diff options
author | Thomas Dinges <blender@dingto.org> | 2016-05-06 23:34:15 +0300 |
---|---|---|
committer | Thomas Dinges <blender@dingto.org> | 2016-05-07 00:13:33 +0300 |
commit | 734d1aec3f93b8757533284330afc2ac651442bd (patch) | |
tree | 98bbf8ea18392b09355c8b041b7727f9af88b2c4 /intern/cycles/device/device_cuda.cpp | |
parent | bd335f13fe3d25e8e6bb472acb9c727d1629e582 (diff) |
Cycles: Make CUDA adaptive feature compile a Debug flag.
If the CUDA Toolkit is installed and the user is on Linux,
adaptive, feature based CUDA runtime compile is now possible to enable via:
* Environment flag CYCLES_CUDA_ADAPTIVE_COMPILE or
* Debug menu (Debug value 256) in the Cycles UI.
Diffstat (limited to 'intern/cycles/device/device_cuda.cpp')
-rw-r--r-- | intern/cycles/device/device_cuda.cpp | 42 |
1 files changed, 23 insertions, 19 deletions
diff --git a/intern/cycles/device/device_cuda.cpp b/intern/cycles/device/device_cuda.cpp index f7c985e787d..5362623247f 100644 --- a/intern/cycles/device/device_cuda.cpp +++ b/intern/cycles/device/device_cuda.cpp @@ -41,11 +41,6 @@ #include "util_types.h" #include "util_time.h" -/* use feature-adaptive kernel compilation. - * Requires CUDA toolkit to be installed and currently only works on Linux. - */ -/* #define KERNEL_USE_ADAPTIVE */ - CCL_NAMESPACE_BEGIN #ifndef WITH_CUDA_DYNLOAD @@ -245,6 +240,11 @@ public: return true; } + bool use_adaptive_compilation() + { + return DebugFlags().cuda.adaptive_compile; + } + string compile_kernel(const DeviceRequestedFeatures& requested_features) { /* compute cubin name */ @@ -252,6 +252,9 @@ public: cuDeviceComputeCapability(&major, &minor, cuDevId); string cubin; + /* adaptive compile */ + bool use_adaptive_compile = use_adaptive_compilation(); + /* attempt to use kernel provided with blender */ cubin = path_get(string_printf("lib/kernel_sm_%d%d.cubin", major, minor)); VLOG(1) << "Testing for pre-compiled kernel " << cubin; @@ -264,17 +267,19 @@ public: string kernel_path = path_get("kernel"); string md5 = path_files_md5_hash(kernel_path); -#ifdef KERNEL_USE_ADAPTIVE - string feature_build_options = requested_features.get_build_options(); - string device_md5 = util_md5_string(feature_build_options); - cubin = string_printf("cycles_kernel_%s_sm%d%d_%s.cubin", - device_md5.c_str(), - major, minor, - md5.c_str()); -#else - (void)requested_features; - cubin = string_printf("cycles_kernel_sm%d%d_%s.cubin", major, minor, md5.c_str()); -#endif + string feature_build_options; + if(use_adaptive_compile) { + feature_build_options = requested_features.get_build_options(); + string device_md5 = util_md5_string(feature_build_options); + cubin = string_printf("cycles_kernel_%s_sm%d%d_%s.cubin", + device_md5.c_str(), + major, minor, + md5.c_str()); + } + else { + (void)requested_features; + cubin = string_printf("cycles_kernel_sm%d%d_%s.cubin", major, minor, md5.c_str()); + } cubin = path_user_get(path_join("cache", cubin)); VLOG(1) << "Testing for locally compiled kernel " << cubin; @@ -331,9 +336,8 @@ public: "-DNVCC -D__KERNEL_CUDA_VERSION__=%d", nvcc, major, minor, machine, kernel.c_str(), cubin.c_str(), include.c_str(), cuda_version); -#ifdef KERNEL_USE_ADAPTIVE - command += " " + feature_build_options; -#endif + if(use_adaptive_compile) + command += " " + feature_build_options; const char* extra_cflags = getenv("CYCLES_CUDA_EXTRA_CFLAGS"); if(extra_cflags) { |