ggml : automatic selection of best CPU backend (#10606)

* ggml : automatic selection of best CPU backend

* amx : minor opt

* add GGML_AVX_VNNI to enable avx-vnni, fix checks

.devops/llama-server.Dockerfile

@@ -3,22 +3,34 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION AS build
 
 RUN apt-get update && \
-    apt-get install -y build-essential git libcurl4-openssl-dev
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
 WORKDIR /app
 
 COPY . .
 
-ENV LLAMA_CURL=1
 
-RUN make -j$(nproc) llama-server
+RUN \
+    # Build multiple versions of the CPU backend
+    scripts/build-cpu.sh avx         -DGGML_AVX=ON -DGGML_AVX2=OFF && \
+    scripts/build-cpu.sh avx2        -DGGML_AVX=ON -DGGML_AVX2=ON && \
+    scripts/build-cpu.sh avx512      -DGGML_AVX=ON -DGGML_AVX2=ON -DGGML_AVX512=ON && \
+    scripts/build-cpu.sh amx         -DGGML_AVX=ON -DGGML_AVX2=ON -DGGML_AVX512=ON -DGGML_AVX_VNNI=ON -DGGML_AVX512_VNNI=ON -DGGML_AMX_TILE=ON -DGGML_AMX_INT8=ON && \
+    # Build llama-server
+    cmake -S . -B build -DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DLLAMA_CURL=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build build --target llama-server -j $(nproc) && \
+    # Copy the built libraries to /app/lib
+    mkdir -p /app/lib && \
+    mv libggml-cpu* /app/lib/ && \
+    find build -name "*.so" -exec cp {} /app/lib/ \;
 
 FROM ubuntu:$UBUNTU_VERSION AS runtime
 
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
-COPY --from=build /app/llama-server /llama-server
+COPY --from=build /app/build/bin/llama-server /llama-server
+COPY --from=build /app/lib/ /
 
 ENV LC_ALL=C.utf8
 # Must be set to 0.0.0.0 so it can listen to requests from host machine

CMakeLists.txt

@@ -96,10 +96,6 @@ if (NOT DEFINED GGML_LLAMAFILE)
     set(GGML_LLAMAFILE_DEFAULT ON)
 endif()
 
-if (NOT DEFINED GGML_AMX)
-    set(GGML_AMX ON)
-endif()
-
 if (NOT DEFINED GGML_CUDA_GRAPHS)
     set(GGML_CUDA_GRAPHS_DEFAULT ON)
 endif()

Package.swift

@@ -88,5 +88,5 @@ let package = Package(
             linkerSettings: linkerSettings
         )
     ],
-    cxxLanguageStandard: .cxx11
+    cxxLanguageStandard: .cxx17
 )

ggml/CMakeLists.txt

@@ -96,6 +96,7 @@ option(GGML_CPU_HBM     "ggml: use memkind for CPU HBM" OFF)
 option(GGML_CPU_AARCH64 "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
 
 option(GGML_AVX         "ggml: enable AVX"              ${INS_ENB})
+option(GGML_AVX_VNNI    "ggml: enable AVX-VNNI"         OFF)
 option(GGML_AVX2        "ggml: enable AVX2"             ${INS_ENB})
 option(GGML_AVX512      "ggml: enable AVX512"           OFF)
 option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI"      OFF)

ggml/src/ggml-backend-impl.h

@@ -211,27 +211,45 @@ extern "C" {
     GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);
 
     // Add backend dynamic loading support to the backend
-    typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
 
-    #ifdef GGML_BACKEND_DL
+    // Initialize the backend
-        #ifdef __cplusplus
+    typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
-        #    define GGML_BACKEND_DL_IMPL(reg_fn)                                 \
+    // Optional: obtain a score for the backend based on the system configuration
-                extern "C" {                                                     \
+    // Higher scores are preferred, 0 means the backend is not supported in the current system
-                    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
+    typedef int                (*ggml_backend_score_t)(void);
-                }                                                                \
+
-                ggml_backend_reg_t ggml_backend_init(void) {                     \
+#ifdef GGML_BACKEND_DL
-                    return reg_fn();                                             \
+#    ifdef __cplusplus
-                }
+#        define GGML_BACKEND_DL_IMPL(reg_fn)                             \
-        #else
+            extern "C" {                                                 \
-        #    define GGML_BACKEND_DL_IMPL(reg_fn)                             \
+            GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
-                GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
+            }                                                            \
-                ggml_backend_reg_t ggml_backend_init(void) {                 \
+            ggml_backend_reg_t ggml_backend_init(void) {                 \
-                    return reg_fn();                                         \
+                return reg_fn();                                         \
-                }
+            }
-        #endif
+#        define GGML_BACKEND_DL_SCORE_IMPL(score_fn)       \
-    #else
+            extern "C" {                                   \
-    #    define GGML_BACKEND_DL_IMPL(reg_fn)
+            GGML_BACKEND_API int ggml_backend_score(void); \
-    #endif
+            }                                              \
+            int ggml_backend_score(void) {                 \
+                return score_fn();                         \
+            }
+#    else
+#        define GGML_BACKEND_DL_IMPL(reg_fn)                              \
+            GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void);  \
+            ggml_backend_reg_t                  ggml_backend_init(void) { \
+                return reg_fn();                                          \
+            }
+#        define GGML_BACKEND_DL_SCORE_IMPL(score_fn)        \
+            GGML_BACKEND_API int ggml_backend_score(void);  \
+            int                  ggml_backend_score(void) { \
+                return score_fn();                          \
+            }
+#    endif
+#else
+#    define GGML_BACKEND_DL_IMPL(reg_fn)
+#    define GGML_BACKEND_DL_SCORE_IMPL(score_fn)
+#endif
 
