--- /dev/null
+diff -Naur ./source/CMakeLists.txt ../x265_apple_patch/source/CMakeLists.txt
+--- ./source/CMakeLists.txt 2021-05-08 13:06:22.000000000 +0100
++++ ../x265_apple_patch/source/CMakeLists.txt 2021-05-08 13:08:01.000000000 +0100
+@@ -40,9 +40,11 @@
+ # System architecture detection
+ string(TOLOWER "${CMAKE_SYSTEM_PROCESSOR}" SYSPROC)
+ set(X86_ALIASES x86 i386 i686 x86_64 amd64)
+-set(ARM_ALIASES armv6l armv7l aarch64)
++set(ARM_ALIASES armv6l armv7l)
++set(ARM64_ALIASES arm64 arm64e aarch64)
+ list(FIND X86_ALIASES "${SYSPROC}" X86MATCH)
+ list(FIND ARM_ALIASES "${SYSPROC}" ARMMATCH)
++list(FIND ARM64_ALIASES "${SYSPROC}" ARM64MATCH)
+ set(POWER_ALIASES ppc64 ppc64le)
+ list(FIND POWER_ALIASES "${SYSPROC}" POWERMATCH)
+ if("${SYSPROC}" STREQUAL "" OR X86MATCH GREATER "-1")
+@@ -79,6 +81,15 @@
+ message(STATUS "Detected ARM target processor")
+ add_definitions(-DX265_ARCH_ARM=1 -DX265_ARCH_ARM64=0 -DHAVE_ARMV6=1)
+ endif()
++elseif(ARM64MATCH GREATER "-1")
++ if(CROSS_COMPILE_ARM64)
++ message(STATUS "Cross compiling for ARM64 arch")
++ else()
++ set(CROSS_COMPILE_ARM64 0)
++ endif()
++ message(STATUS "Detected ARM64 target processor")
++ set(ARM64 1)
++ add_definitions(-DX265_ARCH_ARM64=1 -DHAVE_NEON)
+ 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}")
+@@ -259,6 +270,9 @@
+ endif()
+ endif()
+ endif()
++ if(ARM64 OR CROSS_COMPILE_ARM64)
++ add_definitions(-DHAVE_NEON)
++ endif()
+ add_definitions(${ARM_ARGS})
+ if(FPROFILE_GENERATE)
+ if(INTEL_CXX)
+@@ -350,7 +364,7 @@
+ endif(GCC)
+
+ find_package(Nasm)
+-if(ARM OR CROSS_COMPILE_ARM)
++if(ARM OR CROSS_COMPILE_ARM OR ARM64 OR CROSS_COMPILE_ARM64)
+ option(ENABLE_ASSEMBLY "Enable use of assembly coded primitives" ON)
+ elseif(NASM_FOUND AND X86)
+ if (NASM_VERSION_STRING VERSION_LESS "2.13.0")
+@@ -549,6 +563,32 @@
+ ARGS ${ARM_ARGS} -c ${ASM_SRC} -o ${ASM}.${SUFFIX}
+ DEPENDS ${ASM_SRC})
+ endforeach()
++ elseif(ARM64 OR CROSS_COMPILE_ARM64)
++ # compile ARM arch asm files here
++ enable_language(ASM)
++ foreach(ASM ${ARM_ASMS})
++ set(ASM_SRC ${CMAKE_CURRENT_SOURCE_DIR}/common/arm64/${ASM})
++ list(APPEND ASM_SRCS ${ASM_SRC})
++ list(APPEND ASM_OBJS ${ASM}.${SUFFIX})
++ add_custom_command(
++ OUTPUT ${ASM}.${SUFFIX}
++ COMMAND ${CMAKE_CXX_COMPILER}
++ ARGS ${ARM_ARGS} -c ${ASM_SRC} -o ${ASM}.${SUFFIX}
++ DEPENDS ${ASM_SRC})
++ endforeach()
++ elseif(ARM64 OR CROSS_COMPILE_ARM64)
++ # compile ARM arch asm files here
++ enable_language(ASM)
++ foreach(ASM ${ARM_ASMS})
++ set(ASM_SRC ${CMAKE_CURRENT_SOURCE_DIR}/common/arm64/${ASM})
++ list(APPEND ASM_SRCS ${ASM_SRC})
++ list(APPEND ASM_OBJS ${ASM}.${SUFFIX})
++ add_custom_command(
++ OUTPUT ${ASM}.${SUFFIX}
++ COMMAND ${CMAKE_CXX_COMPILER}
++ ARGS ${ARM_ARGS} -c ${ASM_SRC} -o ${ASM}.${SUFFIX}
++ DEPENDS ${ASM_SRC})
++ endforeach()
+ elseif(X86)
+ # compile X86 arch asm files here
+ foreach(ASM ${MSVC_ASMS})
+diff -Naur ./source/common/CMakeLists.txt ../x265_apple_patch/source/common/CMakeLists.txt
+--- ./source/common/CMakeLists.txt 2021-05-08 13:06:22.000000000 +0100
++++ ../x265_apple_patch/source/common/CMakeLists.txt 2021-05-08 13:08:01.000000000 +0100
+@@ -114,6 +114,22 @@
+ source_group(Assembly FILES ${ASM_PRIMITIVES})
+ endif(ENABLE_ASSEMBLY AND (ARM OR CROSS_COMPILE_ARM))
+
++
++if(ENABLE_ASSEMBLY AND (ARM64 OR CROSS_COMPILE_ARM64))
++ set(C_SRCS asm-primitives.cpp pixel-prim.h pixel-prim.cpp filter-prim.h filter-prim.cpp dct-prim.h dct-prim.cpp loopfilter-prim.cpp loopfilter-prim.h intrapred-prim.cpp arm64-utils.cpp arm64-utils.h)
++ enable_language(ASM)
++ # add ARM assembly/intrinsic files here
++ #set(A_SRCS )
++ #set(VEC_PRIMITIVES)
++
++ #set(ARM64_ASMS "${A_SRCS}" CACHE INTERNAL "ARM64 Assembly Sources")
++ foreach(SRC ${C_SRCS})
++ set(ASM_PRIMITIVES ${ASM_PRIMITIVES} arm64/${SRC})
++ endforeach()
++ source_group(Assembly FILES ${ASM_PRIMITIVES})
++endif(ENABLE_ASSEMBLY AND (ARM64 OR CROSS_COMPILE_ARM64))
++
++
+ if(POWER)
+ set_source_files_properties(version.cpp PROPERTIES COMPILE_FLAGS -DX265_VERSION=${X265_VERSION})
+ if(ENABLE_ALTIVEC)
+diff -Naur ./source/common/arm64/arm64-utils.cpp ../x265_apple_patch/source/common/arm64/arm64-utils.cpp
+--- ./source/common/arm64/arm64-utils.cpp 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/arm64-utils.cpp 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,290 @@
++#include "common.h"
++#include "x265.h"
++#include "arm64-utils.h"
++#include <arm_neon.h>
++
++#define COPY_16(d,s) *(uint8x16_t *)(d) = *(uint8x16_t *)(s)
++namespace X265_NS {
++
++
++
++void transpose8x8(uint8_t* dst, const uint8_t* src, intptr_t dstride, intptr_t sstride)
++{
++ uint8x8_t a0,a1,a2,a3,a4,a5,a6,a7;
++ uint8x8_t b0,b1,b2,b3,b4,b5,b6,b7;
++
++ a0 = *(uint8x8_t *)(src + 0*sstride);
++ a1 = *(uint8x8_t *)(src + 1*sstride);
++ a2 = *(uint8x8_t *)(src + 2*sstride);
++ a3 = *(uint8x8_t *)(src + 3*sstride);
++ a4 = *(uint8x8_t *)(src + 4*sstride);
++ a5 = *(uint8x8_t *)(src + 5*sstride);
++ a6 = *(uint8x8_t *)(src + 6*sstride);
++ a7 = *(uint8x8_t *)(src + 7*sstride);
++
++ b0 = vtrn1_u32(a0,a4);
++ b1 = vtrn1_u32(a1,a5);
++ b2 = vtrn1_u32(a2,a6);
++ b3 = vtrn1_u32(a3,a7);
++ b4 = vtrn2_u32(a0,a4);
++ b5 = vtrn2_u32(a1,a5);
++ b6 = vtrn2_u32(a2,a6);
++ b7 = vtrn2_u32(a3,a7);
++
++ a0 = vtrn1_u16(b0,b2);
++ a1 = vtrn1_u16(b1,b3);
++ a2 = vtrn2_u16(b0,b2);
++ a3 = vtrn2_u16(b1,b3);
++ a4 = vtrn1_u16(b4,b6);
++ a5 = vtrn1_u16(b5,b7);
++ a6 = vtrn2_u16(b4,b6);
++ a7 = vtrn2_u16(b5,b7);
++
++ b0 = vtrn1_u8(a0,a1);
++ b1 = vtrn2_u8(a0,a1);
++ b2 = vtrn1_u8(a2,a3);
++ b3 = vtrn2_u8(a2,a3);
++ b4 = vtrn1_u8(a4,a5);
++ b5 = vtrn2_u8(a4,a5);
++ b6 = vtrn1_u8(a6,a7);
++ b7 = vtrn2_u8(a6,a7);
++
++ *(uint8x8_t *)(dst + 0*dstride) = b0;
++ *(uint8x8_t *)(dst + 1*dstride) = b1;
++ *(uint8x8_t *)(dst + 2*dstride) = b2;
++ *(uint8x8_t *)(dst + 3*dstride) = b3;
++ *(uint8x8_t *)(dst + 4*dstride) = b4;
++ *(uint8x8_t *)(dst + 5*dstride) = b5;
++ *(uint8x8_t *)(dst + 6*dstride) = b6;
++ *(uint8x8_t *)(dst + 7*dstride) = b7;
++}
++
++
++
++
++
++
++void transpose16x16(uint8_t* dst, const uint8_t* src, intptr_t dstride, intptr_t sstride)
++{
++ uint16x8_t a0,a1,a2,a3,a4,a5,a6,a7,a8,a9,aA,aB,aC,aD,aE,aF;
++ uint16x8_t b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,bA,bB,bC,bD,bE,bF;
++ uint16x8_t c0,c1,c2,c3,c4,c5,c6,c7,c8,c9,cA,cB,cC,cD,cE,cF;
++ uint16x8_t d0,d1,d2,d3,d4,d5,d6,d7,d8,d9,dA,dB,dC,dD,dE,dF;
++
++ a0 = *(uint16x8_t *)(src + 0*sstride);
++ a1 = *(uint16x8_t *)(src + 1*sstride);
++ a2 = *(uint16x8_t *)(src + 2*sstride);
++ a3 = *(uint16x8_t *)(src + 3*sstride);
++ a4 = *(uint16x8_t *)(src + 4*sstride);
++ a5 = *(uint16x8_t *)(src + 5*sstride);
++ a6 = *(uint16x8_t *)(src + 6*sstride);
++ a7 = *(uint16x8_t *)(src + 7*sstride);
++ a8 = *(uint16x8_t *)(src + 8*sstride);
++ a9 = *(uint16x8_t *)(src + 9*sstride);
++ aA = *(uint16x8_t *)(src + 10*sstride);
++ aB = *(uint16x8_t *)(src + 11*sstride);
++ aC = *(uint16x8_t *)(src + 12*sstride);
++ aD = *(uint16x8_t *)(src + 13*sstride);
++ aE = *(uint16x8_t *)(src + 14*sstride);
++ aF = *(uint16x8_t *)(src + 15*sstride);
++
++ b0 = vtrn1q_u64(a0, a8);
++ b1 = vtrn1q_u64(a1, a9);
++ b2 = vtrn1q_u64(a2, aA);
++ b3 = vtrn1q_u64(a3, aB);
++ b4 = vtrn1q_u64(a4, aC);
++ b5 = vtrn1q_u64(a5, aD);
++ b6 = vtrn1q_u64(a6, aE);
++ b7 = vtrn1q_u64(a7, aF);
++ b8 = vtrn2q_u64(a0, a8);
++ b9 = vtrn2q_u64(a1, a9);
++ bA = vtrn2q_u64(a2, aA);
++ bB = vtrn2q_u64(a3, aB);
++ bC = vtrn2q_u64(a4, aC);
++ bD = vtrn2q_u64(a5, aD);
++ bE = vtrn2q_u64(a6, aE);
++ bF = vtrn2q_u64(a7, aF);
++
++ c0 = vtrn1q_u32(b0, b4);
++ c1 = vtrn1q_u32(b1, b5);
++ c2 = vtrn1q_u32(b2, b6);
++ c3 = vtrn1q_u32(b3, b7);
++ c4 = vtrn2q_u32(b0, b4);
++ c5 = vtrn2q_u32(b1, b5);
++ c6 = vtrn2q_u32(b2, b6);
++ c7 = vtrn2q_u32(b3, b7);
++ c8 = vtrn1q_u32(b8, bC);
++ c9 = vtrn1q_u32(b9, bD);
++ cA = vtrn1q_u32(bA, bE);
++ cB = vtrn1q_u32(bB, bF);
++ cC = vtrn2q_u32(b8, bC);
++ cD = vtrn2q_u32(b9, bD);
++ cE = vtrn2q_u32(bA, bE);
++ cF = vtrn2q_u32(bB, bF);
++
++ d0 = vtrn1q_u16(c0, c2);
++ d1 = vtrn1q_u16(c1, c3);
++ d2 = vtrn2q_u16(c0, c2);
++ d3 = vtrn2q_u16(c1, c3);
++ d4 = vtrn1q_u16(c4, c6);
++ d5 = vtrn1q_u16(c5, c7);
++ d6 = vtrn2q_u16(c4, c6);
++ d7 = vtrn2q_u16(c5, c7);
++ d8 = vtrn1q_u16(c8, cA);
++ d9 = vtrn1q_u16(c9, cB);
++ dA = vtrn2q_u16(c8, cA);
++ dB = vtrn2q_u16(c9, cB);
++ dC = vtrn1q_u16(cC, cE);
++ dD = vtrn1q_u16(cD, cF);
++ dE = vtrn2q_u16(cC, cE);
++ dF = vtrn2q_u16(cD, cF);
++
++ *(uint16x8_t *)(dst + 0*dstride) = vtrn1q_u8(d0, d1);
++ *(uint16x8_t *)(dst + 1*dstride) = vtrn2q_u8(d0, d1);
++ *(uint16x8_t *)(dst + 2*dstride) = vtrn1q_u8(d2, d3);
++ *(uint16x8_t *)(dst + 3*dstride) = vtrn2q_u8(d2, d3);
++ *(uint16x8_t *)(dst + 4*dstride) = vtrn1q_u8(d4, d5);
++ *(uint16x8_t *)(dst + 5*dstride) = vtrn2q_u8(d4, d5);
++ *(uint16x8_t *)(dst + 6*dstride) = vtrn1q_u8(d6, d7);
++ *(uint16x8_t *)(dst + 7*dstride) = vtrn2q_u8(d6, d7);
++ *(uint16x8_t *)(dst + 8*dstride) = vtrn1q_u8(d8, d9);
++ *(uint16x8_t *)(dst + 9*dstride) = vtrn2q_u8(d8, d9);
++ *(uint16x8_t *)(dst + 10*dstride) = vtrn1q_u8(dA, dB);
++ *(uint16x8_t *)(dst + 11*dstride) = vtrn2q_u8(dA, dB);
++ *(uint16x8_t *)(dst + 12*dstride) = vtrn1q_u8(dC, dD);
++ *(uint16x8_t *)(dst + 13*dstride) = vtrn2q_u8(dC, dD);
++ *(uint16x8_t *)(dst + 14*dstride) = vtrn1q_u8(dE, dF);
++ *(uint16x8_t *)(dst + 15*dstride) = vtrn2q_u8(dE, dF);
++
++
++}
++
++
++void transpose32x32(uint8_t* dst, const uint8_t* src, intptr_t dstride, intptr_t sstride)
++{
++ //assumption: there is no partial overlap
++ transpose16x16(dst,src,dstride,sstride);
++ transpose16x16(dst+16*dstride+16,src+16*sstride+16,dstride,sstride);
++ if (dst == src)
++ {
++ uint8_t tmp[16*16] __attribute__((aligned(64)));
++ transpose16x16(tmp,src + 16,16,sstride);
++ transpose16x16(dst + 16, src + 16*sstride,dstride,sstride);
++ for (int i=0;i<16;i++) COPY_16(dst+(16 + i)*dstride,tmp + 16*i);
++ }
++ else
++ {
++ transpose16x16(dst+16*dstride,src + 16,dstride,sstride);
++ transpose16x16(dst + 16, src + 16*sstride,dstride,sstride);
++ }
++
++}
++
++
++
++void transpose8x8(uint16_t* dst, const uint16_t* src, intptr_t dstride, intptr_t sstride)
++{
++ uint16x8_t a0,a1,a2,a3,a4,a5,a6,a7;
++ uint16x8_t b0,b1,b2,b3,b4,b5,b6,b7;
++
++ a0 = *(uint16x8_t *)(src + 0*sstride);
++ a1 = *(uint16x8_t *)(src + 1*sstride);
++ a2 = *(uint16x8_t *)(src + 2*sstride);
++ a3 = *(uint16x8_t *)(src + 3*sstride);
++ a4 = *(uint16x8_t *)(src + 4*sstride);
++ a5 = *(uint16x8_t *)(src + 5*sstride);
++ a6 = *(uint16x8_t *)(src + 6*sstride);
++ a7 = *(uint16x8_t *)(src + 7*sstride);
++
++ b0 = vtrn1q_u64(a0,a4);
++ b1 = vtrn1q_u64(a1,a5);
++ b2 = vtrn1q_u64(a2,a6);
++ b3 = vtrn1q_u64(a3,a7);
++ b4 = vtrn2q_u64(a0,a4);
++ b5 = vtrn2q_u64(a1,a5);
++ b6 = vtrn2q_u64(a2,a6);
++ b7 = vtrn2q_u64(a3,a7);
++
++ a0 = vtrn1q_u32(b0,b2);
++ a1 = vtrn1q_u32(b1,b3);
++ a2 = vtrn2q_u32(b0,b2);
++ a3 = vtrn2q_u32(b1,b3);
++ a4 = vtrn1q_u32(b4,b6);
++ a5 = vtrn1q_u32(b5,b7);
++ a6 = vtrn2q_u32(b4,b6);
++ a7 = vtrn2q_u32(b5,b7);
++
++ b0 = vtrn1q_u16(a0,a1);
++ b1 = vtrn2q_u16(a0,a1);
++ b2 = vtrn1q_u16(a2,a3);
++ b3 = vtrn2q_u16(a2,a3);
++ b4 = vtrn1q_u16(a4,a5);
++ b5 = vtrn2q_u16(a4,a5);
++ b6 = vtrn1q_u16(a6,a7);
++ b7 = vtrn2q_u16(a6,a7);
++
++ *(uint16x8_t *)(dst + 0*dstride) = b0;
++ *(uint16x8_t *)(dst + 1*dstride) = b1;
++ *(uint16x8_t *)(dst + 2*dstride) = b2;
++ *(uint16x8_t *)(dst + 3*dstride) = b3;
++ *(uint16x8_t *)(dst + 4*dstride) = b4;
++ *(uint16x8_t *)(dst + 5*dstride) = b5;
++ *(uint16x8_t *)(dst + 6*dstride) = b6;
++ *(uint16x8_t *)(dst + 7*dstride) = b7;
++}
++
++void transpose16x16(uint16_t* dst, const uint16_t* src, intptr_t dstride, intptr_t sstride)
++{
++ //assumption: there is no partial overlap
++ transpose8x8(dst,src,dstride,sstride);
++ transpose8x8(dst+8*dstride+8,src+8*sstride+8,dstride,sstride);
++
++ if (dst == src)
++ {
++ uint16_t tmp[8*8];
++ transpose8x8(tmp,src + 8,8,sstride);
++ transpose8x8(dst + 8, src + 8*sstride,dstride,sstride);
++ for (int i=0;i<8;i++) COPY_16(dst+(8 + i)*dstride,tmp + 8*i);
++ }
++ else
++ {
++ transpose8x8(dst+8*dstride,src + 8,dstride,sstride);
++ transpose8x8(dst + 8, src + 8*sstride,dstride,sstride);
++ }
++
++}
++
++
++
++void transpose32x32(uint16_t* dst, const uint16_t* src, intptr_t dstride, intptr_t sstride)
++{
++ //assumption: there is no partial overlap
++ for (int i=0;i<4;i++)
++ {
++ transpose8x8(dst+i*8*(1+dstride),src+i*8*(1+sstride),dstride,sstride);
++ for (int j=i+1;j<4;j++)
++ {
++ if (dst == src)
++ {
++ uint16_t tmp[8*8] __attribute__((aligned(64)));
++ transpose8x8(tmp,src + 8*i + 8*j*sstride,8,sstride);
++ transpose8x8(dst + 8*i + 8*j*dstride, src + 8*j + 8*i*sstride,dstride,sstride);
++ for (int k=0;k<8;k++) COPY_16(dst+ 8*j + (8*i+k)*dstride,tmp + 8*k);
++ }
++ else
++ {
++ transpose8x8(dst + 8*(j + i*dstride),src + 8*(i + j*sstride),dstride,sstride);
++ transpose8x8(dst + 8*(i + j*dstride),src + 8*(j + i*sstride),dstride,sstride);
++ }
++
++ }
++ }
++}
++
++
++
++
++}
++
++
++
+diff -Naur ./source/common/arm64/arm64-utils.h ../x265_apple_patch/source/common/arm64/arm64-utils.h
+--- ./source/common/arm64/arm64-utils.h 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/arm64-utils.h 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,14 @@
++#ifndef __ARM64_UTILS_H__
++#define __ARM64_UTILS_H__
++
++
++namespace X265_NS {
++void transpose8x8(uint8_t* dst, const uint8_t* src, intptr_t dstride, intptr_t sstride);
++void transpose16x16(uint8_t* dst, const uint8_t* src, intptr_t dstride, intptr_t sstride);
++void transpose32x32(uint8_t* dst, const uint8_t* src, intptr_t dstride, intptr_t sstride);
++void transpose8x8(uint16_t* dst, const uint16_t* src, intptr_t dstride, intptr_t sstride);
++void transpose16x16(uint16_t* dst, const uint16_t* src, intptr_t dstride, intptr_t sstride);
++void transpose32x32(uint16_t* dst, const uint16_t* src, intptr_t dstride, intptr_t sstride);
++}
++
++#endif
+diff -Naur ./source/common/arm64/asm-primitives.cpp ../x265_apple_patch/source/common/arm64/asm-primitives.cpp
+--- ./source/common/arm64/asm-primitives.cpp 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/asm-primitives.cpp 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,53 @@
++/*****************************************************************************
++ * Copyright (C) 2013-2017 MulticoreWare, Inc
++ *
++ * Authors: Steve Borho <steve@borho.org>
++ * Praveen Kumar Tiwari <praveen@multicorewareinc.com>
++ * Min Chen <chenm003@163.com> <min.chen@multicorewareinc.com>
++ * Dnyaneshwar Gorade <dnyaneshwar@multicorewareinc.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
++ * the Free Software Foundation; either version 2 of the License, or
++ * (at your option) any later version.
++ *
++ * This program is distributed in the hope that it will be useful,
++ * but WITHOUT ANY WARRANTY; without even the implied warranty of
++ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
++ * GNU General Public License for more details.
++ *
++ * You should have received a copy of the GNU General Public License
++ * along with this program; if not, write to the Free Software
++ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
++ *
++ * This program is also available under a commercial proprietary license.
++ * For more information, contact us at license @ x265.com.
