Pdist

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: the Pdist operator computes the p-norm distances (pairwise p-norm distances) between each pair of row vectors. The output is a one-dimensional vector in upper-triangular row-major order. It is equivalent to torch.nn.functional.pdist.

• Kernel implementation form: PyPTO DSL (op_kernel/pdist.py). Code generation is performed at the configure stage through enable_pypto_kernel, and the result is delivered as a pypto opc (MIX launch, the AIC stays idle, and all computation is performed on the Vector cores).

• Computation formula:

$$y_{idx}={\Vert x_i-x_j\Vert }_p={\left(\sum _{k=0}^{M-1}{\left|x_{i,k}-x_{j,k}\right|}^p\right)}^{1/p},idx=\frac{i\cdot (2N-i-1)}{2}+(j-i-1)$$

$$0\le i<j\le N-1$$

p branch semantics:

p Value	Semantics
p = 0	Counts the number of unequal dimensions (Hamming distance). NaN dimensions are counted as unequal
p = +inf	Maximum absolute difference (Chebyshev distance). NaN dimensions are ignored
0 < p < +inf	The 1/p-th power of the sum of powers (Minkowski distance). p=2 is the Euclidean distance and p=1 is the Manhattan distance

NaN semantics (aligned with GPU/CUDA): when p=0, NaN dimensions are counted as unequal; when p=+inf, NaN dimensions are ignored (the maximum absolute difference of non-NaN dimensions is output, and all-NaN row pairs output 0); for a general p (not 0/0.5/1/2), NaN dimensions are detected pair by pair and NaN is output; for p∈{0.5, 1, 2}, NaN propagates naturally through IEEE arithmetic.

fp16 input: on-chip intermediate accumulation always uses FP32, and the result is narrowed back to FP16 at the output.

Parameter Description

Parameter Name	Input/Output/Attribute	Description	Data Type	Data Format
x	Input	Input tensor, with the shape (N, M), where N≥2 and M≥1. Row i and row j (j>i) form a row pair for computation. Row-major contiguous layout is required.	FLOAT16, FLOAT32	ND
p	Attribute	Norm order, which is p≥0 or +inf. The default value is 2.0 (Euclidean distance). p=0 counts the number of unequal dimensions, p=+inf takes the maximum absolute difference, and 0	FLOAT	-
y	Output	Pairwise distance vector, with the shape (N*(N-1)/2,), arranged in upper-triangular row-major order. The dtype is the same as that of x.	FLOAT16, FLOAT32	ND

Constraint Description

• x must be a two-dimensional (N, M) row-major contiguous tensor, and the dtype of y must be the same as that of x.
• The shape of y must be (N*(N-1)/2,).
• Intermediate accumulation of fp16 input uses FP32, and the result is narrowed back to FP16 at the output.
• Platform differences:
  • When p=+inf and the input contains NaN: on Ascend 950PR/950DT, NaN dimensions are ignored (the maximum absolute difference of non-NaN dimensions is output, and all-NaN row pairs output 0), which is aligned with the GPU/CUDA semantics. On other platforms, NaN is propagated to the output.

Specification Item	Specification	Specification Description
N	2~65535	Number of rows, at least 2 (at least one row pair).
M	1~65535	Number of dimensions per row. When it exceeds the on-chip memory, automatic chunked reduction is performed.
p	[0, +inf)	p≥0 or +inf; p=0 is the Hamming distance, p=+inf is the Chebyshev distance, and p<0 or NaN is rejected as a parameter error.
dtype	FLOAT16/FLOAT32	The input and output have the same dtype; fp16 intermediate computation uses FP32 accumulation.

Invocation Description

Invocation Method	Invocation Sample	Description
aclnn invocation	test_aclnn_pdist	Invokes the pdist operator through the aclnnPdist API, covering all p branches of fp32, multi-M chunking, and fp16.
aclnn invocation (Forward)	test_aclnn_pdist_forward	Invokes the operator through the aclnnPdistForward API (the composite scalar p entry).
GEIR invocation	test_geir_pdist	Invokes the pdist operator in GE IR graph mode (RunGraph).