USD: support importing dome light textures.

Fixed a bug where the active color wasn't being
set on imported meshes, resulting in no colors
displaying in the viewport.

.arcconfig

@@ -1,8 +0,0 @@
-{
-	"project_id" : "Blender",
-	"conduit_uri" : "https://developer.blender.org/",
-	"phabricator.uri" : "https://developer.blender.org/",
-	"git.default-relative-commit" : "origin/master",
-	"arc.land.update.default" : "rebase",
-	"arc.land.onto.default" : "master"
-}

.clang-format

@@ -236,6 +236,8 @@ ForEachMacros:
   - LOOP_UNSELECTED_POINTS
   - LOOP_VISIBLE_KEYS
   - LOOP_VISIBLE_POINTS
+  - LIGHT_FOREACH_BEGIN_DIRECTIONAL
+  - LIGHT_FOREACH_BEGIN_LOCAL
   - LISTBASE_CIRCULAR_BACKWARD_BEGIN
   - LISTBASE_CIRCULAR_FORWARD_BEGIN
   - LISTBASE_FOREACH

.gitea/issue_template/bug.yaml

@@ -1,9 +1,9 @@
 name: Bug Report
 about: File a bug report
 labels:
-  - "type::Report"
+  - "Type/Report"
-  - "status::Needs Triage"
+  - "Status/Needs Triage"
-  - "priority::Normal"
+  - "Priority/Normal"
 body:
   - type: markdown
     attributes:

.gitea/issue_template/design.yaml

@@ -1,7 +1,7 @@
 name: Design
 about: Create a design task (for developers only)
 labels:
-  - "type::Design"
+  - "Type/Design"
 body:
   - type: textarea
     id: body

.gitea/issue_template/todo.yaml

@@ -1,7 +1,7 @@
 name: To Do
 about: Create a to do task (for developers only)
 labels:
-  - "type::To Do"
+  - "Type/To Do"
 body:
   - type: textarea
     id: body

.github/pull_request_template.md

@@ -1,5 +1,4 @@
-This repository is only used as a mirror of git.blender.org. Blender development happens on
+This repository is only used as a mirror. Blender development happens on projects.blender.org.
-https://developer.blender.org.
 
 To get started with contributing code, please see:
 https://wiki.blender.org/wiki/Process/Contributing_Code

.github/stale.yml

@@ -15,8 +15,7 @@ staleLabel: stale
 # Comment to post when closing a stale Issue or Pull Request.
 closeComment: >
   This issue has been automatically closed, because this repository is only
-  used as a mirror of git.blender.org. Blender development happens on
+  used as a mirror. Blender development happens on projects.blender.org.
-  developer.blender.org.
 
   To get started contributing code, please read:
   https://wiki.blender.org/wiki/Process/Contributing_Code

.gitignore

@@ -39,7 +39,7 @@ Desktop.ini
 /doc/python_api/rst/bmesh.ops.rst
 
 # in-source lib downloads
-/build_files/build_environment/downloads
+/build_files/build_environment/downloads/
 
 # in-source buildbot signing configuration
 /build_files/buildbot/codesign/config_server.py
@@ -48,4 +48,20 @@ Desktop.ini
 waveletNoiseTile.bin
 
 # testing environment
-/Testing
+/Testing/
+
+# Translations.
+/locale/user-config.py
+
+# External repositories.
+/scripts/addons/
+/scripts/addons_contrib/
+
+# Ignore old submodules directories.
+# Eventually need to get rid of those, but for the first time of transition
+# avoid indidents when the folders exists after bisect and developers staging
+# them by accident.
+/release/scripts/addons/
+/release/datafiles/locale/
+/release/scripts/addons_contrib/
+/source/tools/

.gitmodules

@@ -1,20 +0,0 @@
-[submodule "release/scripts/addons"]
-	path = release/scripts/addons
-	url = ../blender-addons.git
-	branch = master
-	ignore = all
-[submodule "release/scripts/addons_contrib"]
-	path = release/scripts/addons_contrib
-	url = ../blender-addons-contrib.git
-	branch = master
-	ignore = all
-[submodule "release/datafiles/locale"]
-	path = release/datafiles/locale
-	url = ../blender-translations.git
-	branch = master
-	ignore = all
-[submodule "source/tools"]
-	path = source/tools
-	url = ../blender-dev-tools.git
-	branch = master
-	ignore = all

CMakeLists.txt

@@ -524,7 +524,7 @@ endif()
 if(NOT APPLE)
   option(WITH_CYCLES_DEVICE_HIP        "Enable Cycles AMD HIP support" ON)
   option(WITH_CYCLES_HIP_BINARIES      "Build Cycles AMD HIP binaries" OFF)
-  set(CYCLES_HIP_BINARIES_ARCH gfx1010 gfx1011 gfx1012 gfx1030 gfx1031 gfx1032 gfx1034 gfx1035 gfx1100 gfx1101 gfx1102 CACHE STRING "AMD HIP architectures to build binaries for")
+  set(CYCLES_HIP_BINARIES_ARCH gfx900 gfx906 gfx90c gfx902 gfx1010 gfx1011 gfx1012 gfx1030 gfx1031 gfx1032 gfx1034 gfx1035 gfx1100 gfx1101 gfx1102 CACHE STRING "AMD HIP architectures to build binaries for")
   mark_as_advanced(WITH_CYCLES_DEVICE_HIP)
   mark_as_advanced(CYCLES_HIP_BINARIES_ARCH)
 endif()
@@ -625,8 +625,10 @@ mark_as_advanced(
 
 # Vulkan
 option(WITH_VULKAN_BACKEND "Enable Vulkan as graphics backend (only for development)" OFF)
+option(WITH_VULKAN_GUARDEDALLOC "Use guardedalloc for host allocations done inside Vulkan (development option)" OFF)
 mark_as_advanced(
   WITH_VULKAN_BACKEND
+  WITH_VULKAN_GUARDEDALLOC
 )
 
 # Metal
@@ -952,21 +954,6 @@ endif()
 # -----------------------------------------------------------------------------
 # Check if Sub-modules are Cloned
 
-if(WITH_INTERNATIONAL)
-  file(GLOB RESULT "${CMAKE_SOURCE_DIR}/release/datafiles/locale")
-  list(LENGTH RESULT DIR_LEN)
-  if(DIR_LEN EQUAL 0)
-    message(
-      WARNING
-      "Translation path '${CMAKE_SOURCE_DIR}/release/datafiles/locale' is missing, "
-      "This is a 'git submodule', which are known not to work with bridges to other version "
-      "control systems."
-    )
-    set(TRANSLATIONS_FOUND OFF)
-    set_and_warn_library_found("Translations" TRANSLATIONS_FOUND WITH_INTERNATIONAL)
-  endif()
-endif()
-
 if(WITH_PYTHON)
   # While we have this as an '#error' in 'bpy_capi_utils.h',
   # upgrading Python tends to cause confusion for users who build.
@@ -982,14 +969,14 @@ if(WITH_PYTHON)
     )
   endif()
 
-  file(GLOB RESULT "${CMAKE_SOURCE_DIR}/release/scripts/addons")
+  file(GLOB RESULT "${CMAKE_SOURCE_DIR}/scripts/addons")
   list(LENGTH RESULT DIR_LEN)
   if(DIR_LEN EQUAL 0)
     message(
       WARNING
-      "Addons path '${CMAKE_SOURCE_DIR}/release/scripts/addons' is missing, "
+      "Addons path '${CMAKE_SOURCE_DIR}/scripts/addons' is missing. "
-      "This is a 'git submodule', which are known not to work with bridges to other version "
+      "This is an external repository which needs to be checked out. Use `make update` to do so. "
-      "control systems: * CONTINUING WITHOUT ADDONS *"
+      "* CONTINUING WITHOUT ADDONS *"
     )
   endif()
 endif()

GNUmakefile

@@ -69,7 +69,7 @@ Static Source Code Checking
    * check_cmake:           Runs our own cmake file checker which detects errors in the cmake file list definitions.
    * check_pep8:            Checks all Python script are pep8 which are tagged to use the stricter formatting.
    * check_mypy:            Checks all Python scripts using mypy,
-                            see: source/tools/check_source/check_mypy_config.py scripts which are included.
+                            see: tools/check_source/check_mypy_config.py scripts which are included.
 
 Documentation Checking
 
@@ -85,7 +85,7 @@ Spell Checkers
    * check_spelling_osl:    Check for spelling errors (OSL only).
    * check_spelling_py:     Check for spelling errors (Python only).
 
-   Note: an additional word-list is maintained at: 'source/tools/check_source/check_spelling_c_config.py'
+   Note: an additional word-list is maintained at: 'tools/check_source/check_spelling_c_config.py'
 
    Note: that spell checkers can take a 'CHECK_SPELLING_CACHE' filepath argument,
    so re-running does not need to re-check unchanged files.
@@ -299,7 +299,11 @@ else
 	ifneq ("$(wildcard $(DEPS_BUILD_DIR)/build.ninja)","")
 		DEPS_BUILD_COMMAND:=ninja
 	else
-		DEPS_BUILD_COMMAND:=make -s
+		ifeq ($(OS), Darwin)
+			DEPS_BUILD_COMMAND:=make -s
+		else
+			DEPS_BUILD_COMMAND:="$(BLENDER_DIR)/build_files/build_environment/linux/make_deps_wrapper.sh" -s
+		endif
 	endif
 endif
 
@@ -398,7 +402,7 @@ endif
 
 deps: .FORCE
 	@echo
-	@echo Configuring dependencies in \"$(DEPS_BUILD_DIR)\"
+	@echo Configuring dependencies in \"$(DEPS_BUILD_DIR)\", install to \"$(DEPS_INSTALL_DIR)\"
 
 	@cmake -H"$(DEPS_SOURCE_DIR)" \
 	       -B"$(DEPS_BUILD_DIR)" \
@@ -486,22 +490,22 @@ check_smatch: .FORCE
 	$(PYTHON) "$(BLENDER_DIR)/build_files/cmake/cmake_static_check_smatch.py"
 
 check_mypy: .FORCE
-	@$(PYTHON) "$(BLENDER_DIR)/source/tools/check_source/check_mypy.py"
+	@$(PYTHON) "$(BLENDER_DIR)/tools/check_source/check_mypy.py"
 
 check_wiki_file_structure: .FORCE
 	@PYTHONIOENCODING=utf_8 $(PYTHON) \
-	    "$(BLENDER_DIR)/source/tools/check_wiki/check_wiki_file_structure.py"
+	    "$(BLENDER_DIR)/tools/check_wiki/check_wiki_file_structure.py"
 
 check_spelling_py: .FORCE
 	@cd "$(BUILD_DIR)" ; \
 	PYTHONIOENCODING=utf_8 $(PYTHON) \
-	    "$(BLENDER_DIR)/source/tools/check_source/check_spelling.py" \
+	    "$(BLENDER_DIR)/tools/check_source/check_spelling.py" \
-	    "$(BLENDER_DIR)/release/scripts"
+	    "$(BLENDER_DIR)/scripts"
 
 check_spelling_c: .FORCE
 	@cd "$(BUILD_DIR)" ; \
 	PYTHONIOENCODING=utf_8 $(PYTHON) \
-	    "$(BLENDER_DIR)/source/tools/check_source/check_spelling.py" \
+	    "$(BLENDER_DIR)/tools/check_source/check_spelling.py" \
 	    --cache-file=$(CHECK_SPELLING_CACHE) \
 	    "$(BLENDER_DIR)/source" \
 	    "$(BLENDER_DIR)/intern/cycles" \
@@ -511,21 +515,21 @@ check_spelling_c: .FORCE
 check_spelling_osl: .FORCE
 	@cd "$(BUILD_DIR)" ; \
 	PYTHONIOENCODING=utf_8 $(PYTHON) \
-	    "$(BLENDER_DIR)/source/tools/check_source/check_spelling.py" \
+	    "$(BLENDER_DIR)/tools/check_source/check_spelling.py" \
 	    --cache-file=$(CHECK_SPELLING_CACHE) \
 	    "$(BLENDER_DIR)/intern/cycles/kernel/shaders"
 
 check_descriptions: .FORCE
 	@$(BLENDER_BIN) --background -noaudio --factory-startup --python \
-	    "$(BLENDER_DIR)/source/tools/check_source/check_descriptions.py"
+	    "$(BLENDER_DIR)/tools/check_source/check_descriptions.py"
 
 check_deprecated: .FORCE
 	@PYTHONIOENCODING=utf_8 $(PYTHON) \
-	    source/tools/check_source/check_deprecated.py
+	    tools/check_source/check_deprecated.py
 
 check_licenses: .FORCE
 	@PYTHONIOENCODING=utf_8 $(PYTHON) \
-	    "$(BLENDER_DIR)/source/tools/check_source/check_licenses.py" \
+	    "$(BLENDER_DIR)/tools/check_source/check_licenses.py" \
 	    "--show-headers=$(SHOW_HEADERS)"
 
 check_pep8: .FORCE
@@ -534,7 +538,7 @@ check_pep8: .FORCE
 
 check_cmake: .FORCE
 	@PYTHONIOENCODING=utf_8 $(PYTHON) \
-	    source/tools/check_source/check_cmake_consistency.py
+	    tools/check_source/check_cmake_consistency.py
 
 
 # -----------------------------------------------------------------------------
@@ -572,8 +576,8 @@ update_code: .FORCE
 	@$(PYTHON) ./build_files/utils/make_update.py --no-libraries
 
 format: .FORCE
-	@PATH="${LIBDIR}/llvm/bin/:$(PATH)" $(PYTHON) source/tools/utils_maintenance/clang_format_paths.py $(PATHS)
+	@PATH="${LIBDIR}/llvm/bin/:$(PATH)" $(PYTHON) tools/utils_maintenance/clang_format_paths.py $(PATHS)
-	@$(PYTHON) source/tools/utils_maintenance/autopep8_format_paths.py --autopep8-command="$(AUTOPEP8)" $(PATHS)
+	@$(PYTHON) tools/utils_maintenance/autopep8_format_paths.py --autopep8-command="$(AUTOPEP8)" $(PATHS)
 
 
 # -----------------------------------------------------------------------------

README.md

@@ -24,7 +24,7 @@ Development
 -----------
 
 - [Build Instructions](https://wiki.blender.org/wiki/Building_Blender)
-- [Code Review & Bug Tracker](https://developer.blender.org)
+- [Code Review & Bug Tracker](https://projects.blender.org)
 - [Developer Forum](https://devtalk.blender.org)
 - [Developer Documentation](https://wiki.blender.org)
 

build_files/build_environment/cmake/epoxy.cmake

@@ -10,7 +10,7 @@ ExternalProject_Add(external_epoxy
   URL_HASH ${EPOXY_HASH_TYPE}=${EPOXY_HASH}
   PREFIX ${BUILD_DIR}/epoxy
   PATCH_COMMAND ${PATCH_CMD} -p 1 -N -d ${BUILD_DIR}/epoxy/src/external_epoxy/ < ${PATCH_DIR}/epoxy.diff
-  CONFIGURE_COMMAND ${CONFIGURE_ENV} && ${MESON} setup --prefix ${LIBDIR}/epoxy --default-library ${EPOXY_LIB_TYPE} --libdir lib ${BUILD_DIR}/epoxy/src/external_epoxy-build ${BUILD_DIR}/epoxy/src/external_epoxy -Dtests=false
+  CONFIGURE_COMMAND ${CONFIGURE_ENV} && ${MESON} setup --prefix ${LIBDIR}/epoxy --default-library ${EPOXY_LIB_TYPE} --libdir lib ${BUILD_DIR}/epoxy/src/external_epoxy-build ${BUILD_DIR}/epoxy/src/external_epoxy -Dtests=false ${MESON_BUILD_TYPE}
   BUILD_COMMAND ninja
   INSTALL_COMMAND ninja install
 )

build_files/build_environment/cmake/fribidi.cmake

@@ -9,7 +9,7 @@ ExternalProject_Add(external_fribidi
   URL_HASH ${FRIBIDI_HASH_TYPE}=${FRIBIDI_HASH}
   DOWNLOAD_DIR ${DOWNLOAD_DIR}
   PREFIX ${BUILD_DIR}/fribidi
-  CONFIGURE_COMMAND ${MESON} setup --prefix ${LIBDIR}/fribidi -Ddocs=false --default-library static --libdir lib ${BUILD_DIR}/fribidi/src/external_fribidi-build ${BUILD_DIR}/fribidi/src/external_fribidi
+  CONFIGURE_COMMAND ${MESON} setup --prefix ${LIBDIR}/fribidi ${MESON_BUILD_TYPE} -Ddocs=false --default-library static --libdir lib ${BUILD_DIR}/fribidi/src/external_fribidi-build ${BUILD_DIR}/fribidi/src/external_fribidi
   BUILD_COMMAND ninja
   INSTALL_COMMAND ninja install
   INSTALL_DIR ${LIBDIR}/fribidi

build_files/build_environment/cmake/gmp.cmake

@@ -22,7 +22,7 @@ elseif(UNIX AND NOT APPLE)
   )
 endif()
 
-# Boolean crashes with Arm assembly, see T103423.
+# Boolean crashes with Arm assembly, see #103423.
 if(BLENDER_PLATFORM_ARM)
   set(GMP_OPTIONS
     ${GMP_OPTIONS}

build_files/build_environment/cmake/harfbuzz.cmake

@@ -21,6 +21,7 @@ set(HARFBUZZ_EXTRA_OPTIONS
   # Only used for command line utilities,
   # disable as this would add an addition & unnecessary build-dependency.
   -Dcairo=disabled
+  ${MESON_BUILD_TYPE}
 )
 
 ExternalProject_Add(external_harfbuzz
@@ -59,3 +60,10 @@ if(BUILD_MODE STREQUAL Release AND WIN32)
     DEPENDEES install
   )
 endif()
+
+if(BUILD_MODE STREQUAL Debug AND WIN32)
+  ExternalProject_Add_Step(external_harfbuzz after_install
+    COMMAND ${CMAKE_COMMAND} -E copy ${LIBDIR}/harfbuzz/lib/libharfbuzz.a ${HARVEST_TARGET}/harfbuzz/lib/libharfbuzz_d.lib
+    DEPENDEES install
+  )
+endif()

build_files/build_environment/cmake/igc.cmake

@@ -40,7 +40,8 @@ ExternalProject_Add(external_igc_llvm
     ${PATCH_CMD} -p 1 -d ${IGC_LLVM_SOURCE_DIR} < ${IGC_OPENCL_CLANG_PATCH_DIR}/clang/0004-OpenCL-support-cl_ext_float_atomics.patch &&
     ${PATCH_CMD} -p 1 -d ${IGC_LLVM_SOURCE_DIR} < ${IGC_OPENCL_CLANG_PATCH_DIR}/clang/0005-OpenCL-Add-cl_khr_integer_dot_product.patch &&
     ${PATCH_CMD} -p 1 -d ${IGC_LLVM_SOURCE_DIR} < ${IGC_OPENCL_CLANG_PATCH_DIR}/llvm/0001-Memory-leak-fix-for-Managed-Static-Mutex.patch &&
-    ${PATCH_CMD} -p 1 -d ${IGC_LLVM_SOURCE_DIR} < ${IGC_OPENCL_CLANG_PATCH_DIR}/llvm/0002-Remove-repo-name-in-LLVM-IR.patch
+    ${PATCH_CMD} -p 1 -d ${IGC_LLVM_SOURCE_DIR} < ${IGC_OPENCL_CLANG_PATCH_DIR}/llvm/0002-Remove-repo-name-in-LLVM-IR.patch &&
+    ${PATCH_CMD} -p 1 -d ${IGC_LLVM_SOURCE_DIR} < ${IGC_OPENCL_CLANG_PATCH_DIR}/llvm/0003-Add-missing-include-limit-in-benchmark.patch
 )
 add_dependencies(
   external_igc_llvm
@@ -55,9 +56,6 @@ ExternalProject_Add(external_igc_spirv_translator
   CONFIGURE_COMMAND echo .
   BUILD_COMMAND echo .
   INSTALL_COMMAND echo .
-  PATCH_COMMAND ${PATCH_CMD} -p 1 -d ${IGC_SPIRV_TRANSLATOR_SOURCE_DIR} < ${IGC_OPENCL_CLANG_PATCH_DIR}/spirv/0001-update-SPIR-V-headers-for-SPV_INTEL_split_barrier.patch &&
-    ${PATCH_CMD} -p 1 -d ${IGC_SPIRV_TRANSLATOR_SOURCE_DIR} < ${IGC_OPENCL_CLANG_PATCH_DIR}/spirv/0002-Add-support-for-split-barriers-extension-SPV_INTEL_s.patch &&
-    ${PATCH_CMD} -p 1 -d ${IGC_SPIRV_TRANSLATOR_SOURCE_DIR} < ${IGC_OPENCL_CLANG_PATCH_DIR}/spirv/0003-Support-cl_bf16_conversions.patch
 )
 add_dependencies(
   external_igc_spirv_translator

build_files/build_environment/cmake/mesa.cmake

@@ -15,7 +15,7 @@ llvm-config = '${LIBDIR}/llvm/bin/llvm-config'"
 )
 
 set(MESA_EXTRA_FLAGS
-  -Dbuildtype=release
+  ${MESON_BUILD_TYPE}
   -Dc_args=${MESA_CFLAGS}
   -Dcpp_args=${MESA_CXXFLAGS}
   -Dc_link_args=${MESA_LDFLAGS}

build_files/build_environment/cmake/openvdb.cmake

@@ -44,13 +44,21 @@ set(OPENVDB_EXTRA_ARGS
   # -DLLVM_DIR=${LIBDIR}/llvm/lib/cmake/llvm
 )
 
+set(OPENVDB_PATCH ${PATCH_CMD} -p 1 -d ${BUILD_DIR}/openvdb/src/openvdb < ${PATCH_DIR}/openvdb.diff)
+if(APPLE)
+  set(OPENVDB_PATCH
+    ${OPENVDB_PATCH} &&
+    ${PATCH_CMD} -p 0 -d ${BUILD_DIR}/openvdb/src/openvdb < ${PATCH_DIR}/openvdb_metal.diff
+  )
+endif()
+
 ExternalProject_Add(openvdb
   URL file://${PACKAGE_DIR}/${OPENVDB_FILE}
   DOWNLOAD_DIR ${DOWNLOAD_DIR}
   URL_HASH ${OPENVDB_HASH_TYPE}=${OPENVDB_HASH}
   CMAKE_GENERATOR ${PLATFORM_ALT_GENERATOR}
   PREFIX ${BUILD_DIR}/openvdb
-  PATCH_COMMAND ${PATCH_CMD} -p 1 -d ${BUILD_DIR}/openvdb/src/openvdb < ${PATCH_DIR}/openvdb.diff
+  PATCH_COMMAND ${OPENVDB_PATCH}
   CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${LIBDIR}/openvdb ${DEFAULT_CMAKE_FLAGS} ${OPENVDB_EXTRA_ARGS}
   INSTALL_DIR ${LIBDIR}/openvdb
 )

build_files/build_environment/cmake/options.cmake

@@ -16,8 +16,10 @@ message("BuildMode = ${BUILD_MODE}")
 
 if(BUILD_MODE STREQUAL "Debug")
   set(LIBDIR ${CMAKE_CURRENT_BINARY_DIR}/Debug)
+  set(MESON_BUILD_TYPE -Dbuildtype=debug)
 else()
   set(LIBDIR ${CMAKE_CURRENT_BINARY_DIR}/Release)
+  set(MESON_BUILD_TYPE -Dbuildtype=release)
 endif()
 
 set(DOWNLOAD_DIR "${CMAKE_CURRENT_BINARY_DIR}/downloads" CACHE STRING "Path for downloaded files")

build_files/build_environment/cmake/python.cmake

@@ -88,6 +88,19 @@ else()
     export LDFLAGS=${PYTHON_LDFLAGS} &&
     export PKG_CONFIG_PATH=${LIBDIR}/ffi/lib/pkgconfig)
 
+  # NOTE: untested on APPLE so far.
+  if(NOT APPLE)
+    set(PYTHON_CONFIGURE_EXTRA_ARGS
+      ${PYTHON_CONFIGURE_EXTRA_ARGS}
+      # Used on most release Linux builds (Fedora for e.g.),
+      # increases build times noticeably with the benefit of a modest speedup at runtime.
+      --enable-optimizations
+      # While LTO is OK when building on the same system, it's incompatible across GCC versions,
+      # making it impractical for developers to build against, so keep it disabled.
+      # `--with-lto`
+    )
+  endif()
+
   ExternalProject_Add(external_python
     URL file://${PACKAGE_DIR}/${PYTHON_FILE}
     DOWNLOAD_DIR ${DOWNLOAD_DIR}

build_files/build_environment/cmake/versions.cmake

@@ -668,9 +668,9 @@ set(SPIRV_HEADERS_FILE SPIR-V-Headers-${SPIRV_HEADERS_VERSION}.tar.gz)
 # compiler, the versions used are taken from the following location
 # https://github.com/intel/intel-graphics-compiler/releases
 
-set(IGC_VERSION 1.0.12149.1)
+set(IGC_VERSION 1.0.13064.7)
 set(IGC_URI https://github.com/intel/intel-graphics-compiler/archive/refs/tags/igc-${IGC_VERSION}.tar.gz)
-set(IGC_HASH 44f67f24e3bc5130f9f062533abf8154782a9d0a992bc19b498639a8521ae836)
+set(IGC_HASH a929abd4cca2b293961ec0437ee4b3b2147bd3b2c8a3c423af78c0c359b2e5ae)
 set(IGC_HASH_TYPE SHA256)
 set(IGC_FILE igc-${IGC_VERSION}.tar.gz)
 
@@ -690,15 +690,15 @@ set(IGC_LLVM_FILE ${IGC_LLVM_VERSION}.tar.gz)
 #
 # WARNING WARNING WARNING
 
-set(IGC_OPENCL_CLANG_VERSION 363a5262d8c7cff3fb28f3bdb5d85c8d7e91c1bb)
+set(IGC_OPENCL_CLANG_VERSION ee31812ea8b89d08c2918f045d11a19bd33525c5)
 set(IGC_OPENCL_CLANG_URI https://github.com/intel/opencl-clang/archive/${IGC_OPENCL_CLANG_VERSION}.tar.gz)
-set(IGC_OPENCL_CLANG_HASH aa8cf72bb239722ce8ce44f79413c6887ecc8ca18477dd520aa5c4809756da9a)
+set(IGC_OPENCL_CLANG_HASH 1db6735bbcfaa31e8a9ba39f121d6bafa806ea8919e9f56782d6aaa67771ddda)
 set(IGC_OPENCL_CLANG_HASH_TYPE SHA256)
 set(IGC_OPENCL_CLANG_FILE opencl-clang-${IGC_OPENCL_CLANG_VERSION}.tar.gz)
 
-set(IGC_VCINTRINSICS_VERSION v0.5.0)
+set(IGC_VCINTRINSICS_VERSION v0.11.0)
 set(IGC_VCINTRINSICS_URI https://github.com/intel/vc-intrinsics/archive/refs/tags/${IGC_VCINTRINSICS_VERSION}.tar.gz)
-set(IGC_VCINTRINSICS_HASH 70bb47c5e32173cf61514941e83ae7c7eb4485e6d2fca60cfa1f50d4f42c41f2)
+set(IGC_VCINTRINSICS_HASH e5acd5626ce7fa6d41ce154c50ac805eda734ee66af94ef28e680ac2ad81bb9f)
 set(IGC_VCINTRINSICS_HASH_TYPE SHA256)
 set(IGC_VCINTRINSICS_FILE vc-intrinsics-${IGC_VCINTRINSICS_VERSION}.tar.gz)
 
@@ -714,9 +714,9 @@ set(IGC_SPIRV_TOOLS_HASH 6e19900e948944243024aedd0a201baf3854b377b9cc7a386553bc1
 set(IGC_SPIRV_TOOLS_HASH_TYPE SHA256)
 set(IGC_SPIRV_TOOLS_FILE SPIR-V-Tools-${IGC_SPIRV_TOOLS_VERSION}.tar.gz)
 
