oneDNN/tests/benchdnn/doc/driver_pool.md

Pooling Driver

Usage

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

where pool-knobs are:

• --dir={FWD_D [default], FWD_I, BWD_D} -- dnnl_prop_kind_t. Refer to direction for details.
• --dt={f32:f32:f32 [default], ...} -- source, weights and destination data types. Interface supports broadcasting, when a single input is provided, e.g., --dt=f32, and the value will be applied for all tensors. Refer to data types for details.
• --cfg={f32 [default], ...} -- Deprecated setting. Refer to Configurations below.
• --tag={nchw [default], ...} -- physical src and dst memory layout. Refer to tags for details.
• --alg={max [default], avg_np, avg_p} -- pooling algorithm. max or pooling_max is dnnl_pooling_max; avg_np or pooling_avg_exclude_padding is dnnl_pooling_avg_exclude_padding; avg_p or pooling_avg_include_padding is dnnl_pooling_avg_include_padding; Refer to pooling 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 pool-desc is a problem descriptor. The canonical form is:

    mbXicX_idXihXiwX_odXohXowX_kdXkhXkwX_sdXshXswX_pdXphXpwX_ddXdhXdwX_nS

Refer to descriptor for details. Input shape and kernel size are mandatory inputs. Output shape and padding may be deduced based on the values provided.

Precision Configurations

--cfg option specifies what data types will be used for a problem. It is implicit for the integer type saturation. This option also defines the threshold for computation errors.

The table below shows supported name configurations for this driver:

src	dst	cfg	notes
f32	f32	f32	TBA.
s32	s32	s32
f16	f16	f16
bf16	bf16	bf16
s8	s8	s8
u8	u8	u8
s8	u8	s8u8	Only on GPU engine
u8	s8	u8s8	Only on GPU engine
s8	f16	s8f16	Only on GPU engine
f16	s8	f16s8	Only on GPU engine
u8	f16	u8f16	Only on GPU engine
f16	u8	f16u8	Only on GPU engine
s8	f32	s8f32	Only on GPU engine
f32	s8	f32s8	Only on GPU engine
u8	f32	u8f32	Only on GPU engine
f32	u8	f32u8	Only on GPU engine

Essence of Testing

max algorithm: Fill input data with integers and expect an integer answer. avg_p algorithm: Fill input data with integers, divisible by the kernel size, and expect an integer answer. avg_np algorithm: Fill input data with integers, divisible by the kernel size, but expect a float answer due to boarder points have different kernel shapes applied to the same point.

Examples

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

    ./benchdnn --pool --batch=inputs/pool/shapes_2d

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

1. both blocked memory layouts, where channel blocking equals 8 and 16,
2. both forward training, inference and backward by data prop_kind,
3. all algorithm combinations,
4. using default minibatch of 96 and 5:

    ./benchdnn --pool --dt=f32 --tag=nChw8c,nChw16c \
               --dir=FWD_D,FWD_I,BWD_D --alg=max,avg_np,avg_p --mb=0,5 \
               mb96ic768_ih17oh17_kh3sh1ph1n"googlenet_v3:ave_pool_mixed_4_pool"

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