Optical Flow
Optical Flow Queues
VK_NV_optical_flow
adds an optical flow queue type bit
VK_QUEUE_OPTICAL_FLOW_BIT_NV
to VkQueueFlagBits.
Optical flow operations are supported by queues with an advertised queue
capability of VK_QUEUE_OPTICAL_FLOW_BIT_NV
.
As in the case of other queue types, an application must use
vkGetPhysicalDeviceQueueFamilyProperties to query whether the physical
device has support for the Optical Flow Queue.
When the implementation reports the VK_QUEUE_OPTICAL_FLOW_BIT_NV
bit
for a queue family, it advertises general support for Vulkan queue
operations described in Devices and Queues.
Optical Flow Image Formats
To enumerate the supported image formats for a specific optical flow usage, call:
// Provided by VK_NV_optical_flow
VkResult vkGetPhysicalDeviceOpticalFlowImageFormatsNV(
VkPhysicalDevice physicalDevice,
const VkOpticalFlowImageFormatInfoNV* pOpticalFlowImageFormatInfo,
uint32_t* pFormatCount,
VkOpticalFlowImageFormatPropertiesNV* pImageFormatProperties);
-
physicalDevice
is the physical device being queried. -
pOpticalFlowImageFormatInfo
is a pointer to a VkOpticalFlowImageFormatInfoNV structure specifying the optical flow usage for which information is returned. -
pFormatCount
is a pointer to an integer related to the number of optical flow properties available or queried, as described below. -
pImageFormatProperties
is a pointer to an array of VkOpticalFlowImageFormatPropertiesNV structures in which supported formats and image parameters are returned.
If pImageFormatProperties
is NULL
, then the number of optical flow
properties supported for the given physicalDevice
is returned in
pFormatCount
.
Otherwise, pFormatCount
must point to a variable set by the
application to the number of elements in the pImageFormatProperties
array, and on return the variable is overwritten with the number of values
actually written to pImageFormatProperties
.
If the value of pFormatCount
is less than the number of optical flow
properties supported, at most pFormatCount
values will be written to
pImageFormatProperties
, and VK_INCOMPLETE
will be returned
instead of VK_SUCCESS
, to indicate that not all the available values
were returned.
Before creating an image to be used as an optical flow frame, obtain the
supported image creation parameters by querying with
vkGetPhysicalDeviceFormatProperties2 and
vkGetPhysicalDeviceImageFormatProperties2 using one of the reported
formats and adding VkOpticalFlowImageFormatInfoNV to the pNext
chain of VkPhysicalDeviceImageFormatInfo2.
When querying the parameters with
vkGetPhysicalDeviceImageFormatProperties2 for images used for optical
flow operations, the VkOpticalFlowImageFormatInfoNV::usage
field
must contain one or more of the bits defined in
VkOpticalFlowUsageFlagBitsNV.
|
The VkOpticalFlowImageFormatInfoNV structure is defined as:
// Provided by VK_NV_optical_flow
typedef struct VkOpticalFlowImageFormatInfoNV {
VkStructureType sType;
const void* pNext;
VkOpticalFlowUsageFlagsNV usage;
} VkOpticalFlowImageFormatInfoNV;
-
sType
is a VkStructureType value identifying this structure. -
pNext
isNULL
or a pointer to a structure extending this structure. -
usage
is a bitmask of VkOpticalFlowUsageFlagBitsNV describing the intended optical flow usage of the image.
Bits which can be set in VkOpticalFlowImageFormatInfoNV::usage
,
controlling optical flow usage, are:
// Provided by VK_NV_optical_flow
typedef enum VkOpticalFlowUsageFlagBitsNV {
VK_OPTICAL_FLOW_USAGE_UNKNOWN_NV = 0,
VK_OPTICAL_FLOW_USAGE_INPUT_BIT_NV = 0x00000001,
VK_OPTICAL_FLOW_USAGE_OUTPUT_BIT_NV = 0x00000002,
VK_OPTICAL_FLOW_USAGE_HINT_BIT_NV = 0x00000004,
VK_OPTICAL_FLOW_USAGE_COST_BIT_NV = 0x00000008,
VK_OPTICAL_FLOW_USAGE_GLOBAL_FLOW_BIT_NV = 0x00000010,
} VkOpticalFlowUsageFlagBitsNV;
-
VK_OPTICAL_FLOW_USAGE_INPUT_BIT_NV
specifies that the image can be used as input or reference frame for an optical flow operation. -
VK_OPTICAL_FLOW_USAGE_OUTPUT_BIT_NV
specifies that the image can be used as output flow vector map for an optical flow operation. -
VK_OPTICAL_FLOW_USAGE_HINT_BIT_NV
specifies that the image can be used as hint flow vector map for an optical flow operation. -
VK_OPTICAL_FLOW_USAGE_COST_BIT_NV
specifies that the image can be used as output cost map for an optical flow operation. -
VK_OPTICAL_FLOW_USAGE_GLOBAL_FLOW_BIT_NV
specifies that the image can be used as global flow vector for an optical flow operation.