-set(IGC_SPIRV_TRANSLATOR_VERSION a31ffaeef77e23d500b3ea3d35e0c42ff5648ad9)
+set(IGC_SPIRV_TRANSLATOR_VERSION d739c01d65ec00dee64dedd40deed805216a7193)
 set(IGC_SPIRV_TRANSLATOR_URI https://github.com/KhronosGroup/SPIRV-LLVM-Translator/archive/${IGC_SPIRV_TRANSLATOR_VERSION}.tar.gz)
-set(IGC_SPIRV_TRANSLATOR_HASH 9e26c96a45341b8f8af521bacea20e752623346340addd02af95d669f6e89252)
+set(IGC_SPIRV_TRANSLATOR_HASH ddc0cc9ccbe59dadeaf291012d59de142b2e9f2b124dbb634644d39daddaa13e)
 set(IGC_SPIRV_TRANSLATOR_HASH_TYPE SHA256)
 set(IGC_SPIRV_TRANSLATOR_FILE SPIR-V-Translator-${IGC_SPIRV_TRANSLATOR_VERSION}.tar.gz)
 
@@ -724,15 +724,15 @@ set(IGC_SPIRV_TRANSLATOR_FILE SPIR-V-Translator-${IGC_SPIRV_TRANSLATOR_VERSION}.
 ### Intel Graphics Compiler DEPS END ###
 ########################################
 
-set(GMMLIB_VERSION intel-gmmlib-22.1.8)
+set(GMMLIB_VERSION intel-gmmlib-22.3.0)
 set(GMMLIB_URI https://github.com/intel/gmmlib/archive/refs/tags/${GMMLIB_VERSION}.tar.gz)
-set(GMMLIB_HASH bf23e9a3742b4fb98c7666c9e9b29f3219e4b2fb4d831aaf4eed71f5e2d17368)
+set(GMMLIB_HASH c1f33e1519edfc527127baeb0436b783430dfd256c643130169a3a71dc86aff9)
 set(GMMLIB_HASH_TYPE SHA256)
 set(GMMLIB_FILE ${GMMLIB_VERSION}.tar.gz)
 
-set(OCLOC_VERSION 22.38.24278)
+set(OCLOC_VERSION 22.49.25018.21)
 set(OCLOC_URI https://github.com/intel/compute-runtime/archive/refs/tags/${OCLOC_VERSION}.tar.gz)
-set(OCLOC_HASH db0c542fccd651e6404b15a74d46027f1ce0eda8dc9e25a40cbb6c0faef257ee)
+set(OCLOC_HASH 92362dae08b503a34e5d3820ed284198c452bcd5e7504d90eb69887b20492c06)
 set(OCLOC_HASH_TYPE SHA256)
 set(OCLOC_FILE ocloc-${OCLOC_VERSION}.tar.gz)
 

build_files/build_environment/cmake/wayland.cmake

@@ -13,7 +13,7 @@ ExternalProject_Add(external_wayland
   # NOTE: `-lm` is needed for `libxml2` which is a static library that uses `libm.so`,
   # without this, math symbols such as `floor` aren't found.
   CONFIGURE_COMMAND ${CMAKE_COMMAND} -E env PKG_CONFIG_PATH=${LIBDIR}/expat/lib/pkgconfig:${LIBDIR}/xml2/lib/pkgconfig:${LIBDIR}/ffi/lib/pkgconfig:$PKG_CONFIG_PATH
-                    ${MESON} --prefix ${LIBDIR}/wayland -Ddocumentation=false -Dtests=false -D "c_link_args=-L${LIBDIR}/ffi/lib -lm" . ../external_wayland
+                    ${MESON} --prefix ${LIBDIR}/wayland ${MESON_BUILD_TYPE} -Ddocumentation=false -Dtests=false -D "c_link_args=-L${LIBDIR}/ffi/lib -lm" . ../external_wayland
   BUILD_COMMAND ninja
   INSTALL_COMMAND ninja install
 )

build_files/build_environment/cmake/wayland_protocols.cmake

@@ -7,7 +7,7 @@ ExternalProject_Add(external_wayland_protocols
   PREFIX ${BUILD_DIR}/wayland-protocols
   # Use `-E` so the `PKG_CONFIG_PATH` can be defined to link against our own WAYLAND.
   CONFIGURE_COMMAND ${CMAKE_COMMAND} -E env PKG_CONFIG_PATH=${LIBDIR}/wayland/lib64/pkgconfig:$PKG_CONFIG_PATH
-                    ${MESON} --prefix ${LIBDIR}/wayland-protocols . ../external_wayland_protocols -Dtests=false
+                    ${MESON} --prefix ${LIBDIR}/wayland-protocols ${MESON_BUILD_TYPE} . ../external_wayland_protocols -Dtests=false
   BUILD_COMMAND ninja
   INSTALL_COMMAND ninja install
 )

build_files/build_environment/cmake/xvidcore.cmake

@@ -17,11 +17,13 @@ ExternalProject_Add(external_xvidcore
   INSTALL_DIR ${LIBDIR}/xvidcore
 )
 
-ExternalProject_Add_Step(external_xvidcore after_install
+if(WIN32)
-  COMMAND ${CMAKE_COMMAND} -E rename ${LIBDIR}/xvidcore/lib/xvidcore.a ${LIBDIR}/xvidcore/lib/libxvidcore.a || true
+  ExternalProject_Add_Step(external_xvidcore after_install
-  COMMAND ${CMAKE_COMMAND} -E remove ${LIBDIR}/xvidcore/lib/xvidcore.dll.a
+    COMMAND ${CMAKE_COMMAND} -E rename ${LIBDIR}/xvidcore/lib/xvidcore.a ${LIBDIR}/xvidcore/lib/libxvidcore.a || true
-  DEPENDEES install
+    COMMAND ${CMAKE_COMMAND} -E remove ${LIBDIR}/xvidcore/lib/xvidcore.dll.a
-)
+    DEPENDEES install
+  )
+endif()
 
 if(MSVC)
   set_target_properties(external_xvidcore PROPERTIES FOLDER Mingw)

build_files/build_environment/linux/make_deps_wrapper.sh

@@ -0,0 +1,74 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: GPL-2.0-or-later
+
+# This script ensures:
+# - One dependency is built at a time.
+# - That dependency uses all available cores.
+#
+# Without this, simply calling `make -j$(nproc)` from the `${CMAKE_BUILD_DIR}/deps/`
+# directory will build many projects at once.
+#
+# This is undesirable for the following reasons:
+#
+# - The output from projects is mixed together,
+#   making it difficult to track down the cause of a build failure.
+#
+# - Larger dependencies such as LLVM can bottleneck the build process,
+#   making it necessary to cancel the build and manually run build commands in each directory.
+#
+# - Building many projects at once means canceling (Control-C) can lead to the build being in an undefined state.
+#   It's possible canceling happens as a patch is being applied or files are being copied.
+#   (steps that aren't part of the compilation process where it's typically safe to cancel).
+
+if [[ -z "$MY_MAKE_CALL_LEVEL" ]]; then
+  export MY_MAKE_CALL_LEVEL=0
+  export MY_MAKEFLAGS=$MAKEFLAGS
+
+  # Extract the jobs argument (`-jN`, `-j N`, `--jobs=N`).
+  add_next=0
+  for i in "$@"; do
+    case $i in
+      -j*)
+        export MY_JOBS_ARG=$i
+        if [ "$MY_JOBS_ARG" = "-j" ]; then
+          add_next=1
+        fi
+        ;;
+      --jobs=*)
+        shift # past argument=value
+        MY_JOBS_ARG=$i
+        ;;
+      *)
+        if (( add_next == 1 )); then
+          MY_JOBS_ARG="$MY_JOBS_ARG $i"
+          add_next=0
+        fi
+        ;;
+    esac
+  done
+  unset i add_next
+
+  if [[ -z "$MY_JOBS_ARG" ]]; then
+    MY_JOBS_ARG="-j$(nproc)"
+  fi
+  export MY_JOBS_ARG
+  # Support user defined `MAKEFLAGS`.
+  export MAKEFLAGS="$MY_MAKEFLAGS -j1"
+else
+  export MY_MAKE_CALL_LEVEL=$(( MY_MAKE_CALL_LEVEL + 1 ))
+  if (( MY_MAKE_CALL_LEVEL == 1 )); then
+    # Important to set jobs to 1, otherwise user defined jobs argument is used.
+    export MAKEFLAGS="$MY_MAKEFLAGS -j1"
+  elif (( MY_MAKE_CALL_LEVEL == 2 )); then
+    # This is the level used by each sub-project.
+    export MAKEFLAGS="$MY_MAKEFLAGS $MY_JOBS_ARG"
+  fi
+  # Else leave `MY_MAKEFLAGS` flags as-is, avoids setting a high number of jobs on recursive
+  # calls (which may easily run out of memory). Let the job-server handle the rest.
+fi
+
+# Useful for troubleshooting the wrapper.
+# echo "Call level: $MY_MAKE_CALL_LEVEL, args=$@".
+
+# Call actual make but ensure recursive calls run via this script.
+exec make MAKE="$0" "$@"

build_files/build_environment/patches/igc_opencl_clang.diff

@@ -1,7 +1,7 @@
 diff -Naur external_igc_opencl_clang.orig/CMakeLists.txt external_igc_opencl_clang/CMakeLists.txt
 --- external_igc_opencl_clang.orig/CMakeLists.txt	2022-03-16 05:51:10 -0600
 +++ external_igc_opencl_clang/CMakeLists.txt	2022-05-23 10:40:09 -0600
-@@ -126,22 +126,24 @@
+@@ -147,22 +147,24 @@
          )
      endif()
  

build_files/cmake/buildinfo.cmake

@@ -23,19 +23,19 @@ if(EXISTS ${SOURCE_DIR}/.git)
 
   if(MY_WC_BRANCH STREQUAL "HEAD")
     # Detached HEAD, check whether commit hash is reachable
-    # in the master branch
+    # in the main branch
     execute_process(COMMAND git rev-parse --short=12 HEAD
                     WORKING_DIRECTORY ${SOURCE_DIR}
                     OUTPUT_VARIABLE MY_WC_HASH
                     OUTPUT_STRIP_TRAILING_WHITESPACE)
 
-    execute_process(COMMAND git branch --list master blender-v* --contains ${MY_WC_HASH}
+    execute_process(COMMAND git branch --list main blender-v* --contains ${MY_WC_HASH}
                     WORKING_DIRECTORY ${SOURCE_DIR}
                     OUTPUT_VARIABLE _git_contains_check
                     OUTPUT_STRIP_TRAILING_WHITESPACE)
 
     if(NOT _git_contains_check STREQUAL "")
-      set(MY_WC_BRANCH "master")
+      set(MY_WC_BRANCH "main")
     else()
       execute_process(COMMAND git show-ref --tags -d
                       WORKING_DIRECTORY ${SOURCE_DIR}
@@ -48,7 +48,7 @@ if(EXISTS ${SOURCE_DIR}/.git)
                       OUTPUT_STRIP_TRAILING_WHITESPACE)
 
       if(_git_tag_hashes MATCHES "${_git_head_hash}")
-        set(MY_WC_BRANCH "master")
+        set(MY_WC_BRANCH "main")
       else()
         execute_process(COMMAND git branch --contains ${MY_WC_HASH}
                         WORKING_DIRECTORY ${SOURCE_DIR}

build_files/cmake/config/blender_release.cmake

@@ -85,7 +85,7 @@ if(NOT APPLE)
   set(WITH_CYCLES_DEVICE_OPTIX    ON  CACHE BOOL "" FORCE)
   set(WITH_CYCLES_CUDA_BINARIES   ON  CACHE BOOL "" FORCE)
   set(WITH_CYCLES_CUBIN_COMPILER  OFF CACHE BOOL "" FORCE)
-  set(WITH_CYCLES_HIP_BINARIES    ON  CACHE BOOL "" FORCE)
+  set(WITH_CYCLES_HIP_BINARIES    OFF CACHE BOOL "" FORCE)
   set(WITH_CYCLES_DEVICE_ONEAPI   ON  CACHE BOOL "" FORCE)
   set(WITH_CYCLES_ONEAPI_BINARIES ON  CACHE BOOL "" FORCE)
 endif()

build_files/cmake/example_scripts/cmake_linux_install.sh

@@ -11,11 +11,11 @@
 mkdir ~/blender-git
 cd ~/blender-git
 
-git clone http://git.blender.org/blender.git
+git clone https://projects.blender.org/blender/blender.git
 cd blender
 git submodule update --init --recursive
-git submodule foreach git checkout master
+git submodule foreach git checkout main
-git submodule foreach git pull --rebase origin master
+git submodule foreach git pull --rebase origin main
 
 # create build dir
 mkdir ~/blender-git/build-cmake
@@ -35,7 +35,7 @@ ln -s ~/blender-git/build-cmake/bin/blender ~/blender-git/blender/blender.bin
 echo ""
 echo "* Useful Commands *"
 echo "   Run Blender: ~/blender-git/blender/blender.bin"
-echo "   Update Blender: git pull --rebase; git submodule foreach git pull --rebase origin master"
+echo "   Update Blender: git pull --rebase; git submodule foreach git pull --rebase origin main"
 echo "   Reconfigure Blender: cd ~/blender-git/build-cmake ; cmake ."
 echo "   Build Blender: cd ~/blender-git/build-cmake ; make"
 echo ""

build_files/cmake/macros.cmake

@@ -544,7 +544,7 @@ endfunction()
 function(setup_platform_linker_libs
   target
   )
-  # jemalloc must be early in the list, to be before pthread (see T57998)
+  # jemalloc must be early in the list, to be before pthread (see #57998).
   if(WITH_MEM_JEMALLOC)
     target_link_libraries(${target} ${JEMALLOC_LIBRARIES})
   endif()

build_files/cmake/platform/platform_apple.cmake

@@ -440,7 +440,7 @@ string(APPEND PLATFORM_LINKFLAGS " -stdlib=libc++")
 # Make stack size more similar to Embree, required for Embree.
 string(APPEND PLATFORM_LINKFLAGS_EXECUTABLE " -Wl,-stack_size,0x100000")
 
-# Suppress ranlib "has no symbols" warnings (workaround for T48250)
+# Suppress ranlib "has no symbols" warnings (workaround for #48250).
 set(CMAKE_C_ARCHIVE_CREATE   "<CMAKE_AR> Scr <TARGET> <LINK_FLAGS> <OBJECTS>")
 set(CMAKE_CXX_ARCHIVE_CREATE "<CMAKE_AR> Scr <TARGET> <LINK_FLAGS> <OBJECTS>")
 # llvm-ranlib doesn't support this flag. Xcode's libtool does.

build_files/cmake/platform/platform_win32.cmake

@@ -121,7 +121,7 @@ if(WITH_WINDOWS_BUNDLE_CRT)
   include(InstallRequiredSystemLibraries)
 
   # ucrtbase(d).dll cannot be in the manifest, due to the way windows 10 handles
-  # redirects for this dll, for details see T88813.
+  # redirects for this dll, for details see #88813.
   foreach(lib ${CMAKE_INSTALL_SYSTEM_RUNTIME_LIBS})
     string(FIND ${lib} "ucrtbase" pos)
     if(NOT pos EQUAL -1)
@@ -295,7 +295,7 @@ unset(MATERIALX_LIB_FOLDER_EXISTS)
 if(NOT MSVC_CLANG                  AND # Available with MSVC 15.7+ but not for CLANG.
    NOT WITH_WINDOWS_SCCACHE        AND # And not when sccache is enabled
    NOT VS_CLANG_TIDY)                  # Clang-tidy does not like these options
-  add_compile_options(/experimental:external /external:templates- /external:I "${LIBDIR}" /external:W0)
+  add_compile_options(/experimental:external /external:I "${LIBDIR}" /external:W0)
 endif()
 
 # Add each of our libraries to our cmake_prefix_path so find_package() could work
@@ -901,11 +901,11 @@ endif()
 
 if(WINDOWS_PYTHON_DEBUG)
   # Include the system scripts in the blender_python_system_scripts project.
-  file(GLOB_RECURSE inFiles "${CMAKE_SOURCE_DIR}/release/scripts/*.*" )
+  file(GLOB_RECURSE inFiles "${CMAKE_SOURCE_DIR}/scripts/*.*" )
   add_custom_target(blender_python_system_scripts SOURCES ${inFiles})
   foreach(_source IN ITEMS ${inFiles})
     get_filename_component(_source_path "${_source}" PATH)
-    string(REPLACE "${CMAKE_SOURCE_DIR}/release/scripts/" "" _source_path "${_source_path}")
+    string(REPLACE "${CMAKE_SOURCE_DIR}/scripts/" "" _source_path "${_source_path}")
     string(REPLACE "/" "\\" _group_path "${_source_path}")
     source_group("${_group_path}" FILES "${_source}")
   endforeach()
@@ -940,7 +940,7 @@ if(WINDOWS_PYTHON_DEBUG)
     file(WRITE ${USER_PROPS_FILE} "<?xml version=\"1.0\" encoding=\"utf-8\"?>
 <Project DefaultTargets=\"Build\" xmlns=\"http://schemas.microsoft.com/developer/msbuild/2003\">
   <PropertyGroup>
-    <LocalDebuggerCommandArguments>-con --env-system-scripts \"${CMAKE_SOURCE_DIR}/release/scripts\" </LocalDebuggerCommandArguments>
+    <LocalDebuggerCommandArguments>-con --env-system-scripts \"${CMAKE_SOURCE_DIR}/scripts\" </LocalDebuggerCommandArguments>
   </PropertyGroup>
 </Project>")
   endif()

build_files/cmake/project_info.py

@@ -142,7 +142,7 @@ def cmake_advanced_info() -> Union[Tuple[List[str], List[Tuple[str, str]]], Tupl
 
     make_exe = cmake_cache_var("CMAKE_MAKE_PROGRAM")
     if make_exe is None:
-        print("Make command not found in: %r not found" % project_path)
+        print("Make command not found: CMAKE_MAKE_PROGRAM")
         return None, None
 
     make_exe_basename = os.path.basename(make_exe)

build_files/config/pipeline_config.yaml

@@ -1,53 +1,3 @@
-#
-# Used by Buildbot build pipeline make_update.py script only for now
-# We intended to update the make_update.py in the branches to use this file eventually
-#
-update-code:
-    git:
-        submodules:
-        -   branch: master
-            commit_id: HEAD
-            path: release/scripts/addons
-        -   branch: master
-            commit_id: HEAD
-            path: release/scripts/addons_contrib
-        -   branch: master
-            commit_id: HEAD
-            path: release/datafiles/locale
-        -   branch: master
-            commit_id: HEAD
-            path: source/tools
-    svn:
-        libraries:
-            darwin-arm64:
-                branch: trunk
-                commit_id: HEAD
-                path: lib/darwin_arm64
-            darwin-x86_64:
-                branch: trunk
-                commit_id: HEAD
-                path: lib/darwin
-            linux-x86_64:
-                branch: trunk
-                commit_id: HEAD
-                path: lib/linux_x86_64_glibc_228
-            windows-amd64:
-                branch: trunk
-                commit_id: HEAD
-                path: lib/win64_vc15
-        tests:
-            branch: trunk
-            commit_id: HEAD
-            path: lib/tests
-        benchmarks:
-            branch: trunk
-            commit_id: HEAD
-            path: lib/benchmarks
-        assets:
-            branch: trunk
-            commit_id: HEAD
-            path: lib/assets
-
 #
 # Buildbot only configs
 #

build_files/utils/make_bpy_wheel.py

@@ -58,7 +58,7 @@ Each Blender release supports one Python version, and the package is only compat
 ## Source Code
 
 * [Releases](https://download.blender.org/source/)
-* Repository: [git.blender.org/blender.git](https://git.blender.org/gitweb/gitweb.cgi/blender.git)
+* Repository: [projects.blender.org/blender/blender.git](https://projects.blender.org/blender/blender)
 
 ## Credits
 

build_files/utils/make_source_archive.py

@@ -135,7 +135,7 @@ def submodules_to_manifest(
         submodule = line.split()[1]
 
         # Don't use native slashes as GIT for MS-Windows outputs forward slashes.
-        if skip_addon_contrib and submodule == "release/scripts/addons_contrib":
+        if skip_addon_contrib and submodule == "scripts/addons_contrib":
             continue
 
         for path in git_ls_files(blender_srcdir / submodule):

build_files/utils/make_update.py

@@ -16,14 +16,28 @@ import shutil
 import sys
 
 import make_utils
+from pathlib import Path
 from make_utils import call, check_output
+from urllib.parse import urljoin
 
 from typing import (
     List,
+    Iterable,
     Optional,
 )
 
 
+class Submodule:
+    path: str
+    branch: str
+    branch_fallback: str
+
+    def __init__(self, path: str, branch: str, branch_fallback: str) -> None:
+        self.path = path
+        self.branch = branch
+        self.branch_fallback = branch_fallback
+
+
 def print_stage(text: str) -> None:
     print("")
     print(text)
@@ -42,6 +56,7 @@ def parse_arguments() -> argparse.Namespace:
     parser.add_argument("--svn-branch", default=None)
     parser.add_argument("--git-command", default="git")
     parser.add_argument("--use-linux-libraries", action="store_true")
+    parser.add_argument("--architecture", type=str, choices=("x86_64", "amd64", "arm64",))
     return parser.parse_args()
 
 
@@ -51,6 +66,19 @@ def get_blender_git_root() -> str:
 # Setup for precompiled libraries and tests from svn.
 
 
+def get_effective_architecture(args: argparse.Namespace) -> str:
+    architecture = args.architecture
+    if architecture:
+        assert isinstance(architecture, str)
+        return architecture
+
+    # Check platform.version to detect arm64 with x86_64 python binary.
+    if "ARM64" in platform.version():
+        return "arm64"
+
+    return platform.machine().lower()
+
+
 def svn_update(args: argparse.Namespace, release_version: Optional[str]) -> None:
     svn_non_interactive = [args.svn_command, '--non-interactive']
 
@@ -58,11 +86,11 @@ def svn_update(args: argparse.Namespace, release_version: Optional[str]) -> None
     svn_url = make_utils.svn_libraries_base_url(release_version, args.svn_branch)
 
     # Checkout precompiled libraries
+    architecture = get_effective_architecture(args)
     if sys.platform == 'darwin':
-        # Check platform.version to detect arm64 with x86_64 python binary.
+        if architecture == 'arm64':
-        if platform.machine() == 'arm64' or ('ARM64' in platform.version()):
             lib_platform = "darwin_arm64"
-        elif platform.machine() == 'x86_64':
+        elif architecture == 'x86_64':
             lib_platform = "darwin"
         else:
             lib_platform = None
@@ -170,7 +198,7 @@ def git_update_skip(args: argparse.Namespace, check_remote_exists: bool = True)
         return "rebase or merge in progress, complete it first"
 
     # Abort if uncommitted changes.
-    changes = check_output([args.git_command, 'status', '--porcelain', '--untracked-files=no'])
+    changes = check_output([args.git_command, 'status', '--porcelain', '--untracked-files=no', '--ignore-submodules'])
     if len(changes) != 0:
         return "you have unstaged changes"
 
@@ -184,97 +212,282 @@ def git_update_skip(args: argparse.Namespace, check_remote_exists: bool = True)
     return ""
 
 
+def use_upstream_workflow(args: argparse.Namespace) -> bool:
+    return make_utils.git_remote_exist(args.git_command, "upstream")
+
+
+def work_tree_update_upstream_workflow(args: argparse.Namespace, use_fetch=True) -> str:
+    """
+    Update the Blender repository using the Github style of fork organization
+
+    Returns true if the current local branch has been updated to the upstream state.
+    Otherwise false is returned.
+    """
+
+    branch_name = make_utils.git_branch(args.git_command)
+
+    if use_fetch:
+        call((args.git_command, "fetch", "upstream"))
+
+    upstream_branch = f"upstream/{branch_name}"
+    if not make_utils.git_branch_exists(args.git_command, upstream_branch):
+        return "no_branch"
+
+    retcode = call((args.git_command, "merge", "--ff-only", upstream_branch), exit_on_error=False)
+    if retcode != 0:
+        return "Unable to fast forward\n"
+
+    return ""
+
+
+def work_tree_update(args: argparse.Namespace, use_fetch=True) -> str:
+    """
+    Update the Git working tree using the best strategy
+
+    This function detects whether it is a github style of fork remote organization is used, or
+    is it a repository which origin is an upstream.
+    """
+
+    if use_upstream_workflow(args):
+        message = work_tree_update_upstream_workflow(args, use_fetch)
+        if message != "no_branch":
+            return message
+
+        # If there is upstream configured but the local branch is not in the upstream, try to
+        # update the branch from the fork.
+
+    update_command = [args.git_command, "pull", "--rebase"]
+
+    call(update_command)
+
+    return ""
+
+
 # Update blender repository.
-def blender_update(args: argparse.Namespace) -> None:
+def blender_update(args: argparse.Namespace) -> str:
     print_stage("Updating Blender Git Repository")
-    call([args.git_command, "pull", "--rebase"])
+
+    return work_tree_update(args)
 
 
-# Update submodules.
+def resolve_external_url(blender_url: str, repo_name: str) -> str:
-def submodules_update(
+    return urljoin(blender_url + "/", "../" + repo_name)
-        args: argparse.Namespace,
-        release_version: Optional[str],
-        branch: Optional[str],
-) -> str:
-    print_stage("Updating Submodules")
-    if make_utils.command_missing(args.git_command):
-        sys.stderr.write("git not found, can't update code\n")
-        sys.exit(1)
 
-    # Update submodules to appropriate given branch,
+
-    # falling back to master if none is given and/or found in a sub-repository.
+def external_script_copy_old_submodule_over(args: argparse.Namespace, directory_name: str) -> None:
-    branch_fallback = "master"
+    blender_git_root = Path(get_blender_git_root())
+    scripts_dir = blender_git_root / "scripts"
+    external_dir = scripts_dir / directory_name
+
+    old_submodule_relative_dir = Path("release") / "scripts" / directory_name
+    print(f"Moving {old_submodule_relative_dir} to scripts/{directory_name} ...")
+
+    old_submodule_dir = blender_git_root / old_submodule_relative_dir
+    shutil.move(old_submodule_dir, external_dir)
+
+    # Remove old ".git" which is a file with path to a submodule bare repo inside of main
+    # repo .git/modules directory.
+    (external_dir / ".git").unlink()
+
+    bare_repo_relative_dir = Path(".git") / "modules" / "release" / "scripts" / directory_name
+    print(f"Copying {bare_repo_relative_dir} to scripts/{directory_name}/.git ...")
+    bare_repo_dir = blender_git_root / bare_repo_relative_dir
+    shutil.copytree(bare_repo_dir, external_dir / ".git")
+
+    git_config = external_dir / ".git" / "config"
+    call((args.git_command, "config", "--file", git_config, "--unset", "core.worktree"))
+
+
+def external_script_initialize_if_needed(args: argparse.Namespace,
+                                         repo_name: str,
+                                         directory_name: str) -> None:
+    """Initialize checkout of an external repository scripts directory"""
+
+    blender_git_root = Path(get_blender_git_root())
+    blender_dot_git = blender_git_root / ".git"
+    scripts_dir = blender_git_root / "scripts"
+    external_dir = scripts_dir / directory_name
+
+    if external_dir.exists():
+        return
+
+    print(f"Initializing scripts/{directory_name} ...")
+
+    old_submodule_dot_git = blender_git_root / "release" / "scripts" / directory_name / ".git"
+    if old_submodule_dot_git.exists() and blender_dot_git.is_dir():
+        external_script_copy_old_submodule_over(args, directory_name)
+        return
+
+    origin_name = "upstream" if use_upstream_workflow(args) else "origin"
+    blender_url = make_utils.git_get_remote_url(args.git_command, origin_name)
+    external_url = resolve_external_url(blender_url, repo_name)
+
+    call((args.git_command, "clone", "--origin", origin_name, external_url, external_dir))
+
+
+def external_script_add_origin_if_needed(args: argparse.Namespace,
+                                         repo_name: str,
+                                         directory_name: str) -> str:
+    """
+    Add remote called 'origin' if there is a fork of the external repository available
+
+    This is only done when using Github style upstream workflow in the main repository.
+    """
+
+    if not use_upstream_workflow(args):
+        return ""
+
+    cwd = os.getcwd()
+
+    blender_git_root = Path(get_blender_git_root())
+    scripts_dir = blender_git_root / "scripts"
+    external_dir = scripts_dir / directory_name
+
+    origin_blender_url = make_utils.git_get_remote_url(args.git_command, "origin")
+    origin_external_url = resolve_external_url(origin_blender_url, repo_name)
+
+    try:
+        os.chdir(external_dir)
+
+        if (make_utils.git_remote_exist(args.git_command, "origin") or
+                not make_utils.git_remote_exist(args.git_command, "upstream")):
+            return
+
+        if not make_utils.git_is_remote_repository(args.git_command, origin_external_url):
+            return
+
+        print(f"Adding origin remote to {directory_name} pointing to fork ...")
+
+        # Non-obvious tricks to introduce the new remote called "origin" to the existing
+        # submodule configuration.
+        #
+        # This is all within the content of creating a fork of a submodule after `make update`
+        # has been run and possibly local branches tracking upstream were added.
+        #
+        # The idea here goes as following:
+        #
+        #  - Rename remote "upstream" to "origin", which takes care of changing the names of
+        #    remotes the local branches are tracking.
+        #
+        #  - Change the URL to the "origin", which so was was still pointing to upstream.
+        #
+        #  - Re-introduce the "upstream" remote, with the same URL as it had prior to rename.
+
+        upstream_url = make_utils.git_get_remote_url(args.git_command, "upstream")
+
+        call((args.git_command, "remote", "rename", "upstream", "origin"))
+        make_utils.git_set_config(args.git_command, f"remote.origin.url", origin_external_url)
+
+        call((args.git_command, "remote", "add", "upstream", upstream_url))
+    finally:
+        os.chdir(cwd)
+
+    return ""
+
+
+def external_scripts_update(args: argparse.Namespace,
+                            repo_name: str,
+                            directory_name: str,
+                            branch: Optional[str]) -> str:
+    """Update a single external checkout with the given name in the scripts folder"""
+
+    external_script_initialize_if_needed(args, repo_name, directory_name)
+    external_script_add_origin_if_needed(args, repo_name, directory_name)
+
+    print(f"Updating scripts/{directory_name} ...")
+
+    cwd = os.getcwd()
+
+    blender_git_root = Path(get_blender_git_root())
+    scripts_dir = blender_git_root / "scripts"
+    external_dir = scripts_dir / directory_name
+
+    # Update externals to appropriate given branch, falling back to main if none is given and/or
+    # found in a sub-repository.
+    branch_fallback = "main"
     if not branch:
         branch = branch_fallback
 
