[x265] [PATCH] Add aarch64 support - Part 2

Suyimeng yimeng.su at huawei.com
Tue Mar 17 05:20:34 CET 2020


From: x265-devel [mailto:x265-devel-bounces at videolan.org] On Behalf Of Gopi Satykrishna Akisetty
Sent: Monday, March 16, 2020 9:37 PM
To: Development for x265 <x265-devel at videolan.org>
Subject: Re: [x265] [PATCH] Add aarch64 support - Part 2



On Thu, Feb 27, 2020 at 8:04 AM Xiyuan Wang <wangxiyuan1007 at gmail.com<mailto:wangxiyuan1007 at gmail.com>> wrote:
From: wangxiyuan <wangxiyuan at huawei.com<mailto:wangxiyuan at huawei.com>>

This patch adds aarch64 build & compile support. This patch must be
merged after the patch Part 1.
---
 build/aarch64-linux/crosscompile.cmake  |  15 ++
 build/aarch64-linux/make-Makefiles.bash |   4 +
 source/CMakeLists.txt                   |  38 +++-
 source/common/CMakeLists.txt            |  35 ++-
 source/common/arm/asm-primitives.cpp    | 291 ++++++++++++------------
 source/common/cpu.cpp                   |   4 +
 source/common/pixel.cpp                 |   9 +
 source/common/primitives.h              |  11 +
 source/test/CMakeLists.txt              |  16 +-
 source/test/testbench.cpp               |  16 ++
 source/test/testharness.h               |   5 +
 11 files changed, 274 insertions(+), 170 deletions(-)
 create mode 100644 build/aarch64-linux/crosscompile.cmake
 create mode 100644 build/aarch64-linux/make-Makefiles.bash

diff --git a/build/aarch64-linux/crosscompile.cmake b/build/aarch64-linux/crosscompile.cmake
new file mode 100644
index 000000000..41c8217f2
--- /dev/null
+++ b/build/aarch64-linux/crosscompile.cmake
@@ -0,0 +1,15 @@
+# CMake toolchain file for cross compiling x265 for aarch64
+# This feature is only supported as experimental. Use with caution.
+# Please report bugs on bitbucket
+# Run cmake with: cmake -DCMAKE_TOOLCHAIN_FILE=crosscompile.cmake -G "Unix Makefiles" ../../source && ccmake ../../source
+
+set(CROSS_COMPILE_ARM 1)
+set(CMAKE_SYSTEM_NAME Linux)
+set(CMAKE_SYSTEM_PROCESSOR aarch64)
+
+# specify the cross compiler
+set(CMAKE_C_COMPILER aarch64-linux-gnu-gcc)
+set(CMAKE_CXX_COMPILER aarch64-linux-gnu-g++)
+
+# specify the target environment
+SET(CMAKE_FIND_ROOT_PATH  /usr/aarch64-linux-gnu)
diff --git a/build/aarch64-linux/make-Makefiles.bash b/build/aarch64-linux/make-Makefiles.bash
new file mode 100644
index 000000000..c9582da0a
--- /dev/null
+++ b/build/aarch64-linux/make-Makefiles.bash
@@ -0,0 +1,4 @@
+#!/bin/bash
+# Run this from within a bash shell
+
+cmake -DCMAKE_TOOLCHAIN_FILE="crosscompile.cmake" -G "Unix Makefiles" ../../source && ccmake ../../source
diff --git a/source/CMakeLists.txt b/source/CMakeLists.txt
index 5d2474d97..7734eafbb 100644
--- a/source/CMakeLists.txt
+++ b/source/CMakeLists.txt
@@ -40,7 +40,7 @@ SET(CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake" "${CMAKE_MODULE_PATH}")
 # System architecture detection
 string(TOLOWER "${CMAKE_SYSTEM_PROCESSOR}" SYSPROC)
 set(X86_ALIASES x86 i386 i686 x86_64 amd64)
-set(ARM_ALIASES armv6l armv7l)
+set(ARM_ALIASES armv6l armv7l aarch64)
 list(FIND X86_ALIASES "${SYSPROC}" X86MATCH)
 list(FIND ARM_ALIASES "${SYSPROC}" ARMMATCH)
 set(POWER_ALIASES ppc64 ppc64le)
@@ -70,9 +70,15 @@ elseif(ARMMATCH GREATER "-1")
     else()
         set(CROSS_COMPILE_ARM 0)
     endif()
-    message(STATUS "Detected ARM target processor")
     set(ARM 1)
-    add_definitions(-DX265_ARCH_ARM=1 -DHAVE_ARMV6=1)
+    if("${CMAKE_SIZEOF_VOID_P}" MATCHES 8)
+        message(STATUS "Detected ARM64 target processor")
+        set(ARM64 1)
+        add_definitions(-DX265_ARCH_ARM=1 -DX265_ARCH_ARM64=1 -DHAVE_ARMV6=0)
+    else()
+        message(STATUS "Detected ARM target processor")
+        add_definitions(-DX265_ARCH_ARM=1 -DX265_ARCH_ARM64=0 -DHAVE_ARMV6=1)
+    endif()
 else()
     message(STATUS "CMAKE_SYSTEM_PROCESSOR value `${CMAKE_SYSTEM_PROCESSOR}` is unknown")
     message(STATUS "Please add this value near ${CMAKE_CURRENT_LIST_FILE}:${CMAKE_CURRENT_LIST_LINE}")
@@ -231,14 +237,24 @@ if(GCC)
         endif()
     endif()
     if(ARM AND CROSS_COMPILE_ARM)
-        set(ARM_ARGS -march=armv6 -mfloat-abi=soft -mfpu=vfp -marm -fPIC)
+        if(ARM64)
+            set(ARM_ARGS -fPIC)
+        else()
+            set(ARM_ARGS -march=armv6 -mfloat-abi=soft -mfpu=vfp -marm -fPIC)
+        endif()
+        message(STATUS "cross compile arm")
     elseif(ARM)
-        find_package(Neon)
-        if(CPU_HAS_NEON)
-            set(ARM_ARGS -mcpu=native -mfloat-abi=hard -mfpu=neon -marm -fPIC)
+        if(ARM64)
+            set(ARM_ARGS -fPIC)
             add_definitions(-DHAVE_NEON)
         else()
-            set(ARM_ARGS -mcpu=native -mfloat-abi=hard -mfpu=vfp -marm)
+            find_package(Neon)
+            if(CPU_HAS_NEON)
+                set(ARM_ARGS -mcpu=native -mfloat-abi=hard -mfpu=neon -marm -fPIC)
+                add_definitions(-DHAVE_NEON)
+            else()
+                set(ARM_ARGS -mcpu=native -mfloat-abi=hard -mfpu=vfp -marm)
+            endif()
         endif()
     endif()
     add_definitions(${ARM_ARGS})
@@ -518,7 +534,11 @@ if((MSVC_IDE OR XCODE OR GCC) AND ENABLE_ASSEMBLY)
     # compile ARM arch asm files here
         enable_language(ASM)
         foreach(ASM ${ARM_ASMS})
-            set(ASM_SRC ${CMAKE_CURRENT_SOURCE_DIR}/common/arm/${ASM})
+            if(ARM64)
+                set(ASM_SRC ${CMAKE_CURRENT_SOURCE_DIR}/common/aarch64/${ASM})
+            else()
+                set(ASM_SRC ${CMAKE_CURRENT_SOURCE_DIR}/common/arm/${ASM})
+            endif()
             list(APPEND ASM_SRCS ${ASM_SRC})
             list(APPEND ASM_OBJS ${ASM}.${SUFFIX})
             add_custom_command(
diff --git a/source/common/CMakeLists.txt b/source/common/CMakeLists.txt
index c70bb108c..c021e603e 100644
--- a/source/common/CMakeLists.txt
+++ b/source/common/CMakeLists.txt
@@ -14,7 +14,7 @@ if(EXTRA_LIB)
 endif(EXTRA_LIB)