 #ifdef  __cplusplus
 }

ggml/src/ggml-backend-reg.cpp

@@ -2,8 +2,13 @@
 #include "ggml-backend.h"
 #include "ggml-impl.h"
 #include <algorithm>
+#include <codecvt>
 #include <cstring>
+#include <filesystem>
+#include <locale>
+#include <memory>
 #include <string>
+#include <type_traits>
 #include <vector>
 
 #ifdef _WIN32
@@ -57,9 +62,71 @@
 #include "ggml-kompute.h"
 #endif
 
+#ifdef _WIN32
+
+using dl_handle = std::remove_pointer_t<HMODULE>;
+
+struct dl_handle_deleter {
+    void operator()(HMODULE handle) {
+        FreeLibrary(handle);
+    }
+};
+
+static dl_handle * dl_load_library(const std::wstring & path) {
+    // suppress error dialogs for missing DLLs
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    HMODULE handle = LoadLibraryW(path.c_str());
+
+    SetErrorMode(old_mode);
+
+    return handle;
+}
+
+static dl_handle * dl_load_library(const std::string & path) {
+    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+    return dl_load_library(converter.from_bytes(path));
+}
+
+static void * dl_get_sym(dl_handle * handle, const char * name) {
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    void * p = (void *) GetProcAddress(handle, name);
+
+    SetErrorMode(old_mode);
+
+    return p;
+}
+
+#else
+
+using dl_handle = void;
+
+struct dl_handle_deleter {
+    void operator()(void * handle) {
+        dlclose(handle);
+    }
+};
+
+static void * dl_load_library(const std::string & path) {
+    dl_handle * handle = dlopen(path.c_str(), RTLD_NOW | RTLD_LOCAL);
+
+    return handle;
+}
+
+static void * dl_get_sym(dl_handle * handle, const char * name) {
+    return dlsym(handle, name);
+}
+
+#endif
+
+using dl_handle_ptr = std::unique_ptr<dl_handle, dl_handle_deleter>;
+
 struct ggml_backend_reg_entry {
     ggml_backend_reg_t reg;
-    void * handle;
+    dl_handle_ptr handle;
 };
 
 struct ggml_backend_registry {
@@ -97,13 +164,16 @@ struct ggml_backend_registry {
     }
 
     ~ggml_backend_registry() {
-        while (!backends.empty()) {
+        // FIXME: backends cannot be safely unloaded without a function to destroy all the backend resources,
-            // use silent since the log system may have been destroyed at this point
+        // since backend threads may still be running and accessing resources from the dynamic library
-            unload_backend(backends.back().reg, true);
+        for (auto & entry : backends) {
+            if (entry.handle) {
+                entry.handle.release(); // NOLINT
+            }
         }
     }
 
-    void register_backend(ggml_backend_reg_t reg, void * handle = nullptr) {
+    void register_backend(ggml_backend_reg_t reg, dl_handle_ptr handle = nullptr) {
         if (!reg) {
             return;
         }
@@ -112,7 +182,7 @@ struct ggml_backend_registry {
         GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
             __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
 #endif
-        backends.push_back({ reg, handle });
+        backends.push_back({ reg, std::move(handle) });
         for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
             register_device(ggml_backend_reg_dev_get(reg, i));
         }
@@ -126,79 +196,53 @@ struct ggml_backend_registry {
     }
 
     ggml_backend_reg_t load_backend(const char * path, bool silent) {
-#ifdef _WIN32
+        dl_handle_ptr handle { dl_load_library(path) };
-        // suppress error dialogs for missing DLLs
-        DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
-        SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
-
-        HMODULE handle = LoadLibraryA(path);
-
         if (!handle) {
             if (!silent) {
-                GGML_LOG_ERROR("%s: failed to load %s: %lu\n", __func__, path, GetLastError());
+                GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path);
             }
-            SetErrorMode(old_mode);
             return nullptr;
         }
 
-        ggml_backend_init_t backend_init = (ggml_backend_init_t) GetProcAddress(handle, "ggml_backend_init");
+        auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
-
+        if (score_fn && score_fn() == 0) {
-        SetErrorMode(old_mode);
-
-        if (!backend_init) {
             if (!silent) {
-                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s: %lu\n", __func__, path, GetLastError());
+                GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, path);
             }
-            FreeLibrary(handle);
             return nullptr;
         }
-#else
-        void * handle = dlopen(path, RTLD_NOW | RTLD_LOCAL);
 
-        if (!handle) {
+        auto backend_init_fn = (ggml_backend_init_t) dl_get_sym(handle.get(), "ggml_backend_init");
+        if (!backend_init_fn) {
             if (!silent) {
-                GGML_LOG_ERROR("%s: failed to load %s: %s\n", __func__, path, dlerror());
+                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, path);
             }
             return nullptr;
         }
 