// Provided by VK_NV_optical_flow
typedef VkFlags VkOpticalFlowUsageFlagsNV;
VkOpticalFlowUsageFlagsNV
is a bitmask type for setting a mask of zero
or more VkOpticalFlowUsageFlagBitsNV.
The VkOpticalFlowImageFormatPropertiesNV structure is defined as:
// Provided by VK_NV_optical_flow
typedef struct VkOpticalFlowImageFormatPropertiesNV {
VkStructureType sType;
const void* pNext;
VkFormat format;
} VkOpticalFlowImageFormatPropertiesNV;
-
sType
is a VkStructureType value identifying this structure. -
pNext
isNULL
or a pointer to a structure extending this structure. -
format
is a VkFormat that specifies the format that can be used with the specified optical flow image usages.
Optical Flow Session
Optical Flow Session Object
Optical flow session objects are abstracted and represented by VkOpticalFlowSessionNV handles:
// Provided by VK_NV_optical_flow
VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkOpticalFlowSessionNV)
Creating an Optical Flow Session
To create an optical flow session object, call:
// Provided by VK_NV_optical_flow
VkResult vkCreateOpticalFlowSessionNV(
VkDevice device,
const VkOpticalFlowSessionCreateInfoNV* pCreateInfo,
const VkAllocationCallbacks* pAllocator,
VkOpticalFlowSessionNV* pSession);
-
device
is the logical device that creates the optical flow session object. -
pCreateInfo
is a pointer to a VkOpticalFlowSessionCreateInfoNV structure containing parameters specifying the creation of the optical flow session. -
pAllocator
controls host memory allocation as described in the Memory Allocation chapter. -
pSession
is a pointer to a VkOpticalFlowSessionNV handle specifying the optical flow session object which will be created by this function when it returnsVK_SUCCESS
The VkOpticalFlowSessionCreateInfoNV structure is defined as:
// Provided by VK_NV_optical_flow
typedef struct VkOpticalFlowSessionCreateInfoNV {
VkStructureType sType;
void* pNext;
uint32_t width;
uint32_t height;
VkFormat imageFormat;
VkFormat flowVectorFormat;
VkFormat costFormat;
VkOpticalFlowGridSizeFlagsNV outputGridSize;
VkOpticalFlowGridSizeFlagsNV hintGridSize;
VkOpticalFlowPerformanceLevelNV performanceLevel;
VkOpticalFlowSessionCreateFlagsNV flags;
} VkOpticalFlowSessionCreateInfoNV;
-
sType
is a VkStructureType value identifying this structure. -
pNext
isNULL
or a pointer to a structure extending this structure. -
width
is the width in pixels of the input or reference frame to be bound to this optical flow session. -
height
is the height in pixels of the input or reference frame to be bound to this optical flow session. -
imageFormat
is the VkFormat of the input and reference frame to be bound to this optical flow session. -
flowVectorFormat
is the VkFormat of the flow vector maps (output or hint) to be bound to this optical flow session. -
costFormat
is the VkFormat of the cost maps to be bound to this optical flow session. -
outputGridSize
is exactly one bit of VkOpticalFlowGridSizeFlagsNV specifying the grid size of the output flow and cost maps to be bound to this optical flow session. The size of the output flow and cost maps is determined byVkOpticalFlowSessionCreateInfoNV
::width
andVkOpticalFlowSessionCreateInfoNV
::height
divided byVkOpticalFlowSessionCreateInfoNV
::outputGridSize
. -
hintGridSize
is one exactly bit of VkOpticalFlowGridSizeFlagsNV specifying the grid size of the hint flow vector maps to be bound to this optical flow session. The size of the hint maps is determined byVkOpticalFlowSessionCreateInfoNV
::width
andVkOpticalFlowSessionCreateInfoNV
::height
divided byVkOpticalFlowSessionCreateInfoNV
::hintGridSize
. -
performanceLevel
is the VkOpticalFlowPerformanceLevelNV used for this optical flow session. -
flags
are the VkOpticalFlowSessionCreateFlagsNV used for this optical flow session.
