DepthToSpace#

Functions#

NVCVStatus pvaDepthToSpaceCreate(NVCVOperatorHandle *handle, NVCVTensorRequirements const *const tensorRequirements, int32_t blkSz)

Constructs an instance of the DepthToSpace operator.

NVCVStatus pvaDepthToSpaceSubmit(NVCVOperatorHandle handle, cupvaStream_t stream, NVCVTensorHandle in, NVCVTensorHandle out)

Submits the DepthToSpace operator to a cuPVA stream.

NVCVStatus pvaDepthToSpaceSubmit(NVCVOperatorHandle handle, cudaStream_t stream, NVCVTensorHandle in, NVCVTensorHandle out)

Submits the DepthToSpace operator to a CUDA stream.

Functions#

NVCVStatus pvaDepthToSpaceCreate(
NVCVOperatorHandle *handle,
NVCVTensorRequirements const *const tensorRequirements,
int32_t blkSz,
)#

Constructs an instance of the DepthToSpace operator.

Rearranges data from the depth (channel) dimension into blocks of spatial data, following the ONNX DepthToSpace operator in DCR (depth-column-row) ordering, see https://onnx.ai/onnx/operators/onnx__DepthToSpace.html. The reverse operation is called SpaceToDepth. Given an input tensor of shape [C, H, W] and block size B, the output tensor has shape [C / (B * B), H * B, W * B] and every output element satisfies

out[c, h * B + by, w * B + bx] = in[(by * B + bx) * (C / (B * B)) + c, h, w]
for 0 <= by, bx < B. This is a pure element rearrangement: values are copied unmodified, without conversion, rounding or saturation, and every input element appears exactly once in the output.

Limitations: The number of input channels must be an exact multiple of blkSz * blkSz.

The tensor is rank 3 in CHW layout: shape[0..2] are the channel count, height and width, and strides[0..1] the channel and row strides.

The input tensor may carry padding, but its regions must not overlap:

  • strides[2] must be 1. Elements within a row are read contiguously and an element stride that skips over bytes cannot be expressed.

  • strides[1] must be at least shape[2]. A row may be followed by padding; it may not run into the next one.

  • strides[0] must be at least shape[1] * strides[1]. A channel may be followed by padding; it may not run into the next one.

The extents and the row pitch are also bounded by the widths of the DMA fields they are programmed into:

  • strides[1] must not exceed 65535, the 16-bit surface line pitch.

  • shape[2] * blkSz must not exceed 65535, the output surface row pitch.

  • shape[1] * blkSz must not exceed 65535, the output surface height.

  • shape[0] * strides[0] must be below 2147483648, the range of the 32-bit offsets the transfers address the tensor by. This bounds the output too: it is tightly packed, and the no-overlap rules above hold shape[0] * shape[1] * shape[2] at or below it.

Every shape and stride is itself read as a 32-bit value, so each must be positive and below 2147483648. A field outside that range is rejected rather than truncated, which would describe a different tensor than the one passed.

The output tensor is derived from the input tensor and the block size, and is tightly packed.

Input: Data Layout: [CHW] Channels: [C, must be a multiple of blkSz * blkSz]

Data Type

Allowed

8bit Unsigned

Yes

8bit Signed

Yes

16bit Unsigned

No

16bit Signed

No

32bit Unsigned

No

32bit Signed

No

32bit Float

No

64bit Float

No

Output: Data Layout: [CHW] Channels: [C / (blkSz * blkSz)]

Data Type

Allowed

8bit Unsigned

Yes

8bit Signed

Yes

16bit Unsigned

No

16bit Signed

No

32bit Unsigned

No

32bit Signed

No

32bit Float

No

64bit Float

No

Input/Output Dependency:

Property

Input == Output

Data Layout

Yes

Data Type

Yes

Number

Yes

Channels

No, Output Channels = (Input Channels) / (blkSz^2)

Width

No, Output Width = (Input Width) * (blkSz)

Height

No, Output Height = (Input Height) * (blkSz)

Parameters

 blkSz = 2, 3 or 4

Parameters:
  • handle – [out] Where the operator instance handle will be written to.

    • Must not be NULL.

  • tensorRequirements – [in] Pointer to the NVCVTensorRequirements structure which contains the input Tensor rank, shape, layout and data type information. The output tensor requirements are derived from it and the block size.

    • Must not be NULL.

  • blkSz – [in] The spatial block size. Supported block sizes: 2, 3 and 4.

Return values:
  • NVCV_ERROR_INVALID_ARGUMENT – The possible cases might be the following: 1) handle or tensorRequirements is NULL, or points to an invalid address. 2) Block size is not 2, 3 or 4. 3) The channel count is not a multiple of (blkSz)^2. 4) The tensor does not have rank 3, is not in CHW layout, has a non-positive extent, or its data type is not 8bit signed or unsigned. 5) The tensor violates a documented limitation above: overlapping regions, a DMA field exceeded, or a shape or stride outside 32-bit range. 6) The output tensor requirements could not be derived from the input tensor and the block size.

  • NVCV_ERROR_OUT_OF_MEMORY – Not enough memory to create the operator.

  • NVCV_ERROR_INTERNAL – Internal error in the operator, invalid types passed in.

  • NVCV_SUCCESS – Operation executed successfully.

NVCVStatus pvaDepthToSpaceSubmit(
NVCVOperatorHandle handle,
cupvaStream_t stream,
NVCVTensorHandle in,
NVCVTensorHandle out,
)#

Submits the DepthToSpace operator to a cuPVA stream.

Parameters:
  • handle – [in] Handle to the operator.

    • Must not be NULL.

  • stream – [in] Handle to a valid cuPVA stream.

  • in – [in] Input tensor handle. Must match the tensor requirements passed to pvaDepthToSpaceCreate().

    • Must not be NULL.

  • out – [out] Output tensor handle. Must match the output tensor requirements derived by pvaDepthToSpaceCreate().

    • Must not be NULL.

Return values:
  • NVCV_ERROR_INVALID_ARGUMENT – The possible cases might be the following: 1) handle, stream, in or out is NULL, or points to an invalid address. 2) The input or output tensor does not meet the requirements used to create the operator handle.

  • NVCV_ERROR_INTERNAL – Internal error in the operator, invalid types passed in.

  • NVCV_SUCCESS – Operation executed successfully.

NVCVStatus pvaDepthToSpaceSubmit(
NVCVOperatorHandle handle,
cudaStream_t stream,
NVCVTensorHandle in,
NVCVTensorHandle out,
)#

Submits the DepthToSpace operator to a CUDA stream.

Note

CUDA stream support requirements:

  • PVA SDK 2.7.0 or later

  • Jetpack 7 or later

  • DriveOS 7 or later

  • x86 Emulator is not supported

Parameters:
  • handle – [in] Handle to the operator.

  • stream – [in] Handle to a valid CUDA stream.

  • in – [in] Input tensor handle.

  • out – [out] Output tensor handle.

Return values:
  • NVCV_ERROR_INVALID_ARGUMENT – Some parameter is outside valid range.

  • NVCV_ERROR_INTERNAL – Internal error in the operator, invalid types passed in.

  • NVCV_SUCCESS – Operation executed successfully.