oneAPI Deep Neural Network Library (oneDNN)
Performance library for Deep Learning
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Groups Pages

For performance reasons, extra hints may be provided to oneDNN which enable the just-in-time (JIT) code generation to prefer or avoid certain CPU ISA features.

Build-time Controls

At build-time, support for this feature is controlled via cmake option DNNL_ENABLE_CPU_ISA_HINTS.

CMake Option Supported values (defaults in bold) Description

This results in making oneDNN aware of DNNL_CPU_ISA_HINTS environment variable and corresponding setter utility namely, dnnl::set_cpu_isa_hints routine.

Run-time Controls

During run-time the DNNL_CPU_ISA_HINTS environment variable can be used to specify ISA specific hint to enable oneDNN to dispatch appropriate kernel.

Environment variable Value Description
DNNL_CPU_ISA_HINTS NO_HINTS Use default configuration for ISA
PREFER_YMM Prefer to use YMM registers for vector operations

This feature can also be managed at run-time with the following functions:

  • dnnl::set_cpu_isa_hints function allows changing the CPU ISA hint at run-time. The limitation is that, it is possible to set the value only before the first JIT-ed function is generated - that is, before the primitive creation stage.
  • dnnl::get_cpu_isa_hints function returns the currently used CPU ISA hints that by default poses no restrictions.

Function settings take precedence over environment variables. Moreover, if the hint is not applicable then it would be silently ignored. For instance with SSE41 there are no YMM registers and so hint to prefer YMM registers would be silently bypassed.