fix performance regression on woa

CMakeLists.txt

@@ -72,6 +72,7 @@ else()
     set(INS_ENB ON)
 endif()
 
+option(LLAMA_LSE_ATOMICS                     "llama: enable LSE atomics"                        ON)
 option(LLAMA_SVE                             "llama: enable SVE"                                OFF)
 option(LLAMA_AVX                             "llama: enable AVX"                                ${INS_ENB})
 option(LLAMA_AVX2                            "llama: enable AVX2"                               ${INS_ENB})
@@ -1002,6 +1003,10 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
         add_compile_definitions(__ARM_NEON)
         add_compile_definitions(__ARM_FEATURE_FMA)
 
+        if (LLAMA_LSE_ATOMICS)
+            list(APPEND ARCH_FLAGS /arch:armv8.1)
+        endif()
+
         set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
         string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
         check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)

README.md

@@ -351,6 +351,13 @@ In order to build llama.cpp you have four different options.
       cmake --build build --config Debug
       ```
 
+    **Note**: (MSVC only) for Windows on ARM builds target preceding armv8.0, e.g. snapdragon 835 (ms8998):
+
+    ```bash
+    cmake -B build -DLLAMA_LSE_ATOMICS=OFF
+    cmake --build build --config Release
+    ```
+
 - Using `Zig` (version 0.11 or later):
 
     Building for optimization levels and CPU features can be accomplished using standard build arguments, for example AVX2, FMA, F16C,