++ *****************************************************************************/
++
++#include "common.h"
++#include "primitives.h"
++#include "x265.h"
++#include "cpu.h"
++
++#include "pixel-prim.h"
++#include "filter-prim.h"
++#include "dct-prim.h"
++#include "loopfilter-prim.h"
++#include "intrapred-prim.h"
++
++namespace X265_NS {
++// private x265 namespace
++
++void setupAssemblyPrimitives(EncoderPrimitives &p, int cpuMask)
++{
++ if (cpuMask & X265_CPU_NEON)
++ {
++ setupPixelPrimitives_neon(p);
++ setupFilterPrimitives_neon(p);
++ setupDCTPrimitives_neon(p);
++ setupLoopFilterPrimitives_neon(p);
++ setupIntraPrimitives_neon(p);
++ }
++}
++
++} // namespace X265_NS
+diff -Naur ./source/common/arm64/dct-prim.cpp ../x265_apple_patch/source/common/arm64/dct-prim.cpp
+--- ./source/common/arm64/dct-prim.cpp 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/dct-prim.cpp 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,933 @@
++#include "dct-prim.h"
++
++
++#if HAVE_NEON
++
++#include <arm_neon.h>
++
++
++namespace {
++using namespace X265_NS;
++
++
++static int16x8_t rev16(const int16x8_t a)
++{
++ static const int8x16_t tbl = {14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1};
++ return vqtbx1q_u8(a,a,tbl);
++}
++
++static int32x4_t rev32(const int32x4_t a)
++{
++ static const int8x16_t tbl = {12,13,14,15,8,9,10,11,4,5,6,7,0,1,2,3};
++ return vqtbx1q_u8(a,a,tbl);
++}
++
++static void transpose_4x4x16(int16x4_t& x0,int16x4_t& x1,int16x4_t& x2,int16x4_t& x3)
++{
++ int16x4_t s0,s1,s2,s3;
++ s0 = vtrn1_s32(x0,x2);
++ s1 = vtrn1_s32(x1,x3);
++ s2 = vtrn2_s32(x0,x2);
++ s3 = vtrn2_s32(x1,x3);
++
++ x0 = vtrn1_s16(s0,s1);
++ x1 = vtrn2_s16(s0,s1);
++ x2 = vtrn1_s16(s2,s3);
++ x3 = vtrn2_s16(s2,s3);
++}
++
++
++
++static int scanPosLast_opt(const uint16_t *scan, const coeff_t *coeff, uint16_t *coeffSign, uint16_t *coeffFlag, uint8_t *coeffNum, int numSig, const uint16_t* /*scanCG4x4*/, const int /*trSize*/)
++{
++
++ // This is an optimized function for scanPosLast, which removes the rmw dependency, once integrated into mainline x265, should replace reference implementation
++ // For clarity, left the original reference code in comments
++ int scanPosLast = 0;
++
++ uint16_t cSign = 0;
++ uint16_t cFlag = 0;
++ uint8_t cNum = 0;
++
++ uint32_t prevcgIdx = 0;
++ do
++ {
++ const uint32_t cgIdx = (uint32_t)scanPosLast >> MLS_CG_SIZE;
++
++ const uint32_t posLast = scan[scanPosLast];
++
++ const int curCoeff = coeff[posLast];
++ const uint32_t isNZCoeff = (curCoeff != 0);
++ /*
++ NOTE: the new algorithm is complicated, so I keep reference code here
++ uint32_t posy = posLast >> log2TrSize;
++ uint32_t posx = posLast - (posy << log2TrSize);
++ uint32_t blkIdx0 = ((posy >> MLS_CG_LOG2_SIZE) << codingParameters.log2TrSizeCG) + (posx >> MLS_CG_LOG2_SIZE);
++ const uint32_t blkIdx = ((posLast >> (2 * MLS_CG_LOG2_SIZE)) & ~maskPosXY) + ((posLast >> MLS_CG_LOG2_SIZE) & maskPosXY);
++ sigCoeffGroupFlag64 |= ((uint64_t)isNZCoeff << blkIdx);
++ */
++
++ // get L1 sig map
++ numSig -= isNZCoeff;
++
++ if (scanPosLast % (1<<MLS_CG_SIZE) == 0)
++ {
++ coeffSign[prevcgIdx] = cSign;
++ coeffFlag[prevcgIdx] = cFlag;
++ coeffNum[prevcgIdx] = cNum;
++ cSign = 0;
++ cFlag = 0;
++ cNum = 0;
++ }
++ // TODO: optimize by instruction BTS
++ cSign += (uint16_t)(((curCoeff < 0) ? 1 : 0) << cNum);
++ cFlag = (cFlag << 1) + (uint16_t)isNZCoeff;
++ cNum += (uint8_t)isNZCoeff;
++ prevcgIdx = cgIdx;
++ scanPosLast++;
++ }
++ while (numSig > 0);
++
++ coeffSign[prevcgIdx] = cSign;
++ coeffFlag[prevcgIdx] = cFlag;
++ coeffNum[prevcgIdx] = cNum;
++ return scanPosLast - 1;
++}
++
++
++#if (MLS_CG_SIZE == 4)
++template<int log2TrSize>
++static void nonPsyRdoQuant_neon(int16_t *m_resiDctCoeff, int64_t *costUncoded, int64_t *totalUncodedCost, int64_t *totalRdCost, uint32_t blkPos)
++{
++ const int transformShift = MAX_TR_DYNAMIC_RANGE - X265_DEPTH - log2TrSize; /* Represents scaling through forward transform */
++ const int scaleBits = SCALE_BITS - 2 * transformShift;
++ const uint32_t trSize = 1 << log2TrSize;
++
++ int64x2_t vcost_sum_0 = vdupq_n_s64(0);
++ int64x2_t vcost_sum_1 = vdupq_n_s64(0);
++ for (int y = 0; y < MLS_CG_SIZE; y++)
++ {
++ int16x4_t in = *(int16x4_t *)&m_resiDctCoeff[blkPos];
++ int32x4_t mul = vmull_s16(in,in);
++ int64x2_t cost0, cost1;
++ cost0 = vshll_n_s32(vget_low_s32(mul),scaleBits);
++ cost1 = vshll_high_n_s32(mul,scaleBits);
++ *(int64x2_t *)&costUncoded[blkPos+0] = cost0;
++ *(int64x2_t *)&costUncoded[blkPos+2] = cost1;
++ vcost_sum_0 = vaddq_s64(vcost_sum_0,cost0);
++ vcost_sum_1 = vaddq_s64(vcost_sum_1,cost1);
++ blkPos += trSize;
++ }
++ int64_t sum = vaddvq_s64(vaddq_s64(vcost_sum_0,vcost_sum_1));
++ *totalUncodedCost += sum;
++ *totalRdCost += sum;
++}
++
++template<int log2TrSize>
++static void psyRdoQuant_neon(int16_t *m_resiDctCoeff, int16_t *m_fencDctCoeff, int64_t *costUncoded, int64_t *totalUncodedCost, int64_t *totalRdCost, int64_t *psyScale, uint32_t blkPos)
++{
++ const int transformShift = MAX_TR_DYNAMIC_RANGE - X265_DEPTH - log2TrSize; /* Represents scaling through forward transform */
++ const int scaleBits = SCALE_BITS - 2 * transformShift;
++ const uint32_t trSize = 1 << log2TrSize;
++ //using preprocessor to bypass clang bug
++ const int max = X265_MAX(0, (2 * transformShift + 1));
++
++ int64x2_t vcost_sum_0 = vdupq_n_s64(0);
++ int64x2_t vcost_sum_1 = vdupq_n_s64(0);
++ int32x4_t vpsy = vdupq_n_s32(*psyScale);
++ for (int y = 0; y < MLS_CG_SIZE; y++)
++ {
++ int32x4_t signCoef = vmovl_s16(*(int16x4_t *)&m_resiDctCoeff[blkPos]);
++ int32x4_t predictedCoef = vsubq_s32(vmovl_s16(*(int16x4_t *)&m_fencDctCoeff[blkPos]),signCoef);
++ int64x2_t cost0, cost1;
++ cost0 = vmull_s32(vget_low_s32(signCoef),vget_low_s32(signCoef));
++ cost1 = vmull_high_s32(signCoef,signCoef);
++ cost0 = vshlq_n_s64(cost0,scaleBits);
++ cost1 = vshlq_n_s64(cost1,scaleBits);
++ int64x2_t neg0 = vmull_s32(vget_low_s32(predictedCoef),vget_low_s32(vpsy));
++ int64x2_t neg1 = vmull_high_s32(predictedCoef,vpsy);
++ if (max > 0) {
++ int64x2_t shift = vdupq_n_s64(-max);
++ neg0 = vshlq_s64(neg0,shift);
++ neg1 = vshlq_s64(neg1,shift);
++ }
++ cost0 = vsubq_s64(cost0,neg0);
++ cost1 = vsubq_s64(cost1,neg1);
++ *(int64x2_t *)&costUncoded[blkPos+0] = cost0;
++ *(int64x2_t *)&costUncoded[blkPos+2] = cost1;
++ vcost_sum_0 = vaddq_s64(vcost_sum_0,cost0);
++ vcost_sum_1 = vaddq_s64(vcost_sum_1,cost1);
++
++ blkPos += trSize;
++ }
++ int64_t sum = vaddvq_s64(vaddq_s64(vcost_sum_0,vcost_sum_1));
++ *totalUncodedCost += sum;
++ *totalRdCost += sum;
++}
++
++#else
++ #error "MLS_CG_SIZE must be 4 for neon version"
++#endif
++
++
++
++template<int trSize>
++int count_nonzero_neon(const int16_t* quantCoeff)
++{
++ X265_CHECK(((intptr_t)quantCoeff & 15) == 0, "quant buffer not aligned\n");
++ int count = 0;
++ int16x8_t vcount = vdupq_n_s16(0);
++ const int numCoeff = trSize * trSize;
++ int i = 0;
++ for (; (i + 8) <= numCoeff; i+=8)
++ {
++ int16x8_t in = *(int16x8_t*)&quantCoeff[i];
++ vcount = vaddq_s16(vcount,vtstq_s16(in,in));
++ }
++ for (; i < numCoeff; i++)
++ {
++ count += quantCoeff[i] != 0;
++ }
++
++ return count - vaddvq_s16(vcount);
++}
++
++template<int trSize>
++uint32_t copy_count_neon(int16_t* coeff, const int16_t* residual, intptr_t resiStride)
++{
++ uint32_t numSig = 0;
++ int16x8_t vcount = vdupq_n_s16(0);
++ for (int k = 0; k < trSize; k++)
++ {
++ int j = 0;
++ for (; (j + 8) <= trSize; j+=8)
++ {
++ int16x8_t in = *(int16x8_t*)&residual[j];
++ *(int16x8_t*)&coeff[j] = in;
++ vcount = vaddq_s16(vcount,vtstq_s16(in,in));
++ }
++ for (; j < trSize; j++)
++ {
++ coeff[j] = residual[j];
++ numSig += (residual[j] != 0);
++ }
++ residual += resiStride;
++ coeff += trSize;
++ }
++
++ return numSig - vaddvq_s16(vcount);
++}
++
++
++static void partialButterfly16(const int16_t* src, int16_t* dst, int shift, int line)
++{
++ int j, k;
++ int32x4_t E[2], O[2];
++ int32x4_t EE, EO;
++ int32x2_t EEE, EEO;
++ const int add = 1 << (shift - 1);
++ const int32x4_t _vadd = {add,0};
++
++ for (j = 0; j < line; j++)
++ {
++ int16x8_t in0 = *(int16x8_t *)src;
++ int16x8_t in1 = rev16(*(int16x8_t *)&src[8]);
++
++ E[0] = vaddl_s16(vget_low_s16(in0),vget_low_s16(in1));
++ O[0] = vsubl_s16(vget_low_s16(in0),vget_low_s16(in1));
++ E[1] = vaddl_high_s16(in0,in1);
++ O[1] = vsubl_high_s16(in0,in1);
++
++ for (k = 1; k < 16; k += 2)
++ {
++ int32x4_t c0 = vmovl_s16(*(int16x4_t *)&g_t16[k][0]);
++ int32x4_t c1 = vmovl_s16(*(int16x4_t *)&g_t16[k][4]);
++
++ int32x4_t res = _vadd;
++ res = vmlaq_s32(res,c0,O[0]);
++ res = vmlaq_s32(res,c1,O[1]);
++ dst[k * line] = (int16_t)(vaddvq_s32(res) >> shift);
++ }
++
++ /* EE and EO */
++ EE = vaddq_s32(E[0],rev32(E[1]));
++ EO = vsubq_s32(E[0],rev32(E[1]));
++
++ for (k = 2; k < 16; k += 4)
++ {
++ int32x4_t c0 = vmovl_s16(*(int16x4_t *)&g_t16[k][0]);
++ int32x4_t res = _vadd;
++ res = vmlaq_s32(res,c0,EO);
++ dst[k * line] = (int16_t)(vaddvq_s32(res) >> shift);
++ }
++
++ /* EEE and EEO */
++ EEE[0] = EE[0] + EE[3];
++ EEO[0] = EE[0] - EE[3];
++ EEE[1] = EE[1] + EE[2];
++ EEO[1] = EE[1] - EE[2];
++
++ dst[0] = (int16_t)((g_t16[0][0] * EEE[0] + g_t16[0][1] * EEE[1] + add) >> shift);
++ dst[8 * line] = (int16_t)((g_t16[8][0] * EEE[0] + g_t16[8][1] * EEE[1] + add) >> shift);
++ dst[4 * line] = (int16_t)((g_t16[4][0] * EEO[0] + g_t16[4][1] * EEO[1] + add) >> shift);
++ dst[12 * line] = (int16_t)((g_t16[12][0] * EEO[0] + g_t16[12][1] * EEO[1] + add) >> shift);
++
++
++ src += 16;
++ dst++;
++ }
++}
++
++
++static void partialButterfly32(const int16_t* src, int16_t* dst, int shift, int line)
++{
++ int j, k;
++ const int add = 1 << (shift - 1);
++
++
++ for (j = 0; j < line; j++)
++ {
++ int32x4_t VE[4], VO0,VO1,VO2,VO3;
++ int32x4_t VEE[2], VEO[2];
++ int32x4_t VEEE, VEEO;
++ int EEEE[2], EEEO[2];
++
++ int16x8x4_t inputs;
++ inputs = *(int16x8x4_t *)&src[0];
++ int16x8x4_t in_rev;
++
++ in_rev.val[1] = rev16(inputs.val[2]);
++ in_rev.val[0] = rev16(inputs.val[3]);
++
++ VE[0] = vaddl_s16(vget_low_s16(inputs.val[0]),vget_low_s16(in_rev.val[0]));
++ VE[1] = vaddl_high_s16(inputs.val[0],in_rev.val[0]);
++ VO0 = vsubl_s16(vget_low_s16(inputs.val[0]),vget_low_s16(in_rev.val[0]));
++ VO1 = vsubl_high_s16(inputs.val[0],in_rev.val[0]);
++ VE[2] = vaddl_s16(vget_low_s16(inputs.val[1]),vget_low_s16(in_rev.val[1]));
++ VE[3] = vaddl_high_s16(inputs.val[1],in_rev.val[1]);
++ VO2 = vsubl_s16(vget_low_s16(inputs.val[1]),vget_low_s16(in_rev.val[1]));
++ VO3 = vsubl_high_s16(inputs.val[1],in_rev.val[1]);
++
++ for (k = 1; k < 32; k += 2)
++ {
++ int32x4_t c0 = vmovl_s16(*(int16x4_t *)&g_t32[k][0]);
++ int32x4_t c1 = vmovl_s16(*(int16x4_t *)&g_t32[k][4]);
++ int32x4_t c2 = vmovl_s16(*(int16x4_t *)&g_t32[k][8]);
++ int32x4_t c3 = vmovl_s16(*(int16x4_t *)&g_t32[k][12]);
++ int32x4_t s = vmulq_s32(c0,VO0);
++ s = vmlaq_s32(s,c1,VO1);
++ s = vmlaq_s32(s,c2,VO2);
++ s = vmlaq_s32(s,c3,VO3);
++
++ dst[k * line] = (int16_t)((vaddvq_s32(s) + add) >> shift);
++
++ }
++
++ int32x4_t rev_VE[2];
++
++
++ rev_VE[0] = rev32(VE[3]);
++ rev_VE[1] = rev32(VE[2]);
++
++ /* EE and EO */
++ for (k = 0; k < 2; k++)
++ {
++ VEE[k] = vaddq_s32(VE[k],rev_VE[k]);
++ VEO[k] = vsubq_s32(VE[k],rev_VE[k]);
++ }
++ for (k = 2; k < 32; k += 4)
++ {
++ int32x4_t c0 = vmovl_s16(*(int16x4_t *)&g_t32[k][0]);
++ int32x4_t c1 = vmovl_s16(*(int16x4_t *)&g_t32[k][4]);
++ int32x4_t s = vmulq_s32(c0,VEO[0]);
++ s = vmlaq_s32(s,c1,VEO[1]);
++
++ dst[k * line] = (int16_t)((vaddvq_s32(s) + add) >> shift);
++
++ }
++
++ int32x4_t tmp = rev32(VEE[1]);
++ VEEE = vaddq_s32(VEE[0],tmp);
++ VEEO = vsubq_s32(VEE[0],tmp);
++ for (k = 4; k < 32; k += 8)
++ {
++ int32x4_t c = vmovl_s16(*(int16x4_t *)&g_t32[k][0]);
++ int32x4_t s = vmulq_s32(c,VEEO);
++
++ dst[k * line] = (int16_t)((vaddvq_s32(s) + add) >> shift);
++ }
++
++ /* EEEE and EEEO */
++ EEEE[0] = VEEE[0] + VEEE[3];
++ EEEO[0] = VEEE[0] - VEEE[3];
++ EEEE[1] = VEEE[1] + VEEE[2];
++ EEEO[1] = VEEE[1] - VEEE[2];
++
++ dst[0] = (int16_t)((g_t32[0][0] * EEEE[0] + g_t32[0][1] * EEEE[1] + add) >> shift);
++ dst[16 * line] = (int16_t)((g_t32[16][0] * EEEE[0] + g_t32[16][1] * EEEE[1] + add) >> shift);
++ dst[8 * line] = (int16_t)((g_t32[8][0] * EEEO[0] + g_t32[8][1] * EEEO[1] + add) >> shift);
++ dst[24 * line] = (int16_t)((g_t32[24][0] * EEEO[0] + g_t32[24][1] * EEEO[1] + add) >> shift);
++
++
++
++ src += 32;
++ dst++;
++ }
++}
++
++static void partialButterfly8(const int16_t* src, int16_t* dst, int shift, int line)
++{
++ int j, k;
++ int E[4], O[4];
++ int EE[2], EO[2];
++ int add = 1 << (shift - 1);
++
++ for (j = 0; j < line; j++)
++ {
++ /* E and O*/
++ for (k = 0; k < 4; k++)
++ {
++ E[k] = src[k] + src[7 - k];
++ O[k] = src[k] - src[7 - k];
++ }
++
++ /* EE and EO */
++ EE[0] = E[0] + E[3];
++ EO[0] = E[0] - E[3];
++ EE[1] = E[1] + E[2];
++ EO[1] = E[1] - E[2];
++
++ dst[0] = (int16_t)((g_t8[0][0] * EE[0] + g_t8[0][1] * EE[1] + add) >> shift);
++ dst[4 * line] = (int16_t)((g_t8[4][0] * EE[0] + g_t8[4][1] * EE[1] + add) >> shift);
++ dst[2 * line] = (int16_t)((g_t8[2][0] * EO[0] + g_t8[2][1] * EO[1] + add) >> shift);
++ dst[6 * line] = (int16_t)((g_t8[6][0] * EO[0] + g_t8[6][1] * EO[1] + add) >> shift);
++
++ dst[line] = (int16_t)((g_t8[1][0] * O[0] + g_t8[1][1] * O[1] + g_t8[1][2] * O[2] + g_t8[1][3] * O[3] + add) >> shift);
++ dst[3 * line] = (int16_t)((g_t8[3][0] * O[0] + g_t8[3][1] * O[1] + g_t8[3][2] * O[2] + g_t8[3][3] * O[3] + add) >> shift);
++ dst[5 * line] = (int16_t)((g_t8[5][0] * O[0] + g_t8[5][1] * O[1] + g_t8[5][2] * O[2] + g_t8[5][3] * O[3] + add) >> shift);
++ dst[7 * line] = (int16_t)((g_t8[7][0] * O[0] + g_t8[7][1] * O[1] + g_t8[7][2] * O[2] + g_t8[7][3] * O[3] + add) >> shift);
++
++ src += 8;
++ dst++;
++ }
++}
++
++static void partialButterflyInverse4(const int16_t* src, int16_t* dst, int shift, int line)
++{
++ int j;
++ int E[2], O[2];
++ int add = 1 << (shift - 1);
++
++ for (j = 0; j < line; j++)
++ {
++ /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */
++ O[0] = g_t4[1][0] * src[line] + g_t4[3][0] * src[3 * line];
++ O[1] = g_t4[1][1] * src[line] + g_t4[3][1] * src[3 * line];
++ E[0] = g_t4[0][0] * src[0] + g_t4[2][0] * src[2 * line];
++ E[1] = g_t4[0][1] * src[0] + g_t4[2][1] * src[2 * line];
++
++ /* Combining even and odd terms at each hierarchy levels to calculate the final spatial domain vector */
++ dst[0] = (int16_t)(x265_clip3(-32768, 32767, (E[0] + O[0] + add) >> shift));
++ dst[1] = (int16_t)(x265_clip3(-32768, 32767, (E[1] + O[1] + add) >> shift));
++ dst[2] = (int16_t)(x265_clip3(-32768, 32767, (E[1] - O[1] + add) >> shift));
++ dst[3] = (int16_t)(x265_clip3(-32768, 32767, (E[0] - O[0] + add) >> shift));
++
++ src++;
++ dst += 4;
++ }
++}
++
++
++
++static void partialButterflyInverse16_neon(const int16_t* src, int16_t* orig_dst, int shift, int line)
++{
++#define FMAK(x,l) s[l] = vmlal_lane_s16(s[l],*(int16x4_t*)&src[(x)*line],*(int16x4_t *)&g_t16[x][k],l)
++#define MULK(x,l) vmull_lane_s16(*(int16x4_t*)&src[x*line],*(int16x4_t *)&g_t16[x][k],l);
++#define ODD3_15(k) FMAK(3,k);FMAK(5,k);FMAK(7,k);FMAK(9,k);FMAK(11,k);FMAK(13,k);FMAK(15,k);
++#define EVEN6_14_STEP4(k) FMAK(6,k);FMAK(10,k);FMAK(14,k);
++
++
++ int j, k;
++ int32x4_t E[8], O[8];
++ int32x4_t EE[4], EO[4];
++ int32x4_t EEE[2], EEO[2];
++ const int add = 1 << (shift - 1);
++
++
++#pragma unroll(4)
++ for (j = 0; j < line; j+=4)
++ {
++ /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */
++
++#pragma unroll(2)
++ for (k=0;k<2;k++) {
++ int32x4_t s;
++ s = vmull_s16(vdup_n_s16(g_t16[4][k]),*(int16x4_t*)&src[4*line]);;
++ EEO[k] = vmlal_s16(s,vdup_n_s16(g_t16[12][k]),*(int16x4_t*)&src[(12)*line]);
++ s = vmull_s16(vdup_n_s16(g_t16[0][k]),*(int16x4_t*)&src[0*line]);;
++ EEE[k] = vmlal_s16(s,vdup_n_s16(g_t16[8][k]),*(int16x4_t*)&src[(8)*line]);
++ }
++
++ /* Combining even and odd terms at each hierarchy levels to calculate the final spatial domain vector */
++ EE[0] = vaddq_s32(EEE[0] , EEO[0]);
++ EE[2] = vsubq_s32(EEE[1] , EEO[1]);
++ EE[1] = vaddq_s32(EEE[1] , EEO[1]);
++ EE[3] = vsubq_s32(EEE[0] , EEO[0]);
++
++
++#pragma unroll(1)
++ for (k = 0; k < 4; k+=4)
++ {
++ int32x4_t s[4];
++ s[0] = MULK(2,0);
++ s[1] = MULK(2,1);
++ s[2] = MULK(2,2);
++ s[3] = MULK(2,3);
++
++ EVEN6_14_STEP4(0);
++ EVEN6_14_STEP4(1);
++ EVEN6_14_STEP4(2);
++ EVEN6_14_STEP4(3);
++
++ EO[k] = s[0];
++ EO[k+1] = s[1];
++ EO[k+2] = s[2];
++ EO[k+3] = s[3];
++ }
++
++
++
++ static const int32x4_t min = vdupq_n_s32(-32768);
++ static const int32x4_t max = vdupq_n_s32(32767);
++ const int32x4_t minus_shift = vdupq_n_s32(-shift);
++
++#pragma unroll(4)
++ for (k = 0; k < 4; k++)
++ {
++ E[k] = vaddq_s32(EE[k] , EO[k]);
++ E[k + 4] = vsubq_s32(EE[3 - k] , EO[3 - k]);
++ }
++
++#pragma unroll(2)
++ for (k = 0; k < 8; k+=4)
++ {
++ int32x4_t s[4];
++ s[0] = MULK(1,0);
++ s[1] = MULK(1,1);
++ s[2] = MULK(1,2);
++ s[3] = MULK(1,3);
++ ODD3_15(0);
++ ODD3_15(1);
++ ODD3_15(2);
++ ODD3_15(3);
++ O[k] = s[0];
++ O[k+1] = s[1];
++ O[k+2] = s[2];
++ O[k+3] = s[3];
++ int32x4_t t;
++ int16x4_t x0,x1,x2,x3;
++
++ E[k] = vaddq_s32(vdupq_n_s32(add),E[k]);
++ t = vaddq_s32(E[k],O[k]);
++ t = vshlq_s32(t,minus_shift);
++ t = vmaxq_s32(t,min);
++ t = vminq_s32(t,max);
++ x0 = vmovn_s32(t);
++
++ E[k+1] = vaddq_s32(vdupq_n_s32(add),E[k+1]);
++ t = vaddq_s32(E[k+1],O[k+1]);
++ t = vshlq_s32(t,minus_shift);
++ t = vmaxq_s32(t,min);
++ t = vminq_s32(t,max);
++ x1 = vmovn_s32(t);
++
++ E[k+2] = vaddq_s32(vdupq_n_s32(add),E[k+2]);
++ t = vaddq_s32(E[k+2],O[k+2]);
++ t = vshlq_s32(t,minus_shift);
++ t = vmaxq_s32(t,min);
++ t = vminq_s32(t,max);
++ x2 = vmovn_s32(t);
++
++ E[k+3] = vaddq_s32(vdupq_n_s32(add),E[k+3]);
++ t = vaddq_s32(E[k+3],O[k+3]);
++ t = vshlq_s32(t,minus_shift);
++ t = vmaxq_s32(t,min);
++ t = vminq_s32(t,max);
++ x3 = vmovn_s32(t);
++
++ transpose_4x4x16(x0,x1,x2,x3);
++ *(int16x4_t*)&orig_dst[0*16+k] = x0;
++ *(int16x4_t*)&orig_dst[1*16+k] = x1;
++ *(int16x4_t*)&orig_dst[2*16+k] = x2;
++ *(int16x4_t*)&orig_dst[3*16+k] = x3;
++ }
++
++
++#pragma unroll(2)
++ for (k = 0; k < 8; k+=4)
++ {
++ int32x4_t t;
++ int16x4_t x0,x1,x2,x3;
++
++ t = vsubq_s32(E[7-k],O[7-k]);
++ t = vshlq_s32(t,minus_shift);
++ t = vmaxq_s32(t,min);
++ t = vminq_s32(t,max);
++ x0 = vmovn_s32(t);
++
++ t = vsubq_s32(E[6-k],O[6-k]);
++ t = vshlq_s32(t,minus_shift);
++ t = vmaxq_s32(t,min);
++ t = vminq_s32(t,max);
++ x1 = vmovn_s32(t);
++
++ t = vsubq_s32(E[5-k],O[5-k]);
++
++ t = vshlq_s32(t,minus_shift);
++ t = vmaxq_s32(t,min);
++ t = vminq_s32(t,max);
++ x2 = vmovn_s32(t);
++
++ t = vsubq_s32(E[4-k],O[4-k]);
++ t = vshlq_s32(t,minus_shift);
++ t = vmaxq_s32(t,min);
++ t = vminq_s32(t,max);
++ x3 = vmovn_s32(t);
++
++ transpose_4x4x16(x0,x1,x2,x3);
++ *(int16x4_t*)&orig_dst[0*16+k+8] = x0;
++ *(int16x4_t*)&orig_dst[1*16+k+8] = x1;
++ *(int16x4_t*)&orig_dst[2*16+k+8] = x2;
++ *(int16x4_t*)&orig_dst[3*16+k+8] = x3;
++ }
++ orig_dst += 4*16;
++ src+=4;
++ }
++
++#undef MUL
++#undef FMA
++#undef FMAK
++#undef MULK
++#undef ODD3_15
++#undef EVEN6_14_STEP4
++
++
++}
++
++
++
++static void partialButterflyInverse32_neon(const int16_t* src, int16_t* orig_dst, int shift, int line)
++{
++#define MUL(x) vmull_s16(vdup_n_s16(g_t32[x][k]),*(int16x4_t*)&src[x*line]);
++#define FMA(x) s = vmlal_s16(s,vdup_n_s16(g_t32[x][k]),*(int16x4_t*)&src[(x)*line])
++#define FMAK(x,l) s[l] = vmlal_lane_s16(s[l],*(int16x4_t*)&src[(x)*line],*(int16x4_t *)&g_t32[x][k],l)
++#define MULK(x,l) vmull_lane_s16(*(int16x4_t*)&src[x*line],*(int16x4_t *)&g_t32[x][k],l);
++#define ODD31(k) FMAK(3,k);FMAK(5,k);FMAK(7,k);FMAK(9,k);FMAK(11,k);FMAK(13,k);FMAK(15,k);FMAK(17,k);FMAK(19,k);FMAK(21,k);FMAK(23,k);FMAK(25,k);FMAK(27,k);FMAK(29,k);FMAK(31,k);
++
++#define ODD15(k) FMAK(6,k);FMAK(10,k);FMAK(14,k);FMAK(18,k);FMAK(22,k);FMAK(26,k);FMAK(30,k);
++#define ODD7(k) FMAK(12,k);FMAK(20,k);FMAK(28,k);
++
++
++ int j, k;
++ int32x4_t E[16], O[16];
++ int32x4_t EE[8], EO[8];
++ int32x4_t EEE[4], EEO[4];
++ int32x4_t EEEE[2], EEEO[2];
++ int16x4_t dst[32];
++ int add = 1 << (shift - 1);
++
++#pragma unroll (8)
++ for (j = 0; j < line; j+=4)
++ {
++#pragma unroll (4)
++ for (k = 0; k < 16; k+=4)
++ {
++ int32x4_t s[4];
++ s[0] = MULK(1,0);
++ s[1] = MULK(1,1);
++ s[2] = MULK(1,2);
++ s[3] = MULK(1,3);
++ ODD31(0);
++ ODD31(1);
++ ODD31(2);
++ ODD31(3);
++ O[k] = s[0];
++ O[k+1] = s[1];
++ O[k+2] = s[2];
++ O[k+3] = s[3];
++
++
++ }
++
++
++#pragma unroll (2)
++ for (k = 0; k < 8; k+=4)
++ {
++ int32x4_t s[4];
++ s[0] = MULK(2,0);
++ s[1] = MULK(2,1);
++ s[2] = MULK(2,2);
++ s[3] = MULK(2,3);
++
++ ODD15(0);
++ ODD15(1);
++ ODD15(2);
++ ODD15(3);
++
++ EO[k] = s[0];
++ EO[k+1] = s[1];
++ EO[k+2] = s[2];
++ EO[k+3] = s[3];
++ }
++
++
++ for (k = 0; k < 4; k+=4)
++ {
++ int32x4_t s[4];
++ s[0] = MULK(4,0);
++ s[1] = MULK(4,1);
++ s[2] = MULK(4,2);
++ s[3] = MULK(4,3);
++
++ ODD7(0);
++ ODD7(1);
++ ODD7(2);
++ ODD7(3);
++
++ EEO[k] = s[0];
++ EEO[k+1] = s[1];
++ EEO[k+2] = s[2];
++ EEO[k+3] = s[3];
++ }
++
++#pragma unroll (2)
++ for (k=0;k<2;k++) {
++ int32x4_t s;
++ s = MUL(8);
++ EEEO[k] = FMA(24);
++ s = MUL(0);
++ EEEE[k] = FMA(16);
++ }
++ /* Combining even and odd terms at each hierarchy levels to calculate the final spatial domain vector */
++ EEE[0] = vaddq_s32(EEEE[0],EEEO[0]);
++ EEE[3] = vsubq_s32(EEEE[0],EEEO[0]);
++ EEE[1] = vaddq_s32(EEEE[1],EEEO[1]);
++ EEE[2] = vsubq_s32(EEEE[1],EEEO[1]);
++
++#pragma unroll (4)
++ for (k = 0; k < 4; k++)
++ {
++ EE[k] = vaddq_s32(EEE[k],EEO[k]);
++ EE[k + 4] = vsubq_s32((EEE[3 - k]), (EEO[3 - k]));
++ }
++
++#pragma unroll (8)
++ for (k = 0; k < 8; k++)
++ {
++ E[k] = vaddq_s32(EE[k],EO[k]);
++ E[k + 8] = vsubq_s32((EE[7 - k]),(EO[7 - k]));
++ }
++
++ static const int32x4_t min = vdupq_n_s32(-32768);
++ static const int32x4_t max = vdupq_n_s32(32767);
++
++
++
++#pragma unroll (16)
++ for (k = 0; k < 16; k++)
++ {
++ int32x4_t adde = vaddq_s32(vdupq_n_s32(add),E[k]);
++ int32x4_t s = vaddq_s32(adde,O[k]);
++ s = vshlq_s32(s,vdupq_n_s32(-shift));
++ s = vmaxq_s32(s,min);
++ s = vminq_s32(s,max);
++
++
++
++ dst[k] = vmovn_s32(s);
++ adde = vaddq_s32(vdupq_n_s32(add),(E[15-k]));
++ s =vsubq_s32(adde,(O[15-k]));
++ s = vshlq_s32(s,vdupq_n_s32(-shift));
++ s = vmaxq_s32(s,min);
++ s = vminq_s32(s,max);
++
++ dst[k+16] = vmovn_s32(s);
++ }
++
++
++#pragma unroll (8)
++ for (k = 0; k < 32; k+=4)
++ {
++ int16x4_t x0 = dst[k+0];
++ int16x4_t x1 = dst[k+1];
++ int16x4_t x2 = dst[k+2];
++ int16x4_t x3 = dst[k+3];
++ transpose_4x4x16(x0,x1,x2,x3);
++ *(int16x4_t*)&orig_dst[0*32+k] = x0;
++ *(int16x4_t*)&orig_dst[1*32+k] = x1;
++ *(int16x4_t*)&orig_dst[2*32+k] = x2;
++ *(int16x4_t*)&orig_dst[3*32+k] = x3;
++ }
++ orig_dst += 4*32;
++ src += 4;
++ }
++#undef MUL
++#undef FMA
++#undef FMAK
++#undef MULK
++#undef ODD31
++#undef ODD15
++#undef ODD7
++
++}
++
++
++static void dct8_neon(const int16_t* src, int16_t* dst, intptr_t srcStride)
++{
++ const int shift_1st = 2 + X265_DEPTH - 8;
++ const int shift_2nd = 9;
++
++ ALIGN_VAR_32(int16_t, coef[8 * 8]);
++ ALIGN_VAR_32(int16_t, block[8 * 8]);
++
++ for (int i = 0; i < 8; i++)
++ {
++ memcpy(&block[i * 8], &src[i * srcStride], 8 * sizeof(int16_t));
++ }
++
++ partialButterfly8(block, coef, shift_1st, 8);
++ partialButterfly8(coef, dst, shift_2nd, 8);
++}
++
++static void dct16_neon(const int16_t* src, int16_t* dst, intptr_t srcStride)
++{
++ const int shift_1st = 3 + X265_DEPTH - 8;
++ const int shift_2nd = 10;
++
++ ALIGN_VAR_32(int16_t, coef[16 * 16]);
++ ALIGN_VAR_32(int16_t, block[16 * 16]);
++
++ for (int i = 0; i < 16; i++)
++ {
++ memcpy(&block[i * 16], &src[i * srcStride], 16 * sizeof(int16_t));
++ }
++
++ partialButterfly16(block, coef, shift_1st, 16);
