Merge branch 'master' into draw-viewport-data

This is not needed anymore with the new dependency system.

CMakeLists.txt

@@ -273,11 +273,13 @@ endif()
 
 if(UNIX AND NOT APPLE)
   option(WITH_SYSTEM_GLEW "Use GLEW OpenGL wrapper library provided by the operating system" OFF)
-  option(WITH_SYSTEM_GLES "Use OpenGL ES library provided by the operating system"           ON)
+  option(WITH_SYSTEM_GLEW "Use GLEW OpenGL wrapper library provided by the operating system" OFF)
+  option(WITH_SYSTEM_FREETYPE "Use the freetype library provided by the operating system" OFF)
 else()
   # not an option for other OS's
   set(WITH_SYSTEM_GLEW OFF)
   set(WITH_SYSTEM_GLES OFF)
+  set(WITH_SYSTEM_FREETYPE OFF)
 endif()
 
 
@@ -536,12 +538,14 @@ option(WITH_OPENGL              "When off limits visibility of the opengl header
 option(WITH_GLEW_ES             "Switches to experimental copy of GLEW that has support for OpenGL ES. (temporary option for development purposes)" OFF)
 option(WITH_GL_EGL              "Use the EGL OpenGL system library instead of the platform specific OpenGL system library (CGL, glX, or WGL)"       OFF)
 option(WITH_GL_PROFILE_ES20     "Support using OpenGL ES 2.0. (through either EGL or the AGL/WGL/XGL 'es20' profile)"                               OFF)
+option(WITH_GPU_SHADER_BUILDER  "Shader builder is a developer option enabling linting on GLSL during compilation"                                  OFF)
 
 mark_as_advanced(
   WITH_OPENGL
   WITH_GLEW_ES
   WITH_GL_EGL
   WITH_GL_PROFILE_ES20
+  WITH_GPU_SHADER_BUILDER
 )
 
 if(WIN32)
@@ -559,12 +563,18 @@ if(WIN32)
   set(CPACK_INSTALL_PREFIX ${CMAKE_GENERIC_PROGRAM_FILES}/${})
 endif()
 
-# Compiler toolchain
+# Compiler tool-chain.
-if(CMAKE_COMPILER_IS_GNUCC)
+if(UNIX AND NOT APPLE)
+  if(CMAKE_COMPILER_IS_GNUCC)
     option(WITH_LINKER_GOLD "Use ld.gold linker which is usually faster than ld.bfd" ON)
     mark_as_advanced(WITH_LINKER_GOLD)
     option(WITH_LINKER_LLD "Use ld.lld linker which is usually faster than ld.gold" OFF)
     mark_as_advanced(WITH_LINKER_LLD)
+  endif()
+  if(CMAKE_COMPILER_IS_GNUCC OR CMAKE_C_COMPILER_ID MATCHES "Clang")
+    option(WITH_LINKER_MOLD "Use ld.mold linker which is usually faster than ld.gold & ld.lld." OFF)
+    mark_as_advanced(WITH_LINKER_MOLD)
+  endif()
 endif()
 
 option(WITH_COMPILER_ASAN "Build and link against address sanitizer (only for Debug & RelWithDebInfo targets)." OFF)
@@ -700,9 +710,12 @@ if(UNIX AND NOT APPLE)
 endif()
 
 # Installation process.
-option(POSTINSTALL_SCRIPT "Run given CMake script after installation process" OFF)
+set(POSTINSTALL_SCRIPT "" CACHE FILEPATH "Run given CMake script after installation process")
 mark_as_advanced(POSTINSTALL_SCRIPT)
 
+set(POSTCONFIGURE_SCRIPT "" CACHE FILEPATH "Run given CMake script as the last step of CMake configuration")
+mark_as_advanced(POSTCONFIGURE_SCRIPT)
+
 # end option(...)
 
 
@@ -2066,3 +2079,8 @@ endif()
 if(0)
   print_all_vars()
 endif()
+
+# Should be the last step of configuration.
+if(POSTCONFIGURE_SCRIPT)
+  include(${POSTCONFIGURE_SCRIPT})
+endif()

build_files/build_environment/CMakeLists.txt

@@ -63,6 +63,7 @@ include(cmake/jpeg.cmake)
 include(cmake/blosc.cmake)
 include(cmake/pthreads.cmake)
 include(cmake/openexr.cmake)
+include(cmake/brotli.cmake)
 include(cmake/freetype.cmake)
 include(cmake/freeglut.cmake)
 include(cmake/glew.cmake)

build_files/build_environment/cmake/boost.cmake

@@ -25,8 +25,13 @@ else()
 endif()
 
 if(WIN32)
+  if(MSVC_VERSION GREATER_EQUAL 1920) # 2019
+    set(BOOST_TOOLSET toolset=msvc-14.2)
+    set(BOOST_COMPILER_STRING -vc142)
+  else() # 2017
     set(BOOST_TOOLSET toolset=msvc-14.1)
     set(BOOST_COMPILER_STRING -vc141)
+  endif()
 
   set(BOOST_CONFIGURE_COMMAND bootstrap.bat)
   set(BOOST_BUILD_COMMAND b2)

build_files/build_environment/cmake/brotli.cmake

@@ -0,0 +1,38 @@
+# ***** BEGIN GPL LICENSE BLOCK *****
+#
+# This program is free software; you can redistribute it and/or
+# modify it under the terms of the GNU General Public License
+# as published by the Free Software Foundation; either version 2
+# of the License, or (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with this program; if not, write to the Free Software Foundation,
+# Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
+#
+# ***** END GPL LICENSE BLOCK *****
+
+set(BROTLI_EXTRA_ARGS
+)
+
+ExternalProject_Add(external_brotli
+  URL file://${PACKAGE_DIR}/${BROTLI_FILE}
+  DOWNLOAD_DIR ${DOWNLOAD_DIR}
+  URL_HASH ${BROTLI_HASH_TYPE}=${BROTLI_HASH}
+  PREFIX ${BUILD_DIR}/brotli
+  CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${LIBDIR}/brotli ${DEFAULT_CMAKE_FLAGS} ${BROTLI_EXTRA_ARGS}
+  INSTALL_DIR ${LIBDIR}/brotli
+)
+
+if(BUILD_MODE STREQUAL Release AND WIN32)
+  ExternalProject_Add_Step(external_brotli after_install
+    COMMAND ${CMAKE_COMMAND} -E copy_directory ${LIBDIR}/brotli/include ${HARVEST_TARGET}/brotli/include
+    COMMAND ${CMAKE_COMMAND} -E copy ${LIBDIR}/brotli/lib/brotlidec-static${LIBEXT} ${HARVEST_TARGET}/brotli/lib/brotlidec-static${LIBEXT}
+    COMMAND ${CMAKE_COMMAND} -E copy ${LIBDIR}/brotli/lib/brotlicommon-static${LIBEXT} ${HARVEST_TARGET}/brotli/lib/brotlicommon-static${LIBEXT}
+    DEPENDEES install
+  )
+endif()

build_files/build_environment/cmake/download.cmake

@@ -94,3 +94,4 @@ download_source(POTRACE)
 download_source(HARU)
 download_source(ZSTD)
 download_source(FLEX)
+download_source(BROTLI)

build_files/build_environment/cmake/freetype.cmake

@@ -23,9 +23,12 @@ set(FREETYPE_EXTRA_ARGS
   -DWITH_HarfBuzz=OFF
   -DFT_WITH_HARFBUZZ=OFF
   -DFT_WITH_BZIP2=OFF
+  -DFT_WITH_BROTLI=ON
   -DCMAKE_DISABLE_FIND_PACKAGE_HarfBuzz=TRUE
   -DCMAKE_DISABLE_FIND_PACKAGE_BZip2=TRUE
-  -DCMAKE_DISABLE_FIND_PACKAGE_BrotliDec=TRUE)
+  -DPC_BROTLIDEC_INCLUDEDIR=${LIBDIR}/brotli/include
+  -DPC_BROTLIDEC_LIBDIR=${LIBDIR}/brotli/lib
+  )
 
 ExternalProject_Add(external_freetype
   URL file://${PACKAGE_DIR}/${FREETYPE_FILE}
@@ -36,6 +39,11 @@ ExternalProject_Add(external_freetype
   INSTALL_DIR ${LIBDIR}/freetype
 )
 
+add_dependencies(
+  external_freetype
+  external_brotli
+)
+
 if(BUILD_MODE STREQUAL Release AND WIN32)
   ExternalProject_Add_Step(external_freetype after_install
     COMMAND ${CMAKE_COMMAND} -E copy_directory ${LIBDIR}/freetype ${HARVEST_TARGET}/freetype

build_files/build_environment/cmake/gmp.cmake

@@ -38,13 +38,6 @@ elseif(UNIX AND NOT APPLE)
   )
 endif()
 
-if(BLENDER_PLATFORM_ARM)
-  set(GMP_OPTIONS
-    ${GMP_OPTIONS}
-    --disable-assembly
-  )
-endif()
-
 ExternalProject_Add(external_gmp
   URL file://${PACKAGE_DIR}/${GMP_FILE}
   DOWNLOAD_DIR ${DOWNLOAD_DIR}

build_files/build_environment/cmake/harvest.cmake

@@ -79,6 +79,8 @@ endfunction()
 harvest(alembic/include alembic/include "*.h")
 harvest(alembic/lib/libAlembic.a alembic/lib/libAlembic.a)
 harvest(alembic/bin alembic/bin "*")
+harvest(brotli/include brotli/include "*.h")
+harvest(brotli/lib brotli/lib "*.a")
 harvest(boost/include boost/include "*")
 harvest(boost/lib boost/lib "*.a")
 harvest(ffmpeg/include ffmpeg/include "*.h")

build_files/build_environment/cmake/versions.cmake

@@ -83,9 +83,9 @@ else()
   set(OPENEXR_VERSION_POSTFIX)
 endif()
 
-set(FREETYPE_VERSION 2.10.2)
+set(FREETYPE_VERSION 2.11.0)
 set(FREETYPE_URI http://prdownloads.sourceforge.net/freetype/freetype-${FREETYPE_VERSION}.tar.gz)
-set(FREETYPE_HASH b1cb620e4c875cd4d1bfa04945400945)
+set(FREETYPE_HASH cf09172322f6b50cf8f568bf8fe14bde)
 set(FREETYPE_HASH_TYPE MD5)
 set(FREETYPE_FILE freetype-${FREETYPE_VERSION}.tar.gz)
 
@@ -474,9 +474,9 @@ set(ISPC_HASH 2e3abedbc0ea9aaec17d6562c632454d)
 set(ISPC_HASH_TYPE MD5)
 set(ISPC_FILE ispc-${ISPC_VERSION}.tar.gz)
 
-set(GMP_VERSION 6.2.0)
+set(GMP_VERSION 6.2.1)
 set(GMP_URI https://gmplib.org/download/gmp/gmp-${GMP_VERSION}.tar.xz)
-set(GMP_HASH a325e3f09e6d91e62101e59f9bda3ec1)
+set(GMP_HASH 0b82665c4a92fd2ade7440c13fcaa42b)
 set(GMP_HASH_TYPE MD5)
 set(GMP_FILE gmp-${GMP_VERSION}.tar.xz)
 
@@ -500,3 +500,9 @@ set(ZSTD_FILE zstd-${ZSTD_VERSION}.tar.gz)
 
 set(SSE2NEON_GIT https://github.com/DLTcollab/sse2neon.git)
 set(SSE2NEON_GIT_HASH fe5ff00bb8d19b327714a3c290f3e2ce81ba3525)
+
+set(BROTLI_VERSION v1.0.9)
+set(BROTLI_URI https://github.com/google/brotli/archive/refs/tags/${BROTLI_VERSION}.tar.gz)
+set(BROTLI_HASH f9e8d81d0405ba66d181529af42a3354f838c939095ff99930da6aa9cdf6fe46)
+set(BROTLI_HASH_TYPE SHA256)
+set(BROTLI_FILE brotli-${BROTLI_VERSION}.tar.gz)

build_files/build_environment/install_deps.sh

@@ -492,7 +492,7 @@ OIIO_SKIP=false
 LLVM_VERSION="12.0.0"
 LLVM_VERSION_SHORT="12.0"
 LLVM_VERSION_MIN="11.0"
-LLVM_VERSION_MEX="13.0"
+LLVM_VERSION_MEX="14.0"
 LLVM_VERSION_FOUND=""
 LLVM_FORCE_BUILD=false
 LLVM_FORCE_REBUILD=false
@@ -3620,8 +3620,8 @@ compile_FFmpeg() {
     fi
 
     ./configure --cc="gcc -Wl,--as-needed" \
-        --extra-ldflags="-pthread -static-libgcc" \
+        --extra-ldflags="-pthread" \
-        --prefix=$_inst --enable-static \
+        --prefix=$_inst --enable-shared \
         --disable-ffplay --disable-doc \
         --enable-gray \
         --enable-avfilter --disable-vdpau \
@@ -5721,76 +5721,6 @@ install_OTHER() {
 # ----------------------------------------------------------------------------
 # Printing User Info
 
-print_info_ffmpeglink_DEB() {
-  dpkg -L $_packages | grep -e ".*\/lib[^\/]\+\.so" | gawk '{ printf(nlines ? "'"$_ffmpeg_list_sep"'%s" : "%s", gensub(/.*lib([^\/]+)\.so/, "\\1", "g", $0)); nlines++ }'
-}
-
-print_info_ffmpeglink_RPM() {
-  rpm -ql $_packages | grep -e ".*\/lib[^\/]\+\.so" | gawk '{ printf(nlines ? "'"$_ffmpeg_list_sep"'%s" : "%s", gensub(/.*lib([^\/]+)\.so/, "\\1", "g", $0)); nlines++ }'
-}
-
-print_info_ffmpeglink_ARCH() {
-  pacman -Ql $_packages | grep -e ".*\/lib[^\/]\+\.so$" | gawk '{ printf(nlines ? "'"$_ffmpeg_list_sep"'%s" : "%s", gensub(/.*lib([^\/]+)\.so/, "\\1", "g", $0)); nlines++ }'
-}
-
-print_info_ffmpeglink() {
-  # This func must only print a ';'-separated list of libs...
-  if [ -z "$DISTRO" ]; then
-    ERROR "Failed to detect distribution type"
-    exit 1
-  fi
-
-  # Create list of packages from which to get libs names...
-  _packages=""
-
-  if [ "$THEORA_USE" = true ]; then
-    _packages="$_packages $THEORA_DEV"
-  fi
-
-  if [ "$VORBIS_USE" = true ]; then
-    _packages="$_packages $VORBIS_DEV"
-  fi
-
-  if [ "$OGG_USE" = true ]; then
-    _packages="$_packages $OGG_DEV"
-  fi
-
-  if [ "$XVID_USE" = true ]; then
-    _packages="$_packages $XVID_DEV"
-  fi
-
-  if [ "$VPX_USE" = true ]; then
-    _packages="$_packages $VPX_DEV"
-  fi
-
-  if [ "$OPUS_USE" = true ]; then
-    _packages="$_packages $OPUS_DEV"
-  fi
-
-  if [ "$MP3LAME_USE" = true ]; then
-    _packages="$_packages $MP3LAME_DEV"
-  fi
-
-  if [ "$X264_USE" = true ]; then
-    _packages="$_packages $X264_DEV"
-  fi
-
-  if [ "$OPENJPEG_USE" = true ]; then
-    _packages="$_packages $OPENJPEG_DEV"
-  fi
-
-  if [ "$DISTRO" = "DEB" ]; then
-    print_info_ffmpeglink_DEB
-  elif [ "$DISTRO" = "RPM" ]; then
-    print_info_ffmpeglink_RPM
-  elif [ "$DISTRO" = "ARCH" ]; then
-    print_info_ffmpeglink_ARCH
-  # XXX TODO!
-  else
-    PRINT "<Could not determine additional link libraries needed for ffmpeg, replace this by valid list of libs...>"
-  fi
-}
-
 print_info() {
   PRINT ""
   PRINT ""
@@ -6002,12 +5932,10 @@ print_info() {
 
   if [ "$FFMPEG_SKIP" = false ]; then
     _1="-D WITH_CODEC_FFMPEG=ON"
-    _2="-D FFMPEG_LIBRARIES='avformat;avcodec;avutil;avdevice;swscale;swresample;lzma;rt;`print_info_ffmpeglink`'"
     PRINT "  $_1"
-    PRINT "  $_2"
+    _buildargs="$_buildargs $_1"
-    _buildargs="$_buildargs $_1 $_2"
     if [ -d $INST/ffmpeg ]; then
-      _1="-D FFMPEG=$INST/ffmpeg"
+      _1="-D FFMPEG_ROOT_DIR=$INST/ffmpeg"
       PRINT "  $_1"
       _buildargs="$_buildargs $_1"
     fi

build_files/build_environment/windows/build_deps.cmd

@@ -1,64 +1,39 @@
 @echo off
 if NOT "%1" == "" (
-	if "%1" == "2013" (
-    echo "Building for VS2013"
-    set VSVER=12.0
-    set VSVER_SHORT=12
-    set BuildDir=VS12
-    goto par2
-  )
-	if "%1" == "2015" (
-    echo "Building for VS2015"
-    set VSVER=14.0
-    set VSVER_SHORT=14
-    set BuildDir=VS14
-    goto par2
-  )
 	if "%1" == "2017" (
     echo "Building for VS2017"
     set VSVER=15.0
     set VSVER_SHORT=15
     set BuildDir=VS15
     goto par2
+  )
+	if "%1" == "2019" (
+    echo "Building for VS2019"
+    set VSVER=15.0
+    set VSVER_SHORT=15
+    set BuildDir=VS15
+    goto par2
   )
   
 )
 :usage
 
-Echo Usage build_deps 2013/2015/2017 x64/x86
+Echo Usage build_deps 2017/2019 x64
 goto exit
 :par2
 if NOT "%2" == "" (
-	if "%2" == "x86" (
-    echo "Building for x86"
-    set HARVESTROOT=Windows_vc
-    set ARCH=86
-		if "%1" == "2013" (
-			set CMAKE_BUILDER=Visual Studio 12 2013
-		)
-		if "%1" == "2015" (
-			set CMAKE_BUILDER=Visual Studio 14 2015
-		)
-		if "%1" == "2017" (
-			set CMAKE_BUILDER=Visual Studio 15 2017
-		)
-		
-    goto start
-  )
 	if "%2" == "x64" (
     echo "Building for x64"
     set HARVESTROOT=Win64_vc
     set ARCH=64
-		if "%1" == "2013" (
+		if "%1" == "2019" (
-			set CMAKE_BUILDER=Visual Studio 12 2013 Win64
+			set CMAKE_BUILDER=Visual Studio 16 2019
-		)
+			set CMAKE_BUILD_ARCH=-A x64
-		if "%1" == "2015" (
-			set CMAKE_BUILDER=Visual Studio 14 2015 Win64
 		)
 		if "%1" == "2017" (
 			set CMAKE_BUILDER=Visual Studio 15 2017 Win64
+            set CMAKE_BUILD_ARCH=
 		)
-		
     goto start
   )
 )
@@ -120,7 +95,7 @@ set path=%BUILD_DIR%\downloads\mingw\mingw64\msys\1.0\bin\;%BUILD_DIR%\downloads
 mkdir %STAGING%\%BuildDir%%ARCH%R
 cd %Staging%\%BuildDir%%ARCH%R
 echo %DATE% %TIME% : Start > %StatusFile%
-cmake -G "%CMAKE_BUILDER%" -Thost=x64  %SOURCE_DIR% -DPACKAGE_DIR=%BUILD_DIR%/packages -DDOWNLOAD_DIR=%BUILD_DIR%/downloads -DBUILD_MODE=Release -DHARVEST_TARGET=%HARVEST_DIR%/%HARVESTROOT%%VSVER_SHORT%/
+cmake -G "%CMAKE_BUILDER%" %CMAKE_BUILD_ARCH% -Thost=x64  %SOURCE_DIR% -DPACKAGE_DIR=%BUILD_DIR%/packages -DDOWNLOAD_DIR=%BUILD_DIR%/downloads -DBUILD_MODE=Release -DHARVEST_TARGET=%HARVEST_DIR%/%HARVESTROOT%%VSVER_SHORT%/
 echo %DATE% %TIME% : Release Configuration done >> %StatusFile%
 if "%dobuild%" == "1" (
 	msbuild /m "ll.vcxproj" /p:Configuration=Release /fl /flp:logfile=BlenderDeps_llvm.log;Verbosity=normal
@@ -133,7 +108,7 @@ if "%NODEBUG%" == "1" goto exit
 cd %BUILD_DIR%
 mkdir %STAGING%\%BuildDir%%ARCH%D
 cd %Staging%\%BuildDir%%ARCH%D
-cmake -G "%CMAKE_BUILDER%" -Thost=x64 %SOURCE_DIR% -DPACKAGE_DIR=%BUILD_DIR%/packages -DDOWNLOAD_DIR=%BUILD_DIR%/downloads -DCMAKE_BUILD_TYPE=Debug -DBUILD_MODE=Debug -DHARVEST_TARGET=%HARVEST_DIR%/%HARVESTROOT%%VSVER_SHORT%/  %CMAKE_DEBUG_OPTIONS%
+cmake -G "%CMAKE_BUILDER%" %CMAKE_BUILD_ARCH% -Thost=x64 %SOURCE_DIR% -DPACKAGE_DIR=%BUILD_DIR%/packages -DDOWNLOAD_DIR=%BUILD_DIR%/downloads -DCMAKE_BUILD_TYPE=Debug -DBUILD_MODE=Debug -DHARVEST_TARGET=%HARVEST_DIR%/%HARVESTROOT%%VSVER_SHORT%/  %CMAKE_DEBUG_OPTIONS%
 echo %DATE% %TIME% : Debug Configuration done >> %StatusFile%
 if "%dobuild%" == "1" (
 	msbuild /m "ll.vcxproj" /p:Configuration=Debug /fl /flp:logfile=BlenderDeps_llvm.log;;Verbosity=normal 

build_files/cmake/Modules/FindBrotli.cmake

@@ -0,0 +1,83 @@
+# - Find Brotli library (compression for freetype/woff2).
+# This module defines
+#  BROTLI_INCLUDE_DIRS, where to find Brotli headers, Set when
+#                       BROTLI_INCLUDE_DIR is found.
+#  BROTLI_LIBRARIES, libraries to link against to use Brotli.
+#  BROTLI_ROOT_DIR, The base directory to search for Brotli.
+#                   This can also be an environment variable.
+#  BROTLI_FOUND, If false, do not try to use Brotli.
+#
+
+#=============================================================================
+# Copyright 2022 Blender Foundation.
+#
+# Distributed under the OSI-approved BSD 3-Clause License,
+# see accompanying file BSD-3-Clause-license.txt for details.
+#=============================================================================
+
+# If BROTLI_ROOT_DIR was defined in the environment, use it.
+IF(NOT BROTLI_ROOT_DIR AND NOT $ENV{BROTLI_ROOT_DIR} STREQUAL "")
+  SET(BROTLI_ROOT_DIR $ENV{BROTLI_ROOT_DIR})
+ENDIF()
+
+SET(_BROTLI_SEARCH_DIRS
+  ${BROTLI_ROOT_DIR}
+)
+
+FIND_PATH(BROTLI_INCLUDE_DIR
+  NAMES
+    brotli/decode.h
+  HINTS
+    ${_BROTLI_SEARCH_DIRS}
+  PATH_SUFFIXES
+    include
+  DOC "Brotli header files"
+)
+
+FIND_LIBRARY(BROTLI_LIBRARY_COMMON
+  NAMES
+    # Some builds use a special `-static` postfix in their static libraries names.
+    brotlicommon-static
+    brotlicommon
+  HINTS
+    ${_BROTLI_SEARCH_DIRS}
+  PATH_SUFFIXES
+    lib64 lib lib/static
+  DOC "Brotli static common library"
+)
+FIND_LIBRARY(BROTLI_LIBRARY_DEC
+  NAMES
+    # Some builds use a special `-static` postfix in their static libraries names.
+    brotlidec-static
+    brotlidec
+  HINTS
+    ${_BROTLI_SEARCH_DIRS}
+  PATH_SUFFIXES
+    lib64 lib lib/static
+  DOC "Brotli static decode library"
+)
+
+
+IF(${BROTLI_LIBRARY_COMMON_NOTFOUND} or ${BROTLI_LIBRARY_DEC_NOTFOUND})
+  set(BROTLI_FOUND FALSE)
+ELSE()
+  # handle the QUIETLY and REQUIRED arguments and set BROTLI_FOUND to TRUE if
+  # all listed variables are TRUE
+  INCLUDE(FindPackageHandleStandardArgs)
+  FIND_PACKAGE_HANDLE_STANDARD_ARGS(Brotli DEFAULT_MSG BROTLI_LIBRARY_COMMON BROTLI_LIBRARY_DEC BROTLI_INCLUDE_DIR)
+
+  IF(BROTLI_FOUND)
+    get_filename_component(BROTLI_LIBRARY_DIR ${BROTLI_LIBRARY_COMMON} DIRECTORY)
+    SET(BROTLI_INCLUDE_DIRS ${BROTLI_INCLUDE_DIR})
+    SET(BROTLI_LIBRARIES ${BROTLI_LIBRARY_DEC} ${BROTLI_LIBRARY_COMMON})
+  ENDIF()
+ENDIF()
+
+MARK_AS_ADVANCED(
+  BROTLI_INCLUDE_DIR
+  BROTLI_LIBRARY_COMMON
+  BROTLI_LIBRARY_DEC
+  BROTLI_LIBRARY_DIR
+)
+
+UNSET(_BROTLI_SEARCH_DIRS)

build_files/cmake/Modules/FindFFmpeg.cmake

@@ -33,6 +33,8 @@ if(NOT FFMPEG_FIND_COMPONENTS)
     avfilter
     avformat
     avutil
+    swscale
+    swresample
   )
 endif()
 
@@ -50,9 +52,9 @@ foreach(_component ${FFMPEG_FIND_COMPONENTS})
   string(TOUPPER ${_component} _upper_COMPONENT)
   find_library(FFMPEG_${_upper_COMPONENT}_LIBRARY
     NAMES
-      ${_upper_COMPONENT}
+      ${_component}
     HINTS
-      ${LIBDIR}/ffmpeg
+      ${_ffmpeg_SEARCH_DIRS}
     PATH_SUFFIXES
       lib64 lib
   )

build_files/cmake/Modules/FindOptiX.cmake

@@ -21,7 +21,7 @@ ENDIF()
 
 SET(_optix_SEARCH_DIRS
   ${OPTIX_ROOT_DIR}
-  "$ENV{PROGRAMDATA}/NVIDIA Corporation/OptiX SDK 7.0.0"
+  "$ENV{PROGRAMDATA}/NVIDIA Corporation/OptiX SDK 7.3.0"
 )
 
 FIND_PATH(OPTIX_INCLUDE_DIR

build_files/cmake/macros.cmake

@@ -488,7 +488,6 @@ function(blender_add_test_executable
 
   include_directories(${includes})
   include_directories(${includes_sys})
-  setup_libdirs()
 
   BLENDER_SRC_GTEST_EX(
     NAME ${name}
@@ -525,83 +524,6 @@ function(setup_heavy_lib_pool)
   endif()
 endfunction()
 
-function(SETUP_LIBDIRS)
-
-  # NOTE: For all new libraries, use absolute library paths.
-  # This should eventually be phased out.
-  # APPLE platform uses full paths for linking libraries, and avoids link_directories.
-  if(NOT MSVC AND NOT APPLE)
-    link_directories(${JPEG_LIBPATH} ${PNG_LIBPATH} ${ZLIB_LIBPATH} ${FREETYPE_LIBPATH})
-
-    if(WITH_PYTHON)  #  AND NOT WITH_PYTHON_MODULE  # WIN32 needs
-      link_directories(${PYTHON_LIBPATH})
-    endif()
-    if(WITH_SDL AND NOT WITH_SDL_DYNLOAD)
-      link_directories(${SDL_LIBPATH})
-    endif()
-    if(WITH_CODEC_FFMPEG)
-      link_directories(${FFMPEG_LIBPATH})
-    endif()
-    if(WITH_IMAGE_OPENEXR)
-      link_directories(${OPENEXR_LIBPATH})
-    endif()
-    if(WITH_IMAGE_TIFF)
-      link_directories(${TIFF_LIBPATH})
-    endif()
-    if(WITH_BOOST)
-      link_directories(${BOOST_LIBPATH})
-    endif()
-    if(WITH_OPENIMAGEIO)
-      link_directories(${OPENIMAGEIO_LIBPATH})
-    endif()
-    if(WITH_OPENIMAGEDENOISE)
-      link_directories(${OPENIMAGEDENOISE_LIBPATH})
-    endif()
-    if(WITH_OPENCOLORIO)
-      link_directories(${OPENCOLORIO_LIBPATH})
-    endif()
-    if(WITH_OPENVDB)
-      link_directories(${OPENVDB_LIBPATH})
-    endif()
-    if(WITH_OPENAL)
-      link_directories(${OPENAL_LIBPATH})
-    endif()
-    if(WITH_JACK AND NOT WITH_JACK_DYNLOAD)
-      link_directories(${JACK_LIBPATH})
-    endif()
-    if(WITH_PULSEAUDIO AND NOT WITH_PULSEAUDIO_DYNLOAD)
-      link_directories(${LIBPULSE_LIBPATH})
-    endif()
-    if(WITH_CODEC_SNDFILE)
-      link_directories(${LIBSNDFILE_LIBPATH})
-    endif()
-    if(WITH_FFTW3)
-      link_directories(${FFTW3_LIBPATH})
-    endif()
-    if(WITH_OPENCOLLADA)
-      link_directories(${OPENCOLLADA_LIBPATH})
-      # # Never set
-      # link_directories(${PCRE_LIBPATH})
-      # link_directories(${EXPAT_LIBPATH})
-    endif()
-    if(WITH_LLVM)
-      link_directories(${LLVM_LIBPATH})
-    endif()
-
-    if(WITH_ALEMBIC)
-      link_directories(${ALEMBIC_LIBPATH})
-    endif()
-
-    if(WITH_GMP)
-      link_directories(${GMP_LIBPATH})
-    endif()
-
-    if(WIN32 AND NOT UNIX)
-      link_directories(${PTHREADS_LIBPATH})
-    endif()
-  endif()
-endfunction()
-
 # Platform specific linker flags for targets.
 function(setup_platform_linker_flags
   target)
@@ -1292,29 +1214,6 @@ macro(openmp_delayload
     endif()
 endmacro()
 
