[OpenSora-hpcai] OSv1.2 performance optimization

[hadipash]

TODO:

• Validate accuracy and visual quality on long training.
• Update performance tables in README.

Tests were conducted in dynamic DVM mode, on MS daily from 09.04 with CANN 8.0 RC2. Results include training step average time only (no data loading time):

Changes	Shape (res x frames x batch)	Time (s)	Change (s)	Comment
Original	720p x 51 x 2	30.409
	144p x 204 x 10	19.934
Switch to repeat_interleave_ext_v2	720p x 51 x 2	28.913	-1.496 (-4.9%)
	144p x 204 x 10	19.872	-0.062 (-0.3%)
Remove SiLU & GELU FP32 upcast	720p x 51 x 2	30.346	-0.062 (-0.2%)	No performance improvement,
	144p x 204 x 10	20.506	+0.572 (+2.9%)	will consult with the MS team.
Convert parameters to BF16	720p x 51 x 2	28.957	-1.452 (-4.8%)
	144p x 204 x 10	18.747	-1.187 (-3.9%)
Remove redundant ops.transpose in VAE	720p x 51 x 2	30.448	+0.040 (+0.1%)	No changes due to the kernel fusion.
	144p x 204 x 10	20.103	+0.168 (+0.8%)	Beneficial in KBK & PyNative modes.
Final improvement	720p x 51 x 2	27.896	-2.512 (-8.3%)
	144p x 204 x 10	18.804	-1.130 (-5.7%)

[zhtmike]

seems no code change for Convert parameters to BF16 ?

[hadipash]

seems no code change for Convert parameters to BF16 ?

This refers to the network parameters that are explicitly defined with nn.Parameter(), such as self.scale_shift_table. For some reason, any calculations performed on self.scale_shift_table are upcast to the parameter type (i.e. fp32) and the new type is propagated in the network, even with AMP enabled.