fix: Avoid SVE codegen in SME wrapper

BUILD.bazel

@@ -363,6 +363,64 @@ cc_library(
     alwayslink = True,
 )
 
+#---------------------------------------------------------------------
+# SVE library sources that must not be compiler-vectorized
+#
+# SME does not imply non-streaming SVE. These sources still need the SVE target
+# for explicit Streaming-SVE/SME code, but ordinary C++ must not gain an
+# unintended non-streaming SVE dependency.
+
+cc_library(
+    name = "arm_compute_sve_no_vectorize",
+    srcs = ["//src:arm_compute_sve_no_vectorize_srcs"],
+    copts = [
+                "-march=armv8.2-a+sve+fp16+dotprod",
+            ] + select({
+                "//:debug_flag": [
+                    "-O0",
+                    "-g",
+                    "-gdwarf-2",
+                ],
+                "//conditions:default": ["-O3"],
+            }) +
+            [
+                "-fno-tree-vectorize",
+                "-fno-tree-slp-vectorize",
+            ] +
+            select({
+                "//:openmp_flag": ["-fopenmp"],
+                "//conditions:default": [],
+            }) +
+            select({
+                "//:Werror_flag": ["-Werror"],
+                "//conditions:default": [],
+            }),
+    includes = [
+        "src/core/NEON/kernels/arm_conv",
+        "src/core/NEON/kernels/arm_gemm",
+        "src/core/NEON/kernels/assembly",
+        "src/core/cpu/kernels/assembly",
+        "src/cpu/kernels/assembly",
+    ],
+    linkopts = select({
+        "//:openmp_flag": ["-fopenmp"],
+        "//conditions:default": [],
+    }),
+    local_defines = [
+        "ENABLE_SVE",
+        "ARM_COMPUTE_ENABLE_SVE",
+        "ARM_COMPUTE_ENABLE_BF16",
+    ],
+    deps = [
+        "//:common_defines",
+        "//arm_compute:core_headers",
+        "//arm_compute:runtime_headers",
+        "//include",
+        "//support",
+    ],
+    alwayslink = True,
+)
+
 #---------------------------------------------------------------------
 # Core and Runtime library
 
@@ -426,6 +484,7 @@ cc_library(
         "@@kleidiai//:common",
         "@@kleidiai//kai/ukernels/matmul:matmul",
         "//:arm_compute_sve",
+        "//:arm_compute_sve_no_vectorize",
         "//:arm_compute_sve2"
     ],
     alwayslink = True,

CMakeLists.txt

@@ -136,6 +136,11 @@ else()
   message(STATUS "Using arch: ${ARM_COMPUTE_ARCH}")
 endif()
 
+# SME does not imply non-streaming SVE. SME-only wrappers built in an
+# SVE-capable target must disable compiler vectorization so ordinary C++ cannot
+# acquire an unintended non-streaming SVE dependency.
+set(ARM_COMPUTE_SME_WRAPPER_NO_VECTORIZE_FLAGS "-fno-tree-vectorize;-fno-tree-slp-vectorize")
+
 if(ARM_COMPUTE_ENABLE_OPENMP)
   find_package(OpenMP REQUIRED)
 endif()
@@ -204,6 +209,20 @@ else()
   add_library(arm_compute_sve OBJECT EXCLUDE_FROM_ALL)
 endif()
 