-    submodules = [
-        ("release/scripts/addons", branch, branch_fallback),
-        ("release/scripts/addons_contrib", branch, branch_fallback),
-        ("release/datafiles/locale", branch, branch_fallback),
-        ("source/tools", branch, branch_fallback),
-    ]
-
-    # Initialize submodules only if needed.
-    for submodule_path, submodule_branch, submodule_branch_fallback in submodules:
-        if not os.path.exists(os.path.join(submodule_path, ".git")):
-            call([args.git_command, "submodule", "update", "--init", "--recursive"])
-            break
-
-    # Checkout appropriate branch and pull changes.
     skip_msg = ""
-    for submodule_path, submodule_branch, submodule_branch_fallback in submodules:
-        cwd = os.getcwd()
-        try:
-            os.chdir(submodule_path)
-            msg = git_update_skip(args, check_remote_exists=False)
-            if msg:
-                skip_msg += submodule_path + " skipped: " + msg + "\n"
-            else:
-                # Find a matching branch that exists.
-                call([args.git_command, "fetch", "origin"])
-                if make_utils.git_branch_exists(args.git_command, submodule_branch):
-                    pass
-                elif make_utils.git_branch_exists(args.git_command, submodule_branch_fallback):
-                    submodule_branch = submodule_branch_fallback
-                else:
-                    # Skip.
-                    submodule_branch = ""
 
-                # Switch to branch and pull.
+    try:
-                if submodule_branch:
+        os.chdir(external_dir)
-                    if make_utils.git_branch(args.git_command) != submodule_branch:
+        msg = git_update_skip(args, check_remote_exists=False)
-                        call([args.git_command, "checkout", submodule_branch])
+        if msg:
-                    call([args.git_command, "pull", "--rebase", "origin", submodule_branch])
+            skip_msg += directory_name + " skipped: " + msg + "\n"
-        finally:
+        else:
-            os.chdir(cwd)
+            # Find a matching branch that exists.
+            for remote in ("origin", "upstream"):
+                if make_utils.git_remote_exist(args.git_command, remote):
+                    call([args.git_command, "fetch", remote])
+
+            submodule_branch = branch
+
+            if make_utils.git_branch_exists(args.git_command, submodule_branch):
+                pass
+            elif make_utils.git_branch_exists(args.git_command, branch_fallback):
+                submodule_branch = branch_fallback
+            else:
+                # Skip.
+                submodule_branch = ""
+
+            # Switch to branch and pull.
+            if submodule_branch:
+                if make_utils.git_branch(args.git_command) != submodule_branch:
+                    call([args.git_command, "checkout", submodule_branch])
+                # Don't use extra fetch since all remotes of interest have been already fetched
+                # some lines above.
+                skip_msg += work_tree_update(args, use_fetch=False)
+    finally:
+        os.chdir(cwd)
 
     return skip_msg
 
 
+def scripts_submodules_update(args: argparse.Namespace, branch: Optional[str]) -> str:
+    """Update working trees of addons and addons_contrib within the scripts/ directory"""
+    msg = ""
+
+    msg += external_scripts_update(args, "blender-addons", "addons", branch)
+    msg += external_scripts_update(args, "blender-addons-contrib", "addons_contrib", branch)
+
+    return msg
+
+
+def submodules_update(args: argparse.Namespace, branch: Optional[str]) -> str:
+    """Update submodules or other externally tracked source trees"""
+    msg = ""
+
+    msg += scripts_submodules_update(args, branch)
+
+    return msg
+
+
 if __name__ == "__main__":
     args = parse_arguments()
     blender_skip_msg = ""
     submodules_skip_msg = ""
 
-    # Test if we are building a specific release version.
+    blender_version = make_utils. parse_blender_version()
-    branch = make_utils.git_branch(args.git_command)
+    if blender_version.cycle != 'alpha':
-    if branch == 'HEAD':
+        major = blender_version.version // 100
-        sys.stderr.write('Blender git repository is in detached HEAD state, must be in a branch\n')
+        minor = blender_version.version % 100
-        sys.exit(1)
+        branch = f"blender-v{major}.{minor}-release"
-
+        release_version: Optional[str] = f"{major}.{minor}"
-    tag = make_utils.git_tag(args.git_command)
+    else:
-    release_version = make_utils.git_branch_release_version(branch, tag)
+        branch = 'main'
+        release_version = None
 
     if not args.no_libraries:
         svn_update(args, release_version)
     if not args.no_blender:
         blender_skip_msg = git_update_skip(args)
+        if not blender_skip_msg:
+            blender_skip_msg = blender_update(args)
         if blender_skip_msg:
             blender_skip_msg = "Blender repository skipped: " + blender_skip_msg + "\n"
-        else:
-            blender_update(args)
     if not args.no_submodules:
-        submodules_skip_msg = submodules_update(args, release_version, branch)
+        submodules_skip_msg = submodules_update(args, branch)
 
     # Report any skipped repositories at the end, so it's not as easy to miss.
     skip_msg = blender_skip_msg + submodules_skip_msg

build_files/utils/make_utils.py

@@ -9,7 +9,9 @@ import re
 import shutil
 import subprocess
 import sys
+import os
 from pathlib import Path
+from urllib.parse import urljoin
 
 from typing import (
     Sequence,
@@ -19,7 +21,7 @@ from typing import (
 
 def call(cmd: Sequence[str], exit_on_error: bool = True, silent: bool = False) -> int:
     if not silent:
-        print(" ".join(cmd))
+        print(" ".join([str(x) for x in cmd]))
 
     # Flush to ensure correct order output on Windows.
     sys.stdout.flush()
@@ -55,10 +57,48 @@ def check_output(cmd: Sequence[str], exit_on_error: bool = True) -> str:
 def git_branch_exists(git_command: str, branch: str) -> bool:
     return (
         call([git_command, "rev-parse", "--verify", branch], exit_on_error=False, silent=True) == 0 or
+        call([git_command, "rev-parse", "--verify", "remotes/upstream/" + branch], exit_on_error=False, silent=True) == 0 or
         call([git_command, "rev-parse", "--verify", "remotes/origin/" + branch], exit_on_error=False, silent=True) == 0
     )
 
 
+def git_get_remote_url(git_command: str, remote_name: str) -> bool:
+    return check_output((git_command, "ls-remote", "--get-url", remote_name))
+
+
+def git_remote_exist(git_command: str, remote_name: str) -> bool:
+    """Check whether there is a remote with the given name"""
+    # `git ls-remote --get-url upstream` will print an URL if there is such remote configured, and
+    # otherwise will print "upstream".
+    remote_url = check_output((git_command, "ls-remote", "--get-url", remote_name))
+    return remote_url != remote_name
+
+
+def git_get_resolved_submodule_url(git_command: str, blender_url: str, submodule_path: str) -> str:
+    git_root = check_output([git_command, "rev-parse", "--show-toplevel"])
+    dot_gitmodules = os.path.join(git_root, ".gitmodules")
+
+    submodule_key_prefix = f"submodule.{submodule_path}"
+    submodule_key_url = f"{submodule_key_prefix}.url"
+
+    gitmodule_url = git_get_config(
+        git_command, submodule_key_url, file=dot_gitmodules)
+
+    # A bit of a trickery to construct final URL.
+    # Only works for the relative submodule URLs.
+    #
+    # Note that unless the LHS URL ends up with a slash urljoin treats the last component as a
+    # file.
+    assert gitmodule_url.startswith('..')
+    return urljoin(blender_url + "/", gitmodule_url)
+
+
+def git_is_remote_repository(git_command: str, repo: str) -> bool:
+    """Returns true if the given repository is a valid/clonable git repo"""
+    exit_code = call((git_command, "ls-remote", repo, "HEAD"), exit_on_error=False, silent=True)
+    return exit_code == 0
+
+
 def git_branch(git_command: str) -> str:
     # Get current branch name.
     try:
@@ -70,6 +110,20 @@ def git_branch(git_command: str) -> str:
     return branch.strip().decode('utf8')
 
 
+def git_get_config(git_command: str, key: str, file: Optional[str] = None) -> str:
+    if file:
+        return check_output([git_command, "config", "--file", file, "--get", key])
+
+    return check_output([git_command, "config", "--get", key])
+
+
+def git_set_config(git_command: str, key: str, value: str, file: Optional[str] = None) -> str:
+    if file:
+        return check_output([git_command, "config", "--file", file, key, value])
+
+    return check_output([git_command, "config", key, value])
+
+
 def git_tag(git_command: str) -> Optional[str]:
     # Get current tag name.
     try:

build_files/windows/check_submodules.cmd

@@ -3,9 +3,9 @@ if NOT exist "%BLENDER_DIR%\source\tools\.git" (
 	if not "%GIT%" == "" (
 		"%GIT%" submodule update --init --recursive --progress
 		if errorlevel 1 goto FAIL
-		"%GIT%" submodule foreach git checkout master
+		"%GIT%" submodule foreach git checkout main
 		if errorlevel 1 goto FAIL
-		"%GIT%" submodule foreach git pull --rebase origin master
+		"%GIT%" submodule foreach git pull --rebase origin main
 		if errorlevel 1 goto FAIL
 		goto EOF
 	) else (

build_files/windows/show_hashes.cmd

@@ -4,9 +4,9 @@ if "%GIT%" == "" (
 )
 cd "%BLENDER_DIR%"
 for /f "delims=" %%i in ('"%GIT%" rev-parse HEAD') do echo Branch_hash=%%i
-cd "%BLENDER_DIR%/release/datafiles/locale"
+cd "%BLENDER_DIR%/locale"
 for /f "delims=" %%i in ('"%GIT%" rev-parse HEAD') do echo Locale_hash=%%i
-cd "%BLENDER_DIR%/release/scripts/addons"
+cd "%BLENDER_DIR%/scripts/addons"
 for /f "delims=" %%i in ('"%GIT%" rev-parse HEAD') do echo Addons_Hash=%%i
 cd "%BLENDER_DIR%"
 :EOF

doc/doxygen/Doxyfile

@@ -38,7 +38,7 @@ PROJECT_NAME           = Blender
 # could be handy for archiving the generated documentation or if some version
 # control system is used.
 
-PROJECT_NUMBER         = V3.5
+PROJECT_NUMBER         = V3.6
 
 # Using the PROJECT_BRIEF tag one can provide an optional one line description
 # for a project that appears at the top of each page and should give viewer a

doc/python_api/examples/blf.py

@@ -37,7 +37,7 @@ def draw_callback_px(self, context):
     # BLF drawing routine
     font_id = font_info["font_id"]
     blf.position(font_id, 2, 80, 0)
-    blf.size(font_id, 50, 72)
+    blf.size(font_id, 50)
     blf.draw(font_id, "Hello World")
 
 

doc/python_api/rst/include__bmesh.rst

@@ -31,7 +31,7 @@ For an overview of BMesh data types and how they reference each other see:
 Example Script
 --------------
 
-.. literalinclude:: __/__/__/release/scripts/templates_py/bmesh_simple.py
+.. literalinclude:: __/__/__/scripts/templates_py/bmesh_simple.py
 
 
 Standalone Module

doc/python_api/rst/info_quickstart.rst

@@ -288,7 +288,7 @@ In Python, this is done by defining a class, which is a subclass of an existing
 Example Operator
 ----------------
 
-.. literalinclude:: __/__/__/release/scripts/templates_py/operator_simple.py
+.. literalinclude:: __/__/__/scripts/templates_py/operator_simple.py
 
 Once this script runs, ``SimpleOperator`` is registered with Blender
 and can be called from Operator Search or added to the toolbar.
@@ -320,7 +320,7 @@ Example Panel
 Panels are registered as a class, like an operator.
 Notice the extra ``bl_`` variables used to set the context they display in.
 
-.. literalinclude:: __/__/__/release/scripts/templates_py/ui_panel_simple.py
+.. literalinclude:: __/__/__/scripts/templates_py/ui_panel_simple.py
 
 To run the script:
 

doc/python_api/sphinx_doc_gen.py

@@ -367,13 +367,13 @@ except ImportError:
 # Note that ".." is replaced by "__" in the RST files,
 # to avoid having to match Blender's source tree.
 EXTRA_SOURCE_FILES = (
-    "../../../release/scripts/templates_py/bmesh_simple.py",
+    "../../../scripts/templates_py/bmesh_simple.py",
-    "../../../release/scripts/templates_py/gizmo_operator.py",
+    "../../../scripts/templates_py/gizmo_operator.py",
-    "../../../release/scripts/templates_py/gizmo_operator_target.py",
+    "../../../scripts/templates_py/gizmo_operator_target.py",
-    "../../../release/scripts/templates_py/gizmo_simple.py",
+    "../../../scripts/templates_py/gizmo_simple.py",
-    "../../../release/scripts/templates_py/operator_simple.py",
+    "../../../scripts/templates_py/operator_simple.py",
-    "../../../release/scripts/templates_py/ui_panel_simple.py",
+    "../../../scripts/templates_py/ui_panel_simple.py",
-    "../../../release/scripts/templates_py/ui_previews_custom_icon.py",
+    "../../../scripts/templates_py/ui_previews_custom_icon.py",
     "../examples/bmesh.ops.1.py",
     "../examples/bpy.app.translations.py",
 )
@@ -476,7 +476,7 @@ MODULE_GROUPING = {
 
 # -------------------------------BLENDER----------------------------------------
 
-# converting bytes to strings, due to T30154
+# Converting bytes to strings, due to #30154.
 BLENDER_REVISION = str(bpy.app.build_hash, 'utf_8')
 BLENDER_REVISION_TIMESTAMP = bpy.app.build_commit_timestamp
 
@@ -487,7 +487,7 @@ BLENDER_VERSION_DOTS = "%d.%d" % (bpy.app.version[0], bpy.app.version[1])
 if BLENDER_REVISION != "Unknown":
     # SHA1 Git hash
     BLENDER_VERSION_HASH = BLENDER_REVISION
-    BLENDER_VERSION_HASH_HTML_LINK = "<a href=https://developer.blender.org/rB%s>%s</a>" % (
+    BLENDER_VERSION_HASH_HTML_LINK = "<a href=https://projects.blender.org/blender/blender/commit/%s>%s</a>" % (
         BLENDER_VERSION_HASH, BLENDER_VERSION_HASH,
     )
     BLENDER_VERSION_DATE = time.strftime("%d/%m/%Y", time.localtime(BLENDER_REVISION_TIMESTAMP))
@@ -647,7 +647,7 @@ def undocumented_message(module_name, type_name, identifier):
         module_name, type_name, identifier,
     )
 
-    return "Undocumented, consider `contributing <https://developer.blender.org/T51061>`__."
+    return "Undocumented, consider `contributing <https://developer.blender.org/>`__."
 
 
 def range_str(val):
@@ -1816,9 +1816,9 @@ def pyrna2sphinx(basepath):
 
     # operators
     def write_ops():
-        API_BASEURL = "https://developer.blender.org/diffusion/B/browse/master/release/scripts"
+        API_BASEURL = "https://projects.blender.org/blender/blender/src/branch/main/scripts"
-        API_BASEURL_ADDON = "https://developer.blender.org/diffusion/BA"
+        API_BASEURL_ADDON = "https://projects.blender.org/blender/blender-addons"
-        API_BASEURL_ADDON_CONTRIB = "https://developer.blender.org/diffusion/BAC"
+        API_BASEURL_ADDON_CONTRIB = "https://projects.blender.org/blender/blender-addons-contrib"
 
         op_modules = {}
         op = None
@@ -2200,7 +2200,7 @@ def write_rst_enum_items(basepath, key, key_no_prefix, enum_items):
     Write a single page for a static enum in RST.
 
     This helps avoiding very large lists being in-lined in many places which is an issue
-    especially with icons in ``bpy.types.UILayout``. See T87008.
+    especially with icons in ``bpy.types.UILayout``. See #87008.
     """
     filepath = os.path.join(basepath, "%s.rst" % key_no_prefix)
     with open(filepath, "w", encoding="utf-8") as fh:

doc/python_api/static/js/version_switch.js

@@ -156,7 +156,7 @@ var Popover = function() {
     },
     getNamed : function(v) {
       $.each(all_versions, function(ix, title) {
-        if (ix === "master" || ix === "latest") {
+        if (ix === "master" || ix === "main" || ix === "latest") {
           var m = title.match(/\d\.\d[\w\d\.]*/)[0];
           if (parseFloat(m) == v) {
             v = ix;

extern/hipew/README.blender

@@ -1,5 +1,5 @@
 Project: Blender
-URL: https://git.blender.org/blender.git
+URL: https://projects.blender.org/blender/blender.git
 License: Apache 2.0
 Upstream version: N/A
 Local modifications: None

intern/cycles/blender/addon/properties.py

@@ -12,6 +12,7 @@ from bpy.props import (
     PointerProperty,
     StringProperty,
 )
+from bpy.app.translations import pgettext_iface as iface_
 
 from math import pi
 
@@ -1664,30 +1665,51 @@ class CyclesPreferences(bpy.types.AddonPreferences):
             col.label(text="No compatible GPUs found for Cycles", icon='INFO')
 
             if device_type == 'CUDA':
-                col.label(text="Requires NVIDIA GPU with compute capability 3.0", icon='BLANK1')
+                compute_capability = "3.0"
+                col.label(text=iface_("Requires NVIDIA GPU with compute capability %s") % compute_capability,
+                          icon='BLANK1', translate=False)
             elif device_type == 'OPTIX':
-                col.label(text="Requires NVIDIA GPU with compute capability 5.0", icon='BLANK1')
+                compute_capability = "5.0"
-                col.label(text="and NVIDIA driver version 470 or newer", icon='BLANK1')
+                driver_version = "470"
+                col.label(text=iface_("Requires NVIDIA GPU with compute capability %s") % compute_capability,
+                          icon='BLANK1', translate=False)
+                col.label(text="and NVIDIA driver version %s or newer" % driver_version,
+                          icon='BLANK1', translate=False)
             elif device_type == 'HIP':
-                import sys
+                if True:
-                if sys.platform[:3] == "win":
+                    col.label(text="HIP temporarily disabled due to compiler bugs", icon='BLANK1')
-                    col.label(text="Requires AMD GPU with RDNA architecture", icon='BLANK1')
+                else:
-                    col.label(text="and AMD Radeon Pro 21.Q4 driver or newer", icon='BLANK1')
+                    import sys
-                elif sys.platform.startswith("linux"):
+                    if sys.platform[:3] == "win":
-                    col.label(text="Requires AMD GPU with RDNA architecture", icon='BLANK1')
+                        driver_version = "21.Q4"
-                    col.label(text="and AMD driver version 22.10 or newer", icon='BLANK1')
+                        col.label(text="Requires AMD GPU with Vega or RDNA architecture", icon='BLANK1')
+                        col.label(text=iface_("and AMD Radeon Pro %s driver or newer") % driver_version,
+                                  icon='BLANK1', translate=False)
+                    elif sys.platform.startswith("linux"):
+                        driver_version = "22.10"
+                        col.label(text="Requires AMD GPU with Vega or RDNA architecture", icon='BLANK1')
+                        col.label(text=iface_("and AMD driver version %s or newer") % driver_version, icon='BLANK1',
+                                  translate=False)
             elif device_type == 'ONEAPI':
                 import sys
                 if sys.platform.startswith("win"):
+                    driver_version = "101.4032"
                     col.label(text="Requires Intel GPU with Xe-HPG architecture", icon='BLANK1')
-                    col.label(text="and Windows driver version 101.4032 or newer", icon='BLANK1')
+                    col.label(text=iface_("and Windows driver version %s or newer") % driver_version,
+                              icon='BLANK1', translate=False)
                 elif sys.platform.startswith("linux"):
+                    driver_version = "1.3.24931"
                     col.label(text="Requires Intel GPU with Xe-HPG architecture and", icon='BLANK1')
-                    col.label(text="  - intel-level-zero-gpu version 1.3.24931 or newer", icon='BLANK1')
+                    col.label(text=iface_("  - intel-level-zero-gpu version %s or newer") % driver_version,
+                              icon='BLANK1', translate=False)
                     col.label(text="  - oneAPI Level-Zero Loader", icon='BLANK1')
             elif device_type == 'METAL':
-                col.label(text="Requires Apple Silicon with macOS 12.2 or newer", icon='BLANK1')
+                silicon_mac_version = "12.2"
-                col.label(text="or AMD with macOS 12.3 or newer", icon='BLANK1')
+                amd_mac_version = "12.3"
+                col.label(text=iface_("Requires Apple Silicon with macOS %s or newer") % silicon_mac_version,
+                          icon='BLANK1', translate=False)
+                col.label(text=iface_("or AMD with macOS %s or newer") % amd_mac_version, icon='BLANK1',
+                          translate=False)
             return
 
         for device in devices:
@@ -1723,12 +1745,21 @@ class CyclesPreferences(bpy.types.AddonPreferences):
 
         if compute_device_type == 'METAL':
             import platform
-            # MetalRT only works on Apple Silicon at present, pending argument encoding fixes on AMD
+            import re
-            # Kernel specialization is only viable on Apple Silicon at present due to relative compilation speed
+            is_navi_2 = False
-            if platform.machine() == 'arm64':
+            for device in devices:
+                if re.search(r"((RX)|(Pro)|(PRO))\s+W?6\d00X", device.name):
+                    is_navi_2 = True
+                    break
+
+            # MetalRT only works on Apple Silicon and Navi2.
+            is_arm64 = platform.machine() == 'arm64'
+            if is_arm64 or is_navi_2:
                 col = layout.column()
                 col.use_property_split = True
-                col.prop(self, "kernel_optimization_level")
+                # Kernel specialization is only supported on Apple Silicon
+                if is_arm64:
+                    col.prop(self, "kernel_optimization_level")
                 col.prop(self, "use_metalrt")
 
     def draw(self, context):

intern/cycles/blender/addon/ui.py

@@ -20,7 +20,7 @@ class CyclesPresetPanel(PresetPanel, Panel):
     @staticmethod
     def post_cb(context):
         # Modify an arbitrary built-in scene property to force a depsgraph
-        # update, because add-on properties don't. (see T62325)
+        # update, because add-on properties don't. (see #62325)
         render = context.scene.render
         render.filter_size = render.filter_size
 

intern/cycles/blender/display_driver.cpp

@@ -105,11 +105,12 @@ GPUShader *BlenderFallbackDisplayShader::bind(int width, int height)
 
   /* Bind shader now to enable uniform assignment. */
   GPU_shader_bind(shader_program_);
-  GPU_shader_uniform_int(shader_program_, image_texture_location_, 0);
+  int slot = 0;
+  GPU_shader_uniform_int_ex(shader_program_, image_texture_location_, 1, 1, &slot);
   float size[2];
   size[0] = width;
   size[1] = height;
-  GPU_shader_uniform_vector(shader_program_, fullscreen_location_, 2, 1, size);
+  GPU_shader_uniform_float_ex(shader_program_, fullscreen_location_, 2, 1, size);
   return shader_program_;
 }
 

intern/cycles/blender/image.cpp

@@ -20,7 +20,7 @@ BlenderImageLoader::BlenderImageLoader(BL::Image b_image,
     : b_image(b_image),
       frame(frame),
       tile_number(tile_number),
-      /* Don't free cache for preview render to avoid race condition from T93560, to be fixed
+      /* Don't free cache for preview render to avoid race condition from #93560, to be fixed
        * properly later as we are close to release. */
       free_cache(!is_preview_render && !b_image.has_data())
 {
@@ -72,7 +72,7 @@ bool BlenderImageLoader::load_metadata(const ImageDeviceFeatures &, ImageMetaDat
     metadata.colorspace = u_colorspace_raw;
   }
   else {
-    /* In some cases (e.g. T94135), the colorspace setting in Blender gets updated as part of the
+    /* In some cases (e.g. #94135), the colorspace setting in Blender gets updated as part of the
      * metadata queries in this function, so update the colorspace setting here. */
     PointerRNA colorspace_ptr = b_image.colorspace_settings().ptr;
     metadata.colorspace = get_enum_identifier(colorspace_ptr, "name");

intern/cycles/blender/light.cpp

@@ -24,7 +24,7 @@ void BlenderSync::sync_light(BL::Object &b_parent,
   Light *light = light_map.find(key);
 
   /* Check if the transform was modified, in case a linked collection is moved we do not get a
-   * specific depsgraph update (T88515). This also mimics the behavior for Objects. */
+   * specific depsgraph update (#88515). This also mimics the behavior for Objects. */
   const bool tfm_updated = (light && light->get_tfm() != tfm);
 
   /* Update if either object or light data changed. */

intern/cycles/blender/python.cpp

@@ -94,7 +94,7 @@ void python_thread_state_restore(void **python_thread_state)
   *python_thread_state = NULL;
 }
 
-static const char *PyC_UnicodeAsByte(PyObject *py_str, PyObject **coerce)
+static const char *PyC_UnicodeAsBytes(PyObject *py_str, PyObject **coerce)
 {
   const char *result = PyUnicode_AsUTF8(py_str);
   if (result) {
@@ -131,8 +131,8 @@ static PyObject *init_func(PyObject * /*self*/, PyObject *args)
   }
 
   PyObject *path_coerce = nullptr, *user_path_coerce = nullptr;
-  path_init(PyC_UnicodeAsByte(path, &path_coerce),
+  path_init(PyC_UnicodeAsBytes(path, &path_coerce),
-            PyC_UnicodeAsByte(user_path, &user_path_coerce));
+            PyC_UnicodeAsBytes(user_path, &user_path_coerce));
   Py_XDECREF(path_coerce);
   Py_XDECREF(user_path_coerce);
 

intern/cycles/blender/session.cpp

@@ -404,7 +404,7 @@ void BlenderSession::render(BL::Depsgraph &b_depsgraph_)
      * point we know that we've got everything to render current view layer.
      */
     /* At the moment we only free if we are not doing multi-view
-     * (or if we are rendering the last view). See T58142/D4239 for discussion.
+     * (or if we are rendering the last view). See #58142/D4239 for discussion.
      */
     if (view_index == num_views - 1) {
       free_blender_memory_if_possible();

intern/cycles/blender/sync.cpp

@@ -349,8 +349,7 @@ void BlenderSync::sync_integrator(BL::ViewLayer &b_view_layer, bool background)
 
   bool use_light_tree = get_boolean(cscene, "use_light_tree");
   integrator->set_use_light_tree(use_light_tree);
-  integrator->set_light_sampling_threshold(
+  integrator->set_light_sampling_threshold(get_float(cscene, "light_sampling_threshold"));
-      (use_light_tree) ? 0.0f : get_float(cscene, "light_sampling_threshold"));
 
   if (integrator->use_light_tree_is_modified()) {
     scene->light_manager->tag_update(scene, LightManager::UPDATE_ALL);
@@ -766,7 +765,7 @@ void BlenderSync::free_data_after_sync(BL::Depsgraph &b_depsgraph)
       (BlenderSession::headless || is_interface_locked) &&
       /* Baking re-uses the depsgraph multiple times, clearing crashes
        * reading un-evaluated mesh data which isn't aligned with the
-       * geometry we're baking, see T71012. */
+       * geometry we're baking, see #71012. */
       !scene->bake_manager->get_baking() &&
       /* Persistent data must main caches for performance and correctness. */
       !is_persistent_data;

intern/cycles/cmake/external_libs.cmake

@@ -42,12 +42,15 @@ endif()
 ###########################################################################
 
 if(WITH_CYCLES_HIP_BINARIES AND WITH_CYCLES_DEVICE_HIP)
-  find_package(HIP)
+  set(WITH_CYCLES_HIP_BINARIES OFF)
-  set_and_warn_library_found("HIP compiler" HIP_FOUND WITH_CYCLES_HIP_BINARIES)
+  message(STATUS "HIP temporarily disabled due to compiler bugs")
 
