Data Types

class cvcuda.Image

cpu(self: cvcuda.Image, layout: cvcuda.TensorLayout | None = None) → object: The image on the CPU

cuda(self: cvcuda.Image, layout: cvcuda.TensorLayout | None = None) → object: The image on the CUDA device

property format: Read-only property that returns the format of the image

property height: Read-only property that returns the height of the image

property size: Read-only property that returns the size of the image

property width: Read-only property that returns the width of the image

static zeros(size: tuple[SupportsInt | SupportsIndex, SupportsInt | SupportsIndex], format: cvcuda.Format, rowalign: SupportsInt | SupportsIndex = 0) → cvcuda.Image: Create an image filled with zeros with a given size, format and optional row align

class cvcuda.Tensor

cuda(self: cvcuda.Tensor) → object: Reference to the Tensor on the CUDA device.

property dtype: The data type of the Tensor.

property layout: The TensorLayout of the Tensor.

property ndim: The number of dimensions of the Tensor.

reshape(self: cvcuda.Tensor, shape: tuple, layout: cvcuda.TensorLayout | None = None) → cvcuda.Tensor: Produces a tensor pointing to the same data but with a new shape and layout.

property shape: The shape of the Tensor.

class cvcuda.ImageBatchVarShape

Batch of Images.

property capacity: Return the capacity of the ImageBatchVarShape in number of images.

clear(self: cvcuda.ImageBatchVarShape) → None: Remove all images from the ImageBatchVarShape.

property maxsize: Return the maximum size of the ImageBatchVarShape in bytes.

popback(self: cvcuda.ImageBatchVarShape, count: SupportsInt | SupportsIndex = 1) → None: Remove one or more images from the end of the ImageBatchVarShape.

pushback(*args, **kwargs)

Overloaded function.

pushback(self: cvcuda.ImageBatchVarShape, arg0: cvcuda.Image) -> None

Add a new image to the end of the ImageBatchVarShape.

pushback(self: cvcuda.ImageBatchVarShape, arg0: collections.abc.Sequence[cvcuda.Image]) -> None

Add multiple images to the end of the ImageBatchVarShape.

property uniqueformat: Return True if all the images have the same format, False otherwise.

class cvcuda.TensorBatch

Container for a batch of tensors. The capacity of the container must be specified upfront in the batch initialization. The tensors in the batch may differ in shapes but they must have a uniform dimensionality, data type and layout.

property capacity: Capacity of the tensor batch.

clear(self: cvcuda.TensorBatch) → None: Remove all images from the TensorBatch.

property dtype: Data type of tensors in the tensor batch. None if the batch is empty.

property layout: Layout of the tensors in the tensor batch. None if the batch is empty.

property ndim: Return the number of dimensions of the tensors or -1 for an empty batch

popback(self: cvcuda.TensorBatch, count: SupportsInt | SupportsIndex = 1) → None: Remove one or more images from the end of the TensorBatch.

pushback(*args, **kwargs)

Overloaded function.

pushback(self: cvcuda.TensorBatch, arg0: cvcuda.Tensor) -> None

Add a new image to the end of the TensorBatch.

pushback(self: cvcuda.TensorBatch, arg0: collections.abc.Sequence[cvcuda.Tensor]) -> None

Add multiple images to the end of the TensorBatch.

class cvcuda.Format

Members:

_2S16

_2S16_BL

_2F16

_2F32

_2C64

_2C128

U8

U8_BL

S8

U16

U32

S32

S16

S16_BL

F16

F32

F64

C64

C128

Y8

Y8_BL

Y8_ER

Y8_ER_BL

Y16

Y16_BL

Y16_ER

Y16_ER_BL

NV12

NV12_BL

NV12_ER

NV12_ER_BL

NV21

NV21_ER

NV24

NV24_BL

NV24_ER

NV24_ER_BL

UYVY

UYVY_BL

UYVY_ER

UYVY_ER_BL

VYUY

VYUY_BL

VYUY_ER

VYUY_ER_BL

YUYV

YUYV_BL

YUYV_ER

YUYV_ER_BL

YUV8p

YUV8p_ER

RGB8

BGR8

RGBA8

BGRA8

RGB8p

BGR8p

RGBA8p

BGRA8p

RGBf16

BGRf16

RGBAf16

BGRAf16

RGBf16p

BGRf16p

RGBAf16p

BGRAf16p

RGBf32

BGRf32

RGBAf32

BGRAf32

RGBf32p

BGRf32p

RGBAf32p

BGRAf32p

HSV8

CMYK8

YCCK8

RGBA8_UNASSOCIATED_ALPHA

RGB8_1U_U8

RGB8_7U_U8

RGBA8_3U_U16

RGBA8_3POS3D_U32

RGB8_3D_F32

NONE

BGR8 = nvcv.Format.BGR8

BGR8p = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,UNSIGNED,ZYX1,ASSOCIATED,X8,X8,X8)>

