ggml : move AMX to the CPU backend (llama/10570)

ggml : automatic selection of best CPU backend (llama/10606)

Author: slaren

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)
@@ -161,7 +162,6 @@ set   (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
 set   (GGML_METAL_STD "" CACHE STRING       "ggml: metal standard version (-std flag)")
 option(GGML_OPENMP                          "ggml: use OpenMP"                                ON)
 option(GGML_RPC                             "ggml: use RPC"                                   OFF)
-option(GGML_AMX                             "ggml: use AMX"                                   OFF)
 option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING

ggml/src/CMakeLists.txt

@@ -261,21 +261,15 @@ function(ggml_add_backend backend)
     if (${backend_id})
         string(TOLOWER "ggml-${backend}" backend_target)
         add_subdirectory(${backend_target})
-        # check again in case the backend disabled itself
-        # note that this should NOT be the normal behavior, in case of errors the backend should fail the build
-        # however, currently it is necessary for AMX, since it is enabled by default on llama.cpp
-        if (${backend_id})
-            message(STATUS "Including ${backend} backend")
-            if (NOT GGML_BACKEND_DL)
-                string(TOUPPER "GGML_USE_${backend}" backend_use)
-                target_compile_definitions(ggml PUBLIC ${backend_use})
-            endif()
+        message(STATUS "Including ${backend} backend")
+        if (NOT GGML_BACKEND_DL)
+            string(TOUPPER "GGML_USE_${backend}" backend_use)
+            target_compile_definitions(ggml PUBLIC ${backend_use})
         endif()
     endif()
 endfunction()
 
 ggml_add_backend(CPU)
-ggml_add_backend(AMX)
 ggml_add_backend(BLAS)
 ggml_add_backend(CANN)
 ggml_add_backend(CUDA)
@@ -289,7 +283,7 @@ ggml_add_backend(Vulkan)
 
 foreach (target ggml-base ggml)
     target_include_directories(${target} PUBLIC    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)
-    target_compile_features   (${target} PRIVATE c_std_11) # don't bump
+    target_compile_features   (${target} PRIVATE c_std_11 cxx_std_17) # don't bump
 endforeach()
 
 target_link_libraries(ggml-base PRIVATE Threads::Threads)

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
-        #ifdef __cplusplus
-        #    define GGML_BACKEND_DL_IMPL(reg_fn)                                 \
-                extern "C" {                                                     \
-                    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
-                }                                                                \
-                ggml_backend_reg_t ggml_backend_init(void) {                     \
-                    return reg_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();                                         \
-                }
-        #endif
-    #else
-    #    define GGML_BACKEND_DL_IMPL(reg_fn)
-    #endif
+    // Initialize the backend
+    typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
+    // Optional: obtain a score for the backend based on the system configuration
+    // Higher scores are preferred, 0 means the backend is not supported in the current system
+    typedef int                (*ggml_backend_score_t)(void);
+
+#ifdef GGML_BACKEND_DL
+#    ifdef __cplusplus
+#        define GGML_BACKEND_DL_IMPL(reg_fn)                             \
+            extern "C" {                                                 \
+            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)       \
+            extern "C" {                                   \
+            GGML_BACKEND_API int ggml_backend_score(void); \
+            }                                              \
+            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
 }