-macro(blender_precompile_headers target cpp header)
-  if(MSVC)
-    # get the name for the pch output file
-    get_filename_component(pchbase ${cpp} NAME_WE)
-    set(pchfinal "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_CFG_INTDIR}/${pchbase}.pch")
-
-    # mark the cpp as the one outputting the pch
-    set_property(SOURCE ${cpp} APPEND PROPERTY OBJECT_OUTPUTS "${pchfinal}")
-
-    # get all sources for the target
-    get_target_property(sources ${target} SOURCES)
-
-    # make all sources depend on the pch to enforce the build order
-    foreach(src ${sources})
-      set_property(SOURCE ${src} APPEND PROPERTY OBJECT_DEPENDS "${pchfinal}")
-    endforeach()
-
-    target_sources(${target} PRIVATE ${cpp} ${header})
-    set_target_properties(${target} PROPERTIES COMPILE_FLAGS "/Yu${header} /Fp${pchfinal} /FI${header}")
-    set_source_files_properties(${cpp} PROPERTIES COMPILE_FLAGS "/Yc${header} /Fp${pchfinal}")
-  endif()
-endmacro()
-
 macro(set_and_warn_dependency
   _dependency _setting _val)
   # when $_dependency is disabled, forces $_setting = $_val

build_files/cmake/platform/platform_apple.cmake

@@ -166,13 +166,18 @@ if(WITH_FFTW3)
   find_package(Fftw3)
 endif()
 
+# FreeType compiled with Brotli compression for woff2.
 find_package(Freetype REQUIRED)
+list(APPEND FREETYPE_LIBRARIES
+  ${LIBDIR}/brotli/lib/libbrotlicommon-static.a
+  ${LIBDIR}/brotli/lib/libbrotlidec-static.a)
 
 if(WITH_IMAGE_OPENEXR)
   find_package(OpenEXR)
 endif()
 
 if(WITH_CODEC_FFMPEG)
+  set(FFMPEG_ROOT_DIR ${LIBDIR}/ffmpeg)
   set(FFMPEG_FIND_COMPONENTS
     avcodec avdevice avformat avutil
     mp3lame ogg opus swresample swscale
@@ -479,8 +484,11 @@ string(APPEND PLATFORM_LINKFLAGS " -stdlib=libc++")
 # Suppress ranlib "has no symbols" warnings (workaround for T48250)
 set(CMAKE_C_ARCHIVE_CREATE   "<CMAKE_AR> Scr <TARGET> <LINK_FLAGS> <OBJECTS>")
 set(CMAKE_CXX_ARCHIVE_CREATE "<CMAKE_AR> Scr <TARGET> <LINK_FLAGS> <OBJECTS>")
-set(CMAKE_C_ARCHIVE_FINISH   "<CMAKE_RANLIB> -no_warning_for_no_symbols -c <TARGET>")
+# llvm-ranlib doesn't support this flag. Xcode's libtool does.
-set(CMAKE_CXX_ARCHIVE_FINISH "<CMAKE_RANLIB> -no_warning_for_no_symbols -c <TARGET>")
+if(NOT ${CMAKE_RANLIB} MATCHES ".*llvm-ranlib$")
+  set(CMAKE_C_ARCHIVE_FINISH   "<CMAKE_RANLIB> -no_warning_for_no_symbols -c <TARGET>")
+  set(CMAKE_CXX_ARCHIVE_FINISH "<CMAKE_RANLIB> -no_warning_for_no_symbols -c <TARGET>")
+endif()
 
 if(WITH_COMPILER_CCACHE)
   if(NOT CMAKE_GENERATOR STREQUAL "Xcode")
@@ -507,3 +515,6 @@ list(APPEND CMAKE_BUILD_RPATH "${OpenMP_LIBRARY_DIR}")
 
 set(CMAKE_SKIP_INSTALL_RPATH FALSE)
 list(APPEND CMAKE_INSTALL_RPATH "@loader_path/../Resources/${BLENDER_VERSION}/lib")
+
+# Same as `CFBundleIdentifier` in Info.plist.
+set(CMAKE_XCODE_ATTRIBUTE_PRODUCT_BUNDLE_IDENTIFIER "org.blenderfoundation.blender")

build_files/cmake/platform/platform_apple_xcode.cmake

@@ -96,7 +96,7 @@ else()
     # Detect SDK version to use.
     if(NOT DEFINED OSX_SYSTEM)
       execute_process(
-          COMMAND xcrun --show-sdk-version
+          COMMAND xcrun --sdk macosx --show-sdk-version
           OUTPUT_VARIABLE OSX_SYSTEM
           OUTPUT_STRIP_TRAILING_WHITESPACE)
     endif()

build_files/cmake/platform/platform_unix.cmake

@@ -18,7 +18,7 @@
 # All rights reserved.
 # ***** END GPL LICENSE BLOCK *****
 
-# Libraries configuration for any *nix system including Linux and Unix.
+# Libraries configuration for any *nix system including Linux and Unix (excluding APPLE).
 
 # Detect precompiled library directory
 if(NOT DEFINED LIBDIR)
@@ -48,6 +48,9 @@ if(NOT DEFINED LIBDIR)
   unset(LIBDIR_CENTOS7_ABI)
 endif()
 
+# Support restoring this value once pre-compiled libraries have been handled.
+set(WITH_STATIC_LIBS_INIT ${WITH_STATIC_LIBS})
+
 if(EXISTS ${LIBDIR})
   message(STATUS "Using pre-compiled LIBDIR: ${LIBDIR}")
 
@@ -100,7 +103,22 @@ find_package_wrapper(JPEG REQUIRED)
 find_package_wrapper(PNG REQUIRED)
 find_package_wrapper(ZLIB REQUIRED)
 find_package_wrapper(Zstd REQUIRED)
-find_package_wrapper(Freetype REQUIRED)
+
+if(NOT WITH_SYSTEM_FREETYPE)
+  # FreeType compiled with Brotli compression for woff2.
+  find_package_wrapper(Freetype REQUIRED)
+  if(EXISTS ${LIBDIR})
+    find_package_wrapper(Brotli REQUIRED)
+
+    # NOTE: This is done on WIN32 & APPLE but fails on some Linux systems.
+    # See: https://devtalk.blender.org/t/22536
+    # So `BROTLI_LIBRARIES` need to be added  `FREETYPE_LIBRARIES`.
+    #
+    # list(APPEND FREETYPE_LIBRARIES
+    #   ${BROTLI_LIBRARIES}
+    # )
+  endif()
+endif()
 
 if(WITH_PYTHON)
   # No way to set py35, remove for now.
@@ -178,26 +196,30 @@ endif()
 
 if(WITH_CODEC_FFMPEG)
   if(EXISTS ${LIBDIR})
-    # For precompiled lib directory, all ffmpeg dependencies are in the same folder
+    set(FFMPEG_ROOT_DIR ${LIBDIR}/ffmpeg)
-    file(GLOB ffmpeg_libs ${LIBDIR}/ffmpeg/lib/*.a ${LIBDIR}/sndfile/lib/*.a)
+    # Override FFMPEG components to also include static library dependencies
-    set(FFMPEG ${LIBDIR}/ffmpeg CACHE PATH "FFMPEG Directory")
+    # included with precompiled libraries, and to ensure correct link order.
-    set(FFMPEG_LIBRARIES ${ffmpeg_libs} ${ffmpeg_libs} CACHE STRING "FFMPEG Libraries")
+    set(FFMPEG_FIND_COMPONENTS
-  else()
+      avformat avcodec avdevice avutil swresample swscale
-    set(FFMPEG /usr CACHE PATH "FFMPEG Directory")
+      sndfile
-    set(FFMPEG_LIBRARIES avformat avcodec avutil avdevice swscale CACHE STRING "FFMPEG Libraries")
+      FLAC
+      mp3lame
+      opus
+      theora theoradec theoraenc
+      vorbis vorbisenc vorbisfile ogg
+      vpx
+      x264
+      xvidcore)
+  elseif(FFMPEG)
+    # Old cache variable used for root dir, convert to new standard.
+    set(FFMPEG_ROOT_DIR ${FFMPEG})
   endif()
+  find_package(FFmpeg)
 
-  mark_as_advanced(FFMPEG)
+  if(NOT FFMPEG_FOUND)
-
+    set(WITH_CODEC_FFMPEG OFF)
-  # lame, but until we have proper find module for ffmpeg
+    message(STATUS "FFmpeg not found, disabling it")
-  set(FFMPEG_INCLUDE_DIRS ${FFMPEG}/include)
-  if(EXISTS "${FFMPEG}/include/ffmpeg/")
-    list(APPEND FFMPEG_INCLUDE_DIRS "${FFMPEG}/include/ffmpeg")
   endif()
-  # end lameness
-
-  mark_as_advanced(FFMPEG_LIBRARIES)
-  set(FFMPEG_LIBPATH ${FFMPEG}/lib)
 endif()
 
 if(WITH_FFTW3)
@@ -532,6 +554,21 @@ add_definitions(-D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64 -D_LARGEFILE64_SOURCE
 #
 # Keep last, so indirectly linked libraries don't override our own pre-compiled libs.
 
+if(EXISTS ${LIBDIR})
+  # Clear the prefix path as it causes the `LIBDIR` to override system locations.
+  unset(CMAKE_PREFIX_PATH)
+
+  # Since the pre-compiled `LIBDIR` directories have been handled, don't prefer static libraries.
+  set(WITH_STATIC_LIBS ${WITH_STATIC_LIBS_INIT})
+endif()
+
+if(WITH_SYSTEM_FREETYPE)
+  find_package_wrapper(Freetype)
+  if(NOT FREETYPE_FOUND)
+    message(FATAL_ERROR "Failed finding system FreeType version!")
+  endif()
+endif()
+
 if(WITH_LZO AND WITH_SYSTEM_LZO)
   find_package_wrapper(LZO)
   if(NOT LZO_FOUND)
@@ -644,6 +681,9 @@ endif()
 # ----------------------------------------------------------------------------
 # Compilers
 
+# Only set the linker once.
+set(_IS_LINKER_DEFAULT ON)
+
 # GNU Compiler
 if(CMAKE_COMPILER_IS_GNUCC)
   # ffp-contract=off:
@@ -662,26 +702,89 @@ if(CMAKE_COMPILER_IS_GNUCC)
   string(PREPEND CMAKE_CXX_FLAGS_RELWITHDEBINFO "${GCC_EXTRA_FLAGS_RELEASE} ")
   unset(GCC_EXTRA_FLAGS_RELEASE)
 
-  if(WITH_LINKER_GOLD)
+  # NOTE(@campbellbarton): Eventually mold will be able to use `-fuse-ld=mold`,
+  # however at the moment this only works for GCC 12.1+ (unreleased at time of writing).
+  # So a workaround is used here "-B" which points to another path to find system commands
+  # such as `ld`.
+  if(WITH_LINKER_MOLD AND _IS_LINKER_DEFAULT)
+    find_program(MOLD_BIN "mold")
+    mark_as_advanced(MOLD_BIN)
+    if(NOT MOLD_BIN)
+      message(STATUS "The \"mold\" binary could not be found, using system linker.")
+      set(WITH_LINKER_MOLD OFF)
+    else()
+      # By default mold installs the binary to:
+      # - `{PREFIX}/bin/mold` as well as a symbolic-link in...
+      # - `{PREFIX}/lib/mold/ld`.
+      # (where `PREFIX` is typically `/usr/`).
+      #
+      # This block of code finds `{PREFIX}/lib/mold` from the `mold` binary.
+      # Other methods of searching for the path could also be made to work,
+      # we could even make our own directory and symbolic-link, however it's more
+      # convenient to use the one provided by mold.
+      #
+      # Use the binary path to "mold", to find the common prefix which contains "lib/mold".
+      # The parent directory: e.g. `/usr/bin/mold` -> `/usr/bin/`.
+      get_filename_component(MOLD_PREFIX "${MOLD_BIN}" DIRECTORY)
+      # The common prefix path: e.g. `/usr/bin/` -> `/usr/` to use as a hint.
+      get_filename_component(MOLD_PREFIX "${MOLD_PREFIX}" DIRECTORY)
+      # Find `{PREFIX}/lib/mold/ld`, store the directory component (without the `ld`).
+      # Then pass `-B {PREFIX}/lib/mold` to GCC so the `ld` located there overrides the default.
+      find_path(
+        MOLD_BIN_DIR "ld"
+        HINTS "${MOLD_PREFIX}"
+        # The default path is `libexec`, Arch Linux for e.g.
+        # replaces this with `lib` so check both.
+        PATH_SUFFIXES "libexec/mold" "lib/mold" "lib64/mold"
+        NO_DEFAULT_PATH
+        NO_CACHE
+      )
+      if(NOT MOLD_BIN_DIR)
+        message(STATUS
+          "The mold linker could not find the directory containing the linker command "
+          "(typically "
+          "\"${MOLD_PREFIX}/libexec/mold/ld\") or "
+          "\"${MOLD_PREFIX}/lib/mold/ld\") using system linker.")
+        set(WITH_LINKER_MOLD OFF)
+      endif()
+      unset(MOLD_PREFIX)
+    endif()
+
+    if(WITH_LINKER_MOLD)
+      # GCC will search for `ld` in this directory first.
+      string(APPEND CMAKE_EXE_LINKER_FLAGS    " -B \"${MOLD_BIN_DIR}\"")
+      string(APPEND CMAKE_SHARED_LINKER_FLAGS " -B \"${MOLD_BIN_DIR}\"")
+      string(APPEND CMAKE_MODULE_LINKER_FLAGS " -B \"${MOLD_BIN_DIR}\"")
+      set(_IS_LINKER_DEFAULT OFF)
+    endif()
+    unset(MOLD_BIN)
+    unset(MOLD_BIN_DIR)
+  endif()
+
+  if(WITH_LINKER_GOLD AND _IS_LINKER_DEFAULT)
     execute_process(
       COMMAND ${CMAKE_C_COMPILER} -fuse-ld=gold -Wl,--version
       ERROR_QUIET OUTPUT_VARIABLE LD_VERSION)
     if("${LD_VERSION}" MATCHES "GNU gold")
-      string(APPEND CMAKE_C_FLAGS " -fuse-ld=gold")
+      string(APPEND CMAKE_EXE_LINKER_FLAGS    " -fuse-ld=gold")
-      string(APPEND CMAKE_CXX_FLAGS " -fuse-ld=gold")
+      string(APPEND CMAKE_SHARED_LINKER_FLAGS " -fuse-ld=gold")
+      string(APPEND CMAKE_MODULE_LINKER_FLAGS " -fuse-ld=gold")
+      set(_IS_LINKER_DEFAULT OFF)
     else()
       message(STATUS "GNU gold linker isn't available, using the default system linker.")
     endif()
     unset(LD_VERSION)
   endif()
 
-  if(WITH_LINKER_LLD)
+  if(WITH_LINKER_LLD AND _IS_LINKER_DEFAULT)
     execute_process(
       COMMAND ${CMAKE_C_COMPILER} -fuse-ld=lld -Wl,--version
       ERROR_QUIET OUTPUT_VARIABLE LD_VERSION)
     if("${LD_VERSION}" MATCHES "LLD")
-      string(APPEND CMAKE_C_FLAGS " -fuse-ld=lld")
+      string(APPEND CMAKE_EXE_LINKER_FLAGS    " -fuse-ld=lld")
-      string(APPEND CMAKE_CXX_FLAGS " -fuse-ld=lld")
+      string(APPEND CMAKE_SHARED_LINKER_FLAGS " -fuse-ld=lld")
+      string(APPEND CMAKE_MODULE_LINKER_FLAGS " -fuse-ld=lld")
+      set(_IS_LINKER_DEFAULT OFF)
     else()
       message(STATUS "LLD linker isn't available, using the default system linker.")
     endif()
@@ -691,6 +794,28 @@ if(CMAKE_COMPILER_IS_GNUCC)
 # CLang is the same as GCC for now.
 elseif(CMAKE_C_COMPILER_ID MATCHES "Clang")
   set(PLATFORM_CFLAGS "-pipe -fPIC -funsigned-char -fno-strict-aliasing")
+
+  if(WITH_LINKER_MOLD AND _IS_LINKER_DEFAULT)
+    find_program(MOLD_BIN "mold")
+    mark_as_advanced(MOLD_BIN)
+    if(NOT MOLD_BIN)
+      message(STATUS "The \"mold\" binary could not be found, using system linker.")
+      set(WITH_LINKER_MOLD OFF)
+    else()
+      if(CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 12.0)
+        string(APPEND CMAKE_EXE_LINKER_FLAGS    " --ld-path=\"${MOLD_BIN}\"")
+        string(APPEND CMAKE_SHARED_LINKER_FLAGS " --ld-path=\"${MOLD_BIN}\"")
+        string(APPEND CMAKE_MODULE_LINKER_FLAGS " --ld-path=\"${MOLD_BIN}\"")
+      else()
+        string(APPEND CMAKE_EXE_LINKER_FLAGS    " -fuse-ld=\"${MOLD_BIN}\"")
+        string(APPEND CMAKE_SHARED_LINKER_FLAGS " -fuse-ld=\"${MOLD_BIN}\"")
+        string(APPEND CMAKE_MODULE_LINKER_FLAGS " -fuse-ld=\"${MOLD_BIN}\"")
+      endif()
+      set(_IS_LINKER_DEFAULT OFF)
+    endif()
+    unset(MOLD_BIN)
+  endif()
+
 # Intel C++ Compiler
 elseif(CMAKE_C_COMPILER_ID MATCHES "Intel")
   # think these next two are broken
@@ -714,6 +839,8 @@ elseif(CMAKE_C_COMPILER_ID MATCHES "Intel")
   string(APPEND PLATFORM_LINKFLAGS " -static-intel")
 endif()
 
+unset(_IS_LINKER_DEFAULT)
+
 # Avoid conflicts with Mesa llvmpipe, Luxrender, and other plug-ins that may
 # use the same libraries as Blender with a different version or build options.
 set(PLATFORM_LINKFLAGS

build_files/cmake/platform/platform_win32.cmake

@@ -347,7 +347,11 @@ set(FREETYPE_INCLUDE_DIRS
   ${LIBDIR}/freetype/include
   ${LIBDIR}/freetype/include/freetype2
 )
-set(FREETYPE_LIBRARY ${LIBDIR}/freetype/lib/freetype2ST.lib)
+set(FREETYPE_LIBRARIES
+  ${LIBDIR}/freetype/lib/freetype2ST.lib
+  ${LIBDIR}/brotli/lib/brotlidec-static.lib
+  ${LIBDIR}/brotli/lib/brotlicommon-static.lib
+)
 windows_find_package(freetype REQUIRED)
 
 if(WITH_FFTW3)

build_files/windows/find_dependencies.cmd

@@ -3,7 +3,32 @@ for %%X in (svn.exe) do (set SVN=%%~$PATH:X)
 for %%X in (cmake.exe) do (set CMAKE=%%~$PATH:X)
 for %%X in (ctest.exe) do (set CTEST=%%~$PATH:X)
 for %%X in (git.exe) do (set GIT=%%~$PATH:X)
+REM For python, default on 39 but if that does not exist also check
+REM the 310,311 and 312 folders to see if those are there, it checks
+REM this far ahead to ensure good lib folder compatiblity in the future.
 set PYTHON=%BLENDER_DIR%\..\lib\win64_vc15\python\39\bin\python.exe
+if EXIST %PYTHON% (
+	goto detect_python_done
+)
+set PYTHON=%BLENDER_DIR%\..\lib\win64_vc15\python\310\bin\python.exe
+if EXIST %PYTHON% (
+	goto detect_python_done
+)
+set PYTHON=%BLENDER_DIR%\..\lib\win64_vc15\python\311\bin\python.exe
+if EXIST %PYTHON% (
+	goto detect_python_done
+)
+set PYTHON=%BLENDER_DIR%\..\lib\win64_vc15\python\312\bin\python.exe
+if EXIST %PYTHON% (
+	goto detect_python_done
+)
+
+if NOT EXIST %PYTHON% (
+    echo Warning: Python not found, there is likely an issue with the library folder
+    set PYTHON=""
+)
+
+:detect_python_done
 if NOT "%verbose%" == "" (
 	echo svn    : "%SVN%"
 	echo cmake  : "%CMAKE%"
@@ -11,7 +36,3 @@ if NOT "%verbose%" == "" (
 	echo git    : "%GIT%"
 	echo python : "%PYTHON%"
 )
-if "%CMAKE%" == "" (
-	echo Cmake not found in path, required for building, exiting...
-	exit /b 1
-)

build_files/windows/format.cmd

@@ -9,17 +9,11 @@ exit /b 1
 :detect_done
 echo found clang-format in %CF_PATH%
 
-if EXIST %PYTHON% (
+if NOT EXIST %PYTHON% (
-    set PYTHON=%BLENDER_DIR%\..\lib\win64_vc15\python\39\bin\python.exe
+    echo python not found, required for this operation
-    goto detect_python_done
+    exit /b 1
 )
 
-echo python not found in lib folder
-exit /b 1
-
-:detect_python_done
-echo found python (%PYTHON%)
-
 set FORMAT_PATHS=%BLENDER_DIR%\source\tools\utils_maintenance\clang_format_paths.py
 
 REM The formatting script expects clang-format to be in the current PATH.

build_files/windows/icons.cmd

@@ -1,18 +1,8 @@
-if EXIST "%PYTHON%" (
+if NOT EXIST %PYTHON% (
-    goto detect_python_done
+    echo python not found, required for this operation
+    exit /b 1
 )
 
-set PYTHON=%BLENDER_DIR%\..\lib\win64_vc15\python\39\bin\python.exe
-if EXIST %PYTHON% (
-    goto detect_python_done
-)
-
-echo python not found at %PYTHON%
-exit /b 1
-
-:detect_python_done
-echo found python (%PYTHON%)
-
 call "%~dp0\find_inkscape.cmd"
 
 if EXIST "%INKSCAPE_BIN%" (

build_files/windows/icons_geom.cmd

@@ -1,18 +1,8 @@
-if EXIST %PYTHON% (
+if NOT EXIST %PYTHON% (
-    goto detect_python_done
+    echo python not found, required for this operation
+    exit /b 1
 )
 
-set PYTHON=%BLENDER_DIR%\..\lib\win64_vc15\python\39\bin\python.exe
-if EXIST %PYTHON% (
-    goto detect_python_done
-)
-
-echo python not found at %PYTHON%
-exit /b 1
-
-:detect_python_done
-echo found python (%PYTHON%)
-
 call "%~dp0\find_blender.cmd"
 
 if EXIST "%BLENDER_BIN%" (

build_files/windows/update_sources.cmd

@@ -1,10 +1,7 @@
-if EXIST %PYTHON% (
+if NOT EXIST %PYTHON% (
-	goto detect_python_done
+    echo python not found, required for this operation
+    exit /b 1
 )
-
-echo python not found in lib folder
-exit /b 1
-
 :detect_python_done
 
 REM Use -B to avoid writing __pycache__ in lib directory and causing update conflicts.

doc/doxygen/doxygen.intern.h

@@ -51,9 +51,6 @@
 /** \defgroup intern_mikktspace MikktSpace
  *  \ingroup intern */
 
-/** \defgroup intern_numaapi NUMA (Non Uniform Memory Architecture)
- *  \ingroup intern */
-
 /** \defgroup intern_rigidbody Rigid-Body C-API
  *  \ingroup intern */
 

doc/python_api/examples/bpy.types.Bone.convert_local_to_pose.py

@@ -8,6 +8,7 @@ def set_pose_matrices(obj, matrix_map):
     "Assign pose space matrices of all bones at once, ignoring constraints."
 
     def rec(pbone, parent_matrix):
+        if pbone.name in matrix_map:
             matrix = matrix_map[pbone.name]
 
             ## Instead of:
@@ -29,6 +30,20 @@ def set_pose_matrices(obj, matrix_map):
                     pbone.bone.matrix_local,
                     invert=True
                 )
+        else:
+            # Compute the updated pose matrix from local and new parent matrix
+            if pbone.parent:
+                matrix = pbone.bone.convert_local_to_pose(
+                    pbone.matrix_basis,
+                    pbone.bone.matrix_local,
+                    parent_matrix=parent_matrix,
+                    parent_matrix_local=pbone.parent.bone.matrix_local,
+                )
+            else:
+                matrix = pbone.bone.convert_local_to_pose(
+                    pbone.matrix_basis,
+                    pbone.bone.matrix_local,
+                )
 
         # Recursively process children, passing the new matrix through
         for child in pbone.children:

doc/python_api/rst/info_gotcha.rst

@@ -743,7 +743,7 @@ will re-allocate objects data,
 any references to a meshes vertices/polygons/UVs, armatures bones,
 curves points, etc. cannot be accessed after switching mode.
 
-Only the reference to the data its self can be re-accessed, the following example will crash.
+Only the reference to the data itself can be re-accessed, the following example will crash.
 
 .. code-block:: python
 

doc/python_api/sphinx_doc_gen.py

@@ -1762,6 +1762,7 @@ except ModuleNotFoundError:
     fw("html_show_sphinx = False\n")
     fw("html_baseurl = 'https://docs.blender.org/api/current/'\n")
     fw("html_use_opensearch = 'https://docs.blender.org/api/current'\n")
+    fw("html_show_search_summary = True\n")
     fw("html_split_index = True\n")
     fw("html_static_path = ['static']\n")
     fw("html_extra_path = ['static/favicon.ico', 'static/blender_logo.svg']\n")

extern/audaspace/CMakeLists.txt

@@ -1092,12 +1092,12 @@ if(WITH_PYTHON)
 		configure_file(${PYTHON_SOURCE_DIRECTORY}/setup.py.in ${CMAKE_CURRENT_BINARY_DIR}/setup.py ESCAPE_QUOTES @ONLY)
 
 		if(APPLE)
-			add_custom_command(OUTPUT build COMMAND MACOSX_DEPLOYMENT_TARGET=${CMAKE_OSX_DEPLOYMENT_TARGET} ${PYTHON_EXECUTABLE} setup.py build DEPENDS ${PYTHON_SRC} ${PYTHON_HDR})
+			add_custom_command(OUTPUT build COMMAND MACOSX_DEPLOYMENT_TARGET=${CMAKE_OSX_DEPLOYMENT_TARGET} ${PYTHON_EXECUTABLE} setup.py build DEPENDS ${PYTHON_SRC} ${PYTHON_HDR} setup.py)
 		elseif(WIN32)
 			set(ENV{VS100COMNTOOLS} $ENV{VS120COMNTOOLS})
-			add_custom_command(OUTPUT build COMMAND ${PYTHON_EXECUTABLE} setup.py build DEPENDS ${PYTHON_SRC} ${PYTHON_HDR})
+			add_custom_command(OUTPUT build COMMAND ${PYTHON_EXECUTABLE} setup.py build DEPENDS ${PYTHON_SRC} ${PYTHON_HDR} setup.py)
 		else()
-			add_custom_command(OUTPUT build COMMAND ${PYTHON_EXECUTABLE} setup.py build DEPENDS ${PYTHON_SRC} ${PYTHON_HDR})
+			add_custom_command(OUTPUT build COMMAND ${PYTHON_EXECUTABLE} setup.py build DEPENDS ${PYTHON_SRC} ${PYTHON_HDR} setup.py)
 		endif()
 		add_custom_target(pythonmodule ALL DEPENDS build SOURCES ${PYTHON_SOURCE_DIRECTORY}/setup.py.in ${PYTHON_SRC} ${PYTHON_HDR})
 		add_dependencies(pythonmodule audaspace)

extern/audaspace/bindings/python/setup.py.in

@@ -43,6 +43,7 @@ audaspace = Extension(
                       library_dirs = ['.', 'Release', 'Debug'],
                       language = 'c++',
                       extra_compile_args = extra_args,
+                      define_macros = [('WITH_CONVOLUTION', None)] if '@WITH_FFTW@' == 'ON' else [],
                       sources = [os.path.join(source_directory, file) for file in ['PyAPI.cpp', 'PyDevice.cpp', 'PyHandle.cpp', 'PySound.cpp', 'PySequenceEntry.cpp', 'PySequence.cpp', 'PyPlaybackManager.cpp', 'PyDynamicMusic.cpp', 'PyThreadPool.cpp', 'PySource.cpp'] + (['PyImpulseResponse.cpp', 'PyHRTF.cpp'] if '@WITH_FFTW@' == 'ON' else [])]
 )
 

extern/audaspace/plugins/wasapi/WASAPIDevice.cpp

@@ -95,6 +95,13 @@ void WASAPIDevice::runMixingThread()
 				sleep_duration = std::chrono::milliseconds(buffer_size * 1000 / int(m_specs.rate) / 2);
 			}
 
+			if(m_default_device_changed)
+			{
+				m_default_device_changed = false;
+				result = AUDCLNT_E_DEVICE_INVALIDATED;
+				goto stop_thread;
+			}
+
 			if(FAILED(result = m_audio_client->GetCurrentPadding(&padding)))
 				goto stop_thread;
 
@@ -296,13 +303,78 @@ bool WASAPIDevice::setupDevice(DeviceSpecs &specs)
 	return true;
 }
 
