Math Kernel Library Developer Guide
Intel® oneAPI Math Kernel Library provides a conditional numerical reproducibility (CNR) functionality that enables you to obtain reproducible results from oneMKL routines. When enabling CNR, you choose a specific code branch of Intel® oneAPI Math Kernel Library that corresponds to the instruction set architecture (ISA) that you target. You can specify the code branch and other CNR options using theMKL_CBWR environment variable.
MKL_CBWR="<branch>[,STRICT]" or
MKL_CBWR="BRANCH=<branch>[,STRICT]"
Use the STRICT flag to enable strict CNR mode. For more information, see Reproducibility Conditions.
The <branch> placeholder specifies the CNR branch with one of the following values:
Value |
Description |
|---|---|
AUTO |
CNR mode uses the standard ISA-based dispatching model while ensuring fixed cache sizes, deterministic reductions, and static scheduling |
COMPATIBLE |
Intel® Streaming SIMD Extensions 2 (Intel® SSE2) without rcpps/rsqrtps instructions |
SSE2 |
Intel SSE2 |
SSE3 |
DEPRECATED. Intel® Streaming SIMD Extensions 3 (Intel® SSE3). This setting is kept for backward compatibility and is equivalent to SSE2. |
SSSE3 |
Supplemental Streaming SIMD Extensions 3 (SSSE3) |
SSE4_2 |
Intel® Streaming SIMD Extensions 4.2 (Intel® SSE4.2) |
AVX |
Intel® Advanced Vector Extensions (Intel® AVX) |
AVX2 |
Intel® Advanced Vector Extensions 2 (Intel® AVX2) |
AVX512 |
Intel AVX-512 on Intel® Xeon® processors |
AVX512_E1 |
Intel® Advanced Vector Extensions 512 (Intel® AVX-512) with support for Vector Neural Network Instructions |
AVX512_MIC |
DEPRECATED. Intel® Advanced Vector Extensions 512 (Intel® AVX-512) on Intel® Xeon Phi™ processors. This setting is kept for backward compatibility and is equivalent to AVX2. |
AVX512_MIC_E1 |
DEPRECATED. Intel® Advanced Vector Extensions 512 (Intel® AVX-512) with support for Vector Neural Network Instructions on Intel® Xeon Phi™ processors. This setting is kept for backward compatibility and is equivalent to AVX2. |
When specifying the CNR branch, be aware of the following:
Reproducible results are provided under Reproducibility Conditions.
Settings other than AUTO or COMPATIBLE are available only for Intel processors.
To get the CNR branch optimized for the processor where your program is currently running, choose the value of AUTO or call the mkl_cbwr_get_auto_branch function.
Strict CNR mode is supported only for AVX2, AVX512, AVX512_E1, AVX512_MIC, and AVX512_MIC_E1 branches. You can also use strict CNR mode with the AUTO branch when running on Intel processors that support one of these instruction set architectures (ISAs).
Setting the MKL_CBWR environment variable or a call to an equivalent mkl_cbwr_set function fixes the code branch and sets the reproducibility mode.
If the value of the branch is incorrect or your processor or operating system does not support the specified ISA, CNR ignores this value and uses the AUTO branch without providing any warning messages.
Calls to functions that define the behavior of CNR must precede any of the math library functions that they control.
Settings specified by the functions take precedence over the settings specified by the environment variable.
See the Intel® oneAPI Math Kernel Library Developer Reference for how to specify the branches using functions.
Product and Performance Information |
|---|
Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex. Notice revision #20201201 |