+if(ACL_MULTI_ISA OR ACL_BUILD_SVE)
+  add_library(arm_compute_sve_no_vectorize OBJECT)
+  set_target_properties(
+    arm_compute_sve_no_vectorize
+    PROPERTIES
+      COMPILE_OPTIONS "${ARM_COMPUTE_SVE_ARCH};${ARM_COMPUTE_COMMON_CCXX_FLAGS};${ARM_COMPUTE_SME_WRAPPER_NO_VECTORIZE_FLAGS}"
+      COMPILE_DEFINITIONS "${ARM_COMPUTE_DEFINES}"
+      INCLUDE_DIRECTORIES "${ARM_COMPUTE_SVE_COMMON_INCLUDE}"
+      LINK_LIBRARIES "${ARM_COMPUTE_LINK_LIBS}"
+  )
+else()
+  add_library(arm_compute_sve_no_vectorize OBJECT EXCLUDE_FROM_ALL)
+endif()
+
 if(ACL_MULTI_ISA OR ACL_BUILD_SVE2)
   add_library(arm_compute_sve2 OBJECT)
   set_target_properties(
@@ -257,6 +276,7 @@ if(ACL_MULTI_ISA)
     $<TARGET_OBJECTS:arm_compute_core>
     $<TARGET_OBJECTS:arm_compute_core_fp16>
     $<TARGET_OBJECTS:arm_compute_sve>
+    $<TARGET_OBJECTS:arm_compute_sve_no_vectorize>
     $<TARGET_OBJECTS:arm_compute_sve2>
   )
 else()
@@ -269,6 +289,7 @@ else()
     list(APPEND lib_objs $<TARGET_OBJECTS:arm_compute_core_fp16>)
     if(ACL_BUILD_SVE)
       list(APPEND lib_objs $<TARGET_OBJECTS:arm_compute_sve>)
+      list(APPEND lib_objs $<TARGET_OBJECTS:arm_compute_sve_no_vectorize>)
     endif()
     if(ACL_BUILD_SVE2)
       list(APPEND lib_objs $<TARGET_OBJECTS:arm_compute_sve2>)
@@ -278,6 +299,7 @@ else()
     list(APPEND lib_objs $<TARGET_OBJECTS:arm_compute_core_fp16>)
     if(ACL_BUILD_SVE)
       list(APPEND lib_objs $<TARGET_OBJECTS:arm_compute_sve>)
+      list(APPEND lib_objs $<TARGET_OBJECTS:arm_compute_sve_no_vectorize>)
     endif()
     if(ACL_BUILD_SVE2)
       list(APPEND lib_objs $<TARGET_OBJECTS:arm_compute_sve2>)
@@ -296,6 +318,9 @@ if(CMAKE_SYSTEM_NAME STREQUAL "Darwin")
   if(TARGET arm_compute_core_fp16)
     target_compile_options(arm_compute_core_fp16 PRIVATE -arch arm64)
   endif()
+  if(TARGET arm_compute_sve_no_vectorize)
+    target_compile_options(arm_compute_sve_no_vectorize PRIVATE -arch arm64)
+  endif()
 endif()
 
 
@@ -319,6 +344,9 @@ endif()
 if(TARGET arm_compute_sve)
   list(APPEND ARM_COMPUTE_TARGETS arm_compute_sve)
 endif()
+if(TARGET arm_compute_sve_no_vectorize)
+  list(APPEND ARM_COMPUTE_TARGETS arm_compute_sve_no_vectorize)
+endif()
 if(TARGET arm_compute_sve2)
   list(APPEND ARM_COMPUTE_TARGETS arm_compute_sve2)
 endif()

SConscript

@@ -118,6 +118,8 @@ def build_multiisa_lib_objects():
     # Build the SVE specific files
     lib_static_objs += build_obj_list(filedefs["armv8.2-a-sve"], misa_lib_files_sve, static=True)
     lib_shared_objs += build_obj_list(filedefs["armv8.2-a-sve"], misa_lib_files_sve, static=False)
+    lib_static_objs += build_obj_list(filedefs["armv8.2-a-sve-no-vectorize"], misa_lib_files_sve_no_vectorize, static=True)
+    lib_shared_objs += build_obj_list(filedefs["armv8.2-a-sve-no-vectorize"], misa_lib_files_sve_no_vectorize, static=False)
     lib_static_objs += build_obj_list(filedefs["armv8.2-a-sve"], misa_lib_files_sve_fp16, static=True)
     lib_shared_objs += build_obj_list(filedefs["armv8.2-a-sve"], misa_lib_files_sve_fp16, static=False)
 
@@ -645,11 +647,16 @@ lib_files_sve2 = []
 
 misa_lib_files = lib_files
 misa_lib_files_sve = []
+misa_lib_files_sve_no_vectorize = []
 misa_lib_files_sve2 = []
 misa_lib_files_neon_fp16 = []
 misa_lib_files_sve_fp16 = []
 misa_lib_files_sve2_fp16 = []
 
+sve_no_vectorize_files = [
+    "src/core/NEON/kernels/arm_gemm/interleave_indirect-sve.cpp",
+]
+
 arm_compute_env.Append(CPPPATH = ["src/cpu/kernels/assembly/"])
 
 if env['neon']:
@@ -708,6 +715,12 @@ if env['neon']:
 
