Split

group Split

Functions

NVCVStatus cvcudaSplitCreate(NVCVOperatorHandle *handle)

Creates a stateless Split operator.

Parameters:

handle – [out] Non-null destination for the operator handle.

Return values:

  • NVCV_SUCCESS – Operator successfully created.

  • NVCV_ERROR_INVALID_ARGUMENT – Handle is null.

  • NVCV_ERROR_OUT_OF_MEMORY – Insufficient memory to create the operator.

NVCVStatus cvcudaSplitSubmit(NVCVOperatorHandle handle, cudaStream_t stream, NVCVTensorHandle in, NVCVTensorBatchHandle out)

Copies input channel c to output tensor c, preserving every element’s bits. The output TensorBatch contains channels, not samples: each tensor retains the input’s sample count, height and width, with an explicit channel dimension of one. C=1 still copies data. There is no color conversion or alpha processing. This entry point operates on tensors; see cvcudaSplitVarShapeSubmit for image batches. Submission is asynchronous; all tensor storage must remain alive until completion.

Limitations:

Tensor input: Data Layout: [kNHWC, kHWC, kNCHW, kCHW] Channels: [1, 2, 3, 4]

Data Type

Allowed

8bit Unsigned

Yes

8bit Signed

Yes

16bit Unsigned

Yes

16bit Signed

Yes

32bit Unsigned

Yes

32bit Signed

Yes

16bit Float

Yes

32bit Float

Yes

64bit Float

Yes

Tensor output (each band): Data Layout: [kNHWC, kHWC, kNCHW, kCHW] Channels: [1] Data Type: same scalar type as input; vector/packed/complex types are not supported.

All dimensions must be positive. Input and output ranks must match. Output bands must have identical shape, layout and dtype; their layout may differ from the input’s layout. The number of bands must equal the input channel count. CUDA-accessible, pitch-linear tensors are required. Base addresses and strides must be aligned to the scalar element size, and all strides must be positive. Each non-singleton dimension’s stride must cover its inner dimensions’ byte span in layout order; singleton strides need not. Row/sample padding is allowed. Input/output byte spans and output/output byte spans must not overlap, including padding inside those spans. Padding is not written. No resizing or broadcasting occurs. Floating-point and multichannel non-U8 support extend Pillow’s pixel-copy semantics.

Parameters:

  • handle – [in] Split operator handle, not null.

  • stream – [in] CUDA stream, including the default stream.

  • in – [in] Input tensor, not null.

  • out – [out] Prepopulated batch of destination band tensors, not null.

Return values:

  • NVCV_SUCCESS – Work successfully submitted.

  • NVCV_ERROR_INVALID_ARGUMENT – Invalid metadata, unsupported storage, or overlapping buffers.

NVCVStatus cvcudaSplitVarShapeSubmit(NVCVOperatorHandle handle, cudaStream_t stream, NVCVImageBatchHandle in, const NVCVImageBatchHandle *out, int32_t numOutputs)

Copies physical storage channel c of input image n to image n in output batch c. All element bits are preserved, including floating-point payloads. C=1 still copies data; outputs are not views. Channel order is the component order within an interleaved pixel or plane order for planar images. Color models, swizzles and alpha metadata are not interpreted; no conversion or alpha processing occurs. Submission is asynchronous. All batches, images, image/batch descriptors and backing storage must remain alive and must not be mutated by the caller until completion. The host array of batch handles need only survive this call.

Limitations:

ImageBatchVarShape input: Data Layout: one interleaved plane, or C full-resolution single-channel planes. Channels: [1, 2, 3, 4]

Data Type

Allowed

8bit Unsigned

Yes

8bit Signed

Yes

16bit Unsigned

Yes

16bit Signed

Yes

32bit Unsigned

Yes

32bit Signed

Yes

16bit Float

Yes

32bit Float

Yes

64bit Float

Yes

ImageBatchVarShape output (each band): Data Layout: one single-channel plane. Channels: [1] Data Type: same scalar type as input.

Each batch must be nonempty and have a uniform image format. Every batch has the same number N of images; numOutputs equals the input channel count C. Semantic format tags may differ across band batches if their scalar dtype matches. Image sizes must be positive and match at each corresponding index across batches, but may vary across indices. Pitches and allocations may vary across images and planes. Plane data kinds and packings must describe contiguous, byte-aligned, equal-width components of exactly one of the listed scalar types, without intra-pixel padding. Subsampled, sub-byte, unequal-component and hybrid plane packings are not supported. Extra-channel metadata must have NVCVExtraChannelInfo.numChannels == 0. CUDA-accessible, pitch-linear storage is required; CUDA arrays are not supported. Plane bases and row pitches must be aligned to the scalar size, and each positive row pitch must cover a full row. Every write/read and write/write overlap across the entire submission is rejected, including across image indices, bands and planes and padding inside their byte spans. Repeated read-only sources are permitted. Padding is not written. No resizing or broadcasting occurs.

Parameters:

  • handle – [in] Split operator handle, not null.

  • stream – [in] CUDA stream, including the default stream.

  • in – [in] Nonempty source image batch, not null.

  • out – [out] Host array of numOutputs prepopulated destination band batch handles, none null.

  • numOutputs – [in] Number of band batches, equal to C (1-4).

Return values:

  • NVCV_SUCCESS – Work successfully submitted.

  • NVCV_ERROR_INVALID_ARGUMENT – Invalid metadata, unsupported storage, or overlapping buffers.