diff --git a/CMakeLists.txt b/CMakeLists.txt
index 3b38603..8ba0cd1 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -26,7 +26,7 @@ string(REPLACE "i686" "ia32" ARCH ${ARCH})
 string(REPLACE "i386" "ia32" ARCH ${ARCH})
 string(REPLACE "armv7l" "arm" ARCH ${ARCH})
 string(REPLACE "arm" "arm" ARCH ${ARCH})
-string(REPLACE "arm64ex" "arm64" ARCH ${ARCH})
+string(REPLACE "arm64x" "arm64" ARCH ${ARCH})
 string(REPLACE "aarch64" "arm64" ARCH ${ARCH})
 
 if(DEFINED VARIANT)
@@ -58,6 +58,21 @@ include_directories(${CMAKE_JS_INC})
 # flags: -fPIC
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 
+# VULKAN_SDK
+if (VULKAN_SDK)
+  set(ENV{VULKAN_SDK} ${VULKAN_SDK})
+  find_package(Vulkan REQUIRED)
+endif()
+
+find_program(PATCH patch REQUIRED)
+
+add_custom_target(
+  patch ALL
+  COMMAND ${PATCH} -p1 -n -i ${CMAKE_SOURCE_DIR}/patches/llama.patch
+  WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}/src/llama.cpp
+  COMMENT "Applying patches"
+)
+
 set(LLAMA_STATIC ON CACHE BOOL "Build llama as static library")
 add_subdirectory("src/llama.cpp")
 
diff --git a/README.md b/README.md
index 1599652..9972e2e 100644
--- a/README.md
+++ b/README.md
@@ -47,8 +47,7 @@ console.log('Result:', text)
 
 ## Lib Variants
 
-- [x] `default`: General usage, Supported GPU: Metal (macOS) and Vulkan (Linux-x86_64)
-- [x] `opencl`: OpenCL (build with CLBlast), support Windows and Linux-aarch64
+- [x] `default`: General usage, Supported GPU: Metal (macOS) and Vulkan (Linux / Windows)
 
 ## License
 
diff --git a/patches/llama.patch b/patches/llama.patch
new file mode 100644
index 0000000..1ac4bdd
--- /dev/null
+++ b/patches/llama.patch
@@ -0,0 +1,59 @@
+diff --git a/ggml-vulkan.cpp b/ggml-vulkan.cpp
+index 1736ab73..55831936 100644
+--- a/ggml-vulkan.cpp
++++ b/ggml-vulkan.cpp
+@@ -25,6 +25,7 @@
+ 
+ #define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
+ 
++#define VK_VENDOR_ID_QUALCOMM 0x5143
+ #define VK_VENDOR_ID_AMD 0x1002
+ #define VK_VENDOR_ID_APPLE 0x106b
+ #define VK_VENDOR_ID_INTEL 0x8086
+@@ -1074,19 +1075,21 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
+         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
+         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
+ 
+-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
+-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
+-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
+-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
+-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
+-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
++        if (device->vendor_id != VK_VENDOR_ID_QUALCOMM) {
++            ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
++            ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
++            ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
++            ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
++            ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
++            ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
+ 
+-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
+-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
+-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
+-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
+-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
+-        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
++            ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
++            ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
++            ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
++            ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
++            ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
++            ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
++        }
+ 
+         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
+         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
+@@ -1208,8 +1211,10 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
+     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q8_0], "dequant_q8_0", dequant_q8_0_len, dequant_q8_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
+     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q2_K], "dequant_q2_K", dequant_q2_K_len, dequant_q2_K_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
+     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q3_K], "dequant_q3_K", dequant_q3_K_len, dequant_q3_K_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
+-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q4_K], "dequant_q4_K", dequant_q4_K_len, dequant_q4_K_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1);
+-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q5_K], "dequant_q5_K", dequant_q5_K_len, dequant_q5_K_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
++    if (device->vendor_id != VK_VENDOR_ID_QUALCOMM) {
++        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q4_K], "dequant_q4_K", dequant_q4_K_len, dequant_q4_K_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1);
++        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q5_K], "dequant_q5_K", dequant_q5_K_len, dequant_q5_K_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
++    }
+     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q6_K], "dequant_q6_K", dequant_q6_K_len, dequant_q6_K_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
+ 
+     // get_rows
diff --git a/scripts/build-linux.sh b/scripts/build-linux.sh
index 0fe45b1..cbf8f58 100755
--- a/scripts/build-linux.sh
+++ b/scripts/build-linux.sh
@@ -5,8 +5,5 @@ set -e
 # General
 
