Stateful GPU dispatch engine for VisionSequence execution. More...

#include <VisionGpuDispatch.hpp>

Collaboration diagram for MayaFlux::Yantra::VisionGpuExecutor:

Public Member Functions
bool	is_suspended () const
	True when a deferred step has work outstanding and the next run will poll rather than start a fresh sequence.

VisionGpuExecutor &	operator= (const VisionGpuExecutor &)=delete

VisionGpuExecutor &	operator= (VisionGpuExecutor &&)=delete

void	reset ()
	Abandon outstanding work and clear the resume point.

Kinesis::Vision::VisionResult	run (const Kinesis::Vision::VisionSequence &sequence, const std::shared_ptr< Core::VKImage > &image, uint32_t w, uint32_t h)
	Execute a VisionSequence on the GPU using this instance's own lazily-constructed context set.

Kinesis::Vision::VisionResult	run (VisionGpuContexts &contexts, const Kinesis::Vision::VisionSequence &sequence, const std::shared_ptr< Core::VKImage > &image, uint32_t w, uint32_t h)
	Execute a VisionSequence on the GPU through an explicit context set.

std::shared_ptr< Core::VKImage >	select_label (const std::shared_ptr< Core::VKImage > &source, uint32_t target_label, uint32_t w, uint32_t h)
	select_label() using this instance's own lazily-constructed context set.

std::shared_ptr< Core::VKImage >	select_label (VisionGpuContexts &contexts, const std::shared_ptr< Core::VKImage > &source, uint32_t target_label, uint32_t w, uint32_t h)
	Isolate one ConnectedComponents label's silhouette on GPU, through the explicit context set that produced the label buffer.

	VisionGpuExecutor ()=default

	VisionGpuExecutor (const VisionGpuExecutor &)=delete

	VisionGpuExecutor (VisionGpuExecutor &&)=delete

	~VisionGpuExecutor ()=default

Static Public Member Functions
static GpuComputeConfig	config (Kinesis::Vision::VisionOp op, const Kinesis::Vision::VisionParams &params)
	GpuComputeConfig for a given VisionOp and its parameters.

static std::vector< Kinesis::Vision::Contour >	read_exported_contours (const Kinesis::Vision::VisionResult &result)
	Host contours decoded from result.buffers.contours, or empty.

static std::vector< Kinesis::Vision::Keypoint >	read_exported_keypoints (const Kinesis::Vision::VisionResult &result)
	Host keypoints decoded from result.buffers.keypoints, or empty.

static std::vector< Kinesis::Vision::TrackResult >	read_exported_tracks (const Kinesis::Vision::VisionResult &result)
	Host tracks decoded from result.buffers.tracks, or empty.

Private Member Functions
const std::vector< float > &	gaussian_kernel_2d (uint32_t radius, float sigma)
	2D Gaussian kernel for convolution, cached by (radius, sigma bit pattern).

Kinesis::Vision::GpuVisionPass::Completed	op_canny (VisionGpuContexts &contexts, const Kinesis::Vision::VisionParams &params, const Kinesis::Vision::CannyParams &p)
	Detect edges in the current image with the Canny pipeline.

Kinesis::Vision::GpuVisionPass::Completed	op_confine (VisionGpuContexts &contexts, const Kinesis::Vision::ConfineParams &p)
	Arbitrary rectangular sub-region of the current image, optionally resized back to the frame's own size before this step.

void	op_connected_components (VisionGpuContexts &contexts, const Kinesis::Vision::ConnectedComponentsParams &p)
	Label components in the current image on the GPU.

bool	op_dense_flow (VisionGpuContexts &contexts, const Kinesis::Vision::VisionStep &step)
	Dense optical flow from the previous frame to the current one.

void	op_extract_peaks (VisionGpuContexts &contexts, const Kinesis::Vision::ExtractPeaksParams &p)
	Extract and rank keypoints from the current response image.

bool	op_find_contours (VisionGpuContexts &contexts, const Kinesis::Vision::FindContoursParams &p)
	Trace contours from the immediately preceding component labels.

Kinesis::Vision::GpuVisionPass::Completed	op_harris_response (VisionGpuContexts &contexts, const Kinesis::Vision::HarrisParams &p)
	Compute the Harris response of the current image.

Kinesis::Vision::GpuVisionPass::Completed	op_open_close (VisionGpuContexts &contexts, Kinesis::Vision::VisionOp op, const Kinesis::Vision::MorphParams &p)
	Apply morphological opening or closing to the current image.

Kinesis::Vision::GpuVisionPass::Completed	op_threshold_otsu (VisionGpuContexts &contexts, const Kinesis::Vision::OtsuParams &p)
	Apply an Otsu threshold to the current image.

bool	op_track_keypoints (VisionGpuContexts &contexts, const Kinesis::Vision::VisionStep &step)
	Track the previous frame's keypoints into the current frame.

Static Private Member Functions
static void	after_step (VisionGpuContexts &contexts, size_t index)
	Run any work a finished step owes the flow context.

static void	build_flow_pyramid (VisionGpuContexts &contexts, uint32_t requested_levels, uint32_t src_channel)
	Build the current frame's pyramid atlas from the working image.

static void	op_snapshot (VisionGpuContexts &contexts)
	Copy the working image into this snapshot's own image and append it to VisionResult::snapshots.

Private Attributes
std::unique_ptr< VisionGpuContexts >	m_contexts

Detailed Description

Stateful GPU dispatch engine for VisionSequence execution.

Owns a lazily-constructed VisionGpuContexts (m_contexts) plus any op-specific persistent GPU state that doesn't belong on a shared context (e.g. ConnectedComponents' ping-pong output image). Mirrors VisionExecutor's ownership model on the CPU side: construct one instance per independent pipeline, hold it, call run() every frame.

Immovable: TextureExecutionContext owns GPU resource handles with no copy or move constructor, so VisionGpuContexts and therefore VisionGpuExecutor cannot be copied or moved either. Construct once, hold by reference, pointer, or shared_ptr, reuse indefinitely.

Not safe to call run() concurrently from multiple threads on the same instance. Independent pipelines running concurrently should each own a separate VisionGpuExecutor instance.

Definition at line 261 of file VisionGpuDispatch.hpp.

The documentation for this class was generated from the following files:

Public Member Functions

Static Public Member Functions

Private Member Functions

Static Private Member Functions

Private Attributes

Detailed Description