From 5c8ce17c7ea43608a2c5505b58db009efb0a4421 Mon Sep 17 00:00:00 2001 From: Mike Solar Date: Tue, 14 Jul 2026 15:14:29 +0800 Subject: [PATCH] submodule: change core into nomarl folder, and move KDockWidgets into third_party. --- .gitmodules | 8 +- CMakeLists.txt | 14 +- core/.gitignore | 2 + core/CMakeLists.txt | 100 + core/README.md | 3 + core/cmake/FindFFMPEG.cmake | 197 + core/cmake/FindOpenTimelineIO.cmake | 133 + core/include/olive/core/core.h | 38 + core/include/olive/core/render/audioparams.h | 315 + core/include/olive/core/render/pixelformat.h | 132 + core/include/olive/core/render/samplebuffer.h | 137 + core/include/olive/core/render/sampleformat.h | 288 + core/include/olive/core/util/bezier.h | 138 + core/include/olive/core/util/color.h | 184 + core/include/olive/core/util/cpuoptimize.h | 30 + core/include/olive/core/util/log.h | 68 + core/include/olive/core/util/math.h | 30 + core/include/olive/core/util/rational.h | 168 + core/include/olive/core/util/sse2neon.h | 8736 +++++++++++++++++ core/include/olive/core/util/stringutils.h | 211 + core/include/olive/core/util/tests.h | 62 + .../olive/core/util/timecodefunctions.h | 91 + core/include/olive/core/util/timerange.h | 315 + core/include/olive/core/util/value.h | 96 + core/src/render/audioparams.cpp | 237 + core/src/render/samplebuffer.cpp | 323 + core/src/util/bezier.cpp | 115 + core/src/util/color.cpp | 324 + core/src/util/rational.cpp | 286 + core/src/util/stringutils.cpp | 107 + core/src/util/tests.cpp | 63 + core/src/util/timecodefunctions.cpp | 407 + core/src/util/timerange.cpp | 398 + core/src/util/value.cpp | 27 + core/tests/rational-test.cpp | 109 + core/tests/stringutils-test.cpp | 47 + core/tests/timecode-test.cpp | 65 + core/tests/timerange-test.cpp | 124 + ext/CMakeLists.txt | 24 - ext/core | 1 - {ext => third_party}/KDDockWidgets | 0 41 files changed, 14119 insertions(+), 34 deletions(-) create mode 100644 core/.gitignore create mode 100644 core/CMakeLists.txt create mode 100644 core/README.md create mode 100755 core/cmake/FindFFMPEG.cmake create mode 100644 core/cmake/FindOpenTimelineIO.cmake create mode 100644 core/include/olive/core/core.h create mode 100644 core/include/olive/core/render/audioparams.h create mode 100644 core/include/olive/core/render/pixelformat.h create mode 100644 core/include/olive/core/render/samplebuffer.h create mode 100644 core/include/olive/core/render/sampleformat.h create mode 100644 core/include/olive/core/util/bezier.h create mode 100644 core/include/olive/core/util/color.h create mode 100644 core/include/olive/core/util/cpuoptimize.h create mode 100644 core/include/olive/core/util/log.h create mode 100644 core/include/olive/core/util/math.h create mode 100644 core/include/olive/core/util/rational.h create mode 100644 core/include/olive/core/util/sse2neon.h create mode 100644 core/include/olive/core/util/stringutils.h create mode 100644 core/include/olive/core/util/tests.h create mode 100644 core/include/olive/core/util/timecodefunctions.h create mode 100644 core/include/olive/core/util/timerange.h create mode 100644 core/include/olive/core/util/value.h create mode 100644 core/src/render/audioparams.cpp create mode 100644 core/src/render/samplebuffer.cpp create mode 100644 core/src/util/bezier.cpp create mode 100644 core/src/util/color.cpp create mode 100644 core/src/util/rational.cpp create mode 100644 core/src/util/stringutils.cpp create mode 100644 core/src/util/tests.cpp create mode 100644 core/src/util/timecodefunctions.cpp create mode 100644 core/src/util/timerange.cpp create mode 100644 core/src/util/value.cpp create mode 100644 core/tests/rational-test.cpp create mode 100644 core/tests/stringutils-test.cpp create mode 100644 core/tests/timecode-test.cpp create mode 100644 core/tests/timerange-test.cpp delete mode 100644 ext/CMakeLists.txt delete mode 160000 ext/core rename {ext => third_party}/KDDockWidgets (100%) diff --git a/.gitmodules b/.gitmodules index 7e19cdfab..2a6ada713 100644 --- a/.gitmodules +++ b/.gitmodules @@ -1,8 +1,4 @@ -[submodule "ext/core"] - path = ext/core - url = https://github.com/OliveCommunity/core.git - branch = dev -[submodule "ext/KDDockWidgets"] - path = ext/KDDockWidgets +[submodule "third_party/KDDockWidgets"] + path = third_party/KDDockWidgets url = https://github.com/OliveCommunity/KDDockWidgets.git branch = main diff --git a/CMakeLists.txt b/CMakeLists.txt index 647fbcb62..994733995 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -133,7 +133,7 @@ list(APPEND OLIVE_INCLUDE_DIRS ${OPENEXR_INCLUDES}) # Link Olive list(APPEND OLIVE_LIBRARIES olivecore) -list(APPEND OLIVE_INCLUDE_DIRS ${CMAKE_CURRENT_SOURCE_DIR}/ext/core/include) +list(APPEND OLIVE_INCLUDE_DIRS ${CMAKE_CURRENT_SOURCE_DIR}/core/include) # Link Qt @@ -327,9 +327,17 @@ if(BUILD_DOXYGEN) endif() set(CMAKE_INCLUDE_CURRENT_DIR ON) -list(APPEND OLIVE_INCLUDE_DIRS ${CMAKE_SOURCE_DIR}/ext) +list(APPEND OLIVE_INCLUDE_DIRS ${CMAKE_SOURCE_DIR}/third_party) + +add_subdirectory(core EXCLUDE_FROM_ALL) + +set(KDDockWidgets_STATIC ON CACHE INTERNAL "Force KDDockWidgets to build statically") +set(KDDockWidgets_QT6 ${BUILD_QT6} CACHE INTERNAL "Conform KDDockWidgets' Qt 6 setting to ours") +# Oak only uses the QtWidgets frontend; building the QtQuick frontend causes +# duplicate QML module registration on macOS and pulls in unused dependencies. +set(KDDockWidgets_FRONTENDS "qtwidgets" CACHE INTERNAL "Only build the QtWidgets frontend for Oak") +add_subdirectory(third_party/KDDockWidgets EXCLUDE_FROM_ALL) -add_subdirectory(ext) add_subdirectory(third_party/openfx/HostSupport) add_subdirectory(app) diff --git a/core/.gitignore b/core/.gitignore new file mode 100644 index 000000000..da1d64b79 --- /dev/null +++ b/core/.gitignore @@ -0,0 +1,2 @@ +*.user +build/ diff --git a/core/CMakeLists.txt b/core/CMakeLists.txt new file mode 100644 index 000000000..87b9a6d63 --- /dev/null +++ b/core/CMakeLists.txt @@ -0,0 +1,100 @@ +# libolivecore +# Copyright (C) 2023 Olive Studios LLC +# Modifications Copyright (C) 2025 mikesolar +# +# This program is free software: you can redistribute it and/or modify +# it under the terms of the GNU General Public License as published by +# the Free Software Foundation, either version 3 of the License, or +# (at your option) any later version. +# +# This program is distributed in the hope that it will be useful, +# but WITHOUT ANY WARRANTY; without even the implied warranty of +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +# GNU General Public License for more details. +# +# You should have received a copy of the GNU General Public License +# along with this program. If not, see . + +cmake_minimum_required(VERSION 3.13 FATAL_ERROR) + +project(libolivecore VERSION 1.0.0 LANGUAGES CXX) + +option(OLIVECORE_BUILD_TESTS ON) + +set(CMAKE_CXX_STANDARD 17) +set(CMAKE_CXX_STANDARD_REQUIRED ON) +set(CMAKE_CXX_EXTENSIONS OFF) + +list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake") + +# Link avutil +find_package(FFMPEG 6.0 REQUIRED + COMPONENTS avutil +) + +# Link Imath +find_package(Imath REQUIRED CONFIG) + +# Link OpenGL +if(UNIX AND NOT APPLE AND NOT DEFINED OpenGL_GL_PREFERENCE) + set(OpenGL_GL_PREFERENCE LEGACY) +endif() +find_package(OpenGL REQUIRED) + +add_library(olivecore + src/render/audioparams.cpp + src/render/samplebuffer.cpp + src/util/bezier.cpp + src/util/color.cpp + src/util/rational.cpp + src/util/stringutils.cpp + src/util/tests.cpp + src/util/timecodefunctions.cpp + src/util/timerange.cpp + src/util/value.cpp +) + +target_include_directories(olivecore PRIVATE + ${FFMPEG_INCLUDE_DIRS} + "${CMAKE_CURRENT_SOURCE_DIR}/include/olive/core" + "${CMAKE_SOURCE_DIR}/third_party/openfx/include/" +) + +target_link_libraries(olivecore PRIVATE + OpenGL::GL + Imath::Imath + FFMPEG::avutil +) + +# Link OpenTimelineIO (optional) +find_package(OpenTimelineIO) +if (OpenTimelineIO_FOUND) + target_compile_definitions(olivecore PRIVATE USE_OTIO) + target_include_directories(olivecore PRIVATE ${OTIO_INCLUDE_DIRS}) + target_link_libraries(olivecore PRIVATE ${OTIO_LIBRARIES}) +else() + message(" OpenTimelineIO interchange will be disabled.") +endif() + +install(TARGETS olivecore) +install(DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/include/olive" DESTINATION "${CMAKE_INSTALL_PREFIX}/include") + +if (OLIVECORE_BUILD_TESTS) + enable_testing() + + function(make_test name) + add_executable(${name} + tests/${name}.cpp + ) + target_link_libraries(${name} PRIVATE olivecore) + target_include_directories(${name} PRIVATE + "${CMAKE_CURRENT_SOURCE_DIR}/include/olive/core" + ) + add_test(${name} ${name}) + endfunction() + + make_test(rational-test) + make_test(stringutils-test) + make_test(timecode-test) + make_test(timerange-test) +endif() diff --git a/core/README.md b/core/README.md new file mode 100644 index 000000000..eff359734 --- /dev/null +++ b/core/README.md @@ -0,0 +1,3 @@ +# libolivecore + +GPU core library used for various video-related applications. Currently supports OpenGL, will soon support Vulkan. diff --git a/core/cmake/FindFFMPEG.cmake b/core/cmake/FindFFMPEG.cmake new file mode 100755 index 000000000..3d36bb09a --- /dev/null +++ b/core/cmake/FindFFMPEG.cmake @@ -0,0 +1,197 @@ +#[==[ +Provides the following variables: + + * `FFMPEG_INCLUDE_DIRS`: Include directories necessary to use FFMPEG. + * `FFMPEG_LIBRARIES`: Libraries necessary to use FFMPEG. Note that this only + includes libraries for the components requested. + * `FFMPEG_VERSION`: The version of FFMPEG found. + +The following components are supported: + + * `avcodec` + * `avdevice` + * `avfilter` + * `avformat` + * `avresample` + * `avutil` + * `swresample` + * `swscale` + +For each component, the following are provided: + + * `FFMPEG__FOUND`: Libraries for the component. + * `FFMPEG__INCLUDE_DIRS`: Include directories for + the component. + * `FFMPEG__LIBRARIES`: Libraries for the component. + * `FFMPEG::`: A target to use with `target_link_libraries`. + +Note that only components requested with `COMPONENTS` or `OPTIONAL_COMPONENTS` +are guaranteed to set these variables or provide targets. +#]==] + +function (_ffmpeg_find component headername) + if (${FFMPEG_${component}_FOUND}) + return() + endif() + + find_path("FFMPEG_${component}_INCLUDE_DIR" + NAMES + "lib${component}/${headername}" + PATHS + "${FFMPEG_ROOT}/include" + ~/Library/Frameworks + /Library/Frameworks + /usr/local/include + /usr/include + /sw/include # Fink + /opt/local/include # DarwinPorts + /opt/csw/include # Blastwave + /opt/include + /usr/freeware/include + PATH_SUFFIXES + ffmpeg + DOC "FFMPEG's ${component} include directory") + mark_as_advanced("FFMPEG_${component}_INCLUDE_DIR") + + # On Windows, static FFMPEG is sometimes built as `lib.a`. + if (WIN32) + list(APPEND CMAKE_FIND_LIBRARY_SUFFIXES ".a" ".lib") + list(APPEND CMAKE_FIND_LIBRARY_PREFIXES "" "lib") + endif () + + find_library("FFMPEG_${component}_LIBRARY" + NAMES + "${component}" + PATHS + "${FFMPEG_ROOT}/lib" + ~/Library/Frameworks + /Library/Frameworks + /usr/local/lib + /usr/local/lib64 + /usr/lib + /usr/lib64 + /sw/lib + /opt/local/lib + /opt/csw/lib + /opt/lib + /usr/freeware/lib64 + "${FFMPEG_ROOT}/bin" + DOC "FFMPEG's ${component} library") + mark_as_advanced("FFMPEG_${component}_LIBRARY") + + if (FFMPEG_${component}_LIBRARY AND FFMPEG_${component}_INCLUDE_DIR) + set(_deps_found TRUE) + set(_deps_link) + foreach (_ffmpeg_dep IN LISTS ARGN) + if (TARGET "FFMPEG::${_ffmpeg_dep}") + list(APPEND _deps_link "FFMPEG::${_ffmpeg_dep}") + else () + set(_deps_found FALSE) + endif () + endforeach () + if (_deps_found) + add_library("FFMPEG::${component}" UNKNOWN IMPORTED) + set_target_properties("FFMPEG::${component}" PROPERTIES + IMPORTED_LOCATION "${FFMPEG_${component}_LIBRARY}" + INTERFACE_INCLUDE_DIRECTORIES "${FFMPEG_${component}_INCLUDE_DIR}" + IMPORTED_LINK_INTERFACE_LIBRARIES "${_deps_link}") + set("FFMPEG_${component}_FOUND" 1 + PARENT_SCOPE) + + set(version_header_path "${FFMPEG_${component}_INCLUDE_DIR}/lib${component}/version.h") + if (EXISTS "${version_header_path}") + string(TOUPPER "${component}" component_upper) + file(STRINGS "${version_header_path}" version + REGEX "#define *LIB${component_upper}_VERSION_(MAJOR|MINOR|MICRO) ") + string(REGEX REPLACE ".*_MAJOR *\([0-9]*\).*" "\\1" major "${version}") + string(REGEX REPLACE ".*_MINOR *\([0-9]*\).*" "\\1" minor "${version}") + string(REGEX REPLACE ".*_MICRO *\([0-9]*\).*" "\\1" micro "${version}") + if (NOT major STREQUAL "" AND + NOT minor STREQUAL "" AND + NOT micro STREQUAL "") + set("FFMPEG_${component}_VERSION" "${major}.${minor}.${micro}" + PARENT_SCOPE) + endif () + endif () + else () + set("FFMPEG_${component}_FOUND" 0 + PARENT_SCOPE) + set(what) + if (NOT FFMPEG_${component}_LIBRARY) + set(what "library") + endif () + if (NOT FFMPEG_${component}_INCLUDE_DIR) + if (what) + string(APPEND what " or headers") + else () + set(what "headers") + endif () + endif () + set("FFMPEG_${component}_NOT_FOUND_MESSAGE" + "Could not find the ${what} for ${component}." + PARENT_SCOPE) + endif () + endif () +endfunction () + +_ffmpeg_find(avutil avutil.h) +_ffmpeg_find(avresample avresample.h + avutil) +_ffmpeg_find(swresample swresample.h + avutil) +_ffmpeg_find(swscale swscale.h + avutil) +_ffmpeg_find(avcodec avcodec.h + avutil) +_ffmpeg_find(avformat avformat.h + avcodec avutil) +_ffmpeg_find(avfilter avfilter.h + avutil) +_ffmpeg_find(avdevice avdevice.h + avformat avutil) + +if (TARGET FFMPEG::avutil) + set(_ffmpeg_version_header_path "${FFMPEG_avutil_INCLUDE_DIR}/libavutil/ffversion.h") + if (EXISTS "${_ffmpeg_version_header_path}") + file(STRINGS "${_ffmpeg_version_header_path}" _ffmpeg_version + REGEX "FFMPEG_VERSION") + string(REGEX REPLACE ".*\"n?\(.*\)\"" "\\1" FFMPEG_VERSION "${_ffmpeg_version}") + unset(_ffmpeg_version) + else () + set(FFMPEG_VERSION FFMPEG_VERSION-NOTFOUND) + endif () + unset(_ffmpeg_version_header_path) +endif () + +set(FFMPEG_INCLUDE_DIRS) +set(FFMPEG_LIBRARIES) +set(_ffmpeg_required_vars) +foreach (_ffmpeg_component IN LISTS FFMPEG_FIND_COMPONENTS) + if (TARGET "FFMPEG::${_ffmpeg_component}") + set(FFMPEG_${_ffmpeg_component}_INCLUDE_DIRS + "${FFMPEG_${_ffmpeg_component}_INCLUDE_DIR}") + set(FFMPEG_${_ffmpeg_component}_LIBRARIES + "${FFMPEG_${_ffmpeg_component}_LIBRARY}") + list(APPEND FFMPEG_INCLUDE_DIRS + "${FFMPEG_${_ffmpeg_component}_INCLUDE_DIRS}") + list(APPEND FFMPEG_LIBRARIES + "${FFMPEG_${_ffmpeg_component}_LIBRARIES}") + if (FFMEG_FIND_REQUIRED_${_ffmpeg_component}) + list(APPEND _ffmpeg_required_vars + "FFMPEG_${_ffmpeg_required_vars}_INCLUDE_DIRS" + "FFMPEG_${_ffmpeg_required_vars}_LIBRARIES") + endif () + endif () +endforeach () +unset(_ffmpeg_component) + +if (FFMPEG_INCLUDE_DIRS) + list(REMOVE_DUPLICATES FFMPEG_INCLUDE_DIRS) +endif () + +include(FindPackageHandleStandardArgs) +find_package_handle_standard_args(FFMPEG + REQUIRED_VARS FFMPEG_INCLUDE_DIRS FFMPEG_LIBRARIES ${_ffmpeg_required_vars} + VERSION_VAR FFMPEG_VERSION + HANDLE_COMPONENTS) +unset(_ffmpeg_required_vars) diff --git a/core/cmake/FindOpenTimelineIO.cmake b/core/cmake/FindOpenTimelineIO.cmake new file mode 100644 index 000000000..38563b6f4 --- /dev/null +++ b/core/cmake/FindOpenTimelineIO.cmake @@ -0,0 +1,133 @@ +# Olive - Non-Linear Video Editor +# Copyright (C) 2023 Olive Studios LLC +# +# This program is free software: you can redistribute it and/or modify +# it under the terms of the GNU General Public License as published by +# the Free Software Foundation, either version 3 of the License, or +# (at your option) any later version. +# +# This program is distributed in the hope that it will be useful, +# but WITHOUT ANY WARRANTY; without even the implied warranty of +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +# GNU General Public License for more details. +# +# You should have received a copy of the GNU General Public License +# along with this program. If not, see . + +if(UNIX) + find_path(OTIO_BASE_DIR + include/opentimelineio/timeline.h + HINTS + "${OTIO_LOCATION}" + "$ENV{OTIO_LOCATION}" + "/opt/otio" + ) + find_path(OTIO_LIBRARY_DIR + libopentimelineio.so + HINTS + "${OTIO_LOCATION}" + "$ENV{OTIO_LOCATION}" + "${OTIO_BASE_DIR}" + PATH_SUFFIXES + lib/ + DOC + "OpenTimelineIO library path" + ) +elseif(WIN32) + find_path(OTIO_BASE_DIR + include/opentimelineio/timeline.h + HINTS + "${OTIO_LOCATION}" + "$ENV{OTIO_LOCATION}" + ) + find_path(OTIO_LIBRARY_DIR + opentimelineio.lib + HINTS + "${OTIO_LOCATION}" + "$ENV{OTIO_LOCATION}" + "${OTIO_BASE_DIR}" + PATH_SUFFIXES + lib/ + DOC + "OpenTimelineIO library path" + ) +endif() + +find_path(OTIO_INCLUDE_DIR + opentimelineio/timeline.h + HINTS + "${OTIO_LOCATION}" + "$ENV{OTIO_LOCATION}" + "${OTIO_BASE_DIR}" + PATH_SUFFIXES + include/ + DOC + "OpenTimelineIO headers path" +) + +list(APPEND OTIO_INCLUDE_DIRS ${OTIO_INCLUDE_DIR}) + +find_path(OTIO_DEPS_INCLUDE_DIR + any/any.hpp + HINTS + "${OTIO_LOCATION}" + "$ENV{OTIO_LOCATION}" + "${OTIO_BASE_DIR}" + PATH_SUFFIXES + include/opentimelineio/deps/ + DOC + "OpenTimelineIO headers path" +) + +list(APPEND OTIO_INCLUDE_DIRS ${OTIO_DEPS_INCLUDE_DIR}) + +find_path(OT_INCLUDE_DIR + opentime/rationalTime.h + HINTS + "${OTIO_LOCATION}" + "$ENV{OTIO_LOCATION}" + "${OTIO_BASE_DIR}" + PATH_SUFFIXES + include/ + DOC + "OpenTime headers path" +) + +list(APPEND OTIO_INCLUDE_DIRS ${OT_INCLUDE_DIR}) + +find_library(OTIO_LIBRARY + opentimelineio + HINTS + "${OTIO_LOCATION}" + "$ENV{OTIO_LOCATION}" + "${OTIO_BASE_DIR}" + PATH_SUFFIXES + lib/ + DOC + "OTIO's ${OTIO_LIB} library path" +) + +list(APPEND OTIO_LIBRARIES ${OTIO_LIBRARY}) + +find_library(OT_LIBRARY + opentime + HINTS + "${OTIO_LOCATION}" + "$ENV{OTIO_LOCATION}" + "${OTIO_BASE_DIR}" + PATH_SUFFIXES + lib/ + DOC + "OpenTime's ${OTIO_LIB} library path" +) + +list(APPEND OTIO_LIBRARIES ${OT_LIBRARY}) + +include(FindPackageHandleStandardArgs) + +find_package_handle_standard_args(OpenTimelineIO + REQUIRED_VARS + OTIO_LIBRARIES + OTIO_INCLUDE_DIRS + OTIO_DEPS_INCLUDE_DIR +) diff --git a/core/include/olive/core/core.h b/core/include/olive/core/core.h new file mode 100644 index 000000000..ee62cf567 --- /dev/null +++ b/core/include/olive/core/core.h @@ -0,0 +1,38 @@ +/* + * Olive Community Edition - Non-Linear Video Editor + * Copyright (C) 2025 Olive CE Team + * + * 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 3 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, see . + */ + +#ifndef LIBOLIVECORE_H +#define LIBOLIVECORE_H + +#include "render/audioparams.h" +#include "render/pixelformat.h" +#include "render/samplebuffer.h" +#include "render/sampleformat.h" +#include "util/bezier.h" +#include "util/color.h" +#include "util/cpuoptimize.h" +#include "util/log.h" +#include "util/math.h" +#include "util/rational.h" +#include "util/stringutils.h" +#include "util/tests.h" +#include "util/timecodefunctions.h" +#include "util/timerange.h" +#include "util/value.h" + +#endif // LIBOLIVECORE_H diff --git a/core/include/olive/core/render/audioparams.h b/core/include/olive/core/render/audioparams.h new file mode 100644 index 000000000..046b335df --- /dev/null +++ b/core/include/olive/core/render/audioparams.h @@ -0,0 +1,315 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#ifndef LIBOLIVECORE_AUDIOPARAMS_H +#define LIBOLIVECORE_AUDIOPARAMS_H +#include + +extern "C" { +#include +} + +#include +#include + +#include "sampleformat.h" +#include "../util/rational.h" + +namespace olive::core +{ + +/** + * @brief Audio parameters class managing audio stream configuration + * + * CRITICAL NOTE: This class manages AVChannelLayout which contains dynamic memory + * (custom channel maps via u.map pointer). Prior to the Rule of Three implementation, + * shallow copies could occur when: + * - AudioParams stored in QVector (QVector reallocations) + * - AudioParams passed by value to RenderVideoParams + * - AudioParams copied during Node graph duplication in ProjectCopier + * + * When shallow copies occurred, one copy's set_channel_layout() could free the + * shared u.map pointer, corrupting other copies. This manifested as: + * - channel_layouts=0x0 errors in AudioProcessor::Open() + * - is_valid() returning false unexpectedly + * + * The Rule of Three (copy ctor, copy assignment, destructor) was added to ensure + * proper deep copies of AVChannelLayout using av_channel_layout_copy(). + */ +class AudioParams { +public: + /** + * @brief Default constructor creates invalid AudioParams + * sample_rate=0, channel_layout empty, format=INVALID + */ + AudioParams() + : sample_rate_(0) + , channel_layout_{} + , channel_count_(0) + , format_(SampleFormat::INVALID) + { + set_default_footage_parameters(); + } + + /** + * @brief Constructor from AVChannelLayout (deep copy) + * @param sample_rate Audio sample rate (e.g., 48000) + * @param channel_layout FFmpeg channel layout (copied via av_channel_layout_copy) + * @param format Sample format (e.g., SampleFormat::F32P) + * + * NOTE: The channel_layout parameter is deep-copied. The original can be + * safely uninit'd after this constructor returns. + */ + AudioParams(const int &sample_rate, const AVChannelLayout &channel_layout, + const SampleFormat &format) + : sample_rate_(sample_rate) + , channel_layout_{} + , channel_count_(0) + , format_(format) + { + set_default_footage_parameters(); + timebase_ = sample_rate_as_time_base(); + av_channel_layout_uninit(&channel_layout_); + av_channel_layout_copy(&channel_layout_, &channel_layout); + + // Cache channel count from the copied layout + calculate_channel_count(); + } + + /** + * @brief Constructor from channel layout mask + * @param sample_rate Audio sample rate + * @param channel_layout Channel layout mask (e.g., AV_CH_LAYOUT_STEREO) + * @param format Sample format + * + * This is the most common constructor used in Olive. The mask is converted + * to AVChannelLayout via av_channel_layout_from_mask(). + */ + AudioParams(const int &sample_rate, uint64_t channel_layout, + const SampleFormat &format) + : sample_rate_(sample_rate) + , channel_layout_{} + , channel_count_(0) + , format_(format) + { + set_default_footage_parameters(); + timebase_ = sample_rate_as_time_base(); + av_channel_layout_uninit(&channel_layout_); + av_channel_layout_from_mask(&channel_layout_, channel_layout); + // Cache channel count + calculate_channel_count(); + } + int sample_rate() const + { + return sample_rate_; + } + + void set_sample_rate(int sample_rate) + { + sample_rate_ = sample_rate; + } + + const AVChannelLayout &channel_layout() const + { + return channel_layout_; + } + + /** + * @brief Set channel layout from AVChannelLayout (deep copy) + * @param channel_layout Source channel layout to copy + * + * CRITICAL: This function first uninitializes the current layout (freeing any + * dynamic memory), then deep-copies the new layout. This is safe only if + * copies are properly managed via Rule of Three. + * + * If called on a shallow-copied AudioParams, this would corrupt other copies + * that share the same u.map pointer. + */ + void set_channel_layout(const AVChannelLayout &channel_layout) + { + av_channel_layout_uninit(&channel_layout_); + av_channel_layout_copy(&channel_layout_, &channel_layout); + calculate_channel_count(); + } + + /** + * @brief Set channel layout from mask + * @param mask Channel layout mask (e.g., AV_CH_LAYOUT_STEREO) + */ + void set_channel_layout(uint64_t mask) + { + av_channel_layout_uninit(&channel_layout_); + av_channel_layout_from_mask(&channel_layout_, mask); + calculate_channel_count(); + } + rational time_base() const + { + return timebase_; + } + + void set_time_base(const rational &timebase) + { + timebase_ = timebase; + } + + rational sample_rate_as_time_base() const + { + return rational(1, sample_rate()); + } + + SampleFormat format() const + { + return format_; + } + + void set_format(SampleFormat format) + { + format_ = format; + } + + bool enabled() const + { + return enabled_; + } + + void set_enabled(bool e) + { + enabled_ = e; + } + + int stream_index() const + { + return stream_index_; + } + + void set_stream_index(int s) + { + stream_index_ = s; + } + + int64_t duration() const + { + return duration_; + } + + void set_duration(int64_t duration) + { + duration_ = duration; + } + + int64_t time_to_bytes(const double &time) const; + int64_t time_to_bytes(const rational &time) const; + int64_t time_to_bytes_per_channel(const double &time) const; + int64_t time_to_bytes_per_channel(const rational &time) const; + int64_t time_to_samples(const double &time) const; + int64_t time_to_samples(const rational &time) const; + int64_t samples_to_bytes(const int64_t &samples) const; + int64_t samples_to_bytes_per_channel(const int64_t &samples) const; + rational samples_to_time(const int64_t &samples) const; + int64_t bytes_to_samples(const int64_t &bytes) const; + rational bytes_to_time(const int64_t &bytes) const; + rational bytes_per_channel_to_time(const int64_t &bytes) const; + int channel_count() const; + int bytes_per_sample_per_channel() const; + int bits_per_sample() const; + bool is_valid() const; + + bool operator==(const AudioParams &other) const; + bool operator!=(const AudioParams &other) const; + + /** + * @name Rule of Three Implementation + * + * These are required because AVChannelLayout (FFmpeg >= 5.0) contains a union + * with a pointer member (u.map for custom channel maps). Without proper + * deep copy management: + * + * 1. Default copy constructor: Shallow copies u.map pointer, leading to + * double-free when original and copy are destroyed + * 2. Default copy assignment: Same issue as copy constructor + * 3. Default destructor: Doesn't free u.map, causing memory leaks + * + * The implementations use av_channel_layout_copy() and av_channel_layout_uninit() + * for proper FFmpeg-managed memory handling. + * + * Context where this matters: + * - QVector in FootageDescription (vector reallocations) + * - RenderVideoParams passing AudioParams by value + * - ProjectCopier duplicating node graphs with audio parameters + */ + ///@{ + AudioParams(const AudioParams &other); + AudioParams &operator=(const AudioParams &other); + ~AudioParams(); + ///@} + + static const std::vector kSupportedChannelLayouts; + static const std::vector kSupportedSampleRates; + +private: + void set_default_footage_parameters() + { + enabled_ = true; + stream_index_ = 0; + duration_ = 0; + } + + /** + * @brief Updates channel_count_ from the current channel_layout_ + * Called after any channel layout modification. + */ + void calculate_channel_count(); + + int sample_rate_; ///< Audio sample rate in Hz (e.g., 48000) + + /** + * @brief FFmpeg channel layout structure + * + * WARNING: This struct contains a union with a pointer member (u.map) when + * using custom channel layouts (order == AV_CHANNEL_ORDER_CUSTOM). The pointer + * must be properly managed via av_channel_layout_copy/uninit. + * + * Layout variants: + * - order == AV_CHANNEL_ORDER_UNSPEC: u.mask is undefined, nb_channels valid + * - order == AV_CHANNEL_ORDER_NATIVE: u.mask contains channel bitmask + * - order == AV_CHANNEL_ORDER_CUSTOM: u.map points to AVChannelCustom array + * + * Corruption symptoms: + * - u.mask == 0 when order should be NATIVE + * - av_channel_layout_check() returns false + * - is_valid() returns false + */ + AVChannelLayout channel_layout_; + + int channel_count_; ///< Cached channel count from layout + + SampleFormat format_; ///< Audio sample format + + // Footage-specific parameters (serialized with footage metadata) + int enabled_; // Using int instead of bool fixes GCC 11 stringop-overflow issue (byte alignment) + int stream_index_; ///< Index in the source file's stream list + int64_t duration_; ///< Stream duration in timebase units + rational timebase_; ///< Timebase for this audio stream +}; + +} + +#endif // LIBOLIVECORE_AUDIOPARAMS_H diff --git a/core/include/olive/core/render/pixelformat.h b/core/include/olive/core/render/pixelformat.h new file mode 100644 index 000000000..52550399a --- /dev/null +++ b/core/include/olive/core/render/pixelformat.h @@ -0,0 +1,132 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#ifndef LIBOLIVECORE_PIXELFORMAT_H +#define LIBOLIVECORE_PIXELFORMAT_H +#include "ofxCore.h" +#include +namespace olive::core +{ + +class PixelFormat { +public: + enum Format { INVALID = -1, U8, U10, U16, F16, F32, COUNT }; + + PixelFormat(Format f = INVALID) + { + f_ = f; + } + + operator Format() const + { + return f_; + } + + static PixelFormat from_ofx(std::string ofxFormat){ + if(ofxFormat == kOfxBitDepthByte){ + return PixelFormat::U8; + } + else if (ofxFormat == kOfxBitDepthShort){ + return PixelFormat::U16; + } + else if(ofxFormat == kOfxBitDepthHalf){ + return PixelFormat::F16; + } + else if(ofxFormat == kOfxBitDepthFloat){ + return PixelFormat::F32; + } + return PixelFormat::INVALID; + } + static int byte_count(Format f) + { + switch (f) { + case INVALID: + case COUNT: + break; + case U8: + return 1; + case U10: + return 4; // packed RGBA10A2, treated as 4 bytes per pixel + case U16: + case F16: + return 2; + case F32: + return 4; + } + + return 0; + } + + const char *to_string() const + { + switch (f_) { + case U8: + return "u8"; + case U10: + return "u10"; + case U16: + return "u16"; + case F16: + return "f16"; + case F32: + return "f32"; + case INVALID: + case COUNT: + break; + } + + return ""; + } + + int byte_count() const + { + return byte_count(f_); + } + + static bool is_float(Format f) + { + switch (f) { + case INVALID: + case COUNT: + case U8: + case U10: + case U16: + break; + case F16: + case F32: + return true; + } + + return false; + } + + bool is_float() const + { + return is_float(f_); + } + +private: + Format f_; +}; + +} + +#endif // LIBOLIVECORE_PIXELFORMAT_H diff --git a/core/include/olive/core/render/samplebuffer.h b/core/include/olive/core/render/samplebuffer.h new file mode 100644 index 000000000..f8ac1a70c --- /dev/null +++ b/core/include/olive/core/render/samplebuffer.h @@ -0,0 +1,137 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#ifndef LIBOLIVECORE_SAMPLEBUFFER_H +#define LIBOLIVECORE_SAMPLEBUFFER_H + +#include +#include + +#include "audioparams.h" +#include "../util/rational.h" + +namespace olive::core +{ + +/** + * @brief A buffer of audio samples + * + * Audio samples in this structure are always stored in PLANAR (separated by channel). This is done to simplify audio + * rendering code. This replaces the old system of using QByteArrays (containing packed audio) and while SampleBuffer + * replaces many of those in the rendering/processing side of things, QByteArrays are currently still in use for + * playback, including reading to and from the cache. + */ +class SampleBuffer { +public: + SampleBuffer(); + SampleBuffer(const AudioParams &audio_params, const rational &length); + SampleBuffer(const AudioParams &audio_params, size_t samples_per_channel); + + SampleBuffer rip_channel(int channel) const; + std::vector rip_channel_vector(int channel) const; + + const AudioParams &audio_params() const; + void set_audio_params(const AudioParams ¶ms); + + const size_t &sample_count() const + { + return sample_count_per_channel_; + } + void set_sample_count(const size_t &sample_count); + void set_sample_count(const rational &length) + { + set_sample_count(audio_params_.time_to_samples(length)); + } + + float *data(int channel) + { + return data_[channel].data(); + } + + const float *data(int channel) const + { + return data_.at(channel).data(); + } + + std::vector to_raw_ptrs() + { + std::vector r(data_.size()); + for (size_t i = 0; i < r.size(); i++) { + r[i] = data_[i].data(); + } + return r; + } + + int channel_count() const + { + return data_.size(); + } + + bool is_allocated() const + { + return !data_.empty(); + } + void allocate(); + void destroy(); + + void reverse(); + void speed(double speed); + + void transform_volume(float f); + void transform_volume_for_channel(int channel, float volume); + static void transform_volume(float f, const SampleBuffer *input, + SampleBuffer *output); + static void transform_volume_for_channel(int channel, float volume, + const SampleBuffer *input, + SampleBuffer *output); + + void transform_volume_for_sample(size_t sample_index, float volume); + void transform_volume_for_sample_on_channel(size_t sample_index, + int channel, float volume); + + void clamp(); + + void silence(); + void silence(size_t start_sample, size_t end_sample); + void silence_bytes(size_t start_byte, size_t end_byte); + + void set(int channel, const float *data, size_t sample_offset, + size_t sample_length); + void set(int channel, const float *data, size_t sample_length) + { + set(channel, data, 0, sample_length); + } + + void fast_set(const SampleBuffer &other, int to, int from = -1); + +private: + void clamp_channel(int channel); + + AudioParams audio_params_; + + size_t sample_count_per_channel_; + + std::vector> data_; +}; + +} + +#endif // LIBOLIVECORE_SAMPLEBUFFER_H diff --git a/core/include/olive/core/render/sampleformat.h b/core/include/olive/core/render/sampleformat.h new file mode 100644 index 000000000..e96d482a6 --- /dev/null +++ b/core/include/olive/core/render/sampleformat.h @@ -0,0 +1,288 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#ifndef LIBOLIVECORE_SAMPLEFORMAT_H +#define LIBOLIVECORE_SAMPLEFORMAT_H + +#include +#include + +namespace olive::core +{ + +class SampleFormat { +public: + enum Format { + INVALID = -1, + + U8P, + S16P, + S32P, + S64P, + F32P, + F64P, + + U8, + S16, + S32, + S64, + F32, + F64, + + COUNT, + + PLANAR_START = U8P, + PACKED_START = U8, + PLANAR_END = PACKED_START, + PACKED_END = COUNT, + }; + + SampleFormat(Format f = INVALID) + { + f_ = f; + } + + operator Format() const + { + return f_; + } + + static int byte_count(Format f) + { + switch (f) { + case U8: + case U8P: + return 1; + case S16: + case S16P: + return 2; + case S32: + case F32: + case S32P: + case F32P: + return 4; + case S64: + case F64: + case S64P: + case F64P: + return 8; + case INVALID: + case COUNT: + break; + } + + return 0; + } + + int byte_count() const + { + return byte_count(f_); + } + + static std::string to_string(Format f) + { + switch (f) { + case INVALID: + case COUNT: + break; + case U8: + return "u8"; + case S16: + return "s16"; + case S32: + return "s32"; + case S64: + return "s64"; + case F32: + return "f32"; + case F64: + return "f64"; + case U8P: + return "u8p"; + case S16P: + return "s16p"; + case S32P: + return "s32p"; + case S64P: + return "s64p"; + case F32P: + return "f32p"; + case F64P: + return "f64p"; + } + + return ""; + } + + std::string to_string() const + { + return to_string(f_); + } + + static SampleFormat from_string(const std::string &s) + { + if (s.empty()) { + return INVALID; + } else if (s == "u8") { + return U8; + } else if (s == "s16") { + return S16; + } else if (s == "s32") { + return S32; + } else if (s == "s64") { + return S64; + } else if (s == "f32") { + return F32; + } else if (s == "f64") { + return F64; + } else if (s == "u8p") { + return U8P; + } else if (s == "s16p") { + return S16P; + } else if (s == "s32p") { + return S32P; + } else if (s == "s64p") { + return S64P; + } else if (s == "f32p") { + return F32P; + } else if (s == "f64p") { + return F64P; + } else { + // Deprecated: sample formats used to be serialized as an integer. Handle that here, but we'll + // probably remove that eventually. + try { + int i = std::stoi(s); + if (i > INVALID && i < COUNT) { + return static_cast(i); + } + } catch (const std::invalid_argument &e) { + } + + // Failed to deserialize from string + return INVALID; + } + } + + static bool is_packed(Format f) + { + return f >= PACKED_START && f < PACKED_END; + } + + bool is_packed() const + { + return is_packed(f_); + } + + static bool is_planar(Format f) + { + return f >= PLANAR_START && f < PLANAR_END; + } + + bool is_planar() const + { + return is_planar(f_); + } + + static SampleFormat to_packed_equivalent(SampleFormat fmt) + { + switch (fmt) { + // For packed input, just return input + case U8: + case S16: + case S32: + case S64: + case F32: + case F64: + return fmt; + + // Convert to packed + case U8P: + return U8; + case S16P: + return S16; + case S32P: + return S32; + case S64P: + return S64; + case F32P: + return F32; + case F64P: + return F64; + + case INVALID: + case COUNT: + break; + } + + return INVALID; + } + + SampleFormat to_packed_equivalent() const + { + return to_packed_equivalent(f_); + } + + static SampleFormat to_planar_equivalent(SampleFormat fmt) + { + switch (fmt) { + // Convert to planar + case U8: + return U8P; + case S16: + return S16P; + case S32: + return S32P; + case S64: + return S64P; + case F32: + return F32P; + case F64: + return F64P; + + // For planar input, just return input + case U8P: + case S16P: + case S32P: + case S64P: + case F32P: + case F64P: + return fmt; + + case INVALID: + case COUNT: + break; + } + + return INVALID; + } + + SampleFormat to_planar_equivalent() const + { + return to_planar_equivalent(f_); + } + +private: + Format f_; +}; + +} + +#endif // LIBOLIVECORE_SAMPLEFORMAT_H diff --git a/core/include/olive/core/util/bezier.h b/core/include/olive/core/util/bezier.h new file mode 100644 index 000000000..c5e140e74 --- /dev/null +++ b/core/include/olive/core/util/bezier.h @@ -0,0 +1,138 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#ifndef LIBOLIVECORE_BEZIER_H +#define LIBOLIVECORE_BEZIER_H + +#include + +namespace olive::core +{ + +class Bezier { +public: + Bezier(); + Bezier(double x, double y); + Bezier(double x, double y, double cp1_x, double cp1_y, double cp2_x, + double cp2_y); + + const double &x() const + { + return x_; + } + const double &y() const + { + return y_; + } + const double &cp1_x() const + { + return cp1_x_; + } + const double &cp1_y() const + { + return cp1_y_; + } + const double &cp2_x() const + { + return cp2_x_; + } + const double &cp2_y() const + { + return cp2_y_; + } + + Imath::V2d to_vec() const + { + return Imath::V2d(x_, y_); + } + + Imath::V2d control_point_1_to_vec() const + { + return Imath::V2d(cp1_x_, cp1_y_); + } + + Imath::V2d control_point_2_to_vec() const + { + return Imath::V2d(cp2_x_, cp2_y_); + } + + void set_x(const double &x) + { + x_ = x; + } + void set_y(const double &y) + { + y_ = y; + } + void set_cp1_x(const double &cp1_x) + { + cp1_x_ = cp1_x; + } + void set_cp1_y(const double &cp1_y) + { + cp1_y_ = cp1_y; + } + void set_cp2_x(const double &cp2_x) + { + cp2_x_ = cp2_x; + } + void set_cp2_y(const double &cp2_y) + { + cp2_y_ = cp2_y; + } + + static double QuadraticXtoT(double x, double a, double b, double c); + + static double QuadraticTtoY(double a, double b, double c, double t); + + static double QuadraticXtoY(double x, const Imath::V2d &a, + const Imath::V2d &b, const Imath::V2d &c) + { + return QuadraticTtoY(a.y, b.y, c.y, QuadraticXtoT(x, a.x, b.x, c.x)); + } + + static double CubicXtoT(double x, double a, double b, double c, double d); + + static double CubicTtoY(double a, double b, double c, double d, double t); + + static double CubicXtoY(double x, const Imath::V2d &a, const Imath::V2d &b, + const Imath::V2d &c, const Imath::V2d &d) + { + return CubicTtoY(a.y, b.y, c.y, d.y, CubicXtoT(x, a.x, b.x, c.x, d.x)); + } + +private: + static double CalculateTFromX(bool cubic, double x, double a, double b, + double c, double d); + + double x_; + double y_; + + double cp1_x_; + double cp1_y_; + + double cp2_x_; + double cp2_y_; +}; + +} + +#endif // LIBOLIVECORE_BEZIER_H diff --git a/core/include/olive/core/util/color.h b/core/include/olive/core/util/color.h new file mode 100644 index 000000000..d5fef8c1b --- /dev/null +++ b/core/include/olive/core/util/color.h @@ -0,0 +1,184 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#ifndef LIBOLIVECORE_COLOR_H +#define LIBOLIVECORE_COLOR_H + +#include "../render/pixelformat.h" + +namespace olive::core +{ + +/** + * @brief High precision 32-bit DataType based RGBA color value + */ +class Color { +public: + using DataType = float; + static constexpr unsigned int RGBA = 4; + + Color() + { + for (unsigned int i = 0; i < RGBA; i++) { + data_[i] = 0.0; + } + } + + Color(const DataType &r, const DataType &g, const DataType &b, + const DataType &a = 1.0f) + { + data_[0] = r; + data_[1] = g; + data_[2] = b; + data_[3] = a; + } + + Color(const char *data, const PixelFormat &format, int ch_layout); + + /** + * @brief Creates a Color struct from hue/saturation/value + * + * Hue expects a value between 0.0 and 360.0. Saturation and Value expect a value between 0.0 and 1.0. + */ + static Color fromHsv(const DataType &h, const DataType &s, + const DataType &v); + + const DataType &red() const + { + return data_[0]; + } + const DataType &green() const + { + return data_[1]; + } + const DataType &blue() const + { + return data_[2]; + } + const DataType &alpha() const + { + return data_[3]; + } + + void toHsv(DataType *hue, DataType *sat, DataType *val) const; + DataType hsv_hue() const; + DataType hsv_saturation() const; + DataType value() const; + + void toHsl(DataType *hue, DataType *sat, DataType *lightness) const; + DataType hsl_hue() const; + DataType hsl_saturation() const; + DataType lightness() const; + + void set_red(const DataType &red) + { + data_[0] = red; + } + void set_green(const DataType &green) + { + data_[1] = green; + } + void set_blue(const DataType &blue) + { + data_[2] = blue; + } + void set_alpha(const DataType &alpha) + { + data_[3] = alpha; + } + + DataType *data() + { + return data_; + } + const DataType *data() const + { + return data_; + } + + void toData(char *out, const PixelFormat &format, + unsigned int nb_channels) const; + + static Color fromData(const char *in, const PixelFormat &format, + unsigned int nb_channels); + + // Suuuuper rough luminance value mostly used for UI (determining whether to overlay with black + // or white text) + DataType GetRoughLuminance() const; + + // Assignment math operators + Color &operator+=(const Color &rhs); + Color &operator-=(const Color &rhs); + Color &operator+=(const DataType &rhs); + Color &operator-=(const DataType &rhs); + Color &operator*=(const DataType &rhs); + Color &operator/=(const DataType &rhs); + + // Binary math operators + Color operator+(const Color &rhs) const + { + Color c(*this); + c += rhs; + return c; + } + + Color operator-(const Color &rhs) const + { + Color c(*this); + c -= rhs; + return c; + } + + Color operator+(const DataType &rhs) const + { + Color c(*this); + c += rhs; + return c; + } + + Color operator-(const DataType &rhs) const + { + Color c(*this); + c -= rhs; + return c; + } + + Color operator*(const DataType &rhs) const + { + Color c(*this); + c *= rhs; + return c; + } + + Color operator/(const DataType &rhs) const + { + Color c(*this); + c /= rhs; + return c; + } + +private: + DataType data_[RGBA]; +}; + +} + +#endif // LIBOLIVECORE_COLOR_H diff --git a/core/include/olive/core/util/cpuoptimize.h b/core/include/olive/core/util/cpuoptimize.h new file mode 100644 index 000000000..9e4234994 --- /dev/null +++ b/core/include/olive/core/util/cpuoptimize.h @@ -0,0 +1,30 @@ +/* + * Olive Community Edition - Non-Linear Video Editor + * Copyright (C) 2025 Olive CE Team + * + * 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 3 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, see . + */ + +#ifndef LIBOLIVECORE_CPUOPTIMIZE_H +#define LIBOLIVECORE_CPUOPTIMIZE_H + +#if defined(__x86_64__) || defined(__i386__) +#define OLIVE_PROCESSOR_X86 +#include +#elif defined(__aarch64__) +#define OLIVE_PROCESSOR_ARM +#include "sse2neon.h" +#endif + +#endif // LIBOLIVECORE_CPUOPTIMIZE_H diff --git a/core/include/olive/core/util/log.h b/core/include/olive/core/util/log.h new file mode 100644 index 000000000..65f71cac0 --- /dev/null +++ b/core/include/olive/core/util/log.h @@ -0,0 +1,68 @@ +/* + * Olive Community Edition - Non-Linear Video Editor + * Copyright (C) 2025 Olive CE Team + * + * 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 3 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, see . + */ + +#ifndef LOG_H +#define LOG_H + +#include + +namespace olive::core +{ + +class Log { +public: + Log(const char *type) + { + std::cerr << "[" << type << "]"; + } + + ~Log() + { + std::cerr << std::endl; + } + + template Log &operator<<(const T &t) + { + std::cerr << " " << t; + return *this; + } + + static Log Debug() + { + return Log("DEBUG"); + } + + static Log Info() + { + return Log("INFO"); + } + + static Log Warning() + { + return Log("WARNING"); + } + + static Log Error() + { + return Log("ERROR"); + } +}; + +} + +#endif // LOG_H diff --git a/core/include/olive/core/util/math.h b/core/include/olive/core/util/math.h new file mode 100644 index 000000000..bb15b122d --- /dev/null +++ b/core/include/olive/core/util/math.h @@ -0,0 +1,30 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#ifndef LIBOLIVECORE_MATH_H +#define LIBOLIVECORE_MATH_H + +namespace olive::core +{ + +} + +#endif // LIBOLIVECORE_MATH_H diff --git a/core/include/olive/core/util/rational.h b/core/include/olive/core/util/rational.h new file mode 100644 index 000000000..a9b8573c7 --- /dev/null +++ b/core/include/olive/core/util/rational.h @@ -0,0 +1,168 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#ifndef LIBOLIVECORE_RATIONAL_H +#define LIBOLIVECORE_RATIONAL_H + +extern "C" { +#include +} + +#include + +#ifdef USE_OTIO +#include +#endif + +namespace olive::core +{ + +class rational { +public: + rational(const int &numerator = 0) + { + r_.num = numerator; + r_.den = 1; + } + + rational(const int &numerator, const int &denominator) + { + r_.num = numerator; + r_.den = denominator; + + fix_signs(); + reduce(); + } + + rational(const rational &rhs) = default; + + rational(const AVRational &r) + { + r_ = r; + + fix_signs(); + } + + static rational fromDouble(const double &flt, bool *ok = nullptr); + static rational fromString(const std::string &str, bool *ok = nullptr); + + static const rational NaN; + + //Assignment Operators + const rational &operator=(const rational &rhs); + const rational &operator+=(const rational &rhs); + const rational &operator-=(const rational &rhs); + const rational &operator/=(const rational &rhs); + const rational &operator*=(const rational &rhs); + + //Binary math operators + rational operator+(const rational &rhs) const; + rational operator-(const rational &rhs) const; + rational operator/(const rational &rhs) const; + rational operator*(const rational &rhs) const; + + //Relational and equality operators + bool operator<(const rational &rhs) const; + bool operator<=(const rational &rhs) const; + bool operator>(const rational &rhs) const; + bool operator>=(const rational &rhs) const; + bool operator==(const rational &rhs) const; + bool operator!=(const rational &rhs) const; + + //Unary operators + const rational &operator+() const + { + return *this; + } + rational operator-() const + { + return rational(r_.num, -r_.den); + } + bool operator!() const + { + return !r_.num; + } + + //Function: convert to double + double toDouble() const; + + AVRational toAVRational() const; + +#ifdef USE_OTIO + static rational fromRationalTime(const opentime::RationalTime &t) + { + // Is this the best way to do this? + return fromDouble(t.to_seconds()); + } + + // Convert Olive rationals to opentime rationals with the given framerate (defaults to 24) + opentime::RationalTime toRationalTime(double framerate = 24) const; +#endif + + // Produce "flipped" version + rational flipped() const; + void flip(); + + // Returns whether the rational is valid but equal to zero or not + // + // A NaN is always a null, but a null is not always a NaN + bool isNull() const + { + return r_.num == 0; + } + + // Returns whether this rational is not a valid number (denominator == 0) + bool isNaN() const + { + return r_.den == 0; + } + + const int &numerator() const + { + return r_.num; + } + const int &denominator() const + { + return r_.den; + } + + std::string toString() const; + + friend std::ostream &operator<<(std::ostream &out, const rational &value) + { + out << value.r_.num << '/' << value.r_.den; + + return out; + } + +private: + void fix_signs(); + void reduce(); + + AVRational r_; +}; + +#define RATIONAL_MIN rational(INT_MIN) +#define RATIONAL_MAX rational(INT_MAX) + +} + +#endif // LIBOLIVECORE_RATIONAL_H diff --git a/core/include/olive/core/util/sse2neon.h b/core/include/olive/core/util/sse2neon.h new file mode 100644 index 000000000..62338db9d --- /dev/null +++ b/core/include/olive/core/util/sse2neon.h @@ -0,0 +1,8736 @@ +/* + * Olive Community Edition - Non-Linear Video Editor + * Copyright (C) 2025 Olive CE Team + * + * 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 3 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, see . + */ + +#ifndef SSE2NEON_H +#define SSE2NEON_H + +// This header file provides a simple API translation layer +// between SSE intrinsics to their corresponding Arm/Aarch64 NEON versions +// +// This header file does not yet translate all of the SSE intrinsics. +// +// Contributors to this work are: +// John W. Ratcliff +// Brandon Rowlett +// Ken Fast +// Eric van Beurden +// Alexander Potylitsin +// Hasindu Gamaarachchi +// Jim Huang +// Mark Cheng +// Malcolm James MacLeod +// Devin Hussey (easyaspi314) +// Sebastian Pop +// Developer Ecosystem Engineering +// Danila Kutenin +// François Turban (JishinMaster) +// Pei-Hsuan Hung +// Yang-Hao Yuan +// Syoyo Fujita +// Brecht Van Lommel + +/* + * sse2neon is freely redistributable under the MIT License. + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * in the Software without restriction, including without limitation the rights + * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell + * copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +/* Tunable configurations */ + +/* Enable precise implementation of math operations + * This would slow down the computation a bit, but gives consistent result with + * x86 SSE. (e.g. would solve a hole or NaN pixel in the rendering result) + */ +/* _mm_min|max_ps|ss|pd|sd */ +#ifndef SSE2NEON_PRECISE_MINMAX +#define SSE2NEON_PRECISE_MINMAX (0) +#endif +/* _mm_rcp_ps and _mm_div_ps */ +#ifndef SSE2NEON_PRECISE_DIV +#define SSE2NEON_PRECISE_DIV (0) +#endif +/* _mm_sqrt_ps and _mm_rsqrt_ps */ +#ifndef SSE2NEON_PRECISE_SQRT +#define SSE2NEON_PRECISE_SQRT (0) +#endif +/* _mm_dp_pd */ +#ifndef SSE2NEON_PRECISE_DP +#define SSE2NEON_PRECISE_DP (0) +#endif + +/* compiler specific definitions */ +#if defined(__GNUC__) || defined(__clang__) +#pragma push_macro("FORCE_INLINE") +#pragma push_macro("ALIGN_STRUCT") +#define FORCE_INLINE static inline __attribute__((always_inline)) +#define ALIGN_STRUCT(x) __attribute__((aligned(x))) +#define _sse2neon_likely(x) __builtin_expect(!!(x), 1) +#define _sse2neon_unlikely(x) __builtin_expect(!!