microsoft · hariharans29 · Jul 25, 2025 · Jun 5, 2025 · Jun 6, 2025 · Jun 11, 2025
diff --git a/cmake/CMakeLists.txt b/cmake/CMakeLists.txt
@@ -83,6 +83,11 @@ option(onnxruntime_USE_RKNPU "Build with RKNPU support" OFF)
 option(onnxruntime_USE_DNNL "Build with DNNL support" OFF)
 option(onnxruntime_USE_JSEP "Build with JavaScript implemented kernels support" OFF)
 option(onnxruntime_USE_KLEIDIAI "Build with KleidiAI integration in MLAS" OFF)
+# iOS simulator build explicitly builds targets with USE_KLEIDIAI=ON so attempting to force override if so
+if(APPLE AND CMAKE_OSX_ARCHITECTURES MATCHES "x86_64")
+    message(WARNING "Disabling KleidiAI: not supported on Apple x86_64 platforms")
+  set(onnxruntime_USE_KLEIDIAI OFF CACHE BOOL "" FORCE)
+endif()
 option(onnxruntime_BUILD_UNIT_TESTS "Build ONNXRuntime unit tests" ON)
 option(onnxruntime_BUILD_CSHARP "Build C# library" OFF)
 option(onnxruntime_BUILD_OBJC "Build Objective-C library" OFF)
@@ -275,8 +280,6 @@ if (onnxruntime_ENABLE_TRAINING_APIS)
 endif()
 
 
-
-
 # Single output director for all binaries
 set(RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin CACHE PATH "Single output directory for all binaries.")
 
@@ -648,17 +651,25 @@ else()
   endif()
 endif()
 
-if (onnxruntime_USE_KLEIDIAI AND NOT MSVC AND (
-    (onnxruntime_target_platform STREQUAL "aarch64") OR
-    (onnxruntime_target_platform STREQUAL "ARM64") OR
-    (APPLE AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")))
-  check_cxx_compiler_flag(-march=armv8.2-a+dotprod HAS_ARM64_DOTPROD)
-  check_cxx_compiler_flag(-march=armv8.2-a+i8mm HAS_ARM64_I8MM)
-  if (NOT HAS_ARM64_DOTPROD)
-    message(FATAL_ERROR  "The compiler doesn't support dotprod")
-  endif()
-  if (NOT HAS_ARM64_I8MM)
-    message(FATAL_ERROR  "The compiler doesn't support i8mm")
+if (onnxruntime_USE_KLEIDIAI AND (
+      (onnxruntime_target_platform STREQUAL "aarch64") OR
+      (onnxruntime_target_platform STREQUAL "ARM64") OR
+      (APPLE AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")))
+
+  # TODO Add checks for MSVC Compilation
+  if(NOT MSVC)
+    check_cxx_compiler_flag(-march=armv8.2-a+dotprod HAS_ARM64_DOTPROD)
+    check_cxx_compiler_flag(-march=armv8.2-a+i8mm HAS_ARM64_I8MM)
+    if (NOT HAS_ARM64_DOTPROD)
+      message(FATAL_ERROR "The compiler doesn't support dotprod")
+    endif()
+    if (NOT HAS_ARM64_I8MM)
+      message(FATAL_ERROR "The compiler doesn't support i8mm")
+    endif()
+  else()
+    message(STATUS "Skipping -march= checks on MSVC (not supported), assuming dotprod/i8mm support manually.")
+    set(HAS_ARM64_DOTPROD TRUE)
+    set(HAS_ARM64_I8MM TRUE)
   endif()
 endif()
 
@@ -1008,6 +1019,10 @@ function(onnxruntime_set_compile_flags target_name)
     if (onnxruntime_ENABLE_ATEN)
       target_compile_definitions(${target_name} PRIVATE ENABLE_ATEN)
     endif()
+    # TODO: Narrow scope for Kleidiai compile
+    if (onnxruntime_USE_KLEIDIAI)
+      target_compile_definitions(${target_name} PRIVATE USE_KLEIDIAI)
+    endif()
 
     set_target_properties(${target_name} PROPERTIES COMPILE_WARNING_AS_ERROR ON)
     if (onnxruntime_USE_CUDA)

diff --git a/cmake/deps.txt b/cmake/deps.txt
@@ -56,5 +56,5 @@ extensions;https://github.com/microsoft/onnxruntime-extensions/archive/c24b7bab0
 directx_headers;https://github.com/microsoft/DirectX-Headers/archive/refs/tags/v1.613.1.zip;47653509a3371eabb156360f42faf582f314bf2e
 cudnn_frontend;https://github.com/NVIDIA/cudnn-frontend/archive/refs/tags/v1.12.0.zip;7e733cfdc410d777b76122d64232499205589a96
 dawn;https://github.com/google/dawn/archive/9733be39e18186961d503e064874afe3e9ceb8d1.zip;2a4017c32892b90d072a9102eba90ae691fae36d
-kleidiai;https://github.com/ARM-software/kleidiai/archive/refs/tags/v1.4.0.tar.gz;22d3b57b54a61c194ab256ff11b0353a3b220244
+kleidiai;https://github.com/ARM-software/kleidiai/archive/refs/tags/v1.9.0.tar.gz;a2765979f64efb173a4b8ba4de39dcba9c655786
 duktape;https://github.com/svaarala/duktape/releases/download/v2.7.0/duktape-2.7.0.tar.xz;8200c8e417dbab7adcc12c4dbdef7651cfc55794
diff --git a/cmake/external/onnxruntime_external_deps.cmake b/cmake/external/onnxruntime_external_deps.cmake
@@ -1,3 +1,6 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+
 message(STATUS "Loading Dependencies URLs ...")
 