-  if(HIP_FOUND)
+  # find_package(HIP)
-    message(STATUS "Found HIP ${HIP_HIPCC_EXECUTABLE} (${HIP_VERSION})")
+  # set_and_warn_library_found("HIP compiler" HIP_FOUND WITH_CYCLES_HIP_BINARIES)
-  endif()
+
+  # if(HIP_FOUND)
+  #   message(STATUS "Found HIP ${HIP_HIPCC_EXECUTABLE} (${HIP_VERSION})")
+  # endif()
 endif()
 
 if(NOT WITH_HIP_DYNLOAD)

intern/cycles/device/cuda/device_impl.cpp

@@ -53,8 +53,12 @@ void CUDADevice::set_error(const string &error)
 }
 
 CUDADevice::CUDADevice(const DeviceInfo &info, Stats &stats, Profiler &profiler)
-    : Device(info, stats, profiler), texture_info(this, "texture_info", MEM_GLOBAL)
+    : GPUDevice(info, stats, profiler)
 {
+  /* Verify that base class types can be used with specific backend types */
+  static_assert(sizeof(texMemObject) == sizeof(CUtexObject));
+  static_assert(sizeof(arrayMemObject) == sizeof(CUarray));
+
   first_error = true;
 
   cuDevId = info.num;
@@ -65,12 +69,6 @@ CUDADevice::CUDADevice(const DeviceInfo &info, Stats &stats, Profiler &profiler)
 
   need_texture_info = false;
 
-  device_texture_headroom = 0;
-  device_working_headroom = 0;
-  move_texture_to_host = false;
-  map_host_limit = 0;
-  map_host_used = 0;
-  can_map_host = 0;
   pitch_alignment = 0;
 
   /* Initialize CUDA. */
@@ -91,8 +89,9 @@ CUDADevice::CUDADevice(const DeviceInfo &info, Stats &stats, Profiler &profiler)
   /* CU_CTX_MAP_HOST for mapping host memory when out of device memory.
    * CU_CTX_LMEM_RESIZE_TO_MAX for reserving local memory ahead of render,
    * so we can predict which memory to map to host. */
-  cuda_assert(
+  int value;
-      cuDeviceGetAttribute(&can_map_host, CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY, cuDevice));
+  cuda_assert(cuDeviceGetAttribute(&value, CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY, cuDevice));
+  can_map_host = value != 0;
 
   cuda_assert(cuDeviceGetAttribute(
       &pitch_alignment, CU_DEVICE_ATTRIBUTE_TEXTURE_PITCH_ALIGNMENT, cuDevice));
@@ -499,311 +498,57 @@ void CUDADevice::reserve_local_memory(const uint kernel_features)
 #  endif
 }
 
-void CUDADevice::init_host_memory()
+void CUDADevice::get_device_memory_info(size_t &total, size_t &free)
-{
-  /* Limit amount of host mapped memory, because allocating too much can
-   * cause system instability. Leave at least half or 4 GB of system
-   * memory free, whichever is smaller. */
-  size_t default_limit = 4 * 1024 * 1024 * 1024LL;
-  size_t system_ram = system_physical_ram();
-
-  if (system_ram > 0) {
-    if (system_ram / 2 > default_limit) {
-      map_host_limit = system_ram - default_limit;
-    }
-    else {
-      map_host_limit = system_ram / 2;
-    }
-  }
-  else {
-    VLOG_WARNING << "Mapped host memory disabled, failed to get system RAM";
-    map_host_limit = 0;
-  }
-
-  /* Amount of device memory to keep is free after texture memory
-   * and working memory allocations respectively. We set the working
-   * memory limit headroom lower so that some space is left after all
-   * texture memory allocations. */
-  device_working_headroom = 32 * 1024 * 1024LL;   // 32MB
-  device_texture_headroom = 128 * 1024 * 1024LL;  // 128MB
-
-  VLOG_INFO << "Mapped host memory limit set to " << string_human_readable_number(map_host_limit)
-            << " bytes. (" << string_human_readable_size(map_host_limit) << ")";
-}
-
-void CUDADevice::load_texture_info()
-{
-  if (need_texture_info) {
-    /* Unset flag before copying, so this does not loop indefinitely if the copy below calls
-     * into 'move_textures_to_host' (which calls 'load_texture_info' again). */
-    need_texture_info = false;
-    texture_info.copy_to_device();
-  }
-}
-
-void CUDADevice::move_textures_to_host(size_t size, bool for_texture)
-{
-  /* Break out of recursive call, which can happen when moving memory on a multi device. */
-  static bool any_device_moving_textures_to_host = false;
-  if (any_device_moving_textures_to_host) {
-    return;
-  }
-
-  /* Signal to reallocate textures in host memory only. */
-  move_texture_to_host = true;
-
-  while (size > 0) {
-    /* Find suitable memory allocation to move. */
-    device_memory *max_mem = NULL;
-    size_t max_size = 0;
-    bool max_is_image = false;
-
-    thread_scoped_lock lock(cuda_mem_map_mutex);
-    foreach (CUDAMemMap::value_type &pair, cuda_mem_map) {
-      device_memory &mem = *pair.first;
-      CUDAMem *cmem = &pair.second;
-
-      /* Can only move textures allocated on this device (and not those from peer devices).
-       * And need to ignore memory that is already on the host. */
-      if (!mem.is_resident(this) || cmem->use_mapped_host) {
-        continue;
-      }
-
-      bool is_texture = (mem.type == MEM_TEXTURE || mem.type == MEM_GLOBAL) &&
-                        (&mem != &texture_info);
-      bool is_image = is_texture && (mem.data_height > 1);
-
-      /* Can't move this type of memory. */
-      if (!is_texture || cmem->array) {
-        continue;
-      }
-
-      /* For other textures, only move image textures. */
-      if (for_texture && !is_image) {
-        continue;
-      }
-
-      /* Try to move largest allocation, prefer moving images. */
-      if (is_image > max_is_image || (is_image == max_is_image && mem.device_size > max_size)) {
-        max_is_image = is_image;
-        max_size = mem.device_size;
-        max_mem = &mem;
-      }
-    }
-    lock.unlock();
-
-    /* Move to host memory. This part is mutex protected since
-     * multiple CUDA devices could be moving the memory. The
-     * first one will do it, and the rest will adopt the pointer. */
-    if (max_mem) {
-      VLOG_WORK << "Move memory from device to host: " << max_mem->name;
-
-      static thread_mutex move_mutex;
-      thread_scoped_lock lock(move_mutex);
-
-      any_device_moving_textures_to_host = true;
-
-      /* Potentially need to call back into multi device, so pointer mapping
-       * and peer devices are updated. This is also necessary since the device
-       * pointer may just be a key here, so cannot be accessed and freed directly.
-       * Unfortunately it does mean that memory is reallocated on all other
-       * devices as well, which is potentially dangerous when still in use (since
-       * a thread rendering on another devices would only be caught in this mutex
-       * if it so happens to do an allocation at the same time as well. */
-      max_mem->device_copy_to();
-      size = (max_size >= size) ? 0 : size - max_size;
-
-      any_device_moving_textures_to_host = false;
-    }
-    else {
-      break;
-    }
-  }
-
-  /* Unset flag before texture info is reloaded, since it should stay in device memory. */
-  move_texture_to_host = false;
-
-  /* Update texture info array with new pointers. */
-  load_texture_info();
-}
-
-CUDADevice::CUDAMem *CUDADevice::generic_alloc(device_memory &mem, size_t pitch_padding)
 {
   CUDAContextScope scope(this);
 
-  CUdeviceptr device_pointer = 0;
-  size_t size = mem.memory_size() + pitch_padding;
-
-  CUresult mem_alloc_result = CUDA_ERROR_OUT_OF_MEMORY;
-  const char *status = "";
-
-  /* First try allocating in device memory, respecting headroom. We make
-   * an exception for texture info. It is small and frequently accessed,
-   * so treat it as working memory.
-   *
-   * If there is not enough room for working memory, we will try to move
-   * textures to host memory, assuming the performance impact would have
-   * been worse for working memory. */
-  bool is_texture = (mem.type == MEM_TEXTURE || mem.type == MEM_GLOBAL) && (&mem != &texture_info);
-  bool is_image = is_texture && (mem.data_height > 1);
-
-  size_t headroom = (is_texture) ? device_texture_headroom : device_working_headroom;
-
-  size_t total = 0, free = 0;
   cuMemGetInfo(&free, &total);
-
-  /* Move textures to host memory if needed. */
-  if (!move_texture_to_host && !is_image && (size + headroom) >= free && can_map_host) {
-    move_textures_to_host(size + headroom - free, is_texture);
-    cuMemGetInfo(&free, &total);
-  }
-
-  /* Allocate in device memory. */
-  if (!move_texture_to_host && (size + headroom) < free) {
-    mem_alloc_result = cuMemAlloc(&device_pointer, size);
-    if (mem_alloc_result == CUDA_SUCCESS) {
-      status = " in device memory";
-    }
-  }
-
-  /* Fall back to mapped host memory if needed and possible. */
-
-  void *shared_pointer = 0;
-
-  if (mem_alloc_result != CUDA_SUCCESS && can_map_host && mem.type != MEM_DEVICE_ONLY) {
-    if (mem.shared_pointer) {
-      /* Another device already allocated host memory. */
-      mem_alloc_result = CUDA_SUCCESS;
-      shared_pointer = mem.shared_pointer;
-    }
-    else if (map_host_used + size < map_host_limit) {
-      /* Allocate host memory ourselves. */
-      mem_alloc_result = cuMemHostAlloc(
-          &shared_pointer, size, CU_MEMHOSTALLOC_DEVICEMAP | CU_MEMHOSTALLOC_WRITECOMBINED);
-
-      assert((mem_alloc_result == CUDA_SUCCESS && shared_pointer != 0) ||
-             (mem_alloc_result != CUDA_SUCCESS && shared_pointer == 0));
-    }
-
-    if (mem_alloc_result == CUDA_SUCCESS) {
-      cuda_assert(cuMemHostGetDevicePointer_v2(&device_pointer, shared_pointer, 0));
-      map_host_used += size;
-      status = " in host memory";
-    }
-  }
-
-  if (mem_alloc_result != CUDA_SUCCESS) {
-    if (mem.type == MEM_DEVICE_ONLY) {
-      status = " failed, out of device memory";
-      set_error("System is out of GPU memory");
-    }
-    else {
-      status = " failed, out of device and host memory";
-      set_error("System is out of GPU and shared host memory");
-    }
-  }
-
-  if (mem.name) {
-    VLOG_WORK << "Buffer allocate: " << mem.name << ", "
-              << string_human_readable_number(mem.memory_size()) << " bytes. ("
-              << string_human_readable_size(mem.memory_size()) << ")" << status;
-  }
-
-  mem.device_pointer = (device_ptr)device_pointer;
-  mem.device_size = size;
-  stats.mem_alloc(size);
-
-  if (!mem.device_pointer) {
-    return NULL;
-  }
-
-  /* Insert into map of allocations. */
-  thread_scoped_lock lock(cuda_mem_map_mutex);
-  CUDAMem *cmem = &cuda_mem_map[&mem];
-  if (shared_pointer != 0) {
-    /* Replace host pointer with our host allocation. Only works if
-     * CUDA memory layout is the same and has no pitch padding. Also
-     * does not work if we move textures to host during a render,
-     * since other devices might be using the memory. */
-
-    if (!move_texture_to_host && pitch_padding == 0 && mem.host_pointer &&
-        mem.host_pointer != shared_pointer) {
-      memcpy(shared_pointer, mem.host_pointer, size);
-
-      /* A Call to device_memory::host_free() should be preceded by
-       * a call to device_memory::device_free() for host memory
-       * allocated by a device to be handled properly. Two exceptions
-       * are here and a call in OptiXDevice::generic_alloc(), where
-       * the current host memory can be assumed to be allocated by
-       * device_memory::host_alloc(), not by a device */
-
-      mem.host_free();
-      mem.host_pointer = shared_pointer;
-    }
-    mem.shared_pointer = shared_pointer;
-    mem.shared_counter++;
-    cmem->use_mapped_host = true;
-  }
-  else {
-    cmem->use_mapped_host = false;
-  }
-
-  return cmem;
 }
 
-void CUDADevice::generic_copy_to(device_memory &mem)
+bool CUDADevice::alloc_device(void *&device_pointer, size_t size)
 {
-  if (!mem.host_pointer || !mem.device_pointer) {
+  CUDAContextScope scope(this);
-    return;
-  }
 
-  /* If use_mapped_host of mem is false, the current device only uses device memory allocated by
+  CUresult mem_alloc_result = cuMemAlloc((CUdeviceptr *)&device_pointer, size);
-   * cuMemAlloc regardless of mem.host_pointer and mem.shared_pointer, and should copy data from
+  return mem_alloc_result == CUDA_SUCCESS;
-   * mem.host_pointer. */
-  thread_scoped_lock lock(cuda_mem_map_mutex);
-  if (!cuda_mem_map[&mem].use_mapped_host || mem.host_pointer != mem.shared_pointer) {
-    const CUDAContextScope scope(this);
-    cuda_assert(
-        cuMemcpyHtoD((CUdeviceptr)mem.device_pointer, mem.host_pointer, mem.memory_size()));
-  }
 }
 
-void CUDADevice::generic_free(device_memory &mem)
+void CUDADevice::free_device(void *device_pointer)
 {
-  if (mem.device_pointer) {
+  CUDAContextScope scope(this);
-    CUDAContextScope scope(this);
-    thread_scoped_lock lock(cuda_mem_map_mutex);
-    DCHECK(cuda_mem_map.find(&mem) != cuda_mem_map.end());
-    const CUDAMem &cmem = cuda_mem_map[&mem];
 
-    /* If cmem.use_mapped_host is true, reference counting is used
+  cuda_assert(cuMemFree((CUdeviceptr)device_pointer));
-     * to safely free a mapped host memory. */
+}
 
-    if (cmem.use_mapped_host) {
+bool CUDADevice::alloc_host(void *&shared_pointer, size_t size)
-      assert(mem.shared_pointer);
+{
-      if (mem.shared_pointer) {
+  CUDAContextScope scope(this);
-        assert(mem.shared_counter > 0);
-        if (--mem.shared_counter == 0) {
-          if (mem.host_pointer == mem.shared_pointer) {
-            mem.host_pointer = 0;
-          }
-          cuMemFreeHost(mem.shared_pointer);
-          mem.shared_pointer = 0;
-        }
-      }
-      map_host_used -= mem.device_size;
-    }
-    else {
-      /* Free device memory. */
-      cuda_assert(cuMemFree(mem.device_pointer));
-    }
 
-    stats.mem_free(mem.device_size);
+  CUresult mem_alloc_result = cuMemHostAlloc(
-    mem.device_pointer = 0;
+      &shared_pointer, size, CU_MEMHOSTALLOC_DEVICEMAP | CU_MEMHOSTALLOC_WRITECOMBINED);
-    mem.device_size = 0;
+  return mem_alloc_result == CUDA_SUCCESS;
+}
 
-    cuda_mem_map.erase(cuda_mem_map.find(&mem));
+void CUDADevice::free_host(void *shared_pointer)
-  }
+{
+  CUDAContextScope scope(this);
+
+  cuMemFreeHost(shared_pointer);
+}
+
+bool CUDADevice::transform_host_pointer(void *&device_pointer, void *&shared_pointer)
+{
+  CUDAContextScope scope(this);
+
+  cuda_assert(cuMemHostGetDevicePointer_v2((CUdeviceptr *)&device_pointer, shared_pointer, 0));
+  return true;
+}
+
+void CUDADevice::copy_host_to_device(void *device_pointer, void *host_pointer, size_t size)
+{
+  const CUDAContextScope scope(this);
+
+  cuda_assert(cuMemcpyHtoD((CUdeviceptr)device_pointer, host_pointer, size));
 }
 
 void CUDADevice::mem_alloc(device_memory &mem)
@@ -868,8 +613,8 @@ void CUDADevice::mem_zero(device_memory &mem)
 
   /* If use_mapped_host of mem is false, mem.device_pointer currently refers to device memory
    * regardless of mem.host_pointer and mem.shared_pointer. */
-  thread_scoped_lock lock(cuda_mem_map_mutex);
+  thread_scoped_lock lock(device_mem_map_mutex);
-  if (!cuda_mem_map[&mem].use_mapped_host || mem.host_pointer != mem.shared_pointer) {
+  if (!device_mem_map[&mem].use_mapped_host || mem.host_pointer != mem.shared_pointer) {
     const CUDAContextScope scope(this);
     cuda_assert(cuMemsetD8((CUdeviceptr)mem.device_pointer, 0, mem.memory_size()));
   }
@@ -994,19 +739,19 @@ void CUDADevice::tex_alloc(device_texture &mem)
       return;
   }
 
-  CUDAMem *cmem = NULL;
+  Mem *cmem = NULL;
   CUarray array_3d = NULL;
   size_t src_pitch = mem.data_width * dsize * mem.data_elements;
   size_t dst_pitch = src_pitch;
 
   if (!mem.is_resident(this)) {
-    thread_scoped_lock lock(cuda_mem_map_mutex);
+    thread_scoped_lock lock(device_mem_map_mutex);
-    cmem = &cuda_mem_map[&mem];
+    cmem = &device_mem_map[&mem];
     cmem->texobject = 0;
 
     if (mem.data_depth > 1) {
       array_3d = (CUarray)mem.device_pointer;
-      cmem->array = array_3d;
+      cmem->array = reinterpret_cast<arrayMemObject>(array_3d);
     }
     else if (mem.data_height > 0) {
       dst_pitch = align_up(src_pitch, pitch_alignment);
@@ -1050,10 +795,10 @@ void CUDADevice::tex_alloc(device_texture &mem)
     mem.device_size = size;
     stats.mem_alloc(size);
 
-    thread_scoped_lock lock(cuda_mem_map_mutex);
+    thread_scoped_lock lock(device_mem_map_mutex);
-    cmem = &cuda_mem_map[&mem];
+    cmem = &device_mem_map[&mem];
     cmem->texobject = 0;
-    cmem->array = array_3d;
+    cmem->array = reinterpret_cast<arrayMemObject>(array_3d);
   }
   else if (mem.data_height > 0) {
     /* 2D texture, using pitch aligned linear memory. */
@@ -1137,8 +882,8 @@ void CUDADevice::tex_alloc(device_texture &mem)
     texDesc.filterMode = filter_mode;
     texDesc.flags = CU_TRSF_NORMALIZED_COORDINATES;
 
-    thread_scoped_lock lock(cuda_mem_map_mutex);
+    thread_scoped_lock lock(device_mem_map_mutex);
-    cmem = &cuda_mem_map[&mem];
+    cmem = &device_mem_map[&mem];
 
     cuda_assert(cuTexObjectCreate(&cmem->texobject, &resDesc, &texDesc, NULL));
 
@@ -1153,9 +898,9 @@ void CUDADevice::tex_free(device_texture &mem)
 {
   if (mem.device_pointer) {
     CUDAContextScope scope(this);
-    thread_scoped_lock lock(cuda_mem_map_mutex);
+    thread_scoped_lock lock(device_mem_map_mutex);
-    DCHECK(cuda_mem_map.find(&mem) != cuda_mem_map.end());
+    DCHECK(device_mem_map.find(&mem) != device_mem_map.end());
-    const CUDAMem &cmem = cuda_mem_map[&mem];
+    const Mem &cmem = device_mem_map[&mem];
 
     if (cmem.texobject) {
       /* Free bindless texture. */
@@ -1164,16 +909,16 @@ void CUDADevice::tex_free(device_texture &mem)
 
     if (!mem.is_resident(this)) {
       /* Do not free memory here, since it was allocated on a different device. */
-      cuda_mem_map.erase(cuda_mem_map.find(&mem));
+      device_mem_map.erase(device_mem_map.find(&mem));
     }
     else if (cmem.array) {
       /* Free array. */
-      cuArrayDestroy(cmem.array);
+      cuArrayDestroy(reinterpret_cast<CUarray>(cmem.array));
       stats.mem_free(mem.device_size);
       mem.device_pointer = 0;
       mem.device_size = 0;
 
-      cuda_mem_map.erase(cuda_mem_map.find(&mem));
+      device_mem_map.erase(device_mem_map.find(&mem));
     }
     else {
       lock.unlock();

intern/cycles/device/cuda/device_impl.h

@@ -21,7 +21,7 @@ CCL_NAMESPACE_BEGIN
 
 class DeviceQueue;
 
-class CUDADevice : public Device {
+class CUDADevice : public GPUDevice {
 
   friend class CUDAContextScope;
 
@@ -29,36 +29,11 @@ class CUDADevice : public Device {
   CUdevice cuDevice;
   CUcontext cuContext;
   CUmodule cuModule;
-  size_t device_texture_headroom;
-  size_t device_working_headroom;
-  bool move_texture_to_host;
-  size_t map_host_used;
-  size_t map_host_limit;
-  int can_map_host;
   int pitch_alignment;
   int cuDevId;
   int cuDevArchitecture;
   bool first_error;
 
-  struct CUDAMem {
-    CUDAMem() : texobject(0), array(0), use_mapped_host(false)
-    {
-    }
-
-    CUtexObject texobject;
-    CUarray array;
-
-    /* If true, a mapped host memory in shared_pointer is being used. */
-    bool use_mapped_host;
-  };
-  typedef map<device_memory *, CUDAMem> CUDAMemMap;
-  CUDAMemMap cuda_mem_map;
-  thread_mutex cuda_mem_map_mutex;
-
-  /* Bindless Textures */
-  device_vector<TextureInfo> texture_info;
-  bool need_texture_info;
-
   CUDADeviceKernels kernels;
 
   static bool have_precompiled_kernels();
@@ -88,17 +63,13 @@ class CUDADevice : public Device {
 
   void reserve_local_memory(const uint kernel_features);
 
-  void init_host_memory();
+  virtual void get_device_memory_info(size_t &total, size_t &free) override;
-
+  virtual bool alloc_device(void *&device_pointer, size_t size) override;
-  void load_texture_info();
+  virtual void free_device(void *device_pointer) override;
-
+  virtual bool alloc_host(void *&shared_pointer, size_t size) override;
-  void move_textures_to_host(size_t size, bool for_texture);
+  virtual void free_host(void *shared_pointer) override;
-
+  virtual bool transform_host_pointer(void *&device_pointer, void *&shared_pointer) override;
-  CUDAMem *generic_alloc(device_memory &mem, size_t pitch_padding = 0);
+  virtual void copy_host_to_device(void *device_pointer, void *host_pointer, size_t size) override;
-
-  void generic_copy_to(device_memory &mem);
-
-  void generic_free(device_memory &mem);
 
   void mem_alloc(device_memory &mem) override;
 