(x), 0) +#else /* non-GNU / non-clang compilers */ +#warning "Macro name collisions may happen with unsupported compiler." +#ifndef FORCE_INLINE +#define FORCE_INLINE static inline +#endif +#ifndef ALIGN_STRUCT +#define ALIGN_STRUCT(x) __declspec(align(x)) +#endif +#define _sse2neon_likely(x) (x) +#define _sse2neon_unlikely(x) (x) +#endif + +#include +#include + +/* Architecture-specific build options */ +/* FIXME: #pragma GCC push_options is only available on GCC */ +#if defined(__GNUC__) +#if defined(__arm__) && __ARM_ARCH == 7 +/* According to ARM C Language Extensions Architecture specification, + * __ARM_NEON is defined to a value indicating the Advanced SIMD (NEON) + * architecture supported. + */ +#if !defined(__ARM_NEON) || !defined(__ARM_NEON__) +#error "You must enable NEON instructions (e.g. -mfpu=neon) to use SSE2NEON." +#endif +#if !defined(__clang__) +#pragma GCC push_options +#pragma GCC target("fpu=neon") +#endif +#elif defined(__aarch64__) +#if !defined(__clang__) +#pragma GCC push_options +#pragma GCC target("+simd") +#endif +#else +#error "Unsupported target. Must be either ARMv7-A+NEON or ARMv8-A." +#endif +#endif + +#include + +/* Rounding functions require either Aarch64 instructions or libm failback */ +#if !defined(__aarch64__) +#include +#endif + +/* "__has_builtin" can be used to query support for built-in functions + * provided by gcc/clang and other compilers that support it. + */ +#ifndef __has_builtin /* GCC prior to 10 or non-clang compilers */ +/* Compatibility with gcc <= 9 */ +#if __GNUC__ <= 9 +#define __has_builtin(x) HAS##x +#define HAS__builtin_popcount 1 +#define HAS__builtin_popcountll 1 +#else +#define __has_builtin(x) 0 +#endif +#endif + +/** + * MACRO for shuffle parameter for _mm_shuffle_ps(). + * Argument fp3 is a digit[0123] that represents the fp from argument "b" + * of mm_shuffle_ps that will be placed in fp3 of result. fp2 is the same + * for fp2 in result. fp1 is a digit[0123] that represents the fp from + * argument "a" of mm_shuffle_ps that will be places in fp1 of result. + * fp0 is the same for fp0 of result. + */ +#define _MM_SHUFFLE(fp3, fp2, fp1, fp0) \ + (((fp3) << 6) | ((fp2) << 4) | ((fp1) << 2) | ((fp0))) + +/* Rounding mode macros. */ +#define _MM_FROUND_TO_NEAREST_INT 0x00 +#define _MM_FROUND_TO_NEG_INF 0x01 +#define _MM_FROUND_TO_POS_INF 0x02 +#define _MM_FROUND_TO_ZERO 0x03 +#define _MM_FROUND_CUR_DIRECTION 0x04 +#define _MM_FROUND_NO_EXC 0x08 +#define _MM_FROUND_RAISE_EXC 0x00 +#define _MM_FROUND_NINT (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_RAISE_EXC) +#define _MM_FROUND_FLOOR (_MM_FROUND_TO_NEG_INF | _MM_FROUND_RAISE_EXC) +#define _MM_FROUND_CEIL (_MM_FROUND_TO_POS_INF | _MM_FROUND_RAISE_EXC) +#define _MM_FROUND_TRUNC (_MM_FROUND_TO_ZERO | _MM_FROUND_RAISE_EXC) +#define _MM_FROUND_RINT (_MM_FROUND_CUR_DIRECTION | _MM_FROUND_RAISE_EXC) +#define _MM_FROUND_NEARBYINT (_MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC) +#define _MM_ROUND_NEAREST 0x0000 +#define _MM_ROUND_DOWN 0x2000 +#define _MM_ROUND_UP 0x4000 +#define _MM_ROUND_TOWARD_ZERO 0x6000 +/* Flush zero mode macros. */ +#define _MM_FLUSH_ZERO_MASK 0x8000 +#define _MM_FLUSH_ZERO_ON 0x8000 +#define _MM_FLUSH_ZERO_OFF 0x0000 +/* Denormals are zeros mode macros. */ +#define _MM_DENORMALS_ZERO_MASK 0x0040 +#define _MM_DENORMALS_ZERO_ON 0x0040 +#define _MM_DENORMALS_ZERO_OFF 0x0000 + +/* indicate immediate constant argument in a given range */ +#define __constrange(a, b) const + +/* A few intrinsics accept traditional data types like ints or floats, but + * most operate on data types that are specific to SSE. + * If a vector type ends in d, it contains doubles, and if it does not have + * a suffix, it contains floats. An integer vector type can contain any type + * of integer, from chars to shorts to unsigned long longs. + */ +typedef int64x1_t __m64; +typedef float32x4_t __m128; /* 128-bit vector containing 4 floats */ +// On ARM 32-bit architecture, the float64x2_t is not supported. +// The data type __m128d should be represented in a different way for related +// intrinsic conversion. +#if defined(__aarch64__) +typedef float64x2_t __m128d; /* 128-bit vector containing 2 doubles */ +#else +typedef float32x4_t __m128d; +#endif +typedef int64x2_t __m128i; /* 128-bit vector containing integers */ + +/* type-safe casting between types */ + +#define vreinterpretq_m128_f16(x) vreinterpretq_f32_f16(x) +#define vreinterpretq_m128_f32(x) (x) +#define vreinterpretq_m128_f64(x) vreinterpretq_f32_f64(x) + +#define vreinterpretq_m128_u8(x) vreinterpretq_f32_u8(x) +#define vreinterpretq_m128_u16(x) vreinterpretq_f32_u16(x) +#define vreinterpretq_m128_u32(x) vreinterpretq_f32_u32(x) +#define vreinterpretq_m128_u64(x) vreinterpretq_f32_u64(x) + +#define vreinterpretq_m128_s8(x) vreinterpretq_f32_s8(x) +#define vreinterpretq_m128_s16(x) vreinterpretq_f32_s16(x) +#define vreinterpretq_m128_s32(x) vreinterpretq_f32_s32(x) +#define vreinterpretq_m128_s64(x) vreinterpretq_f32_s64(x) + +#define vreinterpretq_f16_m128(x) vreinterpretq_f16_f32(x) +#define vreinterpretq_f32_m128(x) (x) +#define vreinterpretq_f64_m128(x) vreinterpretq_f64_f32(x) + +#define vreinterpretq_u8_m128(x) vreinterpretq_u8_f32(x) +#define vreinterpretq_u16_m128(x) vreinterpretq_u16_f32(x) +#define vreinterpretq_u32_m128(x) vreinterpretq_u32_f32(x) +#define vreinterpretq_u64_m128(x) vreinterpretq_u64_f32(x) + +#define vreinterpretq_s8_m128(x) vreinterpretq_s8_f32(x) +#define vreinterpretq_s16_m128(x) vreinterpretq_s16_f32(x) +#define vreinterpretq_s32_m128(x) vreinterpretq_s32_f32(x) +#define vreinterpretq_s64_m128(x) vreinterpretq_s64_f32(x) + +#define vreinterpretq_m128i_s8(x) vreinterpretq_s64_s8(x) +#define vreinterpretq_m128i_s16(x) vreinterpretq_s64_s16(x) +#define vreinterpretq_m128i_s32(x) vreinterpretq_s64_s32(x) +#define vreinterpretq_m128i_s64(x) (x) + +#define vreinterpretq_m128i_u8(x) vreinterpretq_s64_u8(x) +#define vreinterpretq_m128i_u16(x) vreinterpretq_s64_u16(x) +#define vreinterpretq_m128i_u32(x) vreinterpretq_s64_u32(x) +#define vreinterpretq_m128i_u64(x) vreinterpretq_s64_u64(x) + +#define vreinterpretq_f32_m128i(x) vreinterpretq_f32_s64(x) +#define vreinterpretq_f64_m128i(x) vreinterpretq_f64_s64(x) + +#define vreinterpretq_s8_m128i(x) vreinterpretq_s8_s64(x) +#define vreinterpretq_s16_m128i(x) vreinterpretq_s16_s64(x) +#define vreinterpretq_s32_m128i(x) vreinterpretq_s32_s64(x) +#define vreinterpretq_s64_m128i(x) (x) + +#define vreinterpretq_u8_m128i(x) vreinterpretq_u8_s64(x) +#define vreinterpretq_u16_m128i(x) vreinterpretq_u16_s64(x) +#define vreinterpretq_u32_m128i(x) vreinterpretq_u32_s64(x) +#define vreinterpretq_u64_m128i(x) vreinterpretq_u64_s64(x) + +#define vreinterpret_m64_s8(x) vreinterpret_s64_s8(x) +#define vreinterpret_m64_s16(x) vreinterpret_s64_s16(x) +#define vreinterpret_m64_s32(x) vreinterpret_s64_s32(x) +#define vreinterpret_m64_s64(x) (x) + +#define vreinterpret_m64_u8(x) vreinterpret_s64_u8(x) +#define vreinterpret_m64_u16(x) vreinterpret_s64_u16(x) +#define vreinterpret_m64_u32(x) vreinterpret_s64_u32(x) +#define vreinterpret_m64_u64(x) vreinterpret_s64_u64(x) + +#define vreinterpret_m64_f16(x) vreinterpret_s64_f16(x) +#define vreinterpret_m64_f32(x) vreinterpret_s64_f32(x) +#define vreinterpret_m64_f64(x) vreinterpret_s64_f64(x) + +#define vreinterpret_u8_m64(x) vreinterpret_u8_s64(x) +#define vreinterpret_u16_m64(x) vreinterpret_u16_s64(x) +#define vreinterpret_u32_m64(x) vreinterpret_u32_s64(x) +#define vreinterpret_u64_m64(x) vreinterpret_u64_s64(x) + +#define vreinterpret_s8_m64(x) vreinterpret_s8_s64(x) +#define vreinterpret_s16_m64(x) vreinterpret_s16_s64(x) +#define vreinterpret_s32_m64(x) vreinterpret_s32_s64(x) +#define vreinterpret_s64_m64(x) (x) + +#define vreinterpret_f32_m64(x) vreinterpret_f32_s64(x) + +#if defined(__aarch64__) +#define vreinterpretq_m128d_s32(x) vreinterpretq_f64_s32(x) +#define vreinterpretq_m128d_s64(x) vreinterpretq_f64_s64(x) + +#define vreinterpretq_m128d_u64(x) vreinterpretq_f64_u64(x) + +#define vreinterpretq_m128d_f32(x) vreinterpretq_f64_f32(x) +#define vreinterpretq_m128d_f64(x) (x) + +#define vreinterpretq_s64_m128d(x) vreinterpretq_s64_f64(x) + +#define vreinterpretq_u32_m128d(x) vreinterpretq_u32_f64(x) +#define vreinterpretq_u64_m128d(x) vreinterpretq_u64_f64(x) + +#define vreinterpretq_f64_m128d(x) (x) +#define vreinterpretq_f32_m128d(x) vreinterpretq_f32_f64(x) +#else +#define vreinterpretq_m128d_s32(x) vreinterpretq_f32_s32(x) +#define vreinterpretq_m128d_s64(x) vreinterpretq_f32_s64(x) + +#define vreinterpretq_m128d_u32(x) vreinterpretq_f32_u32(x) +#define vreinterpretq_m128d_u64(x) vreinterpretq_f32_u64(x) + +#define vreinterpretq_m128d_f32(x) (x) + +#define vreinterpretq_s64_m128d(x) vreinterpretq_s64_f32(x) + +#define vreinterpretq_u32_m128d(x) vreinterpretq_u32_f32(x) +#define vreinterpretq_u64_m128d(x) vreinterpretq_u64_f32(x) + +#define vreinterpretq_f32_m128d(x) (x) +#endif + +// A struct is defined in this header file called 'SIMDVec' which can be used +// by applications which attempt to access the contents of an __m128 struct +// directly. It is important to note that accessing the __m128 struct directly +// is bad coding practice by Microsoft: @see: +// https://docs.microsoft.com/en-us/cpp/cpp/m128 +// +// However, some legacy source code may try to access the contents of an __m128 +// struct directly so the developer can use the SIMDVec as an alias for it. Any +// casting must be done manually by the developer, as you cannot cast or +// otherwise alias the base NEON data type for intrinsic operations. +// +// union intended to allow direct access to an __m128 variable using the names +// that the MSVC compiler provides. This union should really only be used when +// trying to access the members of the vector as integer values. GCC/clang +// allow native access to the float members through a simple array access +// operator (in C since 4.6, in C++ since 4.8). +// +// Ideally direct accesses to SIMD vectors should not be used since it can cause +// a performance hit. If it really is needed however, the original __m128 +// variable can be aliased with a pointer to this union and used to access +// individual components. The use of this union should be hidden behind a macro +// that is used throughout the codebase to access the members instead of always +// declaring this type of variable. +typedef union ALIGN_STRUCT(16) SIMDVec { + float m128_f32[4]; // as floats - DON'T USE. Added for convenience. + int8_t m128_i8[16]; // as signed 8-bit integers. + int16_t m128_i16[8]; // as signed 16-bit integers. + int32_t m128_i32[4]; // as signed 32-bit integers. + int64_t m128_i64[2]; // as signed 64-bit integers. + uint8_t m128_u8[16]; // as unsigned 8-bit integers. + uint16_t m128_u16[8]; // as unsigned 16-bit integers. + uint32_t m128_u32[4]; // as unsigned 32-bit integers. + uint64_t m128_u64[2]; // as unsigned 64-bit integers. +} SIMDVec; + +// casting using SIMDVec +#define vreinterpretq_nth_u64_m128i(x, n) (((SIMDVec *)&x)->m128_u64[n]) +#define vreinterpretq_nth_u32_m128i(x, n) (((SIMDVec *)&x)->m128_u32[n]) +#define vreinterpretq_nth_u8_m128i(x, n) (((SIMDVec *)&x)->m128_u8[n]) + +/* SSE macros */ +#define _MM_GET_FLUSH_ZERO_MODE _sse2neon_mm_get_flush_zero_mode +#define _MM_SET_FLUSH_ZERO_MODE _sse2neon_mm_set_flush_zero_mode +#define _MM_GET_DENORMALS_ZERO_MODE _sse2neon_mm_get_denormals_zero_mode +#define _MM_SET_DENORMALS_ZERO_MODE _sse2neon_mm_set_denormals_zero_mode + +// Function declaration +// SSE +FORCE_INLINE unsigned int _MM_GET_ROUNDING_MODE(); +FORCE_INLINE __m128 _mm_move_ss(__m128, __m128); +FORCE_INLINE __m128 _mm_or_ps(__m128, __m128); +FORCE_INLINE __m128 _mm_set_ps1(float); +FORCE_INLINE __m128 _mm_setzero_ps(void); +// SSE2 +FORCE_INLINE __m128i _mm_and_si128(__m128i, __m128i); +FORCE_INLINE __m128i _mm_castps_si128(__m128); +FORCE_INLINE __m128i _mm_cmpeq_epi32(__m128i, __m128i); +FORCE_INLINE __m128i _mm_cvtps_epi32(__m128); +FORCE_INLINE __m128d _mm_move_sd(__m128d, __m128d); +FORCE_INLINE __m128i _mm_or_si128(__m128i, __m128i); +FORCE_INLINE __m128i _mm_set_epi32(int, int, int, int); +FORCE_INLINE __m128i _mm_set_epi64x(int64_t, int64_t); +FORCE_INLINE __m128d _mm_set_pd(double, double); +FORCE_INLINE __m128i _mm_set1_epi32(int); +FORCE_INLINE __m128i _mm_setzero_si128(); +// SSE4.1 +FORCE_INLINE __m128d _mm_ceil_pd(__m128d); +FORCE_INLINE __m128 _mm_ceil_ps(__m128); +FORCE_INLINE __m128d _mm_floor_pd(__m128d); +FORCE_INLINE __m128 _mm_floor_ps(__m128); +FORCE_INLINE __m128d _mm_round_pd(__m128d, int); +FORCE_INLINE __m128 _mm_round_ps(__m128, int); +// SSE4.2 +FORCE_INLINE uint32_t _mm_crc32_u8(uint32_t, uint8_t); + +/* Backwards compatibility for compilers with lack of specific type support */ + +// Older gcc does not define vld1q_u8_x4 type +#if defined(__GNUC__) && !defined(__clang__) && \ + ((__GNUC__ <= 10 && defined(__arm__)) || \ + (__GNUC__ == 10 && __GNUC_MINOR__ < 3 && defined(__aarch64__)) || \ + (__GNUC__ <= 9 && defined(__aarch64__))) +FORCE_INLINE uint8x16x4_t _sse2neon_vld1q_u8_x4(const uint8_t *p) +{ + uint8x16x4_t ret; + ret.val[0] = vld1q_u8(p + 0); + ret.val[1] = vld1q_u8(p + 16); + ret.val[2] = vld1q_u8(p + 32); + ret.val[3] = vld1q_u8(p + 48); + return ret; +} +#else +// Wraps vld1q_u8_x4 +FORCE_INLINE uint8x16x4_t _sse2neon_vld1q_u8_x4(const uint8_t *p) +{ + return vld1q_u8_x4(p); +} +#endif + +/* Function Naming Conventions + * The naming convention of SSE intrinsics is straightforward. A generic SSE + * intrinsic function is given as follows: + * _mm__ + * + * The parts of this format are given as follows: + * 1. describes the operation performed by the intrinsic + * 2. identifies the data type of the function's primary arguments + * + * This last part, , is a little complicated. It identifies the + * content of the input values, and can be set to any of the following values: + * + ps - vectors contain floats (ps stands for packed single-precision) + * + pd - vectors cantain doubles (pd stands for packed double-precision) + * + epi8/epi16/epi32/epi64 - vectors contain 8-bit/16-bit/32-bit/64-bit + * signed integers + * + epu8/epu16/epu32/epu64 - vectors contain 8-bit/16-bit/32-bit/64-bit + * unsigned integers + * + si128 - unspecified 128-bit vector or 256-bit vector + * + m128/m128i/m128d - identifies input vector types when they are different + * than the type of the returned vector + * + * For example, _mm_setzero_ps. The _mm implies that the function returns + * a 128-bit vector. The _ps at the end implies that the argument vectors + * contain floats. + * + * A complete example: Byte Shuffle - pshufb (_mm_shuffle_epi8) + * // Set packed 16-bit integers. 128 bits, 8 short, per 16 bits + * __m128i v_in = _mm_setr_epi16(1, 2, 3, 4, 5, 6, 7, 8); + * // Set packed 8-bit integers + * // 128 bits, 16 chars, per 8 bits + * __m128i v_perm = _mm_setr_epi8(1, 0, 2, 3, 8, 9, 10, 11, + * 4, 5, 12, 13, 6, 7, 14, 15); + * // Shuffle packed 8-bit integers + * __m128i v_out = _mm_shuffle_epi8(v_in, v_perm); // pshufb + * + * Data (Number, Binary, Byte Index): + +------+------+-------------+------+------+-------------+ + | 1 | 2 | 3 | 4 | Number + +------+------+------+------+------+------+------+------+ + | 0000 | 0001 | 0000 | 0010 | 0000 | 0011 | 0000 | 0100 | Binary + +------+------+------+------+------+------+------+------+ + | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | Index + +------+------+------+------+------+------+------+------+ + + +------+------+------+------+------+------+------+------+ + | 5 | 6 | 7 | 8 | Number + +------+------+------+------+------+------+------+------+ + | 0000 | 0101 | 0000 | 0110 | 0000 | 0111 | 0000 | 1000 | Binary + +------+------+------+------+------+------+------+------+ + | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | Index + +------+------+------+------+------+------+------+------+ + * Index (Byte Index): + +------+------+------+------+------+------+------+------+ + | 1 | 0 | 2 | 3 | 8 | 9 | 10 | 11 | + +------+------+------+------+------+------+------+------+ + + +------+------+------+------+------+------+------+------+ + | 4 | 5 | 12 | 13 | 6 | 7 | 14 | 15 | + +------+------+------+------+------+------+------+------+ + * Result: + +------+------+------+------+------+------+------+------+ + | 1 | 0 | 2 | 3 | 8 | 9 | 10 | 11 | Index + +------+------+------+------+------+------+------+------+ + | 0001 | 0000 | 0000 | 0010 | 0000 | 0101 | 0000 | 0110 | Binary + +------+------+------+------+------+------+------+------+ + | 256 | 2 | 5 | 6 | Number + +------+------+------+------+------+------+------+------+ + + +------+------+------+------+------+------+------+------+ + | 4 | 5 | 12 | 13 | 6 | 7 | 14 | 15 | Index + +------+------+------+------+------+------+------+------+ + | 0000 | 0011 | 0000 | 0111 | 0000 | 0100 | 0000 | 1000 | Binary + +------+------+------+------+------+------+------+------+ + | 3 | 7 | 4 | 8 | Number + +------+------+------+------+------+------+-------------+ + */ + +/* Constants for use with _mm_prefetch. */ +enum _mm_hint { + _MM_HINT_NTA = 0, /* load data to L1 and L2 cache, mark it as NTA */ + _MM_HINT_T0 = 1, /* load data to L1 and L2 cache */ + _MM_HINT_T1 = 2, /* load data to L2 cache only */ + _MM_HINT_T2 = 3, /* load data to L2 cache only, mark it as NTA */ + _MM_HINT_ENTA = 4, /* exclusive version of _MM_HINT_NTA */ + _MM_HINT_ET0 = 5, /* exclusive version of _MM_HINT_T0 */ + _MM_HINT_ET1 = 6, /* exclusive version of _MM_HINT_T1 */ + _MM_HINT_ET2 = 7 /* exclusive version of _MM_HINT_T2 */ +}; + +// The bit field mapping to the FPCR(floating-point control register) +typedef struct { + uint16_t res0; + uint8_t res1 : 6; + uint8_t bit22 : 1; + uint8_t bit23 : 1; + uint8_t bit24 : 1; + uint8_t res2 : 7; +#if defined(__aarch64__) + uint32_t res3; +#endif +} fpcr_bitfield; + +// Takes the upper 64 bits of a and places it in the low end of the result +// Takes the lower 64 bits of b and places it into the high end of the result. +FORCE_INLINE __m128 _mm_shuffle_ps_1032(__m128 a, __m128 b) +{ + float32x2_t a32 = vget_high_f32(vreinterpretq_f32_m128(a)); + float32x2_t b10 = vget_low_f32(vreinterpretq_f32_m128(b)); + return vreinterpretq_m128_f32(vcombine_f32(a32, b10)); +} + +// takes the lower two 32-bit values from a and swaps them and places in high +// end of result takes the higher two 32 bit values from b and swaps them and +// places in low end of result. +FORCE_INLINE __m128 _mm_shuffle_ps_2301(__m128 a, __m128 b) +{ + float32x2_t a01 = vrev64_f32(vget_low_f32(vreinterpretq_f32_m128(a))); + float32x2_t b23 = vrev64_f32(vget_high_f32(vreinterpretq_f32_m128(b))); + return vreinterpretq_m128_f32(vcombine_f32(a01, b23)); +} + +FORCE_INLINE __m128 _mm_shuffle_ps_0321(__m128 a, __m128 b) +{ + float32x2_t a21 = vget_high_f32( + vextq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a), 3)); + float32x2_t b03 = vget_low_f32( + vextq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b), 3)); + return vreinterpretq_m128_f32(vcombine_f32(a21, b03)); +} + +FORCE_INLINE __m128 _mm_shuffle_ps_2103(__m128 a, __m128 b) +{ + float32x2_t a03 = vget_low_f32( + vextq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a), 3)); + float32x2_t b21 = vget_high_f32( + vextq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b), 3)); + return vreinterpretq_m128_f32(vcombine_f32(a03, b21)); +} + +FORCE_INLINE __m128 _mm_shuffle_ps_1010(__m128 a, __m128 b) +{ + float32x2_t a10 = vget_low_f32(vreinterpretq_f32_m128(a)); + float32x2_t b10 = vget_low_f32(vreinterpretq_f32_m128(b)); + return vreinterpretq_m128_f32(vcombine_f32(a10, b10)); +} + +FORCE_INLINE __m128 _mm_shuffle_ps_1001(__m128 a, __m128 b) +{ + float32x2_t a01 = vrev64_f32(vget_low_f32(vreinterpretq_f32_m128(a))); + float32x2_t b10 = vget_low_f32(vreinterpretq_f32_m128(b)); + return vreinterpretq_m128_f32(vcombine_f32(a01, b10)); +} + +FORCE_INLINE __m128 _mm_shuffle_ps_0101(__m128 a, __m128 b) +{ + float32x2_t a01 = vrev64_f32(vget_low_f32(vreinterpretq_f32_m128(a))); + float32x2_t b01 = vrev64_f32(vget_low_f32(vreinterpretq_f32_m128(b))); + return vreinterpretq_m128_f32(vcombine_f32(a01, b01)); +} + +// keeps the low 64 bits of b in the low and puts the high 64 bits of a in the +// high +FORCE_INLINE __m128 _mm_shuffle_ps_3210(__m128 a, __m128 b) +{ + float32x2_t a10 = vget_low_f32(vreinterpretq_f32_m128(a)); + float32x2_t b32 = vget_high_f32(vreinterpretq_f32_m128(b)); + return vreinterpretq_m128_f32(vcombine_f32(a10, b32)); +} + +FORCE_INLINE __m128 _mm_shuffle_ps_0011(__m128 a, __m128 b) +{ + float32x2_t a11 = vdup_lane_f32(vget_low_f32(vreinterpretq_f32_m128(a)), 1); + float32x2_t b00 = vdup_lane_f32(vget_low_f32(vreinterpretq_f32_m128(b)), 0); + return vreinterpretq_m128_f32(vcombine_f32(a11, b00)); +} + +FORCE_INLINE __m128 _mm_shuffle_ps_0022(__m128 a, __m128 b) +{ + float32x2_t a22 = + vdup_lane_f32(vget_high_f32(vreinterpretq_f32_m128(a)), 0); + float32x2_t b00 = vdup_lane_f32(vget_low_f32(vreinterpretq_f32_m128(b)), 0); + return vreinterpretq_m128_f32(vcombine_f32(a22, b00)); +} + +FORCE_INLINE __m128 _mm_shuffle_ps_2200(__m128 a, __m128 b) +{ + float32x2_t a00 = vdup_lane_f32(vget_low_f32(vreinterpretq_f32_m128(a)), 0); + float32x2_t b22 = + vdup_lane_f32(vget_high_f32(vreinterpretq_f32_m128(b)), 0); + return vreinterpretq_m128_f32(vcombine_f32(a00, b22)); +} + +FORCE_INLINE __m128 _mm_shuffle_ps_3202(__m128 a, __m128 b) +{ + float32_t a0 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); + float32x2_t a22 = + vdup_lane_f32(vget_high_f32(vreinterpretq_f32_m128(a)), 0); + float32x2_t a02 = vset_lane_f32(a0, a22, 1); /* TODO: use vzip ?*/ + float32x2_t b32 = vget_high_f32(vreinterpretq_f32_m128(b)); + return vreinterpretq_m128_f32(vcombine_f32(a02, b32)); +} + +FORCE_INLINE __m128 _mm_shuffle_ps_1133(__m128 a, __m128 b) +{ + float32x2_t a33 = + vdup_lane_f32(vget_high_f32(vreinterpretq_f32_m128(a)), 1); + float32x2_t b11 = vdup_lane_f32(vget_low_f32(vreinterpretq_f32_m128(b)), 1); + return vreinterpretq_m128_f32(vcombine_f32(a33, b11)); +} + +FORCE_INLINE __m128 _mm_shuffle_ps_2010(__m128 a, __m128 b) +{ + float32x2_t a10 = vget_low_f32(vreinterpretq_f32_m128(a)); + float32_t b2 = vgetq_lane_f32(vreinterpretq_f32_m128(b), 2); + float32x2_t b00 = vdup_lane_f32(vget_low_f32(vreinterpretq_f32_m128(b)), 0); + float32x2_t b20 = vset_lane_f32(b2, b00, 1); + return vreinterpretq_m128_f32(vcombine_f32(a10, b20)); +} + +FORCE_INLINE __m128 _mm_shuffle_ps_2001(__m128 a, __m128 b) +{ + float32x2_t a01 = vrev64_f32(vget_low_f32(vreinterpretq_f32_m128(a))); + float32_t b2 = vgetq_lane_f32(b, 2); + float32x2_t b00 = vdup_lane_f32(vget_low_f32(vreinterpretq_f32_m128(b)), 0); + float32x2_t b20 = vset_lane_f32(b2, b00, 1); + return vreinterpretq_m128_f32(vcombine_f32(a01, b20)); +} + +FORCE_INLINE __m128 _mm_shuffle_ps_2032(__m128 a, __m128 b) +{ + float32x2_t a32 = vget_high_f32(vreinterpretq_f32_m128(a)); + float32_t b2 = vgetq_lane_f32(b, 2); + float32x2_t b00 = vdup_lane_f32(vget_low_f32(vreinterpretq_f32_m128(b)), 0); + float32x2_t b20 = vset_lane_f32(b2, b00, 1); + return vreinterpretq_m128_f32(vcombine_f32(a32, b20)); +} + +// Kahan summation for accurate summation of floating-point numbers. +// http://blog.zachbjornson.com/2019/08/11/fast-float-summation.html +FORCE_INLINE void _sse2neon_kadd_f32(float *sum, float *c, float y) +{ + y -= *c; + float t = *sum + y; + *c = (t - *sum) - y; + *sum = t; +} + +#if defined(__ARM_FEATURE_CRYPTO) +// Wraps vmull_p64 +FORCE_INLINE uint64x2_t _sse2neon_vmull_p64(uint64x1_t _a, uint64x1_t _b) +{ + poly64_t a = vget_lane_p64(vreinterpret_p64_u64(_a), 0); + poly64_t b = vget_lane_p64(vreinterpret_p64_u64(_b), 0); + return vreinterpretq_u64_p128(vmull_p64(a, b)); +} +#else // ARMv7 polyfill +// ARMv7/some A64 lacks vmull_p64, but it has vmull_p8. +// +// vmull_p8 calculates 8 8-bit->16-bit polynomial multiplies, but we need a +// 64-bit->128-bit polynomial multiply. +// +// It needs some work and is somewhat slow, but it is still faster than all +// known scalar methods. +// +// Algorithm adapted to C from +// https://www.workofard.com/2017/07/ghash-for-low-end-cores/, which is adapted +// from "Fast Software Polynomial Multiplication on ARM Processors Using the +// NEON Engine" by Danilo Camara, Conrado Gouvea, Julio Lopez and Ricardo Dahab +// (https://hal.inria.fr/hal-01506572) +static uint64x2_t _sse2neon_vmull_p64(uint64x1_t _a, uint64x1_t _b) +{ + poly8x8_t a = vreinterpret_p8_u64(_a); + poly8x8_t b = vreinterpret_p8_u64(_b); + + // Masks + uint8x16_t k48_32 = vcombine_u8(vcreate_u8(0x0000ffffffffffff), + vcreate_u8(0x00000000ffffffff)); + uint8x16_t k16_00 = vcombine_u8(vcreate_u8(0x000000000000ffff), + vcreate_u8(0x0000000000000000)); + + // Do the multiplies, rotating with vext to get all combinations + uint8x16_t d = vreinterpretq_u8_p16(vmull_p8(a, b)); // D = A0 * B0 + uint8x16_t e = + vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 1))); // E = A0 * B1 + uint8x16_t f = + vreinterpretq_u8_p16(vmull_p8(vext_p8(a, a, 1), b)); // F = A1 * B0 + uint8x16_t g = + vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 2))); // G = A0 * B2 + uint8x16_t h = + vreinterpretq_u8_p16(vmull_p8(vext_p8(a, a, 2), b)); // H = A2 * B0 + uint8x16_t i = + vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 3))); // I = A0 * B3 + uint8x16_t j = + vreinterpretq_u8_p16(vmull_p8(vext_p8(a, a, 3), b)); // J = A3 * B0 + uint8x16_t k = + vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 4))); // L = A0 * B4 + + // Add cross products + uint8x16_t l = veorq_u8(e, f); // L = E + F + uint8x16_t m = veorq_u8(g, h); // M = G + H + uint8x16_t n = veorq_u8(i, j); // N = I + J + + // Interleave. Using vzip1 and vzip2 prevents Clang from emitting TBL + // instructions. +#if defined(__aarch64__) + uint8x16_t lm_p0 = vreinterpretq_u8_u64( + vzip1q_u64(vreinterpretq_u64_u8(l), vreinterpretq_u64_u8(m))); + uint8x16_t lm_p1 = vreinterpretq_u8_u64( + vzip2q_u64(vreinterpretq_u64_u8(l), vreinterpretq_u64_u8(m))); + uint8x16_t nk_p0 = vreinterpretq_u8_u64( + vzip1q_u64(vreinterpretq_u64_u8(n), vreinterpretq_u64_u8(k))); + uint8x16_t nk_p1 = vreinterpretq_u8_u64( + vzip2q_u64(vreinterpretq_u64_u8(n), vreinterpretq_u64_u8(k))); +#else + uint8x16_t lm_p0 = vcombine_u8(vget_low_u8(l), vget_low_u8(m)); + uint8x16_t lm_p1 = vcombine_u8(vget_high_u8(l), vget_high_u8(m)); + uint8x16_t nk_p0 = vcombine_u8(vget_low_u8(n), vget_low_u8(k)); + uint8x16_t nk_p1 = vcombine_u8(vget_high_u8(n), vget_high_u8(k)); +#endif + // t0 = (L) (P0 + P1) << 8 + // t1 = (M) (P2 + P3) << 16 + uint8x16_t t0t1_tmp = veorq_u8(lm_p0, lm_p1); + uint8x16_t t0t1_h = vandq_u8(lm_p1, k48_32); + uint8x16_t t0t1_l = veorq_u8(t0t1_tmp, t0t1_h); + + // t2 = (N) (P4 + P5) << 24 + // t3 = (K) (P6 + P7) << 32 + uint8x16_t t2t3_tmp = veorq_u8(nk_p0, nk_p1); + uint8x16_t t2t3_h = vandq_u8(nk_p1, k16_00); + uint8x16_t t2t3_l = veorq_u8(t2t3_tmp, t2t3_h); + + // De-interleave +#if defined(__aarch64__) + uint8x16_t t0 = vreinterpretq_u8_u64( + vuzp1q_u64(vreinterpretq_u64_u8(t0t1_l), vreinterpretq_u64_u8(t0t1_h))); + uint8x16_t t1 = vreinterpretq_u8_u64( + vuzp2q_u64(vreinterpretq_u64_u8(t0t1_l), vreinterpretq_u64_u8(t0t1_h))); + uint8x16_t t2 = vreinterpretq_u8_u64( + vuzp1q_u64(vreinterpretq_u64_u8(t2t3_l), vreinterpretq_u64_u8(t2t3_h))); + uint8x16_t t3 = vreinterpretq_u8_u64( + vuzp2q_u64(vreinterpretq_u64_u8(t2t3_l), vreinterpretq_u64_u8(t2t3_h))); +#else + uint8x16_t t1 = vcombine_u8(vget_high_u8(t0t1_l), vget_high_u8(t0t1_h)); + uint8x16_t t0 = vcombine_u8(vget_low_u8(t0t1_l), vget_low_u8(t0t1_h)); + uint8x16_t t3 = vcombine_u8(vget_high_u8(t2t3_l), vget_high_u8(t2t3_h)); + uint8x16_t t2 = vcombine_u8(vget_low_u8(t2t3_l), vget_low_u8(t2t3_h)); +#endif + // Shift the cross products + uint8x16_t t0_shift = vextq_u8(t0, t0, 15); // t0 << 8 + uint8x16_t t1_shift = vextq_u8(t1, t1, 14); // t1 << 16 + uint8x16_t t2_shift = vextq_u8(t2, t2, 13); // t2 << 24 + uint8x16_t t3_shift = vextq_u8(t3, t3, 12); // t3 << 32 + + // Accumulate the products + uint8x16_t cross1 = veorq_u8(t0_shift, t1_shift); + uint8x16_t cross2 = veorq_u8(t2_shift, t3_shift); + uint8x16_t mix = veorq_u8(d, cross1); + uint8x16_t r = veorq_u8(mix, cross2); + return vreinterpretq_u64_u8(r); +} +#endif // ARMv7 polyfill + +// C equivalent: +// __m128i _mm_shuffle_epi32_default(__m128i a, +// __constrange(0, 255) int imm) { +// __m128i ret; +// ret[0] = a[imm & 0x3]; ret[1] = a[(imm >> 2) & 0x3]; +// ret[2] = a[(imm >> 4) & 0x03]; ret[3] = a[(imm >> 6) & 0x03]; +// return ret; +// } +#define _mm_shuffle_epi32_default(a, imm) \ + __extension__({ \ + int32x4_t ret; \ + ret = vmovq_n_s32( \ + vgetq_lane_s32(vreinterpretq_s32_m128i(a), (imm) & (0x3))); \ + ret = vsetq_lane_s32(vgetq_lane_s32(vreinterpretq_s32_m128i(a), \ + ((imm) >> 2) & 0x3), \ + ret, 1); \ + ret = vsetq_lane_s32(vgetq_lane_s32(vreinterpretq_s32_m128i(a), \ + ((imm) >> 4) & 0x3), \ + ret, 2); \ + ret = vsetq_lane_s32(vgetq_lane_s32(vreinterpretq_s32_m128i(a), \ + ((imm) >> 6) & 0x3), \ + ret, 3); \ + vreinterpretq_m128i_s32(ret); \ + }) + +// Takes the upper 64 bits of a and places it in the low end of the result +// Takes the lower 64 bits of a and places it into the high end of the result. +FORCE_INLINE __m128i _mm_shuffle_epi_1032(__m128i a) +{ + int32x2_t a32 = vget_high_s32(vreinterpretq_s32_m128i(a)); + int32x2_t a10 = vget_low_s32(vreinterpretq_s32_m128i(a)); + return vreinterpretq_m128i_s32(vcombine_s32(a32, a10)); +} + +// takes the lower two 32-bit values from a and swaps them and places in low end +// of result takes the higher two 32 bit values from a and swaps them and places +// in high end of result. +FORCE_INLINE __m128i _mm_shuffle_epi_2301(__m128i a) +{ + int32x2_t a01 = vrev64_s32(vget_low_s32(vreinterpretq_s32_m128i(a))); + int32x2_t a23 = vrev64_s32(vget_high_s32(vreinterpretq_s32_m128i(a))); + return vreinterpretq_m128i_s32(vcombine_s32(a01, a23)); +} + +// rotates the least significant 32 bits into the most significant 32 bits, and +// shifts the rest down +FORCE_INLINE __m128i _mm_shuffle_epi_0321(__m128i a) +{ + return vreinterpretq_m128i_s32( + vextq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(a), 1)); +} + +// rotates the most significant 32 bits into the least significant 32 bits, and +// shifts the rest up +FORCE_INLINE __m128i _mm_shuffle_epi_2103(__m128i a) +{ + return vreinterpretq_m128i_s32( + vextq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(a), 3)); +} + +// gets the lower 64 bits of a, and places it in the upper 64 bits +// gets the lower 64 bits of a and places it in the lower 64 bits +FORCE_INLINE __m128i _mm_shuffle_epi_1010(__m128i a) +{ + int32x2_t a10 = vget_low_s32(vreinterpretq_s32_m128i(a)); + return vreinterpretq_m128i_s32(vcombine_s32(a10, a10)); +} + +// gets the lower 64 bits of a, swaps the 0 and 1 elements, and places it in the +// lower 64 bits gets the lower 64 bits of a, and places it in the upper 64 bits +FORCE_INLINE __m128i _mm_shuffle_epi_1001(__m128i a) +{ + int32x2_t a01 = vrev64_s32(vget_low_s32(vreinterpretq_s32_m128i(a))); + int32x2_t a10 = vget_low_s32(vreinterpretq_s32_m128i(a)); + return vreinterpretq_m128i_s32(vcombine_s32(a01, a10)); +} + +// gets the lower 64 bits of a, swaps the 0 and 1 elements and places it in the +// upper 64 bits gets the lower 64 bits of a, swaps the 0 and 1 elements, and +// places it in the lower 64 bits +FORCE_INLINE __m128i _mm_shuffle_epi_0101(__m128i a) +{ + int32x2_t a01 = vrev64_s32(vget_low_s32(vreinterpretq_s32_m128i(a))); + return vreinterpretq_m128i_s32(vcombine_s32(a01, a01)); +} + +FORCE_INLINE __m128i _mm_shuffle_epi_2211(__m128i a) +{ + int32x2_t a11 = vdup_lane_s32(vget_low_s32(vreinterpretq_s32_m128i(a)), 1); + int32x2_t a22 = vdup_lane_s32(vget_high_s32(vreinterpretq_s32_m128i(a)), 0); + return vreinterpretq_m128i_s32(vcombine_s32(a11, a22)); +} + +FORCE_INLINE __m128i _mm_shuffle_epi_0122(__m128i a) +{ + int32x2_t a22 = vdup_lane_s32(vget_high_s32(vreinterpretq_s32_m128i(a)), 0); + int32x2_t a01 = vrev64_s32(vget_low_s32(vreinterpretq_s32_m128i(a))); + return vreinterpretq_m128i_s32(vcombine_s32(a22, a01)); +} + +FORCE_INLINE __m128i _mm_shuffle_epi_3332(__m128i a) +{ + int32x2_t a32 = vget_high_s32(vreinterpretq_s32_m128i(a)); + int32x2_t a33 = vdup_lane_s32(vget_high_s32(vreinterpretq_s32_m128i(a)), 1); + return vreinterpretq_m128i_s32(vcombine_s32(a32, a33)); +} + +// FORCE_INLINE __m128i _mm_shuffle_epi32_splat(__m128i a, __constrange(0,255) +// int imm) +#if defined(__aarch64__) +#define _mm_shuffle_epi32_splat(a, imm) \ + __extension__({ \ + vreinterpretq_m128i_s32( \ + vdupq_laneq_s32(vreinterpretq_s32_m128i(a), (imm))); \ + }) +#else +#define _mm_shuffle_epi32_splat(a, imm) \ + __extension__({ \ + vreinterpretq_m128i_s32( \ + vdupq_n_s32(vgetq_lane_s32(vreinterpretq_s32_m128i(a), (imm)))); \ + }) +#endif + +// NEON does not support a general purpose permute intrinsic +// Selects four specific single-precision, floating-point values from a and b, +// based on the mask i. +// +// C equivalent: +// __m128 _mm_shuffle_ps_default(__m128 a, __m128 b, +// __constrange(0, 255) int imm) { +// __m128 ret; +// ret[0] = a[imm & 0x3]; ret[1] = a[(imm >> 2) & 0x3]; +// ret[2] = b[(imm >> 4) & 0x03]; ret[3] = b[(imm >> 6) & 0x03]; +// return ret; +// } +// +// https://msdn.microsoft.com/en-us/library/vstudio/5f0858x0(v=vs.100).aspx +#define _mm_shuffle_ps_default(a, b, imm) \ + __extension__({ \ + float32x4_t ret; \ + ret = vmovq_n_f32( \ + vgetq_lane_f32(vreinterpretq_f32_m128(a), (imm) & (0x3))); \ + ret = vsetq_lane_f32(vgetq_lane_f32(vreinterpretq_f32_m128(a), \ + ((imm) >> 2) & 0x3), \ + ret, 1); \ + ret = vsetq_lane_f32(vgetq_lane_f32(vreinterpretq_f32_m128(b), \ + ((imm) >> 4) & 0x3), \ + ret, 2); \ + ret = vsetq_lane_f32(vgetq_lane_f32(vreinterpretq_f32_m128(b), \ + ((imm) >> 6) & 0x3), \ + ret, 3); \ + vreinterpretq_m128_f32(ret); \ + }) + +// Shuffles the lower 4 signed or unsigned 16-bit integers in a as specified +// by imm. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/y41dkk37(v=vs.100) +// FORCE_INLINE __m128i _mm_shufflelo_epi16_function(__m128i a, +// __constrange(0,255) int +// imm) +#define _mm_shufflelo_epi16_function(a, imm) \ + __extension__({ \ + int16x8_t ret = vreinterpretq_s16_m128i(a); \ + int16x4_t lowBits = vget_low_s16(ret); \ + ret = vsetq_lane_s16(vget_lane_s16(lowBits, (imm) & (0x3)), ret, 0); \ + ret = vsetq_lane_s16(vget_lane_s16(lowBits, ((imm) >> 2) & 0x3), ret, \ + 1); \ + ret = vsetq_lane_s16(vget_lane_s16(lowBits, ((imm) >> 4) & 0x3), ret, \ + 2); \ + ret = vsetq_lane_s16(vget_lane_s16(lowBits, ((imm) >> 6) & 0x3), ret, \ + 3); \ + vreinterpretq_m128i_s16(ret); \ + }) + +// Shuffles the upper 4 signed or unsigned 16-bit integers in a as specified +// by imm. +// https://msdn.microsoft.com/en-us/library/13ywktbs(v=vs.100).aspx +// FORCE_INLINE __m128i _mm_shufflehi_epi16_function(__m128i a, +// __constrange(0,255) int +// imm) +#define _mm_shufflehi_epi16_function(a, imm) \ + __extension__({ \ + int16x8_t ret = vreinterpretq_s16_m128i(a); \ + int16x4_t highBits = vget_high_s16(ret); \ + ret = vsetq_lane_s16(vget_lane_s16(highBits, (imm) & (0x3)), ret, 4); \ + ret = vsetq_lane_s16(vget_lane_s16(highBits, ((imm) >> 2) & 0x3), ret, \ + 5); \ + ret = vsetq_lane_s16(vget_lane_s16(highBits, ((imm) >> 4) & 0x3), ret, \ + 6); \ + ret = vsetq_lane_s16(vget_lane_s16(highBits, ((imm) >> 6) & 0x3), ret, \ + 7); \ + vreinterpretq_m128i_s16(ret); \ + }) + +/* MMX */ + +//_mm_empty is a no-op on arm +FORCE_INLINE void _mm_empty(void) +{ +} + +/* SSE */ + +// Adds the four single-precision, floating-point values of a and b. +// +// r0 := a0 + b0 +// r1 := a1 + b1 +// r2 := a2 + b2 +// r3 := a3 + b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/c9848chc(v=vs.100).aspx +FORCE_INLINE __m128 _mm_add_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_f32( + vaddq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// adds the scalar single-precision floating point values of a and b. +// https://msdn.microsoft.com/en-us/library/be94x2y6(v=vs.100).aspx +FORCE_INLINE __m128 _mm_add_ss(__m128 a, __m128 b) +{ + float32_t b0 = vgetq_lane_f32(vreinterpretq_f32_m128(b), 0); + float32x4_t value = vsetq_lane_f32(b0, vdupq_n_f32(0), 0); + // the upper values in the result must be the remnants of . + return vreinterpretq_m128_f32(vaddq_f32(a, value)); +} + +// Computes the bitwise AND of the four single-precision, floating-point values +// of a and b. +// +// r0 := a0 & b0 +// r1 := a1 & b1 +// r2 := a2 & b2 +// r3 := a3 & b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/73ck1xc5(v=vs.100).aspx +FORCE_INLINE __m128 _mm_and_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_s32( + vandq_s32(vreinterpretq_s32_m128(a), vreinterpretq_s32_m128(b))); +} + +// Computes the bitwise AND-NOT of the four single-precision, floating-point +// values of a and b. +// +// r0 := ~a0 & b0 +// r1 := ~a1 & b1 +// r2 := ~a2 & b2 +// r3 := ~a3 & b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/68h7wd02(v=vs.100).aspx +FORCE_INLINE __m128 _mm_andnot_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_s32( + vbicq_s32(vreinterpretq_s32_m128(b), + vreinterpretq_s32_m128(a))); // *NOTE* argument swap +} + +// Average packed unsigned 16-bit integers in a and b, and store the results in +// dst. +// +// FOR j := 0 to 3 +// i := j*16 +// dst[i+15:i] := (a[i+15:i] + b[i+15:i] + 1) >> 1 +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_avg_pu16 +FORCE_INLINE __m64 _mm_avg_pu16(__m64 a, __m64 b) +{ + return vreinterpret_m64_u16( + vrhadd_u16(vreinterpret_u16_m64(a), vreinterpret_u16_m64(b))); +} + +// Average packed unsigned 8-bit integers in a and b, and store the results in +// dst. +// +// FOR j := 0 to 7 +// i := j*8 +// dst[i+7:i] := (a[i+7:i] + b[i+7:i] + 1) >> 1 +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_avg_pu8 +FORCE_INLINE __m64 _mm_avg_pu8(__m64 a, __m64 b) +{ + return vreinterpret_m64_u8( + vrhadd_u8(vreinterpret_u8_m64(a), vreinterpret_u8_m64(b))); +} + +// Compares for equality. +// https://msdn.microsoft.com/en-us/library/vstudio/36aectz5(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cmpeq_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32( + vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Compares for equality. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/k423z28e(v=vs.100) +FORCE_INLINE __m128 _mm_cmpeq_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpeq_ps(a, b)); +} + +// Compares for greater than or equal. +// https://msdn.microsoft.com/en-us/library/vstudio/fs813y2t(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cmpge_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32( + vcgeq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Compares for greater than or equal. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/kesh3ddc(v=vs.100) +FORCE_INLINE __m128 _mm_cmpge_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpge_ps(a, b)); +} + +// Compares for greater than. +// +// r0 := (a0 > b0) ? 0xffffffff : 0x0 +// r1 := (a1 > b1) ? 0xffffffff : 0x0 +// r2 := (a2 > b2) ? 0xffffffff : 0x0 +// r3 := (a3 > b3) ? 0xffffffff : 0x0 +// +// https://msdn.microsoft.com/en-us/library/vstudio/11dy102s(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cmpgt_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32( + vcgtq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Compares for greater than. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/1xyyyy9e(v=vs.100) +FORCE_INLINE __m128 _mm_cmpgt_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpgt_ps(a, b)); +} + +// Compares for less than or equal. +// +// r0 := (a0 <= b0) ? 0xffffffff : 0x0 +// r1 := (a1 <= b1) ? 0xffffffff : 0x0 +// r2 := (a2 <= b2) ? 0xffffffff : 0x0 +// r3 := (a3 <= b3) ? 0xffffffff : 0x0 +// +// https://msdn.microsoft.com/en-us/library/vstudio/1s75w83z(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cmple_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32( + vcleq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Compares for less than or equal. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/a7x0hbhw(v=vs.100) +FORCE_INLINE __m128 _mm_cmple_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmple_ps(a, b)); +} + +// Compares for less than +// https://msdn.microsoft.com/en-us/library/vstudio/f330yhc8(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cmplt_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32( + vcltq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Compares for less than +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/fy94wye7(v=vs.100) +FORCE_INLINE __m128 _mm_cmplt_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmplt_ps(a, b)); +} + +// Compares for inequality. +// https://msdn.microsoft.com/en-us/library/sf44thbx(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cmpneq_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32(vmvnq_u32( + vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)))); +} + +// Compares for inequality. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/ekya8fh4(v=vs.100) +FORCE_INLINE __m128 _mm_cmpneq_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpneq_ps(a, b)); +} + +// Compares for not greater than or equal. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/wsexys62(v=vs.100) +FORCE_INLINE __m128 _mm_cmpnge_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32(vmvnq_u32( + vcgeq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)))); +} + +// Compares for not greater than or equal. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/fk2y80s8(v=vs.100) +FORCE_INLINE __m128 _mm_cmpnge_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpnge_ps(a, b)); +} + +// Compares for not greater than. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/d0xh7w0s(v=vs.100) +FORCE_INLINE __m128 _mm_cmpngt_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32(vmvnq_u32( + vcgtq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)))); +} + +// Compares for not greater than. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/z7x9ydwh(v=vs.100) +FORCE_INLINE __m128 _mm_cmpngt_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpngt_ps(a, b)); +} + +// Compares for not less than or equal. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/6a330kxw(v=vs.100) +FORCE_INLINE __m128 _mm_cmpnle_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32(vmvnq_u32( + vcleq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)))); +} + +// Compares for not less than or equal. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/z7x9ydwh(v=vs.100) +FORCE_INLINE __m128 _mm_cmpnle_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpnle_ps(a, b)); +} + +// Compares for not less than. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/4686bbdw(v=vs.100) +FORCE_INLINE __m128 _mm_cmpnlt_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32(vmvnq_u32( + vcltq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)))); +} + +// Compares for not less than. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/56b9z2wf(v=vs.100) +FORCE_INLINE __m128 _mm_cmpnlt_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpnlt_ps(a, b)); +} + +// Compares the four 32-bit floats in a and b to check if any values are NaN. +// Ordered compare between each value returns true for "orderable" and false for +// "not orderable" (NaN). +// https://msdn.microsoft.com/en-us/library/vstudio/0h9w00fx(v=vs.100).aspx see +// also: +// http://stackoverflow.com/questions/8627331/what-does-ordered-unordered-comparison-mean +// http://stackoverflow.com/questions/29349621/neon-isnanval-intrinsics +FORCE_INLINE __m128 _mm_cmpord_ps(__m128 a, __m128 b) +{ + // Note: NEON does not have ordered compare builtin + // Need to compare a eq a and b eq b to check for NaN + // Do AND of results to get final + uint32x4_t ceqaa = + vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a)); + uint32x4_t ceqbb = + vceqq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b)); + return vreinterpretq_m128_u32(vandq_u32(ceqaa, ceqbb)); +} + +// Compares for ordered. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/343t62da(v=vs.100) +FORCE_INLINE __m128 _mm_cmpord_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpord_ps(a, b)); +} + +// Compares for unordered. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/khy6fk1t(v=vs.100) +FORCE_INLINE __m128 _mm_cmpunord_ps(__m128 a, __m128 b) +{ + uint32x4_t f32a = + vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a)); + uint32x4_t f32b = + vceqq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b)); + return vreinterpretq_m128_u32(vmvnq_u32(vandq_u32(f32a, f32b))); +} + +// Compares for unordered. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/2as2387b(v=vs.100) +FORCE_INLINE __m128 _mm_cmpunord_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpunord_ps(a, b)); +} + +// Compares the lower single-precision floating point scalar values of a and b +// using an equality operation. : +// https://msdn.microsoft.com/en-us/library/93yx2h2b(v=vs.100).aspx +FORCE_INLINE int _mm_comieq_ss(__m128 a, __m128 b) +{ + uint32x4_t a_eq_b = + vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); + return vgetq_lane_u32(a_eq_b, 0) & 0x1; +} + +// Compares the lower single-precision floating point scalar values of a and b +// using a greater than or equal operation. : +// https://msdn.microsoft.com/en-us/library/8t80des6(v=vs.100).aspx +FORCE_INLINE int _mm_comige_ss(__m128 a, __m128 b) +{ + uint32x4_t a_ge_b = + vcgeq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); + return vgetq_lane_u32(a_ge_b, 0) & 0x1; +} + +// Compares the lower single-precision floating point scalar values of a and b +// using a greater than operation. : +// https://msdn.microsoft.com/en-us/library/b0738e0t(v=vs.100).aspx +FORCE_INLINE int _mm_comigt_ss(__m128 a, __m128 b) +{ + uint32x4_t a_gt_b = + vcgtq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); + return vgetq_lane_u32(a_gt_b, 0) & 0x1; +} + +// Compares the lower single-precision floating point scalar values of a and b +// using a less than or equal operation. : +// https://msdn.microsoft.com/en-us/library/1w4t7c57(v=vs.90).aspx +FORCE_INLINE int _mm_comile_ss(__m128 a, __m128 b) +{ + uint32x4_t a_le_b = + vcleq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); + return vgetq_lane_u32(a_le_b, 0) & 0x1; +} + +// Compares the lower single-precision floating point scalar values of a and b +// using a less than operation. : +// https://msdn.microsoft.com/en-us/library/2kwe606b(v=vs.90).aspx Important +// note!! The documentation on MSDN is incorrect! If either of the values is a +// NAN the docs say you will get a one, but in fact, it will return a zero!! +FORCE_INLINE int _mm_comilt_ss(__m128 a, __m128 b) +{ + uint32x4_t a_lt_b = + vcltq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); + return vgetq_lane_u32(a_lt_b, 0) & 0x1; +} + +// Compares the lower single-precision floating point scalar values of a and b +// using an inequality operation. : +// https://msdn.microsoft.com/en-us/library/bafh5e0a(v=vs.90).aspx +FORCE_INLINE int _mm_comineq_ss(__m128 a, __m128 b) +{ + return !