 if(ENABLE_ASSEMBLY)
-    set_source_files_properties(threading.cpp primitives.cpp PROPERTIES COMPILE_FLAGS -DENABLE_ASSEMBLY=1)
+    set_source_files_properties(threading.cpp primitives.cpp pixel.cpp PROPERTIES COMPILE_FLAGS -DENABLE_ASSEMBLY=1)
     list(APPEND VFLAGS "-DENABLE_ASSEMBLY=1")
 endif(ENABLE_ASSEMBLY)

@@ -84,16 +84,33 @@ if(ENABLE_ASSEMBLY AND X86)
 endif(ENABLE_ASSEMBLY AND X86)

 if(ENABLE_ASSEMBLY AND (ARM OR CROSS_COMPILE_ARM))
-    set(C_SRCS asm-primitives.cpp pixel.h mc.h ipfilter8.h blockcopy8.h dct8.h loopfilter.h)
+    if(ARM64)
+        if(GCC AND (CMAKE_CXX_FLAGS_RELEASE MATCHES "-O3"))
+            message(STATUS "Detected CXX compiler using -O3 optimization level")
+            add_definitions(-DAUTO_VECTORIZE=1)
+        endif()
+        set(C_SRCS asm-primitives.cpp pixel.h ipfilter8.h)

-    # add ARM assembly/intrinsic files here
-    set(A_SRCS asm.S cpu-a.S mc-a.S sad-a.S pixel-util.S ssd-a.S blockcopy8.S ipfilter8.S dct-a.S)
-    set(VEC_PRIMITIVES)
+        # add ARM assembly/intrinsic files here
+        set(A_SRCS asm.S mc-a.S sad-a.S pixel-util.S ipfilter8.S)
+        set(VEC_PRIMITIVES)

-    set(ARM_ASMS "${A_SRCS}" CACHE INTERNAL "ARM Assembly Sources")
-    foreach(SRC ${C_SRCS})
-        set(ASM_PRIMITIVES ${ASM_PRIMITIVES} arm/${SRC})
-    endforeach()
+        set(ARM_ASMS "${A_SRCS}" CACHE INTERNAL "ARM Assembly Sources")
+        foreach(SRC ${C_SRCS})
+            set(ASM_PRIMITIVES ${ASM_PRIMITIVES} aarch64/${SRC})
+        endforeach()
+    else()
+        set(C_SRCS asm-primitives.cpp pixel.h mc.h ipfilter8.h blockcopy8.h dct8.h loopfilter.h)
+
+        # add ARM assembly/intrinsic files here
+        set(A_SRCS asm.S cpu-a.S mc-a.S sad-a.S pixel-util.S ssd-a.S blockcopy8.S ipfilter8.S dct-a.S)
+        set(VEC_PRIMITIVES)
+
+        set(ARM_ASMS "${A_SRCS}" CACHE INTERNAL "ARM Assembly Sources")
+        foreach(SRC ${C_SRCS})
+            set(ASM_PRIMITIVES ${ASM_PRIMITIVES} arm/${SRC})
+        endforeach()
+    endif()
     source_group(Assembly FILES ${ASM_PRIMITIVES})
 endif(ENABLE_ASSEMBLY AND (ARM OR CROSS_COMPILE_ARM))

diff --git a/source/common/arm/asm-primitives.cpp b/source/common/arm/asm-primitives.cpp
index 422217845..7f11503f9 100644
--- a/source/common/arm/asm-primitives.cpp
+++ b/source/common/arm/asm-primitives.cpp
@@ -5,6 +5,7 @@
  *          Praveen Kumar Tiwari <praveen at multicorewareinc.com<mailto:praveen at multicorewareinc.com>>
  *          Min Chen <chenm003 at 163.com<mailto:chenm003 at 163.com>> <min.chen at multicorewareinc.com<mailto:min.chen at multicorewareinc.com>>
  *          Dnyaneshwar Gorade <dnyaneshwar at multicorewareinc.com<mailto:dnyaneshwar at multicorewareinc.com>>
+ *          Hongbin Liu<liuhongbin1 at huawei.com<mailto:liuhongbin1 at huawei.com>>
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by
@@ -48,77 +49,77 @@ void setupAssemblyPrimitives(EncoderPrimitives &p, int cpuMask)
         p.ssim_4x4x2_core = PFX(ssim_4x4x2_core_neon);

