frozenleaves/oneDNN
1
0
Fork 0
mirror of https://github.com/uxlfoundation/oneDNN.git synced 2025-10-20 18:43:49 +08:00
Code Issues Packages Projects Releases Wiki Activity
Files
main
oneDNN/doc/programming_model/attributes_accumulation_mode.md
Maria Zhukova d559e96879 doc: minor renaming for nicer displayed list
2025-10-03 11:44:28 -07:00

1.2 KiB
Raw Permalink Blame History

Accumulation Mode

Some applications can benefit from using lower precision accumulators to speed up computations without causing noticeable impacts on accuracy.

The default numerical behavior of oneDNN (described in @ref dev_guide_data_types) can be altered to allow the use of low precision accumulators. When passed to a primitive creation, the @ref dnnl::accumulation_mode primitive attribute specifies which datatype can be used for accumulation purposes for that given primitive.

The @ref dnnl::accumulation_mode primitive attribute accepts:

  • strict (default): For floating-point primitives (as defined in @ref dev_guide_data_types), the default accumulation datatype is f32 (or f64 for f64 primitives). For integral primitives (as defined in @ref dev_guide_data_types), the default accumulation datatype is s32.
  • relaxed: Same as strict except some partial accumulators can be rounded to the src/dst datatype in memory.
  • any: Uses the fastest implementation available with one of the src/dst datatypes or a higher precision accumulation datatype.
  • f32, f16 and s32: Uses the specified accumulation datatype.