diff options
author | Brecht Van Lommel <brechtvanlommel@gmail.com> | 2018-12-04 14:34:59 +0300 |
---|---|---|
committer | Brecht Van Lommel <brechtvanlommel@gmail.com> | 2018-12-04 18:03:18 +0300 |
commit | b14ec1860127eddf9a46132c4877bd4a7ee385a2 (patch) | |
tree | 72db903da88d3589c639984c86159ed5e19c1dd2 /intern | |
parent | 3f31ec83984b19efe174319ae1d86995a4c26ecc (diff) |
Cycles: add initial CUDA 10.0 support, but only recommend use for Turing cards.
There may still be rendering errors when used for older graphics cards.
Diffstat (limited to 'intern')
-rw-r--r-- | intern/cycles/CMakeLists.txt | 7 | ||||
-rw-r--r-- | intern/cycles/kernel/CMakeLists.txt | 6 | ||||
-rw-r--r-- | intern/cycles/kernel/kernels/cuda/kernel_config.h | 18 |
3 files changed, 26 insertions, 5 deletions
diff --git a/intern/cycles/CMakeLists.txt b/intern/cycles/CMakeLists.txt index 3028b7c67d3..8f0d838e881 100644 --- a/intern/cycles/CMakeLists.txt +++ b/intern/cycles/CMakeLists.txt @@ -315,6 +315,13 @@ if(WITH_CYCLES_CUDA_BINARIES AND (NOT WITH_CYCLES_CUBIN_COMPILER)) endif() endif() +# NVRTC gives wrong rendering result in CUDA 10.0, so we must use NVCC. +if(WITH_CYCLES_CUDA_BINARIES AND WITH_CYCLES_CUBIN_COMPILER) + if(${CUDA_VERSION} VERSION_GREATER_EQUAL 10.0) + message(STATUS "cycles_cubin_cc not supported for CUDA 10.0+, using nvcc instead.") + set(WITH_CYCLES_CUBIN_COMPILER OFF) + endif() +endif() # Subdirectories diff --git a/intern/cycles/kernel/CMakeLists.txt b/intern/cycles/kernel/CMakeLists.txt index d4145225b77..3ba43a9e0bd 100644 --- a/intern/cycles/kernel/CMakeLists.txt +++ b/intern/cycles/kernel/CMakeLists.txt @@ -343,11 +343,11 @@ if(WITH_CYCLES_CUDA_BINARIES) set(CUDA_VERSION "${CUDA_VERSION_MAJOR}${CUDA_VERSION_MINOR}") # warn for other versions - if(CUDA_VERSION MATCHES "90" OR CUDA_VERSION MATCHES "91") + if(CUDA_VERSION MATCHES "90" OR CUDA_VERSION MATCHES "91" OR CUDA_VERSION MATCHES "100") else() message(WARNING "CUDA version ${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR} detected, " - "build may succeed but only CUDA 9.0 and 9.1 are officially supported") + "build may succeed but only CUDA 9.0, 9.1 and 10.0 are officially supported") endif() # build for each arch @@ -442,6 +442,8 @@ if(WITH_CYCLES_CUDA_BINARIES) foreach(arch ${CYCLES_CUDA_BINARIES_ARCH}) if(${arch} MATCHES "sm_2.") message(STATUS "CUDA binaries for ${arch} are no longer supported, skipped.") + elseif(${arch} MATCHES "sm_7." AND NOT (${CUDA_VERSION} VERSION_GREATER_EQUAL 10.0)) + message(STATUS "CUDA binaries for ${arch} require CUDA 10.0+, skipped.") else() # Compile regular kernel CYCLES_CUDA_KERNEL_ADD(${arch} ${prev_arch} filter "" "${cuda_filter_sources}" FALSE) diff --git a/intern/cycles/kernel/kernels/cuda/kernel_config.h b/intern/cycles/kernel/kernels/cuda/kernel_config.h index 3808898c5ca..6d41dc15785 100644 --- a/intern/cycles/kernel/kernels/cuda/kernel_config.h +++ b/intern/cycles/kernel/kernels/cuda/kernel_config.h @@ -52,8 +52,8 @@ # define CUDA_KERNEL_MAX_REGISTERS 63 # define CUDA_KERNEL_BRANCHED_MAX_REGISTERS 63 -/* 5.0, 5.2, 5.3, 6.0, 6.1 */ -#elif __CUDA_ARCH__ >= 500 +/* 5.x, 6.x */ +#elif __CUDA_ARCH__ <= 699 # define CUDA_MULTIPRESSOR_MAX_REGISTERS 65536 # define CUDA_MULTIPROCESSOR_MAX_BLOCKS 32 # define CUDA_BLOCK_MAX_THREADS 1024 @@ -62,13 +62,25 @@ /* tunable parameters */ # define CUDA_THREADS_BLOCK_WIDTH 16 /* CUDA 9.0 seems to cause slowdowns on high-end Pascal cards unless we increase the number of registers */ -# if __CUDACC_VER_MAJOR__ == 9 && __CUDA_ARCH__ >= 600 +# if __CUDACC_VER_MAJOR__ >= 9 && __CUDA_ARCH__ >= 600 # define CUDA_KERNEL_MAX_REGISTERS 64 # else # define CUDA_KERNEL_MAX_REGISTERS 48 # endif # define CUDA_KERNEL_BRANCHED_MAX_REGISTERS 63 +/* 7.x */ +#elif __CUDA_ARCH__ <= 799 +# define CUDA_MULTIPRESSOR_MAX_REGISTERS 65536 +# define CUDA_MULTIPROCESSOR_MAX_BLOCKS 32 +# define CUDA_BLOCK_MAX_THREADS 1024 +# define CUDA_THREAD_MAX_REGISTERS 255 + +/* tunable parameters */ +# define CUDA_THREADS_BLOCK_WIDTH 16 +# define CUDA_KERNEL_MAX_REGISTERS 64 +# define CUDA_KERNEL_BRANCHED_MAX_REGISTERS 72 + /* unknown architecture */ #else |