         // addAvg
-         p.pu[LUMA_4x4].addAvg   = PFX(addAvg_4x4_neon);
-         p.pu[LUMA_4x8].addAvg   = PFX(addAvg_4x8_neon);
-         p.pu[LUMA_4x16].addAvg  = PFX(addAvg_4x16_neon);
-         p.pu[LUMA_8x4].addAvg   = PFX(addAvg_8x4_neon);
-         p.pu[LUMA_8x8].addAvg   = PFX(addAvg_8x8_neon);
-         p.pu[LUMA_8x16].addAvg  = PFX(addAvg_8x16_neon);
-         p.pu[LUMA_8x32].addAvg  = PFX(addAvg_8x32_neon);
-         p.pu[LUMA_12x16].addAvg = PFX(addAvg_12x16_neon);
-         p.pu[LUMA_16x4].addAvg  = PFX(addAvg_16x4_neon);
-         p.pu[LUMA_16x8].addAvg  = PFX(addAvg_16x8_neon);
-         p.pu[LUMA_16x12].addAvg = PFX(addAvg_16x12_neon);
-         p.pu[LUMA_16x16].addAvg = PFX(addAvg_16x16_neon);
-         p.pu[LUMA_16x32].addAvg = PFX(addAvg_16x32_neon);
-         p.pu[LUMA_16x64].addAvg = PFX(addAvg_16x64_neon);
-         p.pu[LUMA_24x32].addAvg = PFX(addAvg_24x32_neon);
-         p.pu[LUMA_32x8].addAvg  = PFX(addAvg_32x8_neon);
-         p.pu[LUMA_32x16].addAvg = PFX(addAvg_32x16_neon);
-         p.pu[LUMA_32x24].addAvg = PFX(addAvg_32x24_neon);
-         p.pu[LUMA_32x32].addAvg = PFX(addAvg_32x32_neon);
-         p.pu[LUMA_32x64].addAvg = PFX(addAvg_32x64_neon);
-         p.pu[LUMA_48x64].addAvg = PFX(addAvg_48x64_neon);
-         p.pu[LUMA_64x16].addAvg = PFX(addAvg_64x16_neon);
-         p.pu[LUMA_64x32].addAvg = PFX(addAvg_64x32_neon);
-         p.pu[LUMA_64x48].addAvg = PFX(addAvg_64x48_neon);
-         p.pu[LUMA_64x64].addAvg = PFX(addAvg_64x64_neon);
+         p.pu[LUMA_4x4].addAvg[NONALIGNED]   = PFX(addAvg_4x4_neon);
+         p.pu[LUMA_4x8].addAvg[NONALIGNED]   = PFX(addAvg_4x8_neon);
+         p.pu[LUMA_4x16].addAvg[NONALIGNED]  = PFX(addAvg_4x16_neon);
+         p.pu[LUMA_8x4].addAvg[NONALIGNED]   = PFX(addAvg_8x4_neon);
+         p.pu[LUMA_8x8].addAvg[NONALIGNED]   = PFX(addAvg_8x8_neon);
+         p.pu[LUMA_8x16].addAvg[NONALIGNED]  = PFX(addAvg_8x16_neon);
+         p.pu[LUMA_8x32].addAvg[NONALIGNED]  = PFX(addAvg_8x32_neon);
+         p.pu[LUMA_12x16].addAvg[NONALIGNED] = PFX(addAvg_12x16_neon);
+         p.pu[LUMA_16x4].addAvg[NONALIGNED]  = PFX(addAvg_16x4_neon);
+         p.pu[LUMA_16x8].addAvg[NONALIGNED]  = PFX(addAvg_16x8_neon);
+         p.pu[LUMA_16x12].addAvg[NONALIGNED] = PFX(addAvg_16x12_neon);
+         p.pu[LUMA_16x16].addAvg[NONALIGNED] = PFX(addAvg_16x16_neon);
+         p.pu[LUMA_16x32].addAvg[NONALIGNED] = PFX(addAvg_16x32_neon);
+         p.pu[LUMA_16x64].addAvg[NONALIGNED] = PFX(addAvg_16x64_neon);
+         p.pu[LUMA_24x32].addAvg[NONALIGNED] = PFX(addAvg_24x32_neon);
+         p.pu[LUMA_32x8].addAvg[NONALIGNED]  = PFX(addAvg_32x8_neon);
+         p.pu[LUMA_32x16].addAvg[NONALIGNED] = PFX(addAvg_32x16_neon);
+         p.pu[LUMA_32x24].addAvg[NONALIGNED] = PFX(addAvg_32x24_neon);
+         p.pu[LUMA_32x32].addAvg[NONALIGNED] = PFX(addAvg_32x32_neon);
+         p.pu[LUMA_32x64].addAvg[NONALIGNED] = PFX(addAvg_32x64_neon);
+         p.pu[LUMA_48x64].addAvg[NONALIGNED] = PFX(addAvg_48x64_neon);
+         p.pu[LUMA_64x16].addAvg[NONALIGNED] = PFX(addAvg_64x16_neon);
+         p.pu[LUMA_64x32].addAvg[NONALIGNED] = PFX(addAvg_64x32_neon);
+         p.pu[LUMA_64x48].addAvg[NONALIGNED] = PFX(addAvg_64x48_neon);
+         p.pu[LUMA_64x64].addAvg[NONALIGNED] = PFX(addAvg_64x64_neon);

