test using compiled function in math node

This just moves the relations update to a lower level function that is used
by other functions. Eventually, the special case for this node should be
generalized.

.clang-format

@@ -269,9 +269,5 @@ StatementMacros:
   - PyObject_HEAD
   - PyObject_VAR_HEAD
 
-StatementMacros:
-  - GPU_STAGE_INTERFACE_CREATE
-  - GPU_SHADER_DESCRIPTOR
-
 MacroBlockBegin: "^BSDF_CLOSURE_CLASS_BEGIN$"
 MacroBlockEnd: "^BSDF_CLOSURE_CLASS_END$"

CMakeLists.txt

@@ -187,6 +187,13 @@ mark_as_advanced(CPACK_OVERRIDE_PACKAGENAME)
 mark_as_advanced(BUILDINFO_OVERRIDE_DATE)
 mark_as_advanced(BUILDINFO_OVERRIDE_TIME)
 
+if(${CMAKE_VERSION} VERSION_GREATER_EQUAL "3.16")
+  option(WITH_UNITY_BUILD   "Enable unity build for modules that support it to improve compile times" ON)
+  mark_as_advanced(WITH_UNITY_BUILD)
+else()
+  set(WITH_UNITY_BUILD OFF)
+endif()
+
 option(WITH_IK_ITASC      "Enable ITASC IK solver (only disable for development & for incompatible C++ compilers)" ON)
 option(WITH_IK_SOLVER     "Enable Legacy IK solver (only disable for development)" ON)
 option(WITH_FFTW3         "Enable FFTW3 support (Used for smoke, ocean sim, and audio effects)" ON)
@@ -426,30 +433,40 @@ mark_as_advanced(WITH_CYCLES_DEBUG_NAN)
 mark_as_advanced(WITH_CYCLES_NATIVE_ONLY)
 
 # NVIDIA CUDA & OptiX
-option(WITH_CYCLES_DEVICE_CUDA       "Enable Cycles NVIDIA CUDA compute support" ON)
+if(NOT APPLE)
-option(WITH_CYCLES_DEVICE_OPTIX      "Enable Cycles NVIDIA OptiX support" ON)
+  option(WITH_CYCLES_DEVICE_CUDA       "Enable Cycles NVIDIA CUDA compute support" ON)
-mark_as_advanced(WITH_CYCLES_DEVICE_CUDA)
+  option(WITH_CYCLES_DEVICE_OPTIX      "Enable Cycles NVIDIA OptiX support" ON)
+  mark_as_advanced(WITH_CYCLES_DEVICE_CUDA)
 
-option(WITH_CYCLES_CUDA_BINARIES     "Build Cycles NVIDIA CUDA binaries" OFF)
+  option(WITH_CYCLES_CUDA_BINARIES     "Build Cycles NVIDIA CUDA binaries" OFF)
-set(CYCLES_CUDA_BINARIES_ARCH sm_30 sm_35 sm_37 sm_50 sm_52 sm_60 sm_61 sm_70 sm_75 sm_86 compute_75 CACHE STRING "CUDA architectures to build binaries for")
+  set(CYCLES_CUDA_BINARIES_ARCH sm_30 sm_35 sm_37 sm_50 sm_52 sm_60 sm_61 sm_70 sm_75 sm_86 compute_75 CACHE STRING "CUDA architectures to build binaries for")
-option(WITH_CYCLES_CUBIN_COMPILER    "Build cubins with nvrtc based compiler instead of nvcc" OFF)
+  option(WITH_CYCLES_CUBIN_COMPILER    "Build cubins with nvrtc based compiler instead of nvcc" OFF)
-option(WITH_CYCLES_CUDA_BUILD_SERIAL "Build cubins one after another (useful on machines with limited RAM)" OFF)
+  option(WITH_CYCLES_CUDA_BUILD_SERIAL "Build cubins one after another (useful on machines with limited RAM)" OFF)
-option(WITH_CUDA_DYNLOAD             "Dynamically load CUDA libraries at runtime (for developers, makes cuda-gdb work)" ON)
+  option(WITH_CUDA_DYNLOAD             "Dynamically load CUDA libraries at runtime (for developers, makes cuda-gdb work)" ON)
-mark_as_advanced(CYCLES_CUDA_BINARIES_ARCH)
+  mark_as_advanced(CYCLES_CUDA_BINARIES_ARCH)
-mark_as_advanced(WITH_CYCLES_CUBIN_COMPILER)
+  mark_as_advanced(WITH_CYCLES_CUBIN_COMPILER)
-mark_as_advanced(WITH_CYCLES_CUDA_BUILD_SERIAL)
+  mark_as_advanced(WITH_CYCLES_CUDA_BUILD_SERIAL)
-mark_as_advanced(WITH_CUDA_DYNLOAD)
+  mark_as_advanced(WITH_CUDA_DYNLOAD)
+endif()
 
 # AMD HIP
-if(WIN32)
+if(NOT APPLE)
+  if(WIN32)
     option(WITH_CYCLES_DEVICE_HIP        "Enable Cycles AMD HIP support" ON)
-else()
+  else()
     option(WITH_CYCLES_DEVICE_HIP        "Enable Cycles AMD HIP support" OFF)
+  endif()
+
+  option(WITH_CYCLES_HIP_BINARIES      "Build Cycles AMD HIP binaries" OFF)
+  set(CYCLES_HIP_BINARIES_ARCH gfx1010 gfx1011 gfx1012 gfx1030 gfx1031 gfx1032 gfx1034 CACHE STRING "AMD HIP architectures to build binaries for")
+  mark_as_advanced(WITH_CYCLES_DEVICE_HIP)
+  mark_as_advanced(CYCLES_HIP_BINARIES_ARCH)
+endif()
+
+# Apple Metal
+if(APPLE)
+  option(WITH_CYCLES_DEVICE_METAL       "Enable Cycles Apple Metal compute support" ON)
 endif()
-option(WITH_CYCLES_HIP_BINARIES      "Build Cycles AMD HIP binaries" OFF)
-set(CYCLES_HIP_BINARIES_ARCH gfx1010 gfx1011 gfx1012 gfx1030 gfx1031 gfx1032 gfx1034 CACHE STRING "AMD HIP architectures to build binaries for")
-mark_as_advanced(WITH_CYCLES_DEVICE_HIP)
-mark_as_advanced(CYCLES_HIP_BINARIES_ARCH)
 
 # Draw Manager
 option(WITH_DRAW_DEBUG "Add extra debug capabilities to Draw Manager" OFF)
@@ -494,11 +511,10 @@ if(WIN32)
 endif()
 
 # This should be turned off when Blender enter beta/rc/release
-if("${BLENDER_VERSION_CYCLE}" STREQUAL "release" OR
+if("${BLENDER_VERSION_CYCLE}" STREQUAL "alpha")
-   "${BLENDER_VERSION_CYCLE}" STREQUAL "rc")
-  set(WITH_EXPERIMENTAL_FEATURES OFF)
-else()
   set(WITH_EXPERIMENTAL_FEATURES ON)
+else()
+  set(WITH_EXPERIMENTAL_FEATURES OFF)
 endif()
 
 # Unit testsing
@@ -840,7 +856,7 @@ if(WITH_AUDASPACE)
 endif()
 
 # Auto-enable CUDA dynload if toolkit is not found.
-if(NOT WITH_CUDA_DYNLOAD)
+if(WITH_CYCLES AND WITH_CYCLES_DEVICE_CUDA AND NOT WITH_CUDA_DYNLOAD)
   find_package(CUDA)
   if(NOT CUDA_FOUND)
     message(STATUS "CUDA toolkit not found, using dynamic runtime loading of libraries (WITH_CUDA_DYNLOAD) instead")

build_files/build_environment/install_deps.sh

@@ -2083,9 +2083,9 @@ compile_OIIO() {
     cmake_d="$cmake_d -D OPENEXR_VERSION=$OPENEXR_VERSION"
 
     if [ "$_with_built_openexr" = true ]; then
-      cmake_d="$cmake_d -D ILMBASE_HOME=$INST/openexr"
+      cmake_d="$cmake_d -D ILMBASE_ROOT=$INST/openexr"
-      cmake_d="$cmake_d -D OPENEXR_HOME=$INST/openexr"
+      cmake_d="$cmake_d -D OPENEXR_ROOT=$INST/openexr"
-      INFO "ILMBASE_HOME=$INST/openexr"
+      INFO "Ilmbase_ROOT=$INST/openexr"
     fi
 
     # ptex is only needed when nicholas bishop is ready
@@ -2374,9 +2374,9 @@ compile_OSL() {
     #~ cmake_d="$cmake_d -D ILMBASE_VERSION=$ILMBASE_VERSION"
 
     if [ "$_with_built_openexr" = true ]; then
-      INFO "ILMBASE_HOME=$INST/openexr"
+      cmake_d="$cmake_d -D ILMBASE_ROOT=$INST/openexr"
-      cmake_d="$cmake_d -D OPENEXR_ROOT_DIR=$INST/openexr"
+      cmake_d="$cmake_d -D OPENEXR_ROOT=$INST/openexr"
-      cmake_d="$cmake_d -D ILMBASE_ROOT_DIR=$INST/openexr"
+      INFO "Ilmbase_ROOT=$INST/openexr"
       # XXX Temp workaround... sigh, ILMBase really messed the things up by defining their custom names ON by default :(
     fi
 

build_files/build_environment/patches/usd.diff

@@ -197,3 +197,38 @@ index 67ec0d15f..6dc3e85a0 100644
  #else
  #error Unknown architecture.
  #endif
+
+diff --git a/pxr/base/arch/demangle.cpp b/pxr/base/arch/demangle.cpp
+index 67ec0d15f..6dc3e85a0 100644
+--- a/pxr/base/arch/demangle.cpp
++++ b/pxr/base/arch/demangle.cpp
+@@ -36,6 +36,7 @@
+ #if (ARCH_COMPILER_GCC_MAJOR == 3 && ARCH_COMPILER_GCC_MINOR >= 1) || \
+     ARCH_COMPILER_GCC_MAJOR > 3 || defined(ARCH_COMPILER_CLANG)
+ #define _AT_LEAST_GCC_THREE_ONE_OR_CLANG
++#include <cxxabi.h>
+ #endif
+ 
+ PXR_NAMESPACE_OPEN_SCOPE
+@@ -138,7 +139,6 @@
+ #endif
+ 
+ #if defined(_AT_LEAST_GCC_THREE_ONE_OR_CLANG)
+-#include <cxxabi.h>
+ 
+ /*
+  * This routine doesn't work when you get to gcc3.4.
+
+diff --git a/pxr/base/work/singularTask.h b/pxr/base/work/singularTask.h
+index 67ec0d15f..6dc3e85a0 100644
+--- a/pxr/base/work/singularTask.h
++++ b/pxr/base/work/singularTask.h
+@@ -120,7 +120,7 @@
+                     // case we go again to ensure the task can do whatever it
+                     // was awakened to do.  Once we successfully take the count
+                     // to zero, we stop.
+-                    size_t old = count;
++                    std::size_t old = count;
+                     do { _fn(); } while (
+                         !count.compare_exchange_strong(old, 0));
+                 });

build_files/cmake/config/blender_lite.cmake

@@ -19,9 +19,6 @@ set(WITH_CODEC_SNDFILE       OFF CACHE BOOL "" FORCE)
 set(WITH_COMPOSITOR          OFF CACHE BOOL "" FORCE)
 set(WITH_COREAUDIO           OFF CACHE BOOL "" FORCE)
 set(WITH_CYCLES              OFF CACHE BOOL "" FORCE)
-set(WITH_CYCLES_DEVICE_OPTIX OFF CACHE BOOL "" FORCE)
-set(WITH_CYCLES_EMBREE       OFF CACHE BOOL "" FORCE)
-set(WITH_CYCLES_OSL          OFF CACHE BOOL "" FORCE)
 set(WITH_DRACO               OFF CACHE BOOL "" FORCE)
 set(WITH_FFTW3               OFF CACHE BOOL "" FORCE)
 set(WITH_FREESTYLE           OFF CACHE BOOL "" FORCE)

build_files/cmake/config/blender_release.cmake

@@ -61,6 +61,7 @@ set(WITH_MEM_JEMALLOC          ON  CACHE BOOL "" FORCE)
 # platform dependent options
 if(APPLE)
   set(WITH_COREAUDIO           ON  CACHE BOOL "" FORCE)
+  set(WITH_CYCLES_DEVICE_METAL ON  CACHE BOOL "" FORCE)
 endif()
 if(NOT WIN32)
   set(WITH_JACK                ON  CACHE BOOL "" FORCE)

build_files/cmake/platform/platform_apple.cmake

@@ -257,9 +257,6 @@ if(WITH_BOOST)
   if(WITH_INTERNATIONAL)
     list(APPEND _boost_FIND_COMPONENTS locale)
   endif()
-  if(WITH_CYCLES_NETWORK)
-    list(APPEND _boost_FIND_COMPONENTS serialization)
-  endif()
   if(WITH_OPENVDB)
     list(APPEND _boost_FIND_COMPONENTS iostreams)
   endif()
@@ -339,7 +336,7 @@ if(WITH_LLVM)
 
 endif()
 
-if(WITH_CYCLES_OSL)
+if(WITH_CYCLES AND WITH_CYCLES_OSL)
   set(CYCLES_OSL ${LIBDIR}/osl)
 
   find_library(OSL_LIB_EXEC NAMES oslexec PATHS ${CYCLES_OSL}/lib)
@@ -359,7 +356,7 @@ if(WITH_CYCLES_OSL)
   endif()
 endif()
 
-if(WITH_CYCLES_EMBREE)
+if(WITH_CYCLES AND WITH_CYCLES_EMBREE)
   find_package(Embree 3.8.0 REQUIRED)
   # Increase stack size for Embree, only works for executables.
   if(NOT WITH_PYTHON_MODULE)

build_files/cmake/platform/platform_unix.cmake

@@ -241,7 +241,7 @@ if(WITH_INPUT_NDOF)
   endif()
 endif()
 
-if(WITH_CYCLES_OSL)
+if(WITH_CYCLES AND WITH_CYCLES_OSL)
   set(CYCLES_OSL ${LIBDIR}/osl CACHE PATH "Path to OpenShadingLanguage installation")
   if(EXISTS ${CYCLES_OSL} AND NOT OSL_ROOT)
     set(OSL_ROOT ${CYCLES_OSL})
@@ -314,7 +314,7 @@ if(WITH_BOOST)
     endif()
     set(Boost_USE_MULTITHREADED ON)
     set(__boost_packages filesystem regex thread date_time)
-    if(WITH_CYCLES_OSL)
+    if(WITH_CYCLES AND WITH_CYCLES_OSL)
       if(NOT (${OSL_LIBRARY_VERSION_MAJOR} EQUAL "1" AND ${OSL_LIBRARY_VERSION_MINOR} LESS "6"))
         list(APPEND __boost_packages wave)
       else()
@@ -323,9 +323,6 @@ if(WITH_BOOST)
     if(WITH_INTERNATIONAL)
       list(APPEND __boost_packages locale)
     endif()
-    if(WITH_CYCLES_NETWORK)
-      list(APPEND __boost_packages serialization)
-    endif()
     if(WITH_OPENVDB)
       list(APPEND __boost_packages iostreams)
     endif()
@@ -403,7 +400,7 @@ if(WITH_OPENCOLORIO)
   endif()
 endif()
 
-if(WITH_CYCLES_EMBREE)
+if(WITH_CYCLES AND WITH_CYCLES_EMBREE)
   find_package(Embree 3.8.0 REQUIRED)
 endif()
 

build_files/cmake/platform/platform_win32.cmake

@@ -477,7 +477,7 @@ if(WITH_PYTHON)
 endif()
 
 if(WITH_BOOST)
-  if(WITH_CYCLES_OSL)
+  if(WITH_CYCLES AND WITH_CYCLES_OSL)
     set(boost_extra_libs wave)
   endif()
   if(WITH_INTERNATIONAL)
@@ -520,7 +520,7 @@ if(WITH_BOOST)
       debug ${BOOST_LIBPATH}/libboost_thread-${BOOST_DEBUG_POSTFIX}
       debug ${BOOST_LIBPATH}/libboost_chrono-${BOOST_DEBUG_POSTFIX}
     )
-    if(WITH_CYCLES_OSL)
+    if(WITH_CYCLES AND WITH_CYCLES_OSL)
       set(BOOST_LIBRARIES ${BOOST_LIBRARIES}
         optimized ${BOOST_LIBPATH}/libboost_wave-${BOOST_POSTFIX}
         debug ${BOOST_LIBPATH}/libboost_wave-${BOOST_DEBUG_POSTFIX})
@@ -708,7 +708,7 @@ if(WITH_CODEC_SNDFILE)
   set(LIBSNDFILE_LIBRARIES ${LIBSNDFILE_LIBPATH}/libsndfile-1.lib)
 endif()
 
-if(WITH_CYCLES_OSL)
+if(WITH_CYCLES AND WITH_CYCLES_OSL)
   set(CYCLES_OSL ${LIBDIR}/osl CACHE PATH "Path to OpenShadingLanguage installation")
   set(OSL_SHADER_DIR ${CYCLES_OSL}/shaders)
   # Shaders have moved around a bit between OSL versions, check multiple locations
@@ -741,7 +741,7 @@ if(WITH_CYCLES_OSL)
   endif()
 endif()
 
-if(WITH_CYCLES_EMBREE)
+if(WITH_CYCLES AND WITH_CYCLES_EMBREE)
   windows_find_package(Embree)
   if(NOT EMBREE_FOUND)
     set(EMBREE_INCLUDE_DIRS ${LIBDIR}/embree/include)

doc/doxygen/doxygen.intern.h

@@ -6,91 +6,90 @@
  *  as part of the normal development process.
  */
 
-/** \defgroup MEM Guarded memory (de)allocation
+/* TODO: other modules.
- *  \ingroup intern
+ * - `libmv`
+ * - `cycles`
+ * - `opencolorio`
+ * - `opensubdiv`
+ * - `openvdb`
+ * - `quadriflow`
  */
 
-/** \defgroup clog C-Logging (CLOG)
+/** \defgroup intern_atomic Atomic Operations
- *  \ingroup intern
+ *  \ingroup intern */
- */
 
-/** \defgroup ctr container
+/** \defgroup intern_clog C-Logging (CLOG)
- *  \ingroup intern
+ *  \ingroup intern */
- */
 
-/** \defgroup iksolver iksolver
+/** \defgroup intern_eigen Eigen
- *  \ingroup intern
+ *  \ingroup intern */
- */
 
-/** \defgroup itasc itasc
+/** \defgroup intern_glew-mx GLEW with Multiple Rendering Context's
- *  \ingroup intern
+ *  \ingroup intern */
- */
 
-/** \defgroup memutil memutil
+/** \defgroup intern_iksolver Inverse Kinematics (Solver)
- *  \ingroup intern
+ *  \ingroup intern */
- */
 
-/** \defgroup mikktspace mikktspace
+/** \defgroup intern_itasc Inverse Kinematics (ITASC)
- *  \ingroup intern
+ *  \ingroup intern */
- */
 
-/** \defgroup moto moto
+/** \defgroup intern_libc_compat libc Compatibility For Linux
- *  \ingroup intern
+ *  \ingroup intern */
- */
 
-/** \defgroup eigen eigen
+/** \defgroup intern_locale Locale
- *  \ingroup intern
+ *  \ingroup intern */
- */
 
-/** \defgroup smoke smoke
+/** \defgroup intern_mantaflow Manta-Flow Fluid Simulation
- *  \ingroup intern
+ *  \ingroup intern */
- */
 
-/** \defgroup string string
+/** \defgroup intern_mem Guarded Memory (de)allocation
- *  \ingroup intern
+ *  \ingroup intern */
- */
+
+/** \defgroup intern_memutil Memory Utilities (memutil)
+ *  \ingroup intern */
+
+/** \defgroup intern_mikktspace MikktSpace
+ *  \ingroup intern */
+
+/** \defgroup intern_numaapi NUMA (Non Uniform Memory Architecture)
+ *  \ingroup intern */
+
+/** \defgroup intern_rigidbody Rigid-Body C-API
+ *  \ingroup intern */
+
+/** \defgroup intern_sky_model Sky Model
+ *  \ingroup intern */
+
+/** \defgroup intern_utf_conv UTF-8/16 Conversion (utfconv)
+ *  \ingroup intern */
 
 /** \defgroup audaspace Audaspace
  *  \ingroup intern undoc
- *  \todo add to doxygen
+ *  \todo add to doxygen */
- */
 /** \defgroup audcoreaudio Audaspace CoreAudio
- *  \ingroup audaspace
+ *  \ingroup audaspace */
- */
 /** \defgroup audfx Audaspace FX
- *  \ingroup audaspace
+ *  \ingroup audaspace */
- */
 /** \defgroup audopenal Audaspace OpenAL
- *  \ingroup audaspace
+ *  \ingroup audaspace */
- */
 /** \defgroup audpulseaudio Audaspace PulseAudio
- *  \ingroup audaspace
+ *  \ingroup audaspace */
- */
 /** \defgroup audwasapi Audaspace WASAPI
- *  \ingroup audaspace
+ *  \ingroup audaspace */
- */
 /** \defgroup audpython Audaspace Python
- *  \ingroup audaspace
+ *  \ingroup audaspace */
- */
 /** \defgroup audsdl Audaspace SDL
- *  \ingroup audaspace
+ *  \ingroup audaspace */
- */
 /** \defgroup audsrc Audaspace SRC
- *
+ *  \ingroup audaspace */
- *  \ingroup audaspace
- */
 /** \defgroup audffmpeg Audaspace FFMpeg
- *  \ingroup audaspace
+ *  \ingroup audaspace */
- */
 /** \defgroup audfftw Audaspace FFTW
- *  \ingroup audaspace
+ *  \ingroup audaspace */
- */
 /** \defgroup audjack Audaspace Jack
- *  \ingroup audaspace
+ *  \ingroup audaspace */
- */
 /** \defgroup audsndfile Audaspace sndfile
- *  \ingroup audaspace
+ *  \ingroup audaspace */
- */
 
 /** \defgroup GHOST GHOST API
  * \ingroup intern GUI

doc/doxygen/doxygen.source.h

@@ -5,7 +5,8 @@
 /** \defgroup bmesh BMesh
  *  \ingroup blender
  */
-/** \defgroup compositor Compositing */
+/** \defgroup compositor Compositing
+ *  \ingroup blender */
 
 /** \defgroup python Python
  *  \ingroup blender
@@ -78,7 +79,8 @@
  *  \ingroup blender
  */
 
-/** \defgroup data DNA, RNA and .blend access*/
+/** \defgroup data DNA, RNA and .blend access
+ *  \ingroup blender */
 
 /** \defgroup gpu GPU
  *  \ingroup blender
@@ -101,11 +103,12 @@
  *   merged in docs.
  */
 
-/** \defgroup gui GUI */
+/**
+ * \defgroup gui GUI
+ * \ingroup blender */
 
 /** \defgroup wm Window Manager
- *  \ingroup blender gui
+ *  \ingroup gui */
- */
 
 /* ================================ */
 
@@ -279,7 +282,8 @@
  *  \ingroup gui
  */
 
-/** \defgroup externformats External Formats */
+/** \defgroup externformats External Formats
+ *  \ingroup blender */
 
 /** \defgroup collada COLLADA
  *  \ingroup externformats
@@ -308,4 +312,7 @@
 /* ================================ */
 
 /** \defgroup undoc Undocumented
- *  \brief Modules and libraries that are still undocumented, or lacking proper integration into the doxygen system, are marked in this group. */
+ *
+ * \brief Modules and libraries that are still undocumented,
+ * or lacking proper integration into the doxygen system, are marked in this group.
+ */

doc/manpage/blender.1.py

@@ -61,7 +61,7 @@ def blender_extract_info(blender_bin: str) -> Dict[str, str]:
         stdout=subprocess.PIPE,
     ).stdout.decode(encoding="utf-8")
 
-    blender_version_ouput = subprocess.run(
+    blender_version_output = subprocess.run(
         [blender_bin, "--version"],
         env=blender_env,
         check=True,
@@ -73,7 +73,7 @@ def blender_extract_info(blender_bin: str) -> Dict[str, str]:
     # check for each lines prefix to ensure these aren't included.
     blender_version = ""
     blender_date = ""
-    for l in blender_version_ouput.split("\n"):
+    for l in blender_version_output.split("\n"):
         if l.startswith("Blender "):
             # Remove 'Blender' prefix.
             blender_version = l.split(" ", 1)[1].strip()

doc/python_api/sphinx_doc_gen.py

@@ -1103,6 +1103,7 @@ context_type_map = {
     "selectable_objects": ("Object", True),
     "selected_asset_files": ("FileSelectEntry", True),
     "selected_bones": ("EditBone", True),
+    "selected_editable_actions": ("Action", True),
     "selected_editable_bones": ("EditBone", True),
     "selected_editable_fcurves": ("FCurve", True),
     "selected_editable_keyframes": ("Keyframe", True),
@@ -1118,12 +1119,13 @@ context_type_map = {
     "selected_pose_bones": ("PoseBone", True),
     "selected_pose_bones_from_active_object": ("PoseBone", True),
     "selected_sequences": ("Sequence", True),
+    "selected_visible_actions": ("Action", True),
     "selected_visible_fcurves": ("FCurve", True),
     "sequences": ("Sequence", True),
     "soft_body": ("SoftBodyModifier", False),
     "speaker": ("Speaker", False),
     "texture": ("Texture", False),
-    "texture_slot": ("MaterialTextureSlot", False),
+    "texture_slot": ("TextureSlot", False),
     "texture_user": ("ID", False),
     "texture_user_property": ("Property", False),
     "ui_list": ("UIList", False),

extern/hipew/README

@@ -0,0 +1,12 @@
+The HIP Extension Wrangler Library (HIPEW) is a cross-platform open-source
+C/C++ library to dynamically load the HIP library.
+
+HIP (Heterogeneous-Compute Interface for Portability) is an API for C++
+programming on AMD GPUs.
+
+It is maintained as part of the Blender project, but included in extern/
+for consistency with CUEW and CLEW libraries.
+
+LICENSE
+
+HIPEW is released under the Apache 2.0 license.

extern/hipew/README.blender

@@ -0,0 +1,5 @@
+Project: Blender
+URL: https://git.blender.org/blender.git
+License: Apache 2.0
+Upstream version: N/A
+Local modifications: None

extern/hipew/src/hipew.c

@@ -219,17 +219,17 @@ static int hipewHasOldDriver(const char *hip_path) {
   DWORD verHandle = 0;
   DWORD verSize = GetFileVersionInfoSize(hip_path, &verHandle);
   int old_driver = 0;
-  if(verSize != 0) {
+  if (verSize != 0) {
     LPSTR verData = (LPSTR)malloc(verSize);
-    if(GetFileVersionInfo(hip_path, verHandle, verSize, verData)) {
+    if (GetFileVersionInfo(hip_path, verHandle, verSize, verData)) {
       LPBYTE lpBuffer = NULL;
       UINT size = 0;
-      if(VerQueryValue(verData, "\\", (VOID FAR * FAR *)&lpBuffer, &size)) {
+      if (VerQueryValue(verData, "\\", (VOID FAR * FAR *)&lpBuffer, &size)) {
-        if(size) {
+        if (size) {
           VS_FIXEDFILEINFO *verInfo = (VS_FIXEDFILEINFO *)lpBuffer;
           /* Magic value from
            * https://docs.microsoft.com/en-us/windows/win32/api/verrsrc/ns-verrsrc-vs_fixedfileinfo */
-          if(verInfo->dwSignature == 0xfeef04bd) {
+          if (verInfo->dwSignature == 0xfeef04bd) {
             unsigned int fileVersionLS0 = (verInfo->dwFileVersionLS >> 16) & 0xffff;
             unsigned int fileversionLS1 = (verInfo->dwFileVersionLS >> 0) & 0xffff;
             /* Corresponds to versions older than AMD Radeon Pro 21.Q4. */

extern/nanosvg/README.blender

@@ -1,7 +1,7 @@
 Project: NanoSVG
 URL: https://github.com/memononen/nanosvg
 License: zlib
-Upstream version: 
+Upstream version: 3cdd4a9d7886
 Local modifications: Added some functionality to manage grease pencil layers
 
 Added a fix to SVG import arc and float errors (https://developer.blender.org/rB11dc674c78b49fc4e0b7c134c375b6c8b8eacbcc)

intern/atomic/atomic_ops.h

@@ -45,7 +45,7 @@
  */
 
 /** \file
- * \ingroup Atomic
+ * \ingroup intern_atomic
  *
  * \brief Provides wrapper around system-specific atomic primitives,
  * and some extensions (faked-atomic operations over float numbers).

intern/atomic/intern/atomic_ops_ext.h

@@ -44,6 +44,10 @@
  * The Original Code is: adapted from jemalloc.
  */
 
+/** \file
+ * \ingroup intern_atomic
+ */
+
 #ifndef __ATOMIC_OPS_EXT_H__
 #define __ATOMIC_OPS_EXT_H__
 

intern/atomic/intern/atomic_ops_msvc.h

@@ -5,7 +5,7 @@
  * All rights reserved.
  * Copyright (C) 2007-2012 Mozilla Foundation.  All rights reserved.
  * Copyright (C) 2009-2013 Facebook, Inc.  All rights reserved.
-
+ *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are met:
  * 1. Redistributions of source code must retain the above copyright notice(s),
@@ -13,7 +13,7 @@
  * 2. Redistributions in binary form must reproduce the above copyright notice(s),
  *    this list of conditions and the following disclaimer in the documentation
  *    and/or other materials provided with the distribution.
-
+ *
  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY EXPRESS
  * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
  * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO
@@ -26,6 +26,10 @@
  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
 
+/** \file
+ * \ingroup intern_atomic
+ */
+
 #ifndef __ATOMIC_OPS_MSVC_H__
 #define __ATOMIC_OPS_MSVC_H__
 

intern/atomic/intern/atomic_ops_unix.h

@@ -44,6 +44,10 @@
  * The Original Code is: adapted from jemalloc.
  */
 
+/** \file
+ * \ingroup intern_atomic
+ */
+
 #ifndef __ATOMIC_OPS_UNIX_H__
 #define __ATOMIC_OPS_UNIX_H__
 

intern/atomic/intern/atomic_ops_utils.h

@@ -44,6 +44,10 @@
  * The Original Code is: adapted from jemalloc.
  */
 
+/** \file
+ * \ingroup intern_atomic
+ */
+
 #ifndef __ATOMIC_OPS_UTILS_H__
 #define __ATOMIC_OPS_UTILS_H__
 

intern/clog/CLG_log.h

@@ -14,11 +14,8 @@
  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
  */
 
-#ifndef __CLG_LOG_H__
-#define __CLG_LOG_H__
-
 /** \file
- * \ingroup clog
+ * \ingroup intern_clog
  *
  * C Logging Library (clog)
  * ========================
@@ -68,6 +65,9 @@
  * - 4+: May be used for more details than 3, should be avoided but not prevented.
  */
 
+#ifndef __CLG_LOG_H__
+#define __CLG_LOG_H__
+
 #ifdef __cplusplus
 extern "C" {
 #endif /* __cplusplus */

intern/clog/clog.c

@@ -15,7 +15,7 @@
  */
 
 /** \file
- * \ingroup clog
+ * \ingroup intern_clog
  */
 
 #include <assert.h>
@@ -388,7 +388,7 @@ static void clg_ctx_fatal_action(CLogContext *ctx)
 
 static void clg_ctx_backtrace(CLogContext *ctx)
 {
-  /* Note: we avoid writing to 'FILE', for back-trace we make an exception,
+  /* NOTE: we avoid writing to 'FILE', for back-trace we make an exception,
    * if necessary we could have a version of the callback that writes to file
    * descriptor all at once. */
   ctx->callbacks.backtrace_fn(ctx->output_file);

intern/cycles/blender/CMakeLists.txt

@@ -40,6 +40,7 @@ set(SRC
   object_cull.cpp
   output_driver.cpp
   particles.cpp
+  pointcloud.cpp
   curves.cpp
   logging.cpp
   python.cpp
@@ -87,6 +88,7 @@ endif()
 
 set(ADDON_FILES
   addon/__init__.py
+  addon/camera.py
   addon/engine.py
   addon/operators.py
   addon/osl.py
@@ -101,6 +103,11 @@ add_definitions(${GL_DEFINITIONS})
 if(WITH_CYCLES_DEVICE_HIP)
   add_definitions(-DWITH_HIP)
 endif()
+
+if(WITH_CYCLES_DEVICE_METAL)
+  add_definitions(-DWITH_METAL)
+endif()
+
 if(WITH_MOD_FLUID)
   add_definitions(-DWITH_FLUID)
 endif()

intern/cycles/blender/addon/camera.py

@@ -0,0 +1,84 @@
+#
+# Copyright 2011-2021 Blender Foundation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+# <pep8 compliant>
+
+# Fit to match default projective camera with focal_length 50 and sensor_width 36.
+default_fisheye_polynomial = [-1.1735143712967577e-05,
+                              -0.019988736953434998,
+                              -3.3525322965709175e-06,
+                              3.099275275886036e-06,
+                              -2.6064646454854524e-08]
+
+# Utilities to generate lens polynomials to match built-in camera types, only here
+# for reference at the moment, not used by the code.
+def create_grid(sensor_height, sensor_width):
+    import numpy as np
+    if sensor_height is None:
+        sensor_height = sensor_width / (16 / 9)  # Default aspect ration 16:9
+    uu, vv = np.meshgrid(np.linspace(0, 1, 100), np.linspace(0, 1, 100))
+    uu = (uu - 0.5) * sensor_width
+    vv = (vv - 0.5) * sensor_height
+    rr = np.sqrt(uu ** 2 + vv ** 2)
+    return rr
+
+
+def fisheye_lens_polynomial_from_projective(focal_length=50, sensor_width=36, sensor_height=None):
+    import numpy as np
+    rr = create_grid(sensor_height, sensor_width)
+    polynomial = np.polyfit(rr.flat, (-np.arctan(rr / focal_length)).flat, 4)
+    return list(reversed(polynomial))
+
+
+def fisheye_lens_polynomial_from_projective_fov(fov, sensor_width=36, sensor_height=None):
+    import numpy as np
+    f = sensor_width / 2 / np.tan(fov / 2)
+    return fisheye_lens_polynomial_from_projective(f, sensor_width, sensor_height)
+
+
+def fisheye_lens_polynomial_from_equisolid(lens=10.5, sensor_width=36, sensor_height=None):
+    import numpy as np
+    rr = create_grid(sensor_height, sensor_width)
+    x = rr.reshape(-1)
+    x = np.stack([x**i for i in [1, 2, 3, 4]])
+    y = (-2 * np.arcsin(rr / (2 * lens))).reshape(-1)
+    polynomial = np.linalg.lstsq(x.T, y.T, rcond=None)[0]
+    return [0] + list(polynomial)
+
+
+def fisheye_lens_polynomial_from_equidistant(fov=180, sensor_width=36, sensor_height=None):
+    import numpy as np
+    return [0, -np.radians(fov) / sensor_width, 0, 0, 0]
+
+
+def fisheye_lens_polynomial_from_distorted_projective_polynomial(k1, k2, k3, focal_length=50, sensor_width=36, sensor_height=None):
+    import numpy as np
+    rr = create_grid(sensor_height, sensor_width)
+    r2 = (rr / focal_length) ** 2
+    r4 = r2 * r2
+    r6 = r4 * r2
+    r_coeff = 1 + k1 * r2 + k2 * r4 + k3 * r6
+    polynomial = np.polyfit(rr.flat, (-np.arctan(rr / focal_length * r_coeff)).flat, 4)
+    return list(reversed(polynomial))
+
+def fisheye_lens_polynomial_from_distorted_projective_divisions(k1, k2, focal_length=50, sensor_width=36, sensor_height=None):
+    import numpy as np
+    rr = create_grid(sensor_height, sensor_width)
+    r2 = (rr / focal_length) ** 2
+    r4 = r2 * r2
+    r_coeff = 1 + k1 * r2 + k2 * r4
+    polynomial = np.polyfit(rr.flat, (-np.arctan(rr / focal_length / r_coeff)).flat, 4)
+    return list(reversed(polynomial))

intern/cycles/blender/addon/engine.py

@@ -28,7 +28,7 @@ def _configure_argument_parser():
                         action='store_true')
     parser.add_argument("--cycles-device",
                         help="Set the device to use for Cycles, overriding user preferences and the scene setting."
-                             "Valid options are 'CPU', 'CUDA', 'OPTIX', or 'HIP'"
+                             "Valid options are 'CPU', 'CUDA', 'OPTIX', 'HIP' or 'METAL'."
                              "Additionally, you can append '+CPU' to any GPU type for hybrid rendering.",
                         default=None)
     return parser

intern/cycles/blender/addon/properties.py

@@ -33,6 +33,7 @@ from math import pi
 # enums
 
 from . import engine
+from . import camera
 
 enum_devices = (
     ('CPU', "CPU", "Use CPU for rendering"),
@@ -72,6 +73,8 @@ enum_panorama_types = (
     ('FISHEYE_EQUISOLID', "Fisheye Equisolid",
                           "Similar to most fisheye modern lens, takes sensor dimensions into consideration"),
     ('MIRRORBALL', "Mirror Ball", "Uses the mirror ball mapping"),
+    ('FISHEYE_LENS_POLYNOMIAL', "Fisheye Lens Polynomial",
+     "Defines the lens projection as polynomial to allow real world camera lenses to be mimicked."),
 )
 
 enum_curve_shape = (
@@ -111,7 +114,8 @@ enum_device_type = (
     ('CPU', "CPU", "CPU", 0),
     ('CUDA', "CUDA", "CUDA", 1),
     ('OPTIX', "OptiX", "OptiX", 3),
-    ("HIP", "HIP", "HIP", 4)
+    ('HIP', "HIP", "HIP", 4),
+    ('METAL', "Metal", "Metal", 5)
 )
 
 enum_texture_limit = (
@@ -429,7 +433,7 @@ class CyclesRenderSettings(bpy.types.PropertyGroup):
     )
 
     direct_light_sampling_type: EnumProperty(
-        name="Direct Light Sampling Type",
+        name="Direct Light Sampling",
         description="The type of strategy used for sampling direct light contributions",
         items=enum_direct_light_sampling_type,
         default='MULTIPLE_IMPORTANCE_SAMPLING',
@@ -790,7 +794,7 @@ class CyclesRenderSettings(bpy.types.PropertyGroup):
     )
 
     use_auto_tile: BoolProperty(
-        name="Using Tiling",
+        name="Use Tiling",
         description="Render high resolution images in tiles to reduce memory usage, using the specified tile size. Tiles are cached to disk while rendering to save memory",
         default=True,
     )
@@ -890,6 +894,32 @@ class CyclesCameraSettings(bpy.types.PropertyGroup):
         default=pi,
     )
 
