oneDNN/tests/benchdnn/doc/driver_eltwise.md

Element-wise Driver

Usage

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

where eltwise-knobs are:

• --dir={FWD_D [default], BWD_D} -- dnnl_prop_kind_t. Refer to direction for details.
• --dt={f32 [default], bf16, f16, s32, s8} -- src and dst data type. Refer to data types for details.
• --tag={nchw [default], ...} -- physical src and dst memory layout. Refer to tags for details.
• --alg={RELU [default], ...} -- dnnl_eltwise algorithm. Refer to eltwise primitive for details.
• --alpha=FLOAT -- float value corresponding to algorithm operation. Refer to Floating point arguments below.
• --beta=FLOAT -- float value corresponding to algorithm operation. Refer to Floating point arguments below.
• --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.
• --inplace=BOOL -- memory mode for the primitive. If true, it uses input memory as output, otherwise, input and output are separate. The default is false.
• --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 eltwise-desc is a problem descriptor. The canonical form is:

    NxNxNxNxN

where N is an integer number. This represents a 3D spatial problem with the following logical dimensions: N, C, D, H, W. Consider removing each xN from the end to specify fewer dimensions.

Floating point arguments

Some operations support alpha argument. Some also support beta argument. Refer to a full list for details.

The alpha and beta parameters should meet algorithm requirements, otherwise the problem will be silently skipped. For instance:

• An algorithm does not use alpha, but non-zero value is used;
• An algorithm does not use beta, but non-zero value is used;
• An algorithm expects alpha is non-negative, but negative value is used;
• An algorithm expects alpha is less than or equal to beta, but alpha value greater than beta is used;

This behavior allows using , (comma) operator to run multiple configurations without dealing with the corner cases.

The default set for alpha and beta is {0, 0.25, -0.25}.

Essence of Testing

Fill input data in four ranges: positive/negative integers up to 10 and positive/negative fractions up to 1.0. This covers special areas of all algorithm kinds. There is a general threshold; however, it cannot be applied everywhere. That is why there are some special cases. For details, refer to eltwise/eltwise.cpp::compare().

Examples

Run the eltwise set from an input file with the default settings:

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

Run a specific eltwise problem with the f32 data type and in-place memory mode, iterating over memory layouts and forward and backward prop kinds:

    ./benchdnn --eltwise --dir=FWD_D,BWD_D --dt=f32 --tag=nchw,nChw16c \
               --inplace=true 50x192x55x55

More examples with different driver options can be found at inputs/eltwise/test_eltwise_all. Examples with different benchdnn common options can be found at driver_conv.md.