         // chroma addAvg
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].addAvg   = PFX(addAvg_4x2_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].addAvg   = PFX(addAvg_4x4_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].addAvg   = PFX(addAvg_4x8_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].addAvg  = PFX(addAvg_4x16_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].addAvg   = PFX(addAvg_6x8_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].addAvg   = PFX(addAvg_8x2_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].addAvg   = PFX(addAvg_8x4_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].addAvg   = PFX(addAvg_8x6_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].addAvg   = PFX(addAvg_8x8_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].addAvg  = PFX(addAvg_8x16_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].addAvg  = PFX(addAvg_8x32_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].addAvg = PFX(addAvg_12x16_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].addAvg  = PFX(addAvg_16x4_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].addAvg  = PFX(addAvg_16x8_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].addAvg = PFX(addAvg_16x12_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].addAvg = PFX(addAvg_16x16_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].addAvg = PFX(addAvg_16x32_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].addAvg = PFX(addAvg_24x32_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].addAvg  = PFX(addAvg_32x8_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].addAvg = PFX(addAvg_32x16_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].addAvg = PFX(addAvg_32x24_neon);
-        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].addAvg = PFX(addAvg_32x32_neon);
-
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].addAvg   = PFX(addAvg_4x8_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].addAvg  = PFX(addAvg_4x16_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x32].addAvg  = PFX(addAvg_4x32_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].addAvg  = PFX(addAvg_6x16_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].addAvg   = PFX(addAvg_8x4_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].addAvg   = PFX(addAvg_8x8_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].addAvg  = PFX(addAvg_8x12_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].addAvg  = PFX(addAvg_8x16_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].addAvg  = PFX(addAvg_8x32_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].addAvg  = PFX(addAvg_8x64_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].addAvg = PFX(addAvg_12x32_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].addAvg  = PFX(addAvg_16x8_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].addAvg = PFX(addAvg_16x16_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].addAvg = PFX(addAvg_16x24_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].addAvg = PFX(addAvg_16x32_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].addAvg = PFX(addAvg_16x64_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].addAvg = PFX(addAvg_24x64_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].addAvg = PFX(addAvg_32x16_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].addAvg = PFX(addAvg_32x32_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].addAvg = PFX(addAvg_32x48_neon);
-        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].addAvg = PFX(addAvg_32x64_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].addAvg[NONALIGNED]   = PFX(addAvg_4x2_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].addAvg[NONALIGNED]   = PFX(addAvg_4x4_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].addAvg[NONALIGNED]   = PFX(addAvg_4x8_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].addAvg[NONALIGNED]  = PFX(addAvg_4x16_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].addAvg[NONALIGNED]   = PFX(addAvg_6x8_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].addAvg[NONALIGNED]   = PFX(addAvg_8x2_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].addAvg[NONALIGNED]   = PFX(addAvg_8x4_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].addAvg[NONALIGNED]   = PFX(addAvg_8x6_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].addAvg[NONALIGNED]   = PFX(addAvg_8x8_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].addAvg[NONALIGNED]  = PFX(addAvg_8x16_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].addAvg[NONALIGNED]  = PFX(addAvg_8x32_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].addAvg[NONALIGNED] = PFX(addAvg_12x16_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].addAvg[NONALIGNED]  = PFX(addAvg_16x4_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].addAvg[NONALIGNED]  = PFX(addAvg_16x8_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].addAvg[NONALIGNED] = PFX(addAvg_16x12_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].addAvg[NONALIGNED] = PFX(addAvg_16x16_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].addAvg[NONALIGNED] = PFX(addAvg_16x32_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].addAvg[NONALIGNED] = PFX(addAvg_24x32_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].addAvg[NONALIGNED]  = PFX(addAvg_32x8_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].addAvg[NONALIGNED] = PFX(addAvg_32x16_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].addAvg[NONALIGNED] = PFX(addAvg_32x24_neon);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].addAvg[NONALIGNED] = PFX(addAvg_32x32_neon);
+
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].addAvg[NONALIGNED]   = PFX(addAvg_4x8_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].addAvg[NONALIGNED]  = PFX(addAvg_4x16_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x32].addAvg[NONALIGNED]  = PFX(addAvg_4x32_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].addAvg[NONALIGNED]  = PFX(addAvg_6x16_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].addAvg[NONALIGNED]   = PFX(addAvg_8x4_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].addAvg[NONALIGNED]   = PFX(addAvg_8x8_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].addAvg[NONALIGNED]  = PFX(addAvg_8x12_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].addAvg[NONALIGNED]  = PFX(addAvg_8x16_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].addAvg[NONALIGNED]  = PFX(addAvg_8x32_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].addAvg[NONALIGNED]  = PFX(addAvg_8x64_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].addAvg[NONALIGNED] = PFX(addAvg_12x32_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].addAvg[NONALIGNED]  = PFX(addAvg_16x8_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].addAvg[NONALIGNED] = PFX(addAvg_16x16_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].addAvg[NONALIGNED] = PFX(addAvg_16x24_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].addAvg[NONALIGNED] = PFX(addAvg_16x32_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].addAvg[NONALIGNED] = PFX(addAvg_16x64_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].addAvg[NONALIGNED] = PFX(addAvg_24x64_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].addAvg[NONALIGNED] = PFX(addAvg_32x16_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].addAvg[NONALIGNED] = PFX(addAvg_32x32_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].addAvg[NONALIGNED] = PFX(addAvg_32x48_neon);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].addAvg[NONALIGNED] = PFX(addAvg_32x64_neon);

         // quant
          p.quant = PFX(quant_neon);
@@ -402,7 +403,7 @@ void setupAssemblyPrimitives(EncoderPrimitives &p, int cpuMask)
         p.scale2D_64to32  = PFX(scale2D_64to32_neon);

         // scale1D_128to64
-        p.scale1D_128to64 = PFX(scale1D_128to64_neon);
+        p.scale1D_128to64[NONALIGNED] = PFX(scale1D_128to64_neon);

         // copy_count
         p.cu<http://p.cu>[BLOCK_4x4].copy_cnt     = PFX(copy_cnt_4_neon);
@@ -411,37 +412,37 @@ void setupAssemblyPrimitives(EncoderPrimitives &p, int cpuMask)
         p.cu<http://p.cu>[BLOCK_32x32].copy_cnt   = PFX(copy_cnt_32_neon);