++ partialButterfly16(coef, dst, shift_2nd, 16);
++}
++
++static void dct32_neon(const int16_t* src, int16_t* dst, intptr_t srcStride)
++{
++ const int shift_1st = 4 + X265_DEPTH - 8;
++ const int shift_2nd = 11;
++
++ ALIGN_VAR_32(int16_t, coef[32 * 32]);
++ ALIGN_VAR_32(int16_t, block[32 * 32]);
++
++ for (int i = 0; i < 32; i++)
++ {
++ memcpy(&block[i * 32], &src[i * srcStride], 32 * sizeof(int16_t));
++ }
++
++ partialButterfly32(block, coef, shift_1st, 32);
++ partialButterfly32(coef, dst, shift_2nd, 32);
++}
++
++static void idct4_neon(const int16_t* src, int16_t* dst, intptr_t dstStride)
++{
++ const int shift_1st = 7;
++ const int shift_2nd = 12 - (X265_DEPTH - 8);
++
++ ALIGN_VAR_32(int16_t, coef[4 * 4]);
++ ALIGN_VAR_32(int16_t, block[4 * 4]);
++
++ partialButterflyInverse4(src, coef, shift_1st, 4); // Forward DST BY FAST ALGORITHM, block input, coef output
++ partialButterflyInverse4(coef, block, shift_2nd, 4); // Forward DST BY FAST ALGORITHM, coef input, coeff output
++
++ for (int i = 0; i < 4; i++)
++ {
++ memcpy(&dst[i * dstStride], &block[i * 4], 4 * sizeof(int16_t));
++ }
++}
++
++static void idct16_neon(const int16_t* src, int16_t* dst, intptr_t dstStride)
++{
++ const int shift_1st = 7;
++ const int shift_2nd = 12 - (X265_DEPTH - 8);
++
++ ALIGN_VAR_32(int16_t, coef[16 * 16]);
++ ALIGN_VAR_32(int16_t, block[16 * 16]);
++
++ partialButterflyInverse16_neon(src, coef, shift_1st, 16);
++ partialButterflyInverse16_neon(coef, block, shift_2nd, 16);
++
++ for (int i = 0; i < 16; i++)
++ {
++ memcpy(&dst[i * dstStride], &block[i * 16], 16 * sizeof(int16_t));
++ }
++}
++
++static void idct32_neon(const int16_t* src, int16_t* dst, intptr_t dstStride)
++{
++ const int shift_1st = 7;
++ const int shift_2nd = 12 - (X265_DEPTH - 8);
++
++ ALIGN_VAR_32(int16_t, coef[32 * 32]);
++ ALIGN_VAR_32(int16_t, block[32 * 32]);
++
++ partialButterflyInverse32_neon(src, coef, shift_1st, 32);
++ partialButterflyInverse32_neon(coef, block, shift_2nd, 32);
++
++ for (int i = 0; i < 32; i++)
++ {
++ memcpy(&dst[i * dstStride], &block[i * 32], 32 * sizeof(int16_t));
++ }
++}
++
++
++
++}
++
++namespace X265_NS {
++// x265 private namespace
++void setupDCTPrimitives_neon(EncoderPrimitives& p) {
++ p.cu[BLOCK_4x4].nonPsyRdoQuant = nonPsyRdoQuant_neon<2>;
++ p.cu[BLOCK_8x8].nonPsyRdoQuant = nonPsyRdoQuant_neon<3>;
++ p.cu[BLOCK_16x16].nonPsyRdoQuant = nonPsyRdoQuant_neon<4>;
++ p.cu[BLOCK_32x32].nonPsyRdoQuant = nonPsyRdoQuant_neon<5>;
++ p.cu[BLOCK_4x4].psyRdoQuant = psyRdoQuant_neon<2>;
++ p.cu[BLOCK_8x8].psyRdoQuant = psyRdoQuant_neon<3>;
++ p.cu[BLOCK_16x16].psyRdoQuant = psyRdoQuant_neon<4>;
++ p.cu[BLOCK_32x32].psyRdoQuant = psyRdoQuant_neon<5>;
++ p.cu[BLOCK_8x8].dct = dct8_neon;
++ p.cu[BLOCK_16x16].dct = dct16_neon;
++ p.cu[BLOCK_32x32].dct = dct32_neon;
++ p.cu[BLOCK_4x4].idct = idct4_neon;
++ p.cu[BLOCK_16x16].idct = idct16_neon;
++ p.cu[BLOCK_32x32].idct = idct32_neon;
++ p.cu[BLOCK_4x4].count_nonzero = count_nonzero_neon<4>;
++ p.cu[BLOCK_8x8].count_nonzero = count_nonzero_neon<8>;
++ p.cu[BLOCK_16x16].count_nonzero = count_nonzero_neon<16>;
++ p.cu[BLOCK_32x32].count_nonzero = count_nonzero_neon<32>;
++
++ p.cu[BLOCK_4x4].copy_cnt = copy_count_neon<4>;
++ p.cu[BLOCK_8x8].copy_cnt = copy_count_neon<8>;
++ p.cu[BLOCK_16x16].copy_cnt = copy_count_neon<16>;
++ p.cu[BLOCK_32x32].copy_cnt = copy_count_neon<32>;
++ p.cu[BLOCK_4x4].psyRdoQuant_1p = nonPsyRdoQuant_neon<2>;
++ p.cu[BLOCK_4x4].psyRdoQuant_2p = psyRdoQuant_neon<2>;
++ p.cu[BLOCK_8x8].psyRdoQuant_1p = nonPsyRdoQuant_neon<3>;
++ p.cu[BLOCK_8x8].psyRdoQuant_2p = psyRdoQuant_neon<3>;
++ p.cu[BLOCK_16x16].psyRdoQuant_1p = nonPsyRdoQuant_neon<4>;
++ p.cu[BLOCK_16x16].psyRdoQuant_2p = psyRdoQuant_neon<4>;
++ p.cu[BLOCK_32x32].psyRdoQuant_1p = nonPsyRdoQuant_neon<5>;
++ p.cu[BLOCK_32x32].psyRdoQuant_2p = psyRdoQuant_neon<5>;
++
++ p.scanPosLast =scanPosLast_opt;
++
++}
++};
++
++
++
++#endif
+diff -Naur ./source/common/arm64/dct-prim.h ../x265_apple_patch/source/common/arm64/dct-prim.h
+--- ./source/common/arm64/dct-prim.h 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/dct-prim.h 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,18 @@
++#ifndef __DCT_PRIM_NEON_H__
++#define __DCT_PRIM_NEON_H__
++
++
++#include "common.h"
++#include "primitives.h"
++#include "contexts.h" // costCoeffNxN_c
++#include "threading.h" // CLZ
++
++namespace X265_NS {
++// x265 private namespace
++void setupDCTPrimitives_neon(EncoderPrimitives& p);
++};
++
++
++
++#endif
++
+diff -Naur ./source/common/arm64/filter-prim.cpp ../x265_apple_patch/source/common/arm64/filter-prim.cpp
+--- ./source/common/arm64/filter-prim.cpp 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/filter-prim.cpp 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,797 @@
++
++#if HAVE_NEON
++
++#include "filter-prim.h"
++#include <arm_neon.h>
++
++namespace {
++
++using namespace X265_NS;
++
++
++template<int width, int height>
++void filterPixelToShort_neon(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride)
++{
++ const int shift = IF_INTERNAL_PREC - X265_DEPTH;
++ int row, col;
++ const int16x8_t off = vdupq_n_s16(IF_INTERNAL_OFFS);
++ for (row = 0; row < height; row++)
++ {
++
++ for (col = 0; col < width; col+=8)
++ {
++ int16x8_t in;
++
++#if HIGH_BIT_DEPTH
++ in = *(int16x8_t *)&src[col];
++#else
++ in = vmovl_u8(*(uint8x8_t *)&src[col]);
++#endif
++
++ int16x8_t tmp = vshlq_n_s16(in,shift);
++ tmp = vsubq_s16(tmp,off);
++ *(int16x8_t *)&dst[col] = tmp;
++
++ }
++
++ src += srcStride;
++ dst += dstStride;
++ }
++}
++
++
++template<int N, int width, int height>
++void interp_horiz_pp_neon(const pixel* src, intptr_t srcStride, pixel* dst, intptr_t dstStride, int coeffIdx)
++{
++ const int16_t* coeff = (N == 4) ? g_chromaFilter[coeffIdx] : g_lumaFilter[coeffIdx];
++ int headRoom = IF_FILTER_PREC;
++ int offset = (1 << (headRoom - 1));
++ uint16_t maxVal = (1 << X265_DEPTH) - 1;
++ int cStride = 1;
++
++ src -= (N / 2 - 1) * cStride;
++ int16x8_t vc;
++ vc = *(int16x8_t *)coeff;
++ int16x4_t low_vc = vget_low_s16(vc);
++ int16x4_t high_vc = vget_high_s16(vc);
++
++ const int32x4_t voffset = vdupq_n_s32(offset);
++ const int32x4_t vhr = vdupq_n_s32(-headRoom);
++
++ int row, col;
++ for (row = 0; row < height; row++)
++ {
++ for (col = 0; col < width; col+=8)
++ {
++ int32x4_t vsum1,vsum2;
++
++ int16x8_t input[N];
++
++ for (int i=0;i<N;i++)
++ {
++#if HIGH_BIT_DEPTH
++ input[i] = *(int16x8_t *)&src[col+i];
++#else
++ input[i] = vmovl_u8(*(uint8x8_t *)&src[col+i]);
++#endif
++ }
++ vsum1 = voffset;
++ vsum2 = voffset;
++
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[0]),low_vc,0);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[0],low_vc,0);
++
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[1]),low_vc,1);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[1],low_vc,1);
++
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[2]),low_vc,2);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[2],low_vc,2);
++
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[3]),low_vc,3);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[3],low_vc,3);
++
++ if (N == 8)
++ {
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[4]),high_vc,0);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[4],high_vc,0);
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[5]),high_vc,1);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[5],high_vc,1);
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[6]),high_vc,2);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[6],high_vc,2);
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[7]),high_vc,3);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[7],high_vc,3);
++
++ }
++
++ vsum1 = vshlq_s32(vsum1, vhr);
++ vsum2 = vshlq_s32(vsum2, vhr);
++
++ int16x8_t vsum = vuzp1q_s16(vsum1,vsum2);
++ vsum = vminq_s16(vsum,vdupq_n_s16(maxVal));
++ vsum = vmaxq_s16(vsum,vdupq_n_s16(0));
++#if HIGH_BIT_DEPTH
++ *(int16x8_t *)&dst[col] = vsum;
++#else
++ uint8x16_t usum = vuzp1q_u8(vsum,vsum);
++ *(uint8x8_t *)&dst[col] = vget_low_u8(usum);
++#endif
++
++ }
++
++ src += srcStride;
++ dst += dstStride;
++ }
++}
++
++#if HIGH_BIT_DEPTH
++
++template<int N, int width, int height>
++void interp_horiz_ps_neon(const uint16_t * src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt)
++{
++ const int16_t* coeff = (N == 4) ? g_chromaFilter[coeffIdx] : g_lumaFilter[coeffIdx];
++ const int headRoom = IF_INTERNAL_PREC - X265_DEPTH;
++ const int shift = IF_FILTER_PREC - headRoom;
++ const int offset = (unsigned)-IF_INTERNAL_OFFS << shift;
++
++ int blkheight = height;
++ src -= N / 2 - 1;
++
++ if (isRowExt)
++ {
++ src -= (N / 2 - 1) * srcStride;
++ blkheight += N - 1;
++ }
++ int32x4_t vc0 = vmovl_s16(*(int16x4_t *)coeff);
++ int32x4_t vc1;
++
++ if (N ==8) {
++ vc1 = vmovl_s16(*(int16x4_t *)(coeff + 4));
++ }
++
++ const int32x4_t voffset = vdupq_n_s32(offset);
++ const int32x4_t vhr = vdupq_n_s32(-shift);
++
++ int row, col;
++ for (row = 0; row < blkheight; row++)
++ {
++ for (col = 0; col < width; col+=4)
++ {
++ int32x4_t vsum;
++
++ int32x4_t input[N];
++
++ for (int i=0;i<N;i++)
++ {
++ input[i] = vmovl_s16(*(int16x4_t *)&src[col+i]);
++ }
++ vsum = voffset;
++ vsum = vmlaq_laneq_s32(vsum,(input[0]),vc0,0);
++ vsum = vmlaq_laneq_s32(vsum,(input[1]),vc0,1);
++ vsum = vmlaq_laneq_s32(vsum,(input[2]),vc0,2);
++ vsum = vmlaq_laneq_s32(vsum,(input[3]),vc0,3);
++
++
++ if (N == 8)
++ {
++ vsum = vmlaq_laneq_s32(vsum,(input[4]),vc1,0);
++ vsum = vmlaq_laneq_s32(vsum,(input[5]),vc1,1);
++ vsum = vmlaq_laneq_s32(vsum,(input[6]),vc1,2);
++ vsum = vmlaq_laneq_s32(vsum,(input[7]),vc1,3);
++
++ }
++
++ vsum = vshlq_s32(vsum, vhr);
++ *(int16x4_t *)&dst[col] = vmovn_u32(vsum);
++ }
++
++ src += srcStride;
++ dst += dstStride;
++ }
++ }
++
++
++#else
++
++template<int N, int width, int height>
++void interp_horiz_ps_neon(const uint8_t* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt)
++{
++ const int16_t* coeff = (N == 4) ? g_chromaFilter[coeffIdx] : g_lumaFilter[coeffIdx];
++ const int headRoom = IF_INTERNAL_PREC - X265_DEPTH;
++ const int shift = IF_FILTER_PREC - headRoom;
++ const int offset = (unsigned)-IF_INTERNAL_OFFS << shift;
++
++ int blkheight = height;
++ src -= N / 2 - 1;
++
++ if (isRowExt)
++ {
++ src -= (N / 2 - 1) * srcStride;
++ blkheight += N - 1;
++ }
++ int16x8_t vc;
++ vc = *(int16x8_t *)coeff;
++
++ const int16x8_t voffset = vdupq_n_s16(offset);
++ const int16x8_t vhr = vdupq_n_s16(-shift);
++
++ int row, col;
++ for (row = 0; row < blkheight; row++)
++ {
++ for (col = 0; col < width; col+=8)
++ {
++ int16x8_t vsum;
++
++ int16x8_t input[N];
++
++ for (int i=0;i<N;i++)
++ {
++ input[i] = vmovl_u8(*(uint8x8_t *)&src[col+i]);
++ }
++ vsum = voffset;
++ vsum = vmlaq_laneq_s16(vsum,(input[0]),vc,0);
++ vsum = vmlaq_laneq_s16(vsum,(input[1]),vc,1);
++ vsum = vmlaq_laneq_s16(vsum,(input[2]),vc,2);
++ vsum = vmlaq_laneq_s16(vsum,(input[3]),vc,3);
++
++
++ if (N == 8)
++ {
++ vsum = vmlaq_laneq_s16(vsum,(input[4]),vc,4);
++ vsum = vmlaq_laneq_s16(vsum,(input[5]),vc,5);
++ vsum = vmlaq_laneq_s16(vsum,(input[6]),vc,6);
++ vsum = vmlaq_laneq_s16(vsum,(input[7]),vc,7);
++
++ }
++
++ vsum = vshlq_s16(vsum, vhr);
++ *(int16x8_t *)&dst[col] = vsum;
++ }
++
++ src += srcStride;
++ dst += dstStride;
++ }
++ }
++
++#endif
++
++
++template<int N, int width, int height>
++void interp_vert_ss_neon(const int16_t* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx)
++{
++ const int16_t* c = (N == 8 ? g_lumaFilter[coeffIdx] : g_chromaFilter[coeffIdx]);
++ int shift = IF_FILTER_PREC;
++ src -= (N / 2 - 1) * srcStride;
++ int16x8_t vc;
++ vc = *(int16x8_t *)c;
++ int16x4_t low_vc = vget_low_s16(vc);
++ int16x4_t high_vc = vget_high_s16(vc);
++
++ const int32x4_t vhr = vdupq_n_s32(-shift);
++
++ int row, col;
++ for (row = 0; row < height; row++)
++ {
++ for (col = 0; col < width; col+=8)
++ {
++ int32x4_t vsum1,vsum2;
++
++ int16x8_t input[N];
++
++ for (int i=0;i<N;i++)
++ {
++ input[i] = *(int16x8_t *)&src[col+i*srcStride];
++ }
++
++ vsum1 = vmull_lane_s16(vget_low_s16(input[0]),low_vc,0);
++ vsum2 = vmull_high_lane_s16(input[0],low_vc,0);
++
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[1]),low_vc,1);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[1],low_vc,1);
++
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[2]),low_vc,2);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[2],low_vc,2);
++
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[3]),low_vc,3);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[3],low_vc,3);
++
++ if (N == 8)
++ {
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[4]),high_vc,0);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[4],high_vc,0);
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[5]),high_vc,1);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[5],high_vc,1);
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[6]),high_vc,2);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[6],high_vc,2);
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[7]),high_vc,3);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[7],high_vc,3);
++
++ }
++
++ vsum1 = vshlq_s32(vsum1, vhr);
++ vsum2 = vshlq_s32(vsum2, vhr);
++
++ int16x8_t vsum = vuzp1q_s16(vsum1,vsum2);
++ *(int16x8_t *)&dst[col] = vsum;
++ }
++
++ src += srcStride;
++ dst += dstStride;
++ }
++
++}
++
++
++#if HIGH_BIT_DEPTH
++
++template<int N, int width, int height>
++void interp_vert_pp_neon(const uint16_t* src, intptr_t srcStride, uint16_t* dst, intptr_t dstStride, int coeffIdx)
++{
++
++ const int16_t* c = (N == 4) ? g_chromaFilter[coeffIdx] : g_lumaFilter[coeffIdx];
++ int shift = IF_FILTER_PREC;
++ int offset = 1 << (shift - 1);
++ const uint16_t maxVal = (1 << X265_DEPTH) - 1;
++
++ src -= (N / 2 - 1) * srcStride;
++ int16x8_t vc;
++ vc = *(int16x8_t *)c;
++ int32x4_t low_vc = vmovl_s16(vget_low_s16(vc));
++ int32x4_t high_vc = vmovl_s16(vget_high_s16(vc));
++
++ const int32x4_t voffset = vdupq_n_s32(offset);
++ const int32x4_t vhr = vdupq_n_s32(-shift);
++
++ int row, col;
++ for (row = 0; row < height; row++)
++ {
++ for (col = 0; col < width; col+=4)
++ {
++ int32x4_t vsum;
++
++ int32x4_t input[N];
++
++ for (int i=0;i<N;i++)
++ {
++ input[i] = vmovl_u16(*(uint16x4_t *)&src[col+i*srcStride]);
++ }
++ vsum = voffset;
++
++ vsum = vmlaq_laneq_s32(vsum,(input[0]),low_vc,0);
++ vsum = vmlaq_laneq_s32(vsum,(input[1]),low_vc,1);
++ vsum = vmlaq_laneq_s32(vsum,(input[2]),low_vc,2);
++ vsum = vmlaq_laneq_s32(vsum,(input[3]),low_vc,3);
++
++ if (N == 8)
++ {
++ vsum = vmlaq_laneq_s32(vsum,(input[4]),high_vc,0);
++ vsum = vmlaq_laneq_s32(vsum,(input[5]),high_vc,1);
++ vsum = vmlaq_laneq_s32(vsum,(input[6]),high_vc,2);
++ vsum = vmlaq_laneq_s32(vsum,(input[7]),high_vc,3);
++ }
++
++ vsum = vshlq_s32(vsum, vhr);
++ vsum = vminq_s32(vsum,vdupq_n_s32(maxVal));
++ vsum = vmaxq_s32(vsum,vdupq_n_s32(0));
++ *(uint16x4_t *)&dst[col] = vmovn_u32(vsum);
++ }
++ src += srcStride;
++ dst += dstStride;
++ }
++}
++
++
++
++
++#else
++
++template<int N, int width, int height>
++void interp_vert_pp_neon(const uint8_t* src, intptr_t srcStride, uint8_t* dst, intptr_t dstStride, int coeffIdx)
++{
++
++ const int16_t* c = (N == 4) ? g_chromaFilter[coeffIdx] : g_lumaFilter[coeffIdx];
++ int shift = IF_FILTER_PREC;
++ int offset = 1 << (shift - 1);
++ const uint16_t maxVal = (1 << X265_DEPTH) - 1;
++
++ src -= (N / 2 - 1) * srcStride;
++ int16x8_t vc;
++ vc = *(int16x8_t *)c;
++
++ const int16x8_t voffset = vdupq_n_s16(offset);
++ const int16x8_t vhr = vdupq_n_s16(-shift);
++
++ int row, col;
++ for (row = 0; row < height; row++)
++ {
++ for (col = 0; col < width; col+=8)
++ {
++ int16x8_t vsum;
++
++ int16x8_t input[N];
++
++ for (int i=0;i<N;i++)
++ {
++ input[i] = vmovl_u8(*(uint8x8_t *)&src[col+i*srcStride]);
++ }
++ vsum = voffset;
++
++ vsum = vmlaq_laneq_s16(vsum,(input[0]),vc,0);
++ vsum = vmlaq_laneq_s16(vsum,(input[1]),vc,1);
++ vsum = vmlaq_laneq_s16(vsum,(input[2]),vc,2);
++ vsum = vmlaq_laneq_s16(vsum,(input[3]),vc,3);
++
++ if (N == 8)
++ {
++ vsum = vmlaq_laneq_s16(vsum,(input[4]),vc,4);
++ vsum = vmlaq_laneq_s16(vsum,(input[5]),vc,5);
++ vsum = vmlaq_laneq_s16(vsum,(input[6]),vc,6);
++ vsum = vmlaq_laneq_s16(vsum,(input[7]),vc,7);
++
++ }
++
++ vsum = vshlq_s16(vsum, vhr);
++
++ vsum = vminq_s16(vsum,vdupq_n_s16(maxVal));
++ vsum = vmaxq_s16(vsum,vdupq_n_s16(0));
++ uint8x16_t usum = vuzp1q_u8(vsum,vsum);
++ *(uint8x8_t *)&dst[col] = vget_low_u8(usum);
++
++ }
++
++ src += srcStride;
++ dst += dstStride;
++ }
++}
++
++
++#endif
++
++
++#if HIGH_BIT_DEPTH
++
++template<int N, int width, int height>
++void interp_vert_ps_neon(const uint16_t* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx)
++{
++ const int16_t* c = (N == 4) ? g_chromaFilter[coeffIdx] : g_lumaFilter[coeffIdx];
++ int headRoom = IF_INTERNAL_PREC - X265_DEPTH;
++ int shift = IF_FILTER_PREC - headRoom;
++ int offset = (unsigned)-IF_INTERNAL_OFFS << shift;
++ src -= (N / 2 - 1) * srcStride;
++
++ int16x8_t vc;
++ vc = *(int16x8_t *)c;
++ int32x4_t low_vc = vmovl_s16(vget_low_s16(vc));
++ int32x4_t high_vc = vmovl_s16(vget_high_s16(vc));
++
++ const int32x4_t voffset = vdupq_n_s32(offset);
++ const int32x4_t vhr = vdupq_n_s32(-shift);
++
++ int row, col;
++ for (row = 0; row < height; row++)
++ {
++ for (col = 0; col < width; col+=4)
++ {
++ int16x8_t vsum;
++
++ int16x8_t input[N];
++
++ for (int i=0;i<N;i++)
++ {
++ input[i] = vmovl_u16(*(uint16x4_t *)&src[col+i*srcStride]);
++ }
++ vsum = voffset;
++
++ vsum = vmlaq_laneq_s32(vsum,(input[0]),low_vc,0);
++ vsum = vmlaq_laneq_s32(vsum,(input[1]),low_vc,1);
++ vsum = vmlaq_laneq_s32(vsum,(input[2]),low_vc,2);
++ vsum = vmlaq_laneq_s32(vsum,(input[3]),low_vc,3);
++
++ if (N == 8)
++ {
++ int16x8_t vsum1 = vmulq_laneq_s32((input[4]),high_vc,0);
++ vsum1 = vmlaq_laneq_s32(vsum1,(input[5]),high_vc,1);
++ vsum1 = vmlaq_laneq_s32(vsum1,(input[6]),high_vc,2);
++ vsum1 = vmlaq_laneq_s32(vsum1,(input[7]),high_vc,3);
++ vsum = vaddq_s32(vsum,vsum1);
++ }
++
++ vsum = vshlq_s32(vsum, vhr);
++
++ *(uint16x4_t *)&dst[col] = vmovn_s32(vsum);
++ }
++
++ src += srcStride;
++ dst += dstStride;
++ }
++}
++
++#else
++
++template<int N, int width, int height>
++void interp_vert_ps_neon(const uint8_t* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx)
++{
++ const int16_t* c = (N == 4) ? g_chromaFilter[coeffIdx] : g_lumaFilter[coeffIdx];
++ int headRoom = IF_INTERNAL_PREC - X265_DEPTH;
++ int shift = IF_FILTER_PREC - headRoom;
++ int offset = (unsigned)-IF_INTERNAL_OFFS << shift;
++ src -= (N / 2 - 1) * srcStride;
++
++ int16x8_t vc;
++ vc = *(int16x8_t *)c;
++
++ const int16x8_t voffset = vdupq_n_s16(offset);
++ const int16x8_t vhr = vdupq_n_s16(-shift);
++
++ int row, col;
++ for (row = 0; row < height; row++)
++ {
++ for (col = 0; col < width; col+=8)
++ {
++ int16x8_t vsum;
++
++ int16x8_t input[N];
++
++ for (int i=0;i<N;i++)
++ {
++ input[i] = vmovl_u8(*(uint8x8_t *)&src[col+i*srcStride]);
++ }
++ vsum = voffset;
++
++ vsum = vmlaq_laneq_s16(vsum,(input[0]),vc,0);
++ vsum = vmlaq_laneq_s16(vsum,(input[1]),vc,1);
++ vsum = vmlaq_laneq_s16(vsum,(input[2]),vc,2);
++ vsum = vmlaq_laneq_s16(vsum,(input[3]),vc,3);
++
++ if (N == 8)
++ {
++ int16x8_t vsum1 = vmulq_laneq_s16((input[4]),vc,4);
++ vsum1 = vmlaq_laneq_s16(vsum1,(input[5]),vc,5);
++ vsum1 = vmlaq_laneq_s16(vsum1,(input[6]),vc,6);
++ vsum1 = vmlaq_laneq_s16(vsum1,(input[7]),vc,7);
++ vsum = vaddq_s16(vsum,vsum1);
++ }
++
++ vsum = vshlq_s32(vsum, vhr);
++ *(int16x8_t *)&dst[col] = vsum;
++ }
++
++ src += srcStride;
++ dst += dstStride;
++ }
++}
++
++#endif
++
++
++
++template<int N, int width, int height>
++void interp_vert_sp_neon(const int16_t* src, intptr_t srcStride, pixel* dst, intptr_t dstStride, int coeffIdx)
++{
++ int headRoom = IF_INTERNAL_PREC - X265_DEPTH;
++ int shift = IF_FILTER_PREC + headRoom;
++ int offset = (1 << (shift - 1)) + (IF_INTERNAL_OFFS << IF_FILTER_PREC);
++ uint16_t maxVal = (1 << X265_DEPTH) - 1;
++ const int16_t* coeff = (N == 8 ? g_lumaFilter[coeffIdx] : g_chromaFilter[coeffIdx]);
++
++ src -= (N / 2 - 1) * srcStride;
++
++ int16x8_t vc;
++ vc = *(int16x8_t *)coeff;
++ int16x4_t low_vc = vget_low_s16(vc);
++ int16x4_t high_vc = vget_high_s16(vc);
++
++ const int32x4_t voffset = vdupq_n_s32(offset);
++ const int32x4_t vhr = vdupq_n_s32(-shift);
++
++ int row, col;
++ for (row = 0; row < height; row++)
++ {
++ for (col = 0; col < width; col+=8)
++ {
++ int32x4_t vsum1,vsum2;
++
++ int16x8_t input[N];
++
++ for (int i=0;i<N;i++)
++ {
++ input[i] = *(int16x8_t *)&src[col+i*srcStride];
++ }
++ vsum1 = voffset;
++ vsum2 = voffset;
++
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[0]),low_vc,0);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[0],low_vc,0);
++
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[1]),low_vc,1);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[1],low_vc,1);
++
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[2]),low_vc,2);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[2],low_vc,2);
++
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[3]),low_vc,3);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[3],low_vc,3);
++
++ if (N == 8)
++ {
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[4]),high_vc,0);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[4],high_vc,0);
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[5]),high_vc,1);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[5],high_vc,1);
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[6]),high_vc,2);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[6],high_vc,2);
++ vsum1 = vmlal_lane_s16(vsum1,vget_low_s16(input[7]),high_vc,3);
++ vsum2 = vmlal_high_lane_s16(vsum2,input[7],high_vc,3);
++ }
++
++ vsum1 = vshlq_s32(vsum1, vhr);
++ vsum2 = vshlq_s32(vsum2, vhr);
++
++ int16x8_t vsum = vuzp1q_s16(vsum1,vsum2);
++ vsum = vminq_s16(vsum,vdupq_n_s16(maxVal));
++ vsum = vmaxq_s16(vsum,vdupq_n_s16(0));
++#if HIGH_BIT_DEPTH
++ *(int16x8_t *)&dst[col] = vsum;
++#else
++ uint8x16_t usum = vuzp1q_u8(vsum,vsum);
++ *(uint8x8_t *)&dst[col] = vget_low_u8(usum);
++#endif
++
++ }
++
++ src += srcStride;
++ dst += dstStride;
++ }
++}
++
++
++
++
++
++
++template<int N, int width, int height>
++void interp_hv_pp_neon(const pixel* src, intptr_t srcStride, pixel* dst, intptr_t dstStride, int idxX, int idxY)
++{
++ ALIGN_VAR_32(int16_t, immed[width * (height + N - 1)]);
++
++ interp_horiz_ps_neon<N, width, height>(src, srcStride, immed, width, idxX, 1);
++ interp_vert_sp_neon<N,width,height>(immed + (N / 2 - 1) * width, width, dst, dstStride, idxY);
++}
++
++
++
++}
++
++
++
++
++namespace X265_NS {
++ #define CHROMA_420(W, H) \
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_ ## W ## x ## H].filter_hpp = interp_horiz_pp_neon<4, W, H>; \
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_ ## W ## x ## H].filter_hps = interp_horiz_ps_neon<4, W, H>; \
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_ ## W ## x ## H].filter_vpp = interp_vert_pp_neon<4, W, H>; \
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_ ## W ## x ## H].filter_vps = interp_vert_ps_neon<4, W, H>; \
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_ ## W ## x ## H].filter_vsp = interp_vert_sp_neon<4, W, H>; \
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_ ## W ## x ## H].filter_vss = interp_vert_ss_neon<4, W, H>; \
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_ ## W ## x ## H].p2s[NONALIGNED] = filterPixelToShort_neon<W, H>;\
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_ ## W ## x ## H].p2s[ALIGNED] = filterPixelToShort_neon<W, H>;
++
++ #define CHROMA_422(W, H) \