-        auto * backend_init = (ggml_backend_init_t) dlsym(handle, "ggml_backend_init");
+        ggml_backend_reg_t reg = backend_init_fn();
-
-        if (!backend_init) {
-            if (!silent) {
-                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s: %s\n", __func__, path, dlerror());
-            }
-            dlclose(handle);
-            return nullptr;
-        }
-#endif
-        ggml_backend_reg_t reg = backend_init();
-
         if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) {
             if (!silent) {
                 if (!reg) {
                     GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, path);
                 } else {
                     GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
-                                   __func__, path, reg->api_version, GGML_BACKEND_API_VERSION);
+                        __func__, path, reg->api_version, GGML_BACKEND_API_VERSION);
                 }
             }
-#ifdef _WIN32
-            FreeLibrary(handle);
-#else
-            dlclose(handle);
-#endif
             return nullptr;
         }
 
         GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path);
-        register_backend(reg, handle);
+
+        register_backend(reg, std::move(handle));
+
         return reg;
     }
 
     void unload_backend(ggml_backend_reg_t reg, bool silent) {
         auto it = std::find_if(backends.begin(), backends.end(),
-                                [reg](ggml_backend_reg_entry entry) { return entry.reg == reg; });
+                               [reg](const ggml_backend_reg_entry & entry) { return entry.reg == reg; });
 
         if (it == backends.end()) {
             if (!silent) {
@@ -217,15 +261,6 @@ struct ggml_backend_registry {
                             [reg](ggml_backend_dev_t dev) { return ggml_backend_dev_backend_reg(dev) == reg; }),
             devices.end());
 
-        // unload library
-        if (it->handle) {
-#ifdef _WIN32
-            FreeLibrary((HMODULE) it->handle);
-#else
-            dlclose(it->handle);
-#endif
-        }
-
         // remove backend
         backends.erase(it);
     }
@@ -341,12 +376,7 @@ void ggml_backend_unload(ggml_backend_reg_t reg) {
     get_reg().unload_backend(reg, true);
 }
 
-void ggml_backend_load_all() {
+static std::string get_executable_path() {
-    std::vector<std::string> search_prefix;
-
-    // add the executable directory to the search path
-    // FIXME: this is convenient for development, but it should probably be disabled in production
-
 #if defined(__APPLE__)
     // get executable path
     std::vector<char> path;
@@ -364,7 +394,7 @@ void ggml_backend_load_all() {
     if (last_slash != std::string::npos) {
         base_path = base_path.substr(0, last_slash);
     }
-    search_prefix.push_back(base_path + "/");
+    return base_path + "/";
 #elif defined(__linux__)
     std::string base_path = ".";
     std::vector<char> path(1024);
@@ -386,38 +416,104 @@ void ggml_backend_load_all() {
         path.resize(path.size() * 2);
     }
 
-    search_prefix.push_back(base_path + "/");
+    return base_path + "/";
+#elif defined(_WIN32)
+    std::vector<char> path(MAX_PATH);
+    DWORD len = GetModuleFileNameA(NULL, path.data(), path.size());
+    if (len == 0) {
+        return "";
+    }
+    std::string base_path(path.data(), len);
+    // remove executable name
+    auto last_slash = base_path.find_last_of('\\');
+    if (last_slash != std::string::npos) {
+        base_path = base_path.substr(0, last_slash);
+    }
+    return base_path + "\\";
 #endif
+}
 
-    auto & reg = get_reg();
+static std::string backend_filename_prefix() {
-
-    auto try_load = [&](const std::string & name) {
-        std::string os_name;
 #ifdef _WIN32
-        os_name = "ggml-" + name + ".dll";
+    return "ggml-";
 #else
-        os_name = "libggml-" + name + ".so";
+    return "libggml-";
 #endif
-        if (reg.load_backend(os_name.c_str(), true)) {
+}
-            return;
+
+static std::string backend_filename_suffix() {
+#ifdef _WIN32
+    return ".dll";
+#else
+    return ".so";
+#endif
+}
+
+static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent) {
+    // enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths
+     // TODO: search system paths
+    std::vector<std::string> search_paths = { "./", get_executable_path() };
+    std::string file_prefix = backend_filename_prefix() + name + "-";
+
+    int best_score = 0;
+    std::string best_path;
+
+    namespace fs = std::filesystem;
+    for (const auto & search_path : search_paths) {
+        if (!fs::exists(search_path)) {
+            continue;
         }
-        for (const auto & prefix : search_prefix) {
+        for (const auto & entry : fs::directory_iterator(search_path)) {
-            if (reg.load_backend((prefix + os_name).c_str(), true)) {
+            if (entry.is_regular_file()) {
-                return;
+                std::string filename = entry.path().filename().string();
+                std::string ext = entry.path().extension().string();
+                if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) {
+                    dl_handle_ptr handle { dl_load_library(entry.path().c_str()) };
+                    if (!handle && !silent) {
+                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, entry.path().string().c_str());
+                    }
+                    if (handle) {
+                        auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
+                        if (score_fn) {
+                            int s = score_fn();
+#ifndef NDEBUG
+                            GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, entry.path().string().c_str(), s);
+#endif
+                            if (s > best_score) {
+                                best_score = s;
+                                best_path = entry.path().string();
+                            }
+                        }
+                    }
+                }
             }
         }
-    };
+    }
 
