TvResize

group TvResize

Reference: torchvision.transforms.v2.functional.resize (v0.28.0), with the explicit rounding and saturation semantics documented below.

Functions

NVCVStatus cvcudaTvResizeCreate(NVCVOperatorHandle *handle)

Constructs a stateless TvResize operator.

Parameters:

handle – [out] Nonnull location receiving the owned operator handle.

Return values:

  • NVCV_SUCCESS – Success.

  • NVCV_ERROR_INVALID_ARGUMENT – Null handle pointer.

  • NVCV_ERROR_OUT_OF_MEMORY – Allocation failed.

NVCVStatus cvcudaTvResizeTensorGetWorkspaceRequirements(NVCVOperatorHandle handle, int batchSize, int32_t srcHeight, int32_t srcWidth, int32_t dstHeight, int32_t dstWidth, int32_t numChannels, NVCVInterpolationType interpolation, bool antialias, NVCVWorkspaceRequirements *reqOut)

Queries caller-owned scratch for equally sized tensor samples.

Parameters:

  • handle – [in] TvResize operator handle.

  • batchSize – [in] Sample count in [0,65535].

  • srcHeight – [in] Positive input height.

  • srcWidth – [in] Positive input width.

  • dstHeight – [in] Positive output height.

  • dstWidth – [in] Positive output width.

  • numChannels – [in] Positive scalar channel count.

  • interpolation – [in] NEAREST, NEAREST_FLOOR, LINEAR or CUBIC.

  • antialias – [in] Enable independently dilated, clipped and normalized axis filters.

  • reqOut – [out] Nonnull requirements output. All supplied dimensions and modes are validated even for zero workspace. N=0 requires zero workspace. Queries have no layout/dtype and cannot enforce their limits.

Return values:

  • NVCV_SUCCESS – Success.

  • NVCV_ERROR_INVALID_ARGUMENT – Invalid handle, null output, dimensions, mode or size overflow.

NVCVStatus cvcudaTvResizeImageBatchGetWorkspaceRequirements(NVCVOperatorHandle handle, int batchSize, const int32_t *srcHeights, const int32_t *srcWidths, const int32_t *dstHeights, const int32_t *dstWidths, int32_t numChannels, NVCVInterpolationType interpolation, bool antialias, NVCVWorkspaceRequirements *reqOut)

Queries scratch for a uniform-format variable-shape image batch. The four extent arrays contain batchSize positive entries and must be nonnull even when batchSize=0. Other parameters/statuses follow the tensor query.

Parameters:

  • handle – [in] TvResize operator handle.

  • batchSize – [in] Sample count in [0,65535].

  • srcHeights – [in] Input heights.

  • srcWidths – [in] Input widths.

  • dstHeights – [in] Output heights.

  • dstWidths – [in] Output widths.

  • numChannels – [in] Channel count of 1, 3 or 4, including when batchSize=0.

  • interpolation – [in] Supported interpolation mode.

  • antialias – [in] Enable antialiasing.

  • reqOut – [out] Nonnull requirements output.

Return values:

  • NVCV_SUCCESS – Success.

  • NVCV_ERROR_INVALID_ARGUMENT – Invalid handle, null arrays/output, dimensions, mode or size overflow.

NVCVStatus cvcudaTvResizeSubmit(NVCVOperatorHandle handle, cudaStream_t stream, const NVCVWorkspace *workspace, NVCVTensorHandle in, NVCVTensorHandle out, NVCVInterpolationType interpolation, bool antialias)

Enqueues spatial resize on the caller stream without waiting for completion.

Limitations: Input: Data Layout: [HWC, NHWC, CHW, NCHW] Channels: [1, 2, 3, 4] for HWC/NHWC; any positive C with N*C<=65535 for CHW/NCHW. 1<=N<=65535; unbatched layouts imply N=1. H/W/C must be positive.