+ULONG WASAPIDevice::AddRef()
+{
+	return InterlockedIncrement(&m_reference_count);
+}
+
+ULONG WASAPIDevice::Release()
+{
+	ULONG reference_count = InterlockedDecrement(&m_reference_count);
+
+	if(0 == reference_count)
+		delete this;
+
+	return reference_count;
+}
+
+HRESULT WASAPIDevice::QueryInterface(REFIID riid, void **ppvObject)
+{
+	if(riid == __uuidof(IMMNotificationClient))
+	{
+		*ppvObject = reinterpret_cast<IMMNotificationClient*>(this);
+		AddRef();
+	}
+	else if(riid == IID_IUnknown)
+	{
+		*ppvObject = reinterpret_cast<IUnknown*>(this);
+		AddRef();
+	}
+	else
+	{
+		*ppvObject = nullptr;
+		return E_NOINTERFACE;
+	}
+
+	return S_OK;
+}
+
+HRESULT WASAPIDevice::OnDeviceStateChanged(LPCWSTR pwstrDeviceId, DWORD dwNewState)
+{
+	return S_OK;
+}
+
+HRESULT WASAPIDevice::OnDeviceAdded(LPCWSTR pwstrDeviceId)
+{
+	return S_OK;
+}
+
+HRESULT WASAPIDevice::OnDeviceRemoved(LPCWSTR pwstrDeviceId)
+{
+	return S_OK;
+}
+
+HRESULT WASAPIDevice::OnDefaultDeviceChanged(EDataFlow flow, ERole role, LPCWSTR pwstrDeviceId)
+{
+	if(flow != EDataFlow::eCapture)
+		m_default_device_changed = true;
+
+	return S_OK;
+}
+
+HRESULT WASAPIDevice::OnPropertyValueChanged(LPCWSTR pwstrDeviceId, const PROPERTYKEY key)
+{
+	return S_OK;
+}
+
 WASAPIDevice::WASAPIDevice(DeviceSpecs specs, int buffersize) :
 	m_buffersize(buffersize),
 	m_imm_device_enumerator(nullptr),
 	m_imm_device(nullptr),
 	m_audio_client(nullptr),
-
+	m_wave_format_extensible({}),
-	m_wave_format_extensible({})
+	m_default_device_changed(false),
+	m_reference_count(1)
 {
 	// initialize COM if it hasn't happened yet
 	CoInitializeEx(nullptr, COINIT_MULTITHREADED);
@@ -327,6 +399,8 @@ WASAPIDevice::WASAPIDevice(DeviceSpecs specs, int buffersize) :
 
 	create();
 
+	m_imm_device_enumerator->RegisterEndpointNotificationCallback(this);
+
 	return;
 
 	error:
@@ -340,6 +414,8 @@ WASAPIDevice::~WASAPIDevice()
 {
 	stopMixingThread();
 
+	m_imm_device_enumerator->UnregisterEndpointNotificationCallback(this);
+
 	SafeRelease(&m_audio_client);
 	SafeRelease(&m_imm_device);
 	SafeRelease(&m_imm_device_enumerator);

extern/audaspace/plugins/wasapi/WASAPIDevice.h

@@ -40,7 +40,7 @@ AUD_NAMESPACE_BEGIN
 /**
  * This device plays back through WASAPI, the Windows audio API.
  */
-class AUD_PLUGIN_API WASAPIDevice : public ThreadedDevice
+class AUD_PLUGIN_API WASAPIDevice : IMMNotificationClient, public ThreadedDevice
 {
 private:
 	int m_buffersize;
@@ -48,6 +48,8 @@ private:
 	IMMDevice* m_imm_device;
 	IAudioClient* m_audio_client;
 	WAVEFORMATEXTENSIBLE m_wave_format_extensible;
+	bool m_default_device_changed;
+	LONG m_reference_count;
 
 	AUD_LOCAL HRESULT setupRenderClient(IAudioRenderClient*& render_client, UINT32& buffer_size);
 
@@ -58,6 +60,17 @@ private:
 
 	AUD_LOCAL bool setupDevice(DeviceSpecs& specs);
 
+	// IUnknown implementation
+	ULONG STDMETHODCALLTYPE AddRef();
+	ULONG STDMETHODCALLTYPE Release();
+	HRESULT STDMETHODCALLTYPE QueryInterface(REFIID riid, void **ppvObject);
+	// IMMNotificationClient implementation
+	HRESULT STDMETHODCALLTYPE OnDeviceStateChanged(LPCWSTR pwstrDeviceId, DWORD dwNewState);
+	HRESULT STDMETHODCALLTYPE OnDeviceAdded(LPCWSTR pwstrDeviceId);
+	HRESULT STDMETHODCALLTYPE OnDeviceRemoved(LPCWSTR pwstrDeviceId);
+	HRESULT STDMETHODCALLTYPE OnDefaultDeviceChanged(EDataFlow flow, ERole role, LPCWSTR pwstrDeviceId);
+	HRESULT STDMETHODCALLTYPE OnPropertyValueChanged(LPCWSTR pwstrDeviceId, const PROPERTYKEY key);
+
 	// delete copy constructor and operator=
 	WASAPIDevice(const WASAPIDevice&) = delete;
 	WASAPIDevice& operator=(const WASAPIDevice&) = delete;

extern/hipew/src/hipew.c

@@ -257,7 +257,7 @@ static int hipewHipInit(void) {
 #endif
   static int initialized = 0;
   static int result = 0;
-  int error, driver_version;
+  int error;
 
   if (initialized) {
     return result;
@@ -565,8 +565,6 @@ int hipewCompilerVersion(void) {
   const char *path = hipewCompilerPath();
   const char *marker = "Hip compilation tools, release ";
   FILE *pipe;
-  int major, minor;
-  char *versionstr;
   char buf[128];
   char output[65536] = "\0";
   char command[65536] = "\0";

intern/CMakeLists.txt

@@ -25,7 +25,6 @@ add_subdirectory(ghost)
 add_subdirectory(guardedalloc)
 add_subdirectory(libmv)
 add_subdirectory(memutil)
-add_subdirectory(numaapi)
 add_subdirectory(opencolorio)
 add_subdirectory(opensubdiv)
 add_subdirectory(mikktspace)

intern/cycles/app/CMakeLists.txt

@@ -51,8 +51,6 @@ list(APPEND LIBRARIES ${CYCLES_GL_LIBRARIES})
 
 # Common configuration.
 
-cycles_link_directories()
-
 add_definitions(${GL_DEFINITIONS})
 
 include_directories(${INC})

intern/cycles/blender/addon/engine.py

@@ -60,9 +60,8 @@ def init():
 
     path = os.path.dirname(__file__)
     user_path = os.path.dirname(os.path.abspath(bpy.utils.user_resource('CONFIG', path='')))
-    temp_path = bpy.app.tempdir
 
-    _cycles.init(path, user_path, temp_path, bpy.app.background)
+    _cycles.init(path, user_path, bpy.app.background)
     _parse_command_line()
 
 

intern/cycles/blender/addon/properties.py

@@ -667,6 +667,11 @@ class CyclesRenderSettings(bpy.types.PropertyGroup):
         description="Use special type BVH optimized for hair (uses more ram but renders faster)",
         default=True,
     )
+    debug_use_compact_bvh: BoolProperty(
+        name="Use Compact BVH",
+        description="Use compact BVH structure (uses less ram but renders slower)",
+        default=True,
+    )
     debug_bvh_time_steps: IntProperty(
         name="BVH Time Steps",
         description="Split BVH primitives by this number of time steps to speed up render time in cost of memory",
@@ -802,7 +807,7 @@ class CyclesRenderSettings(bpy.types.PropertyGroup):
         name="Tile Size",
         default=2048,
         description="",
-        min=8, max=16384,
+        min=8, max=8192,
     )
 
     # Various fine-tuning debug flags
@@ -1447,6 +1452,19 @@ class CyclesPreferences(bpy.types.AddonPreferences):
                     num += 1
         return num
 
+    def has_multi_device(self):
+        import _cycles
+        compute_device_type = self.get_compute_device_type()
+        device_list = _cycles.available_devices(compute_device_type)
+        for device in device_list:
+            if device[1] == compute_device_type:
+                continue
+            for dev in self.devices:
+                if dev.use and dev.id == device[2]:
+                    return True
+
+        return False
+
     def has_active_device(self):
         return self.get_num_gpu_devices() > 0
 

intern/cycles/blender/addon/ui.py

@@ -118,11 +118,11 @@ def use_optix(context):
 
     return (get_device_type(context) == 'OPTIX' and cscene.device == 'GPU')
 
-
+def use_multi_device(context):
-def use_sample_all_lights(context):
     cscene = context.scene.cycles
-
+    if cscene.device != 'GPU':
-    return cscene.sample_all_lights_direct or cscene.sample_all_lights_indirect
+        return False
+    return context.preferences.addons[__package__].preferences.has_multi_device()
 
 
 def show_device_active(context):
@@ -667,6 +667,10 @@ class CYCLES_RENDER_PT_performance_acceleration_structure(CyclesButtonsPanel, Pa
     bl_label = "Acceleration Structure"
     bl_parent_id = "CYCLES_RENDER_PT_performance"
 
+    @classmethod
+    def poll(cls, context):
+        return not use_optix(context) or has_multi_device(context)
+
     def draw(self, context):
         import _cycles
 
@@ -679,22 +683,34 @@ class CYCLES_RENDER_PT_performance_acceleration_structure(CyclesButtonsPanel, Pa
 
         col = layout.column()
 
-        use_embree = False
-        if use_cpu(context):
         use_embree = _cycles.with_embree
-            if not use_embree:
+
+        if use_cpu(context):
+            col.prop(cscene, "debug_use_spatial_splits")
+            if use_embree:
+                col.prop(cscene, "debug_use_compact_bvh")
+            else:
+                sub = col.column()
+                sub.active = not cscene.debug_use_spatial_splits
+                sub.prop(cscene, "debug_bvh_time_steps")
+
+                col.prop(cscene, "debug_use_hair_bvh")
+
                 sub = col.column(align=True)
                 sub.label(text="Cycles built without Embree support")
                 sub.label(text="CPU raytracing performance will be poor")
-
+        else:
             col.prop(cscene, "debug_use_spatial_splits")
             sub = col.column()
-        sub.active = not use_embree
+            sub.active = not cscene.debug_use_spatial_splits
-        sub.prop(cscene, "debug_use_hair_bvh")
-        sub = col.column()
-        sub.active = not cscene.debug_use_spatial_splits and not use_embree
             sub.prop(cscene, "debug_bvh_time_steps")
 
+            col.prop(cscene, "debug_use_hair_bvh")
+
+            # CPU is used in addition to a GPU
+            if use_multi_device(context) and use_embree:
+                col.prop(cscene, "debug_use_compact_bvh")
+
 
 class CYCLES_RENDER_PT_performance_final_render(CyclesButtonsPanel, Panel):
     bl_label = "Final Render"
@@ -1803,18 +1819,45 @@ class CYCLES_RENDER_PT_bake_output(CyclesButtonsPanel, Panel):
         rd = scene.render
 
         if rd.use_bake_multires:
-            layout.prop(rd, "bake_margin")
             layout.prop(rd, "use_bake_clear", text="Clear Image")
-
             if rd.bake_type == 'DISPLACEMENT':
                 layout.prop(rd, "use_bake_lores_mesh")
         else:
             layout.prop(cbk, "target")
-
             if cbk.target == 'IMAGE_TEXTURES':
-                layout.prop(cbk, "margin")
                 layout.prop(cbk, "use_clear", text="Clear Image")
 
+class CYCLES_RENDER_PT_bake_output_margin(CyclesButtonsPanel, Panel):
+    bl_label = "Margin"
+    bl_context = "render"
+    bl_parent_id = "CYCLES_RENDER_PT_bake_output"
+    COMPAT_ENGINES = {'CYCLES'}
+
+    @classmethod
+    def poll(cls, context):
+        scene = context.scene
+        cbk = scene.render.bake
+        return cbk.target == 'IMAGE_TEXTURES'
+
+    def draw(self, context):
+        layout = self.layout
+        layout.use_property_split = True
+        layout.use_property_decorate = False  # No animation.
+
+        scene = context.scene
+        cscene = scene.cycles
+        cbk = scene.render.bake
+        rd = scene.render
+
+        if rd.use_bake_multires:
+            layout.prop(rd, "bake_margin_type", text="Type")
+            layout.prop(rd, "bake_margin", text="Size")
+        else:
+            if cbk.target == 'IMAGE_TEXTURES':
+                layout.prop(cbk, "margin_type", text="Type")
+                layout.prop(cbk, "margin", text="Size")
+
+
 
 class CYCLES_RENDER_PT_debug(CyclesDebugButtonsPanel, Panel):
     bl_label = "Debug"
@@ -2183,6 +2226,7 @@ classes = (
     CYCLES_RENDER_PT_bake_influence,
     CYCLES_RENDER_PT_bake_selected_to_active,
     CYCLES_RENDER_PT_bake_output,
+    CYCLES_RENDER_PT_bake_output_margin,
     CYCLES_RENDER_PT_debug,
     node_panel(CYCLES_MATERIAL_PT_settings),
     node_panel(CYCLES_MATERIAL_PT_settings_surface),

intern/cycles/blender/display_driver.cpp

@@ -272,12 +272,300 @@ uint BlenderDisplaySpaceShader::get_shader_program()
   return shader_program_;
 }
 
+/* --------------------------------------------------------------------
+ * DrawTile.
+ */
+
+/* Higher level representation of a texture from the graphics library. */
+class GLTexture {
+ public:
+  /* Global counter for all allocated OpenGL textures used by instances of this class. */
+  static inline std::atomic<int> num_used = 0;
+
+  GLTexture() = default;
+
+  ~GLTexture()
+  {
+    assert(gl_id == 0);
+  }
+
+  GLTexture(const GLTexture &other) = delete;
+  GLTexture &operator=(GLTexture &other) = delete;
+
+  GLTexture(GLTexture &&other) noexcept
+      : gl_id(other.gl_id), width(other.width), height(other.height)
+  {
+    other.reset();
+  }
+
+  GLTexture &operator=(GLTexture &&other)
+  {
+    if (this == &other) {
+      return *this;
+    }
+
+    gl_id = other.gl_id;
+    width = other.width;
+    height = other.height;
+
+    other.reset();
+
+    return *this;
+  }
+
+  bool gl_resources_ensure()
+  {
+    if (gl_id) {
+      return true;
+    }
+
+    /* Create texture. */
+    glGenTextures(1, &gl_id);
+    if (!gl_id) {
+      LOG(ERROR) << "Error creating texture.";
+      return false;
+    }
+
+    /* Configure the texture. */
+    glActiveTexture(GL_TEXTURE0);
+    glBindTexture(GL_TEXTURE_2D, gl_id);
+
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+
+    /* Clamp to edge so that precision issues when zoomed out (which forces linear interpolation)
+     * does not cause unwanted repetition. */
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+
+    glBindTexture(GL_TEXTURE_2D, 0);
+
+    ++num_used;
+
+    return true;
+  }
+
+  void gl_resources_destroy()
+  {
+    if (!gl_id) {
+      return;
+    }
+
+    glDeleteTextures(1, &gl_id);
+
+    reset();
+
+    --num_used;
+  }
+
+  /* OpenGL resource IDs of the texture.
+   *
+   * NOTE: Allocated on the render engine's context. */
+  uint gl_id = 0;
+
+  /* Dimensions of the texture in pixels. */
+  int width = 0;
+  int height = 0;
+
+ protected:
+  void reset()
+  {
+    gl_id = 0;
+    width = 0;
+    height = 0;
+  }
+};
+
+/* Higher level representation of a Pixel Buffer Object (PBO) from the graphics library. */
+class GLPixelBufferObject {
+ public:
+  /* Global counter for all allocated OpenGL PBOs used by instances of this class. */
+  static inline std::atomic<int> num_used = 0;
+
+  GLPixelBufferObject() = default;
+
+  ~GLPixelBufferObject()
+  {
+    assert(gl_id == 0);
+  }
+
+  GLPixelBufferObject(const GLPixelBufferObject &other) = delete;
+  GLPixelBufferObject &operator=(GLPixelBufferObject &other) = delete;
+
+  GLPixelBufferObject(GLPixelBufferObject &&other) noexcept
+      : gl_id(other.gl_id), width(other.width), height(other.height)
+  {
+    other.reset();
+  }
+
+  GLPixelBufferObject &operator=(GLPixelBufferObject &&other)
+  {
+    if (this == &other) {
+      return *this;
+    }
+
+    gl_id = other.gl_id;
+    width = other.width;
+    height = other.height;
+
+    other.reset();
+
+    return *this;
+  }
+
+  bool gl_resources_ensure()
+  {
+    if (gl_id) {
+      return true;
+    }
+
+    glGenBuffers(1, &gl_id);
+    if (!gl_id) {
+      LOG(ERROR) << "Error creating texture pixel buffer object.";
+      return false;
+    }
+
+    ++num_used;
+
+    return true;
+  }
+
+  void gl_resources_destroy()
+  {
+    if (!gl_id) {
+      return;
+    }
+
+    glDeleteBuffers(1, &gl_id);
+
+    reset();
+
+    --num_used;
+  }
+
+  /* OpenGL resource IDs of the PBO.
+   *
+   * NOTE: Allocated on the render engine's context. */
+  uint gl_id = 0;
+
+  /* Dimensions of the PBO. */
+  int width = 0;
+  int height = 0;
+
+ protected:
+  void reset()
+  {
+    gl_id = 0;
+    width = 0;
+    height = 0;
+  }
+};
+
+class DrawTile {
+ public:
+  DrawTile() = default;
+  ~DrawTile() = default;
+
+  DrawTile(const DrawTile &other) = delete;
+  DrawTile &operator=(const DrawTile &other) = delete;
+
+  DrawTile(DrawTile &&other) noexcept = default;
+
+  DrawTile &operator=(DrawTile &&other) = default;
+
+  bool gl_resources_ensure()
+  {
+    if (!texture.gl_resources_ensure()) {
+      gl_resources_destroy();
+      return false;
+    }
+
+    if (!gl_vertex_buffer) {
+      glGenBuffers(1, &gl_vertex_buffer);
+      if (!gl_vertex_buffer) {
+        LOG(ERROR) << "Error allocating tile VBO.";
+        gl_resources_destroy();
+        return false;
+      }
+    }
+
+    return true;
+  }
+
+  void gl_resources_destroy()
+  {
+    texture.gl_resources_destroy();
+
+    if (gl_vertex_buffer) {
+      glDeleteBuffers(1, &gl_vertex_buffer);
+      gl_vertex_buffer = 0;
+    }
+  }
+
+  inline bool ready_to_draw() const
+  {
+    return texture.gl_id != 0;
+  }
+
+  /* Texture which contains pixels of the tile. */
+  GLTexture texture;
+
+  /* Display parameters the texture of this tile has been updated for. */
+  BlenderDisplayDriver::Params params;
+
+  /* OpenGL resources needed for drawing. */
+  uint gl_vertex_buffer = 0;
+};
+
+class DrawTileAndPBO {
+ public:
+  bool gl_resources_ensure()
+  {
+    if (!tile.gl_resources_ensure() || !buffer_object.gl_resources_ensure()) {
+      gl_resources_destroy();
+      return false;
+    }
+
+    return true;
+  }
+
+  void gl_resources_destroy()
+  {
+    tile.gl_resources_destroy();
+    buffer_object.gl_resources_destroy();
+  }
+
+  DrawTile tile;
+  GLPixelBufferObject buffer_object;
+};
+
 /* --------------------------------------------------------------------
  * BlenderDisplayDriver.
  */
 
+struct BlenderDisplayDriver::Tiles {
+  /* Resources of a tile which is being currently rendered. */
+  DrawTileAndPBO current_tile;
+
+  /* All tiles which rendering is finished and which content will not be changed. */
+  struct {
+    vector<DrawTile> tiles;
+
+    void gl_resources_destroy_and_clear()
+    {
+      for (DrawTile &tile : tiles) {
+        tile.gl_resources_destroy();
+      }
+
+      tiles.clear();
+    }
+  } finished_tiles;
+};
+
 BlenderDisplayDriver::BlenderDisplayDriver(BL::RenderEngine &b_engine, BL::Scene &b_scene)
-    : b_engine_(b_engine), display_shader_(BlenderDisplayShader::create(b_engine, b_scene))
+    : b_engine_(b_engine),
+      display_shader_(BlenderDisplayShader::create(b_engine, b_scene)),
+      tiles_(make_unique<Tiles>())
 {
   /* Create context while on the main thread. */
   gl_context_create();
@@ -292,6 +580,21 @@ BlenderDisplayDriver::~BlenderDisplayDriver()
  * Update procedure.
  */
 
+void BlenderDisplayDriver::next_tile_begin()
+{
+  if (!tiles_->current_tile.tile.ready_to_draw()) {
+    LOG(ERROR)
+        << "Unexpectedly moving to the next tile without any data provided for current tile.";
+    return;
+  }
+
+  /* Moving to the next tile without giving render data for the current tile is not an expected
+   * situation. */
+  DCHECK(!need_clear_);
+
+  tiles_->finished_tiles.tiles.emplace_back(std::move(tiles_->current_tile.tile));
+}
+
 bool BlenderDisplayDriver::update_begin(const Params &params,
                                         int texture_width,
                                         int texture_height)
@@ -312,24 +615,33 @@ bool BlenderDisplayDriver::update_begin(const Params &params,
     glWaitSync((GLsync)gl_render_sync_, 0, GL_TIMEOUT_IGNORED);
   }
 
-  if (!gl_texture_resources_ensure()) {
+  DrawTile &current_tile = tiles_->current_tile.tile;
+  GLPixelBufferObject &current_tile_buffer_object = tiles_->current_tile.buffer_object;
+
+  /* Clear storage of all finished tiles when display clear is requested.
+   * Do it when new tile data is provided to handle the display clear flag in a single place.
+   * It also makes the logic reliable from the whether drawing did happen or not point of view. */
+  if (need_clear_) {
+    tiles_->finished_tiles.gl_resources_destroy_and_clear();
+    need_clear_ = false;
+  }
+
+  if (!tiles_->current_tile.gl_resources_ensure()) {
+    tiles_->current_tile.gl_resources_destroy();
     gl_context_disable();
     return false;
   }
 
   /* Update texture dimensions if needed. */
-  if (texture_.width != texture_width || texture_.height != texture_height) {
+  if (current_tile.texture.width != texture_width ||
+      current_tile.texture.height != texture_height) {
     glActiveTexture(GL_TEXTURE0);
-    glBindTexture(GL_TEXTURE_2D, texture_.gl_id);
+    glBindTexture(GL_TEXTURE_2D, current_tile.texture.gl_id);
     glTexImage2D(
         GL_TEXTURE_2D, 0, GL_RGBA16F, texture_width, texture_height, 0, GL_RGBA, GL_HALF_FLOAT, 0);
-    texture_.width = texture_width;
+    current_tile.texture.width = texture_width;
-    texture_.height = texture_height;
+    current_tile.texture.height = texture_height;
     glBindTexture(GL_TEXTURE_2D, 0);
-
-    /* Texture did change, and no pixel storage was provided. Tag for an explicit zeroing out to
-     * avoid undefined content. */
-    texture_.need_clear = true;
   }
 
   /* Update PBO dimensions if needed.
@@ -341,29 +653,58 @@ bool BlenderDisplayDriver::update_begin(const Params &params,
    * sending too much data to GPU when resolution divider is not 1. */
   /* TODO(sergey): Investigate whether keeping the PBO exact size of the texture makes non-interop
    * mode faster. */
-  const int buffer_width = params.full_size.x;
+  const int buffer_width = params.size.x;
-  const int buffer_height = params.full_size.y;
+  const int buffer_height = params.size.y;
-  if (texture_.buffer_width != buffer_width || texture_.buffer_height != buffer_height) {
+  if (current_tile_buffer_object.width != buffer_width ||
+      current_tile_buffer_object.height != buffer_height) {
     const size_t size_in_bytes = sizeof(half4) * buffer_width * buffer_height;
-    glBindBuffer(GL_PIXEL_UNPACK_BUFFER, texture_.gl_pbo_id);
+    glBindBuffer(GL_PIXEL_UNPACK_BUFFER, current_tile_buffer_object.gl_id);
     glBufferData(GL_PIXEL_UNPACK_BUFFER, size_in_bytes, 0, GL_DYNAMIC_DRAW);
     glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 
-    texture_.buffer_width = buffer_width;
+    current_tile_buffer_object.width = buffer_width;
-    texture_.buffer_height = buffer_height;
+    current_tile_buffer_object.height = buffer_height;
   }
 
-  /* New content will be provided to the texture in one way or another, so mark this in a
+  /* Store an updated parameters of the current tile.
-   * centralized place. */
+   * In theory it is only needed once per update of the tile, but doing it on every update is
-  texture_.need_update = true;
+   * the easiest and is not expensive. */
-
+  tiles_->current_tile.tile.params = params;
-  texture_.params = params;
 
   return true;
 }
 
+static void update_tile_texture_pixels(const DrawTileAndPBO &tile)
+{
+  const GLTexture &texture = tile.tile.texture;
+
+  DCHECK_NE(tile.buffer_object.gl_id, 0);
+
+  glActiveTexture(GL_TEXTURE0);
+  glBindTexture(GL_TEXTURE_2D, texture.gl_id);
+  glBindBuffer(GL_PIXEL_UNPACK_BUFFER, tile.buffer_object.gl_id);
+
+  glTexSubImage2D(
+      GL_TEXTURE_2D, 0, 0, 0, texture.width, texture.height, GL_RGBA, GL_HALF_FLOAT, 0);
+
+  glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
+  glBindTexture(GL_TEXTURE_2D, 0);
+}
+
 void BlenderDisplayDriver::update_end()
 {
+  /* Unpack the PBO into the texture as soon as the new content is provided.
+   *
+   * This allows to ensure that the unpacking happens while resources like graphics interop (which
+   * lifetime is outside of control of the display driver) are still valid, as well as allows to
+   * move the tile from being current to finished immediately after this call.
+   *
+   * One concern with this approach is that if the update happens more often than drawing then
+   * doing the unpack here occupies GPU transfer for no good reason. However, the render scheduler
+   * takes care of ensuring updates don't happen that often. In regular applications redraw will
+   * happen much more often than this update. */
+  update_tile_texture_pixels(tiles_->current_tile);
+
   gl_upload_sync_ = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
   glFlush();
 
@@ -376,7 +717,11 @@ void BlenderDisplayDriver::update_end()
 
 half4 *BlenderDisplayDriver::map_texture_buffer()
 {
-  glBindBuffer(GL_PIXEL_UNPACK_BUFFER, texture_.gl_pbo_id);
+  const uint pbo_gl_id = tiles_->current_tile.buffer_object.gl_id;
+
+  DCHECK_NE(pbo_gl_id, 0);
+
+  glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo_gl_id);
 
   half4 *mapped_rgba_pixels = reinterpret_cast<half4 *>(
       glMapBuffer(GL_PIXEL_UNPACK_BUFFER, GL_WRITE_ONLY));
@@ -384,15 +729,6 @@ half4 *BlenderDisplayDriver::map_texture_buffer()
     LOG(ERROR) << "Error mapping BlenderDisplayDriver pixel buffer object.";
   }
 
-  if (texture_.need_clear) {
-    const int64_t texture_width = texture_.width;
-    const int64_t texture_height = texture_.height;
-    memset(reinterpret_cast<void *>(mapped_rgba_pixels),
-           0,
-           texture_width * texture_height * sizeof(half4));
-    texture_.need_clear = false;
-  }
-
   return mapped_rgba_pixels;
 }
 
@@ -411,12 +747,9 @@ BlenderDisplayDriver::GraphicsInterop BlenderDisplayDriver::graphics_interop_get
 {
   GraphicsInterop interop_dst;
 
-  interop_dst.buffer_width = texture_.buffer_width;
+  interop_dst.buffer_width = tiles_->current_tile.buffer_object.width;
-  interop_dst.buffer_height = texture_.buffer_height;
+  interop_dst.buffer_height = tiles_->current_tile.buffer_object.height;
-  interop_dst.opengl_pbo_id = texture_.gl_pbo_id;
+  interop_dst.opengl_pbo_id = tiles_->current_tile.buffer_object.gl_id;
-
-  interop_dst.need_clear = texture_.need_clear;
-  texture_.need_clear = false;
 
   return interop_dst;
 }
@@ -437,7 +770,7 @@ void BlenderDisplayDriver::graphics_interop_deactivate()
 
 void BlenderDisplayDriver::clear()
 {
-  texture_.need_clear = true;
+  need_clear_ = true;
 }
 
 void BlenderDisplayDriver::set_zoom(float zoom_x, float zoom_y)
@@ -445,26 +778,155 @@ void BlenderDisplayDriver::set_zoom(float zoom_x, float zoom_y)
   zoom_ = make_float2(zoom_x, zoom_y);
 }
 
+/* Update vertex buffer with new coordinates of vertex positions and texture coordinates.
+ * This buffer is used to render texture in the viewport.
+ *
+ * NOTE: The buffer needs to be bound. */
+static void vertex_buffer_update(const DisplayDriver::Params &params)
+{
+  const int x = params.full_offset.x;
+  const int y = params.full_offset.y;
+
+  const int width = params.size.x;
+  const int height = params.size.y;
+
+  /* Invalidate old contents - avoids stalling if the buffer is still waiting in queue to be
+   * rendered. */
+  glBufferData(GL_ARRAY_BUFFER, 16 * sizeof(float), NULL, GL_STREAM_DRAW);
+
+  float *vpointer = reinterpret_cast<float *>(glMapBuffer(GL_ARRAY_BUFFER, GL_WRITE_ONLY));
+  if (!vpointer) {
+    return;
+  }
+
+  vpointer[0] = 0.0f;
+  vpointer[1] = 0.0f;
+  vpointer[2] = x;
+  vpointer[3] = y;
+
+  vpointer[4] = 1.0f;
+  vpointer[5] = 0.0f;
+  vpointer[6] = x + width;
+  vpointer[7] = y;
+
+  vpointer[8] = 1.0f;
+  vpointer[9] = 1.0f;
+  vpointer[10] = x + width;
+  vpointer[11] = y + height;
+
+  vpointer[12] = 0.0f;
+  vpointer[13] = 1.0f;
+  vpointer[14] = x;
+  vpointer[15] = y + height;
+
+  glUnmapBuffer(GL_ARRAY_BUFFER);
+}
+
+static void draw_tile(const float2 &zoom,
+                      const int texcoord_attribute,
+                      const int position_attribute,
+                      const DrawTile &draw_tile)
+{
+  if (!draw_tile.ready_to_draw()) {
+    return;
+  }
+
+  const GLTexture &texture = draw_tile.texture;
+
+  DCHECK_NE(texture.gl_id, 0);
+  DCHECK_NE(draw_tile.gl_vertex_buffer, 0);
+
+  glBindBuffer(GL_ARRAY_BUFFER, draw_tile.gl_vertex_buffer);
+
+  /* Draw at the parameters for which the texture has been updated for. This allows to always draw
+   * texture during bordered-rendered camera view without flickering. The validness of the display
+   * parameters for a texture is guaranteed by the initial "clear" state which makes drawing to
+   * have an early output.
+   *
+   * Such approach can cause some extra "jelly" effect during panning, but it is not more jelly
+   * than overlay of selected objects. Also, it's possible to redraw texture at an intersection of
+   * the texture draw parameters and the latest updated draw parameters (although, complexity of
+   * doing it might not worth it. */
+  vertex_buffer_update(draw_tile.params);
+
+  glBindTexture(GL_TEXTURE_2D, texture.gl_id);
+
+  /* Trick to keep sharp rendering without jagged edges on all GPUs.
+   *
+   * The idea here is to enforce driver to use linear interpolation when the image is not zoomed
+   * in.
+   * For the render result with a resolution divider in effect we always use nearest interpolation.
+   *
+   * Use explicit MIN assignment to make sure the driver does not have an undefined behavior at
+   * the zoom level 1. The MAG filter is always NEAREST. */
+  const float zoomed_width = draw_tile.params.size.x * zoom.x;
+  const float zoomed_height = draw_tile.params.size.y * zoom.y;
+  if (texture.width != draw_tile.params.size.x || texture.height != draw_tile.params.size.y) {
+    /* Resolution divider is different from 1, force nearest interpolation. */
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+  }
+  else if (zoomed_width - draw_tile.params.size.x > 0.5f ||
+           zoomed_height - draw_tile.params.size.y > 0.5f) {
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+  }
+  else {
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+  }
+
+  glVertexAttribPointer(
+      texcoord_attribute, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (const GLvoid *)0);
+  glVertexAttribPointer(position_attribute,
+                        2,
+                        GL_FLOAT,
+                        GL_FALSE,
+                        4 * sizeof(float),
+                        (const GLvoid *)(sizeof(float) * 2));
+
+  glDrawArrays(GL_TRIANGLE_FAN, 0, 4);
+}
+
+void BlenderDisplayDriver::flush()
+{
+  /* This is called from the render thread that also calls update_begin/end, right before ending
+   * the render loop. We wait for any queued PBO and render commands to be done, before destroying
+   * the render thread and activating the context in the main thread to destroy resources.
+   *
+   * If we don't do this, the NVIDIA driver hangs for a few seconds for when ending 3D viewport
+   * rendering, for unknown reasons. This was found with NVIDIA driver version 470.73 and a Quadro
+   * RTX 6000 on Linux. */
+  if (!gl_context_enable()) {
+    return;
+  }
+
+  if (gl_upload_sync_) {
+    glWaitSync((GLsync)gl_upload_sync_, 0, GL_TIMEOUT_IGNORED);
+  }
+
+  if (gl_render_sync_) {
+    glWaitSync((GLsync)gl_render_sync_, 0, GL_TIMEOUT_IGNORED);
+  }
+
+  gl_context_disable();
+}
+
 void BlenderDisplayDriver::draw(const Params &params)
 {
   /* See do_update_begin() for why no locking is required here. */
   const bool transparent = true;  // TODO(sergey): Derive this from Film.
 
