kannonier8/blender
1
0
Fork 0
forked from bartvdbraak/blender
Code Pull Requests Activity
6365ddcab5
blender/intern/cycles/kernel/CMakeLists.txt
Lukas Stockner fa3d50af95 Cycles: Improve denoising speed on GPUs with small tile sizes 
Previously, the NLM kernels would be launched once per offset with one thread per pixel.
However, with the smaller tile sizes that are now feasible, there wasn't enough work to fully occupy GPUs which results in a significant slowdown.

Therefore, the kernels are now launched in a single call that handles all offsets at once.
This has two downsides: Memory accesses to accumulating buffers are now atomic, and more importantly, the temporary memory now has to be allocated for every shift at once, increasing the required memory.
On the other hand, of course, the smaller tiles significantly reduce the size of the memory.

The main bottleneck right now is the construction of the transformation - there is nothing to be parallelized there, one thread per pixel is the maximum.
I tried to parallelize the SVD implementation by storing the matrix in shared memory and launching one block per pixel, but that wasn't really going anywhere.

To make the new code somewhat readable, the handling of rectangular regions was cleaned up a bit and commented, it should be easier to understand what's going on now.
Also, some variables have been renamed to make the difference between buffer width and stride more apparent, in addition to some general style cleanup.
2017-11-30 07:37:08 +01:00

511 lines
15 KiB
CMake
Raw Blame History

remove_extra_strict_flags()
set(INC
..
)
set(INC_SYS
)
set(SRC_CPU_KERNELS
kernels/cpu/kernel.cpp
kernels/cpu/kernel_sse2.cpp
kernels/cpu/kernel_sse3.cpp
kernels/cpu/kernel_sse41.cpp
kernels/cpu/kernel_avx.cpp
kernels/cpu/kernel_avx2.cpp
kernels/cpu/kernel_split.cpp
kernels/cpu/kernel_split_sse2.cpp
kernels/cpu/kernel_split_sse3.cpp
kernels/cpu/kernel_split_sse41.cpp
kernels/cpu/kernel_split_avx.cpp
kernels/cpu/kernel_split_avx2.cpp
kernels/cpu/filter.cpp
kernels/cpu/filter_sse2.cpp
kernels/cpu/filter_sse3.cpp
kernels/cpu/filter_sse41.cpp
kernels/cpu/filter_avx.cpp
kernels/cpu/filter_avx2.cpp
)
set(SRC_CUDA_KERNELS
kernels/cuda/kernel.cu
kernels/cuda/kernel_split.cu
kernels/cuda/filter.cu
)
set(SRC_OPENCL_KERNELS
kernels/opencl/kernel.cl
kernels/opencl/kernel_state_buffer_size.cl
kernels/opencl/kernel_split.cl
kernels/opencl/kernel_data_init.cl
kernels/opencl/kernel_path_init.cl
kernels/opencl/kernel_queue_enqueue.cl
kernels/opencl/kernel_scene_intersect.cl
kernels/opencl/kernel_lamp_emission.cl
kernels/opencl/kernel_do_volume.cl
kernels/opencl/kernel_indirect_background.cl
kernels/opencl/kernel_shader_setup.cl
kernels/opencl/kernel_shader_sort.cl
kernels/opencl/kernel_shader_eval.cl
kernels/opencl/kernel_holdout_emission_blurring_pathtermination_ao.cl
kernels/opencl/kernel_subsurface_scatter.cl
kernels/opencl/kernel_direct_lighting.cl
kernels/opencl/kernel_shadow_blocked_ao.cl
kernels/opencl/kernel_shadow_blocked_dl.cl
kernels/opencl/kernel_enqueue_inactive.cl
kernels/opencl/kernel_next_iteration_setup.cl
kernels/opencl/kernel_indirect_subsurface.cl
kernels/opencl/kernel_buffer_update.cl
kernels/opencl/filter.cl
)
set(SRC_BVH_HEADERS
bvh/bvh.h
bvh/bvh_nodes.h