_mm_comieq_ss(a, b); +} + +// Convert packed signed 32-bit integers in b to packed single-precision +// (32-bit) floating-point elements, store the results in the lower 2 elements +// of dst, and copy the upper 2 packed elements from a to the upper elements of +// dst. +// +// dst[31:0] := Convert_Int32_To_FP32(b[31:0]) +// dst[63:32] := Convert_Int32_To_FP32(b[63:32]) +// dst[95:64] := a[95:64] +// dst[127:96] := a[127:96] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvt_pi2ps +FORCE_INLINE __m128 _mm_cvt_pi2ps(__m128 a, __m64 b) +{ + return vreinterpretq_m128_f32( + vcombine_f32(vcvt_f32_s32(vreinterpret_s32_m64(b)), + vget_high_f32(vreinterpretq_f32_m128(a)))); +} + +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed 32-bit integers, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// dst[i+31:i] := Convert_FP32_To_Int32(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvt_ps2pi +FORCE_INLINE __m64 _mm_cvt_ps2pi(__m128 a) +{ +#if defined(__aarch64__) + return vreinterpret_m64_s32( + vget_low_s32(vcvtnq_s32_f32(vrndiq_f32(vreinterpretq_f32_m128(a))))); +#else + return vreinterpret_m64_s32(vcvt_s32_f32(vget_low_f32( + vreinterpretq_f32_m128(_mm_round_ps(a, _MM_FROUND_CUR_DIRECTION))))); +#endif +} + +// Convert the signed 32-bit integer b to a single-precision (32-bit) +// floating-point element, store the result in the lower element of dst, and +// copy the upper 3 packed elements from a to the upper elements of dst. +// +// dst[31:0] := Convert_Int32_To_FP32(b[31:0]) +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvt_si2ss +FORCE_INLINE __m128 _mm_cvt_si2ss(__m128 a, int b) +{ + return vreinterpretq_m128_f32( + vsetq_lane_f32((float)b, vreinterpretq_f32_m128(a), 0)); +} + +// Convert the lower single-precision (32-bit) floating-point element in a to a +// 32-bit integer, and store the result in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvt_ss2si +FORCE_INLINE int _mm_cvt_ss2si(__m128 a) +{ +#if defined(__aarch64__) + return vgetq_lane_s32(vcvtnq_s32_f32(vrndiq_f32(vreinterpretq_f32_m128(a))), + 0); +#else + float32_t data = vgetq_lane_f32( + vreinterpretq_f32_m128(_mm_round_ps(a, _MM_FROUND_CUR_DIRECTION)), 0); + return (int32_t)data; +#endif +} + +// Convert packed 16-bit integers in a to packed single-precision (32-bit) +// floating-point elements, and store the results in dst. +// +// FOR j := 0 to 3 +// i := j*16 +// m := j*32 +// dst[m+31:m] := Convert_Int16_To_FP32(a[i+15:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi16_ps +FORCE_INLINE __m128 _mm_cvtpi16_ps(__m64 a) +{ + return vreinterpretq_m128_f32( + vcvtq_f32_s32(vmovl_s16(vreinterpret_s16_m64(a)))); +} + +// Convert packed 32-bit integers in b to packed single-precision (32-bit) +// floating-point elements, store the results in the lower 2 elements of dst, +// and copy the upper 2 packed elements from a to the upper elements of dst. +// +// dst[31:0] := Convert_Int32_To_FP32(b[31:0]) +// dst[63:32] := Convert_Int32_To_FP32(b[63:32]) +// dst[95:64] := a[95:64] +// dst[127:96] := a[127:96] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi32_ps +FORCE_INLINE __m128 _mm_cvtpi32_ps(__m128 a, __m64 b) +{ + return vreinterpretq_m128_f32( + vcombine_f32(vcvt_f32_s32(vreinterpret_s32_m64(b)), + vget_high_f32(vreinterpretq_f32_m128(a)))); +} + +// Convert packed signed 32-bit integers in a to packed single-precision +// (32-bit) floating-point elements, store the results in the lower 2 elements +// of dst, then covert the packed signed 32-bit integers in b to +// single-precision (32-bit) floating-point element, and store the results in +// the upper 2 elements of dst. +// +// dst[31:0] := Convert_Int32_To_FP32(a[31:0]) +// dst[63:32] := Convert_Int32_To_FP32(a[63:32]) +// dst[95:64] := Convert_Int32_To_FP32(b[31:0]) +// dst[127:96] := Convert_Int32_To_FP32(b[63:32]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi32x2_ps +FORCE_INLINE __m128 _mm_cvtpi32x2_ps(__m64 a, __m64 b) +{ + return vreinterpretq_m128_f32(vcvtq_f32_s32( + vcombine_s32(vreinterpret_s32_m64(a), vreinterpret_s32_m64(b)))); +} + +// Convert the lower packed 8-bit integers in a to packed single-precision +// (32-bit) floating-point elements, and store the results in dst. +// +// FOR j := 0 to 3 +// i := j*8 +// m := j*32 +// dst[m+31:m] := Convert_Int8_To_FP32(a[i+7:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi8_ps +FORCE_INLINE __m128 _mm_cvtpi8_ps(__m64 a) +{ + return vreinterpretq_m128_f32(vcvtq_f32_s32( + vmovl_s16(vget_low_s16(vmovl_s8(vreinterpret_s8_m64(a)))))); +} + +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed 16-bit integers, and store the results in dst. Note: this intrinsic +// will generate 0x7FFF, rather than 0x8000, for input values between 0x7FFF and +// 0x7FFFFFFF. +// +// FOR j := 0 to 3 +// i := 16*j +// k := 32*j +// IF a[k+31:k] >= FP32(0x7FFF) && a[k+31:k] <= FP32(0x7FFFFFFF) +// dst[i+15:i] := 0x7FFF +// ELSE +// dst[i+15:i] := Convert_FP32_To_Int16(a[k+31:k]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtps_pi16 +FORCE_INLINE __m64 _mm_cvtps_pi16(__m128 a) +{ + const __m128 i16Min = _mm_set_ps1((float)INT16_MIN); + const __m128 i16Max = _mm_set_ps1((float)INT16_MAX); + const __m128 i32Max = _mm_set_ps1((float)INT32_MAX); + const __m128i maxMask = _mm_castps_si128( + _mm_and_ps(_mm_cmpge_ps(a, i16Max), _mm_cmple_ps(a, i32Max))); + const __m128i betweenMask = _mm_castps_si128( + _mm_and_ps(_mm_cmpgt_ps(a, i16Min), _mm_cmplt_ps(a, i16Max))); + const __m128i minMask = _mm_cmpeq_epi32(_mm_or_si128(maxMask, betweenMask), + _mm_setzero_si128()); + __m128i max = _mm_and_si128(maxMask, _mm_set1_epi32(INT16_MAX)); + __m128i min = _mm_and_si128(minMask, _mm_set1_epi32(INT16_MIN)); + __m128i cvt = _mm_and_si128(betweenMask, _mm_cvtps_epi32(a)); + __m128i res32 = _mm_or_si128(_mm_or_si128(max, min), cvt); + return vreinterpret_m64_s16(vmovn_s32(vreinterpretq_s32_m128i(res32))); +} + +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed 32-bit integers, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// dst[i+31:i] := Convert_FP32_To_Int32(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtps_pi32 +#define _mm_cvtps_pi32(a) _mm_cvt_ps2pi(a) + +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed 8-bit integers, and store the results in lower 4 elements of dst. +// Note: this intrinsic will generate 0x7F, rather than 0x80, for input values +// between 0x7F and 0x7FFFFFFF. +// +// FOR j := 0 to 3 +// i := 8*j +// k := 32*j +// IF a[k+31:k] >= FP32(0x7F) && a[k+31:k] <= FP32(0x7FFFFFFF) +// dst[i+7:i] := 0x7F +// ELSE +// dst[i+7:i] := Convert_FP32_To_Int8(a[k+31:k]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtps_pi8 +FORCE_INLINE __m64 _mm_cvtps_pi8(__m128 a) +{ + const __m128 i8Min = _mm_set_ps1((float)INT8_MIN); + const __m128 i8Max = _mm_set_ps1((float)INT8_MAX); + const __m128 i32Max = _mm_set_ps1((float)INT32_MAX); + const __m128i maxMask = _mm_castps_si128( + _mm_and_ps(_mm_cmpge_ps(a, i8Max), _mm_cmple_ps(a, i32Max))); + const __m128i betweenMask = _mm_castps_si128( + _mm_and_ps(_mm_cmpgt_ps(a, i8Min), _mm_cmplt_ps(a, i8Max))); + const __m128i minMask = _mm_cmpeq_epi32(_mm_or_si128(maxMask, betweenMask), + _mm_setzero_si128()); + __m128i max = _mm_and_si128(maxMask, _mm_set1_epi32(INT8_MAX)); + __m128i min = _mm_and_si128(minMask, _mm_set1_epi32(INT8_MIN)); + __m128i cvt = _mm_and_si128(betweenMask, _mm_cvtps_epi32(a)); + __m128i res32 = _mm_or_si128(_mm_or_si128(max, min), cvt); + int16x4_t res16 = vmovn_s32(vreinterpretq_s32_m128i(res32)); + int8x8_t res8 = vmovn_s16(vcombine_s16(res16, res16)); + uint32_t bitMask[2] = { 0xFFFFFFFF, 0 }; + int8x8_t mask = vreinterpret_s8_u32(vld1_u32(bitMask)); + + return vreinterpret_m64_s8(vorr_s8(vand_s8(mask, res8), vdup_n_s8(0))); +} + +// Convert packed unsigned 16-bit integers in a to packed single-precision +// (32-bit) floating-point elements, and store the results in dst. +// +// FOR j := 0 to 3 +// i := j*16 +// m := j*32 +// dst[m+31:m] := Convert_UInt16_To_FP32(a[i+15:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpu16_ps +FORCE_INLINE __m128 _mm_cvtpu16_ps(__m64 a) +{ + return vreinterpretq_m128_f32( + vcvtq_f32_u32(vmovl_u16(vreinterpret_u16_m64(a)))); +} + +// Convert the lower packed unsigned 8-bit integers in a to packed +// single-precision (32-bit) floating-point elements, and store the results in +// dst. +// +// FOR j := 0 to 3 +// i := j*8 +// m := j*32 +// dst[m+31:m] := Convert_UInt8_To_FP32(a[i+7:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpu8_ps +FORCE_INLINE __m128 _mm_cvtpu8_ps(__m64 a) +{ + return vreinterpretq_m128_f32(vcvtq_f32_u32( + vmovl_u16(vget_low_u16(vmovl_u8(vreinterpret_u8_m64(a)))))); +} + +// Convert the signed 32-bit integer b to a single-precision (32-bit) +// floating-point element, store the result in the lower element of dst, and +// copy the upper 3 packed elements from a to the upper elements of dst. +// +// dst[31:0] := Convert_Int32_To_FP32(b[31:0]) +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi32_ss +#define _mm_cvtsi32_ss(a, b) _mm_cvt_si2ss(a, b) + +// Convert the signed 64-bit integer b to a single-precision (32-bit) +// floating-point element, store the result in the lower element of dst, and +// copy the upper 3 packed elements from a to the upper elements of dst. +// +// dst[31:0] := Convert_Int64_To_FP32(b[63:0]) +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi64_ss +FORCE_INLINE __m128 _mm_cvtsi64_ss(__m128 a, int64_t b) +{ + return vreinterpretq_m128_f32( + vsetq_lane_f32((float)b, vreinterpretq_f32_m128(a), 0)); +} + +// Copy the lower single-precision (32-bit) floating-point element of a to dst. +// +// dst[31:0] := a[31:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtss_f32 +FORCE_INLINE float _mm_cvtss_f32(__m128 a) +{ + return vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); +} + +// Convert the lower single-precision (32-bit) floating-point element in a to a +// 32-bit integer, and store the result in dst. +// +// dst[31:0] := Convert_FP32_To_Int32(a[31:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtss_si32 +#define _mm_cvtss_si32(a) _mm_cvt_ss2si(a) + +// Convert the lower single-precision (32-bit) floating-point element in a to a +// 64-bit integer, and store the result in dst. +// +// dst[63:0] := Convert_FP32_To_Int64(a[31:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtss_si64 +FORCE_INLINE int64_t _mm_cvtss_si64(__m128 a) +{ +#if defined(__aarch64__) + return (int64_t)vgetq_lane_f32(vrndiq_f32(vreinterpretq_f32_m128(a)), 0); +#else + float32_t data = vgetq_lane_f32( + vreinterpretq_f32_m128(_mm_round_ps(a, _MM_FROUND_CUR_DIRECTION)), 0); + return (int64_t)data; +#endif +} + +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed 32-bit integers with truncation, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// dst[i+31:i] := Convert_FP32_To_Int32_Truncate(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtt_ps2pi +FORCE_INLINE __m64 _mm_cvtt_ps2pi(__m128 a) +{ + return vreinterpret_m64_s32( + vget_low_s32(vcvtq_s32_f32(vreinterpretq_f32_m128(a)))); +} + +// Convert the lower single-precision (32-bit) floating-point element in a to a +// 32-bit integer with truncation, and store the result in dst. +// +// dst[31:0] := Convert_FP32_To_Int32_Truncate(a[31:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtt_ss2si +FORCE_INLINE int _mm_cvtt_ss2si(__m128 a) +{ + return vgetq_lane_s32(vcvtq_s32_f32(vreinterpretq_f32_m128(a)), 0); +} + +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed 32-bit integers with truncation, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// dst[i+31:i] := Convert_FP32_To_Int32_Truncate(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttps_pi32 +#define _mm_cvttps_pi32(a) _mm_cvtt_ps2pi(a) + +// Convert the lower single-precision (32-bit) floating-point element in a to a +// 32-bit integer with truncation, and store the result in dst. +// +// dst[31:0] := Convert_FP32_To_Int32_Truncate(a[31:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttss_si32 +#define _mm_cvttss_si32(a) _mm_cvtt_ss2si(a) + +// Convert the lower single-precision (32-bit) floating-point element in a to a +// 64-bit integer with truncation, and store the result in dst. +// +// dst[63:0] := Convert_FP32_To_Int64_Truncate(a[31:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttss_si64 +FORCE_INLINE int64_t _mm_cvttss_si64(__m128 a) +{ + return (int64_t)vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); +} + +// Divides the four single-precision, floating-point values of a and b. +// +// r0 := a0 / b0 +// r1 := a1 / b1 +// r2 := a2 / b2 +// r3 := a3 / b3 +// +// https://msdn.microsoft.com/en-us/library/edaw8147(v=vs.100).aspx +FORCE_INLINE __m128 _mm_div_ps(__m128 a, __m128 b) +{ +#if defined(__aarch64__) && !SSE2NEON_PRECISE_DIV + return vreinterpretq_m128_f32( + vdivq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +#else + float32x4_t recip = vrecpeq_f32(vreinterpretq_f32_m128(b)); + recip = vmulq_f32(recip, vrecpsq_f32(recip, vreinterpretq_f32_m128(b))); +#if SSE2NEON_PRECISE_DIV + // Additional Netwon-Raphson iteration for accuracy + recip = vmulq_f32(recip, vrecpsq_f32(recip, vreinterpretq_f32_m128(b))); +#endif + return vreinterpretq_m128_f32(vmulq_f32(vreinterpretq_f32_m128(a), recip)); +#endif +} + +// Divides the scalar single-precision floating point value of a by b. +// https://msdn.microsoft.com/en-us/library/4y73xa49(v=vs.100).aspx +FORCE_INLINE __m128 _mm_div_ss(__m128 a, __m128 b) +{ + float32_t value = + vgetq_lane_f32(vreinterpretq_f32_m128(_mm_div_ps(a, b)), 0); + return vreinterpretq_m128_f32( + vsetq_lane_f32(value, vreinterpretq_f32_m128(a), 0)); +} + +// Extract a 16-bit integer from a, selected with imm8, and store the result in +// the lower element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_extract_pi16 +#define _mm_extract_pi16(a, imm) \ + (int32_t)vget_lane_u16(vreinterpret_u16_m64(a), (imm)) + +// Free aligned memory that was allocated with _mm_malloc. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_free +FORCE_INLINE void _mm_free(void *addr) +{ + free(addr); +} + +// Macro: Get the flush zero bits from the MXCSR control and status register. +// The flush zero may contain any of the following flags: _MM_FLUSH_ZERO_ON or +// _MM_FLUSH_ZERO_OFF +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_MM_GET_FLUSH_ZERO_MODE +FORCE_INLINE unsigned int _sse2neon_mm_get_flush_zero_mode() +{ + union { + fpcr_bitfield field; +#if defined(__aarch64__) + uint64_t value; +#else + uint32_t value; +#endif + } r; + +#if defined(__aarch64__) + asm volatile("mrs %0, FPCR" : "=r"(r.value)); /* read */ +#else + asm volatile("vmrs %0, FPSCR" : "=r"(r.value)); /* read */ +#endif + + return r.field.bit24 ? _MM_FLUSH_ZERO_ON : _MM_FLUSH_ZERO_OFF; +} + +// Macro: Get the rounding mode bits from the MXCSR control and status register. +// The rounding mode may contain any of the following flags: _MM_ROUND_NEAREST, +// _MM_ROUND_DOWN, _MM_ROUND_UP, _MM_ROUND_TOWARD_ZERO +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_MM_GET_ROUNDING_MODE +FORCE_INLINE unsigned int _MM_GET_ROUNDING_MODE() +{ + union { + fpcr_bitfield field; +#if defined(__aarch64__) + uint64_t value; +#else + uint32_t value; +#endif + } r; + +#if defined(__aarch64__) + asm volatile("mrs %0, FPCR" : "=r"(r.value)); /* read */ +#else + asm volatile("vmrs %0, FPSCR" : "=r"(r.value)); /* read */ +#endif + + if (r.field.bit22) { + return r.field.bit23 ? _MM_ROUND_TOWARD_ZERO : _MM_ROUND_UP; + } else { + return r.field.bit23 ? _MM_ROUND_DOWN : _MM_ROUND_NEAREST; + } +} + +// Copy a to dst, and insert the 16-bit integer i into dst at the location +// specified by imm8. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_insert_pi16 +#define _mm_insert_pi16(a, b, imm) \ + __extension__({ \ + vreinterpret_m64_s16( \ + vset_lane_s16((b), vreinterpret_s16_m64(a), (imm))); \ + }) + +// Loads four single-precision, floating-point values. +// https://msdn.microsoft.com/en-us/library/vstudio/zzd50xxt(v=vs.100).aspx +FORCE_INLINE __m128 _mm_load_ps(const float *p) +{ + return vreinterpretq_m128_f32(vld1q_f32(p)); +} + +// Load a single-precision (32-bit) floating-point element from memory into all +// elements of dst. +// +// dst[31:0] := MEM[mem_addr+31:mem_addr] +// dst[63:32] := MEM[mem_addr+31:mem_addr] +// dst[95:64] := MEM[mem_addr+31:mem_addr] +// dst[127:96] := MEM[mem_addr+31:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load_ps1 +#define _mm_load_ps1 _mm_load1_ps + +// Loads an single - precision, floating - point value into the low word and +// clears the upper three words. +// https://msdn.microsoft.com/en-us/library/548bb9h4%28v=vs.90%29.aspx +FORCE_INLINE __m128 _mm_load_ss(const float *p) +{ + return vreinterpretq_m128_f32(vsetq_lane_f32(*p, vdupq_n_f32(0), 0)); +} + +// Loads a single single-precision, floating-point value, copying it into all +// four words +// https://msdn.microsoft.com/en-us/library/vstudio/5cdkf716(v=vs.100).aspx +FORCE_INLINE __m128 _mm_load1_ps(const float *p) +{ + return vreinterpretq_m128_f32(vld1q_dup_f32(p)); +} + +// Sets the upper two single-precision, floating-point values with 64 +// bits of data loaded from the address p; the lower two values are passed +// through from a. +// +// r0 := a0 +// r1 := a1 +// r2 := *p0 +// r3 := *p1 +// +// https://msdn.microsoft.com/en-us/library/w92wta0x(v%3dvs.100).aspx +FORCE_INLINE __m128 _mm_loadh_pi(__m128 a, __m64 const *p) +{ + return vreinterpretq_m128_f32( + vcombine_f32(vget_low_f32(a), vld1_f32((const float32_t *)p))); +} + +// Sets the lower two single-precision, floating-point values with 64 +// bits of data loaded from the address p; the upper two values are passed +// through from a. +// +// Return Value +// r0 := *p0 +// r1 := *p1 +// r2 := a2 +// r3 := a3 +// +// https://msdn.microsoft.com/en-us/library/s57cyak2(v=vs.100).aspx +FORCE_INLINE __m128 _mm_loadl_pi(__m128 a, __m64 const *p) +{ + return vreinterpretq_m128_f32( + vcombine_f32(vld1_f32((const float32_t *)p), vget_high_f32(a))); +} + +// Load 4 single-precision (32-bit) floating-point elements from memory into dst +// in reverse order. mem_addr must be aligned on a 16-byte boundary or a +// general-protection exception may be generated. +// +// dst[31:0] := MEM[mem_addr+127:mem_addr+96] +// dst[63:32] := MEM[mem_addr+95:mem_addr+64] +// dst[95:64] := MEM[mem_addr+63:mem_addr+32] +// dst[127:96] := MEM[mem_addr+31:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadr_ps +FORCE_INLINE __m128 _mm_loadr_ps(const float *p) +{ + float32x4_t v = vrev64q_f32(vld1q_f32(p)); + return vreinterpretq_m128_f32(vextq_f32(v, v, 2)); +} + +// Loads four single-precision, floating-point values. +// https://msdn.microsoft.com/en-us/library/x1b16s7z%28v=vs.90%29.aspx +FORCE_INLINE __m128 _mm_loadu_ps(const float *p) +{ + // for neon, alignment doesn't matter, so _mm_load_ps and _mm_loadu_ps are + // equivalent for neon + return vreinterpretq_m128_f32(vld1q_f32(p)); +} + +// Load unaligned 16-bit integer from memory into the first element of dst. +// +// dst[15:0] := MEM[mem_addr+15:mem_addr] +// dst[MAX:16] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadu_si16 +FORCE_INLINE __m128i _mm_loadu_si16(const void *p) +{ + return vreinterpretq_m128i_s16( + vsetq_lane_s16(*(const int16_t *)p, vdupq_n_s16(0), 0)); +} + +// Load unaligned 64-bit integer from memory into the first element of dst. +// +// dst[63:0] := MEM[mem_addr+63:mem_addr] +// dst[MAX:64] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadu_si64 +FORCE_INLINE __m128i _mm_loadu_si64(const void *p) +{ + return vreinterpretq_m128i_s64( + vcombine_s64(vld1_s64((const int64_t *)p), vdup_n_s64(0))); +} + +// Allocate aligned blocks of memory. +// https://software.intel.com/en-us/ +// cpp-compiler-developer-guide-and-reference-allocating-and-freeing-aligned-memory-blocks +FORCE_INLINE void *_mm_malloc(size_t size, size_t align) +{ + void *ptr; + if (align == 1) + return malloc(size); + if (align == 2 || (sizeof(void *) == 8 && align == 4)) + align = sizeof(void *); + if (!posix_memalign(&ptr, align, size)) + return ptr; + return NULL; +} + +// Conditionally store 8-bit integer elements from a into memory using mask +// (elements are not stored when the highest bit is not set in the corresponding +// element) and a non-temporal memory hint. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_maskmove_si64 +FORCE_INLINE void _mm_maskmove_si64(__m64 a, __m64 mask, char *mem_addr) +{ + int8x8_t shr_mask = vshr_n_s8(vreinterpret_s8_m64(mask), 7); + __m128 b = _mm_load_ps((const float *)mem_addr); + int8x8_t masked = + vbsl_s8(vreinterpret_u8_s8(shr_mask), vreinterpret_s8_m64(a), + vreinterpret_s8_u64(vget_low_u64(vreinterpretq_u64_m128(b)))); + vst1_s8((int8_t *)mem_addr, masked); +} + +// Conditionally store 8-bit integer elements from a into memory using mask +// (elements are not stored when the highest bit is not set in the corresponding +// element) and a non-temporal memory hint. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_maskmovq +#define _m_maskmovq(a, mask, mem_addr) _mm_maskmove_si64(a, mask, mem_addr) + +// Compare packed signed 16-bit integers in a and b, and store packed maximum +// values in dst. +// +// FOR j := 0 to 3 +// i := j*16 +// dst[i+15:i] := MAX(a[i+15:i], b[i+15:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_pi16 +FORCE_INLINE __m64 _mm_max_pi16(__m64 a, __m64 b) +{ + return vreinterpret_m64_s16( + vmax_s16(vreinterpret_s16_m64(a), vreinterpret_s16_m64(b))); +} + +// Computes the maximums of the four single-precision, floating-point values of +// a and b. +// https://msdn.microsoft.com/en-us/library/vstudio/ff5d607a(v=vs.100).aspx +FORCE_INLINE __m128 _mm_max_ps(__m128 a, __m128 b) +{ +#if SSE2NEON_PRECISE_MINMAX + float32x4_t _a = vreinterpretq_f32_m128(a); + float32x4_t _b = vreinterpretq_f32_m128(b); + return vreinterpretq_m128_f32(vbslq_f32(vcgtq_f32(_a, _b), _a, _b)); +#else + return vreinterpretq_m128_f32( + vmaxq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +#endif +} + +// Compare packed unsigned 8-bit integers in a and b, and store packed maximum +// values in dst. +// +// FOR j := 0 to 7 +// i := j*8 +// dst[i+7:i] := MAX(a[i+7:i], b[i+7:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_pu8 +FORCE_INLINE __m64 _mm_max_pu8(__m64 a, __m64 b) +{ + return vreinterpret_m64_u8( + vmax_u8(vreinterpret_u8_m64(a), vreinterpret_u8_m64(b))); +} + +// Computes the maximum of the two lower scalar single-precision floating point +// values of a and b. +// https://msdn.microsoft.com/en-us/library/s6db5esz(v=vs.100).aspx +FORCE_INLINE __m128 _mm_max_ss(__m128 a, __m128 b) +{ + float32_t value = vgetq_lane_f32(_mm_max_ps(a, b), 0); + return vreinterpretq_m128_f32( + vsetq_lane_f32(value, vreinterpretq_f32_m128(a), 0)); +} + +// Compare packed signed 16-bit integers in a and b, and store packed minimum +// values in dst. +// +// FOR j := 0 to 3 +// i := j*16 +// dst[i+15:i] := MIN(a[i+15:i], b[i+15:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_pi16 +FORCE_INLINE __m64 _mm_min_pi16(__m64 a, __m64 b) +{ + return vreinterpret_m64_s16( + vmin_s16(vreinterpret_s16_m64(a), vreinterpret_s16_m64(b))); +} + +// Computes the minima of the four single-precision, floating-point values of a +// and b. +// https://msdn.microsoft.com/en-us/library/vstudio/wh13kadz(v=vs.100).aspx +FORCE_INLINE __m128 _mm_min_ps(__m128 a, __m128 b) +{ +#if SSE2NEON_PRECISE_MINMAX + float32x4_t _a = vreinterpretq_f32_m128(a); + float32x4_t _b = vreinterpretq_f32_m128(b); + return vreinterpretq_m128_f32(vbslq_f32(vcltq_f32(_a, _b), _a, _b)); +#else + return vreinterpretq_m128_f32( + vminq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +#endif +} + +// Compare packed unsigned 8-bit integers in a and b, and store packed minimum +// values in dst. +// +// FOR j := 0 to 7 +// i := j*8 +// dst[i+7:i] := MIN(a[i+7:i], b[i+7:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_pu8 +FORCE_INLINE __m64 _mm_min_pu8(__m64 a, __m64 b) +{ + return vreinterpret_m64_u8( + vmin_u8(vreinterpret_u8_m64(a), vreinterpret_u8_m64(b))); +} + +// Computes the minimum of the two lower scalar single-precision floating point +// values of a and b. +// https://msdn.microsoft.com/en-us/library/0a9y7xaa(v=vs.100).aspx +FORCE_INLINE __m128 _mm_min_ss(__m128 a, __m128 b) +{ + float32_t value = vgetq_lane_f32(_mm_min_ps(a, b), 0); + return vreinterpretq_m128_f32( + vsetq_lane_f32(value, vreinterpretq_f32_m128(a), 0)); +} + +// Sets the low word to the single-precision, floating-point value of b +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/35hdzazd(v=vs.100) +FORCE_INLINE __m128 _mm_move_ss(__m128 a, __m128 b) +{ + return vreinterpretq_m128_f32( + vsetq_lane_f32(vgetq_lane_f32(vreinterpretq_f32_m128(b), 0), + vreinterpretq_f32_m128(a), 0)); +} + +// Moves the upper two values of B into the lower two values of A. +// +// r3 := a3 +// r2 := a2 +// r1 := b3 +// r0 := b2 +FORCE_INLINE __m128 _mm_movehl_ps(__m128 __A, __m128 __B) +{ + float32x2_t a32 = vget_high_f32(vreinterpretq_f32_m128(__A)); + float32x2_t b32 = vget_high_f32(vreinterpretq_f32_m128(__B)); + return vreinterpretq_m128_f32(vcombine_f32(b32, a32)); +} + +// Moves the lower two values of B into the upper two values of A. +// +// r3 := b1 +// r2 := b0 +// r1 := a1 +// r0 := a0 +FORCE_INLINE __m128 _mm_movelh_ps(__m128 __A, __m128 __B) +{ + float32x2_t a10 = vget_low_f32(vreinterpretq_f32_m128(__A)); + float32x2_t b10 = vget_low_f32(vreinterpretq_f32_m128(__B)); + return vreinterpretq_m128_f32(vcombine_f32(a10, b10)); +} + +// Create mask from the most significant bit of each 8-bit element in a, and +// store the result in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_movemask_pi8 +FORCE_INLINE int _mm_movemask_pi8(__m64 a) +{ + uint8x8_t input = vreinterpret_u8_m64(a); +#if defined(__aarch64__) + static const int8x8_t shift = { 0, 1, 2, 3, 4, 5, 6, 7 }; + uint8x8_t tmp = vshr_n_u8(input, 7); + return vaddv_u8(vshl_u8(tmp, shift)); +#else + // Refer the implementation of `_mm_movemask_epi8` + uint16x4_t high_bits = vreinterpret_u16_u8(vshr_n_u8(input, 7)); + uint32x2_t paired16 = + vreinterpret_u32_u16(vsra_n_u16(high_bits, high_bits, 7)); + uint8x8_t paired32 = + vreinterpret_u8_u32(vsra_n_u32(paired16, paired16, 14)); + return vget_lane_u8(paired32, 0) | ((int)vget_lane_u8(paired32, 4) << 4); +#endif +} + +// NEON does not provide this method +// Creates a 4-bit mask from the most significant bits of the four +// single-precision, floating-point values. +// https://msdn.microsoft.com/en-us/library/vstudio/4490ys29(v=vs.100).aspx +FORCE_INLINE int _mm_movemask_ps(__m128 a) +{ + uint32x4_t input = vreinterpretq_u32_m128(a); +#if defined(__aarch64__) + static const int32x4_t shift = { 0, 1, 2, 3 }; + uint32x4_t tmp = vshrq_n_u32(input, 31); + return vaddvq_u32(vshlq_u32(tmp, shift)); +#else + // Uses the exact same method as _mm_movemask_epi8, see that for details. + // Shift out everything but the sign bits with a 32-bit unsigned shift + // right. + uint64x2_t high_bits = vreinterpretq_u64_u32(vshrq_n_u32(input, 31)); + // Merge the two pairs together with a 64-bit unsigned shift right + add. + uint8x16_t paired = + vreinterpretq_u8_u64(vsraq_n_u64(high_bits, high_bits, 31)); + // Extract the result. + return vgetq_lane_u8(paired, 0) | (vgetq_lane_u8(paired, 8) << 2); +#endif +} + +// Multiplies the four single-precision, floating-point values of a and b. +// +// r0 := a0 * b0 +// r1 := a1 * b1 +// r2 := a2 * b2 +// r3 := a3 * b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/22kbk6t9(v=vs.100).aspx +FORCE_INLINE __m128 _mm_mul_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_f32( + vmulq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Multiply the lower single-precision (32-bit) floating-point element in a and +// b, store the result in the lower element of dst, and copy the upper 3 packed +// elements from a to the upper elements of dst. +// +// dst[31:0] := a[31:0] * b[31:0] +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mul_ss +FORCE_INLINE __m128 _mm_mul_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_mul_ps(a, b)); +} + +// Multiply the packed unsigned 16-bit integers in a and b, producing +// intermediate 32-bit integers, and store the high 16 bits of the intermediate +// integers in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mulhi_pu16 +FORCE_INLINE __m64 _mm_mulhi_pu16(__m64 a, __m64 b) +{ + return vreinterpret_m64_u16(vshrn_n_u32( + vmull_u16(vreinterpret_u16_m64(a), vreinterpret_u16_m64(b)), 16)); +} + +// Computes the bitwise OR of the four single-precision, floating-point values +// of a and b. +// https://msdn.microsoft.com/en-us/library/vstudio/7ctdsyy0(v=vs.100).aspx +FORCE_INLINE __m128 _mm_or_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_s32( + vorrq_s32(vreinterpretq_s32_m128(a), vreinterpretq_s32_m128(b))); +} + +// Average packed unsigned 8-bit integers in a and b, and store the results in +// dst. +// +// FOR j := 0 to 7 +// i := j*8 +// dst[i+7:i] := (a[i+7:i] + b[i+7:i] + 1) >> 1 +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pavgb +#define _m_pavgb(a, b) _mm_avg_pu8(a, b) + +// Average packed unsigned 16-bit integers in a and b, and store the results in +// dst. +// +// FOR j := 0 to 3 +// i := j*16 +// dst[i+15:i] := (a[i+15:i] + b[i+15:i] + 1) >> 1 +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pavgw +#define _m_pavgw(a, b) _mm_avg_pu16(a, b) + +// Extract a 16-bit integer from a, selected with imm8, and store the result in +// the lower element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pextrw +#define _m_pextrw(a, imm) _mm_extract_pi16(a, imm) + +// Copy a to dst, and insert the 16-bit integer i into dst at the location +// specified by imm8. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=m_pinsrw +#define _m_pinsrw(a, i, imm) _mm_insert_pi16(a, i, imm) + +// Compare packed signed 16-bit integers in a and b, and store packed maximum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pmaxsw +#define _m_pmaxsw(a, b) _mm_max_pi16(a, b) + +// Compare packed unsigned 8-bit integers in a and b, and store packed maximum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pmaxub +#define _m_pmaxub(a, b) _mm_max_pu8(a, b) + +// Compare packed signed 16-bit integers in a and b, and store packed minimum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pminsw +#define _m_pminsw(a, b) _mm_min_pi16(a, b) + +// Compare packed unsigned 8-bit integers in a and b, and store packed minimum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pminub +#define _m_pminub(a, b) _mm_min_pu8(a, b) + +// Create mask from the most significant bit of each 8-bit element in a, and +// store the result in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pmovmskb +#define _m_pmovmskb(a) _mm_movemask_pi8(a) + +// Multiply the packed unsigned 16-bit integers in a and b, producing +// intermediate 32-bit integers, and store the high 16 bits of the intermediate +// integers in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pmulhuw +#define _m_pmulhuw(a, b) _mm_mulhi_pu16(a, b) + +// Loads one cache line of data from address p to a location closer to the +// processor. https://msdn.microsoft.com/en-us/library/84szxsww(v=vs.100).aspx +FORCE_INLINE void _mm_prefetch(const void *p, int i) +{ + (void)i; + __builtin_prefetch(p); +} + +// Compute the absolute differences of packed unsigned 8-bit integers in a and +// b, then horizontally sum each consecutive 8 differences to produce four +// unsigned 16-bit integers, and pack these unsigned 16-bit integers in the low +// 16 bits of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=m_psadbw +#define _m_psadbw(a, b) _mm_sad_pu8(a, b) + +// Shuffle 16-bit integers in a using the control in imm8, and store the results +// in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pshufw +#define _m_pshufw(a, imm) _mm_shuffle_pi16(a, imm) + +// Compute the approximate reciprocal of packed single-precision (32-bit) +// floating-point elements in a, and store the results in dst. The maximum +// relative error for this approximation is less than 1.5*2^-12. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_rcp_ps +FORCE_INLINE __m128 _mm_rcp_ps(__m128 in) +{ + float32x4_t recip = vrecpeq_f32(vreinterpretq_f32_m128(in)); + recip = vmulq_f32(recip, vrecpsq_f32(recip, vreinterpretq_f32_m128(in))); +#if SSE2NEON_PRECISE_DIV + // Additional Netwon-Raphson iteration for accuracy + recip = vmulq_f32(recip, vrecpsq_f32(recip, vreinterpretq_f32_m128(in))); +#endif + return vreinterpretq_m128_f32(recip); +} + +// Compute the approximate reciprocal of the lower single-precision (32-bit) +// floating-point element in a, store the result in the lower element of dst, +// and copy the upper 3 packed elements from a to the upper elements of dst. The +// maximum relative error for this approximation is less than 1.5*2^-12. +// +// dst[31:0] := (1.0 / a[31:0]) +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_rcp_ss +FORCE_INLINE __m128 _mm_rcp_ss(__m128 a) +{ + return _mm_move_ss(a, _mm_rcp_ps(a)); +} + +// Computes the approximations of the reciprocal square roots of the four +// single-precision floating point values of in. +// The current precision is 1% error. +// https://msdn.microsoft.com/en-us/library/22hfsh53(v=vs.100).aspx +FORCE_INLINE __m128 _mm_rsqrt_ps(__m128 in) +{ + float32x4_t out = vrsqrteq_f32(vreinterpretq_f32_m128(in)); +#if SSE2NEON_PRECISE_SQRT + // Additional Netwon-Raphson iteration for accuracy + out = vmulq_f32( + out, vrsqrtsq_f32(vmulq_f32(vreinterpretq_f32_m128(in), out), out)); + out = vmulq_f32( + out, vrsqrtsq_f32(vmulq_f32(vreinterpretq_f32_m128(in), out), out)); +#endif + return vreinterpretq_m128_f32(out); +} + +// Compute the approximate reciprocal square root of the lower single-precision +// (32-bit) floating-point element in a, store the result in the lower element +// of dst, and copy the upper 3 packed elements from a to the upper elements of +// dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_rsqrt_ss +FORCE_INLINE __m128 _mm_rsqrt_ss(__m128 in) +{ + return vsetq_lane_f32(vgetq_lane_f32(_mm_rsqrt_ps(in), 0), in, 0); +} + +// Compute the absolute differences of packed unsigned 8-bit integers in a and +// b, then horizontally sum each consecutive 8 differences to produce four +// unsigned 16-bit integers, and pack these unsigned 16-bit integers in the low +// 16 bits of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sad_pu8 +FORCE_INLINE __m64 _mm_sad_pu8(__m64 a, __m64 b) +{ + uint64x1_t t = vpaddl_u32(vpaddl_u16( + vpaddl_u8(vabd_u8(vreinterpret_u8_m64(a), vreinterpret_u8_m64(b))))); + return vreinterpret_m64_u16( + vset_lane_u16(vget_lane_u64(t, 0), vdup_n_u16(0), 0)); +} + +// Macro: Set the flush zero bits of the MXCSR control and status register to +// the value in unsigned 32-bit integer a. The flush zero may contain any of the +// following flags: _MM_FLUSH_ZERO_ON or _MM_FLUSH_ZERO_OFF +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_MM_SET_FLUSH_ZERO_MODE +FORCE_INLINE void _sse2neon_mm_set_flush_zero_mode(unsigned int flag) +{ + // AArch32 Advanced SIMD arithmetic always uses the Flush-to-zero setting, + // regardless of the value of the FZ bit. + union { + fpcr_bitfield field; +#if defined(__aarch64__) + uint64_t value; +#else + uint32_t value; +#endif + } r; + +#if defined(__aarch64__) + asm volatile("mrs %0, FPCR" : "=r"(r.value)); /* read */ +#else + asm volatile("vmrs %0, FPSCR" : "=r"(r.value)); /* read */ +#endif + + r.field.bit24 = (flag & _MM_FLUSH_ZERO_MASK) == _MM_FLUSH_ZERO_ON; + +#if defined(__aarch64__) + asm volatile("msr FPCR, %0" ::"r"(r)); /* write */ +#else + asm volatile("vmsr FPSCR, %0" ::"r"(r)); /* write */ +#endif +} + +// Sets the four single-precision, floating-point values to the four inputs. +// https://msdn.microsoft.com/en-us/library/vstudio/afh0zf75(v=vs.100).aspx +FORCE_INLINE __m128 _mm_set_ps(float w, float z, float y, float x) +{ + float ALIGN_STRUCT(16) data[4] = { x, y, z, w }; + return vreinterpretq_m128_f32(vld1q_f32(data)); +} + +// Sets the four single-precision, floating-point values to w. +// https://msdn.microsoft.com/en-us/library/vstudio/2x1se8ha(v=vs.100).aspx +FORCE_INLINE __m128 _mm_set_ps1(float _w) +{ + return vreinterpretq_m128_f32(vdupq_n_f32(_w)); +} + +// Macro: Set the rounding mode bits of the MXCSR control and status register to +// the value in unsigned 32-bit integer a. The rounding mode may contain any of +// the following flags: _MM_ROUND_NEAREST, _MM_ROUND_DOWN, _MM_ROUND_UP, +// _MM_ROUND_TOWARD_ZERO +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_MM_SET_ROUNDING_MODE +FORCE_INLINE void _MM_SET_ROUNDING_MODE(int rounding) +{ + union { + fpcr_bitfield field; +#if defined(__aarch64__) + uint64_t value; +#else + uint32_t value; +#endif + } r; + +#if defined(__aarch64__) + asm volatile("mrs %0, FPCR" : "=r"(r.value)); /* read */ +#else + asm volatile("vmrs %0, FPSCR" : "=r"(r.value)); /* read */ +#endif + + switch (rounding) { + case _MM_ROUND_TOWARD_ZERO: + r.field.bit22 = 1; + r.field.bit23 = 1; + break; + case _MM_ROUND_DOWN: + r.field.bit22 = 0; + r.field.bit23 = 1; + break; + case _MM_ROUND_UP: + r.field.bit22 = 1; + r.field.bit23 = 0; + break; + default: //_MM_ROUND_NEAREST + r.field.bit22 = 0; + r.field.bit23 = 0; + } + +#if defined(__aarch64__) + asm volatile("msr FPCR, %0" ::"r"(r)); /* write */ +#else + asm volatile("vmsr FPSCR, %0" ::"r"(r)); /* write */ +#endif +} + +// Copy single-precision (32-bit) floating-point element a to the lower element +// of dst, and zero the upper 3 elements. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set_ss +FORCE_INLINE __m128 _mm_set_ss(float a) +{ + float ALIGN_STRUCT(16) data[4] = { a, 0, 0, 0 }; + return vreinterpretq_m128_f32(vld1q_f32(data)); +} + +// Sets the four single-precision, floating-point values to w. +// +// r0 := r1 := r2 := r3 := w +// +// https://msdn.microsoft.com/en-us/library/vstudio/2x1se8ha(v=vs.100).aspx +FORCE_INLINE __m128 _mm_set1_ps(float _w) +{ + return vreinterpretq_m128_f32(vdupq_n_f32(_w)); +} + +// FIXME: _mm_setcsr() implementation supports changing the rounding mode only. +FORCE_INLINE void _mm_setcsr(unsigned int a) +{ + _MM_SET_ROUNDING_MODE(a); +} + +// FIXME: _mm_getcsr() implementation supports reading the rounding mode only. +FORCE_INLINE unsigned int _mm_getcsr() +{ + return _MM_GET_ROUNDING_MODE(); +} + +// Sets the four single-precision, floating-point values to the four inputs in +// reverse order. +// https://msdn.microsoft.com/en-us/library/vstudio/d2172ct3(v=vs.100).aspx +FORCE_INLINE __m128 _mm_setr_ps(float w, float z, float y, float x) +{ + float ALIGN_STRUCT(16) data[4] = { w, z, y, x }; + return vreinterpretq_m128_f32(vld1q_f32(data)); +} + +// Clears the four single-precision, floating-point values. +// https://msdn.microsoft.com/en-us/library/vstudio/tk1t2tbz(v=vs.100).aspx +FORCE_INLINE __m128 _mm_setzero_ps(void) +{ + return vreinterpretq_m128_f32(vdupq_n_f32(0)); +} + +// Shuffle 16-bit integers in a using the control in imm8, and store the results +// in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_shuffle_pi16 +#if __has_builtin(__builtin_shufflevector) +#define _mm_shuffle_pi16(a, imm) \ + __extension__({ \ + vreinterpret_m64_s16(__builtin_shufflevector( \ + vreinterpret_s16_m64(a), vreinterpret_s16_m64(a), (imm & 0x3), \ + ((imm >> 2) & 0x3), ((imm >> 4) & 0x3), ((imm >> 6) & 0x3))); \ + }) +#else +#define _mm_shuffle_pi16(a, imm) \ + __extension__({ \ + int16x4_t ret; \ + ret = \ + vmov_n_s16(vget_lane_s16(vreinterpret_s16_m64(a), (imm) & (0x3))); \ + ret = vset_lane_s16(vget_lane_s16(vreinterpret_s16_m64(a), \ + ((imm) >> 2) & 0x3), \ + ret, 1); \ + ret = vset_lane_s16(vget_lane_s16(vreinterpret_s16_m64(a), \ + ((imm) >> 4) & 0x3), \ + ret, 2); \ + ret = vset_lane_s16(vget_lane_s16(vreinterpret_s16_m64(a), \ + ((imm) >> 6) & 0x3), \ + ret, 3); \ + vreinterpret_m64_s16(ret); \ + }) +#endif + +// Guarantees that every preceding store is globally visible before any +// subsequent store. +// https://msdn.microsoft.com/en-us/library/5h2w73d1%28v=vs.90%29.aspx +FORCE_INLINE void _mm_sfence(void) +{ + __sync_synchronize(); +} + +// FORCE_INLINE __m128 _mm_shuffle_ps(__m128 a, __m128 b, __constrange(0,255) +// int imm) +#if __has_builtin(__builtin_shufflevector) +#define _mm_shuffle_ps(a, b, imm) \ + __extension__({ \ + float32x4_t _input1 = vreinterpretq_f32_m128(a); \ + float32x4_t _input2 = vreinterpretq_f32_m128(b); \ + float32x4_t _shuf = __builtin_shufflevector( \ + _input1, _input2, (imm) & (0x3), ((imm) >> 2) & 0x3, \ + (((imm) >> 4) & 0x3) + 4, (((imm) >> 6) & 0x3) + 4); \ + vreinterpretq_m128_f32(_shuf); \ + }) +#else // generic +#define _mm_shuffle_ps(a, b, imm) \ + __extension__({ \ + __m128 ret; \ + switch (imm) { \ + case _MM_SHUFFLE(1, 0, 3, 2): \ + ret = _mm_shuffle_ps_1032((a), (b)); \ + break; \ + case _MM_SHUFFLE(2, 3, 0, 1): \ + ret = _mm_shuffle_ps_2301((a), (b)); \ + break; \ + case _MM_SHUFFLE(0, 3, 2, 1): \ + ret = _mm_shuffle_ps_0321((a), (b)); \ + break; \ + case _MM_SHUFFLE(2, 1, 0, 3): \ + ret = _mm_shuffle_ps_2103((a), (b)); \ + break; \ + case _MM_SHUFFLE(1, 0, 1, 0): \ + ret = _mm_movelh_ps((a), (b)); \ + break; \ + case _MM_SHUFFLE(1, 0, 0, 1): \ + ret = _mm_shuffle_ps_1001((a), (b)); \ + break; \ + case _MM_SHUFFLE(0, 1, 0, 1): \ + ret = _mm_shuffle_ps_0101((a), (b)); \ + break; \ + case _MM_SHUFFLE(3, 2, 1, 0): \ + ret = _mm_shuffle_ps_3210((a), (b)); \ + break; \ + case _MM_SHUFFLE(0, 0, 1, 1): \ + ret = _mm_shuffle_ps_0011((a), (b)); \ + break; \ + case _MM_SHUFFLE(0, 0, 2, 2): \ + ret = _mm_shuffle_ps_0022((a), (b)); \ + break; \ + case _MM_SHUFFLE(2, 2, 0, 0): \ + ret = _mm_shuffle_ps_2200((a), (b)); \ + break; \ + case _MM_SHUFFLE(3, 2, 0, 2): \ + ret = _mm_shuffle_ps_3202((a), (b)); \ + break; \ + case _MM_SHUFFLE(3, 2, 3, 2): \ + ret = _mm_movehl_ps((b), (a)); \ + break; \ + case _MM_SHUFFLE(1, 1, 3, 3): \ + ret = _mm_shuffle_ps_1133((a), (b)); \ + break; \ + case _MM_SHUFFLE(2, 0, 1, 0): \ + ret = _mm_shuffle_ps_2010((a), (b)); \ + break; \ + case _MM_SHUFFLE(2, 0, 0, 1): \ + ret = _mm_shuffle_ps_2001((a), (b)); \ + break; \ + case _MM_SHUFFLE(2, 0, 3, 2): \ + ret = _mm_shuffle_ps_2032((a), (b)); \ + break; \ + default: \ + ret = _mm_shuffle_ps_default((a), (b), (imm)); \ + break; \ + } \ + ret; \ + }) +#endif + +// Computes the approximations of square roots of the four single-precision, +// floating-point values of a. First computes reciprocal square roots and then +// reciprocals of the four values. +// +// r0 := sqrt(a0) +// r1 := sqrt(a1) +// r2 := sqrt(a2) +// r3 := sqrt(a3) +// +// https://msdn.microsoft.com/en-us/library/vstudio/8z67bwwk(v=vs.100).aspx +FORCE_INLINE __m128 _mm_sqrt_ps(__m128 in) +{ +#if SSE2NEON_PRECISE_SQRT + float32x4_t recip = vrsqrteq_f32(vreinterpretq_f32_m128(in)); + + // Test for vrsqrteq_f32(0) -> positive infinity case. + // Change to zero, so that s * 1/sqrt(s) result is zero too. + const uint32x4_t pos_inf = vdupq_n_u32(0x7F800000); + const uint32x4_t div_by_zero = + vceqq_u32(pos_inf, vreinterpretq_u32_f32(recip)); + recip = vreinterpretq_f32_u32( + vandq_u32(vmvnq_u32(div_by_zero), vreinterpretq_u32_f32(recip))); + + // Additional Netwon-Raphson iteration for accuracy + recip = vmulq_f32(vrsqrtsq_f32(vmulq_f32(recip, recip), + vreinterpretq_f32_m128(in)), + recip); + recip = vmulq_f32(vrsqrtsq_f32(vmulq_f32(recip, recip), + vreinterpretq_f32_m128(in)), + recip); + + // sqrt(s) = s * 1/sqrt(s) + return vreinterpretq_m128_f32(vmulq_f32(vreinterpretq_f32_m128(in), recip)); +#elif defined(__aarch64__) + return vreinterpretq_m128_f32(vsqrtq_f32(vreinterpretq_f32_m128(in))); +#else + float32x4_t recipsq = vrsqrteq_f32(vreinterpretq_f32_m128(in)); + float32x4_t sq = vrecpeq_f32(recipsq); + return vreinterpretq_m128_f32(sq); +#endif +} + +// Computes the approximation of the square root of the scalar single-precision +// floating point value of in. +// https://msdn.microsoft.com/en-us/library/ahfsc22d(v=vs.100).aspx +FORCE_INLINE __m128 _mm_sqrt_ss(__m128 in) +{ + float32_t value = + vgetq_lane_f32(vreinterpretq_f32_m128(_mm_sqrt_ps(in)), 0); + return vreinterpretq_m128_f32( + vsetq_lane_f32(value, vreinterpretq_f32_m128(in), 0)); +} + +// Stores four single-precision, floating-point values. +// https://msdn.microsoft.com/en-us/library/vstudio/s3h4ay6y(v=vs.100).aspx +FORCE_INLINE void _mm_store_ps(float *p, __m128 a) +{ + vst1q_f32(p, vreinterpretq_f32_m128(a)); +} + +// Store the lower single-precision (32-bit) floating-point element from a into +// 4 contiguous elements in memory. mem_addr must be aligned on a 16-byte +// boundary or a general-protection exception may be generated. +// +// MEM[mem_addr+31:mem_addr] := a[31:0] +// MEM[mem_addr+63:mem_addr+32] := a[31:0] +// MEM[mem_addr+95:mem_addr+64] := a[31:0] +// MEM[mem_addr+127:mem_addr+96] := a[31:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_store_ps1 +FORCE_INLINE void _mm_store_ps1(float *p, __m128 a) +{ + float32_t a0 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); + vst1q_f32(p, vdupq_n_f32(a0)); +} + +// Stores the lower single - precision, floating - point value. +// https://msdn.microsoft.com/en-us/library/tzz10fbx(v=vs.100).aspx +FORCE_INLINE void _mm_store_ss(float *p, __m128 a) +{ + vst1q_lane_f32(p, vreinterpretq_f32_m128(a), 0); +} + +// Store the lower single-precision (32-bit) floating-point element from a into +// 4 contiguous elements in memory. mem_addr must be aligned on a 16-byte +// boundary or a general-protection exception may be generated. +// +// MEM[mem_addr+31:mem_addr] := a[31:0] +// MEM[mem_addr+63:mem_addr+32] := a[31:0] +// MEM[mem_addr+95:mem_addr+64] := a[31:0] +// MEM[mem_addr+127:mem_addr+96] := a[31:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_store1_ps +#define _mm_store1_ps _mm_store_ps1 + +// Stores the upper two single-precision, floating-point values of a to the +// address p. +// +// *p0 := a2 +// *p1 := a3 +// +// https://msdn.microsoft.com/en-us/library/a7525fs8(v%3dvs.90).aspx +FORCE_INLINE void _mm_storeh_pi(__m64 *p, __m128 a) +{ + *p = vreinterpret_m64_f32(vget_high_f32(a)); +} + +// Stores the lower two single-precision floating point values of a to the +// address p. +// +// *p0 := a0 +// *p1 := a1 +// +// https://msdn.microsoft.com/en-us/library/h54t98ks(v=vs.90).aspx +FORCE_INLINE void _mm_storel_pi(__m64 *p, __m128 a) +{ + *p = vreinterpret_m64_f32(vget_low_f32(a)); +} + +// Store 4 single-precision (32-bit) floating-point elements from a into memory +// in reverse order. mem_addr must be aligned on a 16-byte boundary or a +// general-protection exception may be generated. +// +// MEM[mem_addr+31:mem_addr] := a[127:96] +// MEM[mem_addr+63:mem_addr+32] := a[95:64] +// MEM[mem_addr+95:mem_addr+64] := a[63:32] +// MEM[mem_addr+127:mem_addr+96] := a[31:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storer_ps +FORCE_INLINE void _mm_storer_ps(float *p, __m128 a) +{ + float32x4_t tmp = vrev64q_f32(vreinterpretq_f32_m128(a)); + float32x4_t rev = vextq_f32(tmp, tmp, 2); + vst1q_f32(p, rev); +} + +// Stores four single-precision, floating-point values. +// https://msdn.microsoft.com/en-us/library/44e30x22(v=vs.100).aspx +FORCE_INLINE void _mm_storeu_ps(float *p, __m128 a) +{ + vst1q_f32(p, vreinterpretq_f32_m128(a)); +} + +// Stores 16-bits of integer data a at the address p. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_si16 +FORCE_INLINE void _mm_storeu_si16(void *p, __m128i a) +{ + vst1q_lane_s16((int16_t *)p, vreinterpretq_s16_m128i(a), 0); +} + +// Stores 64-bits of integer data a at the address p. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_si64 +FORCE_INLINE void _mm_storeu_si64(void *p, __m128i a) +{ + vst1q_lane_s64((int64_t *)p, vreinterpretq_s64_m128i(a), 0); +} + +// Store 64-bits of integer data from a into memory using a non-temporal memory +// hint. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_pi +FORCE_INLINE void _mm_stream_pi(__m64 *p, __m64 a) +{ + vst1_s64((int64_t *)p, vreinterpret_s64_m64(a)); +} + +// Store 128-bits (composed of 4 packed single-precision (32-bit) floating- +// point elements) from a into memory using a non-temporal memory hint. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_ps +FORCE_INLINE void _mm_stream_ps(float *p, __m128 a) +{ +#if __has_builtin(__builtin_nontemporal_store) + __builtin_nontemporal_store(a, (float32x4_t *)p); +#else + vst1q_f32(p, vreinterpretq_f32_m128(a)); +#endif +} + +// Subtracts the four single-precision, floating-point values of a and b. +// +// r0 := a0 - b0 +// r1 := a1 - b1 +// r2 := a2 - b2 +// r3 := a3 - b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/1zad2k61(v=vs.100).aspx +FORCE_INLINE __m128 _mm_sub_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_f32( + vsubq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Subtract the lower single-precision (32-bit) floating-point element in b from +// the lower single-precision (32-bit) floating-point element in a, store the +// result in the lower element of dst, and copy the upper 3 packed elements from +// a to the upper elements of dst. +// +// dst[31:0] := a[31:0] - b[31:0] +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sub_ss +FORCE_INLINE __m128 _mm_sub_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_sub_ps(a, b)); +} + +// Macro: Transpose the 4x4 matrix formed by the 4 rows of single-precision +// (32-bit) floating-point elements in row0, row1, row2, and row3, and store the +// transposed matrix in these vectors (row0 now contains column 0, etc.). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=MM_TRANSPOSE4_PS +#define _MM_TRANSPOSE4_PS(row0, row1, row2, row3) \ + do { \ + float32x4x2_t ROW01 = vtrnq_f32(row0, row1); \ + float32x4x2_t ROW23 = vtrnq_f32(row2, row3); \ + row0 = vcombine_f32(vget_low_f32(ROW01.val[0]), \ + vget_low_f32(ROW23.val[0])); \ + row1 = vcombine_f32(vget_low_f32(ROW01.val[1]), \ + vget_low_f32(ROW23.val[1])); \ + row2 = vcombine_f32(vget_high_f32(ROW01.val[0]), \ + vget_high_f32(ROW23.val[0])); \ + row3 = vcombine_f32(vget_high_f32(ROW01.val[1]), \ + vget_high_f32(ROW23.val[1])); \ + } while (0) + +// according to the documentation, these intrinsics behave the same as the +// non-'u' versions. We'll just alias them here. +#define _mm_ucomieq_ss _mm_comieq_ss +#define _mm_ucomige_ss _mm_comige_ss +#define _mm_ucomigt_ss _mm_comigt_ss +#define _mm_ucomile_ss _mm_comile_ss +#define _mm_ucomilt_ss _mm_comilt_ss +#define _mm_ucomineq_ss _mm_comineq_ss + +// Return vector of type __m128i with undefined elements. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_undefined_si128 +FORCE_INLINE __m128i _mm_undefined_si128(void) +{ +#if defined(__GNUC__) || defined(__clang__) +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wuninitialized" +#endif + __m128i a; + return a; +#if defined(__GNUC__) || defined(__clang__) +#pragma GCC diagnostic pop +#endif +} + +// Return vector of type __m128 with undefined elements. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_undefined_ps +FORCE_INLINE __m128 _mm_undefined_ps(void) +{ +#if defined(__GNUC__) || defined(__clang__) +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wuninitialized" +#endif + __m128 a; + return a; +#if defined(__GNUC__) || defined(__clang__) +#pragma GCC diagnostic pop +#endif +} + +// Selects and interleaves the upper two single-precision, floating-point values +// from a and b. +// +// r0 := a2 +// r1 := b2 +// r2 := a3 +// r3 := b3 +// +// https://msdn.microsoft.com/en-us/library/skccxx7d%28v=vs.90%29.aspx +FORCE_INLINE __m128 _mm_unpackhi_ps(__m128 a, __m128 b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128_f32( + vzip2q_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +#else + float32x2_t a1 = vget_high_f32(vreinterpretq_f32_m128(a)); + float32x2_t b1 = vget_high_f32(vreinterpretq_f32_m128(b)); + float32x2x2_t result = vzip_f32(a1, b1); + return vreinterpretq_m128_f32(vcombine_f32(result.val[0], result.val[1])); +#endif +} + +// Selects and interleaves the lower two single-precision, floating-point values +// from a and b. +// +// r0 := a0 +// r1 := b0 +// r2 := a1 +// r3 := b1 +// +// https://msdn.microsoft.com/en-us/library/25st103b%28v=vs.90%29.aspx +FORCE_INLINE __m128 _mm_unpacklo_ps(__m128 a, __m128 b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128_f32( + vzip1q_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +#else + float32x2_t a1 = vget_low_f32(vreinterpretq_f32_m128(a)); + float32x2_t b1 = vget_low_f32(vreinterpretq_f32_m128(b)); + float32x2x2_t result = vzip_f32(a1, b1); + return vreinterpretq_m128_f32(vcombine_f32(result.val[0], result.val[1])); +#endif +} + +// Computes bitwise EXOR (exclusive-or) of the four single-precision, +// floating-point values of a and b. +// https://msdn.microsoft.com/en-us/library/ss6k3wk8(v=vs.100).aspx +FORCE_INLINE __m128 _mm_xor_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_s32( + veorq_s32(vreinterpretq_s32_m128(a), vreinterpretq_s32_m128(b))); +} + +/* SSE2 */ + +// Adds the 8 signed or unsigned 16-bit integers in a to the 8 signed or +// unsigned 16-bit integers in b. +// https://msdn.microsoft.com/en-us/library/fceha5k4(v=vs.100).aspx +FORCE_INLINE __m128i _mm_add_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vaddq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Adds the 4 signed or unsigned 32-bit integers in a to the 4 signed or +// unsigned 32-bit integers in b. +// +// r0 := a0 + b0 +// r1 := a1 + b1 +// r2 := a2 + b2 +// r3 := a3 + b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/09xs4fkk(v=vs.100).aspx +FORCE_INLINE __m128i _mm_add_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + vaddq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Adds the 4 signed or unsigned 64-bit integers in a to the 4 signed or +// unsigned 32-bit integers in b. +// https://msdn.microsoft.com/en-us/library/vstudio/09xs4fkk(v=vs.100).aspx +FORCE_INLINE __m128i _mm_add_epi64(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s64( + vaddq_s64(vreinterpretq_s64_m128i(a), vreinterpretq_s64_m128i(b))); +} + +// Adds the 16 signed or unsigned 8-bit integers in a to the 16 signed or +// unsigned 8-bit integers in b. +// https://technet.microsoft.com/en-us/subscriptions/yc7tcyzs(v=vs.90) +FORCE_INLINE __m128i _mm_add_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s8( + vaddq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} + +// Add packed double-precision (64-bit) floating-point elements in a and b, and +// store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_add_pd +FORCE_INLINE __m128d _mm_add_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vaddq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + double *da = (double *)&a; + double *db = (double *)&b; + double c[2]; + c[0] = da[0] + db[0]; + c[1] = da[1] + db[1]; + return vld1q_f32((float32_t *)c); +#endif +} + +// Add the lower double-precision (64-bit) floating-point element in a and b, +// store the result in the lower element of dst, and copy the upper element from +// a to the upper element of dst. +// +// dst[63:0] := a[63:0] + b[63:0] +// dst[127:64] := a[127:64] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_add_sd +FORCE_INLINE __m128d _mm_add_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_add_pd(a, b)); +#else + double *da = (double *)&a; + double *db = (double *)&b; + double c[2]; + c[0] = da[0] + db[0]; + c[1] = da[1]; + return vld1q_f32((float32_t *)c); +#endif +} + +// Add 64-bit integers a and b, and store the result in dst. +// +// dst[63:0] := a[63:0] + b[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_add_si64 +FORCE_INLINE __m64 _mm_add_si64(__m64 a, __m64 b) +{ + return vreinterpret_m64_s64( + vadd_s64(vreinterpret_s64_m64(a), vreinterpret_s64_m64(b))); +} + +// Adds the 8 signed 16-bit integers in a to the 8 signed 16-bit integers in b +// and saturates. +// +// r0 := SignedSaturate(a0 + b0) +// r1 := SignedSaturate(a1 + b1) +// ... +// r7 := SignedSaturate(a7 + b7) +// +// https://msdn.microsoft.com/en-us/library/1a306ef8(v=vs.100).aspx +FORCE_INLINE __m128i _mm_adds_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vqaddq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Add packed signed 8-bit integers in a and b using saturation, and store the +// results in dst. +// +// FOR j := 0 to 15 +// i := j*8 +// dst[i+7:i] := Saturate8( a[i+7:i] + b[i+7:i] ) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_adds_epi8 +FORCE_INLINE __m128i _mm_adds_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s8( + vqaddq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} + +// Add packed unsigned 16-bit integers in a and b using saturation, and store +// the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_adds_epu16 +FORCE_INLINE __m128i _mm_adds_epu16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vqaddq_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b))); +} + +// Adds the 16 unsigned 8-bit integers in a to the 16 unsigned 8-bit integers in +// b and saturates.. +// https://msdn.microsoft.com/en-us/library/9hahyddy(v=vs.100).aspx +FORCE_INLINE __m128i _mm_adds_epu8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vqaddq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); +} + +// Compute the bitwise AND of packed double-precision (64-bit) floating-point +// elements in a and b, and store the results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// dst[i+63:i] := a[i+63:i] AND b[i+63:i] +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_and_pd +FORCE_INLINE __m128d _mm_and_pd(__m128d a, __m128d b) +{ + return vreinterpretq_m128d_s64( + vandq_s64(vreinterpretq_s64_m128d(a), vreinterpretq_s64_m128d(b))); +} + +// Computes the bitwise AND of the 128-bit value in a and the 128-bit value in +// b. +// +// r := a & b +// +// https://msdn.microsoft.com/en-us/library/vstudio/6d1txsa8(v=vs.100).aspx +FORCE_INLINE __m128i _mm_and_si128(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + vandq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Compute the bitwise NOT of packed double-precision (64-bit) floating-point +// elements in a and then AND with b, and store the results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// dst[i+63:i] := ((NOT a[i+63:i]) AND b[i+63:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_andnot_pd +FORCE_INLINE __m128d _mm_andnot_pd(__m128d a, __m128d b) +{ + // *NOTE* argument swap + return vreinterpretq_m128d_s64( + vbicq_s64(vreinterpretq_s64_m128d(b), vreinterpretq_s64_m128d(a))); +} + +// Computes the bitwise AND of the 128-bit value in b and the bitwise NOT of the +// 128-bit value in a. +// +// r := (~a) & b +// +// https://msdn.microsoft.com/en-us/library/vstudio/1beaceh8(v=vs.100).aspx +FORCE_INLINE __m128i _mm_andnot_si128(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + vbicq_s32(vreinterpretq_s32_m128i(b), + vreinterpretq_s32_m128i(a))); // *NOTE* argument swap +} + +// Computes the average of the 8 unsigned 16-bit integers in a and the 8 +// unsigned 16-bit integers in b and rounds. +// +// r0 := (a0 + b0) / 2 +// r1 := (a1 + b1) / 2 +// ... +// r7 := (a7 + b7) / 2 +// +// https://msdn.microsoft.com/en-us/library/vstudio/y13ca3c8(v=vs.90).aspx +FORCE_INLINE __m128i _mm_avg_epu16(__m128i a, __m128i b) +{ + return (__m128i)vrhaddq_u16(vreinterpretq_u16_m128i(a), + vreinterpretq_u16_m128i(b)); +} + +// Computes the average of the 16 unsigned 8-bit integers in a and the 16 +// unsigned 8-bit integers in b and rounds. +// +// r0 := (a0 + b0) / 2 +// r1 := (a1 + b1) / 2 +// ... +// r15 := (a15 + b15) / 2 +// +// https://msdn.microsoft.com/en-us/library/vstudio/8zwh554a(v%3dvs.90).aspx +FORCE_INLINE __m128i _mm_avg_epu8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vrhaddq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); +} + +// Shift a left by imm8 bytes while shifting in zeros, and store the results in +// dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_bslli_si128 +#define _mm_bslli_si128(a, imm) _mm_slli_si128(a, imm) + +// Shift a right by imm8 bytes while shifting in zeros, and store the results in +// dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_bsrli_si128 +#define _mm_bsrli_si128(a, imm) _mm_srli_si128(a, imm) + +// Cast vector of type __m128d to type __m128. This intrinsic is only used for +// compilation and does not generate any instructions, thus it has zero latency. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_castpd_ps +FORCE_INLINE __m128 _mm_castpd_ps(__m128d a) +{ + return vreinterpretq_m128_s64(vreinterpretq_s64_m128d(a)); +} + +// Cast vector of type __m128d to type __m128i. This intrinsic is only used for +// compilation and does not generate any instructions, thus it has zero latency. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_castpd_si128 +FORCE_INLINE __m128i _mm_castpd_si128(__m128d a) +{ + return vreinterpretq_m128i_s64(vreinterpretq_s64_m128d(a)); +} + +// Cast vector of type __m128 to type __m128d. This intrinsic is only used for +// compilation and does not generate any instructions, thus it has zero latency. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_castps_pd +FORCE_INLINE __m128d _mm_castps_pd(__m128 a) +{ + return vreinterpretq_m128d_s32(vreinterpretq_s32_m128(a)); +} + +// Applies a type cast to reinterpret four 32-bit floating point values passed +// in as a 128-bit parameter as packed 32-bit integers. +// https://msdn.microsoft.com/en-us/library/bb514099.aspx +FORCE_INLINE __m128i _mm_castps_si128(__m128 a) +{ + return vreinterpretq_m128i_s32(vreinterpretq_s32_m128(a)); +} + +// Cast vector of type __m128i to type __m128d. This intrinsic is only used for +// compilation and does not generate any instructions, thus it has zero latency. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_castsi128_pd +FORCE_INLINE __m128d _mm_castsi128_pd(__m128i a) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vreinterpretq_f64_m128i(a)); +#else + return vreinterpretq_m128d_f32(vreinterpretq_f32_m128i(a)); +#endif +} + +// Applies a type cast to reinterpret four 32-bit integers passed in as a +// 128-bit parameter as packed 32-bit floating point values. +// https://msdn.microsoft.com/en-us/library/bb514029.aspx +FORCE_INLINE __m128 _mm_castsi128_ps(__m128i a) +{ + return vreinterpretq_m128_s32(vreinterpretq_s32_m128i(a)); +} + +// Cache line containing p is flushed and invalidated from all caches in the +// coherency domain. : +// https://msdn.microsoft.com/en-us/library/ba08y07y(v=vs.100).aspx +FORCE_INLINE void _mm_clflush(void const *p) +{ + (void)p; + // no corollary for Neon? +} + +// Compares the 8 signed or unsigned 16-bit integers in a and the 8 signed or +// unsigned 16-bit integers in b for equality. +// https://msdn.microsoft.com/en-us/library/2ay060te(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cmpeq_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vceqq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Compare packed 32-bit integers in a and b for equality, and store the results +// in dst +FORCE_INLINE __m128i _mm_cmpeq_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u32( + vceqq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Compares the 16 signed or unsigned 8-bit integers in a and the 16 signed or +// unsigned 8-bit integers in b for equality. +// https://msdn.microsoft.com/en-us/library/windows/desktop/bz5xk21a(v=vs.90).aspx +FORCE_INLINE __m128i _mm_cmpeq_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vceqq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for equality, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpeq_pd +FORCE_INLINE __m128d _mm_cmpeq_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64( + vceqq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + // (a == b) -> (a_lo == b_lo) && (a_hi == b_hi) + uint32x4_t cmp = + vceqq_u32(vreinterpretq_u32_m128d(a), vreinterpretq_u32_m128d(b)); + uint32x4_t swapped = vrev64q_u32(cmp); + return vreinterpretq_m128d_u32(vandq_u32(cmp, swapped)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for equality, store the result in the lower element of dst, and copy the +// upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpeq_sd +FORCE_INLINE __m128d _mm_cmpeq_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_cmpeq_pd(a, b)); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for greater-than-or-equal, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpge_pd +FORCE_INLINE __m128d _mm_cmpge_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64( + vcgeq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *)&a0) >= (*(double *)&b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = (*(double *)&a1) >= (*(double *)&b1) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for greater-than-or-equal, store the result in the lower element of dst, +// and copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpge_sd +FORCE_INLINE __m128d _mm_cmpge_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_cmpge_pd(a, b)); +#else + // expand "_mm_cmpge_pd()" to reduce unnecessary operations + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *)&a0) >= (*(double *)&b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = a1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compares the 8 signed 16-bit integers in a and the 8 signed 16-bit integers +// in b for greater than. +// +// r0 := (a0 > b0) ? 0xffff : 0x0 +// r1 := (a1 > b1) ? 0xffff : 0x0 +// ... +// r7 := (a7 > b7) ? 0xffff : 0x0 +// +// https://technet.microsoft.com/en-us/library/xd43yfsa(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cmpgt_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vcgtq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Compares the 4 signed 32-bit integers in a and the 4 signed 32-bit integers +// in b for greater than. +// https://msdn.microsoft.com/en-us/library/vstudio/1s9f2z0y(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cmpgt_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u32( + vcgtq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Compares the 16 signed 8-bit integers in a and the 16 signed 8-bit integers +// in b for greater than. +// +// r0 := (a0 > b0) ? 0xff : 0x0 +// r1 := (a1 > b1) ? 0xff : 0x0 +// ... +// r15 := (a15 > b15) ? 0xff : 0x0 +// +// https://msdn.microsoft.com/zh-tw/library/wf45zt2b(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cmpgt_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vcgtq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for greater-than, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpgt_pd +FORCE_INLINE __m128d _mm_cmpgt_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64( + vcgtq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *)&a0) > (*(double *)&b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = (*(double *)&a1) > (*(double *)&b1) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for greater-than, store the result in the lower element of dst, and copy +// the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpgt_sd +FORCE_INLINE __m128d _mm_cmpgt_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_cmpgt_pd(a, b)); +#else + // expand "_mm_cmpge_pd()" to reduce unnecessary operations + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *)&a0) > (*(double *)&b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = a1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for less-than-or-equal, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmple_pd +FORCE_INLINE __m128d _mm_cmple_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64( + vcleq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *)&a0) <= (*(double *)&b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = (*(double *)&a1) <= (*(double *)&b1) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for less-than-or-equal, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmple_sd +FORCE_INLINE __m128d _mm_cmple_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_cmple_pd(a, b)); +#else + // expand "_mm_cmpge_pd()" to reduce unnecessary operations + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *)&a0) <= (*(double *)&b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = a1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compares the 8 signed 16-bit integers in a and the 8 signed 16-bit integers +// in b for less than. +// +// r0 := (a0 < b0) ? 0xffff : 0x0 +// r1 := (a1 < b1) ? 0xffff : 0x0 +// ... +// r7 := (a7 < b7) ? 0xffff : 0x0 +// +// https://technet.microsoft.com/en-us/library/t863edb2(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cmplt_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vcltq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Compares the 4 signed 32-bit integers in a and the 4 signed 32-bit integers +// in b for less than. +// https://msdn.microsoft.com/en-us/library/vstudio/4ak0bf5d(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cmplt_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u32( + vcltq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Compares the 16 signed 8-bit integers in a and the 16 signed 8-bit integers +// in b for lesser than. +// https://msdn.microsoft.com/en-us/library/windows/desktop/9s46csht(v=vs.90).aspx +FORCE_INLINE __m128i _mm_cmplt_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vcltq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for less-than, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmplt_pd +FORCE_INLINE __m128d _mm_cmplt_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64( + vcltq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *)&a0) < (*(double *)&b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = (*(double *)&a1) < (*(double *)&b1) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for less-than, store the result in the lower element of dst, and copy the +// upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmplt_sd +FORCE_INLINE __m128d _mm_cmplt_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_cmplt_pd(a, b)); +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *)&a0) < (*(double *)&b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = a1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for not-equal, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpneq_pd +FORCE_INLINE __m128d _mm_cmpneq_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_s32(vmvnq_s32(vreinterpretq_s32_u64( + vceqq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))))); +#else + // (a == b) -> (a_lo == b_lo) && (a_hi == b_hi) + uint32x4_t cmp = + vceqq_u32(vreinterpretq_u32_m128d(a), vreinterpretq_u32_m128d(b)); + uint32x4_t swapped = vrev64q_u32(cmp); + return vreinterpretq_m128d_u32(vmvnq_u32(vandq_u32(cmp, swapped))); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for not-equal, store the result in the lower element of dst, and copy the +// upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpneq_sd +FORCE_INLINE __m128d _mm_cmpneq_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_cmpneq_pd(a, b)); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for not-greater-than-or-equal, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnge_pd +FORCE_INLINE __m128d _mm_cmpnge_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64(veorq_u64( + vcgeq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b)), + vdupq_n_u64(UINT64_MAX))); +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = !((*(double *)&a0) >= (*(double *)&b0)) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = !((*(double *)&a1) >= (*(double *)&b1)) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for not-greater-than-or-equal, store the result in the lower element of +// dst, and copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnge_sd +FORCE_INLINE __m128d _mm_cmpnge_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_cmpnge_pd(a, b)); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for not-greater-than, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_cmpngt_pd +FORCE_INLINE __m128d _mm_cmpngt_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64(veorq_u64( + vcgtq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b)), + vdupq_n_u64(UINT64_MAX))); +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = !((*(double *)&a0) > (*(double *)&b0)) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = !((*(double *)&a1) > (*(double *)&b1)) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for not-greater-than, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpngt_sd +FORCE_INLINE __m128d _mm_cmpngt_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_cmpngt_pd(a, b)); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for not-less-than-or-equal, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnle_pd +FORCE_INLINE __m128d _mm_cmpnle_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64(veorq_u64( + vcleq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b)), + vdupq_n_u64(UINT64_MAX))); +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = !((*(double *)&a0) <= (*(double *)&b0)) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = !((*(double *)&a1) <= (*(double *)&b1)) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for not-less-than-or-equal, store the result in the lower element of dst, +// and copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnle_sd +FORCE_INLINE __m128d _mm_cmpnle_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_cmpnle_pd(a, b)); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for not-less-than, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnlt_pd +FORCE_INLINE __m128d _mm_cmpnlt_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64(veorq_u64( + vcltq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b)), + vdupq_n_u64(UINT64_MAX))); +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = !((*(double *)&a0) < (*(double *)&b0)) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = !((*(double *)&a1) < (*(double *)&b1)) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for not-less-than, store the result in the lower element of dst, and copy +// the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnlt_sd +FORCE_INLINE __m128d _mm_cmpnlt_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_cmpnlt_pd(a, b)); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// to see if neither is NaN, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpord_pd +FORCE_INLINE __m128d _mm_cmpord_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + // Excluding NaNs, any two floating point numbers can be compared. + uint64x2_t not_nan_a = + vceqq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(a)); + uint64x2_t not_nan_b = + vceqq_f64(vreinterpretq_f64_m128d(b), vreinterpretq_f64_m128d(b)); + return vreinterpretq_m128d_u64(vandq_u64(not_nan_a, not_nan_b)); +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = ((*(double *)&a0) == (*(double *)&a0) && + (*(double *)&b0) == (*(double *)&b0)) ? + ~UINT64_C(0) : + UINT64_C(0); + d[1] = ((*(double *)&a1) == (*(double *)&a1) && + (*(double *)&b1) == (*(double *)&b1)) ? + ~UINT64_C(0) : + UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b to see if neither is NaN, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpord_sd +FORCE_INLINE __m128d _mm_cmpord_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_cmpord_pd(a, b)); +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t d[2]; + d[0] = ((*(double *)&a0) == (*(double *)&a0) && + (*(double *)&b0) == (*(double *)&b0)) ? + ~UINT64_C(0) : + UINT64_C(0); + d[1] = a1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// to see if either is NaN, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpunord_pd +FORCE_INLINE __m128d _mm_cmpunord_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + // Two NaNs are not equal in comparison operation. + uint64x2_t not_nan_a = + vceqq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(a)); + uint64x2_t not_nan_b = + vceqq_f64(vreinterpretq_f64_m128d(b), vreinterpretq_f64_m128d(b)); + return vreinterpretq_m128d_s32( + vmvnq_s32(vreinterpretq_s32_u64(vandq_u64(not_nan_a, not_nan_b)))); +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = ((*(double *)&a0) == (*(double *)&a0) && + (*(double *)&b0) == (*(double *)&b0)) ? + UINT64_C(0) : + ~UINT64_C(0); + d[1] = ((*(double *)&a1) == (*(double *)&a1) && + (*(double *)&b1) == (*(double *)&b1)) ? + UINT64_C(0) : + ~UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b to see if either is NaN, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpunord_sd +FORCE_INLINE __m128d _mm_cmpunord_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_cmpunord_pd(a, b)); +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t d[2]; + d[0] = ((*(double *)&a0) == (*(double *)&a0) && + (*(double *)&b0) == (*(double *)&b0)) ? + UINT64_C(0) : + ~UINT64_C(0); + d[1] = a1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point element in a and b +// for greater-than-or-equal, and return the boolean result (0 or 1). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_comige_sd +FORCE_INLINE int _mm_comige_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vgetq_lane_u64(vcgeq_f64(a, b), 0) & 0x1; +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + + return (*(double *)&a0 >= *(double *)&b0); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point element in a and b +// for greater-than, and return the boolean result (0 or 1). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_comigt_sd +FORCE_INLINE int _mm_comigt_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vgetq_lane_u64(vcgtq_f64(a, b), 0) & 0x1; +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + + return (*(double *)&a0 > *(double *)&b0); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point element in a and b +// for less-than-or-equal, and return the boolean result (0 or 1). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_comile_sd +FORCE_INLINE int _mm_comile_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vgetq_lane_u64(vcleq_f64(a, b), 0) & 0x1; +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + + return (*(double *)&a0 <= *(double *)&b0); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point element in a and b +// for less-than, and return the boolean result (0 or 1). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_comilt_sd +FORCE_INLINE int _mm_comilt_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vgetq_lane_u64(vcltq_f64(a, b), 0) & 0x1; +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + + return (*(double *)&a0 < *(double *)&b0); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point element in a and b +// for equality, and return the boolean result (0 or 1). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_comieq_sd +FORCE_INLINE int _mm_comieq_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vgetq_lane_u64(vceqq_f64(a, b), 0) & 0x1; +#else + uint32x4_t a_not_nan = + vceqq_u32(vreinterpretq_u32_m128d(a), vreinterpretq_u32_m128d(a)); + uint32x4_t b_not_nan = + vceqq_u32(vreinterpretq_u32_m128d(b), vreinterpretq_u32_m128d(b)); + uint32x4_t a_and_b_not_nan = vandq_u32(a_not_nan, b_not_nan); + uint32x4_t a_eq_b = + vceqq_u32(vreinterpretq_u32_m128d(a), vreinterpretq_u32_m128d(b)); + uint64x2_t and_results = vandq_u64(vreinterpretq_u64_u32(a_and_b_not_nan), + vreinterpretq_u64_u32(a_eq_b)); + return vgetq_lane_u64(and_results, 0) & 0x1; +#endif +} + +// Compare the lower double-precision (64-bit) floating-point element in a and b +// for not-equal, and return the boolean result (0 or 1). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_comineq_sd +FORCE_INLINE int _mm_comineq_sd(__m128d a, __m128d b) +{ + return !_mm_comieq_sd(a, b); +} + +// Convert packed signed 32-bit integers in a to packed double-precision +// (64-bit) floating-point elements, and store the results in dst. +// +// FOR j := 0 to 1 +// i := j*32 +// m := j*64 +// dst[m+63:m] := Convert_Int32_To_FP64(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtepi32_pd +FORCE_INLINE __m128d _mm_cvtepi32_pd(__m128i a) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vcvtq_f64_s64(vmovl_s32(vget_low_s32(vreinterpretq_s32_m128i(a))))); +#else + double a0 = (double)vgetq_lane_s32(vreinterpretq_s32_m128i(a), 0); + double a1 = (double)vgetq_lane_s32(vreinterpretq_s32_m128i(a), 1); + return _mm_set_pd(a1, a0); +#endif +} + +// Converts the four signed 32-bit integer values of a to single-precision, +// floating-point values +// https://msdn.microsoft.com/en-us/library/vstudio/36bwxcx5(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cvtepi32_ps(__m128i a) +{ + return vreinterpretq_m128_f32(vcvtq_f32_s32(vreinterpretq_s32_m128i(a))); +} + +// Convert packed double-precision (64-bit) floating-point elements in a to +// packed 32-bit integers, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// k := 64*j +// dst[i+31:i] := Convert_FP64_To_Int32(a[k+63:k]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpd_epi32 +FORCE_INLINE __m128i _mm_cvtpd_epi32(__m128d a) +{ + __m128d rnd = _mm_round_pd(a, _MM_FROUND_CUR_DIRECTION); + double d0 = ((double *)&rnd)[0]; + double d1 = ((double *)&rnd)[1]; + return _mm_set_epi32(0, 0, (int32_t)d1, (int32_t)d0); +} + +// Convert packed double-precision (64-bit) floating-point elements in a to +// packed 32-bit integers, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// k := 64*j +// dst[i+31:i] := Convert_FP64_To_Int32(a[k+63:k]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpd_pi32 +FORCE_INLINE __m64 _mm_cvtpd_pi32(__m128d a) +{ + __m128d rnd = _mm_round_pd(a, _MM_FROUND_CUR_DIRECTION); + double d0 = ((double *)&rnd)[0]; + double d1 = ((double *)&rnd)[1]; + int32_t ALIGN_STRUCT(16) data[2] = { (int32_t)d0, (int32_t)d1 }; + return vreinterpret_m64_s32(vld1_s32(data)); +} + +// Convert packed double-precision (64-bit) floating-point elements in a to +// packed single-precision (32-bit) floating-point elements, and store the +// results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// k := 64*j +// dst[i+31:i] := Convert_FP64_To_FP32(a[k+64:k]) +// ENDFOR +// dst[127:64] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpd_ps +FORCE_INLINE __m128 _mm_cvtpd_ps(__m128d a) +{ +#if defined(__aarch64__) + float32x2_t tmp = vcvt_f32_f64(vreinterpretq_f64_m128d(a)); + return vreinterpretq_m128_f32(vcombine_f32(tmp, vdup_n_f32(0))); +#else + float a0 = (float)((double *)&a)[0]; + float a1 = (float)((double *)&a)[1]; + return _mm_set_ps(0, 0, a1, a0); +#endif +} + +// Convert packed signed 32-bit integers in a to packed double-precision +// (64-bit) floating-point elements, and store the results in dst. +// +// FOR j := 0 to 1 +// i := j*32 +// m := j*64 +// dst[m+63:m] := Convert_Int32_To_FP64(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi32_pd +FORCE_INLINE __m128d _mm_cvtpi32_pd(__m64 a) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vcvtq_f64_s64(vmovl_s32(vreinterpret_s32_m64(a)))); +#else + double a0 = (double)vget_lane_s32(vreinterpret_s32_m64(a), 0); + double a1 = (double)vget_lane_s32(vreinterpret_s32_m64(a), 1); + return _mm_set_pd(a1, a0); +#endif +} + +// Converts the four single-precision, floating-point values of a to signed +// 32-bit integer values. +// +// r0 := (int) a0 +// r1 := (int) a1 +// r2 := (int) a2 +// r3 := (int) a3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/xdc42k5e(v=vs.100).aspx +// *NOTE*. The default rounding mode on SSE is 'round to even', which ARMv7-A +// does not support! It is supported on ARMv8-A however. +FORCE_INLINE __m128i _mm_cvtps_epi32(__m128 a) +{ +#if defined(__aarch64__) + switch (_MM_GET_ROUNDING_MODE()) { + case _MM_ROUND_NEAREST: + return vreinterpretq_m128i_s32(vcvtnq_s32_f32(a)); + case _MM_ROUND_DOWN: + return vreinterpretq_m128i_s32(vcvtmq_s32_f32(a)); + case _MM_ROUND_UP: + return vreinterpretq_m128i_s32(vcvtpq_s32_f32(a)); + default: // _MM_ROUND_TOWARD_ZERO + return vreinterpretq_m128i_s32(vcvtq_s32_f32(a)); + } +#else + float *f = (float *)&a; + switch (_MM_GET_ROUNDING_MODE()) { + case _MM_ROUND_NEAREST: { + uint32x4_t signmask = vdupq_n_u32(0x80000000); + float32x4_t half = vbslq_f32(signmask, vreinterpretq_f32_m128(a), + vdupq_n_f32(0.5f)); /* +/- 0.5 */ + int32x4_t r_normal = vcvtq_s32_f32(vaddq_f32( + vreinterpretq_f32_m128(a), half)); /* round to integer: [a + 0.5]*/ + int32x4_t r_trunc = vcvtq_s32_f32( + vreinterpretq_f32_m128(a)); /* truncate to integer: [a] */ + int32x4_t plusone = vreinterpretq_s32_u32(vshrq_n_u32( + vreinterpretq_u32_s32(vnegq_s32(r_trunc)), 31)); /* 1 or 0 */ + int32x4_t r_even = vbicq_s32(vaddq_s32(r_trunc, plusone), + vdupq_n_s32(1)); /* ([a] + {0,1}) & ~1 */ + float32x4_t delta = vsubq_f32( + vreinterpretq_f32_m128(a), + vcvtq_f32_s32(r_trunc)); /* compute delta: delta = (a - [a]) */ + uint32x4_t is_delta_half = + vceqq_f32(delta, half); /* delta == +/- 0.5 */ + return vreinterpretq_m128i_s32( + vbslq_s32(is_delta_half, r_even, r_normal)); + } + case _MM_ROUND_DOWN: + return _mm_set_epi32(floorf(f[3]), floorf(f[2]), floorf(f[1]), + floorf(f[0])); + case _MM_ROUND_UP: + return _mm_set_epi32(ceilf(f[3]), ceilf(f[2]), ceilf(f[1]), + ceilf(f[0])); + default: // _MM_ROUND_TOWARD_ZERO + return _mm_set_epi32((int32_t)f[3], (int32_t)f[2], (int32_t)f[1], + (int32_t)f[0]); + } +#endif +} + +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed double-precision (64-bit) floating-point elements, and store the +// results in dst. +// +// FOR j := 0 to 1 +// i := 64*j +// k := 32*j +// dst[i+63:i] := Convert_FP32_To_FP64(a[k+31:k]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtps_pd +FORCE_INLINE __m128d _mm_cvtps_pd(__m128 a) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vcvt_f64_f32(vget_low_f32(vreinterpretq_f32_m128(a)))); +#else + double a0 = (double)vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); + double a1 = (double)vgetq_lane_f32(vreinterpretq_f32_m128(a), 1); + return _mm_set_pd(a1, a0); +#endif +} + +// Copy the lower double-precision (64-bit) floating-point element of a to dst. +// +// dst[63:0] := a[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsd_f64 +FORCE_INLINE double _mm_cvtsd_f64(__m128d a) +{ +#if defined(__aarch64__) + return (double)vgetq_lane_f64(vreinterpretq_f64_m128d(a), 0); +#else + return ((double *)&a)[0]; +#endif +} + +// Convert the lower double-precision (64-bit) floating-point element in a to a +// 32-bit integer, and store the result in dst. +// +// dst[31:0] := Convert_FP64_To_Int32(a[63:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsd_si32 +FORCE_INLINE int32_t _mm_cvtsd_si32(__m128d a) +{ +#if defined(__aarch64__) + return (int32_t)vgetq_lane_f64(vrndiq_f64(vreinterpretq_f64_m128d(a)), 0); +#else + __m128d rnd = _mm_round_pd(a, _MM_FROUND_CUR_DIRECTION); + double ret = ((double *)&rnd)[0]; + return (int32_t)ret; +#endif +} + +// Convert the lower double-precision (64-bit) floating-point element in a to a +// 64-bit integer, and store the result in dst. +// +// dst[63:0] := Convert_FP64_To_Int64(a[63:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsd_si64 +FORCE_INLINE int64_t _mm_cvtsd_si64(__m128d a) +{ +#if defined(__aarch64__) + return (int64_t)vgetq_lane_f64(vrndiq_f64(vreinterpretq_f64_m128d(a)), 0); +#else + __m128d rnd = _mm_round_pd(a, _MM_FROUND_CUR_DIRECTION); + double ret = ((double *)&rnd)[0]; + return (int64_t)ret; +#endif +} + +// Convert the lower double-precision (64-bit) floating-point element in a to a +// 64-bit integer, and store the result in dst. +// +// dst[63:0] := Convert_FP64_To_Int64(a[63:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsd_si64x +#define _mm_cvtsd_si64x _mm_cvtsd_si64 + +// Convert the lower double-precision (64-bit) floating-point element in b to a +// single-precision (32-bit) floating-point element, store the result in the +// lower element of dst, and copy the upper 3 packed elements from a to the +// upper elements of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsd_ss +FORCE_INLINE __m128 _mm_cvtsd_ss(__m128 a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128_f32(vsetq_lane_f32( + vget_lane_f32(vcvt_f32_f64(vreinterpretq_f64_m128d(b)), 0), + vreinterpretq_f32_m128(a), 0)); +#else + return vreinterpretq_m128_f32( + vsetq_lane_f32((float)((double *)&b)[0], vreinterpretq_f32_m128(a), 0)); +#endif +} + +// Copy the lower 32-bit integer in a to dst. +// +// dst[31:0] := a[31:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi128_si32 +FORCE_INLINE int _mm_cvtsi128_si32(__m128i a) +{ + return vgetq_lane_s32(vreinterpretq_s32_m128i(a), 0); +} + +// Copy the lower 64-bit integer in a to dst. +// +// dst[63:0] := a[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi128_si64 +FORCE_INLINE int64_t _mm_cvtsi128_si64(__m128i a) +{ + return vgetq_lane_s64(vreinterpretq_s64_m128i(a), 0); +} + +// Copy the lower 64-bit integer in a to dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi128_si64x +#define _mm_cvtsi128_si64x(a) _mm_cvtsi128_si64(a) + +// Convert the signed 32-bit integer b to a double-precision (64-bit) +// floating-point element, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi32_sd +FORCE_INLINE __m128d _mm_cvtsi32_sd(__m128d a, int32_t b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vsetq_lane_f64((double)b, vreinterpretq_f64_m128d(a), 0)); +#else + double bf = (double)b; + return vreinterpretq_m128d_s64( + vsetq_lane_s64(*(int64_t *)&bf, vreinterpretq_s64_m128d(a), 0)); +#endif +} + +// Copy the lower 64-bit integer in a to dst. +// +// dst[63:0] := a[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi128_si64x +#define _mm_cvtsi128_si64x(a) _mm_cvtsi128_si64(a) + +// Moves 32-bit integer a to the least significant 32 bits of an __m128 object, +// zero extending the upper bits. +// +// r0 := a +// r1 := 0x0 +// r2 := 0x0 +// r3 := 0x0 +// +// https://msdn.microsoft.com/en-us/library/ct3539ha%28v=vs.90%29.aspx +FORCE_INLINE __m128i _mm_cvtsi32_si128(int a) +{ + return vreinterpretq_m128i_s32(vsetq_lane_s32(a, vdupq_n_s32(0), 0)); +} + +// Convert the signed 64-bit integer b to a double-precision (64-bit) +// floating-point element, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi64_sd +FORCE_INLINE __m128d _mm_cvtsi64_sd(__m128d a, int64_t b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vsetq_lane_f64((double)b, vreinterpretq_f64_m128d(a), 0)); +#else + double bf = (double)b; + return vreinterpretq_m128d_s64( + vsetq_lane_s64(*(int64_t *)&bf, vreinterpretq_s64_m128d(a), 0)); +#endif +} + +// Moves 64-bit integer a to the least significant 64 bits of an __m128 object, +// zero extending the upper bits. +// +// r0 := a +// r1 := 0x0 +FORCE_INLINE __m128i _mm_cvtsi64_si128(int64_t a) +{ + return vreinterpretq_m128i_s64(vsetq_lane_s64(a, vdupq_n_s64(0), 0)); +} + +// Copy 64-bit integer a to the lower element of dst, and zero the upper +// element. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi64x_si128 +#define _mm_cvtsi64x_si128(a) _mm_cvtsi64_si128(a) + +// Convert the signed 64-bit integer b to a double-precision (64-bit) +// floating-point element, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi64x_sd +#define _mm_cvtsi64x_sd(a, b) _mm_cvtsi64_sd(a, b) + +// Convert the lower single-precision (32-bit) floating-point element in b to a +// double-precision (64-bit) floating-point element, store the result in the +// lower element of dst, and copy the upper element from a to the upper element +// of dst. +// +// dst[63:0] := Convert_FP32_To_FP64(b[31:0]) +// dst[127:64] := a[127:64] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtss_sd +FORCE_INLINE __m128d _mm_cvtss_sd(__m128d a, __m128 b) +{ + double d = (double)vgetq_lane_f32(vreinterpretq_f32_m128(b), 0); +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vsetq_lane_f64(d, vreinterpretq_f64_m128d(a), 0)); +#else + return vreinterpretq_m128d_s64( + vsetq_lane_s64(*(int64_t *)&d, vreinterpretq_s64_m128d(a), 0)); +#endif +} + +// Convert packed double-precision (64-bit) floating-point elements in a to +// packed 32-bit integers with truncation, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttpd_epi32 +FORCE_INLINE __m128i _mm_cvttpd_epi32(__m128d a) +{ + double a0 = ((double *)&a)[0]; + double a1 = ((double *)&a)[1]; + return _mm_set_epi32(0, 0, (int32_t)a1, (int32_t)a0); +} + +// Convert packed double-precision (64-bit) floating-point elements in a to +// packed 32-bit integers with truncation, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttpd_pi32 +FORCE_INLINE __m64 _mm_cvttpd_pi32(__m128d a) +{ + double a0 = ((double *)&a)[0]; + double a1 = ((double *)&a)[1]; + int32_t ALIGN_STRUCT(16) data[2] = { (int32_t)a0, (int32_t)a1 }; + return vreinterpret_m64_s32(vld1_s32(data)); +} + +// Converts the four single-precision, floating-point values of a to signed +// 32-bit integer values using truncate. +// https://msdn.microsoft.com/en-us/library/vstudio/1h005y6x(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cvttps_epi32(__m128 a) +{ + return vreinterpretq_m128i_s32(vcvtq_s32_f32(vreinterpretq_f32_m128(a))); +} + +// Convert the lower double-precision (64-bit) floating-point element in a to a +// 32-bit integer with truncation, and store the result in dst. +// +// dst[63:0] := Convert_FP64_To_Int32_Truncate(a[63:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttsd_si32 +FORCE_INLINE int32_t _mm_cvttsd_si32(__m128d a) +{ + double ret = *((double *)&a); + return (int32_t)ret; +} + +// Convert the lower double-precision (64-bit) floating-point element in a to a +// 64-bit integer with truncation, and store the result in dst. +// +// dst[63:0] := Convert_FP64_To_Int64_Truncate(a[63:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttsd_si64 +FORCE_INLINE int64_t _mm_cvttsd_si64(__m128d a) +{ +#if defined(__aarch64__) + return vgetq_lane_s64(vcvtq_s64_f64(vreinterpretq_f64_m128d(a)), 0); +#else + double ret = *((double *)&a); + return (int64_t)ret; +#endif +} + +// Convert the lower double-precision (64-bit) floating-point element in a to a +// 64-bit integer with truncation, and store the result in dst. +// +// dst[63:0] := Convert_FP64_To_Int64_Truncate(a[63:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttsd_si64x +#define _mm_cvttsd_si64x(a) _mm_cvttsd_si64(a) + +// Divide packed double-precision (64-bit) floating-point elements in a by +// packed elements in b, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 64*j +// dst[i+63:i] := a[i+63:i] / b[i+63:i] +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_div_pd +FORCE_INLINE __m128d _mm_div_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vdivq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + double *da = (double *)&a; + double *db = (double *)&b; + double c[2]; + c[0] = da[0] / db[0]; + c[1] = da[1] / db[1]; + return vld1q_f32((float32_t *)c); +#endif +} + +// Divide the lower double-precision (64-bit) floating-point element in a by the +// lower double-precision (64-bit) floating-point element in b, store the result +// in the lower element of dst, and copy the upper element from a to the upper +// element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_div_sd +FORCE_INLINE __m128d _mm_div_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + float64x2_t tmp = + vdivq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b)); + return vreinterpretq_m128d_f64( + vsetq_lane_f64(vgetq_lane_f64(vreinterpretq_f64_m128d(a), 1), tmp, 1)); +#else + return _mm_move_sd(a, _mm_div_pd(a, b)); +#endif +} + +// Extracts the selected signed or unsigned 16-bit integer from a and zero +// extends. +// https://msdn.microsoft.com/en-us/library/6dceta0c(v=vs.100).aspx +// FORCE_INLINE int _mm_extract_epi16(__m128i a, __constrange(0,8) int imm) +#define _mm_extract_epi16(a, imm) \ + vgetq_lane_u16(vreinterpretq_u16_m128i(a), (imm)) + +// Inserts the least significant 16 bits of b into the selected 16-bit integer +// of a. +// https://msdn.microsoft.com/en-us/library/kaze8hz1%28v=vs.100%29.aspx +// FORCE_INLINE __m128i _mm_insert_epi16(__m128i a, int b, +// __constrange(0,8) int imm) +#define _mm_insert_epi16(a, b, imm) \ + __extension__({ \ + vreinterpretq_m128i_s16( \ + vsetq_lane_s16((b), vreinterpretq_s16_m128i(a), (imm))); \ + }) + +// Loads two double-precision from 16-byte aligned memory, floating-point +// values. +// +// dst[127:0] := MEM[mem_addr+127:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load_pd +FORCE_INLINE __m128d _mm_load_pd(const double *p) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vld1q_f64(p)); +#else + const float *fp = (const float *)p; + float ALIGN_STRUCT(16) data[4] = { fp[0], fp[1], fp[2], fp[3] }; + return vreinterpretq_m128d_f32(vld1q_f32(data)); +#endif +} + +// Load a double-precision (64-bit) floating-point element from memory into both +// elements of dst. +// +// dst[63:0] := MEM[mem_addr+63:mem_addr] +// dst[127:64] := MEM[mem_addr+63:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load_pd1 +#define _mm_load_pd1 _mm_load1_pd + +// Load a double-precision (64-bit) floating-point element from memory into the +// lower of dst, and zero the upper element. mem_addr does not need to be +// aligned on any particular boundary. +// +// dst[63:0] := MEM[mem_addr+63:mem_addr] +// dst[127:64] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load_sd +FORCE_INLINE __m128d _mm_load_sd(const double *p) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vsetq_lane_f64(*p, vdupq_n_f64(0), 0)); +#else + const float *fp = (const float *)p; + float ALIGN_STRUCT(16) data[4] = { fp[0], fp[1], 0, 0 }; + return vreinterpretq_m128d_f32(vld1q_f32(data)); +#endif +} + +// Loads 128-bit value. : +// https://msdn.microsoft.com/en-us/library/atzzad1h(v=vs.80).aspx +FORCE_INLINE __m128i _mm_load_si128(const __m128i *p) +{ + return vreinterpretq_m128i_s32(vld1q_s32((const int32_t *)p)); +} + +// Load a double-precision (64-bit) floating-point element from memory into both +// elements of dst. +// +// dst[63:0] := MEM[mem_addr+63:mem_addr] +// dst[127:64] := MEM[mem_addr+63:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load1_pd +FORCE_INLINE __m128d _mm_load1_pd(const double *p) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vld1q_dup_f64(p)); +#else + return vreinterpretq_m128d_s64(vdupq_n_s64(*(const int64_t *)p)); +#endif +} + +// Load a double-precision (64-bit) floating-point element from memory into the +// upper element of dst, and copy the lower element from a to dst. mem_addr does +// not need to be aligned on any particular boundary. +// +// dst[63:0] := a[63:0] +// dst[127:64] := MEM[mem_addr+63:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadh_pd +FORCE_INLINE __m128d _mm_loadh_pd(__m128d a, const double *p) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vcombine_f64(vget_low_f64(vreinterpretq_f64_m128d(a)), vld1_f64(p))); +#else + return vreinterpretq_m128d_f32(vcombine_f32( + vget_low_f32(vreinterpretq_f32_m128d(a)), vld1_f32((const float *)p))); +#endif +} + +// Load 64-bit integer from memory into the first element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadl_epi64 +FORCE_INLINE __m128i _mm_loadl_epi64(__m128i const *p) +{ + /* Load the lower 64 bits of the value pointed to by p into the + * lower 64 bits of the result, zeroing the upper 64 bits of the result. + */ + return vreinterpretq_m128i_s32( + vcombine_s32(vld1_s32((int32_t const *)p), vcreate_s32(0))); +} + +// Load a double-precision (64-bit) floating-point element from memory into the +// lower element of dst, and copy the upper element from a to dst. mem_addr does +// not need to be aligned on any particular boundary. +// +// dst[63:0] := MEM[mem_addr+63:mem_addr] +// dst[127:64] := a[127:64] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadl_pd +FORCE_INLINE __m128d _mm_loadl_pd(__m128d a, const double *p) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vcombine_f64(vld1_f64(p), vget_high_f64(vreinterpretq_f64_m128d(a)))); +#else + return vreinterpretq_m128d_f32(vcombine_f32( + vld1_f32((const float *)p), vget_high_f32(vreinterpretq_f32_m128d(a)))); +#endif +} + +// Load 2 double-precision (64-bit) floating-point elements from memory into dst +// in reverse order. mem_addr must be aligned on a 16-byte boundary or a +// general-protection exception may be generated. +// +// dst[63:0] := MEM[mem_addr+127:mem_addr+64] +// dst[127:64] := MEM[mem_addr+63:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadr_pd +FORCE_INLINE __m128d _mm_loadr_pd(const double *p) +{ +#if defined(__aarch64__) + float64x2_t v = vld1q_f64(p); + return vreinterpretq_m128d_f64(vextq_f64(v, v, 1)); +#else + int64x2_t v = vld1q_s64((const int64_t *)p); + return vreinterpretq_m128d_s64(vextq_s64(v, v, 1)); +#endif +} + +// Loads two double-precision from unaligned memory, floating-point values. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadu_pd +FORCE_INLINE __m128d _mm_loadu_pd(const double *p) +{ + return _mm_load_pd(p); +} + +// Loads 128-bit value. : +// https://msdn.microsoft.com/zh-cn/library/f4k12ae8(v=vs.90).aspx +FORCE_INLINE __m128i _mm_loadu_si128(const __m128i *p) +{ + return vreinterpretq_m128i_s32(vld1q_s32((const int32_t *)p)); +} + +// Load unaligned 32-bit integer from memory into the first element of dst. +// +// dst[31:0] := MEM[mem_addr+31:mem_addr] +// dst[MAX:32] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadu_si32 +FORCE_INLINE __m128i _mm_loadu_si32(const void *p) +{ + return vreinterpretq_m128i_s32( + vsetq_lane_s32(*(const int32_t *)p, vdupq_n_s32(0), 0)); +} + +// Multiplies the 8 signed 16-bit integers from a by the 8 signed 16-bit +// integers from b. +// +// r0 := (a0 * b0) + (a1 * b1) +// r1 := (a2 * b2) + (a3 * b3) +// r2 := (a4 * b4) + (a5 * b5) +// r3 := (a6 * b6) + (a7 * b7) +// https://msdn.microsoft.com/en-us/library/yht36sa6(v=vs.90).aspx +FORCE_INLINE __m128i _mm_madd_epi16(__m128i a, __m128i b) +{ + int32x4_t low = vmull_s16(vget_low_s16(vreinterpretq_s16_m128i(a)), + vget_low_s16(vreinterpretq_s16_m128i(b))); + int32x4_t high = vmull_s16(vget_high_s16(vreinterpretq_s16_m128i(a)), + vget_high_s16(vreinterpretq_s16_m128i(b))); + + int32x2_t low_sum = vpadd_s32(vget_low_s32(low), vget_high_s32(low)); + int32x2_t high_sum = vpadd_s32(vget_low_s32(high), vget_high_s32(high)); + + return vreinterpretq_m128i_s32(vcombine_s32(low_sum, high_sum)); +} + +// Conditionally store 8-bit integer elements from a into memory using mask +// (elements are not stored when the highest bit is not set in the corresponding +// element) and a non-temporal memory hint. mem_addr does not need to be aligned +// on any particular boundary. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_maskmoveu_si128 +FORCE_INLINE void _mm_maskmoveu_si128(__m128i a, __m128i mask, char *mem_addr) +{ + int8x16_t shr_mask = vshrq_n_s8(vreinterpretq_s8_m128i(mask), 7); + __m128 b = _mm_load_ps((const float *)mem_addr); + int8x16_t masked = vbslq_s8(vreinterpretq_u8_s8(shr_mask), + vreinterpretq_s8_m128i(a), + vreinterpretq_s8_m128(b)); + vst1q_s8((int8_t *)mem_addr, masked); +} + +// Computes the pairwise maxima of the 8 signed 16-bit integers from a and the 8 +// signed 16-bit integers from b. +// https://msdn.microsoft.com/en-us/LIBRary/3x060h7c(v=vs.100).aspx +FORCE_INLINE __m128i _mm_max_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vmaxq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Computes the pairwise maxima of the 16 unsigned 8-bit integers from a and the +// 16 unsigned 8-bit integers from b. +// https://msdn.microsoft.com/en-us/library/st6634za(v=vs.100).aspx +FORCE_INLINE __m128i _mm_max_epu8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vmaxq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b, +// and store packed maximum values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_pd +FORCE_INLINE __m128d _mm_max_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) +#if SSE2NEON_PRECISE_MINMAX + float64x2_t _a = vreinterpretq_f64_m128d(a); + float64x2_t _b = vreinterpretq_f64_m128d(b); + return vreinterpretq_m128d_f64(vbslq_f64(vcgtq_f64(_a, _b), _a, _b)); +#else + return vreinterpretq_m128d_f64( + vmaxq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#endif +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *)&a0) > (*(double *)&b0) ? a0 : b0; + d[1] = (*(double *)&a1) > (*(double *)&b1) ? a1 : b1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b, store the maximum value in the lower element of dst, and copy the upper +// element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_sd +FORCE_INLINE __m128d _mm_max_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_max_pd(a, b)); +#else + double *da = (double *)&a; + double *db = (double *)&b; + double c[2] = { da[0] > db[0] ? da[0] : db[0], da[1] }; + return vreinterpretq_m128d_f32(vld1q_f32((float32_t *)c)); +#endif +} + +// Computes the pairwise minima of the 8 signed 16-bit integers from a and the 8 +// signed 16-bit integers from b. +// https://msdn.microsoft.com/en-us/library/vstudio/6te997ew(v=vs.100).aspx +FORCE_INLINE __m128i _mm_min_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vminq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Computes the pairwise minima of the 16 unsigned 8-bit integers from a and the +// 16 unsigned 8-bit integers from b. +// https://msdn.microsoft.com/ko-kr/library/17k8cf58(v=vs.100).aspxx +FORCE_INLINE __m128i _mm_min_epu8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vminq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b, +// and store packed minimum values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_pd +FORCE_INLINE __m128d _mm_min_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) +#if SSE2NEON_PRECISE_MINMAX + float64x2_t _a = vreinterpretq_f64_m128d(a); + float64x2_t _b = vreinterpretq_f64_m128d(b); + return vreinterpretq_m128d_f64(vbslq_f64(vcltq_f64(_a, _b), _a, _b)); +#else + return vreinterpretq_m128d_f64( + vminq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#endif +#else + uint64_t a0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t)vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t)vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *)&a0) < (*(double *)&b0) ? a0 : b0; + d[1] = (*(double *)&a1) < (*(double *)&b1) ? a1 : b1; + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b, store the minimum value in the lower element of dst, and copy the upper +// element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_sd +FORCE_INLINE __m128d _mm_min_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_min_pd(a, b)); +#else + double *da = (double *)&a; + double *db = (double *)&b; + double c[2] = { da[0] < db[0] ? da[0] : db[0], da[1] }; + return vreinterpretq_m128d_f32(vld1q_f32((float32_t *)c)); +#endif +} + +// Copy the lower 64-bit integer in a to the lower element of dst, and zero the +// upper element. +// +// dst[63:0] := a[63:0] +// dst[127:64] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_move_epi64 +FORCE_INLINE __m128i _mm_move_epi64(__m128i a) +{ + return vreinterpretq_m128i_s64( + vsetq_lane_s64(0, vreinterpretq_s64_m128i(a), 1)); +} + +// Move the lower double-precision (64-bit) floating-point element from b to the +// lower element of dst, and copy the upper element from a to the upper element +// of dst. +// +// dst[63:0] := b[63:0] +// dst[127:64] := a[127:64] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_move_sd +FORCE_INLINE __m128d _mm_move_sd(__m128d a, __m128d b) +{ + return vreinterpretq_m128d_f32( + vcombine_f32(vget_low_f32(vreinterpretq_f32_m128d(b)), + vget_high_f32(vreinterpretq_f32_m128d(a)))); +} + +// NEON does not provide a version of this function. +// Creates a 16-bit mask from the most significant bits of the 16 signed or +// unsigned 8-bit integers in a and zero extends the upper bits. +// https://msdn.microsoft.com/en-us/library/vstudio/s090c8fk(v=vs.100).aspx +FORCE_INLINE int _mm_movemask_epi8(__m128i a) +{ + // Use increasingly wide shifts+adds to collect the sign bits + // together. + // Since the widening shifts would be rather confusing to follow in little + // endian, everything will be illustrated in big endian order instead. This + // has a different result - the bits would actually be reversed on a big + // endian machine. + + // Starting input (only half the elements are shown): + // 89 ff 1d c0 00 10 99 33 + uint8x16_t input = vreinterpretq_u8_m128i(a); + + // Shift out everything but the sign bits with an unsigned shift right. + // + // Bytes of the vector:: + // 89 ff 1d c0 00 10 99 33 + // \ \ \ \ \ \ \ \ high_bits = (uint16x4_t)(input >> 7) + // | | | | | | | | + // 01 01 00 01 00 00 01 00 + // + // Bits of first important lane(s): + // 10001001 (89) + // \______ + // | + // 00000001 (01) + uint16x8_t high_bits = vreinterpretq_u16_u8(vshrq_n_u8(input, 7)); + + // Merge the even lanes together with a 16-bit unsigned shift right + add. + // 'xx' represents garbage data which will be ignored in the final result. + // In the important bytes, the add functions like a binary OR. + // + // 01 01 00 01 00 00 01 00 + // \_ | \_ | \_ | \_ | paired16 = (uint32x4_t)(input + (input >> 7)) + // \| \| \| \| + // xx 03 xx 01 xx 00 xx 02 + // + // 00000001 00000001 (01 01) + // \_______ | + // \| + // xxxxxxxx xxxxxx11 (xx 03) + uint32x4_t paired16 = + vreinterpretq_u32_u16(vsraq_n_u16(high_bits, high_bits, 7)); + + // Repeat with a wider 32-bit shift + add. + // xx 03 xx 01 xx 00 xx 02 + // \____ | \____ | paired32 = (uint64x1_t)(paired16 + (paired16 >> + // 14)) + // \| \| + // xx xx xx 0d xx xx xx 02 + // + // 00000011 00000001 (03 01) + // \\_____ || + // '----.