-  if (!gl_draw_resources_ensure()) {
-    return;
-  }
-
   if (use_gl_context_) {
     gl_context_mutex_.lock();
   }
 
-  if (texture_.need_clear) {
+  if (need_clear_) {
     /* Texture is requested to be cleared and was not yet cleared.
      *
      * Do early return which should be equivalent of drawing all-zero texture.
      * Watch out for the lock though so that the clear happening during update is properly
      * synchronized here. */
+    if (use_gl_context_) {
       gl_context_mutex_.unlock();
+    }
     return;
   }
 
@@ -477,66 +939,37 @@ void BlenderDisplayDriver::draw(const Params &params)
     glBlendFunc(GL_ONE, GL_ONE_MINUS_SRC_ALPHA);
   }
 
-  display_shader_->bind(params.full_size.x, params.full_size.y);
-
   glActiveTexture(GL_TEXTURE0);
-  glBindTexture(GL_TEXTURE_2D, texture_.gl_id);
 
-  /* Trick to keep sharp rendering without jagged edges on all GPUs.
+  /* NOTE: The VAO is to be allocated on the drawing context as it is not shared across contexts.
-   *
+   * Simplest is to allocate it on every redraw so that it is possible to destroy it from a
-   * The idea here is to enforce driver to use linear interpolation when the image is not zoomed
+   * correct context. */
-   * in.
-   * For the render result with a resolution divider in effect we always use nearest interpolation.
-   *
-   * Use explicit MIN assignment to make sure the driver does not have an undefined behavior at
-   * the zoom level 1. The MAG filter is always NEAREST. */
-  const float zoomed_width = params.size.x * zoom_.x;
-  const float zoomed_height = params.size.y * zoom_.y;
-  if (texture_.width != params.size.x || texture_.height != params.size.y) {
-    /* Resolution divider is different from 1, force nearest interpolation. */
-    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-  }
-  else if (zoomed_width - params.size.x > 0.5f || zoomed_height - params.size.y > 0.5f) {
-    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-  }
-  else {
-    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
-  }
-
-  glBindBuffer(GL_ARRAY_BUFFER, vertex_buffer_);
-
-  texture_update_if_needed();
-  vertex_buffer_update(params);
-
-  /* TODO(sergey): Does it make sense/possible to cache/reuse the VAO? */
   GLuint vertex_array_object;
   glGenVertexArrays(1, &vertex_array_object);
   glBindVertexArray(vertex_array_object);
 
+  display_shader_->bind(params.full_size.x, params.full_size.y);
+
   const int texcoord_attribute = display_shader_->get_tex_coord_attrib_location();
   const int position_attribute = display_shader_->get_position_attrib_location();
 
   glEnableVertexAttribArray(texcoord_attribute);
   glEnableVertexAttribArray(position_attribute);
 
-  glVertexAttribPointer(
+  draw_tile(zoom_, texcoord_attribute, position_attribute, tiles_->current_tile.tile);
-      texcoord_attribute, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (const GLvoid *)0);
-  glVertexAttribPointer(position_attribute,
-                        2,
-                        GL_FLOAT,
-                        GL_FALSE,
-                        4 * sizeof(float),
-                        (const GLvoid *)(sizeof(float) * 2));
 
-  glDrawArrays(GL_TRIANGLE_FAN, 0, 4);
+  for (const DrawTile &tile : tiles_->finished_tiles.tiles) {
-
+    draw_tile(zoom_, texcoord_attribute, position_attribute, tile);
-  glBindBuffer(GL_ARRAY_BUFFER, 0);
+  }
-  glBindTexture(GL_TEXTURE_2D, 0);
-
-  glDeleteVertexArrays(1, &vertex_array_object);
 
   display_shader_->unbind();
 
+  glBindTexture(GL_TEXTURE_2D, 0);
+  glBindVertexArray(0);
+  glBindBuffer(GL_ARRAY_BUFFER, 0);
+
+  glDeleteVertexArrays(1, &vertex_array_object);
+
   if (transparent) {
     glDisable(GL_BLEND);
   }
@@ -544,6 +977,11 @@ void BlenderDisplayDriver::draw(const Params &params)
   gl_render_sync_ = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
   glFlush();
 
+  if (VLOG_IS_ON(5)) {
+    VLOG(5) << "Number of textures: " << GLTexture::num_used;
+    VLOG(5) << "Number of PBOs: " << GLPixelBufferObject::num_used;
+  }
+
   if (use_gl_context_) {
     gl_context_mutex_.unlock();
   }
@@ -618,154 +1056,16 @@ void BlenderDisplayDriver::gl_context_dispose()
   }
 }
 
-bool BlenderDisplayDriver::gl_draw_resources_ensure()
-{
-  if (!texture_.gl_id) {
-    /* If there is no texture allocated, there is nothing to draw. Inform the draw call that it can
-     * can not continue. Note that this is not an unrecoverable error, so once the texture is known
-     * we will come back here and create all the GPU resources needed for draw. */
-    return false;
-  }
-
-  if (gl_draw_resource_creation_attempted_) {
-    return gl_draw_resources_created_;
-  }
-  gl_draw_resource_creation_attempted_ = true;
-
-  if (!vertex_buffer_) {
-    glGenBuffers(1, &vertex_buffer_);
-    if (!vertex_buffer_) {
-      LOG(ERROR) << "Error creating vertex buffer.";
-      return false;
-    }
-  }
-
-  gl_draw_resources_created_ = true;
-
-  return true;
-}
-
 void BlenderDisplayDriver::gl_resources_destroy()
 {
   gl_context_enable();
 
-  if (vertex_buffer_ != 0) {
+  tiles_->current_tile.gl_resources_destroy();
-    glDeleteBuffers(1, &vertex_buffer_);
+  tiles_->finished_tiles.gl_resources_destroy_and_clear();
-  }
-
-  if (texture_.gl_pbo_id) {
-    glDeleteBuffers(1, &texture_.gl_pbo_id);
-    texture_.gl_pbo_id = 0;
-  }
-
-  if (texture_.gl_id) {
-    glDeleteTextures(1, &texture_.gl_id);
-    texture_.gl_id = 0;
-  }
 
   gl_context_disable();
 
   gl_context_dispose();
 }
 
-bool BlenderDisplayDriver::gl_texture_resources_ensure()
-{
-  if (texture_.creation_attempted) {
-    return texture_.is_created;
-  }
-  texture_.creation_attempted = true;
-
-  DCHECK(!texture_.gl_id);
-  DCHECK(!texture_.gl_pbo_id);
-
-  /* Create texture. */
-  glGenTextures(1, &texture_.gl_id);
-  if (!texture_.gl_id) {
-    LOG(ERROR) << "Error creating texture.";
-    return false;
-  }
-
-  /* Configure the texture. */
-  glActiveTexture(GL_TEXTURE0);
-  glBindTexture(GL_TEXTURE_2D, texture_.gl_id);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
-  glBindTexture(GL_TEXTURE_2D, 0);
-
-  /* Create PBO for the texture. */
-  glGenBuffers(1, &texture_.gl_pbo_id);
-  if (!texture_.gl_pbo_id) {
-    LOG(ERROR) << "Error creating texture pixel buffer object.";
-    return false;
-  }
-
-  /* Creation finished with a success. */
-  texture_.is_created = true;
-
-  return true;
-}
-
-void BlenderDisplayDriver::texture_update_if_needed()
-{
-  if (!texture_.need_update) {
-    return;
-  }
-
-  glBindBuffer(GL_PIXEL_UNPACK_BUFFER, texture_.gl_pbo_id);
-  glTexSubImage2D(
-      GL_TEXTURE_2D, 0, 0, 0, texture_.width, texture_.height, GL_RGBA, GL_HALF_FLOAT, 0);
-  glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
-
-  texture_.need_update = false;
-}
-
-void BlenderDisplayDriver::vertex_buffer_update(const Params & /*params*/)
-{
-  /* Draw at the parameters for which the texture has been updated for. This allows to always draw
-   * texture during bordered-rendered camera view without flickering. The validness of the display
-   * parameters for a texture is guaranteed by the initial "clear" state which makes drawing to
-   * have an early output.
-   *
-   * Such approach can cause some extra "jelly" effect during panning, but it is not more jelly
-   * than overlay of selected objects. Also, it's possible to redraw texture at an intersection of
-   * the texture draw parameters and the latest updated draw parameters (although, complexity of
-   * doing it might not worth it. */
-  const int x = texture_.params.full_offset.x;
-  const int y = texture_.params.full_offset.y;
-
-  const int width = texture_.params.size.x;
-  const int height = texture_.params.size.y;
-
-  /* Invalidate old contents - avoids stalling if the buffer is still waiting in queue to be
-   * rendered. */
-  glBufferData(GL_ARRAY_BUFFER, 16 * sizeof(float), NULL, GL_STREAM_DRAW);
-
-  float *vpointer = reinterpret_cast<float *>(glMapBuffer(GL_ARRAY_BUFFER, GL_WRITE_ONLY));
-  if (!vpointer) {
-    return;
-  }
-
-  vpointer[0] = 0.0f;
-  vpointer[1] = 0.0f;
-  vpointer[2] = x;
-  vpointer[3] = y;
-
-  vpointer[4] = 1.0f;
-  vpointer[5] = 0.0f;
-  vpointer[6] = x + width;
-  vpointer[7] = y;
-
-  vpointer[8] = 1.0f;
-  vpointer[9] = 1.0f;
-  vpointer[10] = x + width;
-  vpointer[11] = y + height;
-
-  vpointer[12] = 0.0f;
-  vpointer[13] = 1.0f;
-  vpointer[14] = x;
-  vpointer[15] = y + height;
-
-  glUnmapBuffer(GL_ARRAY_BUFFER);
-}
-
 CCL_NAMESPACE_END

intern/cycles/blender/display_driver.h

@@ -26,6 +26,7 @@
 
 #include "util/thread.h"
 #include "util/unique_ptr.h"
+#include "util/vector.h"
 
 CCL_NAMESPACE_BEGIN
 
@@ -112,6 +113,8 @@ class BlenderDisplayDriver : public DisplayDriver {
   void set_zoom(float zoom_x, float zoom_y);
 
  protected:
+  virtual void next_tile_begin() override;
+
   virtual bool update_begin(const Params &params, int texture_width, int texture_height) override;
   virtual void update_end() override;
 
@@ -122,33 +125,17 @@ class BlenderDisplayDriver : public DisplayDriver {
 
   virtual void draw(const Params &params) override;
 
+  virtual void flush() override;
+
   /* Helper function which allocates new GPU context. */
   void gl_context_create();
   bool gl_context_enable();
   void gl_context_disable();
   void gl_context_dispose();
 
-  /* Make sure texture is allocated and its initial configuration is performed. */
-  bool gl_texture_resources_ensure();
-
-  /* Ensure all runtime GPU resources needed for drawing are allocated.
-   * Returns true if all resources needed for drawing are available. */
-  bool gl_draw_resources_ensure();
-
   /* Destroy all GPU resources which are being used by this object. */
   void gl_resources_destroy();
 
-  /* Update GPU texture dimensions and content if needed (new pixel data was provided).
-   *
-   * NOTE: The texture needs to be bound. */
-  void texture_update_if_needed();
-
-  /* Update vertex buffer with new coordinates of vertex positions and texture coordinates.
-   * This buffer is used to render texture in the viewport.
-   *
-   * NOTE: The buffer needs to be bound. */
-  void vertex_buffer_update(const Params &params);
-
   BL::RenderEngine b_engine_;
 
   /* OpenGL context which is used the render engine doesn't have its own. */
@@ -159,50 +146,14 @@ class BlenderDisplayDriver : public DisplayDriver {
   /* Mutex used to guard the `gl_context_`. */
   thread_mutex gl_context_mutex_;
 
-  /* Texture which contains pixels of the render result. */
+  /* Content of the display is to be filled with zeroes. */
-  struct {
+  std::atomic<bool> need_clear_ = true;
-    /* Indicates whether texture creation was attempted and succeeded.
-     * Used to avoid multiple attempts of texture creation on GPU issues or GPU context
-     * misconfiguration. */
-    bool creation_attempted = false;
-    bool is_created = false;
-
-    /* OpenGL resource IDs of the texture itself and Pixel Buffer Object (PBO) used to write
-     * pixels to it.
-     *
-     * NOTE: Allocated on the engine's context. */
-    uint gl_id = 0;
-    uint gl_pbo_id = 0;
-
-    /* Is true when new data was written to the PBO, meaning, the texture might need to be resized
-     * and new data is to be uploaded to the GPU. */
-    bool need_update = false;
-
-    /* Content of the texture is to be filled with zeroes. */
-    std::atomic<bool> need_clear = true;
-
-    /* Dimensions of the texture in pixels. */
-    int width = 0;
-    int height = 0;
-
-    /* Dimensions of the underlying PBO. */
-    int buffer_width = 0;
-    int buffer_height = 0;
-
-    /* Display parameters the texture has been updated for. */
-    Params params;
-  } texture_;
 
   unique_ptr<BlenderDisplayShader> display_shader_;
 
-  /* Special track of whether GPU resources were attempted to be created, to avoid attempts of
+  /* Opaque storage for an internal state and data for tiles. */
-   * their re-creation on failure on every redraw. */
+  struct Tiles;
-  bool gl_draw_resource_creation_attempted_ = false;
+  unique_ptr<Tiles> tiles_;
-  bool gl_draw_resources_created_ = false;
-
-  /* Vertex buffer which hold vertices of a triangle fan which is textures with the texture
-   * holding the render result. */
-  uint vertex_buffer_ = 0;
 
   void *gl_render_sync_ = nullptr;
   void *gl_upload_sync_ = nullptr;

intern/cycles/blender/mesh.cpp

@@ -1071,7 +1071,15 @@ static void create_subd_mesh(Scene *scene,
 
   for (BL::MeshEdge &e : b_mesh.edges) {
     if (e.crease() != 0.0f) {
-      mesh->add_crease(e.vertices()[0], e.vertices()[1], e.crease());
+      mesh->add_edge_crease(e.vertices()[0], e.vertices()[1], e.crease());
+    }
+  }
+
+  for (BL::MeshVertexCreaseLayer &c : b_mesh.vertex_creases) {
+    for (int i = 0; i < c.data.length(); ++i) {
+      if (c.data[i].value() != 0.0f) {
+        mesh->add_vertex_crease(i, c.data[i].value());
+      }
     }
   }
 

intern/cycles/blender/object.cpp

@@ -529,6 +529,17 @@ void BlenderSync::sync_procedural(BL::Object &b_ob,
   string absolute_path = blender_absolute_path(b_data, b_ob, b_mesh_cache.cache_file().filepath());
   procedural->set_filepath(ustring(absolute_path));
 
+  array<ustring> layers;
+  for (BL::CacheFileLayer &layer : cache_file.layers) {
+    if (layer.hide_layer()) {
+      continue;
+    }
+
+    absolute_path = blender_absolute_path(b_data, b_ob, layer.filepath());
+    layers.push_back_slow(ustring(absolute_path));
+  }
+  procedural->set_layers(layers);
+
   procedural->set_scale(cache_file.scale());
 
   procedural->set_use_prefetch(cache_file.use_prefetch());

intern/cycles/blender/output_driver.cpp

@@ -51,8 +51,6 @@ bool BlenderOutputDriver::read_render_tile(const Tile &tile)
 
   BL::RenderLayer b_rlay = *b_single_rlay;
 
-  vector<float> pixels(tile.size.x * tile.size.y * 4);
-
   /* Copy each pass.
    * TODO:copy only the required ones for better performance? */
   for (BL::RenderPass &b_pass : b_rlay.passes) {
@@ -109,7 +107,7 @@ void BlenderOutputDriver::write_render_tile(const Tile &tile)
 
   BL::RenderLayer b_rlay = *b_single_rlay;
 
-  vector<float> pixels(tile.size.x * tile.size.y * 4);
+  vector<float> pixels(static_cast<size_t>(tile.size.x) * tile.size.y * 4);
 
   /* Copy each pass. */
   for (BL::RenderPass &b_pass : b_rlay.passes) {

intern/cycles/blender/python.cpp

@@ -138,20 +138,18 @@ static const char *PyC_UnicodeAsByte(PyObject *py_str, PyObject **coerce)
 
 static PyObject *init_func(PyObject * /*self*/, PyObject *args)
 {
-  PyObject *path, *user_path, *temp_path;
+  PyObject *path, *user_path;
   int headless;
 
-  if (!PyArg_ParseTuple(args, "OOOi", &path, &user_path, &temp_path, &headless)) {
+  if (!PyArg_ParseTuple(args, "OOi", &path, &user_path, &headless)) {
     return nullptr;
   }
 
-  PyObject *path_coerce = nullptr, *user_path_coerce = nullptr, *temp_path_coerce = nullptr;
+  PyObject *path_coerce = nullptr, *user_path_coerce = nullptr;
   path_init(PyC_UnicodeAsByte(path, &path_coerce),
-            PyC_UnicodeAsByte(user_path, &user_path_coerce),
+            PyC_UnicodeAsByte(user_path, &user_path_coerce));
-            PyC_UnicodeAsByte(temp_path, &temp_path_coerce));
   Py_XDECREF(path_coerce);
   Py_XDECREF(user_path_coerce);
-  Py_XDECREF(temp_path_coerce);
 
   BlenderSession::headless = headless;
 

intern/cycles/blender/session.cpp

@@ -502,10 +502,15 @@ void BlenderSession::render_frame_finish()
     path_remove(filename);
   }
 
-  /* Clear driver. */
+  /* Clear output driver. */
   session->set_output_driver(nullptr);
   session->full_buffer_written_cb = function_null;
 
+  /* The display driver holds OpenGL resources which belong to an OpenGL context held by the render
+   * engine on Blender side. Force destruction of those resources. */
+  display_driver_ = nullptr;
+  session->set_display_driver(nullptr);
+
   /* All the files are handled.
    * Clear the list so that this session can be re-used by Persistent Data. */
   full_buffer_files_.clear();

intern/cycles/blender/shader.cpp

@@ -689,6 +689,9 @@ static ShaderNode *add_node(Scene *scene,
   else if (b_node.is_a(&RNA_ShaderNodeHairInfo)) {
     node = graph->create_node<HairInfoNode>();
   }
+  else if (b_node.is_a(&RNA_ShaderNodePointInfo)) {
+    node = graph->create_node<PointInfoNode>();
+  }
   else if (b_node.is_a(&RNA_ShaderNodeVolumeInfo)) {
     node = graph->create_node<VolumeInfoNode>();
   }

intern/cycles/blender/sync.cpp

@@ -787,6 +787,7 @@ SceneParams BlenderSync::get_scene_params(BL::Scene &b_scene, bool background)
     params.bvh_type = BVH_TYPE_DYNAMIC;
 
   params.use_bvh_spatial_split = RNA_boolean_get(&cscene, "debug_use_spatial_splits");
+  params.use_bvh_compact_structure = RNA_boolean_get(&cscene, "debug_use_compact_bvh");
   params.use_bvh_unaligned_nodes = RNA_boolean_get(&cscene, "debug_use_hair_bvh");
   params.num_bvh_time_steps = RNA_int_get(&cscene, "debug_bvh_time_steps");
 
@@ -832,6 +833,14 @@ SessionParams BlenderSync::get_session_params(BL::RenderEngine &b_engine,
   SessionParams params;
   PointerRNA cscene = RNA_pointer_get(&b_scene.ptr, "cycles");
 
+  if (background && !b_engine.is_preview()) {
+    /* Viewport and preview renders do not require temp directory and do request session
+     * parameters more often than the background render.
+     * Optimize RNA-C++ usage and memory allocation a bit by saving string access which we know is
+     * not needed for viewport render. */
+    params.temp_dir = b_engine.temporary_directory();
+  }
+
   /* feature set */
   params.experimental = (get_enum(cscene, "feature_set") != 0);
 

intern/cycles/bvh/build.cpp

@@ -935,7 +935,7 @@ BVHNode *BVHBuild::create_object_leaf_nodes(const BVHReference *ref, int start,
 
 BVHNode *BVHBuild::create_leaf_node(const BVHRange &range, const vector<BVHReference> &references)
 {
-  /* This is a bit overallocating here (considering leaf size into account),
+  /* This is a bit over-allocating here (considering leaf size into account),
    * but chunk-based re-allocation in vector makes it difficult to use small
    * size of stack storage here. Some tweaks are possible tho.
    *

intern/cycles/bvh/embree.cpp

@@ -61,6 +61,26 @@ static_assert(Object::MAX_MOTION_STEPS == Geometry::MAX_MOTION_STEPS,
 
 #  define IS_HAIR(x) (x & 1)
 
+/* This gets called by Embree at every valid ray/object intersection.
+ * Things like recording subsurface or shadow hits for later evaluation
+ * as well as filtering for volume objects happen here.
+ * Cycles' own BVH does that directly inside the traversal calls.
+ */
+static void rtc_filter_intersection_func(const RTCFilterFunctionNArguments *args)
+{
+  /* Current implementation in Cycles assumes only single-ray intersection queries. */
+  assert(args->N == 1);
+
+  RTCHit *hit = (RTCHit *)args->hit;
+  CCLIntersectContext *ctx = ((IntersectContext *)args->context)->userRayExt;
+  const KernelGlobalsCPU *kg = ctx->kg;
+  const Ray *cray = ctx->ray;
+
+  if (kernel_embree_is_self_intersection(kg, hit, cray)) {
+    *args->valid = 0;
+  }
+}
+
 /* This gets called by Embree at every valid ray/object intersection.
  * Things like recording subsurface or shadow hits for later evaluation
  * as well as filtering for volume objects happen here.
@@ -75,12 +95,16 @@ static void rtc_filter_occluded_func(const RTCFilterFunctionNArguments *args)
   RTCHit *hit = (RTCHit *)args->hit;
   CCLIntersectContext *ctx = ((IntersectContext *)args->context)->userRayExt;
   const KernelGlobalsCPU *kg = ctx->kg;
+  const Ray *cray = ctx->ray;
 
   switch (ctx->type) {
     case CCLIntersectContext::RAY_SHADOW_ALL: {
       Intersection current_isect;
       kernel_embree_convert_hit(kg, ray, hit, &current_isect);
-
+      if (intersection_skip_self_shadow(cray->self, current_isect.object, current_isect.prim)) {
+        *args->valid = 0;
+        return;
+      }
       /* If no transparent shadows or max number of hits exceeded, all light is blocked. */
       const int flags = intersection_get_shader_flags(kg, current_isect.prim, current_isect.type);
       if (!(flags & (SD_HAS_TRANSPARENT_SHADOW)) || ctx->num_hits >= ctx->max_hits) {
@@ -160,6 +184,10 @@ static void rtc_filter_occluded_func(const RTCFilterFunctionNArguments *args)
           break;
         }
       }
+      if (intersection_skip_self_local(cray->self, current_isect.prim)) {
+        *args->valid = 0;
+        return;
+      }
 
       /* No intersection information requested, just return a hit. */
       if (ctx->max_hits == 0) {
@@ -225,6 +253,11 @@ static void rtc_filter_occluded_func(const RTCFilterFunctionNArguments *args)
       if (ctx->num_hits < ctx->max_hits) {
         Intersection current_isect;
         kernel_embree_convert_hit(kg, ray, hit, &current_isect);
+        if (intersection_skip_self(cray->self, current_isect.object, current_isect.prim)) {
+          *args->valid = 0;
+          return;
+        }
+
         Intersection *isect = &ctx->isect_s[ctx->num_hits];
         ++ctx->num_hits;
         *isect = current_isect;
@@ -236,12 +269,15 @@ static void rtc_filter_occluded_func(const RTCFilterFunctionNArguments *args)
         }
         /* This tells Embree to continue tracing. */
         *args->valid = 0;
-        break;
       }
+      break;
     }
     case CCLIntersectContext::RAY_REGULAR:
     default:
-      /* Nothing to do here. */
+      if (kernel_embree_is_self_intersection(kg, hit, cray)) {
+        *args->valid = 0;
+        return;
+      }
       break;
   }
 }
@@ -257,6 +293,14 @@ static void rtc_filter_func_backface_cull(const RTCFilterFunctionNArguments *arg
     *args->valid = 0;
     return;
   }
+
+  CCLIntersectContext *ctx = ((IntersectContext *)args->context)->userRayExt;
+  const KernelGlobalsCPU *kg = ctx->kg;
+  const Ray *cray = ctx->ray;
+
+  if (kernel_embree_is_self_intersection(kg, hit, cray)) {
+    *args->valid = 0;
+  }
 }
 
 static void rtc_filter_occluded_func_backface_cull(const RTCFilterFunctionNArguments *args)
@@ -355,10 +399,12 @@ void BVHEmbree::build(Progress &progress, Stats *stats, RTCDevice rtc_device_)
   }
 
   const bool dynamic = params.bvh_type == BVH_TYPE_DYNAMIC;
+  const bool compact = params.use_compact_structure;
 
   scene = rtcNewScene(rtc_device);
   const RTCSceneFlags scene_flags = (dynamic ? RTC_SCENE_FLAG_DYNAMIC : RTC_SCENE_FLAG_NONE) |
-                                    RTC_SCENE_FLAG_COMPACT | RTC_SCENE_FLAG_ROBUST;
+                                    (compact ? RTC_SCENE_FLAG_COMPACT : RTC_SCENE_FLAG_NONE) |
+                                    RTC_SCENE_FLAG_ROBUST;
   rtcSetSceneFlags(scene, scene_flags);
   build_quality = dynamic ? RTC_BUILD_QUALITY_LOW :
                             (params.use_spatial_split ? RTC_BUILD_QUALITY_HIGH :
@@ -503,6 +549,7 @@ void BVHEmbree::add_triangles(const Object *ob, const Mesh *mesh, int i)
 
   rtcSetGeometryUserData(geom_id, (void *)prim_offset);
   rtcSetGeometryOccludedFilterFunction(geom_id, rtc_filter_occluded_func);
+  rtcSetGeometryIntersectFilterFunction(geom_id, rtc_filter_intersection_func);
   rtcSetGeometryMask(geom_id, ob->visibility_for_tracing());
 
   rtcCommitGeometry(geom_id);
@@ -765,6 +812,7 @@ void BVHEmbree::add_curves(const Object *ob, const Hair *hair, int i)
 
   rtcSetGeometryUserData(geom_id, (void *)prim_offset);
   if (hair->curve_shape == CURVE_RIBBON) {
+    rtcSetGeometryIntersectFilterFunction(geom_id, rtc_filter_intersection_func);
     rtcSetGeometryOccludedFilterFunction(geom_id, rtc_filter_occluded_func);
   }
   else {

intern/cycles/bvh/params.h

@@ -97,6 +97,9 @@ class BVHParams {
    */
   bool use_unaligned_nodes;
 