         // filterPixelToShort
-        p.pu[LUMA_4x4].convert_p2s   = PFX(filterPixelToShort_4x4_neon);
-        p.pu[LUMA_4x8].convert_p2s   = PFX(filterPixelToShort_4x8_neon);
-        p.pu[LUMA_4x16].convert_p2s  = PFX(filterPixelToShort_4x16_neon);
-        p.pu[LUMA_8x4].convert_p2s   = PFX(filterPixelToShort_8x4_neon);
-        p.pu[LUMA_8x8].convert_p2s   = PFX(filterPixelToShort_8x8_neon);
-        p.pu[LUMA_8x16].convert_p2s  = PFX(filterPixelToShort_8x16_neon);
-        p.pu[LUMA_8x32].convert_p2s  = PFX(filterPixelToShort_8x32_neon);
-        p.pu[LUMA_12x16].convert_p2s = PFX(filterPixelToShort_12x16_neon);
-        p.pu[LUMA_16x4].convert_p2s  = PFX(filterPixelToShort_16x4_neon);
-        p.pu[LUMA_16x8].convert_p2s  = PFX(filterPixelToShort_16x8_neon);
-        p.pu[LUMA_16x12].convert_p2s = PFX(filterPixelToShort_16x12_neon);
-        p.pu[LUMA_16x16].convert_p2s = PFX(filterPixelToShort_16x16_neon);
-        p.pu[LUMA_16x32].convert_p2s = PFX(filterPixelToShort_16x32_neon);
-        p.pu[LUMA_16x64].convert_p2s = PFX(filterPixelToShort_16x64_neon);
-        p.pu[LUMA_24x32].convert_p2s = PFX(filterPixelToShort_24x32_neon);
-        p.pu[LUMA_32x8].convert_p2s  = PFX(filterPixelToShort_32x8_neon);
-        p.pu[LUMA_32x16].convert_p2s = PFX(filterPixelToShort_32x16_neon);
-        p.pu[LUMA_32x24].convert_p2s = PFX(filterPixelToShort_32x24_neon);
-        p.pu[LUMA_32x32].convert_p2s = PFX(filterPixelToShort_32x32_neon);
-        p.pu[LUMA_32x64].convert_p2s = PFX(filterPixelToShort_32x64_neon);
-        p.pu[LUMA_48x64].convert_p2s = PFX(filterPixelToShort_48x64_neon);
-        p.pu[LUMA_64x16].convert_p2s = PFX(filterPixelToShort_64x16_neon);
-        p.pu[LUMA_64x32].convert_p2s = PFX(filterPixelToShort_64x32_neon);
-        p.pu[LUMA_64x48].convert_p2s = PFX(filterPixelToShort_64x48_neon);
-        p.pu[LUMA_64x64].convert_p2s = PFX(filterPixelToShort_64x64_neon);
+        p.pu[LUMA_4x4].convert_p2s[NONALIGNED]   = PFX(filterPixelToShort_4x4_neon);
+        p.pu[LUMA_4x8].convert_p2s[NONALIGNED]   = PFX(filterPixelToShort_4x8_neon);
+        p.pu[LUMA_4x16].convert_p2s[NONALIGNED]  = PFX(filterPixelToShort_4x16_neon);
+        p.pu[LUMA_8x4].convert_p2s[NONALIGNED]   = PFX(filterPixelToShort_8x4_neon);
+        p.pu[LUMA_8x8].convert_p2s[NONALIGNED]   = PFX(filterPixelToShort_8x8_neon);
+        p.pu[LUMA_8x16].convert_p2s[NONALIGNED]  = PFX(filterPixelToShort_8x16_neon);
+        p.pu[LUMA_8x32].convert_p2s[NONALIGNED]  = PFX(filterPixelToShort_8x32_neon);
+        p.pu[LUMA_12x16].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_12x16_neon);
+        p.pu[LUMA_16x4].convert_p2s[NONALIGNED]  = PFX(filterPixelToShort_16x4_neon);
+        p.pu[LUMA_16x8].convert_p2s[NONALIGNED]  = PFX(filterPixelToShort_16x8_neon);
+        p.pu[LUMA_16x12].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_16x12_neon);
+        p.pu[LUMA_16x16].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_16x16_neon);
+        p.pu[LUMA_16x32].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_16x32_neon);
+        p.pu[LUMA_16x64].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_16x64_neon);
+        p.pu[LUMA_24x32].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_24x32_neon);
+        p.pu[LUMA_32x8].convert_p2s[NONALIGNED]  = PFX(filterPixelToShort_32x8_neon);
+        p.pu[LUMA_32x16].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_32x16_neon);
+        p.pu[LUMA_32x24].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_32x24_neon);
+        p.pu[LUMA_32x32].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_32x32_neon);
+        p.pu[LUMA_32x64].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_32x64_neon);
+        p.pu[LUMA_48x64].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_48x64_neon);
+        p.pu[LUMA_64x16].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_64x16_neon);
+        p.pu[LUMA_64x32].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_64x32_neon);
+        p.pu[LUMA_64x48].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_64x48_neon);
+        p.pu[LUMA_64x64].convert_p2s[NONALIGNED] = PFX(filterPixelToShort_64x64_neon);

         // Block_fill
-        p.cu<http://p.cu>[BLOCK_4x4].blockfill_s   = PFX(blockfill_s_4x4_neon);
-        p.cu<http://p.cu>[BLOCK_8x8].blockfill_s   = PFX(blockfill_s_8x8_neon);
-        p.cu<http://p.cu>[BLOCK_16x16].blockfill_s = PFX(blockfill_s_16x16_neon);
-        p.cu<http://p.cu>[BLOCK_32x32].blockfill_s = PFX(blockfill_s_32x32_neon);
+        p.cu<http://p.cu>[BLOCK_4x4].blockfill_s[NONALIGNED]   = PFX(blockfill_s_4x4_neon);
+        p.cu<http://p.cu>[BLOCK_8x8].blockfill_s[NONALIGNED]   = PFX(blockfill_s_8x8_neon);
+        p.cu<http://p.cu>[BLOCK_16x16].blockfill_s[NONALIGNED] = PFX(blockfill_s_16x16_neon);
+        p.cu<http://p.cu>[BLOCK_32x32].blockfill_s[NONALIGNED] = PFX(blockfill_s_32x32_neon);

         // Blockcopy_ss
         p.cu<http://p.cu>[BLOCK_4x4].copy_ss   = PFX(blockcopy_ss_4x4_neon);
@@ -495,21 +496,21 @@ void setupAssemblyPrimitives(EncoderPrimitives &p, int cpuMask)
         p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].copy_sp = PFX(blockcopy_sp_32x64_neon);

         // pixel_add_ps
-        p.cu<http://p.cu>[BLOCK_4x4].add_ps   = PFX(pixel_add_ps_4x4_neon);
-        p.cu<http://p.cu>[BLOCK_8x8].add_ps   = PFX(pixel_add_ps_8x8_neon);
-        p.cu<http://p.cu>[BLOCK_16x16].add_ps = PFX(pixel_add_ps_16x16_neon);
-        p.cu<http://p.cu>[BLOCK_32x32].add_ps = PFX(pixel_add_ps_32x32_neon);
-        p.cu<http://p.cu>[BLOCK_64x64].add_ps = PFX(pixel_add_ps_64x64_neon);
+        p.cu<http://p.cu>[BLOCK_4x4].add_ps[NONALIGNED]   = PFX(pixel_add_ps_4x4_neon);
+        p.cu<http://p.cu>[BLOCK_8x8].add_ps[NONALIGNED]   = PFX(pixel_add_ps_8x8_neon);
+        p.cu<http://p.cu>[BLOCK_16x16].add_ps[NONALIGNED] = PFX(pixel_add_ps_16x16_neon);
+        p.cu<http://p.cu>[BLOCK_32x32].add_ps[NONALIGNED] = PFX(pixel_add_ps_32x32_neon);
+        p.cu<http://p.cu>[BLOCK_64x64].add_ps[NONALIGNED] = PFX(pixel_add_ps_64x64_neon);

         // chroma add_ps
-        p.chroma[X265_CSP_I420].cu[BLOCK_420_4x4].add_ps   = PFX(pixel_add_ps_4x4_neon);
-        p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].add_ps   = PFX(pixel_add_ps_8x8_neon);
-        p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].add_ps = PFX(pixel_add_ps_16x16_neon);
-        p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].add_ps = PFX(pixel_add_ps_32x32_neon);
-        p.chroma[X265_CSP_I422].cu[BLOCK_422_4x8].add_ps   = PFX(pixel_add_ps_4x8_neon);
-        p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].add_ps  = PFX(pixel_add_ps_8x16_neon);
-        p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].add_ps = PFX(pixel_add_ps_16x32_neon);
-        p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].add_ps = PFX(pixel_add_ps_32x64_neon);
+        p.chroma[X265_CSP_I420].cu[BLOCK_420_4x4].add_ps[NONALIGNED]   = PFX(pixel_add_ps_4x4_neon);
+        p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].add_ps[NONALIGNED]   = PFX(pixel_add_ps_8x8_neon);
+        p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].add_ps[NONALIGNED] = PFX(pixel_add_ps_16x16_neon);
+        p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].add_ps[NONALIGNED] = PFX(pixel_add_ps_32x32_neon);
+        p.chroma[X265_CSP_I422].cu[BLOCK_422_4x8].add_ps[NONALIGNED]   = PFX(pixel_add_ps_4x8_neon);
+        p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].add_ps[NONALIGNED]  = PFX(pixel_add_ps_8x16_neon);
+        p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].add_ps[NONALIGNED] = PFX(pixel_add_ps_16x32_neon);
+        p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].add_ps[NONALIGNED] = PFX(pixel_add_ps_32x64_neon);