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_ ## W ## x ## H].filter_hpp = interp_horiz_pp_neon<4, W, H>; \
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_ ## W ## x ## H].filter_hps = interp_horiz_ps_neon<4, W, H>; \
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_ ## W ## x ## H].filter_vpp = interp_vert_pp_neon<4, W, H>; \
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_ ## W ## x ## H].filter_vps = interp_vert_ps_neon<4, W, H>; \
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_ ## W ## x ## H].filter_vsp = interp_vert_sp_neon<4, W, H>; \
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_ ## W ## x ## H].filter_vss = interp_vert_ss_neon<4, W, H>; \
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_ ## W ## x ## H].p2s[NONALIGNED] = filterPixelToShort_neon<W, H>;\
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_ ## W ## x ## H].p2s[ALIGNED] = filterPixelToShort_neon<W, H>;
++
++ #define CHROMA_444(W, H) \
++ p.chroma[X265_CSP_I444].pu[LUMA_ ## W ## x ## H].filter_hpp = interp_horiz_pp_neon<4, W, H>; \
++ p.chroma[X265_CSP_I444].pu[LUMA_ ## W ## x ## H].filter_hps = interp_horiz_ps_neon<4, W, H>; \
++ p.chroma[X265_CSP_I444].pu[LUMA_ ## W ## x ## H].filter_vpp = interp_vert_pp_neon<4, W, H>; \
++ p.chroma[X265_CSP_I444].pu[LUMA_ ## W ## x ## H].filter_vps = interp_vert_ps_neon<4, W, H>; \
++ p.chroma[X265_CSP_I444].pu[LUMA_ ## W ## x ## H].filter_vsp = interp_vert_sp_neon<4, W, H>; \
++ p.chroma[X265_CSP_I444].pu[LUMA_ ## W ## x ## H].filter_vss = interp_vert_ss_neon<4, W, H>; \
++ p.chroma[X265_CSP_I444].pu[LUMA_ ## W ## x ## H].p2s[NONALIGNED] = filterPixelToShort_neon<W, H>;\
++ p.chroma[X265_CSP_I444].pu[LUMA_ ## W ## x ## H].p2s[ALIGNED] = filterPixelToShort_neon<W, H>;
++
++ #define LUMA(W, H) \
++ p.pu[LUMA_ ## W ## x ## H].luma_hpp = interp_horiz_pp_neon<8, W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].luma_hps = interp_horiz_ps_neon<8, W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].luma_vpp = interp_vert_pp_neon<8, W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].luma_vps = interp_vert_ps_neon<8, W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].luma_vsp = interp_vert_sp_neon<8, W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].luma_vss = interp_vert_ss_neon<8, W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].luma_hvpp = interp_hv_pp_neon<8, W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].convert_p2s[NONALIGNED] = filterPixelToShort_neon<W, H>;\
++ p.pu[LUMA_ ## W ## x ## H].convert_p2s[ALIGNED] = filterPixelToShort_neon<W, H>;
++
++
++void setupFilterPrimitives_neon(EncoderPrimitives &p)
++{
++
++ // All neon functions assume width of multiple of 8, (2,4,12 variants are not optimized)
++
++ LUMA(8, 8);
++ LUMA(8, 4);
++ LUMA(16, 16);
++ CHROMA_420(8, 8);
++ LUMA(16, 8);
++ CHROMA_420(8, 4);
++ LUMA(8, 16);
++ LUMA(16, 12);
++ CHROMA_420(8, 6);
++ LUMA(16, 4);
++ CHROMA_420(8, 2);
++ LUMA(32, 32);
++ CHROMA_420(16, 16);
++ LUMA(32, 16);
++ CHROMA_420(16, 8);
++ LUMA(16, 32);
++ CHROMA_420(8, 16);
++ LUMA(32, 24);
++ CHROMA_420(16, 12);
++ LUMA(24, 32);
++ LUMA(32, 8);
++ CHROMA_420(16, 4);
++ LUMA(8, 32);
++ LUMA(64, 64);
++ CHROMA_420(32, 32);
++ LUMA(64, 32);
++ CHROMA_420(32, 16);
++ LUMA(32, 64);
++ CHROMA_420(16, 32);
++ LUMA(64, 48);
++ CHROMA_420(32, 24);
++ LUMA(48, 64);
++ CHROMA_420(24, 32);
++ LUMA(64, 16);
++ CHROMA_420(32, 8);
++ LUMA(16, 64);
++ CHROMA_420(8, 32);
++ CHROMA_422(8, 16);
++ CHROMA_422(8, 8);
++ CHROMA_422(8, 12);
++ CHROMA_422(8, 4);
++ CHROMA_422(16, 32);
++ CHROMA_422(16, 16);
++ CHROMA_422(8, 32);
++ CHROMA_422(16, 24);
++ CHROMA_422(16, 8);
++ CHROMA_422(32, 64);
++ CHROMA_422(32, 32);
++ CHROMA_422(16, 64);
++ CHROMA_422(32, 48);
++ CHROMA_422(24, 64);
++ CHROMA_422(32, 16);
++ CHROMA_422(8, 64);
++ CHROMA_444(8, 8);
++ CHROMA_444(8, 4);
++ CHROMA_444(16, 16);
++ CHROMA_444(16, 8);
++ CHROMA_444(8, 16);
++ CHROMA_444(16, 12);
++ CHROMA_444(16, 4);
++ CHROMA_444(32, 32);
++ CHROMA_444(32, 16);
++ CHROMA_444(16, 32);
++ CHROMA_444(32, 24);
++ CHROMA_444(24, 32);
++ CHROMA_444(32, 8);
++ CHROMA_444(8, 32);
++ CHROMA_444(64, 64);
++ CHROMA_444(64, 32);
++ CHROMA_444(32, 64);
++ CHROMA_444(64, 48);
++ CHROMA_444(48, 64);
++ CHROMA_444(64, 16);
++ CHROMA_444(16, 64);
++
++}
++
++};
++
++
++#endif
++
++
+diff -Naur ./source/common/arm64/filter-prim.h ../x265_apple_patch/source/common/arm64/filter-prim.h
+--- ./source/common/arm64/filter-prim.h 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/filter-prim.h 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,20 @@
++#ifndef _FILTER_PRIM_ARM64_H__
++#define _FILTER_PRIM_ARM64_H__
++
++
++#include "common.h"
++#include "slicetype.h" // LOWRES_COST_MASK
++#include "primitives.h"
++#include "x265.h"
++
++
++namespace X265_NS {
++
++
++void setupFilterPrimitives_neon(EncoderPrimitives &p);
++
++};
++
++
++#endif
++
+diff -Naur ./source/common/arm64/intrapred-prim.cpp ../x265_apple_patch/source/common/arm64/intrapred-prim.cpp
+--- ./source/common/arm64/intrapred-prim.cpp 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/intrapred-prim.cpp 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,266 @@
++/*****************************************************************************
++ * Copyright (C) 2013-2017 MulticoreWare, Inc
++ *
++ * Authors: Min Chen <chenm003@163.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
++ * the Free Software Foundation; either version 2 of the License, or
++ * (at your option) any later version.
++ *
++ * This program is distributed in the hope that it will be useful,
++ * but WITHOUT ANY WARRANTY; without even the implied warranty of
++ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
++ * GNU General Public License for more details.
++ *
++ * You should have received a copy of the GNU General Public License
++ * along with this program; if not, write to the Free Software
++ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
++ *
++ * This program is also available under a commercial proprietary license.
++ * For more information, contact us at license @ x265.com.
++ *****************************************************************************/
++
++
++#include "common.h"
++#include "primitives.h"
++
++
++#if 1
++#include "arm64-utils.h"
++#include <arm_neon.h>
++
++using namespace X265_NS;
++
++namespace {
++
++
++
++template<int width>
++void intra_pred_ang_neon(pixel* dst, intptr_t dstStride, const pixel *srcPix0, int dirMode, int bFilter)
++{
++ int width2 = width << 1;
++ // Flip the neighbours in the horizontal case.
++ int horMode = dirMode < 18;
++ pixel neighbourBuf[129];
++ const pixel *srcPix = srcPix0;
++
++ if (horMode)
++ {
++ neighbourBuf[0] = srcPix[0];
++ //for (int i = 0; i < width << 1; i++)
++ //{
++ // neighbourBuf[1 + i] = srcPix[width2 + 1 + i];
++ // neighbourBuf[width2 + 1 + i] = srcPix[1 + i];
++ //}
++ memcpy(&neighbourBuf[1],&srcPix[width2+1],sizeof(pixel)*(width << 1));
++ memcpy(&neighbourBuf[width2 + 1],&srcPix[1],sizeof(pixel)*(width << 1));
++ srcPix = neighbourBuf;
++ }
++
++ // Intra prediction angle and inverse angle tables.
++ const int8_t angleTable[17] = { -32, -26, -21, -17, -13, -9, -5, -2, 0, 2, 5, 9, 13, 17, 21, 26, 32 };
++ const int16_t invAngleTable[8] = { 4096, 1638, 910, 630, 482, 390, 315, 256 };
++
++ // Get the prediction angle.
++ int angleOffset = horMode ? 10 - dirMode : dirMode - 26;
++ int angle = angleTable[8 + angleOffset];
++
++ // Vertical Prediction.
++ if (!angle)
++ {
++ for (int y = 0; y < width; y++) {
++ memcpy(&dst[y * dstStride],srcPix + 1,sizeof(pixel)*width);
++ }
++ if (bFilter)
++ {
++ int topLeft = srcPix[0], top = srcPix[1];
++ for (int y = 0; y < width; y++)
++ dst[y * dstStride] = x265_clip((int16_t)(top + ((srcPix[width2 + 1 + y] - topLeft) >> 1)));
++ }
++ }
++ else // Angular prediction.
++ {
++ // Get the reference pixels. The reference base is the first pixel to the top (neighbourBuf[1]).
++ pixel refBuf[64];
++ const pixel *ref;
++
++ // Use the projected left neighbours and the top neighbours.
++ if (angle < 0)
++ {
++ // Number of neighbours projected.
++ int nbProjected = -((width * angle) >> 5) - 1;
++ pixel *ref_pix = refBuf + nbProjected + 1;
++
++ // Project the neighbours.
++ int invAngle = invAngleTable[- angleOffset - 1];
++ int invAngleSum = 128;
++ for (int i = 0; i < nbProjected; i++)
++ {
++ invAngleSum += invAngle;
++ ref_pix[- 2 - i] = srcPix[width2 + (invAngleSum >> 8)];
++ }
++
++ // Copy the top-left and top pixels.
++ //for (int i = 0; i < width + 1; i++)
++ //ref_pix[-1 + i] = srcPix[i];
++
++ memcpy(&ref_pix[-1],srcPix,(width+1)*sizeof(pixel));
++ ref = ref_pix;
++ }
++ else // Use the top and top-right neighbours.
++ ref = srcPix + 1;
++
++ // Pass every row.
++ int angleSum = 0;
++ for (int y = 0; y < width; y++)
++ {
++ angleSum += angle;
++ int offset = angleSum >> 5;
++ int fraction = angleSum & 31;
++
++ if (fraction) // Interpolate
++ {
++ if (width >= 8 && sizeof(pixel) == 1)
++ {
++ const int16x8_t f0 = vdupq_n_s16(32-fraction);
++ const int16x8_t f1 = vdupq_n_s16(fraction);
++ for (int x = 0;x<width;x+=8) {
++ uint8x8_t in0 = *(uint8x8_t *)&ref[offset + x];
++ uint8x8_t in1 = *(uint8x8_t *)&ref[offset+ x + 1];
++ int16x8_t lo = vmlaq_s16(vdupq_n_s16(16),vmovl_u8(in0),f0);
++ lo = vmlaq_s16(lo,vmovl_u8(in1),f1);
++ lo = vshrq_n_s16(lo,5);
++ *(uint8x8_t *)&dst[y * dstStride + x] = vmovn_u16(lo);
++ }
++ }
++ else if (width >= 4 && sizeof(pixel) == 2)
++ {
++ const int32x4_t f0 = vdupq_n_s32(32-fraction);
++ const int32x4_t f1 = vdupq_n_s32(fraction);
++ for (int x = 0;x<width;x+=4) {
++ uint16x4_t in0 = *(uint16x4_t *)&ref[offset + x];
++ uint16x4_t in1 = *(uint16x4_t *)&ref[offset+ x + 1];
++ int32x4_t lo = vmlaq_s32(vdupq_n_s32(16),vmovl_u16(in0),f0);
++ lo = vmlaq_s32(lo,vmovl_u16(in1),f1);
++ lo = vshrq_n_s32(lo,5);
++ *(uint16x4_t *)&dst[y * dstStride + x] = vmovn_u32(lo);
++ }
++ }
++ else {
++ for (int x = 0; x < width; x++)
++ dst[y * dstStride + x] = (pixel)(((32 - fraction) * ref[offset + x] + fraction * ref[offset + x + 1] + 16) >> 5);
++ }
++ }
++ else // Copy.
++ {
++ memcpy(&dst[y * dstStride],&ref[offset],sizeof(pixel)*width);
++ }
++ }
++ }
++
++ // Flip for horizontal.
++ if (horMode)
++ {
++ if (width == 8) transpose8x8(dst,dst,dstStride,dstStride);
++ else if (width == 16) transpose16x16(dst,dst,dstStride,dstStride);
++ else if (width == 32) transpose32x32(dst,dst,dstStride,dstStride);
++ else {
++ for (int y = 0; y < width - 1; y++)
++ {
++ for (int x = y + 1; x < width; x++)
++ {
++ pixel tmp = dst[y * dstStride + x];
++ dst[y * dstStride + x] = dst[x * dstStride + y];
++ dst[x * dstStride + y] = tmp;
++ }
++ }
++ }
++ }
++}
++
++template<int log2Size>
++void all_angs_pred_neon(pixel *dest, pixel *refPix, pixel *filtPix, int bLuma)
++{
++ const int size = 1 << log2Size;
++ for (int mode = 2; mode <= 34; mode++)
++ {
++ pixel *srcPix = (g_intraFilterFlags[mode] & size ? filtPix : refPix);
++ pixel *out = dest + ((mode - 2) << (log2Size * 2));
++
++ intra_pred_ang_neon<size>(out, size, srcPix, mode, bLuma);
++
++ // Optimize code don't flip buffer
++ bool modeHor = (mode < 18);
++
++ // transpose the block if this is a horizontal mode
++ if (modeHor)
++ {
++ if (size == 8) transpose8x8(out,out,size,size);
++ else if (size == 16) transpose16x16(out,out,size,size);
++ else if (size == 32) transpose32x32(out,out,size,size);
++ else {
++ for (int k = 0; k < size - 1; k++)
++ {
++ for (int l = k + 1; l < size; l++)
++ {
++ pixel tmp = out[k * size + l];
++ out[k * size + l] = out[l * size + k];
++ out[l * size + k] = tmp;
++ }
++ }
++ }
++ }
++ }
++}
++}
++
++namespace X265_NS {
++// x265 private namespace
++
++void setupIntraPrimitives_neon(EncoderPrimitives& p)
++{
++// p.cu[BLOCK_4x4].intra_filter = intraFilter<4>;
++// p.cu[BLOCK_8x8].intra_filter = intraFilter<8>;
++// p.cu[BLOCK_16x16].intra_filter = intraFilter<16>;
++// p.cu[BLOCK_32x32].intra_filter = intraFilter<32>;
++
++// p.cu[BLOCK_4x4].intra_pred[PLANAR_IDX] = planar_pred_neon<2>;
++// p.cu[BLOCK_8x8].intra_pred[PLANAR_IDX] = planar_pred_neon<3>;
++// p.cu[BLOCK_16x16].intra_pred[PLANAR_IDX] = planar_pred_neon<4>;
++// p.cu[BLOCK_32x32].intra_pred[PLANAR_IDX] = planar_pred_neon<5>;
++//
++// p.cu[BLOCK_4x4].intra_pred[DC_IDX] = intra_pred_dc_neon<4>;
++// p.cu[BLOCK_8x8].intra_pred[DC_IDX] = intra_pred_dc_neon<8>;
++// p.cu[BLOCK_16x16].intra_pred[DC_IDX] = intra_pred_dc_neon<16>;
++// p.cu[BLOCK_32x32].intra_pred[DC_IDX] = intra_pred_dc_neon<32>;
++
++ for (int i = 2; i < NUM_INTRA_MODE; i++)
++ {
++ p.cu[BLOCK_4x4].intra_pred[i] = intra_pred_ang_neon<4>;
++ p.cu[BLOCK_8x8].intra_pred[i] = intra_pred_ang_neon<8>;
++ p.cu[BLOCK_16x16].intra_pred[i] = intra_pred_ang_neon<16>;
++ p.cu[BLOCK_32x32].intra_pred[i] = intra_pred_ang_neon<32>;
++ }
++
++ p.cu[BLOCK_4x4].intra_pred_allangs = all_angs_pred_neon<2>;
++ p.cu[BLOCK_8x8].intra_pred_allangs = all_angs_pred_neon<3>;
++ p.cu[BLOCK_16x16].intra_pred_allangs = all_angs_pred_neon<4>;
++ p.cu[BLOCK_32x32].intra_pred_allangs = all_angs_pred_neon<5>;
++}
++}
++
++
++
++#else
++
++namespace X265_NS {
++// x265 private namespace
++void setupIntraPrimitives_neon(EncoderPrimitives& p)
++{}
++}
++
++#endif
++
++
++
+diff -Naur ./source/common/arm64/intrapred-prim.h ../x265_apple_patch/source/common/arm64/intrapred-prim.h
+--- ./source/common/arm64/intrapred-prim.h 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/intrapred-prim.h 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,14 @@
++#ifndef INTRAPRED_PRIM_H__
++
++#if defined(__aarch64__)
++
++namespace X265_NS {
++// x265 private namespace
++
++void setupIntraPrimitives_neon(EncoderPrimitives& p);
++}
++
++#endif
++
++#endif
++
+diff -Naur ./source/common/arm64/loopfilter-prim.cpp ../x265_apple_patch/source/common/arm64/loopfilter-prim.cpp
+--- ./source/common/arm64/loopfilter-prim.cpp 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/loopfilter-prim.cpp 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,305 @@
++/*****************************************************************************
++* Copyright (C) 2013-2017 MulticoreWare, Inc
++*
++* Authors: Praveen Kumar Tiwari <praveen@multicorewareinc.com>
++* Dnyaneshwar Gorade <dnyaneshwar@multicorewareinc.com>
++* Min Chen <chenm003@163.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
++* the Free Software Foundation; either version 2 of the License, or
++* (at your option) any later version.
++*
++* This program is distributed in the hope that it will be useful,
++* but WITHOUT ANY WARRANTY; without even the implied warranty of
++* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
++* GNU General Public License for more details.
++*
++* You should have received a copy of the GNU General Public License
++* along with this program; if not, write to the Free Software
++* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
++*
++* This program is also available under a commercial proprietary license.
++* For more information, contact us at license @ x265.com.
++*****************************************************************************/
++#include "loopfilter-prim.h"
++
++#define PIXEL_MIN 0
++
++
++
++#if !(HIGH_BIT_DEPTH) && defined(HAVE_NEON)
++#include<arm_neon.h>
++
++namespace {
++
++
++/* get the sign of input variable (TODO: this is a dup, make common) */
++static inline int8_t signOf(int x)
++{
++ return (x >> 31) | ((int)((((uint32_t)-x)) >> 31));
++}
++
++static inline int8x8_t sign_diff_neon(const uint8x8_t in0, const uint8x8_t in1)
++{
++ int16x8_t in = vsubl_u8(in0,in1);
++ return vmovn_s16(vmaxq_s16(vminq_s16(in,vdupq_n_s16(1)),vdupq_n_s16(-1)));
++}
++
++static void calSign_neon(int8_t *dst, const pixel *src1, const pixel *src2, const int endX)
++{
++ int x = 0;
++ for (; (x + 8) <= endX; x += 8) {
++ *(int8x8_t *)&dst[x] = sign_diff_neon(*(uint8x8_t *)&src1[x],*(uint8x8_t *)&src2[x]);
++ }
++
++ for (; x < endX; x++)
++ dst[x] = signOf(src1[x] - src2[x]);
++}
++
++static void processSaoCUE0_neon(pixel * rec, int8_t * offsetEo, int width, int8_t* signLeft, intptr_t stride)
++{
++
++
++ int y;
++ int8_t signRight, signLeft0;
++ int8_t edgeType;
++
++ for (y = 0; y < 2; y++)
++ {
++ signLeft0 = signLeft[y];
++ int x = 0;
++
++ if (width >= 8) {
++ int8x8_t vsignRight;
++ int8x8x2_t shifter;
++ shifter.val[1][0] = signLeft0;
++ static const int8x8_t index = {8,0,1,2,3,4,5,6};
++ int8x8_t tbl = *(int8x8_t *)offsetEo;
++ for (; (x+8) <= width; x+=8)
++ {
++ uint8x8_t in = *(uint8x8_t *)&rec[x];
++ vsignRight = sign_diff_neon(in,*(uint8x8_t *)&rec[x+1]);
++ shifter.val[0] = vneg_s8(vsignRight);
++ int8x8_t tmp = shifter.val[0];
++ int8x8_t edge = vtbl2_s8(shifter,index);
++ int8x8_t vedgeType = vadd_s8(vadd_s8(vsignRight,edge),vdup_n_s8(2));
++ shifter.val[1][0] = tmp[7];
++ int16x8_t t1 = vmovl_s8(vtbl1_s8(tbl,vedgeType));
++ t1 = vaddw_u8(t1,in);
++ t1 = vmaxq_s16(t1,vdupq_n_s16(0));
++ t1 = vminq_s16(t1,vdupq_n_s16(255));
++ *(uint8x8_t *)&rec[x] = vmovn_u16(t1);
++ }
++ signLeft0 = shifter.val[1][0];
++ }
++ for (; x < width; x++)
++ {
++ signRight = ((rec[x] - rec[x + 1]) < 0) ? -1 : ((rec[x] - rec[x + 1]) > 0) ? 1 : 0;
++ edgeType = signRight + signLeft0 + 2;
++ signLeft0 = -signRight;
++ rec[x] = x265_clip(rec[x] + offsetEo[edgeType]);
++ }
++ rec += stride;
++ }
++}
++
++static void processSaoCUE1_neon(pixel* rec, int8_t* upBuff1, int8_t* offsetEo, intptr_t stride, int width)
++{
++ int x = 0;
++ int8_t signDown;
++ int edgeType;
++
++ if (width >= 8) {
++ int8x8_t tbl = *(int8x8_t *)offsetEo;
++ for (; (x+8) <= width; x+=8)
++ {
++ uint8x8_t in0 = *(uint8x8_t *)&rec[x];
++ uint8x8_t in1 = *(uint8x8_t *)&rec[x+stride];
++ int8x8_t vsignDown = sign_diff_neon(in0,in1);
++ int8x8_t vedgeType = vadd_s8(vadd_s8(vsignDown,*(int8x8_t *)&upBuff1[x]),vdup_n_s8(2));
++ *(int8x8_t *)&upBuff1[x] = vneg_s8(vsignDown);
++ int16x8_t t1 = vmovl_s8(vtbl1_s8(tbl,vedgeType));
++ t1 = vaddw_u8(t1,in0);
++ *(uint8x8_t *)&rec[x] = vqmovun_s16(t1);
++ }
++ }
++ for (; x < width; x++)
++ {
++ signDown = signOf(rec[x] - rec[x + stride]);
++ edgeType = signDown + upBuff1[x] + 2;
++ upBuff1[x] = -signDown;
++ rec[x] = x265_clip(rec[x] + offsetEo[edgeType]);
++ }
++}
++
++static void processSaoCUE1_2Rows_neon(pixel* rec, int8_t* upBuff1, int8_t* offsetEo, intptr_t stride, int width)
++{
++ int y;
++ int8_t signDown;
++ int edgeType;
++
++ for (y = 0; y < 2; y++)
++ {
++ int x=0;
++ if (width >= 8) {
++ int8x8_t tbl = *(int8x8_t *)offsetEo;
++ for (; (x+8) <= width; x+=8)
++ {
++ uint8x8_t in0 = *(uint8x8_t *)&rec[x];
++ uint8x8_t in1 = *(uint8x8_t *)&rec[x+stride];
++ int8x8_t vsignDown = sign_diff_neon(in0,in1);
++ int8x8_t vedgeType = vadd_s8(vadd_s8(vsignDown,*(int8x8_t *)&upBuff1[x]),vdup_n_s8(2));
++ *(int8x8_t *)&upBuff1[x] = vneg_s8(vsignDown);
++ int16x8_t t1 = vmovl_s8(vtbl1_s8(tbl,vedgeType));
++ t1 = vaddw_u8(t1,in0);
++ t1 = vmaxq_s16(t1,vdupq_n_s16(0));
++ t1 = vminq_s16(t1,vdupq_n_s16(255));
++ *(uint8x8_t *)&rec[x] = vmovn_u16(t1);
++
++ }
++ }
++ for (; x < width; x++)
++ {
++ signDown = signOf(rec[x] - rec[x + stride]);
++ edgeType = signDown + upBuff1[x] + 2;
++ upBuff1[x] = -signDown;
++ rec[x] = x265_clip(rec[x] + offsetEo[edgeType]);
++ }
++ rec += stride;
++ }
++}
++
++static void processSaoCUE2_neon(pixel * rec, int8_t * bufft, int8_t * buff1, int8_t * offsetEo, int width, intptr_t stride)
++{
++ int x;
++
++ if (abs(buff1-bufft) < 16)
++ {
++ for (x = 0; x < width; x++)
++ {
++ int8_t signDown = signOf(rec[x] - rec[x + stride + 1]);
++ int edgeType = signDown + buff1[x] + 2;
++ bufft[x + 1] = -signDown;
++ rec[x] = x265_clip(rec[x] + offsetEo[edgeType]);;
++ }
++ }
++ else
++ {
++ int8x8_t tbl = *(int8x8_t *)offsetEo;
++ x=0;
++ for (; (x + 8) <= width; x+=8)
++ {
++ uint8x8_t in0 = *(uint8x8_t *)&rec[x];
++ uint8x8_t in1 = *(uint8x8_t *)&rec[x+stride+1];
++ int8x8_t vsignDown = sign_diff_neon(in0,in1);
++ int8x8_t vedgeType = vadd_s8(vadd_s8(vsignDown,*(int8x8_t *)&buff1[x]),vdup_n_s8(2));
++ *(int8x8_t *)&bufft[x+1] = vneg_s8(vsignDown);
++ int16x8_t t1 = vmovl_s8(vtbl1_s8(tbl,vedgeType));
++ t1 = vaddw_u8(t1,in0);
++ t1 = vmaxq_s16(t1,vdupq_n_s16(0));
++ t1 = vminq_s16(t1,vdupq_n_s16(255));
++ *(uint8x8_t *)&rec[x] = vmovn_u16(t1);
++ }
++ for (; x < width; x++)
++ {
++ int8_t signDown = signOf(rec[x] - rec[x + stride + 1]);
++ int edgeType = signDown + buff1[x] + 2;
++ bufft[x + 1] = -signDown;
++ rec[x] = x265_clip(rec[x] + offsetEo[edgeType]);;
++ }
++
++ }
++}
++
++
++static void processSaoCUE3_neon(pixel *rec, int8_t *upBuff1, int8_t *offsetEo, intptr_t stride, int startX, int endX)
++{
++ int8_t signDown;
++ int8_t edgeType;
++ int8x8_t tbl = *(int8x8_t *)offsetEo;
++
++ int x = startX + 1;
++ for (; (x+8) <= endX; x+=8 )
++ {
++ uint8x8_t in0 = *(uint8x8_t *)&rec[x];
++ uint8x8_t in1 = *(uint8x8_t *)&rec[x+stride];
++ int8x8_t vsignDown = sign_diff_neon(in0,in1);
++ int8x8_t vedgeType = vadd_s8(vadd_s8(vsignDown,*(int8x8_t *)&upBuff1[x]),vdup_n_s8(2));
++ *(int8x8_t *)&upBuff1[x-1] = vneg_s8(vsignDown);
++ int16x8_t t1 = vmovl_s8(vtbl1_s8(tbl,vedgeType));
++ t1 = vaddw_u8(t1,in0);
++ t1 = vmaxq_s16(t1,vdupq_n_s16(0));
++ t1 = vminq_s16(t1,vdupq_n_s16(255));
++ *(uint8x8_t *)&rec[x] = vmovn_u16(t1);
++
++ }
++ for (; x < endX; x++)
++ {
++ signDown = signOf(rec[x] - rec[x + stride]);
++ edgeType = signDown + upBuff1[x] + 2;
++ upBuff1[x - 1] = -signDown;
++ rec[x] = x265_clip(rec[x] + offsetEo[edgeType]);
++ }
++}
++
++static void processSaoCUB0_neon(pixel* rec, const int8_t* offset, int ctuWidth, int ctuHeight, intptr_t stride)
++{
++ #define SAO_BO_BITS 5
++ const int boShift = X265_DEPTH - SAO_BO_BITS;
++ int x, y;
++ int8x8x4_t table;
++ table = *(int8x8x4_t *)offset;
++
++ for (y = 0; y < ctuHeight; y++)
++ {
++
++ for (x = 0; (x+8) <= ctuWidth; x+=8)
++ {
++ int8x8_t in = *(int8x8_t*)&rec[x];
++ int8x8_t offsets = vtbl4_s8(table,vshr_n_u8(in,boShift));
++ int16x8_t tmp = vmovl_s8(offsets);
++ tmp = vaddw_u8(tmp,in);
++ tmp = vmaxq_s16(tmp,vdupq_n_s16(0));
++ tmp = vminq_s16(tmp,vdupq_n_s16(255));
++ *(uint8x8_t *)&rec[x] = vmovn_u16(tmp);
++ }
++ for (; x < ctuWidth; x++)
++ {
++ rec[x] = x265_clip(rec[x] + offset[rec[x] >> boShift]);
++ }
++ rec += stride;
++ }
++}
++
++}
++
++
++
++namespace X265_NS {
++void setupLoopFilterPrimitives_neon(EncoderPrimitives &p)
++{
++ p.saoCuOrgE0 = processSaoCUE0_neon;
++ p.saoCuOrgE1 = processSaoCUE1_neon;
++ p.saoCuOrgE1_2Rows = processSaoCUE1_2Rows_neon;
++ p.saoCuOrgE2[0] = processSaoCUE2_neon;
++ p.saoCuOrgE2[1] = processSaoCUE2_neon;
++ p.saoCuOrgE3[0] = processSaoCUE3_neon;
++ p.saoCuOrgE3[1] = processSaoCUE3_neon;
++ p.saoCuOrgB0 = processSaoCUB0_neon;
++ p.sign = calSign_neon;
++
++}
++
++#else //HIGH_BIT_DEPTH
++
++
++namespace X265_NS {
++void setupLoopFilterPrimitives_neon(EncoderPrimitives &)
++{
++}
++
++#endif
++
++
++}
+diff -Naur ./source/common/arm64/loopfilter-prim.h ../x265_apple_patch/source/common/arm64/loopfilter-prim.h
+--- ./source/common/arm64/loopfilter-prim.h 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/loopfilter-prim.h 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,43 @@
++#ifndef _LOOPFILTER_NEON_H__
++#define _LOOPFILTER_NEON_H__
++
++
++/*****************************************************************************
++* Copyright (C) 2013-2017 MulticoreWare, Inc
++*
++* Authors: Praveen Kumar Tiwari <praveen@multicorewareinc.com>
++* Dnyaneshwar Gorade <dnyaneshwar@multicorewareinc.com>
++* Min Chen <chenm003@163.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
++* the Free Software Foundation; either version 2 of the License, or
++* (at your option) any later version.