+    fisheye_polynomial_k0: FloatProperty(
+        name="Fisheye Polynomial K0",
+        description="Coefficient K0 of the lens polinomial",
+        default=camera.default_fisheye_polynomial[0], precision=6, step=0.1, subtype='ANGLE',
+    )
+    fisheye_polynomial_k1: FloatProperty(
+        name="Fisheye Polynomial K1",
+        description="Coefficient K1 of the lens polinomial",
+        default=camera.default_fisheye_polynomial[1], precision=6, step=0.1, subtype='ANGLE',
+    )
+    fisheye_polynomial_k2: FloatProperty(
+        name="Fisheye Polynomial K2",
+        description="Coefficient K2 of the lens polinomial",
+        default=camera.default_fisheye_polynomial[2], precision=6, step=0.1, subtype='ANGLE',
+    )
+    fisheye_polynomial_k3: FloatProperty(
+        name="Fisheye Polynomial K3",
+        description="Coefficient K3 of the lens polinomial",
+        default=camera.default_fisheye_polynomial[3], precision=6, step=0.1, subtype='ANGLE',
+    )
+    fisheye_polynomial_k4: FloatProperty(
+        name="Fisheye Polynomial K4",
+        description="Coefficient K4 of the lens polinomial",
+        default=camera.default_fisheye_polynomial[4], precision=6, step=0.1, subtype='ANGLE',
+    )
+
     @classmethod
     def register(cls):
         bpy.types.Camera.cycles = PointerProperty(
@@ -1312,8 +1342,7 @@ class CyclesPreferences(bpy.types.AddonPreferences):
 
     def get_device_types(self, context):
         import _cycles
-        has_cuda, has_optix, has_hip = _cycles.get_device_types()
+        has_cuda, has_optix, has_hip, has_metal = _cycles.get_device_types()
-
         list = [('NONE', "None", "Don't use compute device", 0)]
         if has_cuda:
             list.append(('CUDA', "CUDA", "Use CUDA for GPU acceleration", 1))
@@ -1321,6 +1350,8 @@ class CyclesPreferences(bpy.types.AddonPreferences):
             list.append(('OPTIX', "OptiX", "Use OptiX for GPU acceleration", 3))
         if has_hip:
             list.append(('HIP', "HIP", "Use HIP for GPU acceleration", 4))
+        if has_metal:
+            list.append(('METAL', "Metal", "Use Metal for GPU acceleration", 5))
 
         return list
 
@@ -1346,7 +1377,7 @@ class CyclesPreferences(bpy.types.AddonPreferences):
 
     def update_device_entries(self, device_list):
         for device in device_list:
-            if not device[1] in {'CUDA', 'OPTIX', 'CPU', 'HIP'}:
+            if not device[1] in {'CUDA', 'OPTIX', 'CPU', 'HIP', 'METAL'}:
                 continue
             # Try to find existing Device entry
             entry = self.find_existing_device_entry(device)
@@ -1390,7 +1421,7 @@ class CyclesPreferences(bpy.types.AddonPreferences):
         import _cycles
         # Ensure `self.devices` is not re-allocated when the second call to
         # get_devices_for_type is made, freeing items from the first list.
-        for device_type in ('CUDA', 'OPTIX', 'HIP'):
+        for device_type in ('CUDA', 'OPTIX', 'HIP', 'METAL'):
             self.update_device_entries(_cycles.available_devices(device_type))
 
     # Deprecated: use refresh_devices instead.
@@ -1442,6 +1473,8 @@ class CyclesPreferences(bpy.types.AddonPreferences):
                 col.label(text="Requires discrete AMD GPU with RDNA architecture", icon='BLANK1')
                 if sys.platform[:3] == "win":
                     col.label(text="and AMD Radeon Pro 21.Q4 driver or newer", icon='BLANK1')
+            elif device_type == 'METAL':
+                col.label(text="Requires Apple Silicon and macOS 12.0 or newer", icon='BLANK1')
             return
 
         for device in devices:

intern/cycles/blender/addon/ui.py

@@ -97,6 +97,11 @@ def use_cpu(context):
     return (get_device_type(context) == 'NONE' or cscene.device == 'CPU')
 
 
+def use_metal(context):
+    cscene = context.scene.cycles
+
+    return (get_device_type(context) == 'METAL' and cscene.device == 'GPU')
+
 def use_cuda(context):
     cscene = context.scene.cycles
 
@@ -1015,7 +1020,7 @@ class CYCLES_OBJECT_PT_motion_blur(CyclesButtonsPanel, Panel):
     def poll(cls, context):
         ob = context.object
         if CyclesButtonsPanel.poll(context) and ob:
-            if ob.type in {'MESH', 'CURVE', 'CURVE', 'SURFACE', 'FONT', 'META', 'CAMERA'}:
+            if ob.type in {'MESH', 'CURVE', 'CURVE', 'SURFACE', 'FONT', 'META', 'CAMERA', 'HAIR', 'POINTCLOUD'}:
                 return True
             if ob.instance_type == 'COLLECTION' and ob.instance_collection:
                 return True
@@ -1819,37 +1824,38 @@ class CYCLES_RENDER_PT_debug(CyclesDebugButtonsPanel, Panel):
 
     def draw(self, context):
         layout = self.layout
+        layout.use_property_split = True
+        layout.use_property_decorate = False  # No animation.
 
         scene = context.scene
         cscene = scene.cycles
 
-        col = layout.column()
+        col = layout.column(heading="CPU")
 
-        col.label(text="CPU Flags:")
         row = col.row(align=True)
         row.prop(cscene, "debug_use_cpu_sse2", toggle=True)
         row.prop(cscene, "debug_use_cpu_sse3", toggle=True)
         row.prop(cscene, "debug_use_cpu_sse41", toggle=True)
         row.prop(cscene, "debug_use_cpu_avx", toggle=True)
         row.prop(cscene, "debug_use_cpu_avx2", toggle=True)
-        col.prop(cscene, "debug_bvh_layout")
+        col.prop(cscene, "debug_bvh_layout", text="BVH")
 
         col.separator()
 
-        col = layout.column()
+        col = layout.column(heading="CUDA")
-        col.label(text="CUDA Flags:")
         col.prop(cscene, "debug_use_cuda_adaptive_compile")
+        col = layout.column(heading="OptiX")
+        col.prop(cscene, "debug_use_optix_debug", text="Module Debug")
 
         col.separator()
 
-        col = layout.column()
+        col.prop(cscene, "debug_bvh_type", text="Viewport BVH")
-        col.label(text="OptiX Flags:")
-        col.prop(cscene, "debug_use_optix_debug")
 
         col.separator()
 
-        col = layout.column()
+        import _cycles
-        col.prop(cscene, "debug_bvh_type")
+        if _cycles.with_debug:
+            col.prop(cscene, "direct_light_sampling_type")
 
 
 class CYCLES_RENDER_PT_simplify(CyclesButtonsPanel, Panel):

intern/cycles/blender/camera.cpp

@@ -69,6 +69,12 @@ struct BlenderCamera {
   float pole_merge_angle_from;
   float pole_merge_angle_to;
 
+  float fisheye_polynomial_k0;
+  float fisheye_polynomial_k1;
+  float fisheye_polynomial_k2;
+  float fisheye_polynomial_k3;
+  float fisheye_polynomial_k4;
+
   enum { AUTO, HORIZONTAL, VERTICAL } sensor_fit;
   float sensor_width;
   float sensor_height;
@@ -200,6 +206,12 @@ static void blender_camera_from_object(BlenderCamera *bcam,
     bcam->longitude_min = RNA_float_get(&ccamera, "longitude_min");
     bcam->longitude_max = RNA_float_get(&ccamera, "longitude_max");
 
+    bcam->fisheye_polynomial_k0 = RNA_float_get(&ccamera, "fisheye_polynomial_k0");
+    bcam->fisheye_polynomial_k1 = RNA_float_get(&ccamera, "fisheye_polynomial_k1");
+    bcam->fisheye_polynomial_k2 = RNA_float_get(&ccamera, "fisheye_polynomial_k2");
+    bcam->fisheye_polynomial_k3 = RNA_float_get(&ccamera, "fisheye_polynomial_k3");
+    bcam->fisheye_polynomial_k4 = RNA_float_get(&ccamera, "fisheye_polynomial_k4");
+
     bcam->interocular_distance = b_camera.stereo().interocular_distance();
     if (b_camera.stereo().convergence_mode() == BL::CameraStereoData::convergence_mode_PARALLEL) {
       bcam->convergence_distance = FLT_MAX;
@@ -422,7 +434,8 @@ static void blender_camera_sync(Camera *cam,
   cam->set_full_height(height);
 
   /* panorama sensor */
-  if (bcam->type == CAMERA_PANORAMA && bcam->panorama_type == PANORAMA_FISHEYE_EQUISOLID) {
+  if (bcam->type == CAMERA_PANORAMA && (bcam->panorama_type == PANORAMA_FISHEYE_EQUISOLID ||
+                                        bcam->panorama_type == PANORAMA_FISHEYE_LENS_POLYNOMIAL)) {
     float fit_xratio = (float)bcam->render_width * bcam->pixelaspect.x;
     float fit_yratio = (float)bcam->render_height * bcam->pixelaspect.y;
     bool horizontal_fit;
@@ -465,6 +478,12 @@ static void blender_camera_sync(Camera *cam,
   cam->set_latitude_min(bcam->latitude_min);
   cam->set_latitude_max(bcam->latitude_max);
 
+  cam->set_fisheye_polynomial_k0(bcam->fisheye_polynomial_k0);
+  cam->set_fisheye_polynomial_k1(bcam->fisheye_polynomial_k1);
+  cam->set_fisheye_polynomial_k2(bcam->fisheye_polynomial_k2);
+  cam->set_fisheye_polynomial_k3(bcam->fisheye_polynomial_k3);
+  cam->set_fisheye_polynomial_k4(bcam->fisheye_polynomial_k4);
+
   cam->set_longitude_min(bcam->longitude_min);
   cam->set_longitude_max(bcam->longitude_max);
 

intern/cycles/blender/curves.cpp

@@ -819,11 +819,14 @@ void BlenderSync::sync_hair(BL::Depsgraph b_depsgraph, BObjectInfo &b_ob_info, H
   new_hair.set_used_shaders(used_shaders);
 
   if (view_layer.use_hair) {
+#ifdef WITH_HAIR_NODES
     if (b_ob_info.object_data.is_a(&RNA_Hair)) {
       /* Hair object. */
       sync_hair(&new_hair, b_ob_info, false);
     }
-    else {
+    else
+#endif
+    {
       /* Particle hair. */
       bool need_undeformed = new_hair.need_attribute(scene, ATTR_STD_GENERATED);
       BL::Mesh b_mesh = object_to_mesh(
@@ -870,12 +873,15 @@ void BlenderSync::sync_hair_motion(BL::Depsgraph b_depsgraph,
 
   /* Export deformed coordinates. */
   if (ccl::BKE_object_is_deform_modified(b_ob_info, b_scene, preview)) {
+#ifdef WITH_HAIR_NODES
     if (b_ob_info.object_data.is_a(&RNA_Hair)) {
       /* Hair object. */
       sync_hair(hair, b_ob_info, true, motion_step);
       return;
     }
-    else {
+    else
+#endif
+    {
       /* Particle hair. */
       BL::Mesh b_mesh = object_to_mesh(
           b_data, b_ob_info, b_depsgraph, false, Mesh::SUBDIVISION_NONE);

intern/cycles/blender/device.cpp

@@ -27,6 +27,7 @@ enum ComputeDevice {
   COMPUTE_DEVICE_CUDA = 1,
   COMPUTE_DEVICE_OPTIX = 3,
   COMPUTE_DEVICE_HIP = 4,
+  COMPUTE_DEVICE_METAL = 5,
 
   COMPUTE_DEVICE_NUM
 };
@@ -85,6 +86,9 @@ DeviceInfo blender_device_info(BL::Preferences &b_preferences, BL::Scene &b_scen
       else if (compute_device == COMPUTE_DEVICE_HIP) {
         mask |= DEVICE_MASK_HIP;
       }
+      else if (compute_device == COMPUTE_DEVICE_METAL) {
+        mask |= DEVICE_MASK_METAL;
+      }
       vector<DeviceInfo> devices = Device::available_devices(mask);
 
       /* Match device preferences and available devices. */

intern/cycles/blender/geometry.cpp

@@ -19,6 +19,7 @@
 #include "scene/hair.h"
 #include "scene/mesh.h"
 #include "scene/object.h"
+#include "scene/pointcloud.h"
 #include "scene/volume.h"
 
 #include "blender/sync.h"
@@ -31,10 +32,18 @@ CCL_NAMESPACE_BEGIN
 
 static Geometry::Type determine_geom_type(BObjectInfo &b_ob_info, bool use_particle_hair)
 {
+#ifdef WITH_HAIR_NODES
   if (b_ob_info.object_data.is_a(&RNA_Hair) || use_particle_hair) {
+#else
+  if (use_particle_hair) {
+#endif
     return Geometry::HAIR;
   }
 
+  if (b_ob_info.object_data.is_a(&RNA_PointCloud)) {
+    return Geometry::POINTCLOUD;
+  }
+
   if (b_ob_info.object_data.is_a(&RNA_Volume) ||
       (b_ob_info.object_data == b_ob_info.real_object.data() &&
        object_fluid_gas_domain_find(b_ob_info.real_object))) {
@@ -107,6 +116,9 @@ Geometry *BlenderSync::sync_geometry(BL::Depsgraph &b_depsgraph,
     else if (geom_type == Geometry::VOLUME) {
       geom = scene->create_node<Volume>();
     }
+    else if (geom_type == Geometry::POINTCLOUD) {
+      geom = scene->create_node<PointCloud>();
+    }
     else {
       geom = scene->create_node<Mesh>();
     }
@@ -166,6 +178,10 @@ Geometry *BlenderSync::sync_geometry(BL::Depsgraph &b_depsgraph,
       Volume *volume = static_cast<Volume *>(geom);
       sync_volume(b_ob_info, volume);
     }
+    else if (geom_type == Geometry::POINTCLOUD) {
+      PointCloud *pointcloud = static_cast<PointCloud *>(geom);
+      sync_pointcloud(pointcloud, b_ob_info);
+    }
     else {
       Mesh *mesh = static_cast<Mesh *>(geom);
       sync_mesh(b_depsgraph, b_ob_info, mesh);
@@ -215,7 +231,11 @@ void BlenderSync::sync_geometry_motion(BL::Depsgraph &b_depsgraph,
     if (progress.get_cancel())
       return;
 
+#ifdef WITH_HAIR_NODES
     if (b_ob_info.object_data.is_a(&RNA_Hair) || use_particle_hair) {
+#else
+    if (use_particle_hair) {
+#endif
       Hair *hair = static_cast<Hair *>(geom);
       sync_hair_motion(b_depsgraph, b_ob_info, hair, motion_step);
     }
@@ -223,6 +243,10 @@ void BlenderSync::sync_geometry_motion(BL::Depsgraph &b_depsgraph,
              object_fluid_gas_domain_find(b_ob_info.real_object)) {
       /* No volume motion blur support yet. */
     }
+    else if (b_ob_info.object_data.is_a(&RNA_PointCloud)) {
+      PointCloud *pointcloud = static_cast<PointCloud *>(geom);
+      sync_pointcloud_motion(pointcloud, b_ob_info, motion_step);
+    }
     else {
       Mesh *mesh = static_cast<Mesh *>(geom);
       sync_mesh_motion(b_depsgraph, b_ob_info, mesh, motion_step);

intern/cycles/blender/image.cpp

@@ -24,8 +24,14 @@ CCL_NAMESPACE_BEGIN
 
 /* Packed Images */
 
-BlenderImageLoader::BlenderImageLoader(BL::Image b_image, int frame)
+BlenderImageLoader::BlenderImageLoader(BL::Image b_image,
-    : b_image(b_image), frame(frame), free_cache(!b_image.has_data())
+                                       const int frame,
+                                       const bool is_preview_render)
+    : b_image(b_image),
+      frame(frame),
+      /* Don't free cache for preview render to avoid race condition from T93560, to be fixed
+         properly later as we are close to release. */
+      free_cache(!is_preview_render && !b_image.has_data())
 {
 }
 

intern/cycles/blender/image.h

@@ -25,7 +25,7 @@ CCL_NAMESPACE_BEGIN
 
 class BlenderImageLoader : public ImageLoader {
  public:
-  BlenderImageLoader(BL::Image b_image, int frame);
+  BlenderImageLoader(BL::Image b_image, const int frame, const bool is_preview_render);
 
   bool load_metadata(const ImageDeviceFeatures &features, ImageMetaData &metadata) override;
   bool load_pixels(const ImageMetaData &metadata,

intern/cycles/blender/object.cpp

@@ -72,7 +72,8 @@ bool BlenderSync::object_is_geometry(BObjectInfo &b_ob_info)
 
   BL::Object::type_enum type = b_ob_info.iter_object.type();
 
-  if (type == BL::Object::type_VOLUME || type == BL::Object::type_HAIR) {
+  if (type == BL::Object::type_VOLUME || type == BL::Object::type_HAIR ||
+      type == BL::Object::type_POINTCLOUD) {
     /* Will be exported attached to mesh. */
     return true;
   }
@@ -206,7 +207,7 @@ Object *BlenderSync::sync_object(BL::Depsgraph &b_depsgraph,
     return NULL;
   }
 
-  /* only interested in object that we can create meshes from */
+  /* only interested in object that we can create geometry from */
   if (!object_is_geometry(b_ob_info)) {
     return NULL;
   }

intern/cycles/blender/output_driver.cpp

@@ -66,7 +66,7 @@ bool BlenderOutputDriver::read_render_tile(const Tile &tile)
 
 bool BlenderOutputDriver::update_render_tile(const Tile &tile)
 {
-  /* Use final write for preview renders, otherwise render result wouldn't be be updated
+  /* Use final write for preview renders, otherwise render result wouldn't be updated
    * quickly on Blender side. For all other cases we use the display driver. */
   if (b_engine_.is_preview()) {
     write_render_tile(tile);

intern/cycles/blender/pointcloud.cpp

@@ -0,0 +1,253 @@
+/*
+ * Copyright 2011-2013 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "scene/pointcloud.h"
+#include "scene/attribute.h"
+#include "scene/scene.h"
+
+#include "blender/sync.h"
+#include "blender/util.h"
+
+#include "util/foreach.h"
+#include "util/hash.h"
+
+CCL_NAMESPACE_BEGIN
+
+template<typename TypeInCycles, typename GetValueAtIndex>
+static void fill_generic_attribute(BL::PointCloud &b_pointcloud,
+                                   TypeInCycles *data,
+                                   const GetValueAtIndex &get_value_at_index)
+{
+  const int num_points = b_pointcloud.points.length();
+  for (int i = 0; i < num_points; i++) {
+    data[i] = get_value_at_index(i);
+  }
+}
+
+static void copy_attributes(PointCloud *pointcloud, BL::PointCloud b_pointcloud)
+{
+  AttributeSet &attributes = pointcloud->attributes;
+  for (BL::Attribute &b_attribute : b_pointcloud.attributes) {
+    const ustring name{b_attribute.name().c_str()};
+
+    if (attributes.find(name)) {
+      continue;
+    }
+
+    const AttributeElement element = ATTR_ELEMENT_VERTEX;
+    const BL::Attribute::data_type_enum b_data_type = b_attribute.data_type();
+    switch (b_data_type) {
+      case BL::Attribute::data_type_FLOAT: {
+        BL::FloatAttribute b_float_attribute{b_attribute};
+        Attribute *attr = attributes.add(name, TypeFloat, element);
+        float *data = attr->data_float();
+        fill_generic_attribute(
+            b_pointcloud, data, [&](int i) { return b_float_attribute.data[i].value(); });
+        break;
+      }
+      case BL::Attribute::data_type_BOOLEAN: {
+        BL::BoolAttribute b_bool_attribute{b_attribute};
+        Attribute *attr = attributes.add(name, TypeFloat, element);
+        float *data = attr->data_float();
+        fill_generic_attribute(
+            b_pointcloud, data, [&](int i) { return (float)b_bool_attribute.data[i].value(); });
+        break;
+      }
+      case BL::Attribute::data_type_INT: {
+        BL::IntAttribute b_int_attribute{b_attribute};
+        Attribute *attr = attributes.add(name, TypeFloat, element);
+        float *data = attr->data_float();
+        fill_generic_attribute(
+            b_pointcloud, data, [&](int i) { return (float)b_int_attribute.data[i].value(); });
+        break;
+      }
+      case BL::Attribute::data_type_FLOAT_VECTOR: {
+        BL::FloatVectorAttribute b_vector_attribute{b_attribute};
+        Attribute *attr = attributes.add(name, TypeVector, element);
+        float3 *data = attr->data_float3();
+        fill_generic_attribute(b_pointcloud, data, [&](int i) {
+          BL::Array<float, 3> v = b_vector_attribute.data[i].vector();
+          return make_float3(v[0], v[1], v[2]);
+        });
+        break;
+      }
+      case BL::Attribute::data_type_FLOAT_COLOR: {
+        BL::FloatColorAttribute b_color_attribute{b_attribute};
+        Attribute *attr = attributes.add(name, TypeRGBA, element);
+        float4 *data = attr->data_float4();
+        fill_generic_attribute(b_pointcloud, data, [&](int i) {
+          BL::Array<float, 4> v = b_color_attribute.data[i].color();
+          return make_float4(v[0], v[1], v[2], v[3]);
+        });
+        break;
+      }
+      case BL::Attribute::data_type_FLOAT2: {
+        BL::Float2Attribute b_float2_attribute{b_attribute};
+        Attribute *attr = attributes.add(name, TypeFloat2, element);
+        float2 *data = attr->data_float2();
+        fill_generic_attribute(b_pointcloud, data, [&](int i) {
+          BL::Array<float, 2> v = b_float2_attribute.data[i].vector();
+          return make_float2(v[0], v[1]);
+        });
+        break;
+      }
+      default:
+        /* Not supported. */
+        break;
+    }
+  }
+}
+
+static void export_pointcloud(Scene *scene, PointCloud *pointcloud, BL::PointCloud b_pointcloud)
+{
+  /* TODO: optimize so we can straight memcpy arrays from Blender? */
+
+  /* Add requested attributes. */
+  Attribute *attr_random = NULL;
+  if (pointcloud->need_attribute(scene, ATTR_STD_POINT_RANDOM)) {
+    attr_random = pointcloud->attributes.add(ATTR_STD_POINT_RANDOM);
+  }
+
+  /* Reserve memory. */
+  const int num_points = b_pointcloud.points.length();
+  pointcloud->reserve(num_points);
+
+  /* Export points. */
+  BL::PointCloud::points_iterator b_point_iter;
+  for (b_pointcloud.points.begin(b_point_iter); b_point_iter != b_pointcloud.points.end();
+       ++b_point_iter) {
+    BL::Point b_point = *b_point_iter;
+    const float3 co = get_float3(b_point.co());
+    const float radius = b_point.radius();
+    pointcloud->add_point(co, radius);
+
+    /* Random number per point. */
+    if (attr_random != NULL) {
+      attr_random->add(hash_uint2_to_float(b_point.index(), 0));
+    }
+  }
+
+  /* Export attributes */
+  copy_attributes(pointcloud, b_pointcloud);
+}
+
+static void export_pointcloud_motion(PointCloud *pointcloud,
+                                     BL::PointCloud b_pointcloud,
+                                     int motion_step)
+{
+  /* Find or add attribute. */
+  Attribute *attr_mP = pointcloud->attributes.find(ATTR_STD_MOTION_VERTEX_POSITION);
+  bool new_attribute = false;
+
+  if (!attr_mP) {
+    attr_mP = pointcloud->attributes.add(ATTR_STD_MOTION_VERTEX_POSITION);
+    new_attribute = true;
+  }
+
+  /* Export motion points. */
+  const int num_points = pointcloud->num_points();
+  float3 *mP = attr_mP->data_float3() + motion_step * num_points;
+  bool have_motion = false;
+  int num_motion_points = 0;
+  const array<float3> &pointcloud_points = pointcloud->get_points();
+
+  BL::PointCloud::points_iterator b_point_iter;
+  for (b_pointcloud.points.begin(b_point_iter); b_point_iter != b_pointcloud.points.end();
+       ++b_point_iter) {
+    BL::Point b_point = *b_point_iter;
+
+    if (num_motion_points < num_points) {
+      float3 P = get_float3(b_point.co());
+      P.w = b_point.radius();
+      mP[num_motion_points] = P;
+      have_motion = have_motion || (P != pointcloud_points[num_motion_points]);
+      num_motion_points++;
+    }
+  }
+
+  /* In case of new attribute, we verify if there really was any motion. */
+  if (new_attribute) {
+    if (num_motion_points != num_points || !have_motion) {
+      pointcloud->attributes.remove(ATTR_STD_MOTION_VERTEX_POSITION);
+    }
+    else if (motion_step > 0) {
+      /* Motion, fill up previous steps that we might have skipped because
+       * they had no motion, but we need them anyway now. */
+      for (int step = 0; step < motion_step; step++) {
+        pointcloud->copy_center_to_motion_step(step);
+      }
+    }
+  }
+
+  /* Export attributes */
+  copy_attributes(pointcloud, b_pointcloud);
+}
+
+void BlenderSync::sync_pointcloud(PointCloud *pointcloud, BObjectInfo &b_ob_info)
+{
+  size_t old_numpoints = pointcloud->num_points();
+
+  array<Node *> used_shaders = pointcloud->get_used_shaders();
+
+  PointCloud new_pointcloud;
+  new_pointcloud.set_used_shaders(used_shaders);
+
+  /* TODO: add option to filter out points in the view layer. */
+  BL::PointCloud b_pointcloud(b_ob_info.object_data);
+  export_pointcloud(scene, &new_pointcloud, b_pointcloud);
+
+  /* update original sockets */
+  for (const SocketType &socket : new_pointcloud.type->inputs) {
+    /* Those sockets are updated in sync_object, so do not modify them. */
+    if (socket.name == "use_motion_blur" || socket.name == "motion_steps" ||
+        socket.name == "used_shaders") {
+      continue;
+    }
+    pointcloud->set_value(socket, new_pointcloud, socket);
+  }
+
+  pointcloud->attributes.clear();
+  foreach (Attribute &attr, new_pointcloud.attributes.attributes) {
+    pointcloud->attributes.attributes.push_back(std::move(attr));
+  }
+
+  /* tag update */
+  const bool rebuild = (pointcloud && old_numpoints != pointcloud->num_points());
+  pointcloud->tag_update(scene, rebuild);
+}
+
+void BlenderSync::sync_pointcloud_motion(PointCloud *pointcloud,
+                                         BObjectInfo &b_ob_info,
+                                         int motion_step)
+{
+  /* Skip if nothing exported. */
+  if (pointcloud->num_points() == 0) {
+    return;
+  }
+
+  /* Export deformed coordinates. */
+  if (ccl::BKE_object_is_deform_modified(b_ob_info, b_scene, preview)) {
+    /* PointCloud object. */
+    BL::PointCloud b_pointcloud(b_ob_info.object_data);
+    export_pointcloud_motion(pointcloud, b_pointcloud, motion_step);
+  }
+  else {
+    /* No deformation on this frame, copy coordinates if other frames did have it. */
+    pointcloud->copy_center_to_motion_step(motion_step);
+  }
+}
+
+CCL_NAMESPACE_END

intern/cycles/blender/python.cpp

@@ -906,16 +906,18 @@ static PyObject *enable_print_stats_func(PyObject * /*self*/, PyObject * /*args*
 static PyObject *get_device_types_func(PyObject * /*self*/, PyObject * /*args*/)
 {
   vector<DeviceType> device_types = Device::available_types();
-  bool has_cuda = false, has_optix = false, has_hip = false;
+  bool has_cuda = false, has_optix = false, has_hip = false, has_metal = false;
   foreach (DeviceType device_type, device_types) {
     has_cuda |= (device_type == DEVICE_CUDA);
     has_optix |= (device_type == DEVICE_OPTIX);
     has_hip |= (device_type == DEVICE_HIP);
+    has_metal |= (device_type == DEVICE_METAL);
   }
-  PyObject *list = PyTuple_New(3);
+  PyObject *list = PyTuple_New(4);
   PyTuple_SET_ITEM(list, 0, PyBool_FromLong(has_cuda));
   PyTuple_SET_ITEM(list, 1, PyBool_FromLong(has_optix));
   PyTuple_SET_ITEM(list, 2, PyBool_FromLong(has_hip));
+  PyTuple_SET_ITEM(list, 3, PyBool_FromLong(has_metal));
   return list;
 }
 
@@ -944,6 +946,9 @@ static PyObject *set_device_override_func(PyObject * /*self*/, PyObject *arg)
   else if (override == "HIP") {
     BlenderSession::device_override = DEVICE_MASK_HIP;
   }
+  else if (override == "METAL") {
+    BlenderSession::device_override = DEVICE_MASK_METAL;
+  }
   else {
     printf("\nError: %s is not a valid Cycles device.\n", override.c_str());
     Py_RETURN_FALSE;
@@ -1054,5 +1059,13 @@ void *CCL_python_module_init()
     Py_INCREF(Py_False);
   }
 
+#ifdef WITH_CYCLES_DEBUG
+  PyModule_AddObject(mod, "with_debug", Py_True);
+  Py_INCREF(Py_True);
+#else  /* WITH_CYCLES_DEBUG */
+  PyModule_AddObject(mod, "with_debug", Py_False);
+  Py_INCREF(Py_False);
+#endif /* WITH_CYCLES_DEBUG */
+
   return (void *)mod;
 }

intern/cycles/blender/session.cpp

@@ -396,6 +396,13 @@ void BlenderSession::render(BL::Depsgraph &b_depsgraph_)
     /* set the current view */
     b_engine.active_view_set(b_rview_name.c_str());
 
+    /* Force update in this case, since the camera transform on each frame changes
+     * in different views. This could be optimized by somehow storing the animated
+     * camera transforms separate from the fixed stereo transform. */
+    if ((scene->need_motion() != Scene::MOTION_NONE) && view_index > 0) {
+      sync->tag_update();
+    }
+
     /* update scene */
     BL::Object b_camera_override(b_engine.camera_override());
     sync->sync_camera(b_render, b_camera_override, width, height, b_rview_name.c_str());
@@ -629,7 +636,7 @@ void BlenderSession::bake(BL::Depsgraph &b_depsgraph_,
       integrator->set_use_emission((bake_filter & BL::BakeSettings::pass_filter_EMIT) != 0);
     }
 
-    /* Always use transpanent background for baking. */
+    /* Always use transparent background for baking. */
     scene->background->set_transparent(true);
 
     /* Load built-in images from Blender. */

intern/cycles/blender/shader.cpp

@@ -378,11 +378,20 @@ static ShaderNode *add_node(Scene *scene,
   }
   else if (b_node.is_a(&RNA_ShaderNodeMapRange)) {
     BL::ShaderNodeMapRange b_map_range_node(b_node);
+    if (b_map_range_node.data_type() == BL::ShaderNodeMapRange::data_type_FLOAT_VECTOR) {
+      VectorMapRangeNode *vector_map_range_node = graph->create_node<VectorMapRangeNode>();
+      vector_map_range_node->set_use_clamp(b_map_range_node.clamp());
+      vector_map_range_node->set_range_type(
+          (NodeMapRangeType)b_map_range_node.interpolation_type());
+      node = vector_map_range_node;
+    }
+    else {
       MapRangeNode *map_range_node = graph->create_node<MapRangeNode>();
       map_range_node->set_clamp(b_map_range_node.clamp());
       map_range_node->set_range_type((NodeMapRangeType)b_map_range_node.interpolation_type());
       node = map_range_node;
     }
+  }
   else if (b_node.is_a(&RNA_ShaderNodeClamp)) {
     BL::ShaderNodeClamp b_clamp_node(b_node);
     ClampNode *clamp_node = graph->create_node<ClampNode>();
@@ -762,7 +771,8 @@ static ShaderNode *add_node(Scene *scene,
         int scene_frame = b_scene.frame_current();
         int image_frame = image_user_frame_number(b_image_user, b_image, scene_frame);
         image->handle = scene->image_manager->add_image(
-            new BlenderImageLoader(b_image, image_frame), image->image_params());
+            new BlenderImageLoader(b_image, image_frame, b_engine.is_preview()),
+            image->image_params());
       }
       else {
         ustring filename = ustring(
@@ -797,7 +807,8 @@ static ShaderNode *add_node(Scene *scene,
       if (is_builtin) {
         int scene_frame = b_scene.frame_current();
         int image_frame = image_user_frame_number(b_image_user, b_image, scene_frame);
-        env->handle = scene->image_manager->add_image(new BlenderImageLoader(b_image, image_frame),
+        env->handle = scene->image_manager->add_image(
+            new BlenderImageLoader(b_image, image_frame, b_engine.is_preview()),
             env->image_params());
       }
       else {

intern/cycles/blender/sync.cpp

@@ -95,6 +95,11 @@ void BlenderSync::reset(BL::BlendData &b_data, BL::Scene &b_scene)
   this->b_scene = b_scene;
 }
 
+void BlenderSync::tag_update()
+{
+  has_updates_ = true;
+}
+
 /* Sync */
 
 void BlenderSync::sync_recalc(BL::Depsgraph &b_depsgraph, BL::SpaceView3D &b_v3d)

intern/cycles/blender/sync.h

@@ -66,6 +66,8 @@ class BlenderSync {
 
   void reset(BL::BlendData &b_data, BL::Scene &b_scene);
 
+  void tag_update();
+
   /* sync */
   void sync_recalc(BL::Depsgraph &b_depsgraph, BL::SpaceView3D &b_v3d);
   void sync_data(BL::RenderSettings &b_render,
@@ -167,12 +169,16 @@ class BlenderSync {
       Hair *hair, BL::Mesh &b_mesh, BObjectInfo &b_ob_info, bool motion, int motion_step = 0);
   bool object_has_particle_hair(BL::Object b_ob);
 
+  /* Point Cloud */
+  void sync_pointcloud(PointCloud *pointcloud, BObjectInfo &b_ob_info);
+  void sync_pointcloud_motion(PointCloud *pointcloud, BObjectInfo &b_ob_info, int motion_step = 0);
+
   /* Camera */
   void sync_camera_motion(
       BL::RenderSettings &b_render, BL::Object &b_ob, int width, int height, float motion_time);
 
   /* Geometry */
-  Geometry *sync_geometry(BL::Depsgraph &b_depsgrpah,
+  Geometry *sync_geometry(BL::Depsgraph &b_depsgraph,
                           BObjectInfo &b_ob_info,
                           bool object_updated,
                           bool use_particle_hair,
@@ -267,7 +273,6 @@ class BlenderSync {
 
   Progress &progress;
 
- protected:
   /* Indicates that `sync_recalc()` detected changes in the scene.
    * If this flag is false then the data is considered to be up-to-date and will not be
    * synchronized at all. */

intern/cycles/bvh/CMakeLists.txt

@@ -33,6 +33,17 @@ set(SRC
   unaligned.cpp
 )
 
+set(SRC_METAL
+  metal.mm
+)
+
+if(WITH_CYCLES_DEVICE_METAL)
+  list(APPEND SRC
+    ${SRC_METAL}
+  )
+  add_definitions(-DWITH_METAL)
+endif()
+
 set(SRC_HEADERS
   bvh.h
   bvh2.h
@@ -46,6 +57,7 @@ set(SRC_HEADERS
   sort.h
   split.h
   unaligned.h
+  metal.h
 )
 
 set(LIB

intern/cycles/bvh/build.cpp

@@ -26,6 +26,7 @@
 #include "scene/hair.h"
 #include "scene/mesh.h"
 #include "scene/object.h"
+#include "scene/pointcloud.h"
 #include "scene/scene.h"
 
 #include "util/algorithm.h"
@@ -113,9 +114,9 @@ void BVHBuild::add_reference_triangles(BoundBox &root,
     else {
       /* Motion triangles, trace optimized case:  we split triangle
        * primitives into separate nodes for each of the time steps.
-       * This way we minimize overlap of neighbor curve primitives.
+       * This way we minimize overlap of neighbor triangle primitives.
        */
-      const int num_bvh_steps = params.num_motion_curve_steps * 2 + 1;
+      const int num_bvh_steps = params.num_motion_triangle_steps * 2 + 1;
       const float num_bvh_steps_inv_1 = 1.0f / (num_bvh_steps - 1);
       const size_t num_verts = mesh->verts.size();
       const size_t num_steps = mesh->motion_steps;
@@ -269,6 +270,101 @@ void BVHBuild::add_reference_curves(BoundBox &root, BoundBox &center, Hair *hair
   }
 }
 
+void BVHBuild::add_reference_points(BoundBox &root,
+                                    BoundBox &center,
+                                    PointCloud *pointcloud,
+                                    int i)
+{
+  const Attribute *point_attr_mP = NULL;
+  if (pointcloud->has_motion_blur()) {
+    point_attr_mP = pointcloud->attributes.find(ATTR_STD_MOTION_VERTEX_POSITION);
+  }
+
+  const float3 *points_data = &pointcloud->points[0];
+  const float *radius_data = &pointcloud->radius[0];
+  const size_t num_points = pointcloud->num_points();
+  const float3 *motion_data = (point_attr_mP) ? point_attr_mP->data_float3() : NULL;
+  const size_t num_steps = pointcloud->get_motion_steps();
+
+  if (point_attr_mP == NULL) {
+    /* Really simple logic for static points. */
+    for (uint j = 0; j < num_points; j++) {
+      const PointCloud::Point point = pointcloud->get_point(j);
+      BoundBox bounds = BoundBox::empty;
+      point.bounds_grow(points_data, radius_data, bounds);
+      if (bounds.valid()) {
+        references.push_back(BVHReference(bounds, j, i, PRIMITIVE_POINT));
+        root.grow(bounds);
+        center.grow(bounds.center2());
+      }
+    }
+  }
+  else if (params.num_motion_point_steps == 0 || params.use_spatial_split) {
+    /* Simple case of motion points: single node for the whole
+     * shutter time. Lowest memory usage but less optimal
+     * rendering.
+     */
+    /* TODO(sergey): Support motion steps for spatially split BVH. */
+    for (uint j = 0; j < num_points; j++) {
+      const PointCloud::Point point = pointcloud->get_point(j);
+      BoundBox bounds = BoundBox::empty;
+      point.bounds_grow(points_data, radius_data, bounds);
+      for (size_t step = 0; step < num_steps - 1; step++) {
+        point.bounds_grow(motion_data + step * num_points, radius_data, bounds);
+      }
+      if (bounds.valid()) {
+        references.push_back(BVHReference(bounds, j, i, PRIMITIVE_MOTION_POINT));
+        root.grow(bounds);
+        center.grow(bounds.center2());
+      }
+    }
+  }
+  else {
+    /* Motion points, trace optimized case:  we split point
+     * primitives into separate nodes for each of the time steps.
+     * This way we minimize overlap of neighbor point primitives.
+     */
+    const int num_bvh_steps = params.num_motion_point_steps * 2 + 1;
+    const float num_bvh_steps_inv_1 = 1.0f / (num_bvh_steps - 1);
+
+    for (uint j = 0; j < num_points; j++) {
+      const PointCloud::Point point = pointcloud->get_point(j);
+      const size_t num_steps = pointcloud->get_motion_steps();
+      const float3 *point_steps = point_attr_mP->data_float3();
+
+      /* Calculate bounding box of the previous time step.
+       * Will be reused later to avoid duplicated work on
+       * calculating BVH time step boundbox.
+       */
+      float4 prev_key = point.motion_key(
+          points_data, radius_data, point_steps, num_points, num_steps, 0.0f, j);
+      BoundBox prev_bounds = BoundBox::empty;
+      point.bounds_grow(prev_key, prev_bounds);
+      /* Create all primitive time steps, */
+      for (int bvh_step = 1; bvh_step < num_bvh_steps; ++bvh_step) {
+        const float curr_time = (float)(bvh_step)*num_bvh_steps_inv_1;
+        float4 curr_key = point.motion_key(
+            points_data, radius_data, point_steps, num_points, num_steps, curr_time, j);
+        BoundBox curr_bounds = BoundBox::empty;
+        point.bounds_grow(curr_key, curr_bounds);
+        BoundBox bounds = prev_bounds;
+        bounds.grow(curr_bounds);
+        if (bounds.valid()) {
+          const float prev_time = (float)(bvh_step - 1) * num_bvh_steps_inv_1;
+          references.push_back(
+              BVHReference(bounds, j, i, PRIMITIVE_MOTION_POINT, prev_time, curr_time));
+          root.grow(bounds);
+          center.grow(bounds.center2());
+        }
+        /* Current time boundbox becomes previous one for the
+         * next time step.
+         */
+        prev_bounds = curr_bounds;
+      }
+    }
+  }
+}
+
 void BVHBuild::add_reference_geometry(BoundBox &root,
                                       BoundBox &center,
                                       Geometry *geom,
@@ -282,6 +378,10 @@ void BVHBuild::add_reference_geometry(BoundBox &root,
     Hair *hair = static_cast<Hair *>(geom);
     add_reference_curves(root, center, hair, object_index);
   }
+  else if (geom->geometry_type == Geometry::POINTCLOUD) {
+    PointCloud *pointcloud = static_cast<PointCloud *>(geom);
+    add_reference_points(root, center, pointcloud, object_index);
+  }
 }
 
 void BVHBuild::add_reference_object(BoundBox &root, BoundBox &center, Object *ob, int i)
@@ -311,6 +411,10 @@ static size_t count_primitives(Geometry *geom)
     Hair *hair = static_cast<Hair *>(geom);
     return count_curve_segments(hair);
   }
+  else if (geom->geometry_type == Geometry::POINTCLOUD) {
+    PointCloud *pointcloud = static_cast<PointCloud *>(geom);
+    return pointcloud->num_points();
+  }
 
   return 0;
 }
@@ -328,9 +432,10 @@ void BVHBuild::add_references(BVHRange &root)
       if (!ob->get_geometry()->is_instanced()) {
         num_alloc_references += count_primitives(ob->get_geometry());
       }
-      else
+      else {
         num_alloc_references++;
       }
+    }
     else {
       num_alloc_references += count_primitives(ob->get_geometry());
     }
@@ -394,7 +499,7 @@ BVHNode *BVHBuild::run()
   spatial_min_overlap = root.bounds().safe_area() * params.spatial_split_alpha;
   spatial_free_index = 0;
 
