add HAWQ

[Jiang-Stan]

• W: per-channel-symmetric with minmax observer
• A: per-tensor-symmetric with minmax observer
• Layers not searched are set to 8w8f
• eval data num: 5000
• mix-precision metric: greedy by output mse
• calibration batch size: 64
• Matmul input_quantizer bits are set to same for supporting ILP search
• For DeiT, First and last layers are set to 8w8f

Model	cfg	top1	top5	remark
ResNet18	8w8f	69.82	89.4
ResNet18	5w8f	65.5	86.38
ResNet18	5w8f greedy	69.06	88.42	w bit in [2~8]
ResNet18	5w8f hawqv3_modified	68.76	88.60	w bit in [2~8]
ResNet18	5w8f greedy	66.2	87.16	w bit in [2,4,8]
ResNet18	5w8f hawqv3_modified	66.2	87.16	w bit in [2,4,8]
ResNet18	8w6f	67.6	87.96
ResNet18	8w6f greedy	68.84	88.14	a bit in [2~8]
ResNet18	8w6f hawqv3_modified	68.82	88.50	a bit in [2~8]
ResNet18	8w6f greedy	56.14	78.3	a bit in [2,4,8]
ResNet18	8w6f hawqv3_modified	48.76	73.58	a bit in [2,4,8]
DeiT-tiny	8w8f	71.66	90.32
DeiT-tiny	5w8f	70.46	89.98
DeiT-tiny	5w8f greedy	70.82	90.06	w bit in [2~8]
DeiT-tiny	5w8f greedy	69.44	89.36	w bit in [2,4,8]
DeiT-tiny	8w6f	33.62	54.20
DeiT-tiny	8w6f greedy	70.34	89.92	a bit in [2~8]
DeiT-tiny	8w6f hawqv1	68.4	88.72	a bit in [2~8]
DeiT-tiny	8w6f greedy	62.72	85.10	a bit in [2,4,8]
DeiT-tiny	8w6f hawqv1	64.34	86.18	a bit in [2,4,8]
DeiT-small	8w8f	79.74	95.26
DeiT-small	8w6f	42.00	62.80
DeiT-small	8w6f greedy	76.86	93.9	a bit in [2~8]
DeiT-small	8w6f hawqv1	77.52	93.68	a bit in [2~8]
DeiT-small	8w6f greedy	68.56	88.62	a bit in [2,4,8]
DeiT-small	8w6f hawqv1	58.52	82.94	a bit in [2,4,8]

HAWQv2 R50 Sensitivity: