chore(dotprod): dotprod is enabled by default on the android platform

GitOrigin-RevId: d412108732

CMakeLists.txt

@@ -34,7 +34,6 @@ option(MGE_WITH_HALIDE "Build MegEngine with Halide JIT" OFF)
 option(MGE_WITH_MIDOUT_PROFILE "Build MegEngine with Midout profile." OFF)
 option(MGE_WITH_MINIMUM_SIZE "Swith off MGE_ENABLE_RTTI、MGE_ENABLE_EXCEPTIONS、MGE_ENABLE_LOGGING and switch on MGE_INFERENCE_ONLY so that compile minimum load_and_run. Take effect only when MGE_BIN_REDUCE was set" OFF)
 option(MGE_ARMV8_2_FEATURE_FP16 "Enable armv8.2-a+fp16 support" OFF)
-option(MGE_ARMV8_2_FEATURE_DOTPROD "enable armv8.2-a+dotprod support" OFF)
 option(MGE_DISABLE_FLOAT16 "Disable MegEngine float16 support." OFF)
 option(MGE_WITH_CUDA "Enable MegEngine CUDA support." ON)
 option(MGE_CUDA_USE_STATIC "Enable MegEngine CUDA static linking." ON)
@@ -773,6 +772,14 @@ if(MGE_ARCH STREQUAL "x86_64" OR MGE_ARCH STREQUAL "i386")
         set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse4.2 -mfpmath=sse")
     endif()
 endif()
+# dotprod is not enable by default on APPLE, cpuinfo has some problem on APPLE
+if(NOT APPLE)
+    CHECK_CXX_COMPILER_FLAG("-march=armv8.2-a+dotprod" CXX_COMPILER_SUPPORT_DOT)
+    if(CXX_COMPILER_SUPPORT_DOT)
+        message(STATUS "Enable dotprod feature in armv8.2-a using MGB_ENABLE_DOT")
+        set(MGB_ENABLE_DOT 1)
+    endif()
+endif()
 
 if(MGE_ARCH STREQUAL "armv7")
     # -funsafe-math-optimizations to enable neon auto-vectorization (since neon is not fully IEEE 754 compatible, GCC does not turn on neon auto-vectorization by default.
@@ -797,15 +804,6 @@ if(MGE_ARCH STREQUAL "aarch64")
         set(MARCH "-march=armv8.2-a+fp16")
     endif()
 
-    if(MGE_ARMV8_2_FEATURE_DOTPROD)
-        message(STATUS "Enable dotprod feature support in armv8.2")
-        if(MGE_ARMV8_2_FEATURE_FP16)
-            set(MARCH "-march=armv8.2-a+fp16+dotprod")
-        else()
-            set(MARCH "-march=armv8.2-a+dotprod")
-        endif()
-    endif()
-
     if(MGE_WITH_CUDA)
         message(WARNING "aarch64 ld will add -mfix-cortex-a53-843419 and -mfix-cortex-a53-835769,\
         when cuda enable and CMAKE with DEBUG build type,ld will take about 14min+,\

dnn/src/arm_common/simd_macro/marm_neon.h

@@ -10,6 +10,10 @@
  * implied.
  */
 #pragma once
+
+#include "megdnn/arch.h"
+#include "src/common/unroll_macro.h"
+
 #if MGB_ENABLE_DOT
 #if defined(__ARM_FEATURE_DOTPROD)
 #undef __ARM_FEATURE_DOTPROD
@@ -17,8 +21,6 @@
 #define __ARM_FEATURE_DOTPROD 1
 #endif
 #include <arm_neon.h>
-#include "megdnn/arch.h"
-#include "src/common/unroll_macro.h"
 
 // GCC does not support __nodebug__, it reports:
 // '__nodebug__' attribute directive ignored

scripts/cmake-build/cross_build_android_arm_inference.sh

@@ -4,7 +4,6 @@ set -e
 ARCHS=("arm64-v8a" "armeabi-v7a")
 BUILD_TYPE=Release
 MGE_ARMV8_2_FEATURE_FP16=OFF
-MGE_ARMV8_2_FEATURE_DOTPROD=OFF
 MGE_DISABLE_FLOAT16=OFF
 ARCH=arm64-v8a
 REMOVE_OLD_BUILD=false
@@ -15,7 +14,6 @@ function usage() {
     echo "available args detail:"
     echo "-d : Build with Debug mode, default Release mode"
     echo "-f : enable MGE_ARMV8_2_FEATURE_FP16 for ARM64, need toolchain and hardware support"
-    echo "-p : enable MGE_ARMV8_2_FEATURE_DOTPROD for ARM64, need toolchain and hardware support"
     echo "-k : open MGE_DISABLE_FLOAT16 for NEON "
     echo "-a : config build arch available: ${ARCHS[@]}"
     echo "-r : remove old build dir before make, default off"
@@ -25,7 +23,7 @@ function usage() {
     exit -1
 }
 