intern/cycles/device/device.cpp

@@ -452,6 +452,320 @@ void *Device::get_cpu_osl_memory()
   return nullptr;
 }
 
+GPUDevice::~GPUDevice() noexcept(false)
+{
+}
+
+bool GPUDevice::load_texture_info()
+{
+  if (need_texture_info) {
+    /* Unset flag before copying, so this does not loop indefinitely if the copy below calls
+     * into 'move_textures_to_host' (which calls 'load_texture_info' again). */
+    need_texture_info = false;
+    texture_info.copy_to_device();
+    return true;
+  }
+  else {
+    return false;
+  }
+}
+
+void GPUDevice::init_host_memory(size_t preferred_texture_headroom,
+                                 size_t preferred_working_headroom)
+{
+  /* Limit amount of host mapped memory, because allocating too much can
+   * cause system instability. Leave at least half or 4 GB of system
+   * memory free, whichever is smaller. */
+  size_t default_limit = 4 * 1024 * 1024 * 1024LL;
+  size_t system_ram = system_physical_ram();
+
+  if (system_ram > 0) {
+    if (system_ram / 2 > default_limit) {
+      map_host_limit = system_ram - default_limit;
+    }
+    else {
+      map_host_limit = system_ram / 2;
+    }
+  }
+  else {
+    VLOG_WARNING << "Mapped host memory disabled, failed to get system RAM";
+    map_host_limit = 0;
+  }
+
+  /* Amount of device memory to keep free after texture memory
+   * and working memory allocations respectively. We set the working
+   * memory limit headroom lower than the working one so there
+   * is space left for it. */
+  device_working_headroom = preferred_working_headroom > 0 ? preferred_working_headroom :
+                                                             32 * 1024 * 1024LL;  // 32MB
+  device_texture_headroom = preferred_texture_headroom > 0 ? preferred_texture_headroom :
+                                                             128 * 1024 * 1024LL;  // 128MB
+
+  VLOG_INFO << "Mapped host memory limit set to " << string_human_readable_number(map_host_limit)
+            << " bytes. (" << string_human_readable_size(map_host_limit) << ")";
+}
+
+void GPUDevice::move_textures_to_host(size_t size, bool for_texture)
+{
+  /* Break out of recursive call, which can happen when moving memory on a multi device. */
+  static bool any_device_moving_textures_to_host = false;
+  if (any_device_moving_textures_to_host) {
+    return;
+  }
+
+  /* Signal to reallocate textures in host memory only. */
+  move_texture_to_host = true;
+
+  while (size > 0) {
+    /* Find suitable memory allocation to move. */
+    device_memory *max_mem = NULL;
+    size_t max_size = 0;
+    bool max_is_image = false;
+
+    thread_scoped_lock lock(device_mem_map_mutex);
+    foreach (MemMap::value_type &pair, device_mem_map) {
+      device_memory &mem = *pair.first;
+      Mem *cmem = &pair.second;
+
+      /* Can only move textures allocated on this device (and not those from peer devices).
+       * And need to ignore memory that is already on the host. */
+      if (!mem.is_resident(this) || cmem->use_mapped_host) {
+        continue;
+      }
+
+      bool is_texture = (mem.type == MEM_TEXTURE || mem.type == MEM_GLOBAL) &&
+                        (&mem != &texture_info);
+      bool is_image = is_texture && (mem.data_height > 1);
+
+      /* Can't move this type of memory. */
+      if (!is_texture || cmem->array) {
+        continue;
+      }
+
+      /* For other textures, only move image textures. */
+      if (for_texture && !is_image) {
+        continue;
+      }
+
+      /* Try to move largest allocation, prefer moving images. */
+      if (is_image > max_is_image || (is_image == max_is_image && mem.device_size > max_size)) {
+        max_is_image = is_image;
+        max_size = mem.device_size;
+        max_mem = &mem;
+      }
+    }
+    lock.unlock();
+
+    /* Move to host memory. This part is mutex protected since
+     * multiple backend devices could be moving the memory. The
+     * first one will do it, and the rest will adopt the pointer. */
+    if (max_mem) {
+      VLOG_WORK << "Move memory from device to host: " << max_mem->name;
+
+      static thread_mutex move_mutex;
+      thread_scoped_lock lock(move_mutex);
+
+      any_device_moving_textures_to_host = true;
+
+      /* Potentially need to call back into multi device, so pointer mapping
+       * and peer devices are updated. This is also necessary since the device
+       * pointer may just be a key here, so cannot be accessed and freed directly.
+       * Unfortunately it does mean that memory is reallocated on all other
+       * devices as well, which is potentially dangerous when still in use (since
+       * a thread rendering on another devices would only be caught in this mutex
+       * if it so happens to do an allocation at the same time as well. */
+      max_mem->device_copy_to();
+      size = (max_size >= size) ? 0 : size - max_size;
+
+      any_device_moving_textures_to_host = false;
+    }
+    else {
+      break;
+    }
+  }
+
+  /* Unset flag before texture info is reloaded, since it should stay in device memory. */
+  move_texture_to_host = false;
+
+  /* Update texture info array with new pointers. */
+  load_texture_info();
+}
+
+GPUDevice::Mem *GPUDevice::generic_alloc(device_memory &mem, size_t pitch_padding)
+{
+  void *device_pointer = 0;
+  size_t size = mem.memory_size() + pitch_padding;
+
+  bool mem_alloc_result = false;
+  const char *status = "";
+
+  /* First try allocating in device memory, respecting headroom. We make
+   * an exception for texture info. It is small and frequently accessed,
+   * so treat it as working memory.
+   *
+   * If there is not enough room for working memory, we will try to move
+   * textures to host memory, assuming the performance impact would have
+   * been worse for working memory. */
+  bool is_texture = (mem.type == MEM_TEXTURE || mem.type == MEM_GLOBAL) && (&mem != &texture_info);
+  bool is_image = is_texture && (mem.data_height > 1);
+
+  size_t headroom = (is_texture) ? device_texture_headroom : device_working_headroom;
+
+  size_t total = 0, free = 0;
+  get_device_memory_info(total, free);
+
+  /* Move textures to host memory if needed. */
+  if (!move_texture_to_host && !is_image && (size + headroom) >= free && can_map_host) {
+    move_textures_to_host(size + headroom - free, is_texture);
+    get_device_memory_info(total, free);
+  }
+
+  /* Allocate in device memory. */
+  if (!move_texture_to_host && (size + headroom) < free) {
+    mem_alloc_result = alloc_device(device_pointer, size);
+    if (mem_alloc_result) {
+      device_mem_in_use += size;
+      status = " in device memory";
+    }
+  }
+
+  /* Fall back to mapped host memory if needed and possible. */
+
+  void *shared_pointer = 0;
+
+  if (!mem_alloc_result && can_map_host && mem.type != MEM_DEVICE_ONLY) {
+    if (mem.shared_pointer) {
+      /* Another device already allocated host memory. */
+      mem_alloc_result = true;
+      shared_pointer = mem.shared_pointer;
+    }
+    else if (map_host_used + size < map_host_limit) {
+      /* Allocate host memory ourselves. */
+      mem_alloc_result = alloc_host(shared_pointer, size);
+
+      assert((mem_alloc_result && shared_pointer != 0) ||
+             (!mem_alloc_result && shared_pointer == 0));
+    }
+
+    if (mem_alloc_result) {
+      assert(transform_host_pointer(device_pointer, shared_pointer));
+      map_host_used += size;
+      status = " in host memory";
+    }
+  }
+
+  if (!mem_alloc_result) {
+    if (mem.type == MEM_DEVICE_ONLY) {
+      status = " failed, out of device memory";
+      set_error("System is out of GPU memory");
+    }
+    else {
+      status = " failed, out of device and host memory";
+      set_error("System is out of GPU and shared host memory");
+    }
+  }
+
+  if (mem.name) {
+    VLOG_WORK << "Buffer allocate: " << mem.name << ", "
+              << string_human_readable_number(mem.memory_size()) << " bytes. ("
+              << string_human_readable_size(mem.memory_size()) << ")" << status;
+  }
+
+  mem.device_pointer = (device_ptr)device_pointer;
+  mem.device_size = size;
+  stats.mem_alloc(size);
+
+  if (!mem.device_pointer) {
+    return NULL;
+  }
+
+  /* Insert into map of allocations. */
+  thread_scoped_lock lock(device_mem_map_mutex);
+  Mem *cmem = &device_mem_map[&mem];
+  if (shared_pointer != 0) {
+    /* Replace host pointer with our host allocation. Only works if
+     * memory layout is the same and has no pitch padding. Also
+     * does not work if we move textures to host during a render,
+     * since other devices might be using the memory. */
+
+    if (!move_texture_to_host && pitch_padding == 0 && mem.host_pointer &&
+        mem.host_pointer != shared_pointer) {
+      memcpy(shared_pointer, mem.host_pointer, size);
+
+      /* A Call to device_memory::host_free() should be preceded by
+       * a call to device_memory::device_free() for host memory
+       * allocated by a device to be handled properly. Two exceptions
+       * are here and a call in OptiXDevice::generic_alloc(), where
+       * the current host memory can be assumed to be allocated by
+       * device_memory::host_alloc(), not by a device */
+
+      mem.host_free();
+      mem.host_pointer = shared_pointer;
+    }
+    mem.shared_pointer = shared_pointer;
+    mem.shared_counter++;
+    cmem->use_mapped_host = true;
+  }
+  else {
+    cmem->use_mapped_host = false;
+  }
+
+  return cmem;
+}
+
+void GPUDevice::generic_free(device_memory &mem)
+{
+  if (mem.device_pointer) {
+    thread_scoped_lock lock(device_mem_map_mutex);
+    DCHECK(device_mem_map.find(&mem) != device_mem_map.end());
+    const Mem &cmem = device_mem_map[&mem];
+
+    /* If cmem.use_mapped_host is true, reference counting is used
+     * to safely free a mapped host memory. */
+
+    if (cmem.use_mapped_host) {
+      assert(mem.shared_pointer);
+      if (mem.shared_pointer) {
+        assert(mem.shared_counter > 0);
+        if (--mem.shared_counter == 0) {
+          if (mem.host_pointer == mem.shared_pointer) {
+            mem.host_pointer = 0;
+          }
+          free_host(mem.shared_pointer);
+          mem.shared_pointer = 0;
+        }
+      }
+      map_host_used -= mem.device_size;
+    }
+    else {
+      /* Free device memory. */
+      free_device((void *)mem.device_pointer);
+      device_mem_in_use -= mem.device_size;
+    }
+
+    stats.mem_free(mem.device_size);
+    mem.device_pointer = 0;
+    mem.device_size = 0;
+
+    device_mem_map.erase(device_mem_map.find(&mem));
+  }
+}
+
+void GPUDevice::generic_copy_to(device_memory &mem)
+{
+  if (!mem.host_pointer || !mem.device_pointer) {
+    return;
+  }
+
+  /* If use_mapped_host of mem is false, the current device only uses device memory allocated by
+   * backend device allocation regardless of mem.host_pointer and mem.shared_pointer, and should
+   * copy data from mem.host_pointer. */
+  thread_scoped_lock lock(device_mem_map_mutex);
+  if (!device_mem_map[&mem].use_mapped_host || mem.host_pointer != mem.shared_pointer) {
+    copy_host_to_device((void *)mem.device_pointer, mem.host_pointer, mem.memory_size());
+  }
+}
+
 /* DeviceInfo */
 
 CCL_NAMESPACE_END

intern/cycles/device/device.h

@@ -309,6 +309,93 @@ class Device {
   static uint devices_initialized_mask;
 };
 
+/* Device, which is GPU, with some common functionality for GPU backends */
+class GPUDevice : public Device {
+ protected:
+  GPUDevice(const DeviceInfo &info_, Stats &stats_, Profiler &profiler_)
+      : Device(info_, stats_, profiler_),
+        texture_info(this, "texture_info", MEM_GLOBAL),
+        need_texture_info(false),
+        can_map_host(false),
+        map_host_used(0),
+        map_host_limit(0),
+        device_texture_headroom(0),
+        device_working_headroom(0),
+        device_mem_map(),
+        device_mem_map_mutex(),
+        move_texture_to_host(false),
+        device_mem_in_use(0)
+  {
+  }
+
+ public:
+  virtual ~GPUDevice() noexcept(false);
+
+  /* For GPUs that can use bindless textures in some way or another. */
+  device_vector<TextureInfo> texture_info;
+  bool need_texture_info;
+  /* Returns true if the texture info was copied to the device (meaning, some more
+   * re-initialization might be needed). */
+  virtual bool load_texture_info();
+
+ protected:
+  /* Memory allocation, only accessed through device_memory. */
+  friend class device_memory;
+
+  bool can_map_host;
+  size_t map_host_used;
+  size_t map_host_limit;
+  size_t device_texture_headroom;
+  size_t device_working_headroom;
+  typedef unsigned long long texMemObject;
+  typedef unsigned long long arrayMemObject;
+  struct Mem {
+    Mem() : texobject(0), array(0), use_mapped_host(false)
+    {
+    }
+
+    texMemObject texobject;
+    arrayMemObject array;
+
+    /* If true, a mapped host memory in shared_pointer is being used. */
+    bool use_mapped_host;
+  };
+  typedef map<device_memory *, Mem> MemMap;
+  MemMap device_mem_map;
+  thread_mutex device_mem_map_mutex;
+  bool move_texture_to_host;
+  /* Simple counter which will try to track amount of used device memory */
+  size_t device_mem_in_use;
+
+  virtual void init_host_memory(size_t preferred_texture_headroom = 0,
+                                size_t preferred_working_headroom = 0);
+  virtual void move_textures_to_host(size_t size, bool for_texture);
+
+  /* Allocation, deallocation and copy functions, with corresponding
+   * support of device/host allocations. */
+  virtual GPUDevice::Mem *generic_alloc(device_memory &mem, size_t pitch_padding = 0);
+  virtual void generic_free(device_memory &mem);
+  virtual void generic_copy_to(device_memory &mem);
+
+  /* total - amount of device memory, free - amount of available device memory */
+  virtual void get_device_memory_info(size_t &total, size_t &free) = 0;
+
+  virtual bool alloc_device(void *&device_pointer, size_t size) = 0;
+
+  virtual void free_device(void *device_pointer) = 0;
+
+  virtual bool alloc_host(void *&shared_pointer, size_t size) = 0;
+
+  virtual void free_host(void *shared_pointer) = 0;
+
+  /* This function should return device pointer corresponding to shared pointer, which
+   * is host buffer, allocated in `alloc_host`. The function should `true`, if such
+   * address transformation is possible and `false` otherwise. */
+  virtual bool transform_host_pointer(void *&device_pointer, void *&shared_pointer) = 0;
+
+  virtual void copy_host_to_device(void *device_pointer, void *host_pointer, size_t size) = 0;
+};
+
 CCL_NAMESPACE_END
 
 #endif /* __DEVICE_H__ */

intern/cycles/device/hip/device_impl.cpp

@@ -53,8 +53,12 @@ void HIPDevice::set_error(const string &error)
 }
 
 HIPDevice::HIPDevice(const DeviceInfo &info, Stats &stats, Profiler &profiler)
-    : Device(info, stats, profiler), texture_info(this, "texture_info", MEM_GLOBAL)
+    : GPUDevice(info, stats, profiler)
 {
+  /* Verify that base class types can be used with specific backend types */
+  static_assert(sizeof(texMemObject) == sizeof(hipTextureObject_t));
+  static_assert(sizeof(arrayMemObject) == sizeof(hArray));
+
   first_error = true;
 
   hipDevId = info.num;
@@ -65,12 +69,6 @@ HIPDevice::HIPDevice(const DeviceInfo &info, Stats &stats, Profiler &profiler)
 
   need_texture_info = false;
 
-  device_texture_headroom = 0;
-  device_working_headroom = 0;
-  move_texture_to_host = false;
-  map_host_limit = 0;
-  map_host_used = 0;
-  can_map_host = 0;
   pitch_alignment = 0;
 
   /* Initialize HIP. */
@@ -91,7 +89,9 @@ HIPDevice::HIPDevice(const DeviceInfo &info, Stats &stats, Profiler &profiler)
   /* hipDeviceMapHost for mapping host memory when out of device memory.
    * hipDeviceLmemResizeToMax for reserving local memory ahead of render,
    * so we can predict which memory to map to host. */
-  hip_assert(hipDeviceGetAttribute(&can_map_host, hipDeviceAttributeCanMapHostMemory, hipDevice));
+  int value;
+  hip_assert(hipDeviceGetAttribute(&value, hipDeviceAttributeCanMapHostMemory, hipDevice));
+  can_map_host = value != 0;
 
   hip_assert(
       hipDeviceGetAttribute(&pitch_alignment, hipDeviceAttributeTexturePitchAlignment, hipDevice));
@@ -460,305 +460,58 @@ void HIPDevice::reserve_local_memory(const uint kernel_features)
 #  endif
 }
 
-void HIPDevice::init_host_memory()
+void HIPDevice::get_device_memory_info(size_t &total, size_t &free)
-{
-  /* Limit amount of host mapped memory, because allocating too much can
-   * cause system instability. Leave at least half or 4 GB of system
-   * memory free, whichever is smaller. */
-  size_t default_limit = 4 * 1024 * 1024 * 1024LL;
-  size_t system_ram = system_physical_ram();
-
-  if (system_ram > 0) {
-    if (system_ram / 2 > default_limit) {
-      map_host_limit = system_ram - default_limit;
-    }
-    else {
-      map_host_limit = system_ram / 2;
-    }
-  }
-  else {
-    VLOG_WARNING << "Mapped host memory disabled, failed to get system RAM";
-    map_host_limit = 0;
-  }
-
-  /* Amount of device memory to keep is free after texture memory
-   * and working memory allocations respectively. We set the working
-   * memory limit headroom lower so that some space is left after all
-   * texture memory allocations. */
-  device_working_headroom = 32 * 1024 * 1024LL;   // 32MB
-  device_texture_headroom = 128 * 1024 * 1024LL;  // 128MB
-
-  VLOG_INFO << "Mapped host memory limit set to " << string_human_readable_number(map_host_limit)
-            << " bytes. (" << string_human_readable_size(map_host_limit) << ")";
-}
-
-void HIPDevice::load_texture_info()
-{
-  if (need_texture_info) {
-    /* Unset flag before copying, so this does not loop indefinitely if the copy below calls
-     * into 'move_textures_to_host' (which calls 'load_texture_info' again). */
-    need_texture_info = false;
-    texture_info.copy_to_device();
-  }
-}
-
-void HIPDevice::move_textures_to_host(size_t size, bool for_texture)
-{
-  /* Break out of recursive call, which can happen when moving memory on a multi device. */
-  static bool any_device_moving_textures_to_host = false;
-  if (any_device_moving_textures_to_host) {
-    return;
-  }
-
-  /* Signal to reallocate textures in host memory only. */
-  move_texture_to_host = true;
-
-  while (size > 0) {
-    /* Find suitable memory allocation to move. */
-    device_memory *max_mem = NULL;
-    size_t max_size = 0;
-    bool max_is_image = false;
-
-    thread_scoped_lock lock(hip_mem_map_mutex);
-    foreach (HIPMemMap::value_type &pair, hip_mem_map) {
-      device_memory &mem = *pair.first;
-      HIPMem *cmem = &pair.second;
-
-      /* Can only move textures allocated on this device (and not those from peer devices).
-       * And need to ignore memory that is already on the host. */
-      if (!mem.is_resident(this) || cmem->use_mapped_host) {
-        continue;
-      }
-
-      bool is_texture = (mem.type == MEM_TEXTURE || mem.type == MEM_GLOBAL) &&
-                        (&mem != &texture_info);
-      bool is_image = is_texture && (mem.data_height > 1);
-
-      /* Can't move this type of memory. */
-      if (!is_texture || cmem->array) {
-        continue;
-      }
-
-      /* For other textures, only move image textures. */
-      if (for_texture && !is_image) {
-        continue;
-      }
-
-      /* Try to move largest allocation, prefer moving images. */
-      if (is_image > max_is_image || (is_image == max_is_image && mem.device_size > max_size)) {
-        max_is_image = is_image;
-        max_size = mem.device_size;
-        max_mem = &mem;
-      }
-    }
-    lock.unlock();
-
-    /* Move to host memory. This part is mutex protected since
-     * multiple HIP devices could be moving the memory. The
-     * first one will do it, and the rest will adopt the pointer. */
-    if (max_mem) {
-      VLOG_WORK << "Move memory from device to host: " << max_mem->name;
-
-      static thread_mutex move_mutex;
-      thread_scoped_lock lock(move_mutex);
-
-      any_device_moving_textures_to_host = true;
-
-      /* Potentially need to call back into multi device, so pointer mapping
-       * and peer devices are updated. This is also necessary since the device
-       * pointer may just be a key here, so cannot be accessed and freed directly.
-       * Unfortunately it does mean that memory is reallocated on all other
-       * devices as well, which is potentially dangerous when still in use (since
-       * a thread rendering on another devices would only be caught in this mutex
-       * if it so happens to do an allocation at the same time as well. */
-      max_mem->device_copy_to();
-      size = (max_size >= size) ? 0 : size - max_size;
-
-      any_device_moving_textures_to_host = false;
-    }
-    else {
-      break;
-    }
-  }
-
-  /* Unset flag before texture info is reloaded, since it should stay in device memory. */
-  move_texture_to_host = false;
-
-  /* Update texture info array with new pointers. */
-  load_texture_info();
-}
-
-HIPDevice::HIPMem *HIPDevice::generic_alloc(device_memory &mem, size_t pitch_padding)
 {
   HIPContextScope scope(this);
 
-  hipDeviceptr_t device_pointer = 0;
-  size_t size = mem.memory_size() + pitch_padding;
-
-  hipError_t mem_alloc_result = hipErrorOutOfMemory;
-  const char *status = "";
-
-  /* First try allocating in device memory, respecting headroom. We make
-   * an exception for texture info. It is small and frequently accessed,
-   * so treat it as working memory.
-   *
-   * If there is not enough room for working memory, we will try to move
-   * textures to host memory, assuming the performance impact would have
-   * been worse for working memory. */
-  bool is_texture = (mem.type == MEM_TEXTURE || mem.type == MEM_GLOBAL) && (&mem != &texture_info);
-  bool is_image = is_texture && (mem.data_height > 1);
-
-  size_t headroom = (is_texture) ? device_texture_headroom : device_working_headroom;
-
-  size_t total = 0, free = 0;
   hipMemGetInfo(&free, &total);
-
-  /* Move textures to host memory if needed. */
-  if (!move_texture_to_host && !is_image && (size + headroom) >= free && can_map_host) {
-    move_textures_to_host(size + headroom - free, is_texture);
-    hipMemGetInfo(&free, &total);
-  }
-
-  /* Allocate in device memory. */
-  if (!move_texture_to_host && (size + headroom) < free) {
-    mem_alloc_result = hipMalloc(&device_pointer, size);
-    if (mem_alloc_result == hipSuccess) {
-      status = " in device memory";
-    }
-  }
-
-  /* Fall back to mapped host memory if needed and possible. */
-
-  void *shared_pointer = 0;
-
-  if (mem_alloc_result != hipSuccess && can_map_host) {
-    if (mem.shared_pointer) {
-      /* Another device already allocated host memory. */
-      mem_alloc_result = hipSuccess;
-      shared_pointer = mem.shared_pointer;
-    }
-    else if (map_host_used + size < map_host_limit) {
-      /* Allocate host memory ourselves. */
-      mem_alloc_result = hipHostMalloc(
-          &shared_pointer, size, hipHostMallocMapped | hipHostMallocWriteCombined);
-
-      assert((mem_alloc_result == hipSuccess && shared_pointer != 0) ||
-             (mem_alloc_result != hipSuccess && shared_pointer == 0));
-    }
-
-    if (mem_alloc_result == hipSuccess) {
-      hip_assert(hipHostGetDevicePointer(&device_pointer, shared_pointer, 0));
-      map_host_used += size;
-      status = " in host memory";
-    }
-  }
-
-  if (mem_alloc_result != hipSuccess) {
-    status = " failed, out of device and host memory";
-    set_error("System is out of GPU and shared host memory");
-  }
-
-  if (mem.name) {
-    VLOG_WORK << "Buffer allocate: " << mem.name << ", "
-              << string_human_readable_number(mem.memory_size()) << " bytes. ("
-              << string_human_readable_size(mem.memory_size()) << ")" << status;
-  }
-
-  mem.device_pointer = (device_ptr)device_pointer;
-  mem.device_size = size;
-  stats.mem_alloc(size);
-
-  if (!mem.device_pointer) {
-    return NULL;
-  }
-
-  /* Insert into map of allocations. */
-  thread_scoped_lock lock(hip_mem_map_mutex);
-  HIPMem *cmem = &hip_mem_map[&mem];
-  if (shared_pointer != 0) {
-    /* Replace host pointer with our host allocation. Only works if
-     * HIP memory layout is the same and has no pitch padding. Also
-     * does not work if we move textures to host during a render,
-     * since other devices might be using the memory. */
-
-    if (!move_texture_to_host && pitch_padding == 0 && mem.host_pointer &&
-        mem.host_pointer != shared_pointer) {
-      memcpy(shared_pointer, mem.host_pointer, size);
-
-      /* A Call to device_memory::host_free() should be preceded by
-       * a call to device_memory::device_free() for host memory
-       * allocated by a device to be handled properly. Two exceptions
-       * are here and a call in OptiXDevice::generic_alloc(), where
-       * the current host memory can be assumed to be allocated by
-       * device_memory::host_alloc(), not by a device */
-
-      mem.host_free();
-      mem.host_pointer = shared_pointer;
-    }
-    mem.shared_pointer = shared_pointer;
-    mem.shared_counter++;
-    cmem->use_mapped_host = true;
-  }
-  else {
-    cmem->use_mapped_host = false;
-  }
-
-  return cmem;
 }
 
-void HIPDevice::generic_copy_to(device_memory &mem)
+bool HIPDevice::alloc_device(void *&device_pointer, size_t size)
 {
-  if (!mem.host_pointer || !mem.device_pointer) {
+  HIPContextScope scope(this);
-    return;
-  }
 
-  /* If use_mapped_host of mem is false, the current device only uses device memory allocated by
+  hipError_t mem_alloc_result = hipMalloc((hipDeviceptr_t *)&device_pointer, size);
-   * hipMalloc regardless of mem.host_pointer and mem.shared_pointer, and should copy data from
+  return mem_alloc_result == hipSuccess;
-   * mem.host_pointer. */
-  thread_scoped_lock lock(hip_mem_map_mutex);
-  if (!hip_mem_map[&mem].use_mapped_host || mem.host_pointer != mem.shared_pointer) {
-    const HIPContextScope scope(this);
-    hip_assert(
-        hipMemcpyHtoD((hipDeviceptr_t)mem.device_pointer, mem.host_pointer, mem.memory_size()));
-  }
 }
 
-void HIPDevice::generic_free(device_memory &mem)
+void HIPDevice::free_device(void *device_pointer)
 {
-  if (mem.device_pointer) {
+  HIPContextScope scope(this);
-    HIPContextScope scope(this);
-    thread_scoped_lock lock(hip_mem_map_mutex);
-    DCHECK(hip_mem_map.find(&mem) != hip_mem_map.end());
-    const HIPMem &cmem = hip_mem_map[&mem];
 
-    /* If cmem.use_mapped_host is true, reference counting is used
+  hip_assert(hipFree((hipDeviceptr_t)device_pointer));
-     * to safely free a mapped host memory. */
+}
 
-    if (cmem.use_mapped_host) {
+bool HIPDevice::alloc_host(void *&shared_pointer, size_t size)
-      assert(mem.shared_pointer);
+{
-      if (mem.shared_pointer) {
+  HIPContextScope scope(this);
-        assert(mem.shared_counter > 0);
-        if (--mem.shared_counter == 0) {
-          if (mem.host_pointer == mem.shared_pointer) {
-            mem.host_pointer = 0;
-          }
-          hipHostFree(mem.shared_pointer);
-          mem.shared_pointer = 0;
-        }
-      }
-      map_host_used -= mem.device_size;
-    }
-    else {
-      /* Free device memory. */
-      hip_assert(hipFree(mem.device_pointer));
-    }
 
-    stats.mem_free(mem.device_size);
+  hipError_t mem_alloc_result = hipHostMalloc(
-    mem.device_pointer = 0;
+      &shared_pointer, size, hipHostMallocMapped | hipHostMallocWriteCombined);
-    mem.device_size = 0;
 
-    hip_mem_map.erase(hip_mem_map.find(&mem));
+  return mem_alloc_result == hipSuccess;
-  }
+}
+
+void HIPDevice::free_host(void *shared_pointer)
+{
+  HIPContextScope scope(this);
+
+  hipHostFree(shared_pointer);
+}
+
+bool HIPDevice::transform_host_pointer(void *&device_pointer, void *&shared_pointer)
+{
+  HIPContextScope scope(this);
+
+  hip_assert(hipHostGetDevicePointer((hipDeviceptr_t *)&device_pointer, shared_pointer, 0));
+  return true;
+}
+
+void HIPDevice::copy_host_to_device(void *device_pointer, void *host_pointer, size_t size)
+{
+  const HIPContextScope scope(this);
+
+  hip_assert(hipMemcpyHtoD((hipDeviceptr_t)device_pointer, host_pointer, size));
 }
 
 void HIPDevice::mem_alloc(device_memory &mem)
@@ -823,8 +576,8 @@ void HIPDevice::mem_zero(device_memory &mem)
 
   /* If use_mapped_host of mem is false, mem.device_pointer currently refers to device memory
    * regardless of mem.host_pointer and mem.shared_pointer. */
-  thread_scoped_lock lock(hip_mem_map_mutex);
+  thread_scoped_lock lock(device_mem_map_mutex);
-  if (!hip_mem_map[&mem].use_mapped_host || mem.host_pointer != mem.shared_pointer) {
+  if (!device_mem_map[&mem].use_mapped_host || mem.host_pointer != mem.shared_pointer) {
     const HIPContextScope scope(this);
     hip_assert(hipMemsetD8((hipDeviceptr_t)mem.device_pointer, 0, mem.memory_size()));
   }
@@ -951,19 +704,19 @@ void HIPDevice::tex_alloc(device_texture &mem)
       return;
   }
 
-  HIPMem *cmem = NULL;
+  Mem *cmem = NULL;
   hArray array_3d = NULL;
   size_t src_pitch = mem.data_width * dsize * mem.data_elements;
   size_t dst_pitch = src_pitch;
 
   if (!mem.is_resident(this)) {
-    thread_scoped_lock lock(hip_mem_map_mutex);
+    thread_scoped_lock lock(device_mem_map_mutex);
-    cmem = &hip_mem_map[&mem];
+    cmem = &device_mem_map[&mem];
     cmem->texobject = 0;
 
     if (mem.data_depth > 1) {
       array_3d = (hArray)mem.device_pointer;
-      cmem->array = array_3d;
+      cmem->array = reinterpret_cast<arrayMemObject>(array_3d);
     }
     else if (mem.data_height > 0) {
       dst_pitch = align_up(src_pitch, pitch_alignment);
@@ -1007,10 +760,10 @@ void HIPDevice::tex_alloc(device_texture &mem)
     mem.device_size = size;
     stats.mem_alloc(size);
 
-    thread_scoped_lock lock(hip_mem_map_mutex);
+    thread_scoped_lock lock(device_mem_map_mutex);
-    cmem = &hip_mem_map[&mem];
+    cmem = &device_mem_map[&mem];
     cmem->texobject = 0;
-    cmem->array = array_3d;
+    cmem->array = reinterpret_cast<arrayMemObject>(array_3d);
   }
   else if (mem.data_height > 0) {
     /* 2D texture, using pitch aligned linear memory. */
@@ -1095,8 +848,8 @@ void HIPDevice::tex_alloc(device_texture &mem)
     texDesc.filterMode = filter_mode;
     texDesc.flags = HIP_TRSF_NORMALIZED_COORDINATES;
 
-    thread_scoped_lock lock(hip_mem_map_mutex);
+    thread_scoped_lock lock(device_mem_map_mutex);
-    cmem = &hip_mem_map[&mem];
+    cmem = &device_mem_map[&mem];
 
     hip_assert(hipTexObjectCreate(&cmem->texobject, &resDesc, &texDesc, NULL));
 
@@ -1111,9 +864,9 @@ void HIPDevice::tex_free(device_texture &mem)
 {
   if (mem.device_pointer) {
     HIPContextScope scope(this);
-    thread_scoped_lock lock(hip_mem_map_mutex);
+    thread_scoped_lock lock(device_mem_map_mutex);
-    DCHECK(hip_mem_map.find(&mem) != hip_mem_map.end());
+    DCHECK(device_mem_map.find(&mem) != device_mem_map.end());
-    const HIPMem &cmem = hip_mem_map[&mem];
+    const Mem &cmem = device_mem_map[&mem];
 
     if (cmem.texobject) {
       /* Free bindless texture. */
@@ -1122,16 +875,16 @@ void HIPDevice::tex_free(device_texture &mem)
 
     if (!mem.is_resident(this)) {
       /* Do not free memory here, since it was allocated on a different device. */
-      hip_mem_map.erase(hip_mem_map.find(&mem));
+      device_mem_map.erase(device_mem_map.find(&mem));
     }
     else if (cmem.array) {
       /* Free array. */
-      hipArrayDestroy(cmem.array);
+      hipArrayDestroy(reinterpret_cast<hArray>(cmem.array));
       stats.mem_free(mem.device_size);
       mem.device_pointer = 0;
       mem.device_size = 0;
 