-  need_prim_time = params.num_motion_curve_steps > 0 || params.num_motion_triangle_steps > 0;
+  need_prim_time = params.use_motion_steps();
 
   /* init progress updates */
   double build_start_time;
@@ -535,7 +640,8 @@ bool BVHBuild::range_within_max_leaf_size(const BVHRange &range,
                                           const vector<BVHReference> &references) const
 {
   size_t size = range.size();
-  size_t max_leaf_size = max(params.max_triangle_leaf_size, params.max_curve_leaf_size);
+  size_t max_leaf_size = max(max(params.max_triangle_leaf_size, params.max_curve_leaf_size),
+                             params.max_point_leaf_size);
 
   if (size > max_leaf_size)
     return false;
@@ -544,32 +650,44 @@ bool BVHBuild::range_within_max_leaf_size(const BVHRange &range,
   size_t num_motion_triangles = 0;
   size_t num_curves = 0;
   size_t num_motion_curves = 0;
+  size_t num_points = 0;
+  size_t num_motion_points = 0;
 
   for (int i = 0; i < size; i++) {
     const BVHReference &ref = references[range.start() + i];
 
-    if (ref.prim_type() & PRIMITIVE_ALL_CURVE) {
+    if (ref.prim_type() & PRIMITIVE_CURVE) {
-      if (ref.prim_type() & PRIMITIVE_ALL_MOTION) {
+      if (ref.prim_type() & PRIMITIVE_MOTION) {
         num_motion_curves++;
       }
       else {
         num_curves++;
       }
     }
-    else if (ref.prim_type() & PRIMITIVE_ALL_TRIANGLE) {
+    else if (ref.prim_type() & PRIMITIVE_TRIANGLE) {
-      if (ref.prim_type() & PRIMITIVE_ALL_MOTION) {
+      if (ref.prim_type() & PRIMITIVE_MOTION) {
         num_motion_triangles++;
       }
       else {
         num_triangles++;
       }
     }
+    else if (ref.prim_type() & PRIMITIVE_POINT) {
+      if (ref.prim_type() & PRIMITIVE_MOTION) {
+        num_motion_points++;
+      }
+      else {
+        num_points++;
+      }
+    }
   }
 
   return (num_triangles <= params.max_triangle_leaf_size) &&
          (num_motion_triangles <= params.max_motion_triangle_leaf_size) &&
          (num_curves <= params.max_curve_leaf_size) &&
-         (num_motion_curves <= params.max_motion_curve_leaf_size);
+         (num_motion_curves <= params.max_motion_curve_leaf_size) &&
+         (num_points <= params.max_point_leaf_size) &&
+         (num_motion_points <= params.max_motion_point_leaf_size);
 }
 
 /* multithreaded binning builder */
@@ -855,7 +973,7 @@ BVHNode *BVHBuild::create_leaf_node(const BVHRange &range, const vector<BVHRefer
   for (int i = 0; i < range.size(); i++) {
     const BVHReference &ref = references[range.start() + i];
     if (ref.prim_index() != -1) {
-      uint32_t type_index = bitscan((uint32_t)(ref.prim_type() & PRIMITIVE_ALL));
+      uint32_t type_index = PRIMITIVE_INDEX(ref.prim_type() & PRIMITIVE_ALL);
       p_ref[type_index].push_back(ref);
       p_type[type_index].push_back(ref.prim_type());
       p_index[type_index].push_back(ref.prim_index());

intern/cycles/bvh/build.h

@@ -39,6 +39,7 @@ class Geometry;
 class Hair;
 class Mesh;
 class Object;
+class PointCloud;
 class Progress;
 
 /* BVH Builder */
@@ -68,6 +69,7 @@ class BVHBuild {
   /* Adding references. */
   void add_reference_triangles(BoundBox &root, BoundBox &center, Mesh *mesh, int i);
   void add_reference_curves(BoundBox &root, BoundBox &center, Hair *hair, int i);
+  void add_reference_points(BoundBox &root, BoundBox &center, PointCloud *pointcloud, int i);
   void add_reference_geometry(BoundBox &root, BoundBox &center, Geometry *geom, int i);
   void add_reference_object(BoundBox &root, BoundBox &center, Object *ob, int i);
   void add_references(BVHRange &root);

intern/cycles/bvh/bvh.cpp

@@ -19,6 +19,7 @@
 
 #include "bvh/bvh2.h"
 #include "bvh/embree.h"
+#include "bvh/metal.h"
 #include "bvh/multi.h"
 #include "bvh/optix.h"
 
@@ -40,8 +41,12 @@ const char *bvh_layout_name(BVHLayout layout)
       return "EMBREE";
     case BVH_LAYOUT_OPTIX:
       return "OPTIX";
+    case BVH_LAYOUT_METAL:
+      return "METAL";
     case BVH_LAYOUT_MULTI_OPTIX:
+    case BVH_LAYOUT_MULTI_METAL:
     case BVH_LAYOUT_MULTI_OPTIX_EMBREE:
+    case BVH_LAYOUT_MULTI_METAL_EMBREE:
       return "MULTI";
     case BVH_LAYOUT_ALL:
       return "ALL";
@@ -102,9 +107,18 @@ BVH *BVH::create(const BVHParams &params,
 #else
       (void)device;
       break;
+#endif
+    case BVH_LAYOUT_METAL:
+#ifdef WITH_METAL
+      return bvh_metal_create(params, geometry, objects, device);
+#else
+      (void)device;
+      break;
 #endif
     case BVH_LAYOUT_MULTI_OPTIX:
+    case BVH_LAYOUT_MULTI_METAL:
     case BVH_LAYOUT_MULTI_OPTIX_EMBREE:
+    case BVH_LAYOUT_MULTI_METAL_EMBREE:
       return new BVHMulti(params, geometry, objects);
     case BVH_LAYOUT_NONE:
     case BVH_LAYOUT_ALL:

intern/cycles/bvh/bvh2.cpp

@@ -20,6 +20,7 @@
 #include "scene/hair.h"
 #include "scene/mesh.h"
 #include "scene/object.h"
+#include "scene/pointcloud.h"
 
 #include "bvh/build.h"
 #include "bvh/node.h"
@@ -386,7 +387,7 @@ void BVH2::refit_primitives(int start, int end, BoundBox &bbox, uint &visibility
     }
     else {
       /* Primitives. */
-      if (pack.prim_type[prim] & PRIMITIVE_ALL_CURVE) {
+      if (pack.prim_type[prim] & PRIMITIVE_CURVE) {
         /* Curves. */
         const Hair *hair = static_cast<const Hair *>(ob->get_geometry());
         int prim_offset = (params.top_level) ? hair->prim_offset : 0;
@@ -409,6 +410,30 @@ void BVH2::refit_primitives(int start, int end, BoundBox &bbox, uint &visibility
           }
         }
       }
+      else if (pack.prim_type[prim] & PRIMITIVE_POINT) {
+        /* Points. */
+        const PointCloud *pointcloud = static_cast<const PointCloud *>(ob->get_geometry());
+        int prim_offset = (params.top_level) ? pointcloud->prim_offset : 0;
+        const float3 *points = &pointcloud->points[0];
+        const float *radius = &pointcloud->radius[0];
+        PointCloud::Point point = pointcloud->get_point(pidx - prim_offset);
+
+        point.bounds_grow(points, radius, bbox);
+
+        /* Motion points. */
+        if (pointcloud->get_use_motion_blur()) {
+          Attribute *attr = pointcloud->attributes.find(ATTR_STD_MOTION_VERTEX_POSITION);
+
+          if (attr) {
+            size_t pointcloud_size = pointcloud->points.size();
+            size_t steps = pointcloud->get_motion_steps() - 1;
+            float3 *point_steps = attr->data_float3();
+
+            for (size_t i = 0; i < steps; i++)
+              point.bounds_grow(point_steps + i * pointcloud_size, radius, bbox);
+          }
+        }
+      }
       else {
         /* Triangles. */
         const Mesh *mesh = static_cast<const Mesh *>(ob->get_geometry());
@@ -505,7 +530,8 @@ void BVH2::pack_instances(size_t nodes_size, size_t leaf_nodes_size)
   pack.leaf_nodes.resize(leaf_nodes_size);
   pack.object_node.resize(objects.size());
 
-  if (params.num_motion_curve_steps > 0 || params.num_motion_triangle_steps > 0) {
+  if (params.num_motion_curve_steps > 0 || params.num_motion_triangle_steps > 0 ||
+      params.num_motion_point_steps > 0) {
     pack.prim_time.resize(prim_index_size);
   }
 
@@ -564,13 +590,7 @@ void BVH2::pack_instances(size_t nodes_size, size_t leaf_nodes_size)
       float2 *bvh_prim_time = bvh->pack.prim_time.size() ? &bvh->pack.prim_time[0] : NULL;
 
       for (size_t i = 0; i < bvh_prim_index_size; i++) {
-        if (bvh->pack.prim_type[i] & PRIMITIVE_ALL_CURVE) {
         pack_prim_index[pack_prim_index_offset] = bvh_prim_index[i] + geom_prim_offset;
-        }
-        else {
-          pack_prim_index[pack_prim_index_offset] = bvh_prim_index[i] + geom_prim_offset;
-        }
-
         pack_prim_type[pack_prim_index_offset] = bvh_prim_type[i];
         pack_prim_visibility[pack_prim_index_offset] = bvh_prim_visibility[i];
         pack_prim_object[pack_prim_index_offset] = 0;  // unused for instances

intern/cycles/bvh/embree.cpp

@@ -45,6 +45,7 @@
 #  include "scene/hair.h"
 #  include "scene/mesh.h"
 #  include "scene/object.h"
+#  include "scene/pointcloud.h"
 
 #  include "util/foreach.h"
 #  include "util/log.h"
@@ -90,7 +91,7 @@ static void rtc_filter_occluded_func(const RTCFilterFunctionNArguments *args)
       ++ctx->num_hits;
 
       /* Always use baked shadow transparency for curves. */
-      if (current_isect.type & PRIMITIVE_ALL_CURVE) {
+      if (current_isect.type & PRIMITIVE_CURVE) {
         ctx->throughput *= intersection_curve_shadow_transparency(
             kg, current_isect.object, current_isect.prim, current_isect.u);
 
@@ -245,7 +246,7 @@ static void rtc_filter_occluded_func(const RTCFilterFunctionNArguments *args)
   }
 }
 
-static void rtc_filter_func_thick_curve(const RTCFilterFunctionNArguments *args)
+static void rtc_filter_func_backface_cull(const RTCFilterFunctionNArguments *args)
 {
   const RTCRay *ray = (RTCRay *)args->ray;
   RTCHit *hit = (RTCHit *)args->hit;
@@ -258,7 +259,7 @@ static void rtc_filter_func_thick_curve(const RTCFilterFunctionNArguments *args)
   }
 }
 
-static void rtc_filter_occluded_func_thick_curve(const RTCFilterFunctionNArguments *args)
+static void rtc_filter_occluded_func_backface_cull(const RTCFilterFunctionNArguments *args)
 {
   const RTCRay *ray = (RTCRay *)args->ray;
   RTCHit *hit = (RTCHit *)args->hit;
@@ -410,6 +411,12 @@ void BVHEmbree::add_object(Object *ob, int i)
       add_curves(ob, hair, i);
     }
   }
+  else if (geom->geometry_type == Geometry::POINTCLOUD) {
+    PointCloud *pointcloud = static_cast<PointCloud *>(geom);
+    if (pointcloud->num_points() > 0) {
+      add_points(ob, pointcloud, i);
+    }
+  }
 }
 
 void BVHEmbree::add_instance(Object *ob, int i)
@@ -624,6 +631,89 @@ void BVHEmbree::set_curve_vertex_buffer(RTCGeometry geom_id, const Hair *hair, c
   }
 }
 
+void BVHEmbree::set_point_vertex_buffer(RTCGeometry geom_id,
+                                        const PointCloud *pointcloud,
+                                        const bool update)
+{
+  const Attribute *attr_mP = NULL;
+  size_t num_motion_steps = 1;
+  if (pointcloud->has_motion_blur()) {
+    attr_mP = pointcloud->attributes.find(ATTR_STD_MOTION_VERTEX_POSITION);
+    if (attr_mP) {
+      num_motion_steps = pointcloud->get_motion_steps();
+    }
+  }
+
+  const size_t num_points = pointcloud->num_points();
+
+  /* Copy the point data to Embree */
+  const int t_mid = (num_motion_steps - 1) / 2;
+  const float *radius = pointcloud->get_radius().data();
+  for (int t = 0; t < num_motion_steps; ++t) {
+    const float3 *verts;
+    if (t == t_mid || attr_mP == NULL) {
+      verts = pointcloud->get_points().data();
+    }
+    else {
+      int t_ = (t > t_mid) ? (t - 1) : t;
+      verts = &attr_mP->data_float3()[t_ * num_points];
+    }
+
+    float4 *rtc_verts = (update) ? (float4 *)rtcGetGeometryBufferData(
+                                       geom_id, RTC_BUFFER_TYPE_VERTEX, t) :
+                                   (float4 *)rtcSetNewGeometryBuffer(geom_id,
+                                                                     RTC_BUFFER_TYPE_VERTEX,
+                                                                     t,
+                                                                     RTC_FORMAT_FLOAT4,
+                                                                     sizeof(float) * 4,
+                                                                     num_points);
+
+    assert(rtc_verts);
+    if (rtc_verts) {
+      for (size_t j = 0; j < num_points; ++j) {
+        rtc_verts[j] = float3_to_float4(verts[j]);
+        rtc_verts[j].w = radius[j];
+      }
+    }
+
+    if (update) {
+      rtcUpdateGeometryBuffer(geom_id, RTC_BUFFER_TYPE_VERTEX, t);
+    }
+  }
+}
+
+void BVHEmbree::add_points(const Object *ob, const PointCloud *pointcloud, int i)
+{
+  size_t prim_offset = pointcloud->prim_offset;
+
+  const Attribute *attr_mP = NULL;
+  size_t num_motion_steps = 1;
+  if (pointcloud->has_motion_blur()) {
+    attr_mP = pointcloud->attributes.find(ATTR_STD_MOTION_VERTEX_POSITION);
+    if (attr_mP) {
+      num_motion_steps = pointcloud->get_motion_steps();
+    }
+  }
+
+  enum RTCGeometryType type = RTC_GEOMETRY_TYPE_SPHERE_POINT;
+
+  RTCGeometry geom_id = rtcNewGeometry(rtc_device, type);
+
+  rtcSetGeometryBuildQuality(geom_id, build_quality);
+  rtcSetGeometryTimeStepCount(geom_id, num_motion_steps);
+
+  set_point_vertex_buffer(geom_id, pointcloud, false);
+
+  rtcSetGeometryUserData(geom_id, (void *)prim_offset);
+  rtcSetGeometryIntersectFilterFunction(geom_id, rtc_filter_func_backface_cull);
+  rtcSetGeometryOccludedFilterFunction(geom_id, rtc_filter_occluded_func_backface_cull);
+  rtcSetGeometryMask(geom_id, ob->visibility_for_tracing());
+
+  rtcCommitGeometry(geom_id);
+  rtcAttachGeometryByID(scene, geom_id, i * 2);
+  rtcReleaseGeometry(geom_id);
+}
+
 void BVHEmbree::add_curves(const Object *ob, const Hair *hair, int i)
 {
   size_t prim_offset = hair->curve_segment_offset;
@@ -678,8 +768,8 @@ void BVHEmbree::add_curves(const Object *ob, const Hair *hair, int i)
     rtcSetGeometryOccludedFilterFunction(geom_id, rtc_filter_occluded_func);
   }
   else {
-    rtcSetGeometryIntersectFilterFunction(geom_id, rtc_filter_func_thick_curve);
+    rtcSetGeometryIntersectFilterFunction(geom_id, rtc_filter_func_backface_cull);
-    rtcSetGeometryOccludedFilterFunction(geom_id, rtc_filter_occluded_func_thick_curve);
+    rtcSetGeometryOccludedFilterFunction(geom_id, rtc_filter_occluded_func_backface_cull);
   }
   rtcSetGeometryMask(geom_id, ob->visibility_for_tracing());
 
@@ -716,6 +806,14 @@ void BVHEmbree::refit(Progress &progress)
           rtcCommitGeometry(geom);
         }
       }
+      else if (geom->geometry_type == Geometry::POINTCLOUD) {
+        PointCloud *pointcloud = static_cast<PointCloud *>(geom);
+        if (pointcloud->num_points() > 0) {
+          RTCGeometry geom = rtcGetGeometry(scene, geom_id);
+          set_point_vertex_buffer(geom, pointcloud, true);
+          rtcCommitGeometry(geom);
+        }
+      }
     }
     geom_id += 2;
   }

intern/cycles/bvh/embree.h

@@ -33,6 +33,7 @@ CCL_NAMESPACE_BEGIN
 
 class Hair;
 class Mesh;
+class PointCloud;
 
 class BVHEmbree : public BVH {
  public:
@@ -51,11 +52,15 @@ class BVHEmbree : public BVH {
   void add_object(Object *ob, int i);
   void add_instance(Object *ob, int i);
   void add_curves(const Object *ob, const Hair *hair, int i);
+  void add_points(const Object *ob, const PointCloud *pointcloud, int i);
   void add_triangles(const Object *ob, const Mesh *mesh, int i);
 
  private:
   void set_tri_vertex_buffer(RTCGeometry geom_id, const Mesh *mesh, const bool update);
   void set_curve_vertex_buffer(RTCGeometry geom_id, const Hair *hair, const bool update);
+  void set_point_vertex_buffer(RTCGeometry geom_id,
+                               const PointCloud *pointcloud,
+                               const bool update);
 
   RTCDevice rtc_device;
   enum RTCBuildQuality build_quality;

intern/cycles/bvh/metal.h

@@ -0,0 +1,35 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef __BVH_METAL_H__
+#define __BVH_METAL_H__
+
+#ifdef WITH_METAL
+
+#  include "bvh/bvh.h"
+
+CCL_NAMESPACE_BEGIN
+
+BVH *bvh_metal_create(const BVHParams &params,
+                      const vector<Geometry *> &geometry,
+                      const vector<Object *> &objects,
+                      Device *device);
+
+CCL_NAMESPACE_END
+
+#endif /* WITH_METAL */
+
+#endif /* __BVH_METAL_H__ */

intern/cycles/bvh/metal.mm

@@ -0,0 +1,33 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifdef WITH_METAL
+
+#  include "device/metal/bvh.h"
+
+CCL_NAMESPACE_BEGIN
+
+BVH *bvh_metal_create(const BVHParams &params,
+                      const vector<Geometry *> &geometry,
+                      const vector<Object *> &objects,
+                      Device *device)
+{
+  return new BVHMetal(params, geometry, objects, device);
+}
+
+CCL_NAMESPACE_END
+
+#endif /* WITH_METAL */

intern/cycles/bvh/params.h

@@ -83,6 +83,8 @@ class BVHParams {
   int max_motion_triangle_leaf_size;
   int max_curve_leaf_size;
   int max_motion_curve_leaf_size;
+  int max_point_leaf_size;
+  int max_motion_point_leaf_size;
 
   /* object or mesh level bvh */
   bool top_level;
@@ -98,13 +100,13 @@ class BVHParams {
   /* Split time range to this number of steps and create leaf node for each
    * of this time steps.
    *
-   * Speeds up rendering of motion curve primitives in the cost of higher
+   * Speeds up rendering of motion primitives in the cost of higher memory usage.
-   * memory usage.
    */
-  int num_motion_curve_steps;
 
   /* Same as above, but for triangle primitives. */
   int num_motion_triangle_steps;
+  int num_motion_curve_steps;
+  int num_motion_point_steps;
 
   /* Same as in SceneParams. */
   int bvh_type;
@@ -132,6 +134,8 @@ class BVHParams {
     max_motion_triangle_leaf_size = 8;
     max_curve_leaf_size = 1;
     max_motion_curve_leaf_size = 4;
+    max_point_leaf_size = 8;
+    max_motion_point_leaf_size = 8;
 
     top_level = false;
     bvh_layout = BVH_LAYOUT_BVH2;
@@ -139,6 +143,7 @@ class BVHParams {
 
     num_motion_curve_steps = 0;
     num_motion_triangle_steps = 0;
+    num_motion_point_steps = 0;
 
     bvh_type = 0;
 
@@ -166,6 +171,12 @@ class BVHParams {
     return (size <= min_leaf_size || level >= MAX_DEPTH);
   }
 
+  bool use_motion_steps()
+  {
+    return num_motion_curve_steps > 0 || num_motion_triangle_steps > 0 ||
+           num_motion_point_steps > 0;
+  }
+
   /* Gets best matching BVH.
    *
    * If the requested layout is supported by the device, it will be used.

intern/cycles/bvh/split.cpp

@@ -23,6 +23,7 @@
 #include "scene/hair.h"
 #include "scene/mesh.h"
 #include "scene/object.h"
+#include "scene/pointcloud.h"
 
 #include "util/algorithm.h"
 
@@ -426,6 +427,32 @@ void BVHSpatialSplit::split_curve_primitive(const Hair *hair,
   }
 }
 
+void BVHSpatialSplit::split_point_primitive(const PointCloud *pointcloud,
+                                            const Transform *tfm,
+                                            int prim_index,
+                                            int dim,
+                                            float pos,
+                                            BoundBox &left_bounds,
+                                            BoundBox &right_bounds)
+{
+  /* No real splitting support for points, assume they are small enough for it
+   * not to matter. */
+  float3 point = pointcloud->get_points()[prim_index];
+
+  if (tfm != NULL) {
+    point = transform_point(tfm, point);
+  }
+  point = get_unaligned_point(point);
+
+  if (point[dim] <= pos) {
+    left_bounds.grow(point);
+  }
+
+  if (point[dim] >= pos) {
+    right_bounds.grow(point);
+  }
+}
+
 void BVHSpatialSplit::split_triangle_reference(const BVHReference &ref,
                                                const Mesh *mesh,
                                                int dim,
@@ -453,6 +480,16 @@ void BVHSpatialSplit::split_curve_reference(const BVHReference &ref,
                         right_bounds);
 }
 
+void BVHSpatialSplit::split_point_reference(const BVHReference &ref,
+                                            const PointCloud *pointcloud,
+                                            int dim,
+                                            float pos,
+                                            BoundBox &left_bounds,
+                                            BoundBox &right_bounds)
+{
+  split_point_primitive(pointcloud, NULL, ref.prim_index(), dim, pos, left_bounds, right_bounds);
+}
+
 void BVHSpatialSplit::split_object_reference(
     const Object *object, int dim, float pos, BoundBox &left_bounds, BoundBox &right_bounds)
 {
@@ -475,6 +512,13 @@ void BVHSpatialSplit::split_object_reference(
       }
     }
   }
+  else if (geom->geometry_type == Geometry::POINTCLOUD) {
+    PointCloud *pointcloud = static_cast<PointCloud *>(geom);
+    for (int point_idx = 0; point_idx < pointcloud->num_points(); ++point_idx) {
+      split_point_primitive(
+          pointcloud, &object->get_tfm(), point_idx, dim, pos, left_bounds, right_bounds);
+    }
+  }
 }
 
 void BVHSpatialSplit::split_reference(const BVHBuild &builder,
@@ -491,14 +535,18 @@ void BVHSpatialSplit::split_reference(const BVHBuild &builder,
   /* loop over vertices/edges. */
   const Object *ob = builder.objects[ref.prim_object()];
 
-  if (ref.prim_type() & PRIMITIVE_ALL_TRIANGLE) {
+  if (ref.prim_type() & PRIMITIVE_TRIANGLE) {
     Mesh *mesh = static_cast<Mesh *>(ob->get_geometry());
     split_triangle_reference(ref, mesh, dim, pos, left_bounds, right_bounds);
   }
-  else if (ref.prim_type() & PRIMITIVE_ALL_CURVE) {
+  else if (ref.prim_type() & PRIMITIVE_CURVE) {
     Hair *hair = static_cast<Hair *>(ob->get_geometry());
     split_curve_reference(ref, hair, dim, pos, left_bounds, right_bounds);
   }
+  else if (ref.prim_type() & PRIMITIVE_POINT) {
+    PointCloud *pointcloud = static_cast<PointCloud *>(ob->get_geometry());
+    split_point_reference(ref, pointcloud, dim, pos, left_bounds, right_bounds);
+  }
   else {
     split_object_reference(ob, dim, pos, left_bounds, right_bounds);
   }

intern/cycles/bvh/split.h

@@ -26,6 +26,7 @@ CCL_NAMESPACE_BEGIN
 class BVHBuild;
 class Hair;
 class Mesh;
+class PointCloud;
 struct Transform;
 
 /* Object Split */
@@ -123,6 +124,13 @@ class BVHSpatialSplit {
                              float pos,
                              BoundBox &left_bounds,
                              BoundBox &right_bounds);
+  void split_point_primitive(const PointCloud *pointcloud,
+                             const Transform *tfm,
+                             int prim_index,
+                             int dim,
+                             float pos,
+                             BoundBox &left_bounds,
+                             BoundBox &right_bounds);
 
   /* Lower-level functions which calculates boundaries of left and right nodes
    * needed for spatial split.
@@ -141,6 +149,12 @@ class BVHSpatialSplit {
                              float pos,
                              BoundBox &left_bounds,
                              BoundBox &right_bounds);
+  void split_point_reference(const BVHReference &ref,
+                             const PointCloud *pointcloud,
+                             int dim,
+                             float pos,
+                             BoundBox &left_bounds,
+                             BoundBox &right_bounds);
   void split_object_reference(
       const Object *object, int dim, float pos, BoundBox &left_bounds, BoundBox &right_bounds);
 

intern/cycles/bvh/unaligned.cpp

@@ -69,7 +69,7 @@ bool BVHUnaligned::compute_aligned_space(const BVHReference &ref, Transform *ali
   const int packed_type = ref.prim_type();
   const int type = (packed_type & PRIMITIVE_ALL);
   /* No motion blur curves here, we can't fit them to aligned boxes well. */
-  if (type & (PRIMITIVE_CURVE_RIBBON | PRIMITIVE_CURVE_THICK)) {
+  if ((type & PRIMITIVE_CURVE) && !(type & PRIMITIVE_MOTION)) {
     const int curve_index = ref.prim_index();
     const int segment = PRIMITIVE_UNPACK_SEGMENT(packed_type);
     const Hair *hair = static_cast<const Hair *>(object->get_geometry());
@@ -95,7 +95,7 @@ BoundBox BVHUnaligned::compute_aligned_prim_boundbox(const BVHReference &prim,
   const int packed_type = prim.prim_type();
   const int type = (packed_type & PRIMITIVE_ALL);
   /* No motion blur curves here, we can't fit them to aligned boxes well. */
-  if (type & (PRIMITIVE_CURVE_RIBBON | PRIMITIVE_CURVE_THICK)) {
+  if ((type & PRIMITIVE_CURVE) && !(type & PRIMITIVE_MOTION)) {
     const int curve_index = prim.prim_index();
     const int segment = PRIMITIVE_UNPACK_SEGMENT(packed_type);
     const Hair *hair = static_cast<const Hair *>(object->get_geometry());

intern/cycles/cmake/external_libs.cmake

@@ -551,4 +551,23 @@ if(NOT WITH_HIP_DYNLOAD)
   set(WITH_HIP_DYNLOAD ON)
 endif()
 
+###########################################################################
+# Metal
+###########################################################################
+
+if(WITH_CYCLES_DEVICE_METAL)
+  find_library(METAL_LIBRARY Metal)
+
+  # This file was added in the 12.0 SDK, use it as a way to detect the version.
+  if (METAL_LIBRARY AND NOT EXISTS "${METAL_LIBRARY}/Headers/MTLFunctionStitching.h")
+    message(STATUS "Metal version too old, must be SDK 12.0 or newer, disabling WITH_CYCLES_DEVICE_METAL")
+    set(WITH_CYCLES_DEVICE_METAL OFF)
+  elseif (NOT METAL_LIBRARY)
+    message(STATUS "Metal not found, disabling WITH_CYCLES_DEVICE_METAL")
+    set(WITH_CYCLES_DEVICE_METAL OFF)
+  else()
+    message(STATUS "Found Metal: ${METAL_LIBRARY}")
+  endif()
+endif()
+
 unset(_cycles_lib_dir)

intern/cycles/device/CMakeLists.txt

@@ -43,7 +43,7 @@ if(WITH_CYCLES_DEVICE_HIP AND WITH_HIP_DYNLOAD)
   add_definitions(-DWITH_HIP_DYNLOAD)
 endif()
 
-set(SRC
+set(SRC_BASE
   device.cpp
   denoise.cpp
   graphics_interop.cpp
@@ -104,6 +104,21 @@ set(SRC_MULTI
   multi/device.h
 )
 
+set(SRC_METAL
+  metal/bvh.mm
+  metal/bvh.h
+  metal/device.mm
+  metal/device.h
+  metal/device_impl.mm
+  metal/device_impl.h
+  metal/kernel.mm
+  metal/kernel.h
+  metal/queue.mm
+  metal/queue.h
+  metal/util.mm
+  metal/util.h
+)
+
 set(SRC_OPTIX
   optix/device.cpp
   optix/device.h
@@ -123,6 +138,17 @@ set(SRC_HEADERS
   queue.h
 )
 
+set(SRC
+  ${SRC_BASE}
+  ${SRC_CPU}
+  ${SRC_CUDA}
+  ${SRC_HIP}
+  ${SRC_DUMMY}
+  ${SRC_MULTI}
+  ${SRC_OPTIX}
+  ${SRC_HEADERS}
+)
+
 set(LIB
   cycles_kernel
   cycles_util
@@ -158,6 +184,15 @@ endif()
 if(WITH_CYCLES_DEVICE_OPTIX)
   add_definitions(-DWITH_OPTIX)
 endif()
+if(WITH_CYCLES_DEVICE_METAL)
+  list(APPEND LIB
+    ${METAL_LIBRARY}
+  )
+  add_definitions(-DWITH_METAL)
+  list(APPEND SRC
+    ${SRC_METAL}
+  )
+endif()
 
 if(WITH_OPENIMAGEDENOISE)
   list(APPEND LIB
@@ -168,20 +203,12 @@ endif()
 include_directories(${INC})
 include_directories(SYSTEM ${INC_SYS})
 
-cycles_add_library(cycles_device "${LIB}"
+cycles_add_library(cycles_device "${LIB}" ${SRC})
-  ${SRC}
-  ${SRC_CPU}
-  ${SRC_CUDA}
-  ${SRC_HIP}
-  ${SRC_DUMMY}
-  ${SRC_MULTI}
-  ${SRC_OPTIX}
-  ${SRC_HEADERS}
-)
 
 source_group("cpu" FILES ${SRC_CPU})
 source_group("cuda" FILES ${SRC_CUDA})
 source_group("dummy" FILES ${SRC_DUMMY})
 source_group("multi" FILES ${SRC_MULTI})
+source_group("metal" FILES ${SRC_METAL})
 source_group("optix" FILES ${SRC_OPTIX})
 source_group("common" FILES ${SRC} ${SRC_HEADERS})

intern/cycles/device/cpu/device_impl.cpp

@@ -129,8 +129,7 @@ void CPUDevice::mem_alloc(device_memory &mem)
               << string_human_readable_size(mem.memory_size()) << ")";
     }
 
-    if (mem.type == MEM_DEVICE_ONLY) {
+    if (mem.type == MEM_DEVICE_ONLY || !mem.host_pointer) {
-      assert(!mem.host_pointer);
       size_t alignment = MIN_ALIGNMENT_CPU_DATA_TYPES;
       void *data = util_aligned_malloc(mem.memory_size(), alignment);
       mem.device_pointer = (device_ptr)data;
@@ -189,7 +188,7 @@ void CPUDevice::mem_free(device_memory &mem)
     tex_free((device_texture &)mem);
   }
   else if (mem.device_pointer) {
-    if (mem.type == MEM_DEVICE_ONLY) {
+    if (mem.type == MEM_DEVICE_ONLY || !mem.host_pointer) {
       util_aligned_free((void *)mem.device_pointer);
     }
     mem.device_pointer = 0;
@@ -274,7 +273,8 @@ void CPUDevice::build_bvh(BVH *bvh, Progress &progress, bool refit)
 {
 #ifdef WITH_EMBREE
   if (bvh->params.bvh_layout == BVH_LAYOUT_EMBREE ||
-      bvh->params.bvh_layout == BVH_LAYOUT_MULTI_OPTIX_EMBREE) {
+      bvh->params.bvh_layout == BVH_LAYOUT_MULTI_OPTIX_EMBREE ||
+      bvh->params.bvh_layout == BVH_LAYOUT_MULTI_METAL_EMBREE) {
     BVHEmbree *const bvh_embree = static_cast<BVHEmbree *>(bvh);
     if (refit) {
       bvh_embree->refit(progress);

intern/cycles/device/cuda/device_impl.cpp

@@ -477,10 +477,10 @@ void CUDADevice::reserve_local_memory(const uint kernel_features)
      * still to make it faster. */
     CUDADeviceQueue queue(this);
 
-    void *d_path_index = nullptr;
+    device_ptr d_path_index = 0;
-    void *d_render_buffer = nullptr;
+    device_ptr d_render_buffer = 0;
     int d_work_size = 0;
-    void *args[] = {&d_path_index, &d_render_buffer, &d_work_size};
+    DeviceKernelArguments args(&d_path_index, &d_render_buffer, &d_work_size);
 
     queue.init_execution();
     queue.enqueue(test_kernel, 1, args);
@@ -678,7 +678,7 @@ CUDADevice::CUDAMem *CUDADevice::generic_alloc(device_memory &mem, size_t pitch_
 
   void *shared_pointer = 0;
 
-  if (mem_alloc_result != CUDA_SUCCESS && can_map_host) {
+  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;
@@ -701,9 +701,15 @@ CUDADevice::CUDAMem *CUDADevice::generic_alloc(device_memory &mem, size_t pitch_
   }
 
   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(1) << "Buffer allocate: " << mem.name << ", "

intern/cycles/device/cuda/queue.cpp

@@ -89,7 +89,9 @@ bool CUDADeviceQueue::kernel_available(DeviceKernel kernel) const
   return cuda_device_->kernels.available(kernel);
 }
 
-bool CUDADeviceQueue::enqueue(DeviceKernel kernel, const int work_size, void *args[])
+bool CUDADeviceQueue::enqueue(DeviceKernel kernel,
+                              const int work_size,
+                              DeviceKernelArguments const &args)
 {
   if (cuda_device_->have_error()) {
     return false;
@@ -133,7 +135,7 @@ bool CUDADeviceQueue::enqueue(DeviceKernel kernel, const int work_size, void *ar
                                 1,
                                 shared_mem_bytes,
                                 cuda_stream_,
-                                args,
+                                const_cast<void **>(args.values),
                                 0),
                  "enqueue");
 

intern/cycles/device/cuda/queue.h

@@ -42,7 +42,9 @@ class CUDADeviceQueue : public DeviceQueue {
 
   virtual bool kernel_available(DeviceKernel kernel) const override;
 
-  virtual bool enqueue(DeviceKernel kernel, const int work_size, void *args[]) override;
+  virtual bool enqueue(DeviceKernel kernel,
+                       const int work_size,
+                       DeviceKernelArguments const &args) override;
 
   virtual bool synchronize() override;
 

intern/cycles/device/device.cpp

@@ -27,6 +27,7 @@
 #include "device/cuda/device.h"
 #include "device/dummy/device.h"
 #include "device/hip/device.h"
+#include "device/metal/device.h"
 #include "device/multi/device.h"
 #include "device/optix/device.h"
 
@@ -49,6 +50,7 @@ vector<DeviceInfo> Device::cuda_devices;
 vector<DeviceInfo> Device::optix_devices;
 vector<DeviceInfo> Device::cpu_devices;
 vector<DeviceInfo> Device::hip_devices;
+vector<DeviceInfo> Device::metal_devices;
 uint Device::devices_initialized_mask = 0;
 
 /* Device */
@@ -105,6 +107,12 @@ Device *Device::create(const DeviceInfo &info, Stats &stats, Profiler &profiler)
       break;
 #endif
 
+#ifdef WITH_METAL
+    case DEVICE_METAL:
+      if (device_metal_init())
+        device = device_metal_create(info, stats, profiler);
+      break;
+#endif
     default:
       break;
   }
@@ -128,6 +136,8 @@ DeviceType Device::type_from_string(const char *name)
     return DEVICE_MULTI;
   else if (strcmp(name, "HIP") == 0)
     return DEVICE_HIP;
+  else if (strcmp(name, "METAL") == 0)
+    return DEVICE_METAL;
 
   return DEVICE_NONE;
 }
@@ -144,6 +154,8 @@ string Device::string_from_type(DeviceType type)
     return "MULTI";
   else if (type == DEVICE_HIP)
     return "HIP";
+  else if (type == DEVICE_METAL)
+    return "METAL";
 
   return "";
 }
@@ -161,7 +173,9 @@ vector<DeviceType> Device::available_types()
 #ifdef WITH_HIP
   types.push_back(DEVICE_HIP);
 #endif
-
+#ifdef WITH_METAL
+  types.push_back(DEVICE_METAL);
+#endif
   return types;
 }
 
@@ -227,6 +241,20 @@ vector<DeviceInfo> Device::available_devices(uint mask)
     }
   }
 
