Improvements for Windows with Snapdragon X

Author: AndreasKunar

cmake/arm64-windows-llvm.cmake

@@ -9,7 +9,8 @@ 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 -fno-finite-math-only" )
+# march for Snapdragon X should be 8.7-a, but this currently breaks Q_4_0_4_4 acceleration, 8.5 works
+set( arch_c_flags "-march=armv8.5-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
 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}" )

docs/build.md

@@ -16,7 +16,7 @@ In order to build llama.cpp you have four different options.
       make
       ```
 
-  - On Windows:
+  - On Windows (x86/x64 only, arm64 requires cmake):
 
     1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
     2. Extract `w64devkit` on your pc.
@@ -45,6 +45,13 @@ In order to build llama.cpp you have four different options.
     - For `Q4_0_4_4` quantization type build, add the `-DGGML_LLAMAFILE=OFF` cmake option. For example, use `cmake -B build -DGGML_LLAMAFILE=OFF`.
     - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
     - For faster repeated compilation, install [ccache](https://ccache.dev/).
+    - For Windows:
+      - Install cmake e.g. via `winget install cmake`:
+      - As alternative to the w64devkit mentioned in "using make" above, install MSVC (e.g. via Visual Studio 2022 Community Edition).
+      - For Windows on ARM you need MSVC installed and _additonally_:
+        - Install [clang via LLVM for woa64](https://releases.llvm.org) to enable better ARM optimizations (clang needs the MSVC backend).
+        - For using clang, the first build step needs to be `cmake --preset arm64-windows-llvm-release` (instead of the `cmake -B ...` which defaults to MSVC).
+        - Note: Building for ARM can also just be done with MSVC (without installing clang or using the preset), but this e.g. does not support Q_4_0_4_4 acceleration, because the MSVC frontend cannot inline ARM assembly-code.
     - For debug builds, there are two cases:
 
       1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):

ggml/src/ggml-aarch64.c

@@ -392,7 +392,7 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
 #if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
     GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) &&
                 "__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance");
-#elif defined(__ARM_NEON) && defined(__aarch64__)
+#elif defined(__ARM_NEON) && defined(__aarch64__) && ! defined(_MSC_VER)
     const void * b_ptr = vx;
     const void * a_ptr = vy;
     float * res_ptr = s;
@@ -501,7 +501,7 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
-#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) && ! defined(_MSC_VER)
     const void * b_ptr = vx;
     const void * a_ptr = vy;
     float * res_ptr = s;
@@ -613,7 +613,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if defined(__ARM_FEATURE_SVE)
+#if defined(__ARM_FEATURE_SVE) && ! defined(_MSC_VER)
     if (svcntw() == 8) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
@@ -753,7 +753,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
 #if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
     GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) &&
                 "__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance");
-#elif defined(__ARM_NEON) && defined(__aarch64__)
+#elif defined(__ARM_NEON) && defined(__aarch64__) && ! defined(_MSC_VER)
     const void * b_ptr = vx;
     const void * a_ptr = vy;
     float * res_ptr = s;
@@ -1271,7 +1271,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
-#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) && ! defined(_MSC_VER)
     const void * b_ptr = vx;
     const void * a_ptr = vy;
     float * res_ptr = s;
@@ -1727,7 +1727,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) && ! defined(_MSC_VER)
     if (svcntw() == 8) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;