++*
++* This program is distributed in the hope that it will be useful,
++* but WITHOUT ANY WARRANTY; without even the implied warranty of
++* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
++* GNU General Public License for more details.
++*
++* You should have received a copy of the GNU General Public License
++* along with this program; if not, write to the Free Software
++* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
++*
++* This program is also available under a commercial proprietary license.
++* For more information, contact us at license @ x265.com.
++*****************************************************************************/
++
++
++
++#include "common.h"
++#include "primitives.h"
++
++#define PIXEL_MIN 0
++
++namespace X265_NS {
++void setupLoopFilterPrimitives_neon(EncoderPrimitives &p);
++
++};
++
++
++#endif
+diff -Naur ./source/common/arm64/pixel-prim.cpp ../x265_apple_patch/source/common/arm64/pixel-prim.cpp
+--- ./source/common/arm64/pixel-prim.cpp 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/pixel-prim.cpp 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,1940 @@
++#include "common.h"
++#include "slicetype.h" // LOWRES_COST_MASK
++#include "primitives.h"
++#include "x265.h"
++
++#include "pixel-prim.h"
++#include "arm64-utils.h"
++#if HAVE_NEON
++
++#include <arm_neon.h>
++
++using namespace X265_NS;
++
++
++
++namespace {
++
++
++/* SATD SA8D variants - based on x264 */
++static inline void SUMSUB_AB(int16x8_t& sum, int16x8_t& sub, const int16x8_t a, const int16x8_t b)
++{
++ sum = vaddq_s16(a,b);
++ sub = vsubq_s16(a,b);
++}
++
++static inline void transpose_8h(int16x8_t& t1, int16x8_t& t2, const int16x8_t s1, const int16x8_t s2)
++{
++ t1 = vtrn1q_s16(s1, s2);
++ t2 = vtrn2q_s16(s1, s2);
++}
++
++static inline void transpose_4s(int16x8_t& t1, int16x8_t& t2, const int16x8_t s1, const int16x8_t s2)
++{
++ t1 = vtrn1q_s32(s1, s2);
++ t2 = vtrn2q_s32(s1, s2);
++}
++
++#if (X265_DEPTH <= 10)
++static inline void transpose_2d(int16x8_t& t1, int16x8_t& t2, const int16x8_t s1, const int16x8_t s2)
++{
++ t1 = vtrn1q_s64(s1, s2);
++ t2 = vtrn2q_s64(s1, s2);
++}
++#endif
++
++
++static inline void SUMSUB_ABCD(int16x8_t& s1, int16x8_t& d1, int16x8_t& s2, int16x8_t& d2,
++ int16x8_t a,int16x8_t b,int16x8_t c,int16x8_t d)
++{
++ SUMSUB_AB(s1,d1,a,b);
++ SUMSUB_AB(s2,d2,c,d);
++}
++
++static inline void HADAMARD4_V(int16x8_t& r1,int16x8_t& r2,int16x8_t& r3,int16x8_t& r4,
++ int16x8_t& t1,int16x8_t& t2,int16x8_t& t3,int16x8_t& t4)
++{
++ SUMSUB_ABCD(t1, t2, t3, t4, r1, r2, r3, r4);
++ SUMSUB_ABCD(r1, r3, r2, r4, t1, t3, t2, t4);
++}
++
++
++static int _satd_4x8_8x4_end_neon(int16x8_t v0,int16x8_t v1,int16x8_t v2, int16x8_t v3)
++
++{
++
++ int16x8_t v4,v5,v6,v7,v16,v17,v18,v19;
++
++
++ SUMSUB_AB (v16, v17, v0, v1);
++ SUMSUB_AB (v18, v19, v2, v3);
++
++ SUMSUB_AB (v4 , v6 , v16, v18);
++ SUMSUB_AB (v5 , v7 , v17, v19);
++
++ v0 = vtrn1q_s16(v4, v5);
++ v1 = vtrn2q_s16(v4, v5);
++ v2 = vtrn1q_s16(v6, v7);
++ v3 = vtrn2q_s16(v6, v7);
++
++ SUMSUB_AB (v16, v17, v0, v1);
++ SUMSUB_AB (v18, v19, v2, v3);
++
++ v0 = vtrn1q_s32(v16, v18);
++ v1 = vtrn2q_s32(v16, v18);
++ v2 = vtrn1q_s32(v17, v19);
++ v3 = vtrn2q_s32(v17, v19);
++
++ v0 = vabsq_s16(v0);
++ v1 = vabsq_s16(v1);
++ v2 = vabsq_s16(v2);
++ v3 = vabsq_s16(v3);
++
++ v0 = vmaxq_u16(v0, v1);
++ v1 = vmaxq_u16(v2, v3);
++
++ v0 = vaddq_u16(v0, v1);
++ return vaddlvq_u16(v0);
++}
++
++static inline int _satd_4x4_neon(int16x8_t v0, int16x8_t v1)
++{
++ int16x8_t v2,v3;
++ SUMSUB_AB (v2, v3, v0, v1);
++
++ v0 = vzip1q_s64(v2,v3);
++ v1 = vzip2q_s64(v2,v3);
++ SUMSUB_AB (v2, v3, v0, v1);
++
++ v0 = vtrn1q_s16(v2,v3);
++ v1 = vtrn2q_s16(v2,v3);
++ SUMSUB_AB (v2, v3, v0, v1);
++
++ v0 = vtrn1q_s32(v2,v3);
++ v1 = vtrn2q_s32(v2,v3);
++
++ v0 = vabsq_s16(v0);
++ v1 = vabsq_s16(v1);
++ v0 = vmaxq_u16(v0, v1);
++
++ return vaddlvq_s16(v0);
++}
++
++static void _satd_8x4v_8x8h_neon(int16x8_t& v0,int16x8_t& v1, int16x8_t&v2,int16x8_t& v3,int16x8_t& v20,int16x8_t& v21, int16x8_t&v22,int16x8_t& v23)
++{
++ int16x8_t v16,v17,v18,v19,v4,v5,v6,v7;
++
++ SUMSUB_AB(v16, v18, v0, v2);
++ SUMSUB_AB(v17, v19, v1, v3);
++
++ HADAMARD4_V (v20, v21, v22, v23, v0, v1, v2, v3);
++
++ transpose_8h( v0, v1, v16, v17);
++ transpose_8h( v2, v3, v18, v19);
++ transpose_8h( v4, v5, v20, v21);
++ transpose_8h( v6, v7, v22, v23);
++
++ SUMSUB_AB (v16, v17, v0, v1);
++ SUMSUB_AB (v18, v19, v2, v3);
++ SUMSUB_AB (v20, v21, v4, v5);
++ SUMSUB_AB (v22, v23, v6, v7);
++
++ transpose_4s( v0, v2, v16, v18);
++ transpose_4s( v1, v3, v17, v19);
++ transpose_4s( v4, v6, v20, v22);
++ transpose_4s( v5, v7, v21, v23);
++
++ v0 = vabsq_s16(v0);
++ v1 = vabsq_s16(v1);
++ v2 = vabsq_s16(v2);
++ v3 = vabsq_s16(v3);
++ v4 = vabsq_s16(v4);
++ v5 = vabsq_s16(v5);
++ v6 = vabsq_s16(v6);
++ v7 = vabsq_s16(v7);
++
++ v0 = vmaxq_u16(v0,v2);
++ v1 = vmaxq_u16(v1,v3);
++ v2 = vmaxq_u16(v4,v6);
++ v3 = vmaxq_u16(v5,v7);
++
++}
++
++#if HIGH_BIT_DEPTH
++
++#if (X265_DEPTH > 10)
++static inline void transpose_2d(int32x4_t& t1, int32x4_t& t2, const int32x4_t s1, const int32x4_t s2)
++{
++ t1 = vtrn1q_s64(s1, s2);
++ t2 = vtrn2q_s64(s1, s2);
++}
++
++static inline void ISUMSUB_AB(int32x4_t& sum, int32x4_t& sub, const int32x4_t a, const int32x4_t b)
++{
++ sum = vaddq_s32(a,b);
++ sub = vsubq_s32(a,b);
++}
++
++static inline void ISUMSUB_AB_FROM_INT16(int32x4_t& suml, int32x4_t& sumh, int32x4_t& subl, int32x4_t& subh, const int16x8_t a, const int16x8_t b)
++{
++ suml = vaddl_s16(vget_low_s16(a),vget_low_s16(b));
++ sumh = vaddl_high_s16(a,b);
++ subl = vsubl_s16(vget_low_s16(a),vget_low_s16(b));
++ subh = vsubl_high_s16(a, b);
++}
++
++#endif
++
++static inline void _sub_8x8_fly(const uint16_t* pix1, intptr_t stride_pix1, const uint16_t* pix2, intptr_t stride_pix2,
++ int16x8_t& v0,int16x8_t& v1, int16x8_t& v2,int16x8_t& v3,
++ int16x8_t& v20,int16x8_t& v21, int16x8_t& v22,int16x8_t& v23)
++{
++ uint16x8_t r0,r1,r2,r3;
++ uint16x8_t t0,t1,t2,t3;
++ int16x8_t v16,v17;
++ int16x8_t v18,v19;
++
++ r0 = *(uint16x8_t*)(pix1 + 0*stride_pix1);
++ r1 = *(uint16x8_t*)(pix1 + 1*stride_pix1);
++ r2 = *(uint16x8_t*)(pix1 + 2*stride_pix1);
++ r3 = *(uint16x8_t*)(pix1 + 3*stride_pix1);
++
++ t0 = *(uint16x8_t*)(pix2 + 0*stride_pix2);
++ t1 = *(uint16x8_t*)(pix2 + 1*stride_pix2);
++ t2 = *(uint16x8_t*)(pix2 + 2*stride_pix2);
++ t3 = *(uint16x8_t*)(pix2 + 3*stride_pix2);
++
++ v16 = vsubq_u16(r0,t0);
++ v17 = vsubq_u16(r1,t1);
++ v18 = vsubq_u16(r2,t2);
++ v19 = vsubq_u16(r3,t3);
++
++ r0 = *(uint16x8_t*)(pix1 + 4*stride_pix1);
++ r1 = *(uint16x8_t*)(pix1 + 5*stride_pix1);
++ r2 = *(uint16x8_t*)(pix1 + 6*stride_pix1);
++ r3 = *(uint16x8_t*)(pix1 + 7*stride_pix1);
++
++ t0 = *(uint16x8_t*)(pix2 + 4*stride_pix2);
++ t1 = *(uint16x8_t*)(pix2 + 5*stride_pix2);
++ t2 = *(uint16x8_t*)(pix2 + 6*stride_pix2);
++ t3 = *(uint16x8_t*)(pix2 + 7*stride_pix2);
++
++ v20 = vsubq_u16(r0,t0);
++ v21 = vsubq_u16(r1,t1);
++ v22 = vsubq_u16(r2,t2);
++ v23 = vsubq_u16(r3,t3);
++
++ SUMSUB_AB (v0, v1, v16, v17);
++ SUMSUB_AB (v2, v3, v18, v19);
++
++}
++
++
++
++
++static void _satd_16x4_neon(const uint16_t* pix1, intptr_t stride_pix1, const uint16_t* pix2, intptr_t stride_pix2,
++ int16x8_t& v0,int16x8_t&v1, int16x8_t&v2,int16x8_t&v3)
++{
++ uint8x16_t r0,r1,r2,r3;
++ uint8x16_t t0,t1,t2,t3;
++ int16x8_t v16,v17,v20,v21;
++ int16x8_t v18,v19,v22,v23;
++
++ r0 = *(int16x8_t*)(pix1 + 0*stride_pix1);
++ r1 = *(int16x8_t*)(pix1 + 1*stride_pix1);
++ r2 = *(int16x8_t*)(pix1 + 2*stride_pix1);
++ r3 = *(int16x8_t*)(pix1 + 3*stride_pix1);
++
++ t0 = *(int16x8_t*)(pix2 + 0*stride_pix2);
++ t1 = *(int16x8_t*)(pix2 + 1*stride_pix2);
++ t2 = *(int16x8_t*)(pix2 + 2*stride_pix2);
++ t3 = *(int16x8_t*)(pix2 + 3*stride_pix2);
++
++
++ v16 = vsubq_u16((r0),(t0) );
++ v17 = vsubq_u16((r1),(t1) );
++ v18 = vsubq_u16((r2),(t2) );
++ v19 = vsubq_u16((r3),(t3) );
++
++ r0 = *(int16x8_t*)(pix1 + 0*stride_pix1 + 8);
++ r1 = *(int16x8_t*)(pix1 + 1*stride_pix1 + 8);
++ r2 = *(int16x8_t*)(pix1 + 2*stride_pix1 + 8);
++ r3 = *(int16x8_t*)(pix1 + 3*stride_pix1 + 8);
++
++ t0 = *(int16x8_t*)(pix2 + 0*stride_pix2 + 8);
++ t1 = *(int16x8_t*)(pix2 + 1*stride_pix2 + 8);
++ t2 = *(int16x8_t*)(pix2 + 2*stride_pix2 + 8);
++ t3 = *(int16x8_t*)(pix2 + 3*stride_pix2 + 8);
++
++
++ v20 = vsubq_u16(r0,t0);
++ v21 = vsubq_u16(r1,t1);
++ v22 = vsubq_u16(r2,t2);
++ v23 = vsubq_u16(r3,t3);
++
++ SUMSUB_AB (v0, v1, v16, v17);
++ SUMSUB_AB (v2, v3, v18, v19);
++
++ _satd_8x4v_8x8h_neon(v0,v1,v2,v3,v20,v21,v22,v23);
++
++}
++
++
++int pixel_satd_4x4_neon(const uint16_t* pix1, intptr_t stride_pix1, const uint16_t* pix2, intptr_t stride_pix2)
++{
++ uint64x2_t t0,t1,r0,r1;
++ t0[0] = *(uint64_t *)(pix1 + 0*stride_pix1);
++ t1[0] = *(uint64_t *)(pix1 + 1*stride_pix1);
++ t0[1] = *(uint64_t *)(pix1 + 2*stride_pix1);
++ t1[1] = *(uint64_t *)(pix1 + 3*stride_pix1);
++
++ r0[0] = *(uint64_t *)(pix2 + 0*stride_pix1);
++ r1[0] = *(uint64_t *)(pix2 + 1*stride_pix2);
++ r0[1] = *(uint64_t *)(pix2 + 2*stride_pix2);
++ r1[1] = *(uint64_t *)(pix2 + 3*stride_pix2);
++
++ return _satd_4x4_neon(vsubq_u16(t0,r0), vsubq_u16(r1,t1));
++}
++
++
++
++
++
++
++int pixel_satd_8x4_neon(const uint16_t* pix1, intptr_t stride_pix1, const uint16_t* pix2, intptr_t stride_pix2)
++{
++ uint16x8_t i0,i1,i2,i3,i4,i5,i6,i7;
++
++ i0 = *(uint16x8_t *)(pix1 + 0*stride_pix1);
++ i1 = *(uint16x8_t *)(pix2 + 0*stride_pix2);
++ i2 = *(uint16x8_t *)(pix1 + 1*stride_pix1);
++ i3 = *(uint16x8_t *)(pix2 + 1*stride_pix2);
++ i4 = *(uint16x8_t *)(pix1 + 2*stride_pix1);
++ i5 = *(uint16x8_t *)(pix2 + 2*stride_pix2);
++ i6 = *(uint16x8_t *)(pix1 + 3*stride_pix1);
++ i7 = *(uint16x8_t *)(pix2 + 3*stride_pix2);
++
++ int16x8_t v0 = vsubq_u16(i0,i1);
++ int16x8_t v1 = vsubq_u16(i2,i3);
++ int16x8_t v2 = vsubq_u16(i4,i5);
++ int16x8_t v3 = vsubq_u16(i6,i7);
++
++ return _satd_4x8_8x4_end_neon(v0,v1,v2,v3);
++}
++
++
++int pixel_satd_16x16_neon(const uint16_t* pix1, intptr_t stride_pix1, const uint16_t* pix2, intptr_t stride_pix2)
++{
++ int32x4_t v30 = vdupq_n_u32(0),v31= vdupq_n_u32(0);
++ int16x8_t v0,v1,v2,v3;
++
++ _satd_16x4_neon(pix1,stride_pix1,pix2,stride_pix2,v0,v1,v2,v3);
++ v30 = vpadalq_u16(v30,v0);
++ v30 = vpadalq_u16(v30,v1);
++ v31 = vpadalq_u16(v31,v2);
++ v31 = vpadalq_u16(v31,v3);
++
++ _satd_16x4_neon(pix1 + 4*stride_pix1,stride_pix1,pix2+4*stride_pix2,stride_pix2,v0,v1,v2,v3);
++ v30 = vpadalq_u16(v30,v0);
++ v30 = vpadalq_u16(v30,v1);
++ v31 = vpadalq_u16(v31,v2);
++ v31 = vpadalq_u16(v31,v3);
++
++ _satd_16x4_neon(pix1 + 8*stride_pix1,stride_pix1,pix2+8*stride_pix2,stride_pix2,v0,v1,v2,v3);
++ v30 = vpadalq_u16(v30,v0);
++ v30 = vpadalq_u16(v30,v1);
++ v31 = vpadalq_u16(v31,v2);
++ v31 = vpadalq_u16(v31,v3);
++
++ _satd_16x4_neon(pix1 + 12*stride_pix1,stride_pix1,pix2+12*stride_pix2,stride_pix2,v0,v1,v2,v3);
++ v30 = vpadalq_u16(v30,v0);
++ v30 = vpadalq_u16(v30,v1);
++ v31 = vpadalq_u16(v31,v2);
++ v31 = vpadalq_u16(v31,v3);
++
++ return vaddvq_s32(vaddq_s32(v30,v31));
++
++}
++
++#else //HIGH_BIT_DEPTH
++
++static void _satd_16x4_neon(const uint8_t* pix1, intptr_t stride_pix1, const uint8_t* pix2, intptr_t stride_pix2,
++ int16x8_t& v0,int16x8_t&v1, int16x8_t&v2,int16x8_t&v3)
++{
++ uint8x16_t r0,r1,r2,r3;
++ uint8x16_t t0,t1,t2,t3;
++ int16x8_t v16,v17,v20,v21;
++ int16x8_t v18,v19,v22,v23;
++
++ r0 = *(uint8x16_t*)(pix1 + 0*stride_pix1);
++ r1 = *(uint8x16_t*)(pix1 + 1*stride_pix1);
++ r2 = *(uint8x16_t*)(pix1 + 2*stride_pix1);
++ r3 = *(uint8x16_t*)(pix1 + 3*stride_pix1);
++
++ t0 = *(uint8x16_t*)(pix2 + 0*stride_pix2);
++ t1 = *(uint8x16_t*)(pix2 + 1*stride_pix2);
++ t2 = *(uint8x16_t*)(pix2 + 2*stride_pix2);
++ t3 = *(uint8x16_t*)(pix2 + 3*stride_pix2);
++
++
++
++ v16 = vsubl_u8(vget_low_u8(r0),vget_low_u8(t0) );
++ v20 = vsubl_high_u8(r0,t0);
++ v17 = vsubl_u8(vget_low_u8(r1),vget_low_u8(t1) );
++ v21 = vsubl_high_u8(r1,t1);
++ v18 = vsubl_u8(vget_low_u8(r2),vget_low_u8(t2) );
++ v22 = vsubl_high_u8(r2,t2);
++ v19 = vsubl_u8(vget_low_u8(r3),vget_low_u8(t3) );
++ v23 = vsubl_high_u8(r3,t3);
++
++ SUMSUB_AB (v0, v1, v16, v17);
++ SUMSUB_AB (v2, v3, v18, v19);
++
++ _satd_8x4v_8x8h_neon(v0,v1,v2,v3,v20,v21,v22,v23);
++
++}
++
++
++static inline void _sub_8x8_fly(const uint8_t* pix1, intptr_t stride_pix1, const uint8_t* pix2, intptr_t stride_pix2,
++ int16x8_t& v0,int16x8_t& v1, int16x8_t& v2,int16x8_t& v3,
++ int16x8_t& v20,int16x8_t& v21, int16x8_t& v22,int16x8_t& v23)
++{
++ uint8x8_t r0,r1,r2,r3;
++ uint8x8_t t0,t1,t2,t3;
++ int16x8_t v16,v17;
++ int16x8_t v18,v19;
++
++ r0 = *(uint8x8_t*)(pix1 + 0*stride_pix1);
++ r1 = *(uint8x8_t*)(pix1 + 1*stride_pix1);
++ r2 = *(uint8x8_t*)(pix1 + 2*stride_pix1);
++ r3 = *(uint8x8_t*)(pix1 + 3*stride_pix1);
++
++ t0 = *(uint8x8_t*)(pix2 + 0*stride_pix2);
++ t1 = *(uint8x8_t*)(pix2 + 1*stride_pix2);
++ t2 = *(uint8x8_t*)(pix2 + 2*stride_pix2);
++ t3 = *(uint8x8_t*)(pix2 + 3*stride_pix2);
++
++ v16 = vsubl_u8(r0,t0);
++ v17 = vsubl_u8(r1,t1);
++ v18 = vsubl_u8(r2,t2);
++ v19 = vsubl_u8(r3,t3);
++
++ r0 = *(uint8x8_t*)(pix1 + 4*stride_pix1);
++ r1 = *(uint8x8_t*)(pix1 + 5*stride_pix1);
++ r2 = *(uint8x8_t*)(pix1 + 6*stride_pix1);
++ r3 = *(uint8x8_t*)(pix1 + 7*stride_pix1);
++
++ t0 = *(uint8x8_t*)(pix2 + 4*stride_pix2);
++ t1 = *(uint8x8_t*)(pix2 + 5*stride_pix2);
++ t2 = *(uint8x8_t*)(pix2 + 6*stride_pix2);
++ t3 = *(uint8x8_t*)(pix2 + 7*stride_pix2);
++
++ v20 = vsubl_u8(r0,t0);
++ v21 = vsubl_u8(r1,t1);
++ v22 = vsubl_u8(r2,t2);
++ v23 = vsubl_u8(r3,t3);
++
++
++ SUMSUB_AB (v0, v1, v16, v17);
++ SUMSUB_AB (v2, v3, v18, v19);
++
++}
++
++int pixel_satd_4x4_neon(const uint8_t* pix1, intptr_t stride_pix1, const uint8_t* pix2, intptr_t stride_pix2)
++{
++ uint32x2_t t0,t1,r0,r1;
++ t0[0] = *(uint32_t *)(pix1 + 0*stride_pix1);
++ t1[0] = *(uint32_t *)(pix1 + 1*stride_pix1);
++ t0[1] = *(uint32_t *)(pix1 + 2*stride_pix1);
++ t1[1] = *(uint32_t *)(pix1 + 3*stride_pix1);
++
++ r0[0] = *(uint32_t *)(pix2 + 0*stride_pix1);
++ r1[0] = *(uint32_t *)(pix2 + 1*stride_pix2);
++ r0[1] = *(uint32_t *)(pix2 + 2*stride_pix2);
++ r1[1] = *(uint32_t *)(pix2 + 3*stride_pix2);
++
++ return _satd_4x4_neon(vsubl_u8(t0,r0), vsubl_u8(r1,t1));
++}
++
++
++int pixel_satd_8x4_neon(const uint8_t* pix1, intptr_t stride_pix1, const uint8_t* pix2, intptr_t stride_pix2)
++{
++ uint8x8_t i0,i1,i2,i3,i4,i5,i6,i7;
++
++ i0 = *(uint8x8_t *)(pix1 + 0*stride_pix1);
++ i1 = *(uint8x8_t *)(pix2 + 0*stride_pix2);
++ i2 = *(uint8x8_t *)(pix1 + 1*stride_pix1);
++ i3 = *(uint8x8_t *)(pix2 + 1*stride_pix2);
++ i4 = *(uint8x8_t *)(pix1 + 2*stride_pix1);
++ i5 = *(uint8x8_t *)(pix2 + 2*stride_pix2);
++ i6 = *(uint8x8_t *)(pix1 + 3*stride_pix1);
++ i7 = *(uint8x8_t *)(pix2 + 3*stride_pix2);
++
++ int16x8_t v0 = vsubl_u8(i0,i1);
++ int16x8_t v1 = vsubl_u8(i2,i3);
++ int16x8_t v2 = vsubl_u8(i4,i5);
++ int16x8_t v3 = vsubl_u8(i6,i7);
++
++ return _satd_4x8_8x4_end_neon(v0,v1,v2,v3);
++}
++
++int pixel_satd_16x16_neon(const uint8_t* pix1, intptr_t stride_pix1, const uint8_t* pix2, intptr_t stride_pix2)
++{
++ int16x8_t v30,v31;
++ int16x8_t v0,v1,v2,v3;
++
++ _satd_16x4_neon(pix1,stride_pix1,pix2,stride_pix2,v0,v1,v2,v3);
++ v30 = vaddq_s16(v0,v1);
++ v31 = vaddq_s16(v2,v3);
++
++ _satd_16x4_neon(pix1 + 4*stride_pix1,stride_pix1,pix2+4*stride_pix2,stride_pix2,v0,v1,v2,v3);
++ v0 = vaddq_s16(v0,v1);
++ v1 = vaddq_s16(v2,v3);
++ v30 = vaddq_s16(v30, v0);
++ v31 = vaddq_s16(v31, v1);
++
++ _satd_16x4_neon(pix1 + 8*stride_pix1,stride_pix1,pix2+8*stride_pix2,stride_pix2,v0,v1,v2,v3);
++ v0 = vaddq_s16(v0,v1);
++ v1 = vaddq_s16(v2,v3);
++ v30 = vaddq_s16(v30, v0);
++ v31 = vaddq_s16(v31, v1);
++
++ _satd_16x4_neon(pix1 + 12*stride_pix1,stride_pix1,pix2+12*stride_pix2,stride_pix2,v0,v1,v2,v3);
++ v0 = vaddq_s16(v0,v1);
++ v1 = vaddq_s16(v2,v3);
++ v30 = vaddq_s16(v30, v0);
++ v31 = vaddq_s16(v31, v1);
++
++ int32x4_t sum0 = vpaddlq_u16(v30);
++ int32x4_t sum1 = vpaddlq_u16(v31);
++ sum0 = vaddq_s32(sum0,sum1);
++ return vaddvq_s32(sum0);
++
++}
++#endif //HIGH_BIT_DEPTH
++
++
++static inline void _sa8d_8x8_neon_end(int16x8_t& v0,int16x8_t& v1,int16x8_t v2,int16x8_t v3,
++ int16x8_t v20,int16x8_t v21,int16x8_t v22,int16x8_t v23)
++{
++ int16x8_t v16,v17,v18,v19;
++ int16x8_t v4,v5,v6,v7;
++
++ SUMSUB_AB (v16, v18, v0, v2);
++ SUMSUB_AB (v17, v19, v1, v3);
++
++ HADAMARD4_V (v20, v21, v22, v23, v0, v1, v2, v3);
++
++ SUMSUB_AB (v0, v16, v16, v20);
++ SUMSUB_AB (v1, v17, v17, v21);
++ SUMSUB_AB (v2, v18, v18, v22);
++ SUMSUB_AB (v3, v19, v19, v23);
++
++ transpose_8h (v20, v21, v16, v17);
++ transpose_8h (v4, v5, v0, v1);
++ transpose_8h (v22, v23, v18, v19);
++ transpose_8h (v6, v7, v2, v3);
++
++#if (X265_DEPTH <= 10)
++
++ int16x8_t v24,v25;
++
++ SUMSUB_AB (v2, v3, v20, v21);
++ SUMSUB_AB (v24, v25, v4, v5);
++ SUMSUB_AB (v0, v1, v22, v23);
++ SUMSUB_AB (v4, v5, v6, v7);
++
++ transpose_4s (v20, v22, v2, v0);
++ transpose_4s (v21, v23, v3, v1);
++ transpose_4s (v16, v18, v24, v4);
++ transpose_4s (v17, v19, v25, v5);
++
++ SUMSUB_AB (v0, v2, v20, v22);
++ SUMSUB_AB (v1, v3, v21, v23);
++ SUMSUB_AB (v4, v6, v16, v18);
++ SUMSUB_AB (v5, v7, v17, v19);
++
++ transpose_2d (v16, v20, v0, v4);
++ transpose_2d (v17, v21, v1, v5);
++ transpose_2d (v18, v22, v2, v6);
++ transpose_2d (v19, v23, v3, v7);
++
++
++ v16 = vabsq_s16(v16);
++ v17 = vabsq_s16(v17);
++ v18 = vabsq_s16(v18);
++ v19 = vabsq_s16(v19);
++ v20 = vabsq_s16(v20);
++ v21 = vabsq_s16(v21);
++ v22 = vabsq_s16(v22);
++ v23 = vabsq_s16(v23);
++
++ v16 = vmaxq_u16(v16,v20);
++ v17 = vmaxq_u16(v17,v21);
++ v18 = vmaxq_u16(v18,v22);
++ v19 = vmaxq_u16(v19,v23);
++
++#if HIGH_BIT_DEPTH
++ v0 = vpaddlq_u16(v16);
++ v1 = vpaddlq_u16(v17);
++ v0 = vpadalq_u16(v0,v18);
++ v1 = vpadalq_u16(v1,v19);
++
++#else //HIGH_BIT_DEPTH
++
++ v0 = vaddq_u16(v16,v17);
++ v1 = vaddq_u16(v18,v19);
++
++#endif //HIGH_BIT_DEPTH
++
++#else // HIGH_BIT_DEPTH 12 bit only, switching math to int32, each int16x8 is up-convreted to 2 int32x4 (low and high)
++
++ int32x4_t v2l,v2h,v3l,v3h,v24l,v24h,v25l,v25h,v0l,v0h,v1l,v1h;
++ int32x4_t v22l,v22h,v23l,v23h;
++ int32x4_t v4l,v4h,v5l,v5h;
++ int32x4_t v6l,v6h,v7l,v7h;
++ int32x4_t v16l,v16h,v17l,v17h;
++ int32x4_t v18l,v18h,v19l,v19h;
++ int32x4_t v20l,v20h,v21l,v21h;
++
++ ISUMSUB_AB_FROM_INT16(v2l, v2h, v3l, v3h, v20, v21);
++ ISUMSUB_AB_FROM_INT16(v24l, v24h, v25l, v25h, v4, v5);
++
++ v22l = vmovl_s16(vget_low_s16(v22));
++ v22h = vmovl_high_s16(v22);
++ v23l = vmovl_s16(vget_low_s16(v23));
++ v23h = vmovl_high_s16(v23);
++
++ ISUMSUB_AB(v0l, v1l, v22l, v23l);
++ ISUMSUB_AB(v0h, v1h, v22h, v23h);
++
++ v6l = vmovl_s16(vget_low_s16(v6));
++ v6h = vmovl_high_s16(v6);
++ v7l = vmovl_s16(vget_low_s16(v7));
++ v7h = vmovl_high_s16(v7);
++
++ ISUMSUB_AB (v4l, v5l, v6l, v7l);
++ ISUMSUB_AB (v4h, v5h, v6h, v7h);
++
++ transpose_2d (v20l, v22l, v2l, v0l);
++ transpose_2d (v21l, v23l, v3l, v1l);
++ transpose_2d (v16l, v18l, v24l, v4l);
++ transpose_2d (v17l, v19l, v25l, v5l);
++
++ transpose_2d (v20h, v22h, v2h, v0h);
++ transpose_2d (v21h, v23h, v3h, v1h);
++ transpose_2d (v16h, v18h, v24h, v4h);
++ transpose_2d (v17h, v19h, v25h, v5h);
++
++ ISUMSUB_AB (v0l, v2l, v20l, v22l);
++ ISUMSUB_AB (v1l, v3l, v21l, v23l);
++ ISUMSUB_AB (v4l, v6l, v16l, v18l);
++ ISUMSUB_AB (v5l, v7l, v17l, v19l);
++
++ ISUMSUB_AB (v0h, v2h, v20h, v22h);
++ ISUMSUB_AB (v1h, v3h, v21h, v23h);
++ ISUMSUB_AB (v4h, v6h, v16h, v18h);
++ ISUMSUB_AB (v5h, v7h, v17h, v19h);
++
++ v16l = v0l;
++ v16h = v4l;
++ v20l = v0h;
++ v20h = v4h;
++
++ v17l = v1l;
++ v17h = v5l;
++ v21l = v1h;
++ v21h = v5h;
++
++ v18l = v2l;
++ v18h = v6l;
++ v22l = v2h;
++ v22h = v6h;
++
++ v19l = v3l;
++ v19h = v7l;
++ v23l = v3h;
++ v23h = v7h;
++
++ v16l = vabsq_s32(v16l);
++ v17l = vabsq_s32(v17l);
++ v18l = vabsq_s32(v18l);
++ v19l = vabsq_s32(v19l);
++ v20l = vabsq_s32(v20l);
++ v21l = vabsq_s32(v21l);
++ v22l = vabsq_s32(v22l);
++ v23l = vabsq_s32(v23l);
++
++ v16h = vabsq_s32(v16h);
++ v17h = vabsq_s32(v17h);
