Skip to content

Latest commit

 

History

History
82 lines (71 loc) · 3.52 KB

driver_bnorm.md

File metadata and controls

82 lines (71 loc) · 3.52 KB
Raw

Batch Normalization Driver

Usage

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

where bnorm-knobs are:

  • --dir={FWD_D [default], FWD_I, BWD_D, BWD_DW} -- dnnl_prop_kind_t. Refer to direction for details.
  • --dt={f32 [default], s8, 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.
  • --strides=SRC_STRIDES:DST_STRIDES -- physical memory layout specification for src and dst tensors through strides values. Refer to option documentation for details.
  • --flags=[|G|C|H|R|A] -- batch normalization flags, default none; where multiple simultaneous flags are supported. G is dnnl_use_global_stats; C is dnnl_use_scale; H is dnnl_use_shift; R is dnnl_fuse_norm_relu; A is dnnl_fuse_norm_add_relu; Refer to batch normalization primitive for details.
  • --inplace=BOOL -- memory mode for the primitive. If true, it uses input memory as output, otherwise, input and output are separate. Default is false.
  • --debug-check-ws=BOOL -- checks if workspace has correct values. Feature is based on internal knowledge of a library implementation. The default is false.
  • --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 bnorm-desc is a problem descriptor. The canonical form is:

    mbXicX_idXihXiwX_epsF_nS

Refer to descriptor for details. epsF stands for Batch Normalization epsilon value and accepts float F values. The default is 1.f/16.

Essence of Testing

TBA.

Examples

Run a set of bnorms from an input file, using the default settings:

    ./benchdnn --bnorm --batch=inputs/bnorm/shapes_resnet_50

Run a named problem with single precision src/dst, skipping all problems that use the reference implementation, iterating by:

  1. blocked memory layouts, where channel blocking equals 8 and 16,
  2. forward training, backward by data and weights prop_kinds,
  3. some flag combinations:
    ./benchdnn --bnorm --dt=f32 --skip-impl=ref --tag=nChw8c,nChw16c \
               --dir=FWD_D,BWD_D,BWD_DW --flags=CHR,GCH,CH \
               mb96_ic192_ih71iw71_n"googlenet_v3:conv_4_4_batchnorm"

Run a set of 3D spatial bnorms with s8 src/dst data_type, inference prop_kind, plain physical memory layout with dense channels, and some flags specified:

    ./benchdnn --bnorm --dt=s8 --tag=ndhwc --dir=FWD_I \
               --flags=GCHR --batch=shapes_3d

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