Differential D10535 Diff 34709 intern/cycles/device/device_optix.cpp

Changeset View

Standalone View

intern/cycles/device/device_optix.cpp

Show First 20 Lines • Show All 191 Lines • ▼ Show 20 Lines	# endif
OptixDenoiser denoiser = NULL;		OptixDenoiser denoiser = NULL;
device_only_memory<unsigned char> denoiser_state;		device_only_memory<unsigned char> denoiser_state;
int denoiser_input_passes = 0;		int denoiser_input_passes = 0;

public:		public:
OptiXDevice(DeviceInfo &info_, Stats &stats_, Profiler &profiler_, bool background_)		OptiXDevice(DeviceInfo &info_, Stats &stats_, Profiler &profiler_, bool background_)
: CUDADevice(info_, stats_, profiler_, background_),		: CUDADevice(info_, stats_, profiler_, background_),
sbt_data(this, "__sbt", MEM_READ_ONLY),		sbt_data(this, "__sbt", MEM_READ_ONLY),
launch_params(this, "__params"),		launch_params(this, "__params", false),
denoiser_state(this, "__denoiser_state")		denoiser_state(this, "__denoiser_state", true)
{		{
// Store number of CUDA streams in device info		// Store number of CUDA streams in device info
info.cpu_threads = DebugFlags().optix.cuda_streams;		info.cpu_threads = DebugFlags().optix.cuda_streams;

// Make the CUDA context current		// Make the CUDA context current
if (!cuContext) {		if (!cuContext) {
return; // Do not initialize if CUDA context creation failed already		return; // Do not initialize if CUDA context creation failed already
}		}
▲ Show 20 Lines • Show All 663 Lines • ▼ Show 20 Lines	if (task.denoising.type == DENOISER_OPTIX) {
(task.pass_denoising_data + DENOISING_PASS_ALBEDO) * (int)sizeof(float),		(task.pass_denoising_data + DENOISING_PASS_ALBEDO) * (int)sizeof(float),
(task.pass_denoising_data + DENOISING_PASS_NORMAL) * (int)sizeof(float)};		(task.pass_denoising_data + DENOISING_PASS_NORMAL) * (int)sizeof(float)};

// Start with the current tile pointer offset		// Start with the current tile pointer offset
int input_stride = pixel_stride;		int input_stride = pixel_stride;
device_ptr input_ptr = rtile.buffer + pixel_offset;		device_ptr input_ptr = rtile.buffer + pixel_offset;

// Copy tile data into a common buffer if necessary		// Copy tile data into a common buffer if necessary
device_only_memory<float> input(this, "denoiser input");		device_only_memory<float> input(this, "denoiser input", true);
device_vector<TileInfo> tile_info_mem(this, "denoiser tile info", MEM_READ_WRITE);		device_vector<TileInfo> tile_info_mem(this, "denoiser tile info", MEM_READ_ONLY);

bool contiguous_memory = true;		bool contiguous_memory = true;
for (int i = 0; i < RenderTileNeighbors::SIZE; i++) {		for (int i = 0; i < RenderTileNeighbors::SIZE; i++) {
if (neighbors.tiles[i].buffer && neighbors.tiles[i].buffer != rtile.buffer) {		if (neighbors.tiles[i].buffer && neighbors.tiles[i].buffer != rtile.buffer) {
contiguous_memory = false;		contiguous_memory = false;
}		}
}		}

Show All 28 Lines	if (task.denoising.type == DENOISER_OPTIX) {
tile_info_mem.copy_to_device();		tile_info_mem.copy_to_device();

void *args[] = {		void *args[] = {
&input.device_pointer, &tile_info_mem.device_pointer, &rect.x, &task.pass_stride};		&input.device_pointer, &tile_info_mem.device_pointer, &rect.x, &task.pass_stride};
launch_filter_kernel("kernel_cuda_filter_copy_input", rect_size.x, rect_size.y, args);		launch_filter_kernel("kernel_cuda_filter_copy_input", rect_size.x, rect_size.y, args);
}		}

# if OPTIX_DENOISER_NO_PIXEL_STRIDE		# if OPTIX_DENOISER_NO_PIXEL_STRIDE
device_only_memory<float> input_rgb(this, "denoiser input rgb");		device_only_memory<float> input_rgb(this, "denoiser input rgb", true);
input_rgb.alloc_to_device(rect_size.x * rect_size.y * 3 * task.denoising.input_passes);		input_rgb.alloc_to_device(rect_size.x * rect_size.y * 3 * task.denoising.input_passes);

void *input_args[] = {&input_rgb.device_pointer,		void *input_args[] = {&input_rgb.device_pointer,
&input_ptr,		&input_ptr,
&rect_size.x,		&rect_size.x,
&rect_size.y,		&rect_size.y,
&input_stride,		&input_stride,
&task.pass_stride,		&task.pass_stride,
▲ Show 20 Lines • Show All 205 Lines • ▼ Show 20 Lines	# endif
}		}
}		}

bool build_optix_bvh(BVHOptiX *bvh,		bool build_optix_bvh(BVHOptiX *bvh,
OptixBuildOperation operation,		OptixBuildOperation operation,
const OptixBuildInput &build_input,		const OptixBuildInput &build_input,
uint16_t num_motion_steps)		uint16_t num_motion_steps)
{		{
		/* Allocate and build acceleration structures only one at a time, to prevent parallel builds
		* from running out of memory (since both original and compacted acceleration structure memory
		* may be allocated at the same time for the duration of this function). The builds would
		* otherwise happen on the same CUDA stream anyway. */
		static thread_mutex mutex;
		thread_scoped_lock lock(mutex);

const CUDAContextScope scope(cuContext);		const CUDAContextScope scope(cuContext);

