Add support for properly optimized Windows ARM64 builds with LLVM and MSVC (#7191)

* logging: add proper checks for clang to avoid errors and warnings with VA_ARGS

* build: add CMake Presets and toolchian files for Windows ARM64

* matmul-int8: enable matmul-int8 with MSVC and fix Clang warnings

* ci: add support for optimized Windows ARM64 builds with MSVC and LLVM

* matmul-int8: fixed typos in q8_0_q8_0 matmuls

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* matmul-int8: remove unnecessary casts in q8_0_q8_0

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

7 files changed, 138 insertions, 58 deletions

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 7ac0e5f6..2d2fea4a 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -693,26 +693,28 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: 'rpc'
+          - build: 'rpc-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_RPC=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'noavx'
+          - build: 'noavx-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF -DBUILD_SHARED_LIBS=ON'
-          - build: 'avx2'
+          - build: 'avx2-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'avx'
+          - build: 'avx-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
-          - build: 'avx512'
+          - build: 'avx512-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'clblast'
+          - build: 'clblast-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
-          - build: 'openblas'
+          - build: 'openblas-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
-          - build: 'kompute'
+          - build: 'kompute-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'vulkan'
+          - build: 'vulkan-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'arm64'
-            defines: '-A ARM64 -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'llvm-arm64'
+            defines: '-G Ninja -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'msvc-arm64'
+            defines: '-G Ninja -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
 
     steps:
       - name: Clone
@@ -723,13 +725,13 @@ jobs:
 
       - name: Clone Kompute submodule
         id: clone_kompute
-        if: ${{ matrix.build == 'kompute' }}
+        if: ${{ matrix.build == 'kompute-x64' }}
         run: |
           git submodule update --init kompute
 
       - name: Download OpenCL SDK
         id: get_opencl
-        if: ${{ matrix.build == 'clblast' }}
+        if: ${{ matrix.build == 'clblast-x64' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/opencl.zip -L "https://github.com/KhronosGroup/OpenCL-SDK/releases/download/v${env:OPENCL_VERSION}/OpenCL-SDK-v${env:OPENCL_VERSION}-Win-x64.zip"
           mkdir $env:RUNNER_TEMP/opencl
@@ -737,7 +739,7 @@ jobs:
 
       - name: Download CLBlast
         id: get_clblast
-        if: ${{ matrix.build == 'clblast' }}
+        if: ${{ matrix.build == 'clblast-x64' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/clblast.7z -L "https://github.com/CNugteren/CLBlast/releases/download/${env:CLBLAST_VERSION}/CLBlast-${env:CLBLAST_VERSION}-windows-x64.7z"
           curl.exe -o $env:RUNNER_TEMP/CLBlast.LICENSE.txt -L "https://github.com/CNugteren/CLBlast/raw/${env:CLBLAST_VERSION}/LICENSE"
@@ -750,7 +752,7 @@ jobs:
 
       - name: Download OpenBLAS
         id: get_openblas
-        if: ${{ matrix.build == 'openblas' }}
+        if: ${{ matrix.build == 'openblas-x64' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip"
           curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE"
@@ -763,38 +765,41 @@ jobs:
 
       - name: Install Vulkan SDK
         id: get_vulkan
-        if: ${{ matrix.build == 'kompute' || matrix.build == 'vulkan' }}
+        if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
           Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
           Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
 
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
       - name: Build
         id: cmake_build
         run: |
-          mkdir build
-          cd build
-          cmake .. ${{ matrix.defines }}
-          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
+          cmake -S . -B build ${{ matrix.defines }}
+          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Add clblast.dll
         id: add_clblast_dll
-        if: ${{ matrix.build == 'clblast' }}
+        if: ${{ matrix.build == 'clblast-x64' }}
         run: |
           cp $env:RUNNER_TEMP/clblast/lib/clblast.dll ./build/bin/Release
           cp $env:RUNNER_TEMP/CLBlast.LICENSE.txt ./build/bin/Release/CLBlast-${env:CLBLAST_VERSION}.txt
 
       - name: Add libopenblas.dll
         id: add_libopenblas_dll
-        if: ${{ matrix.build == 'openblas' }}
+        if: ${{ matrix.build == 'openblas-x64' }}
         run: |
           cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll
           cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt
 
       - name: Check AVX512F support
         id: check_avx512f
-        if: ${{ matrix.build == 'avx512' }}
+        if: ${{ matrix.build == 'avx512-x64' }}
         continue-on-error: true
         run: |
           cd build
@@ -808,14 +813,14 @@ jobs:
       - name: Test
         id: cmake_test
         # not all machines have native AVX-512
-        if: ${{ matrix.build != 'arm64' && matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
+        if: ${{ matrix.build != 'msvc-arm64' && matrix.build != 'llvm-arm64' && matrix.build != 'clblast-x64' && matrix.build != 'kompute-x64' && matrix.build != 'vulkan-x64' && (matrix.build != 'avx512-x64' || env.HAS_AVX512F == '1') }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
 
