Cycles: Add CUDA 11 build support

With this patch the build system checks whether the "CUDA10_NVCC_EXECUTABLE" CMake
variable is set and if so will use that to build sm_30 kernels. Similarily for sm_8x kernels it
checks "CUDA11_NVCC_EXECUTABLE". All other kernels are built using the default CUDA
toolkit. This makes it possible to use either the CUDA 10 or CUDA 11 toolkit by default and
only selectively use the other for the kernels where its a hard requirement.

Reviewed By: brecht

Differential Revision: https://developer.blender.org/D9179

5 files changed, 47 insertions, 16 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 6f705ffbe44..b15bbb7486b 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -369,7 +369,7 @@ option(WITH_CYCLES_CUDA_BINARIES    "Build Cycles CUDA binaries" 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)
 mark_as_advanced(WITH_CYCLES_CUDA_BUILD_SERIAL)
-set(CYCLES_CUDA_BINARIES_ARCH sm_30 sm_35 sm_37 sm_50 sm_52 sm_60 sm_61 sm_70 sm_75 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")
 mark_as_advanced(CYCLES_CUDA_BINARIES_ARCH)
 unset(PLATFORM_DEFAULT)
 option(WITH_CYCLES_LOGGING  "Build Cycles with logging support" ON)
diff --git a/build_files/buildbot/worker_compile.py b/build_files/buildbot/worker_compile.py
index 340f507df4c..d5482571c11 100644
--- a/build_files/buildbot/worker_compile.py
+++ b/build_files/buildbot/worker_compile.py
@@ -44,13 +44,17 @@ def get_cmake_options(builder):
     optix_sdk_dir = os.path.join(builder.blender_dir, '..', '..', 'NVIDIA-Optix-SDK')
     options.append('-DOPTIX_ROOT_DIR:PATH=' + optix_sdk_dir)
 
-    # Workers have multiple CUDA versions installed. Select 10.1 for Blender 2.83 releases.
+    # Workaround to build sm_30 kernels with CUDA 10, since CUDA 11 no longer supports that architecture
     if builder.platform == 'win':
-        options.append('-DCUDA_TOOLKIT_ROOT_DIR:PATH=C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.1')
-        options.append('-DCUDA_NVCC_EXECUTABLE:FILEPATH=C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.1/bin/nvcc.exe')
+        options.append('-DCUDA10_TOOLKIT_ROOT_DIR:PATH=C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.1')
+        options.append('-DCUDA10_NVCC_EXECUTABLE:FILEPATH=C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.1/bin/nvcc.exe')
+        options.append('-DCUDA11_TOOLKIT_ROOT_DIR:PATH=C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.1')
+        options.append('-DCUDA11_NVCC_EXECUTABLE:FILEPATH=C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.1/bin/nvcc.exe')
     elif builder.platform == 'linux':
-        options.append('-DCUDA_TOOLKIT_ROOT_DIR:PATH=/usr/local/cuda-10.1')
-        options.append('-DCUDA_NVCC_EXECUTABLE:FILEPATH=/usr/local/cuda-10.1/bin/nvcc')
+        options.append('-DCUDA10_TOOLKIT_ROOT_DIR:PATH=/usr/local/cuda-10.1')
+        options.append('-DCUDA10_NVCC_EXECUTABLE:FILEPATH=/usr/local/cuda-10.1/bin/nvcc')
+        options.append('-DCUDA11_TOOLKIT_ROOT_DIR:PATH=/usr/local/cuda-11.1')
+        options.append('-DCUDA11_NVCC_EXECUTABLE:FILEPATH=/usr/local/cuda-11.1/bin/nvcc')
 