 yarn clean && yarn build-native -a x86_64 --CDLLAMA_VULKAN=1
-yarn clean && yarn build-native -a aarch64 --cc aarch64-linux-gnu-gcc --cxx aarch64-linux-gnu-g++
-
-# build with CLBlast
 yarn clean && yarn build-native -a aarch64 --cc aarch64-linux-gnu-gcc --cxx aarch64-linux-gnu-g++ \
-  --CDVARIANT=opencl --CDLLAMA_CLBLAST=ON --CDCMAKE_PREFIX_PATH="$(realpath 'externals/SDK/lib/cmake/CLBlast')"
+  --CDLLAMA_VULKAN=1 --CDVULKAN_SDK="$(realpath 'externals/arm64-Vulkan-SDK')"
diff --git a/scripts/build-windows.ps1 b/scripts/build-windows.ps1
index c68d858..ecd13f1 100644
--- a/scripts/build-windows.ps1
+++ b/scripts/build-windows.ps1
@@ -2,12 +2,5 @@ $ErrorActionPreference='Stop'
 
 # General
 
-yarn clean ; yarn build-native -a x86_64
-yarn clean ; yarn build-native -a arm64
-
-# build with CLBlast
-$X64_CLBlast=Resolve-Path 'externals/win32-x64/SDK/lib/cmake/CLBlast'
-$ARM64_CLBlast=Resolve-Path 'externals/win32-arm64/SDK/lib/cmake/CLBlast'
-
-yarn clean ; yarn build-native -a x86_64 --CDVARIANT=opencl --CDLLAMA_CLBLAST=ON --CDCMAKE_PREFIX_PATH=$X64_CLBlast
-yarn clean ; yarn build-native -a arm64 --CDVARIANT=opencl --CDLLAMA_CLBLAST=ON --CDCMAKE_PREFIX_PATH=$ARM64_CLBlast
+yarn clean ; yarn build-native -a x86_64 --CDCMAKE_PREFIX_PATH=(Resolve-Path 'externals/win32-x64/Vulkan-SDK')
+yarn clean ; yarn build-native -a arm64 --CDCMAKE_PREFIX_PATH=(Resolve-Path 'externals/win32-arm64/Vulkan-SDK')
diff --git a/scripts/prepare-linux.sh b/scripts/prepare-linux.sh
index fb122ee..cb7d2d5 100755
--- a/scripts/prepare-linux.sh
+++ b/scripts/prepare-linux.sh
@@ -33,20 +33,36 @@ if [ ! -d CLBlast-source ]; then
   cd ..
 fi
 