+#ifdef WITH_METAL
+  if (mask & DEVICE_MASK_METAL) {
+    if (!(devices_initialized_mask & DEVICE_MASK_METAL)) {
+      if (device_metal_init()) {
+        device_metal_info(metal_devices);
+      }
+      devices_initialized_mask |= DEVICE_MASK_METAL;
+    }
+    foreach (DeviceInfo &info, metal_devices) {
+      devices.push_back(info);
+    }
+  }
+#endif
+
   return devices;
 }
 
@@ -266,6 +294,15 @@ string Device::device_capabilities(uint mask)
   }
 #endif
 
+#ifdef WITH_METAL
+  if (mask & DEVICE_MASK_METAL) {
+    if (device_metal_init()) {
+      capabilities += "\nMetal device capabilities:\n";
+      capabilities += device_metal_capabilities();
+    }
+  }
+#endif
+
   return capabilities;
 }
 
@@ -354,6 +391,7 @@ void Device::free_memory()
   optix_devices.free_memory();
   hip_devices.free_memory();
   cpu_devices.free_memory();
+  metal_devices.free_memory();
 }
 
 unique_ptr<DeviceQueue> Device::gpu_queue_create()

intern/cycles/device/device.h

@@ -52,6 +52,7 @@ enum DeviceType {
   DEVICE_MULTI,
   DEVICE_OPTIX,
   DEVICE_HIP,
+  DEVICE_METAL,
   DEVICE_DUMMY,
 };
 
@@ -60,6 +61,7 @@ enum DeviceTypeMask {
   DEVICE_MASK_CUDA = (1 << DEVICE_CUDA),
   DEVICE_MASK_OPTIX = (1 << DEVICE_OPTIX),
   DEVICE_MASK_HIP = (1 << DEVICE_HIP),
+  DEVICE_MASK_METAL = (1 << DEVICE_METAL),
   DEVICE_MASK_ALL = ~0
 };
 
@@ -281,6 +283,7 @@ class Device {
   static vector<DeviceInfo> optix_devices;
   static vector<DeviceInfo> cpu_devices;
   static vector<DeviceInfo> hip_devices;
+  static vector<DeviceInfo> metal_devices;
   static uint devices_initialized_mask;
 };
 

intern/cycles/device/hip/device_impl.cpp

@@ -440,10 +440,10 @@ void HIPDevice::reserve_local_memory(const uint kernel_features)
      * still to make it faster. */
     HIPDeviceQueue queue(this);
 
-    void *d_path_index = nullptr;
+    device_ptr d_path_index = 0;
-    void *d_render_buffer = nullptr;
+    device_ptr d_render_buffer = 0;
     int d_work_size = 0;
-    void *args[] = {&d_path_index, &d_render_buffer, &d_work_size};
+    DeviceKernelArguments args(&d_path_index, &d_render_buffer, &d_work_size);
 
     queue.init_execution();
     queue.enqueue(test_kernel, 1, args);

intern/cycles/device/hip/queue.cpp

@@ -89,7 +89,9 @@ bool HIPDeviceQueue::kernel_available(DeviceKernel kernel) const
   return hip_device_->kernels.available(kernel);
 }
 
-bool HIPDeviceQueue::enqueue(DeviceKernel kernel, const int work_size, void *args[])
+bool HIPDeviceQueue::enqueue(DeviceKernel kernel,
+                             const int work_size,
+                             DeviceKernelArguments const &args)
 {
   if (hip_device_->have_error()) {
     return false;
@@ -132,7 +134,7 @@ bool HIPDeviceQueue::enqueue(DeviceKernel kernel, const int work_size, void *arg
                                        1,
                                        shared_mem_bytes,
                                        hip_stream_,
-                                       args,
+                                       const_cast<void **>(args.values),
                                        0),
                  "enqueue");
 

intern/cycles/device/hip/queue.h

@@ -42,7 +42,9 @@ class HIPDeviceQueue : public DeviceQueue {
 
   virtual bool kernel_available(DeviceKernel kernel) const override;
 
-  virtual bool enqueue(DeviceKernel kernel, const int work_size, void *args[]) override;
+  virtual bool enqueue(DeviceKernel kernel,
+                       const int work_size,
+                       DeviceKernelArguments const &args) override;
 
   virtual bool synchronize() override;
 

intern/cycles/device/memory.h

@@ -263,6 +263,7 @@ class device_memory {
   friend class CUDADevice;
   friend class OptiXDevice;
   friend class HIPDevice;
+  friend class MetalDevice;
 
   /* Only create through subclasses. */
   device_memory(Device *device, const char *name, MemoryType type);
@@ -581,7 +582,7 @@ template<typename T> class device_vector : public device_memory {
  * from an already allocated base memory. It is freed automatically when it
  * goes out of scope, which should happen before base memory is freed.
  *
- * Note: some devices require offset and size of the sub_ptr to be properly
+ * NOTE: some devices require offset and size of the sub_ptr to be properly
  * aligned to device->mem_address_alingment(). */
 
 class device_sub_ptr {

intern/cycles/device/metal/bvh.h

@@ -0,0 +1,66 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#ifdef WITH_METAL
+
+#  include "bvh/bvh.h"
+#  include "bvh/params.h"
+#  include "device/memory.h"
+
+#  include <Metal/Metal.h>
+
+CCL_NAMESPACE_BEGIN
+
+class BVHMetal : public BVH {
+ public:
+  API_AVAILABLE(macos(11.0))
+  id<MTLAccelerationStructure> accel_struct = nil;
+  bool accel_struct_building = false;
+
+  API_AVAILABLE(macos(11.0))
+  vector<id<MTLAccelerationStructure>> blas_array;
+
+  bool motion_blur = false;
+
+  Stats &stats;
+
+  bool build(Progress &progress, id<MTLDevice> device, id<MTLCommandQueue> queue, bool refit);
+
+  BVHMetal(const BVHParams &params,
+           const vector<Geometry *> &geometry,
+           const vector<Object *> &objects,
+           Device *device);
+  virtual ~BVHMetal();
+
+  bool build_BLAS(Progress &progress, id<MTLDevice> device, id<MTLCommandQueue> queue, bool refit);
+  bool build_BLAS_mesh(Progress &progress,
+                       id<MTLDevice> device,
+                       id<MTLCommandQueue> queue,
+                       Geometry *const geom,
+                       bool refit);
+  bool build_BLAS_hair(Progress &progress,
+                       id<MTLDevice> device,
+                       id<MTLCommandQueue> queue,
+                       Geometry *const geom,
+                       bool refit);
+  bool build_TLAS(Progress &progress, id<MTLDevice> device, id<MTLCommandQueue> queue, bool refit);
+};
+
+CCL_NAMESPACE_END
+
+#endif /* WITH_METAL */

intern/cycles/device/metal/bvh.mm

@@ -0,0 +1,813 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifdef WITH_METAL
+
+#  include "scene/hair.h"
+#  include "scene/mesh.h"
+#  include "scene/object.h"
+
+#  include "util/progress.h"
+
+#  include "device/metal/bvh.h"
+
+CCL_NAMESPACE_BEGIN
+
+#  define BVH_status(...) \
+    { \
+      string str = string_printf(__VA_ARGS__); \
+      progress.set_substatus(str); \
+    }
+
+BVHMetal::BVHMetal(const BVHParams &params_,
+                   const vector<Geometry *> &geometry_,
+                   const vector<Object *> &objects_,
+                   Device *device)
+    : BVH(params_, geometry_, objects_), stats(device->stats)
+{
+}
+
+BVHMetal::~BVHMetal()
+{
+  if (@available(macos 12.0, *)) {
+    if (accel_struct) {
+      stats.mem_free(accel_struct.allocatedSize);
+      [accel_struct release];
+    }
+  }
+}
+
+bool BVHMetal::build_BLAS_mesh(Progress &progress,
+                               id<MTLDevice> device,
+                               id<MTLCommandQueue> queue,
+                               Geometry *const geom,
+                               bool refit)
+{
+  if (@available(macos 12.0, *)) {
+    /* Build BLAS for triangle primitives */
+    Mesh *const mesh = static_cast<Mesh *const>(geom);
+    if (mesh->num_triangles() == 0) {
+      return false;
+    }
+
+    /*------------------------------------------------*/
+    BVH_status(
+        "Building mesh BLAS | %7d tris | %s", (int)mesh->num_triangles(), geom->name.c_str());
+    /*------------------------------------------------*/
+
+    const bool use_fast_trace_bvh = (params.bvh_type == BVH_TYPE_STATIC);
+
+    const array<float3> &verts = mesh->get_verts();
+    const array<int> &tris = mesh->get_triangles();
+    const size_t num_verts = verts.size();
+    const size_t num_indices = tris.size();
+
+    size_t num_motion_steps = 1;
+    Attribute *motion_keys = mesh->attributes.find(ATTR_STD_MOTION_VERTEX_POSITION);
+    if (motion_blur && mesh->get_use_motion_blur() && motion_keys) {
+      num_motion_steps = mesh->get_motion_steps();
+    }
+
+    MTLResourceOptions storage_mode;
+    if (device.hasUnifiedMemory) {
+      storage_mode = MTLResourceStorageModeShared;
+    }
+    else {
+      storage_mode = MTLResourceStorageModeManaged;
+    }
+
+    /* Upload the mesh data to the GPU */
+    id<MTLBuffer> posBuf = nil;
+    id<MTLBuffer> indexBuf = [device newBufferWithBytes:tris.data()
+                                                 length:num_indices * sizeof(tris.data()[0])
+                                                options:storage_mode];
+
+    if (num_motion_steps == 1) {
+      posBuf = [device newBufferWithBytes:verts.data()
+                                   length:num_verts * sizeof(verts.data()[0])
+                                  options:storage_mode];
+    }
+    else {
+      posBuf = [device newBufferWithLength:num_verts * num_motion_steps * sizeof(verts.data()[0])
+                                   options:storage_mode];
+      float3 *dest_data = (float3 *)[posBuf contents];
+      size_t center_step = (num_motion_steps - 1) / 2;
+      for (size_t step = 0; step < num_motion_steps; ++step) {
+        const float3 *verts = mesh->get_verts().data();
+
+        /* The center step for motion vertices is not stored in the attribute. */
+        if (step != center_step) {
+          verts = motion_keys->data_float3() + (step > center_step ? step - 1 : step) * num_verts;
+        }
+        memcpy(dest_data + num_verts * step, verts, num_verts * sizeof(float3));
+      }
+      if (storage_mode == MTLResourceStorageModeManaged) {
+        [posBuf didModifyRange:NSMakeRange(0, posBuf.length)];
+      }
+    }
+
+    /* Create an acceleration structure. */
+    MTLAccelerationStructureGeometryDescriptor *geomDesc;
+    if (num_motion_steps > 1) {
+      std::vector<MTLMotionKeyframeData *> vertex_ptrs;
+      vertex_ptrs.reserve(num_motion_steps);
+      for (size_t step = 0; step < num_motion_steps; ++step) {
+        MTLMotionKeyframeData *k = [MTLMotionKeyframeData data];
+        k.buffer = posBuf;
+        k.offset = num_verts * step * sizeof(float3);
+        vertex_ptrs.push_back(k);
+      }
+
+      MTLAccelerationStructureMotionTriangleGeometryDescriptor *geomDescMotion =
+          [MTLAccelerationStructureMotionTriangleGeometryDescriptor descriptor];
+      geomDescMotion.vertexBuffers = [NSArray arrayWithObjects:vertex_ptrs.data()
+                                                         count:vertex_ptrs.size()];
+      geomDescMotion.vertexStride = sizeof(verts.data()[0]);
+      geomDescMotion.indexBuffer = indexBuf;
+      geomDescMotion.indexBufferOffset = 0;
+      geomDescMotion.indexType = MTLIndexTypeUInt32;
+      geomDescMotion.triangleCount = num_indices / 3;
+      geomDescMotion.intersectionFunctionTableOffset = 0;
+
+      geomDesc = geomDescMotion;
+    }
+    else {
+      MTLAccelerationStructureTriangleGeometryDescriptor *geomDescNoMotion =
+          [MTLAccelerationStructureTriangleGeometryDescriptor descriptor];
+      geomDescNoMotion.vertexBuffer = posBuf;
+      geomDescNoMotion.vertexBufferOffset = 0;
+      geomDescNoMotion.vertexStride = sizeof(verts.data()[0]);
+      geomDescNoMotion.indexBuffer = indexBuf;
+      geomDescNoMotion.indexBufferOffset = 0;
+      geomDescNoMotion.indexType = MTLIndexTypeUInt32;
+      geomDescNoMotion.triangleCount = num_indices / 3;
+      geomDescNoMotion.intersectionFunctionTableOffset = 0;
+
+      geomDesc = geomDescNoMotion;
+    }
+
+    /* Force a single any-hit call, so shadow record-all behavior works correctly */
+    /* (Match optix behavior: unsigned int build_flags =
+     * OPTIX_GEOMETRY_FLAG_REQUIRE_SINGLE_ANYHIT_CALL;) */
+    geomDesc.allowDuplicateIntersectionFunctionInvocation = false;
+
+    MTLPrimitiveAccelerationStructureDescriptor *accelDesc =
+        [MTLPrimitiveAccelerationStructureDescriptor descriptor];
+    accelDesc.geometryDescriptors = @[ geomDesc ];
+    if (num_motion_steps > 1) {
+      accelDesc.motionStartTime = 0.0f;
+      accelDesc.motionEndTime = 1.0f;
+      accelDesc.motionStartBorderMode = MTLMotionBorderModeClamp;
+      accelDesc.motionEndBorderMode = MTLMotionBorderModeClamp;
+      accelDesc.motionKeyframeCount = num_motion_steps;
+    }
+
+    if (!use_fast_trace_bvh) {
+      accelDesc.usage |= (MTLAccelerationStructureUsageRefit |
+                          MTLAccelerationStructureUsagePreferFastBuild);
+    }
+
+    MTLAccelerationStructureSizes accelSizes = [device
+        accelerationStructureSizesWithDescriptor:accelDesc];
+    id<MTLAccelerationStructure> accel_uncompressed = [device
+        newAccelerationStructureWithSize:accelSizes.accelerationStructureSize];
+    id<MTLBuffer> scratchBuf = [device newBufferWithLength:accelSizes.buildScratchBufferSize
+                                                   options:MTLResourceStorageModePrivate];
+    id<MTLBuffer> sizeBuf = [device newBufferWithLength:8 options:MTLResourceStorageModeShared];
+    id<MTLCommandBuffer> accelCommands = [queue commandBuffer];
+    id<MTLAccelerationStructureCommandEncoder> accelEnc =
+        [accelCommands accelerationStructureCommandEncoder];
+    if (refit) {
+      [accelEnc refitAccelerationStructure:accel_struct
+                                descriptor:accelDesc
+                               destination:accel_uncompressed
+                             scratchBuffer:scratchBuf
+                       scratchBufferOffset:0];
+    }
+    else {
+      [accelEnc buildAccelerationStructure:accel_uncompressed
+                                descriptor:accelDesc
+                             scratchBuffer:scratchBuf
+                       scratchBufferOffset:0];
+    }
+    if (use_fast_trace_bvh) {
+      [accelEnc writeCompactedAccelerationStructureSize:accel_uncompressed
+                                               toBuffer:sizeBuf
+                                                 offset:0
+                                           sizeDataType:MTLDataTypeULong];
+    }
+    [accelEnc endEncoding];
+    [accelCommands addCompletedHandler:^(id<MTLCommandBuffer> command_buffer) {
+      /* free temp resources */
+      [scratchBuf release];
+      [indexBuf release];
+      [posBuf release];
+
+      if (use_fast_trace_bvh) {
+        /* Compact the accel structure */
+        uint64_t compressed_size = *(uint64_t *)sizeBuf.contents;
+
+        dispatch_async(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), ^{
+          id<MTLCommandBuffer> accelCommands = [queue commandBuffer];
+          id<MTLAccelerationStructureCommandEncoder> accelEnc =
+              [accelCommands accelerationStructureCommandEncoder];
+          id<MTLAccelerationStructure> accel = [device
+              newAccelerationStructureWithSize:compressed_size];
+          [accelEnc copyAndCompactAccelerationStructure:accel_uncompressed
+                                toAccelerationStructure:accel];
+          [accelEnc endEncoding];
+          [accelCommands addCompletedHandler:^(id<MTLCommandBuffer> command_buffer) {
+            uint64_t allocated_size = [accel allocatedSize];
+            stats.mem_alloc(allocated_size);
+            accel_struct = accel;
+            [accel_uncompressed release];
+            accel_struct_building = false;
+          }];
+          [accelCommands commit];
+        });
+      }
+      else {
+        /* set our acceleration structure to the uncompressed structure */
+        accel_struct = accel_uncompressed;
+
+        uint64_t allocated_size = [accel_struct allocatedSize];
+        stats.mem_alloc(allocated_size);
+        accel_struct_building = false;
+      }
+      [sizeBuf release];
+    }];
+
+    accel_struct_building = true;
+    [accelCommands commit];
+
+    return true;
+  }
+  return false;
+}
+
+bool BVHMetal::build_BLAS_hair(Progress &progress,
+                               id<MTLDevice> device,
+                               id<MTLCommandQueue> queue,
+                               Geometry *const geom,
+                               bool refit)
+{
+  if (@available(macos 12.0, *)) {
+    /* Build BLAS for hair curves */
+    Hair *hair = static_cast<Hair *>(geom);
+    if (hair->num_curves() == 0) {
+      return false;
+    }
+
+    /*------------------------------------------------*/
+    BVH_status(
+        "Building hair BLAS | %7d curves | %s", (int)hair->num_curves(), geom->name.c_str());
+    /*------------------------------------------------*/
+
+    const bool use_fast_trace_bvh = (params.bvh_type == BVH_TYPE_STATIC);
+    const size_t num_segments = hair->num_segments();
+
+    size_t num_motion_steps = 1;
+    Attribute *motion_keys = hair->attributes.find(ATTR_STD_MOTION_VERTEX_POSITION);
+    if (motion_blur && hair->get_use_motion_blur() && motion_keys) {
+      num_motion_steps = hair->get_motion_steps();
+    }
+
+    const size_t num_aabbs = num_segments * num_motion_steps;
+
+    MTLResourceOptions storage_mode;
+    if (device.hasUnifiedMemory) {
+      storage_mode = MTLResourceStorageModeShared;
+    }
+    else {
+      storage_mode = MTLResourceStorageModeManaged;
+    }
+
+    /* Allocate a GPU buffer for the AABB data and populate it */
+    id<MTLBuffer> aabbBuf = [device
+        newBufferWithLength:num_aabbs * sizeof(MTLAxisAlignedBoundingBox)
+                    options:storage_mode];
+    MTLAxisAlignedBoundingBox *aabb_data = (MTLAxisAlignedBoundingBox *)[aabbBuf contents];
+
+    /* Get AABBs for each motion step */
+    size_t center_step = (num_motion_steps - 1) / 2;
+    for (size_t step = 0; step < num_motion_steps; ++step) {
+      /* The center step for motion vertices is not stored in the attribute */
+      const float3 *keys = hair->get_curve_keys().data();
+      if (step != center_step) {
+        size_t attr_offset = (step > center_step) ? step - 1 : step;
+        /* Technically this is a float4 array, but sizeof(float3) == sizeof(float4) */
+        keys = motion_keys->data_float3() + attr_offset * hair->get_curve_keys().size();
+      }
+
+      for (size_t j = 0, i = 0; j < hair->num_curves(); ++j) {
+        const Hair::Curve curve = hair->get_curve(j);
+
+        for (int segment = 0; segment < curve.num_segments(); ++segment, ++i) {
+          {
+            BoundBox bounds = BoundBox::empty;
+            curve.bounds_grow(segment, keys, hair->get_curve_radius().data(), bounds);
+
+            const size_t index = step * num_segments + i;
+            aabb_data[index].min = (MTLPackedFloat3 &)bounds.min;
+            aabb_data[index].max = (MTLPackedFloat3 &)bounds.max;
+          }
+        }
+      }
+    }
+
+    if (storage_mode == MTLResourceStorageModeManaged) {
+      [aabbBuf didModifyRange:NSMakeRange(0, aabbBuf.length)];
+    }
+
+#  if 0
+    for (size_t i=0; i<num_aabbs && i < 400; i++) {
+      MTLAxisAlignedBoundingBox& bb = aabb_data[i];
+      printf("  %d:   %.1f,%.1f,%.1f -- %.1f,%.1f,%.1f\n", int(i), bb.min.x, bb.min.y, bb.min.z, bb.max.x, bb.max.y, bb.max.z);
+    }
+#  endif
+
+    MTLAccelerationStructureGeometryDescriptor *geomDesc;
+    if (motion_blur) {
+      std::vector<MTLMotionKeyframeData *> aabb_ptrs;
+      aabb_ptrs.reserve(num_motion_steps);
+      for (size_t step = 0; step < num_motion_steps; ++step) {
+        MTLMotionKeyframeData *k = [MTLMotionKeyframeData data];
+        k.buffer = aabbBuf;
+        k.offset = step * num_segments * sizeof(MTLAxisAlignedBoundingBox);
+        aabb_ptrs.push_back(k);
+      }
+
+      MTLAccelerationStructureMotionBoundingBoxGeometryDescriptor *geomDescMotion =
+          [MTLAccelerationStructureMotionBoundingBoxGeometryDescriptor descriptor];
+      geomDescMotion.boundingBoxBuffers = [NSArray arrayWithObjects:aabb_ptrs.data()
+                                                              count:aabb_ptrs.size()];
+      geomDescMotion.boundingBoxCount = num_segments;
+      geomDescMotion.boundingBoxStride = sizeof(aabb_data[0]);
+      geomDescMotion.intersectionFunctionTableOffset = 1;
+
+      /* Force a single any-hit call, so shadow record-all behavior works correctly */
+      /* (Match optix behavior: unsigned int build_flags =
+       * OPTIX_GEOMETRY_FLAG_REQUIRE_SINGLE_ANYHIT_CALL;) */
+      geomDescMotion.allowDuplicateIntersectionFunctionInvocation = false;
+      geomDescMotion.opaque = true;
+      geomDesc = geomDescMotion;
+    }
+    else {
+      MTLAccelerationStructureBoundingBoxGeometryDescriptor *geomDescNoMotion =
+          [MTLAccelerationStructureBoundingBoxGeometryDescriptor descriptor];
+      geomDescNoMotion.boundingBoxBuffer = aabbBuf;
+      geomDescNoMotion.boundingBoxBufferOffset = 0;
+      geomDescNoMotion.boundingBoxCount = int(num_aabbs);
+      geomDescNoMotion.boundingBoxStride = sizeof(aabb_data[0]);
+      geomDescNoMotion.intersectionFunctionTableOffset = 1;
+
+      /* Force a single any-hit call, so shadow record-all behavior works correctly */
+      /* (Match optix behavior: unsigned int build_flags =
+       * OPTIX_GEOMETRY_FLAG_REQUIRE_SINGLE_ANYHIT_CALL;) */
+      geomDescNoMotion.allowDuplicateIntersectionFunctionInvocation = false;
+      geomDescNoMotion.opaque = true;
+      geomDesc = geomDescNoMotion;
+    }
+
+    MTLPrimitiveAccelerationStructureDescriptor *accelDesc =
+        [MTLPrimitiveAccelerationStructureDescriptor descriptor];
+    accelDesc.geometryDescriptors = @[ geomDesc ];
+
+    if (motion_blur) {
+      accelDesc.motionStartTime = 0.0f;
+      accelDesc.motionEndTime = 1.0f;
+      accelDesc.motionStartBorderMode = MTLMotionBorderModeVanish;
+      accelDesc.motionEndBorderMode = MTLMotionBorderModeVanish;
+      accelDesc.motionKeyframeCount = num_motion_steps;
+    }
+
+    if (!use_fast_trace_bvh) {
+      accelDesc.usage |= (MTLAccelerationStructureUsageRefit |
+                          MTLAccelerationStructureUsagePreferFastBuild);
+    }
+
+    MTLAccelerationStructureSizes accelSizes = [device
+        accelerationStructureSizesWithDescriptor:accelDesc];
+    id<MTLAccelerationStructure> accel_uncompressed = [device
+        newAccelerationStructureWithSize:accelSizes.accelerationStructureSize];
+    id<MTLBuffer> scratchBuf = [device newBufferWithLength:accelSizes.buildScratchBufferSize
+                                                   options:MTLResourceStorageModePrivate];
+    id<MTLBuffer> sizeBuf = [device newBufferWithLength:8 options:MTLResourceStorageModeShared];
+    id<MTLCommandBuffer> accelCommands = [queue commandBuffer];
+    id<MTLAccelerationStructureCommandEncoder> accelEnc =
+        [accelCommands accelerationStructureCommandEncoder];
+    if (refit) {
+      [accelEnc refitAccelerationStructure:accel_struct
+                                descriptor:accelDesc
+                               destination:accel_uncompressed
+                             scratchBuffer:scratchBuf
+                       scratchBufferOffset:0];
+    }
+    else {
+      [accelEnc buildAccelerationStructure:accel_uncompressed
+                                descriptor:accelDesc
+                             scratchBuffer:scratchBuf
+                       scratchBufferOffset:0];
+    }
+    if (use_fast_trace_bvh) {
+      [accelEnc writeCompactedAccelerationStructureSize:accel_uncompressed
+                                               toBuffer:sizeBuf
+                                                 offset:0
+                                           sizeDataType:MTLDataTypeULong];
+    }
+    [accelEnc endEncoding];
+    [accelCommands addCompletedHandler:^(id<MTLCommandBuffer> command_buffer) {
+      /* free temp resources */
+      [scratchBuf release];
+      [aabbBuf release];
+
+      if (use_fast_trace_bvh) {
+        /* Compact the accel structure */
+        uint64_t compressed_size = *(uint64_t *)sizeBuf.contents;
+
+        dispatch_async(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), ^{
+          id<MTLCommandBuffer> accelCommands = [queue commandBuffer];
+          id<MTLAccelerationStructureCommandEncoder> accelEnc =
+              [accelCommands accelerationStructureCommandEncoder];
+          id<MTLAccelerationStructure> accel = [device
+              newAccelerationStructureWithSize:compressed_size];
+          [accelEnc copyAndCompactAccelerationStructure:accel_uncompressed
+                                toAccelerationStructure:accel];
+          [accelEnc endEncoding];
+          [accelCommands addCompletedHandler:^(id<MTLCommandBuffer> command_buffer) {
+            uint64_t allocated_size = [accel allocatedSize];
+            stats.mem_alloc(allocated_size);
+            accel_struct = accel;
+            [accel_uncompressed release];
+            accel_struct_building = false;
+          }];
+          [accelCommands commit];
+        });
+      }
+      else {
+        /* set our acceleration structure to the uncompressed structure */
+        accel_struct = accel_uncompressed;
+
+        uint64_t allocated_size = [accel_struct allocatedSize];
+        stats.mem_alloc(allocated_size);
+        accel_struct_building = false;
+      }
+      [sizeBuf release];
+    }];
+
+    accel_struct_building = true;
+    [accelCommands commit];
+    return true;
+  }
+  return false;
+}
+
+bool BVHMetal::build_BLAS(Progress &progress,
+                          id<MTLDevice> device,
+                          id<MTLCommandQueue> queue,
+                          bool refit)
+{
+  if (@available(macos 12.0, *)) {
+    assert(objects.size() == 1 && geometry.size() == 1);
+
+    /* Build bottom level acceleration structures (BLAS) */
+    Geometry *const geom = geometry[0];
+    switch (geom->geometry_type) {
+      case Geometry::VOLUME:
+      case Geometry::MESH:
+        return build_BLAS_mesh(progress, device, queue, geom, refit);
+      case Geometry::HAIR:
+        return build_BLAS_hair(progress, device, queue, geom, refit);
+      default:
+        return false;
+    }
+  }
+  return false;
+}
+
+bool BVHMetal::build_TLAS(Progress &progress,
+                          id<MTLDevice> device,
+                          id<MTLCommandQueue> queue,
+                          bool refit)
+{
+  if (@available(macos 12.0, *)) {
+
+    /* we need to sync here and ensure that all BLAS have completed async generation by both GCD
+     * and Metal */
+    {
+      __block bool complete_bvh = false;
+      while (!complete_bvh) {
+        dispatch_sync(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), ^{
+          complete_bvh = true;
+          for (Object *ob : objects) {
+            /* Skip non-traceable objects */
+            if (!ob->is_traceable())
+              continue;
+
+            Geometry const *geom = ob->get_geometry();
+            BVHMetal const *blas = static_cast<BVHMetal const *>(geom->bvh);
+            if (blas->accel_struct_building) {
+              complete_bvh = false;
+
+              /* We're likely waiting on a command buffer that's in flight to complete.
+               * Queue up a command buffer and wait for it complete before checking the BLAS again
+               */
+              id<MTLCommandBuffer> command_buffer = [queue commandBuffer];
+              [command_buffer commit];
+              [command_buffer waitUntilCompleted];
+              break;
+            }
+          }
+        });
+      }
+    }
+
+    uint32_t num_instances = 0;
+    uint32_t num_motion_transforms = 0;
+    for (Object *ob : objects) {
+      /* Skip non-traceable objects */
+      if (!ob->is_traceable())
+        continue;
+      num_instances++;
+
+      if (ob->use_motion()) {
+        num_motion_transforms += max(1, ob->get_motion().size());
+      }
+      else {
+        num_motion_transforms++;
+      }
+    }
+
+    /*------------------------------------------------*/
+    BVH_status("Building TLAS      | %7d instances", (int)num_instances);
+    /*------------------------------------------------*/
+
+    const bool use_fast_trace_bvh = (params.bvh_type == BVH_TYPE_STATIC);
+
+    NSMutableArray *all_blas = [NSMutableArray array];
+    unordered_map<BVHMetal const *, int> instance_mapping;
+
+    /* Lambda function to build/retrieve the BLAS index mapping */
+    auto get_blas_index = [&](BVHMetal const *blas) {
+      auto it = instance_mapping.find(blas);
+      if (it != instance_mapping.end()) {
+        return it->second;
+      }
+      else {
+        int blas_index = (int)[all_blas count];
+        instance_mapping[blas] = blas_index;
+        if (@available(macos 12.0, *)) {
+          [all_blas addObject:blas->accel_struct];
+        }
+        return blas_index;
+      }
+    };
+
+    MTLResourceOptions storage_mode;
+    if (device.hasUnifiedMemory) {
+      storage_mode = MTLResourceStorageModeShared;
+    }
+    else {
+      storage_mode = MTLResourceStorageModeManaged;
+    }
+
+    size_t instance_size;
+    if (motion_blur) {
+      instance_size = sizeof(MTLAccelerationStructureMotionInstanceDescriptor);
+    }
+    else {
+      instance_size = sizeof(MTLAccelerationStructureUserIDInstanceDescriptor);
+    }
+
+    /* Allocate a GPU buffer for the instance data and populate it */
+    id<MTLBuffer> instanceBuf = [device newBufferWithLength:num_instances * instance_size
+                                                    options:storage_mode];
+    id<MTLBuffer> motion_transforms_buf = nil;
+    MTLPackedFloat4x3 *motion_transforms = nullptr;
+    if (motion_blur && num_motion_transforms) {
+      motion_transforms_buf = [device
+          newBufferWithLength:num_motion_transforms * sizeof(MTLPackedFloat4x3)
+                      options:storage_mode];
+      motion_transforms = (MTLPackedFloat4x3 *)motion_transforms_buf.contents;
+    }
+
+    uint32_t instance_index = 0;
+    uint32_t motion_transform_index = 0;
+    for (Object *ob : objects) {
+      /* Skip non-traceable objects */
+      if (!ob->is_traceable())
+        continue;
+
+      Geometry const *geom = ob->get_geometry();
+
+      BVHMetal const *blas = static_cast<BVHMetal const *>(geom->bvh);
+      uint32_t accel_struct_index = get_blas_index(blas);
+
+      /* Add some of the object visibility bits to the mask.
+       * __prim_visibility contains the combined visibility bits of all instances, so is not
+       * reliable if they differ between instances.
+       *
+       * METAL_WIP: OptiX visibility mask can only contain 8 bits, so have to trade-off here
+       * and select just a few important ones.
+       */
+      uint32_t mask = ob->visibility_for_tracing() & 0xFF;
+
+      /* Have to have at least one bit in the mask, or else instance would always be culled. */
+      if (0 == mask) {
+        mask = 0xFF;
+      }
+
+      /* Set user instance ID to object index */
+      int object_index = ob->get_device_index();
+      uint32_t user_id = uint32_t(object_index);
+
+      /* Bake into the appropriate descriptor */
+      if (motion_blur) {
+        MTLAccelerationStructureMotionInstanceDescriptor *instances =
+            (MTLAccelerationStructureMotionInstanceDescriptor *)[instanceBuf contents];
+        MTLAccelerationStructureMotionInstanceDescriptor &desc = instances[instance_index++];
+
+        desc.accelerationStructureIndex = accel_struct_index;
+        desc.userID = user_id;
+        desc.mask = mask;
+        desc.motionStartTime = 0.0f;
+        desc.motionEndTime = 1.0f;
+        desc.motionTransformsStartIndex = motion_transform_index;
+        desc.motionStartBorderMode = MTLMotionBorderModeVanish;
+        desc.motionEndBorderMode = MTLMotionBorderModeVanish;
+        desc.intersectionFunctionTableOffset = 0;
+
+        int key_count = ob->get_motion().size();
+        if (key_count) {
+          desc.motionTransformsCount = key_count;
+
+          Transform *keys = ob->get_motion().data();
+          for (int i = 0; i < key_count; i++) {
+            float *t = (float *)&motion_transforms[motion_transform_index++];
+            /* Transpose transform */
+            auto src = (float const *)&keys[i];
+            for (int i = 0; i < 12; i++) {
+              t[i] = src[(i / 3) + 4 * (i % 3)];
+            }
+          }
+        }
+        else {
+          desc.motionTransformsCount = 1;
+
+          float *t = (float *)&motion_transforms[motion_transform_index++];
+          if (ob->get_geometry()->is_instanced()) {
+            /* Transpose transform */
+            auto src = (float const *)&ob->get_tfm();
+            for (int i = 0; i < 12; i++) {
+              t[i] = src[(i / 3) + 4 * (i % 3)];
+            }
+          }
+          else {
+            /* Clear transform to identity matrix */
+            t[0] = t[4] = t[8] = 1.0f;
+          }
+        }
+      }
+      else {
+        MTLAccelerationStructureUserIDInstanceDescriptor *instances =
+            (MTLAccelerationStructureUserIDInstanceDescriptor *)[instanceBuf contents];
+        MTLAccelerationStructureUserIDInstanceDescriptor &desc = instances[instance_index++];
+
+        desc.accelerationStructureIndex = accel_struct_index;
+        desc.userID = user_id;
+        desc.mask = mask;
+        desc.intersectionFunctionTableOffset = 0;
+
+        float *t = (float *)&desc.transformationMatrix;
+        if (ob->get_geometry()->is_instanced()) {
+          /* Transpose transform */
+          auto src = (float const *)&ob->get_tfm();
+          for (int i = 0; i < 12; i++) {
+            t[i] = src[(i / 3) + 4 * (i % 3)];
+          }
+        }
+        else {
+          /* Clear transform to identity matrix */
+          t[0] = t[4] = t[8] = 1.0f;
+        }
+      }
+    }
+
+    if (storage_mode == MTLResourceStorageModeManaged) {
+      [instanceBuf didModifyRange:NSMakeRange(0, instanceBuf.length)];
+      if (motion_transforms_buf) {
+        [motion_transforms_buf didModifyRange:NSMakeRange(0, motion_transforms_buf.length)];
+        assert(num_motion_transforms == motion_transform_index);
+      }
+    }
+
+    MTLInstanceAccelerationStructureDescriptor *accelDesc =
+        [MTLInstanceAccelerationStructureDescriptor descriptor];
+    accelDesc.instanceCount = num_instances;
+    accelDesc.instanceDescriptorType = MTLAccelerationStructureInstanceDescriptorTypeUserID;
+    accelDesc.instanceDescriptorBuffer = instanceBuf;
+    accelDesc.instanceDescriptorBufferOffset = 0;
+    accelDesc.instanceDescriptorStride = instance_size;
+    accelDesc.instancedAccelerationStructures = all_blas;
+
+    if (motion_blur) {
+      accelDesc.instanceDescriptorType = MTLAccelerationStructureInstanceDescriptorTypeMotion;
+      accelDesc.motionTransformBuffer = motion_transforms_buf;
+      accelDesc.motionTransformCount = num_motion_transforms;
+    }
+
+    if (!use_fast_trace_bvh) {
+      accelDesc.usage |= (MTLAccelerationStructureUsageRefit |
+                          MTLAccelerationStructureUsagePreferFastBuild);
+    }
+
+    MTLAccelerationStructureSizes accelSizes = [device
+        accelerationStructureSizesWithDescriptor:accelDesc];
+    id<MTLAccelerationStructure> accel = [device
+        newAccelerationStructureWithSize:accelSizes.accelerationStructureSize];
+    id<MTLBuffer> scratchBuf = [device newBufferWithLength:accelSizes.buildScratchBufferSize
+                                                   options:MTLResourceStorageModePrivate];
+    id<MTLCommandBuffer> accelCommands = [queue commandBuffer];
+    id<MTLAccelerationStructureCommandEncoder> accelEnc =
+        [accelCommands accelerationStructureCommandEncoder];
+    if (refit) {
+      [accelEnc refitAccelerationStructure:accel_struct
+                                descriptor:accelDesc
+                               destination:accel
+                             scratchBuffer:scratchBuf
+                       scratchBufferOffset:0];
+    }
+    else {
+      [accelEnc buildAccelerationStructure:accel
+                                descriptor:accelDesc
+                             scratchBuffer:scratchBuf
+                       scratchBufferOffset:0];
+    }
+    [accelEnc endEncoding];
+    [accelCommands commit];
+    [accelCommands waitUntilCompleted];
+
+    if (motion_transforms_buf) {
+      [motion_transforms_buf release];
+    }
+    [instanceBuf release];
+    [scratchBuf release];
+
+    uint64_t allocated_size = [accel allocatedSize];
+    stats.mem_alloc(allocated_size);
+
+    /* Cache top and bottom-level acceleration structs */
+    accel_struct = accel;
+    blas_array.clear();
+    blas_array.reserve(all_blas.count);
+    for (id<MTLAccelerationStructure> blas in all_blas) {
+      blas_array.push_back(blas);
+    }
+
+    return true;
+  }
+  return false;
+}
+
+bool BVHMetal::build(Progress &progress,
+                     id<MTLDevice> device,
+                     id<MTLCommandQueue> queue,
+                     bool refit)
+{
+  if (@available(macos 12.0, *)) {
+    if (refit && params.bvh_type != BVH_TYPE_STATIC) {
+      assert(accel_struct);
+    }
+    else {
+      if (accel_struct) {
+        stats.mem_free(accel_struct.allocatedSize);
+        [accel_struct release];
+        accel_struct = nil;
+      }
+    }
+  }
+
+  if (!params.top_level) {
+    return build_BLAS(progress, device, queue, refit);
+  }
+  else {
+    return build_TLAS(progress, device, queue, refit);
+  }
+}
+
+CCL_NAMESPACE_END
+
+#endif /* WITH_METAL */

