LogSpaceD

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

LogSpaceD receives the auxiliary tensor assist with a fixed shape and generates a logarithmically spaced sequence of length steps according to the TBE single-ended FP32 semantics. assist is fixed to rank-1 [1] and only serves as the ABI input slot and the dtype profile. Its values do not participate in the computation.

Dynamic capability: DynamicRankSupportFlag=true. During graph building, [-2] (unknown rank) can be declared, but the runtime does not relax the contract: assist.shape=[1], the output rank is fixed to 1, and the shape is derived from attributes as [steps].

Let n=steps. If n>1, ratio=(end-start)/(n-1); otherwise, ratio=end-start. All attribute conversions, interpolation, and subsequent mathematical operations are performed in FP32:

$$x_i=\mathrm{F}\mathrm{P}32(i)\times ratio+\mathrm{F}\mathrm{P}32(start),i=0,1,\dots ,steps-1.$$

The output adopts the TBE three-branch continuous semantics:

$$y_i^{(32)}=\{\begin{array}{ll}\exp (x_i\ln (base)),& base>0,\\ \left[1-2\left(|x_i|-2\lfloor |x_i|/2\rfloor \right)\right]\exp (x_i\ln |base|),& base<0,\\ 1,& base=0\wedge x_i=0,\\ 0,& base=0\wedge x_i\ne 0.\end{array}$$

Both dtype=0/1 are compatibility no-ops, and the output type is always the same as that of assist: FP16→FP16, FP32→FP32, and BF16→BF16. When steps=0, the input is still [1] and the output is [0]. When steps=1, the global index is 0, and the result is determined only by start.

Parameter Description

Parameter Name	Input/Output/Attribute	Description	Data Type	Data Format
assist	Input	Fixed rank-1 [1] carrier, whose values do not participate in the computation.	FLOAT16, FLOAT, BFLOAT16	ND
y	Output	Logarithmically spaced sequence with the shape [steps]. The data type is the same as that of assist, and the dtype is a compatibility no-op.	FLOAT16, FLOAT, BFLOAT16	ND
start	Attribute	Start of the exponent domain, required.	FLOAT	-
end	Attribute	End of the exponent domain, required.	FLOAT	-
steps	Optional attribute	Number of output elements. The default value is 100.	INT64	-
base	Optional attribute	Base of the power. The default value is 10.0.	FLOAT	-
dtype	Optional attribute	Historical compatibility attribute. Only 0 or 1 is allowed, and the default value is 1. Neither value changes the output type.	INT64	-

Constraint Description

• Graph building supports DynamicRankSupportFlag=true, and [-2] unknown rank can be declared. At runtime, assist must be [1], and y must be [steps]. Their shapes are not bound and are not reused in place.
• assist supports FLOAT16, FLOAT, and BFLOAT16 inputs, and the output type is the same as the input type. Together with dtype=0/1, they form six valid combinations.
• dtype allows only 0 or 1, both of which are compatibility no-ops and do not change the output type.
• FLOAT64 input, FLOAT64 output, and FLOAT64 mathematical operations are not supported. All combinations use the FP32 main computation chain, and the FLOAT16 output is converted only once at the end.
• start, end, and base must be finite FLOAT values. A negative base uses the continuous remainder sign branch, and a zero base uses the x_i==0 ? 1 : 0 branch. The special value semantics of the general Pow are not used.
• steps must be non-negative and only determines the output length. The input length is fixed to 1. Only the ND format is supported. Broadcasting, endpoint tensors, axis, and endpoint=False are not supported.
• Only the independent GEIR invocation path of the Ascend 950PR&950DT products is supported, and the algorithm Workspace is 0. The result of the ACLNN combination path may be a multi-dimensional tensor. The runtime assist constraint of the independent GEIR cannot be rewritten based on it.

Invocation Description

Invocation Method	Sample Code	Description
GE graph mode	-	For the graph IR definition, refer to the LogSpaceD prototype. The operator prototype is LogSpaceD(assist, start, end, steps=100, base=10.0, dtype=1) -> y.