Skip to content

Conversion Sync Task

The Conversion sync task changes the dtype of a tensor on the GPU. The configured target dtype and strategy determine how samples are interpreted or rescaled before being written to the output tensor.

Supported combinations are:

  • uint8 -> complex32 with the Real strategy (real part copied, imaginary part set to zero)
  • uint16 -> complex32 with the Real strategy
  • float32 -> uint8 with the Scaled strategy
  • float32 -> uint16 with the Scaled strategy
  • complex32 -> float32 with the Modulus strategy
  • complex32 -> float32 with the Argument strategy

The Scaled strategy computes the minimum and maximum value of the input tensor, then linearly maps the range to fill the target integer interval by rescaling the (min, max) interval to the min and max of the target interval. The Modulus and Argument strategies operate element-wise on complex tensors, returning magnitude or phase respectively.

Warning

When using the Scaled strategy, constant inputs (where min == max) lead to an undefined scale factor. Make sure the source tensor spans a non-zero range.

Inputs

This task consumes a single input tensor of any shape.

The dtype of the input tensor must be one of:

  • uint8
  • uint16
  • float32
  • complex32

The memory location of the input tensor must be device memory.

Outputs

This task produces one output tensor with the same shape as the input.

The dtype of the output tensor is determined by the configured target dtype.

The memory location of the output tensor is device memory.

Inplace

This task does not support inplace operation.

Ownership

This task does not own any inputs or outputs.

Settings

Example:

1
2
3
4
{
    "target": "F32",
    "strategy": "Modulus"
}
Property Pattern Type Deprecated Definition Title/Description
+ target No enum (of string) No - Desired output data type. Accepted values:
- U8: 8-bit unsigned integer
- U16: 16-bit unsigned integer
- F32: 32-bit float
- CF32: 32-bit complex float
+ strategy No enum (of string) No - Conversion strategy to apply. Accepted values:
- Real: Extract real component (valid when target is CF32 and the input is U8/U16)
- Scaled: Normalize float input to integer range (valid when target is U8 or U16 and the input is F32)
- Modulus: Emit magnitude (valid when target is F32 and the input is CF32)
- Argument: Emit phase angle (valid when target is F32 and the input is CF32)