Cycles: Add option to directly link against CUDA libraries

The main purpose of such linking is to make Blender compatible with
NVidia's debuggers and profilers which are doing some LD_PRELOAD
magic to intercept some function calls. Such magic conflicts with
our CUDA wrangler magic and causes segmentation faults.

The option is disabled by default, so there's no affect on any of
artists.

In order to make Blender linked directly against CUDA library use
the WITH_CUDA_DYNLOAD CMake option (it's marked as advanced).

4 files changed, 60 insertions, 4 deletions

diff --git a/intern/cycles/app/CMakeLists.txt b/intern/cycles/app/CMakeLists.txt
index 95bc517ae00..41233287279 100644
--- a/intern/cycles/app/CMakeLists.txt
+++ b/intern/cycles/app/CMakeLists.txt
@@ -29,9 +29,14 @@ set(LIBRARIES
 	${TIFF_LIBRARY}
 	${PTHREADS_LIBRARIES}
 	extern_clew
-	extern_cuew
 )
 
+if(WITH_CUDA_DYNLOAD)
+	list(APPEND LIBRARIES extern_cuew)
+else()
+	list(APPEND LIBRARIES ${CUDA_CUDA_LIBRARY})
+endif()
+
 if(WITH_CYCLES_OSL)
 	list(APPEND LIBRARIES cycles_kernel_osl)
 endif()
diff --git a/intern/cycles/cmake/external_libs.cmake b/intern/cycles/cmake/external_libs.cmake
index 56ab8bed6e8..fc75efa7755 100644
--- a/intern/cycles/cmake/external_libs.cmake
+++ b/intern/cycles/cmake/external_libs.cmake
@@ -37,7 +37,7 @@ endif()
 ###########################################################################
 # CUDA
 
-if(WITH_CYCLES_CUDA_BINARIES)
+if(WITH_CYCLES_CUDA_BINARIES OR NOT WITH_CUDA_DUNLOAD)
 	find_package(CUDA) # Try to auto locate CUDA toolkit
 	if(CUDA_FOUND)
 		message(STATUS "CUDA nvcc = ${CUDA_NVCC_EXECUTABLE}")
diff --git a/intern/cycles/device/CMakeLists.txt b/intern/cycles/device/CMakeLists.txt
index b0fa283c1d8..220b16bb45b 100644
--- a/intern/cycles/device/CMakeLists.txt
+++ b/intern/cycles/device/CMakeLists.txt
@@ -11,10 +11,17 @@ set(INC
 
 set(INC_SYS
 	${GLEW_INCLUDE_DIR}
-	../../../extern/cuew/include
 	../../../extern/clew/include
 )
 
+if(WITH_CUDA_DYNLOAD)
+	list(APPEND INC ../../../extern/cuew/include)
+	add_definitions(-DWITH_CUDA_DYNLOAD)
+else()
+	list(APPEND INC_SYS ${CUDA_TOOLKIT_INCLUDE})
+	add_definitions(-DCYCLES_CUDA_NVCC_EXECUTABLE="${CUDA_NVCC_EXECUTABLE}")
+endif()
+
 set(SRC
 	device.cpp
 	device_cpu.cpp
diff --git a/intern/cycles/device/device_cuda.cpp b/intern/cycles/device/device_cuda.cpp
index 5c9ca3454c6..88f1a86d6ac 100644
--- a/intern/cycles/device/device_cuda.cpp
+++ b/intern/cycles/device/device_cuda.cpp
@@ -23,7 +23,13 @@
 
 #include "buffers.h"
 
-#include "cuew.h"
+#ifdef WITH_CUDA_DYNLOAD
+#  include "cuew.h"
+#else
+#  include "util_opengl.h"
+#  include <cuda.h>
+#  include <cudaGL.h>
+#endif
 #include "util_debug.h"
 #include "util_logging.h"
 #include "util_map.h"
@@ -42,6 +48,40 @@
 
 CCL_NAMESPACE_BEGIN
 
+#ifndef WITH_CUDA_DYNLOAD
+
+/* Transparently implement some functions, so majority of the file does not need
+ * to worry about difference between dynamically loaded and linked CUDA at all.
+ */
+
+namespace {
+
+const char *cuewErrorString(CUresult result)
+{
+	/* We can only give error code here without major code duplication, that
+	 * should be enough since dynamic loading is only being disabled by folks
+	 * who knows what they're doing anyway.
+	 *
+	 * NOTE: Avoid call from several threads.
+	 */
+	static string error;
+	error = string_printf("%d", result);
+	return error.c_str();
+}
+
+const char *cuewCompilerPath(void)
+{
+	return CYCLES_CUDA_NVCC_EXECUTABLE;
+}
+
+int cuewCompilerVersion(void)
+{
+	return (CUDA_VERSION / 100) + (CUDA_VERSION % 100 / 10);
+}
+
+}  /* namespace */
+#endif  /* WITH_CUDA_DYNLOAD */
+
 class CUDADevice : public Device
 {
 public:
@@ -1100,6 +1140,7 @@ public:
 
 bool device_cuda_init(void)
 {
+#ifdef WITH_CUDA_DYNLOAD
 	static bool initialized = false;
 	static bool result = false;
 
@@ -1133,6 +1174,9 @@ bool device_cuda_init(void)
 	}
 
 	return result;
+#else  /* WITH_CUDA_DYNLOAD */
+	return true;
+#endif /* WITH_CUDA_DYNLOAD */
 }
 
 Device *device_cuda_create(DeviceInfo& info, Stats &stats, bool background)