+  /* Use compact acceleration structure (Embree)*/
+  bool use_compact_structure;
+
   /* Split time range to this number of steps and create leaf node for each
    * of this time steps.
    *

intern/cycles/cmake/macros.cmake

@@ -84,39 +84,6 @@ macro(cycles_add_library target library_deps)
   cycles_set_solution_folder(${target})
 endmacro()
 
-# Cycles library dependencies common to all executables
-
-function(cycles_link_directories)
-  if(APPLE)
-    # APPLE platform uses full paths for linking libraries, and avoids link_directories.
-    return()
-  endif()
-
-  if(WITH_OPENCOLORIO)
-    link_directories(${OPENCOLORIO_LIBPATH})
-  endif()
-  if(WITH_OPENVDB)
-    link_directories(${OPENVDB_LIBPATH} ${BLOSC_LIBPATH})
-  endif()
-  if(WITH_OPENSUBDIV)
-    link_directories(${OPENSUBDIV_LIBPATH})
-  endif()
-  if(WITH_OPENIMAGEDENOISE)
-    link_directories(${OPENIMAGEDENOISE_LIBPATH})
-  endif()
-
-  link_directories(
-    ${OPENIMAGEIO_LIBPATH}
-    ${BOOST_LIBPATH}
-    ${PNG_LIBPATH}
-    ${JPEG_LIBPATH}
-    ${ZLIB_LIBPATH}
-    ${TIFF_LIBPATH}
-    ${OPENEXR_LIBPATH}
-    ${OPENJPEG_LIBPATH}
-  )
-endfunction()
-
 macro(cycles_target_link_libraries target)
   if(WITH_CYCLES_LOGGING)
     target_link_libraries(${target} ${GLOG_LIBRARIES} ${GFLAGS_LIBRARIES})
@@ -168,12 +135,6 @@ macro(cycles_target_link_libraries target)
     target_link_libraries(${target} extern_hipew)
   endif()
 
-  if(CYCLES_STANDALONE_REPOSITORY)
-    target_link_libraries(${target} extern_numaapi)
-  else()
-    target_link_libraries(${target} bf_intern_numaapi)
-  endif()
-
   if(UNIX AND NOT APPLE)
     if(CYCLES_STANDALONE_REPOSITORY)
       target_link_libraries(${target} extern_libc_compat)

intern/cycles/device/cpu/device_impl.cpp

@@ -72,7 +72,7 @@ CPUDevice::CPUDevice(const DeviceInfo &info_, Stats &stats_, Profiler &profiler_
           << " CPU kernels.";
 
   if (info.cpu_threads == 0) {
-    info.cpu_threads = TaskScheduler::num_threads();
+    info.cpu_threads = TaskScheduler::max_concurrency();
   }
 
 #ifdef WITH_OSL

intern/cycles/device/cuda/graphics_interop.cpp

@@ -45,9 +45,11 @@ void CUDADeviceGraphicsInterop::set_display_interop(
 
   need_clear_ = display_interop.need_clear;
 
+  if (!display_interop.need_recreate) {
     if (opengl_pbo_id_ == display_interop.opengl_pbo_id && buffer_area_ == new_buffer_area) {
       return;
     }
+  }
 
   CUDAContextScope scope(device_);
 

intern/cycles/device/device.cpp

@@ -37,6 +37,7 @@
 #include "util/math.h"
 #include "util/string.h"
 #include "util/system.h"
+#include "util/task.h"
 #include "util/time.h"
 #include "util/types.h"
 #include "util/vector.h"
@@ -333,7 +334,7 @@ DeviceInfo Device::get_multi_device(const vector<DeviceInfo> &subdevices,
     /* Ensure CPU device does not slow down GPU. */
     if (device.type == DEVICE_CPU && subdevices.size() > 1) {
       if (background) {
-        int orig_cpu_threads = (threads) ? threads : system_cpu_thread_count();
+        int orig_cpu_threads = (threads) ? threads : TaskScheduler::max_concurrency();
         int cpu_threads = max(orig_cpu_threads - (subdevices.size() - 1), 0);
 
         VLOG(1) << "CPU render threads reduced from " << orig_cpu_threads << " to " << cpu_threads

intern/cycles/device/metal/bvh.h

@@ -58,6 +58,11 @@ class BVHMetal : public BVH {
                        id<MTLCommandQueue> queue,
                        Geometry *const geom,
                        bool refit);
+  bool build_BLAS_pointcloud(Progress &progress,
+                             id<MTLDevice> device,
+                             id<MTLCommandQueue> queue,
+                             Geometry *const geom,
+                             bool refit);
   bool build_TLAS(Progress &progress, id<MTLDevice> device, id<MTLCommandQueue> queue, bool refit);
 };
 

intern/cycles/device/metal/bvh.mm

@@ -19,6 +19,7 @@
 #  include "scene/hair.h"
 #  include "scene/mesh.h"
 #  include "scene/object.h"
+#  include "scene/pointcloud.h"
 
 #  include "util/progress.h"
 
@@ -475,6 +476,220 @@ bool BVHMetal::build_BLAS_hair(Progress &progress,
   return false;
 }
 
+bool BVHMetal::build_BLAS_pointcloud(Progress &progress,
+                                     id<MTLDevice> device,
+                                     id<MTLCommandQueue> queue,
+                                     Geometry *const geom,
+                                     bool refit)
+{
+  if (@available(macos 12.0, *)) {
+    /* Build BLAS for point cloud */
+    PointCloud *pointcloud = static_cast<PointCloud *>(geom);
+    if (pointcloud->num_points() == 0) {
+      return false;
+    }
+
+    /*------------------------------------------------*/
+    BVH_status("Building pointcloud BLAS | %7d points | %s",
+               (int)pointcloud->num_points(),
+               geom->name.c_str());
+    /*------------------------------------------------*/
+
+    const size_t num_points = pointcloud->get_points().size();
+    const float3 *points = pointcloud->get_points().data();
+    const float *radius = pointcloud->get_radius().data();
+
+    const bool use_fast_trace_bvh = (params.bvh_type == BVH_TYPE_STATIC);
+
+    size_t num_motion_steps = 1;
+    Attribute *motion_keys = pointcloud->attributes.find(ATTR_STD_MOTION_VERTEX_POSITION);
+    if (motion_blur && pointcloud->get_use_motion_blur() && motion_keys) {
+      num_motion_steps = pointcloud->get_motion_steps();
+    }
+
+    const size_t num_aabbs = num_motion_steps;
+
+    MTLResourceOptions storage_mode;
+    if (device.hasUnifiedMemory) {
+      storage_mode = MTLResourceStorageModeShared;
+    }
+    else {
+      storage_mode = MTLResourceStorageModeManaged;
+    }
+
+    /* Allocate a GPU buffer for the AABB data and populate it */
+    id<MTLBuffer> aabbBuf = [device
+        newBufferWithLength:num_aabbs * sizeof(MTLAxisAlignedBoundingBox)
+                    options:storage_mode];
+    MTLAxisAlignedBoundingBox *aabb_data = (MTLAxisAlignedBoundingBox *)[aabbBuf contents];
+
+    /* Get AABBs for each motion step */
+    size_t center_step = (num_motion_steps - 1) / 2;
+    for (size_t step = 0; step < num_motion_steps; ++step) {
+      /* The center step for motion vertices is not stored in the attribute */
+      if (step != center_step) {
+        size_t attr_offset = (step > center_step) ? step - 1 : step;
+        points = motion_keys->data_float3() + attr_offset * num_points;
+      }
+
+      for (size_t j = 0; j < num_points; ++j) {
+        const PointCloud::Point point = pointcloud->get_point(j);
+        BoundBox bounds = BoundBox::empty;
+        point.bounds_grow(points, radius, bounds);
+
+        const size_t index = step * num_points + j;
+        aabb_data[index].min = (MTLPackedFloat3 &)bounds.min;
+        aabb_data[index].max = (MTLPackedFloat3 &)bounds.max;
+      }
+    }
+
+    if (storage_mode == MTLResourceStorageModeManaged) {
+      [aabbBuf didModifyRange:NSMakeRange(0, aabbBuf.length)];
+    }
+
+#  if 0
+    for (size_t i=0; i<num_aabbs && i < 400; i++) {
+      MTLAxisAlignedBoundingBox& bb = aabb_data[i];
+      printf("  %d:   %.1f,%.1f,%.1f -- %.1f,%.1f,%.1f\n", int(i), bb.min.x, bb.min.y, bb.min.z, bb.max.x, bb.max.y, bb.max.z);
+    }
+#  endif
+
+    MTLAccelerationStructureGeometryDescriptor *geomDesc;
+    if (motion_blur) {
+      std::vector<MTLMotionKeyframeData *> aabb_ptrs;
+      aabb_ptrs.reserve(num_motion_steps);
+      for (size_t step = 0; step < num_motion_steps; ++step) {
+        MTLMotionKeyframeData *k = [MTLMotionKeyframeData data];
+        k.buffer = aabbBuf;
+        k.offset = step * num_points * sizeof(MTLAxisAlignedBoundingBox);
+        aabb_ptrs.push_back(k);
+      }
+
+      MTLAccelerationStructureMotionBoundingBoxGeometryDescriptor *geomDescMotion =
+          [MTLAccelerationStructureMotionBoundingBoxGeometryDescriptor descriptor];
+      geomDescMotion.boundingBoxBuffers = [NSArray arrayWithObjects:aabb_ptrs.data()
+                                                              count:aabb_ptrs.size()];
+      geomDescMotion.boundingBoxCount = num_points;
+      geomDescMotion.boundingBoxStride = sizeof(aabb_data[0]);
+      geomDescMotion.intersectionFunctionTableOffset = 2;
+
+      /* Force a single any-hit call, so shadow record-all behavior works correctly */
+      /* (Match optix behavior: unsigned int build_flags =
+       * OPTIX_GEOMETRY_FLAG_REQUIRE_SINGLE_ANYHIT_CALL;) */
+      geomDescMotion.allowDuplicateIntersectionFunctionInvocation = false;
+      geomDescMotion.opaque = true;
+      geomDesc = geomDescMotion;
+    }
+    else {
+      MTLAccelerationStructureBoundingBoxGeometryDescriptor *geomDescNoMotion =
+          [MTLAccelerationStructureBoundingBoxGeometryDescriptor descriptor];
+      geomDescNoMotion.boundingBoxBuffer = aabbBuf;
+      geomDescNoMotion.boundingBoxBufferOffset = 0;
+      geomDescNoMotion.boundingBoxCount = int(num_aabbs);
+      geomDescNoMotion.boundingBoxStride = sizeof(aabb_data[0]);
+      geomDescNoMotion.intersectionFunctionTableOffset = 2;
+
+      /* Force a single any-hit call, so shadow record-all behavior works correctly */
+      /* (Match optix behavior: unsigned int build_flags =
+       * OPTIX_GEOMETRY_FLAG_REQUIRE_SINGLE_ANYHIT_CALL;) */
+      geomDescNoMotion.allowDuplicateIntersectionFunctionInvocation = false;
+      geomDescNoMotion.opaque = true;
+      geomDesc = geomDescNoMotion;
+    }
+
+    MTLPrimitiveAccelerationStructureDescriptor *accelDesc =
+        [MTLPrimitiveAccelerationStructureDescriptor descriptor];
+    accelDesc.geometryDescriptors = @[ geomDesc ];
+
+    if (motion_blur) {
+      accelDesc.motionStartTime = 0.0f;
+      accelDesc.motionEndTime = 1.0f;
+      accelDesc.motionStartBorderMode = MTLMotionBorderModeVanish;
+      accelDesc.motionEndBorderMode = MTLMotionBorderModeVanish;
+      accelDesc.motionKeyframeCount = num_motion_steps;
+    }
+
+    if (!use_fast_trace_bvh) {
+      accelDesc.usage |= (MTLAccelerationStructureUsageRefit |
+                          MTLAccelerationStructureUsagePreferFastBuild);
+    }
+
+    MTLAccelerationStructureSizes accelSizes = [device
+        accelerationStructureSizesWithDescriptor:accelDesc];
+    id<MTLAccelerationStructure> accel_uncompressed = [device
+        newAccelerationStructureWithSize:accelSizes.accelerationStructureSize];
+    id<MTLBuffer> scratchBuf = [device newBufferWithLength:accelSizes.buildScratchBufferSize
+                                                   options:MTLResourceStorageModePrivate];
+    id<MTLBuffer> sizeBuf = [device newBufferWithLength:8 options:MTLResourceStorageModeShared];
+    id<MTLCommandBuffer> accelCommands = [queue commandBuffer];
+    id<MTLAccelerationStructureCommandEncoder> accelEnc =
+        [accelCommands accelerationStructureCommandEncoder];
+    if (refit) {
+      [accelEnc refitAccelerationStructure:accel_struct
+                                descriptor:accelDesc
+                               destination:accel_uncompressed
+                             scratchBuffer:scratchBuf
+                       scratchBufferOffset:0];
+    }
+    else {
+      [accelEnc buildAccelerationStructure:accel_uncompressed
+                                descriptor:accelDesc
+                             scratchBuffer:scratchBuf
+                       scratchBufferOffset:0];
+    }
+    if (use_fast_trace_bvh) {
+      [accelEnc writeCompactedAccelerationStructureSize:accel_uncompressed
+                                               toBuffer:sizeBuf
+                                                 offset:0
+                                           sizeDataType:MTLDataTypeULong];
+    }
+    [accelEnc endEncoding];
+    [accelCommands addCompletedHandler:^(id<MTLCommandBuffer> command_buffer) {
+      /* free temp resources */
+      [scratchBuf release];
+      [aabbBuf release];
+
+      if (use_fast_trace_bvh) {
+        /* Compact the accel structure */
+        uint64_t compressed_size = *(uint64_t *)sizeBuf.contents;
+
+        dispatch_async(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), ^{
+          id<MTLCommandBuffer> accelCommands = [queue commandBuffer];
+          id<MTLAccelerationStructureCommandEncoder> accelEnc =
+              [accelCommands accelerationStructureCommandEncoder];
+          id<MTLAccelerationStructure> accel = [device
+              newAccelerationStructureWithSize:compressed_size];
+          [accelEnc copyAndCompactAccelerationStructure:accel_uncompressed
+                                toAccelerationStructure:accel];
+          [accelEnc endEncoding];
+          [accelCommands addCompletedHandler:^(id<MTLCommandBuffer> command_buffer) {
+            uint64_t allocated_size = [accel allocatedSize];
+            stats.mem_alloc(allocated_size);
+            accel_struct = accel;
+            [accel_uncompressed release];
+            accel_struct_building = false;
+          }];
+          [accelCommands commit];
+        });
+      }
+      else {
+        /* set our acceleration structure to the uncompressed structure */
+        accel_struct = accel_uncompressed;
+
+        uint64_t allocated_size = [accel_struct allocatedSize];
+        stats.mem_alloc(allocated_size);
+        accel_struct_building = false;
+      }
+      [sizeBuf release];
+    }];
+
+    accel_struct_building = true;
+    [accelCommands commit];
+    return true;
+  }
+  return false;
+}
+
 bool BVHMetal::build_BLAS(Progress &progress,
                           id<MTLDevice> device,
                           id<MTLCommandQueue> queue,
@@ -491,6 +706,8 @@ bool BVHMetal::build_BLAS(Progress &progress,
         return build_BLAS_mesh(progress, device, queue, geom, refit);
       case Geometry::HAIR:
         return build_BLAS_hair(progress, device, queue, geom, refit);
+      case Geometry::POINTCLOUD:
+        return build_BLAS_pointcloud(progress, device, queue, geom, refit);
       default:
         return false;
     }

intern/cycles/device/metal/device_impl.h

@@ -115,6 +115,8 @@ class MetalDevice : public Device {
 
   void load_texture_info();
 
+  void erase_allocation(device_memory &mem);
+
   virtual bool should_use_graphics_interop() override;
 
   virtual unique_ptr<DeviceQueue> gpu_queue_create() override;

intern/cycles/device/metal/device_impl.mm

@@ -87,17 +87,14 @@ MetalDevice::MetalDevice(const DeviceInfo &info, Stats &stats, Profiler &profile
     default:
       break;
     case METAL_GPU_INTEL: {
-      use_metalrt = false;
       max_threads_per_threadgroup = 64;
       break;
     }
     case METAL_GPU_AMD: {
-      use_metalrt = false;
       max_threads_per_threadgroup = 128;
       break;
     }
     case METAL_GPU_APPLE: {
-      use_metalrt = true;
       max_threads_per_threadgroup = 512;
       break;
     }
@@ -432,6 +429,25 @@ void MetalDevice::load_texture_info()
   }
 }
 
+void MetalDevice::erase_allocation(device_memory &mem)
+{
+  stats.mem_free(mem.device_size);
+  mem.device_pointer = 0;
+  mem.device_size = 0;
+
+  auto it = metal_mem_map.find(&mem);
+  if (it != metal_mem_map.end()) {
+    MetalMem *mmem = it->second.get();
+
+    /* blank out reference to MetalMem* in the launch params (fixes crash T94736) */
+    if (mmem->pointer_index >= 0) {
+      device_ptr *pointers = (device_ptr *)&launch_params;
+      pointers[mmem->pointer_index] = 0;
+    }
+    metal_mem_map.erase(it);
+  }
+}
+
 MetalDevice::MetalMem *MetalDevice::generic_alloc(device_memory &mem)
 {
   size_t size = mem.memory_size();
@@ -561,11 +577,7 @@ void MetalDevice::generic_free(device_memory &mem)
       mmem.mtlBuffer = nil;
     }
 
-    stats.mem_free(mem.device_size);
+    erase_allocation(mem);
-    mem.device_pointer = 0;
-    mem.device_size = 0;
-
-    metal_mem_map.erase(&mem);
   }
 }
 
@@ -954,10 +966,7 @@ void MetalDevice::tex_free(device_texture &mem)
       delayed_free_list.push_back(mmem.mtlTexture);
       mmem.mtlTexture = nil;
     }
-    stats.mem_free(mem.device_size);
+    erase_allocation(mem);
-    mem.device_pointer = 0;
-    mem.device_size = 0;
-    metal_mem_map.erase(&mem);
   }
 }
 

intern/cycles/device/metal/kernel.h

@@ -36,6 +36,8 @@ enum {
   METALRT_FUNC_CURVE_RIBBON_SHADOW,
   METALRT_FUNC_CURVE_ALL,
   METALRT_FUNC_CURVE_ALL_SHADOW,
+  METALRT_FUNC_POINT,
+  METALRT_FUNC_POINT_SHADOW,
   METALRT_FUNC_NUM
 };
 

intern/cycles/device/metal/kernel.mm

@@ -358,6 +358,8 @@ bool MetalDeviceKernels::load(MetalDevice *device, int kernel_type)
           "__intersection__curve_ribbon_shadow",
           "__intersection__curve_all",
           "__intersection__curve_all_shadow",
+          "__intersection__point",
+          "__intersection__point_shadow",
       };
       assert(sizeof(function_names) / sizeof(function_names[0]) == METALRT_FUNC_NUM);
 
@@ -400,36 +402,50 @@ bool MetalDeviceKernels::load(MetalDevice *device, int kernel_type)
     NSArray *function_list = nil;
 
     if (device->use_metalrt) {
-      id<MTLFunction> box_intersect_default = nil;
+      id<MTLFunction> curve_intersect_default = nil;
-      id<MTLFunction> box_intersect_shadow = nil;
+      id<MTLFunction> curve_intersect_shadow = nil;
+      id<MTLFunction> point_intersect_default = nil;
+      id<MTLFunction> point_intersect_shadow = nil;
       if (device->kernel_features & KERNEL_FEATURE_HAIR) {
         /* Add curve intersection programs. */
         if (device->kernel_features & KERNEL_FEATURE_HAIR_THICK) {
           /* Slower programs for thick hair since that also slows down ribbons.
            * Ideally this should not be needed. */
-          box_intersect_default = rt_intersection_funcs[kernel_type][METALRT_FUNC_CURVE_ALL];
+          curve_intersect_default = rt_intersection_funcs[kernel_type][METALRT_FUNC_CURVE_ALL];
-          box_intersect_shadow = rt_intersection_funcs[kernel_type][METALRT_FUNC_CURVE_ALL_SHADOW];
+          curve_intersect_shadow =
+              rt_intersection_funcs[kernel_type][METALRT_FUNC_CURVE_ALL_SHADOW];
         }
         else {
-          box_intersect_default = rt_intersection_funcs[kernel_type][METALRT_FUNC_CURVE_RIBBON];
+          curve_intersect_default = rt_intersection_funcs[kernel_type][METALRT_FUNC_CURVE_RIBBON];
-          box_intersect_shadow =
+          curve_intersect_shadow =
               rt_intersection_funcs[kernel_type][METALRT_FUNC_CURVE_RIBBON_SHADOW];
         }
       }
+      if (device->kernel_features & KERNEL_FEATURE_POINTCLOUD) {
+        point_intersect_default = rt_intersection_funcs[kernel_type][METALRT_FUNC_POINT];
+        point_intersect_shadow = rt_intersection_funcs[kernel_type][METALRT_FUNC_POINT_SHADOW];
+      }
       table_functions[METALRT_TABLE_DEFAULT] = [NSArray
           arrayWithObjects:rt_intersection_funcs[kernel_type][METALRT_FUNC_DEFAULT_TRI],
-                           box_intersect_default ?
+                           curve_intersect_default ?
-                               box_intersect_default :
+                               curve_intersect_default :
+                               rt_intersection_funcs[kernel_type][METALRT_FUNC_DEFAULT_BOX],
+                           point_intersect_default ?
+                               point_intersect_default :
                                rt_intersection_funcs[kernel_type][METALRT_FUNC_DEFAULT_BOX],
                            nil];
       table_functions[METALRT_TABLE_SHADOW] = [NSArray
           arrayWithObjects:rt_intersection_funcs[kernel_type][METALRT_FUNC_SHADOW_TRI],
-                           box_intersect_shadow ?
+                           curve_intersect_shadow ?
-                               box_intersect_shadow :
+                               curve_intersect_shadow :
+                               rt_intersection_funcs[kernel_type][METALRT_FUNC_SHADOW_BOX],
+                           point_intersect_shadow ?
+                               point_intersect_shadow :
                                rt_intersection_funcs[kernel_type][METALRT_FUNC_SHADOW_BOX],
                            nil];
       table_functions[METALRT_TABLE_LOCAL] = [NSArray
           arrayWithObjects:rt_intersection_funcs[kernel_type][METALRT_FUNC_LOCAL_TRI],
+                           rt_intersection_funcs[kernel_type][METALRT_FUNC_LOCAL_BOX],
                            rt_intersection_funcs[kernel_type][METALRT_FUNC_LOCAL_BOX],
                            nil];
 

intern/cycles/device/optix/device_impl.cpp

@@ -226,7 +226,7 @@ bool OptiXDevice::load_kernels(const uint kernel_features)
   pipeline_options.usesMotionBlur = false;
   pipeline_options.traversableGraphFlags =
       OPTIX_TRAVERSABLE_GRAPH_FLAG_ALLOW_SINGLE_LEVEL_INSTANCING;
-  pipeline_options.numPayloadValues = 6;
+  pipeline_options.numPayloadValues = 8;
   pipeline_options.numAttributeValues = 2; /* u, v */
   pipeline_options.exceptionFlags = OPTIX_EXCEPTION_FLAG_NONE;
   pipeline_options.pipelineLaunchParamsVariableName = "__params"; /* See globals.h */

intern/cycles/integrator/pass_accessor.cpp

@@ -141,6 +141,7 @@ bool PassAccessor::get_render_tile_pixels(const RenderBuffers *render_buffers,
   const PassType type = pass_access_info_.type;
   const PassMode mode = pass_access_info_.mode;
   const PassInfo pass_info = Pass::get_info(type, pass_access_info_.include_albedo);
+  int num_written_components = pass_info.num_components;
 
   if (pass_info.num_components == 1) {
     /* Single channel passes. */
@@ -188,8 +189,10 @@ bool PassAccessor::get_render_tile_pixels(const RenderBuffers *render_buffers,
     else if ((pass_info.divide_type != PASS_NONE || pass_info.direct_type != PASS_NONE ||
               pass_info.indirect_type != PASS_NONE) &&
              mode != PassMode::DENOISED) {
-      /* RGB lighting passes that need to divide out color and/or sum direct and indirect. */
+      /* RGB lighting passes that need to divide out color and/or sum direct and indirect.
+       * These can also optionally write alpha like the combined pass. */
       get_pass_light_path(render_buffers, buffer_params, destination);
+      num_written_components = 4;
     }
     else {
       /* Passes that need no special computation, or denoised passes that already
@@ -215,7 +218,7 @@ bool PassAccessor::get_render_tile_pixels(const RenderBuffers *render_buffers,
     }
   }
 
-  pad_pixels(buffer_params, destination, pass_info.num_components);
+  pad_pixels(buffer_params, destination, num_written_components);
 
   return true;
 }

intern/cycles/integrator/path_trace.cpp

@@ -115,7 +115,9 @@ bool PathTrace::ready_to_reset()
   return false;
 }
 
-void PathTrace::reset(const BufferParams &full_params, const BufferParams &big_tile_params)
+void PathTrace::reset(const BufferParams &full_params,
+                      const BufferParams &big_tile_params,
+                      const bool reset_rendering)
 {
   if (big_tile_params_.modified(big_tile_params)) {
     big_tile_params_ = big_tile_params;
@@ -128,7 +130,7 @@ void PathTrace::reset(const BufferParams &full_params, const BufferParams &big_t
    * It is requires to inform about reset whenever it happens, so that the redraw state tracking is
    * properly updated. */
   if (display_) {
-    display_->reset(full_params);
+    display_->reset(big_tile_params, reset_rendering);
   }
 
   render_state_.has_denoised_result = false;
@@ -594,6 +596,15 @@ void PathTrace::draw()
   did_draw_after_reset_ |= display_->draw();
 }
 
+void PathTrace::flush_display()
+{
+  if (!display_) {
+    return;
+  }
+
+  display_->flush();
+}
+
 void PathTrace::update_display(const RenderWork &render_work)
 {
   if (!render_work.display.update) {
@@ -622,9 +633,8 @@ void PathTrace::update_display(const RenderWork &render_work)
   if (display_) {
     VLOG(3) << "Perform copy to GPUDisplay work.";
 
-    const int resolution_divider = render_work.resolution_divider;
+    const int texture_width = render_state_.effective_big_tile_params.window_width;
-    const int texture_width = max(1, full_params_.width / resolution_divider);
+    const int texture_height = render_state_.effective_big_tile_params.window_height;
-    const int texture_height = max(1, full_params_.height / resolution_divider);
     if (!display_->update_begin(texture_width, texture_height)) {
       LOG(ERROR) << "Error beginning GPUDisplay update.";
       return;
@@ -810,8 +820,15 @@ void PathTrace::tile_buffer_read()
     return;
   }
 
+  /* Read buffers back from device. */
+  tbb::parallel_for_each(path_trace_works_, [&](unique_ptr<PathTraceWork> &path_trace_work) {
+    path_trace_work->copy_render_buffers_from_device();
+  });
+
+  /* Read (subset of) passes from output driver. */
   PathTraceTile tile(*this);
   if (output_driver_->read_render_tile(tile)) {
+    /* Copy buffers to device again. */
     tbb::parallel_for_each(path_trace_works_, [](unique_ptr<PathTraceWork> &path_trace_work) {
       path_trace_work->copy_render_buffers_to_device();
     });

intern/cycles/integrator/path_trace.h

@@ -72,7 +72,9 @@ class PathTrace {
    * render result. */
   bool ready_to_reset();
 
-  void reset(const BufferParams &full_params, const BufferParams &big_tile_params);
+  void reset(const BufferParams &full_params,
+             const BufferParams &big_tile_params,
+             bool reset_rendering);
 
   void device_free();
 
@@ -112,6 +114,9 @@ class PathTrace {
   /* Perform drawing of the current state of the DisplayDriver. */
   void draw();
 
+  /* Flush outstanding display commands before ending the render loop. */
+  void flush_display();
+
   /* Cancel rendering process as soon as possible, without waiting for full tile to be sampled.
    * Used in cases like reset of render session.
    *

intern/cycles/integrator/path_trace_display.cpp

@@ -26,15 +26,20 @@ PathTraceDisplay::PathTraceDisplay(unique_ptr<DisplayDriver> driver) : driver_(m
 {
 }
 
-void PathTraceDisplay::reset(const BufferParams &buffer_params)
+void PathTraceDisplay::reset(const BufferParams &buffer_params, const bool reset_rendering)
 {
   thread_scoped_lock lock(mutex_);
 
-  params_.full_offset = make_int2(buffer_params.full_x, buffer_params.full_y);
+  params_.full_offset = make_int2(buffer_params.full_x + buffer_params.window_x,
+                                  buffer_params.full_y + buffer_params.window_y);
   params_.full_size = make_int2(buffer_params.full_width, buffer_params.full_height);
-  params_.size = make_int2(buffer_params.width, buffer_params.height);
+  params_.size = make_int2(buffer_params.window_width, buffer_params.window_height);
 
   texture_state_.is_outdated = true;
+
+  if (!reset_rendering) {
+    driver_->next_tile_begin();
+  }
 }
 
 void PathTraceDisplay::mark_texture_updated()
@@ -248,4 +253,9 @@ bool PathTraceDisplay::draw()
   return !is_outdated;
 }
 
+void PathTraceDisplay::flush()
+{
+  driver_->flush();
+}
+
 CCL_NAMESPACE_END

intern/cycles/integrator/path_trace_display.h

@@ -38,14 +38,17 @@ class BufferParams;
 
 class PathTraceDisplay {
  public:
-  PathTraceDisplay(unique_ptr<DisplayDriver> driver);
+  explicit PathTraceDisplay(unique_ptr<DisplayDriver> driver);
   virtual ~PathTraceDisplay() = default;
 