         // cpy2Dto1D_shr
         p.cu<http://p.cu>[BLOCK_4x4].cpy2Dto1D_shr   = PFX(cpy2Dto1D_shr_4x4_neon);
@@ -518,10 +519,10 @@ void setupAssemblyPrimitives(EncoderPrimitives &p, int cpuMask)
         p.cu<http://p.cu>[BLOCK_32x32].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_32x32_neon);

         // ssd_s
-        p.cu<http://p.cu>[BLOCK_4x4].ssd_s   = PFX(pixel_ssd_s_4x4_neon);
-        p.cu<http://p.cu>[BLOCK_8x8].ssd_s   = PFX(pixel_ssd_s_8x8_neon);
-        p.cu<http://p.cu>[BLOCK_16x16].ssd_s = PFX(pixel_ssd_s_16x16_neon);
-        p.cu<http://p.cu>[BLOCK_32x32].ssd_s = PFX(pixel_ssd_s_32x32_neon);
+        p.cu<http://p.cu>[BLOCK_4x4].ssd_s[NONALIGNED]   = PFX(pixel_ssd_s_4x4_neon);
+        p.cu<http://p.cu>[BLOCK_8x8].ssd_s[NONALIGNED]   = PFX(pixel_ssd_s_8x8_neon);
+        p.cu<http://p.cu>[BLOCK_16x16].ssd_s[NONALIGNED] = PFX(pixel_ssd_s_16x16_neon);
+        p.cu<http://p.cu>[BLOCK_32x32].ssd_s[NONALIGNED] = PFX(pixel_ssd_s_32x32_neon);

         // sse_ss
         p.cu<http://p.cu>[BLOCK_4x4].sse_ss   = PFX(pixel_sse_ss_4x4_neon);
@@ -548,10 +549,10 @@ void setupAssemblyPrimitives(EncoderPrimitives &p, int cpuMask)
         p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].sub_ps = PFX(pixel_sub_ps_32x64_neon);

         // calc_Residual
-        p.cu<http://p.cu>[BLOCK_4x4].calcresidual   = PFX(getResidual4_neon);
-        p.cu<http://p.cu>[BLOCK_8x8].calcresidual   = PFX(getResidual8_neon);
-        p.cu<http://p.cu>[BLOCK_16x16].calcresidual = PFX(getResidual16_neon);
-        p.cu<http://p.cu>[BLOCK_32x32].calcresidual = PFX(getResidual32_neon);
+        p.cu<http://p.cu>[BLOCK_4x4].calcresidual[NONALIGNED]   = PFX(getResidual4_neon);
+        p.cu<http://p.cu>[BLOCK_8x8].calcresidual[NONALIGNED]   = PFX(getResidual8_neon);
+        p.cu<http://p.cu>[BLOCK_16x16].calcresidual[NONALIGNED] = PFX(getResidual16_neon);
+        p.cu<http://p.cu>[BLOCK_32x32].calcresidual[NONALIGNED] = PFX(getResidual32_neon);

         // sse_pp
         p.cu<http://p.cu>[BLOCK_4x4].sse_pp   = PFX(pixel_sse_pp_4x4_neon);
@@ -722,31 +723,31 @@ void setupAssemblyPrimitives(EncoderPrimitives &p, int cpuMask)
         p.pu[LUMA_64x64].sad_x4 = PFX(sad_x4_64x64_neon);

