# f32
--reset

--inplace=true,false
--skip-impl="ref:any"

--dir=FWD_D,BWD_D
--dt=f32
--tag=abx,axb,aBx8b,aBx16b
--alg=relu,tanh,elu,square,abs,sqrt,linear,brelu,srelu,logistic,exp,gelu_tanh,swish,log,clip,gelu_erf
        4x8x3x3 3x7x4x5 2x16x6x2 3x19x1x2 2x16x6x2x8 3x15x5x2x3
--alg=relu_dst,tanh_dst,elu_dst,sqrt_dst,logistic_dst,exp_dst
        4x8x3x3 3x7x4x5 2x16x6x2 3x19x1x2 2x16x6x2x8 3x15x5x2x3
--alg=pow --beta=-1,0,0.5,1,1.5,2   # various beta for all injector branches
        4x8x3x3 3x7x4x5 2x16x6x2 3x19x1x2 2x16x6x2x8 3x15x5x2x3

# s32, s8
--reset
--inplace=true,false
--skip-impl="ref:any"

--dir=FWD_I
--dt=s32,s8
--tag=axb
--alg=relu --alpha=0
        4x8x3x3 3x7x4x5 2x16x6x2 3x19x1x2 2x16x6x2x8 3x15x5x2x3

# bf16
--batch=test_eltwise_bfloat16