-    try_load("amx");
+    if (best_score == 0) {
-    try_load("blas");
+        // try to load the base backend
-    try_load("cann");
+        for (const auto & search_path : search_paths) {
-    try_load("cuda");
+            std::string path = search_path + backend_filename_prefix() + name + backend_filename_suffix();
-    try_load("hip");
+            if (fs::exists(path)) {
-    try_load("kompute");
+                return get_reg().load_backend(path.c_str(), silent);
-    try_load("metal");
+            }
-    try_load("rpc");
+        }
-    try_load("sycl");
+        return nullptr;
-    try_load("vulkan");
+    }
-    try_load("musa");
+
-    try_load("cpu");
+    return get_reg().load_backend(best_path.c_str(), silent);
+}
+
+void ggml_backend_load_all() {
+    ggml_backend_load_best("blas", true);
+    ggml_backend_load_best("cann", true);
+    ggml_backend_load_best("cuda", true);
+    ggml_backend_load_best("hip", true);
+    ggml_backend_load_best("kompute", true);
+    ggml_backend_load_best("metal", true);
+    ggml_backend_load_best("rpc", true);
+    ggml_backend_load_best("sycl", true);
+    ggml_backend_load_best("vulkan", true);
+    ggml_backend_load_best("musa", true);
+    ggml_backend_load_best("cpu", true);
 }

ggml/src/ggml-cpu/CMakeLists.txt

@@ -217,6 +217,12 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
         elseif (GGML_AVX)
             list(APPEND ARCH_FLAGS /arch:AVX)
         endif()
+        if (GGML_AVX_VNNI)
+            list(APPEND ARCH_DEFINITIONS __AVXVNNI__)
+            if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                list(APPEND ARCH_FLAGS -mavxvnni)
+            endif()
+        endif()
     else()
         if (GGML_NATIVE)
             list(APPEND ARCH_FLAGS -march=native)
@@ -233,6 +239,9 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
         if (GGML_AVX2)
             list(APPEND ARCH_FLAGS -mavx2)
         endif()
+        if (GGML_AVX_VNNI)
+            list(APPEND ARCH_FLAGS -mavxvnni)
+        endif()
         if (GGML_AVX512)
             list(APPEND ARCH_FLAGS -mavx512f)
             list(APPEND ARCH_FLAGS -mavx512dq)
@@ -301,6 +310,10 @@ target_sources(ggml-cpu PRIVATE ${GGML_CPU_SOURCES})
 set_source_files_properties(${GGML_CPU_SOURCES} PROPERTIES COMPILE_OPTIONS     "${ARCH_FLAGS}")
 set_source_files_properties(${GGML_CPU_SOURCES} PROPERTIES COMPILE_DEFINITIONS "${ARCH_DEFINITIONS}")
 
+# the feature detection code must be compiled without any architecture flags
+target_sources(ggml-cpu PRIVATE cpu-feats-x86.cpp)
+# target_sources(ggml-cpu PRIVATE cpu-feats-arm.cpp) # TODO: ARM feature detection
+
 if (EMSCRIPTEN)
     set_target_properties(ggml-cpu PROPERTIES COMPILE_FLAGS "-msimd128")
 endif()

ggml/src/ggml-cpu/amx/common.h

@@ -78,7 +78,6 @@ inline void parallel_for_ggml(const ggml_compute_params * params, int n, const f
     int tbegin, tend;
     balance211(n, params->nth, params->ith, tbegin, tend);
     f(tbegin, tend);
-    ggml_barrier(params->threadpool); // TODO: might not always be needed
 }
 
 // quantized types that have AMX support

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -1340,21 +1340,19 @@ struct tinygemm_kernel_avx<float, ggml_fp16_t, float, BLOCK_M, BLOCK_N, BLOCK_K>
         __m512 vb[COLS];
         __m512 vc[ROWS * COLS];
 
-        auto loadc = [&](int idx) {
+        auto loadc = [&](auto idx) {
             vc[idx] = _mm512_setzero_ps();
         };
         Unroll<ROWS * COLS>{}(loadc);
 
-        auto compute = [&](int idx, int k) {
+        auto compute = [&](auto idx, auto k) {
-            // TODO: use `constexpr` here to get rid of interger div
+            constexpr int row = idx / COLS;
-            // when upgraded to C++17
+            constexpr int col = idx % COLS;
-            const int row = idx / COLS;
-            const int col = idx % COLS;
 
-            if (col == 0) {
+            if constexpr (col == 0) {
                 va = _mm512_loadu_ps(A + row * K + k);
             }
-            if (row == 0) {
+            if constexpr (row == 0) {
                 vb[col] =  _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(B + col * K + k)));
             }
             vc[idx] = _mm512_fmadd_ps(va, vb[col], vc[idx]);
@@ -1364,9 +1362,9 @@ struct tinygemm_kernel_avx<float, ggml_fp16_t, float, BLOCK_M, BLOCK_N, BLOCK_K>
             Unroll<ROWS * COLS>{}(compute, k);
         }
 
-        auto storec = [&](int idx) {
+        auto storec = [&](auto idx) {
-            const int row = idx / COLS;
+            constexpr int row = idx / COLS;
-            const int col = idx % COLS;
+            constexpr int col = idx % COLS;
             C[row * ldc + col] = _mm512_reduce_add_ps(vc[idx]);
         };
         Unroll<ROWS * COLS>{}(storec);
@@ -1429,14 +1427,14 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLO
         const __m512i off = _mm512_set1_epi8(8);
         const __m512i lowMask = _mm512_set1_epi8(0xF);
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
 