         # SVE files only minus FP16
         misa_lib_files_sve = cpu_files.get('sve', [])
+        # SME does not imply non-streaming SVE. This SME-only wrapper still
+        # needs the SVE assembler target for explicit Streaming-SVE/SME code,
+        # but ordinary compiler-generated C++ must not gain an SVE dependency.
+        sve_no_vectorize_set = set(sve_no_vectorize_files)
+        misa_lib_files_sve_no_vectorize = [f for f in misa_lib_files_sve if f in sve_no_vectorize_set]
+        misa_lib_files_sve = [f for f in misa_lib_files_sve if f not in sve_no_vectorize_set]
 
         # SVE2 files only minus FP16
         misa_lib_files_sve2 = cpu_files.get('sve2', [])

filedefs.json

@@ -13,6 +13,11 @@
                 "cppdefines": ["ARM_COMPUTE_ENABLE_FP16", "ARM_COMPUTE_ENABLE_BF16",
                                "ARM_COMPUTE_ENABLE_I8MM", "ARM_COMPUTE_ENABLE_SVEF32MM"]
             },
+            "armv8.2-a-sve-no-vectorize": {
+                "ccflags": ["-march=armv8.2-a+sve+fp16+dotprod", "-fno-tree-vectorize", "-fno-tree-slp-vectorize"],
+                "cppdefines": ["ARM_COMPUTE_ENABLE_FP16", "ARM_COMPUTE_ENABLE_BF16",
+                               "ARM_COMPUTE_ENABLE_I8MM", "ARM_COMPUTE_ENABLE_SVEF32MM"]
+            },
             "armv8.2-a-sve2": {
                 "ccflags": ["-march=armv8.2-a+sve2+fp16+dotprod"],
                 "cppdefines": ["ARM_COMPUTE_ENABLE_FP16", "ARM_COMPUTE_ENABLE_BF16",

scripts/generate_build_files.py

@@ -36,6 +36,10 @@
 import json
 import glob
 
+SVE_NO_VECTORIZE_FILES = [
+    "src/core/NEON/kernels/arm_gemm/interleave_indirect-sve.cpp",
+]
+
 
 def get_operator_backend_files(filelist, operators, backend='', techs=[], attrs=[], include_common=True):
     files = {"common": []}
@@ -117,7 +121,7 @@ def get_template_header():
 # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 # SOFTWARE."""
 
-def build_from_template_bazel(srcs_graph, srcs_sve, srcs_sve2, _srcs_core):
+def build_from_template_bazel(srcs_graph, srcs_sve, srcs_sve_no_vectorize, srcs_sve2, _srcs_core):
     # Bazel does not support targets referencing upper-levels.
     srcs_core = [path for path in _srcs_core if not path.startswith("../")]
 
@@ -152,6 +156,15 @@ def build_from_template_bazel(srcs_graph, srcs_sve, srcs_sve2, _srcs_core):
 		visibility = ["//visibility:public"]
 )
 
+filegroup(
+        name = "arm_compute_sve_no_vectorize_srcs",
+        srcs = ["{line_separator.join(srcs_sve_no_vectorize)}"]  +
+    glob(["**/*.h",
+    "**/*.hpp",
+    "**/*.inl"]),
+		visibility = ["//visibility:public"]
+)
+
 filegroup(
         name = "arm_compute_srcs",
         srcs = ["{line_separator.join(srcs_core)}"]  +
@@ -165,7 +178,7 @@ def build_from_template_bazel(srcs_graph, srcs_sve, srcs_sve2, _srcs_core):
     return template
 
 
-def build_from_template_cmake(srcs_graph, srcs_sve, srcs_sve2, srcs_core, srcs_core_fp16):
+def build_from_template_cmake(srcs_graph, srcs_sve, srcs_sve_no_vectorize, srcs_sve2, srcs_core, srcs_core_fp16):
 
     line_separator = '\n\t'
 
@@ -183,6 +196,12 @@ def build_from_template_cmake(srcs_graph, srcs_sve, srcs_sve2, srcs_core, srcs_c
     {line_separator.join(srcs_sve)}
 )
 
+target_sources(
+    arm_compute_sve_no_vectorize
+    PRIVATE
+    {line_separator.join(srcs_sve_no_vectorize)}
+)
+
 target_sources(
     arm_compute_sve2
     PRIVATE
@@ -266,6 +285,8 @@ def gather_sources():
     # SVE files only
     lib_files_sve = cpu_files.get('sve', [])
     lib_files_sve += fp16_cpu_files.get('sve', [])
+    lib_files_sve_no_vectorize = [path for path in lib_files_sve if path in SVE_NO_VECTORIZE_FILES]
+    lib_files_sve = [path for path in lib_files_sve if path not in SVE_NO_VECTORIZE_FILES]
 
