NN_ROUND() issues in 16bit ops

[majianjia]

I met an issue of 16 bit NN producing bad results when I tried to use round() in NN calculation using ARM_NN_TRUNCATE (undefined ARM_NN_TRUNCATE)
I normally define ARM_NN_TRUNCATE but recently I want to try to use round so I undefined the macro. In 8 bit ops, it brings some accuracy improvement but in 16 bit it fails.

The 16bit ops I use is this one but other which uses NN_ROUND() have similar issues.

CMSIS_5/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15_opt.c

Line 117 in e42a76a

arm_fully_connected_mat_q7_vec_q15_opt(const q15_t * pV,

change #define NN_ROUND(out_shift) ( (0x1u << out_shift) >> 1 ) to
#define NN_ROUND(out_shift) ((0x1 << out_shift) >> 1 ) fixed my issue.

CMSIS_5/CMSIS/NN/Include/arm_nnsupportfunctions.h

Line 652 in e42a76a

#define NN_ROUND(out_shift) ( (0x1u << out_shift) >> 1 )

Also, it doesn't seem right when it adds to a negative number. Should the round be round down in negative number? In most ops, only add this to the bias which means it is always positive.

More information will be provided when I figure it out. Thanks.

[felix-johnny]

@majianjia Apologies for the delay. Would you mind making a PR for the fix?

[majianjia]

@felix-johnny let me remind myself what is this about.
Happy to, thanks :p

[felix-johnny]

Great , thanks!

[majianjia]

close the issue as the related PR is merged.
Thanks