The VkOpticalFlowSessionCreatePrivateDataInfoNV structure is for NV internal use only and is defined as:
// Provided by VK_NV_optical_flow
typedef struct VkOpticalFlowSessionCreatePrivateDataInfoNV {
VkStructureType sType;
void* pNext;
uint32_t id;
uint32_t size;
const void* pPrivateData;
} VkOpticalFlowSessionCreatePrivateDataInfoNV;
-
sType
is a VkStructureType value identifying this structure. -
pNext
isNULL
or a pointer to a structure extending this structure. -
id
is an identifier for data which is passed at a memory location specified inVkOpticalFlowSessionCreatePrivateDataInfoNV
::pPrivateData
. -
size
is the size of data in bytes which is passed at a memory location specified inVkOpticalFlowSessionCreatePrivateDataInfoNV
::pPrivateData
. -
pPrivateData
is a pointer to NV internal data.
Optical flow vectors are generated block-wise, one vector for each block of NxN pixels (referred to as grid).
Bits which can be set in
VkOpticalFlowSessionCreateInfoNV::outputGridSize
and
VkOpticalFlowSessionCreateInfoNV::hintGridSize
, or which are
returned in
VkPhysicalDeviceOpticalFlowPropertiesNV::supportedOutputGridSizes
and
VkPhysicalDeviceOpticalFlowPropertiesNV::supportedHintGridSizes
controlling optical flow grid sizes, are:
// Provided by VK_NV_optical_flow
typedef enum VkOpticalFlowGridSizeFlagBitsNV {
VK_OPTICAL_FLOW_GRID_SIZE_UNKNOWN_NV = 0,
VK_OPTICAL_FLOW_GRID_SIZE_1X1_BIT_NV = 0x00000001,
VK_OPTICAL_FLOW_GRID_SIZE_2X2_BIT_NV = 0x00000002,
VK_OPTICAL_FLOW_GRID_SIZE_4X4_BIT_NV = 0x00000004,
VK_OPTICAL_FLOW_GRID_SIZE_8X8_BIT_NV = 0x00000008,
} VkOpticalFlowGridSizeFlagBitsNV;
-
VK_OPTICAL_FLOW_GRID_SIZE_1X1_BIT_NV
specifies that grid is 1x1 pixel. -
VK_OPTICAL_FLOW_GRID_SIZE_2X2_BIT_NV
specifies that grid is 2x2 pixel. -
VK_OPTICAL_FLOW_GRID_SIZE_4X4_BIT_NV
specifies that grid is 4x4 pixel. -
VK_OPTICAL_FLOW_GRID_SIZE_8X8_BIT_NV
specifies that grid is 8x8 pixel.
// Provided by VK_NV_optical_flow
typedef VkFlags VkOpticalFlowGridSizeFlagsNV;
VkOpticalFlowGridSizeFlagsNV
is a bitmask type for setting a mask of
zero or more VkOpticalFlowGridSizeFlagBitsNV.
Optical flow exposes performance levels which the application can choose based on the desired performance and quality requirement.
The optical flow performance level types are defined with the following:
// Provided by VK_NV_optical_flow
typedef enum VkOpticalFlowPerformanceLevelNV {
VK_OPTICAL_FLOW_PERFORMANCE_LEVEL_UNKNOWN_NV = 0,
VK_OPTICAL_FLOW_PERFORMANCE_LEVEL_SLOW_NV = 1,
VK_OPTICAL_FLOW_PERFORMANCE_LEVEL_MEDIUM_NV = 2,
VK_OPTICAL_FLOW_PERFORMANCE_LEVEL_FAST_NV = 3,
} VkOpticalFlowPerformanceLevelNV;
-
VK_OPTICAL_FLOW_PERFORMANCE_LEVEL_SLOW_NV
is a level with slower performance but higher quality. -
VK_OPTICAL_FLOW_PERFORMANCE_LEVEL_MEDIUM_NV
is a level with medium performance and medium quality. -
VK_OPTICAL_FLOW_PERFORMANCE_LEVEL_FAST_NV
is a preset with higher performance but lower quality.
