Data types#

oneDNN supports multiple data types. However, the 32-bit IEEE single-precision floating-point data type is the fundamental type in oneDNN. It is the only data type that must be supported by an implementation. All the other types discussed below are optional.

Primitives operating on the single-precision floating-point data type consume data, produce, and store intermediate results using the same data type.

Moreover, single-precision floating-point data type is often used for intermediate results in the mixed precision computations because it provides better accuracy. For example, the elementwise primitive and elementwise post-ops always use it internally.

Note

Implicit downconversion can be enabled in order to speedup computations, and are controlled using the fpmath mode controls

oneDNN uses the following enumeration to refer to data types it supports:

enum class dnnl::memory::data_type#

Data type specification.

Values:

enumerator undef#

Undefined data type (used for empty memory descriptors).

enumerator f16#

16-bit/half-precision floating point.

enumerator bf16#

non-standard 16-bit floating point with 7-bit mantissa.

enumerator f32#

32-bit/single-precision floating point.

enumerator s32#

32-bit signed integer.

enumerator s8#

8-bit signed integer.

enumerator u8#

8-bit unsigned integer.

oneDNN supports training and inference with the following data types:

Usage mode

Data types

inference

dnnl::memory::data_type::f32, dnnl::memory::data_type::bf16, dnnl::memory::data_type::f16, dnnl::memory::data_type::s8/dnnl::memory::data_type::u8

training

dnnl::memory::data_type::f32, dnnl::memory::data_type::bf16

Note

Using lower precision arithmetic may require changes in the deep learning model implementation.

Individual primitives may have additional limitations with respect to data type support based on the precision requirements. The list of data types supported by each primitive is included in the corresponding sections of the specification guide.