# main op
--reset
--cfg=bf16bf16bf16
384x4096:4096x1024:384x1024
8x384x4096:1x4096x1024:8x384x1024
8x16x384x64:8x16x64x384:8x16x384x384

# post bias
--reset
--cfg=bf16bf16bf16
--bia_dt=bf16
384x4096:4096x1024:384x1024

# post bias + swish/relu6
--reset
--cfg=bf16bf16bf16
--bia_dt=bf16
--attr-post-ops=swish:1,clip_v2:0:6
384x4096:4096x1024:384x1024

# post eltwise (with and without post bias)
--reset
--cfg=bf16bf16bf16
--bia_dt=undef,bf16
--attr-post-ops=elu,relu,clip_v2:-1:1,logistic
# gelu_erf <- disabled due to accuracy issue
384x4096:4096x1024:384x1024

# post binary-add
--reset
--cfg=bf16bf16bf16
--attr-post-ops=add:bf16:common,add:bf16:per_oc
384x4096:4096x1024:384x1024

# post sum (with and without eltwise)
--reset
--cfg=bf16bf16bf16
--attr-post-ops=sum,sum+relu,sum+logistic
# sum+gelu_erf <- disabled due to accuracy issue
384x4096:4096x1024:384x1024

# post bias + sum (with and without eltwise)
--reset
--cfg=bf16bf16bf16
--bia_dt=bf16
--attr-post-ops=sum+relu,sum
384x4096:4096x1024:384x1024
