Cycles: add initial CUDA 10.0 support, but only recommend use for Turing cards.

There may still be rendering errors when used for older graphics cards.

7 files changed, 33 insertions, 9 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 7359151d9a3..1912b8f6a34 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -408,7 +408,7 @@ 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 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_72 sm_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/slave_compile.py b/build_files/buildbot/slave_compile.py
index d5ec5630e39..375357670d7 100644
--- a/build_files/buildbot/slave_compile.py
+++ b/build_files/buildbot/slave_compile.py
@@ -126,7 +126,6 @@ if 'cmake' in builder:
     # Prepare CMake options needed to configure cuda binaries compilation, 64bit only.
     if bits == 64:
         cuda_cmake_options.append("-DWITH_CYCLES_CUDA_BINARIES=%s" % ('ON' if build_cubins else 'OFF'))
-        cuda_cmake_options.append("-DCYCLES_CUDA_BINARIES_ARCH=sm_30;sm_35;sm_37;sm_50;sm_52;sm_60;sm_61;sm_70")
         if build_cubins or 'cuda' in targets:
             cuda_cmake_options.append("-DCUDA_64_BIT_DEVICE_CODE=ON")
 
diff --git a/build_files/cmake/config/blender_release.cmake b/build_files/cmake/config/blender_release.cmake
index 23d6b55b9c0..a43470f4e1e 100644
--- a/build_files/cmake/config/blender_release.cmake
+++ b/build_files/cmake/config/blender_release.cmake
@@ -54,7 +54,7 @@ set(WITH_X11_XF86VMODE       ON  CACHE BOOL "" FORCE)
 set(WITH_PLAYER              ON  CACHE BOOL "" FORCE)
 set(WITH_MEM_JEMALLOC        ON  CACHE BOOL "" FORCE)
 set(WITH_CYCLES_CUDA_BINARIES ON  CACHE BOOL "" FORCE)
-set(CYCLES_CUDA_BINARIES_ARCH sm_30;sm_35;sm_37;sm_50;sm_52;sm_60;sm_61 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_72;sm_75 CACHE STRING "" FORCE)
 
 # platform dependent options
 if(UNIX AND NOT APPLE)
diff --git a/extern/cuew/src/cuew.c b/extern/cuew/src/cuew.c
index ad216e66452..a22f6fda570 100644
--- a/extern/cuew/src/cuew.c
+++ b/extern/cuew/src/cuew.c
@@ -619,7 +619,11 @@ static int cuewNvrtcInit(void) {
   /* Library paths. */
 #ifdef _WIN32
   /* Expected in c:/windows/system or similar, no path needed. */
-  const char *nvrtc_paths[] = {"nvrtc64_80.dll", "nvrtc64_90.dll", "nvrtc64_91.dll", NULL};
+  const char *nvrtc_paths[] = {"nvrtc64_80.dll",
+                               "nvrtc64_90.dll",
+                               "nvrtc64_91.dll",
+                               "nvrtc64_10_0.dll",
+                               NULL};
 #elif defined(__APPLE__)
   /* Default installation path. */
   const char *nvrtc_paths[] = {"/usr/local/cuda/lib/libnvrtc.dylib", NULL};
diff --git a/intern/cycles/CMakeLists.txt b/intern/cycles/CMakeLists.txt
index 3028b7c67d3..8f0d838e881 100644
--- a/intern/cycles/CMakeLists.txt
+++ b/intern/cycles/CMakeLists.txt
@@ -315,6 +315,13 @@ if(WITH_CYCLES_CUDA_BINARIES AND (NOT WITH_CYCLES_CUBIN_COMPILER))
 	endif()
 endif()
 
+# NVRTC gives wrong rendering result in CUDA 10.0, so we must use NVCC.
+if(WITH_CYCLES_CUDA_BINARIES AND WITH_CYCLES_CUBIN_COMPILER)
+	if(${CUDA_VERSION} VERSION_GREATER_EQUAL 10.0)
+		message(STATUS "cycles_cubin_cc not supported for CUDA 10.0+, using nvcc instead.")
+		set(WITH_CYCLES_CUBIN_COMPILER OFF)
+	endif()
+endif()
 