bvh/bvh_shadow_all.h
bvh/bvh_local.h
bvh/bvh_traversal.h
bvh/bvh_types.h
bvh/bvh_volume.h
bvh/bvh_volume_all.h
bvh/qbvh_nodes.h
bvh/qbvh_shadow_all.h
bvh/qbvh_local.h
bvh/qbvh_traversal.h
bvh/qbvh_volume.h
bvh/qbvh_volume_all.h
)
set(SRC_HEADERS
kernel_accumulate.h
kernel_bake.h
kernel_camera.h
kernel_compat_cpu.h
kernel_compat_cuda.h
kernel_compat_opencl.h
kernel_differential.h
kernel_emission.h
kernel_film.h
kernel_globals.h
kernel_jitter.h
kernel_light.h
kernel_math.h
kernel_montecarlo.h
kernel_passes.h
kernel_path.h
kernel_path_branched.h
kernel_path_common.h
kernel_path_state.h
kernel_path_surface.h
kernel_path_subsurface.h
kernel_path_volume.h
kernel_projection.h
kernel_queues.h
kernel_random.h
kernel_shader.h
kernel_shadow.h
kernel_subsurface.h
kernel_textures.h
kernel_types.h
kernel_volume.h
kernel_work_stealing.h
)
set(SRC_KERNELS_CPU_HEADERS
kernel.h
kernels/cpu/kernel_cpu.h
kernels/cpu/kernel_cpu_impl.h
kernels/cpu/kernel_cpu_image.h
kernels/cpu/filter_cpu.h
kernels/cpu/filter_cpu_impl.h
)
set(SRC_KERNELS_CUDA_HEADERS
kernels/cuda/kernel_config.h
kernels/cuda/kernel_cuda_image.h
)
set(SRC_KERNELS_OPENCL_HEADERS
kernels/opencl/kernel_split_function.h
kernels/opencl/kernel_opencl_image.h
)
set(SRC_CLOSURE_HEADERS
closure/alloc.h
closure/bsdf.h
closure/bsdf_ashikhmin_velvet.h
closure/bsdf_diffuse.h
closure/bsdf_diffuse_ramp.h
closure/bsdf_microfacet.h
closure/bsdf_microfacet_multi.h
closure/bsdf_microfacet_multi_impl.h
closure/bsdf_oren_nayar.h
closure/bsdf_phong_ramp.h
closure/bsdf_reflection.h
closure/bsdf_refraction.h
closure/bsdf_toon.h
closure/bsdf_transparent.h
closure/bsdf_util.h
closure/bsdf_ashikhmin_shirley.h
closure/bsdf_hair.h
closure/bssrdf.h
closure/emissive.h
closure/volume.h
closure/bsdf_principled_diffuse.h
closure/bsdf_principled_sheen.h
)
set(SRC_SVM_HEADERS
svm/svm.h
svm/svm_attribute.h
svm/svm_bevel.h
svm/svm_blackbody.h
svm/svm_bump.h
svm/svm_camera.h
svm/svm_closure.h
svm/svm_convert.h
svm/svm_checker.h
svm/svm_color_util.h
svm/svm_brick.h
svm/svm_displace.h
svm/svm_fresnel.h
svm/svm_wireframe.h
svm/svm_wavelength.h
svm/svm_gamma.h
svm/svm_brightness.h
svm/svm_geometry.h
svm/svm_gradient.h
svm/svm_hsv.h
svm/svm_image.h
svm/svm_invert.h
svm/svm_light_path.h
svm/svm_magic.h
svm/svm_mapping.h
svm/svm_math.h
svm/svm_math_util.h
svm/svm_mix.h
svm/svm_musgrave.h
svm/svm_noise.h
svm/svm_noisetex.h
svm/svm_normal.h
svm/svm_ramp.h
svm/svm_ramp_util.h
svm/svm_sepcomb_hsv.h
svm/svm_sepcomb_vector.h
svm/svm_sky.h
svm/svm_tex_coord.h
svm/svm_texture.h
svm/svm_types.h
svm/svm_value.h
svm/svm_vector_transform.h
svm/svm_voronoi.h
svm/svm_voxel.h
svm/svm_wave.h
)
set(SRC_GEOM_HEADERS
geom/geom.h
geom/geom_attribute.h
geom/geom_curve.h
geom/geom_curve_intersect.h
geom/geom_motion_curve.h
geom/geom_motion_triangle.h
geom/geom_motion_triangle_intersect.h
geom/geom_motion_triangle_shader.h
geom/geom_object.h
geom/geom_patch.h
geom/geom_primitive.h
geom/geom_subd_triangle.h
geom/geom_triangle.h
geom/geom_triangle_intersect.h
geom/geom_volume.h
)
set(SRC_FILTER_HEADERS
filter/filter.h
filter/filter_defines.h
filter/filter_features.h
filter/filter_features_sse.h
filter/filter_kernel.h
filter/filter_nlm_cpu.h
filter/filter_nlm_gpu.h
filter/filter_prefilter.h
filter/filter_reconstruction.h
