Aarch64 paged attention enablement

cmake/developer_package/compile_flags/os_flags.cmake

@@ -4,6 +4,8 @@
 
 include(ProcessorCount)
 include(CheckCXXCompilerFlag)
+include(CheckCSourceCompiles)
+include(CheckCXXSourceCompiles)
 
 #
 # ov_disable_deprecated_warnings()
@@ -91,6 +93,52 @@ macro(ov_dev_package_no_errors)
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${ov_c_cxx_dev_no_errors}")
 endmacro()
 
+#
+# ov_check_compiler_supports_sve(lang flags)
+#
+# Checks whether compiler for passed language supports SVE code compilation
+#
+
+macro(ov_check_compiler_supports_sve flags)
+    # Code to compile
+    set(SVE_CODE "
+    #include <arm_sve.h>
+    int main() {
+        svfloat64_t a;
+        a = svdup_n_f64(0);
+        (void)a; // to avoid warnings
+        return 0;
+    }")
+
+    # Save the current state of required flags
+    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+
+    # Set the flags necessary for compiling the test code with SVE support
+    set(CMAKE_REQUIRED_FLAGS "${CMAKE_CXX_FLAGS_INIT} ${flags}")
+
+    # Check if the source code compiles with the given flags for the specified language (C or C++)
+    CHECK_CXX_SOURCE_COMPILES("${SVE_CODE}" CXX_HAS_SVE)            
+
+    # If the compilation test is successful, set appropriate variables indicating support
+    if(CXX_HAS_SVE)
+        set(CXX_SVE_FOUND TRUE CACHE BOOL "SVE available on host")
+        set(CXX_SVE_FOUND TRUE CACHE BOOL "CXX SVE support")
+        set(CXX_SVE_FLAGS "${flags}" CACHE STRING "CXX SVE flags")
+    endif()
+
+    # Restore the original state of required flags
+    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
+
+    # If the compilation test fails, indicate that the support is not found
+    if(NOT CXX_SVE_FOUND)
+        set(CXX_SVE_FOUND FALSE CACHE BOOL "CXX SVE support")
+        set(CXX_SVE_FLAGS "" CACHE STRING "CXX SVE flags")
+    endif()
+
+    # Mark the variables as advanced to hide them in the default CMake GUI
+    mark_as_advanced(CXX_SVE_FOUND CXX_SVE_FLAGS)
+endmacro()
+
 #
 # ov_sse42_optimization_flags(<output flags>)
 #
@@ -208,6 +256,49 @@ macro(ov_arm_neon_fp16_optimization_flags flags)
     endif()
 endmacro()
 
+#
+# ov_arm_sve_optimization_flags(<output flags>)
+#
+macro(ov_arm_sve_optimization_flags flags)
+    # Check for compiler SVE support
+    ov_check_compiler_supports_sve("-march=armv8-a+sve")
+
+    if(OV_COMPILER_IS_INTEL_LLVM)
+        message(WARNING "Unsupported CXX compiler ${CMAKE_CXX_COMPILER_ID}")
+    elseif(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        # nothing should be required here
+    elseif(ANDROID)
+        if(ANDROID_ABI STREQUAL "arm64-v8a")
+            set(${flags} -Wno-unused-command-line-argument)
+            if(CXX_SVE_FOUND)
+                list(APPEND ${flags} -march=armv8-a+sve)
+            else()
+                message(WARNING "SVE is not supported on this Android ABI: ${ANDROID_ABI}")
+            endif()
+        else()
+            message(WARNING "SVE is not supported on this Android ABI: ${ANDROID_ABI}")
+        endif()
+    else()
+        if(AARCH64)
+            set(${flags} -O2)
+
+            # Add flag for SVE if supported
+            if(CXX_SVE_FOUND)
+                list(APPEND ${flags} -march=armv8-a+sve)
+            endif()
+            if(NOT CMAKE_CL_64)
+                list(APPEND ${flags} -ftree-vectorize)
+            endif()
+
+            set(${flags} ${${flags}})
+        elseif(ARM)
+            message(WARNING "SVE is not supported on 32-bit ARM architectures.")
+        else()
+            message(WARNING "SVE is not supported by architecture ${CMAKE_SYSTEM_PROCESSOR}")
+        endif()
+    endif()
+endmacro()
+
 #
 # ov_disable_all_warnings(<target1 [target2 target3 ...]>)
 #