       - name: Test (Intel SDE)
         id: cmake_test_sde
-        if: ${{ matrix.build == 'avx512' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
+        if: ${{ matrix.build == 'avx512-x64' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
         run: |
           curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz"
           # for some weird reason windows tar doesn't like sde tar.xz
@@ -843,14 +848,14 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip
-          name: llama-bin-win-${{ matrix.build }}-x64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip
+          name: llama-bin-win-${{ matrix.build }}.zip
 
   windows-latest-cmake-cuda:
     runs-on: windows-latest
diff --git a/CMakeLists.txt b/CMakeLists.txt
index feb6f39d..8ab6a45a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1007,6 +1007,11 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
         if (GGML_COMPILER_SUPPORT_DOTPROD)
             add_compile_definitions(__ARM_FEATURE_DOTPROD)
         endif ()
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+        if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
+            add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
+        endif ()
+
         check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
         if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
             add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
diff --git a/CMakePresets.json b/CMakePresets.json
new file mode 100644
index 00000000..ad1af7ec
--- /dev/null
+++ b/CMakePresets.json
@@ -0,0 +1,45 @@
+{
+  "version": 4,
+  "configurePresets": [
+    {
+        "name":  "base",
+        "hidden": true,
+        "generator":   "Ninja",
+        "binaryDir":   "${sourceDir}/build-${presetName}",
+        "cacheVariables": {
+            "CMAKE_EXPORT_COMPILE_COMMANDS": "ON",
+            "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
+        }
+    },
+
+    { "name": "debug",   "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
+    { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
+    { "name": "static",  "hidden": true, "cacheVariables": { "LLAMA_STATIC": "ON" } },
+
+    {
+        "name": "arm64-windows-msvc", "hidden": true,
+        "architecture": { "value": "arm64",       "strategy": "external" },
+        "toolset":      { "value": "host=x86_64", "strategy": "external" },
+        "cacheVariables": {
+            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-msvc.cmake"
+        }
+    },
+
+    {
+        "name": "arm64-windows-llvm", "hidden": true,
+        "architecture": { "value": "arm64",       "strategy": "external" },
+        "toolset":      { "value": "host=x86_64", "strategy": "external" },
+        "cacheVariables": {
+            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-llvm.cmake"
+        }
+    },
+
+    { "name": "arm64-windows-llvm-debug"  , "inherits": [ "base", "arm64-windows-llvm",  "debug"   ] },
+    { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm",  "release" ] },
+    { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm",  "release", "static" ] },
+
+    { "name": "arm64-windows-msvc-debug"  , "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
+    { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "release" ] },
+    { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "release", "static" ] }
+  ]
+}
diff --git a/cmake/arm64-windows-llvm.cmake b/cmake/arm64-windows-llvm.cmake
new file mode 100644
index 00000000..46fba651
--- /dev/null
+++ b/cmake/arm64-windows-llvm.cmake
@@ -0,0 +1,16 @@
+set( CMAKE_SYSTEM_NAME Windows )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+set( target arm64-pc-windows-msvc )
+
+set( CMAKE_C_COMPILER    clang )
+set( CMAKE_CXX_COMPILER  clang++ )
+
+set( CMAKE_C_COMPILER_TARGET   ${target} )
+set( CMAKE_CXX_COMPILER_TARGET ${target} )
+
+set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast" )
+set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )
+
+set( CMAKE_C_FLAGS_INIT   "${arch_c_flags} ${warn_c_flags}" )
+set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" )
diff --git a/cmake/arm64-windows-msvc.cmake b/cmake/arm64-windows-msvc.cmake
new file mode 100644
index 00000000..c7763142
--- /dev/null
+++ b/cmake/arm64-windows-msvc.cmake
@@ -0,0 +1,6 @@
+set( CMAKE_SYSTEM_NAME Windows )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+set( target arm64-pc-windows-msvc )
+set( CMAKE_C_COMPILER_TARGET   ${target} )
+set( CMAKE_CXX_COMPILER_TARGET ${target} )
diff --git a/common/log.h b/common/log.h
index 6934c57b..09fa63c2 100644
--- a/common/log.h
+++ b/common/log.h
@@ -211,7 +211,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
         #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
     #else
         #define LOG_FLF_FMT "[%24s:%5ld][%24s] "
-        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+        #define LOG_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
     #endif
 #else
     #define LOG_FLF_FMT "%s"
@@ -224,7 +224,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
         #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
     #else
         #define LOG_TEE_FLF_FMT "[%24s:%5ld][%24s] "
-        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+        #define LOG_TEE_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
     #endif
 #else
     #define LOG_TEE_FLF_FMT "%s"
@@ -294,7 +294,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // Main LOG macro.
 //  behaves like printf, and supports arguments the exact same way.
 //
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) || defined(__clang__)
     #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
 #else
     #define LOG(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "")
@@ -308,14 +308,14 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // Secondary target can be changed just like LOG_TARGET
 //  by defining LOG_TEE_TARGET
 //
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) || defined(__clang__)
     #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
 #else
     #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "")
 #endif
 