         // pixel_avg_pp
-        p.pu[LUMA_4x4].pixelavg_pp   = PFX(pixel_avg_pp_4x4_neon);
-        p.pu[LUMA_4x8].pixelavg_pp   = PFX(pixel_avg_pp_4x8_neon);
-        p.pu[LUMA_4x16].pixelavg_pp  = PFX(pixel_avg_pp_4x16_neon);
-        p.pu[LUMA_8x4].pixelavg_pp   = PFX(pixel_avg_pp_8x4_neon);
-        p.pu[LUMA_8x8].pixelavg_pp   = PFX(pixel_avg_pp_8x8_neon);
-        p.pu[LUMA_8x16].pixelavg_pp  = PFX(pixel_avg_pp_8x16_neon);
-        p.pu[LUMA_8x32].pixelavg_pp  = PFX(pixel_avg_pp_8x32_neon);
-        p.pu[LUMA_12x16].pixelavg_pp = PFX(pixel_avg_pp_12x16_neon);
-        p.pu[LUMA_16x4].pixelavg_pp  = PFX(pixel_avg_pp_16x4_neon);
-        p.pu[LUMA_16x8].pixelavg_pp  = PFX(pixel_avg_pp_16x8_neon);
-        p.pu[LUMA_16x12].pixelavg_pp = PFX(pixel_avg_pp_16x12_neon);
-        p.pu[LUMA_16x16].pixelavg_pp = PFX(pixel_avg_pp_16x16_neon);
-        p.pu[LUMA_16x32].pixelavg_pp = PFX(pixel_avg_pp_16x32_neon);
-        p.pu[LUMA_16x64].pixelavg_pp = PFX(pixel_avg_pp_16x64_neon);
-        p.pu[LUMA_24x32].pixelavg_pp = PFX(pixel_avg_pp_24x32_neon);
-        p.pu[LUMA_32x8].pixelavg_pp  = PFX(pixel_avg_pp_32x8_neon);
-        p.pu[LUMA_32x16].pixelavg_pp = PFX(pixel_avg_pp_32x16_neon);
-        p.pu[LUMA_32x24].pixelavg_pp = PFX(pixel_avg_pp_32x24_neon);
-        p.pu[LUMA_32x32].pixelavg_pp = PFX(pixel_avg_pp_32x32_neon);
-        p.pu[LUMA_32x64].pixelavg_pp = PFX(pixel_avg_pp_32x64_neon);
-        p.pu[LUMA_48x64].pixelavg_pp = PFX(pixel_avg_pp_48x64_neon);
-        p.pu[LUMA_64x16].pixelavg_pp = PFX(pixel_avg_pp_64x16_neon);
-        p.pu[LUMA_64x32].pixelavg_pp = PFX(pixel_avg_pp_64x32_neon);
-        p.pu[LUMA_64x48].pixelavg_pp = PFX(pixel_avg_pp_64x48_neon);
-        p.pu[LUMA_64x64].pixelavg_pp = PFX(pixel_avg_pp_64x64_neon);
+        p.pu[LUMA_4x4].pixelavg_pp[NONALIGNED]   = PFX(pixel_avg_pp_4x4_neon);
+        p.pu[LUMA_4x8].pixelavg_pp[NONALIGNED]   = PFX(pixel_avg_pp_4x8_neon);
+        p.pu[LUMA_4x16].pixelavg_pp[NONALIGNED]  = PFX(pixel_avg_pp_4x16_neon);
+        p.pu[LUMA_8x4].pixelavg_pp[NONALIGNED]   = PFX(pixel_avg_pp_8x4_neon);
+        p.pu[LUMA_8x8].pixelavg_pp[NONALIGNED]   = PFX(pixel_avg_pp_8x8_neon);
+        p.pu[LUMA_8x16].pixelavg_pp[NONALIGNED]  = PFX(pixel_avg_pp_8x16_neon);
+        p.pu[LUMA_8x32].pixelavg_pp[NONALIGNED]  = PFX(pixel_avg_pp_8x32_neon);
+        p.pu[LUMA_12x16].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_12x16_neon);
+        p.pu[LUMA_16x4].pixelavg_pp[NONALIGNED]  = PFX(pixel_avg_pp_16x4_neon);
+        p.pu[LUMA_16x8].pixelavg_pp[NONALIGNED]  = PFX(pixel_avg_pp_16x8_neon);
+        p.pu[LUMA_16x12].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_16x12_neon);
+        p.pu[LUMA_16x16].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_16x16_neon);
+        p.pu[LUMA_16x32].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_16x32_neon);
+        p.pu[LUMA_16x64].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_16x64_neon);
+        p.pu[LUMA_24x32].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_24x32_neon);
+        p.pu[LUMA_32x8].pixelavg_pp[NONALIGNED]  = PFX(pixel_avg_pp_32x8_neon);
+        p.pu[LUMA_32x16].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_32x16_neon);
+        p.pu[LUMA_32x24].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_32x24_neon);
+        p.pu[LUMA_32x32].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_32x32_neon);
+        p.pu[LUMA_32x64].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_32x64_neon);
+        p.pu[LUMA_48x64].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_48x64_neon);
+        p.pu[LUMA_64x16].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_64x16_neon);
+        p.pu[LUMA_64x32].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_64x32_neon);
+        p.pu[LUMA_64x48].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_64x48_neon);
+        p.pu[LUMA_64x64].pixelavg_pp[NONALIGNED] = PFX(pixel_avg_pp_64x64_neon);

         // planecopy
         p.planecopy_cp = PFX(pixel_planecopy_cp_neon);
diff --git a/source/common/cpu.cpp b/source/common/cpu.cpp
index 26c82ea50..2eacfe4a9 100644
--- a/source/common/cpu.cpp
+++ b/source/common/cpu.cpp
@@ -5,6 +5,8 @@
  *          Laurent Aimar <fenrir at via.ecp.fr<mailto:fenrir at via.ecp.fr>>
  *          Fiona Glaser <fiona at x264.com<mailto:fiona at x264.com>>
  *          Steve Borho <steve at borho.org<mailto:steve at borho.org>>
+ *          Hongbin Liu <liuhongbin1 at huawei.com<mailto:liuhongbin1 at huawei.com>>
+ *          Yimeng Su <yimeng.su at huawei.com<mailto:yimeng.su at huawei.com>>
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by
@@ -367,6 +369,8 @@ uint32_t cpu_detect(bool benableavx512)
     flags |= PFX(cpu_fast_neon_mrc_test)() ? X265_CPU_FAST_NEON_MRC : 0;
 #endif
     // TODO: write dual issue test? currently it's A8 (dual issue) vs. A9 (fast mrc)
+#elif X265_ARCH_ARM64
+    flags |= X265_CPU_NEON;
 #endif // if HAVE_ARMV6
     return flags;
 }
diff --git a/source/common/pixel.cpp b/source/common/pixel.cpp
index 99b84449c..e4f890cd5 100644
--- a/source/common/pixel.cpp
+++ b/source/common/pixel.cpp
@@ -5,6 +5,7 @@
  *          Mandar Gurav <mandar at multicorewareinc.com<mailto:mandar at multicorewareinc.com>>
  *          Mahesh Pittala <mahesh at multicorewareinc.com<mailto:mahesh at multicorewareinc.com>>
  *          Min Chen <min.chen at multicorewareinc.com<mailto:min.chen at multicorewareinc.com>>
+ *          Hongbin Liu<liuhongbin1 at huawei.com<mailto:liuhongbin1 at huawei.com>>
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by
@@ -265,6 +266,10 @@ int satd4(const pixel* pix1, intptr_t stride_pix1, const pixel* pix2, intptr_t s
 {
     int satd = 0;

+#if ENABLE_ASSEMBLY && X265_ARCH_ARM64
+    pixelcmp_t satd_4x4 = x265_pixel_satd_4x4_neon;
+#endif
is there any specific reason why the above code is added?? is this a kind of a temporary fix for the output mismatch between c and asm code?
No, c and asm output is matched. Currently we only complete partial satd primatives. This is a workaround that improve all satd primitives with asm code. Maybe there is a bad code style.