++ v18h = vabsq_s32(v18h);
++ v19h = vabsq_s32(v19h);
++ v20h = vabsq_s32(v20h);
++ v21h = vabsq_s32(v21h);
++ v22h = vabsq_s32(v22h);
++ v23h = vabsq_s32(v23h);
++
++ v16l = vmaxq_u32(v16l,v20l);
++ v17l = vmaxq_u32(v17l,v21l);
++ v18l = vmaxq_u32(v18l,v22l);
++ v19l = vmaxq_u32(v19l,v23l);
++
++ v16h = vmaxq_u32(v16h,v20h);
++ v17h = vmaxq_u32(v17h,v21h);
++ v18h = vmaxq_u32(v18h,v22h);
++ v19h = vmaxq_u32(v19h,v23h);
++
++ v16l = vaddq_u32(v16l,v16h);
++ v17l = vaddq_u32(v17l,v17h);
++ v18l = vaddq_u32(v18l,v18h);
++ v19l = vaddq_u32(v19l,v19h);
++
++ v0 = vaddq_u32(v16l, v17l);
++ v1 = vaddq_u32(v18l,v19l);
++
++
++#endif
++
++}
++
++
++
++static inline void _satd_8x8_neon(const pixel* pix1, intptr_t stride_pix1, const pixel* pix2, intptr_t stride_pix2,
++ int16x8_t& v0,int16x8_t&v1, int16x8_t&v2,int16x8_t&v3)
++{
++
++ int16x8_t v20,v21,v22,v23;
++ _sub_8x8_fly(pix1,stride_pix1,pix2,stride_pix2,v0,v1,v2,v3,v20,v21,v22,v23);
++ _satd_8x4v_8x8h_neon(v0,v1,v2,v3,v20,v21,v22,v23);
++
++}
++
++
++
++int pixel_satd_8x8_neon(const pixel* pix1, intptr_t stride_pix1, const pixel* pix2, intptr_t stride_pix2)
++{
++ int16x8_t v30,v31;
++ int16x8_t v0,v1,v2,v3;
++
++ _satd_8x8_neon(pix1,stride_pix1,pix2,stride_pix2,v0,v1,v2,v3);
++#if !(HIGH_BIT_DEPTH)
++ v30 = vaddq_u16(v0,v1);
++ v31 = vaddq_u16(v2,v3);
++
++ uint16x8_t sum = vaddq_u16(v30,v31);
++ return vaddvq_s32(vpaddlq_u16(sum));
++#else
++
++ v30 = vaddq_u16(v0,v1);
++ v31 = vaddq_u16(v2,v3);
++
++ int32x4_t sum = vpaddlq_u16(v30);
++ sum = vpadalq_u16(sum, v31);
++ return vaddvq_s32(sum);
++#endif
++}
++
++
++int pixel_sa8d_8x8_neon(const pixel* pix1, intptr_t stride_pix1, const pixel* pix2, intptr_t stride_pix2)
++{
++ int16x8_t v0,v1,v2,v3;
++ int16x8_t v20,v21,v22,v23;
++
++ _sub_8x8_fly(pix1, stride_pix1, pix2, stride_pix2, v0, v1, v2, v3, v20, v21, v22, v23);
++ _sa8d_8x8_neon_end(v0, v1, v2, v3, v20, v21, v22, v23);
++
++#if HIGH_BIT_DEPTH
++//#if 1//HIGH_BIT_DEPTH
++ int32x4_t s = vaddq_u32(v0,v1);
++ return (vaddvq_u32(s) + 1) >> 1;
++#else
++ return (vaddlvq_s16(vaddq_u16(v0, v1)) + 1) >> 1;
++#endif
++}
++
++
++
++
++
++int pixel_sa8d_16x16_neon(const pixel* pix1, intptr_t stride_pix1, const pixel* pix2, intptr_t stride_pix2)
++{
++ int16x8_t v0,v1,v2,v3;
++ int16x8_t v20,v21,v22,v23;
++ int32x4_t v30,v31;
++
++ _sub_8x8_fly(pix1, stride_pix1, pix2, stride_pix2, v0, v1, v2, v3, v20, v21, v22, v23);
++ _sa8d_8x8_neon_end(v0, v1, v2, v3, v20, v21, v22, v23);
++
++#if !(HIGH_BIT_DEPTH)
++ v30 = vpaddlq_u16(v0);
++ v31 = vpaddlq_u16(v1);
++#else
++ v30 = vaddq_s32(v0,v1);
++#endif
++
++ _sub_8x8_fly(pix1 + 8, stride_pix1, pix2 + 8, stride_pix2, v0, v1, v2, v3, v20, v21, v22, v23);
++ _sa8d_8x8_neon_end(v0, v1, v2, v3, v20, v21, v22, v23);
++
++#if !(HIGH_BIT_DEPTH)
++ v30 = vpadalq_u16(v30,v0);
++ v31 = vpadalq_u16(v31,v1);
++#else
++ v31 = vaddq_s32(v0,v1);
++#endif
++
++
++ _sub_8x8_fly(pix1 + 8*stride_pix1, stride_pix1, pix2 + 8*stride_pix2, stride_pix2, v0, v1, v2, v3, v20, v21, v22, v23);
++ _sa8d_8x8_neon_end(v0, v1, v2, v3, v20, v21, v22, v23);
++
++#if !(HIGH_BIT_DEPTH)
++ v30 = vpadalq_u16(v30,v0);
++ v31 = vpadalq_u16(v31,v1);
++#else
++ v30 = vaddq_s32(v30,v0);
++ v31 = vaddq_s32(v31,v1);
++#endif
++
++ _sub_8x8_fly(pix1 + 8*stride_pix1 + 8, stride_pix1, pix2 + 8*stride_pix2 + 8, stride_pix2, v0, v1, v2, v3, v20, v21, v22, v23);
++ _sa8d_8x8_neon_end(v0, v1, v2, v3, v20, v21, v22, v23);
++
++#if !(HIGH_BIT_DEPTH)
++ v30 = vpadalq_u16(v30,v0);
++ v31 = vpadalq_u16(v31,v1);
++#else
++ v30 = vaddq_s32(v30,v0);
++ v31 = vaddq_s32(v31,v1);
++#endif
++
++ v30 = vaddq_u32(v30,v31);
++
++ return (vaddvq_u32(v30) + 1) >> 1;
++}
++
++
++
++
++
++
++
++
++template<int size>
++void blockfill_s_neon(int16_t* dst, intptr_t dstride, int16_t val)
++{
++ for (int y = 0; y < size; y++) {
++ int x = 0;
++ int16x8_t v = vdupq_n_s16(val);
++ for (; (x + 8) <= size; x+=8) {
++ *(int16x8_t*)&dst[y * dstride + x] = v;
++ }
++ for (; x < size; x++) {
++ dst[y * dstride + x] = val;
++ }
++ }
++}
++
++template<int lx, int ly>
++int sad_pp_neon(const pixel* pix1, intptr_t stride_pix1, const pixel* pix2, intptr_t stride_pix2)
++{
++ int sum = 0;
++
++
++ for (int y = 0; y < ly; y++)
++ {
++#if HIGH_BIT_DEPTH
++ int x=0;
++ uint16x8_t vsum16_1 = vdupq_n_u16(0);
++ for (; (x + 8) <= lx; x+=8) {
++ uint16x8_t p1 = *(uint16x8_t*)&pix1[x];
++ uint16x8_t p2 = *(uint16x8_t*)&pix2[x];
++ vsum16_1 = vabaq_s16(vsum16_1,p1,p2);
++
++ }
++ if (lx & 4) {
++ uint16x4_t p1 = *(uint16x4_t*)&pix1[x];
++ uint16x4_t p2 = *(uint16x4_t*)&pix2[x];
++ sum += vaddlv_s16(vaba_s16(vdup_n_s16(0),p1,p2));
++ x += 4;
++ }
++ if (lx >= 4) {
++ sum += vaddlvq_s16(vsum16_1);
++ }
++
++#else
++
++ int x=0;
++ uint16x8_t vsum16_1 = vdupq_n_u16(0);
++ uint16x8_t vsum16_2 = vdupq_n_u16(0);
++
++ for (; (x + 16) <= lx; x+=16) {
++ uint8x16_t p1 = *(uint8x16_t*)&pix1[x];
++ uint8x16_t p2 = *(uint8x16_t*)&pix2[x];
++ vsum16_1 = vabal_u8(vsum16_1,vget_low_u8(p1),vget_low_u8(p2));
++ vsum16_2 = vabal_high_u8(vsum16_2,p1,p2);
++ }
++ if (lx & 8) {
++ uint8x8_t p1 = *(uint8x8_t*)&pix1[x];
++ uint8x8_t p2 = *(uint8x8_t*)&pix2[x];
++ vsum16_1 = vabal_u8(vsum16_1,p1,p2);
++ x += 8;
++ }
++ if (lx & 4) {
++ uint32x2_t p1 = vdup_n_u32(0);
++ p1[0] = *(uint32_t*)&pix1[x];
++ uint32x2_t p2 = vdup_n_u32(0);
++ p2[0] = *(uint32_t*)&pix2[x];
++ vsum16_1 = vabal_u8(vsum16_1,p1,p2);
++ x += 4;
++ }
++ if (lx >= 16) {
++ vsum16_1 = vaddq_u16(vsum16_1,vsum16_2);
++ }
++ if (lx >= 4) {
++ sum += vaddvq_u16(vsum16_1);
++ }
++
++#endif
++ if (lx & 3) for (; x < lx; x++) {
++ sum += abs(pix1[x] - pix2[x]);
++ }
++
++ pix1 += stride_pix1;
++ pix2 += stride_pix2;
++ }
++
++ return sum;
++}
++
++template<int lx, int ly>
++void sad_x3_neon(const pixel* pix1, const pixel* pix2, const pixel* pix3, const pixel* pix4, intptr_t frefstride, int32_t* res)
++{
++ res[0] = 0;
++ res[1] = 0;
++ res[2] = 0;
++ for (int y = 0; y < ly; y++)
++ {
++ int x = 0;
++ uint16x8_t vsum16_0 = vdupq_n_u16(0);
++ uint16x8_t vsum16_1 = vdupq_n_u16(0);
++ uint16x8_t vsum16_2 = vdupq_n_u16(0);
++#if HIGH_BIT_DEPTH
++ for (; (x + 8) <= lx; x+=8) {
++ uint16x8_t p1 = *(uint16x8_t*)&pix1[x];
++ uint16x8_t p2 = *(uint16x8_t*)&pix2[x];
++ uint16x8_t p3 = *(uint16x8_t*)&pix3[x];
++ uint16x8_t p4 = *(uint16x8_t*)&pix4[x];
++ vsum16_0 = vabaq_s16(vsum16_0,p1,p2);
++ vsum16_1 = vabaq_s16(vsum16_1,p1,p3);
++ vsum16_2 = vabaq_s16(vsum16_2,p1,p4);
++
++ }
++ if (lx & 4) {
++ uint16x4_t p1 = *(uint16x4_t*)&pix1[x];
++ uint16x4_t p2 = *(uint16x4_t*)&pix2[x];
++ uint16x4_t p3 = *(uint16x4_t*)&pix3[x];
++ uint16x4_t p4 = *(uint16x4_t*)&pix4[x];
++ res[0] += vaddlv_s16(vaba_s16(vdup_n_s16(0),p1,p2));
++ res[1] += vaddlv_s16(vaba_s16(vdup_n_s16(0),p1,p3));
++ res[2] += vaddlv_s16(vaba_s16(vdup_n_s16(0),p1,p4));
++ x += 4;
++ }
++ if (lx >= 4) {
++ res[0] += vaddlvq_s16(vsum16_0);
++ res[1] += vaddlvq_s16(vsum16_1);
++ res[2] += vaddlvq_s16(vsum16_2);
++ }
++#else
++
++ for (; (x + 16) <= lx; x+=16) {
++ uint8x16_t p1 = *(uint8x16_t*)&pix1[x];
++ uint8x16_t p2 = *(uint8x16_t*)&pix2[x];
++ uint8x16_t p3 = *(uint8x16_t*)&pix3[x];
++ uint8x16_t p4 = *(uint8x16_t*)&pix4[x];
++ vsum16_0 = vabal_u8(vsum16_0,vget_low_u8(p1),vget_low_u8(p2));
++ vsum16_0 = vabal_high_u8(vsum16_0,p1,p2);
++ vsum16_1 = vabal_u8(vsum16_1,vget_low_u8(p1),vget_low_u8(p3));
++ vsum16_1 = vabal_high_u8(vsum16_1,p1,p3);
++ vsum16_2 = vabal_u8(vsum16_2,vget_low_u8(p1),vget_low_u8(p4));
++ vsum16_2 = vabal_high_u8(vsum16_2,p1,p4);
++ }
++ if (lx & 8) {
++ uint8x8_t p1 = *(uint8x8_t*)&pix1[x];
++ uint8x8_t p2 = *(uint8x8_t*)&pix2[x];
++ uint8x8_t p3 = *(uint8x8_t*)&pix3[x];
++ uint8x8_t p4 = *(uint8x8_t*)&pix4[x];
++ vsum16_0 = vabal_u8(vsum16_0,p1,p2);
++ vsum16_1 = vabal_u8(vsum16_1,p1,p3);
++ vsum16_2 = vabal_u8(vsum16_2,p1,p4);
++ x += 8;
++ }
++ if (lx & 4) {
++ uint32x2_t p1 = vdup_n_u32(0);
++ p1[0] = *(uint32_t*)&pix1[x];
++ uint32x2_t p2 = vdup_n_u32(0);
++ p2[0] = *(uint32_t*)&pix2[x];
++ uint32x2_t p3 = vdup_n_u32(0);
++ p3[0] = *(uint32_t*)&pix3[x];
++ uint32x2_t p4 = vdup_n_u32(0);
++ p4[0] = *(uint32_t*)&pix4[x];
++ vsum16_0 = vabal_u8(vsum16_0,p1,p2);
++ vsum16_1 = vabal_u8(vsum16_1,p1,p3);
++ vsum16_2 = vabal_u8(vsum16_2,p1,p4);
++ x += 4;
++ }
++ if (lx >= 4) {
++ res[0] += vaddvq_u16(vsum16_0);
++ res[1] += vaddvq_u16(vsum16_1);
++ res[2] += vaddvq_u16(vsum16_2);
++ }
++
++#endif
++ if (lx & 3) for (; x < lx; x++)
++ {
++ res[0] += abs(pix1[x] - pix2[x]);
++ res[1] += abs(pix1[x] - pix3[x]);
++ res[2] += abs(pix1[x] - pix4[x]);
++ }
++
++ pix1 += FENC_STRIDE;
++ pix2 += frefstride;
++ pix3 += frefstride;
++ pix4 += frefstride;
++ }
++}
++
++template<int lx, int ly>
++void sad_x4_neon(const pixel* pix1, const pixel* pix2, const pixel* pix3, const pixel* pix4, const pixel* pix5, intptr_t frefstride, int32_t* res)
++{
++ res[0] = 0;
++ res[1] = 0;
++ res[2] = 0;
++ res[3] = 0;
++ for (int y = 0; y < ly; y++)
++ {
++ int x=0;
++ uint16x8_t vsum16_0 = vdupq_n_u16(0);
++ uint16x8_t vsum16_1 = vdupq_n_u16(0);
++ uint16x8_t vsum16_2 = vdupq_n_u16(0);
++ uint16x8_t vsum16_3 = vdupq_n_u16(0);
++#if HIGH_BIT_DEPTH
++ for (; (x + 8) <= lx; x+=8) {
++ uint16x8_t p1 = *(uint16x8_t*)&pix1[x];
++ uint16x8_t p2 = *(uint16x8_t*)&pix2[x];
++ uint16x8_t p3 = *(uint16x8_t*)&pix3[x];
++ uint16x8_t p4 = *(uint16x8_t*)&pix4[x];
++ uint16x8_t p5 = *(uint16x8_t*)&pix5[x];
++ vsum16_0 = vabaq_s16(vsum16_0,p1,p2);
++ vsum16_1 = vabaq_s16(vsum16_1,p1,p3);
++ vsum16_2 = vabaq_s16(vsum16_2,p1,p4);
++ vsum16_3 = vabaq_s16(vsum16_3,p1,p5);
++
++ }
++ if (lx & 4) {
++ uint16x4_t p1 = *(uint16x4_t*)&pix1[x];
++ uint16x4_t p2 = *(uint16x4_t*)&pix2[x];
++ uint16x4_t p3 = *(uint16x4_t*)&pix3[x];
++ uint16x4_t p4 = *(uint16x4_t*)&pix4[x];
++ uint16x4_t p5 = *(uint16x4_t*)&pix5[x];
++ res[0] += vaddlv_s16(vaba_s16(vdup_n_s16(0),p1,p2));
++ res[1] += vaddlv_s16(vaba_s16(vdup_n_s16(0),p1,p3));
++ res[2] += vaddlv_s16(vaba_s16(vdup_n_s16(0),p1,p4));
++ res[3] += vaddlv_s16(vaba_s16(vdup_n_s16(0),p1,p5));
++ x += 4;
++ }
++ if (lx >= 4) {
++ res[0] += vaddlvq_s16(vsum16_0);
++ res[1] += vaddlvq_s16(vsum16_1);
++ res[2] += vaddlvq_s16(vsum16_2);
++ res[3] += vaddlvq_s16(vsum16_3);
++ }
++
++#else
++
++ for (; (x + 16) <= lx; x+=16) {
++ uint8x16_t p1 = *(uint8x16_t*)&pix1[x];
++ uint8x16_t p2 = *(uint8x16_t*)&pix2[x];
++ uint8x16_t p3 = *(uint8x16_t*)&pix3[x];
++ uint8x16_t p4 = *(uint8x16_t*)&pix4[x];
++ uint8x16_t p5 = *(uint8x16_t*)&pix5[x];
++ vsum16_0 = vabal_u8(vsum16_0,vget_low_u8(p1),vget_low_u8(p2));
++ vsum16_0 = vabal_high_u8(vsum16_0,p1,p2);
++ vsum16_1 = vabal_u8(vsum16_1,vget_low_u8(p1),vget_low_u8(p3));
++ vsum16_1 = vabal_high_u8(vsum16_1,p1,p3);
++ vsum16_2 = vabal_u8(vsum16_2,vget_low_u8(p1),vget_low_u8(p4));
++ vsum16_2 = vabal_high_u8(vsum16_2,p1,p4);
++ vsum16_3 = vabal_u8(vsum16_3,vget_low_u8(p1),vget_low_u8(p5));
++ vsum16_3 = vabal_high_u8(vsum16_3,p1,p5);
++ }
++ if (lx & 8) {
++ uint8x8_t p1 = *(uint8x8_t*)&pix1[x];
++ uint8x8_t p2 = *(uint8x8_t*)&pix2[x];
++ uint8x8_t p3 = *(uint8x8_t*)&pix3[x];
++ uint8x8_t p4 = *(uint8x8_t*)&pix4[x];
++ uint8x8_t p5 = *(uint8x8_t*)&pix5[x];
++ vsum16_0 = vabal_u8(vsum16_0,p1,p2);
++ vsum16_1 = vabal_u8(vsum16_1,p1,p3);
++ vsum16_2 = vabal_u8(vsum16_2,p1,p4);
++ vsum16_3 = vabal_u8(vsum16_3,p1,p5);
++ x += 8;
++ }
++ if (lx & 4) {
++ uint32x2_t p1 = vdup_n_u32(0);
++ p1[0] = *(uint32_t*)&pix1[x];
++ uint32x2_t p2 = vdup_n_u32(0);
++ p2[0] = *(uint32_t*)&pix2[x];
++ uint32x2_t p3 = vdup_n_u32(0);
++ p3[0] = *(uint32_t*)&pix3[x];
++ uint32x2_t p4 = vdup_n_u32(0);
++ p4[0] = *(uint32_t*)&pix4[x];
++ uint32x2_t p5 = vdup_n_u32(0);
++ p5[0] = *(uint32_t*)&pix5[x];
++ vsum16_0 = vabal_u8(vsum16_0,p1,p2);
++ vsum16_1 = vabal_u8(vsum16_1,p1,p3);
++ vsum16_2 = vabal_u8(vsum16_2,p1,p4);
++ vsum16_3 = vabal_u8(vsum16_3,p1,p5);
++ x += 4;
++ }
++ if (lx >= 4) {
++ res[0] += vaddvq_u16(vsum16_0);
++ res[1] += vaddvq_u16(vsum16_1);
++ res[2] += vaddvq_u16(vsum16_2);
++ res[3] += vaddvq_u16(vsum16_3);
++ }
++
++#endif
++ if (lx & 3) for (; x < lx; x++)
++ {
++ res[0] += abs(pix1[x] - pix2[x]);
++ res[1] += abs(pix1[x] - pix3[x]);
++ res[2] += abs(pix1[x] - pix4[x]);
++ res[3] += abs(pix1[x] - pix5[x]);
++ }
++
++ pix1 += FENC_STRIDE;
++ pix2 += frefstride;
++ pix3 += frefstride;
++ pix4 += frefstride;
++ pix5 += frefstride;
++ }
++}
++
++
++template<int lx, int ly, class T1, class T2>
++sse_t sse_neon(const T1* pix1, intptr_t stride_pix1, const T2* pix2, intptr_t stride_pix2)
++{
++ sse_t sum = 0;
++
++ int32x4_t vsum1 = vdupq_n_s32(0);
++ int32x4_t vsum2 = vdupq_n_s32(0);
++ for (int y = 0; y < ly; y++)
++ {
++ int x = 0;
++ for (; (x+8) <= lx; x+=8)
++ {
++ int16x8_t tmp;
++ if (sizeof(T1) == 2 && sizeof(T2) == 2) {
++ tmp = vsubq_s16(*(int16x8_t *)&pix1[x],*(int16x8_t *)&pix2[x]);
++ } else if (sizeof(T1) == 1 && sizeof(T2) == 1){
++ tmp = vsubl_u8(*(uint8x8_t *)&pix1[x],*(uint8x8_t *)&pix2[x]);
++ }
++ else {
++ X265_CHECK(false,"unsupported sse");
++ }
++ vsum1 = vmlal_s16(vsum1,vget_low_s16(tmp),vget_low_s16(tmp));
++ vsum2 = vmlal_high_s16(vsum2,tmp,tmp);
++ }
++ for (; x < lx; x++)
++ {
++ int tmp = pix1[x] - pix2[x];
++ sum += (tmp * tmp);
++ }
++
++ if (sizeof(T1) == 2 && sizeof(T2) == 2)
++ {
++ int32x4_t vsum = vaddq_u32(vsum1,vsum2);;
++ sum += vaddvq_u32(vsum);
++ vsum1 = vsum2 = vdupq_n_u16(0);
++ }
++
++ pix1 += stride_pix1;
++ pix2 += stride_pix2;
++ }
++ int32x4_t vsum = vaddq_u32(vsum1,vsum2);
++
++ return sum + vaddvq_u32(vsum);
++}
++
++
++template<int bx, int by>
++void blockcopy_ps_neon(int16_t* a, intptr_t stridea, const pixel* b, intptr_t strideb)
++{
++ for (int y = 0; y < by; y++)
++ {
++ int x= 0;
++ for (; (x + 8) <= bx; x+=8)
++ {
++#if HIGH_BIT_DEPTH
++ *(int16x8_t *)&a[x] = *(int16x8_t *)&b[x];
++#else
++ *(int16x8_t *)&a[x] = vmovl_u8(*(int8x8_t *)&b[x]);
++#endif
++ }
++ for (; x < bx; x++) {
++ a[x] = (int16_t)b[x];
++ }
++
++ a += stridea;
++ b += strideb;
++ }
++}
++
++
++template<int bx, int by>
++void blockcopy_pp_neon(pixel* a, intptr_t stridea, const pixel* b, intptr_t strideb)
++{
++ for (int y = 0; y < by; y++)
++ {
++ int x = 0;
++#if HIGH_BIT_DEPTH
++ for (; (x + 8) <= bx; x+=8)
++ {
++ *(int16x8_t *)&a[x] = *(int16x8_t *)&b[x];
++ }
++ if (bx & 4)
++ {
++ *(uint64_t *)&a[x] = *(uint64_t *)&b[x];
++ x += 4;
++ }
++#else
++ for (; (x + 16) <= bx; x+=16)
++ {
++ *(uint8x16_t *)&a[x] = *(uint8x16_t *)&b[x];
++ }
++ if (bx & 8)
++ {
++ *(uint8x8_t *)&a[x] = *(uint8x8_t *)&b[x];
++ x += 8;
++ }
++ if (bx & 4)
++ {
++ *(uint32_t *)&a[x] = *(uint32_t *)&b[x];
++ x += 4;
++ }
++#endif
++ for (; x < bx; x++) {
++ a[x] = b[x];
++ }
++
++ a += stridea;
++ b += strideb;
++ }
++}
++
++
++template<int bx, int by>
++void pixel_sub_ps_neon(int16_t* a, intptr_t dstride, const pixel* b0, const pixel* b1, intptr_t sstride0, intptr_t sstride1)
++{
++ for (int y = 0; y < by; y++)
++ {
++ int x = 0;
++ for (; (x + 8) <= bx; x+=8) {
++#if HIGH_BIT_DEPTH
++ *(int16x8_t *)&a[x] = vsubq_s16(*(int16x8_t *)&b0[x], *(int16x8_t *)&b1[x]);
++#else
++ *(int16x8_t *)&a[x] = vsubl_u8(*(uint8x8_t *)&b0[x], *(uint8x8_t *)&b1[x]);
++#endif
++ }
++ for (; x < bx; x++)
++ a[x] = (int16_t)(b0[x] - b1[x]);
++
++ b0 += sstride0;
++ b1 += sstride1;
++ a += dstride;
++ }
++}
++
++template<int bx, int by>
++void pixel_add_ps_neon(pixel* a, intptr_t dstride, const pixel* b0, const int16_t* b1, intptr_t sstride0, intptr_t sstride1)
++{
++ for (int y = 0; y < by; y++)
++ {
++ int x = 0;
++ for (; (x + 8) <= bx; x+=8) {
++ int16x8_t t;
++ int16x8_t b1e = *(int16x8_t *)&b1[x];
++ int16x8_t b0e;
++#if HIGH_BIT_DEPTH
++ b0e = *(int16x8_t *)&b0[x];
++ t = vaddq_s16(b0e,b1e);
++ t = vminq_s16(t,vdupq_n_s16((1 << X265_DEPTH) - 1));
++ t = vmaxq_s16(t,vdupq_n_s16(0));
++ *(int16x8_t *)&a[x] = t;
++#else
++ b0e = vmovl_u8(*(uint8x8_t *)&b0[x]);
++ t = vaddq_s16(b0e,b1e);
++ *(uint8x8_t *)&a[x] = vqmovun_s16(t);
++#endif
++ }
++ for (; x < bx; x++)
++ a[x] = (int16_t)x265_clip(b0[x] + b1[x]);
++
++ b0 += sstride0;
++ b1 += sstride1;
++ a += dstride;
++ }
++}
++
++template<int bx, int by>
++void addAvg_neon(const int16_t* src0, const int16_t* src1, pixel* dst, intptr_t src0Stride, intptr_t src1Stride, intptr_t dstStride)
++{
++
++ const int shiftNum = IF_INTERNAL_PREC + 1 - X265_DEPTH;
++ const int offset = (1 << (shiftNum - 1)) + 2 * IF_INTERNAL_OFFS;
++
++ const int32x4_t addon = vdupq_n_s32(offset);
++ for (int y = 0; y < by; y++)
++ {
++ int x = 0;
++
++ for (; (x + 8) <= bx; x += 8)
++ {
++ int16x8_t in0 = *(int16x8_t*)&src0[x];
++ int16x8_t in1 = *(int16x8_t*)&src1[x];
++ int32x4_t t1 = vaddl_s16(vget_low_s16(in0),vget_low_s16(in1));
++ int32x4_t t2 = vaddl_high_s16(in0,in1);
++ t1 = vaddq_s32(t1,addon);
++ t2 = vaddq_s32(t2,addon);
++ t1 = vshrq_n_s32(t1,shiftNum);
++ t2 = vshrq_n_s32(t2,shiftNum);
++ int16x8_t t = vuzp1q_s16(t1,t2);
++#if HIGH_BIT_DEPTH
++ t = vminq_s16(t,vdupq_n_s16((1 << X265_DEPTH) - 1));
++ t = vmaxq_s16(t,vdupq_n_s16(0));
++ *(int16x8_t *)&dst[x] = t;
++#else
++ *(uint8x8_t *)&dst[x] = vqmovun_s16(t);
++#endif
++ }
++ for (; x < bx; x += 2)
++ {
++ dst[x + 0] = x265_clip((src0[x + 0] + src1[x + 0] + offset) >> shiftNum);
++ dst[x + 1] = x265_clip((src0[x + 1] + src1[x + 1] + offset) >> shiftNum);
++ }
++
++ src0 += src0Stride;
++ src1 += src1Stride;
++ dst += dstStride;
++ }
++}
++
++template<int lx, int ly>
++void pixelavg_pp_neon(pixel* dst, intptr_t dstride, const pixel* src0, intptr_t sstride0, const pixel* src1, intptr_t sstride1, int)
++{
++ for (int y = 0; y < ly; y++)
++ {
++ int x = 0;
++ for (; (x+8) <= lx; x+=8) {
++#if HIGH_BIT_DEPTH
++ int16x8_t in0 = *(int16x8_t *)&src0[x];
++ int16x8_t in1 = *(int16x8_t *)&src1[x];
++ int16x8_t t = vaddq_s16(in0,in1);
++ t = vaddq_s16(t,vdupq_n_s16(1));
++ t = vshrq_n_s16(t,1);
++ *(int16x8_t *)&dst[x] = t;
++#else
++ int16x8_t in0 = vmovl_u8(*(uint8x8_t *)&src0[x]);
++ int16x8_t in1 = vmovl_u8(*(uint8x8_t *)&src1[x]);
++ int16x8_t t = vaddq_s16(in0,in1);
++ t = vaddq_s16(t,vdupq_n_s16(1));
++ t = vshrq_n_s16(t,1);
++ *(uint8x8_t *)&dst[x] = vmovn_u16(t);
++#endif
++ }
++ for (; x < lx; x++)
++ dst[x] = (src0[x] + src1[x] + 1) >> 1;
++
++ src0 += sstride0;
++ src1 += sstride1;
++ dst += dstride;
++ }
++}
++
++
++template<int size>
++void cpy1Dto2D_shl_neon(int16_t* dst, const int16_t* src, intptr_t dstStride, int shift)
++{
++ X265_CHECK((((intptr_t)dst | (dstStride * sizeof(*dst))) & 15) == 0 || size == 4, "dst alignment error\n");
++ X265_CHECK(((intptr_t)src & 15) == 0, "src alignment error\n");
++ X265_CHECK(shift >= 0, "invalid shift\n");
++
++ for (int i = 0; i < size; i++)
++ {
++ int j = 0;
++ for (; (j+8) <= size; j+=8)
++ {
++ *(int16x8_t *)&dst[j] = vshlq_s16(*(int16x8_t*)&src[j],vdupq_n_s16(shift));
++ }
++ for (; j < size; j++)
++ {
++ dst[j] = src[j] << shift;
++ }
++ src += size;
++ dst += dstStride;
++ }
++}
++
++
++template<int size>
++uint64_t pixel_var_neon(const uint8_t* pix, intptr_t i_stride)
++{
++ uint32_t sum = 0, sqr = 0;
++
++ int32x4_t vsqr = vdupq_n_s32(0);
++ for (int y = 0; y < size; y++)
++ {
++ int x = 0;
++ int16x8_t vsum = vdupq_n_s16(0);
++ for (; (x + 8) <= size; x+=8)
++ {
++ int16x8_t in;
++ in = vmovl_u8(*(uint8x8_t*)&pix[x]);
++ vsum = vaddq_u16(vsum,in);
++ vsqr = vmlal_s16(vsqr,vget_low_s16(in),vget_low_s16(in));
++ vsqr = vmlal_high_s16(vsqr,in,in);
++ }
++ for (; x < size; x++)
++ {
++ sum += pix[x];
++ sqr += pix[x] * pix[x];
++ }
++ sum += vaddvq_s16(vsum);
++
++ pix += i_stride;
++ }
++ sqr += vaddvq_u32(vsqr);
++ return sum + ((uint64_t)sqr << 32);
++}
++
++template<int blockSize>
++void getResidual_neon(const pixel* fenc, const pixel* pred, int16_t* residual, intptr_t stride)
++{
++ for (int y = 0; y < blockSize; y++)
++ {
++ int x = 0;
++ for (; (x + 8) < blockSize; x+=8) {
++ int16x8_t vfenc,vpred;
++#if HIGH_BIT_DEPTH
++ vfenc = *(int16x8_t *)&fenc[x];
++ vpred = *(int16x8_t *)&pred[x];
++#else
++ vfenc = vmovl_u8(*(uint8x8_t *)&fenc[x]);
++ vpred = vmovl_u8(*(uint8x8_t *)&pred[x]);