-while getopts "rkhdfpa:" arg
+while getopts "rkhdfa:" arg
 do
     case $arg in
         d)
@@ -36,10 +34,6 @@ do
             echo "enable MGE_ARMV8_2_FEATURE_FP16 for ARM64"
             MGE_ARMV8_2_FEATURE_FP16=ON
             ;;
-        p)
-            echo "enable MGE_ARMV8_2_FEATURE_DOTPROD for ARM64"
-            MGE_ARMV8_2_FEATURE_DOTPROD=ON
-            ;;
         k)
             echo "open MGE_DISABLE_FLOAT16 for NEON"
             MGE_DISABLE_FLOAT16=ON
@@ -78,7 +72,6 @@ echo "----------------------------------------------------"
 echo "build config summary:"
 echo "BUILD_TYPE: $BUILD_TYPE"
 echo "MGE_ARMV8_2_FEATURE_FP16: $MGE_ARMV8_2_FEATURE_FP16"
-echo "MGE_ARMV8_2_FEATURE_DOTPROD: $MGE_ARMV8_2_FEATURE_DOTPROD"
 echo "MGE_DISABLE_FLOAT16: $MGE_DISABLE_FLOAT16"
 echo "ARCH: $ARCH"
 echo "----------------------------------------------------"
@@ -129,7 +122,6 @@ function cmake_build() {
         -DMGE_INFERENCE_ONLY=ON \
         -DMGE_WITH_CUDA=OFF \
         -DMGE_ARMV8_2_FEATURE_FP16= $MGE_ARMV8_2_FEATURE_FP16 \
-        -DMGE_ARMV8_2_FEATURE_DOTPROD=$MGE_ARMV8_2_FEATURE_DOTPROD \
         -DMGE_DISABLE_FLOAT16=$MGE_DISABLE_FLOAT16 \
         -DCMAKE_INSTALL_PREFIX=$INSTALL_DIR \
         ${EXTRA_CMAKE_ARGS} \

scripts/cmake-build/cross_build_ios_arm_inference.sh

@@ -4,7 +4,6 @@ set -e
 ARCHS=("arm64" "armv7")
 BUILD_TYPE=Release
 MGE_ARMV8_2_FEATURE_FP16=OFF
-MGE_ARMV8_2_FEATURE_DOTPROD=OFF
 MGE_DISABLE_FLOAT16=OFF
 ARCH=arm64
 REMOVE_OLD_BUILD=false
@@ -15,7 +14,6 @@ function usage() {
     echo "available args detail:"
     echo "-d : Build with Debug mode, default Release mode"
     echo "-f : enable MGE_ARMV8_2_FEATURE_FP16 for ARM64, need toolchain and hardware support"
-    echo "-p : enable MGE_ARMV8_2_FEATURE_DOTPROD for ARM64, need toolchain and hardware support"
     echo "-k : open MGE_DISABLE_FLOAT16 for NEON "
     echo "-a : config build arch available: ${ARCHS[@]}"
     echo "-r : remove old build dir before make, default off"
@@ -25,7 +23,7 @@ function usage() {
     exit -1
 }
 
-while getopts "rkhdfpa:" arg
+while getopts "rkhdfa:" arg
 do
     case $arg in
         d)
@@ -36,10 +34,6 @@ do
             echo "enable MGE_ARMV8_2_FEATURE_FP16 for ARM64"
             MGE_ARMV8_2_FEATURE_FP16=ON
             ;;
-        p)
-            echo "enable MGE_ARMV8_2_FEATURE_DOTPROD for ARM64"
-            MGE_ARMV8_2_FEATURE_DOTPROD=ON
-            ;;
         k)
             echo "open MGE_DISABLE_FLOAT16 for NEON"
             MGE_DISABLE_FLOAT16=ON
@@ -78,7 +72,6 @@ echo "----------------------------------------------------"
 echo "build config summary:"
 echo "BUILD_TYPE: $BUILD_TYPE"
 echo "MGE_ARMV8_2_FEATURE_FP16: $MGE_ARMV8_2_FEATURE_FP16"
-echo "MGE_ARMV8_2_FEATURE_DOTPROD: $MGE_ARMV8_2_FEATURE_DOTPROD"
 echo "MGE_DISABLE_FLOAT16: $MGE_DISABLE_FLOAT16"
 echo "ARCH: $ARCH"
 echo "----------------------------------------------------"
@@ -126,7 +119,6 @@ function cmake_build() {
         -DPYTHON_EXECUTABLE=/usr/local/bin/python3 \
         -DMGE_WITH_CUDA=OFF \
         -DMGE_ARMV8_2_FEATURE_FP16= $MGE_ARMV8_2_FEATURE_FP16 \
-        -DMGE_ARMV8_2_FEATURE_DOTPROD=$MGE_ARMV8_2_FEATURE_DOTPROD \
         -DMGE_DISABLE_FLOAT16=$MGE_DISABLE_FLOAT16 \
         -DCMAKE_INSTALL_PREFIX=$INSTALL_DIR \
         ${EXTRA_CMAKE_ARGS} \