-SDK_ROOT=$(realpath ./SDK)
+# build Vulkan SDK from source for arm64
 
-if [ ! -d SDK ]; then
-  # build from source
-  export CC=aarch64-linux-gnu-gcc
-  export CXX=aarch64-linux-gnu-g++
-  cd OpenCL-SDK-source
-  rm -rf build
-  cmake -B build -DBUILD_DOCS=OFF -DBUILD_EXAMPLES=OFF -DBUILD_TESTING=OFF -DOPENCL_SDK_BUILD_SAMPLES=OFF -DOPENCL_SDK_TEST_SAMPLES=OFF -DCMAKE_INSTALL_PREFIX=$SDK_ROOT
-  cmake --build build --config Release
-  cmake --install build
-  cd ../CLBlast-source
-  rm -rf build
-  cmake -B build -DBUILD_SHARED_LIBS=ON -DOVERRIDE_MSVC_FLAGS_TO_MT=OFF -DTUNERS=OFF -DOPENCL_ROOT="$SDK_ROOT" -DCMAKE_INSTALL_PREFIX=$SDK_ROOT
-  cmake --build build --config Release
-  cmake --install build
+if [ ! -d arm64-Vulkan-SDK ]; then
+  VULKAN_ROOT=$(realpath ./arm64-Vulkan-SDK)
+  if [ ! -d Vulkan-Headers-source ]; then
+    git clone "https://github.com/KhronosGroup/Vulkan-Headers.git" "Vulkan-Headers-source"
+    cd "Vulkan-Headers-source"
+    git checkout "sdk-1.3.261"
+    CC=aarch64-linux-gnu-gcc CXX=aarch64-linux-gnu-g++ cmake -B build
+    cmake --build build --config Release
+    cmake --install build --prefix $VULKAN_ROOT
+    cd ..
+  fi
+  if [ ! -d Vulkan-Loader-source ]; then
+    git clone "https://github.com/KhronosGroup/Vulkan-Loader.git" "Vulkan-Loader-source"
+    cd "Vulkan-Loader-source"
+    git checkout "sdk-1.3.261"
+    CC=aarch64-linux-gnu-gcc CXX=aarch64-linux-gnu-g++ cmake -B build -DVULKAN_HEADERS_INSTALL_DIR="$VULKAN_ROOT" -DUSE_MASM=OFF
+    cmake --build build --config Release
+    cmake --install build --prefix $VULKAN_ROOT
+    cd ".."
+  fi
+  if [ ! -d Vulkan-Hpp-source ]; then
+    git clone "https://github.com/KhronosGroup/Vulkan-Hpp.git" "Vulkan-Hpp-source"
+    cd "Vulkan-Hpp-source"
+    git checkout "v1.3.261"
+    git submodule update --init --recursive
+    cmake -B build -DVULKAN_HPP_INSTALL=ON -DVULKAN_HPP_RUN_GENERATOR=ON
+    cmake --build build --config Release
+    cmake --install build --prefix $VULKAN_ROOT
+    cd ".."
+  fi
 fi
diff --git a/scripts/prepare-windows.ps1 b/scripts/prepare-windows.ps1
index 5f6356e..d4dddeb 100644
--- a/scripts/prepare-windows.ps1
+++ b/scripts/prepare-windows.ps1
@@ -11,54 +11,72 @@ New-Item -ItemType Directory -Force -Path "externals" | Out-Null
 
 cd "externals"
 
-if (-not (Test-Path "OpenCL-SDK-source")) {
-  git clone "https://github.com/KhronosGroup/OpenCL-SDK.git" "OpenCL-SDK-source" --recursive
-  cd "OpenCL-SDK-source"
-  git checkout v2023.12.14
-  cd ".."
-}
-
-if (-not (Test-Path "CLBlast-source")) {
-  git clone "https://github.com/CNugteren/CLBlast.git" "CLBlast-source" --recursive
-  cd "CLBlast-source"
-  git checkout 1.6.2
-  cd ".."
+if (-not (Test-Path "win32-x64/Vulkan-SDK") -or -not (Test-Path "win32-arm64/Vulkan-SDK")) {
+  if (-not (Test-Path "Vulkan-Headers-source")) {
+    git clone "https://github.com/KhronosGroup/Vulkan-Headers.git" "Vulkan-Headers-source"
+    cd "Vulkan-Headers-source"
+    git checkout sdk-1.3.261
+    cd ".."
+  }
+  if (-not (Test-Path "Vulkan-Loader-source")) {
+    git clone "https://github.com/KhronosGroup/Vulkan-Loader.git" "Vulkan-Loader-source"
+    cd "Vulkan-Loader-source"
+    git checkout sdk-1.3.261
+    cd ".."
+  }
+  if (-not (Test-Path "Vulkan-Hpp-source")) {
+    git clone "https://github.com/KhronosGroup/Vulkan-Hpp.git" "Vulkan-Hpp-source"
+    cd "Vulkan-Hpp-source"
+    git checkout v1.3.261
+    git submodule update --init --recursive
+    cd ".."
+  }
 }
 