filter/filter_transform.h
filter/filter_transform_gpu.h
filter/filter_transform_sse.h
)
set(SRC_UTIL_HEADERS
../util/util_atomic.h
../util/util_color.h
../util/util_defines.h
../util/util_half.h
../util/util_hash.h
../util/util_math.h
../util/util_math_fast.h
../util/util_math_intersect.h
../util/util_math_float2.h
../util/util_math_float3.h
../util/util_math_float4.h
../util/util_math_int2.h
../util/util_math_int3.h
../util/util_math_int4.h
../util/util_math_matrix.h
../util/util_rect.h
../util/util_static_assert.h
../util/util_transform.h
../util/util_texture.h
../util/util_types.h
../util/util_types_float2.h
../util/util_types_float2_impl.h
../util/util_types_float3.h
../util/util_types_float3_impl.h
../util/util_types_float4.h
../util/util_types_float4_impl.h
../util/util_types_int2.h
../util/util_types_int2_impl.h
../util/util_types_int3.h
../util/util_types_int3_impl.h
../util/util_types_int4.h
../util/util_types_int4_impl.h
../util/util_types_uchar2.h
../util/util_types_uchar2_impl.h
../util/util_types_uchar3.h
../util/util_types_uchar3_impl.h
../util/util_types_uchar4.h
../util/util_types_uchar4_impl.h
../util/util_types_uint2.h
../util/util_types_uint2_impl.h
../util/util_types_uint3.h
../util/util_types_uint3_impl.h
../util/util_types_uint4.h
../util/util_types_uint4_impl.h
../util/util_types_vector3.h
../util/util_types_vector3_impl.h
)
set(SRC_SPLIT_HEADERS
split/kernel_branched.h
split/kernel_buffer_update.h
split/kernel_data_init.h
split/kernel_direct_lighting.h
split/kernel_do_volume.h
split/kernel_enqueue_inactive.h
split/kernel_holdout_emission_blurring_pathtermination_ao.h
split/kernel_indirect_background.h
split/kernel_indirect_subsurface.h
split/kernel_lamp_emission.h
split/kernel_next_iteration_setup.h
split/kernel_path_init.h
split/kernel_queue_enqueue.h
split/kernel_scene_intersect.h
split/kernel_shader_setup.h
split/kernel_shader_sort.h
split/kernel_shader_eval.h
split/kernel_shadow_blocked_ao.h
split/kernel_shadow_blocked_dl.h
split/kernel_split_common.h
split/kernel_split_data.h
split/kernel_split_data_types.h
split/kernel_subsurface_scatter.h
)
# CUDA module
if(WITH_CYCLES_CUDA_BINARIES)
# 32 bit or 64 bit
if(CUDA_64_BIT_DEVICE_CODE)
set(CUDA_BITS 64)
else()
set(CUDA_BITS 32)
endif()
# CUDA version
execute_process(COMMAND ${CUDA_NVCC_EXECUTABLE} "--version" OUTPUT_VARIABLE NVCC_OUT)
string(REGEX REPLACE ".*release ([0-9]+)\\.([0-9]+).*" "\\1" CUDA_VERSION_MAJOR "${NVCC_OUT}")
string(REGEX REPLACE ".*release ([0-9]+)\\.([0-9]+).*" "\\2" CUDA_VERSION_MINOR "${NVCC_OUT}")
set(CUDA_VERSION "${CUDA_VERSION_MAJOR}${CUDA_VERSION_MINOR}")
# warn for other versions
if(CUDA_VERSION MATCHES "80" OR CUDA_VERSION MATCHES "90")
else()
message(WARNING
"CUDA version ${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR} detected, "
"build may succeed but only CUDA 8.0 is officially supported")
endif()
# build for each arch
set(cuda_sources kernels/cuda/kernel.cu kernels/cuda/kernel_split.cu
${SRC_HEADERS}
${SRC_KERNELS_CUDA_HEADERS}
${SRC_BVH_HEADERS}
${SRC_SVM_HEADERS}
${SRC_GEOM_HEADERS}
${SRC_CLOSURE_HEADERS}
${SRC_UTIL_HEADERS}
)
set(cuda_filter_sources kernels/cuda/filter.cu
${SRC_HEADERS}
${SRC_KERNELS_CUDA_HEADERS}
${SRC_FILTER_HEADERS}
${SRC_UTIL_HEADERS}
)
set(cuda_cubins)
