diff options
Diffstat (limited to 'intern/cycles/device/optix')
-rw-r--r-- | intern/cycles/device/optix/device.cpp | 105 | ||||
-rw-r--r-- | intern/cycles/device/optix/device.h | 35 | ||||
-rw-r--r-- | intern/cycles/device/optix/device_impl.cpp | 1573 | ||||
-rw-r--r-- | intern/cycles/device/optix/device_impl.h | 186 | ||||
-rw-r--r-- | intern/cycles/device/optix/queue.cpp | 144 | ||||
-rw-r--r-- | intern/cycles/device/optix/queue.h | 39 | ||||
-rw-r--r-- | intern/cycles/device/optix/util.h | 45 |
7 files changed, 2127 insertions, 0 deletions
diff --git a/intern/cycles/device/optix/device.cpp b/intern/cycles/device/optix/device.cpp new file mode 100644 index 00000000000..13f23bd229a --- /dev/null +++ b/intern/cycles/device/optix/device.cpp @@ -0,0 +1,105 @@ +/* + * Copyright 2019, NVIDIA Corporation. + * Copyright 2019, Blender Foundation. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "device/optix/device.h" + +#include "device/cuda/device.h" +#include "device/optix/device_impl.h" +#include "util/util_logging.h" + +#ifdef WITH_OPTIX +# include <optix_function_table_definition.h> +#endif + +CCL_NAMESPACE_BEGIN + +bool device_optix_init() +{ +#ifdef WITH_OPTIX + if (g_optixFunctionTable.optixDeviceContextCreate != NULL) { + /* Already initialized function table. */ + return true; + } + + /* Need to initialize CUDA as well. */ + if (!device_cuda_init()) { + return false; + } + + const OptixResult result = optixInit(); + + if (result == OPTIX_ERROR_UNSUPPORTED_ABI_VERSION) { + VLOG(1) << "OptiX initialization failed because the installed NVIDIA driver is too old. " + "Please update to the latest driver first!"; + return false; + } + else if (result != OPTIX_SUCCESS) { + VLOG(1) << "OptiX initialization failed with error code " << (unsigned int)result; + return false; + } + + /* Loaded OptiX successfully! */ + return true; +#else + return false; +#endif +} + +void device_optix_info(const vector<DeviceInfo> &cuda_devices, vector<DeviceInfo> &devices) +{ +#ifdef WITH_OPTIX + devices.reserve(cuda_devices.size()); + + /* Simply add all supported CUDA devices as OptiX devices again. */ + for (DeviceInfo info : cuda_devices) { + assert(info.type == DEVICE_CUDA); + + int major; + cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, info.num); + if (major < 5) { + /* Only Maxwell and up are supported by OptiX. */ + continue; + } + + info.type = DEVICE_OPTIX; + info.id += "_OptiX"; + info.denoisers |= DENOISER_OPTIX; + + devices.push_back(info); + } +#else + (void)cuda_devices; + (void)devices; +#endif +} + +Device *device_optix_create(const DeviceInfo &info, Stats &stats, Profiler &profiler) +{ +#ifdef WITH_OPTIX + return new OptiXDevice(info, stats, profiler); +#else + (void)info; + (void)stats; + (void)profiler; + + LOG(FATAL) << "Request to create OptiX device without compiled-in support. Should never happen."; + + return nullptr; +#endif +} + +CCL_NAMESPACE_END diff --git a/intern/cycles/device/optix/device.h b/intern/cycles/device/optix/device.h new file mode 100644 index 00000000000..29fa729c2e4 --- /dev/null +++ b/intern/cycles/device/optix/device.h @@ -0,0 +1,35 @@ +/* + * Copyright 2011-2021 Blender Foundation + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#pragma once + +#include "util/util_string.h" +#include "util/util_vector.h" + +CCL_NAMESPACE_BEGIN + +class Device; +class DeviceInfo; +class Profiler; +class Stats; + +bool device_optix_init(); + +Device *device_optix_create(const DeviceInfo &info, Stats &stats, Profiler &profiler); + +void device_optix_info(const vector<DeviceInfo> &cuda_devices, vector<DeviceInfo> &devices); + +CCL_NAMESPACE_END diff --git a/intern/cycles/device/optix/device_impl.cpp b/intern/cycles/device/optix/device_impl.cpp new file mode 100644 index 00000000000..cd16b8c9f01 --- /dev/null +++ b/intern/cycles/device/optix/device_impl.cpp @@ -0,0 +1,1573 @@ +/* + * Copyright 2019, NVIDIA Corporation. + * Copyright 2019, Blender Foundation. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifdef WITH_OPTIX + +# include "device/optix/device_impl.h" + +# include "bvh/bvh.h" +# include "bvh/bvh_optix.h" +# include "integrator/pass_accessor_gpu.h" +# include "render/buffers.h" +# include "render/hair.h" +# include "render/mesh.h" +# include "render/object.h" +# include "render/pass.h" +# include "render/scene.h" + +# include "util/util_debug.h" +# include "util/util_logging.h" +# include "util/util_md5.h" +# include "util/util_path.h" +# include "util/util_progress.h" +# include "util/util_time.h" + +# undef __KERNEL_CPU__ +# define __KERNEL_OPTIX__ +# include "kernel/device/optix/globals.h" + +CCL_NAMESPACE_BEGIN + +OptiXDevice::Denoiser::Denoiser(OptiXDevice *device) + : device(device), queue(device), state(device, "__denoiser_state") +{ +} + +OptiXDevice::Denoiser::~Denoiser() +{ + const CUDAContextScope scope(device); + if (optix_denoiser != nullptr) { + optixDenoiserDestroy(optix_denoiser); + } +} + +OptiXDevice::OptiXDevice(const DeviceInfo &info, Stats &stats, Profiler &profiler) + : CUDADevice(info, stats, profiler), + sbt_data(this, "__sbt", MEM_READ_ONLY), + launch_params(this, "__params"), + denoiser_(this) +{ + /* Make the CUDA context current. */ + if (!cuContext) { + /* Do not initialize if CUDA context creation failed already. */ + return; + } + const CUDAContextScope scope(this); + + /* Create OptiX context for this device. */ + OptixDeviceContextOptions options = {}; +# ifdef WITH_CYCLES_LOGGING + options.logCallbackLevel = 4; /* Fatal = 1, Error = 2, Warning = 3, Print = 4. */ + options.logCallbackFunction = [](unsigned int level, const char *, const char *message, void *) { + switch (level) { + case 1: + LOG_IF(FATAL, VLOG_IS_ON(1)) << message; + break; + case 2: + LOG_IF(ERROR, VLOG_IS_ON(1)) << message; + break; + case 3: + LOG_IF(WARNING, VLOG_IS_ON(1)) << message; + break; + case 4: + LOG_IF(INFO, VLOG_IS_ON(1)) << message; + break; + } + }; +# endif + if (DebugFlags().optix.use_debug) { + options.validationMode = OPTIX_DEVICE_CONTEXT_VALIDATION_MODE_ALL; + } + optix_assert(optixDeviceContextCreate(cuContext, &options, &context)); +# ifdef WITH_CYCLES_LOGGING + optix_assert(optixDeviceContextSetLogCallback( + context, options.logCallbackFunction, options.logCallbackData, options.logCallbackLevel)); +# endif + + /* Fix weird compiler bug that assigns wrong size. */ + launch_params.data_elements = sizeof(KernelParamsOptiX); + + /* Allocate launch parameter buffer memory on device. */ + launch_params.alloc_to_device(1); +} + +OptiXDevice::~OptiXDevice() +{ + /* Make CUDA context current. */ + const CUDAContextScope scope(this); + + free_bvh_memory_delayed(); + + sbt_data.free(); + texture_info.free(); + launch_params.free(); + + /* Unload modules. */ + if (optix_module != NULL) { + optixModuleDestroy(optix_module); + } + for (unsigned int i = 0; i < 2; ++i) { + if (builtin_modules[i] != NULL) { + optixModuleDestroy(builtin_modules[i]); + } + } + for (unsigned int i = 0; i < NUM_PIPELINES; ++i) { + if (pipelines[i] != NULL) { + optixPipelineDestroy(pipelines[i]); + } + } + + optixDeviceContextDestroy(context); +} + +unique_ptr<DeviceQueue> OptiXDevice::gpu_queue_create() +{ + return make_unique<OptiXDeviceQueue>(this); +} + +BVHLayoutMask OptiXDevice::get_bvh_layout_mask() const +{ + /* OptiX has its own internal acceleration structure format. */ + return BVH_LAYOUT_OPTIX; +} + +string OptiXDevice::compile_kernel_get_common_cflags(const uint kernel_features) +{ + string common_cflags = CUDADevice::compile_kernel_get_common_cflags(kernel_features); + + /* Add OptiX SDK include directory to include paths. */ + const char *optix_sdk_path = getenv("OPTIX_ROOT_DIR"); + if (optix_sdk_path) { + common_cflags += string_printf(" -I\"%s/include\"", optix_sdk_path); + } + + /* Specialization for shader raytracing. */ + if (kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) { + common_cflags += " --keep-device-functions"; + } + + return common_cflags; +} + +bool OptiXDevice::load_kernels(const uint kernel_features) +{ + if (have_error()) { + /* Abort early if context creation failed already. */ + return false; + } + + /* Load CUDA modules because we need some of the utility kernels. */ + if (!CUDADevice::load_kernels(kernel_features)) { + return false; + } + + /* Skip creating OptiX module if only doing denoising. */ + if (!