[Bf-blender-cvs] [b3ac3d13a2a] master: Cycles: Add option for building CUDA kernels sequentially

[Lukas Stockner]

Commit: b3ac3d13a2abdb3bd15b677b4e478a83118f4c0a
Author: Lukas Stockner
Date:   Mon Aug 20 01:17:34 2018 -0700
Branches: master
https://developer.blender.org/rBb3ac3d13a2abdb3bd15b677b4e478a83118f4c0a

Cycles: Add option for building CUDA kernels sequentially

Building the CUDA kernels takes quite a bit of memory, and when building all of
them the combined usage can be too much on some systems (especially VMs).

Therefore, this patch adds an option to force the build system to build them
sequentially by making each build step depend on the previous kernel.

Reviewers: brecht, sergey

Differential Revision: https://developer.blender.org/D3623

===================================================================

M	CMakeLists.txt
M	intern/cycles/kernel/CMakeLists.txt

===================================================================

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 05ad76996a3..9e88c89cb0e 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -411,6 +411,8 @@ option(WITH_CYCLES_OSL				"Build Cycles with OSL support" ${_init_CYCLES_OSL})
 option(WITH_CYCLES_OPENSUBDIV		"Build Cycles with OpenSubdiv support" ${_init_CYCLES_OPENSUBDIV})
 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")
 mark_as_advanced(CYCLES_CUDA_BINARIES_ARCH)
 unset(PLATFORM_DEFAULT)
diff --git a/intern/cycles/kernel/CMakeLists.txt b/intern/cycles/kernel/CMakeLists.txt
index c4cad17429d..118e18374a1 100644
--- a/intern/cycles/kernel/CMakeLists.txt
+++ b/intern/cycles/kernel/CMakeLists.txt
@@ -357,8 +357,14 @@ if(WITH_CYCLES_CUDA_BINARIES)
 	)
 	set(cuda_cubins)
 
-	macro(CYCLES_CUDA_KERNEL_ADD arch name flags sources experimental)
+	macro(CYCLES_CUDA_KERNEL_ADD arch prev_arch name flags sources experimental)
 		set(cuda_cubin ${name}_${arch}.cubin)
+
+		set(kernel_sources ${sources})
+		if(NOT ${prev_arch} STREQUAL "none")
+			set(kernel_sources ${kernel_sources} ${name}_${prev_arch}.cubin)
+		endif()
+
 		set(cuda_kernel_src "/kernels/cuda/${name}.cu")
 
 		set(cuda_flags
@@ -402,7 +408,7 @@ if(WITH_CYCLES_CUDA_BINARIES)
 						${cuda_flags}
 						-v
 						-cuda-toolkit-dir "${CUDA_TOOLKIT_ROOT_DIR}"
-				DEPENDS ${sources} cycles_cubin_cc)
+				DEPENDS ${kernel_sources} cycles_cubin_cc)
 		else()
 			add_custom_command(
 				OUTPUT ${cuda_cubin}
@@ -413,7 +419,7 @@ if(WITH_CYCLES_CUDA_BINARIES)
 						${CMAKE_CURRENT_SOURCE_DIR}${cuda_kernel_src}
 						--ptxas-options="-v"
 						${cuda_flags}
-				DEPENDS ${sources})
+				DEPENDS ${kernel_sources})
 		endif()
 		delayed_install("${CMAKE_CURRENT_BINARY_DIR}" "${cuda_cubin}" ${CYCLES_INSTALL_PATH}/lib)
 		list(APPEND cuda_cubins ${cuda_cubin})
@@ -421,18 +427,23 @@ if(WITH_CYCLES_CUDA_BINARIES)
 		unset(cuda_debug_flags)
 	endmacro()
 
+	set(prev_arch "none")
 	foreach(arch ${CYCLES_CUDA_BINARIES_ARCH})
 		if(${arch} MATCHES "sm_2.")
 			message(STATUS "CUDA binaries for ${arch} are no longer supported, skipped.")
 		else()
 			# Compile regular kernel
-			CYCLES_CUDA_KERNEL_ADD(${arch} filter "" "${cuda_filter_sources}" FALSE)
-			CYCLES_CUDA_KERNEL_ADD(${arch} kernel "" "${cuda_sources}" FALSE)
-		endif()
+			CYCLES_CUDA_KERNEL_ADD(${arch} ${prev_arch} filter "" "${cuda_filter_sources}" FALSE)
+			CYCLES_CUDA_KERNEL_ADD(${arch} ${prev_arch} kernel "" "${cuda_sources}" FALSE)
 
-		if(WITH_CYCLES_CUDA_SPLIT_KERNEL_BINARIES)
-			# Compile split kernel
-			CYCLES_CUDA_KERNEL_ADD(${arch} kernel_split "-D __SPLIT__" ${cuda_sources} FALSE)
+			if(WITH_CYCLES_CUDA_SPLIT_KERNEL_BINARIES)
+				# Compile split kernel
+				CYCLES_CUDA_KERNEL_ADD(${arch} ${prev_arch} kernel_split "-D __SPLIT__" "${cuda_sources}" FALSE)
+			endif()
+
+			if(WITH_CYCLES_CUDA_BUILD_SERIAL)
+				set(prev_arch ${arch})
+			endif()
 		endif()
 	endforeach()