BGRA8 = nvcv.Format.BGRA8

BGRA8p = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,UNSIGNED,ZYXW,ASSOCIATED,X8,X8,X8,X8)>

BGRAf16 = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,ZYXW,ASSOCIATED,X16_Y16_Z16_W16)>

BGRAf16p = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,ZYXW,ASSOCIATED,X16,X16,X16,X16)>

BGRAf32 = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,ZYXW,ASSOCIATED,X32_Y32_Z32_W32)>

BGRAf32p = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,ZYXW,ASSOCIATED,X32,X32,X32,X32)>

BGRf16 = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,ZYX1,ASSOCIATED,X16_Y16_Z16)>

BGRf16p = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,ZYX1,ASSOCIATED,X16,X16,X16)>

BGRf32 = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,ZYX1,ASSOCIATED,X32_Y32_Z32)>

BGRf32p = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,ZYX1,ASSOCIATED,X32,X32,X32)>

C128 = nvcv.Format.C128

C64 = nvcv.Format.C64

CMYK8 = nvcv.Format.CMYK8

F16 = nvcv.<Unknown image format: NVCVImageFormat(UNDEFINED,PITCH_LINEAR,FLOAT,X000,ASSOCIATED,X16)>

F32 = nvcv.Format.F32

F64 = nvcv.Format.F64

HSV8 = nvcv.Format.HSV8

NONE = nvcv.Format.NONE

NV12 = nvcv.Format.NV12

NV12_BL = nvcv.Format.NV12_BL

NV12_ER = nvcv.Format.NV12_ER

NV12_ER_BL = nvcv.Format.NV12_ER_BL

NV21 = nvcv.Format.NV12

NV21_ER = nvcv.Format.NV12_ER

NV24 = nvcv.Format.NV24

NV24_BL = nvcv.Format.NV24_BL

NV24_ER = nvcv.Format.NV24_ER

NV24_ER_BL = nvcv.Format.NV24_ER_BL

RGB8 = nvcv.Format.RGB8

RGB8_1U_U8 = nvcv.Format.RGB8_1U_U8

RGB8_3D_F32 = nvcv.Format.RGB8_3D_F32

RGB8_7U_U8 = nvcv.Format.RGB8_7U_U8

RGB8p = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,UNSIGNED,XYZ0,ASSOCIATED,X8,X8,X8)>

RGBA8 = nvcv.Format.RGBA8

RGBA8_3POS3D_U32 = nvcv.Format.RGBA8_3POS3D_U32

RGBA8_3U_U16 = nvcv.Format.RGBA8_3U_U16

RGBA8_UNASSOCIATED_ALPHA = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,UNSIGNED,XYZW,UNASSOCIATED,X8_Y8_Z8_W8)>

RGBA8p = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,UNSIGNED,XYZW,ASSOCIATED,X8,X8,X8,X8)>

RGBAf16 = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,XYZW,ASSOCIATED,X16_Y16_Z16_W16)>

RGBAf16p = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,XYZW,ASSOCIATED,X16,X16,X16,X16)>

RGBAf32 = nvcv.Format.RGBAf32

RGBAf32p = nvcv.Format.RGBAf32p

RGBf16 = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,XYZ1,ASSOCIATED,X16_Y16_Z16)>

RGBf16p = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,XYZ0,ASSOCIATED,X16,X16,X16)>

RGBf32 = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,XYZ1,ASSOCIATED,X32_Y32_Z32)>

RGBf32p = nvcv.<Unknown image format: NVCVImageFormat(RGB,UNDEFINED,PITCH_LINEAR,FLOAT,XYZ0,ASSOCIATED,X32,X32,X32)>

S16 = nvcv.Format.S16

S16_BL = nvcv.Format.S16_BL

S32 = nvcv.Format.S32

S8 = nvcv.Format.S8

U16 = nvcv.Format.U16

U32 = nvcv.Format.U32

U8 = nvcv.Format.U8

U8_BL = nvcv.Format.U8_BL

UYVY = nvcv.Format.UYVY

UYVY_BL = nvcv.Format.UYVY_BL

UYVY_ER = nvcv.Format.UYVY_ER

UYVY_ER_BL = nvcv.Format.UYVY_ER_BL

VYUY = nvcv.Format.VYUY

VYUY_BL = nvcv.Format.VYUY_BL

VYUY_ER = nvcv.Format.VYUY_ER

VYUY_ER_BL = nvcv.Format.VYUY_ER_BL

Y16 = nvcv.Format.Y16

Y16_BL = nvcv.Format.Y16_BL

Y16_ER = nvcv.Format.Y16_ER

Y16_ER_BL = nvcv.Format.Y16_ER_BL

Y8 = nvcv.Format.Y8

Y8_BL = nvcv.Format.Y8_BL

Y8_ER = nvcv.Format.Y8_ER

Y8_ER_BL = nvcv.Format.Y8_ER_BL

YCCK8 = nvcv.Format.YCCK8

YUV8p = nvcv.Format.YUV8p

YUV8p_ER = nvcv.Format.YUV8p_ER

YUYV = nvcv.Format.YUYV

YUYV_BL = nvcv.Format.YUYV_BL

YUYV_ER = nvcv.Format.YUYV_ER

YUYV_ER_BL = nvcv.Format.YUYV_ER_BL

property channels: Read-only property that returns the number of color channels in the image

property name

property planes: Read-only property that returns the number of planes in the image

property value

class cvcuda.ColorSpec

Members:

BT601 : Color spec defining ITU-R BT.601 standard.

BT709 : Color spec defining ITU-R BT.709 standard.

BT2020 : Color spec defining ITU-R BT.2020 standard.

BT2020 = <ColorSpec.BT2020: 21930>

BT601 = <ColorSpec.BT601: 21665>

BT709 = <ColorSpec.BT709: 21793>

property name

property value

class cvcuda.TensorLayout

C = C

CDHW = CDHW

CFDHW = CFDHW

CFHW = CFHW

CHW = CHW

CW = CW

D = D

DHW = DHW

DHWC = DHWC

F = F

FCDHW = FCDHW

FCHW = FCHW

FDHW = FDHW

FDHWC = FDHWC

FHW = FHW

FHWC = FHWC

H = H

HW = HW

HWC = HWC

N = N

NCDHW = NCDHW

NCFDHW = NCFDHW

NCFHW = NCFHW

NCHW = NCHW

NCW = NCW

NDHW = NDHW

NDHWC = NDHWC

NFCDHW = NFCDHW

NFCHW = NFCHW

NFDHW = NFDHW

NFDHWC = NFDHWC

NFHW = NFHW

NFHWC = NFHWC

NHW = NHW

NHWC = NHWC

NONE

NW = NW

NWC = NWC

W = W

WC = WC

class cvcuda.RectI

property height: Height of the rectangle

property width: Width of the rectangle

property x: X coordinate of the rectangle’s top-left corner

property y: Y coordinate of the rectangle’s top-left corner

class cvcuda.Stream

A CUDA stream that cvcuda operators can be submitted to.

Streams created via cvcuda.Stream() are non-blocking (cudaStreamNonBlocking) so they can run concurrently with CUDA’s legacy default stream and with other non-blocking streams.

Cross-library stream safety is handled automatically via the CUDA Array Interface (CAI) v3 stream field:

On input, cvcuda.as_tensor / cvcuda.as_image read __cuda_array_interface__["stream"] from the producer (cupy, torch, jax, etc.) and arrange for the first cvcuda operator to insert the proper cudaStreamWaitEvent before reading the data. No user-side synchronize() call is required.
On output, tensors returned by cvcuda.Tensor.cuda() / cvcuda.Image.cuda() populate __cuda_array_interface__ ["stream"] with the cvcuda stream the data was last written on, so downstream consumers can sync themselves.

Advanced usage:

Opt out of the implicit wait: export your buffer with stream: -1 in its CAI dict to tell cvcuda you’ve already synchronized; the first op then runs without an extra event.
Legacy default stream: if your producer uses v2 CAI (no stream field) or advertises stream: 1, cvcuda waits on the legacy default stream. This is the safe conservative default but forfeits concurrency with stream-0 work.
Use the default stream: cvcuda.Stream.default wraps stream 0 and has the usual legacy-default blocking semantics; use it when you want implicit sync with everything else on stream 0 and don’t need the concurrency of a dedicated stream.

property handle: Get the integer handle for the CUDA stream object.

property id: Get the unique ID for the CUDA stream object.

sync(self: cvcuda.Stream) → None: Wait for all preceding CUDA calls in the current stream to complete.

wait_stream(self: cvcuda.Stream, other: cvcuda.Stream) → None: Insert a dependency on ‘other’ into this stream. All subsequent work enqueued on this stream will wait until all work currently enqueued on ‘other’ has completed. Calling wait_stream(self) is a no-op. Cross-device usage is not supported and will raise a CUDA error.

class cvcuda.Resource

property id: Unique resource instance identifier

submitStreamSync(self: cvcuda.Resource, stream: cvcuda.Stream) → None: Syncs object on new Stream

class cvcuda.ExternalBuffer

property dtype: Get the data type of the buffer

property shape: Get the shape of the buffer as an array

property strides: Get the strides of the buffer

class cvcuda.Container

class cvcuda.Elements

class cvcuda.ThreadScope

Members:

GLOBAL : Global scope.

LOCAL : Thread-local scope.

GLOBAL = <ThreadScope.GLOBAL: 0>

LOCAL = <ThreadScope.LOCAL: 1>

property name

property value