AdaCast

English | 简体中文

Product Support Status

Product	Supported
Ascend 950PR&950DT products	√
Atlas A3 products	√
Atlas A2 products	√
Atlas 200I/500 A2 inference products	×
Atlas inference products	×
Atlas training products	×

Function Description

• Operator function: an input preprocessing operator for the HDRnet (High Dynamic Range Network) real-time image enhancement network. It converts the uint16 HDR image data output by the sensor into a float16 tensor and completes normalization based on the white level (pixel), which is equivalent to fusing the two steps of integer-to-float type conversion and scaling by the divisor into a single element-wise operation.
• Computation formula:

$$ou{t}_i={\mathrm{C}\mathrm{a}\mathrm{s}\mathrm{t}\_\mathrm{F}\mathrm{P}16}^{sat+round}(\mathrm{C}\mathrm{a}\mathrm{s}\mathrm{t}\_\mathrm{F}\mathrm{P}32(sel{f}_i)\times (1/pixel))$$

In the final step, saturation (clipping to the fp16 value range [-65504, 65504]) is performed first, and then the value is rounded and converted to float16 in round-to-nearest-even mode.

Parameter Description

Parameter Name	Input/Output/Attribute	Description	Data Type	Data Format
x	Input	The HDR uint16 image data to be normalized, which is self_i in the formula. The shape supports 1D to 4D.	UINT16	ND
y	Output	The normalized float16 tensor, which is out_i in the formula. Its shape is exactly the same as that of x.	FLOAT16	ND
pixel	Optional attribute	The white level normalization base, and the scaling factor is 1/pixel. The default value is 65535. The value must be a positive integer (pixel > 0). If the value is not positive, operator compilation reports an error.	INT64	-

Constraint Description

• Only the ND data format is supported.
• The rank of the input x ranges from [1, 4] (0-D scalars and rank greater than 4 are both rejected). The shape of the output y must be exactly the same as that of x (element-wise semantics, broadcast is not supported).
• The type combination is fixed to UINT16 → FLOAT16, and other dtype combinations are not supported.
• pixel is an optional attribute. When it is not passed in, the default value 65535 is used. The value must be a positive integer (pixel > 0). If the value is not positive, operator compilation reports an error.
• Empty tensors (with 0 elements) are supported. The host side returns early and does not dispatch the kernel.
• A deterministic implementation is used by default, and the same input always produces the same output.

Invocation Description

Invocation Method	Sample Code	Description
GE graph mode	test_geir_ada_cast.cpp	Invokes the AdaCast operator by building a graph through GE-IR (ge::Session / RunGraph). The operator IR prototype is defined in ada_cast_proto.h. The sample does not participate in the build by default (ENABLE_GEIR_EXAMPLE is not defined in the repository CMake). You need to explicitly compile the sample in the client project as described in the sample comments and link against libgraph / libge_runner.

Note: This operator provides only the GE graph mode API. In CMakeLists.txt, ACLNNTYPE aclnn_exclude is configured and no op_api/ directory exists. Therefore, the ACLNN and PyTorch/TensorFlow APIs are not provided.