Skip to content

Latest commit

 

History

History
59 lines (48 loc) · 2.28 KB

driver_shuffle.md

File metadata and controls

59 lines (48 loc) · 2.28 KB
Raw

Shuffle Driver

Usage

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

where shuffle-knobs are:

  • --dir={FWD_D [default], BWD_D} -- dnnl_prop_kind_t. Refer to direction for details.
  • --dt={f32 [default], s32, s8, u8, bf16, f16} -- 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.
  • --axis=INT -- dimension on which operation will be performed, also called the shuffle axis. The default is 1 and corresponds to channels in logical memory layout.
  • --group=INT -- group size along the shuffle axis. Must evenly divide the shuffle axis. The default is 1.
  • --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 shuffle-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.

Essence of Testing

Fill input data with integers so that an output will not overflow in the f16 or bf16 data types. As no compute operations are executed, we set a zero threshold and expect a precise answer with the reference implementation.

Examples

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

    ./benchdnn --shuffle --batch=inputs/shuffle/test_shuffle_all

Run a specific shuffle problem with forward prop_kind and plain physical memory layout. Group elements by 4 and over h dimension, iterating by all listed data types:

    ./benchdnn --shuffle --dir=FWD_D --dt=f32,s32,s8,u8,bf16 \
               --tag=nchw --group=4 --axis=2 1x68x56x56

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