-      hip_mem_map.erase(hip_mem_map.find(&mem));
+      device_mem_map.erase(device_mem_map.find(&mem));
     }
     else {
       lock.unlock();
@@ -1153,7 +906,7 @@ bool HIPDevice::should_use_graphics_interop()
    * possible, but from the empiric measurements it can be considerably slower than using naive
    * pixels copy. */
 
-  /* Disable graphics interop for now, because of driver bug in 21.40. See T92972 */
+  /* Disable graphics interop for now, because of driver bug in 21.40. See #92972 */
 #  if 0
   HIPContextScope scope(this);
 

intern/cycles/device/hip/device_impl.h

@@ -18,7 +18,7 @@ CCL_NAMESPACE_BEGIN
 
 class DeviceQueue;
 
-class HIPDevice : public Device {
+class HIPDevice : public GPUDevice {
 
   friend class HIPContextScope;
 
@@ -26,36 +26,11 @@ class HIPDevice : public Device {
   hipDevice_t hipDevice;
   hipCtx_t hipContext;
   hipModule_t hipModule;
-  size_t device_texture_headroom;
-  size_t device_working_headroom;
-  bool move_texture_to_host;
-  size_t map_host_used;
-  size_t map_host_limit;
-  int can_map_host;
   int pitch_alignment;
   int hipDevId;
   int hipDevArchitecture;
   bool first_error;
 
-  struct HIPMem {
-    HIPMem() : texobject(0), array(0), use_mapped_host(false)
-    {
-    }
-
-    hipTextureObject_t texobject;
-    hArray array;
-
-    /* If true, a mapped host memory in shared_pointer is being used. */
-    bool use_mapped_host;
-  };
-  typedef map<device_memory *, HIPMem> HIPMemMap;
-  HIPMemMap hip_mem_map;
-  thread_mutex hip_mem_map_mutex;
-
-  /* Bindless Textures */
-  device_vector<TextureInfo> texture_info;
-  bool need_texture_info;
-
   HIPDeviceKernels kernels;
 
   static bool have_precompiled_kernels();
@@ -81,17 +56,13 @@ class HIPDevice : public Device {
   virtual bool load_kernels(const uint kernel_features) override;
   void reserve_local_memory(const uint kernel_features);
 
-  void init_host_memory();
+  virtual void get_device_memory_info(size_t &total, size_t &free) override;
-
+  virtual bool alloc_device(void *&device_pointer, size_t size) override;
-  void load_texture_info();
+  virtual void free_device(void *device_pointer) override;
-
+  virtual bool alloc_host(void *&shared_pointer, size_t size) override;
-  void move_textures_to_host(size_t size, bool for_texture);
+  virtual void free_host(void *shared_pointer) override;
-
+  virtual bool transform_host_pointer(void *&device_pointer, void *&shared_pointer) override;
-  HIPMem *generic_alloc(device_memory &mem, size_t pitch_padding = 0);
+  virtual void copy_host_to_device(void *device_pointer, void *host_pointer, size_t size) override;
-
-  void generic_copy_to(device_memory &mem);
-
-  void generic_free(device_memory &mem);
 
   void mem_alloc(device_memory &mem) override;
 

intern/cycles/device/hip/util.h

@@ -51,7 +51,7 @@ static inline bool hipSupportsDevice(const int hipDevId)
   hipDeviceGetAttribute(&major, hipDeviceAttributeComputeCapabilityMajor, hipDevId);
   hipDeviceGetAttribute(&minor, hipDeviceAttributeComputeCapabilityMinor, hipDevId);
 
-  return (major >= 10);
+  return (major >= 9);
 }
 
 CCL_NAMESPACE_END

intern/cycles/device/kernel.cpp

@@ -73,6 +73,10 @@ const char *device_kernel_as_string(DeviceKernel kernel)
       return "integrator_terminated_paths_array";
     case DEVICE_KERNEL_INTEGRATOR_SORTED_PATHS_ARRAY:
       return "integrator_sorted_paths_array";
+    case DEVICE_KERNEL_INTEGRATOR_SORT_BUCKET_PASS:
+      return "integrator_sort_bucket_pass";
+    case DEVICE_KERNEL_INTEGRATOR_SORT_WRITE_PASS:
+      return "integrator_sort_write_pass";
     case DEVICE_KERNEL_INTEGRATOR_COMPACT_PATHS_ARRAY:
       return "integrator_compact_paths_array";
     case DEVICE_KERNEL_INTEGRATOR_COMPACT_STATES:

intern/cycles/device/memory.h

@@ -247,6 +247,8 @@ class device_memory {
   bool is_resident(Device *sub_device) const;
 
  protected:
+  friend class Device;
+  friend class GPUDevice;
   friend class CUDADevice;
   friend class OptiXDevice;
   friend class HIPDevice;

intern/cycles/device/metal/bvh.h

@@ -21,6 +21,7 @@ class BVHMetal : public BVH {
 
   API_AVAILABLE(macos(11.0))
   vector<id<MTLAccelerationStructure>> blas_array;
+  vector<uint32_t> blas_lookup;
 
   bool motion_blur = false;
 

intern/cycles/device/metal/bvh.mm

@@ -816,6 +816,11 @@ bool BVHMetal::build_TLAS(Progress &progress,
 
     uint32_t instance_index = 0;
     uint32_t motion_transform_index = 0;
+
+    // allocate look up buffer for wost case scenario
+    uint64_t count = objects.size();
+    blas_lookup.resize(count);
+
     for (Object *ob : objects) {
       /* Skip non-traceable objects */
       if (!ob->is_traceable())
@@ -843,12 +848,15 @@ bool BVHMetal::build_TLAS(Progress &progress,
       /* Set user instance ID to object index */
       int object_index = ob->get_device_index();
       uint32_t user_id = uint32_t(object_index);
+      int currIndex = instance_index++;
+      assert(user_id < blas_lookup.size());
+      blas_lookup[user_id] = accel_struct_index;
 
       /* Bake into the appropriate descriptor */
       if (motion_blur) {
         MTLAccelerationStructureMotionInstanceDescriptor *instances =
             (MTLAccelerationStructureMotionInstanceDescriptor *)[instanceBuf contents];
-        MTLAccelerationStructureMotionInstanceDescriptor &desc = instances[instance_index++];
+        MTLAccelerationStructureMotionInstanceDescriptor &desc = instances[currIndex];
 
         desc.accelerationStructureIndex = accel_struct_index;
         desc.userID = user_id;
@@ -894,7 +902,7 @@ bool BVHMetal::build_TLAS(Progress &progress,
       else {
         MTLAccelerationStructureUserIDInstanceDescriptor *instances =
             (MTLAccelerationStructureUserIDInstanceDescriptor *)[instanceBuf contents];
-        MTLAccelerationStructureUserIDInstanceDescriptor &desc = instances[instance_index++];
+        MTLAccelerationStructureUserIDInstanceDescriptor &desc = instances[currIndex];
 
         desc.accelerationStructureIndex = accel_struct_index;
         desc.userID = user_id;

intern/cycles/device/metal/device.mm

@@ -55,6 +55,9 @@ void device_metal_info(vector<DeviceInfo> &devices)
     info.denoisers = DENOISER_NONE;
     info.id = id;
 
+    info.has_nanovdb = MetalInfo::get_device_vendor(device) == METAL_GPU_APPLE;
+    info.has_light_tree = MetalInfo::get_device_vendor(device) != METAL_GPU_AMD;
+
     devices.push_back(info);
     device_index++;
   }

intern/cycles/device/metal/device_impl.h

@@ -67,13 +67,21 @@ class MetalDevice : public Device {
   std::recursive_mutex metal_mem_map_mutex;
 
   /* Bindless Textures */
+  bool is_texture(const TextureInfo &tex);
   device_vector<TextureInfo> texture_info;
   bool need_texture_info;
   id<MTLArgumentEncoder> mtlTextureArgEncoder = nil;
+  id<MTLArgumentEncoder> mtlBufferArgEncoder = nil;
+  id<MTLBuffer> buffer_bindings_1d = nil;
   id<MTLBuffer> texture_bindings_2d = nil;
   id<MTLBuffer> texture_bindings_3d = nil;
   std::vector<id<MTLTexture>> texture_slot_map;
 
+  /* BLAS encoding & lookup */
+  id<MTLArgumentEncoder> mtlBlasArgEncoder = nil;
+  id<MTLBuffer> blas_buffer = nil;
+  id<MTLBuffer> blas_lookup_buffer = nil;
+
   bool use_metalrt = false;
   MetalPipelineType kernel_specialization_level = PSO_GENERIC;
 
@@ -105,6 +113,8 @@ class MetalDevice : public Device {
 
   bool use_adaptive_compilation();
 
+  bool use_local_atomic_sort() const;
+
   bool make_source_and_check_if_compile_needed(MetalPipelineType pso_type);
 
   void make_source(MetalPipelineType pso_type, const uint kernel_features);

intern/cycles/device/metal/device_impl.mm

@@ -91,11 +91,6 @@ MetalDevice::MetalDevice(const DeviceInfo &info, Stats &stats, Profiler &profile
     }
   }
 
-  texture_bindings_2d = [mtlDevice newBufferWithLength:4096 options:default_storage_mode];
-  texture_bindings_3d = [mtlDevice newBufferWithLength:4096 options:default_storage_mode];
-
-  stats.mem_alloc(texture_bindings_2d.allocatedSize + texture_bindings_3d.allocatedSize);
-
   switch (device_vendor) {
     default:
       break;
@@ -105,6 +100,7 @@ MetalDevice::MetalDevice(const DeviceInfo &info, Stats &stats, Profiler &profile
     }
     case METAL_GPU_AMD: {
       max_threads_per_threadgroup = 128;
+      use_metalrt = info.use_metalrt;
       break;
     }
     case METAL_GPU_APPLE: {
@@ -155,6 +151,16 @@ MetalDevice::MetalDevice(const DeviceInfo &info, Stats &stats, Profiler &profile
   arg_desc_texture.dataType = MTLDataTypeTexture;
   arg_desc_texture.access = MTLArgumentAccessReadOnly;
   mtlTextureArgEncoder = [mtlDevice newArgumentEncoderWithArguments:@[ arg_desc_texture ]];
+  MTLArgumentDescriptor *arg_desc_buffer = [[MTLArgumentDescriptor alloc] init];
+  arg_desc_buffer.dataType = MTLDataTypePointer;
+  arg_desc_buffer.access = MTLArgumentAccessReadOnly;
+  mtlBufferArgEncoder = [mtlDevice newArgumentEncoderWithArguments:@[ arg_desc_buffer ]];
+
+  buffer_bindings_1d = [mtlDevice newBufferWithLength:8192 options:default_storage_mode];
+  texture_bindings_2d = [mtlDevice newBufferWithLength:8192 options:default_storage_mode];
+  texture_bindings_3d = [mtlDevice newBufferWithLength:8192 options:default_storage_mode];
+  stats.mem_alloc(buffer_bindings_1d.allocatedSize + texture_bindings_2d.allocatedSize +
+                  texture_bindings_3d.allocatedSize);
 
   /* command queue for non-tracing work on the GPU */
   mtlGeneralCommandQueue = [mtlDevice newCommandQueue];
@@ -179,6 +185,8 @@ MetalDevice::MetalDevice(const DeviceInfo &info, Stats &stats, Profiler &profile
     arg_desc_tex.dataType = MTLDataTypePointer;
     arg_desc_tex.access = MTLArgumentAccessReadOnly;
 
+    arg_desc_tex.index = index++;
+    [ancillary_desc addObject:[arg_desc_tex copy]]; /* metal_buf_1d */
     arg_desc_tex.index = index++;
     [ancillary_desc addObject:[arg_desc_tex copy]]; /* metal_tex_2d */
     arg_desc_tex.index = index++;
@@ -192,6 +200,10 @@ MetalDevice::MetalDevice(const DeviceInfo &info, Stats &stats, Profiler &profile
         arg_desc_as.dataType = MTLDataTypeInstanceAccelerationStructure;
         arg_desc_as.access = MTLArgumentAccessReadOnly;
 
+        MTLArgumentDescriptor *arg_desc_ptrs = [[MTLArgumentDescriptor alloc] init];
+        arg_desc_ptrs.dataType = MTLDataTypePointer;
+        arg_desc_ptrs.access = MTLArgumentAccessReadOnly;
+
         MTLArgumentDescriptor *arg_desc_ift = [[MTLArgumentDescriptor alloc] init];
         arg_desc_ift.dataType = MTLDataTypeIntersectionFunctionTable;
         arg_desc_ift.access = MTLArgumentAccessReadOnly;
@@ -204,14 +216,32 @@ MetalDevice::MetalDevice(const DeviceInfo &info, Stats &stats, Profiler &profile
         [ancillary_desc addObject:[arg_desc_ift copy]]; /* ift_shadow */
         arg_desc_ift.index = index++;
         [ancillary_desc addObject:[arg_desc_ift copy]]; /* ift_local */
+        arg_desc_ift.index = index++;
+        [ancillary_desc addObject:[arg_desc_ift copy]]; /* ift_local_prim */
+        arg_desc_ptrs.index = index++;
+        [ancillary_desc addObject:[arg_desc_ptrs copy]]; /* blas array */
+        arg_desc_ptrs.index = index++;
+        [ancillary_desc addObject:[arg_desc_ptrs copy]]; /* look up table for blas */
 
         [arg_desc_ift release];
         [arg_desc_as release];
+        [arg_desc_ptrs release];
       }
     }
 
     mtlAncillaryArgEncoder = [mtlDevice newArgumentEncoderWithArguments:ancillary_desc];
 
+    // preparing the blas arg encoder
+    if (@available(macos 11.0, *)) {
+      if (use_metalrt) {
+        MTLArgumentDescriptor *arg_desc_blas = [[MTLArgumentDescriptor alloc] init];
+        arg_desc_blas.dataType = MTLDataTypeInstanceAccelerationStructure;
+        arg_desc_blas.access = MTLArgumentAccessReadOnly;
+        mtlBlasArgEncoder = [mtlDevice newArgumentEncoderWithArguments:@[ arg_desc_blas ]];
+        [arg_desc_blas release];
+      }
+    }
+
     for (int i = 0; i < ancillary_desc.count; i++) {
       [ancillary_desc[i] release];
     }
@@ -230,22 +260,26 @@ MetalDevice::~MetalDevice()
    * existing_devices_mutex). */
   thread_scoped_lock lock(existing_devices_mutex);
 
-  for (auto &tex : texture_slot_map) {
+  int num_resources = texture_info.size();
-    if (tex) {
+  for (int res = 0; res < num_resources; res++) {
-      [tex release];
+    if (is_texture(texture_info[res])) {
-      tex = nil;
+      [texture_slot_map[res] release];
+      texture_slot_map[res] = nil;
     }
   }
+
   flush_delayed_free_list();
 
   if (texture_bindings_2d) {
-    stats.mem_free(texture_bindings_2d.allocatedSize + texture_bindings_3d.allocatedSize);
+    stats.mem_free(buffer_bindings_1d.allocatedSize + texture_bindings_2d.allocatedSize +
-
+                   texture_bindings_3d.allocatedSize);
+    [buffer_bindings_1d release];
     [texture_bindings_2d release];
     [texture_bindings_3d release];
   }
   [mtlTextureArgEncoder release];
   [mtlBufferKernelParamsEncoder release];
+  [mtlBufferArgEncoder release];
   [mtlASArgEncoder release];
   [mtlAncillaryArgEncoder release];
   [mtlGeneralCommandQueue release];
@@ -271,6 +305,11 @@ bool MetalDevice::use_adaptive_compilation()
   return DebugFlags().metal.adaptive_compile;
 }
 
+bool MetalDevice::use_local_atomic_sort() const
+{
+  return DebugFlags().metal.use_local_atomic_sort;
+}
+
 void MetalDevice::make_source(MetalPipelineType pso_type, const uint kernel_features)
 {
   string global_defines;
@@ -278,6 +317,10 @@ void MetalDevice::make_source(MetalPipelineType pso_type, const uint kernel_feat
     global_defines += "#define __KERNEL_FEATURES__ " + to_string(kernel_features) + "\n";
   }
 
+  if (use_local_atomic_sort()) {
+    global_defines += "#define __KERNEL_LOCAL_ATOMIC_SORT__\n";
+  }
+
   if (use_metalrt) {
     global_defines += "#define __METALRT__\n";
     if (motion_blur) {
@@ -300,6 +343,9 @@ void MetalDevice::make_source(MetalPipelineType pso_type, const uint kernel_feat
       break;
     case METAL_GPU_APPLE:
       global_defines += "#define __KERNEL_METAL_APPLE__\n";
+#  ifdef WITH_NANOVDB
+      global_defines += "#define WITH_NANOVDB\n";
+#  endif
       break;
   }
 
@@ -514,6 +560,11 @@ void MetalDevice::compile_and_load(int device_id, MetalPipelineType pso_type)
   }
 }
 
+bool MetalDevice::is_texture(const TextureInfo &tex)
+{
+  return (tex.depth > 0 || tex.height > 0);
+}
+
 void MetalDevice::load_texture_info()
 {
   if (need_texture_info) {
@@ -525,21 +576,20 @@ void MetalDevice::load_texture_info()
 
     for (int tex = 0; tex < num_textures; tex++) {
       uint64_t offset = tex * sizeof(void *);
-
+      if (is_texture(texture_info[tex]) && texture_slot_map[tex]) {
-      id<MTLTexture> metal_texture = texture_slot_map[tex];
+        id<MTLTexture> metal_texture = texture_slot_map[tex];
-      if (!metal_texture) {
-        [mtlTextureArgEncoder setArgumentBuffer:texture_bindings_2d offset:offset];
-        [mtlTextureArgEncoder setTexture:nil atIndex:0];
-        [mtlTextureArgEncoder setArgumentBuffer:texture_bindings_3d offset:offset];
-        [mtlTextureArgEncoder setTexture:nil atIndex:0];
-      }
-      else {
         MTLTextureType type = metal_texture.textureType;
         [mtlTextureArgEncoder setArgumentBuffer:texture_bindings_2d offset:offset];
         [mtlTextureArgEncoder setTexture:type == MTLTextureType2D ? metal_texture : nil atIndex:0];
         [mtlTextureArgEncoder setArgumentBuffer:texture_bindings_3d offset:offset];
         [mtlTextureArgEncoder setTexture:type == MTLTextureType3D ? metal_texture : nil atIndex:0];
       }
+      else {
+        [mtlTextureArgEncoder setArgumentBuffer:texture_bindings_2d offset:offset];
+        [mtlTextureArgEncoder setTexture:nil atIndex:0];
+        [mtlTextureArgEncoder setArgumentBuffer:texture_bindings_3d offset:offset];
+        [mtlTextureArgEncoder setTexture:nil atIndex:0];
+      }
     }
     if (default_storage_mode == MTLResourceStorageModeManaged) {
       [texture_bindings_2d didModifyRange:NSMakeRange(0, num_textures * sizeof(void *))];
@@ -558,7 +608,7 @@ void MetalDevice::erase_allocation(device_memory &mem)
   if (it != metal_mem_map.end()) {
     MetalMem *mmem = it->second.get();
 
-    /* blank out reference to MetalMem* in the launch params (fixes crash T94736) */
+    /* blank out reference to MetalMem* in the launch params (fixes crash #94736) */
     if (mmem->pointer_index >= 0) {
       device_ptr *pointers = (device_ptr *)&launch_params;
       pointers[mmem->pointer_index] = 0;
@@ -712,7 +762,6 @@ void MetalDevice::generic_free(device_memory &mem)
       mem.shared_pointer = 0;
 
       /* Free device memory. */
-      delayed_free_list.push_back(mmem.mtlBuffer);
       mmem.mtlBuffer = nil;
     }
 
@@ -947,7 +996,7 @@ void MetalDevice::global_free(device_memory &mem)
 
 void MetalDevice::tex_alloc_as_buffer(device_texture &mem)
 {
-  generic_alloc(mem);
+  MetalDevice::MetalMem *mmem = generic_alloc(mem);
   generic_copy_to(mem);
 
   /* Resize once */
@@ -956,27 +1005,32 @@ void MetalDevice::tex_alloc_as_buffer(device_texture &mem)
     /* Allocate some slots in advance, to reduce amount
      * of re-allocations. */
     texture_info.resize(round_up(slot + 1, 128));
+    texture_slot_map.resize(round_up(slot + 1, 128));
   }
 
-  mem.info.data = (uint64_t)mem.device_pointer;
-
-  /* Set Mapping and tag that we need to (re-)upload to device */
   texture_info[slot] = mem.info;
+  uint64_t offset = slot * sizeof(void *);
+  [mtlBufferArgEncoder setArgumentBuffer:buffer_bindings_1d offset:offset];
+  [mtlBufferArgEncoder setBuffer:mmem->mtlBuffer offset:0 atIndex:0];
+  texture_info[slot].data = *(uint64_t *)((uint64_t)buffer_bindings_1d.contents + offset);
+  texture_slot_map[slot] = nil;
   need_texture_info = true;
 }
 
 void MetalDevice::tex_alloc(device_texture &mem)
 {
   /* Check that dimensions fit within maximum allowable size.
+   * If 1D texture is allocated, use 1D buffer.
    * See: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf */
-  if (mem.data_width > 16384 || mem.data_height > 16384) {
+  if (mem.data_height > 0) {
-    set_error(string_printf(
+    if (mem.data_width > 16384 || mem.data_height > 16384) {
-        "Texture exceeds maximum allowed size of 16384 x 16384 (requested: %zu x %zu)",
+      set_error(string_printf(
-        mem.data_width,
+          "Texture exceeds maximum allowed size of 16384 x 16384 (requested: %zu x %zu)",
-        mem.data_height));
+          mem.data_width,
-    return;
+          mem.data_height));
+      return;
+    }
   }
-
   MTLStorageMode storage_mode = MTLStorageModeManaged;
   if (@available(macos 10.15, *)) {
     if ([mtlDevice hasUnifiedMemory] &&
@@ -1116,8 +1170,9 @@ void MetalDevice::tex_alloc(device_texture &mem)
                   bytesPerRow:src_pitch];
   }
   else {
-    assert(0);
     /* 1D texture, using linear memory. */
+    tex_alloc_as_buffer(mem);
+    return;
   }
 
   mem.device_pointer = (device_ptr)mtlTexture;
@@ -1141,17 +1196,22 @@ void MetalDevice::tex_alloc(device_texture &mem)
     ssize_t min_buffer_length = sizeof(void *) * texture_info.size();
     if (!texture_bindings_2d || (texture_bindings_2d.length < min_buffer_length)) {
       if (texture_bindings_2d) {
+        delayed_free_list.push_back(buffer_bindings_1d);
         delayed_free_list.push_back(texture_bindings_2d);
         delayed_free_list.push_back(texture_bindings_3d);
 
-        stats.mem_free(texture_bindings_2d.allocatedSize + texture_bindings_3d.allocatedSize);
+        stats.mem_free(buffer_bindings_1d.allocatedSize + texture_bindings_2d.allocatedSize +
+                       texture_bindings_3d.allocatedSize);
       }
+      buffer_bindings_1d = [mtlDevice newBufferWithLength:min_buffer_length
+                                                  options:default_storage_mode];
       texture_bindings_2d = [mtlDevice newBufferWithLength:min_buffer_length
                                                    options:default_storage_mode];
       texture_bindings_3d = [mtlDevice newBufferWithLength:min_buffer_length
                                                    options:default_storage_mode];
 
-      stats.mem_alloc(texture_bindings_2d.allocatedSize + texture_bindings_3d.allocatedSize);
+      stats.mem_alloc(buffer_bindings_1d.allocatedSize + texture_bindings_2d.allocatedSize +
+                      texture_bindings_3d.allocatedSize);
     }
   }
 
@@ -1178,12 +1238,18 @@ void MetalDevice::tex_alloc(device_texture &mem)
 
 void MetalDevice::tex_free(device_texture &mem)
 {
+  if (mem.data_depth == 0 && mem.data_height == 0) {
+    generic_free(mem);
+    return;
+  }
+
   if (metal_mem_map.count(&mem)) {
     std::lock_guard<std::recursive_mutex> lock(metal_mem_map_mutex);
     MetalMem &mmem = *metal_mem_map.at(&mem);
 
     assert(texture_slot_map[mem.slot] == mmem.mtlTexture);
-    texture_slot_map[mem.slot] = nil;
+    if (texture_slot_map[mem.slot] == mmem.mtlTexture)
+      texture_slot_map[mem.slot] = nil;
 
     if (mmem.mtlTexture) {
       /* Free bindless texture. */
@@ -1231,6 +1297,33 @@ void MetalDevice::build_bvh(BVH *bvh, Progress &progress, bool refit)
     if (@available(macos 11.0, *)) {
       if (bvh->params.top_level) {
         bvhMetalRT = bvh_metal;
+
+        // allocate required buffers for BLAS array
+        uint64_t count = bvhMetalRT->blas_array.size();
+        uint64_t bufferSize = mtlBlasArgEncoder.encodedLength * count;
+        blas_buffer = [mtlDevice newBufferWithLength:bufferSize options:default_storage_mode];
+        stats.mem_alloc(blas_buffer.allocatedSize);
+
+        for (uint64_t i = 0; i < count; ++i) {
+          [mtlBlasArgEncoder setArgumentBuffer:blas_buffer
+                                        offset:i * mtlBlasArgEncoder.encodedLength];
+          [mtlBlasArgEncoder setAccelerationStructure:bvhMetalRT->blas_array[i] atIndex:0];
+        }
+
+        count = bvhMetalRT->blas_lookup.size();
+        bufferSize = sizeof(uint32_t) * count;
+        blas_lookup_buffer = [mtlDevice newBufferWithLength:bufferSize
+                                                    options:default_storage_mode];
+        stats.mem_alloc(blas_lookup_buffer.allocatedSize);
+
+        memcpy([blas_lookup_buffer contents],
+               bvhMetalRT -> blas_lookup.data(),
+               blas_lookup_buffer.allocatedSize);
+
+        if (default_storage_mode == MTLResourceStorageModeManaged) {
+          [blas_buffer didModifyRange:NSMakeRange(0, blas_buffer.length)];
+          [blas_lookup_buffer didModifyRange:NSMakeRange(0, blas_lookup_buffer.length)];
+        }
       }
     }
   }

intern/cycles/device/metal/kernel.h

@@ -19,6 +19,8 @@ enum {
   METALRT_FUNC_SHADOW_BOX,
   METALRT_FUNC_LOCAL_TRI,
   METALRT_FUNC_LOCAL_BOX,
+  METALRT_FUNC_LOCAL_TRI_PRIM,
+  METALRT_FUNC_LOCAL_BOX_PRIM,
   METALRT_FUNC_CURVE_RIBBON,
   METALRT_FUNC_CURVE_RIBBON_SHADOW,
   METALRT_FUNC_CURVE_ALL,
@@ -28,7 +30,13 @@ enum {
   METALRT_FUNC_NUM
 };
 
-enum { METALRT_TABLE_DEFAULT, METALRT_TABLE_SHADOW, METALRT_TABLE_LOCAL, METALRT_TABLE_NUM };
+enum {
+  METALRT_TABLE_DEFAULT,
+  METALRT_TABLE_SHADOW,
+  METALRT_TABLE_LOCAL,
+  METALRT_TABLE_LOCAL_PRIM,
+  METALRT_TABLE_NUM
+};
 
 /* Pipeline State Object types */
 enum MetalPipelineType {

intern/cycles/device/metal/kernel.mm

@@ -87,6 +87,9 @@ struct ShaderCache {
           break;
       }
     }
+
+    occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_SORT_BUCKET_PASS] = {1024, 1024};
+    occupancy_tuning[DEVICE_KERNEL_INTEGRATOR_SORT_WRITE_PASS] = {1024, 1024};
   }
   ~ShaderCache();
 
