diff options
author | Sergey Sharybin <sergey.vfx@gmail.com> | 2015-11-21 20:16:01 +0300 |
---|---|---|
committer | Sergey Sharybin <sergey.vfx@gmail.com> | 2015-11-21 20:16:01 +0300 |
commit | c08727ebabae1d84c9d4e3096a61ef131177d4d5 (patch) | |
tree | 57d093f0a17d9ef6d20ed7236922ec0aabf30a88 | |
parent | f8ab3fd30f163fad9f96c889d1989faab5a471e0 (diff) |
Cycles: Quick experiment with using feature-adaptive kernels for CUDA
Gives few percent of memory improvement for regular feature set kernel
and could give significant memory improvement for Experimental kernel.
It could also give some degree of performance improvement, but this I
didn't really measure reliably yet.
Code is ifdef-ed for now, since it's only working on Linux and requires
CUDA toolkit to be installed (other platform only use precompiled
kernels).
This is just an experiment for now and a base for the proper feature
support in the future (with runtime compilation using CUDA 7?).
-rw-r--r-- | intern/cycles/device/device_cuda.cpp | 23 |
1 files changed, 22 insertions, 1 deletions
diff --git a/intern/cycles/device/device_cuda.cpp b/intern/cycles/device/device_cuda.cpp index 7f21cc5e036..ea023b4af66 100644 --- a/intern/cycles/device/device_cuda.cpp +++ b/intern/cycles/device/device_cuda.cpp @@ -27,6 +27,7 @@ #include "util_debug.h" #include "util_logging.h" #include "util_map.h" +#include "util_md5.h" #include "util_opengl.h" #include "util_path.h" #include "util_string.h" @@ -34,6 +35,11 @@ #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 class CUDADevice : public Device @@ -221,10 +227,20 @@ 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 if(requested_features.experimental) cubin = string_printf("cycles_kernel_experimental_sm%d%d_%s.cubin", major, minor, md5.c_str()); else cubin = string_printf("cycles_kernel_sm%d%d_%s.cubin", major, minor, md5.c_str()); +#endif + cubin = path_user_get(path_join("cache", cubin)); VLOG(1) << "Testing for locally compiled kernel " << cubin; /* if exists already, use it */ @@ -280,8 +296,13 @@ public: "-DNVCC -D__KERNEL_CUDA_VERSION__=%d", nvcc, major, minor, machine, kernel.c_str(), cubin.c_str(), include.c_str(), cuda_version); - if(requested_features.experimental) +#ifdef KERNEL_USE_ADAPTIVE + command += " " + feature_build_options; +#else + if(requested_features.experimental) { command += " -D__KERNEL_EXPERIMENTAL__"; + } +#endif if(getenv("CYCLES_CUDA_EXTRA_CFLAGS")) { command += string(" ") + getenv("CYCLES_CUDA_EXTRA_CFLAGS"); |