 include(external/helper_functions.cmake)
@@ -819,6 +822,14 @@ if(onnxruntime_USE_COREML)
 
 endif()
 
+if(onnxruntime_USE_KLEIDIAI)
+  # Disable the KleidiAI tests
+  set(KLEIDIAI_BUILD_TESTS  OFF)
+
+  onnxruntime_fetchcontent_declare(kleidiai URL ${DEP_URL_kleidiai} URL_HASH SHA1=${DEP_SHA1_kleidiai} EXCLUDE_FROM_ALL)
+  onnxruntime_fetchcontent_makeavailable(kleidiai)
+endif()
+
 set(onnxruntime_LINK_DIRS)
 if (onnxruntime_USE_CUDA)
   find_package(CUDAToolkit REQUIRED)

diff --git a/cmake/onnxruntime_mlas.cmake b/cmake/onnxruntime_mlas.cmake
@@ -267,24 +267,23 @@ function(setup_mlas_source_for_windows)
 endfunction()
 
 function(setup_kleidiai)
-  target_compile_definitions(onnxruntime_mlas PRIVATE USE_KLEIDIAI)
-
-  # Disable the KleidiAI tests
-  set(KLEIDIAI_BUILD_TESTS  OFF)
-
-  # Fetch KleidiAI sources:
-  if (NOT TARGET kleidiai)
-    onnxruntime_fetchcontent_declare(kleidiai URL ${DEP_URL_kleidiai} URL_HASH SHA1=${DEP_SHA1_kleidiai} EXCLUDE_FROM_ALL)
-  endif()
-  onnxruntime_fetchcontent_makeavailable(kleidiai)
-
   target_sources(onnxruntime_mlas PRIVATE
     ${MLAS_SRC_DIR}/kai_ukernel_interface.cpp
+    ${MLAS_SRC_DIR}/kleidiai/sgemm_kleidiai.cpp
+    ${MLAS_SRC_DIR}/kleidiai/convolve_kleidiai.cpp
+    ${MLAS_SRC_DIR}/kleidiai/qgemm_kleidiai.cpp
   )
   target_link_libraries(onnxruntime_mlas PRIVATE kleidiai)
-
   list(APPEND onnxruntime_EXTERNAL_LIBRARIES kleidiai)
   set(onnxruntime_EXTERNAL_LIBRARIES ${onnxruntime_EXTERNAL_LIBRARIES} PARENT_SCOPE)
+
+  if (NOT onnxruntime_BUILD_SHARED_LIB)
+    install(TARGETS kleidiai EXPORT ${PROJECT_NAME}Targets
+    ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
+    LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
+    RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
+    FRAMEWORK DESTINATION ${CMAKE_INSTALL_BINDIR})
+  endif()
 endfunction()
 
 if (CMAKE_SYSTEM_NAME STREQUAL "Emscripten")
@@ -311,7 +310,6 @@ if (CMAKE_SYSTEM_NAME STREQUAL "Emscripten")
 elseif(MSVC)
   setup_mlas_source_for_windows()
 else()
-
     if(APPLE)
         get_target_property(ONNXRUNTIME_MLAS_OSX_ARCH onnxruntime_mlas OSX_ARCHITECTURES)
 

diff --git a/onnxruntime/contrib_ops/cpu/bert/attention.cc b/onnxruntime/contrib_ops/cpu/bert/attention.cc
@@ -71,7 +71,7 @@ bool Attention<T>::IsPackWeightsSuccessful(int qkv_index,
                                            const T* weights_data,
                                            size_t weight_matrix_col_size,
                                            /*out*/ PrePackedWeights* prepacked_weights) {
-  size_t packb_size = MlasGemmPackBSize(head_size, input_hidden_size);
+  size_t packb_size = MlasGemmPackBSize(CblasNoTrans, CblasNoTrans, head_size, input_hidden_size);
   if (packb_size == 0) {
     return false;
   }
@@ -87,7 +87,7 @@ bool Attention<T>::IsPackWeightsSuccessful(int qkv_index,
   memset(packed_weights_data, 0, packed_weights_data_size);
 
   for (size_t i = 0; i < loop_len; i++) {
-    MlasGemmPackB(CblasNoTrans, head_size, input_hidden_size, weights_data, weight_matrix_col_size, packed_weights_data);
+    MlasGemmPackB(CblasNoTrans, CblasNoTrans, head_size, input_hidden_size, weights_data, weight_matrix_col_size, packed_weights_data);
     packed_weights_data += packb_size;
     weights_data += head_size;
   }