cmake/developer_package/cross_compile/cross_compiled_disp_gen.cmake

@@ -18,6 +18,7 @@ set(_CPU_CHECK_ANY         "true")
 set(_CPU_CHECK_SSE42       "with_cpu_x86_sse42()")
 set(_CPU_CHECK_AVX         "with_cpu_x86_avx()")
 set(_CPU_CHECK_NEON_FP16   "with_cpu_neon_fp16()")
+set(_CPU_CHECK_SVE         "with_cpu_sve()")
 set(_CPU_CHECK_AVX2        "with_cpu_x86_avx2()")
 set(_CPU_CHECK_AVX512F     "with_cpu_x86_avx512f()")
 

cmake/developer_package/cross_compile/cross_compiled_func.cmake

@@ -3,14 +3,15 @@
 #
 
 ## list of available instruction sets
-set(_ARCH_LIST ANY SSE42 AVX AVX2 AVX512F NEON_FP16)
+set(_ARCH_LIST ANY SSE42 AVX AVX2 AVX512F NEON_FP16 SVE)
 
 set(_ACCEPTED_ARCHS_ANY     "^(ANY)$")
 set(_ACCEPTED_ARCHS_SSE42   "^(ANY|SSE42)$")
 set(_ACCEPTED_ARCHS_AVX     "^(ANY|SSE42|AVX)$")
 set(_ACCEPTED_ARCHS_AVX2    "^(ANY|SSE42|AVX|AVX2)$")
 set(_ACCEPTED_ARCHS_AVX512F "^(ANY|SSE42|AVX|AVX2|AVX512F)$")
 set(_ACCEPTED_ARCHS_NEON_FP16 "^(ANY|NEON_FP16)$")
+set(_ACCEPTED_ARCHS_SVE     "^(ANY|SVE)$")
 
 ## Arch specific definitions
 set(_DEFINE_ANY       "")
@@ -19,12 +20,14 @@ set(_DEFINE_AVX       "HAVE_AVX"      ${_DEFINE_SSE42})
 set(_DEFINE_AVX2      "HAVE_AVX2"     ${_DEFINE_AVX})
 set(_DEFINE_AVX512F   "HAVE_AVX512F"  ${_DEFINE_AVX2})
 set(_DEFINE_NEON_FP16 "HAVE_NEON_FP16" ${_DEFINE_ANY})
+set(_DEFINE_SVE       "HAVE_SVE"      ${_DEFINE_ANY})
 
 ## Arch specific compile options
 ov_avx512_optimization_flags(_FLAGS_AVX512F)
 ov_avx2_optimization_flags  (_FLAGS_AVX2)
 ov_sse42_optimization_flags (_FLAGS_SSE42)
 ov_arm_neon_fp16_optimization_flags(_FLAGS_NEON_FP16)
+ov_arm_sve_optimization_flags(_FLAGS_SVE)
 set(_FLAGS_AVX "")  ## TBD is not defined for OV project yet
 set(_FLAGS_ANY "")  ##
 
@@ -183,7 +186,9 @@ endfunction()
 #  Return currently requested ARCH id
 #
 function(_currently_requested_top_arch VAR)
-    if(ENABLE_NEON_FP16)
+    if(ENABLE_SVE)
+        set(RES SVE)
+    elseif(ENABLE_NEON_FP16)
         set(RES NEON_FP16)
     elseif(ENABLE_AVX512F)
         set(RES AVX512F)

cmake/developer_package/features.cmake

@@ -51,6 +51,8 @@ ov_dependent_option (ENABLE_AVX512F "Enable AVX512 optimizations" ON "X86_64 OR
 
 ov_dependent_option(ENABLE_NEON_FP16 "Enable ARM FP16 optimizations" ON "AARCH64" OFF)
 
+ov_dependent_option(ENABLE_SVE "Enable SVE optimizations" ON "AARCH64" OFF)
+
 # Type of build, we add this as an explicit option to default it to ON
 get_property(BUILD_SHARED_LIBS_DEFAULT GLOBAL PROPERTY TARGET_SUPPORTS_SHARED_LIBS)
 ov_option (BUILD_SHARED_LIBS "Build as a shared library" ${BUILD_SHARED_LIBS_DEFAULT})

src/inference/dev_api/openvino/runtime/system_conf.hpp

@@ -83,6 +83,13 @@ OPENVINO_RUNTIME_API bool with_cpu_x86_sse42();
  */
 OPENVINO_RUNTIME_API bool with_cpu_neon_fp16();
 