     options.append("-C" + os.path.join(builder.blender_dir, config_file))
     options.append("-DCMAKE_INSTALL_PREFIX=%s" % (builder.install_dir))
diff --git a/build_files/cmake/config/blender_release.cmake b/build_files/cmake/config/blender_release.cmake
index 2d52fb22c86..e1b7560e4fc 100644
--- a/build_files/cmake/config/blender_release.cmake
+++ b/build_files/cmake/config/blender_release.cmake
@@ -52,7 +52,7 @@ set(WITH_USD                 ON  CACHE BOOL "" FORCE)
 set(WITH_MEM_JEMALLOC          ON  CACHE BOOL "" FORCE)
 set(WITH_CYCLES_CUDA_BINARIES  ON  CACHE BOOL "" FORCE)
 set(WITH_CYCLES_CUBIN_COMPILER OFF CACHE BOOL "" FORCE)
-set(CYCLES_CUDA_BINARIES_ARCH sm_30;sm_35;sm_37;sm_50;sm_52;sm_60;sm_61;sm_70;sm_75;compute_75 CACHE STRING "" FORCE)
+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 "" FORCE)
 set(WITH_CYCLES_DEVICE_OPTIX   ON CACHE BOOL "" FORCE)
 
 # platform dependent options
diff --git a/intern/cycles/CMakeLists.txt b/intern/cycles/CMakeLists.txt
index 121c8bdad6e..762d37996cf 100644
--- a/intern/cycles/CMakeLists.txt
+++ b/intern/cycles/CMakeLists.txt
@@ -313,7 +313,7 @@ if(WITH_CYCLES_CUDA_BINARIES AND (NOT WITH_CYCLES_CUBIN_COMPILER))
       set(MAX_MSVC 1910)
     elseif(${CUDA_VERSION} EQUAL "9.1")
       set(MAX_MSVC 1911)
-    elseif(${CUDA_VERSION} LESS "11.0")
+    elseif(${CUDA_VERSION} VERSION_GREATER_EQUAL 10.0)
       set(MAX_MSVC 1999)
     endif()
     if(NOT MSVC_VERSION LESS ${MAX_MSVC} OR CMAKE_C_COMPILER_ID MATCHES "Clang")
diff --git a/intern/cycles/kernel/CMakeLists.txt b/intern/cycles/kernel/CMakeLists.txt
index 5e408c55328..d2fbff29776 100644
--- a/intern/cycles/kernel/CMakeLists.txt
+++ b/intern/cycles/kernel/CMakeLists.txt
@@ -375,11 +375,11 @@ if(WITH_CYCLES_CUDA_BINARIES)
   set(CUDA_VERSION "${CUDA_VERSION_MAJOR}${CUDA_VERSION_MINOR}")
 
   # warn for other versions
-  if((CUDA_VERSION MATCHES "101") OR (CUDA_VERSION MATCHES "102"))
+  if((CUDA_VERSION MATCHES "101") OR (CUDA_VERSION MATCHES "102") OR (CUDA_VERSION MATCHES "111"))
   else()
     message(WARNING
       "CUDA version ${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR} detected, "
-      "build may succeed but only CUDA 10.1 and 10.2 are officially supported")
+      "build may succeed but only CUDA 10.1, 10.2 and 11.1 are officially supported")
   endif()
 
   # build for each arch
@@ -445,10 +445,10 @@ if(WITH_CYCLES_CUDA_BINARIES)
       # cycles_cubin_cc since the env variable is read before main()
       if(APPLE)
         set(CUBIN_CC_ENV ${CMAKE_COMMAND}
-          -E env DYLD_LIBRARY_PATH="${CUDA_TOOLKIT_ROOT_DIR}/lib")
+          -E env DYLD_LIBRARY_PATH="${cuda_toolkit_root_dir}/lib")
       elseif(UNIX)
         set(CUBIN_CC_ENV ${CMAKE_COMMAND}
-          -E env LD_LIBRARY_PATH="${CUDA_TOOLKIT_ROOT_DIR}/lib64")
+          -E env LD_LIBRARY_PATH="${cuda_toolkit_root_dir}/lib64")
       endif()
 