-        auto compute = [&](int col, int i) {
+        auto compute = [&](auto col, auto i) {
             // load a and compute compensation
-            if (col == 0) {
+            if constexpr (col == 0) {
                 const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
                 vcomp = _mm512_setzero_si512();
                 for (int k = 0; k < 8; ++k) {
@@ -1468,7 +1466,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLO
         }
 
         //store to C
-        auto storec = [&](int col) {
+        auto storec = [&](auto col) {
             _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
         };
         Unroll<COLS>{}(storec);
@@ -1492,14 +1490,14 @@ struct tinygemm_kernel_vnni<block_q8_1, block_q4_1, float, 1, BLOCK_N, BLOCK_K>
 
         const __m512i lowMask = _mm512_set1_epi8(0xF);
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
 
-        auto compute = [&](int col, int i) {
+        auto compute = [&](auto col, auto i) {
             // load a
-            if (col == 0) {
+            if constexpr (col == 0) {
                 const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
                 for (int k = 0; k < 8; ++k) {
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
@@ -1533,7 +1531,7 @@ struct tinygemm_kernel_vnni<block_q8_1, block_q4_1, float, 1, BLOCK_N, BLOCK_K>
         }
 
         //store to C
-        auto storec = [&](int col) {
+        auto storec = [&](auto col) {
             _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
         };
         Unroll<COLS>{}(storec);
@@ -1564,14 +1562,14 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q8_0, float, BLOCK_M, BLOCK_N, BLO
         //
         const __m512i off = _mm512_set1_epi8(static_cast<char>(0x80));
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
 
-        auto compute = [&](int col, int i) {
+        auto compute = [&](auto col, auto i) {
             // load a and add offset 128
-            if (col == 0) {
+            if constexpr (col == 0) {
                 const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
                 for (int k = 0; k < 8; ++k) {
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
@@ -1604,7 +1602,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q8_0, float, BLOCK_M, BLOCK_N, BLO
         }
 
         //store to C
-        auto storec = [&](int col) {
+        auto storec = [&](auto col) {
             _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
         };
         Unroll<COLS>{}(storec);
@@ -1636,7 +1634,7 @@ struct tinygemm_kernel_vnni<block_q8_K, block_q4_K, float, BLOCK_M, BLOCK_N, BLO
 
         const __m512i lowMask = _mm512_set1_epi8(0xF);
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
@@ -1650,9 +1648,9 @@ struct tinygemm_kernel_vnni<block_q8_K, block_q4_K, float, BLOCK_M, BLOCK_N, BLO
         //     int16 {k/2, n, 2}, viewed as 2d {k/2, 2n}, k = 8
         //     from {16,  8} to {4, 32}
         //
-        auto compute = [&](int col, int i) {
+        auto compute = [&](auto col, auto i) {
             // load a
-            if (col == 0) {
+            if constexpr (col == 0) {
                 for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
                     va[k_group] = _mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)(A[0 * KB + i].qs + k_group * 32)));
                 }
@@ -1704,7 +1702,7 @@ struct tinygemm_kernel_vnni<block_q8_K, block_q4_K, float, BLOCK_M, BLOCK_N, BLO
         }
 
         //store to C
-        auto storec = [&](int col) {
+        auto storec = [&](auto col) {
             _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
         };
         Unroll<COLS>{}(storec);
@@ -1737,15 +1735,15 @@ struct tinygemm_kernel_vnni<block_q8_K, block_q5_K, float, BLOCK_M, BLOCK_N, BLO
 
         const __m512i lowMask = _mm512_set1_epi8(0xF);
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
 
         // Q5_K and Q4_K shares the same vnni formats, refer to notes above.
-        auto compute = [&](int col, int i) {
+        auto compute = [&](auto col, auto i) {
             // load a
-            if (col == 0) {
+            if constexpr (col == 0) {
                 for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
                     va[k_group] = _mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)(A[0 * KB + i].qs + k_group * 32)));
                 }
@@ -1810,7 +1808,7 @@ struct tinygemm_kernel_vnni<block_q8_K, block_q5_K, float, BLOCK_M, BLOCK_N, BLO
         }
 
         //store to C
-        auto storec = [&](int col) {
+        auto storec = [&](auto col) {
             _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
         };
         Unroll<COLS>{}(storec);
@@ -1843,13 +1841,13 @@ struct tinygemm_kernel_vnni<block_q8_K, block_q6_K, float, BLOCK_M, BLOCK_N, BLO
         const __m512i m32s = _mm512_set1_epi32(32);
         const __m512i lowMask = _mm512_set1_epi8(0xF);
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
 
-        auto compute = [&](int col, int i) {
+        auto compute = [&](auto col, auto i) {
-            if (col == 0) {
+            if constexpr (col == 0) {
                 // load a
                 va[0] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs +   0));
                 va[1] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs +  64));
@@ -1961,13 +1959,13 @@ struct tinygemm_kernel_vnni<block_q8_K, block_iq4_xs, float, BLOCK_M, BLOCK_N, B
         const __m512i off = _mm512_set1_epi8(static_cast<char>(0x80));
         const __m512i values256 = _mm512_add_epi8(values128, off);
 
-        auto loadc = [&](int col) {
+        auto loadc = [&](auto col) {
             vc[col] = _mm512_setzero_ps();
         };
         Unroll<COLS>{}(loadc);
 
