cycles: Add an nvrtc based cubin cli compiler.

nvcc is very picky regarding compiler versions, severely limiting the compiler we can use, this commit adds a nvrtc based compiler that'll allow us to build the cubins even if the host compiler is unsupported. for details see D2913.

Differential Revision: http://developer.blender.org/D2913

1 files changed, 51 insertions, 39 deletions

diff --git a/intern/cycles/kernel/CMakeLists.txt b/intern/cycles/kernel/CMakeLists.txt
index 8f7bc7996a4..3b76b3403e7 100644
--- a/intern/cycles/kernel/CMakeLists.txt
+++ b/intern/cycles/kernel/CMakeLists.txt
@@ -356,55 +356,67 @@ if(WITH_CYCLES_CUDA_BINARIES)
 	set(cuda_cubins)
 
 	macro(CYCLES_CUDA_KERNEL_ADD arch name flags sources experimental)
+		set(cuda_cubin ${name}_${arch}.cubin)
+		set(cuda_kernel_src "/kernels/cuda/${name}.cu")
+
+		set(cuda_flags
+			-D CCL_NAMESPACE_BEGIN=
+			-D CCL_NAMESPACE_END=
+			-D NVCC
+			-m ${CUDA_BITS}
+			-I ${CMAKE_CURRENT_SOURCE_DIR}/..
+			-I ${CMAKE_CURRENT_SOURCE_DIR}/kernels/cuda
+			--use_fast_math
+			-o ${CMAKE_CURRENT_BINARY_DIR}/${cuda_cubin})
+
 		if(${experimental})
-			set(flags ${flags} -D__KERNEL_EXPERIMENTAL__)
+			set(cuda_flags ${cuda_flags} -D __KERNEL_EXPERIMENTAL__)
 			set(name ${name}_experimental)
 		endif()
 
-		set(cuda_cubin ${name}_${arch}.cubin)
-
 		if(WITH_CYCLES_DEBUG)
-			set(cuda_debug_flags "-D__KERNEL_DEBUG__")
-		else()
-			set(cuda_debug_flags "")
+			set(cuda_flags ${cuda_flags} -D __KERNEL_DEBUG__)
 		endif()
 
-		set(cuda_nvcc_command ${CUDA_NVCC_EXECUTABLE})
-		set(cuda_nvcc_version ${CUDA_VERSION})
-
-		set(cuda_version_flags "-D__KERNEL_CUDA_VERSION__=${cuda_nvcc_version}")
-		set(cuda_math_flags "--use_fast_math")
-
-		set(cuda_kernel_src "/kernels/cuda/${name}.cu")
-
-		add_custom_command(
-			OUTPUT ${cuda_cubin}
-			COMMAND ${cuda_nvcc_command}
-					-arch=${arch}
-					${CUDA_NVCC_FLAGS}
-					-m${CUDA_BITS}
-					--cubin ${CMAKE_CURRENT_SOURCE_DIR}${cuda_kernel_src}
-					-o ${CMAKE_CURRENT_BINARY_DIR}/${cuda_cubin}
-					--ptxas-options="-v"
-					${cuda_arch_flags}
-					${cuda_version_flags}
-					${cuda_math_flags}
-					${flags}
-					${cuda_debug_flags}
-					-I${CMAKE_CURRENT_SOURCE_DIR}/..
-					-DCCL_NAMESPACE_BEGIN=
-					-DCCL_NAMESPACE_END=
-					-DNVCC
-			DEPENDS ${sources})
+		if(WITH_CYCLES_CUBIN_COMPILER)
+			string(SUBSTRING ${arch} 3 -1 CUDA_ARCH)
+
+			# Needed to find libnvrtc-builtins.so. Can't do it from inside
+			# 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")
+			elseif(UNIX)
+				set(CUBIN_CC_ENV ${CMAKE_COMMAND}
+					-E env LD_LIBRARY_PATH="${CUDA_TOOLKIT_ROOT_DIR}/lib64")
+			endif()
 
+			add_custom_command(
+				OUTPUT ${cuda_cubin}
+				COMMAND ${CUBIN_CC_ENV}
+						"$<TARGET_FILE:cycles_cubin_cc>"
+						-target ${CUDA_ARCH}
+						-i ${CMAKE_CURRENT_SOURCE_DIR}${cuda_kernel_src}
+						${cuda_flags}
+						-v
+						-cuda-toolkit-dir "${CUDA_TOOLKIT_ROOT_DIR}"
+				DEPENDS ${sources} cycles_cubin_cc)
+		else()
+			add_custom_command(
+				OUTPUT ${cuda_cubin}
+				COMMAND ${CUDA_NVCC_EXECUTABLE}
+						-arch=${arch}
+						${CUDA_NVCC_FLAGS}
+						--cubin
+						${CMAKE_CURRENT_SOURCE_DIR}${cuda_kernel_src}
+						--ptxas-options="-v"
+						${cuda_flags}
+				DEPENDS ${sources})
+		endif()
 		delayed_install("${CMAKE_CURRENT_BINARY_DIR}" "${cuda_cubin}" ${CYCLES_INSTALL_PATH}/lib)
 		list(APPEND cuda_cubins ${cuda_cubin})
 
-		unset(cuda_extra_flags)
 		unset(cuda_debug_flags)
-
-		unset(cuda_nvcc_command)
-		unset(cuda_nvcc_version)
 	endmacro()
 
 	foreach(arch ${CYCLES_CUDA_BINARIES_ARCH})
@@ -412,12 +424,12 @@ if(WITH_CYCLES_CUDA_BINARIES)
 			message(STATUS "CUDA binaries for ${arch} disabled, not supported by CUDA 9.")
 		else()
 			# Compile regular kernel
-			CYCLES_CUDA_KERNEL_ADD(${arch} kernel "" "${cuda_sources}" FALSE)
 			CYCLES_CUDA_KERNEL_ADD(${arch} filter "" "${cuda_filter_sources}" FALSE)
+			CYCLES_CUDA_KERNEL_ADD(${arch} kernel "" "${cuda_sources}" FALSE)
 
 			if(WITH_CYCLES_CUDA_SPLIT_KERNEL_BINARIES)
 				# Compile split kernel
-				CYCLES_CUDA_KERNEL_ADD(${arch} kernel_split "-D__SPLIT__" ${cuda_sources} FALSE)
+				CYCLES_CUDA_KERNEL_ADD(${arch} kernel_split "-D __SPLIT__" ${cuda_sources} FALSE)
 			endif()
 		endif()
 	endforeach()