   /* Reset the display for the new state of render session. Is called whenever session is reset,
    * which happens on changes like viewport navigation or viewport dimension change.
    *
-   * This call will configure parameters for a changed buffer and reset the texture state. */
+   * This call will configure parameters for a changed buffer and reset the texture state.
-  void reset(const BufferParams &buffer_params);
+   *
+   * When the `reset_rendering` a complete display reset happens. When it is false reset happens
+   * for a new state of the buffer parameters which is assumed to correspond to the next tile. */
+  void reset(const BufferParams &buffer_params, bool reset_rendering);
 
   /* --------------------------------------------------------------------
    * Update procedure.
@@ -151,6 +154,9 @@ class PathTraceDisplay {
    * Returns true if this call did draw an updated state of the texture. */
   bool draw();
 
+  /* Flush outstanding display commands before ending the render loop. */
+  void flush();
+
  private:
   /* Display driver implemented by the host application. */
   unique_ptr<DisplayDriver> driver_;

intern/cycles/integrator/path_trace_work.cpp

@@ -194,10 +194,10 @@ PassAccessor::Destination PathTraceWork::get_display_destination_template(
   PassAccessor::Destination destination(film_->get_display_pass());
 
   const int2 display_texture_size = display->get_texture_size();
-  const int texture_x = effective_buffer_params_.full_x - effective_full_params_.full_x +
+  const int texture_x = effective_buffer_params_.full_x - effective_big_tile_params_.full_x +
-                        effective_buffer_params_.window_x;
+                        effective_buffer_params_.window_x - effective_big_tile_params_.window_x;
-  const int texture_y = effective_buffer_params_.full_y - effective_full_params_.full_y +
+  const int texture_y = effective_buffer_params_.full_y - effective_big_tile_params_.full_y +
-                        effective_buffer_params_.window_y;
+                        effective_buffer_params_.window_y - effective_big_tile_params_.window_y;
 
   destination.offset = texture_y * display_texture_size.x + texture_x;
   destination.stride = display_texture_size.x;

intern/cycles/integrator/path_trace_work_gpu.cpp

@@ -875,8 +875,10 @@ void PathTraceWorkGPU::copy_to_display_naive(PathTraceDisplay *display,
   const int final_width = buffers_->params.window_width;
   const int final_height = buffers_->params.window_height;
 
-  const int texture_x = full_x - effective_full_params_.full_x + effective_buffer_params_.window_x;
+  const int texture_x = full_x - effective_big_tile_params_.full_x +
-  const int texture_y = full_y - effective_full_params_.full_y + effective_buffer_params_.window_y;
+                        effective_buffer_params_.window_x - effective_big_tile_params_.window_x;
+  const int texture_y = full_y - effective_big_tile_params_.full_y +
+                        effective_buffer_params_.window_y - effective_big_tile_params_.window_y;
 
   /* Re-allocate display memory if needed, and make sure the device pointer is allocated.
    *

intern/cycles/integrator/render_scheduler.cpp

@@ -406,9 +406,6 @@ bool RenderScheduler::set_postprocess_render_work(RenderWork *render_work)
     any_scheduled = true;
   }
 
-  /* Force update. */
-  any_scheduled = true;
-
   if (any_scheduled) {
     render_work->display.update = true;
   }

intern/cycles/integrator/render_scheduler.h

@@ -283,7 +283,7 @@ class RenderScheduler {
   /* Check whether timing report about the given work need to reset accumulated average time. */
   bool work_report_reset_average(const RenderWork &render_work);
 
-  /* CHeck whether render time limit has been reached (or exceeded), and if so store related
+  /* Check whether render time limit has been reached (or exceeded), and if so store related
    * information in the state so that rendering is considered finished, and is possible to report
    * average render time information. */
   void check_time_limit_reached();

intern/cycles/kernel/bvh/bvh.h

@@ -173,15 +173,16 @@ ccl_device_intersect bool scene_intersect(KernelGlobals kg,
   uint p3 = 0;
   uint p4 = visibility;
   uint p5 = PRIMITIVE_NONE;
+  uint p6 = ((uint64_t)ray) & 0xFFFFFFFF;
+  uint p7 = (((uint64_t)ray) >> 32) & 0xFFFFFFFF;
 
   uint ray_mask = visibility & 0xFF;
-  uint ray_flags = OPTIX_RAY_FLAG_NONE;
+  uint ray_flags = OPTIX_RAY_FLAG_ENFORCE_ANYHIT;
   if (0 == ray_mask && (visibility & ~0xFF) != 0) {
     ray_mask = 0xFF;
-    ray_flags = OPTIX_RAY_FLAG_ENFORCE_ANYHIT;
   }
   else if (visibility & PATH_RAY_SHADOW_OPAQUE) {
-    ray_flags = OPTIX_RAY_FLAG_TERMINATE_ON_FIRST_HIT;
+    ray_flags |= OPTIX_RAY_FLAG_TERMINATE_ON_FIRST_HIT;
   }
 
   optixTrace(scene_intersect_valid(ray) ? kernel_data.bvh.scene : 0,
@@ -200,7 +201,9 @@ ccl_device_intersect bool scene_intersect(KernelGlobals kg,
              p2,
              p3,
              p4,
-             p5);
+             p5,
+             p6,
+             p7);
 
   isect->t = __uint_as_float(p0);
   isect->u = __uint_as_float(p1);
@@ -242,6 +245,7 @@ ccl_device_intersect bool scene_intersect(KernelGlobals kg,
   }
 
   MetalRTIntersectionPayload payload;
+  payload.self = ray->self;
   payload.u = 0.0f;
   payload.v = 0.0f;
   payload.visibility = visibility;
@@ -309,6 +313,7 @@ ccl_device_intersect bool scene_intersect(KernelGlobals kg,
     CCLIntersectContext ctx(kg, CCLIntersectContext::RAY_REGULAR);
     IntersectContext rtc_ctx(&ctx);
     RTCRayHit ray_hit;
+    ctx.ray = ray;
     kernel_embree_setup_rayhit(*ray, ray_hit, visibility);
     rtcIntersect1(kernel_data.bvh.scene, &rtc_ctx.context, &ray_hit);
     if (ray_hit.hit.geomID != RTC_INVALID_GEOMETRY_ID &&
@@ -356,6 +361,9 @@ ccl_device_intersect bool scene_intersect_local(KernelGlobals kg,
   uint p2 = pointer_pack_to_uint_0(local_isect);
   uint p3 = pointer_pack_to_uint_1(local_isect);
   uint p4 = local_object;
+  uint p6 = ((uint64_t)ray) & 0xFFFFFFFF;
+  uint p7 = (((uint64_t)ray) >> 32) & 0xFFFFFFFF;
+
   /* Is set to zero on miss or if ray is aborted, so can be used as return value. */
   uint p5 = max_hits;
 
@@ -379,7 +387,9 @@ ccl_device_intersect bool scene_intersect_local(KernelGlobals kg,
              p2,
              p3,
              p4,
-             p5);
+             p5,
+             p6,
+             p7);
 
   return p5;
 #  elif defined(__METALRT__)
@@ -417,6 +427,7 @@ ccl_device_intersect bool scene_intersect_local(KernelGlobals kg,
   }
 
   MetalRTIntersectionLocalPayload payload;
+  payload.self = ray->self;
   payload.local_object = local_object;
   payload.max_hits = max_hits;
   payload.local_isect.num_hits = 0;
@@ -460,6 +471,7 @@ ccl_device_intersect bool scene_intersect_local(KernelGlobals kg,
         kg, has_bvh ? CCLIntersectContext::RAY_SSS : CCLIntersectContext::RAY_LOCAL);
     ctx.lcg_state = lcg_state;
     ctx.max_hits = max_hits;
+    ctx.ray = ray;
     ctx.local_isect = local_isect;
     if (local_isect) {
       local_isect->num_hits = 0;
@@ -532,6 +544,8 @@ ccl_device_intersect bool scene_intersect_shadow_all(KernelGlobals kg,
   uint p3 = max_hits;
   uint p4 = visibility;
   uint p5 = false;
+  uint p6 = ((uint64_t)ray) & 0xFFFFFFFF;
+  uint p7 = (((uint64_t)ray) >> 32) & 0xFFFFFFFF;
 
   uint ray_mask = visibility & 0xFF;
   if (0 == ray_mask && (visibility & ~0xFF) != 0) {
@@ -555,7 +569,9 @@ ccl_device_intersect bool scene_intersect_shadow_all(KernelGlobals kg,
              p2,
              p3,
              p4,
-             p5);
+             p5,
+             p6,
+             p7);
 
   *num_recorded_hits = uint16_unpack_from_uint_0(p2);
   *throughput = __uint_as_float(p1);
@@ -588,6 +604,7 @@ ccl_device_intersect bool scene_intersect_shadow_all(KernelGlobals kg,
   }
 
   MetalRTIntersectionShadowPayload payload;
+  payload.self = ray->self;
   payload.visibility = visibility;
   payload.max_hits = max_hits;
   payload.num_hits = 0;
@@ -634,6 +651,7 @@ ccl_device_intersect bool scene_intersect_shadow_all(KernelGlobals kg,
     Intersection *isect_array = (Intersection *)state->shadow_isect;
     ctx.isect_s = isect_array;
     ctx.max_hits = max_hits;
+    ctx.ray = ray;
     IntersectContext rtc_ctx(&ctx);
     RTCRay rtc_ray;
     kernel_embree_setup_ray(*ray, rtc_ray, visibility);
@@ -685,6 +703,8 @@ ccl_device_intersect bool scene_intersect_volume(KernelGlobals kg,
   uint p3 = 0;
   uint p4 = visibility;
   uint p5 = PRIMITIVE_NONE;
+  uint p6 = ((uint64_t)ray) & 0xFFFFFFFF;
+  uint p7 = (((uint64_t)ray) >> 32) & 0xFFFFFFFF;
 
   uint ray_mask = visibility & 0xFF;
   if (0 == ray_mask && (visibility & ~0xFF) != 0) {
@@ -708,7 +728,9 @@ ccl_device_intersect bool scene_intersect_volume(KernelGlobals kg,
              p2,
              p3,
              p4,
-             p5);
+             p5,
+             p6,
+             p7);
 
   isect->t = __uint_as_float(p0);
   isect->u = __uint_as_float(p1);
@@ -744,6 +766,7 @@ ccl_device_intersect bool scene_intersect_volume(KernelGlobals kg,
   }
 
   MetalRTIntersectionPayload payload;
+  payload.self = ray->self;
   payload.visibility = visibility;
 
   typename metalrt_intersector_type::result_type intersection;
@@ -820,6 +843,7 @@ ccl_device_intersect uint scene_intersect_volume_all(KernelGlobals kg,
     ctx.isect_s = isect;
     ctx.max_hits = max_hits;
     ctx.num_hits = 0;
+    ctx.ray = ray;
     IntersectContext rtc_ctx(&ctx);
     RTCRay rtc_ray;
     kernel_embree_setup_ray(*ray, rtc_ray, visibility);

intern/cycles/kernel/bvh/embree.h

@@ -22,6 +22,8 @@
 #include "kernel/device/cpu/compat.h"
 #include "kernel/device/cpu/globals.h"
 
+#include "kernel/bvh/util.h"
+
 #include "util/vector.h"
 
 CCL_NAMESPACE_BEGIN
@@ -38,6 +40,9 @@ struct CCLIntersectContext {
   KernelGlobals kg;
   RayType type;
 
+  /* For avoiding self intersections */
+  const Ray *ray;
+
   /* for shadow rays */
   Intersection *isect_s;
   uint max_hits;
@@ -56,6 +61,7 @@ struct CCLIntersectContext {
   {
     kg = kg_;
     type = type_;
+    ray = NULL;
     max_hits = 1;
     num_hits = 0;
     num_recorded_hits = 0;
@@ -102,7 +108,34 @@ ccl_device_inline void kernel_embree_setup_rayhit(const Ray &ray,
 {
   kernel_embree_setup_ray(ray, rayhit.ray, visibility);
   rayhit.hit.geomID = RTC_INVALID_GEOMETRY_ID;
-  rayhit.hit.primID = RTC_INVALID_GEOMETRY_ID;
+  rayhit.hit.instID[0] = RTC_INVALID_GEOMETRY_ID;
+}
+
+ccl_device_inline bool kernel_embree_is_self_intersection(const KernelGlobals kg,
+                                                          const RTCHit *hit,
+                                                          const Ray *ray)
+{
+  bool status = false;
+  if (hit->instID[0] != RTC_INVALID_GEOMETRY_ID) {
+    const int oID = hit->instID[0] / 2;
+    if ((ray->self.object == oID) || (ray->self.light_object == oID)) {
+      RTCScene inst_scene = (RTCScene)rtcGetGeometryUserData(
+          rtcGetGeometry(kernel_data.bvh.scene, hit->instID[0]));
+      const int pID = hit->primID +
+                      (intptr_t)rtcGetGeometryUserData(rtcGetGeometry(inst_scene, hit->geomID));
+      status = intersection_skip_self_shadow(ray->self, oID, pID);
+    }
+  }
+  else {
+    const int oID = hit->geomID / 2;
+    if ((ray->self.object == oID) || (ray->self.light_object == oID)) {
+      const int pID = hit->primID + (intptr_t)rtcGetGeometryUserData(
+                                        rtcGetGeometry(kernel_data.bvh.scene, hit->geomID));
+      status = intersection_skip_self_shadow(ray->self, oID, pID);
+    }
+  }
+
+  return status;
 }
 
 ccl_device_inline void kernel_embree_convert_hit(KernelGlobals kg,

intern/cycles/kernel/bvh/local.h

@@ -148,12 +148,27 @@ ccl_device_inline
             /* intersect ray against primitive */
             for (; prim_addr < prim_addr2; prim_addr++) {
               kernel_assert(kernel_tex_fetch(__prim_type, prim_addr) == type);
+
+              /* Only intersect with matching object, for instanced objects we
+               * already know we are only intersecting the right object. */
+              if (object == OBJECT_NONE) {
+                if (kernel_tex_fetch(__prim_object, prim_addr) != local_object) {
+                  continue;
+                }
+              }
+
+              /* Skip self intersection. */
+              const int prim = kernel_tex_fetch(__prim_index, prim_addr);
+              if (intersection_skip_self_local(ray->self, prim)) {
+                continue;
+              }
+
               if (triangle_intersect_local(kg,
                                            local_isect,
                                            P,
                                            dir,
-                                           object,
                                            local_object,
+                                           prim,
                                            prim_addr,
                                            isect_t,
                                            lcg_state,
@@ -168,13 +183,28 @@ ccl_device_inline
             /* intersect ray against primitive */
             for (; prim_addr < prim_addr2; prim_addr++) {
               kernel_assert(kernel_tex_fetch(__prim_type, prim_addr) == type);
+
+              /* Only intersect with matching object, for instanced objects we
+               * already know we are only intersecting the right object. */
+              if (object == OBJECT_NONE) {
+                if (kernel_tex_fetch(__prim_object, prim_addr) != local_object) {
+                  continue;
+                }
+              }
+
+              /* Skip self intersection. */
+              const int prim = kernel_tex_fetch(__prim_index, prim_addr);
+              if (intersection_skip_self_local(ray->self, prim)) {
+                continue;
+              }
+
               if (motion_triangle_intersect_local(kg,
                                                   local_isect,
                                                   P,
                                                   dir,
                                                   ray->time,
-                                                  object,
                                                   local_object,
+                                                  prim,
                                                   prim_addr,
                                                   isect_t,
                                                   lcg_state,

intern/cycles/kernel/bvh/metal.h

@@ -15,6 +15,7 @@
  */
 
 struct MetalRTIntersectionPayload {
+  RaySelfPrimitives self;
   uint visibility;
   float u, v;
   int prim;
@@ -25,6 +26,7 @@ struct MetalRTIntersectionPayload {
 };
 
 struct MetalRTIntersectionLocalPayload {
+  RaySelfPrimitives self;
   uint local_object;
   uint lcg_state;
   short max_hits;
@@ -34,6 +36,7 @@ struct MetalRTIntersectionLocalPayload {
 };
 
 struct MetalRTIntersectionShadowPayload {
+  RaySelfPrimitives self;
   uint visibility;
 #if defined(__METALRT_MOTION__)
   float time;

intern/cycles/kernel/bvh/shadow_all.h

@@ -146,7 +146,7 @@ ccl_device_inline
           --stack_ptr;
 
           /* primitive intersection */
-          while (prim_addr < prim_addr2) {
+          for (; prim_addr < prim_addr2; prim_addr++) {
             kernel_assert((kernel_tex_fetch(__prim_type, prim_addr) & PRIMITIVE_ALL) ==
                           (type & PRIMITIVE_ALL));
             bool hit;
@@ -156,16 +156,32 @@ ccl_device_inline
              * might give a few % performance improvement */
             Intersection isect ccl_optional_struct_init;
 
+            const int prim_object = (object == OBJECT_NONE) ?
+                                        kernel_tex_fetch(__prim_object, prim_addr) :
+                                        object;
+            const int prim = kernel_tex_fetch(__prim_index, prim_addr);
+            if (intersection_skip_self_shadow(ray->self, prim_object, prim)) {
+              continue;
+            }
+
             switch (type & PRIMITIVE_ALL) {
               case PRIMITIVE_TRIANGLE: {
                 hit = triangle_intersect(
-                    kg, &isect, P, dir, t_max_current, visibility, object, prim_addr);
+                    kg, &isect, P, dir, t_max_current, visibility, prim_object, prim, prim_addr);
                 break;
               }
 #if BVH_FEATURE(BVH_MOTION)
               case PRIMITIVE_MOTION_TRIANGLE: {
-                hit = motion_triangle_intersect(
+                hit = motion_triangle_intersect(kg,
-                    kg, &isect, P, dir, t_max_current, ray->time, visibility, object, prim_addr);
+                                                &isect,
+                                                P,
+                                                dir,
+                                                t_max_current,
+                                                ray->time,
+                                                visibility,
+                                                prim_object,
+                                                prim,
+                                                prim_addr);
                 break;
               }
 #endif
@@ -182,20 +198,9 @@ ccl_device_inline
                   }
                 }
 
-                const int curve_object = (object == OBJECT_NONE) ?
-                                             kernel_tex_fetch(__prim_object, prim_addr) :
-                                             object;
                 const int curve_type = kernel_tex_fetch(__prim_type, prim_addr);
-                const int curve_prim = kernel_tex_fetch(__prim_index, prim_addr);
+                hit = curve_intersect(
-                hit = curve_intersect(kg,
+                    kg, &isect, P, dir, t_max_current, prim_object, prim, ray->time, curve_type);
-                                      &isect,
-                                      P,
-                                      dir,
-                                      t_max_current,
-                                      curve_object,
-                                      curve_prim,
-                                      ray->time,
-                                      curve_type);
 
                 break;
               }
@@ -211,20 +216,9 @@ ccl_device_inline
                   }
                 }
 
-                const int point_object = (object == OBJECT_NONE) ?
-                                             kernel_tex_fetch(__prim_object, prim_addr) :
-                                             object;
-                const int point_prim = kernel_tex_fetch(__prim_index, prim_addr);
                 const int point_type = kernel_tex_fetch(__prim_type, prim_addr);
-                hit = point_intersect(kg,
+                hit = point_intersect(
-                                      &isect,
+                    kg, &isect, P, dir, t_max_current, prim_object, prim, ray->time, point_type);
-                                      P,
-                                      dir,
-                                      t_max_current,
-                                      point_object,
-                                      point_prim,
-                                      ray->time,
-                                      point_type);
                 break;
               }
 #endif /* BVH_FEATURE(BVH_POINTCLOUD) */
@@ -301,8 +295,6 @@ ccl_device_inline
                 integrator_state_write_shadow_isect(state, &isect, record_index);
               }
             }
-
-            prim_addr++;
           }
         }
         else {

intern/cycles/kernel/bvh/traversal.h

@@ -133,30 +133,43 @@ ccl_device_noinline bool BVH_FUNCTION_FULL_NAME(BVH)(KernelGlobals kg,
           --stack_ptr;
 
           /* primitive intersection */
-          switch (type & PRIMITIVE_ALL) {
-            case PRIMITIVE_TRIANGLE: {
           for (; prim_addr < prim_addr2; prim_addr++) {
             kernel_assert(kernel_tex_fetch(__prim_type, prim_addr) == type);
+
+            const int prim_object = (object == OBJECT_NONE) ?
+                                        kernel_tex_fetch(__prim_object, prim_addr) :
+                                        object;
+            const int prim = kernel_tex_fetch(__prim_index, prim_addr);
+            if (intersection_skip_self_shadow(ray->self, prim_object, prim)) {
+              continue;
+            }
+
+            switch (type & PRIMITIVE_ALL) {
+              case PRIMITIVE_TRIANGLE: {
                 if (triangle_intersect(
-                        kg, isect, P, dir, isect->t, visibility, object, prim_addr)) {
+                        kg, isect, P, dir, isect->t, visibility, prim_object, prim, prim_addr)) {
                   /* shadow ray early termination */
                   if (visibility & PATH_RAY_SHADOW_OPAQUE)
                     return true;
                 }
-              }
                 break;
               }
 #if BVH_FEATURE(BVH_MOTION)
               case PRIMITIVE_MOTION_TRIANGLE: {
-              for (; prim_addr < prim_addr2; prim_addr++) {
+                if (motion_triangle_intersect(kg,
-                kernel_assert(kernel_tex_fetch(__prim_type, prim_addr) == type);
+                                              isect,
-                if (motion_triangle_intersect(
+                                              P,
-                        kg, isect, P, dir, isect->t, ray->time, visibility, object, prim_addr)) {
+                                              dir,
+                                              isect->t,
+                                              ray->time,
+                                              visibility,
+                                              prim_object,
+                                              prim,
+                                              prim_addr)) {
                   /* shadow ray early termination */
                   if (visibility & PATH_RAY_SHADOW_OPAQUE)
                     return true;
                 }
-              }
                 break;
               }
 #endif /* BVH_FEATURE(BVH_MOTION) */
@@ -165,59 +178,48 @@ ccl_device_noinline bool BVH_FUNCTION_FULL_NAME(BVH)(KernelGlobals kg,
               case PRIMITIVE_MOTION_CURVE_THICK:
               case PRIMITIVE_CURVE_RIBBON:
               case PRIMITIVE_MOTION_CURVE_RIBBON: {
-              for (; prim_addr < prim_addr2; prim_addr++) {
                 if ((type & PRIMITIVE_MOTION) && kernel_data.bvh.use_bvh_steps) {
                   const float2 prim_time = kernel_tex_fetch(__prim_time, prim_addr);
                   if (ray->time < prim_time.x || ray->time > prim_time.y) {
-                    continue;
+                    break;
                   }
                 }
 
-                const int curve_object = (object == OBJECT_NONE) ?
-                                             kernel_tex_fetch(__prim_object, prim_addr) :
-                                             object;
-                const int curve_prim = kernel_tex_fetch(__prim_index, prim_addr);
                 const int curve_type = kernel_tex_fetch(__prim_type, prim_addr);
                 const bool hit = curve_intersect(
-                    kg, isect, P, dir, isect->t, curve_object, curve_prim, ray->time, curve_type);
+                    kg, isect, P, dir, isect->t, prim_object, prim, ray->time, curve_type);
                 if (hit) {
                   /* shadow ray early termination */
                   if (visibility & PATH_RAY_SHADOW_OPAQUE)
                     return true;
                 }
-              }
                 break;
               }
 #endif /* BVH_FEATURE(BVH_HAIR) */
 #if BVH_FEATURE(BVH_POINTCLOUD)
               case PRIMITIVE_POINT:
               case PRIMITIVE_MOTION_POINT: {
-              for (; prim_addr < prim_addr2; prim_addr++) {
                 if ((type & PRIMITIVE_MOTION) && kernel_data.bvh.use_bvh_steps) {
                   const float2 prim_time = kernel_tex_fetch(__prim_time, prim_addr);
                   if (ray->time < prim_time.x || ray->time > prim_time.y) {
-                    continue;
+                    break;
                   }
                 }
 
-                const int point_object = (object == OBJECT_NONE) ?
-                                             kernel_tex_fetch(__prim_object, prim_addr) :
-                                             object;
-                const int point_prim = kernel_tex_fetch(__prim_index, prim_addr);
                 const int point_type = kernel_tex_fetch(__prim_type, prim_addr);
                 const bool hit = point_intersect(
-                    kg, isect, P, dir, isect->t, point_object, point_prim, ray->time, point_type);
+                    kg, isect, P, dir, isect->t, prim_object, prim, ray->time, point_type);
                 if (hit) {
                   /* shadow ray early termination */
                   if (visibility & PATH_RAY_SHADOW_OPAQUE)
                     return true;
                 }
-              }
                 break;
               }
 #endif /* BVH_FEATURE(BVH_POINTCLOUD) */
             }
           }
+        }
         else {
           /* instance push */
           object = kernel_tex_fetch(__prim_object, -prim_addr - 1);

intern/cycles/kernel/bvh/util.h

@@ -21,54 +21,22 @@ CCL_NAMESPACE_BEGIN
 /* Ray offset to avoid self intersection.
  *
  * This function should be used to compute a modified ray start position for
- * rays leaving from a surface. */
+ * rays leaving from a surface. This is from "A Fast and Robust Method for Avoiding
-
+ * Self-Intersection" see https://research.nvidia.com/publication/2019-03_A-Fast-and
+ */
 ccl_device_inline float3 ray_offset(float3 P, float3 Ng)
 {
-#ifdef __INTERSECTION_REFINE__
+  const float int_scale = 256.0f;
-  const float epsilon_f = 1e-5f;
+  int3 of_i = make_int3((int)(int_scale * Ng.x), (int)(int_scale * Ng.y), (int)(int_scale * Ng.z));
-  /* ideally this should match epsilon_f, but instancing and motion blur
-   * precision makes it problematic */
-  const float epsilon_test = 1.0f;
-  const int epsilon_i = 32;
 
-  float3 res;
+  float3 p_i = make_float3(__int_as_float(__float_as_int(P.x) + ((P.x < 0) ? -of_i.x : of_i.x)),
-
+                           __int_as_float(__float_as_int(P.y) + ((P.y < 0) ? -of_i.y : of_i.y)),
-  /* x component */
+                           __int_as_float(__float_as_int(P.z) + ((P.z < 0) ? -of_i.z : of_i.z)));
-  if (fabsf(P.x) < epsilon_test) {
+  const float origin = 1.0f / 32.0f;
-    res.x = P.x + Ng.x * epsilon_f;
+  const float float_scale = 1.0f / 65536.0f;
-  }
+  return make_float3(fabsf(P.x) < origin ? P.x + float_scale * Ng.x : p_i.x,
-  else {
+                     fabsf(P.y) < origin ? P.y + float_scale * Ng.y : p_i.y,
-    uint ix = __float_as_uint(P.x);
+                     fabsf(P.z) < origin ? P.z + float_scale * Ng.z : p_i.z);
-    ix += ((ix ^ __float_as_uint(Ng.x)) >> 31) ? -epsilon_i : epsilon_i;
-    res.x = __uint_as_float(ix);
-  }
-
-  /* y component */
-  if (fabsf(P.y) < epsilon_test) {
-    res.y = P.y + Ng.y * epsilon_f;
-  }
-  else {
-    uint iy = __float_as_uint(P.y);
-    iy += ((iy ^ __float_as_uint(Ng.y)) >> 31) ? -epsilon_i : epsilon_i;
-    res.y = __uint_as_float(iy);
-  }
-
-  /* z component */
-  if (fabsf(P.z) < epsilon_test) {
-    res.z = P.z + Ng.z * epsilon_f;
-  }
-  else {
-    uint iz = __float_as_uint(P.z);
-    iz += ((iz ^ __float_as_uint(Ng.z)) >> 31) ? -epsilon_i : epsilon_i;
-    res.z = __uint_as_float(iz);
-  }
-
-  return res;
-#else
-  const float epsilon_f = 1e-4f;
-  return P + epsilon_f * Ng;
-#endif
 }
 
 #if defined(__KERNEL_CPU__)
@@ -227,4 +195,25 @@ ccl_device_inline float intersection_curve_shadow_transparency(KernelGlobals kg,
   return (1.0f - u) * f0 + u * f1;
 }
 
