xref: /aosp_15_r20/external/XNNPACK/doc/microkernel-naming-conventions.md (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1# Microkernel naming conventions
2
3This documents deciphers XNNPACK's microkernels naming convention.
4
5## General conventions
6
7Microkernel function names follow this convention:
8
9`xnn_<datatype>_<microkernel><activation?>_ukernel_<parameters>__<arch>`
10
11Where `<datatype>` can be:
12
13-   `cs16`
14-   `f16` - 16-bit half precision float
15-   `f32` - 32-bit single precision float
16-   `qc8`
17-   `qs8` - quantized signed 8 bit
18-   `qu8` - quantized unsigned 8 bit
19-   `s16`
20-   `u32`
21-   `x8`
22-   `x16`
23-   `x24`
24-   `x32`
25-   `xx`
26
27`<microkernel>` is the type of microkernel, such as:
28
29-   `gemm`
30-   `igemm`
31-   `avgpool`
32
33`<activation>` if supported for the microkernel is activation that is fused into
34the microkernel:
35
36-   `linear`
37-   `minmax`
38-   `relu`
39
40`<parameters>` are microkernel specific, and can mean different things depending
41on the microkernel (see below for details).
42
43`<arch>` is the architecture the microkernel is optimized for, and can contain
44further subdivisions for additional instruction sets supported on the specified
45architecture, or processor information:
46
47-   `scalar`
48-   `aarch32_neon_cortex_a55`
49-   `neonv8_mlal`
50-   `wasm`
51-   `avx512`
52-   `avx512skx`
53
54## GEMM and IGEMM microkernels
55
56The `<parameters>` for GEMM and IGEMM microkernels represent the `mr` and `nr`
57of the microkernel. You can think of it as the number of rows and columns of the
58output calculated by the microkernel.
59
60E.g. `xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a7` processes 32
61elements of the output matrix.
62
63## Average Pooling and Global Average Pooling
64
65These microkernels come in 2 varieties, uni-pass and multi-pass.
66
67Uni-pass have `Cx` in their name, where `C` is a number. This microkernel
68processes up to and including `C` elements.
69
70Multi-pass have `CpDx` in their name, where `C` and `D` are numbers. This
71microkernel processes `D` elements in the first pass, and middle pass (which can
72run multiple times), and up to `C` elements in the last pass.
73
74E.g. `xnn_f32_avgpool_minmax_ukernel_9x__neon_c4` can process up to 9 elements.
75