\|| + // xxxxxxxx xxxx1101 (xx 0d) + uint64x2_t paired32 = + vreinterpretq_u64_u32(vsraq_n_u32(paired16, paired16, 14)); + + // Last, an even wider 64-bit shift + add to get our result in the low 8 bit + // lanes. xx xx xx 0d xx xx xx 02 + // \_________ | paired64 = (uint8x8_t)(paired32 + (paired32 >> + // 28)) + // \| + // xx xx xx xx xx xx xx d2 + // + // 00001101 00000010 (0d 02) + // \ \___ | | + // '---. \| | + // xxxxxxxx 11010010 (xx d2) + uint8x16_t paired64 = + vreinterpretq_u8_u64(vsraq_n_u64(paired32, paired32, 28)); + + // Extract the low 8 bits from each 64-bit lane with 2 8-bit extracts. + // xx xx xx xx xx xx xx d2 + // || return paired64[0] + // d2 + // Note: Little endian would return the correct value 4b (01001011) instead. + return vgetq_lane_u8(paired64, 0) | ((int)vgetq_lane_u8(paired64, 8) << 8); +} + +// Set each bit of mask dst based on the most significant bit of the +// corresponding packed double-precision (64-bit) floating-point element in a. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_movemask_pd +FORCE_INLINE int _mm_movemask_pd(__m128d a) +{ + uint64x2_t input = vreinterpretq_u64_m128d(a); + uint64x2_t high_bits = vshrq_n_u64(input, 63); + return vgetq_lane_u64(high_bits, 0) | (vgetq_lane_u64(high_bits, 1) << 1); +} + +// Copy the lower 64-bit integer in a to dst. +// +// dst[63:0] := a[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_movepi64_pi64 +FORCE_INLINE __m64 _mm_movepi64_pi64(__m128i a) +{ + return vreinterpret_m64_s64(vget_low_s64(vreinterpretq_s64_m128i(a))); +} + +// Copy the 64-bit integer a to the lower element of dst, and zero the upper +// element. +// +// dst[63:0] := a[63:0] +// dst[127:64] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_movpi64_epi64 +FORCE_INLINE __m128i _mm_movpi64_epi64(__m64 a) +{ + return vreinterpretq_m128i_s64( + vcombine_s64(vreinterpret_s64_m64(a), vdup_n_s64(0))); +} + +// Multiply the low unsigned 32-bit integers from each packed 64-bit element in +// a and b, and store the unsigned 64-bit results in dst. +// +// r0 := (a0 & 0xFFFFFFFF) * (b0 & 0xFFFFFFFF) +// r1 := (a2 & 0xFFFFFFFF) * (b2 & 0xFFFFFFFF) +FORCE_INLINE __m128i _mm_mul_epu32(__m128i a, __m128i b) +{ + // vmull_u32 upcasts instead of masking, so we downcast. + uint32x2_t a_lo = vmovn_u64(vreinterpretq_u64_m128i(a)); + uint32x2_t b_lo = vmovn_u64(vreinterpretq_u64_m128i(b)); + return vreinterpretq_m128i_u64(vmull_u32(a_lo, b_lo)); +} + +// Multiply packed double-precision (64-bit) floating-point elements in a and b, +// and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mul_pd +FORCE_INLINE __m128d _mm_mul_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vmulq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + double *da = (double *)&a; + double *db = (double *)&b; + double c[2]; + c[0] = da[0] * db[0]; + c[1] = da[1] * db[1]; + return vld1q_f32((float32_t *)c); +#endif +} + +// Multiply the lower double-precision (64-bit) floating-point element in a and +// b, store the result in the lower element of dst, and copy the upper element +// from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_mul_sd +FORCE_INLINE __m128d _mm_mul_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_mul_pd(a, b)); +} + +// Multiply the low unsigned 32-bit integers from a and b, and store the +// unsigned 64-bit result in dst. +// +// dst[63:0] := a[31:0] * b[31:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mul_su32 +FORCE_INLINE __m64 _mm_mul_su32(__m64 a, __m64 b) +{ + return vreinterpret_m64_u64(vget_low_u64( + vmull_u32(vreinterpret_u32_m64(a), vreinterpret_u32_m64(b)))); +} + +// Multiplies the 8 signed 16-bit integers from a by the 8 signed 16-bit +// integers from b. +// +// r0 := (a0 * b0)[31:16] +// r1 := (a1 * b1)[31:16] +// ... +// r7 := (a7 * b7)[31:16] +// +// https://msdn.microsoft.com/en-us/library/vstudio/59hddw1d(v=vs.100).aspx +FORCE_INLINE __m128i _mm_mulhi_epi16(__m128i a, __m128i b) +{ + /* FIXME: issue with large values because of result saturation */ + // int16x8_t ret = vqdmulhq_s16(vreinterpretq_s16_m128i(a), + // vreinterpretq_s16_m128i(b)); /* =2*a*b */ return + // vreinterpretq_m128i_s16(vshrq_n_s16(ret, 1)); + int16x4_t a3210 = vget_low_s16(vreinterpretq_s16_m128i(a)); + int16x4_t b3210 = vget_low_s16(vreinterpretq_s16_m128i(b)); + int32x4_t ab3210 = vmull_s16(a3210, b3210); /* 3333222211110000 */ + int16x4_t a7654 = vget_high_s16(vreinterpretq_s16_m128i(a)); + int16x4_t b7654 = vget_high_s16(vreinterpretq_s16_m128i(b)); + int32x4_t ab7654 = vmull_s16(a7654, b7654); /* 7777666655554444 */ + uint16x8x2_t r = + vuzpq_u16(vreinterpretq_u16_s32(ab3210), vreinterpretq_u16_s32(ab7654)); + return vreinterpretq_m128i_u16(r.val[1]); +} + +// Multiply the packed unsigned 16-bit integers in a and b, producing +// intermediate 32-bit integers, and store the high 16 bits of the intermediate +// integers in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mulhi_epu16 +FORCE_INLINE __m128i _mm_mulhi_epu16(__m128i a, __m128i b) +{ + uint16x4_t a3210 = vget_low_u16(vreinterpretq_u16_m128i(a)); + uint16x4_t b3210 = vget_low_u16(vreinterpretq_u16_m128i(b)); + uint32x4_t ab3210 = vmull_u16(a3210, b3210); +#if defined(__aarch64__) + uint32x4_t ab7654 = + vmull_high_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b)); + uint16x8_t r = vuzp2q_u16(vreinterpretq_u16_u32(ab3210), + vreinterpretq_u16_u32(ab7654)); + return vreinterpretq_m128i_u16(r); +#else + uint16x4_t a7654 = vget_high_u16(vreinterpretq_u16_m128i(a)); + uint16x4_t b7654 = vget_high_u16(vreinterpretq_u16_m128i(b)); + uint32x4_t ab7654 = vmull_u16(a7654, b7654); + uint16x8x2_t r = + vuzpq_u16(vreinterpretq_u16_u32(ab3210), vreinterpretq_u16_u32(ab7654)); + return vreinterpretq_m128i_u16(r.val[1]); +#endif +} + +// Multiplies the 8 signed or unsigned 16-bit integers from a by the 8 signed or +// unsigned 16-bit integers from b. +// +// r0 := (a0 * b0)[15:0] +// r1 := (a1 * b1)[15:0] +// ... +// r7 := (a7 * b7)[15:0] +// +// https://msdn.microsoft.com/en-us/library/vstudio/9ks1472s(v=vs.100).aspx +FORCE_INLINE __m128i _mm_mullo_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vmulq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Compute the bitwise OR of packed double-precision (64-bit) floating-point +// elements in a and b, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_or_pd +FORCE_INLINE __m128d _mm_or_pd(__m128d a, __m128d b) +{ + return vreinterpretq_m128d_s64( + vorrq_s64(vreinterpretq_s64_m128d(a), vreinterpretq_s64_m128d(b))); +} + +// Computes the bitwise OR of the 128-bit value in a and the 128-bit value in b. +// +// r := a | b +// +// https://msdn.microsoft.com/en-us/library/vstudio/ew8ty0db(v=vs.100).aspx +FORCE_INLINE __m128i _mm_or_si128(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + vorrq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Packs the 16 signed 16-bit integers from a and b into 8-bit integers and +// saturates. +// https://msdn.microsoft.com/en-us/library/k4y4f7w5%28v=vs.90%29.aspx +FORCE_INLINE __m128i _mm_packs_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s8( + vcombine_s8(vqmovn_s16(vreinterpretq_s16_m128i(a)), + vqmovn_s16(vreinterpretq_s16_m128i(b)))); +} + +// Packs the 8 signed 32-bit integers from a and b into signed 16-bit integers +// and saturates. +// +// r0 := SignedSaturate(a0) +// r1 := SignedSaturate(a1) +// r2 := SignedSaturate(a2) +// r3 := SignedSaturate(a3) +// r4 := SignedSaturate(b0) +// r5 := SignedSaturate(b1) +// r6 := SignedSaturate(b2) +// r7 := SignedSaturate(b3) +// +// https://msdn.microsoft.com/en-us/library/393t56f9%28v=vs.90%29.aspx +FORCE_INLINE __m128i _mm_packs_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vcombine_s16(vqmovn_s32(vreinterpretq_s32_m128i(a)), + vqmovn_s32(vreinterpretq_s32_m128i(b)))); +} + +// Packs the 16 signed 16 - bit integers from a and b into 8 - bit unsigned +// integers and saturates. +// +// r0 := UnsignedSaturate(a0) +// r1 := UnsignedSaturate(a1) +// ... +// r7 := UnsignedSaturate(a7) +// r8 := UnsignedSaturate(b0) +// r9 := UnsignedSaturate(b1) +// ... +// r15 := UnsignedSaturate(b7) +// +// https://msdn.microsoft.com/en-us/library/07ad1wx4(v=vs.100).aspx +FORCE_INLINE __m128i _mm_packus_epi16(const __m128i a, const __m128i b) +{ + return vreinterpretq_m128i_u8( + vcombine_u8(vqmovun_s16(vreinterpretq_s16_m128i(a)), + vqmovun_s16(vreinterpretq_s16_m128i(b)))); +} + +// Pause the processor. This is typically used in spin-wait loops and depending +// on the x86 processor typical values are in the 40-100 cycle range. The +// 'yield' instruction isn't a good fit beacuse it's effectively a nop on most +// Arm cores. Experience with several databases has shown has shown an 'isb' is +// a reasonable approximation. +FORCE_INLINE void _mm_pause() +{ + __asm__ __volatile__("isb\n"); +} + +// Compute the absolute differences of packed unsigned 8-bit integers in a and +// b, then horizontally sum each consecutive 8 differences to produce two +// unsigned 16-bit integers, and pack these unsigned 16-bit integers in the low +// 16 bits of 64-bit elements in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sad_epu8 +FORCE_INLINE __m128i _mm_sad_epu8(__m128i a, __m128i b) +{ + uint16x8_t t = vpaddlq_u8(vabdq_u8((uint8x16_t)a, (uint8x16_t)b)); + return vreinterpretq_m128i_u64(vpaddlq_u32(vpaddlq_u16(t))); +} + +// Sets the 8 signed 16-bit integer values. +// https://msdn.microsoft.com/en-au/library/3e0fek84(v=vs.90).aspx +FORCE_INLINE __m128i _mm_set_epi16(short i7, short i6, short i5, short i4, + short i3, short i2, short i1, short i0) +{ + int16_t ALIGN_STRUCT(16) data[8] = { i0, i1, i2, i3, i4, i5, i6, i7 }; + return vreinterpretq_m128i_s16(vld1q_s16(data)); +} + +// Sets the 4 signed 32-bit integer values. +// https://msdn.microsoft.com/en-us/library/vstudio/019beekt(v=vs.100).aspx +FORCE_INLINE __m128i _mm_set_epi32(int i3, int i2, int i1, int i0) +{ + int32_t ALIGN_STRUCT(16) data[4] = { i0, i1, i2, i3 }; + return vreinterpretq_m128i_s32(vld1q_s32(data)); +} + +// Returns the __m128i structure with its two 64-bit integer values +// initialized to the values of the two 64-bit integers passed in. +// https://msdn.microsoft.com/en-us/library/dk2sdw0h(v=vs.120).aspx +FORCE_INLINE __m128i _mm_set_epi64(__m64 i1, __m64 i2) +{ + return _mm_set_epi64x((int64_t)i1, (int64_t)i2); +} + +// Returns the __m128i structure with its two 64-bit integer values +// initialized to the values of the two 64-bit integers passed in. +// https://msdn.microsoft.com/en-us/library/dk2sdw0h(v=vs.120).aspx +FORCE_INLINE __m128i _mm_set_epi64x(int64_t i1, int64_t i2) +{ + return vreinterpretq_m128i_s64( + vcombine_s64(vcreate_s64(i2), vcreate_s64(i1))); +} + +// Sets the 16 signed 8-bit integer values. +// https://msdn.microsoft.com/en-us/library/x0cx8zd3(v=vs.90).aspx +FORCE_INLINE __m128i +_mm_set_epi8(signed char b15, signed char b14, signed char b13, signed char b12, + signed char b11, signed char b10, signed char b9, signed char b8, + signed char b7, signed char b6, signed char b5, signed char b4, + signed char b3, signed char b2, signed char b1, signed char b0) +{ + int8_t ALIGN_STRUCT(16) + data[16] = { (int8_t)b0, (int8_t)b1, (int8_t)b2, (int8_t)b3, + (int8_t)b4, (int8_t)b5, (int8_t)b6, (int8_t)b7, + (int8_t)b8, (int8_t)b9, (int8_t)b10, (int8_t)b11, + (int8_t)b12, (int8_t)b13, (int8_t)b14, (int8_t)b15 }; + return (__m128i)vld1q_s8(data); +} + +// Set packed double-precision (64-bit) floating-point elements in dst with the +// supplied values. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set_pd +FORCE_INLINE __m128d _mm_set_pd(double e1, double e0) +{ + double ALIGN_STRUCT(16) data[2] = { e0, e1 }; +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vld1q_f64((float64_t *)data)); +#else + return vreinterpretq_m128d_f32(vld1q_f32((float32_t *)data)); +#endif +} + +// Broadcast double-precision (64-bit) floating-point value a to all elements of +// dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set_pd1 +#define _mm_set_pd1 _mm_set1_pd + +// Copy double-precision (64-bit) floating-point element a to the lower element +// of dst, and zero the upper element. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set_sd +FORCE_INLINE __m128d _mm_set_sd(double a) +{ + return _mm_set_pd(0, a); +} + +// Sets the 8 signed 16-bit integer values to w. +// +// r0 := w +// r1 := w +// ... +// r7 := w +// +// https://msdn.microsoft.com/en-us/library/k0ya3x0e(v=vs.90).aspx +FORCE_INLINE __m128i _mm_set1_epi16(short w) +{ + return vreinterpretq_m128i_s16(vdupq_n_s16(w)); +} + +// Sets the 4 signed 32-bit integer values to i. +// +// r0 := i +// r1 := i +// r2 := i +// r3 := I +// +// https://msdn.microsoft.com/en-us/library/vstudio/h4xscxat(v=vs.100).aspx +FORCE_INLINE __m128i _mm_set1_epi32(int _i) +{ + return vreinterpretq_m128i_s32(vdupq_n_s32(_i)); +} + +// Sets the 2 signed 64-bit integer values to i. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/whtfzhzk(v=vs.100) +FORCE_INLINE __m128i _mm_set1_epi64(__m64 _i) +{ + return vreinterpretq_m128i_s64(vdupq_n_s64((int64_t)_i)); +} + +// Sets the 2 signed 64-bit integer values to i. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set1_epi64x +FORCE_INLINE __m128i _mm_set1_epi64x(int64_t _i) +{ + return vreinterpretq_m128i_s64(vdupq_n_s64(_i)); +} + +// Sets the 16 signed 8-bit integer values to b. +// +// r0 := b +// r1 := b +// ... +// r15 := b +// +// https://msdn.microsoft.com/en-us/library/6e14xhyf(v=vs.100).aspx +FORCE_INLINE __m128i _mm_set1_epi8(signed char w) +{ + return vreinterpretq_m128i_s8(vdupq_n_s8(w)); +} + +// Broadcast double-precision (64-bit) floating-point value a to all elements of +// dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set1_pd +FORCE_INLINE __m128d _mm_set1_pd(double d) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vdupq_n_f64(d)); +#else + return vreinterpretq_m128d_s64(vdupq_n_s64(*(int64_t *)&d)); +#endif +} + +// Sets the 8 signed 16-bit integer values in reverse order. +// +// Return Value +// r0 := w0 +// r1 := w1 +// ... +// r7 := w7 +FORCE_INLINE __m128i _mm_setr_epi16(short w0, short w1, short w2, short w3, + short w4, short w5, short w6, short w7) +{ + int16_t ALIGN_STRUCT(16) data[8] = { w0, w1, w2, w3, w4, w5, w6, w7 }; + return vreinterpretq_m128i_s16(vld1q_s16((int16_t *)data)); +} + +// Sets the 4 signed 32-bit integer values in reverse order +// https://technet.microsoft.com/en-us/library/security/27yb3ee5(v=vs.90).aspx +FORCE_INLINE __m128i _mm_setr_epi32(int i3, int i2, int i1, int i0) +{ + int32_t ALIGN_STRUCT(16) data[4] = { i3, i2, i1, i0 }; + return vreinterpretq_m128i_s32(vld1q_s32(data)); +} + +// Set packed 64-bit integers in dst with the supplied values in reverse order. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_setr_epi64 +FORCE_INLINE __m128i _mm_setr_epi64(__m64 e1, __m64 e0) +{ + return vreinterpretq_m128i_s64(vcombine_s64(e1, e0)); +} + +// Sets the 16 signed 8-bit integer values in reverse order. +// https://msdn.microsoft.com/en-us/library/2khb9c7k(v=vs.90).aspx +FORCE_INLINE __m128i _mm_setr_epi8( + signed char b0, signed char b1, signed char b2, signed char b3, + signed char b4, signed char b5, signed char b6, signed char b7, + signed char b8, signed char b9, signed char b10, signed char b11, + signed char b12, signed char b13, signed char b14, signed char b15) +{ + int8_t ALIGN_STRUCT(16) + data[16] = { (int8_t)b0, (int8_t)b1, (int8_t)b2, (int8_t)b3, + (int8_t)b4, (int8_t)b5, (int8_t)b6, (int8_t)b7, + (int8_t)b8, (int8_t)b9, (int8_t)b10, (int8_t)b11, + (int8_t)b12, (int8_t)b13, (int8_t)b14, (int8_t)b15 }; + return (__m128i)vld1q_s8(data); +} + +// Set packed double-precision (64-bit) floating-point elements in dst with the +// supplied values in reverse order. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_setr_pd +FORCE_INLINE __m128d _mm_setr_pd(double e1, double e0) +{ + return _mm_set_pd(e0, e1); +} + +// Return vector of type __m128d with all elements set to zero. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_setzero_pd +FORCE_INLINE __m128d _mm_setzero_pd(void) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vdupq_n_f64(0)); +#else + return vreinterpretq_m128d_f32(vdupq_n_f32(0)); +#endif +} + +// Sets the 128-bit value to zero +// https://msdn.microsoft.com/en-us/library/vstudio/ys7dw0kh(v=vs.100).aspx +FORCE_INLINE __m128i _mm_setzero_si128(void) +{ + return vreinterpretq_m128i_s32(vdupq_n_s32(0)); +} + +// Shuffles the 4 signed or unsigned 32-bit integers in a as specified by imm. +// https://msdn.microsoft.com/en-us/library/56f67xbk%28v=vs.90%29.aspx +// FORCE_INLINE __m128i _mm_shuffle_epi32(__m128i a, +// __constrange(0,255) int imm) +#if __has_builtin(__builtin_shufflevector) +#define _mm_shuffle_epi32(a, imm) \ + __extension__({ \ + int32x4_t _input = vreinterpretq_s32_m128i(a); \ + int32x4_t _shuf = __builtin_shufflevector( \ + _input, _input, (imm) & (0x3), ((imm) >> 2) & 0x3, \ + ((imm) >> 4) & 0x3, ((imm) >> 6) & 0x3); \ + vreinterpretq_m128i_s32(_shuf); \ + }) +#else // generic +#define _mm_shuffle_epi32(a, imm) \ + __extension__({ \ + __m128i ret; \ + switch (imm) { \ + case _MM_SHUFFLE(1, 0, 3, 2): \ + ret = _mm_shuffle_epi_1032((a)); \ + break; \ + case _MM_SHUFFLE(2, 3, 0, 1): \ + ret = _mm_shuffle_epi_2301((a)); \ + break; \ + case _MM_SHUFFLE(0, 3, 2, 1): \ + ret = _mm_shuffle_epi_0321((a)); \ + break; \ + case _MM_SHUFFLE(2, 1, 0, 3): \ + ret = _mm_shuffle_epi_2103((a)); \ + break; \ + case _MM_SHUFFLE(1, 0, 1, 0): \ + ret = _mm_shuffle_epi_1010((a)); \ + break; \ + case _MM_SHUFFLE(1, 0, 0, 1): \ + ret = _mm_shuffle_epi_1001((a)); \ + break; \ + case _MM_SHUFFLE(0, 1, 0, 1): \ + ret = _mm_shuffle_epi_0101((a)); \ + break; \ + case _MM_SHUFFLE(2, 2, 1, 1): \ + ret = _mm_shuffle_epi_2211((a)); \ + break; \ + case _MM_SHUFFLE(0, 1, 2, 2): \ + ret = _mm_shuffle_epi_0122((a)); \ + break; \ + case _MM_SHUFFLE(3, 3, 3, 2): \ + ret = _mm_shuffle_epi_3332((a)); \ + break; \ + case _MM_SHUFFLE(0, 0, 0, 0): \ + ret = _mm_shuffle_epi32_splat((a), 0); \ + break; \ + case _MM_SHUFFLE(1, 1, 1, 1): \ + ret = _mm_shuffle_epi32_splat((a), 1); \ + break; \ + case _MM_SHUFFLE(2, 2, 2, 2): \ + ret = _mm_shuffle_epi32_splat((a), 2); \ + break; \ + case _MM_SHUFFLE(3, 3, 3, 3): \ + ret = _mm_shuffle_epi32_splat((a), 3); \ + break; \ + default: \ + ret = _mm_shuffle_epi32_default((a), (imm)); \ + break; \ + } \ + ret; \ + }) +#endif + +// Shuffle double-precision (64-bit) floating-point elements using the control +// in imm8, and store the results in dst. +// +// dst[63:0] := (imm8[0] == 0) ? a[63:0] : a[127:64] +// dst[127:64] := (imm8[1] == 0) ? b[63:0] : b[127:64] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_shuffle_pd +#if __has_builtin(__builtin_shufflevector) +#define _mm_shuffle_pd(a, b, imm8) \ + vreinterpretq_m128d_s64(__builtin_shufflevector( \ + vreinterpretq_s64_m128d(a), vreinterpretq_s64_m128d(b), imm8 & 0x1, \ + ((imm8 & 0x2) >> 1) + 2)) +#else +#define _mm_shuffle_pd(a, b, imm8) \ + _mm_castsi128_pd(_mm_set_epi64x( \ + vgetq_lane_s64(vreinterpretq_s64_m128d(b), (imm8 & 0x2) >> 1), \ + vgetq_lane_s64(vreinterpretq_s64_m128d(a), imm8 & 0x1))) +#endif + +// FORCE_INLINE __m128i _mm_shufflehi_epi16(__m128i a, +// __constrange(0,255) int imm) +#if __has_builtin(__builtin_shufflevector) +#define _mm_shufflehi_epi16(a, imm) \ + __extension__({ \ + int16x8_t _input = vreinterpretq_s16_m128i(a); \ + int16x8_t _shuf = __builtin_shufflevector(_input, _input, 0, 1, 2, 3, \ + ((imm) & (0x3)) + 4, \ + (((imm) >> 2) & 0x3) + 4, \ + (((imm) >> 4) & 0x3) + 4, \ + (((imm) >> 6) & 0x3) + 4); \ + vreinterpretq_m128i_s16(_shuf); \ + }) +#else // generic +#define _mm_shufflehi_epi16(a, imm) _mm_shufflehi_epi16_function((a), (imm)) +#endif + +// FORCE_INLINE __m128i _mm_shufflelo_epi16(__m128i a, +// __constrange(0,255) int imm) +#if __has_builtin(__builtin_shufflevector) +#define _mm_shufflelo_epi16(a, imm) \ + __extension__({ \ + int16x8_t _input = vreinterpretq_s16_m128i(a); \ + int16x8_t _shuf = __builtin_shufflevector( \ + _input, _input, ((imm) & (0x3)), (((imm) >> 2) & 0x3), \ + (((imm) >> 4) & 0x3), (((imm) >> 6) & 0x3), 4, 5, 6, 7); \ + vreinterpretq_m128i_s16(_shuf); \ + }) +#else // generic +#define _mm_shufflelo_epi16(a, imm) _mm_shufflelo_epi16_function((a), (imm)) +#endif + +// Shift packed 16-bit integers in a left by count while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// IF count[63:0] > 15 +// dst[i+15:i] := 0 +// ELSE +// dst[i+15:i] := ZeroExtend16(a[i+15:i] << count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sll_epi16 +FORCE_INLINE __m128i _mm_sll_epi16(__m128i a, __m128i count) +{ + uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); + if (_sse2neon_unlikely(c & ~15)) + return _mm_setzero_si128(); + + int16x8_t vc = vdupq_n_s16((int16_t)c); + return vreinterpretq_m128i_s16(vshlq_s16(vreinterpretq_s16_m128i(a), vc)); +} + +// Shift packed 32-bit integers in a left by count while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// IF count[63:0] > 31 +// dst[i+31:i] := 0 +// ELSE +// dst[i+31:i] := ZeroExtend32(a[i+31:i] << count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sll_epi32 +FORCE_INLINE __m128i _mm_sll_epi32(__m128i a, __m128i count) +{ + uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); + if (_sse2neon_unlikely(c & ~31)) + return _mm_setzero_si128(); + + int32x4_t vc = vdupq_n_s32((int32_t)c); + return vreinterpretq_m128i_s32(vshlq_s32(vreinterpretq_s32_m128i(a), vc)); +} + +// Shift packed 64-bit integers in a left by count while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// IF count[63:0] > 63 +// dst[i+63:i] := 0 +// ELSE +// dst[i+63:i] := ZeroExtend64(a[i+63:i] << count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sll_epi64 +FORCE_INLINE __m128i _mm_sll_epi64(__m128i a, __m128i count) +{ + uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); + if (_sse2neon_unlikely(c & ~63)) + return _mm_setzero_si128(); + + int64x2_t vc = vdupq_n_s64((int64_t)c); + return vreinterpretq_m128i_s64(vshlq_s64(vreinterpretq_s64_m128i(a), vc)); +} + +// Shift packed 16-bit integers in a left by imm8 while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// IF imm8[7:0] > 15 +// dst[i+15:i] := 0 +// ELSE +// dst[i+15:i] := ZeroExtend16(a[i+15:i] << imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_slli_epi16 +FORCE_INLINE __m128i _mm_slli_epi16(__m128i a, int imm) +{ + if (_sse2neon_unlikely(imm & ~15)) + return _mm_setzero_si128(); + return vreinterpretq_m128i_s16( + vshlq_s16(vreinterpretq_s16_m128i(a), vdupq_n_s16(imm))); +} + +// Shift packed 32-bit integers in a left by imm8 while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// IF imm8[7:0] > 31 +// dst[i+31:i] := 0 +// ELSE +// dst[i+31:i] := ZeroExtend32(a[i+31:i] << imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_slli_epi32 +FORCE_INLINE __m128i _mm_slli_epi32(__m128i a, int imm) +{ + if (_sse2neon_unlikely(imm & ~31)) + return _mm_setzero_si128(); + return vreinterpretq_m128i_s32( + vshlq_s32(vreinterpretq_s32_m128i(a), vdupq_n_s32(imm))); +} + +// Shift packed 64-bit integers in a left by imm8 while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// IF imm8[7:0] > 63 +// dst[i+63:i] := 0 +// ELSE +// dst[i+63:i] := ZeroExtend64(a[i+63:i] << imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_slli_epi64 +FORCE_INLINE __m128i _mm_slli_epi64(__m128i a, int imm) +{ + if (_sse2neon_unlikely(imm & ~63)) + return _mm_setzero_si128(); + return vreinterpretq_m128i_s64( + vshlq_s64(vreinterpretq_s64_m128i(a), vdupq_n_s64(imm))); +} + +// Shift a left by imm8 bytes while shifting in zeros, and store the results in +// dst. +// +// tmp := imm8[7:0] +// IF tmp > 15 +// tmp := 16 +// FI +// dst[127:0] := a[127:0] << (tmp*8) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_slli_si128 +FORCE_INLINE __m128i _mm_slli_si128(__m128i a, int imm) +{ + if (_sse2neon_unlikely(imm & ~15)) + return _mm_setzero_si128(); + uint8x16_t tmp[2] = { vdupq_n_u8(0), vreinterpretq_u8_m128i(a) }; + return vreinterpretq_m128i_u8( + vld1q_u8(((uint8_t const *)tmp) + (16 - imm))); +} + +// Compute the square root of packed double-precision (64-bit) floating-point +// elements in a, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sqrt_pd +FORCE_INLINE __m128d _mm_sqrt_pd(__m128d a) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vsqrtq_f64(vreinterpretq_f64_m128d(a))); +#else + double a0 = sqrt(((double *)&a)[0]); + double a1 = sqrt(((double *)&a)[1]); + return _mm_set_pd(a1, a0); +#endif +} + +// Compute the square root of the lower double-precision (64-bit) floating-point +// element in b, store the result in the lower element of dst, and copy the +// upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sqrt_sd +FORCE_INLINE __m128d _mm_sqrt_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_sqrt_pd(b)); +#else + return _mm_set_pd(((double *)&a)[1], sqrt(((double *)&b)[0])); +#endif +} + +// Shift packed 16-bit integers in a right by count while shifting in sign bits, +// and store the results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// IF count[63:0] > 15 +// dst[i+15:i] := (a[i+15] ? 0xFFFF : 0x0) +// ELSE +// dst[i+15:i] := SignExtend16(a[i+15:i] >> count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sra_epi16 +FORCE_INLINE __m128i _mm_sra_epi16(__m128i a, __m128i count) +{ + int64_t c = (int64_t)vget_low_s64((int64x2_t)count); + if (_sse2neon_unlikely(c & ~15)) + return _mm_cmplt_epi16(a, _mm_setzero_si128()); + return vreinterpretq_m128i_s16(vshlq_s16((int16x8_t)a, vdupq_n_s16(-c))); +} + +// Shift packed 32-bit integers in a right by count while shifting in sign bits, +// and store the results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// IF count[63:0] > 31 +// dst[i+31:i] := (a[i+31] ? 0xFFFFFFFF : 0x0) +// ELSE +// dst[i+31:i] := SignExtend32(a[i+31:i] >> count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sra_epi32 +FORCE_INLINE __m128i _mm_sra_epi32(__m128i a, __m128i count) +{ + int64_t c = (int64_t)vget_low_s64((int64x2_t)count); + if (_sse2neon_unlikely(c & ~31)) + return _mm_cmplt_epi32(a, _mm_setzero_si128()); + return vreinterpretq_m128i_s32(vshlq_s32((int32x4_t)a, vdupq_n_s32(-c))); +} + +// Shift packed 16-bit integers in a right by imm8 while shifting in sign +// bits, and store the results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// IF imm8[7:0] > 15 +// dst[i+15:i] := (a[i+15] ? 0xFFFF : 0x0) +// ELSE +// dst[i+15:i] := SignExtend16(a[i+15:i] >> imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srai_epi16 +FORCE_INLINE __m128i _mm_srai_epi16(__m128i a, int imm) +{ + const int count = (imm & ~15) ? 15 : imm; + return (__m128i)vshlq_s16((int16x8_t)a, vdupq_n_s16(-count)); +} + +// Shift packed 32-bit integers in a right by imm8 while shifting in sign bits, +// and store the results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// IF imm8[7:0] > 31 +// dst[i+31:i] := (a[i+31] ? 0xFFFFFFFF : 0x0) +// ELSE +// dst[i+31:i] := SignExtend32(a[i+31:i] >> imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srai_epi32 +// FORCE_INLINE __m128i _mm_srai_epi32(__m128i a, __constrange(0,255) int imm) +#define _mm_srai_epi32(a, imm) \ + __extension__({ \ + __m128i ret; \ + if (_sse2neon_unlikely((imm) == 0)) { \ + ret = a; \ + } else if (_sse2neon_likely(0 < (imm) && (imm) < 32)) { \ + ret = vreinterpretq_m128i_s32( \ + vshlq_s32(vreinterpretq_s32_m128i(a), vdupq_n_s32(-imm))); \ + } else { \ + ret = vreinterpretq_m128i_s32( \ + vshrq_n_s32(vreinterpretq_s32_m128i(a), 31)); \ + } \ + ret; \ + }) + +// Shift packed 16-bit integers in a right by count while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// IF count[63:0] > 15 +// dst[i+15:i] := 0 +// ELSE +// dst[i+15:i] := ZeroExtend16(a[i+15:i] >> count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srl_epi16 +FORCE_INLINE __m128i _mm_srl_epi16(__m128i a, __m128i count) +{ + uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); + if (_sse2neon_unlikely(c & ~15)) + return _mm_setzero_si128(); + + int16x8_t vc = vdupq_n_s16(-(int16_t)c); + return vreinterpretq_m128i_u16(vshlq_u16(vreinterpretq_u16_m128i(a), vc)); +} + +// Shift packed 32-bit integers in a right by count while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// IF count[63:0] > 31 +// dst[i+31:i] := 0 +// ELSE +// dst[i+31:i] := ZeroExtend32(a[i+31:i] >> count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srl_epi32 +FORCE_INLINE __m128i _mm_srl_epi32(__m128i a, __m128i count) +{ + uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); + if (_sse2neon_unlikely(c & ~31)) + return _mm_setzero_si128(); + + int32x4_t vc = vdupq_n_s32(-(int32_t)c); + return vreinterpretq_m128i_u32(vshlq_u32(vreinterpretq_u32_m128i(a), vc)); +} + +// Shift packed 64-bit integers in a right by count while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// IF count[63:0] > 63 +// dst[i+63:i] := 0 +// ELSE +// dst[i+63:i] := ZeroExtend64(a[i+63:i] >> count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srl_epi64 +FORCE_INLINE __m128i _mm_srl_epi64(__m128i a, __m128i count) +{ + uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); + if (_sse2neon_unlikely(c & ~63)) + return _mm_setzero_si128(); + + int64x2_t vc = vdupq_n_s64(-(int64_t)c); + return vreinterpretq_m128i_u64(vshlq_u64(vreinterpretq_u64_m128i(a), vc)); +} + +// Shift packed 16-bit integers in a right by imm8 while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// IF imm8[7:0] > 15 +// dst[i+15:i] := 0 +// ELSE +// dst[i+15:i] := ZeroExtend16(a[i+15:i] >> imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srli_epi16 +#define _mm_srli_epi16(a, imm) \ + __extension__({ \ + __m128i ret; \ + if (_sse2neon_unlikely(imm & ~15)) { \ + ret = _mm_setzero_si128(); \ + } else { \ + ret = vreinterpretq_m128i_u16( \ + vshlq_u16(vreinterpretq_u16_m128i(a), vdupq_n_s16(-imm))); \ + } \ + ret; \ + }) + +// Shift packed 32-bit integers in a right by imm8 while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// IF imm8[7:0] > 31 +// dst[i+31:i] := 0 +// ELSE +// dst[i+31:i] := ZeroExtend32(a[i+31:i] >> imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srli_epi32 +// FORCE_INLINE __m128i _mm_srli_epi32(__m128i a, __constrange(0,255) int imm) +#define _mm_srli_epi32(a, imm) \ + __extension__({ \ + __m128i ret; \ + if (_sse2neon_unlikely(imm & ~31)) { \ + ret = _mm_setzero_si128(); \ + } else { \ + ret = vreinterpretq_m128i_u32( \ + vshlq_u32(vreinterpretq_u32_m128i(a), vdupq_n_s32(-imm))); \ + } \ + ret; \ + }) + +// Shift packed 64-bit integers in a right by imm8 while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// IF imm8[7:0] > 63 +// dst[i+63:i] := 0 +// ELSE +// dst[i+63:i] := ZeroExtend64(a[i+63:i] >> imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srli_epi64 +#define _mm_srli_epi64(a, imm) \ + __extension__({ \ + __m128i ret; \ + if (_sse2neon_unlikely(imm & ~63)) { \ + ret = _mm_setzero_si128(); \ + } else { \ + ret = vreinterpretq_m128i_u64( \ + vshlq_u64(vreinterpretq_u64_m128i(a), vdupq_n_s64(-imm))); \ + } \ + ret; \ + }) + +// Shift a right by imm8 bytes while shifting in zeros, and store the results in +// dst. +// +// tmp := imm8[7:0] +// IF tmp > 15 +// tmp := 16 +// FI +// dst[127:0] := a[127:0] >> (tmp*8) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srli_si128 +FORCE_INLINE __m128i _mm_srli_si128(__m128i a, int imm) +{ + if (_sse2neon_unlikely(imm & ~15)) + return _mm_setzero_si128(); + uint8x16_t tmp[2] = { vreinterpretq_u8_m128i(a), vdupq_n_u8(0) }; + return vreinterpretq_m128i_u8(vld1q_u8(((uint8_t const *)tmp) + imm)); +} + +// Store 128-bits (composed of 2 packed double-precision (64-bit) floating-point +// elements) from a into memory. mem_addr must be aligned on a 16-byte boundary +// or a general-protection exception may be generated. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_store_pd +FORCE_INLINE void _mm_store_pd(double *mem_addr, __m128d a) +{ +#if defined(__aarch64__) + vst1q_f64((float64_t *)mem_addr, vreinterpretq_f64_m128d(a)); +#else + vst1q_f32((float32_t *)mem_addr, vreinterpretq_f32_m128d(a)); +#endif +} + +// Store the lower double-precision (64-bit) floating-point element from a into +// 2 contiguous elements in memory. mem_addr must be aligned on a 16-byte +// boundary or a general-protection exception may be generated. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_store_pd1 +FORCE_INLINE void _mm_store_pd1(double *mem_addr, __m128d a) +{ +#if defined(__aarch64__) + float64x1_t a_low = vget_low_f64(vreinterpretq_f64_m128d(a)); + vst1q_f64((float64_t *)mem_addr, + vreinterpretq_f64_m128d(vcombine_f64(a_low, a_low))); +#else + float32x2_t a_low = vget_low_f32(vreinterpretq_f32_m128d(a)); + vst1q_f32((float32_t *)mem_addr, + vreinterpretq_f32_m128d(vcombine_f32(a_low, a_low))); +#endif +} + +// Store the lower double-precision (64-bit) floating-point element from a into +// memory. mem_addr does not need to be aligned on any particular boundary. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_store_sd +FORCE_INLINE void _mm_store_sd(double *mem_addr, __m128d a) +{ +#if defined(__aarch64__) + vst1_f64((float64_t *)mem_addr, vget_low_f64(vreinterpretq_f64_m128d(a))); +#else + vst1_u64((uint64_t *)mem_addr, vget_low_u64(vreinterpretq_u64_m128d(a))); +#endif +} + +// Stores four 32-bit integer values as (as a __m128i value) at the address p. +// https://msdn.microsoft.com/en-us/library/vstudio/edk11s13(v=vs.100).aspx +FORCE_INLINE void _mm_store_si128(__m128i *p, __m128i a) +{ + vst1q_s32((int32_t *)p, vreinterpretq_s32_m128i(a)); +} + +// Store the lower double-precision (64-bit) floating-point element from a into +// 2 contiguous elements in memory. mem_addr must be aligned on a 16-byte +// boundary or a general-protection exception may be generated. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#expand=9,526,5601&text=_mm_store1_pd +#define _mm_store1_pd _mm_store_pd1 + +// Store the upper double-precision (64-bit) floating-point element from a into +// memory. +// +// MEM[mem_addr+63:mem_addr] := a[127:64] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeh_pd +FORCE_INLINE void _mm_storeh_pd(double *mem_addr, __m128d a) +{ +#if defined(__aarch64__) + vst1_f64((float64_t *)mem_addr, vget_high_f64(vreinterpretq_f64_m128d(a))); +#else + vst1_f32((float32_t *)mem_addr, vget_high_f32(vreinterpretq_f32_m128d(a))); +#endif +} + +// Reads the lower 64 bits of b and stores them into the lower 64 bits of a. +// https://msdn.microsoft.com/en-us/library/hhwf428f%28v=vs.90%29.aspx +FORCE_INLINE void _mm_storel_epi64(__m128i *a, __m128i b) +{ + uint64x1_t hi = vget_high_u64(vreinterpretq_u64_m128i(*a)); + uint64x1_t lo = vget_low_u64(vreinterpretq_u64_m128i(b)); + *a = vreinterpretq_m128i_u64(vcombine_u64(lo, hi)); +} + +// Store the lower double-precision (64-bit) floating-point element from a into +// memory. +// +// MEM[mem_addr+63:mem_addr] := a[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storel_pd +FORCE_INLINE void _mm_storel_pd(double *mem_addr, __m128d a) +{ +#if defined(__aarch64__) + vst1_f64((float64_t *)mem_addr, vget_low_f64(vreinterpretq_f64_m128d(a))); +#else + vst1_f32((float32_t *)mem_addr, vget_low_f32(vreinterpretq_f32_m128d(a))); +#endif +} + +// Store 2 double-precision (64-bit) floating-point elements from a into memory +// in reverse order. mem_addr must be aligned on a 16-byte boundary or a +// general-protection exception may be generated. +// +// MEM[mem_addr+63:mem_addr] := a[127:64] +// MEM[mem_addr+127:mem_addr+64] := a[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storer_pd +FORCE_INLINE void _mm_storer_pd(double *mem_addr, __m128d a) +{ + float32x4_t f = vreinterpretq_f32_m128d(a); + _mm_store_pd(mem_addr, vreinterpretq_m128d_f32(vextq_f32(f, f, 2))); +} + +// Store 128-bits (composed of 2 packed double-precision (64-bit) floating-point +// elements) from a into memory. mem_addr does not need to be aligned on any +// particular boundary. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_pd +FORCE_INLINE void _mm_storeu_pd(double *mem_addr, __m128d a) +{ + _mm_store_pd(mem_addr, a); +} + +// Stores 128-bits of integer data a at the address p. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_si128 +FORCE_INLINE void _mm_storeu_si128(__m128i *p, __m128i a) +{ + vst1q_s32((int32_t *)p, vreinterpretq_s32_m128i(a)); +} + +// Stores 32-bits of integer data a at the address p. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_si32 +FORCE_INLINE void _mm_storeu_si32(void *p, __m128i a) +{ + vst1q_lane_s32((int32_t *)p, vreinterpretq_s32_m128i(a), 0); +} + +// Store 128-bits (composed of 2 packed double-precision (64-bit) floating-point +// elements) from a into memory using a non-temporal memory hint. mem_addr must +// be aligned on a 16-byte boundary or a general-protection exception may be +// generated. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_pd +FORCE_INLINE void _mm_stream_pd(double *p, __m128d a) +{ +#if __has_builtin(__builtin_nontemporal_store) + __builtin_nontemporal_store(a, (float32x4_t *)p); +#elif defined(__aarch64__) + vst1q_f64(p, vreinterpretq_f64_m128d(a)); +#else + vst1q_s64((int64_t *)p, vreinterpretq_s64_m128d(a)); +#endif +} + +// Stores the data in a to the address p without polluting the caches. If the +// cache line containing address p is already in the cache, the cache will be +// updated. +// https://msdn.microsoft.com/en-us/library/ba08y07y%28v=vs.90%29.aspx +FORCE_INLINE void _mm_stream_si128(__m128i *p, __m128i a) +{ +#if __has_builtin(__builtin_nontemporal_store) + __builtin_nontemporal_store(a, p); +#else + vst1q_s64((int64_t *)p, vreinterpretq_s64_m128i(a)); +#endif +} + +// Store 32-bit integer a into memory using a non-temporal hint to minimize +// cache pollution. If the cache line containing address mem_addr is already in +// the cache, the cache will be updated. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_si32 +FORCE_INLINE void _mm_stream_si32(int *p, int a) +{ + vst1q_lane_s32((int32_t *)p, vdupq_n_s32(a), 0); +} + +// Subtract packed 16-bit integers in b from packed 16-bit integers in a, and +// store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sub_epi16 +FORCE_INLINE __m128i _mm_sub_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vsubq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Subtracts the 4 signed or unsigned 32-bit integers of b from the 4 signed or +// unsigned 32-bit integers of a. +// +// r0 := a0 - b0 +// r1 := a1 - b1 +// r2 := a2 - b2 +// r3 := a3 - b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/fhh866h0(v=vs.100).aspx +FORCE_INLINE __m128i _mm_sub_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + vsubq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Subtract 2 packed 64-bit integers in b from 2 packed 64-bit integers in a, +// and store the results in dst. +// r0 := a0 - b0 +// r1 := a1 - b1 +FORCE_INLINE __m128i _mm_sub_epi64(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s64( + vsubq_s64(vreinterpretq_s64_m128i(a), vreinterpretq_s64_m128i(b))); +} + +// Subtract packed 8-bit integers in b from packed 8-bit integers in a, and +// store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sub_epi8 +FORCE_INLINE __m128i _mm_sub_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s8( + vsubq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} + +// Subtract packed double-precision (64-bit) floating-point elements in b from +// packed double-precision (64-bit) floating-point elements in a, and store the +// results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// dst[i+63:i] := a[i+63:i] - b[i+63:i] +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_sub_pd +FORCE_INLINE __m128d _mm_sub_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vsubq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + double *da = (double *)&a; + double *db = (double *)&b; + double c[2]; + c[0] = da[0] - db[0]; + c[1] = da[1] - db[1]; + return vld1q_f32((float32_t *)c); +#endif +} + +// Subtract the lower double-precision (64-bit) floating-point element in b from +// the lower double-precision (64-bit) floating-point element in a, store the +// result in the lower element of dst, and copy the upper element from a to the +// upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sub_sd +FORCE_INLINE __m128d _mm_sub_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_sub_pd(a, b)); +} + +// Subtract 64-bit integer b from 64-bit integer a, and store the result in dst. +// +// dst[63:0] := a[63:0] - b[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sub_si64 +FORCE_INLINE __m64 _mm_sub_si64(__m64 a, __m64 b) +{ + return vreinterpret_m64_s64( + vsub_s64(vreinterpret_s64_m64(a), vreinterpret_s64_m64(b))); +} + +// Subtracts the 8 signed 16-bit integers of b from the 8 signed 16-bit integers +// of a and saturates. +// +// r0 := SignedSaturate(a0 - b0) +// r1 := SignedSaturate(a1 - b1) +// ... +// r7 := SignedSaturate(a7 - b7) +// +// https://technet.microsoft.com/en-us/subscriptions/3247z5b8(v=vs.90) +FORCE_INLINE __m128i _mm_subs_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vqsubq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Subtracts the 16 signed 8-bit integers of b from the 16 signed 8-bit integers +// of a and saturates. +// +// r0 := SignedSaturate(a0 - b0) +// r1 := SignedSaturate(a1 - b1) +// ... +// r15 := SignedSaturate(a15 - b15) +// +// https://technet.microsoft.com/en-us/subscriptions/by7kzks1(v=vs.90) +FORCE_INLINE __m128i _mm_subs_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s8( + vqsubq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} + +// Subtracts the 8 unsigned 16-bit integers of bfrom the 8 unsigned 16-bit +// integers of a and saturates.. +// https://technet.microsoft.com/en-us/subscriptions/index/f44y0s19(v=vs.90).aspx +FORCE_INLINE __m128i _mm_subs_epu16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vqsubq_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b))); +} + +// Subtracts the 16 unsigned 8-bit integers of b from the 16 unsigned 8-bit +// integers of a and saturates. +// +// r0 := UnsignedSaturate(a0 - b0) +// r1 := UnsignedSaturate(a1 - b1) +// ... +// r15 := UnsignedSaturate(a15 - b15) +// +// https://technet.microsoft.com/en-us/subscriptions/yadkxc18(v=vs.90) +FORCE_INLINE __m128i _mm_subs_epu8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vqsubq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); +} + +#define _mm_ucomieq_sd _mm_comieq_sd +#define _mm_ucomige_sd _mm_comige_sd +#define _mm_ucomigt_sd _mm_comigt_sd +#define _mm_ucomile_sd _mm_comile_sd +#define _mm_ucomilt_sd _mm_comilt_sd +#define _mm_ucomineq_sd _mm_comineq_sd + +// Return vector of type __m128d with undefined elements. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_undefined_pd +FORCE_INLINE __m128d _mm_undefined_pd(void) +{ +#if defined(__GNUC__) || defined(__clang__) +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wuninitialized" +#endif + __m128d a; + return a; +#if defined(__GNUC__) || defined(__clang__) +#pragma GCC diagnostic pop +#endif +} + +// Interleaves the upper 4 signed or unsigned 16-bit integers in a with the +// upper 4 signed or unsigned 16-bit integers in b. +// +// r0 := a4 +// r1 := b4 +// r2 := a5 +// r3 := b5 +// r4 := a6 +// r5 := b6 +// r6 := a7 +// r7 := b7 +// +// https://msdn.microsoft.com/en-us/library/03196cz7(v=vs.100).aspx +FORCE_INLINE __m128i _mm_unpackhi_epi16(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_s16( + vzip2q_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +#else + int16x4_t a1 = vget_high_s16(vreinterpretq_s16_m128i(a)); + int16x4_t b1 = vget_high_s16(vreinterpretq_s16_m128i(b)); + int16x4x2_t result = vzip_s16(a1, b1); + return vreinterpretq_m128i_s16(vcombine_s16(result.val[0], result.val[1])); +#endif +} + +// Interleaves the upper 2 signed or unsigned 32-bit integers in a with the +// upper 2 signed or unsigned 32-bit integers in b. +// https://msdn.microsoft.com/en-us/library/65sa7cbs(v=vs.100).aspx +FORCE_INLINE __m128i _mm_unpackhi_epi32(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_s32( + vzip2q_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +#else + int32x2_t a1 = vget_high_s32(vreinterpretq_s32_m128i(a)); + int32x2_t b1 = vget_high_s32(vreinterpretq_s32_m128i(b)); + int32x2x2_t result = vzip_s32(a1, b1); + return vreinterpretq_m128i_s32(vcombine_s32(result.val[0], result.val[1])); +#endif +} + +// Interleaves the upper signed or unsigned 64-bit integer in a with the +// upper signed or unsigned 64-bit integer in b. +// +// r0 := a1 +// r1 := b1 +FORCE_INLINE __m128i _mm_unpackhi_epi64(__m128i a, __m128i b) +{ + int64x1_t a_h = vget_high_s64(vreinterpretq_s64_m128i(a)); + int64x1_t b_h = vget_high_s64(vreinterpretq_s64_m128i(b)); + return vreinterpretq_m128i_s64(vcombine_s64(a_h, b_h)); +} + +// Interleaves the upper 8 signed or unsigned 8-bit integers in a with the upper +// 8 signed or unsigned 8-bit integers in b. +// +// r0 := a8 +// r1 := b8 +// r2 := a9 +// r3 := b9 +// ... +// r14 := a15 +// r15 := b15 +// +// https://msdn.microsoft.com/en-us/library/t5h7783k(v=vs.100).aspx +FORCE_INLINE __m128i _mm_unpackhi_epi8(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_s8( + vzip2q_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +#else + int8x8_t a1 = + vreinterpret_s8_s16(vget_high_s16(vreinterpretq_s16_m128i(a))); + int8x8_t b1 = + vreinterpret_s8_s16(vget_high_s16(vreinterpretq_s16_m128i(b))); + int8x8x2_t result = vzip_s8(a1, b1); + return vreinterpretq_m128i_s8(vcombine_s8(result.val[0], result.val[1])); +#endif +} + +// Unpack and interleave double-precision (64-bit) floating-point elements from +// the high half of a and b, and store the results in dst. +// +// DEFINE INTERLEAVE_HIGH_QWORDS(src1[127:0], src2[127:0]) { +// dst[63:0] := src1[127:64] +// dst[127:64] := src2[127:64] +// RETURN dst[127:0] +// } +// dst[127:0] := INTERLEAVE_HIGH_QWORDS(a[127:0], b[127:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_unpackhi_pd +FORCE_INLINE __m128d _mm_unpackhi_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vzip2q_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + return vreinterpretq_m128d_s64( + vcombine_s64(vget_high_s64(vreinterpretq_s64_m128d(a)), + vget_high_s64(vreinterpretq_s64_m128d(b)))); +#endif +} + +// Interleaves the lower 4 signed or unsigned 16-bit integers in a with the +// lower 4 signed or unsigned 16-bit integers in b. +// +// r0 := a0 +// r1 := b0 +// r2 := a1 +// r3 := b1 +// r4 := a2 +// r5 := b2 +// r6 := a3 +// r7 := b3 +// +// https://msdn.microsoft.com/en-us/library/btxb17bw%28v=vs.90%29.aspx +FORCE_INLINE __m128i _mm_unpacklo_epi16(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_s16( + vzip1q_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +#else + int16x4_t a1 = vget_low_s16(vreinterpretq_s16_m128i(a)); + int16x4_t b1 = vget_low_s16(vreinterpretq_s16_m128i(b)); + int16x4x2_t result = vzip_s16(a1, b1); + return vreinterpretq_m128i_s16(vcombine_s16(result.val[0], result.val[1])); +#endif +} + +// Interleaves the lower 2 signed or unsigned 32 - bit integers in a with the +// lower 2 signed or unsigned 32 - bit integers in b. +// +// r0 := a0 +// r1 := b0 +// r2 := a1 +// r3 := b1 +// +// https://msdn.microsoft.com/en-us/library/x8atst9d(v=vs.100).aspx +FORCE_INLINE __m128i _mm_unpacklo_epi32(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_s32( + vzip1q_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +#else + int32x2_t a1 = vget_low_s32(vreinterpretq_s32_m128i(a)); + int32x2_t b1 = vget_low_s32(vreinterpretq_s32_m128i(b)); + int32x2x2_t result = vzip_s32(a1, b1); + return vreinterpretq_m128i_s32(vcombine_s32(result.val[0], result.val[1])); +#endif +} + +FORCE_INLINE __m128i _mm_unpacklo_epi64(__m128i a, __m128i b) +{ + int64x1_t a_l = vget_low_s64(vreinterpretq_s64_m128i(a)); + int64x1_t b_l = vget_low_s64(vreinterpretq_s64_m128i(b)); + return vreinterpretq_m128i_s64(vcombine_s64(a_l, b_l)); +} + +// Interleaves the lower 8 signed or unsigned 8-bit integers in a with the lower +// 8 signed or unsigned 8-bit integers in b. +// +// r0 := a0 +// r1 := b0 +// r2 := a1 +// r3 := b1 +// ... +// r14 := a7 +// r15 := b7 +// +// https://msdn.microsoft.com/en-us/library/xf7k860c%28v=vs.90%29.aspx +FORCE_INLINE __m128i _mm_unpacklo_epi8(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_s8( + vzip1q_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +#else + int8x8_t a1 = vreinterpret_s8_s16(vget_low_s16(vreinterpretq_s16_m128i(a))); + int8x8_t b1 = vreinterpret_s8_s16(vget_low_s16(vreinterpretq_s16_m128i(b))); + int8x8x2_t result = vzip_s8(a1, b1); + return vreinterpretq_m128i_s8(vcombine_s8(result.val[0], result.val[1])); +#endif +} + +// Unpack and interleave double-precision (64-bit) floating-point elements from +// the low half of a and b, and store the results in dst. +// +// DEFINE INTERLEAVE_QWORDS(src1[127:0], src2[127:0]) { +// dst[63:0] := src1[63:0] +// dst[127:64] := src2[63:0] +// RETURN dst[127:0] +// } +// dst[127:0] := INTERLEAVE_QWORDS(a[127:0], b[127:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_unpacklo_pd +FORCE_INLINE __m128d _mm_unpacklo_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vzip1q_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + return vreinterpretq_m128d_s64( + vcombine_s64(vget_low_s64(vreinterpretq_s64_m128d(a)), + vget_low_s64(vreinterpretq_s64_m128d(b)))); +#endif +} + +// Compute the bitwise XOR of packed double-precision (64-bit) floating-point +// elements in a and b, and store the results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// dst[i+63:i] := a[i+63:i] XOR b[i+63:i] +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_xor_pd +FORCE_INLINE __m128d _mm_xor_pd(__m128d a, __m128d b) +{ + return vreinterpretq_m128d_s64( + veorq_s64(vreinterpretq_s64_m128d(a), vreinterpretq_s64_m128d(b))); +} + +// Computes the bitwise XOR of the 128-bit value in a and the 128-bit value in +// b. https://msdn.microsoft.com/en-us/library/fzt08www(v=vs.100).aspx +FORCE_INLINE __m128i _mm_xor_si128(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + veorq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +/* SSE3 */ + +// Alternatively add and subtract packed double-precision (64-bit) +// floating-point elements in a to/from packed elements in b, and store the +// results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// IF ((j & 1) == 0) +// dst[i+63:i] := a[i+63:i] - b[i+63:i] +// ELSE +// dst[i+63:i] := a[i+63:i] + b[i+63:i] +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_addsub_pd +FORCE_INLINE __m128d _mm_addsub_pd(__m128d a, __m128d b) +{ + __m128d mask = _mm_set_pd(1.0f, -1.0f); +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vfmaq_f64(vreinterpretq_f64_m128d(a), + vreinterpretq_f64_m128d(b), + vreinterpretq_f64_m128d(mask))); +#else + return _mm_add_pd(_mm_mul_pd(b, mask), a); +#endif +} + +// Alternatively add and subtract packed single-precision (32-bit) +// floating-point elements in a to/from packed elements in b, and store the +// results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=addsub_ps +FORCE_INLINE __m128 _mm_addsub_ps(__m128 a, __m128 b) +{ + __m128 mask = { -1.0f, 1.0f, -1.0f, 1.0f }; +#if defined(__aarch64__) || defined(__ARM_FEATURE_FMA) /* VFPv4+ */ + return vreinterpretq_m128_f32(vfmaq_f32(vreinterpretq_f32_m128(a), + vreinterpretq_f32_m128(mask), + vreinterpretq_f32_m128(b))); +#else + return _mm_add_ps(_mm_mul_ps(b, mask), a); +#endif +} + +// Horizontally add adjacent pairs of double-precision (64-bit) floating-point +// elements in a and b, and pack the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hadd_pd +FORCE_INLINE __m128d _mm_hadd_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vpaddq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + double *da = (double *)&a; + double *db = (double *)&b; + double c[] = { da[0] + da[1], db[0] + db[1] }; + return vreinterpretq_m128d_u64(vld1q_u64((uint64_t *)c)); +#endif +} + +// Computes pairwise add of each argument as single-precision, floating-point +// values a and b. +// https://msdn.microsoft.com/en-us/library/yd9wecaa.aspx +FORCE_INLINE __m128 _mm_hadd_ps(__m128 a, __m128 b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128_f32( + vpaddq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +#else + float32x2_t a10 = vget_low_f32(vreinterpretq_f32_m128(a)); + float32x2_t a32 = vget_high_f32(vreinterpretq_f32_m128(a)); + float32x2_t b10 = vget_low_f32(vreinterpretq_f32_m128(b)); + float32x2_t b32 = vget_high_f32(vreinterpretq_f32_m128(b)); + return vreinterpretq_m128_f32( + vcombine_f32(vpadd_f32(a10, a32), vpadd_f32(b10, b32))); +#endif +} + +// Horizontally subtract adjacent pairs of double-precision (64-bit) +// floating-point elements in a and b, and pack the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsub_pd +FORCE_INLINE __m128d _mm_hsub_pd(__m128d _a, __m128d _b) +{ +#if defined(__aarch64__) + float64x2_t a = vreinterpretq_f64_m128d(_a); + float64x2_t b = vreinterpretq_f64_m128d(_b); + return vreinterpretq_m128d_f64( + vsubq_f64(vuzp1q_f64(a, b), vuzp2q_f64(a, b))); +#else + double *da = (double *)&_a; + double *db = (double *)&_b; + double c[] = { da[0] - da[1], db[0] - db[1] }; + return vreinterpretq_m128d_u64(vld1q_u64((uint64_t *)c)); +#endif +} + +// Horizontally substract adjacent pairs of single-precision (32-bit) +// floating-point elements in a and b, and pack the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsub_ps +FORCE_INLINE __m128 _mm_hsub_ps(__m128 _a, __m128 _b) +{ + float32x4_t a = vreinterpretq_f32_m128(_a); + float32x4_t b = vreinterpretq_f32_m128(_b); +#if defined(__aarch64__) + return vreinterpretq_m128_f32( + vsubq_f32(vuzp1q_f32(a, b), vuzp2q_f32(a, b))); +#else + float32x4x2_t c = vuzpq_f32(a, b); + return vreinterpretq_m128_f32(vsubq_f32(c.val[0], c.val[1])); +#endif +} + +// Load 128-bits of integer data from unaligned memory into dst. This intrinsic +// may perform better than _mm_loadu_si128 when the data crosses a cache line +// boundary. +// +// dst[127:0] := MEM[mem_addr+127:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_lddqu_si128 +#define _mm_lddqu_si128 _mm_loadu_si128 + +// Load a double-precision (64-bit) floating-point element from memory into both +// elements of dst. +// +// dst[63:0] := MEM[mem_addr+63:mem_addr] +// dst[127:64] := MEM[mem_addr+63:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loaddup_pd +#define _mm_loaddup_pd _mm_load1_pd + +// Duplicate the low double-precision (64-bit) floating-point element from a, +// and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_movedup_pd +FORCE_INLINE __m128d _mm_movedup_pd(__m128d a) +{ +#if (__aarch64__) + return vreinterpretq_m128d_f64( + vdupq_laneq_f64(vreinterpretq_f64_m128d(a), 0)); +#else + return vreinterpretq_m128d_u64( + vdupq_n_u64(vgetq_lane_u64(vreinterpretq_u64_m128d(a), 0))); +#endif +} + +// Duplicate odd-indexed single-precision (32-bit) floating-point elements +// from a, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_movehdup_ps +FORCE_INLINE __m128 _mm_movehdup_ps(__m128 a) +{ +#if __has_builtin(__builtin_shufflevector) + return vreinterpretq_m128_f32(__builtin_shufflevector( + vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a), 1, 1, 3, 3)); +#else + float32_t a1 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 1); + float32_t a3 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 3); + float ALIGN_STRUCT(16) data[4] = { a1, a1, a3, a3 }; + return vreinterpretq_m128_f32(vld1q_f32(data)); +#endif +} + +// Duplicate even-indexed single-precision (32-bit) floating-point elements +// from a, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_moveldup_ps +FORCE_INLINE __m128 _mm_moveldup_ps(__m128 a) +{ +#if __has_builtin(__builtin_shufflevector) + return vreinterpretq_m128_f32(__builtin_shufflevector( + vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a), 0, 0, 2, 2)); +#else + float32_t a0 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); + float32_t a2 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 2); + float ALIGN_STRUCT(16) data[4] = { a0, a0, a2, a2 }; + return vreinterpretq_m128_f32(vld1q_f32(data)); +#endif +} + +/* SSSE3 */ + +// Compute the absolute value of packed signed 16-bit integers in a, and store +// the unsigned results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// dst[i+15:i] := ABS(a[i+15:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_epi16 +FORCE_INLINE __m128i _mm_abs_epi16(__m128i a) +{ + return vreinterpretq_m128i_s16(vabsq_s16(vreinterpretq_s16_m128i(a))); +} + +// Compute the absolute value of packed signed 32-bit integers in a, and store +// the unsigned results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// dst[i+31:i] := ABS(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_epi32 +FORCE_INLINE __m128i _mm_abs_epi32(__m128i a) +{ + return vreinterpretq_m128i_s32(vabsq_s32(vreinterpretq_s32_m128i(a))); +} + +// Compute the absolute value of packed signed 8-bit integers in a, and store +// the unsigned results in dst. +// +// FOR j := 0 to 15 +// i := j*8 +// dst[i+7:i] := ABS(a[i+7:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_epi8 +FORCE_INLINE __m128i _mm_abs_epi8(__m128i a) +{ + return vreinterpretq_m128i_s8(vabsq_s8(vreinterpretq_s8_m128i(a))); +} + +// Compute the absolute value of packed signed 16-bit integers in a, and store +// the unsigned results in dst. +// +// FOR j := 0 to 3 +// i := j*16 +// dst[i+15:i] := ABS(a[i+15:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_pi16 +FORCE_INLINE __m64 _mm_abs_pi16(__m64 a) +{ + return vreinterpret_m64_s16(vabs_s16(vreinterpret_s16_m64(a))); +} + +// Compute the absolute value of packed signed 32-bit integers in a, and store +// the unsigned results in dst. +// +// FOR j := 0 to 1 +// i := j*32 +// dst[i+31:i] := ABS(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_pi32 +FORCE_INLINE __m64 _mm_abs_pi32(__m64 a) +{ + return vreinterpret_m64_s32(vabs_s32(vreinterpret_s32_m64(a))); +} + +// Compute the absolute value of packed signed 8-bit integers in a, and store +// the unsigned results in dst. +// +// FOR j := 0 to 7 +// i := j*8 +// dst[i+7:i] := ABS(a[i+7:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_pi8 +FORCE_INLINE __m64 _mm_abs_pi8(__m64 a) +{ + return vreinterpret_m64_s8(vabs_s8(vreinterpret_s8_m64(a))); +} + +// Concatenate 16-byte blocks in a and b into a 32-byte temporary result, shift +// the result right by imm8 bytes, and store the low 16 bytes in dst. +// +// tmp[255:0] := ((a[127:0] << 128)[255:0] OR b[127:0]) >> (imm8*8) +// dst[127:0] := tmp[127:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_alignr_epi8 +FORCE_INLINE __m128i _mm_alignr_epi8(__m128i a, __m128i b, int imm) +{ + if (_sse2neon_unlikely(imm & ~31)) + return _mm_setzero_si128(); + int idx; + uint8x16_t tmp[2]; + if (imm >= 16) { + idx = imm - 16; + tmp[0] = vreinterpretq_u8_m128i(a); + tmp[1] = vdupq_n_u8(0); + } else { + idx = imm; + tmp[0] = vreinterpretq_u8_m128i(b); + tmp[1] = vreinterpretq_u8_m128i(a); + } + return vreinterpretq_m128i_u8(vld1q_u8(((uint8_t const *)tmp) + idx)); +} + +// Concatenate 8-byte blocks in a and b into a 16-byte temporary result, shift +// the result right by imm8 bytes, and store the low 8 bytes in dst. +// +// tmp[127:0] := ((a[63:0] << 64)[127:0] OR b[63:0]) >> (imm8*8) +// dst[63:0] := tmp[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_alignr_pi8 +#define _mm_alignr_pi8(a, b, imm) \ + __extension__({ \ + __m64 ret; \ + if (_sse2neon_unlikely((imm) >= 16)) { \ + ret = vreinterpret_m64_s8(vdup_n_s8(0)); \ + } else { \ + uint8x8_t tmp_low, tmp_high; \ + if (imm >= 8) { \ + const int idx = imm - 8; \ + tmp_low = vreinterpret_u8_m64(a); \ + tmp_high = vdup_n_u8(0); \ + ret = vreinterpret_m64_u8(vext_u8(tmp_low, tmp_high, idx)); \ + } else { \ + const int idx = imm; \ + tmp_low = vreinterpret_u8_m64(b); \ + tmp_high = vreinterpret_u8_m64(a); \ + ret = vreinterpret_m64_u8(vext_u8(tmp_low, tmp_high, idx)); \ + } \ + } \ + ret; \ + }) + +// Computes pairwise add of each argument as a 16-bit signed or unsigned integer +// values a and b. +FORCE_INLINE __m128i _mm_hadd_epi16(__m128i _a, __m128i _b) +{ + int16x8_t a = vreinterpretq_s16_m128i(_a); + int16x8_t b = vreinterpretq_s16_m128i(_b); +#if defined(__aarch64__) + return vreinterpretq_m128i_s16(vpaddq_s16(a, b)); +#else + return vreinterpretq_m128i_s16( + vcombine_s16(vpadd_s16(vget_low_s16(a), vget_high_s16(a)), + vpadd_s16(vget_low_s16(b), vget_high_s16(b)))); +#endif +} + +// Computes pairwise add of each argument as a 32-bit signed or unsigned integer +// values a and b. +FORCE_INLINE __m128i _mm_hadd_epi32(__m128i _a, __m128i _b) +{ + int32x4_t a = vreinterpretq_s32_m128i(_a); + int32x4_t b = vreinterpretq_s32_m128i(_b); + return vreinterpretq_m128i_s32( + vcombine_s32(vpadd_s32(vget_low_s32(a), vget_high_s32(a)), + vpadd_s32(vget_low_s32(b), vget_high_s32(b)))); +} + +// Horizontally add adjacent pairs of 16-bit integers in a and b, and pack the +// signed 16-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hadd_pi16 +FORCE_INLINE __m64 _mm_hadd_pi16(__m64 a, __m64 b) +{ + return vreinterpret_m64_s16( + vpadd_s16(vreinterpret_s16_m64(a), vreinterpret_s16_m64(b))); +} + +// Horizontally add adjacent pairs of 32-bit integers in a and b, and pack the +// signed 32-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hadd_pi32 +FORCE_INLINE __m64 _mm_hadd_pi32(__m64 a, __m64 b) +{ + return vreinterpret_m64_s32( + vpadd_s32(vreinterpret_s32_m64(a), vreinterpret_s32_m64(b))); +} + +// Computes saturated pairwise sub of each argument as a 16-bit signed +// integer values a and b. +FORCE_INLINE __m128i _mm_hadds_epi16(__m128i _a, __m128i _b) +{ +#if defined(__aarch64__) + int16x8_t a = vreinterpretq_s16_m128i(_a); + int16x8_t b = vreinterpretq_s16_m128i(_b); + return vreinterpretq_s64_s16( + vqaddq_s16(vuzp1q_s16(a, b), vuzp2q_s16(a, b))); +#else + int32x4_t a = vreinterpretq_s32_m128i(_a); + int32x4_t b = vreinterpretq_s32_m128i(_b); + // Interleave using vshrn/vmovn + // [a0|a2|a4|a6|b0|b2|b4|b6] + // [a1|a3|a5|a7|b1|b3|b5|b7] + int16x8_t ab0246 = vcombine_s16(vmovn_s32(a), vmovn_s32(b)); + int16x8_t ab1357 = vcombine_s16(vshrn_n_s32(a, 16), vshrn_n_s32(b, 16)); + // Saturated add + return vreinterpretq_m128i_s16(vqaddq_s16(ab0246, ab1357)); +#endif +} + +// Horizontally add adjacent pairs of signed 16-bit integers in a and b using +// saturation, and pack the signed 16-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hadds_pi16 +FORCE_INLINE __m64 _mm_hadds_pi16(__m64 _a, __m64 _b) +{ + int16x4_t a = vreinterpret_s16_m64(_a); + int16x4_t b = vreinterpret_s16_m64(_b); +#if defined(__aarch64__) + return vreinterpret_s64_s16(vqadd_s16(vuzp1_s16(a, b), vuzp2_s16(a, b))); +#else + int16x4x2_t res = vuzp_s16(a, b); + return vreinterpret_s64_s16(vqadd_s16(res.val[0], res.val[1])); +#endif +} + +// Horizontally subtract adjacent pairs of 16-bit integers in a and b, and pack +// the signed 16-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsub_epi16 +FORCE_INLINE __m128i _mm_hsub_epi16(__m128i _a, __m128i _b) +{ + int16x8_t a = vreinterpretq_s16_m128i(_a); + int16x8_t b = vreinterpretq_s16_m128i(_b); +#if defined(__aarch64__) + return vreinterpretq_m128i_s16( + vsubq_s16(vuzp1q_s16(a, b), vuzp2q_s16(a, b))); +#else + int16x8x2_t c = vuzpq_s16(a, b); + return vreinterpretq_m128i_s16(vsubq_s16(c.val[0], c.val[1])); +#endif +} + +// Horizontally subtract adjacent pairs of 32-bit integers in a and b, and pack +// the signed 32-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsub_epi32 +FORCE_INLINE __m128i _mm_hsub_epi32(__m128i _a, __m128i _b) +{ + int32x4_t a = vreinterpretq_s32_m128i(_a); + int32x4_t b = vreinterpretq_s32_m128i(_b); +#if defined(__aarch64__) + return vreinterpretq_m128i_s32( + vsubq_s32(vuzp1q_s32(a, b), vuzp2q_s32(a, b))); +#else + int32x4x2_t c = vuzpq_s32(a, b); + return vreinterpretq_m128i_s32(vsubq_s32(c.val[0], c.val[1])); +#endif +} + +// Horizontally subtract adjacent pairs of 16-bit integers in a and b, and pack +// the signed 16-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsub_pi16 +FORCE_INLINE __m64 _mm_hsub_pi16(__m64 _a, __m64 _b) +{ + int16x4_t a = vreinterpret_s16_m64(_a); + int16x4_t b = vreinterpret_s16_m64(_b); +#if defined(__aarch64__) + return vreinterpret_m64_s16(vsub_s16(vuzp1_s16(a, b), vuzp2_s16(a, b))); +#else + int16x4x2_t c = vuzp_s16(a, b); + return vreinterpret_m64_s16(vsub_s16(c.val[0], c.val[1])); +#endif +} + +// Horizontally subtract adjacent pairs of 32-bit integers in a and b, and pack +// the signed 32-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_hsub_pi32 +FORCE_INLINE __m64 _mm_hsub_pi32(__m64 _a, __m64 _b) +{ + int32x2_t a = vreinterpret_s32_m64(_a); + int32x2_t b = vreinterpret_s32_m64(_b); +#if defined(__aarch64__) + return vreinterpret_m64_s32(vsub_s32(vuzp1_s32(a, b), vuzp2_s32(a, b))); +#else + int32x2x2_t c = vuzp_s32(a, b); + return vreinterpret_m64_s32(vsub_s32(c.val[0], c.val[1])); +#endif +} + +// Computes saturated pairwise difference of each argument as a 16-bit signed +// integer values a and b. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsubs_epi16 +FORCE_INLINE __m128i _mm_hsubs_epi16(__m128i _a, __m128i _b) +{ + int16x8_t a = vreinterpretq_s16_m128i(_a); + int16x8_t b = vreinterpretq_s16_m128i(_b); +#if defined(__aarch64__) + return vreinterpretq_m128i_s16( + vqsubq_s16(vuzp1q_s16(a, b), vuzp2q_s16(a, b))); +#else + int16x8x2_t c = vuzpq_s16(a, b); + return vreinterpretq_m128i_s16(vqsubq_s16(c.val[0], c.val[1])); +#endif +} + +// Horizontally subtract adjacent pairs of signed 16-bit integers in a and b +// using saturation, and pack the signed 16-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsubs_pi16 +FORCE_INLINE __m64 _mm_hsubs_pi16(__m64 _a, __m64 _b) +{ + int16x4_t a = vreinterpret_s16_m64(_a); + int16x4_t b = vreinterpret_s16_m64(_b); +#if defined(__aarch64__) + return vreinterpret_m64_s16(vqsub_s16(vuzp1_s16(a, b), vuzp2_s16(a, b))); +#else + int16x4x2_t c = vuzp_s16(a, b); + return vreinterpret_m64_s16(vqsub_s16(c.val[0], c.val[1])); +#endif +} + +// Vertically multiply each unsigned 8-bit integer from a with the corresponding +// signed 8-bit integer from b, producing intermediate signed 16-bit integers. +// Horizontally add adjacent pairs of intermediate signed 16-bit integers, +// and pack the saturated results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// dst[i+15:i] := Saturate_To_Int16( a[i+15:i+8]*b[i+15:i+8] + +// a[i+7:i]*b[i+7:i] ) +// ENDFOR +FORCE_INLINE __m128i _mm_maddubs_epi16(__m128i _a, __m128i _b) +{ +#if defined(__aarch64__) + uint8x16_t a = vreinterpretq_u8_m128i(_a); + int8x16_t b = vreinterpretq_s8_m128i(_b); + int16x8_t tl = vmulq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(a))), + vmovl_s8(vget_low_s8(b))); + int16x8_t th = vmulq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(a))), + vmovl_s8(vget_high_s8(b))); + return vreinterpretq_m128i_s16( + vqaddq_s16(vuzp1q_s16(tl, th), vuzp2q_s16(tl, th))); +#else + // This would be much simpler if x86 would choose to zero extend OR sign + // extend, not both. This could probably be optimized better. + uint16x8_t a = vreinterpretq_u16_m128i(_a); + int16x8_t b = vreinterpretq_s16_m128i(_b); + + // Zero extend a + int16x8_t a_odd = vreinterpretq_s16_u16(vshrq_n_u16(a, 8)); + int16x8_t a_even = vreinterpretq_s16_u16(vbicq_u16(a, vdupq_n_u16(0xff00))); + + // Sign extend by shifting left then shifting right. + int16x8_t b_even = vshrq_n_s16(vshlq_n_s16(b, 8), 8); + int16x8_t b_odd = vshrq_n_s16(b, 8); + + // multiply + int16x8_t prod1 = vmulq_s16(a_even, b_even); + int16x8_t prod2 = vmulq_s16(a_odd, b_odd); + + // saturated add + return vreinterpretq_m128i_s16(vqaddq_s16(prod1, prod2)); +#endif +} + +// Vertically multiply each unsigned 8-bit integer from a with the corresponding +// signed 8-bit integer from b, producing intermediate signed 16-bit integers. +// Horizontally add adjacent pairs of intermediate signed 16-bit integers, and +// pack the saturated results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_maddubs_pi16 +FORCE_INLINE __m64 _mm_maddubs_pi16(__m64 _a, __m64 _b) +{ + uint16x4_t a = vreinterpret_u16_m64(_a); + int16x4_t b = vreinterpret_s16_m64(_b); + + // Zero extend a + int16x4_t a_odd = vreinterpret_s16_u16(vshr_n_u16(a, 8)); + int16x4_t a_even = vreinterpret_s16_u16(vand_u16(a, vdup_n_u16(0xff))); + + // Sign extend by shifting left then shifting right. + int16x4_t b_even = vshr_n_s16(vshl_n_s16(b, 8), 8); + int16x4_t b_odd = vshr_n_s16(b, 8); + + // multiply + int16x4_t prod1 = vmul_s16(a_even, b_even); + int16x4_t prod2 = vmul_s16(a_odd, b_odd); + + // saturated add + return vreinterpret_m64_s16(vqadd_s16(prod1, prod2)); +} + +// Multiply packed signed 16-bit integers in a and b, producing intermediate +// signed 32-bit integers. Shift right by 15 bits while rounding up, and store +// the packed 16-bit integers in dst. +// +// r0 := Round(((int32_t)a0 * (int32_t)b0) >> 15) +// r1 := Round(((int32_t)a1 * (int32_t)b1) >> 15) +// r2 := Round(((int32_t)a2 * (int32_t)b2) >> 15) +// ... +// r7 := Round(((int32_t)a7 * (int32_t)b7) >> 15) +FORCE_INLINE __m128i _mm_mulhrs_epi16(__m128i a, __m128i b) +{ + // Has issues due to saturation + // return vreinterpretq_m128i_s16(vqrdmulhq_s16(a, b)); + + // Multiply + int32x4_t mul_lo = vmull_s16(vget_low_s16(vreinterpretq_s16_m128i(a)), + vget_low_s16(vreinterpretq_s16_m128i(b))); + int32x4_t mul_hi = vmull_s16(vget_high_s16(vreinterpretq_s16_m128i(a)), + vget_high_s16(vreinterpretq_s16_m128i(b))); + + // Rounding narrowing shift right + // narrow = (int16_t)((mul + 16384) >> 15); + int16x4_t narrow_lo = vrshrn_n_s32(mul_lo, 15); + int16x4_t narrow_hi = vrshrn_n_s32(mul_hi, 15); + + // Join together + return vreinterpretq_m128i_s16(vcombine_s16(narrow_lo, narrow_hi)); +} + +// Multiply packed signed 16-bit integers in a and b, producing intermediate +// signed 32-bit integers. Truncate each intermediate integer to the 18 most +// significant bits, round by adding 1, and store bits [16:1] to dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mulhrs_pi16 +FORCE_INLINE __m64 _mm_mulhrs_pi16(__m64 a, __m64 b) +{ + int32x4_t mul_extend = + vmull_s16((vreinterpret_s16_m64(a)), (vreinterpret_s16_m64(b))); + + // Rounding narrowing shift right + return vreinterpret_m64_s16(vrshrn_n_s32(mul_extend, 15)); +} + +// Shuffle packed 8-bit integers in a according to shuffle control mask in the +// corresponding 8-bit element of b, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_shuffle_epi8 +FORCE_INLINE __m128i _mm_shuffle_epi8(__m128i a, __m128i b) +{ + int8x16_t tbl = vreinterpretq_s8_m128i(a); // input a + uint8x16_t idx = vreinterpretq_u8_m128i(b); // input b + uint8x16_t idx_masked = + vandq_u8(idx, vdupq_n_u8(0x8F)); // avoid using meaningless bits +#if defined(__aarch64__) + return vreinterpretq_m128i_s8(vqtbl1q_s8(tbl, idx_masked)); +#elif defined(__GNUC__) + int8x16_t ret; + // %e and %f represent the even and odd D registers + // respectively. + __asm__ __volatile__("vtbl.8 %e[ret], {%e[tbl], %f[tbl]}, %e[idx]\n" + "vtbl.8 %f[ret], {%e[tbl], %f[tbl]}, %f[idx]\n" + : [ret] "=&w"(ret) + : [tbl] "w"(tbl), [idx] "w"(idx_masked)); + return vreinterpretq_m128i_s8(ret); +#else + // use this line if testing on aarch64 + int8x8x2_t a_split = { vget_low_s8(tbl), vget_high_s8(tbl) }; + return vreinterpretq_m128i_s8( + vcombine_s8(vtbl2_s8(a_split, vget_low_u8(idx_masked)), + vtbl2_s8(a_split, vget_high_u8(idx_masked)))); +#endif +} + +// Shuffle packed 8-bit integers in a according to shuffle control mask in the +// corresponding 8-bit element of b, and store the results in dst. +// +// FOR j := 0 to 7 +// i := j*8 +// IF b[i+7] == 1 +// dst[i+7:i] := 0 +// ELSE +// index[2:0] := b[i+2:i] +// dst[i+7:i] := a[index*8+7:index*8] +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_shuffle_pi8 +FORCE_INLINE __m64 _mm_shuffle_pi8(__m64 a, __m64 b) +{ + const int8x8_t controlMask = + vand_s8(vreinterpret_s8_m64(b), vdup_n_s8((int8_t)(0x1 << 7 | 0x07))); + int8x8_t res = vtbl1_s8(vreinterpret_s8_m64(a), controlMask); + return vreinterpret_m64_s8(res); +} + +// Negate packed 16-bit integers in a when the corresponding signed +// 16-bit integer in b is negative, and store the results in dst. +// Element in dst are zeroed out when the corresponding element +// in b is zero. +// +// for i in 0..7 +// if b[i] < 0 +// r[i] := -a[i] +// else if b[i] == 0 +// r[i] := 0 +// else +// r[i] := a[i] +// fi +// done +FORCE_INLINE __m128i _mm_sign_epi16(__m128i _a, __m128i _b) +{ + int16x8_t a = vreinterpretq_s16_m128i(_a); + int16x8_t b = vreinterpretq_s16_m128i(_b); + + // signed shift right: faster than vclt + // (b < 0) ? 0xFFFF : 0 + uint16x8_t ltMask = vreinterpretq_u16_s16(vshrq_n_s16(b, 15)); + // (b == 0) ? 0xFFFF : 0 +#if defined(__aarch64__) + int16x8_t zeroMask = vreinterpretq_s16_u16(vceqzq_s16(b)); +#else + int16x8_t zeroMask = vreinterpretq_s16_u16(vceqq_s16(b, vdupq_n_s16(0))); +#endif + + // bitwise select either a or negative 'a' (vnegq_s16(a) equals to negative + // 'a') based on ltMask + int16x8_t masked = vbslq_s16(ltMask, vnegq_s16(a), a); + // res = masked & (~zeroMask) + int16x8_t res = vbicq_s16(masked, zeroMask); + return vreinterpretq_m128i_s16(res); +} + +// Negate packed 32-bit integers in a when the corresponding signed +// 32-bit integer in b is negative, and store the results in dst. +// Element in dst are zeroed out when the corresponding element +// in b is zero. +// +// for i in 0..3 +// if b[i] < 0 +// r[i] := -a[i] +// else if b[i] == 0 +// r[i] := 0 +// else +// r[i] := a[i] +// fi +// done +FORCE_INLINE __m128i _mm_sign_epi32(__m128i _a, __m128i _b) +{ + int32x4_t a = vreinterpretq_s32_m128i(_a); + int32x4_t b = vreinterpretq_s32_m128i(_b); + + // signed shift right: faster than vclt + // (b < 0) ? 0xFFFFFFFF : 0 + uint32x4_t ltMask = vreinterpretq_u32_s32(vshrq_n_s32(b, 31)); + + // (b == 0) ? 0xFFFFFFFF : 0 +#if defined(__aarch64__) + int32x4_t zeroMask = vreinterpretq_s32_u32(vceqzq_s32(b)); +#else + int32x4_t zeroMask = vreinterpretq_s32_u32(vceqq_s32(b, vdupq_n_s32(0))); +#endif + + // bitwise select either a or negative 'a' (vnegq_s32(a) equals to negative + // 'a') based on ltMask + int32x4_t masked = vbslq_s32(ltMask, vnegq_s32(a), a); + // res = masked & (~zeroMask) + int32x4_t res = vbicq_s32(masked, zeroMask); + return vreinterpretq_m128i_s32(res); +} + +// Negate packed 8-bit integers in a when the corresponding signed +// 8-bit integer in b is negative, and store the results in dst. +// Element in dst are zeroed out when the corresponding element +// in b is zero. +// +// for i in 0..15 +// if b[i] < 0 +// r[i] := -a[i] +// else if b[i] == 0 +// r[i] := 0 +// else +// r[i] := a[i] +// fi +// done +FORCE_INLINE __m128i _mm_sign_epi8(__m128i _a, __m128i _b) +{ + int8x16_t a = vreinterpretq_s8_m128i(_a); + int8x16_t b = vreinterpretq_s8_m128i(_b); + + // signed shift right: faster than vclt + // (b < 0) ? 0xFF : 0 + uint8x16_t ltMask = vreinterpretq_u8_s8(vshrq_n_s8(b, 7)); + + // (b == 0) ? 0xFF : 0 +#if defined(__aarch64__) + int8x16_t zeroMask = vreinterpretq_s8_u8(vceqzq_s8(b)); +#else + int8x16_t zeroMask = vreinterpretq_s8_u8(vceqq_s8(b, vdupq_n_s8(0))); +#endif + + // bitwise select either a or nagative 'a' (vnegq_s8(a) return nagative 'a') + // based on ltMask + int8x16_t masked = vbslq_s8(ltMask, vnegq_s8(a), a); + // res = masked & (~zeroMask) + int8x16_t res = vbicq_s8(masked, zeroMask); + + return vreinterpretq_m128i_s8(res); +} + +// Negate packed 16-bit integers in a when the corresponding signed 16-bit +// integer in b is negative, and store the results in dst. Element in dst are +// zeroed out when the corresponding element in b is zero. +// +// FOR j := 0 to 3 +// i := j*16 +// IF b[i+15:i] < 0 +// dst[i+15:i] := -(a[i+15:i]) +// ELSE IF b[i+15:i] == 0 +// dst[i+15:i] := 0 +// ELSE +// dst[i+15:i] := a[i+15:i] +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sign_pi16 +FORCE_INLINE __m64 _mm_sign_pi16(__m64 _a, __m64 _b) +{ + int16x4_t a = vreinterpret_s16_m64(_a); + int16x4_t b = vreinterpret_s16_m64(_b); + + // signed shift right: faster than vclt + // (b < 0) ? 0xFFFF : 0 + uint16x4_t ltMask = vreinterpret_u16_s16(vshr_n_s16(b, 15)); + + // (b == 0) ? 0xFFFF : 0 +#if defined(__aarch64__) + int16x4_t zeroMask = vreinterpret_s16_u16(vceqz_s16(b)); +#else + int16x4_t zeroMask = vreinterpret_s16_u16(vceq_s16(b, vdup_n_s16(0))); +#endif + + // bitwise select either a or nagative 'a' (vneg_s16(a) return nagative 'a') + // based on ltMask + int16x4_t masked = vbsl_s16(ltMask, vneg_s16(a), a); + // res = masked & (~zeroMask) + int16x4_t res = vbic_s16(masked, zeroMask); + + return vreinterpret_m64_s16(res); +} + +// Negate packed 32-bit integers in a when the corresponding signed 32-bit +// integer in b is negative, and store the results in dst. Element in dst are +// zeroed out when the corresponding element in b is zero. +// +// FOR j := 0 to 1 +// i := j*32 +// IF b[i+31:i] < 0 +// dst[i+31:i] := -(a[i+31:i]) +// ELSE IF b[i+31:i] == 0 +// dst[i+31:i] := 0 +// ELSE +// dst[i+31:i] := a[i+31:i] +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sign_pi32 +FORCE_INLINE __m64 _mm_sign_pi32(__m64 _a, __m64 _b) +{ + int32x2_t a = vreinterpret_s32_m64(_a); + int32x2_t b = vreinterpret_s32_m64(_b); + + // signed shift right: faster than vclt + // (b < 0) ? 0xFFFFFFFF : 0 + uint32x2_t ltMask = vreinterpret_u32_s32(vshr_n_s32(b, 31)); + + // (b == 0) ? 0xFFFFFFFF : 0 +#if defined(__aarch64__) + int32x2_t zeroMask = vreinterpret_s32_u32(vceqz_s32(b)); +#else + int32x2_t zeroMask = vreinterpret_s32_u32(vceq_s32(b, vdup_n_s32(0))); +#endif + + // bitwise select either a or nagative 'a' (vneg_s32(a) return nagative 'a') + // based on ltMask + int32x2_t masked = vbsl_s32(ltMask, vneg_s32(a), a); + // res = masked & (~zeroMask) + int32x2_t res = vbic_s32(masked, zeroMask); + + return vreinterpret_m64_s32(res); +} + +// Negate packed 8-bit integers in a when the corresponding signed 8-bit integer +// in b is negative, and store the results in dst. Element in dst are zeroed out +// when the corresponding element in b is zero. +// +// FOR j := 0 to 7 +// i := j*8 +// IF b[i+7:i] < 0 +// dst[i+7:i] := -(a[i+7:i]) +// ELSE IF b[i+7:i] == 0 +// dst[i+7:i] := 0 +// ELSE +// dst[i+7:i] := a[i+7:i] +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sign_pi8 +FORCE_INLINE __m64 _mm_sign_pi8(__m64 _a, __m64 _b) +{ + int8x8_t a = vreinterpret_s8_m64(_a); + int8x8_t b = vreinterpret_s8_m64(_b); + + // signed shift right: faster than vclt + // (b < 0) ? 0xFF : 0 + uint8x8_t ltMask = vreinterpret_u8_s8(vshr_n_s8(b, 7)); + + // (b == 0) ? 0xFF : 0 +#if defined(__aarch64__) + int8x8_t zeroMask = vreinterpret_s8_u8(vceqz_s8(b)); +#else + int8x8_t zeroMask = vreinterpret_s8_u8(vceq_s8(b, vdup_n_s8(0))); +#endif + + // bitwise select either a or nagative 'a' (vneg_s8(a) return nagative 'a') + // based on ltMask + int8x8_t masked = vbsl_s8(ltMask, vneg_s8(a), a); + // res = masked & (~zeroMask) + int8x8_t res = vbic_s8(masked, zeroMask); + + return vreinterpret_m64_s8(res); +} + +/* SSE4.1 */ + +// Blend packed 16-bit integers from a and b using control mask imm8, and store +// the results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// IF imm8[j] +// dst[i+15:i] := b[i+15:i] +// ELSE +// dst[i+15:i] := a[i+15:i] +// FI +// ENDFOR +// FORCE_INLINE __m128i _mm_blend_epi16(__m128i a, __m128i b, +// __constrange(0,255) int imm) +#define _mm_blend_epi16(a, b, imm) \ + __extension__({ \ + const uint16_t _mask[8] = { ((imm) & (1 << 0)) ? (uint16_t)-1 : 0x0, \ + ((imm) & (1 << 1)) ? (uint16_t)-1 : 0x0, \ + ((imm) & (1 << 2)) ? (uint16_t)-1 : 0x0, \ + ((imm) & (1 << 3)) ? (uint16_t)-1 : 0x0, \ + ((imm) & (1 << 4)) ? (uint16_t)-1 : 0x0, \ + ((imm) & (1 << 5)) ? (uint16_t)-1 : 0x0, \ + ((imm) & (1 << 6)) ? (uint16_t)-1 : 0x0, \ + ((imm) & (1 << 7)) ? (uint16_t)-1 : 0x0 }; \ + uint16x8_t _mask_vec = vld1q_u16(_mask); \ + uint16x8_t _a = vreinterpretq_u16_m128i(a); \ + uint16x8_t _b = vreinterpretq_u16_m128i(b); \ + vreinterpretq_m128i_u16(vbslq_u16(_mask_vec, _b, _a)); \ + }) + +// Blend packed double-precision (64-bit) floating-point elements from a and b +// using control mask imm8, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_blend_pd +#define _mm_blend_pd(a, b, imm) \ + __extension__({ \ + const uint64_t _mask[2] = { \ + ((imm) & (1 << 0)) ? ~UINT64_C(0) : UINT64_C(0), \ + ((imm) & (1 << 1)) ? ~UINT64_C(0) : UINT64_C(0) \ + }; \ + uint64x2_t _mask_vec = vld1q_u64(_mask); \ + uint64x2_t _a = vreinterpretq_u64_m128d(a); \ + uint64x2_t _b = vreinterpretq_u64_m128d(b); \ + vreinterpretq_m128d_u64(vbslq_u64(_mask_vec, _b, _a)); \ + }) + +// Blend packed single-precision (32-bit) floating-point elements from a and b +// using mask, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_blend_ps +FORCE_INLINE __m128 _mm_blend_ps(__m128 _a, __m128 _b, const char imm8) +{ + const uint32_t ALIGN_STRUCT(16) + data[4] = { ((imm8) & (1 << 0)) ? UINT32_MAX : 0, + ((imm8) & (1 << 1)) ? UINT32_MAX : 0, + ((imm8) & (1 << 2)) ? UINT32_MAX : 0, + ((imm8) & (1 << 3)) ? UINT32_MAX : 0 }; + uint32x4_t mask = vld1q_u32(data); + float32x4_t a = vreinterpretq_f32_m128(_a); + float32x4_t b = vreinterpretq_f32_m128(_b); + return vreinterpretq_m128_f32(vbslq_f32(mask, b, a)); +} + +// Blend packed 8-bit integers from a and b using mask, and store the results in +// dst. +// +// FOR j := 0 to 15 +// i := j*8 +// IF mask[i+7] +// dst[i+7:i] := b[i+7:i] +// ELSE +// dst[i+7:i] := a[i+7:i] +// FI +// ENDFOR +FORCE_INLINE __m128i _mm_blendv_epi8(__m128i _a, __m128i _b, __m128i _mask) +{ + // Use a signed shift right to create a mask with the sign bit + uint8x16_t mask = + vreinterpretq_u8_s8(vshrq_n_s8(vreinterpretq_s8_m128i(_mask), 7)); + uint8x16_t a = vreinterpretq_u8_m128i(_a); + uint8x16_t b = vreinterpretq_u8_m128i(_b); + return vreinterpretq_m128i_u8(vbslq_u8(mask, b, a)); +} + +// Blend packed double-precision (64-bit) floating-point elements from a and b +// using mask, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_blendv_pd +FORCE_INLINE __m128d _mm_blendv_pd(__m128d _a, __m128d _b, __m128d _mask) +{ + uint64x2_t mask = + vreinterpretq_u64_s64(vshrq_n_s64(vreinterpretq_s64_m128d(_mask), 63)); +#if defined(__aarch64__) + float64x2_t a = vreinterpretq_f64_m128d(_a); + float64x2_t b = vreinterpretq_f64_m128d(_b); + return vreinterpretq_m128d_f64(vbslq_f64(mask, b, a)); +#else + uint64x2_t a = vreinterpretq_u64_m128d(_a); + uint64x2_t b = vreinterpretq_u64_m128d(_b); + return vreinterpretq_m128d_u64(vbslq_u64(mask, b, a)); +#endif +} + +// Blend packed single-precision (32-bit) floating-point elements from a and b +// using mask, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_blendv_ps +FORCE_INLINE __m128 _mm_blendv_ps(__m128 _a, __m128 _b, __m128 _mask) +{ + // Use a signed shift right to create a mask with the sign bit + uint32x4_t mask = + vreinterpretq_u32_s32(vshrq_n_s32(vreinterpretq_s32_m128(_mask), 31)); + float32x4_t a = vreinterpretq_f32_m128(_a); + float32x4_t b = vreinterpretq_f32_m128(_b); + return vreinterpretq_m128_f32(vbslq_f32(mask, b, a)); +} + +// Round the packed double-precision (64-bit) floating-point elements in a up +// to an integer value, and store the results as packed double-precision +// floating-point elements in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_ceil_pd +FORCE_INLINE __m128d _mm_ceil_pd(__m128d a) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vrndpq_f64(vreinterpretq_f64_m128d(a))); +#else + double *f = (double *)&a; + return _mm_set_pd(ceil(f[1]), ceil(f[0])); +#endif +} + +// Round the packed single-precision (32-bit) floating-point elements in a up to +// an integer value, and store the results as packed single-precision +// floating-point elements in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_ceil_ps +FORCE_INLINE __m128 _mm_ceil_ps(__m128 a) +{ +#if defined(__aarch64__) + return vreinterpretq_m128_f32(vrndpq_f32(vreinterpretq_f32_m128(a))); +#else + float *f = (float *)&a; + return _mm_set_ps(ceilf(f[3]), ceilf(f[2]), ceilf(f[1]), ceilf(f[0])); +#endif +} + +// Round the lower double-precision (64-bit) floating-point element in b up to +// an integer value, store the result as a double-precision floating-point +// element in the lower element of dst, and copy the upper element from a to the +// upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_ceil_sd +FORCE_INLINE __m128d _mm_ceil_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_ceil_pd(b)); +} + +// Round the lower single-precision (32-bit) floating-point element in b up to +// an integer value, store the result as a single-precision floating-point +// element in the lower element of dst, and copy the upper 3 packed elements +// from a to the upper elements of dst. +// +// dst[31:0] := CEIL(b[31:0]) +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_ceil_ss +FORCE_INLINE __m128 _mm_ceil_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_ceil_ps(b)); +} + +// Compare packed 64-bit integers in a and b for equality, and store the results +// in dst +FORCE_INLINE __m128i _mm_cmpeq_epi64(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_u64( + vceqq_u64(vreinterpretq_u64_m128i(a), vreinterpretq_u64_m128i(b))); +#else + // ARMv7 lacks vceqq_u64 + // (a == b) -> (a_lo == b_lo) && (a_hi == b_hi) + uint32x4_t cmp = + vceqq_u32(vreinterpretq_u32_m128i(a), vreinterpretq_u32_m128i(b)); + uint32x4_t swapped = vrev64q_u32(cmp); + return vreinterpretq_m128i_u32(vandq_u32(cmp, swapped)); +#endif +} + +// Converts the four signed 16-bit integers in the lower 64 bits to four signed +// 32-bit integers. +FORCE_INLINE __m128i _mm_cvtepi16_epi32(__m128i a) +{ + return vreinterpretq_m128i_s32( + vmovl_s16(vget_low_s16(vreinterpretq_s16_m128i(a)))); +} + +// Converts the two signed 16-bit integers in the lower 32 bits two signed +// 32-bit integers. +FORCE_INLINE __m128i _mm_cvtepi16_epi64(__m128i a) +{ + int16x8_t s16x8 = vreinterpretq_s16_m128i(a); /* xxxx xxxx xxxx 0B0A */ + int32x4_t s32x4 = vmovl_s16(vget_low_s16(s16x8)); /* 000x 000x 000B 000A */ + int64x2_t s64x2 = vmovl_s32(vget_low_s32(s32x4)); /* 0000 000B 0000 000A */ + return vreinterpretq_m128i_s64(s64x2); +} + +// Converts the two signed 32-bit integers in the lower 64 bits to two signed +// 64-bit integers. +FORCE_INLINE __m128i _mm_cvtepi32_epi64(__m128i a) +{ + return vreinterpretq_m128i_s64( + vmovl_s32(vget_low_s32(vreinterpretq_s32_m128i(a)))); +} + +// Converts the four unsigned 8-bit integers in the lower 16 bits to four +// unsigned 32-bit integers. +FORCE_INLINE __m128i _mm_cvtepi8_epi16(__m128i a) +{ + int8x16_t s8x16 = vreinterpretq_s8_m128i(a); /* xxxx xxxx xxxx DCBA */ + int16x8_t s16x8 = vmovl_s8(vget_low_s8(s8x16)); /* 0x0x 0x0x 0D0C 0B0A */ + return vreinterpretq_m128i_s16(s16x8); +} + +// Converts the four unsigned 8-bit integers in the lower 32 bits to four +// unsigned 32-bit integers. +FORCE_INLINE __m128i _mm_cvtepi8_epi32(__m128i a) +{ + int8x16_t s8x16 = vreinterpretq_s8_m128i(a); /* xxxx xxxx xxxx DCBA */ + int16x8_t s16x8 = vmovl_s8(vget_low_s8(s8x16)); /* 0x0x 0x0x 0D0C 0B0A */ + int32x4_t s32x4 = vmovl_s16(vget_low_s16(s16x8)); /* 000D 000C 000B 000A */ + return vreinterpretq_m128i_s32(s32x4); +} + +// Converts the two signed 8-bit integers in the lower 32 bits to four +// signed 64-bit integers. +FORCE_INLINE __m128i _mm_cvtepi8_epi64(__m128i a) +{ + int8x16_t s8x16 = vreinterpretq_s8_m128i(a); /* xxxx xxxx xxxx xxBA */ + int16x8_t s16x8 = vmovl_s8(vget_low_s8(s8x16)); /* 0x0x 0x0x 0x0x 0B0A */ + int32x4_t s32x4 = vmovl_s16(vget_low_s16(s16x8)); /* 000x 000x 000B 000A */ + int64x2_t s64x2 = vmovl_s32(vget_low_s32(s32x4)); /* 0000 000B 0000 000A */ + return vreinterpretq_m128i_s64(s64x2); +} + +// Converts the four unsigned 16-bit integers in the lower 64 bits to four +// unsigned 32-bit integers. +FORCE_INLINE __m128i _mm_cvtepu16_epi32(__m128i a) +{ + return vreinterpretq_m128i_u32( + vmovl_u16(vget_low_u16(vreinterpretq_u16_m128i(a)))); +} + +// Converts the two unsigned 16-bit integers in the lower 32 bits to two +// unsigned 64-bit integers. +FORCE_INLINE __m128i _mm_cvtepu16_epi64(__m128i a) +{ + uint16x8_t u16x8 = vreinterpretq_u16_m128i(a); /* xxxx xxxx xxxx 0B0A */ + uint32x4_t u32x4 = vmovl_u16(vget_low_u16(u16x8)); /* 000x 000x 000B 000A */ + uint64x2_t u64x2 = vmovl_u32(vget_low_u32(u32x4)); /* 0000 000B 0000 000A */ + return vreinterpretq_m128i_u64(u64x2); +} + +// Converts the two unsigned 32-bit integers in the lower 64 bits to two +// unsigned 64-bit integers. +FORCE_INLINE __m128i _mm_cvtepu32_epi64(__m128i a) +{ + return vreinterpretq_m128i_u64( + vmovl_u32(vget_low_u32(vreinterpretq_u32_m128i(a)))); +} + +// Zero extend packed unsigned 8-bit integers in a to packed 16-bit integers, +// and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtepu8_epi16 +FORCE_INLINE __m128i _mm_cvtepu8_epi16(__m128i a) +{ + uint8x16_t u8x16 = vreinterpretq_u8_m128i(a); /* xxxx xxxx HGFE DCBA */ + uint16x8_t u16x8 = vmovl_u8(vget_low_u8(u8x16)); /* 0H0G 0F0E 0D0C 0B0A */ + return vreinterpretq_m128i_u16(u16x8); +} + +// Converts the four unsigned 8-bit integers in the lower 32 bits to four +// unsigned 32-bit integers. +// https://msdn.microsoft.com/en-us/library/bb531467%28v=vs.100%29.aspx +FORCE_INLINE __m128i _mm_cvtepu8_epi32(__m128i a) +{ + uint8x16_t u8x16 = vreinterpretq_u8_m128i(a); /* xxxx xxxx xxxx DCBA */ + uint16x8_t u16x8 = vmovl_u8(vget_low_u8(u8x16)); /* 0x0x 0x0x 0D0C 0B0A */ + uint32x4_t u32x4 = vmovl_u16(vget_low_u16(u16x8)); /* 000D 000C 000B 000A */ + return vreinterpretq_m128i_u32(u32x4); +} + +// Converts the two unsigned 8-bit integers in the lower 16 bits to two +// unsigned 64-bit integers. +FORCE_INLINE __m128i _mm_cvtepu8_epi64(__m128i a) +{ + uint8x16_t u8x16 = vreinterpretq_u8_m128i(a); /* xxxx xxxx xxxx xxBA */ + uint16x8_t u16x8 = vmovl_u8(vget_low_u8(u8x16)); /* 0x0x 0x0x 0x0x 0B0A */ + uint32x4_t u32x4 = vmovl_u16(vget_low_u16(u16x8)); /* 000x 000x 000B 000A */ + uint64x2_t u64x2 = vmovl_u32(vget_low_u32(u32x4)); /* 0000 000B 0000 000A */ + return vreinterpretq_m128i_u64(u64x2); +} + +// Conditionally multiply the packed double-precision (64-bit) floating-point +// elements in a and b using the high 4 bits in imm8, sum the four products, and +// conditionally store the sum in dst using the low 4 bits of imm8. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_dp_pd +FORCE_INLINE __m128d _mm_dp_pd(__m128d a, __m128d b, const int imm) +{ + // Generate mask value from constant immediate bit value + const int64_t bit0Mask = imm & 0x01 ? UINT64_MAX : 0; + const int64_t bit1Mask = imm & 0x02 ? UINT64_MAX : 0; +#if !SSE2NEON_PRECISE_DP + const int64_t bit4Mask = imm & 0x10 ? UINT64_MAX : 0; + const int64_t bit5Mask = imm & 0x20 ? UINT64_MAX : 0; +#endif + // Conditional multiplication +#if !SSE2NEON_PRECISE_DP + __m128d mul = _mm_mul_pd(a, b); + const __m128d mulMask = + _mm_castsi128_pd(_mm_set_epi64x(bit5Mask, bit4Mask)); + __m128d tmp = _mm_and_pd(mul, mulMask); +#else +#if defined(__aarch64__) + double d0 = (imm & 0x10) ? + vgetq_lane_f64(vreinterpretq_f64_m128d(a), 0) * + vgetq_lane_f64(vreinterpretq_f64_m128d(b), 0) : + 0; + double d1 = (imm & 0x20) ? + vgetq_lane_f64(vreinterpretq_f64_m128d(a), 1) * + vgetq_lane_f64(vreinterpretq_f64_m128d(b), 1) : + 0; +#else + double d0 = (imm & 0x10) ? ((double *)&a)[0] * ((double *)&b)[0] : 0; + double d1 = (imm & 0x20) ? ((double *)&a)[1] * ((double *)&b)[1] : 0; +#endif + __m128d tmp = _mm_set_pd(d1, d0); +#endif + // Sum the products +#if defined(__aarch64__) + double sum = vpaddd_f64(vreinterpretq_f64_m128d(tmp)); +#else + double sum = *((double *)&tmp) + *(((double *)&tmp) + 1); +#endif + // Conditionally store the sum + const __m128d sumMask = + _mm_castsi128_pd(_mm_set_epi64x(bit1Mask, bit0Mask)); + __m128d res = _mm_and_pd(_mm_set_pd1(sum), sumMask); + return res; +} + +// Conditionally multiply the packed single-precision (32-bit) floating-point +// elements in a and b using the high 4 bits in imm8, sum the four products, +// and conditionally store the sum in dst using the low 4 bits of imm. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_dp_ps +FORCE_INLINE __m128 _mm_dp_ps(__m128 a, __m128 b, const int imm) +{ +#if defined(__aarch64__) + /* shortcuts */ + if (imm == 0xFF) { + return _mm_set1_ps(vaddvq_f32(_mm_mul_ps(a, b))); + } + if (imm == 0x7F) { + float32x4_t m = _mm_mul_ps(a, b); + m[3] = 0; + return _mm_set1_ps(vaddvq_f32(m)); + } +#endif + + float s = 0, c = 0; + float32x4_t f32a = vreinterpretq_f32_m128(a); + float32x4_t f32b = vreinterpretq_f32_m128(b); + + /* To improve the accuracy of floating-point summation, Kahan algorithm + * is used for each operation. + */ + if (imm & (1 << 4)) + _sse2neon_kadd_f32(&s, &c, f32a[0] * f32b[0]); + if (imm & (1 << 5)) + _sse2neon_kadd_f32(&s, &c, f32a[1] * f32b[1]); + if (imm & (1 << 6)) + _sse2neon_kadd_f32(&s, &c, f32a[2] * f32b[2]); + if (imm & (1 << 7)) + _sse2neon_kadd_f32(&s, &c, f32a[3] * f32b[3]); + s += c; + + float32x4_t res = { + (imm & 0x1) ? s : 0, + (imm & 0x2) ? s : 0, + (imm & 0x4) ? s : 0, + (imm & 0x8) ? s : 0, + }; + return vreinterpretq_m128_f32(res); +} + +// Extracts the selected signed or unsigned 32-bit integer from a and zero +// extends. +// FORCE_INLINE int _mm_extract_epi32(__m128i a, __constrange(0,4) int imm) +#define _mm_extract_epi32(a, imm) \ + vgetq_lane_s32(vreinterpretq_s32_m128i(a), (imm)) + +// Extracts the selected signed or unsigned 64-bit integer from a and zero +// extends. +// FORCE_INLINE __int64 _mm_extract_epi64(__m128i a, __constrange(0,2) int imm) +#define _mm_extract_epi64(a, imm) \ + vgetq_lane_s64(vreinterpretq_s64_m128i(a), (imm)) + +// Extracts the selected signed or unsigned 8-bit integer from a and zero +// extends. +// FORCE_INLINE int _mm_extract_epi8(__m128i a, __constrange(0,16) int imm) +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_extract_epi8 +#define _mm_extract_epi8(a, imm) vgetq_lane_u8(vreinterpretq_u8_m128i(a), (imm)) + +// Extracts the selected single-precision (32-bit) floating-point from a. +// FORCE_INLINE int _mm_extract_ps(__m128 a, __constrange(0,4) int imm) +#define _mm_extract_ps(a, imm) vgetq_lane_s32(vreinterpretq_s32_m128(a), (imm)) + +// Round the packed double-precision (64-bit) floating-point elements in a down +// to an integer value, and store the results as packed double-precision +// floating-point elements in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_floor_pd +FORCE_INLINE __m128d _mm_floor_pd(__m128d a) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vrndmq_f64(vreinterpretq_f64_m128d(a))); +#else + double *f = (double *)&a; + return _mm_set_pd(floor(f[1]), floor(f[0])); +#endif +} + +// Round the packed single-precision (32-bit) floating-point elements in a down +// to an integer value, and store the results as packed single-precision +// floating-point elements in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_floor_ps +FORCE_INLINE __m128 _mm_floor_ps(__m128 a) +{ +#if defined(__aarch64__) + return vreinterpretq_m128_f32(vrndmq_f32(vreinterpretq_f32_m128(a))); +#else + float *f = (float *)&a; + return _mm_set_ps(floorf(f[3]), floorf(f[2]), floorf(f[1]), floorf(f[0])); +#endif +} + +// Round the lower double-precision (64-bit) floating-point element in b down to +// an integer value, store the result as a double-precision floating-point +// element in the lower element of dst, and copy the upper element from a to the +// upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_floor_sd +FORCE_INLINE __m128d _mm_floor_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_floor_pd(b)); +} + +// Round the lower single-precision (32-bit) floating-point element in b down to +// an integer value, store the result as a single-precision floating-point +// element in the lower element of dst, and copy the upper 3 packed elements +// from a to the upper elements of dst. +// +// dst[31:0] := FLOOR(b[31:0]) +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_floor_ss +FORCE_INLINE __m128 _mm_floor_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_floor_ps(b)); +} + +// Inserts the least significant 32 bits of b into the selected 32-bit integer +// of a. +// FORCE_INLINE __m128i _mm_insert_epi32(__m128i a, int b, +// __constrange(0,4) int imm) +#define _mm_insert_epi32(a, b, imm) \ + __extension__({ \ + vreinterpretq_m128i_s32( \ + vsetq_lane_s32((b), vreinterpretq_s32_m128i(a), (imm))); \ + }) + +// Inserts the least significant 64 bits of b into the selected 64-bit integer +// of a. +// FORCE_INLINE __m128i _mm_insert_epi64(__m128i a, __int64 b, +// __constrange(0,2) int imm) +#define _mm_insert_epi64(a, b, imm) \ + __extension__({ \ + vreinterpretq_m128i_s64( \ + vsetq_lane_s64((b), vreinterpretq_s64_m128i(a), (imm))); \ + }) + +// Inserts the least significant 8 bits of b into the selected 8-bit integer +// of a. +// FORCE_INLINE __m128i _mm_insert_epi8(__m128i a, int b, +// __constrange(0,16) int imm) +#define _mm_insert_epi8(a, b, imm) \ + __extension__({ \ + vreinterpretq_m128i_s8( \ + vsetq_lane_s8((b), vreinterpretq_s8_m128i(a), (imm))); \ + }) + +// Copy a to tmp, then insert a single-precision (32-bit) floating-point +// element from b into tmp using the control in imm8. Store tmp to dst using +// the mask in imm8 (elements are zeroed out when the corresponding bit is set). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=insert_ps +#define _mm_insert_ps(a, b, imm8) \ + __extension__({ \ + float32x4_t tmp1 = \ + vsetq_lane_f32(vgetq_lane_f32(b, (imm8 >> 6) & 0x3), \ + vreinterpretq_f32_m128(a), 0); \ + float32x4_t tmp2 = vsetq_lane_f32(vgetq_lane_f32(tmp1, 0), \ + vreinterpretq_f32_m128(a), \ + ((imm8 >> 4) & 0x3)); \ + const uint32_t data[4] = { ((imm8) & (1 << 0)) ? UINT32_MAX : 0, \ + ((imm8) & (1 << 1)) ? UINT32_MAX : 0, \ + ((imm8) & (1 << 2)) ? UINT32_MAX : 0, \ + ((imm8) & (1 << 3)) ? UINT32_MAX : 0 }; \ + uint32x4_t mask = vld1q_u32(data); \ + float32x4_t all_zeros = vdupq_n_f32(0); \ + \ + vreinterpretq_m128_f32( \ + vbslq_f32(mask, all_zeros, vreinterpretq_f32_m128(tmp2))); \ + }) + +// epi versions of min/max +// Computes the pariwise maximums of the four signed 32-bit integer values of a +// and b. +// +// A 128-bit parameter that can be defined with the following equations: +// r0 := (a0 > b0) ? a0 : b0 +// r1 := (a1 > b1) ? a1 : b1 +// r2 := (a2 > b2) ? a2 : b2 +// r3 := (a3 > b3) ? a3 : b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/bb514055(v=vs.100).aspx +FORCE_INLINE __m128i _mm_max_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + vmaxq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Compare packed signed 8-bit integers in a and b, and store packed maximum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_epi8 +FORCE_INLINE __m128i _mm_max_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s8( + vmaxq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} + +// Compare packed unsigned 16-bit integers in a and b, and store packed maximum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_epu16 +FORCE_INLINE __m128i _mm_max_epu16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vmaxq_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b))); +} + +// Compare packed unsigned 32-bit integers in a and b, and store packed maximum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_epu32 +FORCE_INLINE __m128i _mm_max_epu32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u32( + vmaxq_u32(vreinterpretq_u32_m128i(a), vreinterpretq_u32_m128i(b))); +} + +// Computes the pariwise minima of the four signed 32-bit integer values of a +// and b. +// +// A 128-bit parameter that can be defined with the following equations: +// r0 := (a0 < b0) ? a0 : b0 +// r1 := (a1 < b1) ? a1 : b1 +// r2 := (a2 < b2) ? a2 : b2 +// r3 := (a3 < b3) ? a3 : b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/bb531476(v=vs.100).aspx +FORCE_INLINE __m128i _mm_min_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + vminq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Compare packed signed 8-bit integers in a and b, and store packed minimum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_epi8 +FORCE_INLINE __m128i _mm_min_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s8( + vminq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} + +// Compare packed unsigned 16-bit integers in a and b, and store packed minimum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_epu16 +FORCE_INLINE __m128i _mm_min_epu16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vminq_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b))); +} + +// Compare packed unsigned 32-bit integers in a and b, and store packed minimum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_epu32 +FORCE_INLINE __m128i _mm_min_epu32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u32( + vminq_u32(vreinterpretq_u32_m128i(a), vreinterpretq_u32_m128i(b))); +} + +// Horizontally compute the minimum amongst the packed unsigned 16-bit integers +// in a, store the minimum and index in dst, and zero the remaining bits in dst. +// +// index[2:0] := 0 +// min[15:0] := a[15:0] +// FOR j := 0 to 7 +// i := j*16 +// IF a[i+15:i] < min[15:0] +// index[2:0] := j +// min[15:0] := a[i+15:i] +// FI +// ENDFOR +// dst[15:0] := min[15:0] +// dst[18:16] := index[2:0] +// dst[127:19] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_minpos_epu16 +FORCE_INLINE __m128i _mm_minpos_epu16(__m128i a) +{ + __m128i dst; + uint16_t min, idx = 0; + // Find the minimum value +#if defined(__aarch64__) + min = vminvq_u16(vreinterpretq_u16_m128i(a)); +#else + __m64 tmp; + tmp = vreinterpret_m64_u16( + vmin_u16(vget_low_u16(vreinterpretq_u16_m128i(a)), + vget_high_u16(vreinterpretq_u16_m128i(a)))); + tmp = vreinterpret_m64_u16( + vpmin_u16(vreinterpret_u16_m64(tmp), vreinterpret_u16_m64(tmp))); + tmp = vreinterpret_m64_u16( + vpmin_u16(vreinterpret_u16_m64(tmp), vreinterpret_u16_m64(tmp))); + min = vget_lane_u16(vreinterpret_u16_m64(tmp), 0); +#endif + // Get the index of the minimum value + int i; + for (i = 0; i < 8; i++) { + if (min == vgetq_lane_u16(vreinterpretq_u16_m128i(a), 0)) { + idx = (uint16_t)i; + break; + } + a = _mm_srli_si128(a, 2); + } + // Generate result + dst = _mm_setzero_si128(); + dst = vreinterpretq_m128i_u16( + vsetq_lane_u16(min, vreinterpretq_u16_m128i(dst), 0)); + dst = vreinterpretq_m128i_u16( + vsetq_lane_u16(idx, vreinterpretq_u16_m128i(dst), 1)); + return dst; +} + +// Compute the sum of absolute differences (SADs) of quadruplets of unsigned +// 8-bit integers in a compared to those in b, and store the 16-bit results in +// dst. Eight SADs are performed using one quadruplet from b and eight +// quadruplets from a. One quadruplet is selected from b starting at on the +// offset specified in imm8. Eight quadruplets are formed from sequential 8-bit +// integers selected from a starting at the offset specified in imm8. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mpsadbw_epu8 +FORCE_INLINE __m128i _mm_mpsadbw_epu8(__m128i a, __m128i b, const int imm) +{ + uint8x16_t _a, _b; + + switch (imm & 0x4) { + case 0: + // do nothing + _a = vreinterpretq_u8_m128i(a); + break; + case 4: + _a = vreinterpretq_u8_u32(vextq_u32(vreinterpretq_u32_m128i(a), + vreinterpretq_u32_m128i(a), 1)); + break; + default: +#if defined(__GNUC__) || defined(__clang__) + __builtin_unreachable(); +#endif + break; + } + + switch (imm & 0x3) { + case 0: + _b = vreinterpretq_u8_u32( + vdupq_n_u32(vgetq_lane_u32(vreinterpretq_u32_m128i(b), 0))); + break; + case 1: + _b = vreinterpretq_u8_u32( + vdupq_n_u32(vgetq_lane_u32(vreinterpretq_u32_m128i(b), 1))); + break; + case 2: + _b = vreinterpretq_u8_u32( + vdupq_n_u32(vgetq_lane_u32(vreinterpretq_u32_m128i(b), 2))); + break; + case 3: + _b = vreinterpretq_u8_u32( + vdupq_n_u32(vgetq_lane_u32(vreinterpretq_u32_m128i(b), 3))); + break; + default: +#if defined(__GNUC__) || defined(__clang__) + __builtin_unreachable(); +#endif + break; + } + + int16x8_t c04, c15, c26, c37; + uint8x8_t low_b = vget_low_u8(_b); + c04 = vabsq_s16(vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(_a), low_b))); + _a = vextq_u8(_a, _a, 1); + c15 = vabsq_s16(vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(_a), low_b))); + _a = vextq_u8(_a, _a, 1); + c26 = vabsq_s16(vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(_a), low_b))); + _a = vextq_u8(_a, _a, 1); + c37 = vabsq_s16(vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(_a), low_b))); +#if defined(__aarch64__) + // |0|4|2|6| + c04 = vpaddq_s16(c04, c26); + // |1|5|3|7| + c15 = vpaddq_s16(c15, c37); + + int32x4_t trn1_c = + vtrn1q_s32(vreinterpretq_s32_s16(c04), vreinterpretq_s32_s16(c15)); + int32x4_t trn2_c = + vtrn2q_s32(vreinterpretq_s32_s16(c04), vreinterpretq_s32_s16(c15)); + return vreinterpretq_m128i_s16(vpaddq_s16(vreinterpretq_s16_s32(trn1_c), + vreinterpretq_s16_s32(trn2_c))); +#else + int16x4_t c01, c23, c45, c67; + c01 = vpadd_s16(vget_low_s16(c04), vget_low_s16(c15)); + c23 = vpadd_s16(vget_low_s16(c26), vget_low_s16(c37)); + c45 = vpadd_s16(vget_high_s16(c04), vget_high_s16(c15)); + c67 = vpadd_s16(vget_high_s16(c26), vget_high_s16(c37)); + + return vreinterpretq_m128i_s16( + vcombine_s16(vpadd_s16(c01, c23), vpadd_s16(c45, c67))); +#endif +} + +// Multiply the low signed 32-bit integers from each packed 64-bit element in +// a and b, and store the signed 64-bit results in dst. +// +// r0 := (int64_t)(int32_t)a0 * (int64_t)(int32_t)b0 +// r1 := (int64_t)(int32_t)a2 * (int64_t)(int32_t)b2 +FORCE_INLINE __m128i _mm_mul_epi32(__m128i a, __m128i b) +{ + // vmull_s32 upcasts instead of masking, so we downcast. + int32x2_t a_lo = vmovn_s64(vreinterpretq_s64_m128i(a)); + int32x2_t b_lo = vmovn_s64(vreinterpretq_s64_m128i(b)); + return vreinterpretq_m128i_s64(vmull_s32(a_lo, b_lo)); +} + +// Multiplies the 4 signed or unsigned 32-bit integers from a by the 4 signed or +// unsigned 32-bit integers from b. +// https://msdn.microsoft.com/en-us/library/vstudio/bb531409(v=vs.100).aspx +FORCE_INLINE __m128i _mm_mullo_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + vmulq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Packs the 8 unsigned 32-bit integers from a and b into unsigned 16-bit +// integers and saturates. +// +// r0 := UnsignedSaturate(a0) +// r1 := UnsignedSaturate(a1) +// r2 := UnsignedSaturate(a2) +// r3 := UnsignedSaturate(a3) +// r4 := UnsignedSaturate(b0) +// r5 := UnsignedSaturate(b1) +// r6 := UnsignedSaturate(b2) +// r7 := UnsignedSaturate(b3) +FORCE_INLINE __m128i _mm_packus_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vcombine_u16(vqmovun_s32(vreinterpretq_s32_m128i(a)), + vqmovun_s32(vreinterpretq_s32_m128i(b)))); +} + +// Round the packed double-precision (64-bit) floating-point elements in a using +// the rounding parameter, and store the results as packed double-precision +// floating-point elements in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_round_pd +FORCE_INLINE __m128d _mm_round_pd(__m128d a, int rounding) +{ +#if defined(__aarch64__) + switch (rounding) { + case (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC): + return vreinterpretq_m128d_f64(vrndnq_f64(vreinterpretq_f64_m128d(a))); + case (_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC): + return _mm_floor_pd(a); + case (_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC): + return _mm_ceil_pd(a); + case (_MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC): + return vreinterpretq_m128d_f64(vrndq_f64(vreinterpretq_f64_m128d(a))); + default: //_MM_FROUND_CUR_DIRECTION + return vreinterpretq_m128d_f64(vrndiq_f64(vreinterpretq_f64_m128d(a))); + } +#else + double *v_double = (double *)&a; + + if (rounding == (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC) || + (rounding == _MM_FROUND_CUR_DIRECTION && + _MM_GET_ROUNDING_MODE() == _MM_ROUND_NEAREST)) { + double res[2], tmp; + for (int i = 0; i < 2; i++) { + tmp = (v_double[i] < 0) ? -v_double[i] : v_double[i]; + double roundDown = floor(tmp); // Round down value + double roundUp = ceil(tmp); // Round up value + double diffDown = tmp - roundDown; + double diffUp = roundUp - tmp; + if (diffDown < diffUp) { + /* If it's closer to the round down value, then use it */ + res[i] = roundDown; + } else if (diffDown > diffUp) { + /* If it's closer to the round up value, then use it */ + res[i] = roundUp; + } else { + /* If it's equidistant between round up and round down value, + * pick the one which is an even number */ + double half = roundDown / 2; + if (half != floor(half)) { + /* If the round down value is odd, return the round up value + */ + res[i] = roundUp; + } else { + /* If the round up value is odd, return the round down value + */ + res[i] = roundDown; + } + } + res[i] = (v_double[i] < 0) ? -res[i] : res[i]; + } + return _mm_set_pd(res[1], res[0]); + } else if (rounding == (_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC) || + (rounding == _MM_FROUND_CUR_DIRECTION && + _MM_GET_ROUNDING_MODE() == _MM_ROUND_DOWN)) { + return _mm_floor_pd(a); + } else if (rounding == (_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC) || + (rounding == _MM_FROUND_CUR_DIRECTION && + _MM_GET_ROUNDING_MODE() == _MM_ROUND_UP)) { + return _mm_ceil_pd(a); + } + return _mm_set_pd(v_double[1] > 0 ? floor(v_double[1]) : ceil(v_double[1]), + v_double[0] > 0 ? floor(v_double[0]) : ceil(v_double[0])); +#endif +} + +// Round the packed single-precision (32-bit) floating-point elements in a using +// the rounding parameter, and store the results as packed single-precision +// floating-point elements in dst. +// software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_round_ps +FORCE_INLINE __m128 _mm_round_ps(__m128 a, int rounding) +{ +#if defined(__aarch64__) + switch (rounding) { + case (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC): + return vreinterpretq_m128_f32(vrndnq_f32(vreinterpretq_f32_m128(a))); + case (_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC): + return _mm_floor_ps(a); + case (_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC): + return _mm_ceil_ps(a); + case (_MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC): + return vreinterpretq_m128_f32(vrndq_f32(vreinterpretq_f32_m128(a))); + default: //_MM_FROUND_CUR_DIRECTION + return vreinterpretq_m128_f32(vrndiq_f32(vreinterpretq_f32_m128(a))); + } +#else + float *v_float = (float *)&a; + + if (rounding == (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC) || + (rounding == _MM_FROUND_CUR_DIRECTION && + _MM_GET_ROUNDING_MODE() == _MM_ROUND_NEAREST)) { + uint32x4_t signmask = vdupq_n_u32(0x80000000); + float32x4_t half = vbslq_f32(signmask, vreinterpretq_f32_m128(a), + vdupq_n_f32(0.5f)); /* +/- 0.5 */ + int32x4_t r_normal = vcvtq_s32_f32(vaddq_f32( + vreinterpretq_f32_m128(a), half)); /* round to integer: [a + 0.5]*/ + int32x4_t r_trunc = vcvtq_s32_f32( + vreinterpretq_f32_m128(a)); /* truncate to integer: [a] */ + int32x4_t plusone = vreinterpretq_s32_u32(vshrq_n_u32( + vreinterpretq_u32_s32(vnegq_s32(r_trunc)), 31)); /* 1 or 0 */ + int32x4_t r_even = vbicq_s32(vaddq_s32(r_trunc, plusone), + vdupq_n_s32(1)); /* ([a] + {0,1}) & ~1 */ + float32x4_t delta = vsubq_f32( + vreinterpretq_f32_m128(a), + vcvtq_f32_s32(r_trunc)); /* compute delta: delta = (a - [a]) */ + uint32x4_t is_delta_half = + vceqq_f32(delta, half); /* delta == +/- 0.5 */ + return vreinterpretq_m128_f32( + vcvtq_f32_s32(vbslq_s32(is_delta_half, r_even, r_normal))); + } else if (rounding == (_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC) || + (rounding == _MM_FROUND_CUR_DIRECTION && + _MM_GET_ROUNDING_MODE() == _MM_ROUND_DOWN)) { + return _mm_floor_ps(a); + } else if (rounding == (_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC) || + (rounding == _MM_FROUND_CUR_DIRECTION && + _MM_GET_ROUNDING_MODE() == _MM_ROUND_UP)) { + return _mm_ceil_ps(a); + } + return _mm_set_ps(v_float[3] > 0 ? floorf(v_float[3]) : ceilf(v_float[3]), + v_float[2] > 0 ? floorf(v_float[2]) : ceilf(v_float[2]), + v_float[1] > 0 ? floorf(v_float[1]) : ceilf(v_float[1]), + v_float[0] > 0 ? floorf(v_float[0]) : ceilf(v_float[0])); +#endif +} + +// Round the lower double-precision (64-bit) floating-point element in b using +// the rounding parameter, store the result as a double-precision floating-point +// element in the lower element of dst, and copy the upper element from a to the +// upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_round_sd +FORCE_INLINE __m128d _mm_round_sd(__m128d a, __m128d b, int rounding) +{ + return _mm_move_sd(a, _mm_round_pd(b, rounding)); +} + +// Round the lower single-precision (32-bit) floating-point element in b using +// the rounding parameter, store the result as a single-precision floating-point +// element in the lower element of dst, and copy the upper 3 packed elements +// from a to the upper elements of dst. Rounding is done according to the +// rounding[3:0] parameter, which can be one of: +// (_MM_FROUND_TO_NEAREST_INT |_MM_FROUND_NO_EXC) // round to nearest, and +// suppress exceptions +// (_MM_FROUND_TO_NEG_INF |_MM_FROUND_NO_EXC) // round down, and +// suppress exceptions +// (_MM_FROUND_TO_POS_INF |_MM_FROUND_NO_EXC) // round up, and suppress +// exceptions +// (_MM_FROUND_TO_ZERO |_MM_FROUND_NO_EXC) // truncate, and suppress +// exceptions _MM_FROUND_CUR_DIRECTION // use MXCSR.RC; see +// _MM_SET_ROUNDING_MODE +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_round_ss +FORCE_INLINE __m128 _mm_round_ss(__m128 a, __m128 b, int rounding) +{ + return _mm_move_ss(a, _mm_round_ps(b, rounding)); +} + +// Load 128-bits of integer data from memory into dst using a non-temporal +// memory hint. mem_addr must be aligned on a 16-byte boundary or a +// general-protection exception may be generated. +// +// dst[127:0] := MEM[mem_addr+127:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_load_si128 +FORCE_INLINE __m128i _mm_stream_load_si128(__m128i *p) +{ +#if __has_builtin(__builtin_nontemporal_store) + return __builtin_nontemporal_load(p); +#else + return vreinterpretq_m128i_s64(vld1q_s64((int64_t *)p)); +#endif +} + +// Compute the bitwise NOT of a and then AND with a 128-bit vector containing +// all 1's, and return 1 if the result is zero, otherwise return 0. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_test_all_ones +FORCE_INLINE int _mm_test_all_ones(__m128i a) +{ + return (uint64_t)(vgetq_lane_s64(a, 0) & vgetq_lane_s64(a, 1)) == + ~(uint64_t)0; +} + +// Compute the bitwise AND of 128 bits (representing integer data) in a and +// mask, and return 1 if the result is zero, otherwise return 0. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_test_all_zeros +FORCE_INLINE int _mm_test_all_zeros(__m128i a, __m128i mask) +{ + int64x2_t a_and_mask = + vandq_s64(vreinterpretq_s64_m128i(a), vreinterpretq_s64_m128i(mask)); + return !(vgetq_lane_s64(a_and_mask, 0) | vgetq_lane_s64(a_and_mask, 1)); +} + +// Compute the bitwise AND of 128 bits (representing integer data) in a and +// mask, and set ZF to 1 if the result is zero, otherwise set ZF to 0. Compute +// the bitwise NOT of a and then AND with mask, and set CF to 1 if the result is +// zero, otherwise set CF to 0. Return 1 if both the ZF and CF values are zero, +// otherwise return 0. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_test_mix_ones_zero +FORCE_INLINE int _mm_test_mix_ones_zeros(__m128i a, __m128i mask) +{ + uint64x2_t zf = + vandq_u64(vreinterpretq_u64_m128i(mask), vreinterpretq_u64_m128i(a)); + uint64x2_t cf = + vbicq_u64(vreinterpretq_u64_m128i(mask), vreinterpretq_u64_m128i(a)); + uint64x2_t result = vandq_u64(zf, cf); + return !(vgetq_lane_u64(result, 0) | vgetq_lane_u64(result, 1)); +} + +// Compute the bitwise AND of 128 bits (representing integer data) in a and b, +// and set ZF to 1 if the result is zero, otherwise set ZF to 0. Compute the +// bitwise NOT of a and then AND with b, and set CF to 1 if the result is zero, +// otherwise set CF to 0. Return the CF value. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_testc_si128 +FORCE_INLINE int _mm_testc_si128(__m128i a, __m128i b) +{ + int64x2_t s64 = + vandq_s64(vreinterpretq_s64_s32(vmvnq_s32(vreinterpretq_s32_m128i(a))), + vreinterpretq_s64_m128i(b)); + return !(vgetq_lane_s64(s64, 0) | vgetq_lane_s64(s64, 1)); +} + +// Compute the bitwise AND of 128 bits (representing integer data) in a and b, +// and set ZF to 1 if the result is zero, otherwise set ZF to 0. Compute the +// bitwise NOT of a and then AND with b, and set CF to 1 if the result is zero, +// otherwise set CF to 0. Return 1 if both the ZF and CF values are zero, +// otherwise return 0. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_testnzc_si128 +#define _mm_testnzc_si128(a, b) _mm_test_mix_ones_zeros(a, b) + +// Compute the bitwise AND of 128 bits (representing integer data) in a and b, +// and set ZF to 1 if the result is zero, otherwise set ZF to 0. Compute the +// bitwise NOT of a and then AND with b, and set CF to 1 if the result is zero, +// otherwise set CF to 0. Return the ZF value. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_testz_si128 +FORCE_INLINE int _mm_testz_si128(__m128i a, __m128i b) +{ + int64x2_t s64 = + vandq_s64(vreinterpretq_s64_m128i(a), vreinterpretq_s64_m128i(b)); + return !(vgetq_lane_s64(s64, 0) | vgetq_lane_s64(s64, 1)); +} + +/* SSE4.2 */ + +// Compares the 2 signed 64-bit integers in a and the 2 signed 64-bit integers +// in b for greater than. +FORCE_INLINE __m128i _mm_cmpgt_epi64(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_u64( + vcgtq_s64(vreinterpretq_s64_m128i(a), vreinterpretq_s64_m128i(b))); +#else + return vreinterpretq_m128i_s64(vshrq_n_s64( + vqsubq_s64(vreinterpretq_s64_m128i(b), vreinterpretq_s64_m128i(a)), + 63)); +#endif +} + +// Starting with the initial value in crc, accumulates a CRC32 value for +// unsigned 16-bit integer v. +// https://msdn.microsoft.com/en-us/library/bb531411(v=vs.100) +FORCE_INLINE uint32_t _mm_crc32_u16(uint32_t crc, uint16_t v) +{ +#if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32) + __asm__ __volatile__("crc32ch %w[c], %w[c], %w[v]\n\t" + : [c] "+r"(crc) + : [v] "r"(v)); +#else + crc = _mm_crc32_u8(crc, v & 0xff); + crc = _mm_crc32_u8(crc, (v >> 8) & 0xff); +#endif + return crc; +} + +// Starting with the initial value in crc, accumulates a CRC32 value for +// unsigned 32-bit integer v. +// https://msdn.microsoft.com/en-us/library/bb531394(v=vs.100) +FORCE_INLINE uint32_t _mm_crc32_u32(uint32_t crc, uint32_t v) +{ +#if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32) + __asm__ __volatile__("crc32cw %w[c], %w[c], %w[v]\n\t" + : [c] "+r"(crc) + : [v] "r"(v)); +#else + crc = _mm_crc32_u16(crc, v & 0xffff); + crc = _mm_crc32_u16(crc, (v >> 16) & 0xffff); +#endif + return crc; +} + +// Starting with the initial value in crc, accumulates a CRC32 value for +// unsigned 64-bit integer v. +// https://msdn.microsoft.com/en-us/library/bb514033(v=vs.100) +FORCE_INLINE uint64_t _mm_crc32_u64(uint64_t crc, uint64_t v) +{ +#if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32) + __asm__ __volatile__("crc32cx %w[c], %w[c], %x[v]\n\t" + : [c] "+r"(crc) + : [v] "r"(v)); +#else + crc = _mm_crc32_u32((uint32_t)(crc), v & 0xffffffff); + crc = _mm_crc32_u32((uint32_t)(crc), (v >> 32) & 0xffffffff); +#endif + return crc; +} + +// Starting with the initial value in crc, accumulates a CRC32 value for +// unsigned 8-bit integer v. +// https://msdn.microsoft.com/en-us/library/bb514036(v=vs.100) +FORCE_INLINE uint32_t _mm_crc32_u8(uint32_t crc, uint8_t v) +{ +#if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32) + __asm__ __volatile__("crc32cb %w[c], %w[c], %w[v]\n\t" + : [c] "+r"(crc) + : [v] "r"(v)); +#else + crc ^= v; + for (int bit = 0; bit < 8; bit++) { + if (crc & 1) + crc = (crc >> 1) ^ UINT32_C(0x82f63b78); + else + crc = (crc >> 1); + } +#endif + return crc; +} + +/* AES */ + +#if !defined(__ARM_FEATURE_CRYPTO) +/* clang-format off */ +#define SSE2NEON_AES_DATA(w) \ + { \ + w(0x63), w(0x7c), w(0x77), w(0x7b), w(0xf2), w(0x6b), w(0x6f), \ + w(0xc5), w(0x30), w(0x01), w(0x67), w(0x2b), w(0xfe), w(0xd7), \ + w(0xab), w(0x76), w(0xca), w(0x82), w(0xc9), w(0x7d), w(0xfa), \ + w(0x59), w(0x47), w(0xf0), w(0xad), w(0xd4), w(0xa2), w(0xaf), \ + w(0x9c), w(0xa4), w(0x72), w(0xc0), w(0xb7), w(0xfd), w(0x93), \ + w(0x26), w(0x36), w(0x3f), w(0xf7), w(0xcc), w(0x34), w(0xa5), \ + w(0xe5), w(0xf1), w(0x71), w(0xd8), w(0x31), w(0x15), w(0x04), \ + w(0xc7), w(0x23), w(0xc3), w(0x18), w(0x96), w(0x05), w(0x9a), \ + w(0x07), w(0x12), w(0x80), w(0xe2), w(0xeb), w(0x27), w(0xb2), \ + w(0x75), w(0x09), w(0x83), w(0x2c), w(0x1a), w(0x1b), w(0x6e), \ + w(0x5a), w(0xa0), w(0x52), w(0x3b), w(0xd6), w(0xb3), w(0x29), \ + w(0xe3), w(0x2f), w(0x84), w(0x53), w(0xd1), w(0x00), w(0xed), \ + w(0x20), w(0xfc), w(0xb1), w(0x5b), w(0x6a), w(0xcb), w(0xbe), \ + w(0x39), w(0x4a), w(0x4c), w(0x58), w(0xcf), w(0xd0), w(0xef), \ + w(0xaa), w(0xfb), w(0x43), w(0x4d), w(0x33), w(0x85), w(0x45), \ + w(0xf9), w(0x02), w(0x7f), w(0x50), w(0x3c), w(0x9f), w(0xa8), \ + w(0x51), w(0xa3), w(0x40), w(0x8f), w(0x92), w(0x9d), w(0x38), \ + w(0xf5), w(0xbc), w(0xb6), w(0xda), w(0x21), w(0x10), w(0xff), \ + w(0xf3), w(0xd2), w(0xcd), w(0x0c), w(0x13), w(0xec), w(0x5f), \ + w(0x97), w(0x44), w(0x17), w(0xc4), w(0xa7), w(0x7e), w(0x3d), \ + w(0x64), w(0x5d), w(0x19), w(0x73), w(0x60), w(0x81), w(0x4f), \ + w(0xdc), w(0x22), w(0x2a), w(0x90), w(0x88), w(0x46), w(0xee), \ + w(0xb8), w(0x14), w(0xde), w(0x5e), w(0x0b), w(0xdb), w(0xe0), \ + w(0x32), w(0x3a), w(0x0a), w(0x49), w(0x06), w(0x24), w(0x5c), \ + w(0xc2), w(0xd3), w(0xac), w(0x62), w(0x91), w(0x95), w(0xe4), \ + w(0x79), w(0xe7), w(0xc8), w(0x37), w(0x6d), w(0x8d), w(0xd5), \ + w(0x4e), w(0xa9), w(0x6c), w(0x56), w(0xf4), w(0xea), w(0x65), \ + w(0x7a), w(0xae), w(0x08), w(0xba), w(0x78), w(0x25), w(0x2e), \ + w(0x1c), w(0xa6), w(0xb4), w(0xc6), w(0xe8), w(0xdd), w(0x74), \ + w(0x1f), w(0x4b), w(0xbd), w(0x8b), w(0x8a), w(0x70), w(0x3e), \ + w(0xb5), w(0x66), w(0x48), w(0x03), w(0xf6), w(0x0e), w(0x61), \ + w(0x35), w(0x57), w(0xb9), w(0x86), w(0xc1), w(0x1d), w(0x9e), \ + w(0xe1), w(0xf8), w(0x98), w(0x11), w(0x69), w(0xd9), w(0x8e), \ + w(0x94), w(0x9b), w(0x1e), w(0x87), w(0xe9), w(0xce), w(0x55), \ + w(0x28), w(0xdf), w(0x8c), w(0xa1), w(0x89), w(0x0d), w(0xbf), \ + w(0xe6), w(0x42), w(0x68), w(0x41), w(0x99), w(0x2d), w(0x0f), \ + w(0xb0), w(0x54), w(0xbb), w(0x16) \ + } +/* clang-format on */ + +/* X Macro trick. See https://en.wikipedia.org/wiki/X_Macro */ +#define SSE2NEON_AES_H0(x) (x) +static const uint8_t SSE2NEON_sbox[256] = SSE2NEON_AES_DATA(SSE2NEON_AES_H0); +#undef SSE2NEON_AES_H0 + +// In the absence of crypto extensions, implement aesenc using regular neon +// intrinsics instead. See: +// https://www.workofard.com/2017/01/accelerated-aes-for-the-arm64-linux-kernel/ +// https://www.workofard.com/2017/07/ghash-for-low-end-cores/ and +// https://github.com/ColinIanKing/linux-next-mirror/blob/b5f466091e130caaf0735976648f72bd5e09aa84/crypto/aegis128-neon-inner.c#L52 +// for more information Reproduced with permission of the author. +FORCE_INLINE __m128i _mm_aesenc_si128(__m128i EncBlock, __m128i RoundKey) +{ +#if defined(__aarch64__) + static const uint8_t shift_rows[] = { 0x0, 0x5, 0xa, 0xf, 0x4, 0x9, + 0xe, 0x3, 0x8, 0xd, 0x2, 0x7, + 0xc, 0x1, 0x6, 0xb }; + static const uint8_t ror32by8[] = { 0x1, 0x2, 0x3, 0x0, 0x5, 0x6, 0x7, 0x4, + 0x9, 0xa, 0xb, 0x8, 0xd, 0xe, 0xf, 0xc }; + + uint8x16_t v; + uint8x16_t w = vreinterpretq_u8_m128i(EncBlock); + + // shift rows + w = vqtbl1q_u8(w, vld1q_u8(shift_rows)); + + // sub bytes + v = vqtbl4q_u8(_sse2neon_vld1q_u8_x4(SSE2NEON_sbox), w); + v = vqtbx4q_u8(v, _sse2neon_vld1q_u8_x4(SSE2NEON_sbox + 0x40), w - 0x40); + v = vqtbx4q_u8(v, _sse2neon_vld1q_u8_x4(SSE2NEON_sbox + 0x80), w - 0x80); + v = vqtbx4q_u8(v, _sse2neon_vld1q_u8_x4(SSE2NEON_sbox + 0xc0), w - 0xc0); + + // mix columns + w = (v << 1) ^ (uint8x16_t)(((int8x16_t)v >> 7) & 0x1b); + w ^= (uint8x16_t)vrev32q_u16((uint16x8_t)v); + w ^= vqtbl1q_u8(v ^ w, vld1q_u8(ror32by8)); + + // add round key + return vreinterpretq_m128i_u8(w) ^ RoundKey; + +#else /* ARMv7-A NEON implementation */ +#define SSE2NEON_AES_B2W(b0, b1, b2, b3) \ + (((uint32_t)(b3) << 24) | ((uint32_t)(b2) << 16) | ((uint32_t)(b1) << 8) | \ + (b0)) +#define SSE2NEON_AES_F2(x) ((x << 1) ^ (((x >> 7) & 1) * 0x011b /* WPOLY */)) +#define SSE2NEON_AES_F3(x) (SSE2NEON_AES_F2(x) ^ x) +#define SSE2NEON_AES_U0(p) \ + SSE2NEON_AES_B2W(SSE2NEON_AES_F2(p), p, p, SSE2NEON_AES_F3(p)) +#define SSE2NEON_AES_U1(p) \ + SSE2NEON_AES_B2W(SSE2NEON_AES_F3(p), SSE2NEON_AES_F2(p), p, p) +#define SSE2NEON_AES_U2(p) \ + SSE2NEON_AES_B2W(p, SSE2NEON_AES_F3(p), SSE2NEON_AES_F2(p), p) +#define SSE2NEON_AES_U3(p) \ + SSE2NEON_AES_B2W(p, p, SSE2NEON_AES_F3(p), SSE2NEON_AES_F2(p)) + static const uint32_t ALIGN_STRUCT(16) aes_table[4][256] = { + SSE2NEON_AES_DATA(SSE2NEON_AES_U0), + SSE2NEON_AES_DATA(SSE2NEON_AES_U1), + SSE2NEON_AES_DATA(SSE2NEON_AES_U2), + SSE2NEON_AES_DATA(SSE2NEON_AES_U3), + }; +#undef SSE2NEON_AES_B2W +#undef SSE2NEON_AES_F2 +#undef SSE2NEON_AES_F3 +#undef SSE2NEON_AES_U0 +#undef SSE2NEON_AES_U1 +#undef SSE2NEON_AES_U2 +#undef SSE2NEON_AES_U3 + + uint32_t x0 = _mm_cvtsi128_si32(EncBlock); + uint32_t x1 = _mm_cvtsi128_si32(_mm_shuffle_epi32(EncBlock, 0x55)); + uint32_t x2 = _mm_cvtsi128_si32(_mm_shuffle_epi32(EncBlock, 0xAA)); + uint32_t x3 = _mm_cvtsi128_si32(_mm_shuffle_epi32(EncBlock, 0xFF)); + + __m128i out = _mm_set_epi32( + (aes_table[0][x3 & 0xff] ^ aes_table[1][(x0 >> 8) & 0xff] ^ + aes_table[2][(x1 >> 16) & 0xff] ^ aes_table[3][x2 >> 24]), + (aes_table[0][x2 & 0xff] ^ aes_table[1][(x3 >> 8) & 0xff] ^ + aes_table[2][(x0 >> 16) & 0xff] ^ aes_table[3][x1 >> 24]), + (aes_table[0][x1 & 0xff] ^ aes_table[1][(x2 >> 8) & 0xff] ^ + aes_table[2][(x3 >> 16) & 0xff] ^ aes_table[3][x0 >> 24]), + (aes_table[0][x0 & 0xff] ^ aes_table[1][(x1 >> 8) & 0xff] ^ + aes_table[2][(x2 >> 16) & 0xff] ^ aes_table[3][x3 >> 24])); + + return _mm_xor_si128(out, RoundKey); +#endif +} + +// Perform the last round of an AES encryption flow on data (state) in a using +// the round key in RoundKey, and store the result in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_aesenclast_si128 +FORCE_INLINE __m128i _mm_aesenclast_si128(__m128i a, __m128i RoundKey) +{ + /* FIXME: optimized for NEON */ + uint8_t v[4][4] = { + { SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 0)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 5)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 10)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 15)] }, + { SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 4)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 9)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 14)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 3)] }, + { SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 8)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 13)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 2)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 7)] }, + { SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 12)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 1)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 6)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 11)] }, + }; + for (int i = 0; i < 16; i++) + vreinterpretq_nth_u8_m128i(a, i) = + v[i / 4][i % 4] ^ vreinterpretq_nth_u8_m128i(RoundKey, i); + return a; +} + +// Emits the Advanced Encryption Standard (AES) instruction aeskeygenassist. +// This instruction generates a round key for AES encryption. See +// https://kazakov.life/2017/11/01/cryptocurrency-mining-on-ios-devices/ +// for details. +// +// https://msdn.microsoft.com/en-us/library/cc714138(v=vs.120).aspx +FORCE_INLINE __m128i _mm_aeskeygenassist_si128(__m128i key, const int rcon) +{ + uint32_t X1 = _mm_cvtsi128_si32(_mm_shuffle_epi32(key, 0x55)); + uint32_t X3 = _mm_cvtsi128_si32(_mm_shuffle_epi32(key, 0xFF)); + for (int i = 0; i < 4; ++i) { + ((uint8_t *)&X1)[i] = SSE2NEON_sbox[((uint8_t *)&X1)[i]]; + ((uint8_t *)&X3)[i] = SSE2NEON_sbox[((uint8_t *)&X3)[i]]; + } + return _mm_set_epi32(((X3 >> 8) | (X3 << 24)) ^ rcon, X3, + ((X1 >> 8) | (X1 << 24)) ^ rcon, X1); +} +#undef SSE2NEON_AES_DATA + +#else /* __ARM_FEATURE_CRYPTO */ +// Implements equivalent of 'aesenc' by combining AESE (with an empty key) and +// AESMC and then manually applying the real key as an xor operation. This +// unfortunately means an additional xor op; the compiler should be able to +// optimize this away for repeated calls however. See +// https://blog.michaelbrase.com/2018/05/08/emulating-x86-aes-intrinsics-on-armv8-a +// for more details. +FORCE_INLINE __m128i _mm_aesenc_si128(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vaesmcq_u8(vaeseq_u8(vreinterpretq_u8_m128i(a), vdupq_n_u8(0))) ^ + vreinterpretq_u8_m128i(b)); +} + +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_aesenclast_si128 +FORCE_INLINE __m128i _mm_aesenclast_si128(__m128i a, __m128i RoundKey) +{ + return _mm_xor_si128(vreinterpretq_m128i_u8(vaeseq_u8( + vreinterpretq_u8_m128i(a), vdupq_n_u8(0))), + RoundKey); +} + +FORCE_INLINE __m128i _mm_aeskeygenassist_si128(__m128i a, const int rcon) +{ + // AESE does ShiftRows and SubBytes on A + uint8x16_t u8 = vaeseq_u8(vreinterpretq_u8_m128i(a), vdupq_n_u8(0)); + + uint8x16_t dest = { + // Undo ShiftRows step from AESE and extract X1 and X3 + u8[0x4], u8[0x1], u8[0xE], u8[0xB], // SubBytes(X1) + u8[0x1], u8[0xE], u8[0xB], u8[0x4], // ROT(SubBytes(X1)) + u8[0xC], u8[0x9], u8[0x6], u8[0x3], // SubBytes(X3) + u8[0x9], u8[0x6], u8[0x3], u8[0xC], // ROT(SubBytes(X3)) + }; + uint32x4_t r = { 0, (unsigned)rcon, 0, (unsigned)rcon }; + return vreinterpretq_m128i_u8(dest) ^ vreinterpretq_m128i_u32(r); +} +#endif + +/* Others */ + +// Perform a carry-less multiplication of two 64-bit integers, selected from a +// and b according to imm8, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_clmulepi64_si128 +FORCE_INLINE __m128i _mm_clmulepi64_si128(__m128i _a, __m128i _b, const int imm) +{ + uint64x2_t a = vreinterpretq_u64_m128i(_a); + uint64x2_t b = vreinterpretq_u64_m128i(_b); + switch (imm & 0x11) { + case 0x00: + return vreinterpretq_m128i_u64( + _sse2neon_vmull_p64(vget_low_u64(a), vget_low_u64(b))); + case 0x01: + return vreinterpretq_m128i_u64( + _sse2neon_vmull_p64(vget_high_u64(a), vget_low_u64(b))); + case 0x10: + return vreinterpretq_m128i_u64( + _sse2neon_vmull_p64(vget_low_u64(a), vget_high_u64(b))); + case 0x11: + return vreinterpretq_m128i_u64( + _sse2neon_vmull_p64(vget_high_u64(a), vget_high_u64(b))); + default: + abort(); + } +} + +FORCE_INLINE unsigned int _sse2neon_mm_get_denormals_zero_mode() +{ + union { + fpcr_bitfield field; +#if defined(__aarch64__) + uint64_t value; +#else + uint32_t value; +#endif + } r; + +#if defined(__aarch64__) + asm volatile("mrs %0, FPCR" : "=r"(r.value)); /* read */ +#else + asm volatile("vmrs %0, FPSCR" : "=r"(r.value)); /* read */ +#endif + + return r.field.bit24 ? _MM_DENORMALS_ZERO_ON : _MM_DENORMALS_ZERO_OFF; +} + +// Count the number of bits set to 1 in unsigned 32-bit integer a, and +// return that count in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_popcnt_u32 +FORCE_INLINE int _mm_popcnt_u32(unsigned int a) +{ +#if defined(__aarch64__) +#if __has_builtin(__builtin_popcount) + return __builtin_popcount(a); +#else + return (int)vaddlv_u8(vcnt_u8(vcreate_u8((uint64_t)a))); +#endif +#else + uint32_t count = 0; + uint8x8_t input_val, count8x8_val; + uint16x4_t count16x4_val; + uint32x2_t count32x2_val; + + input_val = vld1_u8((uint8_t *)&a); + count8x8_val = vcnt_u8(input_val); + count16x4_val = vpaddl_u8(count8x8_val); + count32x2_val = vpaddl_u16(count16x4_val); + + vst1_u32(&count, count32x2_val); + return count; +#endif +} + +// Count the number of bits set to 1 in unsigned 64-bit integer a, and +// return that count in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_popcnt_u64 +FORCE_INLINE int64_t _mm_popcnt_u64(uint64_t a) +{ +#if defined(__aarch64__) +#if __has_builtin(__builtin_popcountll) + return __builtin_popcountll(a); +#else + return (int64_t)vaddlv_u8(vcnt_u8(vcreate_u8(a))); +#endif +#else + uint64_t count = 0; + uint8x8_t input_val, count8x8_val; + uint16x4_t count16x4_val; + uint32x2_t count32x2_val; + uint64x1_t count64x1_val; + + input_val = vld1_u8((uint8_t *)&a); + count8x8_val = vcnt_u8(input_val); + count16x4_val = vpaddl_u8(count8x8_val); + count32x2_val = vpaddl_u16(count16x4_val); + count64x1_val = vpaddl_u32(count32x2_val); + vst1_u64(&count, count64x1_val); + return count; +#endif +} + +FORCE_INLINE void _sse2neon_mm_set_denormals_zero_mode(unsigned int flag) +{ + // AArch32 Advanced SIMD arithmetic always uses the Flush-to-zero setting, + // regardless of the value of the FZ bit. + union { + fpcr_bitfield field; +#if defined(__aarch64__) + uint64_t value; +#else + uint32_t value; +#endif + } r; + +#if defined(__aarch64__) + asm volatile("mrs %0, FPCR" : "=r"(r.value)); /* read */ +#else + asm volatile("vmrs %0, FPSCR" : "=r"(r.value)); /* read */ +#endif + + r.field.bit24 = (flag & _MM_DENORMALS_ZERO_MASK) == _MM_DENORMALS_ZERO_ON; + +#if defined(__aarch64__) + asm volatile("msr FPCR, %0" ::"r"(r)); /* write */ +#else + asm volatile("vmsr FPSCR, %0" ::"r"(r)); /* write */ +#endif +} + +#if defined(__GNUC__) || defined(__clang__) +#pragma pop_macro("ALIGN_STRUCT") +#pragma pop_macro("FORCE_INLINE") +#endif + +#if defined(__GNUC__) && !defined(__clang__) +#pragma GCC pop_options +#endif + +#endif diff --git a/core/include/olive/core/util/stringutils.h b/core/include/olive/core/util/stringutils.h new file mode 100644 index 000000000..8ac0dad97 --- /dev/null +++ b/core/include/olive/core/util/stringutils.h @@ -0,0 +1,211 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#ifndef LIBOLIVECORE_STRINGUTILS_H +#define LIBOLIVECORE_STRINGUTILS_H + +#include +#include +#include +#include + +namespace olive::core +{ + +class StringUtils { +public: + /** + * @brief Split a string into a list of strings using a specific delimiter + * + * @param s + * + * The string to split. + * + * @param separator + * + * The character to split the string on. + * + * @return + * + * A vector of strings split by the specified delimiter. + */ + static std::vector split(const std::string &s, char separator); + + /** + * @brief Splits a string into a list of strings using regular expressions. + * + * @param s + * + * The string to split. + * + * @param regex + * + * The regular expression to split the string on. + * + * @return + * + * A vector of strings split wherever the regular expression matched. + */ + static std::vector split_regex(const std::string &s, + const std::regex ®ex); + + /** + * @brief Convert a string to int using a bool pointer to determine success rather than an exception + * + * @param s + * + * The string to parse an int from. + * + * @param base + * + * The base of the number in the string (usually 10, or 16 for hex). + * + * @param ok + * + * (Optional) a boolean output parameter specifying whether the conversion was successful or not. + * + * @return + * + * Either the int parsed from the string, or 0 (with *ok set to false) on parser error. + */ + static int to_int(const std::string &s, int base, bool *ok = nullptr); + + /** + * @brief Overloaded function + * + * @param s + * + * The string to parse an int from. + * + * @param ok + * + * (Optional) a boolean output parameter specifying whether the conversion was successful or not. + * + * @return + * + * Either the int parsed from the string, or 0 (with *ok set to false) on parser error. + */ + static int to_int(const std::string &s, bool *ok = nullptr) + { + return to_int(s, 10, ok); + } + + /** + * @brief Convert a number to a string with left padding + * + * Usually used for converting a number to a string with leading zeroes. + * + * @param val + * + * The number to convert. This is a templated function and will accept any type, e.g. + * int/long/float/double/etc. + * + * @param padding + * + * Total desired length of the string. For example, setting this to `2` will ensure the string + * is at least 2 characters in size, using `c` to pad the left side where necessary. + * + * @param c + * + * The character to pad with. This defaults to `0` assuming you'll be using this function to + * create leading zeroes. + * + * @return + * + * The padded string. + */ + template + static std::string to_string_leftpad(T val, size_t padding, char c = '0') + { + std::string s = std::to_string(val); + + if (s.size() < padding) { + s.insert(0, padding - s.size(), c); + } + + return s; + } + + /** + * @brief Format a string + * + * A sprintf wrapper that returns a std::string. + * + * @param fmt + * + * The format to use. + * + * @return + * + * A formatted string in std::string form. + */ + static std::string format(const char *fmt, ...); + + // trim from start (in place) + static inline void ltrim(std::string &s) + { + s.erase(s.begin(), + std::find_if(s.begin(), s.end(), [](unsigned char ch) { + return !std::isspace(ch); + })); + } + + // trim from end (in place) + static inline void rtrim(std::string &s) + { + s.erase(std::find_if(s.rbegin(), s.rend(), + [](unsigned char ch) { return !std::isspace(ch); }) + .base(), + s.end()); + } + + // trim from both ends (in place) + static inline void trim(std::string &s) + { + rtrim(s); + ltrim(s); + } + + // trim from start (copying) + static inline std::string ltrimmed(std::string s) + { + ltrim(s); + return s; + } + + // trim from end (copying) + static inline std::string rtrimmed(std::string s) + { + rtrim(s); + return s; + } + + // trim from both ends (copying) + static inline std::string trimmed(std::string s) + { + trim(s); + return s; + } +}; + +} + +#endif // LIBOLIVECORE_STRINGUTILS_H diff --git a/core/include/olive/core/util/tests.h b/core/include/olive/core/util/tests.h new file mode 100644 index 000000000..09296aff4 --- /dev/null +++ b/core/include/olive/core/util/tests.h @@ -0,0 +1,62 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#ifndef LIBOLIVECORE_TESTS_H +#define LIBOLIVECORE_TESTS_H + +#include + +namespace olive::core +{ + +class Tester { +public: + Tester() = default; + + typedef bool (*test_t)(); + + void add(const char *name, test_t test_function) + { + test_names_.push_back(name); + test_functions_.push_back(test_function); + } + + bool run(); + + int exec() + { + if (run()) { + return 0; + } else { + return 1; + } + } + + static void echo(const char *fmt, ...); + +private: + std::list test_names_; + std::list test_functions_; +}; + +} + +#endif // LIBOLIVECORE_TESTS_H diff --git a/core/include/olive/core/util/timecodefunctions.h b/core/include/olive/core/util/timecodefunctions.h new file mode 100644 index 000000000..301c6d5eb --- /dev/null +++ b/core/include/olive/core/util/timecodefunctions.h @@ -0,0 +1,91 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#ifndef LIBOLIVECORE_TIMECODEFUNCTIONS_H +#define LIBOLIVECORE_TIMECODEFUNCTIONS_H + +#include "rational.h" + +namespace olive::core +{ + +/** + * @brief Functions for converting times/timecodes/timestamps + * + * Olive uses the following terminology through its code: + * + * `time` - time in seconds presented in a rational form + * `timebase` - the base time unit of an audio/video stream in seconds + * `timestamp` - an integer representation of a time in timebase units (in many cases is used like a frame number) + * `timecode` a user-friendly string representation of a time according to Timecode::Display + */ +class Timecode { +public: + enum Display { + kTimecodeDropFrame, + kTimecodeNonDropFrame, + kTimecodeSeconds, + kFrames, + kMilliseconds + }; + + enum Rounding { kCeil, kFloor, kRound }; + + /** + * @brief Convert a timestamp (according to a rational timebase) to a user-friendly string representation + */ + static std::string time_to_timecode(const rational &time, + const rational &timebase, + const Display &display, + bool show_plus_if_positive = false); + static rational timecode_to_time(std::string timecode, + const rational &timebase, + const Display &display, + bool *ok = nullptr); + + static std::string time_to_string(int64_t ms); + + static rational snap_time_to_timebase(const rational &time, + const rational &timebase, + Rounding floor = kRound); + + static int64_t time_to_timestamp(const rational &time, + const rational &timebase, + Rounding floor = kRound); + static int64_t time_to_timestamp(const double &time, + const rational &timebase, + Rounding floor = kRound); + + static int64_t rescale_timestamp(const int64_t &ts, const rational &source, + const rational &dest); + static int64_t rescale_timestamp_ceil(const int64_t &ts, + const rational &source, + const rational &dest); + + static rational timestamp_to_time(const int64_t ×tamp, + const rational &timebase); + + static bool timebase_is_drop_frame(const rational &timebase); +}; + +} + +#endif // LIBOLIVECORE_TIMECODEFUNCTIONS_H diff --git a/core/include/olive/core/util/timerange.h b/core/include/olive/core/util/timerange.h new file mode 100644 index 000000000..72918d90c --- /dev/null +++ b/core/include/olive/core/util/timerange.h @@ -0,0 +1,315 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#ifndef LIBOLIVECORE_TIMERANGE_H +#define LIBOLIVECORE_TIMERANGE_H + +#include +#include + +#include "rational.h" + +namespace olive::core +{ + +class TimeRange { +public: + TimeRange() = default; + TimeRange(const rational &in, const rational &out); + TimeRange(const TimeRange &r) + : TimeRange(r.in(), r.out()) + { + } + + TimeRange &operator=(const TimeRange &r) + { + set_range(r.in(), r.out()); + return *this; + } + + const rational &in() const; + const rational &out() const; + const rational &length() const; + + void set_in(const rational &in); + void set_out(const rational &out); + void set_range(const rational &in, const rational &out); + + bool operator==(const TimeRange &r) const; + bool operator!=(const TimeRange &r) const; + + bool OverlapsWith(const TimeRange &a, bool in_inclusive = true, + bool out_inclusive = true) const; + bool Contains(const TimeRange &a, bool in_inclusive = true, + bool out_inclusive = true) const; + bool Contains(const rational &r) const; + + TimeRange Combined(const TimeRange &a) const; + static TimeRange Combine(const TimeRange &a, const TimeRange &b); + TimeRange Intersected(const TimeRange &a) const; + static TimeRange Intersect(const TimeRange &a, const TimeRange &b); + + TimeRange operator+(const rational &rhs) const; + TimeRange operator-(const rational &rhs) const; + + const TimeRange &operator+=(const rational &rhs); + const TimeRange &operator-=(const rational &rhs); + + std::list Split(const int &chunk_size) const; + +private: + void normalize(); + + rational in_; + rational out_; + rational length_; +}; + +class TimeRangeList { +public: + TimeRangeList() = default; + + TimeRangeList(std::initializer_list r) + : array_(r) + { + } + + void insert(const TimeRangeList &list_to_add); + void insert(TimeRange range_to_add); + + void remove(const TimeRange &remove); + void remove(const TimeRangeList &list); + + template + static void util_remove(std::vector *list, const TimeRange &remove) + { + std::vector additions; + + for (auto it = list->begin(); it != list->end();) { + T &compare = *it; + + if (remove.Contains(compare)) { + // This element is entirely encompassed in this range, remove it + it = list->erase(it); + } else { + if (compare.Contains(remove, false, false)) { + // The remove range is within this element, only choice is to split the element into two + T new_range = compare; + new_range.set_in(remove.out()); + compare.set_out(remove.in()); + + additions.push_back(new_range); + break; + } else { + if (compare.in() < remove.in() && + compare.out() > remove.in()) { + // This element's out point overlaps the range's in, we'll trim it + compare.set_out(remove.in()); + } else if (compare.in() < remove.out() && + compare.out() > remove.out()) { + // This element's in point overlaps the range's out, we'll trim it + compare.set_in(remove.out()); + } + + it++; + } + } + } + + list->insert(list->end(), additions.begin(), additions.end()); + } + + bool contains(const TimeRange &range, bool in_inclusive = true, + bool out_inclusive = true) const; + + bool contains(const rational &r) const + { + for (const TimeRange &range : array_) { + if (range.Contains(r)) { + return true; + } + } + + return false; + } + + bool OverlapsWith(const TimeRange &r, bool in_inclusive = true, + bool out_inclusive = true) const + { + for (const TimeRange &range : array_) { + if (range.OverlapsWith(r, in_inclusive, out_inclusive)) { + return true; + } + } + + return false; + } + + bool isEmpty() const + { + return array_.empty(); + } + + void clear() + { + array_.clear(); + } + + int size() const + { + return array_.size(); + } + + void shift(const rational &diff); + + void trim_in(const rational &diff); + + void trim_out(const rational &diff); + + TimeRangeList Intersects(const TimeRange &range) const; + + using const_iterator = std::vector::const_iterator; + + const_iterator begin() const + { + return array_.cbegin(); + } + + const_iterator end() const + { + return array_.cend(); + } + + const_iterator cbegin() const + { + return begin(); + } + + const_iterator cend() const + { + return end(); + } + + const TimeRange &first() const + { + return array_.front(); + } + + const TimeRange &last() const + { + return array_.back(); + } + + const TimeRange &at(int index) const + { + return array_.at(index); + } + + const std::vector &internal_array() const + { + return array_; + } + + bool operator==(const TimeRangeList &rhs) const + { + return array_ == rhs.array_; + } + +private: + std::vector array_; +}; + +class TimeRangeListFrameIterator { +public: + TimeRangeListFrameIterator(); + TimeRangeListFrameIterator(const TimeRangeList &list, + const rational &timebase); + + rational Snap(const rational &r) const; + + bool GetNext(rational *out); + + bool HasNext() const; + + std::vector ToVector() const + { + TimeRangeListFrameIterator copy(list_, timebase_); + std::vector times; + rational r; + while (copy.GetNext(&r)) { + times.push_back(r); + } + return times; + } + + int size(); + + void reset() + { + *this = TimeRangeListFrameIterator(); + } + + void insert(const TimeRange &range) + { + list_.insert(range); + } + + void insert(const TimeRangeList &list) + { + list_.insert(list); + } + + bool IsCustomRange() const + { + return custom_range_; + } + + void SetCustomRange(bool e) + { + custom_range_ = e; + } + + int frame_index() const + { + return frame_index_; + } + +private: + void UpdateIndexIfNecessary(); + + TimeRangeList list_; + + rational timebase_; + + rational current_; + + int range_index_; + + int size_; + + int frame_index_; + + bool custom_range_; +}; + +} + +#endif // LIBOLIVECORE_TIMERANGE_H diff --git a/core/include/olive/core/util/value.h b/core/include/olive/core/util/value.h new file mode 100644 index 000000000..bb9023544 --- /dev/null +++ b/core/include/olive/core/util/value.h @@ -0,0 +1,96 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#ifndef LIBOLIVECORE_VALUE_H +#define LIBOLIVECORE_VALUE_H + +#include +#include +#include +#include +#include + +namespace olive::core +{ + +/** + * @brief Generic type container + */ +class Value { +public: + enum Type { + /// Null/no data + NONE, + + /// Signed int64 + INT, + + /// Double-precision float + FLOAT, + + /// UTF-8 string + STRING + }; + + Value() + { + type_ = NONE; + } + + Value(int64_t v) + { + data_.resize(sizeof(int64_t)); + memcpy(data_.data(), &v, sizeof(int64_t)); + type_ = INT; + } + + Value(double v) + { + data_.resize(sizeof(double)); + memcpy(data_.data(), &v, sizeof(int64_t)); + type_ = FLOAT; + } + + Value(const char *s) + { + size_t sz = strlen(s); + data_.resize(sz); + memcpy(data_.data(), s, sz); + type_ = STRING; + } + + Value(const std::string &s) + { + data_.resize(s.size()); + memcpy(data_.data(), s.data(), data_.size()); + type_ = STRING; + } + +private: + std::vector data_; + Type type_; +}; + +using ValueMap = std::map; + +} + +#endif // LIBOLIVECORE_VALUE_H diff --git a/core/src/render/audioparams.cpp b/core/src/render/audioparams.cpp new file mode 100644 index 000000000..6ab50cf76 --- /dev/null +++ b/core/src/render/audioparams.cpp @@ -0,0 +1,237 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include "render/audioparams.h" + +#include + +namespace olive::core +{ + +const std::vector AudioParams::kSupportedSampleRates = { + 8000, // 8000 Hz + 11025, // 11025 Hz + 16000, // 16000 Hz + 22050, // 22050 Hz + 24000, // 24000 Hz + 32000, // 32000 Hz + 44100, // 44100 Hz + 48000, // 48000 Hz + 88200, // 88200 Hz + 96000 // 96000 Hz +}; + +const std::vector AudioParams::kSupportedChannelLayouts = { + AV_CH_LAYOUT_MONO, AV_CH_LAYOUT_STEREO, AV_CH_LAYOUT_2_1, + AV_CH_LAYOUT_5POINT1, AV_CH_LAYOUT_7POINT1 +}; + +bool AudioParams::operator==(const AudioParams &other) const +{ + return format() == other.format() && sample_rate() == other.sample_rate() && + time_base() == other.time_base() && + av_channel_layout_compare(&channel_layout_, + &other.channel_layout()) == 0; +} + +bool AudioParams::operator!=(const AudioParams &other) const +{ + return !