intern/cycles/device/metal/device.h

@@ -0,0 +1,37 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#include "util/string.h"
+#include "util/vector.h"
+
+CCL_NAMESPACE_BEGIN
+
+class Device;
+class DeviceInfo;
+class Profiler;
+class Stats;
+
+bool device_metal_init();
+
+Device *device_metal_create(const DeviceInfo &info, Stats &stats, Profiler &profiler);
+
+void device_metal_info(vector<DeviceInfo> &devices);
+
+string device_metal_capabilities();
+
+CCL_NAMESPACE_END

intern/cycles/device/metal/device.mm

@@ -0,0 +1,136 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifdef WITH_METAL
+
+#  include "device/metal/device.h"
+#  include "device/metal/device_impl.h"
+
+#endif
+
+#include "util/debug.h"
+#include "util/set.h"
+#include "util/system.h"
+
+CCL_NAMESPACE_BEGIN
+
+#ifdef WITH_METAL
+
+Device *device_metal_create(const DeviceInfo &info, Stats &stats, Profiler &profiler)
+{
+  return new MetalDevice(info, stats, profiler);
+}
+
+bool device_metal_init()
+{
+  return true;
+}
+
+static int device_metal_get_num_devices_safe(uint32_t *num_devices)
+{
+  *num_devices = MTLCopyAllDevices().count;
+  return 0;
+}
+
+void device_metal_info(vector<DeviceInfo> &devices)
+{
+  uint32_t num_devices = 0;
+  device_metal_get_num_devices_safe(&num_devices);
+  if (num_devices == 0) {
+    return;
+  }
+
+  vector<MetalPlatformDevice> usable_devices;
+  MetalInfo::get_usable_devices(&usable_devices);
+  /* Devices are numbered consecutively across platforms. */
+  set<string> unique_ids;
+  int device_index = 0;
+  for (MetalPlatformDevice &device : usable_devices) {
+    /* Compute unique ID for persistent user preferences. */
+    const string &device_name = device.device_name;
+    string id = string("METAL_") + device_name;
+
+    /* Hardware ID might not be unique, add device number in that case. */
+    if (unique_ids.find(id) != unique_ids.end()) {
+      id += string_printf("_ID_%d", num_devices);
+    }
+    unique_ids.insert(id);
+
+    /* Create DeviceInfo. */
+    DeviceInfo info;
+    info.type = DEVICE_METAL;
+    info.description = string_remove_trademark(string(device_name));
+
+    /* Ensure unique naming on Apple Silicon / SoC devices which return the same string for CPU and
+     * GPU */
+    if (info.description == system_cpu_brand_string()) {
+      info.description += " (GPU)";
+    }
+
+    info.num = device_index;
+    /* We don't know if it's used for display, but assume it is. */
+    info.display_device = true;
+    info.denoisers = DENOISER_NONE;
+    info.id = id;
+
+    devices.push_back(info);
+    device_index++;
+  }
+}
+
+string device_metal_capabilities()
+{
+  string result = "";
+  string error_msg = "";
+  uint32_t num_devices = 0;
+  assert(device_metal_get_num_devices_safe(&num_devices));
+  if (num_devices == 0) {
+    return "No Metal devices found\n";
+  }
+  result += string_printf("Number of devices: %u\n", num_devices);
+
+  NSArray<id<MTLDevice>> *allDevices = MTLCopyAllDevices();
+  for (id<MTLDevice> device in allDevices) {
+    result += string_printf("\t\tDevice: %s\n", [device.name UTF8String]);
+  }
+
+  return result;
+}
+
+#else
+
+Device *device_metal_create(const DeviceInfo &info, Stats &stats, Profiler &profiler)
+{
+  return nullptr;
+}
+
+bool device_metal_init()
+{
+  return false;
+}
+
+void device_metal_info(vector<DeviceInfo> &devices)
+{
+}
+
+string device_metal_capabilities()
+{
+  return "";
+}
+
+#endif
+
+CCL_NAMESPACE_END

intern/cycles/device/metal/device_impl.h

@@ -0,0 +1,166 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#ifdef WITH_METAL
+
+#  include "bvh/bvh.h"
+#  include "device/device.h"
+#  include "device/metal/bvh.h"
+#  include "device/metal/device.h"
+#  include "device/metal/kernel.h"
+#  include "device/metal/queue.h"
+#  include "device/metal/util.h"
+
+#  include <Metal/Metal.h>
+
+CCL_NAMESPACE_BEGIN
+
+class DeviceQueue;
+
+class MetalDevice : public Device {
+ public:
+  id<MTLDevice> mtlDevice = nil;
+  id<MTLLibrary> mtlLibrary[PSO_NUM] = {nil};
+  id<MTLArgumentEncoder> mtlBufferKernelParamsEncoder =
+      nil; /* encoder used for fetching device pointers from MTLBuffers */
+  id<MTLCommandQueue> mtlGeneralCommandQueue = nil;
+  id<MTLArgumentEncoder> mtlAncillaryArgEncoder =
+      nil; /* encoder used for fetching device pointers from MTLBuffers */
+  string source_used_for_compile[PSO_NUM];
+
+  KernelParamsMetal launch_params = {0};
+
+  /* MetalRT members ----------------------------------*/
+  BVHMetal *bvhMetalRT = nullptr;
+  bool motion_blur = false;
+  id<MTLArgumentEncoder> mtlASArgEncoder =
+      nil; /* encoder used for fetching device pointers from MTLAccelerationStructure */
+  /*---------------------------------------------------*/
+
+  string device_name;
+  MetalGPUVendor device_vendor;
+
+  uint kernel_features;
+  MTLResourceOptions default_storage_mode;
+  int max_threads_per_threadgroup;
+
+  int mtlDevId = 0;
+  bool first_error = true;
+
+  struct MetalMem {
+    device_memory *mem = nullptr;
+    int pointer_index = -1;
+    id<MTLBuffer> mtlBuffer = nil;
+    id<MTLTexture> mtlTexture = nil;
+    uint64_t offset = 0;
+    uint64_t size = 0;
+    void *hostPtr = nullptr;
+    bool use_UMA = false; /* If true, UMA memory in shared_pointer is being used. */
+  };
+  typedef map<device_memory *, unique_ptr<MetalMem>> MetalMemMap;
+  MetalMemMap metal_mem_map;
+  std::vector<id<MTLResource>> delayed_free_list;
+  std::recursive_mutex metal_mem_map_mutex;
+
+  /* Bindless Textures */
+  device_vector<TextureInfo> texture_info;
+  bool need_texture_info;
+  id<MTLArgumentEncoder> mtlTextureArgEncoder = nil;
+  id<MTLBuffer> texture_bindings_2d = nil;
+  id<MTLBuffer> texture_bindings_3d = nil;
+  std::vector<id<MTLTexture>> texture_slot_map;
+
+  MetalDeviceKernels kernels;
+  bool use_metalrt = false;
+  bool use_function_specialisation = false;
+
+  virtual BVHLayoutMask get_bvh_layout_mask() const override;
+
+  void set_error(const string &error) override;
+
+  MetalDevice(const DeviceInfo &info, Stats &stats, Profiler &profiler);
+
+  virtual ~MetalDevice();
+
+  bool support_device(const uint /*kernel_features*/);
+
+  bool check_peer_access(Device *peer_device) override;
+
+  bool use_adaptive_compilation();
+
+  string get_source(const uint kernel_features);
+
+  string compile_kernel(const uint kernel_features, const char *name);
+
+  virtual bool load_kernels(const uint kernel_features) override;
+
+  void reserve_local_memory(const uint kernel_features);
+
+  void init_host_memory();
+
+  void load_texture_info();
+
+  virtual bool should_use_graphics_interop() override;
+
+  virtual unique_ptr<DeviceQueue> gpu_queue_create() override;
+
+  virtual void build_bvh(BVH *bvh, Progress &progress, bool refit) override;
+
+  /* ------------------------------------------------------------------ */
+  /* low-level memory management */
+
+  MetalMem *generic_alloc(device_memory &mem);
+
+  void generic_copy_to(device_memory &mem);
+
+  void generic_free(device_memory &mem);
+
+  void mem_alloc(device_memory &mem) override;
+
+  void mem_copy_to(device_memory &mem) override;
+
+  void mem_copy_from(device_memory &mem)
+  {
+    mem_copy_from(mem, -1, -1, -1, -1);
+  }
+  void mem_copy_from(device_memory &mem, size_t y, size_t w, size_t h, size_t elem) override;
+
+  void mem_zero(device_memory &mem) override;
+
+  void mem_free(device_memory &mem) override;
+
+  device_ptr mem_alloc_sub_ptr(device_memory &mem, size_t offset, size_t /*size*/) override;
+
+  virtual void const_copy_to(const char *name, void *host, size_t size) override;
+
+  void global_alloc(device_memory &mem);
+
+  void global_free(device_memory &mem);
+
+  void tex_alloc(device_texture &mem);
+
+  void tex_alloc_as_buffer(device_texture &mem);
+
+  void tex_free(device_texture &mem);
+
+  void flush_delayed_free_list();
+};
+
+CCL_NAMESPACE_END
+
+#endif

intern/cycles/device/metal/kernel.h

@@ -0,0 +1,168 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#ifdef WITH_METAL
+
+#  include "device/kernel.h"
+#  include <Metal/Metal.h>
+
+CCL_NAMESPACE_BEGIN
+
+class MetalDevice;
+
+enum {
+  METALRT_FUNC_DEFAULT_TRI,
+  METALRT_FUNC_DEFAULT_BOX,
+  METALRT_FUNC_SHADOW_TRI,
+  METALRT_FUNC_SHADOW_BOX,
+  METALRT_FUNC_LOCAL_TRI,
+  METALRT_FUNC_LOCAL_BOX,
+  METALRT_FUNC_CURVE_RIBBON,
+  METALRT_FUNC_CURVE_RIBBON_SHADOW,
+  METALRT_FUNC_CURVE_ALL,
+  METALRT_FUNC_CURVE_ALL_SHADOW,
+  METALRT_FUNC_NUM
+};
+
+enum { METALRT_TABLE_DEFAULT, METALRT_TABLE_SHADOW, METALRT_TABLE_LOCAL, METALRT_TABLE_NUM };
+
+/* Pipeline State Object types */
+enum {
+  /* A kernel that can be used with all scenes, supporting all features.
+   * It is slow to compile, but only needs to be compiled once and is then
+   * cached for future render sessions. This allows a render to get underway
+   * on the GPU quickly.
+   */
+  PSO_GENERIC,
+
+  /* A kernel that is relatively quick to compile, but is specialized for the
+   * scene being rendered. It only contains the functionality and even baked in
+   * constants for values that means it needs to be recompiled whenever a
+   * dependent setting is changed. The render performance of this kernel is
+   * significantly faster though, and justifies the extra compile time.
+   */
+  /* METAL_WIP: This isn't used and will require more changes to enable. */
+  PSO_SPECIALISED,
+
+  PSO_NUM
+};
+
+const char *kernel_type_as_string(int kernel_type);
+
+struct MetalKernelPipeline {
+  void release()
+  {
+    if (pipeline) {
+      [pipeline release];
+      pipeline = nil;
+      if (@available(macOS 11.0, *)) {
+        for (int i = 0; i < METALRT_TABLE_NUM; i++) {
+          if (intersection_func_table[i]) {
+            [intersection_func_table[i] release];
+            intersection_func_table[i] = nil;
+          }
+        }
+      }
+    }
+    if (function) {
+      [function release];
+      function = nil;
+    }
+    if (@available(macOS 11.0, *)) {
+      for (int i = 0; i < METALRT_TABLE_NUM; i++) {
+        if (intersection_func_table[i]) {
+          [intersection_func_table[i] release];
+        }
+      }
+    }
+  }
+
+  bool loaded = false;
+  id<MTLFunction> function = nil;
+  id<MTLComputePipelineState> pipeline = nil;
+
+  API_AVAILABLE(macos(11.0))
+  id<MTLIntersectionFunctionTable> intersection_func_table[METALRT_TABLE_NUM] = {nil};
+};
+
+struct MetalKernelLoadDesc {
+  int pso_index = 0;
+  const char *function_name = nullptr;
+  int kernel_index = 0;
+  int threads_per_threadgroup = 0;
+  MTLFunctionConstantValues *constant_values = nullptr;
+  NSArray *linked_functions = nullptr;
+
+  struct IntersectorFunctions {
+    NSArray *defaults;
+    NSArray *shadow;
+    NSArray *local;
+    NSArray *operator[](int index) const
+    {
+      if (index == METALRT_TABLE_DEFAULT)
+        return defaults;
+      if (index == METALRT_TABLE_SHADOW)
+        return shadow;
+      return local;
+    }
+  } intersector_functions = {nullptr};
+};
+
+/* Metal kernel and associate occupancy information. */
+class MetalDeviceKernel {
+ public:
+  ~MetalDeviceKernel();
+
+  bool load(MetalDevice *device, MetalKernelLoadDesc const &desc, class MD5Hash const &md5);
+
+  void mark_loaded(int pso_index)
+  {
+    pso[pso_index].loaded = true;
+  }
+
+  int get_num_threads_per_block() const
+  {
+    return num_threads_per_block;
+  }
+  const MetalKernelPipeline &get_pso() const;
+
+  double load_duration = 0.0;
+
+ private:
+  MetalKernelPipeline pso[PSO_NUM];
+
+  int num_threads_per_block = 0;
+};
+
+/* Cache of Metal kernels for each DeviceKernel. */
+class MetalDeviceKernels {
+ public:
+  bool load(MetalDevice *device, int kernel_type);
+  bool available(DeviceKernel kernel) const;
+  const MetalDeviceKernel &get(DeviceKernel kernel) const;
+
+  MetalDeviceKernel kernels_[DEVICE_KERNEL_NUM];
+
+  id<MTLFunction> rt_intersection_funcs[PSO_NUM][METALRT_FUNC_NUM] = {{nil}};
+
+  string loaded_md5[PSO_NUM];
+};
+
+CCL_NAMESPACE_END
+
+#endif /* WITH_METAL */

intern/cycles/device/metal/kernel.mm

@@ -0,0 +1,525 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifdef WITH_METAL
+
+#  include "device/metal/kernel.h"
+#  include "device/metal/device_impl.h"
+#  include "util/md5.h"
+#  include "util/path.h"
+#  include "util/tbb.h"
+#  include "util/time.h"
+
+CCL_NAMESPACE_BEGIN
+
+/* limit to 2 MTLCompiler instances */
+int max_mtlcompiler_threads = 2;
+
+const char *kernel_type_as_string(int kernel_type)
+{
+  switch (kernel_type) {
+    case PSO_GENERIC:
+      return "PSO_GENERIC";
+    case PSO_SPECIALISED:
+      return "PSO_SPECIALISED";
+    default:
+      assert(0);
+  }
+  return "";
+}
+
+MetalDeviceKernel::~MetalDeviceKernel()
+{
+  for (int i = 0; i < PSO_NUM; i++) {
+    pso[i].release();
+  }
+}
+
+bool MetalDeviceKernel::load(MetalDevice *device,
+                             MetalKernelLoadDesc const &desc_in,
+                             MD5Hash const &md5)
+{
+  __block MetalKernelLoadDesc const desc(desc_in);
+  if (desc.kernel_index == DEVICE_KERNEL_INTEGRATOR_MEGAKERNEL) {
+    /* skip megakernel */
+    return true;
+  }
+
+  bool use_binary_archive = true;
+  if (getenv("CYCLES_METAL_DISABLE_BINARY_ARCHIVES")) {
+    use_binary_archive = false;
+  }
+
+  id<MTLBinaryArchive> archive = nil;
+  string metalbin_path;
+  if (use_binary_archive) {
+    NSProcessInfo *processInfo = [NSProcessInfo processInfo];
+    string osVersion = [[processInfo operatingSystemVersionString] UTF8String];
+    MD5Hash local_md5(md5);
+    local_md5.append(osVersion);
+    string metalbin_name = string(desc.function_name) + "." + local_md5.get_hex() +
+                           to_string(desc.pso_index) + ".bin";
+    metalbin_path = path_cache_get(path_join("kernels", metalbin_name));
+    path_create_directories(metalbin_path);
+
+    if (path_exists(metalbin_path) && use_binary_archive) {
+      if (@available(macOS 11.0, *)) {
+        MTLBinaryArchiveDescriptor *archiveDesc = [[MTLBinaryArchiveDescriptor alloc] init];
+        archiveDesc.url = [NSURL fileURLWithPath:@(metalbin_path.c_str())];
+        archive = [device->mtlDevice newBinaryArchiveWithDescriptor:archiveDesc error:nil];
+        [archiveDesc release];
+      }
+    }
+  }
+
+  NSString *entryPoint = [@(desc.function_name) copy];
+
+  NSError *error = NULL;
+  if (@available(macOS 11.0, *)) {
+    MTLFunctionDescriptor *func_desc = [MTLIntersectionFunctionDescriptor functionDescriptor];
+    func_desc.name = entryPoint;
+    if (desc.constant_values) {
+      func_desc.constantValues = desc.constant_values;
+    }
+    pso[desc.pso_index].function = [device->mtlLibrary[desc.pso_index]
+        newFunctionWithDescriptor:func_desc
+                            error:&error];
+  }
+  [entryPoint release];
+
+  if (pso[desc.pso_index].function == nil) {
+    NSString *err = [error localizedDescription];
+    string errors = [err UTF8String];
+
+    device->set_error(
+        string_printf("Error getting function \"%s\": %s", desc.function_name, errors.c_str()));
+    return false;
+  }
+
+  pso[desc.pso_index].function.label = [@(desc.function_name) copy];
+
+  __block MTLComputePipelineDescriptor *computePipelineStateDescriptor =
+      [[MTLComputePipelineDescriptor alloc] init];
+
+  computePipelineStateDescriptor.buffers[0].mutability = MTLMutabilityImmutable;
+  computePipelineStateDescriptor.buffers[1].mutability = MTLMutabilityImmutable;
+  computePipelineStateDescriptor.buffers[2].mutability = MTLMutabilityImmutable;
+
+  if (@available(macos 10.14, *)) {
+    computePipelineStateDescriptor.maxTotalThreadsPerThreadgroup = desc.threads_per_threadgroup;
+  }
+  computePipelineStateDescriptor.threadGroupSizeIsMultipleOfThreadExecutionWidth = true;
+
+  computePipelineStateDescriptor.computeFunction = pso[desc.pso_index].function;
+  if (@available(macOS 11.0, *)) {
+    /* Attach the additional functions to an MTLLinkedFunctions object */
+    if (desc.linked_functions) {
+      computePipelineStateDescriptor.linkedFunctions = [[MTLLinkedFunctions alloc] init];
+      computePipelineStateDescriptor.linkedFunctions.functions = desc.linked_functions;
+    }
+
+    computePipelineStateDescriptor.maxCallStackDepth = 1;
+  }
+
+  /* Create a new Compute pipeline state object */
+  MTLPipelineOption pipelineOptions = MTLPipelineOptionNone;
+
+  bool creating_new_archive = false;
+  if (@available(macOS 11.0, *)) {
+    if (use_binary_archive) {
+      if (!archive) {
+        MTLBinaryArchiveDescriptor *archiveDesc = [[MTLBinaryArchiveDescriptor alloc] init];
+        archiveDesc.url = nil;
+        archive = [device->mtlDevice newBinaryArchiveWithDescriptor:archiveDesc error:nil];
+        creating_new_archive = true;
+
+        double starttime = time_dt();
+
+        if (![archive addComputePipelineFunctionsWithDescriptor:computePipelineStateDescriptor
+                                                          error:&error]) {
+          NSString *errStr = [error localizedDescription];
+          metal_printf("Failed to add PSO to archive:\n%s\n",
+                       errStr ? [errStr UTF8String] : "nil");
+        }
+        else {
+          double duration = time_dt() - starttime;
+          metal_printf("%2d | %-55s | %7.2fs\n",
+                       desc.kernel_index,
+                       device_kernel_as_string((DeviceKernel)desc.kernel_index),
+                       duration);
+
+          if (desc.pso_index == PSO_GENERIC) {
+            this->load_duration = duration;
+          }
+        }
+      }
+      computePipelineStateDescriptor.binaryArchives = [NSArray arrayWithObjects:archive, nil];
+      pipelineOptions = MTLPipelineOptionFailOnBinaryArchiveMiss;
+    }
+  }
+
+  double starttime = time_dt();
+
+  MTLNewComputePipelineStateWithReflectionCompletionHandler completionHandler = ^(
+      id<MTLComputePipelineState> computePipelineState,
+      MTLComputePipelineReflection *reflection,
+      NSError *error) {
+    bool recreate_archive = false;
+    if (computePipelineState == nil && archive && !creating_new_archive) {
+
+      assert(0);
+
+      NSString *errStr = [error localizedDescription];
+      metal_printf(
+          "Failed to create compute pipeline state \"%s\" from archive - attempting recreation... "
+          "(error: %s)\n",
+          device_kernel_as_string((DeviceKernel)desc.kernel_index),
+          errStr ? [errStr UTF8String] : "nil");
+      computePipelineState = [device->mtlDevice
+          newComputePipelineStateWithDescriptor:computePipelineStateDescriptor
+                                        options:MTLPipelineOptionNone
+                                     reflection:nullptr
+                                          error:&error];
+      recreate_archive = true;
+    }
+
+    double duration = time_dt() - starttime;
+
+    if (computePipelineState == nil) {
+      NSString *errStr = [error localizedDescription];
+      device->set_error(string_printf("Failed to create compute pipeline state \"%s\", error: \n",
+                                      device_kernel_as_string((DeviceKernel)desc.kernel_index)) +
+                        (errStr ? [errStr UTF8String] : "nil"));
+      metal_printf("%2d | %-55s | %7.2fs | FAILED!\n",
+                   desc.kernel_index,
+                   device_kernel_as_string((DeviceKernel)desc.kernel_index),
+                   duration);
+      return;
+    }
+
+    pso[desc.pso_index].pipeline = computePipelineState;
+    num_threads_per_block = round_down(computePipelineState.maxTotalThreadsPerThreadgroup,
+                                       computePipelineState.threadExecutionWidth);
+    num_threads_per_block = std::max(num_threads_per_block,
+                                     (int)computePipelineState.threadExecutionWidth);
+
+    if (!use_binary_archive) {
+      metal_printf("%2d | %-55s | %7.2fs\n",
+                   desc.kernel_index,
+                   device_kernel_as_string((DeviceKernel)desc.kernel_index),
+                   duration);
+
+      if (desc.pso_index == PSO_GENERIC) {
+        this->load_duration = duration;
+      }
+    }
+
+    if (@available(macOS 11.0, *)) {
+      if (creating_new_archive || recreate_archive) {
+        if (![archive serializeToURL:[NSURL fileURLWithPath:@(metalbin_path.c_str())]
+                               error:&error]) {
+          metal_printf("Failed to save binary archive, error:\n%s\n",
+                       [[error localizedDescription] UTF8String]);
+        }
+      }
+    }
+
+    [computePipelineStateDescriptor release];
+    computePipelineStateDescriptor = nil;
+
+    if (device->use_metalrt && desc.linked_functions) {
+      for (int table = 0; table < METALRT_TABLE_NUM; table++) {
+        if (@available(macOS 11.0, *)) {
+          MTLIntersectionFunctionTableDescriptor *ift_desc =
+              [[MTLIntersectionFunctionTableDescriptor alloc] init];
+          ift_desc.functionCount = desc.intersector_functions[table].count;
+
+          pso[desc.pso_index].intersection_func_table[table] = [pso[desc.pso_index].pipeline
+              newIntersectionFunctionTableWithDescriptor:ift_desc];
+
+          /* Finally write the function handles into this pipeline's table */
+          for (int i = 0; i < 2; i++) {
+            id<MTLFunctionHandle> handle = [pso[desc.pso_index].pipeline
+                functionHandleWithFunction:desc.intersector_functions[table][i]];
+            [pso[desc.pso_index].intersection_func_table[table] setFunction:handle atIndex:i];
+          }
+        }
+      }
+    }
+
+    mark_loaded(desc.pso_index);
+  };
+
+  if (desc.pso_index == PSO_SPECIALISED) {
+    /* Asynchronous load */
+    dispatch_async(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), ^{
+      NSError *error;
+      id<MTLComputePipelineState> pipeline = [device->mtlDevice
+          newComputePipelineStateWithDescriptor:computePipelineStateDescriptor
+                                        options:pipelineOptions
+                                     reflection:nullptr
+                                          error:&error];
+      completionHandler(pipeline, nullptr, error);
+    });
+  }
+  else {
+    /* Block on load to ensure we continue with a valid kernel function */
+    id<MTLComputePipelineState> pipeline = [device->mtlDevice
+        newComputePipelineStateWithDescriptor:computePipelineStateDescriptor
+                                      options:pipelineOptions
+                                   reflection:nullptr
+                                        error:&error];
+    completionHandler(pipeline, nullptr, error);
+  }
+
+  return true;
+}
+
+const MetalKernelPipeline &MetalDeviceKernel::get_pso() const
+{
+  if (pso[PSO_SPECIALISED].loaded) {
+    return pso[PSO_SPECIALISED];
+  }
+
+  assert(pso[PSO_GENERIC].loaded);
+  return pso[PSO_GENERIC];
+}
+
+bool MetalDeviceKernels::load(MetalDevice *device, int kernel_type)
+{
+  bool any_error = false;
+
+  MD5Hash md5;
+
+  /* Build the function constant table */
+  MTLFunctionConstantValues *constant_values = nullptr;
+  if (kernel_type == PSO_SPECIALISED) {
+    constant_values = [MTLFunctionConstantValues new];
+
+#  define KERNEL_FILM(_type, name) \
+    [constant_values setConstantValue:&data.film.name \
+                                 type:get_MTLDataType_##_type() \
+                              atIndex:KernelData_film_##name]; \
+    md5.append((uint8_t *)&data.film.name, sizeof(data.film.name));
+
+#  define KERNEL_BACKGROUND(_type, name) \
+    [constant_values setConstantValue:&data.background.name \
+                                 type:get_MTLDataType_##_type() \
+                              atIndex:KernelData_background_##name]; \
+    md5.append((uint8_t *)&data.background.name, sizeof(data.background.name));
+
+#  define KERNEL_INTEGRATOR(_type, name) \
+    [constant_values setConstantValue:&data.integrator.name \
+                                 type:get_MTLDataType_##_type() \
+                              atIndex:KernelData_integrator_##name]; \
+    md5.append((uint8_t *)&data.integrator.name, sizeof(data.integrator.name));
+
+#  define KERNEL_BVH(_type, name) \
+    [constant_values setConstantValue:&data.bvh.name \
+                                 type:get_MTLDataType_##_type() \
+                              atIndex:KernelData_bvh_##name]; \
+    md5.append((uint8_t *)&data.bvh.name, sizeof(data.bvh.name));
+
+    /* METAL_WIP: populate constant_values based on KernelData */
+    assert(0);
+    /*
+        const KernelData &data = device->launch_params.data;
+    #    include "kernel/types/background.h"
+    #    include "kernel/types/bvh.h"
+    #    include "kernel/types/film.h"
+    #    include "kernel/types/integrator.h"
+    */
+  }
+
+  if (device->use_metalrt) {
+    if (@available(macOS 11.0, *)) {
+      /* create the id<MTLFunction> for each intersection function */
+      const char *function_names[] = {
+          "__anyhit__cycles_metalrt_visibility_test_tri",
+          "__anyhit__cycles_metalrt_visibility_test_box",
+          "__anyhit__cycles_metalrt_shadow_all_hit_tri",
+          "__anyhit__cycles_metalrt_shadow_all_hit_box",
+          "__anyhit__cycles_metalrt_local_hit_tri",
+          "__anyhit__cycles_metalrt_local_hit_box",
+          "__intersection__curve_ribbon",
+          "__intersection__curve_ribbon_shadow",
+          "__intersection__curve_all",
+          "__intersection__curve_all_shadow",
+      };
+      assert(sizeof(function_names) / sizeof(function_names[0]) == METALRT_FUNC_NUM);
+
+      MTLFunctionDescriptor *desc = [MTLIntersectionFunctionDescriptor functionDescriptor];
+      if (kernel_type == PSO_SPECIALISED) {
+        desc.constantValues = constant_values;
+      }
+      for (int i = 0; i < METALRT_FUNC_NUM; i++) {
+        const char *function_name = function_names[i];
+        desc.name = [@(function_name) copy];
+
+        NSError *error = NULL;
+        rt_intersection_funcs[kernel_type][i] = [device->mtlLibrary[kernel_type]
+            newFunctionWithDescriptor:desc
+                                error:&error];
+
+        if (rt_intersection_funcs[kernel_type][i] == nil) {
+          NSString *err = [error localizedDescription];
+          string errors = [err UTF8String];
+
+          device->set_error(string_printf(
+              "Error getting intersection function \"%s\": %s", function_name, errors.c_str()));
+          any_error = true;
+          break;
+        }
+
+        rt_intersection_funcs[kernel_type][i].label = [@(function_name) copy];
+      }
+    }
+  }
+  md5.append(device->source_used_for_compile[kernel_type]);
+
+  string hash = md5.get_hex();
+  if (loaded_md5[kernel_type] == hash) {
+    return true;
+  }
+
+  if (!any_error) {
+    NSArray *table_functions[METALRT_TABLE_NUM] = {nil};
+    NSArray *function_list = nil;
+
+    if (device->use_metalrt) {
+      id<MTLFunction> box_intersect_default = nil;
+      id<MTLFunction> box_intersect_shadow = nil;
+      if (device->kernel_features & KERNEL_FEATURE_HAIR) {
+        /* Add curve intersection programs. */
+        if (device->kernel_features & KERNEL_FEATURE_HAIR_THICK) {
+          /* Slower programs for thick hair since that also slows down ribbons.
+           * Ideally this should not be needed. */
+          box_intersect_default = rt_intersection_funcs[kernel_type][METALRT_FUNC_CURVE_ALL];
+          box_intersect_shadow = rt_intersection_funcs[kernel_type][METALRT_FUNC_CURVE_ALL_SHADOW];
+        }
+        else {
+          box_intersect_default = rt_intersection_funcs[kernel_type][METALRT_FUNC_CURVE_RIBBON];
+          box_intersect_shadow =
+              rt_intersection_funcs[kernel_type][METALRT_FUNC_CURVE_RIBBON_SHADOW];
+        }
+      }
+      table_functions[METALRT_TABLE_DEFAULT] = [NSArray
+          arrayWithObjects:rt_intersection_funcs[kernel_type][METALRT_FUNC_DEFAULT_TRI],
+                           box_intersect_default ?
+                               box_intersect_default :
+                               rt_intersection_funcs[kernel_type][METALRT_FUNC_DEFAULT_BOX],
+                           nil];
+      table_functions[METALRT_TABLE_SHADOW] = [NSArray
+          arrayWithObjects:rt_intersection_funcs[kernel_type][METALRT_FUNC_SHADOW_TRI],
+                           box_intersect_shadow ?
+                               box_intersect_shadow :
+                               rt_intersection_funcs[kernel_type][METALRT_FUNC_SHADOW_BOX],
+                           nil];
+      table_functions[METALRT_TABLE_LOCAL] = [NSArray
+          arrayWithObjects:rt_intersection_funcs[kernel_type][METALRT_FUNC_LOCAL_TRI],
+                           rt_intersection_funcs[kernel_type][METALRT_FUNC_LOCAL_BOX],
+                           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]];
+
+      function_list = [[NSArray arrayWithArray:[unique_functions allObjects]]
+          sortedArrayUsingComparator:^NSComparisonResult(id<MTLFunction> f1, id<MTLFunction> f2) {
+            return [f1.label compare:f2.label];
+          }];
+
+      unique_functions = nil;
+    }
+
+    metal_printf("Starting %s \"cycles_metal_...\" pipeline builds\n",
+                 kernel_type_as_string(kernel_type));
+
+    tbb::task_arena local_arena(max_mtlcompiler_threads);
+    local_arena.execute([&]() {
+      tbb::parallel_for(int(0), int(DEVICE_KERNEL_NUM), [&](int i) {
+        /* skip megakernel */
+        if (i == DEVICE_KERNEL_INTEGRATOR_MEGAKERNEL) {
+          return;
+        }
+
+        /* Only specialize kernels where it can make an impact. */
+        if (kernel_type == PSO_SPECIALISED) {
+          if (i < DEVICE_KERNEL_INTEGRATOR_INTERSECT_CLOSEST ||
+              i > DEVICE_KERNEL_INTEGRATOR_MEGAKERNEL) {
+            return;
+          }
+        }
+
+        MetalDeviceKernel &kernel = kernels_[i];
+
+        const std::string function_name = std::string("cycles_metal_") +
+                                          device_kernel_as_string((DeviceKernel)i);
+        int threads_per_threadgroup = device->max_threads_per_threadgroup;
+        if (i > DEVICE_KERNEL_INTEGRATOR_MEGAKERNEL && i < DEVICE_KERNEL_INTEGRATOR_RESET) {
+          /* Always use 512 for the sorting kernels */
+          threads_per_threadgroup = 512;
+        }
+
+        NSArray *kernel_function_list = nil;
+
+        if (i == DEVICE_KERNEL_INTEGRATOR_INTERSECT_CLOSEST ||
+            i == DEVICE_KERNEL_INTEGRATOR_INTERSECT_SHADOW ||
+            i == DEVICE_KERNEL_INTEGRATOR_INTERSECT_SUBSURFACE ||
+            i == DEVICE_KERNEL_INTEGRATOR_INTERSECT_VOLUME_STACK ||
+            i == DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE) {
+          kernel_function_list = function_list;
+        }
+
+        MetalKernelLoadDesc desc;
+        desc.pso_index = kernel_type;
+        desc.kernel_index = i;
+        desc.linked_functions = kernel_function_list;
+        desc.intersector_functions.defaults = table_functions[METALRT_TABLE_DEFAULT];
+        desc.intersector_functions.shadow = table_functions[METALRT_TABLE_SHADOW];
+        desc.intersector_functions.local = table_functions[METALRT_TABLE_LOCAL];
+        desc.constant_values = constant_values;
+        desc.threads_per_threadgroup = threads_per_threadgroup;
+        desc.function_name = function_name.c_str();
+
+        bool success = kernel.load(device, desc, md5);
+
+        any_error |= !success;
+      });
+    });
+  }
+
+  bool loaded = !any_error;
+  if (loaded) {
+    loaded_md5[kernel_type] = hash;
+  }
+  return loaded;
+}
+
+const MetalDeviceKernel &MetalDeviceKernels::get(DeviceKernel kernel) const
+{
+  return kernels_[(int)kernel];
+}
+
+bool MetalDeviceKernels::available(DeviceKernel kernel) const
+{
+  return kernels_[(int)kernel].get_pso().function != nil;
+}
+
+CCL_NAMESPACE_END
+
+#endif /* WITH_METAL*/

intern/cycles/device/metal/queue.h

@@ -0,0 +1,99 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#ifdef WITH_METAL
+
+#  include "device/kernel.h"
+#  include "device/memory.h"
+#  include "device/queue.h"
+
+#  include "device/metal/util.h"
+#  include "kernel/device/metal/globals.h"
+
+#  define metal_printf VLOG(4) << string_printf
+
+CCL_NAMESPACE_BEGIN
+
+class MetalDevice;
+
+/* Base class for Metal queues. */
+class MetalDeviceQueue : public DeviceQueue {
+ public:
+  MetalDeviceQueue(MetalDevice *device);
+  ~MetalDeviceQueue();
+
+  virtual int num_concurrent_states(const size_t) const override;
+  virtual int num_concurrent_busy_states() const override;
+
+  virtual void init_execution() override;
+
+  virtual bool enqueue(DeviceKernel kernel,
+                       const int work_size,
+                       DeviceKernelArguments const &args) override;
+
+  virtual bool synchronize() override;
+
+  virtual void zero_to_device(device_memory &mem) override;
+  virtual void copy_to_device(device_memory &mem) override;
+  virtual void copy_from_device(device_memory &mem) override;
+
+  virtual bool kernel_available(DeviceKernel kernel) const override;
+
+ protected:
+  void prepare_resources(DeviceKernel kernel);
+
+  id<MTLComputeCommandEncoder> get_compute_encoder(DeviceKernel kernel);
+  id<MTLBlitCommandEncoder> get_blit_encoder();
+
+  MetalDevice *metal_device;
+  MetalBufferPool temp_buffer_pool;
+
+  API_AVAILABLE(macos(11.0), ios(14.0))
+  MTLCommandBufferDescriptor *command_buffer_desc = nullptr;
+  id<MTLDevice> mtlDevice = nil;
+  id<MTLCommandQueue> mtlCommandQueue = nil;
+  id<MTLCommandBuffer> mtlCommandBuffer = nil;
+  id<MTLComputeCommandEncoder> mtlComputeEncoder = nil;
+  id<MTLBlitCommandEncoder> mtlBlitEncoder = nil;
+  API_AVAILABLE(macos(10.14), ios(14.0))
+  id<MTLSharedEvent> shared_event = nil;
+  API_AVAILABLE(macos(10.14), ios(14.0))
+  MTLSharedEventListener *shared_event_listener = nil;
+
+  dispatch_queue_t event_queue;
+  dispatch_semaphore_t wait_semaphore;
+
+  struct CopyBack {
+    void *host_pointer;
+    void *gpu_mem;
+    uint64_t size;
+  };
+  std::vector<CopyBack> copy_back_mem;
+
+  uint64_t shared_event_id;
+  uint64_t command_buffers_submitted = 0;
+  uint64_t command_buffers_completed = 0;
+  Stats &stats;
+
+  void close_compute_encoder();
+  void close_blit_encoder();
+};
+
+CCL_NAMESPACE_END
+
+#endif /* WITH_METAL */

