE LIBRARY INCLUDE)
+    set(CMAKE_FIND_ROOT_PATH_MODE_${mode_kind} ${OLD_CMAKE_FIND_ROOT_PATH_MODE_${mode_kind}})
+  endforeach()
+endmacro()

build/build_android_llm_demo.sh

@@ -7,6 +7,12 @@
 
 set -ex
 
+if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
+  PYTHON_EXECUTABLE=python3
+fi
+which "${PYTHON_EXECUTABLE}"
+CMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')"
+
 build_jar() {
   pushd extension/android
   ./gradlew build
@@ -36,6 +42,7 @@ build_android_native_library() {
   fi
 
   cmake . -DCMAKE_INSTALL_PREFIX="${CMAKE_OUT}" \
+    -DCMAKE_PREFIX_PATH="${CMAKE_PREFIX_PATH}" \
     -DCMAKE_TOOLCHAIN_FILE="${ANDROID_NDK}/build/cmake/android.toolchain.cmake" \
     -DANDROID_ABI="${ANDROID_ABI}" \
     -DANDROID_PLATFORM=android-26 \
@@ -69,6 +76,7 @@ build_android_native_library() {
     -DANDROID_ABI="${ANDROID_ABI}" \
     -DANDROID_PLATFORM=android-26 \
     -DCMAKE_INSTALL_PREFIX="${CMAKE_OUT}" \
+    -DCMAKE_PREFIX_PATH="${CMAKE_PREFIX_PATH}" \
     -DEXECUTORCH_ENABLE_LOGGING=ON \
     -DEXECUTORCH_LOG_LEVEL=Info \
     -DCMAKE_FIND_ROOT_PATH_MODE_PACKAGE=BOTH \

kernels/optimized/CMakeLists.txt

@@ -61,6 +61,8 @@ message("Generated files ${gen_command_sources}")
 
 list(TRANSFORM _optimized_kernels__srcs PREPEND "${EXECUTORCH_ROOT}/")
 add_library(optimized_kernels ${_optimized_kernels__srcs})
+find_package_torch_headers()
+target_include_directories(optimized_kernels PRIVATE ${TORCH_INCLUDE_DIRS})
 target_link_libraries(
   optimized_kernels PRIVATE executorch_core cpublas extension_threadpool
 )

kernels/optimized/cpu/op_gelu.cpp

@@ -13,6 +13,7 @@
 
 #include <cmath>
 
+#include <ATen/native/cpu/Gelu.h>
 #include <executorch/kernels/portable/cpu/util/activation_ops_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
@@ -47,48 +48,26 @@ void gelu(
   CTYPE* out_data = output.mutable_data_ptr<CTYPE>();
   size_t lim = input.numel();
 
-  // TODO: Add fast path for tanh using sleef's tanh
   if (approximate == "tanh") {
-    // 0.5 * x * (1 + Tanh(sqrt(2 / pi) * (x + 0.044715 * x^3))
-    for (size_t i = 0; i < lim; ++i) {
-      const CTYPE x = in_data[i];
-      const CTYPE kBeta = M_SQRT2 * M_2_SQRTPI * 0.5;
-      const CTYPE kKappa = 0.044715;
-      auto x_cube = x * x * x;
-      auto inner = kBeta * (x + kKappa * x_cube);
-      out_data[i] = CTYPE(0.5) * x * (CTYPE(1) + std::tanh(inner));
+    using Vec = at::vec::Vectorized<CTYPE>;
+    int i = 0;
+    for (; i < lim - (lim % Vec::size()); i += Vec::size()) {
+      Vec x = Vec::loadu(in_data + i);
+      at::native::vectorized_gelu_approximated_with_tanh(x).store(out_data + i);
     }
-  } else if (approximate == "none") { // dont appx
-    // GELU(x) = x * Φ(x) where Φ(x) is the is the Cumulative Distribution
-    // Function for Gaussian Distribution.
-
-#ifndef __aarch64__
-    for (size_t i = 0; i < lim; ++i) {
-      const CTYPE x = in_data[i];
-      out_data[i] = CTYPE(0.5) * x * (CTYPE(1) + std::erf(x * M_SQRT1_2));
+    for (; i < lim; ++i) {
+      out_data[i] = at::native::scalar_gelu_approximated_with_tanh(in_data[i]);
     }
-#else
-    size_t i = 0;
-    if constexpr (std::is_same_v<CTYPE, float>) {
-      for (; i + 4 < lim; i += 4) {
-        const float32x4_t in =
-            vld1q_f32(static_cast<const float*>(&in_data[i]));
-        const float32x4_t m_sqrt1_2x4 = {
-            M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, M_SQRT1_2};
-        const float32x4_t ones = vmovq_n_f32(1.0);
-        const float32x4_t halves = vmovq_n_f32(0.5);
-        float32x4_t out = Sleef_erff4_u10(vmulq_f32(in, m_sqrt1_2x4));
-        vst1q_f32(
-            static_cast<float*>(&out_data[i]),
-            vmulq_f32(vmulq_f32(vaddq_f32(out, ones), in), halves));
-      }
+  } else if (approximate == "none") {
+    using Vec = at::vec::Vectorized<CTYPE>;
+    int i = 0;
+    for (; i < lim - (lim % Vec::size()); i += Vec::size()) {
+      Vec x = Vec::loadu(in_data + i);
+      at::native::vectorized_gelu(x).store(out_data + i);
     }
     for (; i < lim; ++i) {
-      const CTYPE x = in_data[i];
-      out_data[i] = CTYPE(0.5) * x * (CTYPE(1) + std::erf(x * M_SQRT1_2));
+      out_data[i] = at::native::scalar_gelu(in_data[i]);
     }
-#endif // __aarch64__
-
   } else {
     ET_KERNEL_CHECK_MSG(
         context,