 # x64
 
-if (-not (Test-Path "win32-x64/SDK")) {
-  New-Item -ItemType Directory -Force -Path "win32-x64/SDK" | Out-Null
-  $SDK_ROOT = (Resolve-Path "win32-x64/SDK").Path
+if (-not (Test-Path "win32-x64/Vulkan-SDK") {
+  New-Item -ItemType Directory -Force -Path "win32-x64/Vulkan-SDK" | Out-Null
+  $VULKAN_ROOT = (Resolve-Path "win32-x64/Vulkan-SDK").Path
   # build from source
-  cd "OpenCL-SDK-source"
+  cd "Vulkan-Headers-source"
+  DelIfExist "build"
+  cmake -B build -G "Visual Studio 17 2022" -A x64
+  cmake --build build --config Release
+  cmake --install build --prefix $VULKAN_ROOT
+  cd "../Vulkan-Loader-source"
   DelIfExist "build"
-  cmake -B build -DBUILD_DOCS=OFF -DBUILD_EXAMPLES=OFF -DBUILD_TESTING=OFF -DOPENCL_SDK_BUILD_SAMPLES=OFF -DOPENCL_SDK_TEST_SAMPLES=OFF -G "Visual Studio 17 2022"
+  cmake -B build -G "Visual Studio 17 2022" -DVULKAN_HEADERS_INSTALL_DIR="$VULKAN_ROOT" -DUSE_MASM=OFF -A x64
   cmake --build build --config Release
-  cmake --install build --prefix $SDK_ROOT
-  cd "../CLBlast-source"
+  cmake --install build --prefix $VULKAN_ROOT
+  cd "../Vulkan-Hpp-source"
   DelIfExist "build"
-  cmake -B build -DBUILD_SHARED_LIBS=ON -DOVERRIDE_MSVC_FLAGS_TO_MT=OFF -DTUNERS=OFF -DOPENCL_ROOT="$SDK_ROOT" -G "Visual Studio 17 2022"
+  cmake -B build -G "Visual Studio 17 2022" -DVULKAN_HPP_INSTALL=ON -DVULKAN_HPP_RUN_GENERATOR=ON
   cmake --build build --config Release
-  cmake --install build --prefix $SDK_ROOT
+  cmake --install build --prefix $VULKAN_ROOT
   cd ".."
 }
 
 # ARM64
 
-if (-not (Test-Path "win32-arm64/SDK")) {
-  New-Item -ItemType Directory -Force -Path "win32-arm64/SDK" | Out-Null
-  $SDK_ROOT = (Resolve-Path "win32-arm64/SDK").Path
+if (-not (Test-Path "win32-arm64/Vulkan-SDK")) {
+  New-Item -ItemType Directory -Force -Path "win32-arm64/Vulkan-SDK" | Out-Null
+  $VULKAN_ROOT = (Resolve-Path "win32-arm64/Vulkan-SDK").Path
   # build from source
-  cd "OpenCL-SDK-source"
+  cd "Vulkan-Headers-source"
+  DelIfExist "build"
+  cmake -B build -G "Visual Studio 17 2022" -A arm64
+  cmake --build build --config Release
+  cmake --install build --prefix $VULKAN_ROOT
+  cd "../Vulkan-Loader-source"
   DelIfExist "build"
-  cmake -B build -DBUILD_DOCS=OFF -DBUILD_EXAMPLES=OFF -DBUILD_TESTING=OFF -DOPENCL_SDK_BUILD_SAMPLES=OFF -DOPENCL_SDK_TEST_SAMPLES=OFF -G "Visual Studio 17 2022" -A arm64
+  cmake -B build -G "Visual Studio 17 2022" -DVULKAN_HEADERS_INSTALL_DIR="$VULKAN_ROOT" -DUSE_MASM=OFF -A arm64
   cmake --build build --config Release
-  cmake --install build --prefix $SDK_ROOT
-  cd "../CLBlast-source"
+  cmake --install build --prefix $VULKAN_ROOT
+  cd "../Vulkan-Hpp-source"
   DelIfExist "build"
-  cmake -B build -DBUILD_SHARED_LIBS=ON -DOVERRIDE_MSVC_FLAGS_TO_MT=OFF -DTUNERS=OFF -DOPENCL_ROOT="$SDK_ROOT" -G "Visual Studio 17 2022" -A arm64
+  cmake -B build -G "Visual Studio 17 2022" -DVULKAN_HPP_INSTALL=ON -DVULKAN_HPP_RUN_GENERATOR=ON
   cmake --build build --config Release
-  cmake --install build --prefix $SDK_ROOT
+  cmake --install build --prefix $VULKAN_ROOT
   cd ".."
 }