intern/cycles/device/metal/queue.mm

@@ -0,0 +1,610 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifdef WITH_METAL
+
+#  include "device/metal/queue.h"
+
+#  include "device/metal/device_impl.h"
+#  include "device/metal/kernel.h"
+
+#  include "util/path.h"
+#  include "util/string.h"
+#  include "util/time.h"
+
+CCL_NAMESPACE_BEGIN
+
+/* MetalDeviceQueue */
+
+MetalDeviceQueue::MetalDeviceQueue(MetalDevice *device)
+    : DeviceQueue(device), metal_device(device), stats(device->stats)
+{
+  if (@available(macos 11.0, *)) {
+    command_buffer_desc = [[MTLCommandBufferDescriptor alloc] init];
+    command_buffer_desc.errorOptions = MTLCommandBufferErrorOptionEncoderExecutionStatus;
+  }
+
+  mtlDevice = device->mtlDevice;
+  mtlCommandQueue = [mtlDevice newCommandQueue];
+
+  if (@available(macos 10.14, *)) {
+    shared_event = [mtlDevice newSharedEvent];
+    shared_event_id = 1;
+
+    /* Shareable event listener */
+    event_queue = dispatch_queue_create("com.cycles.metal.event_queue", NULL);
+    shared_event_listener = [[MTLSharedEventListener alloc] initWithDispatchQueue:event_queue];
+  }
+
+  wait_semaphore = dispatch_semaphore_create(0);
+}
+
+MetalDeviceQueue::~MetalDeviceQueue()
+{
+  /* Tidying up here isn't really practical - we should expect and require the work
+   * queue to be empty here. */
+  assert(mtlCommandBuffer == nil);
+  assert(command_buffers_submitted == command_buffers_completed);
+
+  if (@available(macos 10.14, *)) {
+    [shared_event_listener release];
+    [shared_event release];
+  }
+
+  if (@available(macos 11.0, *)) {
+    [command_buffer_desc release];
+  }
+  if (mtlCommandQueue) {
+    [mtlCommandQueue release];
+    mtlCommandQueue = nil;
+  }
+}
+
+int MetalDeviceQueue::num_concurrent_states(const size_t /*state_size*/) const
+{
+  /* METAL_WIP */
+  /* TODO: compute automatically. */
+  /* TODO: must have at least num_threads_per_block. */
+  int result = 1048576;
+  if (metal_device->device_vendor == METAL_GPU_AMD) {
+    result *= 2;
+  }
+  else if (metal_device->device_vendor == METAL_GPU_APPLE) {
+    result *= 4;
+  }
+  return result;
+}
+
+int MetalDeviceQueue::num_concurrent_busy_states() const
+{
+  /* METAL_WIP */
+  /* TODO: compute automatically. */
+  int result = 65536;
+  if (metal_device->device_vendor == METAL_GPU_AMD) {
+    result *= 2;
+  }
+  else if (metal_device->device_vendor == METAL_GPU_APPLE) {
+    result *= 4;
+  }
+  return result;
+}
+
+void MetalDeviceQueue::init_execution()
+{
+  /* Synchronize all textures and memory copies before executing task. */
+  metal_device->load_texture_info();
+
+  synchronize();
+}
+
+bool MetalDeviceQueue::enqueue(DeviceKernel kernel,
+                               const int work_size,
+                               DeviceKernelArguments const &args)
+{
+  if (metal_device->have_error()) {
+    return false;
+  }
+
+  VLOG(3) << "Metal queue launch " << device_kernel_as_string(kernel) << ", work_size "
+          << work_size;
+
+  const MetalDeviceKernel &metal_kernel = metal_device->kernels.get(kernel);
+  const MetalKernelPipeline &metal_kernel_pso = metal_kernel.get_pso();
+
+  id<MTLComputeCommandEncoder> mtlComputeCommandEncoder = get_compute_encoder(kernel);
+
+  /* Determine size requirement for argument buffer. */
+  size_t arg_buffer_length = 0;
+  for (size_t i = 0; i < args.count; i++) {
+    size_t size_in_bytes = args.sizes[i];
+    arg_buffer_length = round_up(arg_buffer_length, size_in_bytes) + size_in_bytes;
+  }
+  /* 256 is the Metal offset alignment for constant address space bindings */
+  arg_buffer_length = round_up(arg_buffer_length, 256);
+
+  /* Globals placed after "vanilla" arguments. */
+  size_t globals_offsets = arg_buffer_length;
+  arg_buffer_length += sizeof(KernelParamsMetal);
+  arg_buffer_length = round_up(arg_buffer_length, 256);
+
+  /* Metal ancillary bindless pointers. */
+  size_t metal_offsets = arg_buffer_length;
+  arg_buffer_length += metal_device->mtlAncillaryArgEncoder.encodedLength;
+  arg_buffer_length = round_up(arg_buffer_length, metal_device->mtlAncillaryArgEncoder.alignment);
+
+  /* Temporary buffer used to prepare arg_buffer */
+  uint8_t *init_arg_buffer = (uint8_t *)alloca(arg_buffer_length);
+  memset(init_arg_buffer, 0, arg_buffer_length);
+
+  /* Prepare the non-pointer "enqueue" arguments */
+  size_t bytes_written = 0;
+  for (size_t i = 0; i < args.count; i++) {
+    size_t size_in_bytes = args.sizes[i];
+    bytes_written = round_up(bytes_written, size_in_bytes);
+    if (args.types[i] != DeviceKernelArguments::POINTER) {
+      memcpy(init_arg_buffer + bytes_written, args.values[i], size_in_bytes);
+    }
+    bytes_written += size_in_bytes;
+  }
+
+  /* Prepare any non-pointer (i.e. plain-old-data) KernelParamsMetal data */
+  /* The plain-old-data is contiguous, continuing to the end of KernelParamsMetal */
+  size_t plain_old_launch_data_offset = offsetof(KernelParamsMetal, __integrator_state) +
+                                        sizeof(IntegratorStateGPU);
+  size_t plain_old_launch_data_size = sizeof(KernelParamsMetal) - plain_old_launch_data_offset;
+  memcpy(init_arg_buffer + globals_offsets + plain_old_launch_data_offset,
+         (uint8_t *)&metal_device->launch_params + plain_old_launch_data_offset,
+         plain_old_launch_data_size);
+
+  /* Allocate an argument buffer. */
+  MTLResourceOptions arg_buffer_options = MTLResourceStorageModeManaged;
+  if (@available(macOS 11.0, *)) {
+    if ([mtlDevice hasUnifiedMemory]) {
+      arg_buffer_options = MTLResourceStorageModeShared;
+    }
+  }
+
+  id<MTLBuffer> arg_buffer = temp_buffer_pool.get_buffer(
+      mtlDevice, mtlCommandBuffer, arg_buffer_length, arg_buffer_options, init_arg_buffer, stats);
+
+  /* Encode the pointer "enqueue" arguments */
+  bytes_written = 0;
+  for (size_t i = 0; i < args.count; i++) {
+    size_t size_in_bytes = args.sizes[i];
+    bytes_written = round_up(bytes_written, size_in_bytes);
+    if (args.types[i] == DeviceKernelArguments::POINTER) {
+      [metal_device->mtlBufferKernelParamsEncoder setArgumentBuffer:arg_buffer
+                                                             offset:bytes_written];
+      if (MetalDevice::MetalMem *mmem = *(MetalDevice::MetalMem **)args.values[i]) {
+        [mtlComputeCommandEncoder useResource:mmem->mtlBuffer
+                                        usage:MTLResourceUsageRead | MTLResourceUsageWrite];
+        [metal_device->mtlBufferKernelParamsEncoder setBuffer:mmem->mtlBuffer offset:0 atIndex:0];
+      }
+      else {
+        if (@available(macos 12.0, *)) {
+          [metal_device->mtlBufferKernelParamsEncoder setBuffer:nil offset:0 atIndex:0];
+        }
+      }
+    }
+    bytes_written += size_in_bytes;
+  }
+
+  /* Encode KernelParamsMetal buffers */
+  [metal_device->mtlBufferKernelParamsEncoder setArgumentBuffer:arg_buffer offset:globals_offsets];
+
+  /* this relies on IntegratorStateGPU layout being contiguous device_ptrs  */
+  const size_t pointer_block_end = offsetof(KernelParamsMetal, __integrator_state) +
+                                   sizeof(IntegratorStateGPU);
+  for (size_t offset = 0; offset < pointer_block_end; offset += sizeof(device_ptr)) {
+    int pointer_index = offset / sizeof(device_ptr);
+    MetalDevice::MetalMem *mmem = *(
+        MetalDevice::MetalMem **)((uint8_t *)&metal_device->launch_params + offset);
+    if (mmem && (mmem->mtlBuffer || mmem->mtlTexture)) {
+      [metal_device->mtlBufferKernelParamsEncoder setBuffer:mmem->mtlBuffer
+                                                     offset:0
+                                                    atIndex:pointer_index];
+    }
+    else {
+      if (@available(macos 12.0, *)) {
+        [metal_device->mtlBufferKernelParamsEncoder setBuffer:nil offset:0 atIndex:pointer_index];
+      }
+    }
+  }
+  bytes_written = globals_offsets + sizeof(KernelParamsMetal);
+
+  /* Encode ancillaries */
+  [metal_device->mtlAncillaryArgEncoder setArgumentBuffer:arg_buffer offset:metal_offsets];
+  [metal_device->mtlAncillaryArgEncoder setBuffer:metal_device->texture_bindings_2d
+                                           offset:0
+                                          atIndex:0];
+  [metal_device->mtlAncillaryArgEncoder setBuffer:metal_device->texture_bindings_3d
+                                           offset:0
+                                          atIndex:1];
+  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];
+      }
+
+      for (int table = 0; table < METALRT_TABLE_NUM; table++) {
+        if (metal_kernel_pso.intersection_func_table[table]) {
+          [metal_kernel_pso.intersection_func_table[table] setBuffer:arg_buffer
+                                                              offset:globals_offsets
+                                                             atIndex:1];
+          [metal_device->mtlAncillaryArgEncoder
+              setIntersectionFunctionTable:metal_kernel_pso.intersection_func_table[table]
+                                   atIndex:3 + table];
+          [mtlComputeCommandEncoder useResource:metal_kernel_pso.intersection_func_table[table]
+                                          usage:MTLResourceUsageRead];
+        }
+        else {
+          [metal_device->mtlAncillaryArgEncoder setIntersectionFunctionTable:nil
+                                                                     atIndex:3 + table];
+        }
+      }
+    }
+    bytes_written = metal_offsets + metal_device->mtlAncillaryArgEncoder.encodedLength;
+  }
+
+  if (arg_buffer.storageMode == MTLStorageModeManaged) {
+    [arg_buffer didModifyRange:NSMakeRange(0, bytes_written)];
+  }
+
+  [mtlComputeCommandEncoder setBuffer:arg_buffer offset:0 atIndex:0];
+  [mtlComputeCommandEncoder setBuffer:arg_buffer offset:globals_offsets atIndex:1];
+  [mtlComputeCommandEncoder setBuffer:arg_buffer offset:metal_offsets atIndex:2];
+
+  if (metal_device->use_metalrt) {
+    if (@available(macos 12.0, *)) {
+
+      auto bvhMetalRT = metal_device->bvhMetalRT;
+      switch (kernel) {
+        case DEVICE_KERNEL_INTEGRATOR_INTERSECT_CLOSEST:
+        case DEVICE_KERNEL_INTEGRATOR_INTERSECT_SHADOW:
+        case DEVICE_KERNEL_INTEGRATOR_INTERSECT_SUBSURFACE:
+        case DEVICE_KERNEL_INTEGRATOR_INTERSECT_VOLUME_STACK:
+        case DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE:
+          break;
+        default:
+          bvhMetalRT = nil;
+          break;
+      }
+
+      if (bvhMetalRT) {
+        /* Mark all Accelerations resources as used */
+        [mtlComputeCommandEncoder useResource:bvhMetalRT->accel_struct usage:MTLResourceUsageRead];
+        [mtlComputeCommandEncoder useResources:bvhMetalRT->blas_array.data()
+                                         count:bvhMetalRT->blas_array.size()
+                                         usage:MTLResourceUsageRead];
+      }
+    }
+  }
+
+  [mtlComputeCommandEncoder setComputePipelineState:metal_kernel_pso.pipeline];
+
+  /* Compute kernel launch parameters. */
+  const int num_threads_per_block = metal_kernel.get_num_threads_per_block();
+
+  int shared_mem_bytes = 0;
+
+  switch (kernel) {
+    case DEVICE_KERNEL_INTEGRATOR_QUEUED_PATHS_ARRAY:
+    case DEVICE_KERNEL_INTEGRATOR_QUEUED_SHADOW_PATHS_ARRAY:
+    case DEVICE_KERNEL_INTEGRATOR_ACTIVE_PATHS_ARRAY:
+    case DEVICE_KERNEL_INTEGRATOR_TERMINATED_PATHS_ARRAY:
+    case DEVICE_KERNEL_INTEGRATOR_SORTED_PATHS_ARRAY:
+    case DEVICE_KERNEL_INTEGRATOR_COMPACT_PATHS_ARRAY:
+    case DEVICE_KERNEL_INTEGRATOR_TERMINATED_SHADOW_PATHS_ARRAY:
+    case DEVICE_KERNEL_INTEGRATOR_COMPACT_SHADOW_PATHS_ARRAY:
+      /* See parallel_active_index.h for why this amount of shared memory is needed.
+       * Rounded up to 16 bytes for Metal */
+      shared_mem_bytes = round_up((num_threads_per_block + 1) * sizeof(int), 16);
+      [mtlComputeCommandEncoder setThreadgroupMemoryLength:shared_mem_bytes atIndex:0];
+      break;
+
+    default:
+      break;
+  }
+
+  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);
+  [mtlComputeCommandEncoder dispatchThreadgroups:size_threadgroups_per_dispatch
+                           threadsPerThreadgroup:size_threads_per_threadgroup];
+
+  [mtlCommandBuffer addCompletedHandler:^(id<MTLCommandBuffer> command_buffer) {
+    NSString *kernel_name = metal_kernel_pso.function.label;
+
+    /* Enhanced command buffer errors are only available in 11.0+ */
+    if (@available(macos 11.0, *)) {
+      if (command_buffer.status == MTLCommandBufferStatusError && command_buffer.error != nil) {
+        printf("CommandBuffer Failed: %s\n", [kernel_name UTF8String]);
+        NSArray<id<MTLCommandBufferEncoderInfo>> *encoderInfos = [command_buffer.error.userInfo
+            valueForKey:MTLCommandBufferEncoderInfoErrorKey];
+        if (encoderInfos != nil) {
+          for (id<MTLCommandBufferEncoderInfo> encoderInfo : encoderInfos) {
+            NSLog(@"%@", encoderInfo);
+          }
+        }
+        id<MTLLogContainer> logs = command_buffer.logs;
+        for (id<MTLFunctionLog> log in logs) {
+          NSLog(@"%@", log);
+        }
+      }
+      else if (command_buffer.error) {
+        printf("CommandBuffer Failed: %s\n", [kernel_name UTF8String]);
+      }
+    }
+  }];
+
+  return !(metal_device->have_error());
+}
+
+bool MetalDeviceQueue::synchronize()
+{
+  if (metal_device->have_error()) {
+    return false;
+  }
+
+  if (mtlComputeEncoder) {
+    close_compute_encoder();
+  }
+  close_blit_encoder();
+
+  if (mtlCommandBuffer) {
+    uint64_t shared_event_id = this->shared_event_id++;
+
+    if (@available(macos 10.14, *)) {
+      __block dispatch_semaphore_t block_sema = wait_semaphore;
+      [shared_event notifyListener:shared_event_listener
+                           atValue:shared_event_id
+                             block:^(id<MTLSharedEvent> sharedEvent, uint64_t value) {
+                               dispatch_semaphore_signal(block_sema);
+                             }];
+
+      [mtlCommandBuffer encodeSignalEvent:shared_event value:shared_event_id];
+      [mtlCommandBuffer commit];
+      dispatch_semaphore_wait(wait_semaphore, DISPATCH_TIME_FOREVER);
+    }
+
+    [mtlCommandBuffer release];
+
+    for (const CopyBack &mmem : copy_back_mem) {
+      memcpy((uchar *)mmem.host_pointer, (uchar *)mmem.gpu_mem, mmem.size);
+    }
+    copy_back_mem.clear();
+
+    temp_buffer_pool.process_command_buffer_completion(mtlCommandBuffer);
+    metal_device->flush_delayed_free_list();
+
+    mtlCommandBuffer = nil;
+  }
+
+  return !(metal_device->have_error());
+}
+
+void MetalDeviceQueue::zero_to_device(device_memory &mem)
+{
+  assert(mem.type != MEM_GLOBAL && mem.type != MEM_TEXTURE);
+
+  if (mem.memory_size() == 0) {
+    return;
+  }
+
+  /* Allocate on demand. */
+  if (mem.device_pointer == 0) {
+    metal_device->mem_alloc(mem);
+  }
+
+  /* Zero memory on device. */
+  assert(mem.device_pointer != 0);
+
+  std::lock_guard<std::recursive_mutex> lock(metal_device->metal_mem_map_mutex);
+  MetalDevice::MetalMem &mmem = *metal_device->metal_mem_map.at(&mem);
+  if (mmem.mtlBuffer) {
+    id<MTLBlitCommandEncoder> blitEncoder = get_blit_encoder();
+    [blitEncoder fillBuffer:mmem.mtlBuffer range:NSMakeRange(mmem.offset, mmem.size) value:0];
+  }
+  else {
+    metal_device->mem_zero(mem);
+  }
+}
+
+void MetalDeviceQueue::copy_to_device(device_memory &mem)
+{
+  if (mem.memory_size() == 0) {
+    return;
+  }
+
+  /* Allocate on demand. */
+  if (mem.device_pointer == 0) {
+    metal_device->mem_alloc(mem);
+  }
+
+  assert(mem.device_pointer != 0);
+  assert(mem.host_pointer != nullptr);
+
+  std::lock_guard<std::recursive_mutex> lock(metal_device->metal_mem_map_mutex);
+  auto result = metal_device->metal_mem_map.find(&mem);
+  if (result != metal_device->metal_mem_map.end()) {
+    if (mem.host_pointer == mem.shared_pointer) {
+      return;
+    }
+
+    MetalDevice::MetalMem &mmem = *result->second;
+    id<MTLBlitCommandEncoder> blitEncoder = get_blit_encoder();
+
+    id<MTLBuffer> buffer = temp_buffer_pool.get_buffer(mtlDevice,
+                                                       mtlCommandBuffer,
+                                                       mmem.size,
+                                                       MTLResourceStorageModeShared,
+                                                       mem.host_pointer,
+                                                       stats);
+
+    [blitEncoder copyFromBuffer:buffer
+                   sourceOffset:0
+                       toBuffer:mmem.mtlBuffer
+              destinationOffset:mmem.offset
+                           size:mmem.size];
+  }
+  else {
+    metal_device->mem_copy_to(mem);
+  }
+}
+
+void MetalDeviceQueue::copy_from_device(device_memory &mem)
+{
+  assert(mem.type != MEM_GLOBAL && mem.type != MEM_TEXTURE);
+
+  if (mem.memory_size() == 0) {
+    return;
+  }
+
+  assert(mem.device_pointer != 0);
+  assert(mem.host_pointer != nullptr);
+
+  std::lock_guard<std::recursive_mutex> lock(metal_device->metal_mem_map_mutex);
+  MetalDevice::MetalMem &mmem = *metal_device->metal_mem_map.at(&mem);
+  if (mmem.mtlBuffer) {
+    const size_t size = mem.memory_size();
+
+    if (mem.device_pointer) {
+      if ([mmem.mtlBuffer storageMode] == MTLStorageModeManaged) {
+        id<MTLBlitCommandEncoder> blitEncoder = get_blit_encoder();
+        [blitEncoder synchronizeResource:mmem.mtlBuffer];
+      }
+      if (mem.host_pointer != mmem.hostPtr) {
+        if (mtlCommandBuffer) {
+          copy_back_mem.push_back({mem.host_pointer, mmem.hostPtr, size});
+        }
+        else {
+          memcpy((uchar *)mem.host_pointer, (uchar *)mmem.hostPtr, size);
+        }
+      }
+    }
+    else {
+      memset((char *)mem.host_pointer, 0, size);
+    }
+  }
+  else {
+    metal_device->mem_copy_from(mem);
+  }
+}
+
+bool MetalDeviceQueue::kernel_available(DeviceKernel kernel) const
+{
+  return metal_device->kernels.available(kernel);
+}
+
+void MetalDeviceQueue::prepare_resources(DeviceKernel kernel)
+{
+  std::lock_guard<std::recursive_mutex> lock(metal_device->metal_mem_map_mutex);
+
+  /* declare resource usage */
+  for (auto &it : metal_device->metal_mem_map) {
+    device_memory *mem = it.first;
+
+    MTLResourceUsage usage = MTLResourceUsageRead;
+    if (mem->type != MEM_GLOBAL && mem->type != MEM_READ_ONLY && mem->type != MEM_TEXTURE) {
+      usage |= MTLResourceUsageWrite;
+    }
+
+    if (it.second->mtlBuffer) {
+      /* METAL_WIP - use array version (i.e. useResources) */
+      [mtlComputeEncoder useResource:it.second->mtlBuffer usage:usage];
+    }
+    else if (it.second->mtlTexture) {
+      /* METAL_WIP - use array version (i.e. useResources) */
+      [mtlComputeEncoder useResource:it.second->mtlTexture usage:usage | MTLResourceUsageSample];
+    }
+  }
+
+  /* ancillaries */
+  [mtlComputeEncoder useResource:metal_device->texture_bindings_2d usage:MTLResourceUsageRead];
+  [mtlComputeEncoder useResource:metal_device->texture_bindings_3d usage:MTLResourceUsageRead];
+}
+
+id<MTLComputeCommandEncoder> MetalDeviceQueue::get_compute_encoder(DeviceKernel kernel)
+{
+  bool concurrent = (kernel < DEVICE_KERNEL_INTEGRATOR_NUM);
+
+  if (@available(macos 10.14, *)) {
+    if (mtlComputeEncoder) {
+      if (mtlComputeEncoder.dispatchType == concurrent ? MTLDispatchTypeConcurrent :
+                                                         MTLDispatchTypeSerial) {
+        /* declare usage of MTLBuffers etc */
+        prepare_resources(kernel);
+
+        return mtlComputeEncoder;
+      }
+      close_compute_encoder();
+    }
+
+    close_blit_encoder();
+
+    if (!mtlCommandBuffer) {
+      mtlCommandBuffer = [mtlCommandQueue commandBuffer];
+      [mtlCommandBuffer retain];
+    }
+
+    mtlComputeEncoder = [mtlCommandBuffer
+        computeCommandEncoderWithDispatchType:concurrent ? MTLDispatchTypeConcurrent :
+                                                           MTLDispatchTypeSerial];
+
+    /* declare usage of MTLBuffers etc */
+    prepare_resources(kernel);
+  }
+
+  return mtlComputeEncoder;
+}
+
+id<MTLBlitCommandEncoder> MetalDeviceQueue::get_blit_encoder()
+{
+  if (mtlBlitEncoder) {
+    return mtlBlitEncoder;
+  }
+
+  if (mtlComputeEncoder) {
+    close_compute_encoder();
+  }
+
+  if (!mtlCommandBuffer) {
+    mtlCommandBuffer = [mtlCommandQueue commandBuffer];
+    [mtlCommandBuffer retain];
+  }
+
+  mtlBlitEncoder = [mtlCommandBuffer blitCommandEncoder];
+  return mtlBlitEncoder;
+}
+
+void MetalDeviceQueue::close_compute_encoder()
+{
+  [mtlComputeEncoder endEncoding];
+  mtlComputeEncoder = nil;
+}
+
+void MetalDeviceQueue::close_blit_encoder()
+{
+  if (mtlBlitEncoder) {
+    [mtlBlitEncoder endEncoding];
+    mtlBlitEncoder = nil;
+  }
+}
+
+CCL_NAMESPACE_END
+
+#endif /* WITH_METAL */

intern/cycles/device/metal/util.h

@@ -0,0 +1,101 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#ifdef WITH_METAL
+
+#  include <Metal/Metal.h>
+#  include <string>
+
+#  include "device/metal/device.h"
+#  include "device/metal/kernel.h"
+#  include "device/queue.h"
+
+#  include "util/thread.h"
+
+CCL_NAMESPACE_BEGIN
+
+enum MetalGPUVendor {
+  METAL_GPU_UNKNOWN = 0,
+  METAL_GPU_APPLE = 1,
+  METAL_GPU_AMD = 2,
+  METAL_GPU_INTEL = 3,
+};
+
+/* Retains a named MTLDevice for device enumeration. */
+struct MetalPlatformDevice {
+  MetalPlatformDevice(id<MTLDevice> device, const string &device_name)
+      : device_id(device), device_name(device_name)
+  {
+    [device_id retain];
+  }
+  ~MetalPlatformDevice()
+  {
+    [device_id release];
+  }
+  id<MTLDevice> device_id;
+  string device_name;
+};
+
+/* Contains static Metal helper functions. */
+struct MetalInfo {
+  static bool device_version_check(id<MTLDevice> device);
+  static void get_usable_devices(vector<MetalPlatformDevice> *usable_devices);
+  static MetalGPUVendor get_vendor_from_device_name(string const &device_name);
+
+  /* Platform information. */
+  static bool get_num_devices(uint32_t *num_platforms);
+  static uint32_t get_num_devices();
+
+  static bool get_device_name(id<MTLDevice> device_id, string *device_name);
+  static string get_device_name(id<MTLDevice> device_id);
+};
+
+/* Pool of MTLBuffers whose lifetime is linked to a single MTLCommandBuffer */
+class MetalBufferPool {
+  struct MetalBufferListEntry {
+    MetalBufferListEntry(id<MTLBuffer> buffer, id<MTLCommandBuffer> command_buffer)
+        : buffer(buffer), command_buffer(command_buffer)
+    {
+    }
+
+    MetalBufferListEntry() = delete;
+
+    id<MTLBuffer> buffer;
+    id<MTLCommandBuffer> command_buffer;
+  };
+  std::vector<MetalBufferListEntry> buffer_free_list;
+  std::vector<MetalBufferListEntry> buffer_in_use_list;
+  thread_mutex buffer_mutex;
+  size_t total_temp_mem_size = 0;
+
+ public:
+  MetalBufferPool() = default;
+  ~MetalBufferPool();
+
+  id<MTLBuffer> get_buffer(id<MTLDevice> device,
+                           id<MTLCommandBuffer> command_buffer,
+                           NSUInteger length,
+                           MTLResourceOptions options,
+                           const void *pointer,
+                           Stats &stats);
+  void process_command_buffer_completion(id<MTLCommandBuffer> command_buffer);
+};
+
+CCL_NAMESPACE_END
+
+#endif /* WITH_METAL */

intern/cycles/device/metal/util.mm

@@ -0,0 +1,218 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifdef WITH_METAL
+
+#  include "device/metal/util.h"
+#  include "device/metal/device_impl.h"
+#  include "util/md5.h"
+#  include "util/path.h"
+#  include "util/string.h"
+#  include "util/time.h"
+
+#  include <pwd.h>
+#  include <sys/shm.h>
+#  include <time.h>
+
+CCL_NAMESPACE_BEGIN
+
+MetalGPUVendor MetalInfo::get_vendor_from_device_name(string const &device_name)
+{
+  if (device_name.find("Intel") != string::npos) {
+    return METAL_GPU_INTEL;
+  }
+  else if (device_name.find("AMD") != string::npos) {
+    return METAL_GPU_AMD;
+  }
+  else if (device_name.find("Apple") != string::npos) {
+    return METAL_GPU_APPLE;
+  }
+  return METAL_GPU_UNKNOWN;
+}
+
+bool MetalInfo::device_version_check(id<MTLDevice> device)
+{
+  /* Metal Cycles doesn't work correctly on macOS versions older than 12.0 */
+  if (@available(macos 12.0, *)) {
+    MetalGPUVendor vendor = get_vendor_from_device_name([[device name] UTF8String]);
+
+    /* Metal Cycles works on Apple Silicon GPUs at present */
+    return (vendor == METAL_GPU_APPLE);
+  }
+
+  return false;
+}
+
+void MetalInfo::get_usable_devices(vector<MetalPlatformDevice> *usable_devices)
+{
+  static bool first_time = true;
+#  define FIRST_VLOG(severity) \
+    if (first_time) \
+    VLOG(severity)
+
+  usable_devices->clear();
+
+  NSArray<id<MTLDevice>> *allDevices = MTLCopyAllDevices();
+  for (id<MTLDevice> device in allDevices) {
+    string device_name;
+    if (!get_device_name(device, &device_name)) {
+      FIRST_VLOG(2) << "Failed to get device name, ignoring.";
+      continue;
+    }
+
+    static const char *forceIntelStr = getenv("CYCLES_METAL_FORCE_INTEL");
+    bool forceIntel = forceIntelStr ? (atoi(forceIntelStr) != 0) : false;
+    if (forceIntel && device_name.find("Intel") == string::npos) {
+      FIRST_VLOG(2) << "CYCLES_METAL_FORCE_INTEL causing non-Intel device " << device_name
+                    << " to be ignored.";
+      continue;
+    }
+
+    if (!device_version_check(device)) {
+      FIRST_VLOG(2) << "Ignoring device " << device_name << " due to too old compiler version.";
+      continue;
+    }
+    FIRST_VLOG(2) << "Adding new device " << device_name << ".";
+    string hardware_id;
+    usable_devices->push_back(MetalPlatformDevice(device, device_name));
+  }
+  first_time = false;
+}
+
+bool MetalInfo::get_num_devices(uint32_t *num_devices)
+{
+  *num_devices = MTLCopyAllDevices().count;
+  return true;
+}
+
+uint32_t MetalInfo::get_num_devices()
+{
+  uint32_t num_devices;
+  if (!get_num_devices(&num_devices)) {
+    return 0;
+  }
+  return num_devices;
+}
+
+bool MetalInfo::get_device_name(id<MTLDevice> device, string *platform_name)
+{
+  *platform_name = [device.name UTF8String];
+  return true;
+}
+
+string MetalInfo::get_device_name(id<MTLDevice> device)
+{
+  string platform_name;
+  if (!get_device_name(device, &platform_name)) {
+    return "";
+  }
+  return platform_name;
+}
+
+id<MTLBuffer> MetalBufferPool::get_buffer(id<MTLDevice> device,
+                                          id<MTLCommandBuffer> command_buffer,
+                                          NSUInteger length,
+                                          MTLResourceOptions options,
+                                          const void *pointer,
+                                          Stats &stats)
+{
+  id<MTLBuffer> buffer;
+
+  MTLStorageMode storageMode = MTLStorageMode((options & MTLResourceStorageModeMask) >>
+                                              MTLResourceStorageModeShift);
+  MTLCPUCacheMode cpuCacheMode = MTLCPUCacheMode((options & MTLResourceCPUCacheModeMask) >>
+                                                 MTLResourceCPUCacheModeShift);
+
+  buffer_mutex.lock();
+  for (auto entry = buffer_free_list.begin(); entry != buffer_free_list.end(); entry++) {
+    MetalBufferListEntry bufferEntry = *entry;
+
+    /* Check if buffer matches size and storage mode and is old enough to reuse */
+    if (bufferEntry.buffer.length == length && storageMode == bufferEntry.buffer.storageMode &&
+        cpuCacheMode == bufferEntry.buffer.cpuCacheMode) {
+      buffer = bufferEntry.buffer;
+      buffer_free_list.erase(entry);
+      bufferEntry.command_buffer = command_buffer;
+      buffer_in_use_list.push_back(bufferEntry);
+      buffer_mutex.unlock();
+
+      /* Copy over data */
+      if (pointer) {
+        memcpy(buffer.contents, pointer, length);
+        if (bufferEntry.buffer.storageMode == MTLStorageModeManaged) {
+          [buffer didModifyRange:NSMakeRange(0, length)];
+        }
+      }
+
+      return buffer;
+    }
+  }
+  // NSLog(@"Creating buffer of length %lu (%lu)", length, frameCount);
+  if (pointer) {
+    buffer = [device newBufferWithBytes:pointer length:length options:options];
+  }
+  else {
+    buffer = [device newBufferWithLength:length options:options];
+  }
+
+  MetalBufferListEntry buffer_entry(buffer, command_buffer);
+
+  stats.mem_alloc(buffer.allocatedSize);
+
+  total_temp_mem_size += buffer.allocatedSize;
+  buffer_in_use_list.push_back(buffer_entry);
+  buffer_mutex.unlock();
+
+  return buffer;
+}
+
+void MetalBufferPool::process_command_buffer_completion(id<MTLCommandBuffer> command_buffer)
+{
+  assert(command_buffer);
+  thread_scoped_lock lock(buffer_mutex);
+  /* Release all buffers that have not been recently reused back into the free pool */
+  for (auto entry = buffer_in_use_list.begin(); entry != buffer_in_use_list.end();) {
+    MetalBufferListEntry buffer_entry = *entry;
+    if (buffer_entry.command_buffer == command_buffer) {
+      entry = buffer_in_use_list.erase(entry);
+      buffer_entry.command_buffer = nil;
+      buffer_free_list.push_back(buffer_entry);
+    }
+    else {
+      entry++;
+    }
+  }
+}
+
+MetalBufferPool::~MetalBufferPool()
+{
+  thread_scoped_lock lock(buffer_mutex);
+  /* Release all buffers that have not been recently reused */
+  for (auto entry = buffer_free_list.begin(); entry != buffer_free_list.end();) {
+    MetalBufferListEntry buffer_entry = *entry;
+
+    id<MTLBuffer> buffer = buffer_entry.buffer;
+    // NSLog(@"Releasing buffer of length %lu (%lu) (%lu outstanding)", buffer.length, frameCount,
+    // bufferFreeList.size());
+    total_temp_mem_size -= buffer.allocatedSize;
+    [buffer release];
+    entry = buffer_free_list.erase(entry);
+  }
+}
+
+CCL_NAMESPACE_END
+
+#endif /* WITH_METAL */

intern/cycles/device/multi/device.cpp

@@ -124,11 +124,20 @@ class MultiDevice : public Device {
       return BVH_LAYOUT_MULTI_OPTIX;
     }
 
+    /* With multiple Metal devices, every device needs its own acceleration structure */
+    if (bvh_layout_mask == BVH_LAYOUT_METAL) {
+      return BVH_LAYOUT_MULTI_METAL;
+    }
+
     /* When devices do not share a common BVH layout, fall back to creating one for each */
     const BVHLayoutMask BVH_LAYOUT_OPTIX_EMBREE = (BVH_LAYOUT_OPTIX | BVH_LAYOUT_EMBREE);
     if ((bvh_layout_mask_all & BVH_LAYOUT_OPTIX_EMBREE) == BVH_LAYOUT_OPTIX_EMBREE) {
       return BVH_LAYOUT_MULTI_OPTIX_EMBREE;
     }
+    const BVHLayoutMask BVH_LAYOUT_METAL_EMBREE = (BVH_LAYOUT_METAL | BVH_LAYOUT_EMBREE);
+    if ((bvh_layout_mask_all & BVH_LAYOUT_METAL_EMBREE) == BVH_LAYOUT_METAL_EMBREE) {
+      return BVH_LAYOUT_MULTI_METAL_EMBREE;
+    }
 
     return bvh_layout_mask;
   }
@@ -151,7 +160,9 @@ class MultiDevice : public Device {
     }
 
     assert(bvh->params.bvh_layout == BVH_LAYOUT_MULTI_OPTIX ||
-           bvh->params.bvh_layout == BVH_LAYOUT_MULTI_OPTIX_EMBREE);
+           bvh->params.bvh_layout == BVH_LAYOUT_MULTI_METAL ||
+           bvh->params.bvh_layout == BVH_LAYOUT_MULTI_OPTIX_EMBREE ||
+           bvh->params.bvh_layout == BVH_LAYOUT_MULTI_METAL_EMBREE);
 
     BVHMulti *const bvh_multi = static_cast<BVHMulti *>(bvh);
     bvh_multi->sub_bvhs.resize(devices.size());
@@ -174,9 +185,14 @@ class MultiDevice : public Device {
         BVHParams params = bvh->params;
         if (bvh->params.bvh_layout == BVH_LAYOUT_MULTI_OPTIX)
           params.bvh_layout = BVH_LAYOUT_OPTIX;
+        else if (bvh->params.bvh_layout == BVH_LAYOUT_MULTI_METAL)
+          params.bvh_layout = BVH_LAYOUT_METAL;
         else if (bvh->params.bvh_layout == BVH_LAYOUT_MULTI_OPTIX_EMBREE)
           params.bvh_layout = sub.device->info.type == DEVICE_OPTIX ? BVH_LAYOUT_OPTIX :
                                                                       BVH_LAYOUT_EMBREE;
+        else if (bvh->params.bvh_layout == BVH_LAYOUT_MULTI_METAL_EMBREE)
+          params.bvh_layout = sub.device->info.type == DEVICE_METAL ? BVH_LAYOUT_METAL :
+                                                                      BVH_LAYOUT_EMBREE;
 
         /* Skip building a bottom level acceleration structure for non-instanced geometry on Embree
          * (since they are put into the top level directly, see bvh_embree.cpp) */

intern/cycles/device/optix/device_impl.cpp

@@ -28,6 +28,7 @@
 #  include "scene/mesh.h"
 #  include "scene/object.h"
 #  include "scene/pass.h"
+#  include "scene/pointcloud.h"
 #  include "scene/scene.h"
 
 #  include "util/debug.h"
@@ -41,17 +42,19 @@
 #  define __KERNEL_OPTIX__
 #  include "kernel/device/optix/globals.h"
 
+#  include <optix_denoiser_tiling.h>
+
 CCL_NAMESPACE_BEGIN
 
 OptiXDevice::Denoiser::Denoiser(OptiXDevice *device)
-    : device(device), queue(device), state(device, "__denoiser_state")
+    : device(device), queue(device), state(device, "__denoiser_state", true)
 {
 }
 
 OptiXDevice::OptiXDevice(const DeviceInfo &info, Stats &stats, Profiler &profiler)
     : CUDADevice(info, stats, profiler),
       sbt_data(this, "__sbt", MEM_READ_ONLY),
-      launch_params(this, "__params"),
+      launch_params(this, "__params", false),
       denoiser_(this)
 {
   /* Make the CUDA context current. */
@@ -208,11 +211,15 @@ bool OptiXDevice::load_kernels(const uint kernel_features)
   }
   else {
     module_options.optLevel = OPTIX_COMPILE_OPTIMIZATION_LEVEL_3;
-    module_options.debugLevel = OPTIX_COMPILE_DEBUG_LEVEL_LINEINFO;
+    module_options.debugLevel = OPTIX_COMPILE_DEBUG_LEVEL_NONE;
   }
 
   module_options.boundValues = nullptr;
   module_options.numBoundValues = 0;
+#  if OPTIX_ABI_VERSION >= 55
+  module_options.payloadTypes = nullptr;
+  module_options.numPayloadTypes = 0;
+#  endif
 
   OptixPipelineCompileOptions pipeline_options = {};
   /* Default to no motion blur and two-level graph, since it is the fastest option. */
@@ -227,11 +234,18 @@ bool OptiXDevice::load_kernels(const uint kernel_features)
   pipeline_options.usesPrimitiveTypeFlags = OPTIX_PRIMITIVE_TYPE_FLAGS_TRIANGLE;
   if (kernel_features & KERNEL_FEATURE_HAIR) {
     if (kernel_features & KERNEL_FEATURE_HAIR_THICK) {
+#  if OPTIX_ABI_VERSION >= 55
+      pipeline_options.usesPrimitiveTypeFlags |= OPTIX_PRIMITIVE_TYPE_FLAGS_ROUND_CATMULLROM;
+#  else
       pipeline_options.usesPrimitiveTypeFlags |= OPTIX_PRIMITIVE_TYPE_FLAGS_ROUND_CUBIC_BSPLINE;
+#  endif
     }
     else
       pipeline_options.usesPrimitiveTypeFlags |= OPTIX_PRIMITIVE_TYPE_FLAGS_CUSTOM;
   }
+  if (kernel_features & KERNEL_FEATURE_POINTCLOUD) {
+    pipeline_options.usesPrimitiveTypeFlags |= OPTIX_PRIMITIVE_TYPE_FLAGS_CUSTOM;
+  }
 
   /* Keep track of whether motion blur is enabled, so to enable/disable motion in BVH builds
    * This is necessary since objects may be reported to have motion if the Vector pass is
@@ -324,7 +338,13 @@ bool OptiXDevice::load_kernels(const uint kernel_features)
     if (kernel_features & KERNEL_FEATURE_HAIR_THICK) {
       /* Built-in thick curve intersection. */
       OptixBuiltinISOptions builtin_options = {};
+#  if OPTIX_ABI_VERSION >= 55
+      builtin_options.builtinISModuleType = OPTIX_PRIMITIVE_TYPE_ROUND_CATMULLROM;
+      builtin_options.buildFlags = OPTIX_BUILD_FLAG_PREFER_FAST_TRACE;
+      builtin_options.curveEndcapFlags = OPTIX_CURVE_ENDCAP_DEFAULT; /* Disable end-caps. */
+#  else
       builtin_options.builtinISModuleType = OPTIX_PRIMITIVE_TYPE_ROUND_CUBIC_BSPLINE;
+#  endif
       builtin_options.usesMotionBlur = false;
 
       optix_assert(optixBuiltinISModuleGet(
@@ -356,6 +376,18 @@ bool OptiXDevice::load_kernels(const uint kernel_features)
     }
   }
 
