Intel® oneAPI Deep Neural Network Developer Guide and Reference
A newer version of this document is available. Customers should click here to go to the newest version.
Visible to Intel only — GUID: GUID-68B5107E-7757-46F4-B74D-DB0D11204676
Visible to Intel only — GUID: GUID-68B5107E-7757-46F4-B74D-DB0D11204676
enum dnnl::accumulation_mode
Overview
Accumulation mode. More…
#include <dnnl_common.hpp> enum accumulation_mode { strict = dnnl_accumulation_mode_strict, relaxed = dnnl_accumulation_mode_relaxed, any = dnnl_accumulation_mode_any, s32 = dnnl_accumulation_mode_s32, f32 = dnnl_accumulation_mode_f32, f16 = dnnl_accumulation_mode_f16, };
Detailed Documentation
Accumulation mode.
Enum Values
strict
Default behavior, f32 for floating point computation, s32 for integer.
relaxed
same as strict except some partial accumulators can be rounded to src/dst datatype in memory.
any
uses fastest implementation, could use src/dst datatype or wider datatype for accumulators
s32
use s32 accumulators during computation
f32
use f32 accumulators during computation
f16
use f16 accumulators during computation