     # SVE2 files only
     lib_files_sve2 = cpu_files.get('sve2', [])
@@ -280,11 +301,12 @@ def strip_prefix(filename, prefix = "src/"):
 
     graph_files = sorted([strip_prefix(path, "src/") for path in graph_files])
     lib_files_sve = sorted([strip_prefix(path, "src/") for path in lib_files_sve])
+    lib_files_sve_no_vectorize = sorted([strip_prefix(path, "src/") for path in lib_files_sve_no_vectorize])
     lib_files_sve2 = sorted([strip_prefix(path, "src/") for path in lib_files_sve2])
     lib_files = sorted([strip_prefix(path, "src/") for path in lib_files])
     lib_files_neon_fp16 = sorted([strip_prefix(path, "src/") for path in lib_files_neon_fp16])
 
-    return (graph_files, lib_files_sve, lib_files_sve2, lib_files, lib_files_neon_fp16)
+    return (graph_files, lib_files_sve, lib_files_sve_no_vectorize, lib_files_sve2, lib_files, lib_files_neon_fp16)
 
 
 if "__main__" in __name__:
@@ -294,20 +316,20 @@ def strip_prefix(filename, prefix = "src/"):
     parser.add_argument("--cmake", action="store_true")
     args = parser.parse_args()
 
-    (graph_files, lib_files_sve, lib_files_sve2, lib_files, lib_files_neon_fp16) = gather_sources()
+    (graph_files, lib_files_sve, lib_files_sve_no_vectorize, lib_files_sve2, lib_files, lib_files_neon_fp16) = gather_sources()
 
     if args.bazel:
         # 8562a4ec: Remove CommonGraphOptions from Utils target and warnings
         graph_files += ["//utils:CommonGraphOptions.cpp"]
 
         bazel_build_string = build_from_template_bazel(
-            graph_files, lib_files_sve, lib_files_sve2, lib_files + lib_files_neon_fp16)
+            graph_files, lib_files_sve, lib_files_sve_no_vectorize, lib_files_sve2, lib_files + lib_files_neon_fp16)
         with open("src/BUILD.bazel", "w") as fp:
             fp.write(bazel_build_string)
 
     if args.cmake:
         cmake_build_string = build_from_template_cmake(
-            graph_files, lib_files_sve, lib_files_sve2, lib_files, lib_files_neon_fp16)
+            graph_files, lib_files_sve, lib_files_sve_no_vectorize, lib_files_sve2, lib_files, lib_files_neon_fp16)
         with open("src/CMakeLists.txt", "w") as fp:
             fp.write(cmake_build_string)
 

src/BUILD.bazel

@@ -250,7 +250,6 @@ filegroup(
 	"core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp",
 	"core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
 	"core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_max_generic_depthfirst/generic.cpp",
-	"core/NEON/kernels/arm_gemm/interleave_indirect-sve.cpp",
 	"core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp",
 	"core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp16fp32fp16_dot_16VL/generic.cpp",
 	"core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32_mla_16VL/generic.cpp",
@@ -407,6 +406,15 @@ filegroup(
 		visibility = ["//visibility:public"]
 )
 
+filegroup(
+        name = "arm_compute_sve_no_vectorize_srcs",
+        srcs = ["core/NEON/kernels/arm_gemm/interleave_indirect-sve.cpp"]  +
+    glob(["**/*.h",
+    "**/*.hpp",
+    "**/*.inl"]),
+		visibility = ["//visibility:public"]
+)
+
 filegroup(
         name = "arm_compute_srcs",
         srcs = ["c/AclContext.cpp",

src/CMakeLists.txt

@@ -219,7 +219,6 @@ target_sources(
 	core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp
 	core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_avg_generic_depthfirst/generic.cpp
 	core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_max_generic_depthfirst/generic.cpp
-	core/NEON/kernels/arm_gemm/interleave_indirect-sve.cpp
 	core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp
 	core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp16fp32fp16_dot_16VL/generic.cpp
 	core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32_mla_16VL/generic.cpp
@@ -372,6 +371,12 @@ target_sources(
 	cpu/kernels/topkv/generic/sve/qasymm8_signed.cpp
 )
 
+target_sources(
+    arm_compute_sve_no_vectorize
+    PRIVATE
+    core/NEON/kernels/arm_gemm/interleave_indirect-sve.cpp
+)
+
 target_sources(
     arm_compute_sve2
     PRIVATE