+  /* Pointclouds */
+  if (kernel_features & KERNEL_FEATURE_POINTCLOUD) {
+    group_descs[PG_HITD_POINTCLOUD] = group_descs[PG_HITD];
+    group_descs[PG_HITD_POINTCLOUD].kind = OPTIX_PROGRAM_GROUP_KIND_HITGROUP;
+    group_descs[PG_HITD_POINTCLOUD].hitgroup.moduleIS = optix_module;
+    group_descs[PG_HITD_POINTCLOUD].hitgroup.entryFunctionNameIS = "__intersection__point";
+    group_descs[PG_HITS_POINTCLOUD] = group_descs[PG_HITS];
+    group_descs[PG_HITS_POINTCLOUD].kind = OPTIX_PROGRAM_GROUP_KIND_HITGROUP;
+    group_descs[PG_HITS_POINTCLOUD].hitgroup.moduleIS = optix_module;
+    group_descs[PG_HITS_POINTCLOUD].hitgroup.entryFunctionNameIS = "__intersection__point";
+  }
+
   if (kernel_features & (KERNEL_FEATURE_SUBSURFACE | KERNEL_FEATURE_NODE_RAYTRACE)) {
     /* Add hit group for local intersections. */
     group_descs[PG_HITL].kind = OPTIX_PROGRAM_GROUP_KIND_HITGROUP;
@@ -403,6 +435,10 @@ bool OptiXDevice::load_kernels(const uint kernel_features)
                        stack_size[PG_HITD_MOTION].cssIS + stack_size[PG_HITD_MOTION].cssAH);
   trace_css = std::max(trace_css,
                        stack_size[PG_HITS_MOTION].cssIS + stack_size[PG_HITS_MOTION].cssAH);
+  trace_css = std::max(
+      trace_css, stack_size[PG_HITD_POINTCLOUD].cssIS + stack_size[PG_HITD_POINTCLOUD].cssAH);
+  trace_css = std::max(
+      trace_css, stack_size[PG_HITS_POINTCLOUD].cssIS + stack_size[PG_HITS_POINTCLOUD].cssAH);
 
   OptixPipelineLinkOptions link_options = {};
   link_options.maxTraceDepth = 1;
@@ -411,7 +447,7 @@ bool OptiXDevice::load_kernels(const uint kernel_features)
     link_options.debugLevel = OPTIX_COMPILE_DEBUG_LEVEL_FULL;
   }
   else {
-    link_options.debugLevel = OPTIX_COMPILE_DEBUG_LEVEL_LINEINFO;
+    link_options.debugLevel = OPTIX_COMPILE_DEBUG_LEVEL_NONE;
   }
 
   if (kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) {
@@ -428,6 +464,10 @@ bool OptiXDevice::load_kernels(const uint kernel_features)
       pipeline_groups.push_back(groups[PG_HITD_MOTION]);
       pipeline_groups.push_back(groups[PG_HITS_MOTION]);
     }
+    if (kernel_features & KERNEL_FEATURE_POINTCLOUD) {
+      pipeline_groups.push_back(groups[PG_HITD_POINTCLOUD]);
+      pipeline_groups.push_back(groups[PG_HITS_POINTCLOUD]);
+    }
     pipeline_groups.push_back(groups[PG_CALL_SVM_AO]);
     pipeline_groups.push_back(groups[PG_CALL_SVM_BEVEL]);
 
@@ -467,6 +507,10 @@ bool OptiXDevice::load_kernels(const uint kernel_features)
       pipeline_groups.push_back(groups[PG_HITD_MOTION]);
       pipeline_groups.push_back(groups[PG_HITS_MOTION]);
     }
+    if (kernel_features & KERNEL_FEATURE_POINTCLOUD) {
+      pipeline_groups.push_back(groups[PG_HITD_POINTCLOUD]);
+      pipeline_groups.push_back(groups[PG_HITS_POINTCLOUD]);
+    }
 
     optix_assert(optixPipelineCreate(context,
                                      &pipeline_options,
@@ -507,7 +551,7 @@ class OptiXDevice::DenoiseContext {
       : denoise_params(task.params),
         render_buffers(task.render_buffers),
         buffer_params(task.buffer_params),
-        guiding_buffer(device, "denoiser guiding passes buffer"),
+        guiding_buffer(device, "denoiser guiding passes buffer", true),
         num_samples(task.num_samples)
   {
     num_input_passes = 1;
@@ -522,9 +566,9 @@ class OptiXDevice::DenoiseContext {
       }
     }
 
-    const int num_guiding_passes = num_input_passes - 1;
+    use_guiding_passes = (num_input_passes - 1) > 0;
 
-    if (num_guiding_passes) {
+    if (use_guiding_passes) {
       if (task.allow_inplace_modification) {
         guiding_params.device_pointer = render_buffers->buffer.device_pointer;
 
@@ -577,6 +621,7 @@ class OptiXDevice::DenoiseContext {
 
   /* Number of input passes. Including the color and extra auxiliary passes. */
   int num_input_passes = 0;
+  bool use_guiding_passes = false;
   bool use_pass_albedo = false;
   bool use_pass_normal = false;
 
@@ -653,22 +698,22 @@ bool OptiXDevice::denoise_filter_guiding_preprocess(DenoiseContext &context)
 
   const int work_size = buffer_params.width * buffer_params.height;
 
-  void *args[] = {const_cast<device_ptr *>(&context.guiding_params.device_pointer),
+  DeviceKernelArguments args(&context.guiding_params.device_pointer,
-                  const_cast<int *>(&context.guiding_params.pass_stride),
+                             &context.guiding_params.pass_stride,
-                  const_cast<int *>(&context.guiding_params.pass_albedo),
+                             &context.guiding_params.pass_albedo,
-                  const_cast<int *>(&context.guiding_params.pass_normal),
+                             &context.guiding_params.pass_normal,
                              &context.render_buffers->buffer.device_pointer,
-                  const_cast<int *>(&buffer_params.offset),
+                             &buffer_params.offset,
-                  const_cast<int *>(&buffer_params.stride),
+                             &buffer_params.stride,
-                  const_cast<int *>(&buffer_params.pass_stride),
+                             &buffer_params.pass_stride,
-                  const_cast<int *>(&context.pass_sample_count),
+                             &context.pass_sample_count,
-                  const_cast<int *>(&context.pass_denoising_albedo),
+                             &context.pass_denoising_albedo,
-                  const_cast<int *>(&context.pass_denoising_normal),
+                             &context.pass_denoising_normal,
-                  const_cast<int *>(&buffer_params.full_x),
+                             &buffer_params.full_x,
-                  const_cast<int *>(&buffer_params.full_y),
+                             &buffer_params.full_y,
-                  const_cast<int *>(&buffer_params.width),
+                             &buffer_params.width,
-                  const_cast<int *>(&buffer_params.height),
+                             &buffer_params.height,
-                  const_cast<int *>(&context.num_samples)};
+                             &context.num_samples);
 
   return denoiser_.queue.enqueue(DEVICE_KERNEL_FILTER_GUIDING_PREPROCESS, work_size, args);
 }
@@ -679,11 +724,11 @@ bool OptiXDevice::denoise_filter_guiding_set_fake_albedo(DenoiseContext &context
 
   const int work_size = buffer_params.width * buffer_params.height;
 
-  void *args[] = {const_cast<device_ptr *>(&context.guiding_params.device_pointer),
+  DeviceKernelArguments args(&context.guiding_params.device_pointer,
-                  const_cast<int *>(&context.guiding_params.pass_stride),
+                             &context.guiding_params.pass_stride,
-                  const_cast<int *>(&context.guiding_params.pass_albedo),
+                             &context.guiding_params.pass_albedo,
-                  const_cast<int *>(&buffer_params.width),
+                             &buffer_params.width,
-                  const_cast<int *>(&buffer_params.height)};
+                             &buffer_params.height);
 
   return denoiser_.queue.enqueue(DEVICE_KERNEL_FILTER_GUIDING_SET_FAKE_ALBEDO, work_size, args);
 }
@@ -708,7 +753,7 @@ void OptiXDevice::denoise_pass(DenoiseContext &context, PassType pass_type)
       return;
     }
   }
-  else if (!context.albedo_replaced_with_fake) {
+  else if (context.use_guiding_passes && !context.albedo_replaced_with_fake) {
     context.albedo_replaced_with_fake = true;
     if (!denoise_filter_guiding_set_fake_albedo(context)) {
       LOG(ERROR) << "Error replacing real albedo with the fake one.";
@@ -779,15 +824,15 @@ bool OptiXDevice::denoise_filter_color_preprocess(DenoiseContext &context, const
 
   const int work_size = buffer_params.width * buffer_params.height;
 
-  void *args[] = {&context.render_buffers->buffer.device_pointer,
+  DeviceKernelArguments args(&context.render_buffers->buffer.device_pointer,
-                  const_cast<int *>(&buffer_params.full_x),
+                             &buffer_params.full_x,
-                  const_cast<int *>(&buffer_params.full_y),
+                             &buffer_params.full_y,
-                  const_cast<int *>(&buffer_params.width),
+                             &buffer_params.width,
-                  const_cast<int *>(&buffer_params.height),
+                             &buffer_params.height,
-                  const_cast<int *>(&buffer_params.offset),
+                             &buffer_params.offset,
-                  const_cast<int *>(&buffer_params.stride),
+                             &buffer_params.stride,
-                  const_cast<int *>(&buffer_params.pass_stride),
+                             &buffer_params.pass_stride,
-                  const_cast<int *>(&pass.denoised_offset)};
+                             &pass.denoised_offset);
 
   return denoiser_.queue.enqueue(DEVICE_KERNEL_FILTER_COLOR_PREPROCESS, work_size, args);
 }
@@ -799,20 +844,20 @@ bool OptiXDevice::denoise_filter_color_postprocess(DenoiseContext &context,
 
   const int work_size = buffer_params.width * buffer_params.height;
 
-  void *args[] = {&context.render_buffers->buffer.device_pointer,
+  DeviceKernelArguments args(&context.render_buffers->buffer.device_pointer,
-                  const_cast<int *>(&buffer_params.full_x),
+                             &buffer_params.full_x,
-                  const_cast<int *>(&buffer_params.full_y),
+                             &buffer_params.full_y,
-                  const_cast<int *>(&buffer_params.width),
+                             &buffer_params.width,
-                  const_cast<int *>(&buffer_params.height),
+                             &buffer_params.height,
-                  const_cast<int *>(&buffer_params.offset),
+                             &buffer_params.offset,
-                  const_cast<int *>(&buffer_params.stride),
+                             &buffer_params.stride,
-                  const_cast<int *>(&buffer_params.pass_stride),
+                             &buffer_params.pass_stride,
-                  const_cast<int *>(&context.num_samples),
+                             &context.num_samples,
-                  const_cast<int *>(&pass.noisy_offset),
+                             &pass.noisy_offset,
-                  const_cast<int *>(&pass.denoised_offset),
+                             &pass.denoised_offset,
-                  const_cast<int *>(&context.pass_sample_count),
+                             &context.pass_sample_count,
-                  const_cast<int *>(&pass.num_components),
+                             &pass.num_components,
-                  const_cast<bool *>(&pass.use_compositing)};
+                             &pass.use_compositing);
 
   return denoiser_.queue.enqueue(DEVICE_KERNEL_FILTER_COLOR_POSTPROCESS, work_size, args);
 }
@@ -870,35 +915,33 @@ bool OptiXDevice::denoise_create_if_needed(DenoiseContext &context)
 
 bool OptiXDevice::denoise_configure_if_needed(DenoiseContext &context)
 {
-  if (denoiser_.is_configured && (denoiser_.configured_size.x == context.buffer_params.width &&
+  /* Limit maximum tile size denoiser can be invoked with. */
-                                  denoiser_.configured_size.y == context.buffer_params.height)) {
+  const int2 tile_size = make_int2(min(context.buffer_params.width, 4096),
+                                   min(context.buffer_params.height, 4096));
+
+  if (denoiser_.is_configured &&
+      (denoiser_.configured_size.x == tile_size.x && denoiser_.configured_size.y == tile_size.y)) {
     return true;
   }
 
-  const BufferParams &buffer_params = context.buffer_params;
-
-  OptixDenoiserSizes sizes = {};
   optix_assert(optixDenoiserComputeMemoryResources(
-      denoiser_.optix_denoiser, buffer_params.width, buffer_params.height, &sizes));
+      denoiser_.optix_denoiser, tile_size.x, tile_size.y, &denoiser_.sizes));
-
-  /* Denoiser is invoked on whole images only, so no overlap needed (would be used for tiling). */
-  denoiser_.scratch_size = sizes.withoutOverlapScratchSizeInBytes;
-  denoiser_.scratch_offset = sizes.stateSizeInBytes;
 
   /* Allocate denoiser state if tile size has changed since last setup. */
-  denoiser_.state.alloc_to_device(denoiser_.scratch_offset + denoiser_.scratch_size);
+  denoiser_.state.alloc_to_device(denoiser_.sizes.stateSizeInBytes +
+                                  denoiser_.sizes.withOverlapScratchSizeInBytes);
 
   /* Initialize denoiser state for the current tile size. */
   const OptixResult result = optixDenoiserSetup(
       denoiser_.optix_denoiser,
       0, /* Work around bug in r495 drivers that causes artifacts when denoiser setup is called
             on a stream that is not the default stream */
-      buffer_params.width,
+      tile_size.x + denoiser_.sizes.overlapWindowSizeInPixels * 2,
-      buffer_params.height,
+      tile_size.y + denoiser_.sizes.overlapWindowSizeInPixels * 2,
       denoiser_.state.device_pointer,
-      denoiser_.scratch_offset,
+      denoiser_.sizes.stateSizeInBytes,
-      denoiser_.state.device_pointer + denoiser_.scratch_offset,
+      denoiser_.state.device_pointer + denoiser_.sizes.stateSizeInBytes,
-      denoiser_.scratch_size);
+      denoiser_.sizes.withOverlapScratchSizeInBytes);
   if (result != OPTIX_SUCCESS) {
     set_error("Failed to set up OptiX denoiser");
     return false;
@@ -907,8 +950,7 @@ bool OptiXDevice::denoise_configure_if_needed(DenoiseContext &context)
   cuda_assert(cuCtxSynchronize());
 
   denoiser_.is_configured = true;
-  denoiser_.configured_size.x = buffer_params.width;
+  denoiser_.configured_size = tile_size;
-  denoiser_.configured_size.y = buffer_params.height;
 
   return true;
 }
@@ -979,18 +1021,20 @@ bool OptiXDevice::denoise_run(DenoiseContext &context, const DenoisePass &pass)
   guide_layers.albedo = albedo_layer;
   guide_layers.normal = normal_layer;
 
-  optix_assert(optixDenoiserInvoke(denoiser_.optix_denoiser,
+  optix_assert(optixUtilDenoiserInvokeTiled(denoiser_.optix_denoiser,
                                             denoiser_.queue.stream(),
                                             &params,
                                             denoiser_.state.device_pointer,
-                                   denoiser_.scratch_offset,
+                                            denoiser_.sizes.stateSizeInBytes,
                                             &guide_layers,
                                             &image_layers,
                                             1,
-                                   0,
+                                            denoiser_.state.device_pointer +
-                                   0,
+                                                denoiser_.sizes.stateSizeInBytes,
-                                   denoiser_.state.device_pointer + denoiser_.scratch_offset,
+                                            denoiser_.sizes.withOverlapScratchSizeInBytes,
-                                   denoiser_.scratch_size));
+                                            denoiser_.sizes.overlapWindowSizeInPixels,
+                                            denoiser_.configured_size.x,
+                                            denoiser_.configured_size.y));
 
   return true;
 }
@@ -1000,6 +1044,13 @@ bool OptiXDevice::build_optix_bvh(BVHOptiX *bvh,
                                   const OptixBuildInput &build_input,
                                   uint16_t num_motion_steps)
 {
+  /* Allocate and build acceleration structures only one at a time, to prevent parallel builds
+   * from running out of memory (since both original and compacted acceleration structure memory
+   * may be allocated at the same time for the duration of this function). The builds would
+   * otherwise happen on the same CUDA stream anyway. */
+  static thread_mutex mutex;
+  thread_scoped_lock lock(mutex);
+
   const CUDAContextScope scope(this);
 
   const bool use_fast_trace_bvh = (bvh->params.bvh_type == BVH_TYPE_STATIC);
@@ -1025,13 +1076,14 @@ bool OptiXDevice::build_optix_bvh(BVHOptiX *bvh,
   optix_assert(optixAccelComputeMemoryUsage(context, &options, &build_input, 1, &sizes));
 
   /* Allocate required output buffers. */
-  device_only_memory<char> temp_mem(this, "optix temp as build mem");
+  device_only_memory<char> temp_mem(this, "optix temp as build mem", true);
   temp_mem.alloc_to_device(align_up(sizes.tempSizeInBytes, 8) + 8);
   if (!temp_mem.device_pointer) {
     /* Make sure temporary memory allocation succeeded. */
     return false;
   }
 
+  /* Acceleration structure memory has to be allocated on the device (not allowed on the host). */
   device_only_memory<char> &out_data = *bvh->as_data;
   if (operation == OPTIX_BUILD_OPERATION_BUILD) {
     assert(out_data.device == this);
@@ -1080,12 +1132,13 @@ bool OptiXDevice::build_optix_bvh(BVHOptiX *bvh,
 
     /* There is no point compacting if the size does not change. */
     if (compacted_size < sizes.outputSizeInBytes) {
-      device_only_memory<char> compacted_data(this, "optix compacted as");
+      device_only_memory<char> compacted_data(this, "optix compacted as", false);
       compacted_data.alloc_to_device(compacted_size);
-      if (!compacted_data.device_pointer)
+      if (!compacted_data.device_pointer) {
         /* Do not compact if memory allocation for compacted acceleration structure fails.
          * Can just use the uncompacted one then, so succeed here regardless. */
         return !have_error();
+      }
 
       optix_assert(optixAccelCompact(
           context, NULL, out_handle, compacted_data.device_pointer, compacted_size, &out_handle));
@@ -1096,6 +1149,8 @@ bool OptiXDevice::build_optix_bvh(BVHOptiX *bvh,
 
       std::swap(out_data.device_size, compacted_data.device_size);
       std::swap(out_data.device_pointer, compacted_data.device_pointer);
+      /* Original acceleration structure memory is freed when 'compacted_data' goes out of scope.
+       */
     }
   }
 
@@ -1178,20 +1233,27 @@ void OptiXDevice::build_bvh(BVH *bvh, Progress &progress, bool refit)
               int ka = max(k0 - 1, curve.first_key);
               int kb = min(k1 + 1, curve.first_key + curve.num_keys - 1);
 
+              index_data[i] = i * 4;
+              float4 *const v = vertex_data.data() + step * num_vertices + index_data[i];
+
+#  if OPTIX_ABI_VERSION >= 55
+              v[0] = make_float4(keys[ka].x, keys[ka].y, keys[ka].z, curve_radius[ka]);
+              v[1] = make_float4(keys[k0].x, keys[k0].y, keys[k0].z, curve_radius[k0]);
+              v[2] = make_float4(keys[k1].x, keys[k1].y, keys[k1].z, curve_radius[k1]);
+              v[3] = make_float4(keys[kb].x, keys[kb].y, keys[kb].z, curve_radius[kb]);
+#  else
               const float4 px = make_float4(keys[ka].x, keys[k0].x, keys[k1].x, keys[kb].x);
               const float4 py = make_float4(keys[ka].y, keys[k0].y, keys[k1].y, keys[kb].y);
               const float4 pz = make_float4(keys[ka].z, keys[k0].z, keys[k1].z, keys[kb].z);
               const float4 pw = make_float4(
                   curve_radius[ka], curve_radius[k0], curve_radius[k1], curve_radius[kb]);
 
-              /* Convert Catmull-Rom data to Bezier spline. */
+              /* Convert Catmull-Rom data to B-spline. */
               static const float4 cr2bsp0 = make_float4(+7, -4, +5, -2) / 6.f;
               static const float4 cr2bsp1 = make_float4(-2, 11, -4, +1) / 6.f;
               static const float4 cr2bsp2 = make_float4(+1, -4, 11, -2) / 6.f;
               static const float4 cr2bsp3 = make_float4(-2, +5, -4, +7) / 6.f;
 
-              index_data[i] = i * 4;
-              float4 *const v = vertex_data.data() + step * num_vertices + index_data[i];
               v[0] = make_float4(
                   dot(cr2bsp0, px), dot(cr2bsp0, py), dot(cr2bsp0, pz), dot(cr2bsp0, pw));
               v[1] = make_float4(
@@ -1200,6 +1262,7 @@ void OptiXDevice::build_bvh(BVH *bvh, Progress &progress, bool refit)
                   dot(cr2bsp2, px), dot(cr2bsp2, py), dot(cr2bsp2, pz), dot(cr2bsp2, pw));
               v[3] = make_float4(
                   dot(cr2bsp3, px), dot(cr2bsp3, py), dot(cr2bsp3, pz), dot(cr2bsp3, pw));
+#  endif
             }
             else {
               BoundBox bounds = BoundBox::empty;
@@ -1241,7 +1304,11 @@ void OptiXDevice::build_bvh(BVH *bvh, Progress &progress, bool refit)
       OptixBuildInput build_input = {};
       if (hair->curve_shape == CURVE_THICK) {
         build_input.type = OPTIX_BUILD_INPUT_TYPE_CURVES;
+#  if OPTIX_ABI_VERSION >= 55
+        build_input.curveArray.curveType = OPTIX_PRIMITIVE_TYPE_ROUND_CATMULLROM;
+#  else
         build_input.curveArray.curveType = OPTIX_PRIMITIVE_TYPE_ROUND_CUBIC_BSPLINE;
+#  endif
         build_input.curveArray.numPrimitives = num_segments;
         build_input.curveArray.vertexBuffers = (CUdeviceptr *)vertex_ptrs.data();
         build_input.curveArray.numVertices = num_vertices;
@@ -1255,7 +1322,7 @@ void OptiXDevice::build_bvh(BVH *bvh, Progress &progress, bool refit)
       }
       else {
         /* Disable visibility test any-hit program, since it is already checked during
-         * intersection. Those trace calls that require anyhit can force it with a ray flag. */
+         * intersection. Those trace calls that require any-hit can force it with a ray flag. */
         build_flags |= OPTIX_GEOMETRY_FLAG_DISABLE_ANYHIT;
 
         build_input.type = OPTIX_BUILD_INPUT_TYPE_CUSTOM_PRIMITIVES;
@@ -1335,6 +1402,86 @@ void OptiXDevice::build_bvh(BVH *bvh, Progress &progress, bool refit)
       build_input.triangleArray.numSbtRecords = 1;
       build_input.triangleArray.primitiveIndexOffset = mesh->prim_offset;
 
+      if (!build_optix_bvh(bvh_optix, operation, build_input, num_motion_steps)) {
+        progress.set_error("Failed to build OptiX acceleration structure");
+      }
+    }
+    else if (geom->geometry_type == Geometry::POINTCLOUD) {
+      /* Build BLAS for points primitives. */
+      PointCloud *const pointcloud = static_cast<PointCloud *const>(geom);
+      const size_t num_points = pointcloud->num_points();
+      if (num_points == 0) {
+        return;
+      }
+
+      size_t num_motion_steps = 1;
+      Attribute *motion_points = pointcloud->attributes.find(ATTR_STD_MOTION_VERTEX_POSITION);
+      if (motion_blur && pointcloud->get_use_motion_blur() && motion_points) {
+        num_motion_steps = pointcloud->get_motion_steps();
+      }
+
+      device_vector<OptixAabb> aabb_data(this, "optix temp aabb data", MEM_READ_ONLY);
+      aabb_data.alloc(num_points * num_motion_steps);
+
+      /* Get AABBs for each motion step. */
+      for (size_t step = 0; step < num_motion_steps; ++step) {
+        /* The center step for motion vertices is not stored in the attribute. */
+        const float3 *points = pointcloud->get_points().data();
+        const float *radius = pointcloud->get_radius().data();
+        size_t center_step = (num_motion_steps - 1) / 2;
+        if (step != center_step) {
+          size_t attr_offset = (step > center_step) ? step - 1 : step;
+          /* Technically this is a float4 array, but sizeof(float3) == sizeof(float4). */
+          points = motion_points->data_float3() + attr_offset * num_points;
+        }
+
+        for (size_t i = 0; i < num_points; ++i) {
+          const PointCloud::Point point = pointcloud->get_point(i);
+          BoundBox bounds = BoundBox::empty;
+          point.bounds_grow(points, radius, bounds);
+
+          const size_t index = step * num_points + i;
+          aabb_data[index].minX = bounds.min.x;
+          aabb_data[index].minY = bounds.min.y;
+          aabb_data[index].minZ = bounds.min.z;
+          aabb_data[index].maxX = bounds.max.x;
+          aabb_data[index].maxY = bounds.max.y;
+          aabb_data[index].maxZ = bounds.max.z;
+        }
+      }
+
+      /* Upload AABB data to GPU. */
+      aabb_data.copy_to_device();
+
+      vector<device_ptr> aabb_ptrs;
+      aabb_ptrs.reserve(num_motion_steps);
+      for (size_t step = 0; step < num_motion_steps; ++step) {
+        aabb_ptrs.push_back(aabb_data.device_pointer + step * num_points * sizeof(OptixAabb));
+      }
+
+      /* Disable visibility test any-hit program, since it is already checked during
+       * intersection. Those trace calls that require anyhit can force it with a ray flag.
+       * For those, force a single any-hit call, so shadow record-all behavior works correctly. */
+      unsigned int build_flags = OPTIX_GEOMETRY_FLAG_DISABLE_ANYHIT |
+                                 OPTIX_GEOMETRY_FLAG_REQUIRE_SINGLE_ANYHIT_CALL;
+      OptixBuildInput build_input = {};
+      build_input.type = OPTIX_BUILD_INPUT_TYPE_CUSTOM_PRIMITIVES;
+#  if OPTIX_ABI_VERSION < 23
+      build_input.aabbArray.aabbBuffers = (CUdeviceptr *)aabb_ptrs.data();
+      build_input.aabbArray.numPrimitives = num_points;
+      build_input.aabbArray.strideInBytes = sizeof(OptixAabb);
+      build_input.aabbArray.flags = &build_flags;
+      build_input.aabbArray.numSbtRecords = 1;
+      build_input.aabbArray.primitiveIndexOffset = pointcloud->prim_offset;
+#  else
+      build_input.customPrimitiveArray.aabbBuffers = (CUdeviceptr *)aabb_ptrs.data();
+      build_input.customPrimitiveArray.numPrimitives = num_points;
+      build_input.customPrimitiveArray.strideInBytes = sizeof(OptixAabb);
+      build_input.customPrimitiveArray.flags = &build_flags;
+      build_input.customPrimitiveArray.numSbtRecords = 1;
+      build_input.customPrimitiveArray.primitiveIndexOffset = pointcloud->prim_offset;
+#  endif
+
       if (!build_optix_bvh(bvh_optix, operation, build_input, num_motion_steps)) {
         progress.set_error("Failed to build OptiX acceleration structure");
       }
@@ -1422,9 +1569,22 @@ void OptiXDevice::build_bvh(BVH *bvh, Progress &progress, bool refit)
           instance.sbtOffset = PG_HITD_MOTION - PG_HITD;
         }
       }
-      else {
+      else if (ob->get_geometry()->geometry_type == Geometry::POINTCLOUD) {
+        /* Use the hit group that has an intersection program for point clouds. */
+        instance.sbtOffset = PG_HITD_POINTCLOUD - PG_HITD;
+
+        /* Also skip point clouds in local trace calls. */
+        instance.visibilityMask |= 4;
+      }
+
+#  if OPTIX_ABI_VERSION < 55
+      /* Cannot disable any-hit program for thick curves, since it needs to filter out end-caps. */
+      else
+#  endif
+      {
         /* Can disable __anyhit__kernel_optix_visibility_test by default (except for thick curves,
          * since it needs to filter out end-caps there).
+
          * It is enabled where necessary (visibility mask exceeds 8 bits or the other any-hit
          * programs like __anyhit__kernel_optix_shadow_all_hit) via OPTIX_RAY_FLAG_ENFORCE_ANYHIT.
          */
@@ -1494,9 +1654,6 @@ void OptiXDevice::build_bvh(BVH *bvh, Progress &progress, bool refit)
         cuMemcpyHtoD(motion_transform_gpu, &motion_transform, motion_transform_size);
         delete[] reinterpret_cast<uint8_t *>(&motion_transform);
 
-        /* Disable instance transform if object uses motion transform already. */
-        instance.flags |= OPTIX_INSTANCE_FLAG_DISABLE_TRANSFORM;
-
         /* Get traversable handle to motion transform. */
         optixConvertPointerToTraversableHandle(context,
                                                motion_transform_gpu,
@@ -1510,10 +1667,6 @@ void OptiXDevice::build_bvh(BVH *bvh, Progress &progress, bool refit)
           /* Set transform matrix. */
           memcpy(instance.transform, &ob->get_tfm(), sizeof(instance.transform));
         }
-        else {
-          /* Disable instance transform if geometry already has it applied to vertex data. */
-          instance.flags |= OPTIX_INSTANCE_FLAG_DISABLE_TRANSFORM;
-        }
       }
     }
 

intern/cycles/device/optix/device_impl.h

@@ -44,6 +44,8 @@ enum {
   PG_HITV, /* __VOLUME__ hit group. */
   PG_HITD_MOTION,
   PG_HITS_MOTION,
+  PG_HITD_POINTCLOUD,
+  PG_HITS_POINTCLOUD,
   PG_CALL_SVM_AO,
   PG_CALL_SVM_BEVEL,
   NUM_PROGRAM_GROUPS
@@ -52,9 +54,9 @@ enum {
 static const int MISS_PROGRAM_GROUP_OFFSET = PG_MISS;
 static const int NUM_MIS_PROGRAM_GROUPS = 1;
 static const int HIT_PROGAM_GROUP_OFFSET = PG_HITD;
-static const int NUM_HIT_PROGRAM_GROUPS = 6;
+static const int NUM_HIT_PROGRAM_GROUPS = 8;
 static const int CALLABLE_PROGRAM_GROUPS_BASE = PG_CALL_SVM_AO;
-static const int NUM_CALLABLE_PROGRAM_GROUPS = 3;
+static const int NUM_CALLABLE_PROGRAM_GROUPS = 2;
 
 /* List of OptiX pipelines. */
 enum { PIP_SHADE_RAYTRACE, PIP_INTERSECT, NUM_PIPELINES };
@@ -98,8 +100,7 @@ class OptiXDevice : public CUDADevice {
     /* OptiX denoiser state and scratch buffers, stored in a single memory buffer.
      * The memory layout goes as following: [denoiser state][scratch buffer]. */
     device_only_memory<unsigned char> state;
-    size_t scratch_offset = 0;
+    OptixDenoiserSizes sizes = {};
-    size_t scratch_size = 0;
 
     bool use_pass_albedo = false;
     bool use_pass_normal = false;

intern/cycles/device/optix/queue.cpp

@@ -47,7 +47,9 @@ static bool is_optix_specific_kernel(DeviceKernel kernel)
           kernel == DEVICE_KERNEL_INTEGRATOR_INTERSECT_VOLUME_STACK);
 }
 
-bool OptiXDeviceQueue::enqueue(DeviceKernel kernel, const int work_size, void *args[])
+bool OptiXDeviceQueue::enqueue(DeviceKernel kernel,
+                               const int work_size,
+                               DeviceKernelArguments const &args)
 {
   if (!is_optix_specific_kernel(kernel)) {
     return CUDADeviceQueue::enqueue(kernel, work_size, args);
@@ -69,7 +71,7 @@ bool OptiXDeviceQueue::enqueue(DeviceKernel kernel, const int work_size, void *a
   cuda_device_assert(
       cuda_device_,
       cuMemcpyHtoDAsync(launch_params_ptr + offsetof(KernelParamsOptiX, path_index_array),
-                        args[0],  // &d_path_index
+                        args.values[0],  // &d_path_index
                         sizeof(device_ptr),
                         cuda_stream_));
 
@@ -78,7 +80,7 @@ bool OptiXDeviceQueue::enqueue(DeviceKernel kernel, const int work_size, void *a
     cuda_device_assert(
         cuda_device_,
         cuMemcpyHtoDAsync(launch_params_ptr + offsetof(KernelParamsOptiX, render_buffer),
-                          args[1],  // &d_render_buffer
+                          args.values[1],  // &d_render_buffer
                           sizeof(device_ptr),
                           cuda_stream_));
   }

intern/cycles/device/optix/queue.h

@@ -31,7 +31,9 @@ class OptiXDeviceQueue : public CUDADeviceQueue {
 
   virtual void init_execution() override;
 
-  virtual bool enqueue(DeviceKernel kernel, const int work_size, void *args[]) override;
+  virtual bool enqueue(DeviceKernel kernel,
+                       const int work_size,
+                       DeviceKernelArguments const &args) override;
 };
 
 CCL_NAMESPACE_END

intern/cycles/device/queue.h

@@ -31,6 +31,72 @@ class device_memory;
 
 struct KernelWorkTile;
 
+/* Container for device kernel arguments with type correctness ensured by API. */
+struct DeviceKernelArguments {
+
+  enum Type {
+    POINTER,
+    INT32,
+    FLOAT32,
+    BOOLEAN,
+    KERNEL_FILM_CONVERT,
+  };
+
+  static const int MAX_ARGS = 16;
+  Type types[MAX_ARGS];
+  void *values[MAX_ARGS];
+  size_t sizes[MAX_ARGS];
+  size_t count = 0;
+
+  DeviceKernelArguments()
+  {
+  }
+
+  template<class T> DeviceKernelArguments(const T *arg)
+  {
+    add(arg);
+  }
+
+  template<class T, class... Args> DeviceKernelArguments(const T *first, Args... args)
+  {
+    add(first);
+    add(args...);
+  }
+
+  void add(const KernelFilmConvert *value)
+  {
+    add(KERNEL_FILM_CONVERT, value, sizeof(KernelFilmConvert));
+  }
+  void add(const device_ptr *value)
+  {
+    add(POINTER, value, sizeof(device_ptr));
+  }
+  void add(const int32_t *value)
+  {
+    add(INT32, value, sizeof(int32_t));
+  }
+  void add(const float *value)
+  {
+    add(FLOAT32, value, sizeof(float));
+  }
+  void add(const bool *value)
+  {
+    add(BOOLEAN, value, 4);
+  }
+  void add(const Type type, const void *value, size_t size)
+  {
+    types[count] = type;
+    values[count] = (void *)value;
+    sizes[count] = size;
+    count++;
+  }
+  template<typename T, typename... Args> void add(const T *first, Args... args)
+  {
+    add(first);
+    add(args...);
+  }
+};
+
 /* Abstraction of a command queue for a device.
  * Provides API to schedule kernel execution in a specific queue with minimal possible overhead
  * from driver side.
@@ -66,7 +132,9 @@ class DeviceQueue {
    * - int: pass pointer to the int
    * - device memory: pass pointer to device_memory.device_pointer
    * Return false if there was an error executing this or a previous kernel. */
-  virtual bool enqueue(DeviceKernel kernel, const int work_size, void *args[]) = 0;
+  virtual bool enqueue(DeviceKernel kernel,
+                       const int work_size,
+                       DeviceKernelArguments const &args) = 0;
 
   /* Wait unit all enqueued kernels have finished execution.
    * Return false if there was an error executing any of the enqueued kernels. */

intern/cycles/graph/node.h

@@ -31,7 +31,7 @@ struct Node;
 struct NodeType;
 struct Transform;
 
-/* Note: in the following macros we use "type const &" instead of "const type &"
+/* NOTE: in the following macros we use "type const &" instead of "const type &"
  * to avoid issues when pasting a pointer type. */
 #define NODE_SOCKET_API_BASE_METHODS(type_, name, string_name) \
   const SocketType *get_##name##_socket() const \

intern/cycles/integrator/pass_accessor_gpu.cpp

@@ -54,30 +54,30 @@ void PassAccessorGPU::run_film_convert_kernels(DeviceKernel kernel,
   if (destination.d_pixels) {
     DCHECK_EQ(destination.stride, 0) << "Custom stride for float destination is not implemented.";
 
-    void *args[] = {const_cast<KernelFilmConvert *>(&kfilm_convert),
+    DeviceKernelArguments args(&kfilm_convert,
-                    const_cast<device_ptr *>(&destination.d_pixels),
+                               &destination.d_pixels,
-                    const_cast<device_ptr *>(&render_buffers->buffer.device_pointer),
+                               &render_buffers->buffer.device_pointer,
-                    const_cast<int *>(&work_size),
+                               &work_size,
-                    const_cast<int *>(&buffer_params.window_width),
+                               &buffer_params.window_width,
-                    const_cast<int *>(&offset),
+                               &offset,
-                    const_cast<int *>(&buffer_params.stride),
+                               &buffer_params.stride,
-                    const_cast<int *>(&destination.offset),
+                               &destination.offset,
-                    const_cast<int *>(&destination_stride)};
+                               &destination_stride);
 
     queue_->enqueue(kernel, work_size, args);
   }
   if (destination.d_pixels_half_rgba) {
     const DeviceKernel kernel_half_float = static_cast<DeviceKernel>(kernel + 1);
 
-    void *args[] = {const_cast<KernelFilmConvert *>(&kfilm_convert),
+    DeviceKernelArguments args(&kfilm_convert,
-                    const_cast<device_ptr *>(&destination.d_pixels_half_rgba),
+                               &destination.d_pixels_half_rgba,
-                    const_cast<device_ptr *>(&render_buffers->buffer.device_pointer),
+                               &render_buffers->buffer.device_pointer,
-                    const_cast<int *>(&work_size),
+                               &work_size,
-                    const_cast<int *>(&buffer_params.window_width),
+                               &buffer_params.window_width,
-                    const_cast<int *>(&offset),
+                               &offset,
-                    const_cast<int *>(&buffer_params.stride),
+                               &buffer_params.stride,
-                    const_cast<int *>(&destination.offset),
+                               &destination.offset,
-                    const_cast<int *>(&destination_stride)};
+                               &destination_stride);
 
     queue_->enqueue(kernel_half_float, work_size, args);
   }

intern/cycles/integrator/path_trace.cpp

@@ -482,7 +482,11 @@ void PathTrace::set_denoiser_params(const DenoiseParams &params)
   }
 
   denoiser_ = Denoiser::create(device_, params);
-  denoiser_->is_cancelled_cb = [this]() { return is_cancel_requested(); };
+
+  /* Only take into account the "immediate" cancel to have interactive rendering responding to
+   * navigation as quickly as possible, but allow to run denoiser after user hit Esc button while
+   * doing offline rendering. */
+  denoiser_->is_cancelled_cb = [this]() { return render_cancel_.is_requested; };
 }
 
 void PathTrace::set_adaptive_sampling(const AdaptiveSampling &adaptive_sampling)
@@ -1089,6 +1093,8 @@ static const char *device_type_for_description(const DeviceType type)
       return "Dummy";
     case DEVICE_MULTI:
       return "Multi";
+    case DEVICE_METAL:
+      return "Metal";
   }
 
   return "UNKNOWN";

intern/cycles/integrator/path_trace_work_gpu.cpp

@@ -258,7 +258,8 @@ void PathTraceWorkGPU::render_samples(RenderStatistics &statistics,
    * become busy after adding new tiles). This is especially important for the shadow catcher which
    * schedules work in halves of available number of paths. */
   work_tile_scheduler_.set_max_num_path_states(max_num_paths_ / 8);
-
+  work_tile_scheduler_.set_accelerated_rt((device_->get_bvh_layout_mask() & BVH_LAYOUT_OPTIX) !=
+                                          0);
   work_tile_scheduler_.reset(effective_buffer_params_,
                              start_sample,
                              samples_num,
@@ -333,7 +334,8 @@ DeviceKernel PathTraceWorkGPU::get_most_queued_kernel() const
 
 void PathTraceWorkGPU::enqueue_reset()
 {
-  void *args[] = {&max_num_paths_};
+  DeviceKernelArguments args(&max_num_paths_);
+
   queue_->enqueue(DEVICE_KERNEL_INTEGRATOR_RESET, max_num_paths_, args);
   queue_->zero_to_device(integrator_queue_counter_);
   queue_->zero_to_device(integrator_shader_sort_counter_);
@@ -404,7 +406,7 @@ bool PathTraceWorkGPU::enqueue_path_iteration()
 
 void PathTraceWorkGPU::enqueue_path_iteration(DeviceKernel kernel, const int num_paths_limit)
 {
-  void *d_path_index = (void *)NULL;
+  device_ptr d_path_index = 0;
 