Bits which can be set in
VkOpticalFlowSessionCreateInfoNV::flags
, controlling optical
flow session operations, are:
// Provided by VK_NV_optical_flow
typedef enum VkOpticalFlowSessionCreateFlagBitsNV {
VK_OPTICAL_FLOW_SESSION_CREATE_ENABLE_HINT_BIT_NV = 0x00000001,
VK_OPTICAL_FLOW_SESSION_CREATE_ENABLE_COST_BIT_NV = 0x00000002,
VK_OPTICAL_FLOW_SESSION_CREATE_ENABLE_GLOBAL_FLOW_BIT_NV = 0x00000004,
VK_OPTICAL_FLOW_SESSION_CREATE_ALLOW_REGIONS_BIT_NV = 0x00000008,
VK_OPTICAL_FLOW_SESSION_CREATE_BOTH_DIRECTIONS_BIT_NV = 0x00000010,
} VkOpticalFlowSessionCreateFlagBitsNV;
-
VK_OPTICAL_FLOW_SESSION_CREATE_ENABLE_HINT_BIT_NV
specifies that a VkImageView with external flow vectors will be used as hints in performing the motion search and must be bound toVK_OPTICAL_FLOW_SESSION_BINDING_POINT_HINT_NV
. -
VK_OPTICAL_FLOW_SESSION_CREATE_ENABLE_COST_BIT_NV
specifies that the cost for the forward flow is generated in a VkImageView which must be bound toVK_OPTICAL_FLOW_SESSION_BINDING_POINT_COST_NV
. Additionally, ifVK_OPTICAL_FLOW_SESSION_CREATE_BOTH_DIRECTIONS_BIT_NV
is also set, the cost for backward flow is generated in a VkImageView which must be bound toVK_OPTICAL_FLOW_SESSION_BINDING_POINT_BACKWARD_COST_NV
. The cost is the confidence level of the flow vector for each grid in the frame. The Cost implies how (in)accurate the flow vector is. Higher cost value implies the flow vector to be less accurate and vice-versa. -
VK_OPTICAL_FLOW_SESSION_CREATE_ENABLE_GLOBAL_FLOW_BIT_NV
specifies that a global flow vector is estimated from forward flow in a single pixel VkImageView which must be bound toVK_OPTICAL_FLOW_SESSION_BINDING_POINT_GLOBAL_FLOW_NV
. -
VK_OPTICAL_FLOW_SESSION_CREATE_ALLOW_REGIONS_BIT_NV
specifies that regions of interest can be specified in VkOpticalFlowExecuteInfoNV. -
VK_OPTICAL_FLOW_SESSION_CREATE_BOTH_DIRECTIONS_BIT_NV
specifies that backward flow is generated in addition to forward flow which is always generated.
VkOpticalFlowSessionCreateFlagsNV
is a bitmask type for setting a mask
of zero or more VkOpticalFlowSessionCreateFlagBitsNV.
// Provided by VK_NV_optical_flow
typedef VkFlags VkOpticalFlowSessionCreateFlagsNV;
Destroying an Optical Flow Session
To destroy an optical flow session object, call:
// Provided by VK_NV_optical_flow
void vkDestroyOpticalFlowSessionNV(
VkDevice device,
VkOpticalFlowSessionNV session,
const VkAllocationCallbacks* pAllocator);
-
device
is the device that was used for the creation of the optical flow session. -
session
is the optical flow session to be destroyed. -
pAllocator
controls host memory allocation as described in the Memory Allocation chapter.
Binding Vulkan Image Views to an Optical Flow Session
To bind a vulkan image to an optical flow session object, call:
// Provided by VK_NV_optical_flow
VkResult vkBindOpticalFlowSessionImageNV(
VkDevice device,
VkOpticalFlowSessionNV session,
VkOpticalFlowSessionBindingPointNV bindingPoint,
VkImageView view,
VkImageLayout layout);
-
device
is the device which owns the optical flow session objectsession
. -
session
is the optical flow session object to which the image view is to be bound. -
bindingPoint
specifies the binding point VkOpticalFlowSessionBindingPointNV to which the image view is bound. -
view
is a VkImageView to be bound. -
layout must specify the layout that the image subresources accessible from
view
will be in at the time the optical flow vectors are calculated with vkCmdOpticalFlowExecuteNV on aVkDevice
.