+ccl_device_inline bool intersection_skip_self(ccl_private const RaySelfPrimitives &self,
+                                              const int object,
+                                              const int prim)
+{
+  return (self.prim == prim) && (self.object == object);
+}
+
+ccl_device_inline bool intersection_skip_self_shadow(ccl_private const RaySelfPrimitives &self,
+                                                     const int object,
+                                                     const int prim)
+{
+  return ((self.prim == prim) && (self.object == object)) ||
+         ((self.light_prim == prim) && (self.light_object == object));
+}
+
+ccl_device_inline bool intersection_skip_self_local(ccl_private const RaySelfPrimitives &self,
+                                                    const int prim)
+{
+  return (self.prim == prim);
+}
+
 CCL_NAMESPACE_END

intern/cycles/kernel/bvh/volume.h

@@ -140,14 +140,20 @@ ccl_device_inline
               for (; prim_addr < prim_addr2; prim_addr++) {
                 kernel_assert(kernel_tex_fetch(__prim_type, prim_addr) == type);
                 /* only primitives from volume object */
-                uint tri_object = (object == OBJECT_NONE) ?
+                const int prim_object = (object == OBJECT_NONE) ?
                                             kernel_tex_fetch(__prim_object, prim_addr) :
                                             object;
-                int object_flag = kernel_tex_fetch(__object_flag, tri_object);
+                const int prim = kernel_tex_fetch(__prim_index, prim_addr);
+                if (intersection_skip_self(ray->self, prim_object, prim)) {
+                  continue;
+                }
+
+                int object_flag = kernel_tex_fetch(__object_flag, prim_object);
                 if ((object_flag & SD_OBJECT_HAS_VOLUME) == 0) {
                   continue;
                 }
-                triangle_intersect(kg, isect, P, dir, isect->t, visibility, object, prim_addr);
+                triangle_intersect(
+                    kg, isect, P, dir, isect->t, visibility, prim_object, prim, prim_addr);
               }
               break;
             }
@@ -157,15 +163,27 @@ ccl_device_inline
               for (; prim_addr < prim_addr2; prim_addr++) {
                 kernel_assert(kernel_tex_fetch(__prim_type, prim_addr) == type);
                 /* only primitives from volume object */
-                uint tri_object = (object == OBJECT_NONE) ?
+                const int prim_object = (object == OBJECT_NONE) ?
                                             kernel_tex_fetch(__prim_object, prim_addr) :
                                             object;
-                int object_flag = kernel_tex_fetch(__object_flag, tri_object);
+                const int prim = kernel_tex_fetch(__prim_index, prim_addr);
+                if (intersection_skip_self(ray->self, prim_object, prim)) {
+                  continue;
+                }
+                int object_flag = kernel_tex_fetch(__object_flag, prim_object);
                 if ((object_flag & SD_OBJECT_HAS_VOLUME) == 0) {
                   continue;
                 }
-                motion_triangle_intersect(
+                motion_triangle_intersect(kg,
-                    kg, isect, P, dir, isect->t, ray->time, visibility, object, prim_addr);
+                                          isect,
+                                          P,
+                                          dir,
+                                          isect->t,
+                                          ray->time,
+                                          visibility,
+                                          prim_object,
+                                          prim,
+                                          prim_addr);
               }
               break;
             }

intern/cycles/kernel/bvh/volume_all.h

@@ -143,15 +143,19 @@ ccl_device_inline
               for (; prim_addr < prim_addr2; prim_addr++) {
                 kernel_assert(kernel_tex_fetch(__prim_type, prim_addr) == type);
                 /* only primitives from volume object */
-                uint tri_object = (object == OBJECT_NONE) ?
+                const int prim_object = (object == OBJECT_NONE) ?
                                             kernel_tex_fetch(__prim_object, prim_addr) :
                                             object;
-                int object_flag = kernel_tex_fetch(__object_flag, tri_object);
+                const int prim = kernel_tex_fetch(__prim_index, prim_addr);
+                if (intersection_skip_self(ray->self, prim_object, prim)) {
+                  continue;
+                }
+                int object_flag = kernel_tex_fetch(__object_flag, prim_object);
                 if ((object_flag & SD_OBJECT_HAS_VOLUME) == 0) {
                   continue;
                 }
                 hit = triangle_intersect(
-                    kg, isect_array, P, dir, isect_t, visibility, object, prim_addr);
+                    kg, isect_array, P, dir, isect_t, visibility, prim_object, prim, prim_addr);
                 if (hit) {
                   /* Move on to next entry in intersections array. */
                   isect_array++;
@@ -183,15 +187,27 @@ ccl_device_inline
               for (; prim_addr < prim_addr2; prim_addr++) {
                 kernel_assert(kernel_tex_fetch(__prim_type, prim_addr) == type);
                 /* only primitives from volume object */
-                uint tri_object = (object == OBJECT_NONE) ?
+                const int prim_object = (object == OBJECT_NONE) ?
                                             kernel_tex_fetch(__prim_object, prim_addr) :
                                             object;
-                int object_flag = kernel_tex_fetch(__object_flag, tri_object);
+                const int prim = kernel_tex_fetch(__prim_index, prim_addr);
+                if (intersection_skip_self(ray->self, prim_object, prim)) {
+                  continue;
+                }
+                int object_flag = kernel_tex_fetch(__object_flag, prim_object);
                 if ((object_flag & SD_OBJECT_HAS_VOLUME) == 0) {
                   continue;
                 }
-                hit = motion_triangle_intersect(
+                hit = motion_triangle_intersect(kg,
-                    kg, isect_array, P, dir, isect_t, ray->time, visibility, object, prim_addr);
+                                                isect_array,
+                                                P,
+                                                dir,
+                                                isect_t,
+                                                ray->time,
+                                                visibility,
+                                                prim_object,
+                                                prim,
+                                                prim_addr);
                 if (hit) {
                   /* Move on to next entry in intersections array. */
                   isect_array++;

intern/cycles/kernel/closure/volume.h

@@ -199,22 +199,18 @@ ccl_device int volume_sample_channel(float3 albedo,
    *  Tracing". Matt Jen-Yuan Chiang, Peter Kutz, Brent Burley. SIGGRAPH 2016. */
   float3 weights = fabs(throughput * albedo);
   float sum_weights = weights.x + weights.y + weights.z;
-  float3 weights_pdf;
 
   if (sum_weights > 0.0f) {
-    weights_pdf = weights / sum_weights;
+    *pdf = weights / sum_weights;
   }
   else {
-    weights_pdf = make_float3(1.0f / 3.0f, 1.0f / 3.0f, 1.0f / 3.0f);
+    *pdf = make_float3(1.0f / 3.0f, 1.0f / 3.0f, 1.0f / 3.0f);
   }
 
-  *pdf = weights_pdf;
+  if (rand < pdf->x) {
-
-  /* OpenCL does not support -> on float3, so don't use pdf->x. */
-  if (rand < weights_pdf.x) {
     return 0;
   }
-  else if (rand < weights_pdf.x + weights_pdf.y) {
+  else if (rand < pdf->x + pdf->y) {
     return 1;
   }
   else {

intern/cycles/kernel/device/gpu/kernel.h

@@ -243,6 +243,10 @@ ccl_gpu_kernel(GPU_KERNEL_BLOCK_NUM_THREADS, GPU_KERNEL_MAX_REGISTERS)
   }
 }
 
+#ifdef __KERNEL_METAL__
+constant int __dummy_constant [[function_constant(0)]];
+#endif
+
 ccl_gpu_kernel(GPU_KERNEL_BLOCK_NUM_THREADS, GPU_KERNEL_MAX_REGISTERS)
     ccl_gpu_kernel_signature(integrator_shade_surface_raytrace,
                              ccl_global const int *path_index_array,
@@ -253,7 +257,16 @@ ccl_gpu_kernel(GPU_KERNEL_BLOCK_NUM_THREADS, GPU_KERNEL_MAX_REGISTERS)
 
   if (global_index < work_size) {
     const int state = (path_index_array) ? path_index_array[global_index] : global_index;
+
+#ifdef __KERNEL_METAL__
+    KernelGlobals kg = NULL;
+    /* Workaround Ambient Occlusion and Bevel nodes not working with Metal.
+     * Dummy offset should not affect result, but somehow fixes bug! */
+    kg += __dummy_constant;
+    ccl_gpu_kernel_call(integrator_shade_surface_raytrace(kg, state, render_buffer));
+#else
     ccl_gpu_kernel_call(integrator_shade_surface_raytrace(NULL, state, render_buffer));
+#endif
   }
 }
 
@@ -821,8 +834,8 @@ ccl_gpu_kernel(GPU_KERNEL_BLOCK_NUM_THREADS, GPU_KERNEL_MAX_REGISTERS)
   if (guiding_pass_flow != PASS_UNUSED) {
     kernel_assert(render_pass_motion != PASS_UNUSED);
 
-    const float *motion_in = buffer + render_pass_motion;
+    ccl_global const float *motion_in = buffer + render_pass_motion;
-    float *flow_out = guiding_pixel + guiding_pass_flow;
+    ccl_global float *flow_out = guiding_pixel + guiding_pass_flow;
 
     flow_out[0] = -motion_in[0] * pixel_scale;
     flow_out[1] = -motion_in[1] * pixel_scale;

intern/cycles/kernel/device/metal/compat.h

@@ -98,8 +98,12 @@ using namespace metal::raytracing;
 #define FN14(p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11, p12, p13, p14) p1; p2; p3; p4; p5; p6; p7; p8; p9; p10; p11; p12; p13; p14;
 #define FN15(p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11, p12, p13, p14, p15) p1; p2; p3; p4; p5; p6; p7; p8; p9; p10; p11; p12; p13; p14; p15;
 #define FN16(p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11, p12, p13, p14, p15, p16) p1; p2; p3; p4; p5; p6; p7; p8; p9; p10; p11; p12; p13; p14; p15; p16;
-#define GET_LAST_ARG(p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11, p12, p13, p14, p15, p16, ...) p16
+#define FN17(p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11, p12, p13, p14, p15, p16, p17) p1; p2; p3; p4; p5; p6; p7; p8; p9; p10; p11; p12; p13; p14; p15; p16; p17;
-#define PARAMS_MAKER(...) GET_LAST_ARG(__VA_ARGS__, FN16, FN15, FN14, FN13, FN12, FN11, FN10, FN9, FN8, FN7, FN6, FN5, FN4, FN3, FN2, FN1, FN0)
+#define FN18(p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11, p12, p13, p14, p15, p16, p17, p18) p1; p2; p3; p4; p5; p6; p7; p8; p9; p10; p11; p12; p13; p14; p15; p16; p17; p18;
+#define FN19(p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11, p12, p13, p14, p15, p16, p17, p18, p19) p1; p2; p3; p4; p5; p6; p7; p8; p9; p10; p11; p12; p13; p14; p15; p16; p17; p18; p19;
+#define FN20(p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11, p12, p13, p14, p15, p16, p17, p18, p19, p20) p1; p2; p3; p4; p5; p6; p7; p8; p9; p10; p11; p12; p13; p14; p15; p16; p17; p18; p19; p20;
+#define GET_LAST_ARG(p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11, p12, p13, p14, p15, p16, p17, p18, p19, p20, ...) p20
+#define PARAMS_MAKER(...) GET_LAST_ARG(__VA_ARGS__, FN20, FN19, FN18, FN17, FN16, FN15, FN14, FN13, FN12, FN11, FN10, FN9, FN8, FN7, FN6, FN5, FN4, FN3, FN2, FN1, FN0)
 
 /* Generate a struct containing the entry-point parameters and a "run"
  * method which can access them implicitly via this-> */

intern/cycles/kernel/device/metal/kernel.metal

@@ -40,6 +40,27 @@ struct TriangleIntersectionResult
 
 enum { METALRT_HIT_TRIANGLE, METALRT_HIT_BOUNDING_BOX };
 
+ccl_device_inline bool intersection_skip_self(ray_data const RaySelfPrimitives& self,
+                                              const int object,
+                                              const int prim)
+{
+  return (self.prim == prim) && (self.object == object);
+}
+
+ccl_device_inline bool intersection_skip_self_shadow(ray_data const RaySelfPrimitives& self,
+                                                     const int object,
+                                                     const int prim)
+{
+  return ((self.prim == prim) && (self.object == object)) ||
+         ((self.light_prim == prim) && (self.light_object == object));
+}
+
+ccl_device_inline bool intersection_skip_self_local(ray_data const RaySelfPrimitives& self,
+                                                    const int prim)
+{
+  return (self.prim == prim);
+}
+
 template<typename TReturn, uint intersection_type>
 TReturn metalrt_local_hit(constant KernelParamsMetal &launch_params_metal,
                           ray_data MetalKernelContext::MetalRTIntersectionLocalPayload &payload,
@@ -53,8 +74,8 @@ TReturn metalrt_local_hit(constant KernelParamsMetal &launch_params_metal,
 #ifdef __BVH_LOCAL__
   uint prim = primitive_id + kernel_tex_fetch(__object_prim_offset, object);
 
-  if (object != payload.local_object) {
+  if ((object != payload.local_object) || intersection_skip_self_local(payload.self, prim)) {
-    /* Only intersect with matching object */
+    /* Only intersect with matching object and skip self-intersecton. */
     result.accept = false;
     result.continue_search = true;
     return result;
@@ -166,6 +187,11 @@ bool metalrt_shadow_all_hit(constant KernelParamsMetal &launch_params_metal,
   }
 #  endif
 
+  if (intersection_skip_self_shadow(payload.self, object, prim)) {
+    /* continue search */
+    return true;
+  }
+
   float u = 0.0f, v = 0.0f;
   int type = 0;
   if (intersection_type == METALRT_HIT_TRIANGLE) {
@@ -322,21 +348,35 @@ inline TReturnType metalrt_visibility_test(constant KernelParamsMetal &launch_pa
   }
 #  endif
 
-#  ifdef __VISIBILITY_FLAG__
   uint visibility = payload.visibility;
+#  ifdef __VISIBILITY_FLAG__
   if ((kernel_tex_fetch(__objects, object).visibility & visibility) == 0) {
     result.accept = false;
     result.continue_search = true;
     return result;
   }
+#  endif
 
   /* Shadow ray early termination. */
   if (visibility & PATH_RAY_SHADOW_OPAQUE) {
+    if (intersection_skip_self_shadow(payload.self, object, prim)) {
+      result.accept = false;
+      result.continue_search = true;
+      return result;
+    }
+    else {
       result.accept = true;
       result.continue_search = false;
       return result;
     }
-#  endif
+  }
+  else {
+    if (intersection_skip_self(payload.self, object, prim)) {
+      result.accept = false;
+      result.continue_search = true;
+      return result;
+    }
+  }
 
   result.accept = true;
   result.continue_search = true;
@@ -576,6 +616,150 @@ __intersection__curve_all_shadow(constant KernelParamsMetal &launch_params_metal
 
   return result;
 }
-
 #endif /* __HAIR__ */
+
+#ifdef __POINTCLOUD__
+ccl_device_inline
+void metalrt_intersection_point(constant KernelParamsMetal &launch_params_metal,
+                                ray_data MetalKernelContext::MetalRTIntersectionPayload &payload,
+                                const uint object,
+                                const uint prim,
+                                const uint type,
+                                const float3 ray_origin,
+                                const float3 ray_direction,
+                                float time,
+                                const float ray_tmax,
+                                thread BoundingBoxIntersectionResult &result)
+{
+#  ifdef __VISIBILITY_FLAG__
+  const uint visibility = payload.visibility;
+  if ((kernel_tex_fetch(__objects, object).visibility & visibility) == 0) {
+    return;
+  }
+#  endif
+
+  float3 P = ray_origin;
+  float3 dir = ray_direction;
+
+  /* The direction is not normalized by default, but the point intersection routine expects that */
+  float len;
+  dir = normalize_len(dir, &len);
+
+  Intersection isect;
+  isect.t = ray_tmax;
+  /* Transform maximum distance into object space. */
+  if (isect.t != FLT_MAX)
+    isect.t *= len;
+
+  MetalKernelContext context(launch_params_metal);
+  if (context.point_intersect(NULL, &isect, P, dir, isect.t, object, prim, time, type)) {
+    result = metalrt_visibility_test<BoundingBoxIntersectionResult, METALRT_HIT_BOUNDING_BOX>(
+                  launch_params_metal, payload, object, prim, isect.u);
+    if (result.accept) {
+      result.distance = isect.t / len;
+      payload.u = isect.u;
+      payload.v = isect.v;
+      payload.prim = prim;
+      payload.type = type;
+    }
+  }
+}
+
+ccl_device_inline
+void metalrt_intersection_point_shadow(constant KernelParamsMetal &launch_params_metal,
+                                       ray_data MetalKernelContext::MetalRTIntersectionShadowPayload &payload,
+                                       const uint object,
+                                       const uint prim,
+                                       const uint type,
+                                       const float3 ray_origin,
+                                       const float3 ray_direction,
+                                       float time,
+                                       const float ray_tmax,
+                                       thread BoundingBoxIntersectionResult &result)
+{
+  const uint visibility = payload.visibility;
+
+  float3 P = ray_origin;
+  float3 dir = ray_direction;
+
+  /* The direction is not normalized by default, but the point intersection routine expects that */
+  float len;
+  dir = normalize_len(dir, &len);
+
+  Intersection isect;
+  isect.t = ray_tmax;
+  /* Transform maximum distance into object space */
+  if (isect.t != FLT_MAX)
+    isect.t *= len;
+
+  MetalKernelContext context(launch_params_metal);
+  if (context.point_intersect(NULL, &isect, P, dir, isect.t, object, prim, time, type)) {
+    result.continue_search = metalrt_shadow_all_hit<METALRT_HIT_BOUNDING_BOX>(
+                launch_params_metal, payload, object, prim, float2(isect.u, isect.v), ray_tmax);
+    result.accept = !result.continue_search;
+
+    if (result.accept) {
+      result.distance = isect.t / len;
+    }
+  }
+}
+
+[[intersection(bounding_box, triangle_data, METALRT_TAGS)]]
+BoundingBoxIntersectionResult
+__intersection__point(constant KernelParamsMetal &launch_params_metal [[buffer(1)]],
+                             ray_data MetalKernelContext::MetalRTIntersectionPayload &payload [[payload]],
+                             const uint object [[user_instance_id]],
+                             const uint primitive_id [[primitive_id]],
+                             const float3 ray_origin [[origin]],
+                             const float3 ray_direction [[direction]],
+                             const float ray_tmax [[max_distance]])
+{
+  const uint prim = primitive_id + kernel_tex_fetch(__object_prim_offset, object);
+  const int type = kernel_tex_fetch(__objects, object).primitive_type;
+
+  BoundingBoxIntersectionResult result;
+  result.accept = false;
+  result.continue_search = true;
+  result.distance = ray_tmax;
+
+  metalrt_intersection_point(launch_params_metal, payload, object, prim, type, ray_origin, ray_direction,
+#  if defined(__METALRT_MOTION__)
+                             payload.time,
+#  else
+                             0.0f,
+#  endif
+                             ray_tmax, result);
+
+  return result;
+}
+
+[[intersection(bounding_box, triangle_data, METALRT_TAGS)]]
+BoundingBoxIntersectionResult
+__intersection__point_shadow(constant KernelParamsMetal &launch_params_metal [[buffer(1)]],
+                                    ray_data MetalKernelContext::MetalRTIntersectionShadowPayload &payload [[payload]],
+                                    const uint object [[user_instance_id]],
+                                    const uint primitive_id [[primitive_id]],
+                                    const float3 ray_origin [[origin]],
+                                    const float3 ray_direction [[direction]],
+                                    const float ray_tmax [[max_distance]])
+{
+  const uint prim = primitive_id + kernel_tex_fetch(__object_prim_offset, object);
+  const int type = kernel_tex_fetch(__objects, object).primitive_type;
+
+  BoundingBoxIntersectionResult result;
+  result.accept = false;
+  result.continue_search = true;
+  result.distance = ray_tmax;
+
+  metalrt_intersection_point_shadow(launch_params_metal, payload, object, prim, type, ray_origin, ray_direction,
+#  if defined(__METALRT_MOTION__)
+                             payload.time,
+#  else
+                             0.0f,
+#  endif
+                             ray_tmax, result);
+
+  return result;
+}
+#endif /* __POINTCLOUD__ */
 #endif /* __METALRT__ */

intern/cycles/kernel/device/optix/kernel.cu

@@ -45,6 +45,11 @@ template<typename T> ccl_device_forceinline T *get_payload_ptr_2()
   return pointer_unpack_from_uint<T>(optixGetPayload_2(), optixGetPayload_3());
 }
 
+template<typename T> ccl_device_forceinline T *get_payload_ptr_6()
+{
+  return (T *)(((uint64_t)optixGetPayload_7() << 32) | optixGetPayload_6());
+}
+
 ccl_device_forceinline int get_object_id()
 {
 #ifdef __OBJECT_MOTION__
@@ -111,6 +116,12 @@ extern "C" __global__ void __anyhit__kernel_optix_local_hit()
     return optixIgnoreIntersection();
   }
 
+  const int prim = optixGetPrimitiveIndex();
+  ccl_private Ray *const ray = get_payload_ptr_6<Ray>();
+  if (intersection_skip_self_local(ray->self, prim)) {
+    return optixIgnoreIntersection();
+  }
+
   const uint max_hits = optixGetPayload_5();
   if (max_hits == 0) {
     /* Special case for when no hit information is requested, just report that something was hit */
@@ -149,8 +160,6 @@ extern "C" __global__ void __anyhit__kernel_optix_local_hit()
     local_isect->num_hits = 1;
   }
 
-  const int prim = optixGetPrimitiveIndex();
-
   Intersection *isect = &local_isect->hits[hit];
   isect->t = optixGetRayTmax();
   isect->prim = prim;
@@ -185,6 +194,11 @@ extern "C" __global__ void __anyhit__kernel_optix_shadow_all_hit()
   }
 #  endif
 
+  ccl_private Ray *const ray = get_payload_ptr_6<Ray>();
+  if (intersection_skip_self_shadow(ray->self, object, prim)) {
+    return optixIgnoreIntersection();
+  }
+
   float u = 0.0f, v = 0.0f;
   int type = 0;
   if (optixIsTriangleHit()) {
@@ -314,6 +328,12 @@ extern "C" __global__ void __anyhit__kernel_optix_volume_test()
   if ((kernel_tex_fetch(__object_flag, object) & SD_OBJECT_HAS_VOLUME) == 0) {
     return optixIgnoreIntersection();
   }
+
+  const int prim = optixGetPrimitiveIndex();
+  ccl_private Ray *const ray = get_payload_ptr_6<Ray>();
+  if (intersection_skip_self(ray->self, object, prim)) {
+    return optixIgnoreIntersection();
+  }
 }
 
 extern "C" __global__ void __anyhit__kernel_optix_visibility_test()
@@ -330,18 +350,31 @@ extern "C" __global__ void __anyhit__kernel_optix_visibility_test()
 #  endif
 #endif
 
-#ifdef __VISIBILITY_FLAG__
   const uint object = get_object_id();
   const uint visibility = optixGetPayload_4();
+#ifdef __VISIBILITY_FLAG__
   if ((kernel_tex_fetch(__objects, object).visibility & visibility) == 0) {
     return optixIgnoreIntersection();
   }
+#endif
+
+  const int prim = optixGetPrimitiveIndex();
+  ccl_private Ray *const ray = get_payload_ptr_6<Ray>();
 
-  /* Shadow ray early termination. */
   if (visibility & PATH_RAY_SHADOW_OPAQUE) {
+    if (intersection_skip_self_shadow(ray->self, object, prim)) {
+      return optixIgnoreIntersection();
+    }
+    else {
+      /* Shadow ray early termination. */
       return optixTerminateRay();
     }
-#endif
+  }
+  else {
+    if (intersection_skip_self(ray->self, object, prim)) {
+      return optixIgnoreIntersection();
+    }
+  }
 }
 
 extern "C" __global__ void __closesthit__kernel_optix_hit()

intern/cycles/kernel/film/passes.h

@@ -92,6 +92,14 @@ ccl_device_forceinline void kernel_write_denoising_features_surface(
     else if (sc->type == CLOSURE_BSDF_HAIR_PRINCIPLED_ID) {
       closure_albedo *= bsdf_principled_hair_albedo(sc);
     }
+    else if (sc->type == CLOSURE_BSDF_PRINCIPLED_DIFFUSE_ID) {
+      /* BSSRDF already accounts for weight, retro-reflection would double up. */
+      ccl_private const PrincipledDiffuseBsdf *bsdf = (ccl_private const PrincipledDiffuseBsdf *)
+          sc;
+      if (bsdf->components == PRINCIPLED_DIFFUSE_RETRO_REFLECTION) {
+        continue;
+      }
+    }
 
     if (bsdf_get_specular_roughness_squared(sc) > sqr(0.075f)) {
       diffuse_albedo += closure_albedo;

intern/cycles/kernel/film/read.h

@@ -214,6 +214,21 @@ ccl_device_inline void film_get_pass_pixel_light_path(
   pixel[0] = f.x;
   pixel[1] = f.y;
   pixel[2] = f.z;
+
+  /* Optional alpha channel. */
+  if (kfilm_convert->num_components >= 4) {
+    if (kfilm_convert->pass_combined != PASS_UNUSED) {
+      float scale, scale_exposure;
+      film_get_scale_and_scale_exposure(kfilm_convert, buffer, &scale, &scale_exposure);
+
+      ccl_global const float *in_combined = buffer + kfilm_convert->pass_combined;
+      const float alpha = in_combined[3] * scale;
+      pixel[3] = film_transparency_to_alpha(alpha);
+    }
+    else {
+      pixel[3] = 1.0f;
+    }
+  }
 }
 
 ccl_device_inline void film_get_pass_pixel_float3(ccl_global const KernelFilmConvert *ccl_restrict

intern/cycles/kernel/geom/curve.h

@@ -226,6 +226,18 @@ ccl_device float curve_thickness(KernelGlobals kg, ccl_private const ShaderData
   return r * 2.0f;
 }
 
+/* Curve random */
+
+ccl_device float curve_random(KernelGlobals kg, ccl_private const ShaderData *sd)
+{
+  if (sd->type & PRIMITIVE_CURVE) {
+    const AttributeDescriptor desc = find_attribute(kg, sd, ATTR_STD_CURVE_RANDOM);
+    return (desc.offset != ATTR_STD_NOT_FOUND) ? curve_attribute_float(kg, sd, desc, NULL, NULL) :
+                                                 0.0f;
+  }
+  return 0.0f;
+}
+
 /* Curve location for motion pass, linear interpolation between keys and
  * ignoring radius because we do the same for the motion keys */
 

intern/cycles/kernel/geom/motion_triangle.h

@@ -116,6 +116,52 @@ ccl_device_inline void motion_triangle_vertices(
   verts[2] = (1.0f - t) * verts[2] + t * next_verts[2];
 }
 
+ccl_device_inline void motion_triangle_vertices_and_normals(
+    KernelGlobals kg, int object, int prim, float time, float3 verts[3], float3 normals[3])
+{
+  /* get motion info */
+  int numsteps, numverts;
+  object_motion_info(kg, object, &numsteps, &numverts, NULL);
+
+  /* Figure out which steps we need to fetch and their interpolation factor. */
+  int maxstep = numsteps * 2;
+  int step = min((int)(time * maxstep), maxstep - 1);
+  float t = time * maxstep - step;
+
+  /* Find attribute. */
+  int offset = intersection_find_attribute(kg, object, ATTR_STD_MOTION_VERTEX_POSITION);
+  kernel_assert(offset != ATTR_STD_NOT_FOUND);
+
+  /* Fetch vertex coordinates. */
+  float3 next_verts[3];
+  uint4 tri_vindex = kernel_tex_fetch(__tri_vindex, prim);
+
+  motion_triangle_verts_for_step(kg, tri_vindex, offset, numverts, numsteps, step, verts);
+  motion_triangle_verts_for_step(kg, tri_vindex, offset, numverts, numsteps, step + 1, next_verts);
+
+  /* Interpolate between steps. */
+  verts[0] = (1.0f - t) * verts[0] + t * next_verts[0];
+  verts[1] = (1.0f - t) * verts[1] + t * next_verts[1];
+  verts[2] = (1.0f - t) * verts[2] + t * next_verts[2];
+
+  /* Compute smooth normal. */
+
+  /* Find attribute. */
+  offset = intersection_find_attribute(kg, object, ATTR_STD_MOTION_VERTEX_NORMAL);
+  kernel_assert(offset != ATTR_STD_NOT_FOUND);
+
+  /* Fetch vertex coordinates. */
+  float3 next_normals[3];
+  motion_triangle_normals_for_step(kg, tri_vindex, offset, numverts, numsteps, step, normals);
+  motion_triangle_normals_for_step(
+      kg, tri_vindex, offset, numverts, numsteps, step + 1, next_normals);
+
+  /* Interpolate between steps. */
+  normals[0] = (1.0f - t) * normals[0] + t * next_normals[0];
+  normals[1] = (1.0f - t) * normals[1] + t * next_normals[1];
+  normals[2] = (1.0f - t) * normals[2] + t * next_normals[2];
+}
+
 ccl_device_inline float3 motion_triangle_smooth_normal(
     KernelGlobals kg, float3 Ng, int object, int prim, float u, float v, float time)
 {

intern/cycles/kernel/geom/motion_triangle_intersect.h

@@ -29,46 +29,19 @@
 
 CCL_NAMESPACE_BEGIN
 
-/* Refine triangle intersection to more precise hit point. For rays that travel
+/**
- * far the precision is often not so good, this reintersects the primitive from
+ * Use the barycentric coordinates to get the intersection location
- * a closer distance.
  */
-
+ccl_device_inline float3 motion_triangle_point_from_uv(KernelGlobals kg,
-ccl_device_inline float3 motion_triangle_refine(KernelGlobals kg,
                                                        ccl_private ShaderData *sd,
-                                                float3 P,
-                                                float3 D,
-                                                float t,
                                                        const int isect_object,
                                                        const int isect_prim,
+                                                       const float u,
+                                                       const float v,
                                                        float3 verts[3])
 {
-#ifdef __INTERSECTION_REFINE__
+  float w = 1.0f - u - v;
-  if (!(sd->object_flag & SD_OBJECT_TRANSFORM_APPLIED)) {
+  float3 P = u * verts[0] + v * verts[1] + w * verts[2];
-    if (UNLIKELY(t == 0.0f)) {
-      return P;
-    }
-    const Transform tfm = object_get_inverse_transform(kg, sd);
-
-    P = transform_point(&tfm, P);
-    D = transform_direction(&tfm, D * t);
-    D = normalize_len(D, &t);
-  }
-
-  P = P + D * t;
-
-  /* Compute refined intersection distance. */
-  const float3 e1 = verts[0] - verts[2];
-  const float3 e2 = verts[1] - verts[2];
-  const float3 s1 = cross(D, e2);
-
-  const float invdivisor = 1.0f / dot(s1, e1);
-  const float3 d = P - verts[2];
-  const float3 s2 = cross(d, e1);
-  float rt = dot(e2, s2) * invdivisor;
-
-  /* Compute refined position. */
-  P = P + D * rt;
 
   if (!(sd->object_flag & SD_OBJECT_TRANSFORM_APPLIED)) {
     const Transform tfm = object_get_transform(kg, sd);
@@ -76,71 +49,8 @@ ccl_device_inline float3 motion_triangle_refine(KernelGlobals kg,
   }
 
   return P;
-#else
-  return P + D * t;
-#endif
 }
 
-/* Same as above, except that t is assumed to be in object space
- * for instancing.
- */
-
-#ifdef __BVH_LOCAL__
-#  if defined(__KERNEL_CUDA__) && (defined(i386) || defined(_M_IX86))
-ccl_device_noinline
-#  else
-ccl_device_inline
-#  endif
-    float3
-    motion_triangle_refine_local(KernelGlobals kg,
-                                 ccl_private ShaderData *sd,
-                                 float3 P,
-                                 float3 D,
-                                 float t,
-                                 const int isect_object,
-                                 const int isect_prim,
-                                 float3 verts[3])
-{
-#  if defined(__KERNEL_GPU_RAYTRACING__)
-  /* t is always in world space with OptiX and MetalRT. */
-  return motion_triangle_refine(kg, sd, P, D, t, isect_object, isect_prim, verts);
-#  else
-#    ifdef __INTERSECTION_REFINE__
-  if (!(sd->object_flag & SD_OBJECT_TRANSFORM_APPLIED)) {
-    const Transform tfm = object_get_inverse_transform(kg, sd);
-
-    P = transform_point(&tfm, P);
-    D = transform_direction(&tfm, D);
-    D = normalize(D);
-  }
-
-  P = P + D * t;
-
-  /* compute refined intersection distance */
-  const float3 e1 = verts[0] - verts[2];
-  const float3 e2 = verts[1] - verts[2];
-  const float3 s1 = cross(D, e2);
-
-  const float invdivisor = 1.0f / dot(s1, e1);
-  const float3 d = P - verts[2];
-  const float3 s2 = cross(d, e1);
-  float rt = dot(e2, s2) * invdivisor;
-
-  P = P + D * rt;
-
-  if (!(sd->object_flag & SD_OBJECT_TRANSFORM_APPLIED)) {
-    const Transform tfm = object_get_transform(kg, sd);
-    P = transform_point(&tfm, P);
-  }
-
-  return P;
-#    else  /* __INTERSECTION_REFINE__ */
-  return P + D * t;
-#    endif /* __INTERSECTION_REFINE__ */
-#  endif
-}
-#endif /* __BVH_LOCAL__ */
-
 /* Ray intersection. We simply compute the vertex positions at the given ray
  * time and do a ray intersection with the resulting triangle.
  */
@@ -153,14 +63,12 @@ ccl_device_inline bool motion_triangle_intersect(KernelGlobals kg,
                                                  float time,
                                                  uint visibility,
                                                  int object,
+                                                 int prim,
                                                  int prim_addr)
 {
-  /* Primitive index for vertex location lookup. */
-  int prim = kernel_tex_fetch(__prim_index, prim_addr);
-  int fobject = (object == OBJECT_NONE) ? kernel_tex_fetch(__prim_object, prim_addr) : object;
   /* Get vertex locations for intersection. */
   float3 verts[3];
-  motion_triangle_vertices(kg, fobject, prim, time, verts);
+  motion_triangle_vertices(kg, object, prim, time, verts);
   /* Ray-triangle intersection, unoptimized. */
   float t, u, v;
   if (ray_triangle_intersect(P, dir, tmax, verts[0], verts[1], verts[2], &u, &v, &t)) {
@@ -175,8 +83,7 @@ ccl_device_inline bool motion_triangle_intersect(KernelGlobals kg,
       isect->u = u;
       isect->v = v;
       isect->prim = prim;
-      isect->object = (object == OBJECT_NONE) ? kernel_tex_fetch(__prim_object, prim_addr) :
+      isect->object = object;
-                                                object;
       isect->type = PRIMITIVE_MOTION_TRIANGLE;
       return true;
     }
@@ -196,25 +103,15 @@ ccl_device_inline bool motion_triangle_intersect_local(KernelGlobals kg,
                                                        float3 dir,
                                                        float time,
                                                        int object,
-                                                       int local_object,
+                                                       int prim,
                                                        int prim_addr,
                                                        float tmax,
                                                        ccl_private uint *lcg_state,
                                                        int max_hits)
 {
-  /* Only intersect with matching object, for instanced objects we
-   * already know we are only intersecting the right object. */
-  if (object == OBJECT_NONE) {
-    if (kernel_tex_fetch(__prim_object, prim_addr) != local_object) {
-      return false;
-    }
-  }
-
-  /* Primitive index for vertex location lookup. */
-  int prim = kernel_tex_fetch(__prim_index, prim_addr);
   /* Get vertex locations for intersection. */
   float3 verts[3];
-  motion_triangle_vertices(kg, local_object, prim, time, verts);
+  motion_triangle_vertices(kg, object, prim, time, verts);
   /* Ray-triangle intersection, unoptimized. */
   float t, u, v;
   if (!ray_triangle_intersect(P, dir, tmax, verts[0], verts[1], verts[2], &u, &v, &t)) {
@@ -266,7 +163,7 @@ ccl_device_inline bool motion_triangle_intersect_local(KernelGlobals kg,
   isect->u = u;
   isect->v = v;
   isect->prim = prim;
-  isect->object = local_object;
+  isect->object = object;
   isect->type = PRIMITIVE_MOTION_TRIANGLE;
 