@@ -521,6 +524,8 @@ void MetalKernelPipeline::compile()
           "__anyhit__cycles_metalrt_shadow_all_hit_box",
           "__anyhit__cycles_metalrt_local_hit_tri",
           "__anyhit__cycles_metalrt_local_hit_box",
+          "__anyhit__cycles_metalrt_local_hit_tri_prim",
+          "__anyhit__cycles_metalrt_local_hit_box_prim",
           "__intersection__curve_ribbon",
           "__intersection__curve_ribbon_shadow",
           "__intersection__curve_all",
@@ -611,11 +616,17 @@ void MetalKernelPipeline::compile()
                          rt_intersection_function[METALRT_FUNC_LOCAL_BOX],
                          rt_intersection_function[METALRT_FUNC_LOCAL_BOX],
                          nil];
+    table_functions[METALRT_TABLE_LOCAL_PRIM] = [NSArray
+        arrayWithObjects:rt_intersection_function[METALRT_FUNC_LOCAL_TRI_PRIM],
+                         rt_intersection_function[METALRT_FUNC_LOCAL_BOX_PRIM],
+                         rt_intersection_function[METALRT_FUNC_LOCAL_BOX_PRIM],
+                         nil];
 
     NSMutableSet *unique_functions = [NSMutableSet
         setWithArray:table_functions[METALRT_TABLE_DEFAULT]];
     [unique_functions addObjectsFromArray:table_functions[METALRT_TABLE_SHADOW]];
     [unique_functions addObjectsFromArray:table_functions[METALRT_TABLE_LOCAL]];
+    [unique_functions addObjectsFromArray:table_functions[METALRT_TABLE_LOCAL_PRIM]];
 
     if (kernel_has_intersection(device_kernel)) {
       linked_functions = [[NSArray arrayWithArray:[unique_functions allObjects]]

intern/cycles/device/metal/queue.h

@@ -25,6 +25,7 @@ class MetalDeviceQueue : public DeviceQueue {
   virtual int num_concurrent_states(const size_t) const override;
   virtual int num_concurrent_busy_states(const size_t) const override;
   virtual int num_sort_partition_elements() const override;
+  virtual bool supports_local_atomic_sort() const override;
 
   virtual void init_execution() override;
 

intern/cycles/device/metal/queue.mm

@@ -315,6 +315,11 @@ int MetalDeviceQueue::num_sort_partition_elements() const
   return MetalInfo::optimal_sort_partition_elements(metal_device_->mtlDevice);
 }
 
+bool MetalDeviceQueue::supports_local_atomic_sort() const
+{
+  return metal_device_->use_local_atomic_sort();
+}
+
 void MetalDeviceQueue::init_execution()
 {
   /* Synchronize all textures and memory copies before executing task. */
@@ -472,11 +477,21 @@ bool MetalDeviceQueue::enqueue(DeviceKernel kernel,
   [metal_device_->mtlAncillaryArgEncoder setBuffer:metal_device_->texture_bindings_3d
                                             offset:0
                                            atIndex:1];
+  [metal_device_->mtlAncillaryArgEncoder setBuffer:metal_device_->buffer_bindings_1d
+                                            offset:0
+                                           atIndex:2];
+
   if (@available(macos 12.0, *)) {
     if (metal_device_->use_metalrt) {
       if (metal_device_->bvhMetalRT) {
         id<MTLAccelerationStructure> accel_struct = metal_device_->bvhMetalRT->accel_struct;
-        [metal_device_->mtlAncillaryArgEncoder setAccelerationStructure:accel_struct atIndex:2];
+        [metal_device_->mtlAncillaryArgEncoder setAccelerationStructure:accel_struct atIndex:3];
+        [metal_device_->mtlAncillaryArgEncoder setBuffer:metal_device_->blas_buffer
+                                                  offset:0
+                                                 atIndex:8];
+        [metal_device_->mtlAncillaryArgEncoder setBuffer:metal_device_->blas_lookup_buffer
+                                                  offset:0
+                                                 atIndex:9];
       }
 
       for (int table = 0; table < METALRT_TABLE_NUM; table++) {
@@ -486,13 +501,13 @@ bool MetalDeviceQueue::enqueue(DeviceKernel kernel,
                                                               atIndex:1];
           [metal_device_->mtlAncillaryArgEncoder
               setIntersectionFunctionTable:metal_kernel_pso->intersection_func_table[table]
-                                   atIndex:3 + table];
+                                   atIndex:4 + table];
           [mtlComputeCommandEncoder useResource:metal_kernel_pso->intersection_func_table[table]
                                           usage:MTLResourceUsageRead];
         }
         else {
           [metal_device_->mtlAncillaryArgEncoder setIntersectionFunctionTable:nil
-                                                                      atIndex:3 + table];
+                                                                      atIndex:4 + table];
         }
       }
     }
@@ -527,6 +542,10 @@ bool MetalDeviceQueue::enqueue(DeviceKernel kernel,
       if (bvhMetalRT) {
         /* Mark all Accelerations resources as used */
         [mtlComputeCommandEncoder useResource:bvhMetalRT->accel_struct usage:MTLResourceUsageRead];
+        [mtlComputeCommandEncoder useResource:metal_device_->blas_buffer
+                                        usage:MTLResourceUsageRead];
+        [mtlComputeCommandEncoder useResource:metal_device_->blas_lookup_buffer
+                                        usage:MTLResourceUsageRead];
         [mtlComputeCommandEncoder useResources:bvhMetalRT->blas_array.data()
                                          count:bvhMetalRT->blas_array.size()
                                          usage:MTLResourceUsageRead];
@@ -553,13 +572,24 @@ bool MetalDeviceQueue::enqueue(DeviceKernel kernel,
       /* See parallel_active_index.h for why this amount of shared memory is needed.
        * Rounded up to 16 bytes for Metal */
       shared_mem_bytes = (int)round_up((num_threads_per_block + 1) * sizeof(int), 16);
-      [mtlComputeCommandEncoder setThreadgroupMemoryLength:shared_mem_bytes atIndex:0];
       break;
 
+    case DEVICE_KERNEL_INTEGRATOR_SORT_BUCKET_PASS:
+    case DEVICE_KERNEL_INTEGRATOR_SORT_WRITE_PASS: {
+      int key_count = metal_device_->launch_params.data.max_shaders;
+      shared_mem_bytes = (int)round_up(key_count * sizeof(int), 16);
+      break;
+    }
+
     default:
       break;
   }
 
+  if (shared_mem_bytes) {
+    assert(shared_mem_bytes <= 32 * 1024);
+    [mtlComputeCommandEncoder setThreadgroupMemoryLength:shared_mem_bytes atIndex:0];
+  }
+
   MTLSize size_threadgroups_per_dispatch = MTLSizeMake(
       divide_up(work_size, num_threads_per_block), 1, 1);
   MTLSize size_threads_per_threadgroup = MTLSizeMake(num_threads_per_block, 1, 1);
@@ -848,6 +878,7 @@ void MetalDeviceQueue::prepare_resources(DeviceKernel kernel)
   /* ancillaries */
   [mtlComputeEncoder_ useResource:metal_device_->texture_bindings_2d usage:MTLResourceUsageRead];
   [mtlComputeEncoder_ useResource:metal_device_->texture_bindings_3d usage:MTLResourceUsageRead];
+  [mtlComputeEncoder_ useResource:metal_device_->buffer_bindings_1d usage:MTLResourceUsageRead];
 }
 
 id<MTLComputeCommandEncoder> MetalDeviceQueue::get_compute_encoder(DeviceKernel kernel)

intern/cycles/device/metal/util.mm

@@ -64,6 +64,12 @@ MetalGPUVendor MetalInfo::get_device_vendor(id<MTLDevice> device)
     return METAL_GPU_INTEL;
   }
   else if (strstr(device_name, "AMD")) {
+    /* Setting this env var hides AMD devices thus exposing any integrated Intel devices. */
+    if (auto str = getenv("CYCLES_METAL_FORCE_INTEL")) {
+      if (atoi(str)) {
+        return METAL_GPU_UNKNOWN;
+      }
+    }
     return METAL_GPU_AMD;
   }
   else if (strstr(device_name, "Apple")) {
@@ -96,6 +102,15 @@ vector<id<MTLDevice>> const &MetalInfo::get_usable_devices()
     return usable_devices;
   }
 
+  /* If the system has both an AMD GPU (discrete) and an Intel one (integrated), prefer the AMD
+   * one. This can be overridden with CYCLES_METAL_FORCE_INTEL. */
+  bool has_usable_amd_gpu = false;
+  if (@available(macos 12.3, *)) {
+    for (id<MTLDevice> device in MTLCopyAllDevices()) {
+      has_usable_amd_gpu |= (get_device_vendor(device) == METAL_GPU_AMD);
+    }
+  }
+
   metal_printf("Usable Metal devices:\n");
   for (id<MTLDevice> device in MTLCopyAllDevices()) {
     string device_name = get_device_name(device);
@@ -111,8 +126,10 @@ vector<id<MTLDevice>> const &MetalInfo::get_usable_devices()
     }
 
 #  if defined(MAC_OS_VERSION_13_0)
-    if (@available(macos 13.0, *)) {
+    if (!has_usable_amd_gpu) {
-      usable |= (vendor == METAL_GPU_INTEL);
+      if (@available(macos 13.0, *)) {
+        usable |= (vendor == METAL_GPU_INTEL);
+      }
     }
 #  endif
 

intern/cycles/device/optix/device_impl.cpp

@@ -854,12 +854,14 @@ bool OptiXDevice::load_osl_kernels()
         context, group_descs, 2, &group_options, nullptr, 0, &osl_groups[i * 2]));
   }
 
+  OptixStackSizes stack_size[NUM_PROGRAM_GROUPS] = {};
   vector<OptixStackSizes> osl_stack_size(osl_groups.size());
 
   /* Update SBT with new entries. */
   sbt_data.alloc(NUM_PROGRAM_GROUPS + osl_groups.size());
   for (int i = 0; i < NUM_PROGRAM_GROUPS; ++i) {
     optix_assert(optixSbtRecordPackHeader(groups[i], &sbt_data[i]));
+    optix_assert(optixProgramGroupGetStackSize(groups[i], &stack_size[i]));
   }
   for (size_t i = 0; i < osl_groups.size(); ++i) {
     if (osl_groups[i] != NULL) {
@@ -907,13 +909,15 @@ bool OptiXDevice::load_osl_kernels()
                                      0,
                                      &pipelines[PIP_SHADE]));
 
+    const unsigned int css = std::max(stack_size[PG_RGEN_SHADE_SURFACE_RAYTRACE].cssRG,
+                                      stack_size[PG_RGEN_SHADE_SURFACE_MNEE].cssRG);
     unsigned int dss = 0;
     for (unsigned int i = 0; i < osl_stack_size.size(); ++i) {
       dss = std::max(dss, osl_stack_size[i].dssDC);
     }
 
     optix_assert(optixPipelineSetStackSize(
-        pipelines[PIP_SHADE], 0, dss, 0, pipeline_options.usesMotionBlur ? 3 : 2));
+        pipelines[PIP_SHADE], 0, dss, css, pipeline_options.usesMotionBlur ? 3 : 2));
   }
 
   return !have_error();

intern/cycles/device/queue.h

@@ -112,6 +112,13 @@ class DeviceQueue {
     return 65536;
   }
 
+  /* Does device support local atomic sorting kernels (INTEGRATOR_SORT_BUCKET_PASS and
+   * INTEGRATOR_SORT_WRITE_PASS)? */
+  virtual bool supports_local_atomic_sort() const
+  {
+    return false;
+  }
+
   /* Initialize execution of kernels on this queue.
    *
    * Will, for example, load all data required by the kernels from Device to global or path state.

intern/cycles/integrator/path_trace_work_gpu.cpp

@@ -71,6 +71,8 @@ PathTraceWorkGPU::PathTraceWorkGPU(Device *device,
           device, "integrator_shader_mnee_sort_counter", MEM_READ_WRITE),
       integrator_shader_sort_prefix_sum_(
           device, "integrator_shader_sort_prefix_sum", MEM_READ_WRITE),
+      integrator_shader_sort_partition_key_offsets_(
+          device, "integrator_shader_sort_partition_key_offsets", MEM_READ_WRITE),
       integrator_next_main_path_index_(device, "integrator_next_main_path_index", MEM_READ_WRITE),
       integrator_next_shadow_path_index_(
           device, "integrator_next_shadow_path_index", MEM_READ_WRITE),
@@ -207,33 +209,45 @@ void PathTraceWorkGPU::alloc_integrator_sorting()
   integrator_state_gpu_.sort_partition_divisor = (int)divide_up(max_num_paths_,
                                                                 num_sort_partitions_);
 
-  /* Allocate arrays for shader sorting. */
+  if (num_sort_partitions_ > 1 && queue_->supports_local_atomic_sort()) {
-  const int sort_buckets = device_scene_->data.max_shaders * num_sort_partitions_;
+    /* Allocate array for partitioned shader sorting using local atomics. */
-  if (integrator_shader_sort_counter_.size() < sort_buckets) {
+    const int num_offsets = (device_scene_->data.max_shaders + 1) * num_sort_partitions_;
-    integrator_shader_sort_counter_.alloc(sort_buckets);
+    if (integrator_shader_sort_partition_key_offsets_.size() < num_offsets) {
-    integrator_shader_sort_counter_.zero_to_device();
+      integrator_shader_sort_partition_key_offsets_.alloc(num_offsets);
-    integrator_state_gpu_.sort_key_counter[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE] =
+      integrator_shader_sort_partition_key_offsets_.zero_to_device();
-        (int *)integrator_shader_sort_counter_.device_pointer;
-
-    integrator_shader_sort_prefix_sum_.alloc(sort_buckets);
-    integrator_shader_sort_prefix_sum_.zero_to_device();
-  }
-
-  if (device_scene_->data.kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) {
-    if (integrator_shader_raytrace_sort_counter_.size() < sort_buckets) {
-      integrator_shader_raytrace_sort_counter_.alloc(sort_buckets);
-      integrator_shader_raytrace_sort_counter_.zero_to_device();
-      integrator_state_gpu_.sort_key_counter[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE] =
-          (int *)integrator_shader_raytrace_sort_counter_.device_pointer;
     }
+    integrator_state_gpu_.sort_partition_key_offsets =
+        (int *)integrator_shader_sort_partition_key_offsets_.device_pointer;
   }
+  else {
+    /* Allocate arrays for shader sorting. */
+    const int sort_buckets = device_scene_->data.max_shaders * num_sort_partitions_;
+    if (integrator_shader_sort_counter_.size() < sort_buckets) {
+      integrator_shader_sort_counter_.alloc(sort_buckets);
+      integrator_shader_sort_counter_.zero_to_device();
+      integrator_state_gpu_.sort_key_counter[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE] =
+          (int *)integrator_shader_sort_counter_.device_pointer;
 
-  if (device_scene_->data.kernel_features & KERNEL_FEATURE_MNEE) {
+      integrator_shader_sort_prefix_sum_.alloc(sort_buckets);
-    if (integrator_shader_mnee_sort_counter_.size() < sort_buckets) {
+      integrator_shader_sort_prefix_sum_.zero_to_device();
-      integrator_shader_mnee_sort_counter_.alloc(sort_buckets);
+    }
-      integrator_shader_mnee_sort_counter_.zero_to_device();
+
-      integrator_state_gpu_.sort_key_counter[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_MNEE] =
+    if (device_scene_->data.kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) {
-          (int *)integrator_shader_mnee_sort_counter_.device_pointer;
+      if (integrator_shader_raytrace_sort_counter_.size() < sort_buckets) {
+        integrator_shader_raytrace_sort_counter_.alloc(sort_buckets);
+        integrator_shader_raytrace_sort_counter_.zero_to_device();
+        integrator_state_gpu_.sort_key_counter[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE] =
+            (int *)integrator_shader_raytrace_sort_counter_.device_pointer;
+      }
+    }
+
+    if (device_scene_->data.kernel_features & KERNEL_FEATURE_MNEE) {
+      if (integrator_shader_mnee_sort_counter_.size() < sort_buckets) {
+        integrator_shader_mnee_sort_counter_.alloc(sort_buckets);
+        integrator_shader_mnee_sort_counter_.zero_to_device();
+        integrator_state_gpu_.sort_key_counter[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_MNEE] =
+            (int *)integrator_shader_mnee_sort_counter_.device_pointer;
+      }
     }
   }
 }
@@ -451,8 +465,7 @@ void PathTraceWorkGPU::enqueue_path_iteration(DeviceKernel kernel, const int num
     work_size = num_queued;
     d_path_index = queued_paths_.device_pointer;
 
-    compute_sorted_queued_paths(
+    compute_sorted_queued_paths(kernel, num_paths_limit);
-        DEVICE_KERNEL_INTEGRATOR_SORTED_PATHS_ARRAY, kernel, num_paths_limit);
   }
   else if (num_queued < work_size) {
     work_size = num_queued;
@@ -511,11 +524,26 @@ void PathTraceWorkGPU::enqueue_path_iteration(DeviceKernel kernel, const int num
   }
 }
 
-void PathTraceWorkGPU::compute_sorted_queued_paths(DeviceKernel kernel,
+void PathTraceWorkGPU::compute_sorted_queued_paths(DeviceKernel queued_kernel,
-                                                   DeviceKernel queued_kernel,
                                                    const int num_paths_limit)
 {
   int d_queued_kernel = queued_kernel;
+
+  /* Launch kernel to fill the active paths arrays. */
+  if (num_sort_partitions_ > 1 && queue_->supports_local_atomic_sort()) {
+    const int work_size = kernel_max_active_main_path_index(queued_kernel);
+    device_ptr d_queued_paths = queued_paths_.device_pointer;
+
+    int partition_size = (int)integrator_state_gpu_.sort_partition_divisor;
+
+    DeviceKernelArguments args(
+        &work_size, &partition_size, &num_paths_limit, &d_queued_paths, &d_queued_kernel);
+
+    queue_->enqueue(DEVICE_KERNEL_INTEGRATOR_SORT_BUCKET_PASS, 1024 * num_sort_partitions_, args);
+    queue_->enqueue(DEVICE_KERNEL_INTEGRATOR_SORT_WRITE_PASS, 1024 * num_sort_partitions_, args);
+    return;
+  }
+
   device_ptr d_counter = (device_ptr)integrator_state_gpu_.sort_key_counter[d_queued_kernel];
   device_ptr d_prefix_sum = integrator_shader_sort_prefix_sum_.device_pointer;
   assert(d_counter != 0 && d_prefix_sum != 0);
@@ -552,7 +580,7 @@ void PathTraceWorkGPU::compute_sorted_queued_paths(DeviceKernel kernel,
                                &d_prefix_sum,
                                &d_queued_kernel);
 
-    queue_->enqueue(kernel, work_size, args);
+    queue_->enqueue(DEVICE_KERNEL_INTEGRATOR_SORTED_PATHS_ARRAY, work_size, args);
   }
 }
 

intern/cycles/integrator/path_trace_work_gpu.h

@@ -70,9 +70,7 @@ class PathTraceWorkGPU : public PathTraceWork {
   void enqueue_path_iteration(DeviceKernel kernel, const int num_paths_limit = INT_MAX);
 
   void compute_queued_paths(DeviceKernel kernel, DeviceKernel queued_kernel);
-  void compute_sorted_queued_paths(DeviceKernel kernel,
+  void compute_sorted_queued_paths(DeviceKernel queued_kernel, const int num_paths_limit);
-                                   DeviceKernel queued_kernel,
-                                   const int num_paths_limit);
 
   void compact_main_paths(const int num_active_paths);
   void compact_shadow_paths();
@@ -135,6 +133,7 @@ class PathTraceWorkGPU : public PathTraceWork {
   device_vector<int> integrator_shader_raytrace_sort_counter_;
   device_vector<int> integrator_shader_mnee_sort_counter_;
   device_vector<int> integrator_shader_sort_prefix_sum_;
+  device_vector<int> integrator_shader_sort_partition_key_offsets_;
   /* Path split. */
   device_vector<int> integrator_next_main_path_index_;
   device_vector<int> integrator_next_shadow_path_index_;

intern/cycles/integrator/render_scheduler.cpp

@@ -886,7 +886,7 @@ int RenderScheduler::get_num_samples_during_navigation(int resolution_divider) c
 {
   /* Special trick for fast navigation: schedule multiple samples during fast navigation
    * (which will prefer to use lower resolution to keep up with refresh rate). This gives more
-   * usable visual feedback for artists. There are a couple of tricks though. */
+   * usable visual feedback for artists. */
 
   if (is_denoise_active_during_update()) {
     /* When denoising is used during navigation prefer using a higher resolution with less samples
@@ -896,25 +896,12 @@ int RenderScheduler::get_num_samples_during_navigation(int resolution_divider) c
     return 1;
   }
 
-  if (resolution_divider <= pixel_size_) {
+  /* Schedule samples equal to the resolution divider up to a maximum of 4.
-    /* When resolution divider is at or below pixel size, schedule one sample. This doesn't effect
+   * The idea is to have enough information on the screen by increasing the sample count as the
-     * the sample count at this resolution division, but instead assists in the calculation of
+   * resolution is decreased. */
-     * the resolution divider. */
+  /* NOTE: Changing this formula will change the formula in
-    return 1;
+   * `RenderScheduler::calculate_resolution_divider_for_time()`. */
-  }
+  return min(max(1, resolution_divider / pixel_size_), 4);
-
-  if (resolution_divider == pixel_size_ * 2) {
-    /* When resolution divider is the previous step to the final resolution, schedule two samples.
-     * This is so that rendering on lower resolution does not exceed time that it takes to render
-     * first sample at the full resolution. */
-    return 2;
-  }
-
-  /* Always render 4 samples, even if scene is configured for less.
-   * The idea here is to have enough information on the screen. Resolution divider of 2 allows us
-   * to have 4 time extra samples, so overall worst case timing is the same as the final resolution
-   * at one sample. */
-  return 4;
 }
 
 bool RenderScheduler::work_need_adaptive_filter() const
@@ -1100,9 +1087,10 @@ void RenderScheduler::update_start_resolution_divider()
   /* TODO(sergey): Need to add hysteresis to avoid resolution divider bouncing around when actual
    * render time is somewhere on a boundary between two resolutions. */
 
-  /* Never increase resolution to higher than the pixel size (which is possible if the scene is
+  /* Don't let resolution drop below the desired one. It's better to be slow than provide an
-   * simple and compute device is fast). */
+   * unreadable viewport render. */
-  start_resolution_divider_ = max(resolution_divider_for_update, pixel_size_);
+  start_resolution_divider_ = min(resolution_divider_for_update,
+                                  default_start_resolution_divider_);
 
   VLOG_WORK << "Calculated resolution divider is " << start_resolution_divider_;
 }
@@ -1187,24 +1175,24 @@ void RenderScheduler::check_time_limit_reached()
 
 int RenderScheduler::calculate_resolution_divider_for_time(double desired_time, double actual_time)
 {
-  /* TODO(sergey): There should a non-iterative analytical formula here. */
+  const double ratio_between_times = actual_time / desired_time;
 
-  int resolution_divider = 1;
+  /* We can pass `ratio_between_times` to `get_num_samples_during_navigation()` to get our
+   * navigation samples because the equation for calculating the resolution divider is as follows:
+   * `actual_time / desired_time = sqr(resolution_divider) / sample_count`.
+   * While `resolution_divider` is less than or equal to 4, `resolution_divider = sample_count`
+   * (This relationship is determined in `get_num_samples_during_navigation()`). With some
+   * substitution we end up with `actual_time / desired_time = resolution_divider` while the
+   * resolution divider is less than or equal to 4. Once the resolution divider increases above 4,
+   * the relationship of `actual_time / desired_time = resolution_divider` is no longer true,
+   * however the sample count retrieved from `get_num_samples_during_navigation()` is still
+   * accurate if we continue using this assumption. It should be noted that the interaction between
+   * `pixel_size`, sample count, and resolution divider are automatically accounted for and that's
+   * why `pixel_size` isn't included in any of the equations. */
+  const int navigation_samples = get_num_samples_during_navigation(
+      ceil_to_int(ratio_between_times));
 
-  /* This algorithm iterates through resolution dividers until a divider is found that achieves
+  return ceil_to_int(sqrt(navigation_samples * ratio_between_times));
-   * the desired render time. A limit of default_start_resolution_divider_ is put in place as the
-   * maximum resolution divider to avoid an unreadable viewport due to a low resolution.
-   * pre_resolution_division_samples and post_resolution_division_samples are used in this
-   * calculation to better predict the performance impact of changing resolution divisions as
-   * the sample count can also change between resolution divisions. */
-  while (actual_time > desired_time && resolution_divider < default_start_resolution_divider_) {
-    int pre_resolution_division_samples = get_num_samples_during_navigation(resolution_divider);
-    resolution_divider = resolution_divider * 2;
-    int post_resolution_division_samples = get_num_samples_during_navigation(resolution_divider);
-    actual_time /= 4.0 * pre_resolution_division_samples / post_resolution_division_samples;
-  }
-
-  return resolution_divider;
 }
 
 int calculate_resolution_divider_for_resolution(int width, int height, int resolution)

intern/cycles/kernel/CMakeLists.txt

@@ -412,11 +412,12 @@ if(WITH_CYCLES_CUDA_BINARIES)
   # warn for other versions
   if((CUDA_VERSION STREQUAL "101") OR
      (CUDA_VERSION STREQUAL "102") OR
-     (CUDA_VERSION_MAJOR STREQUAL "11"))
+     (CUDA_VERSION_MAJOR STREQUAL "11") OR
+     (CUDA_VERSION_MAJOR STREQUAL "12"))
   else()
     message(WARNING
       "CUDA version ${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR} detected, "
-      "build may succeed but only CUDA 11, 10.2 and 10.1 have been tested")
+      "build may succeed but only CUDA 12, 11, 10.2 and 10.1 have been tested")
   endif()
 
   # build for each arch
@@ -514,6 +515,16 @@ if(WITH_CYCLES_CUDA_BINARIES)
       else()
         message(STATUS "CUDA binaries for ${arch} require CUDA 10 or earlier, skipped.")
       endif()
+    elseif(${arch} MATCHES ".*_3.")
+      if(DEFINED CUDA11_NVCC_EXECUTABLE)
+        set(cuda_nvcc_executable ${CUDA11_NVCC_EXECUTABLE})
+        set(cuda_toolkit_root_dir ${CUDA11_TOOLKIT_ROOT_DIR})
+      elseif("${CUDA_VERSION}" LESS 120) # Support for sm_35, sm_37 was removed in CUDA 12
+        set(cuda_nvcc_executable ${CUDA_NVCC_EXECUTABLE})
+        set(cuda_toolkit_root_dir ${CUDA_TOOLKIT_ROOT_DIR})
+      else()
+        message(STATUS "CUDA binaries for ${arch} require CUDA 11 or earlier, skipped.")
+      endif()
     elseif(${arch} MATCHES ".*_7." AND "${CUDA_VERSION}" LESS 100)
       message(STATUS "CUDA binaries for ${arch} require CUDA 10.0+, skipped.")
     elseif(${arch} MATCHES ".*_8.")

intern/cycles/kernel/closure/bsdf.h

@@ -661,7 +661,8 @@ ccl_device void bsdf_blur(KernelGlobals kg, ccl_private ShaderClosure *sc, float
 #endif
 }
 
-ccl_device_inline Spectrum bsdf_albedo(ccl_private const ShaderData *sd, ccl_private const ShaderClosure *sc)
+ccl_device_inline Spectrum bsdf_albedo(ccl_private const ShaderData *sd,
+                                       ccl_private const ShaderClosure *sc)
 {
   Spectrum albedo = sc->weight;
   /* Some closures include additional components such as Fresnel terms that cause their albedo to
@@ -685,7 +686,7 @@ ccl_device_inline Spectrum bsdf_albedo(ccl_private const ShaderData *sd, ccl_pri
       albedo *= ((ccl_private const PrincipledSheenBsdf *)sc)->avg_value;
       break;
     case CLOSURE_BSDF_HAIR_PRINCIPLED_ID:
-      albedo *= bsdf_principled_hair_albedo(sc);
+      albedo *= bsdf_principled_hair_albedo(sd, sc);
       break;
     default:
       break;

intern/cycles/kernel/closure/bsdf_hair_principled.h

@@ -478,10 +478,18 @@ ccl_device_inline float bsdf_principled_hair_albedo_roughness_scale(
   return (((((0.245f * x) + 5.574f) * x - 10.73f) * x + 2.532f) * x - 0.215f) * x + 5.969f;
 }
 