 // LOG macro variants with auto endline.
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) || defined(__clang__)
     #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
     #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
 #else
diff --git a/ggml-quants.c b/ggml-quants.c
index 9e62a3f3..f13599f6 100644
--- a/ggml-quants.c
+++ b/ggml-quants.c
@@ -3487,10 +3487,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
 #if defined(__ARM_FEATURE_MATMUL_INT8)
     if (nrc == 2) {
         const block_q4_0 * restrict vx0 = vx;
-        const block_q4_0 * restrict vx1 = vx + bx;
-
+        const block_q4_0 * restrict vx1 = (const block_q4_0 *) ((const uint8_t*)vx + bx);
         const block_q8_0 * restrict vy0 = vy;
-        const block_q8_0 * restrict vy1 = vy + by;
+        const block_q8_0 * restrict vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by);
 
         float32x4_t sumv0 = vdupq_n_f32(0.0f);
 
@@ -3524,10 +3523,12 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
             const int8x16_t y1_l = vld1q_s8(b_y1->qs);
             const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
 
-            float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                                 GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                                 GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                                 GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
+            float32_t _scale[4] = { GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                    GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                                    GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                    GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
+
+            float32x4_t scale = vld1q_f32(_scale);
 
             int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
             int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
@@ -3894,9 +3895,9 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 #if defined(__ARM_FEATURE_MATMUL_INT8)
     if (nrc == 2) {
         const block_q4_1 * restrict vx0 = vx;
-        const block_q4_1 * restrict vx1 = vx + bx;
+        const block_q4_1 * restrict vx1 = (const block_q4_1 *) ((const uint8_t*)vx + bx);
         const block_q8_1 * restrict vy0 = vy;
-        const block_q8_1 * restrict vy1 = vy + by;
+        const block_q8_1 * restrict vy1 = (const block_q8_1 *) ((const uint8_t*)vy + by);
 
         float32x4_t sumv0 = vdupq_n_f32(0.0f);
         float32x4_t summs0 = vdupq_n_f32(0.0f);
@@ -3907,11 +3908,11 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
             const block_q8_1 * restrict b_y0 = &vy0[i];
             const block_q8_1 * restrict b_y1 = &vy1[i];
 
-            float32x4_t summs_t = {GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
-                                   GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
-                                   GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
-                                   GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)};
-            summs0 += summs_t;
+            float32_t summs_t[4] = {GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
+                                    GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
+                                    GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
+                                    GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)};
+            summs0 = vaddq_f32(summs0, vld1q_f32(summs_t));
 
             const uint8x16_t m4b = vdupq_n_u8(0x0F);
 
@@ -3931,10 +3932,11 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
             const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
 
             // mmla into int32x4_t
-            float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*b_y0->d,
-                                 GGML_FP16_TO_FP32(b_x0->d)*b_y1->d,
-                                 GGML_FP16_TO_FP32(b_x1->d)*b_y0->d,
-                                 GGML_FP16_TO_FP32(b_x1->d)*b_y1->d};
+            float32_t _scale[4] = {GGML_FP16_TO_FP32(b_x0->d)*b_y0->d,
+                                   GGML_FP16_TO_FP32(b_x0->d)*b_y1->d,
+                                   GGML_FP16_TO_FP32(b_x1->d)*b_y0->d,
+                                   GGML_FP16_TO_FP32(b_x1->d)*b_y1->d};
+            float32x4_t scale = vld1q_f32(_scale);
 
             int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
             int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
@@ -3953,7 +3955,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 
         float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
         float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
-        sumv2 = sumv2 + summs0;
+        sumv2 = vaddq_f32(sumv2, summs0);
 
         vst1_f32(s, vget_low_f32(sumv2));
         vst1_f32(s + bs, vget_high_f32(sumv2));
@@ -4837,9 +4839,9 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
 #if defined(__ARM_FEATURE_MATMUL_INT8)
     if (nrc == 2) {
         const block_q8_0 * restrict vx0 = vx;
-        const block_q8_0 * restrict vx1 = vx + bx;
+        const block_q8_0 * restrict vx1 = (const block_q8_0 *) ((const uint8_t*)vx + bx);
         const block_q8_0 * restrict vy0 = vy;
-        const block_q8_0 * restrict vy1 = vy + by;
+        const block_q8_0 * restrict vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by);
 
         float32x4_t sumv0 = vdupq_n_f32(0.0f);
 
@@ -4861,10 +4863,11 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
             const int8x16_t y1_l = vld1q_s8(b_y1->qs);
             const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
 
-            float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                             GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                             GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                             GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
+            float32_t _scale[4] = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                   GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                                   GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                   GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
+            float32x4_t scale = vld1q_f32(_scale);
 
             int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
             int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));