Differential D12604 Diff 42257 intern/cycles/device/optix/device_impl.cpp

Changeset View

Standalone View

intern/cycles/device/optix/device_impl.cpp

Show First 20 Lines • Show All 309 Lines • ▼ Show 20 Lines	bool OptiXDevice::load_kernels(const uint kernel_features)
group_descs[PG_HITD].kind = OPTIX_PROGRAM_GROUP_KIND_HITGROUP;		group_descs[PG_HITD].kind = OPTIX_PROGRAM_GROUP_KIND_HITGROUP;
group_descs[PG_HITD].hitgroup.moduleCH = optix_module;		group_descs[PG_HITD].hitgroup.moduleCH = optix_module;
group_descs[PG_HITD].hitgroup.entryFunctionNameCH = "__closesthit__kernel_optix_hit";		group_descs[PG_HITD].hitgroup.entryFunctionNameCH = "__closesthit__kernel_optix_hit";
group_descs[PG_HITD].hitgroup.moduleAH = optix_module;		group_descs[PG_HITD].hitgroup.moduleAH = optix_module;
group_descs[PG_HITD].hitgroup.entryFunctionNameAH = "__anyhit__kernel_optix_visibility_test";		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].kind = OPTIX_PROGRAM_GROUP_KIND_HITGROUP;
group_descs[PG_HITS].hitgroup.moduleAH = optix_module;		group_descs[PG_HITS].hitgroup.moduleAH = optix_module;
group_descs[PG_HITS].hitgroup.entryFunctionNameAH = "__anyhit__kernel_optix_shadow_all_hit";		group_descs[PG_HITS].hitgroup.entryFunctionNameAH = "__anyhit__kernel_optix_shadow_all_hit";
		group_descs[PG_HITV].kind = OPTIX_PROGRAM_GROUP_KIND_HITGROUP;
		group_descs[PG_HITV].hitgroup.moduleCH = optix_module;
		group_descs[PG_HITV].hitgroup.entryFunctionNameCH = "__closesthit__kernel_optix_hit";
		group_descs[PG_HITV].hitgroup.moduleAH = optix_module;
		group_descs[PG_HITV].hitgroup.entryFunctionNameAH = "__anyhit__kernel_optix_volume_test";

if (kernel_features & KERNEL_FEATURE_HAIR) {		if (kernel_features & KERNEL_FEATURE_HAIR) {
if (kernel_features & KERNEL_FEATURE_HAIR_THICK) {		if (kernel_features & KERNEL_FEATURE_HAIR_THICK) {
/* Built-in thick curve intersection. */		/* Built-in thick curve intersection. */
OptixBuiltinISOptions builtin_options = {};		OptixBuiltinISOptions builtin_options = {};
builtin_options.builtinISModuleType = OPTIX_PRIMITIVE_TYPE_ROUND_CUBIC_BSPLINE;		builtin_options.builtinISModuleType = OPTIX_PRIMITIVE_TYPE_ROUND_CUBIC_BSPLINE;
builtin_options.usesMotionBlur = false;		builtin_options.usesMotionBlur = false;

▲ Show 20 Lines • Show All 66 Lines • ▼ Show 20 Lines	bool OptiXDevice::load_kernels(const uint kernel_features)
sbt_data.copy_to_device(); /* Upload SBT to device. */		sbt_data.copy_to_device(); /* Upload SBT to device. */

/* Calculate maximum trace continuation stack size. */		/* Calculate maximum trace continuation stack size. */
unsigned int trace_css = stack_size[PG_HITD].cssCH;		unsigned int trace_css = stack_size[PG_HITD].cssCH;
/* This is based on the maximum of closest-hit and any-hit/intersection programs. */		/* 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_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_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_HITL].cssIS + stack_size[PG_HITL].cssAH);
		trace_css = std::max(trace_css, stack_size[PG_HITV].cssIS + stack_size[PG_HITV].cssAH);
trace_css = std::max(trace_css,		trace_css = std::max(trace_css,
stack_size[PG_HITD_MOTION].cssIS + stack_size[PG_HITD_MOTION].cssAH);		stack_size[PG_HITD_MOTION].cssIS + stack_size[PG_HITD_MOTION].cssAH);
trace_css = std::max(trace_css,		trace_css = std::max(trace_css,
stack_size[PG_HITS_MOTION].cssIS + stack_size[PG_HITS_MOTION].cssAH);		stack_size[PG_HITS_MOTION].cssIS + stack_size[PG_HITS_MOTION].cssAH);

