diff options
| author | Brecht Van Lommel <brechtvanlommel@gmail.com> | 2014-04-30 12:54:17 +0400 |
|---|---|---|
| committer | Brecht Van Lommel <brechtvanlommel@gmail.com> | 2014-04-30 18:07:27 +0400 |
| commit | 741f17f05b8272abfaf000a403e44b73692ba4c7 (patch) | |
| tree | dafcf851ca51b51b56e947bfb3879d224381cd6d /intern/cycles/kernel | |
| parent | 4d1a109ddec02ad7e527d8b65a5cdc8d4a7528a9 (diff) | |
Cycles CUDA: make CUDA toolkit 6.0 the official supported version.
This also updates the configurations to build kernels for compute capability
5.0 cards, when using and older CUDA toolkit version this will be skipped.
Also includes tweaks to improve performance with this version:
* Increase max registers on sm_30, sm_35 and sm_50
* No longer use texture storage on sm_30
Diffstat (limited to 'intern/cycles/kernel')
| -rw-r--r-- | intern/cycles/kernel/CMakeLists.txt | 10 | ||||
| -rw-r--r-- | intern/cycles/kernel/SConscript | 8 | ||||
| -rw-r--r-- | intern/cycles/kernel/kernel.cu | 8 | ||||
| -rw-r--r-- | intern/cycles/kernel/kernel_compat_cuda.h | 2 |
4 files changed, 17 insertions, 11 deletions
diff --git a/intern/cycles/kernel/CMakeLists.txt b/intern/cycles/kernel/CMakeLists.txt index 1527d154c86..d18f4fa2998 100644 --- a/intern/cycles/kernel/CMakeLists.txt +++ b/intern/cycles/kernel/CMakeLists.txt @@ -146,11 +146,11 @@ if(WITH_CYCLES_CUDA_BINARIES) set(CUDA_VERSION "${CUDA_VERSION_MAJOR}${CUDA_VERSION_MINOR}") # warn for other versions - if(CUDA_VERSION MATCHES "50") + if(CUDA_VERSION MATCHES "60") else() message(WARNING "CUDA version ${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR} detected, " - "build may succeed but only CUDA 5.0 is officially supported") + "build may succeed but only CUDA 6.0 is officially supported") endif() # build for each arch @@ -162,8 +162,10 @@ if(WITH_CYCLES_CUDA_BINARIES) set(cuda_version_flags "-D__KERNEL_CUDA_VERSION__=${CUDA_VERSION}") set(cuda_math_flags "--use_fast_math") - - if(CUDA_VERSION LESS 50 AND ${arch} MATCHES "sm_35") + + if(CUDA_VERSION LESS 60 AND ${arch} MATCHES "sm_50") + message(WARNING "Can't build kernel for CUDA sm_50 architecture, skipping") + elseif(CUDA_VERSION LESS 50 AND ${arch} MATCHES "sm_35") message(WARNING "Can't build kernel for CUDA sm_35 architecture, skipping") else() add_custom_command( diff --git a/intern/cycles/kernel/SConscript b/intern/cycles/kernel/SConscript index 5316ec9926b..04e1bad7538 100644 --- a/intern/cycles/kernel/SConscript +++ b/intern/cycles/kernel/SConscript @@ -69,8 +69,8 @@ if env['WITH_BF_CYCLES_CUDA_BINARIES']: cuda_major_minor = re.findall(r'release (\d+).(\d+)', output)[0] cuda_version = int(cuda_major_minor[0])*10 + int(cuda_major_minor[1]) - if cuda_version != 50: - print("CUDA version %d.%d detected, build may succeed but only CUDA 5.0 is officially supported." % (cuda_version/10, cuda_version%10)) + if cuda_version != 60: + print("CUDA version %d.%d detected, build may succeed but only CUDA 6.0 is officially supported." % (cuda_version/10, cuda_version%10)) # nvcc flags nvcc_flags = "-m%s" % (bits) @@ -85,6 +85,10 @@ if env['WITH_BF_CYCLES_CUDA_BINARIES']: # add command for each cuda architecture for arch in cuda_archs: + if cuda_version < 60 and arch == "sm_50": + print("Can't build kernel for CUDA sm_50 architecture, skipping") + continue + cubin_file = os.path.join(build_dir, "kernel_%s.cubin" % arch) if env['BF_CYCLES_CUDA_ENV']: diff --git a/intern/cycles/kernel/kernel.cu b/intern/cycles/kernel/kernel.cu index d91c6d97123..636e48b5456 100644 --- a/intern/cycles/kernel/kernel.cu +++ b/intern/cycles/kernel/kernel.cu @@ -49,8 +49,8 @@ /* tunable parameters */ #define CUDA_THREADS_BLOCK_WIDTH 16 -#define CUDA_KERNEL_MAX_REGISTERS 32 -#define CUDA_KERNEL_BRANCHED_MAX_REGISTERS 40 +#define CUDA_KERNEL_MAX_REGISTERS 63 +#define CUDA_KERNEL_BRANCHED_MAX_REGISTERS 63 /* 5.0 */ #elif __CUDA_ARCH__ == 500 @@ -61,8 +61,8 @@ /* tunable parameters */ #define CUDA_THREADS_BLOCK_WIDTH 16 -#define CUDA_KERNEL_MAX_REGISTERS 32 -#define CUDA_KERNEL_BRANCHED_MAX_REGISTERS 40 +#define CUDA_KERNEL_MAX_REGISTERS 63 +#define CUDA_KERNEL_BRANCHED_MAX_REGISTERS 63 /* unknown architecture */ #else diff --git a/intern/cycles/kernel/kernel_compat_cuda.h b/intern/cycles/kernel/kernel_compat_cuda.h index 15e7353ec38..e4c20d26ff1 100644 --- a/intern/cycles/kernel/kernel_compat_cuda.h +++ b/intern/cycles/kernel/kernel_compat_cuda.h @@ -60,7 +60,7 @@ typedef texture<uchar4, 2, cudaReadModeNormalizedFloat> texture_image_uchar4; /* In order to use full 6GB of memory on Titan cards, use arrays instead * of textures. On earlier cards this seems slower, but on Titan it is * actually slightly faster in tests. */ -#if __CUDA_ARCH__ < 350 +#if __CUDA_ARCH__ < 300 #define __KERNEL_CUDA_TEX_STORAGE__ #endif |