// Compute memory usage		// Compute memory usage
OptixAccelBufferSizes sizes = {};		OptixAccelBufferSizes sizes = {};
OptixAccelBuildOptions options;		OptixAccelBuildOptions options;
options.operation = operation;		options.operation = operation;
if (background) {		if (background) {
// Prefer best performance and lowest memory consumption in background		// Prefer best performance and lowest memory consumption in background
options.buildFlags = OPTIX_BUILD_FLAG_PREFER_FAST_TRACE \| OPTIX_BUILD_FLAG_ALLOW_COMPACTION;		options.buildFlags = OPTIX_BUILD_FLAG_PREFER_FAST_TRACE \| OPTIX_BUILD_FLAG_ALLOW_COMPACTION;
}		}
else {		else {
// Prefer fast updates in viewport		// Prefer fast updates in viewport
options.buildFlags = OPTIX_BUILD_FLAG_PREFER_FAST_BUILD \| OPTIX_BUILD_FLAG_ALLOW_UPDATE;		options.buildFlags = OPTIX_BUILD_FLAG_PREFER_FAST_BUILD \| OPTIX_BUILD_FLAG_ALLOW_UPDATE;
}		}

options.motionOptions.numKeys = num_motion_steps;		options.motionOptions.numKeys = num_motion_steps;
options.motionOptions.flags = OPTIX_MOTION_FLAG_START_VANISH \| OPTIX_MOTION_FLAG_END_VANISH;		options.motionOptions.flags = OPTIX_MOTION_FLAG_START_VANISH \| OPTIX_MOTION_FLAG_END_VANISH;
options.motionOptions.timeBegin = 0.0f;		options.motionOptions.timeBegin = 0.0f;
options.motionOptions.timeEnd = 1.0f;		options.motionOptions.timeEnd = 1.0f;

check_result_optix_ret(		check_result_optix_ret(
optixAccelComputeMemoryUsage(context, &options, &build_input, 1, &sizes));		optixAccelComputeMemoryUsage(context, &options, &build_input, 1, &sizes));

// Allocate required output buffers		// Allocate required output buffers
device_only_memory<char> temp_mem(this, "optix temp as build mem");		device_only_memory<char> temp_mem(this, "optix temp as build mem", true);
temp_mem.alloc_to_device(align_up(sizes.tempSizeInBytes, 8) + 8);		temp_mem.alloc_to_device(align_up(sizes.tempSizeInBytes, 8) + 8);
if (!temp_mem.device_pointer)		if (!temp_mem.device_pointer)
return false; // Make sure temporary memory allocation succeeded		return false; // Make sure temporary memory allocation succeeded

		// Acceleration structure memory has to be allocated on the device (not allowed to be on host)
device_only_memory<char> &out_data = bvh->as_data;		device_only_memory<char> &out_data = bvh->as_data;
if (operation == OPTIX_BUILD_OPERATION_BUILD) {		if (operation == OPTIX_BUILD_OPERATION_BUILD) {
assert(out_data.device == this);		assert(out_data.device == this);
out_data.alloc_to_device(sizes.outputSizeInBytes);		out_data.alloc_to_device(sizes.outputSizeInBytes);
if (!out_data.device_pointer)		if (!out_data.device_pointer)
return false;		return false;
}		}
else {		else {
Show All 31 Lines	if (background) {
check_result_cuda_ret(		check_result_cuda_ret(
cuMemcpyDtoH(&compacted_size, compacted_size_prop.result, sizeof(compacted_size)));		cuMemcpyDtoH(&compacted_size, compacted_size_prop.result, sizeof(compacted_size)));

// Temporary memory is no longer needed, so free it now to make space		// Temporary memory is no longer needed, so free it now to make space
temp_mem.free();		temp_mem.free();

// There is no point compacting if the size does not change		// There is no point compacting if the size does not change
if (compacted_size < sizes.outputSizeInBytes) {		if (compacted_size < sizes.outputSizeInBytes) {
device_only_memory<char> compacted_data(this, "optix compacted as");		device_only_memory<char> compacted_data(this, "optix compacted as", false);
compacted_data.alloc_to_device(compacted_size);		compacted_data.alloc_to_device(compacted_size);
if (!compacted_data.device_pointer)		if (!compacted_data.device_pointer)
// Do not compact if memory allocation for compacted acceleration structure fails		// Do not compact if memory allocation for compacted acceleration structure fails
// Can just use the uncompacted one then, so succeed here regardless		// Can just use the uncompacted one then, so succeed here regardless
return true;		return true;

check_result_optix_ret(optixAccelCompact(context,		check_result_optix_ret(optixAccelCompact(context,
NULL,		NULL,
out_handle,		out_handle,
compacted_data.device_pointer,		compacted_data.device_pointer,
compacted_size,		compacted_size,
&out_handle));		&out_handle));
bvh->traversable_handle = static_cast<uint64_t>(out_handle);		bvh->traversable_handle = static_cast<uint64_t>(out_handle);

// Wait for compaction to finish		// Wait for compaction to finish
check_result_cuda_ret(cuStreamSynchronize(NULL));		check_result_cuda_ret(cuStreamSynchronize(NULL));

std::swap(out_data.device_size, compacted_data.device_size);		std::swap(out_data.device_size, compacted_data.device_size);
std::swap(out_data.device_pointer, compacted_data.device_pointer);		std::swap(out_data.device_pointer, compacted_data.device_pointer);
		// Original acceleration structure memory is freed when 'compacted_data' goes out of scope
}		}
}		}

return true;		return true;
}		}

void build_bvh(BVH *bvh, Progress &progress, bool refit) override		void build_bvh(BVH *bvh, Progress &progress, bool refit) override
{		{
▲ Show 20 Lines • Show All 621 Lines • Show Last 20 Lines