From 64abb77269ccf1cb40311675d515a65f7a1dae38 Mon Sep 17 00:00:00 2001
From: Hans <hans.chen@bricks.tools>
Date: Mon, 7 Oct 2024 20:37:07 +0800
Subject: [PATCH] chore: tune compile options

---
 .github/workflows/release.yml                 | 14 ++++++++------
 build.rs                                      |  3 ++-
 cmake/mingw-w64-aarch64.clang.toolchain.cmake |  8 +++-----
 3 files changed, 13 insertions(+), 12 deletions(-)

diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index ecad423..9c19821 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -125,30 +125,32 @@ jobs:
           if [[ "$PLATFORM" == "windows" ]]; then
             EXTRA_CMAKE_ARGS="-DCMAKE_TOOLCHAIN_FILE=../cmake/mingw-w64-${ARCH}.clang.toolchain.cmake"
             if [[ "$ARCH" == "aarch64" ]]; then
-              EXTRA_CMAKE_ARGS="${EXTRA_CMAKE_ARGS} -DEXECUTORCH_BUILD_QNN=ON"
+              EXTRA_CMAKE_ARGS="${EXTRA_CMAKE_ARGS} -DEXECUTORCH_BUILD_QNN=ON -DQNN_SDK_ROOT=$QNN_SDK_ROOT"
             fi
           elif [[ "$PLATFORM" == "macos" ]]; then
             EXTRA_CMAKE_ARGS="-DEXECUTORCH_BUILD_COREML=ON -DCMAKE_OSX_ARCHITECTURES=${ARCH}"
           elif [[ "$PLATFORM" == "linux" ]] && [[ "$ARCH" == "aarch64" ]]; then
             EXTRA_CMAKE_ARGS="-DCMAKE_TOOLCHAIN_FILE=../cmake/aarch64-linux-gnu.gcc.toolchain.cmake"
           fi
+          if [[ "$ARCH" == "aarch64" ]]; then
+            EXTRA_CMAKE_ARGS="$EXTRA_CMAKE_ARGS -DEXECUTORCH_XNNPACK_ENABLE_KLEIDI=ON"
+          fi
           cmake \
             -S . \
             -B cmake-out \
+            -DCMAKE_BUILD_TYPE=Release \
             -DCMAKE_INSTALL_PREFIX=cmake-out \
-            -DEXECUTORCH_BUILD_XNNPACK=ON \
+            -DEXECUTORCH_ENABLE_LOGGING=ON \
             -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
             -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
             -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
             -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
             -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
             -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
-            -DQNN_SDK_ROOT=$QNN_SDK_ROOT \
+            -DEXECUTORCH_BUILD_PTHREADPOOL=ON \
             -DEXECUTORCH_BUILD_CPUINFO=ON \
             -DEXECUTORCH_BUILD_XNNPACK=ON \
-            -DEXECUTORCH_BUILD_PTHREADPOOL=ON \
-            -DEXECUTORCH_BUILD_SDK=ON \
-            -DEXECUTORCH_ENABLE_LOGGING=ON \
+            -DEXECUTORCH_XNNPACK_SHARED_WORKSPACE=ON \
             $EXTRA_CMAKE_ARGS
           cmake --build cmake-out --target install --config Release -j$(nproc)
         fi
diff --git a/build.rs b/build.rs
index 00ffea3..3f242a2 100644
--- a/build.rs
+++ b/build.rs
@@ -115,7 +115,8 @@ fn main() {
     // XNNPACK
     if link_lib(&lib_path, "xnnpack_backend", true).is_ok() {
         assert!(link_lib(&lib_path, "XNNPACK", false).is_ok());
-        let _ = link_lib(&lib_path, "microkernels-prod", false);
+        assert!(link_lib(&lib_path, "microkernels-prod", false).is_ok());
+        let _ = link_lib(&lib_path, "kleidiai", false);
     }
 
     // Vulkan
diff --git a/cmake/mingw-w64-aarch64.clang.toolchain.cmake b/cmake/mingw-w64-aarch64.clang.toolchain.cmake
index bb20bcd..94ad13c 100644
--- a/cmake/mingw-w64-aarch64.clang.toolchain.cmake
+++ b/cmake/mingw-w64-aarch64.clang.toolchain.cmake
@@ -9,9 +9,7 @@ set(CMAKE_RANLIB aarch64-w64-mingw32-ranlib)
 set(CMAKE_STRIP aarch64-w64-mingw32-strip)
 set(CMAKE_LINKER aarch64-w64-mingw32-ld)
 
-set(CMAKE_C_FLAGS "-march=armv8.4-a")
-set(CMAKE_CXX_FLAGS "-march=armv8.4-a")
+set(arch_c_flags "-march=armv8.7-a")
 
-# cache flags
-set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}" CACHE STRING "c flags")
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}" CACHE STRING "c++ flags")
+set( CMAKE_C_FLAGS_INIT   "${arch_c_flags}" )
+set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags}" )