   /* Record geometric normal. */

intern/cycles/kernel/geom/motion_triangle_shader.h

@@ -68,15 +68,7 @@ ccl_device_noinline void motion_triangle_shader_setup(KernelGlobals kg,
   verts[1] = (1.0f - t) * verts[1] + t * next_verts[1];
   verts[2] = (1.0f - t) * verts[2] + t * next_verts[2];
   /* Compute refined position. */
-#ifdef __BVH_LOCAL__
+  sd->P = motion_triangle_point_from_uv(kg, sd, isect_object, isect_prim, sd->u, sd->v, verts);
-  if (is_local) {
-    sd->P = motion_triangle_refine_local(kg, sd, P, D, ray_t, isect_object, isect_prim, verts);
-  }
-  else
-#endif /* __BVH_LOCAL__*/
-  {
-    sd->P = motion_triangle_refine(kg, sd, P, D, ray_t, isect_object, isect_prim, verts);
-  }
   /* Compute face normal. */
   float3 Ng;
   if (sd->object_flag & SD_OBJECT_NEGATIVE_SCALE_APPLIED) {

intern/cycles/kernel/geom/point.h

@@ -81,7 +81,7 @@ ccl_device float3 point_attribute_float3(KernelGlobals kg,
 #  endif
 
   if (desc.element == ATTR_ELEMENT_VERTEX) {
-    return float4_to_float3(kernel_tex_fetch(__attributes_float4, desc.offset + sd->prim));
+    return kernel_tex_fetch(__attributes_float3, desc.offset + sd->prim);
   }
   else {
     return make_float3(0.0f, 0.0f, 0.0f);
@@ -109,17 +109,59 @@ ccl_device float4 point_attribute_float4(KernelGlobals kg,
   }
 }
 
+/* Point position */
+
+ccl_device float3 point_position(KernelGlobals kg, ccl_private const ShaderData *sd)
+{
+  if (sd->type & PRIMITIVE_POINT) {
+    /* World space center. */
+    float3 P = (sd->type & PRIMITIVE_MOTION) ?
+                   float4_to_float3(motion_point(kg, sd->object, sd->prim, sd->time)) :
+                   float4_to_float3(kernel_tex_fetch(__points, sd->prim));
+
+    if (!(sd->object_flag & SD_OBJECT_TRANSFORM_APPLIED)) {
+      object_position_transform(kg, sd, &P);
+    }
+
+    return P;
+  }
+
+  return zero_float3();
+}
+
 /* Point radius */
 
 ccl_device float point_radius(KernelGlobals kg, ccl_private const ShaderData *sd)
 {
   if (sd->type & PRIMITIVE_POINT) {
-    return kernel_tex_fetch(__points, sd->prim).w;
+    /* World space radius. */
+    const float r = kernel_tex_fetch(__points, sd->prim).w;
+
+    if (sd->object_flag & SD_OBJECT_TRANSFORM_APPLIED) {
+      return r;
+    }
+    else {
+      float3 dir = make_float3(r, r, r);
+      object_dir_transform(kg, sd, &dir);
+      return average(dir);
+    }
   }
 
   return 0.0f;
 }
 
+/* Point random */
+
+ccl_device float point_random(KernelGlobals kg, ccl_private const ShaderData *sd)
+{
+  if (sd->type & PRIMITIVE_POINT) {
+    const AttributeDescriptor desc = find_attribute(kg, sd, ATTR_STD_POINT_RANDOM);
+    return (desc.offset != ATTR_STD_NOT_FOUND) ? point_attribute_float(kg, sd, desc, NULL, NULL) :
+                                                 0.0f;
+  }
+  return 0.0f;
+}
+
 /* Point location for motion pass, linear interpolation between keys and
  * ignoring radius because we do the same for the motion keys */
 

intern/cycles/kernel/geom/shader_data.h

@@ -89,7 +89,7 @@ ccl_device_inline void shader_setup_from_ray(KernelGlobals kg,
       sd->shader = kernel_tex_fetch(__tri_shader, sd->prim);
 
       /* vectors */
-      sd->P = triangle_refine(kg, sd, ray->P, ray->D, isect->t, isect->object, isect->prim);
+      sd->P = triangle_point_from_uv(kg, sd, isect->object, isect->prim, isect->u, isect->v);
       sd->Ng = Ng;
       sd->N = Ng;
 

intern/cycles/kernel/geom/triangle_intersect.h

@@ -33,9 +33,9 @@ ccl_device_inline bool triangle_intersect(KernelGlobals kg,
                                           float tmax,
                                           uint visibility,
                                           int object,
+                                          int prim,
                                           int prim_addr)
 {
-  const int prim = kernel_tex_fetch(__prim_index, prim_addr);
   const uint tri_vindex = kernel_tex_fetch(__tri_vindex, prim).w;
   const float3 tri_a = kernel_tex_fetch(__tri_verts, tri_vindex + 0),
                tri_b = kernel_tex_fetch(__tri_verts, tri_vindex + 1),
@@ -49,8 +49,7 @@ ccl_device_inline bool triangle_intersect(KernelGlobals kg,
     if (kernel_tex_fetch(__prim_visibility, prim_addr) & visibility)
 #endif
     {
-      isect->object = (object == OBJECT_NONE) ? kernel_tex_fetch(__prim_object, prim_addr) :
+      isect->object = object;
-                                                object;
       isect->prim = prim;
       isect->type = PRIMITIVE_TRIANGLE;
       isect->u = u;
@@ -74,21 +73,12 @@ ccl_device_inline bool triangle_intersect_local(KernelGlobals kg,
                                                 float3 P,
                                                 float3 dir,
                                                 int object,
-                                                int local_object,
+                                                int prim,
                                                 int prim_addr,
                                                 float tmax,
                                                 ccl_private uint *lcg_state,
                                                 int max_hits)
 {
-  /* Only intersect with matching object, for instanced objects we
-   * already know we are only intersecting the right object. */
-  if (object == OBJECT_NONE) {
-    if (kernel_tex_fetch(__prim_object, prim_addr) != local_object) {
-      return false;
-    }
-  }
-
-  const int prim = kernel_tex_fetch(__prim_index, prim_addr);
   const uint tri_vindex = kernel_tex_fetch(__tri_vindex, prim).w;
   const float3 tri_a = kernel_tex_fetch(__tri_verts, tri_vindex + 0),
                tri_b = kernel_tex_fetch(__tri_verts, tri_vindex + 1),
@@ -139,7 +129,7 @@ ccl_device_inline bool triangle_intersect_local(KernelGlobals kg,
   /* Record intersection. */
   ccl_private Intersection *isect = &local_isect->hits[hit];
   isect->prim = prim;
-  isect->object = local_object;
+  isect->object = object;
   isect->type = PRIMITIVE_TRIANGLE;
   isect->u = u;
   isect->v = v;
@@ -152,58 +142,23 @@ ccl_device_inline bool triangle_intersect_local(KernelGlobals kg,
 }
 #endif /* __BVH_LOCAL__ */
 
-/* Refine triangle intersection to more precise hit point. For rays that travel
+/**
- * far the precision is often not so good, this reintersects the primitive from
+ * Use the barycentric coordinates to get the intersection location
- * a closer distance. */
-
-/* Reintersections uses the paper:
- *
- * Tomas Moeller
- * Fast, minimum storage ray/triangle intersection
- * http://www.cs.virginia.edu/~gfx/Courses/2003/ImageSynthesis/papers/Acceleration/Fast%20MinimumStorage%20RayTriangle%20Intersection.pdf
  */
-
+ccl_device_inline float3 triangle_point_from_uv(KernelGlobals kg,
-ccl_device_inline float3 triangle_refine(KernelGlobals kg,
                                                 ccl_private ShaderData *sd,
-                                         float3 P,
-                                         float3 D,
-                                         float t,
                                                 const int isect_object,
-                                         const int isect_prim)
+                                                const int isect_prim,
+                                                const float u,
+                                                const float v)
 {
-#ifdef __INTERSECTION_REFINE__
-  if (!(sd->object_flag & SD_OBJECT_TRANSFORM_APPLIED)) {
-    if (UNLIKELY(t == 0.0f)) {
-      return P;
-    }
-    const Transform tfm = object_get_inverse_transform(kg, sd);
-
-    P = transform_point(&tfm, P);
-    D = transform_direction(&tfm, D * t);
-    D = normalize_len(D, &t);
-  }
-
-  P = P + D * t;
-
   const uint tri_vindex = kernel_tex_fetch(__tri_vindex, isect_prim).w;
   const packed_float3 tri_a = kernel_tex_fetch(__tri_verts, tri_vindex + 0),
                       tri_b = kernel_tex_fetch(__tri_verts, tri_vindex + 1),
                       tri_c = kernel_tex_fetch(__tri_verts, tri_vindex + 2);
-  float3 edge1 = make_float3(tri_a.x - tri_c.x, tri_a.y - tri_c.y, tri_a.z - tri_c.z);
+  float w = 1.0f - u - v;
-  float3 edge2 = make_float3(tri_b.x - tri_c.x, tri_b.y - tri_c.y, tri_b.z - tri_c.z);
+
-  float3 tvec = make_float3(P.x - tri_c.x, P.y - tri_c.y, P.z - tri_c.z);
+  float3 P = u * tri_a + v * tri_b + w * tri_c;
-  float3 qvec = cross(tvec, edge1);
-  float3 pvec = cross(D, edge2);
-  float det = dot(edge1, pvec);
-  if (det != 0.0f) {
-    /* If determinant is zero it means ray lies in the plane of
-     * the triangle. It is possible in theory due to watertight
-     * nature of triangle intersection. For such cases we simply
-     * don't refine intersection hoping it'll go all fine.
-     */
-    float rt = dot(edge2, qvec) / det;
-    P = P + D * rt;
-  }
 
   if (!(sd->object_flag & SD_OBJECT_TRANSFORM_APPLIED)) {
     const Transform tfm = object_get_transform(kg, sd);
@@ -211,65 +166,6 @@ ccl_device_inline float3 triangle_refine(KernelGlobals kg,
   }
 
   return P;
-#else
-  return P + D * t;
-#endif
-}
-
-/* Same as above, except that t is assumed to be in object space for
- * instancing.
- */
-ccl_device_inline float3 triangle_refine_local(KernelGlobals kg,
-                                               ccl_private ShaderData *sd,
-                                               float3 P,
-                                               float3 D,
-                                               float t,
-                                               const int isect_object,
-                                               const int isect_prim)
-{
-#if defined(__KERNEL_GPU_RAYTRACING__)
-  /* t is always in world space with OptiX and MetalRT. */
-  return triangle_refine(kg, sd, P, D, t, isect_object, isect_prim);
-#else
-  if (!(sd->object_flag & SD_OBJECT_TRANSFORM_APPLIED)) {
-    const Transform tfm = object_get_inverse_transform(kg, sd);
-
-    P = transform_point(&tfm, P);
-    D = transform_direction(&tfm, D);
-    D = normalize(D);
-  }
-
-  P = P + D * t;
-
-#  ifdef __INTERSECTION_REFINE__
-  const uint tri_vindex = kernel_tex_fetch(__tri_vindex, isect_prim).w;
-  const packed_float3 tri_a = kernel_tex_fetch(__tri_verts, tri_vindex + 0),
-                      tri_b = kernel_tex_fetch(__tri_verts, tri_vindex + 1),
-                      tri_c = kernel_tex_fetch(__tri_verts, tri_vindex + 2);
-  float3 edge1 = make_float3(tri_a.x - tri_c.x, tri_a.y - tri_c.y, tri_a.z - tri_c.z);
-  float3 edge2 = make_float3(tri_b.x - tri_c.x, tri_b.y - tri_c.y, tri_b.z - tri_c.z);
-  float3 tvec = make_float3(P.x - tri_c.x, P.y - tri_c.y, P.z - tri_c.z);
-  float3 qvec = cross(tvec, edge1);
-  float3 pvec = cross(D, edge2);
-  float det = dot(edge1, pvec);
-  if (det != 0.0f) {
-    /* If determinant is zero it means ray lies in the plane of
-     * the triangle. It is possible in theory due to watertight
-     * nature of triangle intersection. For such cases we simply
-     * don't refine intersection hoping it'll go all fine.
-     */
-    float rt = dot(edge2, qvec) / det;
-    P = P + D * rt;
-  }
-#  endif /* __INTERSECTION_REFINE__ */
-
-  if (!(sd->object_flag & SD_OBJECT_TRANSFORM_APPLIED)) {
-    const Transform tfm = object_get_transform(kg, sd);
-    P = transform_point(&tfm, P);
-  }
-
-  return P;
-#endif
 }
 
 CCL_NAMESPACE_END

intern/cycles/kernel/integrator/intersect_closest.h

@@ -328,6 +328,12 @@ ccl_device void integrator_intersect_closest(KernelGlobals kg,
 
   /* Scene Intersection. */
   Intersection isect ccl_optional_struct_init;
+  isect.object = OBJECT_NONE;
+  isect.prim = PRIM_NONE;
+  ray.self.object = last_isect_object;
+  ray.self.prim = last_isect_prim;
+  ray.self.light_object = OBJECT_NONE;
+  ray.self.light_prim = PRIM_NONE;
   bool hit = scene_intersect(kg, &ray, visibility, &isect);
 
   /* TODO: remove this and do it in the various intersection functions instead. */

intern/cycles/kernel/integrator/intersect_shadow.h

@@ -156,7 +156,10 @@ ccl_device void integrator_intersect_shadow(KernelGlobals kg, IntegratorShadowSt
   /* Read ray from integrator state into local memory. */
   Ray ray ccl_optional_struct_init;
   integrator_state_read_shadow_ray(kg, state, &ray);
-
+  ray.self.object = INTEGRATOR_STATE_ARRAY(state, shadow_isect, 0, object);
+  ray.self.prim = INTEGRATOR_STATE_ARRAY(state, shadow_isect, 0, prim);
+  ray.self.light_object = INTEGRATOR_STATE_ARRAY(state, shadow_isect, 1, object);
+  ray.self.light_prim = INTEGRATOR_STATE_ARRAY(state, shadow_isect, 1, prim);
   /* Compute visibility. */
   const uint visibility = integrate_intersect_shadow_visibility(kg, state);
 

intern/cycles/kernel/integrator/intersect_volume_stack.h

@@ -38,7 +38,10 @@ ccl_device void integrator_volume_stack_update_for_subsurface(KernelGlobals kg,
   Ray volume_ray ccl_optional_struct_init;
   volume_ray.P = from_P;
   volume_ray.D = normalize_len(to_P - from_P, &volume_ray.t);
-
+  volume_ray.self.object = INTEGRATOR_STATE(state, isect, object);
+  volume_ray.self.prim = INTEGRATOR_STATE(state, isect, prim);
+  volume_ray.self.light_object = OBJECT_NONE;
+  volume_ray.self.light_prim = PRIM_NONE;
   /* Store to avoid global fetches on every intersection step. */
   const uint volume_stack_size = kernel_data.volume_stack_size;
 
@@ -68,7 +71,7 @@ ccl_device void integrator_volume_stack_update_for_subsurface(KernelGlobals kg,
     volume_stack_enter_exit(kg, state, stack_sd);
 
     /* Move ray forward. */
-    volume_ray.P = ray_offset(stack_sd->P, -stack_sd->Ng);
+    volume_ray.P = stack_sd->P;
     if (volume_ray.t != FLT_MAX) {
       volume_ray.D = normalize_len(to_P - volume_ray.P, &volume_ray.t);
     }
@@ -91,6 +94,10 @@ ccl_device void integrator_volume_stack_init(KernelGlobals kg, IntegratorState s
    * fewest hits. */
   volume_ray.D = make_float3(0.0f, 0.0f, 1.0f);
   volume_ray.t = FLT_MAX;
+  volume_ray.self.object = OBJECT_NONE;
+  volume_ray.self.prim = PRIM_NONE;
+  volume_ray.self.light_object = OBJECT_NONE;
+  volume_ray.self.light_prim = PRIM_NONE;
 
   int stack_index = 0, enclosed_index = 0;
 
@@ -203,7 +210,7 @@ ccl_device void integrator_volume_stack_init(KernelGlobals kg, IntegratorState s
     }
 
     /* Move ray forward. */
-    volume_ray.P = ray_offset(stack_sd->P, -stack_sd->Ng);
+    volume_ray.P = stack_sd->P;
     ++step;
   }
 #endif

intern/cycles/kernel/integrator/shade_light.h

@@ -37,8 +37,9 @@ ccl_device_inline void integrate_light(KernelGlobals kg,
 
   /* Advance ray beyond light. */
   /* TODO: can we make this more numerically robust to avoid reintersecting the
-   * same light in some cases? */
+   * same light in some cases? Ray should not intersect surface anymore as the
-  const float3 new_ray_P = ray_offset(ray_P + ray_D * isect.t, ray_D);
+   * object and prim ids will prevent self intersection. */
+  const float3 new_ray_P = ray_P + ray_D * isect.t;
   INTEGRATOR_STATE_WRITE(state, ray, P) = new_ray_P;
   INTEGRATOR_STATE_WRITE(state, ray, t) -= isect.t;
 
@@ -46,7 +47,7 @@ ccl_device_inline void integrate_light(KernelGlobals kg,
   const float mis_ray_t = INTEGRATOR_STATE(state, path, mis_ray_t);
   ray_P -= ray_D * mis_ray_t;
   isect.t += mis_ray_t;
-  INTEGRATOR_STATE_WRITE(state, path, mis_ray_t) = mis_ray_t + isect.t;
+  INTEGRATOR_STATE_WRITE(state, path, mis_ray_t) = isect.t;
 
   LightSample ls ccl_optional_struct_init;
   const bool use_light_sample = light_sample_from_intersection(kg, &isect, ray_P, ray_D, &ls);

intern/cycles/kernel/integrator/shade_shadow.h

@@ -83,7 +83,10 @@ ccl_device_inline void integrate_transparent_volume_shadow(KernelGlobals kg,
   /* Setup shader data. */
   Ray ray ccl_optional_struct_init;
   integrator_state_read_shadow_ray(kg, state, &ray);
-
+  ray.self.object = OBJECT_NONE;
+  ray.self.prim = PRIM_NONE;
+  ray.self.light_object = OBJECT_NONE;
+  ray.self.light_prim = PRIM_NONE;
   /* Modify ray position and length to match current segment. */
   const float start_t = (hit == 0) ? 0.0f :
                                      INTEGRATOR_STATE_ARRAY(state, shadow_isect, hit - 1, t);
@@ -149,7 +152,7 @@ ccl_device_inline bool integrate_transparent_shadow(KernelGlobals kg,
     const float last_hit_t = INTEGRATOR_STATE_ARRAY(state, shadow_isect, num_recorded_hits - 1, t);
     const float3 ray_P = INTEGRATOR_STATE(state, shadow_ray, P);
     const float3 ray_D = INTEGRATOR_STATE(state, shadow_ray, D);
-    INTEGRATOR_STATE_WRITE(state, shadow_ray, P) = ray_offset(ray_P + last_hit_t * ray_D, ray_D);
+    INTEGRATOR_STATE_WRITE(state, shadow_ray, P) = ray_P + last_hit_t * ray_D;
     INTEGRATOR_STATE_WRITE(state, shadow_ray, t) -= last_hit_t;
   }
 

intern/cycles/kernel/integrator/shade_surface.h

@@ -182,13 +182,18 @@ ccl_device_forceinline void integrate_surface_direct_light(KernelGlobals kg,
 
   /* Write shadow ray and associated state to global memory. */
   integrator_state_write_shadow_ray(kg, shadow_state, &ray);
+  // Save memory by storing the light and object indices in the shadow_isect
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 0, object) = ray.self.object;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 0, prim) = ray.self.prim;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 1, object) = ray.self.light_object;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 1, prim) = ray.self.light_prim;
 
   /* Copy state from main path to shadow path. */
   const uint16_t bounce = INTEGRATOR_STATE(state, path, bounce);
   const uint16_t transparent_bounce = INTEGRATOR_STATE(state, path, transparent_bounce);
   uint32_t shadow_flag = INTEGRATOR_STATE(state, path, flag);
   shadow_flag |= (is_light) ? PATH_RAY_SHADOW_FOR_LIGHT : 0;
-  shadow_flag |= PATH_RAY_SURFACE_PASS;
+  shadow_flag |= (shadow_flag & PATH_RAY_ANY_PASS) ? 0 : PATH_RAY_SURFACE_PASS;
   const float3 throughput = INTEGRATOR_STATE(state, path, throughput) * bsdf_eval_sum(&bsdf_eval);
 
   if (kernel_data.kernel_features & KERNEL_FEATURE_LIGHT_PASSES) {
@@ -266,13 +271,11 @@ ccl_device_forceinline int integrate_surface_bsdf_bssrdf_bounce(
   }
 
   /* Setup ray. Note that clipping works through transparent bounces. */
-  INTEGRATOR_STATE_WRITE(state, ray, P) = ray_offset(sd->P,
+  INTEGRATOR_STATE_WRITE(state, ray, P) = sd->P;
-                                                     (label & LABEL_TRANSMIT) ? -sd->Ng : sd->Ng);
   INTEGRATOR_STATE_WRITE(state, ray, D) = normalize(bsdf_omega_in);
   INTEGRATOR_STATE_WRITE(state, ray, t) = (label & LABEL_TRANSPARENT) ?
                                               INTEGRATOR_STATE(state, ray, t) - sd->ray_length :
                                               FLT_MAX;
-
 #ifdef __RAY_DIFFERENTIALS__
   INTEGRATOR_STATE_WRITE(state, ray, dP) = differential_make_compact(sd->dP);
   INTEGRATOR_STATE_WRITE(state, ray, dD) = differential_make_compact(bsdf_domega_in);
@@ -316,7 +319,7 @@ ccl_device_forceinline bool integrate_surface_volume_only_bounce(IntegratorState
   }
 
   /* Setup ray position, direction stays unchanged. */
-  INTEGRATOR_STATE_WRITE(state, ray, P) = ray_offset(sd->P, -sd->Ng);
+  INTEGRATOR_STATE_WRITE(state, ray, P) = sd->P;
 
   /* Clipping works through transparent. */
   INTEGRATOR_STATE_WRITE(state, ray, t) -= sd->ray_length;
@@ -360,10 +363,14 @@ ccl_device_forceinline void integrate_surface_ao(KernelGlobals kg,
   }
 
   Ray ray ccl_optional_struct_init;
-  ray.P = ray_offset(sd->P, sd->Ng);
+  ray.P = sd->P;
   ray.D = ao_D;
   ray.t = kernel_data.integrator.ao_bounces_distance;
   ray.time = sd->time;
+  ray.self.object = sd->object;
+  ray.self.prim = sd->prim;
+  ray.self.light_object = OBJECT_NONE;
+  ray.self.light_prim = PRIM_NONE;
   ray.dP = differential_zero_compact();
   ray.dD = differential_zero_compact();
 
@@ -375,6 +382,10 @@ ccl_device_forceinline void integrate_surface_ao(KernelGlobals kg,
 
   /* Write shadow ray and associated state to global memory. */
   integrator_state_write_shadow_ray(kg, shadow_state, &ray);
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 0, object) = ray.self.object;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 0, prim) = ray.self.prim;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 1, object) = ray.self.light_object;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 1, prim) = ray.self.light_prim;
 
   /* Copy state from main path to shadow path. */
   const uint16_t bounce = INTEGRATOR_STATE(state, path, bounce);

intern/cycles/kernel/integrator/shade_volume.h

@@ -791,13 +791,17 @@ ccl_device_forceinline void integrate_volume_direct_light(
 
   /* Write shadow ray and associated state to global memory. */
   integrator_state_write_shadow_ray(kg, shadow_state, &ray);
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 0, object) = ray.self.object;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 0, prim) = ray.self.prim;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 1, object) = ray.self.light_object;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 1, prim) = ray.self.light_prim;
 
   /* Copy state from main path to shadow path. */
   const uint16_t bounce = INTEGRATOR_STATE(state, path, bounce);
   const uint16_t transparent_bounce = INTEGRATOR_STATE(state, path, transparent_bounce);
   uint32_t shadow_flag = INTEGRATOR_STATE(state, path, flag);
   shadow_flag |= (is_light) ? PATH_RAY_SHADOW_FOR_LIGHT : 0;
-  shadow_flag |= PATH_RAY_VOLUME_PASS;
+  shadow_flag |= (shadow_flag & PATH_RAY_ANY_PASS) ? 0 : PATH_RAY_VOLUME_PASS;
   const float3 throughput_phase = throughput * bsdf_eval_sum(&phase_eval);
 
   if (kernel_data.kernel_features & KERNEL_FEATURE_LIGHT_PASSES) {
@@ -873,11 +877,13 @@ ccl_device_forceinline bool integrate_volume_phase_scatter(
   INTEGRATOR_STATE_WRITE(state, ray, P) = sd->P;
   INTEGRATOR_STATE_WRITE(state, ray, D) = normalize(phase_omega_in);
   INTEGRATOR_STATE_WRITE(state, ray, t) = FLT_MAX;
-
 #  ifdef __RAY_DIFFERENTIALS__
   INTEGRATOR_STATE_WRITE(state, ray, dP) = differential_make_compact(sd->dP);
   INTEGRATOR_STATE_WRITE(state, ray, dD) = differential_make_compact(phase_domega_in);
 #  endif
+  // Save memory by storing last hit prim and object in isect
+  INTEGRATOR_STATE_WRITE(state, isect, prim) = sd->prim;
+  INTEGRATOR_STATE_WRITE(state, isect, object) = sd->object;
 
   /* Update throughput. */
   const float3 throughput = INTEGRATOR_STATE(state, path, throughput);