(kernel_features & (KERNEL_FEATURE_PATH_TRACING | KERNEL_FEATURE_BAKING))) { + return true; + } + + const CUDAContextScope scope(this); + + /* Unload existing OptiX module and pipelines first. */ + if (optix_module != NULL) { + optixModuleDestroy(optix_module); + optix_module = NULL; + } + for (unsigned int i = 0; i < 2; ++i) { + if (builtin_modules[i] != NULL) { + optixModuleDestroy(builtin_modules[i]); + builtin_modules[i] = NULL; + } + } + for (unsigned int i = 0; i < NUM_PIPELINES; ++i) { + if (pipelines[i] != NULL) { + optixPipelineDestroy(pipelines[i]); + pipelines[i] = NULL; + } + } + + OptixModuleCompileOptions module_options = {}; + module_options.maxRegisterCount = 0; /* Do not set an explicit register limit. */ + + if (DebugFlags().optix.use_debug) { + module_options.optLevel = OPTIX_COMPILE_OPTIMIZATION_LEVEL_0; + module_options.debugLevel = OPTIX_COMPILE_DEBUG_LEVEL_FULL; + } + else { + module_options.optLevel = OPTIX_COMPILE_OPTIMIZATION_LEVEL_3; + module_options.debugLevel = OPTIX_COMPILE_DEBUG_LEVEL_LINEINFO; + } + + module_options.boundValues = nullptr; + module_options.numBoundValues = 0; + + OptixPipelineCompileOptions pipeline_options = {}; + /* Default to no motion blur and two-level graph, since it is the fastest option. */ + pipeline_options.usesMotionBlur = false; + pipeline_options.traversableGraphFlags = + OPTIX_TRAVERSABLE_GRAPH_FLAG_ALLOW_SINGLE_LEVEL_INSTANCING; + pipeline_options.numPayloadValues = 6; + pipeline_options.numAttributeValues = 2; /* u, v */ + pipeline_options.exceptionFlags = OPTIX_EXCEPTION_FLAG_NONE; + pipeline_options.pipelineLaunchParamsVariableName = "__params"; /* See globals.h */ + + pipeline_options.usesPrimitiveTypeFlags = OPTIX_PRIMITIVE_TYPE_FLAGS_TRIANGLE; + if (kernel_features & KERNEL_FEATURE_HAIR) { + if (kernel_features & KERNEL_FEATURE_HAIR_THICK) { + pipeline_options.usesPrimitiveTypeFlags |= OPTIX_PRIMITIVE_TYPE_FLAGS_ROUND_CUBIC_BSPLINE; + } + else + pipeline_options.usesPrimitiveTypeFlags |= OPTIX_PRIMITIVE_TYPE_FLAGS_CUSTOM; + } + + /* Keep track of whether motion blur is enabled, so to enable/disable motion in BVH builds + * This is necessary since objects may be reported to have motion if the Vector pass is + * active, but may still need to be rendered without motion blur if that isn't active as well. */ + motion_blur = (kernel_features & KERNEL_FEATURE_OBJECT_MOTION) != 0; + + if (motion_blur) { + pipeline_options.usesMotionBlur = true; + /* Motion blur can insert motion transforms into the traversal graph. + * It is no longer a two-level graph then, so need to set flags to allow any configuration. */ + pipeline_options.traversableGraphFlags = OPTIX_TRAVERSABLE_GRAPH_FLAG_ALLOW_ANY; + } + + { /* Load and compile PTX module with OptiX kernels. */ + string ptx_data, ptx_filename = path_get((kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) ? + "lib/kernel_optix_shader_raytrace.ptx" : + "lib/kernel_optix.ptx"); + if (use_adaptive_compilation() || path_file_size(ptx_filename) == -1) { + if (!getenv("OPTIX_ROOT_DIR")) { + set_error( + "Missing OPTIX_ROOT_DIR environment variable (which must be set with the path to " + "the Optix SDK to be able to compile Optix kernels on demand)."); + return false; + } + ptx_filename = compile_kernel( + kernel_features, + (kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) ? "kernel_shader_raytrace" : "kernel", + "optix", + true); + } + if (ptx_filename.empty() || !path_read_text(ptx_filename, ptx_data)) { + set_error(string_printf("Failed to load OptiX kernel from '%s'", ptx_filename.c_str())); + return false; + } + + const OptixResult result = optixModuleCreateFromPTX(context, + &module_options, + &pipeline_options, + ptx_data.data(), + ptx_data.size(), + nullptr, + 0, + &optix_module); + if (result != OPTIX_SUCCESS) { + set_error(string_printf("Failed to load OptiX kernel from '%s' (%s)", + ptx_filename.c_str(), + optixGetErrorName(result))); + return false; + } + } + + /* Create program groups. */ + OptixProgramGroup groups[NUM_PROGRAM_GROUPS] = {}; + OptixProgramGroupDesc group_descs[NUM_PROGRAM_GROUPS] = {}; + OptixProgramGroupOptions group_options = {}; /* There are no options currently. */ + group_descs[PG_RGEN_INTERSECT_CLOSEST].kind = OPTIX_PROGRAM_GROUP_KIND_RAYGEN; + group_descs[PG_RGEN_INTERSECT_CLOSEST].raygen.module = optix_module; + group_descs[PG_RGEN_INTERSECT_CLOSEST].raygen.entryFunctionName = + "__raygen__kernel_optix_integrator_intersect_closest"; + group_descs[PG_RGEN_INTERSECT_SHADOW].kind = OPTIX_PROGRAM_GROUP_KIND_RAYGEN; + group_descs[PG_RGEN_INTERSECT_SHADOW].raygen.module = optix_module; + group_descs[PG_RGEN_INTERSECT_SHADOW].raygen.entryFunctionName = + "__raygen__kernel_optix_integrator_intersect_shadow"; + group_descs[PG_RGEN_INTERSECT_SUBSURFACE].kind = OPTIX_PROGRAM_GROUP_KIND_RAYGEN; + group_descs[PG_RGEN_INTERSECT_SUBSURFACE].raygen.module = optix_module; + group_descs[PG_RGEN_INTERSECT_SUBSURFACE].raygen.entryFunctionName = + "__raygen__kernel_optix_integrator_intersect_subsurface"; + group_descs[PG_RGEN_INTERSECT_VOLUME_STACK].kind = OPTIX_PROGRAM_GROUP_KIND_RAYGEN; + group_descs[PG_RGEN_INTERSECT_VOLUME_STACK].raygen.module = optix_module; + group_descs[PG_RGEN_INTERSECT_VOLUME_STACK].raygen.entryFunctionName = + "__raygen__kernel_optix_integrator_intersect_volume_stack"; + group_descs[PG_MISS].kind = OPTIX_PROGRAM_GROUP_KIND_MISS; + group_descs[PG_MISS].miss.module = optix_module; + group_descs[PG_MISS].miss.entryFunctionName = "__miss__kernel_optix_miss"; + group_descs[PG_HITD].kind = OPTIX_PROGRAM_GROUP_KIND_HITGROUP; + group_descs[PG_HITD].hitgroup.moduleCH = optix_module; + group_descs[PG_HITD].hitgroup.entryFunctionNameCH = "__closesthit__kernel_optix_hit"; + group_descs[PG_HITD].hitgroup.moduleAH = optix_module; + group_descs[PG_HITD].hitgroup.entryFunctionNameAH = "__anyhit__kernel_optix_visibility_test"; + group_descs[PG_HITS].kind = OPTIX_PROGRAM_GROUP_KIND_HITGROUP; + group_descs[PG_HITS].hitgroup.moduleAH = optix_module; + group_descs[PG_HITS].hitgroup.entryFunctionNameAH = "__anyhit__kernel_optix_shadow_all_hit"; + + if (kernel_features & KERNEL_FEATURE_HAIR) { + if (kernel_features & KERNEL_FEATURE_HAIR_THICK) { + /* Built-in thick curve intersection. */ + OptixBuiltinISOptions builtin_options = {}; + builtin_options.builtinISModuleType = OPTIX_PRIMITIVE_TYPE_ROUND_CUBIC_BSPLINE; + builtin_options.usesMotionBlur = false; + + optix_assert(optixBuiltinISModuleGet( + context, &module_options, &pipeline_options, &builtin_options, &builtin_modules[0])); + + group_descs[PG_HITD].hitgroup.moduleIS = builtin_modules[0]; + group_descs[PG_HITD].hitgroup.entryFunctionNameIS = nullptr; + group_descs[PG_HITS].hitgroup.moduleIS = builtin_modules[0]; + group_descs[PG_HITS].hitgroup.entryFunctionNameIS = nullptr; + + if (motion_blur) { + builtin_options.usesMotionBlur = true; + + optix_assert(optixBuiltinISModuleGet( + context, &module_options, &pipeline_options, &builtin_options, &builtin_modules[1])); + + group_descs[PG_HITD_MOTION] = group_descs[PG_HITD]; + group_descs[PG_HITD_MOTION].hitgroup.moduleIS = builtin_modules[1]; + group_descs[PG_HITS_MOTION] = group_descs[PG_HITS]; + group_descs[PG_HITS_MOTION].hitgroup.moduleIS = builtin_modules[1]; + } + } + else { + /* Custom ribbon intersection. */ + group_descs[PG_HITD].hitgroup.moduleIS = optix_module; + group_descs[PG_HITS].hitgroup.moduleIS = optix_module; + group_descs[PG_HITD].hitgroup.entryFunctionNameIS = "__intersection__curve_ribbon"; + group_descs[PG_HITS].hitgroup.entryFunctionNameIS = "__intersection__curve_ribbon"; + } + } + + if (kernel_features & (KERNEL_FEATURE_SUBSURFACE | KERNEL_FEATURE_NODE_RAYTRACE)) { + /* Add hit group for local intersections. */ + group_descs[PG_HITL].kind = OPTIX_PROGRAM_GROUP_KIND_HITGROUP; + group_descs[PG_HITL].hitgroup.moduleAH = optix_module; + group_descs[PG_HITL].hitgroup.entryFunctionNameAH = "__anyhit__kernel_optix_local_hit"; + } + + /* Shader raytracing replaces some functions with direct callables. */ + if (kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) { + group_descs[PG_RGEN_SHADE_SURFACE_RAYTRACE].kind = OPTIX_PROGRAM_GROUP_KIND_RAYGEN; + group_descs[PG_RGEN_SHADE_SURFACE_RAYTRACE].raygen.module = optix_module; + group_descs[PG_RGEN_SHADE_SURFACE_RAYTRACE].raygen.entryFunctionName = + "__raygen__kernel_optix_integrator_shade_surface_raytrace"; + group_descs[PG_CALL_SVM_AO].kind = OPTIX_PROGRAM_GROUP_KIND_CALLABLES; + group_descs[PG_CALL_SVM_AO].callables.moduleDC = optix_module; + group_descs[PG_CALL_SVM_AO].callables.entryFunctionNameDC = "__direct_callable__svm_node_ao"; + group_descs[PG_CALL_SVM_BEVEL].kind = OPTIX_PROGRAM_GROUP_KIND_CALLABLES; + group_descs[PG_CALL_SVM_BEVEL].callables.moduleDC = optix_module; + group_descs[PG_CALL_SVM_BEVEL].callables.entryFunctionNameDC = + "__direct_callable__svm_node_bevel"; + group_descs[PG_CALL_AO_PASS].kind = OPTIX_PROGRAM_GROUP_KIND_CALLABLES; + group_descs[PG_CALL_AO_PASS].callables.moduleDC = optix_module; + group_descs[PG_CALL_AO_PASS].callables.entryFunctionNameDC = "__direct_callable__ao_pass"; + } + + optix_assert(optixProgramGroupCreate( + context, group_descs, NUM_PROGRAM_GROUPS, &group_options, nullptr, 0, groups)); + + /* Get program stack sizes. */ + OptixStackSizes stack_size[NUM_PROGRAM_GROUPS] = {}; + /* Set up SBT, which in this case is used only to select between different programs. */ + sbt_data.alloc(NUM_PROGRAM_GROUPS); + memset(sbt_data.host_pointer, 0, sizeof(SbtRecord) * NUM_PROGRAM_GROUPS); + for (unsigned int i = 0; i < NUM_PROGRAM_GROUPS; ++i) { + optix_assert(optixSbtRecordPackHeader(groups[i], &sbt_data[i])); + optix_assert(optixProgramGroupGetStackSize(groups[i], &stack_size[i])); + } + sbt_data.copy_to_device(); /* Upload SBT to device. */ + + /* Calculate maximum trace continuation stack size. */ + unsigned int trace_css = stack_size[PG_HITD].cssCH; + /* This is based on the maximum of closest-hit and any-hit/intersection programs. */ + trace_css = std::max(trace_css, stack_size[PG_HITD].cssIS + stack_size[PG_HITD].cssAH); + trace_css = std::max(trace_css, stack_size[PG_HITS].cssIS + stack_size[PG_HITS].cssAH); + trace_css = std::max(trace_css, stack_size[PG_HITL].cssIS + stack_size[PG_HITL].cssAH); + trace_css = std::max(trace_css, + stack_size[PG_HITD_MOTION].cssIS + stack_size[PG_HITD_MOTION].cssAH); + trace_css = std::max(trace_css, + stack_size[PG_HITS_MOTION].cssIS + stack_size[PG_HITS_MOTION].cssAH); + + OptixPipelineLinkOptions link_options = {}; + link_options.maxTraceDepth = 1; + + if (DebugFlags().optix.use_debug) { + link_options.debugLevel = OPTIX_COMPILE_DEBUG_LEVEL_FULL; + } + else { + link_options.debugLevel = OPTIX_COMPILE_DEBUG_LEVEL_LINEINFO; + } + + if (kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) { + /* Create shader raytracing pipeline. */ + vector<OptixProgramGroup> pipeline_groups; + pipeline_groups.reserve(NUM_PROGRAM_GROUPS); + pipeline_groups.push_back(groups[PG_RGEN_SHADE_SURFACE_RAYTRACE]); + pipeline_groups.push_back(groups[PG_MISS]); + pipeline_groups.push_back(groups[PG_HITD]); + pipeline_groups.push_back(groups[PG_HITS]); + pipeline_groups.push_back(groups[PG_HITL]); + if (motion_blur) { + pipeline_groups.push_back(groups[PG_HITD_MOTION]); + pipeline_groups.push_back(groups[PG_HITS_MOTION]); + } + pipeline_groups.push_back(groups[PG_CALL_SVM_AO]); + pipeline_groups.push_back(groups[PG_CALL_SVM_BEVEL]); + + optix_assert(optixPipelineCreate(context, + &pipeline_options, + &link_options, + pipeline_groups.data(), + pipeline_groups.size(), + nullptr, + 0, + &pipelines[PIP_SHADE_RAYTRACE])); + + /* Combine ray generation and trace continuation stack size. */ + const unsigned int css = stack_size[PG_RGEN_SHADE_SURFACE_RAYTRACE].cssRG + + link_options.maxTraceDepth * trace_css; + const unsigned int dss = std::max(stack_size[PG_CALL_SVM_AO].dssDC, + stack_size[PG_CALL_SVM_BEVEL].dssDC); + + /* Set stack size depending on pipeline options. */ + optix_assert(optixPipelineSetStackSize( + pipelines[PIP_SHADE_RAYTRACE], 0, dss, css, motion_blur ? 3 : 2)); + } + + { /* Create intersection-only pipeline. */ + vector<OptixProgramGroup> pipeline_groups; + pipeline_groups.reserve(NUM_PROGRAM_GROUPS); + pipeline_groups.push_back(groups[PG_RGEN_INTERSECT_CLOSEST]); + pipeline_groups.push_back(groups[PG_RGEN_INTERSECT_SHADOW]); + pipeline_groups.push_back(groups[PG_RGEN_INTERSECT_SUBSURFACE]); + pipeline_groups.push_back(groups[PG_RGEN_INTERSECT_VOLUME_STACK]); + pipeline_groups.push_back(groups[PG_MISS]); + pipeline_groups.push_back(groups[PG_HITD]); + pipeline_groups.push_back(groups[PG_HITS]); + pipeline_groups.push_back(groups[PG_HITL]); + if (motion_blur) { + pipeline_groups.push_back(groups[PG_HITD_MOTION]); + pipeline_groups.push_back(groups[PG_HITS_MOTION]); + } + + optix_assert(optixPipelineCreate(context, + &pipeline_options, + &link_options, + pipeline_groups.data(), + pipeline_groups.size(), + nullptr, + 0, + &pipelines[PIP_INTERSECT])); + + /* Calculate continuation stack size based on the maximum of all ray generation stack sizes. */ + const unsigned int css = + std::max(stack_size[PG_RGEN_INTERSECT_CLOSEST].cssRG, + std::max(stack_size[PG_RGEN_INTERSECT_SHADOW].cssRG, + std::max(stack_size[PG_RGEN_INTERSECT_SUBSURFACE].cssRG, + stack_size[PG_RGEN_INTERSECT_VOLUME_STACK].cssRG))) + + link_options.maxTraceDepth * trace_css; + + optix_assert( + optixPipelineSetStackSize(pipelines[PIP_INTERSECT], 0, 0, css, motion_blur ? 3 : 2)); + } + + /* Clean up program group objects. */ + for (unsigned int i = 0; i < NUM_PROGRAM_GROUPS; ++i) { + optixProgramGroupDestroy(groups[i]); + } + + return true; +} + +/* -------------------------------------------------------------------- + * Buffer denoising. + */ + +class OptiXDevice::DenoiseContext { + public: + explicit DenoiseContext(OptiXDevice *device, const DeviceDenoiseTask &task) + : denoise_params(task.params), + render_buffers(task.render_buffers), + buffer_params(task.buffer_params), + guiding_buffer(device, "denoiser guiding passes buffer"), + num_samples(task.num_samples) + { + num_input_passes = 1; + if (denoise_params.use_pass_albedo) { + num_input_passes += 1; + use_pass_albedo = true; + pass_denoising_albedo = buffer_params.get_pass_offset(PASS_DENOISING_ALBEDO); + if (denoise_params.use_pass_normal) { + num_input_passes += 1; + use_pass_normal = true; + pass_denoising_normal = buffer_params.get_pass_offset(PASS_DENOISING_NORMAL); + } + } + + const int num_guiding_passes = num_input_passes - 1; + + if (num_guiding_passes) { + if (task.allow_inplace_modification) { + guiding_params.device_pointer = render_buffers->buffer.device_pointer; + + guiding_params.pass_albedo = pass_denoising_albedo; + guiding_params.pass_normal = pass_denoising_normal; + + guiding_params.stride = buffer_params.stride; + guiding_params.pass_stride = buffer_params.pass_stride; + } + else { + guiding_params.pass_stride = 0; + if (use_pass_albedo) { + guiding_params.pass_albedo = guiding_params.pass_stride; + guiding_params.pass_stride += 3; + } + if (use_pass_normal) { + guiding_params.pass_normal = guiding_params.pass_stride; + guiding_params.pass_stride += 3; + } + + guiding_params.stride = buffer_params.width; + + guiding_buffer.alloc_to_device(buffer_params.width * buffer_params.height * + guiding_params.pass_stride); + guiding_params.device_pointer = guiding_buffer.device_pointer; + } + } + + pass_sample_count = buffer_params.get_pass_offset(PASS_SAMPLE_COUNT); + } + + const DenoiseParams &denoise_params; + + RenderBuffers *render_buffers = nullptr; + const BufferParams &buffer_params; + + /* Device-side storage of the guiding passes. */ + device_only_memory<float> guiding_buffer; + + struct { + device_ptr device_pointer = 0; + + /* NOTE: Are only initialized when the corresponding guiding pass is enabled. */ + int pass_albedo = PASS_UNUSED; + int pass_normal = PASS_UNUSED; + + int stride = -1; + int pass_stride = -1; + } guiding_params; + + /* Number of input passes. Including the color and extra auxillary passes. */ + int num_input_passes = 0; + bool use_pass_albedo = false; + bool use_pass_normal = false; + + int num_samples = 0; + + int pass_sample_count = PASS_UNUSED; + + /* NOTE: Are only initialized when the corresponding guiding pass is enabled. */ + int pass_denoising_albedo = PASS_UNUSED; + int pass_denoising_normal = PASS_UNUSED; + + /* For passes which don't need albedo channel for denoising we replace the actual albedo with + * the (0.5, 0.5, 0.5). This flag indicates that the real albedo pass has been replaced with + * the fake values and denoising of passes which do need albedo can no longer happen. */ + bool albedo_replaced_with_fake = false; +}; + +class OptiXDevice::DenoisePass { + public: + DenoisePass(const PassType type, const BufferParams &buffer_params) : type(type) + { + noisy_offset = buffer_params.get_pass_offset(type, PassMode::NOISY); + denoised_offset = buffer_params.get_pass_offset(type, PassMode::DENOISED); + + const PassInfo pass_info = Pass::get_info(type); + num_components = pass_info.num_components; + use_compositing = pass_info.use_compositing; + use_denoising_albedo = pass_info.use_denoising_albedo; + } + + PassType type; + + int noisy_offset; + int denoised_offset; + + int num_components; + bool use_compositing; + bool use_denoising_albedo; +}; + +bool OptiXDevice::denoise_buffer(const DeviceDenoiseTask &task) +{ + const CUDAContextScope scope(this); + + DenoiseContext context(this, task); + + if (!denoise_ensure(context)) { + return false; + } + + if (!denoise_filter_guiding_preprocess(context)) { + LOG(ERROR) << "Error preprocessing guiding passes."; + return false; + } + + /* Passes which will use real albedo when it is available. */ + denoise_pass(context, PASS_COMBINED); + denoise_pass(context, PASS_SHADOW_CATCHER_MATTE); + + /* Passes which do not need albedo and hence if real is present it needs to become fake. */ + denoise_pass(context, PASS_SHADOW_CATCHER); + + return true; +} + +DeviceQueue *OptiXDevice::get_denoise_queue() +{ + return &denoiser_.queue; +} + +bool OptiXDevice::denoise_filter_guiding_preprocess(DenoiseContext &context) +{ + const BufferParams &buffer_params = context.buffer_params; + + const int work_size = buffer_params.width * buffer_params.height; + + void *args[] = {const_cast<device_ptr *>(&context.guiding_params.device_pointer), + const_cast<int *>(&context.guiding_params.pass_stride), + const_cast<int *>(&context.guiding_params.pass_albedo), + const_cast<int *>(&context.guiding_params.pass_normal), + &context.render_buffers->buffer.device_pointer, + const_cast<int *>(&buffer_params.offset), + const_cast<int *>(&buffer_params.stride), + const_cast<int *>(&buffer_params.pass_stride), + const_cast<int *>(&context.pass_sample_count), + const_cast<int *>(&context.pass_denoising_albedo), + const_cast<int *>(&context.pass_denoising_normal), + const_cast<int *>(&buffer_params.full_x), + const_cast<int *>(&buffer_params.full_y), + const_cast<int *>(&buffer_params.width), + const_cast<int *>(&buffer_params.height), + const_cast<int *>(&context.num_samples)}; + + return denoiser_.queue.enqueue(DEVICE_KERNEL_FILTER_GUIDING_PREPROCESS, work_size, args); +} + +bool