macro(CYCLES_CUDA_KERNEL_ADD arch name flags sources experimental)
if(${experimental})
set(flags ${flags} -D__KERNEL_EXPERIMENTAL__)
set(name ${name}_experimental)
endif()
set(cuda_cubin ${name}_${arch}.cubin)
if(WITH_CYCLES_DEBUG)
set(cuda_debug_flags "-D__KERNEL_DEBUG__")
else()
set(cuda_debug_flags "")
endif()
set(cuda_nvcc_command ${CUDA_NVCC_EXECUTABLE})
set(cuda_nvcc_version ${CUDA_VERSION})
set(cuda_version_flags "-D__KERNEL_CUDA_VERSION__=${cuda_nvcc_version}")
set(cuda_math_flags "--use_fast_math")
set(cuda_kernel_src "/kernels/cuda/${name}.cu")
add_custom_command(
OUTPUT ${cuda_cubin}
COMMAND ${cuda_nvcc_command}
-arch=${arch}
${CUDA_NVCC_FLAGS}
-m${CUDA_BITS}
--cubin ${CMAKE_CURRENT_SOURCE_DIR}${cuda_kernel_src}
-o ${CMAKE_CURRENT_BINARY_DIR}/${cuda_cubin}
--ptxas-options="-v"
${cuda_arch_flags}
${cuda_version_flags}
${cuda_math_flags}
${flags}
${cuda_debug_flags}
-I${CMAKE_CURRENT_SOURCE_DIR}/..
-DCCL_NAMESPACE_BEGIN=
-DCCL_NAMESPACE_END=
-DNVCC
DEPENDS ${sources})
delayed_install("${CMAKE_CURRENT_BINARY_DIR}" "${cuda_cubin}" ${CYCLES_INSTALL_PATH}/lib)
list(APPEND cuda_cubins ${cuda_cubin})
unset(cuda_extra_flags)
unset(cuda_debug_flags)
unset(cuda_nvcc_command)
unset(cuda_nvcc_version)
endmacro()
foreach(arch ${CYCLES_CUDA_BINARIES_ARCH})
if(CUDA_VERSION MATCHES "90" AND ${arch} MATCHES "sm_2.")
message(STATUS "CUDA binaries for ${arch} disabled, not supported by CUDA 9.")
else()
# Compile regular kernel
CYCLES_CUDA_KERNEL_ADD(${arch} kernel "" "${cuda_sources}" FALSE)
CYCLES_CUDA_KERNEL_ADD(${arch} filter "" "${cuda_filter_sources}" FALSE)
if(WITH_CYCLES_CUDA_SPLIT_KERNEL_BINARIES)
# Compile split kernel
CYCLES_CUDA_KERNEL_ADD(${arch} kernel_split "-D__SPLIT__" ${cuda_sources} FALSE)
endif()
endif()
endforeach()
add_custom_target(cycles_kernel_cuda ALL DEPENDS ${cuda_cubins})
endif()
# OSL module
if(WITH_CYCLES_OSL)
add_subdirectory(osl)
add_subdirectory(shaders)
endif()
# CPU module
include_directories(${INC})
include_directories(SYSTEM ${INC_SYS})
set_source_files_properties(kernels/cpu/kernel.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/kernel_split.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/filter.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_KERNEL_FLAGS}")
if(CXX_HAS_SSE)
set_source_files_properties(kernels/cpu/kernel_sse2.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE2_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/kernel_sse3.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE3_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/kernel_sse41.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE41_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/kernel_split_sse2.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE2_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/kernel_split_sse3.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE3_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/kernel_split_sse41.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE41_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/filter_sse2.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE2_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/filter_sse3.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE3_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/filter_sse41.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE41_KERNEL_FLAGS}")