+
     for (int row = 0; row < h; row += 4)
         for (int col = 0; col < w; col += 4)
             satd += satd_4x4(pix1 + row * stride_pix1 + col, stride_pix1,
@@ -279,6 +284,10 @@ int satd8(const pixel* pix1, intptr_t stride_pix1, const pixel* pix2, intptr_t s
 {
     int satd = 0;

+#if ENABLE_ASSEMBLY && X265_ARCH_ARM64
+    pixelcmp_t satd_8x4 = x265_pixel_satd_8x4_neon;
+#endif
+
Same comment as above.
Same response.
     for (int row = 0; row < h; row += 4)
         for (int col = 0; col < w; col += 8)
             satd += satd_8x4(pix1 + row * stride_pix1 + col, stride_pix1,
diff --git a/source/common/primitives.h b/source/common/primitives.h
index 5c64952fb..0b52f84de 100644
--- a/source/common/primitives.h
+++ b/source/common/primitives.h
@@ -8,6 +8,8 @@
  *          Rajesh Paulraj <rajesh at multicorewareinc.com<mailto:rajesh at multicorewareinc.com>>
  *          Praveen Kumar Tiwari <praveen at multicorewareinc.com<mailto:praveen at multicorewareinc.com>>
  *          Min Chen <chenm003 at 163.com<mailto:chenm003 at 163.com>>
+ *          Hongbin Liu<liuhongbin1 at huawei.com<mailto:liuhongbin1 at huawei.com>>
+ *          Yimeng Su <yimeng.su at huawei.com<mailto:yimeng.su at huawei.com>>
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by
@@ -467,6 +469,9 @@ void setupCPrimitives(EncoderPrimitives &p);
 void setupInstrinsicPrimitives(EncoderPrimitives &p, int cpuMask);
 void setupAssemblyPrimitives(EncoderPrimitives &p, int cpuMask);
 void setupAliasPrimitives(EncoderPrimitives &p);
+#if X265_ARCH_ARM64
+void setupAliasCPrimitives(EncoderPrimitives &cp, EncoderPrimitives &asmp, int cpuMask);
+#endif
 #if HAVE_ALTIVEC
 void setupPixelPrimitives_altivec(EncoderPrimitives &p);
 void setupDCTPrimitives_altivec(EncoderPrimitives &p);
@@ -481,4 +486,10 @@ extern const char* PFX(version_str);
 extern const char* PFX(build_info_str);
 #endif

+#if ENABLE_ASSEMBLY && X265_ARCH_ARM64
+extern "C" {
+#include "aarch64/pixel-util.h"
+}
+#endif
+
 #endif // ifndef X265_PRIMITIVES_H
diff --git a/source/test/CMakeLists.txt b/source/test/CMakeLists.txt
index 260195f53..9abaf31ff 100644
--- a/source/test/CMakeLists.txt
+++ b/source/test/CMakeLists.txt
@@ -23,13 +23,15 @@ endif(X86)

 # add ARM assembly files
 if(ARM OR CROSS_COMPILE_ARM)
-    enable_language(ASM)
-    set(NASM_SRC checkasm-arm.S)
-    add_custom_command(
-        OUTPUT checkasm-arm.obj
-        COMMAND ${CMAKE_CXX_COMPILER}
-        ARGS ${NASM_FLAGS} ${CMAKE_CURRENT_SOURCE_DIR}/checkasm-arm.S -o checkasm-arm.obj
-        DEPENDS checkasm-arm.S)
+    if(NOT ARM64)
+        enable_language(ASM)
+        set(NASM_SRC checkasm-arm.S)
+        add_custom_command(
+            OUTPUT checkasm-arm.obj
+            COMMAND ${CMAKE_CXX_COMPILER}
+            ARGS ${NASM_FLAGS} ${CMAKE_CURRENT_SOURCE_DIR}/checkasm-arm.S -o checkasm-arm.obj
+            DEPENDS checkasm-arm.S)
+    endif()
 endif(ARM OR CROSS_COMPILE_ARM)

 # add PowerPC assembly files
diff --git a/source/test/testbench.cpp b/source/test/testbench.cpp
index ac14f9710..8db8c0c25 100644
--- a/source/test/testbench.cpp
+++ b/source/test/testbench.cpp
@@ -5,6 +5,7 @@
  *          Mandar Gurav <mandar at multicorewareinc.com<mailto:mandar at multicorewareinc.com>>
  *          Mahesh Pittala <mahesh at multicorewareinc.com<mailto:mahesh at multicorewareinc.com>>
  *          Min Chen <chenm003 at 163.com<mailto:chenm003 at 163.com>>
+ *          Yimeng Su <yimeng.su at huawei.com<mailto:yimeng.su at huawei.com>>
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by
@@ -208,6 +209,14 @@ int main(int argc, char *argv[])
         EncoderPrimitives asmprim;
         memset(&asmprim, 0, sizeof(asmprim));
         setupAssemblyPrimitives(asmprim, test_arch[i].flag);
+
+#if X265_ARCH_ARM64
+        /* Temporary workaround because luma_vsp assembly primitive has not been completed
+         * but interp_8tap_hv_pp_cpu uses mixed C primitive and assembly primitive.
+         * Otherwise, segment fault occurs. */
+        setupAliasCPrimitives(cprim, asmprim, test_arch[i].flag);
+#endif
+
         setupAliasPrimitives(asmprim);
         memcpy(&primitives, &asmprim, sizeof(EncoderPrimitives));
         for (size_t h = 0; h < sizeof(harness) / sizeof(TestHarness*); h++)
@@ -232,6 +241,13 @@ int main(int argc, char *argv[])
 #endif
     setupAssemblyPrimitives(optprim, cpuid);

+#if X265_ARCH_ARM64
+    /* Temporary workaround because luma_vsp assembly primitive has not been completed
+     * but interp_8tap_hv_pp_cpu uses mixed C primitive and assembly primitive.
+     * Otherwise, segment fault occurs. */
+    setupAliasCPrimitives(cprim, optprim, cpuid);
+#endif
+
     /* Note that we do not setup aliases for performance tests, that would be
      * redundant. The testbench only verifies they are correctly aliased */

diff --git a/source/test/testharness.h b/source/test/testharness.h
index 771551583..6e680953f 100644
--- a/source/test/testharness.h
+++ b/source/test/testharness.h
@@ -3,6 +3,7 @@
  *
  * Authors: Steve Borho <steve at borho.org<mailto:steve at borho.org>>
  *          Min Chen <chenm003 at 163.com<mailto:chenm003 at 163.com>>
+ *          Yimeng Su <yimeng.su at huawei.com<mailto:yimeng.su at huawei.com>>
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by
@@ -81,11 +82,15 @@ static inline uint32_t __rdtsc(void)
 #if X265_ARCH_X86
     asm volatile("rdtsc" : "=a" (a) ::"edx");
 #elif X265_ARCH_ARM
+#if X265_ARCH_ARM64
+    asm volatile("mrs %0, cntvct_el0" : "=r"(a));
+#else
     // TOD-DO: verify following inline asm to get cpu Timestamp Counter for ARM arch
     // asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r"(a));

     // TO-DO: replace clock() function with appropriate ARM cpu instructions
     a = clock();
+#endif
 #endif
     return a;
 }
--
2.21.0.windows.1
_______________________________________________
x265-devel mailing list
x265-devel at videolan.org<mailto:x265-devel at videolan.org>
https://mailman.videolan.org/listinfo/x265-devel
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20200317/50c964f4/attachment-0001.html>


More information about the x265-devel mailing list