OptixPipelineLinkOptions link_options = {};		OptixPipelineLinkOptions link_options = {};
link_options.maxTraceDepth = 1;		link_options.maxTraceDepth = 1;

if (DebugFlags().optix.use_debug) {		if (DebugFlags().optix.use_debug) {
link_options.debugLevel = OPTIX_COMPILE_DEBUG_LEVEL_FULL;		link_options.debugLevel = OPTIX_COMPILE_DEBUG_LEVEL_FULL;
}		}
else {		else {
link_options.debugLevel = OPTIX_COMPILE_DEBUG_LEVEL_LINEINFO;		link_options.debugLevel = OPTIX_COMPILE_DEBUG_LEVEL_LINEINFO;
}		}

if (kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) {		if (kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) {
/* Create shader raytracing pipeline. */		/* Create shader raytracing pipeline. */
vector<OptixProgramGroup> pipeline_groups;		vector<OptixProgramGroup> pipeline_groups;
pipeline_groups.reserve(NUM_PROGRAM_GROUPS);		pipeline_groups.reserve(NUM_PROGRAM_GROUPS);
pipeline_groups.push_back(groups[PG_RGEN_SHADE_SURFACE_RAYTRACE]);		pipeline_groups.push_back(groups[PG_RGEN_SHADE_SURFACE_RAYTRACE]);
pipeline_groups.push_back(groups[PG_MISS]);		pipeline_groups.push_back(groups[PG_MISS]);
pipeline_groups.push_back(groups[PG_HITD]);		pipeline_groups.push_back(groups[PG_HITD]);
pipeline_groups.push_back(groups[PG_HITS]);		pipeline_groups.push_back(groups[PG_HITS]);
pipeline_groups.push_back(groups[PG_HITL]);		pipeline_groups.push_back(groups[PG_HITL]);
		pipeline_groups.push_back(groups[PG_HITV]);
if (motion_blur) {		if (motion_blur) {
pipeline_groups.push_back(groups[PG_HITD_MOTION]);		pipeline_groups.push_back(groups[PG_HITD_MOTION]);
pipeline_groups.push_back(groups[PG_HITS_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_AO]);
pipeline_groups.push_back(groups[PG_CALL_SVM_BEVEL]);		pipeline_groups.push_back(groups[PG_CALL_SVM_BEVEL]);

optix_assert(optixPipelineCreate(context,		optix_assert(optixPipelineCreate(context,
Show All 22 Lines	if (kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) {
pipeline_groups.push_back(groups[PG_RGEN_INTERSECT_CLOSEST]);		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_SHADOW]);
pipeline_groups.push_back(groups[PG_RGEN_INTERSECT_SUBSURFACE]);		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_RGEN_INTERSECT_VOLUME_STACK]);
pipeline_groups.push_back(groups[PG_MISS]);		pipeline_groups.push_back(groups[PG_MISS]);
pipeline_groups.push_back(groups[PG_HITD]);		pipeline_groups.push_back(groups[PG_HITD]);
pipeline_groups.push_back(groups[PG_HITS]);		pipeline_groups.push_back(groups[PG_HITS]);
pipeline_groups.push_back(groups[PG_HITL]);		pipeline_groups.push_back(groups[PG_HITL]);
		pipeline_groups.push_back(groups[PG_HITV]);
if (motion_blur) {		if (motion_blur) {
pipeline_groups.push_back(groups[PG_HITD_MOTION]);		pipeline_groups.push_back(groups[PG_HITD_MOTION]);
pipeline_groups.push_back(groups[PG_HITS_MOTION]);		pipeline_groups.push_back(groups[PG_HITS_MOTION]);
}		}

optix_assert(optixPipelineCreate(context,		optix_assert(optixPipelineCreate(context,
&pipeline_options,		&pipeline_options,
&link_options,		&link_options,
▲ Show 20 Lines • Show All 915 Lines • ▼ Show 20 Lines	for (Object *ob : bvh->objects) {
/* Clear transform to identity matrix. */		/* Clear transform to identity matrix. */
instance.transform[0] = 1.0f;		instance.transform[0] = 1.0f;
instance.transform[5] = 1.0f;		instance.transform[5] = 1.0f;
instance.transform[10] = 1.0f;		instance.transform[10] = 1.0f;

/* Set user instance ID to object index (but leave low bit blank). */		/* Set user instance ID to object index (but leave low bit blank). */
instance.instanceId = ob->get_device_index() << 1;		instance.instanceId = ob->get_device_index() << 1;