scripts/cmake-build/cross_build_linux_arm_inference.sh

@@ -5,7 +5,6 @@ ARCHS=("arm64-v8a" "armeabi-v7a-softfp" "armeabi-v7a-hardfp")
 BUILD_TYPE=Release
 MGE_WITH_CUDA=OFF
 MGE_ARMV8_2_FEATURE_FP16=OFF
-MGE_ARMV8_2_FEATURE_DOTPROD=OFF
 MGE_DISABLE_FLOAT16=OFF
 ARCH=arm64-v8a
 REMOVE_OLD_BUILD=false
@@ -19,7 +18,6 @@ function usage() {
     echo "-d : Build with Debug mode, default Release mode"
     echo "-c : Build with CUDA, default without CUDA(for arm with cuda, example tx1)"
     echo "-f : enable MGE_ARMV8_2_FEATURE_FP16 for ARM64, need toolchain and hardware support"
-    echo "-p : enable MGE_ARMV8_2_FEATURE_DOTPROD for ARM64, need toolchain and hardware support"
     echo "-k : open MGE_DISABLE_FLOAT16 for NEON "
     echo "-a : config build arch available: ${ARCHS[@]}"
     echo "-r : remove old build dir before make, default off"
@@ -29,7 +27,7 @@ function usage() {
     exit -1
 }
 
-while getopts "rkhdcfpa:" arg
+while getopts "rkhdcfa:" arg
 do
     case $arg in
         d)
@@ -44,10 +42,6 @@ do
             echo "enable MGE_ARMV8_2_FEATURE_FP16 for ARM64"
             MGE_ARMV8_2_FEATURE_FP16=ON
             ;;
-        p)
-            echo "enable MGE_ARMV8_2_FEATURE_DOTPROD for ARM64"
-            MGE_ARMV8_2_FEATURE_DOTPROD=ON
-            ;;
         k)
             echo "open MGE_DISABLE_FLOAT16 for NEON"
             MGE_DISABLE_FLOAT16=ON
@@ -87,7 +81,6 @@ echo "build config summary:"
 echo "BUILD_TYPE: $BUILD_TYPE"
 echo "MGE_WITH_CUDA: $MGE_WITH_CUDA"
 echo "MGE_ARMV8_2_FEATURE_FP16: $MGE_ARMV8_2_FEATURE_FP16"
-echo "MGE_ARMV8_2_FEATURE_DOTPROD: $MGE_ARMV8_2_FEATURE_DOTPROD"
 echo "MGE_DISABLE_FLOAT16: $MGE_DISABLE_FLOAT16"
 echo "ARCH: $ARCH"
 echo "----------------------------------------------------"
@@ -147,7 +140,6 @@ function cmake_build() {
         -DMGE_INFERENCE_ONLY=ON \
         -DMGE_WITH_CUDA=$MGE_WITH_CUDA \
         -DMGE_ARMV8_2_FEATURE_FP16= $MGE_ARMV8_2_FEATURE_FP16 \
-        -DMGE_ARMV8_2_FEATURE_DOTPROD=$MGE_ARMV8_2_FEATURE_DOTPROD \
         -DMGE_DISABLE_FLOAT16=$MGE_DISABLE_FLOAT16 \
         -DCMAKE_INSTALL_PREFIX=$INSTALL_DIR \
         ${EXTRA_CMAKE_ARGS} \

src/megbrain_build_config.h.in

@@ -93,11 +93,6 @@
 #define MGB_ENABLE_CPUINFO      1
 #endif
 
-#ifdef IOS
-#undef MGB_ENABLE_CPUINFO
-#define MGB_ENABLE_CPUINFO      0
-#endif
-
 //! use one MACRO indicate enable_arm_dotprod
 #if __ARM_FEATURE_DOTPROD
 #ifdef MGB_ENABLE_DOT
@@ -117,7 +112,12 @@
 #endif
 #endif
 
-
+//! IOS disabled cpuinfo and dotprod, cpuinfo has some problem on ios
+#ifdef IOS
+#undef MGB_ENABLE_CPUINFO
+#define MGB_ENABLE_CPUINFO      0
+#undef MGB_ENABLE_DOT
+#endif
 
 // whether to include actual class name in mgb::Typeinfo object; if this is
 // disabled, mgb::serialization::OprRegistry::find_opr_by_name would not work.