OptiXDevice::denoise_filter_guiding_set_fake_albedo(DenoiseContext &context) +{ + const BufferParams &buffer_params = context.buffer_params; + + const int work_size = buffer_params.width * buffer_params.height; + + void *args[] = {const_cast<device_ptr *>(&context.guiding_params.device_pointer), + const_cast<int *>(&context.guiding_params.pass_stride), + const_cast<int *>(&context.guiding_params.pass_albedo), + const_cast<int *>(&buffer_params.width), + const_cast<int *>(&buffer_params.height)}; + + return denoiser_.queue.enqueue(DEVICE_KERNEL_FILTER_GUIDING_SET_FAKE_ALBEDO, work_size, args); +} + +void OptiXDevice::denoise_pass(DenoiseContext &context, PassType pass_type) +{ + const BufferParams &buffer_params = context.buffer_params; + + const DenoisePass pass(pass_type, buffer_params); + + if (pass.noisy_offset == PASS_UNUSED) { + return; + } + if (pass.denoised_offset == PASS_UNUSED) { + LOG(DFATAL) << "Missing denoised pass " << pass_type_as_string(pass_type); + return; + } + + if (pass.use_denoising_albedo) { + if (context.albedo_replaced_with_fake) { + LOG(ERROR) << "Pass which requires albedo is denoised after fake albedo has been set."; + return; + } + } + else if (!context.albedo_replaced_with_fake) { + context.albedo_replaced_with_fake = true; + if (!denoise_filter_guiding_set_fake_albedo(context)) { + LOG(ERROR) << "Error replacing real albedo with the fake one."; + return; + } + } + + /* Read and preprocess noisy color input pass. */ + denoise_color_read(context, pass); + if (!denoise_filter_color_preprocess(context, pass)) { + LOG(ERROR) << "Error connverting denoising passes to RGB buffer."; + return; + } + + if (!denoise_run(context, pass)) { + LOG(ERROR) << "Error running OptiX denoiser."; + return; + } + + /* Store result in the combined pass of the render buffer. + * + * This will scale the denoiser result up to match the number of, possibly per-pixel, samples. */ + if (!denoise_filter_color_postprocess(context, pass)) { + LOG(ERROR) << "Error copying denoiser result to the denoised pass."; + return; + } + + denoiser_.queue.synchronize(); +} + +void OptiXDevice::denoise_color_read(DenoiseContext &context, const DenoisePass &pass) +{ + PassAccessor::PassAccessInfo pass_access_info; + pass_access_info.type = pass.type; + pass_access_info.mode = PassMode::NOISY; + pass_access_info.offset = pass.noisy_offset; + + /* Denoiser operates on passes which are used to calculate the approximation, and is never used + * on the approximation. The latter is not even possible because OptiX does not support + * denoising of semi-transparent pixels. */ + pass_access_info.use_approximate_shadow_catcher = false; + pass_access_info.use_approximate_shadow_catcher_background = false; + pass_access_info.show_active_pixels = false; + + /* TODO(sergey): Consider adding support of actual exposure, to avoid clamping in extreme cases. + */ + const PassAccessorGPU pass_accessor( + &denoiser_.queue, pass_access_info, 1.0f, context.num_samples); + + PassAccessor::Destination destination(pass_access_info.type); + destination.d_pixels = context.render_buffers->buffer.device_pointer + + pass.denoised_offset * sizeof(float); + destination.num_components = 3; + destination.pixel_stride = context.buffer_params.pass_stride; + + pass_accessor.get_render_tile_pixels(context.render_buffers, context.buffer_params, destination); +} + +bool OptiXDevice::denoise_filter_color_preprocess(DenoiseContext &context, const DenoisePass &pass) +{ + const BufferParams &buffer_params = context.buffer_params; + + const int work_size = buffer_params.width * buffer_params.height; + + void *args[] = {&context.render_buffers->buffer.device_pointer, + const_cast<int *>(&buffer_params.full_x), + const_cast<int *>(&buffer_params.full_y), + const_cast<int *>(&buffer_params.width), + const_cast<int *>(&buffer_params.height), + const_cast<int *>(&buffer_params.offset), + const_cast<int *>(&buffer_params.stride), + const_cast<int *>(&buffer_params.pass_stride), + const_cast<int *>(&pass.denoised_offset)}; + + return denoiser_.queue.enqueue(DEVICE_KERNEL_FILTER_COLOR_PREPROCESS, work_size, args); +} + +bool OptiXDevice::denoise_filter_color_postprocess(DenoiseContext &context, + const DenoisePass &pass) +{ + const BufferParams &buffer_params = context.buffer_params; + + const int work_size = buffer_params.width * buffer_params.height; + + void *args[] = {&context.render_buffers->buffer.device_pointer, + const_cast<int *>(&buffer_params.full_x), + const_cast<int *>(&buffer_params.full_y), + const_cast<int *>(&buffer_params.width), + const_cast<int *>(&buffer_params.height), + const_cast<int *>(&buffer_params.offset), + const_cast<int *>(&buffer_params.stride), + const_cast<int *>(&buffer_params.pass_stride), + const_cast<int *>(&context.num_samples), + const_cast<int *>(&pass.noisy_offset), + const_cast<int *>(&pass.denoised_offset), + const_cast<int *>(&context.pass_sample_count), + const_cast<int *>(&pass.num_components), + const_cast<bool *>(&pass.use_compositing)}; + + return denoiser_.queue.enqueue(DEVICE_KERNEL_FILTER_COLOR_POSTPROCESS, work_size, args); +} + +bool OptiXDevice::denoise_ensure(DenoiseContext &context) +{ + if (!denoise_create_if_needed(context)) { + LOG(ERROR) << "OptiX denoiser creation has failed."; + return false; + } + + if (!denoise_configure_if_needed(context)) { + LOG(ERROR) << "OptiX denoiser configuration has failed."; + return false; + } + + return true; +} + +bool OptiXDevice::denoise_create_if_needed(DenoiseContext &context) +{ + const bool recreate_denoiser = (denoiser_.optix_denoiser == nullptr) || + (denoiser_.use_pass_albedo != context.use_pass_albedo) || + (denoiser_.use_pass_normal != context.use_pass_normal); + if (!recreate_denoiser) { + return true; + } + + /* Destroy existing handle before creating new one. */ + if (denoiser_.optix_denoiser) { + optixDenoiserDestroy(denoiser_.optix_denoiser); + } + + /* Create OptiX denoiser handle on demand when it is first used. */ + OptixDenoiserOptions denoiser_options = {}; + denoiser_options.guideAlbedo = context.use_pass_albedo; + denoiser_options.guideNormal = context.use_pass_normal; + const OptixResult result = optixDenoiserCreate( + this->context, OPTIX_DENOISER_MODEL_KIND_HDR, &denoiser_options, &denoiser_.optix_denoiser); + + if (result != OPTIX_SUCCESS) { + set_error("Failed to create OptiX denoiser"); + return false; + } + + /* OptiX denoiser handle was created with the requested number of input passes. */ + denoiser_.use_pass_albedo = context.use_pass_albedo; + denoiser_.use_pass_normal = context.use_pass_normal; + + /* OptiX denoiser has been created, but it needs configuration. */ + denoiser_.is_configured = false; + + return true; +} + +bool OptiXDevice::denoise_configure_if_needed(DenoiseContext &context) +{ + if (denoiser_.is_configured && (denoiser_.configured_size.x == context.buffer_params.width && + denoiser_.configured_size.y == context.buffer_params.height)) { + return true; + } + + const BufferParams &buffer_params = context.buffer_params; + + OptixDenoiserSizes sizes = {}; + optix_assert(optixDenoiserComputeMemoryResources( + denoiser_.optix_denoiser, buffer_params.width, buffer_params.height, &sizes)); + + denoiser_.scratch_size = sizes.withOverlapScratchSizeInBytes; + denoiser_.scratch_offset = sizes.stateSizeInBytes; + + /* Allocate denoiser state if tile size has changed since last setup. */ + denoiser_.state.alloc_to_device(denoiser_.scratch_offset + denoiser_.scratch_size); + + /* Initialize denoiser state for the current tile size. */ + const OptixResult result = optixDenoiserSetup(denoiser_.optix_denoiser, + denoiser_.queue.stream(), + buffer_params.width, + buffer_params.height, + denoiser_.state.device_pointer, + denoiser_.scratch_offset, + denoiser_.state.device_pointer + + denoiser_.scratch_offset, + denoiser_.scratch_size); + if (result != OPTIX_SUCCESS) { + set_error("Failed to set up OptiX denoiser"); + return false; + } + + denoiser_.is_configured = true; + denoiser_.configured_size.x = buffer_params.width; + denoiser_.configured_size.y = buffer_params.height; + + return true; +} + +bool OptiXDevice::denoise_run(DenoiseContext &context, const DenoisePass &pass) +{ + const BufferParams &buffer_params = context.buffer_params; + const int width = buffer_params.width; + const int height = buffer_params.height; + + /* Set up input and output layer information. */ + OptixImage2D color_layer = {0}; + OptixImage2D albedo_layer = {0}; + OptixImage2D normal_layer = {0}; + + OptixImage2D output_layer = {0}; + + /* Color pass. */ + { + const int pass_denoised = pass.denoised_offset; + const int64_t pass_stride_in_bytes = context.buffer_params.pass_stride * sizeof(float); + + color_layer.data = context.render_buffers->buffer.device_pointer + + pass_denoised * sizeof(float); + color_layer.width = width; + color_layer.height = height; + color_layer.rowStrideInBytes = pass_stride_in_bytes * context.buffer_params.stride; + color_layer.pixelStrideInBytes = pass_stride_in_bytes; + color_layer.format = OPTIX_PIXEL_FORMAT_FLOAT3; + } + + device_vector<float> fake_albedo(this, "fake_albedo", MEM_READ_WRITE); + + /* Optional albedo and color passes. */ + if (context.num_input_passes > 1) { + const device_ptr d_guiding_buffer = context.guiding_params.device_pointer; + const int64_t pixel_stride_in_bytes = context.guiding_params.pass_stride * sizeof(float); + const int64_t row_stride_in_bytes = context.guiding_params.stride * pixel_stride_in_bytes; + + if (context.use_pass_albedo) { + albedo_layer.data = d_guiding_buffer + context.guiding_params.pass_albedo * sizeof(float); + albedo_layer.width = width; + albedo_layer.height = height; + albedo_layer.rowStrideInBytes = row_stride_in_bytes; + albedo_layer.pixelStrideInBytes = pixel_stride_in_bytes; + albedo_layer.format = OPTIX_PIXEL_FORMAT_FLOAT3; + } + + if (context.use_pass_normal) { + normal_layer.data = d_guiding_buffer + context.guiding_params.pass_normal * sizeof(float); + normal_layer.width = width; + normal_layer.height = height; + normal_layer.rowStrideInBytes = row_stride_in_bytes; + normal_layer.pixelStrideInBytes = pixel_stride_in_bytes; + normal_layer.format = OPTIX_PIXEL_FORMAT_FLOAT3; + } + } + + /* Denoise in-place of the noisy input in the render buffers. */ + output_layer = color_layer; + + /* Finally run denonising. */ + OptixDenoiserParams params = {}; /* All parameters are disabled/zero. */ + OptixDenoiserLayer image_layers = {}; + image_layers.input = color_layer; + image_layers.output = output_layer; + + OptixDenoiserGuideLayer guide_layers = {}; + guide_layers.albedo = albedo_layer; + guide_layers.normal = normal_layer; + + optix_assert(optixDenoiserInvoke(denoiser_.optix_denoiser, + denoiser_.queue.stream(), + ¶ms, + denoiser_.state.device_pointer, + denoiser_.scratch_offset, + &guide_layers, + &image_layers, + 1, + 0, + 0, + denoiser_.state.device_pointer + denoiser_.scratch_offset, + denoiser_.scratch_size)); + + return true; +} + +bool OptiXDevice::build_optix_bvh(BVHOptiX *bvh, + OptixBuildOperation operation, + const OptixBuildInput &build_input, + uint16_t num_motion_steps) +{ + const CUDAContextScope scope(this); + + const bool use_fast_trace_bvh = (bvh->params.bvh_type == BVH_TYPE_STATIC); + + /* Compute memory usage. */ + OptixAccelBufferSizes sizes = {}; + OptixAccelBuildOptions options = {}; + options.operation = operation; + if (use_fast_trace_bvh) { + VLOG(2) << "Using fast to trace OptiX BVH"; + options.buildFlags = OPTIX_BUILD_FLAG_PREFER_FAST_TRACE | OPTIX_BUILD_FLAG_ALLOW_COMPACTION; + } + else { + VLOG(2) << "Using fast to update OptiX BVH"; + options.buildFlags = OPTIX_BUILD_FLAG_PREFER_FAST_BUILD | OPTIX_BUILD_FLAG_ALLOW_UPDATE; + } + + options.motionOptions.numKeys = num_motion_steps; + options.motionOptions.flags = OPTIX_MOTION_FLAG_START_VANISH | OPTIX_MOTION_FLAG_END_VANISH; + options.motionOptions.timeBegin = 0.0f; + options.motionOptions.timeEnd = 1.0f; + + optix_assert(optixAccelComputeMemoryUsage(context, &options, &build_input, 1, &sizes)); + + /* Allocate required output buffers. */ + device_only_memory<char> temp_mem(this, "optix temp as build mem"); + temp_mem.alloc_to_device(align_up(sizes.tempSizeInBytes, 8) + 8); + if (!temp_mem.device_pointer) { + /* Make sure temporary memory allocation succeeded. */ + return false; + } + + device_only_memory<char> &out_data = bvh->as_data; + if (operation == OPTIX_BUILD_OPERATION_BUILD) { + assert(out_data.device == this); + out_data.alloc_to_device(sizes.outputSizeInBytes); + if (!out_data.device_pointer) { + return false; + } + } + else { + assert(out_data.device_pointer && out_data.device_size >= sizes.outputSizeInBytes); + } + + /* Finally build the acceleration structure. */ + OptixAccelEmitDesc compacted_size_prop = {}; + compacted_size_prop.type = OPTIX_PROPERTY_TYPE_COMPACTED_SIZE; + /* A tiny space was allocated for this property at the end of the temporary buffer above. + * Make sure this pointer is 8-byte aligned. */ + compacted_size_prop.result = align_up(temp_mem.device_pointer + sizes.tempSizeInBytes, 8); + + OptixTraversableHandle out_handle = 0; + optix_assert(optixAccelBuild(context, + NULL, + &options, + &build_input, + 1, + temp_mem.device_pointer, + sizes.tempSizeInBytes, + out_data.device_pointer, + sizes.outputSizeInBytes, + &out_handle, + use_fast_trace_bvh ? &compacted_size_prop : NULL, + use_fast_trace_bvh ? 1 : 0)); + bvh->traversable_handle = static_cast<uint64_t>(out_handle); + + /* Wait for all operations to finish. */ + cuda_assert(cuStreamSynchronize(NULL)); + + /* Compact acceleration structure to save memory (do not do this in viewport for faster builds). + */ + if (use_fast_trace_bvh) { + uint64_t compacted_size = sizes.outputSizeInBytes; + cuda_assert(cuMemcpyDtoH(&compacted_size, compacted_size_prop.result, sizeof(compacted_size))); + + /* Temporary memory is no longer needed, so free it now to make space. */ + temp_mem.free(); + + /* There is no point compacting if the size does not change. */ + if (compacted_size < sizes.outputSizeInBytes) { + device_only_memory<char> compacted_data(this, "optix compacted as"); + compacted_data.alloc_to_device(compacted_size); + if (!compacted_data.device_pointer) + /* Do not compact if memory allocation for compacted acceleration structure fails. + * Can just use the uncompacted one then, so succeed here regardless. */ + return !have_error(); + + optix_assert(optixAccelCompact( + context, NULL, out_handle, compacted_data.device_pointer, compacted_size, &out_handle)); + bvh->traversable_handle = static_cast<uint64_t>(out_handle); + + /* Wait for compaction to finish. */ + cuda_assert(cuStreamSynchronize(NULL)); + + std::swap(out_data.device_size, compacted_data.device_size); + std::swap(out_data.device_pointer, compacted_data.device_pointer); + } + } + + return !have_error(); +} + +void OptiXDevice::build_bvh(BVH *bvh, Progress &progress, bool refit) +{ + const bool use_fast_trace_bvh = (bvh->params.bvh_type == BVH_TYPE_STATIC); + + free_bvh_memory_delayed(); + + BVHOptiX *const bvh_optix = static_cast<BVHOptiX *>(bvh); + + progress.set_substatus("Building OptiX acceleration structure"); + + if (!bvh->params.top_level) { + assert(bvh->objects.size() == 1 && bvh->geometry.size() == 1); + + /* Refit is only possible in viewport for now (because AS is built with + * OPTIX_BUILD_FLAG_ALLOW_UPDATE only there, see above). */ + OptixBuildOperation operation = OPTIX_BUILD_OPERATION_BUILD; + if (refit && !use_fast_trace_bvh) { + assert(bvh_optix->traversable_handle != 0); + operation = OPTIX_BUILD_OPERATION_UPDATE; + } + else { + bvh_optix->as_data.free(); + bvh_optix->traversable_handle = 0; + } + + /* Build bottom level acceleration structures (BLAS). */ + Geometry *const geom = bvh->geometry[0]; + if (geom->geometry_type == Geometry::HAIR) { + /* Build BLAS for curve primitives. */ + Hair *const hair = static_cast<Hair *const>(geom); + if (hair->num_curves() == 0) { + return; + } + + const size_t num_segments = hair->num_segments(); + + size_t num_motion_steps = 1; + Attribute *motion_keys = hair->attributes.find(ATTR_STD_MOTION_VERTEX_POSITION); + if (motion_blur && hair->get_use_motion_blur() && motion_keys) { + num_motion_steps = hair->get_motion_steps(); + } + + device_vector<OptixAabb> aabb_data(this, "optix temp aabb data", MEM_READ_ONLY); + device_vector<int> index_data(this, "optix temp index data", MEM_READ_ONLY); + device_vector<float4> vertex_data(this, "optix temp vertex data", MEM_READ_ONLY); + /* Four control points for each curve segment. */ + const size_t num_vertices = num_segments * 4; + if (hair->curve_shape == CURVE_THICK) { + index_data.alloc(num_segments); + vertex_data.alloc(num_vertices * num_motion_steps); + } + else + aabb_data.alloc(num_segments * num_motion_steps); + + /* Get AABBs for each motion step. */ + for (size_t step = 0; step < num_motion_steps; ++step) { + /* The center step for motion vertices is not stored in the attribute. */ + const float3 *keys = hair->get_curve_keys().data(); + size_t center_step = (num_motion_steps - 1) / 2; + if (step != center_step) { + size_t attr_offset = (step > center_step) ? step - 1 : step; + /* Technically this is a float4 array, but sizeof(float3) == sizeof(float4). */ + keys = motion_keys->data_float3() + attr_offset * hair->get_curve_keys().size(); + } + + for (size_t j = 0, i = 0; j < hair->num_curves(); ++j) { + const Hair::Curve curve = hair->get_curve(j); + const array<float> &curve_radius = hair->get_curve_radius(); + + for (int segment = 0; segment < curve.num_segments(); ++segment, ++i) { + if (hair->curve_shape == CURVE_THICK) { + int k0 = curve.first_key + segment; + int k1 = k0 + 1; + int ka = max(k0 - 1, curve.first_key); + int kb = min(k1 + 1, curve.first_key + curve.num_keys - 1); + + const float4 px = make_float4(keys[ka].x, keys[k0].x, keys[k1].x, keys[kb].x); + const float4 py = make_float4(keys[ka].y, keys[k0].y, keys[k1].y, keys[kb].y); + const float4 pz = make_float4(keys[ka].z, keys[k0].z, keys[k1].z, keys[kb].z); + const float4 pw = make_float4( + curve_radius[ka], curve_radius[k0], curve_radius[k1], curve_radius[kb]); + + /* Convert Catmull-Rom data to Bezier spline. */ + static const float4 cr2bsp0 = make_float4(+7, -4, +5, -2) / 6.f; + static const float4 cr2bsp1 = make_float4(-2, 11, -4, +1) / 6.f; + static const float4 cr2bsp2 = make_float4(+1, -4, 11, -2) / 6.f; + static