Data Type

Allowed

8bit Unsigned

Yes

8bit Signed

No

16bit Unsigned

Yes

16bit Signed

Yes

16bit Float

Yes

32bit Unsigned

No

32bit Signed

No

32bit Float

Yes

64bit Float

No

Output: same layouts, channels and scalar types as input. Vector element types are not supported. CUDA-accessible positive padded strides are supported with checked 64-bit address arithmetic; negative/overlapping-element strides are invalid.

Input/Output dependency:

Property

Input == Output

Data Layout

Yes

Data Type

Yes

Number

Yes

Channels

Yes

Width

No

Height

No

NEAREST gathers floor((dst+0.5)*scale), NEAREST_FLOOR gathers floor(dst*scale), with float32 scale=S/D and coordinates, clamped to S-1. Both preserve sampled bits and ignore antialias. LINEAR/CUBIC use half-pixel centers and float32 accumulation. Plain cubic uses Keys a=-0.75 and clamped taps. Antialias uses per-axis dilation max(1,S/D), clips taps to the image, and normalizes surviving weights; cubic uses a=-0.5 even when enlarging. Singleton axes are supported; identity is bit-exact. Integers are rounded ties-to-even and saturated, F16 rounds to nearest-even. Interpolation guarantees apply to finite inputs with finite representable results; no nonfinite pixel scan is performed. Unlike external TorchVision casting, U16/S16 cubic overshoot saturates rather than wraps.

Source/destination storage must not overlap, even for identity (caller precondition). The shared Tensor API rejects zero extents at construction; empty Tensor submits are not representable. N=0 workspace queries still return zero workspace. All metadata is validated before output work. The caller keeps inputs, outputs and workspace alive until stream completion. No hidden device allocation or host synchronization.

Parameters:

  • handle – [in] TvResize operator handle.

  • stream – [in] CUDA stream.

  • workspace – [in] Queried scratch with sufficient size/alignment; may be null only when zero scratch is needed.

  • in – [in] Source tensor.

  • out – [out] Destination tensor.

  • interpolation – [in] NEAREST, NEAREST_FLOOR, LINEAR or CUBIC.

  • antialias – [in] Enable antialiasing (ignored for nearest).

Return values:

  • NVCV_SUCCESS – Work enqueued successfully.

  • NVCV_ERROR_INVALID_ARGUMENT – Invalid handle, metadata, mode, workspace or resource-size overflow.

  • NVCV_ERROR_OUT_OF_MEMORY – Resource allocation failed.

  • NVCV_ERROR_INTERNAL – Internal error. Device execution errors may be asynchronous.

NVCVStatus cvcudaTvResizeImageBatchSubmit(NVCVOperatorHandle handle, cudaStream_t stream, const NVCVWorkspace *workspace, NVCVImageBatchHandle in, NVCVImageBatchHandle out, NVCVInterpolationType interpolation, bool antialias)

Enqueues variable-shape resize with the sampling/workspace/lifetime contract of cvcudaTvResizeSubmit. Limitations: CUDA-accessible pitch-linear single-plane interleaved images, channels [1,3,4], scalar U8/U16/S16/F16/F32, N<=65535. Each batch must have a uniform format, matching between input and output. Counts match; only spatial sizes may differ. Two-channel, multiplane and mixed-format batches are rejected. Matched empty batches succeed with zero workspace and no launches.

Parameters:

  • handle – [in] TvResize operator handle.

  • stream – [in] CUDA stream.

  • workspace – [in] Queried scratch, or null if no scratch is needed.

  • in – [in] Source image batch.

  • out – [out] Destination image batch.

  • interpolation – [in] Supported interpolation mode.

  • antialias – [in] Enable antialiasing.

Return values:

  • NVCV_SUCCESS – Work enqueued successfully.

  • NVCV_ERROR_INVALID_ARGUMENT – Invalid handle, metadata, mode, workspace or size overflow.

  • NVCV_ERROR_OUT_OF_MEMORY – Resource allocation failed.

  • NVCV_ERROR_INTERNAL – Internal error; device execution errors may be asynchronous.