ggml-cpu: Implement GGML_CPU_ALL_VARIANTS for ARM

Like x86, however to pass around arch flags within cmake, we use
GGML_INTERNAL_<FEAT> as we don't have GGML_<FEAT>.

Some features are optional, so we may need to build multiple backends
per arch version (armv8.2_1, armv8.2_2, ...), and let the scoring
function sort out which one can be used.

Author: ckastner

ggml/src/CMakeLists.txt

@@ -269,17 +269,23 @@ endfunction()
 function(ggml_add_cpu_backend_variant tag_name)
     set(GGML_CPU_TAG_NAME ${tag_name})
     # other: OPENMP LLAMAFILE CPU_HBM
-    foreach (feat NATIVE
-                  SSE42
-                  AVX AVX2 BMI2 AVX_VNNI FMA F16C
-                  AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
-                  AMX_TILE AMX_INT8 AMX_BF16)
-        set(GGML_${feat} OFF)
-    endforeach()
-
-    foreach (feat ${ARGN})
-        set(GGML_${feat} ON)
-    endforeach()
+    if (GGML_SYSTEM_ARCH STREQUAL "x86")
+        foreach (feat NATIVE
+                      SSE42
+                      AVX AVX2 BMI2 AVX_VNNI FMA F16C
+                      AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
+                      AMX_TILE AMX_INT8 AMX_BF16)
+            set(GGML_${feat} OFF)
+        endforeach()
+
+        foreach (feat ${ARGN})
+            set(GGML_${feat} ON)
+        endforeach()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "ARM")
+        foreach (feat ${ARGN})
+            set(GGML_INTERNAL_${feat} ON)
+        endforeach()
+    endif()
 
     ggml_add_cpu_backend_variant_impl(${tag_name})
 endfunction()
@@ -302,6 +308,21 @@ if (GGML_CPU_ALL_VARIANTS)
             # MSVC doesn't support AMX
             ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
         endif()
+    elseif(GGML_SYSTEM_ARCH STREQUAL "ARM")
+        if (GGML_CPU_ARM_ARCH)
+            message(FATAL_ERROR "Cannot use both GGML_CPU_ARM_ARCH and GGML_CPU_ALL_VARIANTS")
+        endif()
+        # Many of these features are optional so we build versions with popular
+        # combinations and name the backends based on the version they were
+        # first released with
+        ggml_add_cpu_backend_variant(armv8.0_1)
+        ggml_add_cpu_backend_variant(armv8.2_1    DOTPROD)
+        ggml_add_cpu_backend_variant(armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
+        ggml_add_cpu_backend_variant(armv8.2_3    DOTPROD FP16_VECTOR_ARITHMETIC SVE)
+        ggml_add_cpu_backend_variant(armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8)
+        ggml_add_cpu_backend_variant(armv8.6_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2)
+        ggml_add_cpu_backend_variant(armv9.2_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SME)
+        ggml_add_cpu_backend_variant(armv9.2_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2 SME)
     else()
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported on ${GGML_SYSTEM_ARCH}")
     endif()

ggml/src/ggml-cpu/CMakeLists.txt

@@ -152,6 +152,47 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             else()
                 if (GGML_CPU_ARM_ARCH)
                     list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
+                else()
+                    # Begin with the lowest baseline
+                    set(ARM_MCPU "armv8-a")
+                    set(ARCH_TAGS "")
+            set(ARCH_DEFINITIONS "")
+
+                    # When a feature is selected, bump the MCPU to the first
+                    # version that supported it
+                    if (GGML_INTERNAL_DOTPROD)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+dotprod")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD)
+                    endif()
+                    if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+fp16")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC)
+                    endif()
+                    if (GGML_INTERNAL_SVE)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sve")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE)
+                    endif()
+                    if (GGML_INTERNAL_MATMUL_INT8)
+                        set(ARM_MCPU "armv8.6-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+i8mm")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8)
+                    endif()
+                    if (GGML_INTERNAL_SVE2)
+                        set(ARM_MCPU "armv8.6-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sve2")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
+                    endif()
+                    if (GGML_INTERNAL_SME)
+                        set(ARM_MCPU "armv9.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sme")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SME)
+                    endif()
+
+                    list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}")
+                    ggml_add_cpu_backend_features(${GGML_CPU_NAME} aarch64 ${ARCH_DEFINITIONS})
                 endif()
             endif()