oneDNN/tests/benchdnn/doc/driver_lrn.md

Local Response Normalization Driver

Usage

    ./benchdnn --lrn [benchdnn-knobs] [lrn-knobs] [lrn-desc] ...

where lrn-knobs are:

• --dir={FWD_D [default], BWD_D} -- dnnl_prop_kind_t. Refer to direction for details.
• --dt={f32 [default], bf16, f16} -- src and dst data types. Refer to data types for details.
• --tag={nchw [default], ...} -- physical src and dst memory layout. Refer to tags for details.
• --alg={ACROSS [default], WITHIN} -- lrn algorithm. ACROSS is dnnl_lrn_across_channels; WITHIN is dnnl_lrn_within_channel; Refer to LRN primitive for details.
• --mb=INT -- override minibatch size specified in the problem description. When set to 0, use minibatch size as defined by the individual problem descriptor. The default is 0.
• --match=REGEX -- skip problems not matching the regular expression in REGEX. By default no pattern is applied (run everything). Note: Windows may interpret only string arguments surrounded by double quotation marks.
• Any attributes options. Refer to attributes for details.

and lrn-desc is a problem descriptor. The canonical form is:

    mbXicX_idXihXiwX_lsX_alphaF_betaF_kF_nS

Refer to descriptor for details. lsX stands for LRN local size value and accepts integer X values. The default is 5. alphaF stands for LRN alpha scale and accepts float F values. The default is 1.f / 8192. betaF stands for LRN beta power and accepts float F values. The default is 0.75f. kF stands for LRN k shift and accept float F values. The default is 1.f. Refer to LRN primitive for details.

Essence of Testing

Fill input data with integers so that an output will not overflow in the f16 or bf16 data types.

Examples

Run a set of lrns from an input file with the default settings:

    ./benchdnn --lrn --batch=inputs/lrn/shapes_ci

Run a named problem with single precision src/dst, iterating by:

1. memory layouts, plain and blocked, where channel blocking equals 8 and 16,
2. both forward training and backward by data,
3. both algorithms supported:

    ./benchdnn --lrn --dt=f32 --tag=nchw,nChw8c,nChw16c \
               --dir=FWD_D,BWD_D --alg=ACROSS,WITHIN \
               mb256ic96_ih55n"alexnet:norm1"

More examples with different driver options can be found at inputs/lrn/test_*. Examples with different problem descriptors can be found at inputs/lrn/shapes_*. Examples with different benchdnn common options can be found at driver_conv.md.