-        auto compute = [&](int col, int i) {
+        auto compute = [&](auto col, auto i) {
-            if (col == 0) {
+            if constexpr (col == 0) {
                 // load a
                 va[0] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs +   0));
                 va[1] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs +  64));
@@ -2017,7 +2015,7 @@ struct tinygemm_kernel_vnni<block_q8_K, block_iq4_xs, float, BLOCK_M, BLOCK_N, B
         }
 
         //store to C
-        auto storec = [&](int col) {
+        auto storec = [&](auto col) {
             _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
         };
         Unroll<COLS>{}(storec);

ggml/src/ggml-cpu/cpu-feats-x86.cpp

@@ -0,0 +1,298 @@
+#include "ggml-cpu.h"
+#include "ggml-backend-impl.h"
+
+#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))
+
+#ifdef _MSC_VER
+#include <intrin.h>
+#endif
+
+#include <cstring>
+#include <vector>
+#include <bitset>
+#include <array>
+#include <string>
+
+struct cpuid_x86 {
+    bool SSE3(void) { return f_1_ecx[0]; }
+    bool PCLMULQDQ(void) { return f_1_ecx[1]; }
+    bool MONITOR(void) { return f_1_ecx[3]; }
+    bool SSSE3(void) { return f_1_ecx[9]; }
+    bool FMA(void) { return f_1_ecx[12]; }
+    bool CMPXCHG16B(void) { return f_1_ecx[13]; }
+    bool SSE41(void) { return f_1_ecx[19]; }
+    bool SSE42(void) { return f_1_ecx[20]; }
+    bool MOVBE(void) { return f_1_ecx[22]; }
+    bool POPCNT(void) { return f_1_ecx[23]; }
+    bool AES(void) { return f_1_ecx[25]; }
+    bool XSAVE(void) { return f_1_ecx[26]; }
+    bool OSXSAVE(void) { return f_1_ecx[27]; }
+    bool AVX(void) { return f_1_ecx[28]; }
+    bool F16C(void) { return f_1_ecx[29]; }
+    bool RDRAND(void) { return f_1_ecx[30]; }
+
+    bool MSR(void) { return f_1_edx[5]; }
+    bool CX8(void) { return f_1_edx[8]; }
+    bool SEP(void) { return f_1_edx[11]; }
+    bool CMOV(void) { return f_1_edx[15]; }
+    bool CLFSH(void) { return f_1_edx[19]; }
+    bool MMX(void) { return f_1_edx[23]; }
+    bool FXSR(void) { return f_1_edx[24]; }
+    bool SSE(void) { return f_1_edx[25]; }
+    bool SSE2(void) { return f_1_edx[26]; }
+
+    bool FSGSBASE(void) { return f_7_ebx[0]; }
+    bool BMI1(void) { return f_7_ebx[3]; }
+    bool HLE(void) { return is_intel && f_7_ebx[4]; }
+    bool AVX2(void) { return f_7_ebx[5]; }
+    bool BMI2(void) { return f_7_ebx[8]; }
+    bool ERMS(void) { return f_7_ebx[9]; }
+    bool INVPCID(void) { return f_7_ebx[10]; }
+    bool RTM(void) { return is_intel && f_7_ebx[11]; }
+    bool AVX512F(void) { return f_7_ebx[16]; }
+    bool RDSEED(void) { return f_7_ebx[18]; }
+    bool ADX(void) { return f_7_ebx[19]; }
+    bool AVX512PF(void) { return f_7_ebx[26]; }
+    bool AVX512ER(void) { return f_7_ebx[27]; }
+    bool AVX512CD(void) { return f_7_ebx[28]; }
+    bool SHA(void) { return f_7_ebx[29]; }
+
+    bool PREFETCHWT1(void) { return f_7_ecx[0]; }
+
+    bool LAHF(void) { return f_81_ecx[0]; }
+    bool LZCNT(void) { return is_intel && f_81_ecx[5]; }
+    bool ABM(void) { return is_amd && f_81_ecx[5]; }
+    bool SSE4a(void) { return is_amd && f_81_ecx[6]; }
+    bool XOP(void) { return is_amd && f_81_ecx[11]; }
+    bool TBM(void) { return is_amd && f_81_ecx[21]; }
+
+    bool SYSCALL(void) { return is_intel && f_81_edx[11]; }
+    bool MMXEXT(void) { return is_amd && f_81_edx[22]; }
+    bool RDTSCP(void) { return is_intel && f_81_edx[27]; }
+    bool _3DNOWEXT(void) { return is_amd && f_81_edx[30]; }
+    bool _3DNOW(void) { return is_amd && f_81_edx[31]; }
+
+    bool AVX512_VBMI(void) { return f_7_ecx[1]; }
+    bool AVX512_VNNI(void) { return f_7_ecx[11]; }
+    bool AVX512_FP16(void) { return f_7_edx[23]; }
+    bool AVX512_BF16(void) { return f_7_1_eax[5]; }
+    bool AVX_VNNI(void) { return f_7_1_eax[4]; }
+
+    bool AMX_TILE(void) { return f_7_edx[24]; }
+    bool AMX_INT8(void) { return f_7_edx[25]; }
+    bool AMX_FP16(void) { return f_7_1_eax[21]; }
+    bool AMX_BF16(void) { return f_7_edx[22]; }
+
+#ifdef _MSC_VER
+    static void cpuid(int cpu_info[4], int eax) {
+        __cpuid(cpu_info, eax);
+    }
+    static void cpuidex(int cpu_info[4], int eax, int ecx) {
+        __cpuidex(cpu_info, eax, ecx);
+    }
+#else