 # Subdirectories
 
diff --git a/intern/cycles/kernel/CMakeLists.txt b/intern/cycles/kernel/CMakeLists.txt
index d4145225b77..3ba43a9e0bd 100644
--- a/intern/cycles/kernel/CMakeLists.txt
+++ b/intern/cycles/kernel/CMakeLists.txt
@@ -343,11 +343,11 @@ if(WITH_CYCLES_CUDA_BINARIES)
 	set(CUDA_VERSION "${CUDA_VERSION_MAJOR}${CUDA_VERSION_MINOR}")
 
 	# warn for other versions
-	if(CUDA_VERSION MATCHES "90" OR CUDA_VERSION MATCHES "91")
+	if(CUDA_VERSION MATCHES "90" OR CUDA_VERSION MATCHES "91" OR CUDA_VERSION MATCHES "100")
 	else()
 		message(WARNING
 			"CUDA version ${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR} detected, "
-			"build may succeed but only CUDA 9.0 and 9.1 are officially supported")
+			"build may succeed but only CUDA 9.0, 9.1 and 10.0 are officially supported")
 	endif()
 
 	# build for each arch
@@ -442,6 +442,8 @@ if(WITH_CYCLES_CUDA_BINARIES)
 	foreach(arch ${CYCLES_CUDA_BINARIES_ARCH})
 		if(${arch} MATCHES "sm_2.")
 			message(STATUS "CUDA binaries for ${arch} are no longer supported, skipped.")
+		elseif(${arch} MATCHES "sm_7." AND NOT (${CUDA_VERSION} VERSION_GREATER_EQUAL 10.0))
+			message(STATUS "CUDA binaries for ${arch} require CUDA 10.0+, skipped.")
 		else()
 			# Compile regular kernel
 			CYCLES_CUDA_KERNEL_ADD(${arch} ${prev_arch} filter "" "${cuda_filter_sources}" FALSE)
diff --git a/intern/cycles/kernel/kernels/cuda/kernel_config.h b/intern/cycles/kernel/kernels/cuda/kernel_config.h
index 3808898c5ca..6d41dc15785 100644
--- a/intern/cycles/kernel/kernels/cuda/kernel_config.h
+++ b/intern/cycles/kernel/kernels/cuda/kernel_config.h
@@ -52,8 +52,8 @@
 #  define CUDA_KERNEL_MAX_REGISTERS 63
 #  define CUDA_KERNEL_BRANCHED_MAX_REGISTERS 63
 
-/* 5.0, 5.2, 5.3, 6.0, 6.1 */
-#elif __CUDA_ARCH__ >= 500
+/* 5.x, 6.x */
+#elif __CUDA_ARCH__ <= 699
 #  define CUDA_MULTIPRESSOR_MAX_REGISTERS 65536
 #  define CUDA_MULTIPROCESSOR_MAX_BLOCKS 32
 #  define CUDA_BLOCK_MAX_THREADS 1024
@@ -62,13 +62,25 @@
 /* tunable parameters */
 #  define CUDA_THREADS_BLOCK_WIDTH 16
 /* CUDA 9.0 seems to cause slowdowns on high-end Pascal cards unless we increase the number of registers */
-#  if __CUDACC_VER_MAJOR__ == 9 && __CUDA_ARCH__ >= 600
+#  if __CUDACC_VER_MAJOR__ >= 9 && __CUDA_ARCH__ >= 600
 #    define CUDA_KERNEL_MAX_REGISTERS 64
 #  else
 #    define CUDA_KERNEL_MAX_REGISTERS 48
 #  endif
 #  define CUDA_KERNEL_BRANCHED_MAX_REGISTERS 63
 
+/* 7.x */
+#elif __CUDA_ARCH__ <= 799
+#  define CUDA_MULTIPRESSOR_MAX_REGISTERS 65536
+#  define CUDA_MULTIPROCESSOR_MAX_BLOCKS 32
+#  define CUDA_BLOCK_MAX_THREADS 1024
+#  define CUDA_THREAD_MAX_REGISTERS 255
+
+/* tunable parameters */
+#  define CUDA_THREADS_BLOCK_WIDTH 16
+#  define CUDA_KERNEL_MAX_REGISTERS 64
+#  define CUDA_KERNEL_BRANCHED_MAX_REGISTERS 72
+
 
 /* unknown architecture */
 #else