The optical flow session binding points are defined with the following:
// Provided by VK_NV_optical_flow
typedef enum VkOpticalFlowSessionBindingPointNV {
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_UNKNOWN_NV = 0,
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_INPUT_NV = 1,
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_REFERENCE_NV = 2,
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_HINT_NV = 3,
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_FLOW_VECTOR_NV = 4,
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_BACKWARD_FLOW_VECTOR_NV = 5,
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_COST_NV = 6,
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_BACKWARD_COST_NV = 7,
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_GLOBAL_FLOW_NV = 8,
} VkOpticalFlowSessionBindingPointNV;
-
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_INPUT_NV
specifies the binding point for the input frame. -
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_REFERENCE_NV
specifies the binding point for the input reference frame. -
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_HINT_NV
specifies the binding point for the optional external hint flow vectors. -
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_FLOW_VECTOR_NV
specifies the binding point for output flow vectors of default forward flow calculation. -
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_BACKWARD_FLOW_VECTOR_NV
specifies the binding point for the optional output flow vector map of optional backward flow calculation. -
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_COST_NV
specifies the binding point for the optional output cost map of default forward flow calculation. -
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_BACKWARD_COST_NV
specifies the binding point for the optional output cost map of optional backward flow calculation. -
VK_OPTICAL_FLOW_SESSION_BINDING_POINT_GLOBAL_FLOW_NV
specifies the binding point for the optional global flow value of default forward flow calculation.
Optical Flow Execution
Default direction of flow estimation is forward which calculates the optical flow from input frame to reference frame. Optionally backward flow estimation can be additionally calculated. An output flow vector (Vx, Vy) means that current pixel (x, y) of input frame can be found at location (x+Vx, y+Vy) in reference frame. A backward flow vector (Vx, Vy) means that current pixel (x, y) of reference frame can be found at location (x+Vx, y+Vy) in input frame.
To calculate optical flow vectors from two input frames, call:
// Provided by VK_NV_optical_flow
void vkCmdOpticalFlowExecuteNV(
VkCommandBuffer commandBuffer,
VkOpticalFlowSessionNV session,
const VkOpticalFlowExecuteInfoNV* pExecuteInfo);
-
commandBuffer
is the command buffer into which the command will be recorded. -
session
is the optical flow session object on which this command is operating. -
pExecuteInfo
Info is a pointer to a VkOpticalFlowExecuteInfoNV.
The VkOpticalFlowExecuteInfoNV structure is defined as:
// Provided by VK_NV_optical_flow
typedef struct VkOpticalFlowExecuteInfoNV {
VkStructureType sType;
void* pNext;
VkOpticalFlowExecuteFlagsNV flags;
uint32_t regionCount;
const VkRect2D* pRegions;
} VkOpticalFlowExecuteInfoNV;
-
sType
is a VkStructureType value identifying this structure. -
pNext
isNULL
or a pointer to a structure extending this structure. -
flags
are the VkOpticalFlowExecuteFlagsNV used for this command. -
regionCount
is the number of regions of interest specified inpRegions
. -
pRegions
is a pointer toregionCount
VkRect2D
regions of interest.
Bits which can be set in VkOpticalFlowExecuteInfoNV::flags
,
controlling optical flow execution, are:
// Provided by VK_NV_optical_flow
typedef enum VkOpticalFlowExecuteFlagBitsNV {
VK_OPTICAL_FLOW_EXECUTE_DISABLE_TEMPORAL_HINTS_BIT_NV = 0x00000001,
} VkOpticalFlowExecuteFlagBitsNV;
-
VK_OPTICAL_FLOW_EXECUTE_DISABLE_TEMPORAL_HINTS_BIT_NV
specifies that temporal hints from previously generated flow vectors are not used. If temporal hints are enabled, optical flow vectors from previous vkCmdOpticalFlowExecuteNV call are automatically used as hints for the current vkCmdOpticalFlowExecuteNV call, to take advantage of temporal correlation in a video sequence. Temporal hints should be disabled if there is a-priori knowledge of no temporal correlation (e.g. a scene change, independent successive frame pairs).
VkOpticalFlowExecuteFlagsNV
is a bitmask type for setting a mask of
zero or more VkOpticalFlowExecuteFlagBitsNV.
// Provided by VK_NV_optical_flow
typedef VkFlags VkOpticalFlowExecuteFlagsNV;