fix i8mm

Author: cocoa-xu

.github/workflows/linux-precompile.yml

@@ -16,6 +16,7 @@ jobs:
     runs-on: ubuntu-20.04
     env:
       TFLITE_BEAM_PREFER_PRECOMPILED: NO
+      TFBEAM_XNNPACK_ENABLE_ARM_I8MM: "OFF"
     name: Linux ${{ matrix.pair.target_arch }}-${{ matrix.pair.target_os }}-${{ matrix.pair.target_abi }} - OTP ${{ matrix.pair.otp }}
 
     strategy:
@@ -116,7 +117,7 @@ jobs:
               cxx: "/usr/bin/riscv64-linux-gnu-g++"
 
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
 
       - uses: erlef/setup-beam@v1
         with:

CMakeLists.txt

@@ -69,8 +69,13 @@ if(DEFINED ENV{TARGET_ARCH})
         set(TFLITE_ENABLE_XNNPACK OFF CACHE BOOL "Disable XNNPACK in TensorFlow Lite for armv7l")
     elseif("$ENV{TARGET_ARCH}" MATCHES "arm")
         if(DEFINED ENV{TARGET_CPU})
-            if("$ENV{TARGET_CPU}" STREQUAL "cortex_a53" OR "$ENV{TARGET_CPU}" STREQUAL "cortex_a57")
+            if("$ENV{TARGET_CPU}" STREQUAL "cortex_a35" OR "$ENV{TARGET_CPU}" STREQUAL "cortex_a53" OR "$ENV{TARGET_CPU}" STREQUAL "cortex_a57" OR "$ENV{TARGET_CPU}" STREQUAL "cortex_a72" OR "$ENV{TARGET_CPU}" STREQUAL "cortex_a73")
                 set(CMAKE_SYSTEM_PROCESSOR "armv8l" CACHE STRING "Updates CMAKE_SYSTEM_PROCESSOR")
+                set(XNNPACK_ENABLE_ARM_FP16_SCALAR OFF CACHE BOOL "Disable XNNPACK fp16 scalar in TensorFlow Lite as it only available on armv8.2+")
+                set(XNNPACK_ENABLE_ARM_FP16_VECTOR OFF CACHE BOOL "Disable XNNPACK fp16 vector in TensorFlow Lite as it only available on armv8.2+")
+                set(XNNPACK_ENABLE_ARM_BF16 OFF CACHE BOOL "Disable XNNPACK bf16 in TensorFlow Lite as it only available on armv8.2+")
+                set(XNNPACK_ENABLE_ARM_DOTPROD OFF CACHE BOOL "Disable XNNPACK dotprod in TensorFlow Lite as it only available on armv8.2+")
+                set(XNNPACK_ENABLE_ARM_I8MM OFF CACHE BOOL "Disable XNNPACK i8mm in TensorFlow Lite as it only available on armv8.2+")
             elseif("$ENV{TARGET_CPU}" STREQUAL "cortex_a8" OR "$ENV{TARGET_CPU}" STREQUAL "cortex_a7")
                 set(CMAKE_SYSTEM_PROCESSOR "armv7l" CACHE STRING "Updates CMAKE_SYSTEM_PROCESSOR")
                 set(TFLITE_ENABLE_XNNPACK OFF CACHE BOOL "Disable XNNPACK in TensorFlow Lite for armv7l")
@@ -86,6 +91,22 @@ if(DEFINED ENV{TARGET_ARCH})
     endif()
 endif()
 
+if(DEFINED ENV{TARGET_ARCH})
+    if(NOT "$ENV{TARGET_ARCH}" STREQUAL "arm")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DTF_HAS_vcvtaq_s32_f32")
+    endif()
+endif()
+
+if(DEFINED ENV{TFBEAM_XNNPACK_ENABLE_ARM_I8MM})
+    if("$ENV{TFBEAM_XNNPACK_ENABLE_ARM_I8MM}" STREQUAL "OFF")
+        set(XNNPACK_ENABLE_ARM_I8MM OFF CACHE BOOL "Disable XNNPACK i8mm in TensorFlow Lite as TFBEAM_XNNPACK_ENABLE_ARM_I8MM is set to OFF")
+    else()
+        set(XNNPACK_ENABLE_ARM_I8MM ON CACHE BOOL "Enable XNNPACK i8mm in TensorFlow Lite as TFBEAM_XNNPACK_ENABLE_ARM_I8MM is set to ON")
+    endif()
+else()
+    set(XNNPACK_ENABLE_ARM_I8MM OFF CACHE BOOL "Disable XNNPACK i8mm in TensorFlow Lite as TFBEAM_XNNPACK_ENABLE_ARM_I8MM is default to OFF")
+endif()
+
 if(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
     if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 10.0)
         set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-gnu-inline-cpp-without-extern")

README.md

@@ -36,6 +36,12 @@ bash "3rd_party/cache/${TFLITE_BEAM_CORAL_LIBEDGETPU_RUNTIME}/edgetpu_runtime/in
 
   Defaults to `true`.
 
+- `TFBEAM_XNNPACK_ENABLE_ARM_I8MM`
+  
+  Set to 'ON' so that XNNPACK can use 8-bit Integer Matrix Multiply instructions. This also enables Advanced SIMD and floating-point instructions. Note that this only available on armv8.2-a and later. 
+
+  Defaults to `OFF`.
+
 - `TFLITE_BEAM_CORAL_SUPPORT`
 
   Enable Coral Support.