/* Have to have at least one bit in the mask, or else instance would always be culled. */		/* Add some of the object visibility bits to the mask.
instance.visibilityMask = 1;		* __prim_visibility contains the combined visibility bits of all instances, so is not
		* reliable if they differ between instances. But the OptiX visibility mask can only contain
if (ob->get_geometry()->has_volume) {		* 8 bits, so have to trade-off here and select just a few important ones.
/* Volumes have a special bit set in the visibility mask so a trace can mask only volumes.
*/		*/
instance.visibilityMask \|= 2;		instance.visibilityMask = ob->visibility_for_tracing() & 0xFF;
}

if (ob->get_geometry()->geometry_type == Geometry::HAIR) {		/* Have to have at least one bit in the mask, or else instance would always be culled. */
/* Same applies to curves (so they can be skipped in local trace calls). */		if (0 == instance.visibilityMask) {
instance.visibilityMask \|= 4;		instance.visibilityMask = 0xFF;
		}

if (motion_blur && ob->get_geometry()->has_motion_blur() &&		if (ob->get_geometry()->geometry_type == Geometry::HAIR &&
static_cast<const Hair *>(ob->get_geometry())->curve_shape == CURVE_THICK) {		static_cast<const Hair *>(ob->get_geometry())->curve_shape == CURVE_THICK) {
		if (motion_blur && ob->get_geometry()->has_motion_blur()) {
/* Select between motion blur and non-motion blur built-in intersection module. */		/* Select between motion blur and non-motion blur built-in intersection module. */
instance.sbtOffset = PG_HITD_MOTION - PG_HITD;		instance.sbtOffset = PG_HITD_MOTION - PG_HITD;
}		}
}		}
		else {
		/* Can disable __anyhit__kernel_optix_visibility_test by default (except for thick curves,
		* since it needs to filter out endcaps there).
		* It is enabled where necessary (visibility mask exceeds 8 bits or the other any-hit
		* programs like __anyhit__kernel_optix_shadow_all_hit) via OPTIX_RAY_FLAG_ENFORCE_ANYHIT.
		*/
		instance.flags = OPTIX_INSTANCE_FLAG_DISABLE_ANYHIT;
		}

/* Insert motion traversable if object has motion. */		/* Insert motion traversable if object has motion. */
if (motion_blur && ob->use_motion()) {		if (motion_blur && ob->use_motion()) {
size_t motion_keys = max(ob->get_motion().size(), 2) - 2;		size_t motion_keys = max(ob->get_motion().size(), 2) - 2;
size_t motion_transform_size = sizeof(OptixSRTMotionTransform) +		size_t motion_transform_size = sizeof(OptixSRTMotionTransform) +
motion_keys * sizeof(OptixSRTData);		motion_keys * sizeof(OptixSRTData);

const CUDAContextScope scope(this);		const CUDAContextScope scope(this);
▲ Show 20 Lines • Show All 49 Lines • ▼ Show 20 Lines	for (Object *ob : bvh->objects) {
srt_data[i].tz = decomp[i].y.z;		srt_data[i].tz = decomp[i].y.z;
}		}

/* Upload motion transform to GPU. */		/* Upload motion transform to GPU. */
cuMemcpyHtoD(motion_transform_gpu, &motion_transform, motion_transform_size);		cuMemcpyHtoD(motion_transform_gpu, &motion_transform, motion_transform_size);
delete[] reinterpret_cast<uint8_t *>(&motion_transform);		delete[] reinterpret_cast<uint8_t *>(&motion_transform);

/* Disable instance transform if object uses motion transform already. */		/* Disable instance transform if object uses motion transform already. */
instance.flags = OPTIX_INSTANCE_FLAG_DISABLE_TRANSFORM;		instance.flags \|= OPTIX_INSTANCE_FLAG_DISABLE_TRANSFORM;

/* Get traversable handle to motion transform. */		/* Get traversable handle to motion transform. */
optixConvertPointerToTraversableHandle(context,		optixConvertPointerToTraversableHandle(context,
motion_transform_gpu,		motion_transform_gpu,
OPTIX_TRAVERSABLE_TYPE_SRT_MOTION_TRANSFORM,		OPTIX_TRAVERSABLE_TYPE_SRT_MOTION_TRANSFORM,
&instance.traversableHandle);		&instance.traversableHandle);
}		}
else {		else {
instance.traversableHandle = handle;		instance.traversableHandle = handle;

if (ob->get_geometry()->is_instanced()) {		if (ob->get_geometry()->is_instanced()) {
/* Set transform matrix. */		/* Set transform matrix. */
memcpy(instance.transform, &ob->get_tfm(), sizeof(instance.transform));		memcpy(instance.transform, &ob->get_tfm(), sizeof(instance.transform));
}		}
else {		else {
/* Disable instance transform if geometry already has it applied to vertex data. */		/* Disable instance transform if geometry already has it applied to vertex data. */
instance.flags = OPTIX_INSTANCE_FLAG_DISABLE_TRANSFORM;		instance.flags \|= OPTIX_INSTANCE_FLAG_DISABLE_TRANSFORM;
/* Non-instanced objects read ID from 'prim_object', so distinguish		/* Non-instanced objects read ID from 'prim_object', so distinguish
* them from instanced objects with the low bit set. */		* them from instanced objects with the low bit set. */
instance.instanceId \|= 1;		instance.instanceId \|= 1;
}		}
}		}
}		}

/* Upload instance descriptions. */		/* Upload instance descriptions. */
▲ Show 20 Lines • Show All 71 Lines • Show Last 20 Lines