+    static void cpuid(int cpu_info[4], int eax) {
+        __asm__ __volatile__(
+            "cpuid"
+            : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
+            : "a"(eax), "c"(0));
+    }
+    static void cpuidex(int cpu_info[4], int eax, int ecx) {
+        __asm__ __volatile__(
+            "cpuid"
+            : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
+            : "a"(eax), "c"(ecx));
+    }
+#endif
+
+    cpuid_x86() {
+        std::array<int, 4> cpui;
+        std::vector<std::array<int, 4>> data;
+
+        // calling __cpuid with 0x0 as the function_id argument
+        // gets the number of the highest valid function ID.
+        cpuid(cpui.data(), 0);
+        int n_ids = cpui[0];
+
+        for (int i = 0; i <= n_ids; ++i) {
+            cpuidex(cpui.data(), i, 0);
+            data.push_back(cpui);
+        }
+
+        // capture vendor string
+        char vendor[0x20] = {};
+        *reinterpret_cast<int *>(vendor)     = data[0][1];
+        *reinterpret_cast<int *>(vendor + 4) = data[0][3];
+        *reinterpret_cast<int *>(vendor + 8) = data[0][2];
+        this->vendor = vendor;
+        if (this->vendor == "GenuineIntel") {
+            is_intel = true;
+        } else if (this->vendor == "AuthenticAMD") {
+            is_amd = true;
+        }
+
+        // load bitset with flags for function 0x00000001
+        if (n_ids >= 1) {
+            f_1_ecx = data[1][2];
+            f_1_edx = data[1][3];
+        }
+
+        // load bitset with flags for function 0x00000007
+        if (n_ids >= 7) {
+            f_7_ebx = data[7][1];
+            f_7_ecx = data[7][2];
+            f_7_edx = data[7][3];
+            cpuidex(cpui.data(), 7, 1);
+            f_7_1_eax = cpui[0];
+        }
+
+        // calling __cpuid with 0x80000000 as the function_id argument
+        // gets the number of the highest valid extended ID.
+        cpuid(cpui.data(), 0x80000000);
+        unsigned int n_ex_ids = cpui[0];
+
+        std::vector<std::array<int, 4>> ext_data;
+        for (unsigned int i = 0x80000000; i <= n_ex_ids; ++i) {
+            cpuidex(cpui.data(), i, 0);
+            ext_data.push_back(cpui);
+        }
+
+        // load bitset with flags for function 0x80000001
+        if (n_ex_ids >= 0x80000001) {
+            f_81_ecx = ext_data[1][2];
+            f_81_edx = ext_data[1][3];
+        }
+
+        // interpret CPU brand string if reported
+        char brand[0x40] = {};
+        if (n_ex_ids >= 0x80000004) {
+            std::memcpy(brand, ext_data[2].data(), sizeof(cpui));
+            std::memcpy(brand + 16, ext_data[3].data(), sizeof(cpui));
+            std::memcpy(brand + 32, ext_data[4].data(), sizeof(cpui));
+            this->brand = brand;
+        }
+    }
+
+    bool is_intel = false;
+    bool is_amd = false;
+    std::string vendor;
+    std::string brand;
+    std::bitset<32> f_1_ecx;
+    std::bitset<32> f_1_edx;
+    std::bitset<32> f_7_ebx;
+    std::bitset<32> f_7_ecx;
+    std::bitset<32> f_7_edx;
+    std::bitset<32> f_7_1_eax;
+    std::bitset<32> f_81_ecx;
+    std::bitset<32> f_81_edx;
+};
+
+#if 0
+void test_x86_is() {
+    cpuid_x86 is;
+    printf("CPU Vendor: %s\n", is.vendor.c_str());
+    printf("Brand: %s\n", is.brand.c_str());
+    printf("is_intel: %d\n", is.is_intel);
+    printf("is_amd: %d\n", is.is_amd);
+    printf("sse3: %d\n", is.SSE3());
+    printf("pclmulqdq: %d\n", is.PCLMULQDQ());
+    printf("ssse3: %d\n", is.SSSE3());
+    printf("fma: %d\n", is.FMA());
+    printf("cmpxchg16b: %d\n", is.CMPXCHG16B());
+    printf("sse41: %d\n", is.SSE41());
+    printf("sse42: %d\n", is.SSE42());
+    printf("movbe: %d\n", is.MOVBE());
+    printf("popcnt: %d\n", is.POPCNT());
+    printf("aes: %d\n", is.AES());
+    printf("xsave: %d\n", is.XSAVE());
+    printf("osxsave: %d\n", is.OSXSAVE());
+    printf("avx: %d\n", is.AVX());
+    printf("f16c: %d\n", is.F16C());
+    printf("rdrand: %d\n", is.RDRAND());
+    printf("msr: %d\n", is.MSR());
+    printf("cx8: %d\n", is.CX8());
+    printf("sep: %d\n", is.SEP());
+    printf("cmov: %d\n", is.CMOV());
+    printf("clflush: %d\n", is.CLFSH());
+    printf("mmx: %d\n", is.MMX());
+    printf("fxsr: %d\n", is.FXSR());