++#endif
++ *(int16x8_t*)&residual[x] = vsubq_s16(vfenc,vpred);
++ }
++ for (; x < blockSize; x++) {
++ residual[x] = static_cast<int16_t>(fenc[x]) - static_cast<int16_t>(pred[x]);
++ }
++ fenc += stride;
++ residual += stride;
++ pred += stride;
++ }
++}
++
++#if 1//!(HIGH_BIT_DEPTH)
++template<int size>
++int psyCost_pp_neon(const pixel* source, intptr_t sstride, const pixel* recon, intptr_t rstride)
++{
++ static pixel zeroBuf[8] /* = { 0 } */;
++
++ if (size)
++ {
++ int dim = 1 << (size + 2);
++ uint32_t totEnergy = 0;
++ for (int i = 0; i < dim; i += 8)
++ {
++ for (int j = 0; j < dim; j+= 8)
++ {
++ /* AC energy, measured by sa8d (AC + DC) minus SAD (DC) */
++ int sourceEnergy = pixel_sa8d_8x8_neon(source + i * sstride + j, sstride, zeroBuf, 0) -
++ (sad_pp_neon<8, 8>(source + i * sstride + j, sstride, zeroBuf, 0) >> 2);
++ int reconEnergy = pixel_sa8d_8x8_neon(recon + i * rstride + j, rstride, zeroBuf, 0) -
++ (sad_pp_neon<8, 8>(recon + i * rstride + j, rstride, zeroBuf, 0) >> 2);
++
++ totEnergy += abs(sourceEnergy - reconEnergy);
++ }
++ }
++ return totEnergy;
++ }
++ else
++ {
++ /* 4x4 is too small for sa8d */
++ int sourceEnergy = pixel_satd_4x4_neon(source, sstride, zeroBuf, 0) - (sad_pp_neon<4, 4>(source, sstride, zeroBuf, 0) >> 2);
++ int reconEnergy = pixel_satd_4x4_neon(recon, rstride, zeroBuf, 0) - (sad_pp_neon<4, 4>(recon, rstride, zeroBuf, 0) >> 2);
++ return abs(sourceEnergy - reconEnergy);
++ }
++}
++
++
++template<int w, int h>
++// Calculate sa8d in blocks of 8x8
++int sa8d8(const pixel* pix1, intptr_t i_pix1, const pixel* pix2, intptr_t i_pix2)
++{
++ int cost = 0;
++
++ for (int y = 0; y < h; y += 8)
++ for (int x = 0; x < w; x += 8)
++ cost += pixel_sa8d_8x8_neon(pix1 + i_pix1 * y + x, i_pix1, pix2 + i_pix2 * y + x, i_pix2);
++
++ return cost;
++}
++
++template<int w, int h>
++// Calculate sa8d in blocks of 16x16
++int sa8d16(const pixel* pix1, intptr_t i_pix1, const pixel* pix2, intptr_t i_pix2)
++{
++ int cost = 0;
++
++ for (int y = 0; y < h; y += 16)
++ for (int x = 0; x < w; x += 16)
++ cost += pixel_sa8d_16x16_neon(pix1 + i_pix1 * y + x, i_pix1, pix2 + i_pix2 * y + x, i_pix2);
++
++ return cost;
++}
++#endif
++
++template<int size>
++void cpy2Dto1D_shl_neon(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift)
++{
++ X265_CHECK(((intptr_t)dst & 15) == 0, "dst alignment error\n");
++ X265_CHECK((((intptr_t)src | (srcStride * sizeof(*src))) & 15) == 0 || size == 4, "src alignment error\n");
++ X265_CHECK(shift >= 0, "invalid shift\n");
++
++ for (int i = 0; i < size; i++)
++ {
++ for (int j = 0; j < size; j++)
++ dst[j] = src[j] << shift;
++
++ src += srcStride;
++ dst += size;
++ }
++}
++
++
++#if 1//!(HIGH_BIT_DEPTH)
++template<int w, int h>
++// calculate satd in blocks of 4x4
++int satd4(const pixel* pix1, intptr_t stride_pix1, const pixel* pix2, intptr_t stride_pix2)
++{
++ int satd = 0;
++
++ for (int row = 0; row < h; row += 4)
++ for (int col = 0; col < w; col += 4)
++ satd += pixel_satd_4x4_neon(pix1 + row * stride_pix1 + col, stride_pix1,
++ pix2 + row * stride_pix2 + col, stride_pix2);
++
++ return satd;
++}
++
++template<int w, int h>
++// calculate satd in blocks of 8x4
++int satd8(const pixel* pix1, intptr_t stride_pix1, const pixel* pix2, intptr_t stride_pix2)
++{
++ int satd = 0;
++
++ if (((w | h) & 15) == 0)
++ {
++ for (int row = 0; row < h; row += 16)
++ for (int col = 0; col < w; col += 16)
++ satd += pixel_satd_16x16_neon(pix1 + row * stride_pix1 + col, stride_pix1,
++ pix2 + row * stride_pix2 + col, stride_pix2);
++
++ }
++ else
++ if (((w | h) & 7) == 0)
++ {
++ for (int row = 0; row < h; row += 8)
++ for (int col = 0; col < w; col += 8)
++ satd += pixel_satd_8x8_neon(pix1 + row * stride_pix1 + col, stride_pix1,
++ pix2 + row * stride_pix2 + col, stride_pix2);
++
++ }
++ else
++ {
++ for (int row = 0; row < h; row += 4)
++ for (int col = 0; col < w; col += 8)
++ satd += pixel_satd_8x4_neon(pix1 + row * stride_pix1 + col, stride_pix1,
++ pix2 + row * stride_pix2 + col, stride_pix2);
++ }
++
++ return satd;
++}
++#endif
++
++
++template<int blockSize>
++void transpose_neon(pixel* dst, const pixel* src, intptr_t stride)
++{
++ for (int k = 0; k < blockSize; k++)
++ for (int l = 0; l < blockSize; l++)
++ dst[k * blockSize + l] = src[l * stride + k];
++}
++
++
++template<>
++void transpose_neon<8>(pixel* dst, const pixel* src, intptr_t stride)
++{
++ transpose8x8(dst,src,8,stride);
++}
++
++template<>
++void transpose_neon<16>(pixel* dst, const pixel* src, intptr_t stride)
++{
++ transpose16x16(dst,src,16,stride);
++}
++
++template<>
++void transpose_neon<32>(pixel* dst, const pixel* src, intptr_t stride)
++{
++ transpose32x32(dst,src,32,stride);
++}
++
++
++template<>
++void transpose_neon<64>(pixel* dst, const pixel* src, intptr_t stride)
++{
++ transpose32x32(dst,src,64,stride);
++ transpose32x32(dst+32*64+32,src+32*stride+32,64,stride);
++ transpose32x32(dst+32*64,src+32,64,stride);
++ transpose32x32(dst+32,src+32*stride,64,stride);
++}
++
++
++template<int size>
++sse_t pixel_ssd_s_neon(const int16_t* a, intptr_t dstride)
++{
++ sse_t sum = 0;
++
++
++ int32x4_t vsum = vdupq_n_s32(0);
++
++ for (int y = 0; y < size; y++)
++ {
++ int x = 0;
++
++ for (; (x + 8) <= size; x+=8) {
++ int16x8_t in = *(int16x8_t*)&a[x];
++ vsum = vmlal_s16(vsum,vget_low_s16(in),vget_low_s16(in));
++ vsum = vmlal_high_s16(vsum,(in),(in));
++ }
++ for (; x < size; x++) {
++ sum += a[x] * a[x];
++ }
++
++ a += dstride;
++ }
++ return sum + vaddvq_s32(vsum);
++}
++
++
++};
++
++
++
++
++namespace X265_NS {
++
++
++void setupPixelPrimitives_neon(EncoderPrimitives &p)
++{
++ #define LUMA_PU(W, H) \
++ p.pu[LUMA_ ## W ## x ## H].copy_pp = blockcopy_pp_neon<W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].addAvg[NONALIGNED] = addAvg_neon<W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].addAvg[ALIGNED] = addAvg_neon<W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].sad = sad_pp_neon<W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].sad_x3 = sad_x3_neon<W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].sad_x4 = sad_x4_neon<W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].pixelavg_pp[NONALIGNED] = pixelavg_pp_neon<W, H>; \
++ p.pu[LUMA_ ## W ## x ## H].pixelavg_pp[ALIGNED] = pixelavg_pp_neon<W, H>;
++
++#if !(HIGH_BIT_DEPTH)
++
++#define LUMA_CU(W, H) \
++ p.cu[BLOCK_ ## W ## x ## H].sub_ps = pixel_sub_ps_neon<W, H>; \
++ p.cu[BLOCK_ ## W ## x ## H].add_ps[NONALIGNED] = pixel_add_ps_neon<W, H>; \
++ p.cu[BLOCK_ ## W ## x ## H].add_ps[ALIGNED] = pixel_add_ps_neon<W, H>; \
++ p.cu[BLOCK_ ## W ## x ## H].copy_ps = blockcopy_ps_neon<W, H>; \
++ p.cu[BLOCK_ ## W ## x ## H].copy_pp = blockcopy_pp_neon<W, H>; \
++ p.cu[BLOCK_ ## W ## x ## H].blockfill_s[NONALIGNED] = blockfill_s_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].blockfill_s[ALIGNED] = blockfill_s_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].cpy2Dto1D_shl = cpy2Dto1D_shl_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].cpy1Dto2D_shl[NONALIGNED] = cpy1Dto2D_shl_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].cpy1Dto2D_shl[ALIGNED] = cpy1Dto2D_shl_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].psy_cost_pp = psyCost_pp_neon<BLOCK_ ## W ## x ## H>; \
++ p.cu[BLOCK_ ## W ## x ## H].transpose = transpose_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].var = pixel_var_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].calcresidual[NONALIGNED] = getResidual_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].calcresidual[ALIGNED] = getResidual_neon<W>; \
++
++#else
++
++ #define LUMA_CU(W, H) \
++ p.cu[BLOCK_ ## W ## x ## H].sub_ps = pixel_sub_ps_neon<W, H>; \
++ p.cu[BLOCK_ ## W ## x ## H].add_ps[NONALIGNED] = pixel_add_ps_neon<W, H>; \
++ p.cu[BLOCK_ ## W ## x ## H].add_ps[ALIGNED] = pixel_add_ps_neon<W, H>; \
++ p.cu[BLOCK_ ## W ## x ## H].copy_pp = blockcopy_pp_neon<W, H>; \
++ p.cu[BLOCK_ ## W ## x ## H].copy_ps = blockcopy_ps_neon<W, H>; \
++ p.cu[BLOCK_ ## W ## x ## H].copy_pp = blockcopy_pp_neon<W, H>; \
++ p.cu[BLOCK_ ## W ## x ## H].blockfill_s[NONALIGNED] = blockfill_s_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].blockfill_s[ALIGNED] = blockfill_s_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].cpy2Dto1D_shl = cpy2Dto1D_shl_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].cpy1Dto2D_shl[NONALIGNED] = cpy1Dto2D_shl_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].cpy1Dto2D_shl[ALIGNED] = cpy1Dto2D_shl_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].psy_cost_pp = psyCost_pp_neon<BLOCK_ ## W ## x ## H>; \
++ p.cu[BLOCK_ ## W ## x ## H].transpose = transpose_neon<W>; \
++ /*p.cu[BLOCK_ ## W ## x ## H].var = pixel_var_neon<W>;*/ \
++ p.cu[BLOCK_ ## W ## x ## H].calcresidual[NONALIGNED] = getResidual_neon<W>; \
++ p.cu[BLOCK_ ## W ## x ## H].calcresidual[ALIGNED] = getResidual_neon<W>; \
++
++
++
++#endif
++
++
++ LUMA_PU(4, 4);
++ LUMA_PU(8, 8);
++ LUMA_PU(16, 16);
++ LUMA_PU(32, 32);
++ LUMA_PU(64, 64);
++ LUMA_PU(4, 8);
++ LUMA_PU(8, 4);
++ LUMA_PU(16, 8);
++ LUMA_PU(8, 16);
++ LUMA_PU(16, 12);
++ LUMA_PU(12, 16);
++ LUMA_PU(16, 4);
++ LUMA_PU(4, 16);
++ LUMA_PU(32, 16);
++ LUMA_PU(16, 32);
++ LUMA_PU(32, 24);
++ LUMA_PU(24, 32);
++ LUMA_PU(32, 8);
++ LUMA_PU(8, 32);
++ LUMA_PU(64, 32);
++ LUMA_PU(32, 64);
++ LUMA_PU(64, 48);
++ LUMA_PU(48, 64);
++ LUMA_PU(64, 16);
++ LUMA_PU(16, 64);
++
++ p.pu[LUMA_4x4].satd = pixel_satd_4x4_neon;
++ p.pu[LUMA_8x8].satd = satd8<8, 8>;
++ p.pu[LUMA_8x4].satd = pixel_satd_8x4_neon;
++ p.pu[LUMA_4x8].satd = satd4<4, 8>;
++ p.pu[LUMA_16x16].satd = satd8<16, 16>;
++ p.pu[LUMA_16x8].satd = satd8<16, 8>;
++ p.pu[LUMA_8x16].satd = satd8<8, 16>;
++ p.pu[LUMA_16x12].satd = satd8<16, 12>;
++ p.pu[LUMA_12x16].satd = satd4<12, 16>;
++ p.pu[LUMA_16x4].satd = satd8<16, 4>;
++ p.pu[LUMA_4x16].satd = satd4<4, 16>;
++ p.pu[LUMA_32x32].satd = satd8<32, 32>;
++ p.pu[LUMA_32x16].satd = satd8<32, 16>;
++ p.pu[LUMA_16x32].satd = satd8<16, 32>;
++ p.pu[LUMA_32x24].satd = satd8<32, 24>;
++ p.pu[LUMA_24x32].satd = satd8<24, 32>;
++ p.pu[LUMA_32x8].satd = satd8<32, 8>;
++ p.pu[LUMA_8x32].satd = satd8<8, 32>;
++ p.pu[LUMA_64x64].satd = satd8<64, 64>;
++ p.pu[LUMA_64x32].satd = satd8<64, 32>;
++ p.pu[LUMA_32x64].satd = satd8<32, 64>;
++ p.pu[LUMA_64x48].satd = satd8<64, 48>;
++ p.pu[LUMA_48x64].satd = satd8<48, 64>;
++ p.pu[LUMA_64x16].satd = satd8<64, 16>;
++ p.pu[LUMA_16x64].satd = satd8<16, 64>;
++
++
++ LUMA_CU(4, 4);
++ LUMA_CU(8, 8);
++ LUMA_CU(16, 16);
++ LUMA_CU(32, 32);
++ LUMA_CU(64, 64);
++
++
++ p.cu[BLOCK_4x4].sa8d = pixel_satd_4x4_neon;
++ p.cu[BLOCK_8x8].sa8d = pixel_sa8d_8x8_neon;
++ p.cu[BLOCK_16x16].sa8d = pixel_sa8d_16x16_neon;
++ p.cu[BLOCK_32x32].sa8d = sa8d16<32, 32>;
++ p.cu[BLOCK_64x64].sa8d = sa8d16<64, 64>;
++
++
++ #define CHROMA_PU_420(W, H) \
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_ ## W ## x ## H].addAvg[NONALIGNED] = addAvg_neon<W, H>; \
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_ ## W ## x ## H].addAvg[ALIGNED] = addAvg_neon<W, H>; \
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_ ## W ## x ## H].copy_pp = blockcopy_pp_neon<W, H>; \
++
++
++ CHROMA_PU_420(4, 4);
++ CHROMA_PU_420(8, 8);
++ CHROMA_PU_420(16, 16);
++ CHROMA_PU_420(32, 32);
++ CHROMA_PU_420(4, 2);
++ CHROMA_PU_420(8, 4);
++ CHROMA_PU_420(4, 8);
++ CHROMA_PU_420(8, 6);
++ CHROMA_PU_420(6, 8);
++ CHROMA_PU_420(8, 2);
++ CHROMA_PU_420(2, 8);
++ CHROMA_PU_420(16, 8);
++ CHROMA_PU_420(8, 16);
++ CHROMA_PU_420(16, 12);
++ CHROMA_PU_420(12, 16);
++ CHROMA_PU_420(16, 4);
++ CHROMA_PU_420(4, 16);
++ CHROMA_PU_420(32, 16);
++ CHROMA_PU_420(16, 32);
++ CHROMA_PU_420(32, 24);
++ CHROMA_PU_420(24, 32);
++ CHROMA_PU_420(32, 8);
++ CHROMA_PU_420(8, 32);
++
++
++
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_2x2].satd = NULL;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].satd = pixel_satd_4x4_neon;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].satd = satd8<8, 8>;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].satd = satd8<16, 16>;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].satd = satd8<32, 32>;
++
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].satd = NULL;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_2x4].satd = NULL;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].satd = pixel_satd_8x4_neon;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].satd = satd4<4, 8>;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].satd = satd8<16, 8>;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].satd = satd8<8, 16>;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].satd = satd8<32, 16>;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].satd = satd8<16, 32>;
++
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].satd = NULL;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].satd = NULL;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].satd = NULL;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_2x8].satd = NULL;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].satd = satd4<16, 12>;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].satd = satd4<12, 16>;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].satd = satd4<16, 4>;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].satd = satd4<4, 16>;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].satd = satd8<32, 24>;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].satd = satd8<24, 32>;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].satd = satd8<32, 8>;
++ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].satd = satd8<8, 32>;
++
++
++ #define CHROMA_CU_420(W, H) \
++ p.chroma[X265_CSP_I420].cu[BLOCK_420_ ## W ## x ## H].sse_pp = sse_neon<W, H, pixel, pixel>; \
++ p.chroma[X265_CSP_I420].cu[BLOCK_420_ ## W ## x ## H].copy_pp = blockcopy_pp_neon<W, H>; \
++ p.chroma[X265_CSP_I420].cu[BLOCK_420_ ## W ## x ## H].copy_ps = blockcopy_ps_neon<W, H>; \
++ p.chroma[X265_CSP_I420].cu[BLOCK_420_ ## W ## x ## H].sub_ps = pixel_sub_ps_neon<W, H>; \
++ p.chroma[X265_CSP_I420].cu[BLOCK_420_ ## W ## x ## H].add_ps[NONALIGNED] = pixel_add_ps_neon<W, H>; \
++ p.chroma[X265_CSP_I420].cu[BLOCK_420_ ## W ## x ## H].add_ps[ALIGNED] = pixel_add_ps_neon<W, H>;
++
++
++ CHROMA_CU_420(4, 4)
++ CHROMA_CU_420(8, 8)
++ CHROMA_CU_420(16, 16)
++ CHROMA_CU_420(32, 32)
++
++
++ p.chroma[X265_CSP_I420].cu[BLOCK_8x8].sa8d = p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].satd;
++ p.chroma[X265_CSP_I420].cu[BLOCK_16x16].sa8d = sa8d8<8, 8>;
++ p.chroma[X265_CSP_I420].cu[BLOCK_32x32].sa8d = sa8d16<16, 16>;
++ p.chroma[X265_CSP_I420].cu[BLOCK_64x64].sa8d = sa8d16<32, 32>;
++
++
++ #define CHROMA_PU_422(W, H) \
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_ ## W ## x ## H].addAvg[NONALIGNED] = addAvg_neon<W, H>; \
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_ ## W ## x ## H].addAvg[ALIGNED] = addAvg_neon<W, H>; \
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_ ## W ## x ## H].copy_pp = blockcopy_pp_neon<W, H>; \
++
++
++ CHROMA_PU_422(4, 8);
++ CHROMA_PU_422(8, 16);
++ CHROMA_PU_422(16, 32);
++ CHROMA_PU_422(32, 64);
++ CHROMA_PU_422(4, 4);
++ CHROMA_PU_422(2, 8);
++ CHROMA_PU_422(8, 8);
++ CHROMA_PU_422(4, 16);
++ CHROMA_PU_422(8, 12);
++ CHROMA_PU_422(6, 16);
++ CHROMA_PU_422(8, 4);
++ CHROMA_PU_422(2, 16);
++ CHROMA_PU_422(16, 16);
++ CHROMA_PU_422(8, 32);
++ CHROMA_PU_422(16, 24);
++ CHROMA_PU_422(12, 32);
++ CHROMA_PU_422(16, 8);
++ CHROMA_PU_422(4, 32);
++ CHROMA_PU_422(32, 32);
++ CHROMA_PU_422(16, 64);
++ CHROMA_PU_422(32, 48);
++ CHROMA_PU_422(24, 64);
++ CHROMA_PU_422(32, 16);
++ CHROMA_PU_422(8, 64);
++
++
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_2x4].satd = NULL;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].satd = satd4<4, 8>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].satd = satd8<8, 16>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].satd = satd8<16, 32>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].satd = satd8<32, 64>;
++
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_4x4].satd = pixel_satd_4x4_neon;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_2x8].satd = NULL;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].satd = satd8<8, 8>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].satd = satd4<4, 16>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].satd = satd8<16, 16>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].satd = satd8<8, 32>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].satd = satd8<32, 32>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].satd = satd8<16, 64>;
++
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].satd = satd4<8, 12>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].satd = NULL;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].satd = satd4<8, 4>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_2x16].satd = NULL;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].satd = satd8<16, 24>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].satd = satd4<12, 32>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].satd = satd8<16, 8>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_4x32].satd = satd4<4, 32>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].satd = satd8<32, 48>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].satd = satd8<24, 64>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].satd = satd8<32, 16>;
++ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].satd = satd8<8, 64>;
++
++
++ #define CHROMA_CU_422(W, H) \
++ p.chroma[X265_CSP_I422].cu[BLOCK_422_ ## W ## x ## H].sse_pp = sse_neon<W, H, pixel, pixel>; \
++ p.chroma[X265_CSP_I422].cu[BLOCK_422_ ## W ## x ## H].copy_pp = blockcopy_pp_neon<W, H>; \
++ p.chroma[X265_CSP_I422].cu[BLOCK_422_ ## W ## x ## H].copy_ps = blockcopy_ps_neon<W, H>; \
++ p.chroma[X265_CSP_I422].cu[BLOCK_422_ ## W ## x ## H].sub_ps = pixel_sub_ps_neon<W, H>; \
++ p.chroma[X265_CSP_I422].cu[BLOCK_422_ ## W ## x ## H].add_ps[NONALIGNED] = pixel_add_ps_neon<W, H>; \
++ p.chroma[X265_CSP_I422].cu[BLOCK_422_ ## W ## x ## H].add_ps[ALIGNED] = pixel_add_ps_neon<W, H>;
++
++
++ CHROMA_CU_422(4, 8)
++ CHROMA_CU_422(8, 16)
++ CHROMA_CU_422(16, 32)
++ CHROMA_CU_422(32, 64)
++
++ p.chroma[X265_CSP_I422].cu[BLOCK_8x8].sa8d = p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].satd;
++ p.chroma[X265_CSP_I422].cu[BLOCK_16x16].sa8d = sa8d8<8, 16>;
++ p.chroma[X265_CSP_I422].cu[BLOCK_32x32].sa8d = sa8d16<16, 32>;
++ p.chroma[X265_CSP_I422].cu[BLOCK_64x64].sa8d = sa8d16<32, 64>;
++
++
++}
++
++
++}
++
++
++#endif
++
+diff -Naur ./source/common/arm64/pixel-prim.h ../x265_apple_patch/source/common/arm64/pixel-prim.h
+--- ./source/common/arm64/pixel-prim.h 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/pixel-prim.h 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,22 @@
++#ifndef PIXEL_PRIM_NEON_H__
++#define PIXEL_PRIM_NEON_H__
++
++#include "common.h"
++#include "slicetype.h" // LOWRES_COST_MASK
++#include "primitives.h"
++#include "x265.h"
++
++
++
++namespace X265_NS {
++
++
++
++void setupPixelPrimitives_neon(EncoderPrimitives &p);
++
++
++}
++
++
++#endif
++
+diff -Naur ./source/common/arm64/pixel.h ../x265_apple_patch/source/common/arm64/pixel.h
+--- ./source/common/arm64/pixel.h 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/common/arm64/pixel.h 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,134 @@
++/*****************************************************************************
++ * pixel.h: aarch64 pixel metrics
++ *****************************************************************************
++ * Copyright (C) 2009-2019 x265 project
++ *
++ * Authors: David Conrad <lessen42@gmail.com>
++ * Janne Grunau <janne-x265@jannau.net>
++ *
++ * 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
++ * the Free Software Foundation; either version 2 of the License, or
++ * (at your option) any later version.