       add_custom_command(
@@ -459,12 +459,12 @@ if(WITH_CYCLES_CUDA_BINARIES)
             -i ${CMAKE_CURRENT_SOURCE_DIR}${cuda_kernel_src}
             ${cuda_flags}
             -v
-            -cuda-toolkit-dir "${CUDA_TOOLKIT_ROOT_DIR}"
+            -cuda-toolkit-dir "${cuda_toolkit_root_dir}"
         DEPENDS ${kernel_sources} cycles_cubin_cc)
     else()
       add_custom_command(
         OUTPUT ${cuda_file}
-        COMMAND ${CUDA_NVCC_EXECUTABLE}
+        COMMAND ${cuda_nvcc_executable}
             -arch=${arch}
             ${CUDA_NVCC_FLAGS}
             --${format}
@@ -481,11 +481,35 @@ if(WITH_CYCLES_CUDA_BINARIES)
 
   set(prev_arch "none")
   foreach(arch ${CYCLES_CUDA_BINARIES_ARCH})
-    if(${arch} MATCHES "sm_2.")
+    if(${arch} MATCHES ".*_2.")
       message(STATUS "CUDA binaries for ${arch} are no longer supported, skipped.")
-    elseif(${arch} MATCHES "sm_7." AND ${CUDA_VERSION} LESS 100)
+    elseif(${arch} MATCHES ".*_30")
+      if(DEFINED CUDA10_NVCC_EXECUTABLE)
+        set(cuda_nvcc_executable ${CUDA10_NVCC_EXECUTABLE})
+        set(cuda_toolkit_root_dir ${CUDA10_TOOLKIT_ROOT_DIR})
+      elseif(${CUDA_VERSION} LESS 110) # Support for sm_30 was removed in CUDA 11
+        set(cuda_nvcc_executable ${CUDA_NVCC_EXECUTABLE})
+        set(cuda_toolkit_root_dir ${CUDA_TOOLKIT_ROOT_DIR})
+      else()
+        message(STATUS "CUDA binaries for ${arch} require CUDA 10 or earlier, skipped.")
+      endif()
+    elseif(${arch} MATCHES ".*_7." AND ${CUDA_VERSION} LESS 100)
       message(STATUS "CUDA binaries for ${arch} require CUDA 10.0+, skipped.")
+    elseif(${arch} MATCHES ".*_8.")
+      if(DEFINED CUDA11_NVCC_EXECUTABLE)
+        set(cuda_nvcc_executable ${CUDA11_NVCC_EXECUTABLE})
+        set(cuda_toolkit_root_dir ${CUDA11_TOOLKIT_ROOT_DIR})
+      elseif(${CUDA_VERSION} GREATER_EQUAL 111) # Support for sm_86 was introduced in CUDA 11
+        set(cuda_nvcc_executable ${CUDA_NVCC_EXECUTABLE})
+        set(cuda_toolkit_root_dir ${CUDA_TOOLKIT_ROOT_DIR})
+      else()
+        message(STATUS "CUDA binaries for ${arch} require CUDA 11.1+, skipped.")
+      endif()
     else()
+      set(cuda_nvcc_executable ${CUDA_NVCC_EXECUTABLE})
+      set(cuda_toolkit_root_dir ${CUDA_TOOLKIT_ROOT_DIR})
+    endif()
+    if(DEFINED cuda_nvcc_executable AND DEFINED cuda_toolkit_root_dir)
       # Compile regular kernel
       CYCLES_CUDA_KERNEL_ADD(${arch} ${prev_arch} filter "" "${cuda_filter_sources}" FALSE)
       CYCLES_CUDA_KERNEL_ADD(${arch} ${prev_arch} kernel "" "${cuda_sources}" FALSE)
@@ -498,6 +522,9 @@ if(WITH_CYCLES_CUDA_BINARIES)
       if(WITH_CYCLES_CUDA_BUILD_SERIAL)
         set(prev_arch ${arch})
       endif()
+
+      unset(cuda_nvcc_executable)
+      unset(cuda_toolkit_root_dir)
     endif()
   endforeach()