+    printf("sse: %d\n", is.SSE());
+    printf("sse2: %d\n", is.SSE2());
+    printf("fsgsbase: %d\n", is.FSGSBASE());
+    printf("bmi1: %d\n", is.BMI1());
+    printf("hle: %d\n", is.HLE());
+    printf("avx2: %d\n", is.AVX2());
+    printf("bmi2: %d\n", is.BMI2());
+    printf("erms: %d\n", is.ERMS());
+    printf("invpcid: %d\n", is.INVPCID());
+    printf("rtm: %d\n", is.RTM());
+    printf("avx512f: %d\n", is.AVX512F());
+    printf("rdseed: %d\n", is.RDSEED());
+    printf("adx: %d\n", is.ADX());
+    printf("avx512pf: %d\n", is.AVX512PF());
+    printf("avx512er: %d\n", is.AVX512ER());
+    printf("avx512cd: %d\n", is.AVX512CD());
+    printf("sha: %d\n", is.SHA());
+    printf("prefetchwt1: %d\n", is.PREFETCHWT1());
+    printf("lahf: %d\n", is.LAHF());
+    printf("lzcnt: %d\n", is.LZCNT());
+    printf("abm: %d\n", is.ABM());
+    printf("sse4a: %d\n", is.SSE4a());
+    printf("xop: %d\n", is.XOP());
+    printf("tbm: %d\n", is.TBM());
+    printf("syscall: %d\n", is.SYSCALL());
+    printf("mmxext: %d\n", is.MMXEXT());
+    printf("rdtscp: %d\n", is.RDTSCP());
+    printf("3dnowext: %d\n", is._3DNOWEXT());
+    printf("3dnow: %d\n", is._3DNOW());
+    printf("avx512_vbmi: %d\n", is.AVX512_VBMI());
+    printf("avx512_vnni: %d\n", is.AVX512_VNNI());
+    printf("avx512_fp16: %d\n", is.AVX512_FP16());
+    printf("avx512_bf16: %d\n", is.AVX512_BF16());
+    printf("amx_tile: %d\n", is.AMX_TILE());
+    printf("amx_int8: %d\n", is.AMX_INT8());
+    printf("amx_fp16: %d\n", is.AMX_FP16());
+    printf("amx_bf16: %d\n", is.AMX_BF16());
+}
+#endif
+
+static int ggml_backend_cpu_x86_score() {
+    // FIXME: this does not check for OS support
+
+    cpuid_x86 is;
+    // if the CPU backend was built with any features not supported by the current CPU, it cannot be used
+    if (ggml_cpu_has_fma() && !is.FMA()) { return 0; }
+    if (ggml_cpu_has_f16c() && !is.F16C()) { return 0; }
+    if (ggml_cpu_has_ssse3() && !is.SSSE3()) { return 0; }
+    if (ggml_cpu_has_sse3() && !is.SSE3()) { return 0; }
+    if (ggml_cpu_has_avx() && !is.AVX()) { return 0; }
+    if (ggml_cpu_has_avx_vnni() && !is.AVX_VNNI()) { return 0; }
+    if (ggml_cpu_has_avx2() && !is.AVX2()) { return 0; }
+    if (ggml_cpu_has_avx512() && !is.AVX512F()) { return 0; }
+    if (ggml_cpu_has_avx512_vbmi() && !is.AVX512_VBMI()) { return 0; }
+    if (ggml_cpu_has_avx512_bf16() && !is.AVX512_BF16()) { return 0; }
+    if (ggml_cpu_has_avx512_vnni() && !is.AVX512_VNNI()) { return 0; }
+    if (ggml_cpu_has_amx_int8() && !is.AMX_INT8()) { return 0; }
+
+    // calculate a backend score based on the supported features
+    // more important features have a higher weight
+    int score = 0;
+    score +=  ggml_cpu_has_fma        () * 1;
+    score +=  ggml_cpu_has_f16c       () * 1<<1;
+    score +=  ggml_cpu_has_ssse3      () * 1<<2;
+    score +=  ggml_cpu_has_sse3       () * 1<<3;
+    score +=  ggml_cpu_has_avx_vnni   () * 1<<4;
+    score +=  ggml_cpu_has_avx        () * 1<<5;
+    score +=  ggml_cpu_has_avx2       () * 1<<6;
+    score +=  ggml_cpu_has_avx512     () * 1<<7;
+    // score +=  ggml_cpu_has_avx512_vbmi() * 1<<8; // not used
+    score +=  ggml_cpu_has_avx512_bf16() * 1<<9;
+    score +=  ggml_cpu_has_avx512_vnni() * 1<<10;
+    score +=  ggml_cpu_has_amx_int8   () * 1<<11;
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_x86_score)
+
+#endif // defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))

ggml/src/ggml-cpu/ggml-cpu-aarch64.c

@@ -128,7 +128,7 @@ static inline __m512i sum_i16_pairs_int_32x16(const __m512i x) {
 }
 
 static inline __m512i mul_sum_us8_pairs_int32x16(const __m512i ax, const __m512i sy) {
-#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+#if defined(__AVX512VNNI__)
     const __m512i zero = _mm512_setzero_si512();
     return _mm512_dpbusd_epi32(zero, ax, sy);
 #else

scripts/build-cpu.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+
+name="$1"
+args="${@:2}"
+
+echo "Building $name with args: $args"
+
+rm -fr build-cpu-$1
+cmake -S . -B build-cpu-$1 -DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF $args
+cmake --build build-cpu-$1 --config Release -t ggml-cpu -j $(nproc)
+cp build-cpu-$1/bin/libggml-cpu.so ./libggml-cpu-$1.so
+rm -fr build-cpu-$1