(*this == other); +} + +int64_t AudioParams::time_to_bytes(const double &time) const +{ + return time_to_bytes_per_channel(time) * channel_count(); +} + +int64_t AudioParams::time_to_bytes(const rational &time) const +{ + return time_to_bytes(time.toDouble()); +} + +int64_t AudioParams::time_to_bytes_per_channel(const double &time) const +{ + assert(is_valid()); + + return int64_t(time_to_samples(time)) * bytes_per_sample_per_channel(); +} + +int64_t AudioParams::time_to_bytes_per_channel(const rational &time) const +{ + return time_to_bytes_per_channel(time.toDouble()); +} + +int64_t AudioParams::time_to_samples(const double &time) const +{ + assert(is_valid()); + + return std::round(double(sample_rate()) * time); +} + +int64_t AudioParams::time_to_samples(const rational &time) const +{ + return time_to_samples(time.toDouble()); +} + +int64_t AudioParams::samples_to_bytes(const int64_t &samples) const +{ + assert(is_valid()); + + return samples_to_bytes_per_channel(samples) * channel_count(); +} + +int64_t AudioParams::samples_to_bytes_per_channel(const int64_t &samples) const +{ + assert(is_valid()); + + return samples * bytes_per_sample_per_channel(); +} + +rational AudioParams::samples_to_time(const int64_t &samples) const +{ + return sample_rate_as_time_base() * samples; +} + +int64_t AudioParams::bytes_to_samples(const int64_t &bytes) const +{ + assert(is_valid()); + + return bytes / (channel_count() * bytes_per_sample_per_channel()); +} + +rational AudioParams::bytes_to_time(const int64_t &bytes) const +{ + assert(is_valid()); + + return samples_to_time(bytes_to_samples(bytes)); +} + +rational AudioParams::bytes_per_channel_to_time(const int64_t &bytes) const +{ + assert(is_valid()); + + return samples_to_time(bytes_to_samples(bytes * channel_count())); +} + +int AudioParams::channel_count() const +{ + return channel_count_; +} + +int AudioParams::bytes_per_sample_per_channel() const +{ + return format_.byte_count(); +} + +int AudioParams::bits_per_sample() const +{ + return bytes_per_sample_per_channel() * 8; +} + +bool AudioParams::is_valid() const +{ + return (!time_base().isNull() && + av_channel_layout_check(&channel_layout_) && + format_ > SampleFormat::INVALID && format_ < SampleFormat::COUNT); +} + +void AudioParams::calculate_channel_count() +{ + channel_count_ = channel_layout().nb_channels; +} + +/** + * @brief Copy constructor - deep copies AVChannelLayout + * + * This is critical because AVChannelLayout::u.map is a pointer for custom + * channel layouts. Default copy would share the pointer, leading to double-free. + * + * The member initializer list initializes channel_layout_ to zero ({}), + * then av_channel_layout_copy performs the deep copy from other. + * + * @param other Source AudioParams to copy from + */ +AudioParams::AudioParams(const AudioParams &other) + : sample_rate_(other.sample_rate_) + , channel_layout_{} // Zero-initialize before FFmpeg copy + , channel_count_(other.channel_count_) + , format_(other.format_) + , enabled_(other.enabled_) + , stream_index_(other.stream_index_) + , duration_(other.duration_) + , timebase_(other.timebase_) +{ + // Deep copy AVChannelLayout using FFmpeg API + // This handles all layout types: unspecified, native (mask), and custom (map) + av_channel_layout_copy(&channel_layout_, &other.channel_layout_); +} + +/** + * @brief Copy assignment - cleans up existing layout before copying + * + * CRITICAL ORDER OF OPERATIONS: + * 1. Check for self-assignment (this != &other) + * 2. Copy all scalar members + * 3. Uninitialize current channel_layout_ (frees old u.map if present) + * 4. Deep copy from other's channel_layout_ + * + * Step 3 must happen before step 4 to avoid memory leaks. If we copied first, + * we'd lose the pointer to the old u.map that needs to be freed. + * + * @param other Source AudioParams to copy from + * @return Reference to this for chaining + */ +AudioParams &AudioParams::operator=(const AudioParams &other) +{ + if (this != &other) { + // Copy scalar members first (no dependencies) + sample_rate_ = other.sample_rate_; + format_ = other.format_; + channel_count_ = other.channel_count_; + enabled_ = other.enabled_; + stream_index_ = other.stream_index_; + duration_ = other.duration_; + timebase_ = other.timebase_; + + // Free current layout's dynamic memory (u.map if custom) + av_channel_layout_uninit(&channel_layout_); + + // Deep copy from other (includes allocating new u.map if needed) + av_channel_layout_copy(&channel_layout_, &other.channel_layout_); + } + return *this; +} + +/** + * @brief Destructor - frees AVChannelLayout dynamic memory + * + * av_channel_layout_uninit() handles all cases: + * - Unspecified/Native: No-op (no dynamic memory) + * - Custom: Frees u.map array + * + * Without this, custom channel layouts would leak memory. + */ +AudioParams::~AudioParams() +{ + av_channel_layout_uninit(&channel_layout_); +} + +} diff --git a/core/src/render/samplebuffer.cpp b/core/src/render/samplebuffer.cpp new file mode 100644 index 000000000..62c664369 --- /dev/null +++ b/core/src/render/samplebuffer.cpp @@ -0,0 +1,323 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include "render/samplebuffer.h" + +#include +#include +#include +#include + +#include "util/cpuoptimize.h" +#include "util/log.h" + +namespace olive::core +{ + +SampleBuffer::SampleBuffer() + : sample_count_per_channel_(0) +{ +} + +SampleBuffer::SampleBuffer(const AudioParams &audio_params, + const rational &length) + : audio_params_(audio_params) +{ + sample_count_per_channel_ = audio_params_.time_to_samples(length); + allocate(); +} + +SampleBuffer::SampleBuffer(const AudioParams &audio_params, + size_t samples_per_channel) + : audio_params_(audio_params) + , sample_count_per_channel_(samples_per_channel) +{ + allocate(); +} + +SampleBuffer SampleBuffer::rip_channel(int channel) const +{ + AudioParams p = this->audio_params_; + AVChannelLayout layout; + av_channel_layout_from_mask(&layout, AV_CH_LAYOUT_MONO); + p.set_channel_layout(layout); + av_channel_layout_uninit(&layout); + + SampleBuffer b(p, this->sample_count_per_channel_); + b.fast_set(*this, 0, channel); + return b; +} + +std::vector SampleBuffer::rip_channel_vector(int channel) const +{ + return data_.at(channel); +} + +const AudioParams &SampleBuffer::audio_params() const +{ + return audio_params_; +} + +void SampleBuffer::set_audio_params(const AudioParams ¶ms) +{ + if (is_allocated()) { + Log::Warning() << "Tried to set parameters on allocated sample buffer"; + return; + } + + audio_params_ = params; +} + +void SampleBuffer::set_sample_count(const size_t &sample_count) +{ + if (is_allocated()) { + Log::Warning() + << "Tried to set sample count on allocated sample buffer"; + return; + } + + sample_count_per_channel_ = sample_count; +} + +void SampleBuffer::allocate() +{ + if (!audio_params_.is_valid()) { + Log::Warning() + << "Tried to allocate sample buffer with invalid audio parameters"; + return; + } + + if (!sample_count_per_channel_) { + Log::Warning() + << "Tried to allocate sample buffer with zero sample count"; + return; + } + + if (is_allocated()) { + Log::Warning() << "Tried to allocate already allocated sample buffer"; + return; + } + + data_.resize(audio_params_.channel_count()); + for (int i = 0; i < audio_params_.channel_count(); i++) { + data_[i].resize(sample_count_per_channel_); + } +} + +void SampleBuffer::destroy() +{ + data_.clear(); +} + +void SampleBuffer::reverse() +{ + if (!is_allocated()) { + Log::Warning() << "Tried to reverse an unallocated sample buffer"; + return; + } + + size_t half_nb_sample = sample_count_per_channel_ / 2; + + for (size_t i = 0; i < half_nb_sample; i++) { + size_t opposite_ind = sample_count_per_channel_ - i - 1; + + for (int j = 0; j < audio_params_.channel_count(); j++) { + std::swap(data_[j][i], data_[j][opposite_ind]); + } + } +} + +void SampleBuffer::speed(double speed) +{ + if (!is_allocated()) { + Log::Warning() << "Tried to speed an unallocated sample buffer"; + return; + } + + sample_count_per_channel_ = + std::llround(static_cast(sample_count_per_channel_) / speed); + + std::vector> output_data; + + output_data.resize(audio_params_.channel_count()); + for (int i = 0; i < audio_params_.channel_count(); i++) { + output_data[i].resize(sample_count_per_channel_); + } + + for (size_t i = 0; i < sample_count_per_channel_; i++) { + size_t input_index = std::floor(static_cast(i) * speed); + + for (int j = 0; j < audio_params_.channel_count(); j++) { + output_data[j][i] = data_[j][input_index]; + } + } + + data_ = output_data; +} + +void SampleBuffer::transform_volume(float f) +{ + transform_volume(f, this, this); +} + +void SampleBuffer::transform_volume_for_channel(int channel, float volume) +{ + transform_volume_for_channel(channel, volume, this, this); +} + +void SampleBuffer::transform_volume(float f, const SampleBuffer *input, + SampleBuffer *output) +{ + assert(input->channel_count() == output->channel_count()); + assert(input->sample_count_per_channel_ == + output->sample_count_per_channel_); + + for (int i = 0; i < input->audio_params().channel_count(); i++) { + transform_volume_for_channel(i, f, input, output); + } +} + +void SampleBuffer::transform_volume_for_channel(int channel, float volume, + const SampleBuffer *input, + SampleBuffer *output) +{ + const float *cdat = input->data_[channel].data(); + float *odat = output->data_[channel].data(); + size_t unopt_start = 0; + + assert(input->channel_count() == output->channel_count()); + assert(input->sample_count_per_channel_ == + output->sample_count_per_channel_); + +#if defined(OLIVE_PROCESSOR_X86) || defined(OLIVE_PROCESSOR_ARM) + __m128 mult = _mm_load1_ps(&volume); + unopt_start = (input->sample_count_per_channel_ / 4) * 4; + for (size_t j = 0; j < unopt_start; j += 4) { + const float *in_here = cdat + j; + float *out_here = odat + j; + __m128 samples = _mm_loadu_ps(in_here); + __m128 multiplied = _mm_mul_ps(samples, mult); + _mm_storeu_ps(out_here, multiplied); + } +#endif + + for (size_t j = unopt_start; j < input->sample_count_per_channel_; j++) { + odat[j] = cdat[j] * volume; + } +} + +void SampleBuffer::transform_volume_for_sample(size_t sample_index, + float volume) +{ + for (int i = 0; i < audio_params().channel_count(); i++) { + transform_volume_for_sample_on_channel(sample_index, i, volume); + } +} + +void SampleBuffer::transform_volume_for_sample_on_channel(size_t sample_index, + int channel, + float volume) +{ + data_[channel][sample_index] *= volume; +} + +void SampleBuffer::clamp() +{ + for (int i = 0; i < channel_count(); i++) { + clamp_channel(i); + } +} + +void SampleBuffer::silence() +{ + silence(0, sample_count_per_channel_); +} + +void SampleBuffer::silence(size_t start_sample, size_t end_sample) +{ + silence_bytes(start_sample * sizeof(float), end_sample * sizeof(float)); +} + +void SampleBuffer::silence_bytes(size_t start_byte, size_t end_byte) +{ + if (!is_allocated()) { + Log::Warning() << "Tried to fill an unallocated sample buffer"; + return; + } + + for (int i = 0; i < audio_params().channel_count(); i++) { + memset(reinterpret_cast(data_[i].data()) + start_byte, 0, + end_byte - start_byte); + } +} + +void SampleBuffer::set(int channel, const float *data, size_t sample_offset, + size_t sample_length) +{ + if (!is_allocated()) { + Log::Warning() << "Tried to fill an unallocated sample buffer"; + return; + } + + memcpy(&data_[channel].data()[sample_offset], data, + sizeof(float) * sample_length); +} + +void SampleBuffer::fast_set(const SampleBuffer &other, int to, int from) +{ + if (from == -1) { + from = to; + } + + data_[to] = other.data_[from]; +} + +void SampleBuffer::clamp_channel(int channel) +{ + const float min = -1.0f; + const float max = 1.0f; + + float *cdat = data_[channel].data(); + size_t unopt_start = 0; + +#if defined(OLIVE_PROCESSOR_X86) || defined(OLIVE_PROCESSOR_ARM) + __m128 min_sse = _mm_load1_ps(&min); + __m128 max_sse = _mm_load1_ps(&max); + + unopt_start = (sample_count_per_channel_ / 4) * 4; + for (size_t j = 0; j < unopt_start; j += 4) { + float *here = cdat + j; + __m128 samples = _mm_loadu_ps(here); + + samples = _mm_max_ps(samples, min_sse); + samples = _mm_min_ps(samples, max_sse); + + _mm_storeu_ps(here, samples); + } +#endif + + for (size_t sample = unopt_start; sample < sample_count(); sample++) { + float &s = data(channel)[sample]; + s = std::clamp(s, min, max); + } +} + +} diff --git a/core/src/util/bezier.cpp b/core/src/util/bezier.cpp new file mode 100644 index 000000000..6182a93b4 --- /dev/null +++ b/core/src/util/bezier.cpp @@ -0,0 +1,115 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include "util/bezier.h" + +#include + +namespace olive::core +{ + +Bezier::Bezier() + : x_(0) + , y_(0) + , cp1_x_(0) + , cp1_y_(0) + , cp2_x_(0) + , cp2_y_(0) +{ +} + +Bezier::Bezier(double x, double y) + : x_(x) + , y_(y) + , cp1_x_(0) + , cp1_y_(0) + , cp2_x_(0) + , cp2_y_(0) +{ +} + +Bezier::Bezier(double x, double y, double cp1_x, double cp1_y, double cp2_x, + double cp2_y) + : x_(x) + , y_(y) + , cp1_x_(cp1_x) + , cp1_y_(cp1_y) + , cp2_x_(cp2_x) + , cp2_y_(cp2_y) +{ +} + +double Bezier::QuadraticXtoT(double x, double a, double b, double c) +{ + // Clamp to prevent infinite loop + x = std::clamp(x, a, c); + + return CalculateTFromX(false, x, a, b, c, 0); +} + +double Bezier::QuadraticTtoY(double a, double b, double c, double t) +{ + return std::pow(1.0 - t, 2) * a + 2 * (1.0 - t) * t * b + + std::pow(t, 2) * c; +} + +double Bezier::CubicXtoT(double x, double a, double b, double c, double d) +{ + // Clamp to prevent infinite loop + x = std::clamp(x, a, d); + + return CalculateTFromX(true, x, a, b, c, d); +} + +double Bezier::CubicTtoY(double a, double b, double c, double d, double t) +{ + return std::pow(1.0 - t, 3) * a + 3 * std::pow(1.0 - t, 2) * t * b + + 3 * (1.0 - t) * std::pow(t, 2) * c + std::pow(t, 3) * d; +} + +double Bezier::CalculateTFromX(bool cubic, double x, double a, double b, + double c, double d) +{ + double bottom = 0.0; + double top = 1.0; + + while (true) { + if (bottom == top) { + return bottom; + } + + double mid = (bottom + top) * 0.5; + double test = cubic ? CubicTtoY(a, b, c, d, mid) : + QuadraticTtoY(a, b, c, mid); + + if (std::abs(test - x) < 0.000001) { + return mid; + } else if (x > test) { + bottom = mid; + } else { + top = mid; + } + } + + return NAN; +} + +} diff --git a/core/src/util/color.cpp b/core/src/util/color.cpp new file mode 100644 index 000000000..eda50d67c --- /dev/null +++ b/core/src/util/color.cpp @@ -0,0 +1,324 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include "util/color.h" + +#include +#include +#include +#include +#include + +namespace olive::core +{ + +Color Color::fromHsv(const DataType &h, const DataType &s, const DataType &v) +{ + DataType C = s * v; + DataType X = C * (1.0 - std::abs(std::fmod(h / 60.0, 2.0) - 1.0)); + DataType m = v - C; + DataType Rs, Gs, Bs; + + if (h >= 0.0 && h < 60.0) { + Rs = C; + Gs = X; + Bs = 0.0; + } else if (h >= 60.0 && h < 120.0) { + Rs = X; + Gs = C; + Bs = 0.0; + } else if (h >= 120.0 && h < 180.0) { + Rs = 0.0; + Gs = C; + Bs = X; + } else if (h >= 180.0 && h < 240.0) { + Rs = 0.0; + Gs = X; + Bs = C; + } else if (h >= 240.0 && h < 300.0) { + Rs = X; + Gs = 0.0; + Bs = C; + } else { + Rs = C; + Gs = 0.0; + Bs = X; + } + + return Color(Rs + m, Gs + m, Bs + m); +} + +Color::Color(const char *data, const PixelFormat &format, int ch_layout) +{ + *this = fromData(data, format, ch_layout); +} + +void Color::toHsv(DataType *hue, DataType *sat, DataType *val) const +{ + DataType fCMax = std::max(std::max(red(), green()), blue()); + DataType fCMin = std::min(std::min(red(), green()), blue()); + DataType fDelta = fCMax - fCMin; + + if (fDelta > 0) { + if (fCMax == red()) { + *hue = 60 * (fmod(((green() - blue()) / fDelta), 6)); + } else if (fCMax == green()) { + *hue = 60 * (((blue() - red()) / fDelta) + 2); + } else if (fCMax == blue()) { + *hue = 60 * (((red() - green()) / fDelta) + 4); + } + + if (fCMax > 0) { + *sat = fDelta / fCMax; + } else { + *sat = 0; + } + + *val = fCMax; + } else { + *hue = 0; + *sat = 0; + *val = fCMax; + } + + if (*hue < 0) { + *hue = 360 + *hue; + } +} + +Color::DataType Color::hsv_hue() const +{ + DataType h, s, v; + toHsv(&h, &s, &v); + return h; +} + +Color::DataType Color::hsv_saturation() const +{ + DataType h, s, v; + toHsv(&h, &s, &v); + return s; +} + +Color::DataType Color::value() const +{ + DataType h, s, v; + toHsv(&h, &s, &v); + return v; +} + +void Color::toHsl(DataType *hue, DataType *sat, DataType *lightness) const +{ + DataType fCMin = std::min(red(), std::min(green(), blue())); + DataType fCMax = std::max(red(), std::max(green(), blue())); + + *lightness = 0.5 * (fCMin + fCMax); + + if (fCMin == fCMax) { + *sat = 0; + *hue = 0; + return; + + } else if (*lightness < 0.5) { + *sat = (fCMax - fCMin) / (fCMax + fCMin); + } else { + *sat = (fCMax - fCMin) / (2.0 - fCMax - fCMin); + } + + if (fCMax == red()) { + *hue = 60 * (green() - blue()) / (fCMax - fCMin); + } + if (fCMax == green()) { + *hue = 60 * (blue() - red()) / (fCMax - fCMin) + 120; + } + if (fCMax == blue()) { + *hue = 60 * (red() - green()) / (fCMax - fCMin) + 240; + } + if (*hue < 0) { + *hue = *hue + 360; + } +} + +Color::DataType Color::hsl_hue() const +{ + DataType h, s, l; + toHsl(&h, &s, &l); + return h; +} + +Color::DataType Color::hsl_saturation() const +{ + DataType h, s, l; + toHsl(&h, &s, &l); + return s; +} + +Color::DataType Color::lightness() const +{ + DataType h, s, l; + toHsl(&h, &s, &l); + return l; +} + +void Color::toData(char *out, const PixelFormat &format, + unsigned int nb_channels) const +{ + unsigned int count = std::min(RGBA, nb_channels); + + if (format == PixelFormat::U10 && count == 4) { + const uint32_t r = static_cast(std::clamp(data_[0], DataType(0.0), DataType(1.0)) * 1023.0 + 0.5); + const uint32_t g = static_cast(std::clamp(data_[1], DataType(0.0), DataType(1.0)) * 1023.0 + 0.5); + const uint32_t b = static_cast(std::clamp(data_[2], DataType(0.0), DataType(1.0)) * 1023.0 + 0.5); + const uint32_t a = static_cast(std::clamp(data_[3], DataType(0.0), DataType(1.0)) * 3.0 + 0.5); + reinterpret_cast(out)[0] = r | (g << 10) | (b << 20) | (a << 30); + return; + } + + for (unsigned int i = 0; i < count; i++) { + DataType f = data_[i]; + + switch (format) { + case PixelFormat::INVALID: + case PixelFormat::COUNT: + break; + case PixelFormat::U8: + reinterpret_cast(out)[i] = f * 255.0; + break; + case PixelFormat::U10: + // handled above + break; + case PixelFormat::U16: + reinterpret_cast(out)[i] = f * 65535.0; + break; + case PixelFormat::F16: + reinterpret_cast(out)[i] = f; + break; + case PixelFormat::F32: + reinterpret_cast(out)[i] = f; + break; + } + } +} + +Color Color::fromData(const char *in, const PixelFormat &format, + unsigned int nb_channels) +{ + Color c; + + unsigned int count = std::min(RGBA, nb_channels); + + if (format == PixelFormat::U10 && count == 4) { + const uint32_t word = reinterpret_cast(in)[0]; + c.data_[0] = DataType((word & 0x3ff) / 1023.0); + c.data_[1] = DataType(((word >> 10) & 0x3ff) / 1023.0); + c.data_[2] = DataType(((word >> 20) & 0x3ff) / 1023.0); + c.data_[3] = DataType(((word >> 30) & 0x3) / 3.0); + return c; + } + + for (unsigned int i = 0; i < count; i++) { + DataType &f = c.data_[i]; + + switch (format) { + case PixelFormat::INVALID: + case PixelFormat::COUNT: + break; + case PixelFormat::U8: + f = DataType(reinterpret_cast(in)[i]) / 255.0; + break; + case PixelFormat::U10: + // handled above + break; + case PixelFormat::U16: + f = DataType(reinterpret_cast(in)[i]) / 65535.0; + break; + case PixelFormat::F16: + f = DataType(reinterpret_cast(in)[i]); + break; + case PixelFormat::F32: + f = DataType(reinterpret_cast(in)[i]); + break; + } + } + + return c; +} + +Color::DataType Color::GetRoughLuminance() const +{ + return (2 * red() + blue() + 3 * green()) / 6.0; +} + +Color &Color::operator+=(const Color &rhs) +{ + for (int i = 0; i < RGBA; i++) { + data_[i] += rhs.data_[i]; + } + + return *this; +} + +Color &Color::operator-=(const Color &rhs) +{ + for (int i = 0; i < RGBA; i++) { + data_[i] -= rhs.data_[i]; + } + + return *this; +} + +Color &Color::operator+=(const DataType &rhs) +{ + for (int i = 0; i < RGBA; i++) { + data_[i] += rhs; + } + + return *this; +} + +Color &Color::operator-=(const DataType &rhs) +{ + for (int i = 0; i < RGBA; i++) { + data_[i] -= rhs; + } + + return *this; +} + +Color &Color::operator*=(const DataType &rhs) +{ + for (int i = 0; i < RGBA; i++) { + data_[i] *= rhs; + } + + return *this; +} + +Color &Color::operator/=(const DataType &rhs) +{ + for (int i = 0; i < RGBA; i++) { + data_[i] /= rhs; + } + + return *this; +} + +} diff --git a/core/src/util/rational.cpp b/core/src/util/rational.cpp new file mode 100644 index 000000000..a0aae381f --- /dev/null +++ b/core/src/util/rational.cpp @@ -0,0 +1,286 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include "util/rational.h" + +#include + +#include "util/stringutils.h" + +namespace olive::core +{ + +const rational rational::NaN = rational(0, 0); + +rational rational::fromDouble(const double &flt, bool *ok) +{ + if (isnan(flt)) { + // Return NaN rational + if (ok) + *ok = false; + return NaN; + } + + // Use FFmpeg function for the time being + AVRational r = av_d2q(flt, INT_MAX); + + if (r.den == 0) { + // If den == 0, we were unable to convert to a rational + if (ok) { + *ok = false; + } + } else { + // Otherwise, assume we received a real rational + if (ok) { + *ok = true; + } + } + + return r; +} + +rational rational::fromString(const std::string &str, bool *ok) +{ + std::vector elements = StringUtils::split(str, '/'); + + switch (elements.size()) { + case 1: + return rational(StringUtils::to_int(elements.front(), ok)); + case 2: + return rational(StringUtils::to_int(elements.at(0), ok), + StringUtils::to_int(elements.at(1), ok)); + default: + // Returns NaN with ok set to false + if (ok) { + *ok = false; + } + return NaN; + } +} + +//Function: convert to double + +double rational::toDouble() const +{ + if (r_.den != 0) { + return av_q2d(r_); + } else { + return std::numeric_limits::quiet_NaN(); + } +} + +AVRational rational::toAVRational() const +{ + return r_; +} + +#ifdef USE_OTIO +opentime::RationalTime rational::toRationalTime(double framerate) const +{ + // Is this the best way of doing this? + // Olive can store rationals as 0/0 which causes errors in OTIO + opentime::RationalTime time = + opentime::RationalTime(r_.num, r_.den == 0 ? 1 : r_.den); + return time.rescaled_to(framerate); +} +#endif + +rational rational::flipped() const +{ + rational r = *this; + r.flip(); + return r; +} + +void rational::flip() +{ + if (!isNull()) { + std::swap(r_.den, r_.num); + fix_signs(); + } +} + +std::string rational::toString() const +{ + return StringUtils::format("%d/%d", r_.num, r_.den); +} + +void rational::fix_signs() +{ + if (r_.den < 0) { + // Normalize so that denominator is always positive + r_.den = -r_.den; + r_.num = -r_.num; + } else if (r_.den == 0) { + // Normalize to 0/0 (aka NaN) if denominator is zero + r_.num = 0; + } else if (r_.num == 0) { + // Normalize to 0/1 if numerator is zero + r_.den = 1; + } +} + +void rational::reduce() +{ + av_reduce(&r_.num, &r_.den, r_.num, r_.den, INT_MAX); +} + +//Assignment Operators + +const rational &rational::operator=(const rational &rhs) +{ + r_ = rhs.r_; + return *this; +} + +const rational &rational::operator+=(const rational &rhs) +{ + if (*this == RATIONAL_MIN || *this == RATIONAL_MAX || rhs == RATIONAL_MIN || + rhs == RATIONAL_MAX) { + *this = NaN; + } else if (!isNaN()) { + if (rhs.isNaN()) { + *this = NaN; + } else { + r_ = av_add_q(r_, rhs.r_); + fix_signs(); + } + } + + return *this; +} + +const rational &rational::operator-=(const rational &rhs) +{ + if (*this == RATIONAL_MIN || *this == RATIONAL_MAX || rhs == RATIONAL_MIN || + rhs == RATIONAL_MAX) { + *this = NaN; + } else if (!isNaN()) { + if (rhs.isNaN()) { + *this = NaN; + } else { + r_ = av_sub_q(r_, rhs.r_); + fix_signs(); + } + } + + return *this; +} + +const rational &rational::operator*=(const rational &rhs) +{ + if (*this == RATIONAL_MIN || *this == RATIONAL_MAX || rhs == RATIONAL_MIN || + rhs == RATIONAL_MAX) { + *this = NaN; + } else if (!isNaN()) { + if (rhs.isNaN()) { + *this = NaN; + } else { + r_ = av_mul_q(r_, rhs.r_); + fix_signs(); + } + } + + return *this; +} + +const rational &rational::operator/=(const rational &rhs) +{ + if (*this == RATIONAL_MIN || *this == RATIONAL_MAX || rhs == RATIONAL_MIN || + rhs == RATIONAL_MAX) { + *this = NaN; + } else if (!isNaN()) { + if (rhs.isNaN()) { + *this = NaN; + } else { + r_ = av_div_q(r_, rhs.r_); + fix_signs(); + } + } + + return *this; +} + +//Binary math operators + +rational rational::operator+(const rational &rhs) const +{ + rational answer(*this); + answer += rhs; + return answer; +} + +rational rational::operator-(const rational &rhs) const +{ + rational answer(*this); + answer -= rhs; + return answer; +} + +rational rational::operator/(const rational &rhs) const +{ + rational answer(*this); + answer /= rhs; + return answer; +} + +rational rational::operator*(const rational &rhs) const +{ + rational answer(*this); + answer *= rhs; + return answer; +} + +//Relational and equality operators + +bool rational::operator<(const rational &rhs) const +{ + return av_cmp_q(r_, rhs.r_) == -1; +} + +bool rational::operator<=(const rational &rhs) const +{ + int cmp = av_cmp_q(r_, rhs.r_); + return cmp == 0 || cmp == -1; +} + +bool rational::operator>(const rational &rhs) const +{ + return av_cmp_q(r_, rhs.r_) == 1; +} + +bool rational::operator>=(const rational &rhs) const +{ + int cmp = av_cmp_q(r_, rhs.r_); + return cmp == 0 || cmp == 1; +} + +bool rational::operator==(const rational &rhs) const +{ + return av_cmp_q(r_, rhs.r_) == 0; +} + +bool rational::operator!=(const rational &rhs) const +{ + return !(*this == rhs); +} + +} diff --git a/core/src/util/stringutils.cpp b/core/src/util/stringutils.cpp new file mode 100644 index 000000000..b12061648 --- /dev/null +++ b/core/src/util/stringutils.cpp @@ -0,0 +1,107 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include "util/stringutils.h" + +#include +#include + +namespace olive::core +{ + +std::vector StringUtils::split(const std::string &s, + char separator) +{ + std::vector output; + + std::string::size_type prev_pos = 0, pos = 0; + + while ((pos = s.find(separator, pos)) != std::string::npos) { + std::string substring(s.substr(prev_pos, pos - prev_pos)); + + output.push_back(substring); + + prev_pos = ++pos; + } + + output.push_back(s.substr(prev_pos, pos - prev_pos)); // Last word + + return output; +} + +std::vector StringUtils::split_regex(const std::string &s, + const std::regex ®ex) +{ + std::vector output; + + std::sregex_token_iterator iter(s.begin(), s.end(), regex, -1); + std::sregex_token_iterator end; + for (; iter != end; iter++) { + output.push_back(*iter); + } + + return output; +} + +int StringUtils::to_int(const std::string &s, int base, bool *ok) +{ + try { + int x = std::stoi(s, nullptr, base); + if (ok) { + *ok = true; + } + return x; + } catch (const std::invalid_argument &e) { + if (ok) { + *ok = false; + } + return 0; + } +} + +std::string StringUtils::format(const char *fmt, ...) +{ + va_list ap1, ap2; + va_start(ap1, fmt); + + // Need to duplicate because we call vsnprintf twice and it consumes the va_list each time + va_copy(ap2, ap1); + + int s = std::vsnprintf(nullptr, 0, fmt, ap1); + + // Create string with size, adding 1 because vsnprintf will want to write a null terminator + std::string r; + s++; + r.resize(s); + + // Write into string + std::vsnprintf(r.data(), s, fmt, ap2); + + // Pop null terminator + r.pop_back(); + + va_end(ap2); + va_end(ap1); + + return r; +} + +} diff --git a/core/src/util/tests.cpp b/core/src/util/tests.cpp new file mode 100644 index 000000000..8ba83d2fd --- /dev/null +++ b/core/src/util/tests.cpp @@ -0,0 +1,63 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include "util/tests.h" + +#include +#include +#include + +namespace olive::core +{ + +bool Tester::run() +{ + size_t index = 1; + size_t count = test_functions_.size(); + + while (!test_functions_.empty()) { + echo("[%lu/%lu] %s :: ", index, count, test_names_.front()); + + if (test_functions_.front()()) { + echo("PASSED\n"); + } else { + echo("FAILED\n"); + return false; + } + + test_names_.pop_front(); + test_functions_.pop_front(); + } + + return true; +} + +void Tester::echo(const char *fmt, ...) +{ + va_list a; + va_start(a, fmt); + + vfprintf(stderr, fmt, a); + + va_end(a); +} + +} diff --git a/core/src/util/timecodefunctions.cpp b/core/src/util/timecodefunctions.cpp new file mode 100644 index 000000000..88311f9c2 --- /dev/null +++ b/core/src/util/timecodefunctions.cpp @@ -0,0 +1,407 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include "util/timecodefunctions.h" + +extern "C" { +#include +} + +#include "util/stringutils.h" + +namespace olive::core +{ + +std::string Timecode::time_to_timecode(const rational &time, + const rational &timebase, + const Timecode::Display &display, + bool show_plus_if_positive) +{ + if (timebase.isNull() || timebase.flipped().toDouble() < 1) { + return "INVALID TIMEBASE"; + } + + double time_dbl = time.toDouble(); + + switch (display) { + case kTimecodeNonDropFrame: + case kTimecodeDropFrame: + case kTimecodeSeconds: { + const char *prefix = ""; + + if (time_dbl < 0) { + prefix = "-"; + } else if (show_plus_if_positive) { + prefix = "+"; + } + + if (display == kTimecodeSeconds) { + time_dbl = std::abs(time_dbl); + + int64_t total_seconds = std::floor(time_dbl); + + int64_t hours = total_seconds / 3600; + int64_t mins = total_seconds / 60 - hours * 60; + int64_t secs = total_seconds - mins * 60; + int64_t fraction = std::llround( + (time_dbl - static_cast(total_seconds)) * 1000); + + return StringUtils::format( + "%s%s:%s:%s.%s", prefix, + StringUtils::to_string_leftpad(hours, 2).c_str(), + StringUtils::to_string_leftpad(mins, 2).c_str(), + StringUtils::to_string_leftpad(secs, 2).c_str(), + StringUtils::to_string_leftpad(fraction, 3).c_str()); + } else { + // Determine what symbol to separate frames (";" is used for drop frame, ":" is non-drop frame) + const char *frame_token; + double frame_rate = timebase.flipped().toDouble(); + int rounded_frame_rate = std::llround(frame_rate); + int64_t frames, secs, mins, hours; + int64_t f = std::abs(time_to_timestamp(time, timebase)); + + if (display == kTimecodeDropFrame && + timebase_is_drop_frame(timebase)) { + frame_token = ";"; + + /** + * CONVERT A FRAME NUMBER TO DROP FRAME TIMECODE + * + * Code by David Heidelberger, adapted from Andrew Duncan, further adapted for Olive by Olive Team + * Given an int called framenumber and a double called framerate + * Framerate should be 29.97, 59.94, or 23.976, otherwise the calculations will be off. + */ + + // If frame number is greater than 24 hrs, next operation will rollover clock + f %= (std::llround(frame_rate * 3600) * 24); + + // Number of frames per ten minutes + int64_t framesPer10Minutes = std::llround(frame_rate * 600); + int64_t d = f / framesPer10Minutes; + int64_t m = f % framesPer10Minutes; + + // Number of frames to drop on the minute marks is the nearest integer to 6% of the framerate + int64_t dropFrames = std::llround(frame_rate * (2.0 / 30.0)); + + // Number of frames per minute is the round of the framerate * 60 minus the number of dropped frames + f += dropFrames * 9 * d; + if (m > dropFrames) { + f += dropFrames * + ((m - dropFrames) / + (std::llround(frame_rate) * 60 - dropFrames)); + } + } else { + frame_token = ":"; + } + + // non-drop timecode + hours = f / (3600 * rounded_frame_rate); + mins = f / (60 * rounded_frame_rate) % 60; + secs = f / rounded_frame_rate % 60; + frames = f % rounded_frame_rate; + + return StringUtils::format( + "%s%s:%s:%s%s%s", prefix, + StringUtils::to_string_leftpad(hours, 2).c_str(), + StringUtils::to_string_leftpad(mins, 2).c_str(), + StringUtils::to_string_leftpad(secs, 2).c_str(), frame_token, + StringUtils::to_string_leftpad(frames, 2).c_str()); + } + } + case kFrames: + return std::to_string(time_to_timestamp(time, timebase)); + case kMilliseconds: + return std::to_string(std::llround(time_dbl * 1000)); + } + + return "INVALID TIMECODE MODE"; +} + +int64_t StrToInt64EmptyTolerant(const std::string &s, bool *ok) +{ + if (s.empty()) { + if (ok) + *ok = true; + return 0; + } else { + try { + int64_t ll = std::stoll(s); + if (ok) + *ok = true; + return ll; + } catch (const std::invalid_argument &e) { + if (ok) + *ok = false; + return 0; + } + } +} + +double StrToDoubleEmptyTolerant(const std::string &s, bool *ok) +{ + if (s.empty()) { + if (ok) + *ok = true; + return 0; + } else { + try { + double d = std::stod(s); + if (ok) + *ok = true; + return d; + } catch (const std::invalid_argument &e) { + if (ok) + *ok = false; + return 0; + } + } +} + +rational Timecode::timecode_to_time(std::string timecode, + const rational &timebase, + const Timecode::Display &display, bool *ok) +{ + StringUtils::trim(timecode); + if (timecode.empty()) { + goto err_fatal; + } + + switch (display) { + case kTimecodeNonDropFrame: + case kTimecodeDropFrame: + case kTimecodeSeconds: { + std::vector timecode_split = + StringUtils::split_regex(timecode, std::regex("(:)|(;)")); + + const int element_count = display == kTimecodeSeconds ? 3 : 4; + + // Remove excess tokens (we're only interested in HH:MM:SS.FF) + if (timecode_split.size() > element_count) { + timecode_split.resize(element_count); + } + + // For easier index calculations, ensure minimum size + if (timecode_split.size() < element_count) { + timecode_split.insert(timecode_split.begin(), + element_count - timecode_split.size(), + std::string()); + } + + bool negative = (timecode.at(0) == '-'); + + double frame_rate = timebase.flipped().toDouble(); + int rounded_frame_rate = std::lround(frame_rate); + + bool valid; + rational time; + + int64_t hours = StrToInt64EmptyTolerant(timecode_split.at(0), &valid); + if (!valid) + goto err_fatal; + int64_t mins = StrToInt64EmptyTolerant(timecode_split.at(1), &valid); + if (!valid) + goto err_fatal; + + if (display == kTimecodeSeconds) { + double secs = + StrToDoubleEmptyTolerant(timecode_split.at(2), &valid); + if (!valid) + goto err_fatal; + + time = rational::fromDouble(hours * 3600 + mins * 60 + secs); + } else { + int64_t secs = + StrToInt64EmptyTolerant(timecode_split.at(2), &valid); + if (!valid) + goto err_fatal; + int64_t frames = + StrToInt64EmptyTolerant(timecode_split.at(3), &valid); + if (!valid) + goto err_fatal; + + int64_t sec_count = (hours * 3600 + mins * 60 + secs); + int64_t frame_count = sec_count * rounded_frame_rate + frames; + + if (display == kTimecodeDropFrame && + timebase_is_drop_frame(timebase)) { + // Number of frames to drop on the minute marks is the nearest integer to 6% of the framerate + int64_t dropFrames = std::llround(frame_rate * (2.0 / 30.0)); + + // d and m need to be calculated from + int64_t real_fr_ts = + std::llround(static_cast(sec_count) * frame_rate) + + frames; + + int64_t framesPer10Minutes = std::llround(frame_rate * 600); + int64_t d = real_fr_ts / framesPer10Minutes; + int64_t m = real_fr_ts % framesPer10Minutes; + + if (m > dropFrames) { + frame_count -= + dropFrames * + ((m - dropFrames) / + (std::llround(frame_rate) * 60 - dropFrames)); + } + frame_count -= dropFrames * 9 * d; + } + + time = timestamp_to_time(frame_count, timebase); + } + + if (ok) + *ok = true; + + if (negative) + time = -time; + + return time; + } + case kMilliseconds: { + try { + double timecode_secs = std::stod(timecode); + + // Convert milliseconds to seconds + timecode_secs *= 0.001; + + // Convert seconds to rational + return rational::fromDouble(timecode_secs, ok); + } catch (const std::invalid_argument &e) { + goto err_fatal; + } + } + case kFrames: { + try { + int64_t ts = std::stoll(timecode); + if (ok) + *ok = true; + return timestamp_to_time(ts, timebase); + } catch (const std::invalid_argument &e) { + goto err_fatal; + } + } + } + +err_fatal: + if (ok) + *ok = false; + return 0; +} + +std::string Timecode::time_to_string(int64_t ms) +{ + int64_t total_seconds = ms / 1000; + int64_t ss = total_seconds % 60; + int64_t mm = (total_seconds / 60) % 60; + int64_t hh = total_seconds / 3600; + + return StringUtils::format("%s:%s:%s", + StringUtils::to_string_leftpad(hh, 2).c_str(), + StringUtils::to_string_leftpad(mm, 2).c_str(), + StringUtils::to_string_leftpad(ss, 2).c_str()); +} + +rational Timecode::snap_time_to_timebase(const rational &time, + const rational &timebase, + Rounding floor) +{ + // Just convert to a timestamp in timebase units and back + int64_t timestamp = time_to_timestamp(time, timebase, floor); + + return timestamp_to_time(timestamp, timebase); +} + +rational Timecode::timestamp_to_time(const int64_t ×tamp, + const rational &timebase) +{ + int64_t num = int64_t(timebase.numerator()) * timestamp; + int64_t den = timebase.denominator(); + + int num_r, den_r; + + av_reduce(&num_r, &den_r, num, den, INT_MAX); + + return rational(num_r, den_r); +} + +bool Timecode::timebase_is_drop_frame(const rational &timebase) +{ + return (timebase.numerator() != 1); +} + +int64_t Timecode::time_to_timestamp(const rational &time, + const rational &timebase, Rounding floor) +{ + return time_to_timestamp(time.toDouble(), timebase, floor); +} + +int64_t Timecode::time_to_timestamp(const double &time, + const rational &timebase, Rounding floor) +{ + const double d = time * timebase.flipped().toDouble(); + + if (std::isnan(d)) { + return 0; + } + + const double eps = 0.000000000001; + + switch (floor) { + case kRound: + default: + return std::llround(d); + case kFloor: + if (d > std::ceil(d) - eps) { + return std::ceil(d); + } else { + return std::floor(d); + } + case kCeil: + if (d < std::floor(d) + eps) { + return std::floor(d); + } else { + return std::ceil(d); + } + } +} + +int64_t Timecode::rescale_timestamp(const int64_t &ts, const rational &source, + const rational &dest) +{ + if (source == dest) { + return ts; + } + + return av_rescale_q(ts, source.toAVRational(), dest.toAVRational()); +} + +int64_t Timecode::rescale_timestamp_ceil(const int64_t &ts, + const rational &source, + const rational &dest) +{ + if (source == dest) { + return ts; + } + + return av_rescale_q_rnd(ts, source.toAVRational(), dest.toAVRational(), + AV_ROUND_UP); +} + +} diff --git a/core/src/util/timerange.cpp b/core/src/util/timerange.cpp new file mode 100644 index 000000000..34730d7ac --- /dev/null +++ b/core/src/util/timerange.cpp @@ -0,0 +1,398 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include "util/timerange.h" + +#include +#include +#include + +#include "util/timecodefunctions.h" + +namespace olive::core +{ + +TimeRange::TimeRange(const rational &in, const rational &out) + : in_(in) + , out_(out) +{ + normalize(); +} + +const rational &TimeRange::in() const +{ + return in_; +} + +const rational &TimeRange::out() const +{ + return out_; +} + +const rational &TimeRange::length() const +{ + return length_; +} + +void TimeRange::set_in(const rational &in) +{ + in_ = in; + normalize(); +} + +void TimeRange::set_out(const rational &out) +{ + out_ = out; + normalize(); +} + +void TimeRange::set_range(const rational &in, const rational &out) +{ + in_ = in; + out_ = out; + normalize(); +} + +bool TimeRange::operator==(const TimeRange &r) const +{ + return in() == r.in() && out() == r.out(); +} + +bool TimeRange::operator!=(const TimeRange &r) const +{ + return in() != r.in() || out() != r.out(); +} + +bool TimeRange::OverlapsWith(const TimeRange &a, bool in_inclusive, + bool out_inclusive) const +{ + bool doesnt_overlap_in = (in_inclusive) ? (a.out() < in()) : + (a.out() <= in()); + + bool doesnt_overlap_out = (out_inclusive) ? (a.in() > out()) : + (a.in() >= out()); + + return !doesnt_overlap_in && !doesnt_overlap_out; +} + +TimeRange TimeRange::Combined(const TimeRange &a) const +{ + return Combine(a, *this); +} + +bool TimeRange::Contains(const TimeRange &compare, bool in_inclusive, + bool out_inclusive) const +{ + bool contains_in = (in_inclusive) ? (compare.in() >= in()) : + (compare.in() > in()); + + bool contains_out = (out_inclusive) ? (compare.out() <= out()) : + (compare.out() < out()); + + return contains_in && contains_out; +} + +bool TimeRange::Contains(const rational &r) const +{ + return r >= in_ && r < out_; +} + +TimeRange TimeRange::Combine(const TimeRange &a, const TimeRange &b) +{ + return TimeRange(std::min(a.in(), b.in()), std::max(a.out(), b.out())); +} + +TimeRange TimeRange::Intersected(const TimeRange &a) const +{ + return Intersect(a, *this); +} + +TimeRange TimeRange::Intersect(const TimeRange &a, const TimeRange &b) +{ + return TimeRange(std::max(a.in(), b.in()), std::min(a.out(), b.out())); +} + +TimeRange TimeRange::operator+(const rational &rhs) const +{ + TimeRange answer(*this); + answer += rhs; + return answer; +} + +TimeRange TimeRange::operator-(const rational &rhs) const +{ + TimeRange answer(*this); + answer -= rhs; + return answer; +} + +const TimeRange &TimeRange::operator+=(const rational &rhs) +{ + set_range(in_ + rhs, out_ + rhs); + + return *this; +} + +const TimeRange &TimeRange::operator-=(const rational &rhs) +{ + set_range(in_ - rhs, out_ - rhs); + + return *this; +} + +std::list TimeRange::Split(const int &chunk_size) const +{ + std::list split_ranges; + + int start_time = + std::floor(this->in().toDouble() / static_cast(chunk_size)) * + chunk_size; + int end_time = + std::ceil(this->out().toDouble() / static_cast(chunk_size)) * + chunk_size; + + for (int i = start_time; i < end_time; i += chunk_size) { + split_ranges.push_back( + TimeRange(std::max(this->in(), rational(i)), + std::min(this->out(), rational(i + chunk_size)))); + } + + return split_ranges; +} + +void TimeRange::normalize() +{ + // If `out` is earlier than `in`, swap them + if (out_ < in_) { + std::swap(out_, in_); + } + + // Calculate length + if (out_ == RATIONAL_MIN || out_ == RATIONAL_MAX || in_ == RATIONAL_MIN || + in_ == RATIONAL_MAX) { + length_ = rational::NaN; + } else { + length_ = out_ - in_; + } +} + +void TimeRangeList::insert(const TimeRangeList &list_to_add) +{ + for (auto it = list_to_add.cbegin(); it != list_to_add.cend(); it++) { + insert(*it); + } +} + +void TimeRangeList::insert(TimeRange range_to_add) +{ + // See if list contains this range + if (contains(range_to_add)) { + return; + } + + // Does not contain range, so we'll almost certainly be adding it in some way + for (auto it = array_.begin(); it != array_.end();) { + const TimeRange &compare = *it; + + if (compare.OverlapsWith(range_to_add)) { + range_to_add = TimeRange::Combine(range_to_add, compare); + it = array_.erase(it); + } else { + it++; + } + } + + array_.push_back(range_to_add); +} + +void TimeRangeList::remove(const TimeRange &remove) +{ + util_remove(&array_, remove); +} + +void TimeRangeList::remove(const TimeRangeList &list) +{ + for (const TimeRange &r : list) { + remove(r); + } +} + +bool TimeRangeList::contains(const TimeRange &range, bool in_inclusive, + bool out_inclusive) const +{ + for (int i = 0; i < size(); i++) { + if (array_.at(i).Contains(range, in_inclusive, out_inclusive)) { + return true; + } + } + + return false; +} + +void TimeRangeList::shift(const rational &diff) +{ + for (int i = 0; i < array_.size(); i++) { + array_[i] += diff; + } +} + +void TimeRangeList::trim_in(const rational &diff) +{ + // Re-do list since we want to handle overlaps + TimeRangeList temp = *this; + + clear(); + + for (auto it = temp.array_.begin(); it != temp.array_.end(); it++) { + TimeRange &r = *it; + r.set_in(r.in() + diff); + insert(r); + } +} + +void TimeRangeList::trim_out(const rational &diff) +{ + // Re-do list since we want to handle overlaps + TimeRangeList temp = *this; + + clear(); + + for (auto it = temp.array_.begin(); it != temp.array_.end(); it++) { + TimeRange &r = *it; + r.set_out(r.out() + diff); + insert(r); + } +} + +TimeRangeList TimeRangeList::Intersects(const TimeRange &range) const +{ + TimeRangeList intersect_list; + + for (int i = 0; i < size(); i++) { + const TimeRange &compare = array_.at(i); + + if (compare.out() <= range.in() || compare.in() >= range.out()) { + // No intersect + continue; + } else { + // Crop the time range to the range and add it to the list + TimeRange cropped(std::max(range.in(), compare.in()), + std::min(range.out(), compare.out())); + + intersect_list.insert(cropped); + } + } + + return intersect_list; +} + +TimeRangeListFrameIterator::TimeRangeListFrameIterator() + : TimeRangeListFrameIterator(TimeRangeList(), rational::NaN) +{ +} + +TimeRangeListFrameIterator::TimeRangeListFrameIterator( + const TimeRangeList &list, const rational &timebase) + : list_(list) + , timebase_(timebase) + , range_index_(-1) + , size_(-1) + , frame_index_(0) + , custom_range_(false) +{ + if (!list_.isEmpty() && timebase_.isNull()) { + std::cerr + << "TimeRangeListFrameIterator created with null timebase but non-empty list, this will likely lead to infinite loops" + << std::endl; + } + + UpdateIndexIfNecessary(); +} + +rational TimeRangeListFrameIterator::Snap(const rational &r) const +{ + return Timecode::snap_time_to_timebase(r, timebase_, Timecode::kFloor); +} + +bool TimeRangeListFrameIterator::GetNext(rational *out) +{ + if (!HasNext()) { + return false; + } + + // Output current value + *out = current_; + + // Determine next value by adding timebase + current_ += timebase_; + + // If this time is outside the current range, jump to the next one + UpdateIndexIfNecessary(); + + // Increment frame index + frame_index_++; + + return true; +} + +bool TimeRangeListFrameIterator::HasNext() const +{ + return range_index_ < list_.size(); +} + +int TimeRangeListFrameIterator::size() +{ + if (size_ == -1) { + // Size isn't calculated automatically for optimization, so we'll calculate it now + size_ = 0; + + for (const TimeRange &range : list_) { + rational start = Snap(range.in()); + rational end = Timecode::snap_time_to_timebase( + range.out(), timebase_, Timecode::kFloor); + + if (end == range.out()) { + end -= timebase_; + } + + int64_t start_ts = Timecode::time_to_timestamp(start, timebase_); + int64_t end_ts = Timecode::time_to_timestamp(end, timebase_); + + size_ += 1 + (end_ts - start_ts); + } + } + + return size_; +} + +void TimeRangeListFrameIterator::UpdateIndexIfNecessary() +{ + while (range_index_ < list_.size() && + (range_index_ == -1 || current_ >= list_.at(range_index_).out())) { + range_index_++; + + if (range_index_ < list_.size()) { + current_ = Snap(list_.at(range_index_).in()); + } + } +} + +} diff --git a/core/src/util/value.cpp b/core/src/util/value.cpp new file mode 100644 index 000000000..8e6564381 --- /dev/null +++ b/core/src/util/value.cpp @@ -0,0 +1,27 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include "util/value.h" + +namespace olive::core +{ + +} diff --git a/core/tests/rational-test.cpp b/core/tests/rational-test.cpp new file mode 100644 index 000000000..b20dcfd97 --- /dev/null +++ b/core/tests/rational-test.cpp @@ -0,0 +1,109 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include + +#include "util/rational.h" +#include "util/tests.h" + +using namespace olive::core; + +bool rational_to_from_string_test() +{ + rational r(1, 30); + + std::string s = r.toString(); + + rational r2 = rational::fromString(s); + + return r == r2; +} + +bool rational_to_from_string_test2() +{ + rational r(69, 420); + + std::string s = r.toString(); + + rational r2 = rational::fromString(s); + + return r == r2; +} + +bool rational_defaults() +{ + // By default, rationals are valid 0/1 + rational basic_constructor; + + if (!basic_constructor.isNull()) { + return false; + } + + if (basic_constructor.isNaN()) { + return false; + } + + return true; +} + +bool rational_nan() +{ + // Create a NaN with a 0 denominator + rational nan = rational(0, 0); + if (!nan.isNaN()) + return false; + if (!nan.isNull()) + return false; + + // Create a non-NaN with a zero numerator + rational zero_nonnan(0, 999); + if (!zero_nonnan.isNull()) + return false; + if (zero_nonnan.isNaN()) + return false; + + // Create a non-NaN with a non-zero numerator + rational nonzer_nonnan(1, 30); + if (nonzer_nonnan.isNull()) + return false; + if (nonzer_nonnan.isNaN()) + return false; + + return true; +} + +bool rational_nan_constant() +{ + return rational::NaN.isNaN(); +} + +int main() +{ + Tester t; + + t.add("rational::defaults", rational_defaults); + t.add("rational::NaN", rational_nan); + t.add("rational::NaN_constant", rational_nan_constant); + t.add("rational::toString/fromString", rational_to_from_string_test); + t.add("rational::toString/fromString2", rational_to_from_string_test2); + + return t.exec(); +} diff --git a/core/tests/stringutils-test.cpp b/core/tests/stringutils-test.cpp new file mode 100644 index 000000000..684b83cf6 --- /dev/null +++ b/core/tests/stringutils-test.cpp @@ -0,0 +1,47 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include + +#include "util/stringutils.h" +#include "util/tests.h" + +using namespace olive::core; + +bool stringutils_format_test() +{ + const char *expected = "Hello, world!"; + std::string f = StringUtils::format("%s, %s!", "Hello", "world"); + if (strcmp(f.c_str(), expected) != 0) { + return false; + } + + return true; +} + +int main() +{ + Tester t; + + t.add("StringUtils::format", stringutils_format_test); + + return t.exec(); +} diff --git a/core/tests/timecode-test.cpp b/core/tests/timecode-test.cpp new file mode 100644 index 000000000..42b5a09cf --- /dev/null +++ b/core/tests/timecode-test.cpp @@ -0,0 +1,65 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include + +#include "util/timecodefunctions.h" +#include "util/tests.h" + +using namespace olive::core; + +bool timecodefunctions_time_to_timecode_test() +{ + rational drop_frame_30(1001, 30000); + + std::string timecode = Timecode::time_to_timecode( + rational(1), drop_frame_30, Timecode::kTimecodeDropFrame); + if (strcmp(timecode.c_str(), "00:00:01;00") != 0) { + return false; + } + + return true; +} + +bool timecodefunctions_time_to_timecode_test2() +{ + rational bizarre_timebase(156632219); + + std::string timecode = Timecode::time_to_timecode( + rational(0), bizarre_timebase, Timecode::kTimecodeDropFrame); + if (strcmp(timecode.c_str(), "INVALID TIMEBASE") != 0) { + return false; + } + + return true; +} + +int main() +{ + Tester t; + + t.add("Timecode::time_to_timecode", + timecodefunctions_time_to_timecode_test); + t.add("Timecode::time_to_timecode2", + timecodefunctions_time_to_timecode_test2); + + return t.exec(); +} diff --git a/core/tests/timerange-test.cpp b/core/tests/timerange-test.cpp new file mode 100644 index 000000000..b7f719049 --- /dev/null +++ b/core/tests/timerange-test.cpp @@ -0,0 +1,124 @@ +/*** + + Olive - Non-Linear Video Editor + Copyright (C) 2023 Olive Studios LLC + Modifications Copyright (C) 2025 mikesolar + + 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 3 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, see . + +***/ + +#include + +#include "util/timerange.h" +#include "util/tests.h" + +using namespace olive::core; + +bool timerangelist_remove_test() +{ + TimeRangeList t; + + t.insert(TimeRange(0, 30)); + t.remove(TimeRange(2, 5)); + + return true; +} + +bool timerangelist_mergeadjacent_test() +{ + TimeRangeList t; + + // TimeRangeList should merge 1 and 3 together since they're adjacent + t.insert(TimeRange(0, 6)); + t.insert(TimeRange(20, 30)); + t.insert(TimeRange(6, 10)); + + if (!(t.size() == 2)) { + return false; + } + if (!(t.first() == TimeRange(20, 30))) { + return false; + } + if (!(t.at(1) == TimeRange(0, 10))) { + return false; + } + + // TimeRangeList should ignore these because it's already contained + TimeRangeList noop_test = t; + + noop_test.insert(TimeRange(4, 7)); + if (!(noop_test == t)) { + return false; + } + + noop_test.insert(TimeRange(0, 3)); + if (!(noop_test == t)) { + return false; + } + + noop_test.insert(TimeRange(25, 30)); + if (!(noop_test == t)) { + return false; + } + + // TimeRangeList should combine all these together + TimeRangeList combine_test_no_overlap = t; + combine_test_no_overlap.insert(TimeRange(10, 20)); + if (!(combine_test_no_overlap.size() == 1)) { + return false; + } + if (!(combine_test_no_overlap.first() == TimeRange(0, 30))) { + return false; + } + + TimeRangeList combine_test_in_overlap = t; + combine_test_in_overlap.insert(TimeRange(9, 20)); + if (!(combine_test_in_overlap.size() == 1)) { + return false; + } + if (!(combine_test_in_overlap.first() == TimeRange(0, 30))) { + return false; + } + + TimeRangeList combine_test_out_overlap = t; + combine_test_out_overlap.insert(TimeRange(10, 21)); + if (!(combine_test_out_overlap.size() == 1)) { + return false; + } + if (!(combine_test_out_overlap.first() == TimeRange(0, 30))) { + return false; + } + + TimeRangeList combine_test_both_overlap = t; + combine_test_both_overlap.insert(TimeRange(9, 21)); + if (!(combine_test_both_overlap.size() == 1)) { + return false; + } + if (!(combine_test_both_overlap.first() == TimeRange(0, 30))) { + return false; + } + + return true; +} + +int main() +{ + Tester t; + + t.add("TimeRangeList::remove", timerangelist_remove_test); + t.add("TimeRangeList::merge_adjacent", timerangelist_mergeadjacent_test); + + return t.exec(); +} diff --git a/ext/CMakeLists.txt b/ext/CMakeLists.txt deleted file mode 100644 index 448a97685..000000000 --- a/ext/CMakeLists.txt +++ /dev/null @@ -1,24 +0,0 @@ -# Olive - Non-Linear Video Editor -# Copyright (C) 2023 Olive Studios LLC -# -# 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 3 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, see . - -add_subdirectory(core EXCLUDE_FROM_ALL) - -set(KDDockWidgets_STATIC ON CACHE INTERNAL "Force KDDockWidgets to build statically") -set(KDDockWidgets_QT6 ${BUILD_QT6} CACHE INTERNAL "Conform KDDockWidgets' Qt 6 setting to ours") -# Oak only uses the QtWidgets frontend; building the QtQuick frontend causes -# duplicate QML module registration on macOS and pulls in unused dependencies. -set(KDDockWidgets_FRONTENDS "qtwidgets" CACHE INTERNAL "Only build the QtWidgets frontend for Oak") -add_subdirectory(KDDockWidgets EXCLUDE_FROM_ALL) diff --git a/ext/core b/ext/core deleted file mode 160000 index c4f8f7bde..000000000 --- a/ext/core +++ /dev/null @@ -1 +0,0 @@ -Subproject commit c4f8f7bde16a9e3807382de0c7f60c52d0e7799d diff --git a/ext/KDDockWidgets b/third_party/KDDockWidgets similarity index 100% rename from ext/KDDockWidgets rename to third_party/KDDockWidgets