+/**
+ * @brief      Checks whether CPU supports ARM SVE capability
+ * @ingroup    ov_dev_api_system_conf
+ * @return     `True` if ARM SVE instructions are available, `false` otherwise
+ */
+OPENVINO_RUNTIME_API bool with_cpu_sve();
+
 /**
  * @brief      Checks whether CPU supports AVX capability
  * @ingroup    ov_dev_api_system_conf

src/inference/src/system_conf.cpp

@@ -22,6 +22,7 @@
 #    include <sys/auxv.h>
 #    define ARM_COMPUTE_CPU_FEATURE_HWCAP_FPHP    (1 << 9)
 #    define ARM_COMPUTE_CPU_FEATURE_HWCAP_ASIMDHP (1 << 10)
+#    define ARM_COMPUTE_CPU_FEATURE_HWCAP_SVE     (1 << 24)
 #elif defined(__APPLE__) && defined(__aarch64__)
 #    include <sys/sysctl.h>
 #    include <sys/types.h>
@@ -114,6 +115,10 @@ bool with_cpu_neon_fp16() {
     return false;
 }
 
+bool with_cpu_sve() {
+    return false;
+}
+
 #else  // OPENVINO_ARCH_X86 || OPENVINO_ARCH_X86_64
 
 bool with_cpu_x86_sse42() {
@@ -173,6 +178,20 @@ bool with_cpu_neon_fp16() {
     return false;
 #    endif
 }
+bool with_cpu_sve() {
+#    if !defined(_WIN64) && !defined(BARE_METAL) && !defined(__APPLE__) && !defined(__OpenBSD__) && \
+        !defined(__arm__) && defined(__aarch64__)
+    const uint32_t hwcaps = getauxval(AT_HWCAP);
+    return hwcaps & ARM_COMPUTE_CPU_FEATURE_HWCAP_SVE;
+#    elif !defined(_WIN64) && !defined(BARE_METAL) && !defined(__APPLE__) && !defined(__OpenBSD__) && \
+        !defined(__aarch64__) && defined(__arm__)
+    return false;
+#    elif defined(__aarch64__) && defined(__APPLE__)
+    return false;
+#    else
+    return false;
+#    endif
+}
 #endif  // OPENVINO_ARCH_X86 || OPENVINO_ARCH_X86_64
 
 bool check_open_mp_env_vars(bool include_omp_num_threads) {

src/plugins/intel_cpu/CMakeLists.txt

@@ -302,21 +302,21 @@ cross_compiled_file(${TARGET_NAME}
         NAMESPACE   ov::Extensions::Cpu::XARCH
 )
 cross_compiled_file(${TARGET_NAME}
-        ARCH AVX512F AVX2 ANY
+        ARCH AVX512F AVX2 SVE ANY
                     src/nodes/kernels/scaled_attn/executor_pa.cpp
         API         src/nodes/kernels/scaled_attn/executor_pa.hpp
         NAME        make_pa_executor
         NAMESPACE   ov::Extensions::Cpu::XARCH
 )
 cross_compiled_file(${TARGET_NAME}
-        ARCH AVX512F AVX2 ANY
+        ARCH AVX512F AVX2 SVE ANY
                     src/nodes/kernels/scaled_attn/attn_memcpy.cpp
         API         src/nodes/kernels/scaled_attn/attn_memcpy.hpp
         NAME        attn_memcpy paged_attn_memcpy attn_memcpy2d_kernel
         NAMESPACE   ov::Extensions::Cpu::XARCH
 )
 cross_compiled_file(${TARGET_NAME}
-        ARCH AVX512F AVX2 ANY
+        ARCH AVX512F AVX2 SVE ANY
                     src/nodes/kernels/scaled_attn/attn_quant.cpp
         API         src/nodes/kernels/scaled_attn/attn_quant.hpp
         NAME        attn_quantkv paged_attn_quantkv attn_quant_u8 attn_dequant_u8