const float4 cr2bsp3 = make_float4(-2, +5, -4, +7) / 6.f; + + index_data[i] = i * 4; + float4 *const v = vertex_data.data() + step * num_vertices + index_data[i]; + v[0] = make_float4( + dot(cr2bsp0, px), dot(cr2bsp0, py), dot(cr2bsp0, pz), dot(cr2bsp0, pw)); + v[1] = make_float4( + dot(cr2bsp1, px), dot(cr2bsp1, py), dot(cr2bsp1, pz), dot(cr2bsp1, pw)); + v[2] = make_float4( + dot(cr2bsp2, px), dot(cr2bsp2, py), dot(cr2bsp2, pz), dot(cr2bsp2, pw)); + v[3] = make_float4( + dot(cr2bsp3, px), dot(cr2bsp3, py), dot(cr2bsp3, pz), dot(cr2bsp3, pw)); + } + else { + BoundBox bounds = BoundBox::empty; + curve.bounds_grow(segment, keys, hair->get_curve_radius().data(), bounds); + + const size_t index = step * num_segments + i; + aabb_data[index].minX = bounds.min.x; + aabb_data[index].minY = bounds.min.y; + aabb_data[index].minZ = bounds.min.z; + aabb_data[index].maxX = bounds.max.x; + aabb_data[index].maxY = bounds.max.y; + aabb_data[index].maxZ = bounds.max.z; + } + } + } + } + + /* Upload AABB data to GPU. */ + aabb_data.copy_to_device(); + index_data.copy_to_device(); + vertex_data.copy_to_device(); + + vector<device_ptr> aabb_ptrs; + aabb_ptrs.reserve(num_motion_steps); + vector<device_ptr> width_ptrs; + vector<device_ptr> vertex_ptrs; + width_ptrs.reserve(num_motion_steps); + vertex_ptrs.reserve(num_motion_steps); + for (size_t step = 0; step < num_motion_steps; ++step) { + aabb_ptrs.push_back(aabb_data.device_pointer + step * num_segments * sizeof(OptixAabb)); + const device_ptr base_ptr = vertex_data.device_pointer + + step * num_vertices * sizeof(float4); + width_ptrs.push_back(base_ptr + 3 * sizeof(float)); /* Offset by vertex size. */ + vertex_ptrs.push_back(base_ptr); + } + + /* Force a single any-hit call, so shadow record-all behavior works correctly. */ + unsigned int build_flags = OPTIX_GEOMETRY_FLAG_REQUIRE_SINGLE_ANYHIT_CALL; + OptixBuildInput build_input = {}; + if (hair->curve_shape == CURVE_THICK) { + build_input.type = OPTIX_BUILD_INPUT_TYPE_CURVES; + build_input.curveArray.curveType = OPTIX_PRIMITIVE_TYPE_ROUND_CUBIC_BSPLINE; + build_input.curveArray.numPrimitives = num_segments; + build_input.curveArray.vertexBuffers = (CUdeviceptr *)vertex_ptrs.data(); + build_input.curveArray.numVertices = num_vertices; + build_input.curveArray.vertexStrideInBytes = sizeof(float4); + build_input.curveArray.widthBuffers = (CUdeviceptr *)width_ptrs.data(); + build_input.curveArray.widthStrideInBytes = sizeof(float4); + build_input.curveArray.indexBuffer = (CUdeviceptr)index_data.device_pointer; + build_input.curveArray.indexStrideInBytes = sizeof(int); + build_input.curveArray.flag = build_flags; + build_input.curveArray.primitiveIndexOffset = hair->optix_prim_offset; + } + else { + /* Disable visibility test any-hit program, since it is already checked during + * intersection. Those trace calls that require anyhit can force it with a ray flag. */ + build_flags |= OPTIX_GEOMETRY_FLAG_DISABLE_ANYHIT; + + build_input.type = OPTIX_BUILD_INPUT_TYPE_CUSTOM_PRIMITIVES; + build_input.customPrimitiveArray.aabbBuffers = (CUdeviceptr *)aabb_ptrs.data(); + build_input.customPrimitiveArray.numPrimitives = num_segments; + build_input.customPrimitiveArray.strideInBytes = sizeof(OptixAabb); + build_input.customPrimitiveArray.flags = &build_flags; + build_input.customPrimitiveArray.numSbtRecords = 1; + build_input.customPrimitiveArray.primitiveIndexOffset = hair->optix_prim_offset; + } + + if (!build_optix_bvh(bvh_optix, operation, build_input, num_motion_steps)) { + progress.set_error("Failed to build OptiX acceleration structure"); + } + } + else if (geom->geometry_type == Geometry::MESH || geom->geometry_type == Geometry::VOLUME) { + /* Build BLAS for triangle primitives. */ + Mesh *const mesh = static_cast<Mesh *const>(geom); + if (mesh->num_triangles() == 0) { + return; + } + + const size_t num_verts = mesh->get_verts().size(); + + size_t num_motion_steps = 1; + Attribute *motion_keys = mesh->attributes.find(ATTR_STD_MOTION_VERTEX_POSITION); + if (motion_blur && mesh->get_use_motion_blur() && motion_keys) { + num_motion_steps = mesh->get_motion_steps(); + } + + device_vector<int> index_data(this, "optix temp index data", MEM_READ_ONLY); + index_data.alloc(mesh->get_triangles().size()); + memcpy(index_data.data(), + mesh->get_triangles().data(), + mesh->get_triangles().size() * sizeof(int)); + device_vector<float4> vertex_data(this, "optix temp vertex data", MEM_READ_ONLY); + vertex_data.alloc(num_verts * num_motion_steps); + + for (size_t step = 0; step < num_motion_steps; ++step) { + const float3 *verts = mesh->get_verts().data(); + + size_t center_step = (num_motion_steps - 1) / 2; + /* The center step for motion vertices is not stored in the attribute. */ + if (step != center_step) { + verts = motion_keys->data_float3() + (step > center_step ? step - 1 : step) * num_verts; + } + + memcpy(vertex_data.data() + num_verts * step, verts, num_verts * sizeof(float3)); + } + + /* Upload triangle data to GPU. */ + index_data.copy_to_device(); + vertex_data.copy_to_device(); + + vector<device_ptr> vertex_ptrs; + vertex_ptrs.reserve(num_motion_steps); + for (size_t step = 0; step < num_motion_steps; ++step) { + vertex_ptrs.push_back(vertex_data.device_pointer + num_verts * step * sizeof(float3)); + } + + /* Force a single any-hit call, so shadow record-all behavior works correctly. */ + unsigned int build_flags = OPTIX_GEOMETRY_FLAG_REQUIRE_SINGLE_ANYHIT_CALL; + OptixBuildInput build_input = {}; + build_input.type = OPTIX_BUILD_INPUT_TYPE_TRIANGLES; + build_input.triangleArray.vertexBuffers = (CUdeviceptr *)vertex_ptrs.data(); + build_input.triangleArray.numVertices = num_verts; + build_input.triangleArray.vertexFormat = OPTIX_VERTEX_FORMAT_FLOAT3; + build_input.triangleArray.vertexStrideInBytes = sizeof(float4); + build_input.triangleArray.indexBuffer = index_data.device_pointer; + build_input.triangleArray.numIndexTriplets = mesh->num_triangles(); + build_input.triangleArray.indexFormat = OPTIX_INDICES_FORMAT_UNSIGNED_INT3; + build_input.triangleArray.indexStrideInBytes = 3 * sizeof(int); + build_input.triangleArray.flags = &build_flags; + /* The SBT does not store per primitive data since Cycles already allocates separate + * buffers for that purpose. OptiX does not allow this to be zero though, so just pass in + * one and rely on that having the same meaning in this case. */ + build_input.triangleArray.numSbtRecords = 1; + build_input.triangleArray.primitiveIndexOffset = mesh->optix_prim_offset; + + if (!build_optix_bvh(bvh_optix, operation, build_input, num_motion_steps)) { + progress.set_error("Failed to build OptiX acceleration structure"); + } + } + } + else { + unsigned int num_instances = 0; + unsigned int max_num_instances = 0xFFFFFFFF; + + bvh_optix->as_data.free(); + bvh_optix->traversable_handle = 0; + bvh_optix->motion_transform_data.free(); + + optixDeviceContextGetProperty(context, + OPTIX_DEVICE_PROPERTY_LIMIT_MAX_INSTANCE_ID, + &max_num_instances, + sizeof(max_num_instances)); + /* Do not count first bit, which is used to distinguish instanced and non-instanced objects. */ + max_num_instances >>= 1; + if (bvh->objects.size() > max_num_instances) { + progress.set_error( + "Failed to build OptiX acceleration structure because there are too many instances"); + return; + } + + /* Fill instance descriptions. */ + device_vector<OptixInstance> instances(this, "optix tlas instances", MEM_READ_ONLY); + instances.alloc(bvh->objects.size()); + + /* Calculate total motion transform size and allocate memory for them. */ + size_t motion_transform_offset = 0; + if (motion_blur) { + size_t total_motion_transform_size = 0; + for (Object *const ob : bvh->objects) { + if (ob->is_traceable() && ob->use_motion()) { + total_motion_transform_size = align_up(total_motion_transform_size, + OPTIX_TRANSFORM_BYTE_ALIGNMENT); + const size_t motion_keys = max(ob->get_motion().size(), 2) - 2; + total_motion_transform_size = total_motion_transform_size + + sizeof(OptixSRTMotionTransform) + + motion_keys * sizeof(OptixSRTData); + } + } + + assert(bvh_optix->motion_transform_data.device == this); + bvh_optix->motion_transform_data.alloc_to_device(total_motion_transform_size); + } + + for (Object *ob : bvh->objects) { + /* Skip non-traceable objects. */ + if (!ob->is_traceable()) { + continue; + } + + BVHOptiX *const blas = static_cast<BVHOptiX *>(ob->get_geometry()->bvh); + OptixTraversableHandle handle = blas->traversable_handle; + + OptixInstance &instance = instances[num_instances++]; + memset(&instance, 0, sizeof(instance)); + + /* Clear transform to identity matrix. */ + instance.transform[0] = 1.0f; + instance.transform[5] = 1.0f; + instance.transform[10] = 1.0f; + + /* Set user instance ID to object index (but leave low bit blank). */ + instance.instanceId = ob->get_device_index() << 1; + + /* Have to have at least one bit in the mask, or else instance would always be culled. */ + instance.visibilityMask = 1; + + if (ob->get_geometry()->has_volume) { + /* Volumes have a special bit set in the visibility mask so a trace can mask only volumes. + */ + instance.visibilityMask |= 2; + } + + if (ob->get_geometry()->geometry_type == Geometry::HAIR) { + /* Same applies to curves (so they can be skipped in local trace calls). */ + instance.visibilityMask |= 4; + + if (motion_blur && ob->get_geometry()->has_motion_blur() && + static_cast<const Hair *>(ob->get_geometry())->curve_shape == CURVE_THICK) { + /* Select between motion blur and non-motion blur built-in intersection module. */ + instance.sbtOffset = PG_HITD_MOTION - PG_HITD; + } + } + + /* Insert motion traversable if object has motion. */ + if (motion_blur && ob->use_motion()) { + size_t motion_keys = max(ob->get_motion().size(), 2) - 2; + size_t motion_transform_size = sizeof(OptixSRTMotionTransform) + + motion_keys * sizeof(OptixSRTData); + + const CUDAContextScope scope(this); + + motion_transform_offset = align_up(motion_transform_offset, + OPTIX_TRANSFORM_BYTE_ALIGNMENT); + CUdeviceptr motion_transform_gpu = bvh_optix->motion_transform_data.device_pointer + + motion_transform_offset; + motion_transform_offset += motion_transform_size; + + /* Allocate host side memory for motion transform and fill it with transform data. */ + OptixSRTMotionTransform &motion_transform = *reinterpret_cast<OptixSRTMotionTransform *>( + new uint8_t[motion_transform_size]); + motion_transform.child = handle; + motion_transform.motionOptions.numKeys = ob->get_motion().size(); + motion_transform.motionOptions.flags = OPTIX_MOTION_FLAG_NONE; + motion_transform.motionOptions.timeBegin = 0.0f; + motion_transform.motionOptions.timeEnd = 1.0f; + + OptixSRTData *const srt_data = motion_transform.srtData; + array<DecomposedTransform> decomp(ob->get_motion().size()); + transform_motion_decompose( + decomp.data(), ob->get_motion().data(), ob->get_motion().size()); + + for (size_t i = 0; i < ob->get_motion().size(); ++i) { + /* Scale. */ + srt_data[i].sx = decomp[i].y.w; /* scale.x.x */ + srt_data[i].sy = decomp[i].z.w; /* scale.y.y */ + srt_data[i].sz = decomp[i].w.w; /* scale.z.z */ + + /* Shear. */ + srt_data[i].a = decomp[i].z.x; /* scale.x.y */ + srt_data[i].b = decomp[i].z.y; /* scale.x.z */ + srt_data[i].c = decomp[i].w.x; /* scale.y.z */ + assert(decomp[i].z.z == 0.0f); /* scale.y.x */ + assert(decomp[i].w.y == 0.0f); /* scale.z.x */ + assert(decomp[i].w.z == 0.0f); /* scale.z.y */ + + /* Pivot point. */ + srt_data[i].pvx = 0.0f; + srt_data[i].pvy = 0.0f; + srt_data[i].pvz = 0.0f; + + /* Rotation. */ + srt_data[i].qx = decomp[i].x.x; + srt_data[i].qy = decomp[i].x.y; + srt_data[i].qz = decomp[i].x.z; + srt_data[i].qw = decomp[i].x.w; + + /* Translation. */ + srt_data[i].tx = decomp[i].y.x; + srt_data[i].ty = decomp[i].y.y; + srt_data[i].tz = decomp[i].y.z; + } + + /* Upload motion transform to GPU. */ + cuMemcpyHtoD(motion_transform_gpu, &motion_transform, motion_transform_size); + delete[] reinterpret_cast<uint8_t *>(&motion_transform); + + /* Disable instance transform if object uses motion transform already. */ + instance.flags = OPTIX_INSTANCE_FLAG_DISABLE_TRANSFORM; + + /* Get traversable handle to motion transform. */ + optixConvertPointerToTraversableHandle(context, + motion_transform_gpu, + OPTIX_TRAVERSABLE_TYPE_SRT_MOTION_TRANSFORM, + &instance.traversableHandle); + } + else { + instance.traversableHandle = handle; + + if (ob->get_geometry()->is_instanced()) { + /* Set transform matrix. */ + memcpy(instance.transform, &ob->get_tfm(), sizeof(instance.transform)); + } + else { + /* Disable instance transform if geometry already has it applied to vertex data. */ + instance.flags = OPTIX_INSTANCE_FLAG_DISABLE_TRANSFORM; + /* Non-instanced objects read ID from 'prim_object', so distinguish + * them from instanced objects with the low bit set. */ + instance.instanceId |= 1; + } + } + } + + /* Upload instance descriptions. */ + instances.resize(num_instances); + instances.copy_to_device(); + + /* Build top-level acceleration structure (TLAS) */ + OptixBuildInput build_input = {}; + build_input.type = OPTIX_BUILD_INPUT_TYPE_INSTANCES; + build_input.instanceArray.instances = instances.device_pointer; + build_input.instanceArray.numInstances = num_instances; + + if (!build_optix_bvh(bvh_optix, OPTIX_BUILD_OPERATION_BUILD, build_input, 0)) { + progress.set_error("Failed to build OptiX acceleration structure"); + } + tlas_handle = bvh_optix->traversable_handle; + } +} + +void OptiXDevice::release_optix_bvh(BVH *bvh) +{ + thread_scoped_lock lock(delayed_free_bvh_mutex); + /* Do delayed free of BVH memory, since geometry holding BVH might be deleted + * while GPU is still rendering. */ + BVHOptiX *const bvh_optix = static_cast<BVHOptiX *>(bvh); + + delayed_free_bvh_memory.emplace_back(std::move(bvh_optix->as_data)); + delayed_free_bvh_memory.emplace_back(std::move(bvh_optix->motion_transform_data)); + bvh_optix->traversable_handle = 0; +} + +void OptiXDevice::free_bvh_memory_delayed() +{ + thread_scoped_lock lock(delayed_free_bvh_mutex); + delayed_free_bvh_memory.free_memory(); +} + +void OptiXDevice::const_copy_to(const char *name, void *host, size_t size) +{ + /* Set constant memory for CUDA module. */ + CUDADevice::const_copy_to(name, host, size); + + if (strcmp(name, "__data") == 0) { + assert(size <= sizeof(KernelData)); + + /* Update traversable handle (since it is different for each device on multi devices). */ + KernelData *const data = (KernelData *)host; + *(OptixTraversableHandle *)&data->bvh.scene = tlas_handle; + + update_launch_params(offsetof(KernelParamsOptiX, data), host, size); + return; + } + + /* Update data storage pointers in launch parameters. */ +# define KERNEL_TEX(data_type, tex_name) \ + if (strcmp(name, #tex_name) == 0) { \ + update_launch_params(offsetof(KernelParamsOptiX, tex_name), host, size); \ + return; \ + } + KERNEL_TEX(IntegratorStateGPU, __integrator_state) +# include "kernel/kernel_textures.h" +# undef KERNEL_TEX +} + +void OptiXDevice::update_launch_params(size_t offset, void *data, size_t data_size) +{ + const CUDAContextScope scope(this); + + cuda_assert(cuMemcpyHtoD(launch_params.device_pointer + offset, data, data_size)); +} + +CCL_NAMESPACE_END + +#endif /* WITH_OPTIX */ diff --git a/intern/cycles/device/optix/device_impl.h b/intern/cycles/device/optix/device_impl.h new file mode 100644 index 00000000000..742ae0f1bab --- /dev/null +++ b/intern/cycles/device/optix/device_impl.h @@ -0,0 +1,186 @@ +/* + * Copyright 2019, NVIDIA Corporation. + * Copyright 2019, Blender Foundation. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#pragma once + +#ifdef WITH_OPTIX + +# include "device/cuda/device_impl.h" +# include "device/optix/queue.h" +# include "device/optix/util.h" +# include "kernel/kernel_types.h" + +CCL_NAMESPACE_BEGIN + +class BVHOptiX; +struct KernelParamsOptiX; + +/* List of OptiX program groups. */ +enum { + PG_RGEN_INTERSECT_CLOSEST, + PG_RGEN_INTERSECT_SHADOW, + PG_RGEN_INTERSECT_SUBSURFACE, + PG_RGEN_INTERSECT_VOLUME_STACK, + PG_RGEN_SHADE_SURFACE_RAYTRACE, + PG_MISS, + PG_HITD, /* Default hit group. */ + PG_HITS, /* __SHADOW_RECORD_ALL__ hit group. */ + PG_HITL, /* __BVH_LOCAL__ hit group (only used for triangles). */ + PG_HITD_MOTION, + PG_HITS_MOTION, + PG_CALL_SVM_AO, + PG_CALL_SVM_BEVEL, + PG_CALL_AO_PASS, + NUM_PROGRAM_GROUPS +}; + +static const int MISS_PROGRAM_GROUP_OFFSET = PG_MISS; +static const int NUM_MIS_PROGRAM_GROUPS = 1; +static const int HIT_PROGAM_GROUP_OFFSET = PG_HITD; +static const int NUM_HIT_PROGRAM_GROUPS = 5; +static const int CALLABLE_PROGRAM_GROUPS_BASE = PG_CALL_SVM_AO; +static const int NUM_CALLABLE_PROGRAM_GROUPS = 3; + +/* List of OptiX pipelines. */ +enum { PIP_SHADE_RAYTRACE, PIP_INTERSECT, NUM_PIPELINES }; + +/* A single shader binding table entry. */ +struct SbtRecord { + char header[OPTIX_SBT_RECORD_HEADER_SIZE]; +}; + +class OptiXDevice : public CUDADevice { + public: + OptixDeviceContext context = NULL; + + OptixModule optix_module = NULL; /* All necessary OptiX kernels are in one module. */ + OptixModule builtin_modules[2] = {}; + OptixPipeline pipelines[NUM_PIPELINES] = {}; + + bool motion_blur = false; + device_vector<SbtRecord> sbt_data; + device_only_memory<KernelParamsOptiX> launch_params; + OptixTraversableHandle tlas_handle = 0; + + vector<device_only_memory<char>> delayed_free_bvh_memory; + thread_mutex delayed_free_bvh_mutex; + + class Denoiser { + public: + explicit Denoiser(OptiXDevice *device); + ~Denoiser(); + + OptiXDevice *device; + OptiXDeviceQueue queue; + + OptixDenoiser optix_denoiser = nullptr; + + /* Configuration size, as provided to `optixDenoiserSetup`. + * If the `optixDenoiserSetup()` was never used on the current `optix_denoiser` the + * `is_configured` will be false. */ + bool is_configured = false; + int2 configured_size = make_int2(0, 0); + + /* OptiX denoiser state and scratch buffers, stored in a single memory buffer. + * The memory layout goes as following: [denoiser state][scratch buffer]. */ + device_only_memory<unsigned char> state; + size_t scratch_offset = 0; + size_t scratch_size = 0; + + bool use_pass_albedo = false; + bool use_pass_normal = false; + }; + Denoiser denoiser_; + + public: + OptiXDevice(const DeviceInfo &info, Stats &stats, Profiler &profiler); + ~OptiXDevice(); + + private: + BVHLayoutMask get_bvh_layout_mask() const override; + + string compile_kernel_get_common_cflags(const