   /* Create array of path indices for which this kernel is queued to be executed. */
   int work_size = kernel_max_active_main_path_index(kernel);
@@ -415,14 +417,14 @@ void PathTraceWorkGPU::enqueue_path_iteration(DeviceKernel kernel, const int num
   if (kernel_uses_sorting(kernel)) {
     /* Compute array of active paths, sorted by shader. */
     work_size = num_queued;
-    d_path_index = (void *)queued_paths_.device_pointer;
+    d_path_index = queued_paths_.device_pointer;
 
     compute_sorted_queued_paths(
         DEVICE_KERNEL_INTEGRATOR_SORTED_PATHS_ARRAY, kernel, num_paths_limit);
   }
   else if (num_queued < work_size) {
     work_size = num_queued;
-    d_path_index = (void *)queued_paths_.device_pointer;
+    d_path_index = queued_paths_.device_pointer;
 
     if (kernel_is_shadow_path(kernel)) {
       /* Compute array of active shadow paths for specific kernel. */
@@ -441,8 +443,7 @@ void PathTraceWorkGPU::enqueue_path_iteration(DeviceKernel kernel, const int num
   switch (kernel) {
     case DEVICE_KERNEL_INTEGRATOR_INTERSECT_CLOSEST: {
       /* Closest ray intersection kernels with integrator state and render buffer. */
-      void *d_render_buffer = (void *)buffers_->buffer.device_pointer;
+      DeviceKernelArguments args(&d_path_index, &buffers_->buffer.device_pointer, &work_size);
-      void *args[] = {&d_path_index, &d_render_buffer, const_cast<int *>(&work_size)};
 
       queue_->enqueue(kernel, work_size, args);
       break;
@@ -452,7 +453,7 @@ void PathTraceWorkGPU::enqueue_path_iteration(DeviceKernel kernel, const int num
     case DEVICE_KERNEL_INTEGRATOR_INTERSECT_SUBSURFACE:
     case DEVICE_KERNEL_INTEGRATOR_INTERSECT_VOLUME_STACK: {
       /* Ray intersection kernels with integrator state. */
-      void *args[] = {&d_path_index, const_cast<int *>(&work_size)};
+      DeviceKernelArguments args(&d_path_index, &work_size);
 
       queue_->enqueue(kernel, work_size, args);
       break;
@@ -464,8 +465,7 @@ void PathTraceWorkGPU::enqueue_path_iteration(DeviceKernel kernel, const int num
     case DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE:
     case DEVICE_KERNEL_INTEGRATOR_SHADE_VOLUME: {
       /* Shading kernels with integrator state and render buffer. */
-      void *d_render_buffer = (void *)buffers_->buffer.device_pointer;
+      DeviceKernelArguments args(&d_path_index, &buffers_->buffer.device_pointer, &work_size);
-      void *args[] = {&d_path_index, &d_render_buffer, const_cast<int *>(&work_size)};
 
       queue_->enqueue(kernel, work_size, args);
       break;
@@ -483,15 +483,17 @@ void PathTraceWorkGPU::compute_sorted_queued_paths(DeviceKernel kernel,
                                                    const int num_paths_limit)
 {
   int d_queued_kernel = queued_kernel;
-  void *d_counter = integrator_state_gpu_.sort_key_counter[d_queued_kernel];
+  device_ptr d_counter = (device_ptr)integrator_state_gpu_.sort_key_counter[d_queued_kernel];
-  void *d_prefix_sum = (void *)integrator_shader_sort_prefix_sum_.device_pointer;
+  device_ptr d_prefix_sum = integrator_shader_sort_prefix_sum_.device_pointer;
-  assert(d_counter != nullptr && d_prefix_sum != nullptr);
+  assert(d_counter != 0 && d_prefix_sum != 0);
 
   /* Compute prefix sum of number of active paths with each shader. */
   {
     const int work_size = 1;
     int max_shaders = device_scene_->data.max_shaders;
-    void *args[] = {&d_counter, &d_prefix_sum, &max_shaders};
+
+    DeviceKernelArguments args(&d_counter, &d_prefix_sum, &max_shaders);
+
     queue_->enqueue(DEVICE_KERNEL_PREFIX_SUM, work_size, args);
   }
 
@@ -506,15 +508,16 @@ void PathTraceWorkGPU::compute_sorted_queued_paths(DeviceKernel kernel,
      * end of the array since compaction would need to do less work. */
     const int work_size = kernel_max_active_main_path_index(queued_kernel);
 
-    void *d_queued_paths = (void *)queued_paths_.device_pointer;
+    device_ptr d_queued_paths = queued_paths_.device_pointer;
-    void *d_num_queued_paths = (void *)num_queued_paths_.device_pointer;
+    device_ptr d_num_queued_paths = num_queued_paths_.device_pointer;
-    void *args[] = {const_cast<int *>(&work_size),
+
-                    const_cast<int *>(&num_paths_limit),
+    DeviceKernelArguments args(&work_size,
+                               &num_paths_limit,
                                &d_queued_paths,
                                &d_num_queued_paths,
                                &d_counter,
                                &d_prefix_sum,
-                    &d_queued_kernel};
+                               &d_queued_kernel);
 
     queue_->enqueue(kernel, work_size, args);
   }
@@ -526,10 +529,10 @@ void PathTraceWorkGPU::compute_queued_paths(DeviceKernel kernel, DeviceKernel qu
 
   /* Launch kernel to fill the active paths arrays. */
   const int work_size = kernel_max_active_main_path_index(queued_kernel);
-  void *d_queued_paths = (void *)queued_paths_.device_pointer;
+  device_ptr d_queued_paths = queued_paths_.device_pointer;
-  void *d_num_queued_paths = (void *)num_queued_paths_.device_pointer;
+  device_ptr d_num_queued_paths = num_queued_paths_.device_pointer;
-  void *args[] = {
+
-      const_cast<int *>(&work_size), &d_queued_paths, &d_num_queued_paths, &d_queued_kernel};
+  DeviceKernelArguments args(&work_size, &d_queued_paths, &d_num_queued_paths, &d_queued_kernel);
 
   queue_->zero_to_device(num_queued_paths_);
   queue_->enqueue(kernel, work_size, args);
@@ -605,15 +608,17 @@ void PathTraceWorkGPU::compact_paths(const int num_active_paths,
 {
   /* Compact fragmented path states into the start of the array, moving any paths
    * with index higher than the number of active paths into the gaps. */
-  void *d_compact_paths = (void *)queued_paths_.device_pointer;
+  device_ptr d_compact_paths = queued_paths_.device_pointer;
-  void *d_num_queued_paths = (void *)num_queued_paths_.device_pointer;
+  device_ptr d_num_queued_paths = num_queued_paths_.device_pointer;
 
   /* Create array with terminated paths that we can write to. */
   {
     /* TODO: can the work size be reduced here? */
     int offset = num_active_paths;
     int work_size = num_active_paths;
-    void *args[] = {&work_size, &d_compact_paths, &d_num_queued_paths, &offset};
+
+    DeviceKernelArguments args(&work_size, &d_compact_paths, &d_num_queued_paths, &offset);
+
     queue_->zero_to_device(num_queued_paths_);
     queue_->enqueue(terminated_paths_kernel, work_size, args);
   }
@@ -622,8 +627,10 @@ void PathTraceWorkGPU::compact_paths(const int num_active_paths,
    * than the number of active paths. */
   {
     int work_size = max_active_path_index;
-    void *args[] = {
+
-        &work_size, &d_compact_paths, &d_num_queued_paths, const_cast<int *>(&num_active_paths)};
+    DeviceKernelArguments args(
+        &work_size, &d_compact_paths, &d_num_queued_paths, &num_active_paths);
+
     queue_->zero_to_device(num_queued_paths_);
     queue_->enqueue(compact_paths_kernel, work_size, args);
   }
@@ -638,8 +645,10 @@ void PathTraceWorkGPU::compact_paths(const int num_active_paths,
     int work_size = num_compact_paths;
     int active_states_offset = 0;
     int terminated_states_offset = num_active_paths;
-    void *args[] = {
+
-        &d_compact_paths, &active_states_offset, &terminated_states_offset, &work_size};
+    DeviceKernelArguments args(
+        &d_compact_paths, &active_states_offset, &terminated_states_offset, &work_size);
+
     queue_->enqueue(compact_kernel, work_size, args);
   }
 }
@@ -768,14 +777,12 @@ void PathTraceWorkGPU::enqueue_work_tiles(DeviceKernel kernel,
 
   queue_->copy_to_device(work_tiles_);
 
-  void *d_work_tiles = (void *)work_tiles_.device_pointer;
+  device_ptr d_work_tiles = work_tiles_.device_pointer;
-  void *d_render_buffer = (void *)buffers_->buffer.device_pointer;
+  device_ptr d_render_buffer = buffers_->buffer.device_pointer;
 
   /* Launch kernel. */
-  void *args[] = {&d_work_tiles,
+  DeviceKernelArguments args(
-                  const_cast<int *>(&num_work_tiles),
+      &d_work_tiles, &num_work_tiles, &d_render_buffer, &max_tile_work_size);
-                  &d_render_buffer,
-                  const_cast<int *>(&max_tile_work_size)};
 
   queue_->enqueue(kernel, max_tile_work_size * num_work_tiles, args);
 
@@ -965,16 +972,16 @@ int PathTraceWorkGPU::adaptive_sampling_convergence_check_count_active(float thr
 
   const int work_size = effective_buffer_params_.width * effective_buffer_params_.height;
 
-  void *args[] = {&buffers_->buffer.device_pointer,
+  DeviceKernelArguments args(&buffers_->buffer.device_pointer,
-                  const_cast<int *>(&effective_buffer_params_.full_x),
+                             &effective_buffer_params_.full_x,
-                  const_cast<int *>(&effective_buffer_params_.full_y),
+                             &effective_buffer_params_.full_y,
-                  const_cast<int *>(&effective_buffer_params_.width),
+                             &effective_buffer_params_.width,
-                  const_cast<int *>(&effective_buffer_params_.height),
+                             &effective_buffer_params_.height,
                              &threshold,
                              &reset,
                              &effective_buffer_params_.offset,
                              &effective_buffer_params_.stride,
-                  &num_active_pixels.device_pointer};
+                             &num_active_pixels.device_pointer);
 
   queue_->enqueue(DEVICE_KERNEL_ADAPTIVE_SAMPLING_CONVERGENCE_CHECK, work_size, args);
 
@@ -988,13 +995,13 @@ void PathTraceWorkGPU::enqueue_adaptive_sampling_filter_x()
 {
   const int work_size = effective_buffer_params_.height;
 
-  void *args[] = {&buffers_->buffer.device_pointer,
+  DeviceKernelArguments args(&buffers_->buffer.device_pointer,
                              &effective_buffer_params_.full_x,
                              &effective_buffer_params_.full_y,
                              &effective_buffer_params_.width,
                              &effective_buffer_params_.height,
                              &effective_buffer_params_.offset,
-                  &effective_buffer_params_.stride};
+                             &effective_buffer_params_.stride);
 
   queue_->enqueue(DEVICE_KERNEL_ADAPTIVE_SAMPLING_CONVERGENCE_FILTER_X, work_size, args);
 }
@@ -1003,13 +1010,13 @@ void PathTraceWorkGPU::enqueue_adaptive_sampling_filter_y()
 {
   const int work_size = effective_buffer_params_.width;
 
-  void *args[] = {&buffers_->buffer.device_pointer,
+  DeviceKernelArguments args(&buffers_->buffer.device_pointer,
                              &effective_buffer_params_.full_x,
                              &effective_buffer_params_.full_y,
                              &effective_buffer_params_.width,
                              &effective_buffer_params_.height,
                              &effective_buffer_params_.offset,
-                  &effective_buffer_params_.stride};
+                             &effective_buffer_params_.stride);
 
   queue_->enqueue(DEVICE_KERNEL_ADAPTIVE_SAMPLING_CONVERGENCE_FILTER_Y, work_size, args);
 }
@@ -1018,10 +1025,10 @@ void PathTraceWorkGPU::cryptomatte_postproces()
 {
   const int work_size = effective_buffer_params_.width * effective_buffer_params_.height;
 
-  void *args[] = {&buffers_->buffer.device_pointer,
+  DeviceKernelArguments args(&buffers_->buffer.device_pointer,
-                  const_cast<int *>(&work_size),
+                             &work_size,
                              &effective_buffer_params_.offset,
-                  &effective_buffer_params_.stride};
+                             &effective_buffer_params_.stride);
 
   queue_->enqueue(DEVICE_KERNEL_CRYPTOMATTE_POSTPROCESS, work_size, args);
 }
@@ -1070,8 +1077,9 @@ int PathTraceWorkGPU::shadow_catcher_count_possible_splits()
   queue_->zero_to_device(num_queued_paths_);
 
   const int work_size = max_active_main_path_index_;
-  void *d_num_queued_paths = (void *)num_queued_paths_.device_pointer;
+  device_ptr d_num_queued_paths = num_queued_paths_.device_pointer;
-  void *args[] = {const_cast<int *>(&work_size), &d_num_queued_paths};
+
+  DeviceKernelArguments args(&work_size, &d_num_queued_paths);
 
   queue_->enqueue(DEVICE_KERNEL_INTEGRATOR_SHADOW_CATCHER_COUNT_POSSIBLE_SPLITS, work_size, args);
   queue_->copy_from_device(num_queued_paths_);

intern/cycles/integrator/render_scheduler.cpp

@@ -112,7 +112,7 @@ int RenderScheduler::get_rendered_sample() const
 {
   DCHECK_GT(get_num_rendered_samples(), 0);
 
-  return start_sample_ + get_num_rendered_samples() - 1;
+  return start_sample_ + get_num_rendered_samples() - 1 - sample_offset_;
 }
 
 int RenderScheduler::get_num_rendered_samples() const
@@ -406,6 +406,9 @@ bool RenderScheduler::set_postprocess_render_work(RenderWork *render_work)
     any_scheduled = true;
   }
 
+  /* Force update. */
+  any_scheduled = true;
+
   if (any_scheduled) {
     render_work->display.update = true;
   }
@@ -874,7 +877,8 @@ int RenderScheduler::get_num_samples_to_path_trace() const
    * is to ensure that the final render is pixel-matched regardless of how many samples per second
    * compute device can do. */
 
-  return adaptive_sampling_.align_samples(path_trace_start_sample, num_samples_to_render);
+  return adaptive_sampling_.align_samples(path_trace_start_sample - sample_offset_,
+                                          num_samples_to_render);
 }
 
 int RenderScheduler::get_num_samples_during_navigation(int resolution_divider) const

intern/cycles/integrator/shader_eval.cpp

@@ -158,14 +158,16 @@ bool ShaderEval::eval_gpu(Device *device,
 
   /* Execute work on GPU in chunk, so we can cancel.
    * TODO : query appropriate size from device.*/
-  const int64_t chunk_size = 65536;
+  const int32_t chunk_size = 65536;
 
-  void *d_input = (void *)input.device_pointer;
+  device_ptr d_input = input.device_pointer;
-  void *d_output = (void *)output.device_pointer;
+  device_ptr d_output = output.device_pointer;
 
-  for (int64_t d_offset = 0; d_offset < work_size; d_offset += chunk_size) {
+  assert(work_size <= 0x7fffffff);
-    int64_t d_work_size = std::min(chunk_size, work_size - d_offset);
+  for (int32_t d_offset = 0; d_offset < int32_t(work_size); d_offset += chunk_size) {
-    void *args[] = {&d_input, &d_output, &d_offset, &d_work_size};
+    int32_t d_work_size = std::min(chunk_size, int32_t(work_size) - d_offset);
+
+    DeviceKernelArguments args(&d_input, &d_output, &d_offset, &d_work_size);
 
     queue->enqueue(kernel, d_work_size, args);
     queue->synchronize();

intern/cycles/integrator/tile.cpp

@@ -46,7 +46,8 @@ ccl_device_inline uint round_up_to_power_of_two(uint x)
   return next_power_of_two(x);
 }
 
-TileSize tile_calculate_best_size(const int2 &image_size,
+TileSize tile_calculate_best_size(const bool accel_rt,
+                                  const int2 &image_size,
                                   const int num_samples,
                                   const int max_num_path_states,
                                   const float scrambling_distance)
@@ -73,7 +74,7 @@ TileSize tile_calculate_best_size(const int2 &image_size,
 
   TileSize tile_size;
   const int num_path_states_per_sample = max_num_path_states / num_samples;
-  if (scrambling_distance < 0.9f) {
+  if (scrambling_distance < 0.9f && accel_rt) {
     /* Prefer large tiles for scrambling distance, bounded by max num path states. */
     tile_size.width = min(image_size.x, max_num_path_states);
     tile_size.height = min(image_size.y, max(max_num_path_states / tile_size.width, 1));

intern/cycles/integrator/tile.h

@@ -49,7 +49,8 @@ std::ostream &operator<<(std::ostream &os, const TileSize &tile_size);
  * of active path states.
  * Will attempt to provide best guess to keep path tracing threads of a device as localized as
  * possible, and have as many threads active for every tile as possible. */
-TileSize tile_calculate_best_size(const int2 &image_size,
+TileSize tile_calculate_best_size(const bool accel_rt,
+                                  const int2 &image_size,
                                   const int num_samples,
                                   const int max_num_path_states,
                                   const float scrambling_distance);

intern/cycles/integrator/work_tile_scheduler.cpp

@@ -28,6 +28,11 @@ WorkTileScheduler::WorkTileScheduler()
 {
 }
 
+void WorkTileScheduler::set_accelerated_rt(bool accelerated_rt)
+{
+  accelerated_rt_ = accelerated_rt;
+}
+
 void WorkTileScheduler::set_max_num_path_states(int max_num_path_states)
 {
   max_num_path_states_ = max_num_path_states;
@@ -61,7 +66,7 @@ void WorkTileScheduler::reset(const BufferParams &buffer_params,
 void WorkTileScheduler::reset_scheduler_state()
 {
   tile_size_ = tile_calculate_best_size(
-      image_size_px_, samples_num_, max_num_path_states_, scrambling_distance_);
+      accelerated_rt_, image_size_px_, samples_num_, max_num_path_states_, scrambling_distance_);
 
   VLOG(3) << "Will schedule tiles of size " << tile_size_;
 

intern/cycles/integrator/work_tile_scheduler.h

@@ -31,6 +31,9 @@ class WorkTileScheduler {
  public:
   WorkTileScheduler();
 
+  /* To indicate if there is accelerated RT support. */
+  void set_accelerated_rt(bool state);
+
   /* MAximum path states which are allowed to be used by a single scheduled work tile.
    *
    * Affects the scheduled work size: the work size will be as big as possible, but will not exceed
@@ -55,6 +58,9 @@ class WorkTileScheduler {
  protected:
   void reset_scheduler_state();
 
+  /* Used to indicate if there is accelerated ray tracing. */
+  bool accelerated_rt_ = false;
+
   /* Maximum allowed path states to be used.
    *
    * TODO(sergey): Naming can be improved. The fact that this is a limiting factor based on the

intern/cycles/kernel/CMakeLists.txt

@@ -179,11 +179,14 @@ set(SRC_KERNEL_GEOM_HEADERS
   geom/curve.h
   geom/curve_intersect.h
   geom/motion_curve.h
+  geom/motion_point.h
   geom/motion_triangle.h
   geom/motion_triangle_intersect.h
   geom/motion_triangle_shader.h
   geom/object.h
   geom/patch.h
+  geom/point.h
+  geom/point_intersect.h
   geom/primitive.h
   geom/shader_data.h
   geom/subd_triangle.h
@@ -207,6 +210,7 @@ set(SRC_KERNEL_BVH_HEADERS
   bvh/volume.h
   bvh/volume_all.h
   bvh/embree.h
+  bvh/metal.h
 )
 
 set(SRC_KERNEL_CAMERA_HEADERS

intern/cycles/kernel/bvh/bvh.h

@@ -31,6 +31,10 @@
 #  include "kernel/bvh/embree.h"
 #endif
 
+#ifdef __METALRT__
+#  include "kernel/bvh/metal.h"
+#endif
+
 #include "kernel/bvh/types.h"
 #include "kernel/bvh/util.h"
 
@@ -38,31 +42,31 @@
 
 CCL_NAMESPACE_BEGIN
 
-#ifndef __KERNEL_OPTIX__
+#if !defined(__KERNEL_GPU_RAYTRACING__)
 
 /* Regular BVH traversal */
 
 #  include "kernel/bvh/nodes.h"
 
 #  define BVH_FUNCTION_NAME bvh_intersect
-#  define BVH_FUNCTION_FEATURES 0
+#  define BVH_FUNCTION_FEATURES BVH_POINTCLOUD
 #  include "kernel/bvh/traversal.h"
 
 #  if defined(__HAIR__)
 #    define BVH_FUNCTION_NAME bvh_intersect_hair
-#    define BVH_FUNCTION_FEATURES BVH_HAIR
+#    define BVH_FUNCTION_FEATURES BVH_HAIR | BVH_POINTCLOUD
 #    include "kernel/bvh/traversal.h"
 #  endif
 
 #  if defined(__OBJECT_MOTION__)
 #    define BVH_FUNCTION_NAME bvh_intersect_motion
-#    define BVH_FUNCTION_FEATURES BVH_MOTION
+#    define BVH_FUNCTION_FEATURES BVH_MOTION | BVH_POINTCLOUD
 #    include "kernel/bvh/traversal.h"
 #  endif
 
 #  if defined(__HAIR__) && defined(__OBJECT_MOTION__)
 #    define BVH_FUNCTION_NAME bvh_intersect_hair_motion
-#    define BVH_FUNCTION_FEATURES BVH_HAIR | BVH_MOTION
+#    define BVH_FUNCTION_FEATURES BVH_HAIR | BVH_MOTION | BVH_POINTCLOUD
 #    include "kernel/bvh/traversal.h"
 #  endif
 
@@ -98,26 +102,27 @@ CCL_NAMESPACE_BEGIN
 
 #  if defined(__SHADOW_RECORD_ALL__)
 #    define BVH_FUNCTION_NAME bvh_intersect_shadow_all
-#    define BVH_FUNCTION_FEATURES 0
+#    define BVH_FUNCTION_FEATURES BVH_POINTCLOUD
 #    include "kernel/bvh/shadow_all.h"
 
 #    if defined(__HAIR__)
 #      define BVH_FUNCTION_NAME bvh_intersect_shadow_all_hair
-#      define BVH_FUNCTION_FEATURES BVH_HAIR
+#      define BVH_FUNCTION_FEATURES BVH_HAIR | BVH_POINTCLOUD
 #      include "kernel/bvh/shadow_all.h"
 #    endif
 
 #    if defined(__OBJECT_MOTION__)
 #      define BVH_FUNCTION_NAME bvh_intersect_shadow_all_motion
-#      define BVH_FUNCTION_FEATURES BVH_MOTION
+#      define BVH_FUNCTION_FEATURES BVH_MOTION | BVH_POINTCLOUD
 #      include "kernel/bvh/shadow_all.h"
 #    endif
 
 #    if defined(__HAIR__) && defined(__OBJECT_MOTION__)
 #      define BVH_FUNCTION_NAME bvh_intersect_shadow_all_hair_motion
-#      define BVH_FUNCTION_FEATURES BVH_HAIR | BVH_MOTION
+#      define BVH_FUNCTION_FEATURES BVH_HAIR | BVH_MOTION | BVH_POINTCLOUD
 #      include "kernel/bvh/shadow_all.h"
 #    endif
+
 #  endif /* __SHADOW_RECORD_ALL__ */
 
 /* Record all intersections - Volume BVH traversal. */
@@ -139,7 +144,7 @@ CCL_NAMESPACE_BEGIN
 #  undef BVH_NAME_EVAL
 #  undef BVH_FUNCTION_FULL_NAME
 
-#endif /* __KERNEL_OPTIX__ */
+#endif /* !defined(__KERNEL_GPU_RAYTRACING__) */
 
 ccl_device_inline bool scene_intersect_valid(ccl_private const Ray *ray)
 {
@@ -205,7 +210,95 @@ ccl_device_intersect bool scene_intersect(KernelGlobals kg,
   isect->type = p5;
 
   return p5 != PRIMITIVE_NONE;
-#else /* __KERNEL_OPTIX__ */
+#elif defined(__METALRT__)
+
+  if (!scene_intersect_valid(ray)) {
+    isect->t = ray->t;
+    isect->type = PRIMITIVE_NONE;
+    return false;
+  }
+
+#  if defined(__KERNEL_DEBUG__)
+  if (is_null_instance_acceleration_structure(metal_ancillaries->accel_struct)) {
+    isect->t = ray->t;
+    isect->type = PRIMITIVE_NONE;
+    kernel_assert(!"Invalid metal_ancillaries->accel_struct pointer");
+    return false;
+  }
+
+  if (is_null_intersection_function_table(metal_ancillaries->ift_default)) {
+    isect->t = ray->t;
+    isect->type = PRIMITIVE_NONE;
+    kernel_assert(!"Invalid ift_default");
+    return false;
+  }
+#  endif
+
+  metal::raytracing::ray r(ray->P, ray->D, 0.0f, ray->t);
+  metalrt_intersector_type metalrt_intersect;
+
+  if (!kernel_data.bvh.have_curves) {
+    metalrt_intersect.assume_geometry_type(metal::raytracing::geometry_type::triangle);
+  }
+
+  MetalRTIntersectionPayload payload;
+  payload.u = 0.0f;
+  payload.v = 0.0f;
+  payload.visibility = visibility;
+
+  typename metalrt_intersector_type::result_type intersection;
+
+  uint ray_mask = visibility & 0xFF;
+  if (0 == ray_mask && (visibility & ~0xFF) != 0) {
+    ray_mask = 0xFF;
+    /* No further intersector setup required: Default MetalRT behavior is any-hit. */
+  }
+  else if (visibility & PATH_RAY_SHADOW_OPAQUE) {
+    /* No further intersector setup required: Shadow ray early termination is controlled by the
+     * intersection handler */
+  }
+
+#  if defined(__METALRT_MOTION__)
+  payload.time = ray->time;
+  intersection = metalrt_intersect.intersect(r,
+                                             metal_ancillaries->accel_struct,
+                                             ray_mask,
+                                             ray->time,
+                                             metal_ancillaries->ift_default,
+                                             payload);
+#  else
+  intersection = metalrt_intersect.intersect(
+      r, metal_ancillaries->accel_struct, ray_mask, metal_ancillaries->ift_default, payload);
+#  endif
+
+  if (intersection.type == intersection_type::none) {
+    isect->t = ray->t;
+    isect->type = PRIMITIVE_NONE;
+
+    return false;
+  }
+
+  isect->t = intersection.distance;
+
+  isect->prim = payload.prim;
+  isect->type = payload.type;
+  isect->object = intersection.user_instance_id;
+
+  isect->t = intersection.distance;
+  if (intersection.type == intersection_type::triangle) {
+    isect->u = 1.0f - intersection.triangle_barycentric_coord.y -
+               intersection.triangle_barycentric_coord.x;
+    isect->v = intersection.triangle_barycentric_coord.x;
+  }
+  else {
+    isect->u = payload.u;
+    isect->v = payload.v;
+  }
+
+  return isect->type != PRIMITIVE_NONE;
+
+#else
+
   if (!scene_intersect_valid(ray)) {
     return false;
   }
@@ -289,7 +382,69 @@ ccl_device_intersect bool scene_intersect_local(KernelGlobals kg,
              p5);
 
   return p5;
-#  else /* __KERNEL_OPTIX__ */
+#  elif defined(__METALRT__)
+  if (!scene_intersect_valid(ray)) {
+    if (local_isect) {
+      local_isect->num_hits = 0;
+    }
+    return false;
+  }
+
+#    if defined(__KERNEL_DEBUG__)
+  if (is_null_instance_acceleration_structure(metal_ancillaries->accel_struct)) {
+    if (local_isect) {
+      local_isect->num_hits = 0;
+    }
+    kernel_assert(!"Invalid metal_ancillaries->accel_struct pointer");
+    return false;
+  }
+
+  if (is_null_intersection_function_table(metal_ancillaries->ift_local)) {
+    if (local_isect) {
+      local_isect->num_hits = 0;
+    }
+    kernel_assert(!"Invalid ift_local");
+    return false;
+  }
+#    endif
+
+  metal::raytracing::ray r(ray->P, ray->D, 0.0f, ray->t);
+  metalrt_intersector_type metalrt_intersect;
+
+  metalrt_intersect.force_opacity(metal::raytracing::forced_opacity::non_opaque);
+  if (!kernel_data.bvh.have_curves) {
+    metalrt_intersect.assume_geometry_type(metal::raytracing::geometry_type::triangle);
+  }
+
+  MetalRTIntersectionLocalPayload payload;
+  payload.local_object = local_object;
+  payload.max_hits = max_hits;
+  payload.local_isect.num_hits = 0;
+  if (lcg_state) {
+    payload.has_lcg_state = true;
+    payload.lcg_state = *lcg_state;
+  }
+  payload.result = false;
+
+  typename metalrt_intersector_type::result_type intersection;
+
+#    if defined(__METALRT_MOTION__)
+  intersection = metalrt_intersect.intersect(
+      r, metal_ancillaries->accel_struct, 0xFF, ray->time, metal_ancillaries->ift_local, payload);
+#    else
+  intersection = metalrt_intersect.intersect(
+      r, metal_ancillaries->accel_struct, 0xFF, metal_ancillaries->ift_local, payload);
+#    endif
+
+  if (lcg_state) {
+    *lcg_state = payload.lcg_state;
+  }
+  *local_isect = payload.local_isect;
+
+  return payload.result;
+
+#  else
+
   if (!scene_intersect_valid(ray)) {
     if (local_isect) {
       local_isect->num_hits = 0;
@@ -406,7 +561,67 @@ ccl_device_intersect bool scene_intersect_shadow_all(KernelGlobals kg,
   *throughput = __uint_as_float(p1);
 
   return p5;
-#  else /* __KERNEL_OPTIX__ */
+#  elif defined(__METALRT__)
+
+  if (!scene_intersect_valid(ray)) {
+    return false;
+  }
+
+#    if defined(__KERNEL_DEBUG__)
+  if (is_null_instance_acceleration_structure(metal_ancillaries->accel_struct)) {
+    kernel_assert(!"Invalid metal_ancillaries->accel_struct pointer");
+    return false;
+  }
+
+  if (is_null_intersection_function_table(metal_ancillaries->ift_shadow)) {
+    kernel_assert(!"Invalid ift_shadow");
+    return false;
+  }
+#    endif
+
+  metal::raytracing::ray r(ray->P, ray->D, 0.0f, ray->t);
+  metalrt_intersector_type metalrt_intersect;
+
+  metalrt_intersect.force_opacity(metal::raytracing::forced_opacity::non_opaque);
+  if (!kernel_data.bvh.have_curves) {
+    metalrt_intersect.assume_geometry_type(metal::raytracing::geometry_type::triangle);
+  }
+
+  MetalRTIntersectionShadowPayload payload;
+  payload.visibility = visibility;
+  payload.max_hits = max_hits;
+  payload.num_hits = 0;
+  payload.num_recorded_hits = 0;
+  payload.throughput = 1.0f;
+  payload.result = false;
+  payload.state = state;
+
+  uint ray_mask = visibility & 0xFF;
+  if (0 == ray_mask && (visibility & ~0xFF) != 0) {
+    ray_mask = 0xFF;
+  }
+
+  typename metalrt_intersector_type::result_type intersection;
+
+#    if defined(__METALRT_MOTION__)
+  payload.time = ray->time;
+  intersection = metalrt_intersect.intersect(r,
+                                             metal_ancillaries->accel_struct,
+                                             ray_mask,
+                                             ray->time,
+                                             metal_ancillaries->ift_shadow,
+                                             payload);
+#    else
+  intersection = metalrt_intersect.intersect(
+      r, metal_ancillaries->accel_struct, ray_mask, metal_ancillaries->ift_shadow, payload);
+#    endif
+
+  *num_recorded_hits = payload.num_recorded_hits;
+  *throughput = payload.throughput;
+
+  return payload.result;
+
+#  else
   if (!scene_intersect_valid(ray)) {
     *num_recorded_hits = 0;
     *throughput = 1.0f;
@@ -503,7 +718,76 @@ ccl_device_intersect bool scene_intersect_volume(KernelGlobals kg,
   isect->type = p5;
 
   return p5 != PRIMITIVE_NONE;
-#  else /* __KERNEL_OPTIX__ */
+#  elif defined(__METALRT__)
+
+  if (!scene_intersect_valid(ray)) {
+    return false;
+  }
+#    if defined(__KERNEL_DEBUG__)
+  if (is_null_instance_acceleration_structure(metal_ancillaries->accel_struct)) {
+    kernel_assert(!"Invalid metal_ancillaries->accel_struct pointer");
+    return false;
+  }
+
+  if (is_null_intersection_function_table(metal_ancillaries->ift_default)) {
+    kernel_assert(!"Invalid ift_default");
+    return false;
+  }
+#    endif
+
+  metal::raytracing::ray r(ray->P, ray->D, 0.0f, ray->t);
+  metalrt_intersector_type metalrt_intersect;
+
+  metalrt_intersect.force_opacity(metal::raytracing::forced_opacity::non_opaque);
+  if (!kernel_data.bvh.have_curves) {
+    metalrt_intersect.assume_geometry_type(metal::raytracing::geometry_type::triangle);
+  }
+
+  MetalRTIntersectionPayload payload;
+  payload.visibility = visibility;
+
+  typename metalrt_intersector_type::result_type intersection;
+
+  uint ray_mask = visibility & 0xFF;
+  if (0 == ray_mask && (visibility & ~0xFF) != 0) {
+    ray_mask = 0xFF;
+  }
+
+#    if defined(__METALRT_MOTION__)
+  payload.time = ray->time;
+  intersection = metalrt_intersect.intersect(r,
+                                             metal_ancillaries->accel_struct,
+                                             ray_mask,
+                                             ray->time,
+                                             metal_ancillaries->ift_default,
+                                             payload);
+#    else
+  intersection = metalrt_intersect.intersect(
+      r, metal_ancillaries->accel_struct, ray_mask, metal_ancillaries->ift_default, payload);
+#    endif
+
+  if (intersection.type == intersection_type::none) {
+    return false;
+  }
+
+  isect->prim = payload.prim;
+  isect->type = payload.type;
+  isect->object = intersection.user_instance_id;
+
+  isect->t = intersection.distance;
+  if (intersection.type == intersection_type::triangle) {
+    isect->u = 1.0f - intersection.triangle_barycentric_coord.y -
+               intersection.triangle_barycentric_coord.x;
+    isect->v = intersection.triangle_barycentric_coord.x;
+  }
+  else {
+    isect->u = payload.u;
+    isect->v = payload.v;
+  }
+
+  return isect->type != PRIMITIVE_NONE;
+
+#  else
   if (!scene_intersect_valid(ray)) {
     return false;
   }

intern/cycles/kernel/bvh/metal.h

@@ -0,0 +1,47 @@
+/*
+ * Copyright 2021 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+struct MetalRTIntersectionPayload {
+  uint visibility;
+  float u, v;
+  int prim;
+  int type;
+#if defined(__METALRT_MOTION__)
+  float time;
+#endif
+};
+
+struct MetalRTIntersectionLocalPayload {
+  uint local_object;
+  uint lcg_state;
+  short max_hits;
+  bool has_lcg_state;
+  bool result;
+  LocalIntersection local_isect;
+};
+
+struct MetalRTIntersectionShadowPayload {
+  uint visibility;
+#if defined(__METALRT_MOTION__)
+  float time;
+#endif
+  int state;
+  float throughput;
+  short max_hits;
+  short num_hits;
+  short num_recorded_hits;
+  bool result;
+};

intern/cycles/kernel/bvh/shadow_all.h

@@ -28,6 +28,7 @@
  * without new features slowing things down.
  *
  * BVH_HAIR: hair curve rendering
+ * BVH_POINTCLOUD: point cloud rendering
  * BVH_MOTION: motion blur rendering
  */
 
@@ -173,7 +174,7 @@ ccl_device_inline
               case PRIMITIVE_MOTION_CURVE_THICK:
               case PRIMITIVE_CURVE_RIBBON:
               case PRIMITIVE_MOTION_CURVE_RIBBON: {
-                if ((type & PRIMITIVE_ALL_MOTION) && kernel_data.bvh.use_bvh_steps) {
+                if ((type & PRIMITIVE_MOTION) && kernel_data.bvh.use_bvh_steps) {
                   const float2 prim_time = kernel_tex_fetch(__prim_time, prim_addr);
                   if (ray->time < prim_time.x || ray->time > prim_time.y) {
                     hit = false;
@@ -199,6 +200,34 @@ ccl_device_inline
                 break;
               }
 #endif
+#if BVH_FEATURE(BVH_POINTCLOUD)
+              case PRIMITIVE_POINT:
+              case PRIMITIVE_MOTION_POINT: {
+                if ((type & PRIMITIVE_MOTION) && kernel_data.bvh.use_bvh_steps) {
+                  const float2 prim_time = kernel_tex_fetch(__prim_time, prim_addr);
+                  if (ray->time < prim_time.x || ray->time > prim_time.y) {
+                    hit = false;
+                    break;
+                  }
+                }
+
+                const int point_object = (object == OBJECT_NONE) ?
+                                             kernel_tex_fetch(__prim_object, prim_addr) :
+                                             object;
+                const int point_prim = kernel_tex_fetch(__prim_index, prim_addr);
+                const int point_type = kernel_tex_fetch(__prim_type, prim_addr);
+                hit = point_intersect(kg,
+                                      &isect,
+                                      P,
+                                      dir,
+                                      t_max_current,
+                                      point_object,
+                                      point_prim,
+                                      ray->time,
+                                      point_type);
+                break;
+              }
+#endif /* BVH_FEATURE(BVH_POINTCLOUD) */
               default: {
                 hit = false;
                 break;
@@ -226,7 +255,7 @@ ccl_device_inline
               bool record_intersection = true;
 
               /* Always use baked shadow transparency for curves. */
-              if (isect.type & PRIMITIVE_ALL_CURVE) {
+              if (isect.type & PRIMITIVE_CURVE) {
                 *throughput *= intersection_curve_shadow_transparency(
                     kg, isect.object, isect.prim, isect.u);