kernels/optimized/cpu/targets.bzl

@@ -28,13 +28,9 @@ _OPTIMIZED_ATEN_OPS = (
     op_target(name = "op_sigmoid"),
     op_target(
         name = "op_gelu",
-        deps = select({
-            "DEFAULT": [],
-            "ovr_config//cpu:arm64": [
-                "fbsource//third-party/sleef:sleef_arm",
-            ],
-        }) + [
+        deps = [
             "//executorch/kernels/portable/cpu/util:activation_ops_util",
+            "//executorch/runtime/core/portable_type/c10:aten_headers_for_executorch",
         ],
     ),
     op_target(
@@ -96,6 +92,13 @@ _OPTIMIZED_ATEN_OPS = (
     ),
 )
 
+
+def get_sleef_preprocessor_flags():
+    if runtime.is_oss:
+        return []
+    return ["-DAT_BUILD_ARM_VEC256_WITH_SLEEF"]
+
+
 def define_common_targets():
     """Defines targets that should be shared between fbcode and xplat.
 

kernels/optimized/optimized-oss.yaml

@@ -1,8 +1,8 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 #
 # This yaml file contains operators that have optimized kernels available.
-# Note that this is a copy of optimized.yaml that does not include gelu and
-# log_softmax, due to the OSS build not currently including sleef.
+# Note that this is a copy of optimized.yaml that does not include log_softmax,
+# due to the OSS build not currently including sleef.
 # TODO (T183193812)
 
 - op: add.out
@@ -40,6 +40,11 @@
     - arg_meta: null
       kernel_name: torch::executor::opt_sigmoid_out
 
+- op: gelu.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::opt_gelu_out
+
 - op: le.Scalar_out
   kernels:
     - arg_meta: null

shim/xplat/executorch/kernels/optimized/op_registration_util.bzl

@@ -134,5 +134,5 @@ def define_op_target(name, deps):
 
 def is_op_disabled(name):
     # TODO (gjcomer) Enable ops with sleef dependency in OSS
-    disabled_ops = ["op_gelu", "op_log_softmax"]
+    disabled_ops = ["op_log_softmax"]
     return name in disabled_ops

test/run_oss_cpp_tests.sh

@@ -22,13 +22,20 @@ elif [[ $(uname) == "Linux" ]]; then
   export LLVM_COV="${LLVM_COV:-llvm-cov}"
 fi
 
+if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
+  PYTHON_EXECUTABLE=python3
+fi
+which "${PYTHON_EXECUTABLE}"
+
 build_executorch() {
   BUILD_VULKAN="OFF"
   if [ -x "$(command -v glslc)" ]; then
     BUILD_VULKAN="ON"
   fi
+  CMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')"
   cmake . \
     -DCMAKE_INSTALL_PREFIX=cmake-out \
+    -DCMAKE_PREFIX_PATH="${CMAKE_PREFIX_PATH}" \
     -DEXECUTORCH_USE_CPP_CODE_COVERAGE=ON \
     -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
     -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \