# bfloat16
--reset
--mb=2
--dir=FWD_B
--cfg=bf16bf16f32  --batch=shapes_resnet_50
--cfg=bf16bf16bf16 --batch=set_conv_all

--dir=FWD_D
--cfg=bf16bf16bf16 --batch=shapes_resnet_50 --batch=shapes_regression_1x1

--dir=BWD_D
--cfg=f32bf16bf16  --batch=shapes_resnet_50 --batch=shapes_regression_1x1
--cfg=bf16bf16bf16 --batch=set_conv_all

--dir=BWD_WB
--cfg=bf16f32bf16
--batch=set_conv_all --batch=set_dilated-conv --batch=shapes_regression_1x1
--cfg=bf16bf16bf16 --batch=shapes_resnet_50 --batch=set_dilated-conv

--reset
--mb=2
--skip-impl="ref:gemm"
--dir=FWD_D --cfg=bf16bf16bf16
--batch=shapes_3d_2d_strided_padding --batch=shapes_dilated_3d_strided_padding
--dir=BWD_D --cfg=f32bf16bf16 --batch=shapes_3d_2d_strided_padding
--dir=BWD_WB --cfg=bf16f32bf16
--batch=set_conv_3d --batch=shapes_dilated_3d_unit-stride_no-padding

# Attributes
--reset
--mb=2
--skip-impl="ref:gemm"
--dir=FWD_B
--cfg=bf16bf16bf16
## Post-Ops
--attr-post-ops='sum' --batch=shapes_resnet_50
--attr-post-ops='relu' --batch=shapes_googlenet_v3
--attr-post-ops='add:bf16:per_oc' --batch=shapes_tails
--attr-post-ops='mul:f32;sum;tanh:1:1:2.5' --batch=shapes_tails

## Depthwise fusion stride 1
--attr-post-ops='dw_k3s1p1:bf16;relu'
--batch=shapes_fused_mobilenet_stride_1
--attr-post-ops='relu:0.5;dw_k3s1p1:f32;tanh:0:0:2.5'
--batch=shapes_fused_mobilenet_stride_1

## Depthwise fusion stride 2
--attr-post-ops='relu;dw_k3s2p1:bf16;tanh'
--batch=shapes_fused_mobilenet_stride_2
--attr-post-ops='relu;dw_k3s2p1:f32;tanh'
--batch=shapes_fused_mobilenet_stride_2

--batch=harness_conv_dw_bfloat16

# Test src-transpose padding handling in bf16 bwd-w convolution
--reset --mb=2
--skip-impl="ref:gemm"
--cfg=bf16bf16bf16 --dir=BWD_W
--batch=shapes_src-transpose_padding

# Test plain and blocked layout in non-1x1 jit kernels
# (amx non-1x1 kernel supports both, so need to always test both)
--reset
--mb=2
--skip-impl=ref:gemm
--dir=FWD_B
--cfg=bf16bf16bf16
--match=.*k[dhw][^1].*

--stag=axb --dtag=axb
--batch=shapes_1d --batch=shapes_googlenet_v1 --batch=shapes_resnet_50

--stag=aBx16b --dtag=aBx16b
--batch=shapes_1d --batch=shapes_googlenet_v1 --batch=shapes_resnet_50