endif()
if(CXX_HAS_AVX)
set_source_files_properties(kernels/cpu/kernel_avx.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_AVX_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/kernel_split_avx.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_AVX_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/filter_avx.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_AVX_KERNEL_FLAGS}")
endif()
if(CXX_HAS_AVX2)
set_source_files_properties(kernels/cpu/kernel_avx2.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_AVX2_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/kernel_split_avx2.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_AVX2_KERNEL_FLAGS}")
set_source_files_properties(kernels/cpu/filter_avx2.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_AVX2_KERNEL_FLAGS}")
endif()
add_library(cycles_kernel
${SRC_CPU_KERNELS}
${SRC_CUDA_KERNELS}
${SRC_OPENCL_KERNELS}
${SRC_HEADERS}
${SRC_KERNELS_CPU_HEADERS}
${SRC_KERNELS_CUDA_HEADERS}
${SRC_KERNELS_OPENCL_HEADERS}
${SRC_BVH_HEADERS}
${SRC_CLOSURE_HEADERS}
${SRC_FILTER_HEADERS}
${SRC_SVM_HEADERS}
${SRC_GEOM_HEADERS}
${SRC_SPLIT_HEADERS}
)
if(WITH_CYCLES_CUDA)
add_dependencies(cycles_kernel cycles_kernel_cuda)
endif()
# OpenCL kernel
#set(KERNEL_PREPROCESSED ${CMAKE_CURRENT_BINARY_DIR}/kernel_preprocessed.cl)
#add_custom_command(
# OUTPUT ${KERNEL_PREPROCESSED}
# COMMAND gcc -x c++ -E ${CMAKE_CURRENT_SOURCE_DIR}/kernel.cl -I ${CMAKE_CURRENT_SOURCE_DIR}/../util/ -DCCL_NAMESPACE_BEGIN= -DCCL_NAMESPACE_END= -o ${KERNEL_PREPROCESSED}
# DEPENDS ${SRC_KERNEL} ${SRC_UTIL_HEADERS})
#add_custom_target(cycles_kernel_preprocess ALL DEPENDS ${KERNEL_PREPROCESSED})
#delayed_install(${CMAKE_CURRENT_SOURCE_DIR} "${KERNEL_PREPROCESSED}" ${CYCLES_INSTALL_PATH}/kernel)
delayed_install(${CMAKE_CURRENT_SOURCE_DIR} "${SRC_OPENCL_KERNELS}" ${CYCLES_INSTALL_PATH}/source/kernel/kernels/opencl)
delayed_install(${CMAKE_CURRENT_SOURCE_DIR} "${SRC_CUDA_KERNELS}" ${CYCLES_INSTALL_PATH}/source/kernel/kernels/cuda)
delayed_install(${CMAKE_CURRENT_SOURCE_DIR} "${SRC_HEADERS}" ${CYCLES_INSTALL_PATH}/source/kernel)
delayed_install(${CMAKE_CURRENT_SOURCE_DIR} "${SRC_KERNELS_OPENCL_HEADERS}" ${CYCLES_INSTALL_PATH}/source/kernel/kernels/opencl)
delayed_install(${CMAKE_CURRENT_SOURCE_DIR} "${SRC_KERNELS_CUDA_HEADERS}" ${CYCLES_INSTALL_PATH}/source/kernel/kernels/cuda)
delayed_install(${CMAKE_CURRENT_SOURCE_DIR} "${SRC_BVH_HEADERS}" ${CYCLES_INSTALL_PATH}/source/kernel/bvh)
delayed_install(${CMAKE_CURRENT_SOURCE_DIR} "${SRC_CLOSURE_HEADERS}" ${CYCLES_INSTALL_PATH}/source/kernel/closure)
delayed_install(${CMAKE_CURRENT_SOURCE_DIR} "${SRC_FILTER_HEADERS}" ${CYCLES_INSTALL_PATH}/source/kernel/filter)
delayed_install(${CMAKE_CURRENT_SOURCE_DIR} "${SRC_SVM_HEADERS}" ${CYCLES_INSTALL_PATH}/source/kernel/svm)
delayed_install(${CMAKE_CURRENT_SOURCE_DIR} "${SRC_GEOM_HEADERS}" ${CYCLES_INSTALL_PATH}/source/kernel/geom)
delayed_install(${CMAKE_CURRENT_SOURCE_DIR} "${SRC_UTIL_HEADERS}" ${CYCLES_INSTALL_PATH}/source/util)
delayed_install(${CMAKE_CURRENT_SOURCE_DIR} "${SRC_SPLIT_HEADERS}" ${CYCLES_INSTALL_PATH}/source/kernel/split)
View Git Blame Copy Permalink