++ *
++ * This program is distributed in the hope that it will be useful,
++ * but WITHOUT ANY WARRANTY; without even the implied warranty of
++ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
++ * GNU General Public License for more details.
++ *
++ * You should have received a copy of the GNU General Public License
++ * along with this program; if not, write to the Free Software
++ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
++ *
++ * This program is also available under a commercial proprietary license.
++ * For more information, contact us at licensing@x265.com.
++ *****************************************************************************/
++
++#ifndef x265_AARCH64_PIXEL_H
++#define x265_AARCH64_PIXEL_H
++
++#define x265_pixel_sad_16x16_neon x265_template(pixel_sad_16x16_neon)
++#define x265_pixel_sad_16x8_neon x265_template(pixel_sad_16x8_neon)
++#define x265_pixel_sad_4x16_neon x265_template(pixel_sad_4x16_neon)
++#define x265_pixel_sad_4x4_neon x265_template(pixel_sad_4x4_neon)
++#define x265_pixel_sad_4x8_neon x265_template(pixel_sad_4x8_neon)
++#define x265_pixel_sad_8x16_neon x265_template(pixel_sad_8x16_neon)
++#define x265_pixel_sad_8x4_neon x265_template(pixel_sad_8x4_neon)
++#define x265_pixel_sad_8x8_neon x265_template(pixel_sad_8x8_neon)
++#define x265_pixel_sad_x3_16x16_neon x265_template(pixel_sad_x3_16x16_neon)
++#define x265_pixel_sad_x3_16x8_neon x265_template(pixel_sad_x3_16x8_neon)
++#define x265_pixel_sad_x3_4x4_neon x265_template(pixel_sad_x3_4x4_neon)
++#define x265_pixel_sad_x3_4x8_neon x265_template(pixel_sad_x3_4x8_neon)
++#define x265_pixel_sad_x3_8x16_neon x265_template(pixel_sad_x3_8x16_neon)
++#define x265_pixel_sad_x3_8x4_neon x265_template(pixel_sad_x3_8x4_neon)
++#define x265_pixel_sad_x3_8x8_neon x265_template(pixel_sad_x3_8x8_neon)
++#define x265_pixel_sad_x4_16x16_neon x265_template(pixel_sad_x4_16x16_neon)
++#define x265_pixel_sad_x4_16x8_neon x265_template(pixel_sad_x4_16x8_neon)
++#define x265_pixel_sad_x4_4x4_neon x265_template(pixel_sad_x4_4x4_neon)
++#define x265_pixel_sad_x4_4x8_neon x265_template(pixel_sad_x4_4x8_neon)
++#define x265_pixel_sad_x4_8x16_neon x265_template(pixel_sad_x4_8x16_neon)
++#define x265_pixel_sad_x4_8x4_neon x265_template(pixel_sad_x4_8x4_neon)
++#define x265_pixel_sad_x4_8x8_neon x265_template(pixel_sad_x4_8x8_neon)
++#define x265_pixel_satd_16x16_neon x265_template(pixel_satd_16x16_neon)
++#define x265_pixel_satd_16x8_neon x265_template(pixel_satd_16x8_neon)
++#define x265_pixel_satd_4x16_neon x265_template(pixel_satd_4x16_neon)
++#define x265_pixel_satd_4x4_neon x265_template(pixel_satd_4x4_neon)
++#define x265_pixel_satd_4x8_neon x265_template(pixel_satd_4x8_neon)
++#define x265_pixel_satd_8x16_neon x265_template(pixel_satd_8x16_neon)
++#define x265_pixel_satd_8x4_neon x265_template(pixel_satd_8x4_neon)
++#define x265_pixel_satd_8x8_neon x265_template(pixel_satd_8x8_neon)
++#define x265_pixel_ssd_16x16_neon x265_template(pixel_ssd_16x16_neon)
++#define x265_pixel_ssd_16x8_neon x265_template(pixel_ssd_16x8_neon)
++#define x265_pixel_ssd_4x16_neon x265_template(pixel_ssd_4x16_neon)
++#define x265_pixel_ssd_4x4_neon x265_template(pixel_ssd_4x4_neon)
++#define x265_pixel_ssd_4x8_neon x265_template(pixel_ssd_4x8_neon)
++#define x265_pixel_ssd_8x16_neon x265_template(pixel_ssd_8x16_neon)
++#define x265_pixel_ssd_8x4_neon x265_template(pixel_ssd_8x4_neon)
++#define x265_pixel_ssd_8x8_neon x265_template(pixel_ssd_8x8_neon)
++#define DECL_PIXELS( ret, name, suffix, args ) \
++ ret x265_pixel_##name##_16x16_##suffix args;\
++ ret x265_pixel_##name##_16x8_##suffix args;\
++ ret x265_pixel_##name##_8x16_##suffix args;\
++ ret x265_pixel_##name##_8x8_##suffix args;\
++ ret x265_pixel_##name##_8x4_##suffix args;\
++ ret x265_pixel_##name##_4x16_##suffix args;\
++ ret x265_pixel_##name##_4x8_##suffix args;\
++ ret x265_pixel_##name##_4x4_##suffix args;\
++
++#define DECL_X1( name, suffix ) \
++ DECL_PIXELS( int, name, suffix, ( uint8_t *, intptr_t, uint8_t *, intptr_t ) )
++
++#define DECL_X4( name, suffix ) \
++ DECL_PIXELS( void, name##_x3, suffix, ( uint8_t *, uint8_t *, uint8_t *, uint8_t *, intptr_t, int * ) )\
++ DECL_PIXELS( void, name##_x4, suffix, ( uint8_t *, uint8_t *, uint8_t *, uint8_t *, uint8_t *, intptr_t, int * ) )
++
++DECL_X1( sad, neon )
++DECL_X4( sad, neon )
++DECL_X1( satd, neon )
++DECL_X1( ssd, neon )
++
++
++#define x265_pixel_ssd_nv12_core_neon x265_template(pixel_ssd_nv12_core_neon)
++void x265_pixel_ssd_nv12_core_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, int, int, uint64_t *, uint64_t * );
++
++#define x265_pixel_vsad_neon x265_template(pixel_vsad_neon)
++int x265_pixel_vsad_neon( uint8_t *, intptr_t, int );
++
++#define x265_pixel_sa8d_8x8_neon x265_template(pixel_sa8d_8x8_neon)
++int x265_pixel_sa8d_8x8_neon ( uint8_t *, intptr_t, uint8_t *, intptr_t );
++#define x265_pixel_sa8d_16x16_neon x265_template(pixel_sa8d_16x16_neon)
++int x265_pixel_sa8d_16x16_neon( uint8_t *, intptr_t, uint8_t *, intptr_t );
++#define x265_pixel_sa8d_satd_16x16_neon x265_template(pixel_sa8d_satd_16x16_neon)
++uint64_t x265_pixel_sa8d_satd_16x16_neon( uint8_t *, intptr_t, uint8_t *, intptr_t );
++
++#define x265_pixel_var_8x8_neon x265_template(pixel_var_8x8_neon)
++uint64_t x265_pixel_var_8x8_neon ( uint8_t *, intptr_t );
++#define x265_pixel_var_8x16_neon x265_template(pixel_var_8x16_neon)
++uint64_t x265_pixel_var_8x16_neon ( uint8_t *, intptr_t );
++#define x265_pixel_var_16x16_neon x265_template(pixel_var_16x16_neon)
++uint64_t x265_pixel_var_16x16_neon( uint8_t *, intptr_t );
++#define x265_pixel_var2_8x8_neon x265_template(pixel_var2_8x8_neon)
++int x265_pixel_var2_8x8_neon ( uint8_t *, uint8_t *, int * );
++#define x265_pixel_var2_8x16_neon x265_template(pixel_var2_8x16_neon)
++int x265_pixel_var2_8x16_neon( uint8_t *, uint8_t *, int * );
++
++#define x265_pixel_hadamard_ac_8x8_neon x265_template(pixel_hadamard_ac_8x8_neon)
++uint64_t x265_pixel_hadamard_ac_8x8_neon ( uint8_t *, intptr_t );
++#define x265_pixel_hadamard_ac_8x16_neon x265_template(pixel_hadamard_ac_8x16_neon)
++uint64_t x265_pixel_hadamard_ac_8x16_neon ( uint8_t *, intptr_t );
++#define x265_pixel_hadamard_ac_16x8_neon x265_template(pixel_hadamard_ac_16x8_neon)
++uint64_t x265_pixel_hadamard_ac_16x8_neon ( uint8_t *, intptr_t );
++#define x265_pixel_hadamard_ac_16x16_neon x265_template(pixel_hadamard_ac_16x16_neon)
++uint64_t x265_pixel_hadamard_ac_16x16_neon( uint8_t *, intptr_t );
++
++#define x265_pixel_ssim_4x4x2_core_neon x265_template(pixel_ssim_4x4x2_core_neon)
++void x265_pixel_ssim_4x4x2_core_neon( const uint8_t *, intptr_t,
++ const uint8_t *, intptr_t,
++ int sums[2][4] );
++#define x265_pixel_ssim_end4_neon x265_template(pixel_ssim_end4_neon)
++float x265_pixel_ssim_end4_neon( int sum0[5][4], int sum1[5][4], int width );
++
++#define x265_pixel_asd8_neon x265_template(pixel_asd8_neon)
++int x265_pixel_asd8_neon( uint8_t *, intptr_t, uint8_t *, intptr_t, int );
++
++#endif
+diff -Naur ./source/common/cpu.cpp ../x265_apple_patch/source/common/cpu.cpp
+--- ./source/common/cpu.cpp 2021-05-08 13:06:22.000000000 +0100
++++ ../x265_apple_patch/source/common/cpu.cpp 2021-05-08 13:08:01.000000000 +0100
+@@ -104,7 +104,8 @@
+ { "ARMv6", X265_CPU_ARMV6 },
+ { "NEON", X265_CPU_NEON },
+ { "FastNeonMRC", X265_CPU_FAST_NEON_MRC },
+-
++#elif X265_ARCH_ARM64
++ { "NEON", X265_CPU_NEON },
+ #elif X265_ARCH_POWER8
+ { "Altivec", X265_CPU_ALTIVEC },
+
+@@ -374,6 +375,18 @@
+ #endif // if HAVE_ARMV6
+ return flags;
+ }
++#elif X265_ARCH_ARM64
++
++uint32_t cpu_detect(bool benableavx512)
++{
++ int flags = 0;
++
++#if HAVE_NEON
++ flags |= X265_CPU_NEON;
++#endif
++
++ return flags;
++}
+
+ #elif X265_ARCH_POWER8
+
+diff -Naur ./source/common/pixel.cpp ../x265_apple_patch/source/common/pixel.cpp
+--- ./source/common/pixel.cpp 2021-05-08 13:06:22.000000000 +0100
++++ ../x265_apple_patch/source/common/pixel.cpp 2021-05-08 13:08:01.000000000 +0100
+@@ -266,7 +266,7 @@
+ {
+ int satd = 0;
+
+-#if ENABLE_ASSEMBLY && X265_ARCH_ARM64
++#if ENABLE_ASSEMBLY && X265_ARCH_ARM64 && 0
+ pixelcmp_t satd_4x4 = x265_pixel_satd_4x4_neon;
+ #endif
+
+@@ -284,7 +284,7 @@
+ {
+ int satd = 0;
+
+-#if ENABLE_ASSEMBLY && X265_ARCH_ARM64
++#if ENABLE_ASSEMBLY && X265_ARCH_ARM64 && 0
+ pixelcmp_t satd_8x4 = x265_pixel_satd_8x4_neon;
+ #endif
+
+diff -Naur ./source/common/version.cpp ../x265_apple_patch/source/common/version.cpp
+--- ./source/common/version.cpp 2021-05-08 13:06:22.000000000 +0100
++++ ../x265_apple_patch/source/common/version.cpp 2021-05-08 13:47:38.000000000 +0100
+@@ -31,7 +31,7 @@
+
+ #if defined(__clang__)
+ #define COMPILEDBY "[clang " XSTR(__clang_major__) "." XSTR(__clang_minor__) "." XSTR(__clang_patchlevel__) "]"
+-#ifdef __IA64__
++#ifdef __IA64__ || __arm64__ || __aarch64__
+ #define ONARCH "[on 64-bit] "
+ #else
+ #define ONARCH "[on 32-bit] "
+@@ -71,7 +71,7 @@
+ #define ONOS "[Unk-OS]"
+ #endif
+
+-#if X86_64
++#if X86_64 || __arm64__ || __aarch64__
+ #define BITS "[64 bit]"
+ #else
+ #define BITS "[32 bit]"
+diff -Naur ./source/test/testharness.h ../x265_apple_patch/source/test/testharness.h
+--- ./source/test/testharness.h 2021-05-08 13:06:22.000000000 +0100
++++ ../x265_apple_patch/source/test/testharness.h 2021-05-08 13:08:01.000000000 +0100
+@@ -64,7 +64,6 @@
+
+ uint64_t m_rand;
+ };
+-
+ #ifdef _MSC_VER
+ #include <intrin.h>
+ #elif HAVE_RDTSC
+@@ -73,7 +72,7 @@
+ #include <x86intrin.h>
+ #elif ( !defined(__APPLE__) && defined (__GNUC__) && defined(__ARM_NEON__))
+ #include <arm_neon.h>
+-#elif defined(__GNUC__) && (!defined(__clang__) || __clang_major__ < 4)
++#else
+ /* fallback for older GCC/MinGW */
+ static inline uint32_t __rdtsc(void)
+ {
+@@ -90,6 +89,12 @@
+
+ // TO-DO: replace clock() function with appropriate ARM cpu instructions
+ a = clock();
++#elif X265_ARCH_ARM64
++ // 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;
+@@ -140,7 +145,7 @@
+ * needs an explicit asm check because it only sometimes crashes in normal use. */
+ intptr_t PFX(checkasm_call)(intptr_t (*func)(), int *ok, ...);
+ float PFX(checkasm_call_float)(float (*func)(), int *ok, ...);
+-#elif X265_ARCH_ARM == 0
++#elif (X265_ARCH_ARM == 0 && X265_ARCH_ARM64 == 0)
+ #define PFX(stack_pagealign)(func, align) func()
+ #endif
+
+diff -Naur ./source/test/testharness.h.orig ../x265_apple_patch/source/test/testharness.h.orig
+--- ./source/test/testharness.h.orig 1970-01-01 01:00:00.000000000 +0100
++++ ../x265_apple_patch/source/test/testharness.h.orig 2021-05-08 13:08:01.000000000 +0100
+@@ -0,0 +1,184 @@
++/*****************************************************************************
++ * Copyright (C) 2013-2020 MulticoreWare, Inc
++ *
++ * Authors: Steve Borho <steve@borho.org>
++ * Min Chen <chenm003@163.com>
++ * Yimeng Su <yimeng.su@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
++ * the Free Software Foundation; either version 2 of the License, or
++ * (at your option) any later version.
++ *
++ * This program is distributed in the hope that it will be useful,
++ * but WITHOUT ANY WARRANTY; without even the implied warranty of
++ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
++ * GNU General Public License for more details.
++ *
++ * You should have received a copy of the GNU General Public License
++ * along with this program; if not, write to the Free Software
++ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
++ *
++ * This program is also available under a commercial proprietary license.
++ * For more information, contact us at license @ x265.com.
++ *****************************************************************************/
++
++#ifndef _TESTHARNESS_H_
++#define _TESTHARNESS_H_ 1
++
++#include "common.h"
++#include "primitives.h"
++
++#if _MSC_VER
++#pragma warning(disable: 4324) // structure was padded due to __declspec(align())
++#endif
++
++#define PIXEL_MIN 0
++#define SHORT_MAX 32767
++#define SHORT_MIN -32767
++#define UNSIGNED_SHORT_MAX 65535
++
++using namespace X265_NS;
++
++extern const char* lumaPartStr[NUM_PU_SIZES];
++extern const char* const* chromaPartStr[X265_CSP_COUNT];
++
++class TestHarness
++{
++public:
++
++ TestHarness() {}
++
++ virtual ~TestHarness() {}
++
++ virtual bool testCorrectness(const EncoderPrimitives& ref, const EncoderPrimitives& opt) = 0;
++
++ virtual void measureSpeed(const EncoderPrimitives& ref, const EncoderPrimitives& opt) = 0;
++
++ virtual const char *getName() const = 0;
++
++protected:
++
++ /* Temporary variables for stack checks */
++ int m_ok;
++
++ uint64_t m_rand;
++};
++
++#ifdef _MSC_VER
++#include <intrin.h>
++#elif HAVE_RDTSC
++#include <intrin.h>
++#elif (!defined(__APPLE__) && (defined (__GNUC__) && (defined(__x86_64__) || defined(__i386__))))
++#include <x86intrin.h>
++#elif ( !defined(__APPLE__) && defined (__GNUC__) && defined(__ARM_NEON__))
++#include <arm_neon.h>
++#elif defined(__GNUC__) && (!defined(__clang__) || __clang_major__ < 4)
++/* fallback for older GCC/MinGW */
++static inline uint32_t __rdtsc(void)
++{
++ uint32_t a = 0;
++
++#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;
++}
++#endif // ifdef _MSC_VER
++
++#define BENCH_RUNS 2000
++
++/* Adapted from checkasm.c, runs each optimized primitive four times, measures rdtsc
++ * and discards invalid times. Repeats BENCH_RUNS times to get a good average.
++ * Then measures the C reference with BENCH_RUNS / 4 runs and reports X factor and average cycles.*/
++#define REPORT_SPEEDUP(RUNOPT, RUNREF, ...) \
++ { \
++ uint32_t cycles = 0; int runs = 0; \
++ RUNOPT(__VA_ARGS__); \
++ for (int ti = 0; ti < BENCH_RUNS; ti++) { \
++ uint32_t t0 = (uint32_t)__rdtsc(); \
++ RUNOPT(__VA_ARGS__); \
++ RUNOPT(__VA_ARGS__); \
++ RUNOPT(__VA_ARGS__); \
++ RUNOPT(__VA_ARGS__); \
++ uint32_t t1 = (uint32_t)__rdtsc() - t0; \
++ if (t1 * runs <= cycles * 4 && ti > 0) { cycles += t1; runs++; } \
++ } \
++ uint32_t refcycles = 0; int refruns = 0; \
++ RUNREF(__VA_ARGS__); \
++ for (int ti = 0; ti < BENCH_RUNS / 4; ti++) { \
++ uint32_t t0 = (uint32_t)__rdtsc(); \
++ RUNREF(__VA_ARGS__); \
++ RUNREF(__VA_ARGS__); \
++ RUNREF(__VA_ARGS__); \
++ RUNREF(__VA_ARGS__); \
++ uint32_t t1 = (uint32_t)__rdtsc() - t0; \
++ if (t1 * refruns <= refcycles * 4 && ti > 0) { refcycles += t1; refruns++; } \
++ } \
++ x265_emms(); \
++ float optperf = (10.0f * cycles / runs) / 4; \
++ float refperf = (10.0f * refcycles / refruns) / 4; \
++ printf("\t%3.2fx ", refperf / optperf); \
++ printf("\t %-8.2lf \t %-8.2lf\n", optperf, refperf); \
++ }
++
++extern "C" {
++#if X265_ARCH_X86
++int PFX(stack_pagealign)(int (*func)(), int align);
++
++/* detect when callee-saved regs aren't saved
++ * needs an explicit asm check because it only sometimes crashes in normal use. */
++intptr_t PFX(checkasm_call)(intptr_t (*func)(), int *ok, ...);
++float PFX(checkasm_call_float)(float (*func)(), int *ok, ...);
++#elif X265_ARCH_ARM == 0
++#define PFX(stack_pagealign)(func, align) func()
++#endif
++
++#if X86_64
++
++/* Evil hack: detect incorrect assumptions that 32-bit ints are zero-extended to 64-bit.
++ * This is done by clobbering the stack with junk around the stack pointer and calling the
++ * assembly function through x265_checkasm_call with added dummy arguments which forces all
++ * real arguments to be passed on the stack and not in registers. For 32-bit argument the
++ * upper half of the 64-bit register location on the stack will now contain junk. Note that
++ * this is dependent on compiler behavior and that interrupts etc. at the wrong time may
++ * overwrite the junk written to the stack so there's no guarantee that it will always
++ * detect all functions that assumes zero-extension.
++ */
++void PFX(checkasm_stack_clobber)(uint64_t clobber, ...);
++#define checked(func, ...) ( \
++ m_ok = 1, m_rand = (rand() & 0xffff) * 0x0001000100010001ULL, \
++ PFX(checkasm_stack_clobber)(m_rand, m_rand, m_rand, m_rand, m_rand, m_rand, m_rand, m_rand, \
++ m_rand, m_rand, m_rand, m_rand, m_rand, m_rand, m_rand, m_rand, \
++ m_rand, m_rand, m_rand, m_rand, m_rand), /* max_args+6 */ \
++ PFX(checkasm_call)((intptr_t(*)())func, &m_ok, 0, 0, 0, 0, __VA_ARGS__))
++
++#define checked_float(func, ...) ( \
++ m_ok = 1, m_rand = (rand() & 0xffff) * 0x0001000100010001ULL, \
++ PFX(checkasm_stack_clobber)(m_rand, m_rand, m_rand, m_rand, m_rand, m_rand, m_rand, m_rand, \
++ m_rand, m_rand, m_rand, m_rand, m_rand, m_rand, m_rand, m_rand, \
++ m_rand, m_rand, m_rand, m_rand, m_rand), /* max_args+6 */ \
++ PFX(checkasm_call_float)((float(*)())func, &m_ok, 0, 0, 0, 0, __VA_ARGS__))
++#define reportfail() if (!m_ok) { fflush(stdout); fprintf(stderr, "stack clobber check failed at %s:%d", __FILE__, __LINE__); abort(); }
++#elif ARCH_X86
++#define checked(func, ...) PFX(checkasm_call)((intptr_t(*)())func, &m_ok, __VA_ARGS__);
++#define checked_float(func, ...) PFX(checkasm_call_float)((float(*)())func, &m_ok, __VA_ARGS__);
++
++#else // if X86_64
++#define checked(func, ...) func(__VA_ARGS__)
++#define checked_float(func, ...) func(__VA_ARGS__)
++#define reportfail()
++#endif // if X86_64
++}
++
++#endif // ifndef _TESTHARNESS_H_
projects / goodguy / cinelerra.git / commitdiff