-ccl_device Spectrum bsdf_principled_hair_albedo(ccl_private const ShaderClosure *sc)
+ccl_device Spectrum bsdf_principled_hair_albedo(ccl_private const ShaderData *sd,
+                                                ccl_private const ShaderClosure *sc)
 {
   ccl_private PrincipledHairBSDF *bsdf = (ccl_private PrincipledHairBSDF *)sc;
-  return exp(-sqrt(bsdf->sigma) * bsdf_principled_hair_albedo_roughness_scale(bsdf->v));
+
+  const float cos_theta_o = cos_from_sin(dot(sd->wi, safe_normalize(sd->dPdu)));
+  const float cos_gamma_o = cos_from_sin(bsdf->extra->geom.w);
+  const float f = fresnel_dielectric_cos(cos_theta_o * cos_gamma_o, bsdf->eta);
+
+  const float roughness_scale = bsdf_principled_hair_albedo_roughness_scale(bsdf->v);
+  /* TODO(lukas): Adding the Fresnel term here as a workaround until the proper refactor. */
+  return exp(-sqrt(bsdf->sigma) * roughness_scale) + make_spectrum(f);
 }
 
 ccl_device_inline Spectrum

intern/cycles/kernel/closure/bsdf_microfacet.h

@@ -519,14 +519,6 @@ ccl_device int bsdf_microfacet_ggx_setup(ccl_private MicrofacetBsdf *bsdf)
   return SD_BSDF | SD_BSDF_HAS_EVAL;
 }
 
-/* Required to maintain OSL interface. */
-ccl_device int bsdf_microfacet_ggx_isotropic_setup(ccl_private MicrofacetBsdf *bsdf)
-{
-  bsdf->alpha_y = bsdf->alpha_x;
-
-  return bsdf_microfacet_ggx_setup(bsdf);
-}
-
 ccl_device int bsdf_microfacet_ggx_fresnel_setup(ccl_private MicrofacetBsdf *bsdf,
                                                  ccl_private const ShaderData *sd)
 {
@@ -613,14 +605,6 @@ ccl_device int bsdf_microfacet_beckmann_setup(ccl_private MicrofacetBsdf *bsdf)
   return SD_BSDF | SD_BSDF_HAS_EVAL;
 }
 
-/* Required to maintain OSL interface. */
-ccl_device int bsdf_microfacet_beckmann_isotropic_setup(ccl_private MicrofacetBsdf *bsdf)
-{
-  bsdf->alpha_y = bsdf->alpha_x;
-
-  return bsdf_microfacet_beckmann_setup(bsdf);
-}
-
 ccl_device int bsdf_microfacet_beckmann_refraction_setup(ccl_private MicrofacetBsdf *bsdf)
 {
   bsdf->alpha_x = saturatef(bsdf->alpha_x);

intern/cycles/kernel/closure/bsdf_util.h

@@ -90,8 +90,10 @@ ccl_device float schlick_fresnel(float u)
 }
 
 /* Calculate the fresnel color, which is a blend between white and the F0 color */
-ccl_device_forceinline Spectrum
+ccl_device_forceinline Spectrum interpolate_fresnel_color(float3 L,
-interpolate_fresnel_color(float3 L, float3 H, float ior, Spectrum F0)
+                                                          float3 H,
+                                                          float ior,
+                                                          Spectrum F0)
 {
   /* Compute the real Fresnel term and remap it from real_F0..1 to F0..1.
    * The reason why we use this remapping instead of directly doing the

intern/cycles/kernel/device/cpu/kernel_avx2.cpp

@@ -10,7 +10,7 @@
 #ifndef WITH_CYCLES_OPTIMIZED_KERNEL_AVX2
 #  define KERNEL_STUB
 #else
-/* SSE optimization disabled for now on 32 bit, see bug T36316. */
+/* SSE optimization disabled for now on 32 bit, see bug #36316. */
 #  if !(defined(__GNUC__) && (defined(i386) || defined(_M_IX86)))
 #    define __KERNEL_SSE__
 #    define __KERNEL_SSE2__

intern/cycles/kernel/device/cpu/kernel_sse2.cpp

@@ -10,7 +10,7 @@
 #ifndef WITH_CYCLES_OPTIMIZED_KERNEL_SSE2
 #  define KERNEL_STUB
 #else
-/* SSE optimization disabled for now on 32 bit, see bug T36316. */
+/* SSE optimization disabled for now on 32 bit, see bug #36316. */
 #  if !(defined(__GNUC__) && (defined(i386) || defined(_M_IX86)))
 #    define __KERNEL_SSE2__
 #  endif

intern/cycles/kernel/device/cpu/kernel_sse41.cpp

@@ -10,7 +10,7 @@
 #ifndef WITH_CYCLES_OPTIMIZED_KERNEL_SSE41
 #  define KERNEL_STUB
 #else
-/* SSE optimization disabled for now on 32 bit, see bug T36316. */
+/* SSE optimization disabled for now on 32 bit, see bug #36316. */
 #  if !(defined(__GNUC__) && (defined(i386) || defined(_M_IX86)))
 #    define __KERNEL_SSE2__
 #    define __KERNEL_SSE3__

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

@@ -5,13 +5,14 @@
 
 CCL_NAMESPACE_BEGIN
 
-#ifdef WITH_NANOVDB
+#if !defined __KERNEL_METAL__
-#  define NDEBUG /* Disable "assert" in device code */
+#  ifdef WITH_NANOVDB
-#  define NANOVDB_USE_INTRINSICS
+#    define NDEBUG /* Disable "assert" in device code */
-#  include "nanovdb/NanoVDB.h"
+#    define NANOVDB_USE_INTRINSICS
-#  include "nanovdb/util/SampleFromVoxels.h"
+#    include "nanovdb/NanoVDB.h"
+#    include "nanovdb/util/SampleFromVoxels.h"
+#  endif
 #endif
-
 /* w0, w1, w2, and w3 are the four cubic B-spline basis functions. */
 ccl_device float cubic_w0(float a)
 {
@@ -126,7 +127,7 @@ kernel_tex_image_interp_tricubic(ccl_global const TextureInfo &info, float x, fl
 #ifdef WITH_NANOVDB
 template<typename T, typename S>
 ccl_device typename nanovdb::NanoGrid<T>::ValueType kernel_tex_image_interp_tricubic_nanovdb(
-    S &s, float x, float y, float z)
+    ccl_private S &s, float x, float y, float z)
 {
   float px = floorf(x);
   float py = floorf(y);
@@ -157,13 +158,19 @@ ccl_device typename nanovdb::NanoGrid<T>::ValueType kernel_tex_image_interp_tric
                 g1y * (g0x * s(Vec3f(x0, y1, z1)) + g1x * s(Vec3f(x1, y1, z1))));
 }
 
+#  if defined(__KERNEL_METAL__)
+template<typename T>
+__attribute__((noinline)) typename nanovdb::NanoGrid<T>::ValueType kernel_tex_image_interp_nanovdb(
+    ccl_global const TextureInfo &info, float x, float y, float z, uint interpolation)
+#  else
 template<typename T>
 ccl_device_noinline typename nanovdb::NanoGrid<T>::ValueType kernel_tex_image_interp_nanovdb(
     ccl_global const TextureInfo &info, float x, float y, float z, uint interpolation)
+#  endif
 {
   using namespace nanovdb;
 
-  NanoGrid<T> *const grid = (NanoGrid<T> *)info.data;
+  ccl_global NanoGrid<T> *const grid = (ccl_global NanoGrid<T> *)info.data;
   typedef typename nanovdb::NanoGrid<T>::AccessorType AccessorType;
   AccessorType acc = grid->getAccessor();
 

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

@@ -401,6 +401,72 @@ ccl_gpu_kernel_threads(GPU_PARALLEL_SORTED_INDEX_DEFAULT_BLOCK_SIZE)
 }
 ccl_gpu_kernel_postfix
 
+ccl_gpu_kernel_threads(GPU_PARALLEL_SORT_BLOCK_SIZE)
+    ccl_gpu_kernel_signature(integrator_sort_bucket_pass,
+                             int num_states,
+                             int partition_size,
+                             int num_states_limit,
+                             ccl_global int *indices,
+                             int kernel_index)
+{
+#if defined(__KERNEL_LOCAL_ATOMIC_SORT__)
+  int max_shaders = context.launch_params_metal.data.max_shaders;
+  ccl_global ushort *d_queued_kernel = (ccl_global ushort *)
+                                           kernel_integrator_state.path.queued_kernel;
+  ccl_global uint *d_shader_sort_key = (ccl_global uint *)
+                                           kernel_integrator_state.path.shader_sort_key;
+  ccl_global int *key_offsets = (ccl_global int *)
+                                    kernel_integrator_state.sort_partition_key_offsets;
+
+  gpu_parallel_sort_bucket_pass(num_states,
+                                partition_size,
+                                max_shaders,
+                                kernel_index,
+                                d_queued_kernel,
+                                d_shader_sort_key,
+                                key_offsets,
+                                (threadgroup int *)threadgroup_array,
+                                metal_local_id,
+                                metal_local_size,
+                                metal_grid_id);
+#endif
+}
+ccl_gpu_kernel_postfix
+
+ccl_gpu_kernel_threads(GPU_PARALLEL_SORT_BLOCK_SIZE)
+    ccl_gpu_kernel_signature(integrator_sort_write_pass,
+                             int num_states,
+                             int partition_size,
+                             int num_states_limit,
+                             ccl_global int *indices,
+                             int kernel_index)
+{
+#if defined(__KERNEL_LOCAL_ATOMIC_SORT__)
+  int max_shaders = context.launch_params_metal.data.max_shaders;
+  ccl_global ushort *d_queued_kernel = (ccl_global ushort *)
+                                           kernel_integrator_state.path.queued_kernel;
+  ccl_global uint *d_shader_sort_key = (ccl_global uint *)
+                                           kernel_integrator_state.path.shader_sort_key;
+  ccl_global int *key_offsets = (ccl_global int *)
+                                    kernel_integrator_state.sort_partition_key_offsets;
+
+  gpu_parallel_sort_write_pass(num_states,
+                               partition_size,
+                               max_shaders,
+                               kernel_index,
+                               num_states_limit,
+                               indices,
+                               d_queued_kernel,
+                               d_shader_sort_key,
+                               key_offsets,
+                               (threadgroup int *)threadgroup_array,
+                               metal_local_id,
+                               metal_local_size,
+                               metal_grid_id);
+#endif
+}
+ccl_gpu_kernel_postfix
+
 ccl_gpu_kernel_threads(GPU_PARALLEL_ACTIVE_INDEX_DEFAULT_BLOCK_SIZE)
     ccl_gpu_kernel_signature(integrator_compact_paths_array,
                              int num_states,
@@ -579,7 +645,7 @@ ccl_device_inline void kernel_gpu_film_convert_half_write(ccl_global uchar4 *rgb
                                                           const int y,
                                                           const half4 half_pixel)
 {
-  /* Work around HIP issue with half float display, see T92972. */
+  /* Work around HIP issue with half float display, see #92972. */
 #ifdef __KERNEL_HIP__
   ccl_global half *out = ((ccl_global half *)rgba) + (rgba_offset + y * rgba_stride + x) * 4;
   out[0] = half_pixel.x;

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

@@ -178,7 +178,7 @@ __device__
                                          simd_lane_index, \
                                          simd_group_index, \
                                          num_simd_groups, \
-                                         simdgroup_offset)
+                                         (threadgroup int *)threadgroup_array)
 #elif defined(__KERNEL_ONEAPI__)
 
 #  define gpu_parallel_active_index_array(num_states, indices, num_indices, is_active_op) \

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

@@ -19,6 +19,115 @@ CCL_NAMESPACE_BEGIN
 #  define GPU_PARALLEL_SORTED_INDEX_DEFAULT_BLOCK_SIZE 512
 #endif
 #define GPU_PARALLEL_SORTED_INDEX_INACTIVE_KEY (~0)
+#define GPU_PARALLEL_SORT_BLOCK_SIZE 1024
+
+#if defined(__KERNEL_LOCAL_ATOMIC_SORT__)
+
+#  define atomic_store_local(p, x) \
+    atomic_store_explicit((threadgroup atomic_int *)p, x, memory_order_relaxed)
+#  define atomic_load_local(p) \
+    atomic_load_explicit((threadgroup atomic_int *)p, memory_order_relaxed)
+
+ccl_device_inline void gpu_parallel_sort_bucket_pass(const uint num_states,
+                                                     const uint partition_size,
+                                                     const uint max_shaders,
+                                                     const uint queued_kernel,
+                                                     ccl_global ushort *d_queued_kernel,
+                                                     ccl_global uint *d_shader_sort_key,
+                                                     ccl_global int *partition_key_offsets,
+                                                     ccl_gpu_shared int *buckets,
+                                                     const ushort local_id,
+                                                     const ushort local_size,
+                                                     const ushort grid_id)
+{
+  /* Zero the bucket sizes. */
+  if (local_id < max_shaders) {
+    atomic_store_local(&buckets[local_id], 0);
+  }
+
+  ccl_gpu_syncthreads();
+
+  /* Determine bucket sizes within the partitions. */
+
+  const uint partition_start = partition_size * uint(grid_id);
+  const uint partition_end = min(num_states, partition_start + partition_size);
+
+  for (int state_index = partition_start + uint(local_id); state_index < partition_end;
+       state_index += uint(local_size)) {
+    ushort kernel_index = d_queued_kernel[state_index];
+    if (kernel_index == queued_kernel) {
+      uint key = d_shader_sort_key[state_index] % max_shaders;
+      atomic_fetch_and_add_uint32(&buckets[key], 1);
+    }
+  }
+
+  ccl_gpu_syncthreads();
+
+  /* Calculate the partition's local offsets from the prefix sum of bucket sizes. */
+
+  if (local_id == 0) {
+    int offset = 0;
+    for (int i = 0; i < max_shaders; i++) {
+      partition_key_offsets[i + uint(grid_id) * (max_shaders + 1)] = offset;
+      offset = offset + atomic_load_local(&buckets[i]);
+    }
+
+    /* Store the number of active states in this partition. */
+    partition_key_offsets[max_shaders + uint(grid_id) * (max_shaders + 1)] = offset;
+  }
+}
+
+ccl_device_inline void gpu_parallel_sort_write_pass(const uint num_states,
+                                                    const uint partition_size,
+                                                    const uint max_shaders,
+                                                    const uint queued_kernel,
+                                                    const int num_states_limit,
+                                                    ccl_global int *indices,
+                                                    ccl_global ushort *d_queued_kernel,
+                                                    ccl_global uint *d_shader_sort_key,
+                                                    ccl_global int *partition_key_offsets,
+                                                    ccl_gpu_shared int *local_offset,
+                                                    const ushort local_id,
+                                                    const ushort local_size,
+                                                    const ushort grid_id)
+{
+  /* Calculate each partition's global offset from the prefix sum of the active state counts per
+   * partition. */
+
+  if (local_id < max_shaders) {
+    int partition_offset = 0;
+    for (int i = 0; i < uint(grid_id); i++) {
+      int partition_key_count = partition_key_offsets[max_shaders + uint(i) * (max_shaders + 1)];
+      partition_offset += partition_key_count;
+    }
+
+    ccl_global int *key_offsets = partition_key_offsets + (uint(grid_id) * (max_shaders + 1));
+    atomic_store_local(&local_offset[local_id], key_offsets[local_id] + partition_offset);
+  }
+
+  ccl_gpu_syncthreads();
+
+  /* Write the sorted active indices. */
+
+  const uint partition_start = partition_size * uint(grid_id);
+  const uint partition_end = min(num_states, partition_start + partition_size);
+
+  ccl_global int *key_offsets = partition_key_offsets + (uint(grid_id) * max_shaders);
+
+  for (int state_index = partition_start + uint(local_id); state_index < partition_end;
+       state_index += uint(local_size)) {
+    ushort kernel_index = d_queued_kernel[state_index];
+    if (kernel_index == queued_kernel) {
+      uint key = d_shader_sort_key[state_index] % max_shaders;
+      int index = atomic_fetch_and_add_uint32(&local_offset[key], 1);
+      if (index < num_states_limit) {
+        indices[index] = state_index;
+      }
+    }
+  }
+}
+
+#endif /* __KERNEL_LOCAL_ATOMIC_SORT__ */
 
 template<typename GetKeyOp>
 __device__ void gpu_parallel_sorted_index_array(const uint state_index,

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

@@ -172,17 +172,14 @@ ccl_device_intersect bool scene_intersect_local(KernelGlobals kg,
     kernel_assert(!"Invalid ift_local");
     return false;
   }
-#  endif
+  if (is_null_intersection_function_table(metal_ancillaries->ift_local_prim)) {
-
+    if (local_isect) {
-  metal::raytracing::ray r(ray->P, ray->D, ray->tmin, ray->tmax);
+      local_isect->num_hits = 0;
-  metalrt_intersector_type metalrt_intersect;
+    }
-
+    kernel_assert(!"Invalid ift_local_prim");
-  metalrt_intersect.force_opacity(metal::raytracing::forced_opacity::non_opaque);
+    return false;
-
-  bool triangle_only = !kernel_data.bvh.have_curves && !kernel_data.bvh.have_points;
-  if (triangle_only) {
-    metalrt_intersect.assume_geometry_type(metal::raytracing::geometry_type::triangle);
   }
+#  endif
 
   MetalRTIntersectionLocalPayload payload;
   payload.self = ray->self;
@@ -195,14 +192,48 @@ ccl_device_intersect bool scene_intersect_local(KernelGlobals kg,
   }
   payload.result = false;
 
-  typename metalrt_intersector_type::result_type intersection;
+  metal::raytracing::ray r(ray->P, ray->D, ray->tmin, ray->tmax);
 
 #  if defined(__METALRT_MOTION__)
+  metalrt_intersector_type metalrt_intersect;
+  typename metalrt_intersector_type::result_type intersection;
+
+  metalrt_intersect.force_opacity(metal::raytracing::forced_opacity::non_opaque);
+  bool triangle_only = !kernel_data.bvh.have_curves && !kernel_data.bvh.have_points;
+  if (triangle_only) {
+    metalrt_intersect.assume_geometry_type(metal::raytracing::geometry_type::triangle);
+  }
+
   intersection = metalrt_intersect.intersect(
       r, metal_ancillaries->accel_struct, 0xFF, ray->time, metal_ancillaries->ift_local, payload);
 #  else
+
+  metalrt_blas_intersector_type metalrt_intersect;
+  typename metalrt_blas_intersector_type::result_type intersection;
+
+  metalrt_intersect.force_opacity(metal::raytracing::forced_opacity::non_opaque);
+  bool triangle_only = !kernel_data.bvh.have_curves && !kernel_data.bvh.have_points;
+  if (triangle_only) {
+    metalrt_intersect.assume_geometry_type(metal::raytracing::geometry_type::triangle);
+  }
+
+  // if we know we are going to get max one hit, like for random-sss-walk we can
+  // optimize and accept the first hit
+  if (max_hits == 1) {
+    metalrt_intersect.accept_any_intersection(true);
+  }
+
+  int blas_index = metal_ancillaries->blas_userID_to_index_lookUp[local_object];
+  // transform the ray into object's local space
+  Transform itfm = kernel_data_fetch(objects, local_object).itfm;
+  r.origin = transform_point(&itfm, r.origin);
+  r.direction = transform_direction(&itfm, r.direction);
+
   intersection = metalrt_intersect.intersect(
-      r, metal_ancillaries->accel_struct, 0xFF, metal_ancillaries->ift_local, payload);
+      r,
+      metal_ancillaries->blas_accel_structs[blas_index].blas,
+      metal_ancillaries->ift_local_prim,
+      payload);
 #  endif
 
   if (lcg_state) {

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

@@ -105,10 +105,11 @@ struct kernel_gpu_##name \
 { \
   PARAMS_MAKER(__VA_ARGS__)(__VA_ARGS__) \
   void run(thread MetalKernelContext& context, \
-           threadgroup int *simdgroup_offset, \
+           threadgroup atomic_int *threadgroup_array, \
            const uint metal_global_id, \
            const ushort metal_local_id, \
            const ushort metal_local_size, \
+           const ushort metal_grid_id, \
            uint simdgroup_size, \
            uint simd_lane_index, \
            uint simd_group_index, \
@@ -117,22 +118,24 @@ struct kernel_gpu_##name \
 kernel void cycles_metal_##name(device const kernel_gpu_##name *params_struct, \
                                 constant KernelParamsMetal &ccl_restrict   _launch_params_metal, \
                                 constant MetalAncillaries *_metal_ancillaries, \
-                                threadgroup int *simdgroup_offset[[ threadgroup(0) ]], \
+                                threadgroup atomic_int *threadgroup_array[[ threadgroup(0) ]], \
                                 const uint metal_global_id [[thread_position_in_grid]], \
                                 const ushort metal_local_id   [[thread_position_in_threadgroup]], \
                                 const ushort metal_local_size [[threads_per_threadgroup]], \
+                                const ushort metal_grid_id    [[threadgroup_position_in_grid]], \
                                 uint simdgroup_size [[threads_per_simdgroup]], \
                                 uint simd_lane_index [[thread_index_in_simdgroup]], \
                                 uint simd_group_index [[simdgroup_index_in_threadgroup]], \
                                 uint num_simd_groups [[simdgroups_per_threadgroup]]) { \
   MetalKernelContext context(_launch_params_metal, _metal_ancillaries); \
-  params_struct->run(context, simdgroup_offset, metal_global_id, metal_local_id, metal_local_size, simdgroup_size, simd_lane_index, simd_group_index, num_simd_groups); \
+  params_struct->run(context, threadgroup_array, metal_global_id, metal_local_id, metal_local_size, metal_grid_id, simdgroup_size, simd_lane_index, simd_group_index, num_simd_groups); \
 } \
 void kernel_gpu_##name::run(thread MetalKernelContext& context, \
-                  threadgroup int *simdgroup_offset, \
+                  threadgroup atomic_int *threadgroup_array, \
                   const uint metal_global_id, \
                   const ushort metal_local_id, \
                   const ushort metal_local_size, \
+                  const ushort metal_grid_id, \
                   uint simdgroup_size, \
                   uint simd_lane_index, \
                   uint simd_group_index, \
@@ -263,18 +266,34 @@ ccl_device_forceinline uchar4 make_uchar4(const uchar x,
 
 #  if defined(__METALRT_MOTION__)
 #    define METALRT_TAGS instancing, instance_motion, primitive_motion
+#    define METALRT_BLAS_TAGS , primitive_motion
 #  else
 #    define METALRT_TAGS instancing
+#    define METALRT_BLAS_TAGS
 #  endif /* __METALRT_MOTION__ */
 
 typedef acceleration_structure<METALRT_TAGS> metalrt_as_type;
 typedef intersection_function_table<triangle_data, METALRT_TAGS> metalrt_ift_type;
 typedef metal::raytracing::intersector<triangle_data, METALRT_TAGS> metalrt_intersector_type;
+#  if defined(__METALRT_MOTION__)
+typedef acceleration_structure<primitive_motion> metalrt_blas_as_type;
+typedef intersection_function_table<triangle_data, primitive_motion> metalrt_blas_ift_type;
+typedef metal::raytracing::intersector<triangle_data, primitive_motion>
+    metalrt_blas_intersector_type;
+#  else
+typedef acceleration_structure<> metalrt_blas_as_type;
+typedef intersection_function_table<triangle_data> metalrt_blas_ift_type;
+typedef metal::raytracing::intersector<triangle_data> metalrt_blas_intersector_type;
+#  endif
 
 #endif /* __METALRT__ */
 
 /* texture bindings and sampler setup */
 
+struct Buffer1DParamsMetal {
+  device float *buf;
+};
+
 struct Texture2DParamsMetal {
   texture2d<float, access::sample> tex;
 };
@@ -282,15 +301,25 @@ struct Texture3DParamsMetal {
   texture3d<float, access::sample> tex;
 };
 
+#ifdef __METALRT__
+struct MetalRTBlasWrapper {
+  metalrt_blas_as_type blas;
+};
+#endif
+
 struct MetalAncillaries {
   device Texture2DParamsMetal *textures_2d;
   device Texture3DParamsMetal *textures_3d;
+  device Buffer1DParamsMetal *buffers;
 
 #ifdef __METALRT__
   metalrt_as_type accel_struct;
   metalrt_ift_type ift_default;
   metalrt_ift_type ift_shadow;
   metalrt_ift_type ift_local;
+  metalrt_blas_ift_type ift_local_prim;
+  constant MetalRTBlasWrapper *blas_accel_structs;
+  constant int *blas_userID_to_index_lookUp;
 #endif
 };
 

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

@@ -3,6 +3,13 @@
 
 // clang-format off
 
+#ifdef WITH_NANOVDB
+#  define NDEBUG /* Disable "assert" in device code */
+#  define NANOVDB_USE_INTRINSICS
+#  include "nanovdb/NanoVDB.h"
+#  include "nanovdb/util/SampleFromVoxels.h"
+#endif
+
 /* Open the Metal kernel context class
  * Necessary to access resource bindings */
 class MetalKernelContext {

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

@@ -139,6 +139,20 @@ TReturn metalrt_local_hit(constant KernelParamsMetal &launch_params_metal,
 #endif
 }
 
+[[intersection(triangle, triangle_data )]] TriangleIntersectionResult
+__anyhit__cycles_metalrt_local_hit_tri_prim(
+    constant KernelParamsMetal &launch_params_metal [[buffer(1)]],
+    ray_data MetalKernelContext::MetalRTIntersectionLocalPayload &payload [[payload]],
+    uint primitive_id [[primitive_id]],
+    float2 barycentrics [[barycentric_coord]],
+    float ray_tmax [[distance]])
+{
+  //instance_id, aka the user_id has been removed. If we take this function we optimized the
+  //SSS for starting traversal from a primitive acceleration structure instead of the root of the global AS.
+  //this means we will always be intersecting the correct object no need for the userid to check
+  return metalrt_local_hit<TriangleIntersectionResult, METALRT_HIT_TRIANGLE>(
+      launch_params_metal, payload, payload.local_object, primitive_id, barycentrics, ray_tmax);
+}
 [[intersection(triangle, triangle_data, METALRT_TAGS)]] TriangleIntersectionResult
 __anyhit__cycles_metalrt_local_hit_tri(
     constant KernelParamsMetal &launch_params_metal [[buffer(1)]],
@@ -163,6 +177,17 @@ __anyhit__cycles_metalrt_local_hit_box(const float ray_tmax [[max_distance]])
   return result;
 }
 
+[[intersection(bounding_box, triangle_data )]] BoundingBoxIntersectionResult
+__anyhit__cycles_metalrt_local_hit_box_prim(const float ray_tmax [[max_distance]])
+{
+  /* unused function */
+  BoundingBoxIntersectionResult result;
+  result.distance = ray_tmax;
+  result.accept = false;
+  result.continue_search = false;
+  return result;
+}
+
 template<uint intersection_type>
 bool metalrt_shadow_all_hit(constant KernelParamsMetal &launch_params_metal,
                             ray_data MetalKernelContext::MetalRTIntersectionShadowPayload &payload,

intern/cycles/kernel/device/oneapi/kernel.cpp

@@ -372,6 +372,16 @@ bool oneapi_enqueue_kernel(KernelContext *kernel_context,
               kg, cgh, global_size, local_size, args, oneapi_kernel_integrator_sorted_paths_array);
           break;
         }
+        case DEVICE_KERNEL_INTEGRATOR_SORT_BUCKET_PASS: {
+          oneapi_call(
+              kg, cgh, global_size, local_size, args, oneapi_kernel_integrator_sort_bucket_pass);
+          break;
+        }
+        case DEVICE_KERNEL_INTEGRATOR_SORT_WRITE_PASS: {
+          oneapi_call(
+              kg, cgh, global_size, local_size, args, oneapi_kernel_integrator_sort_write_pass);
+          break;
+        }
         case DEVICE_KERNEL_INTEGRATOR_COMPACT_PATHS_ARRAY: {
           oneapi_call(kg,
                       cgh,

intern/cycles/kernel/integrator/state.h

@@ -132,6 +132,9 @@ typedef struct IntegratorStateGPU {
   /* Index of main path which will be used by a next shadow catcher split.  */
   ccl_global int *next_main_path_index;
 
+  /* Partition/key offsets used when writing sorted active indices. */
+  ccl_global int *sort_partition_key_offsets;
+
   /* Divisor used to partition active indices by locality when sorting by material.  */
   uint sort_partition_divisor;
 } IntegratorStateGPU;