uint kernel_features) override; + + bool load_kernels(const uint kernel_features) override; + + bool build_optix_bvh(BVHOptiX *bvh, + OptixBuildOperation operation, + const OptixBuildInput &build_input, + uint16_t num_motion_steps); + + void build_bvh(BVH *bvh, Progress &progress, bool refit) override; + + void release_optix_bvh(BVH *bvh) override; + void free_bvh_memory_delayed(); + + void const_copy_to(const char *name, void *host, size_t size) override; + + void update_launch_params(size_t offset, void *data, size_t data_size); + + virtual unique_ptr<DeviceQueue> gpu_queue_create() override; + + /* -------------------------------------------------------------------- + * Denoising. + */ + + class DenoiseContext; + class DenoisePass; + + virtual bool denoise_buffer(const DeviceDenoiseTask &task) override; + virtual DeviceQueue *get_denoise_queue() override; + + /* Read guiding passes from the render buffers, preprocess them in a way which is expected by + * OptiX and store in the guiding passes memory within the given context. + * + * Pre=-processing of the guiding passes is to only hapopen once per context lifetime. DO not + * preprocess them for every pass which is being denoised. */ + bool denoise_filter_guiding_preprocess(DenoiseContext &context); + + /* Set fake albedo pixels in the albedo guiding pass storage. + * After this point only passes which do not need albedo for denoising can be processed. */ + bool denoise_filter_guiding_set_fake_albedo(DenoiseContext &context); + + void denoise_pass(DenoiseContext &context, PassType pass_type); + + /* Read input color pass from the render buffer into the memory which corresponds to the noisy + * input within the given context. Pixels are scaled to the number of samples, but are not + * preprocessed yet. */ + void denoise_color_read(DenoiseContext &context, const DenoisePass &pass); + + /* Run corresponding filter kernels, preparing data for the denoiser or copying data from the + * denoiser result to the render buffer. */ + bool denoise_filter_color_preprocess(DenoiseContext &context, const DenoisePass &pass); + bool denoise_filter_color_postprocess(DenoiseContext &context, const DenoisePass &pass); + + /* Make sure the OptiX denoiser is created and configured. */ + bool denoise_ensure(DenoiseContext &context); + + /* Create OptiX denoiser descriptor if needed. + * Will do nothing if the current OptiX descriptor is usable for the given parameters. + * If the OptiX denoiser descriptor did re-allocate here it is left unconfigured. */ + bool denoise_create_if_needed(DenoiseContext &context); + + /* Configure existing OptiX denoiser descriptor for the use for the given task. */ + bool denoise_configure_if_needed(DenoiseContext &context); + + /* Run configured denoiser. */ + bool denoise_run(DenoiseContext &context, const DenoisePass &pass); +}; + +CCL_NAMESPACE_END + +#endif /* WITH_OPTIX */ diff --git a/intern/cycles/device/optix/queue.cpp b/intern/cycles/device/optix/queue.cpp new file mode 100644 index 00000000000..458ed70baa8 --- /dev/null +++ b/intern/cycles/device/optix/queue.cpp @@ -0,0 +1,144 @@ +/* + * Copyright 2011-2013 Blender Foundation + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifdef WITH_OPTIX + +# include "device/optix/queue.h" +# include "device/optix/device_impl.h" + +# include "util/util_time.h" + +# undef __KERNEL_CPU__ +# define __KERNEL_OPTIX__ +# include "kernel/device/optix/globals.h" + +CCL_NAMESPACE_BEGIN + +/* CUDADeviceQueue */ + +OptiXDeviceQueue::OptiXDeviceQueue(OptiXDevice *device) : CUDADeviceQueue(device) +{ +} + +void OptiXDeviceQueue::init_execution() +{ + CUDADeviceQueue::init_execution(); +} + +static bool is_optix_specific_kernel(DeviceKernel kernel) +{ + return (kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE || + kernel == DEVICE_KERNEL_INTEGRATOR_INTERSECT_CLOSEST || + kernel == DEVICE_KERNEL_INTEGRATOR_INTERSECT_SHADOW || + kernel == DEVICE_KERNEL_INTEGRATOR_INTERSECT_SUBSURFACE || + kernel == DEVICE_KERNEL_INTEGRATOR_INTERSECT_VOLUME_STACK); +} + +bool OptiXDeviceQueue::enqueue(DeviceKernel kernel, const int work_size, void *args[]) +{ + if (!is_optix_specific_kernel(kernel)) { + return CUDADeviceQueue::enqueue(kernel, work_size, args); + } + + if (cuda_device_->have_error()) { + return false; + } + + debug_enqueue(kernel, work_size); + + const CUDAContextScope scope(cuda_device_); + + OptiXDevice *const optix_device = static_cast<OptiXDevice *>(cuda_device_); + + const device_ptr sbt_data_ptr = optix_device->sbt_data.device_pointer; + const device_ptr launch_params_ptr = optix_device->launch_params.device_pointer; + + cuda_device_assert( + cuda_device_, + cuMemcpyHtoDAsync(launch_params_ptr + offsetof(KernelParamsOptiX, path_index_array), + args[0], // &d_path_index + sizeof(device_ptr), + cuda_stream_)); + + if (kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE) { + cuda_device_assert( + cuda_device_, + cuMemcpyHtoDAsync(launch_params_ptr + offsetof(KernelParamsOptiX, render_buffer), + args[1], // &d_render_buffer + sizeof(device_ptr), + cuda_stream_)); + } + + cuda_device_assert(cuda_device_, cuStreamSynchronize(cuda_stream_)); + + OptixPipeline pipeline = nullptr; + OptixShaderBindingTable sbt_params = {}; + + switch (kernel) { + case DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE: + pipeline = optix_device->pipelines[PIP_SHADE_RAYTRACE]; + sbt_params.raygenRecord = sbt_data_ptr + PG_RGEN_SHADE_SURFACE_RAYTRACE * sizeof(SbtRecord); + break; + case DEVICE_KERNEL_INTEGRATOR_INTERSECT_CLOSEST: + pipeline = optix_device->pipelines[PIP_INTERSECT]; + sbt_params.raygenRecord = sbt_data_ptr + PG_RGEN_INTERSECT_CLOSEST * sizeof(SbtRecord); + break; + case DEVICE_KERNEL_INTEGRATOR_INTERSECT_SHADOW: + pipeline = optix_device->pipelines[PIP_INTERSECT]; + sbt_params.raygenRecord = sbt_data_ptr + PG_RGEN_INTERSECT_SHADOW * sizeof(SbtRecord); + break; + case DEVICE_KERNEL_INTEGRATOR_INTERSECT_SUBSURFACE: + pipeline = optix_device->pipelines[PIP_INTERSECT]; + sbt_params.raygenRecord = sbt_data_ptr + PG_RGEN_INTERSECT_SUBSURFACE * sizeof(SbtRecord); + break; + case DEVICE_KERNEL_INTEGRATOR_INTERSECT_VOLUME_STACK: + pipeline = optix_device->pipelines[PIP_INTERSECT]; + sbt_params.raygenRecord = sbt_data_ptr + PG_RGEN_INTERSECT_VOLUME_STACK * sizeof(SbtRecord); + break; + + default: + LOG(ERROR) << "Invalid kernel " << device_kernel_as_string(kernel) + << " is attempted to be enqueued."; + return false; + } + + sbt_params.missRecordBase = sbt_data_ptr + MISS_PROGRAM_GROUP_OFFSET * sizeof(SbtRecord); + sbt_params.missRecordStrideInBytes = sizeof(SbtRecord); + sbt_params.missRecordCount = NUM_MIS_PROGRAM_GROUPS; + sbt_params.hitgroupRecordBase = sbt_data_ptr + HIT_PROGAM_GROUP_OFFSET * sizeof(SbtRecord); + sbt_params.hitgroupRecordStrideInBytes = sizeof(SbtRecord); + sbt_params.hitgroupRecordCount = NUM_HIT_PROGRAM_GROUPS; + sbt_params.callablesRecordBase = sbt_data_ptr + CALLABLE_PROGRAM_GROUPS_BASE * sizeof(SbtRecord); + sbt_params.callablesRecordCount = NUM_CALLABLE_PROGRAM_GROUPS; + sbt_params.callablesRecordStrideInBytes = sizeof(SbtRecord); + + /* Launch the ray generation program. */ + optix_device_assert(optix_device, + optixLaunch(pipeline, + cuda_stream_, + launch_params_ptr, + optix_device->launch_params.data_elements, + &sbt_params, + work_size, + 1, + 1)); + + return !(optix_device->have_error()); +} + +CCL_NAMESPACE_END + +#endif /* WITH_OPTIX */ diff --git a/intern/cycles/device/optix/queue.h b/intern/cycles/device/optix/queue.h new file mode 100644 index 00000000000..0de422ccc71 --- /dev/null +++ b/intern/cycles/device/optix/queue.h @@ -0,0 +1,39 @@ +/* + * Copyright 2011-2021 Blender Foundation + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#pragma once + +#ifdef WITH_OPTIX + +# include "device/cuda/queue.h" + +CCL_NAMESPACE_BEGIN + +class OptiXDevice; + +/* Base class for CUDA queues. */ +class OptiXDeviceQueue : public CUDADeviceQueue { + public: + OptiXDeviceQueue(OptiXDevice *device); + + virtual void init_execution() override; + + virtual bool enqueue(DeviceKernel kernel, const int work_size, void *args[]) override; +}; + +CCL_NAMESPACE_END + +#endif /* WITH_OPTIX */ diff --git a/intern/cycles/device/optix/util.h b/intern/cycles/device/optix/util.h new file mode 100644 index 00000000000..34ae5bb5609 --- /dev/null +++ b/intern/cycles/device/optix/util.h @@ -0,0 +1,45 @@ +/* + * Copyright 2011-2021 Blender Foundation + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#pragma once + +#ifdef WITH_OPTIX + +# include "device/cuda/util.h" + +# ifdef WITH_CUDA_DYNLOAD +# include <cuew.h> +// Do not use CUDA SDK headers when using CUEW +# define OPTIX_DONT_INCLUDE_CUDA +# endif + +# include <optix_stubs.h> + +/* Utility for checking return values of OptiX function calls. */ +# define optix_device_assert(optix_device, stmt) \ + { \ + OptixResult result = stmt; \ + if (result != OPTIX_SUCCESS) { \ + const char *name = optixGetErrorName(result); \ + optix_device->set_error( \ + string_printf("%s in %s (%s:%d)", name, #stmt, __FILE__, __LINE__)); \ + } \ + } \ + (void)0 + +# define optix_assert(stmt) optix_device_assert(this, stmt) + +#endif /* WITH_OPTIX */ |