benchdnn: brgemm: support f16 b matrix vnni layout for non-amx

tests/benchdnn/brgemm/brgemm.cpp

@@ -142,6 +142,7 @@ dnnl_status_t brgemm_attr_init(
         // PROCESS_KEY_VAL(bd_mask_level);
         PROCESS_KEY_VAL(use_uker);
         PROCESS_KEY_VAL(use_interleave_stores);
+        PROCESS_KEY_VAL(b_is_vnni);
         PROCESS_KEY_VAL(postops_only);
         PROCESS_KEY_VAL(hint_bd_block);
         PROCESS_KEY_VAL(hint_bd_block2);

tests/benchdnn/inputs/brgemm/test_brgemm_f16

@@ -1,27 +1,58 @@
 --reset
+--dt=f16,f16:f16:f32,f16:f32:f32,f32:f16:f32
+--bia_dt=undef,f32,f16
+--beta=0,1
+--attr-post-ops=,sum:2,relu
+--brgemm-attr=b_is_vnni:0
+# f16 uses f32 blocking for avx512_core_fp16
+--batch=option_set_f32
 
+--reset
 --dt=f16,f16:f16:f32,f16:f32:f32,f32:f16:f32
 --bia_dt=undef,f32,f16
 --beta=0,1
 --attr-post-ops=,sum:2,relu
---batch=option_set_f32 # f16 uses f32 blocking for avx512_core_fp16
+--brgemm-attr=b_is_vnni:1
+# f16 with b_is_vnni=1 uses vnni blocking for avx512_core_fp16
+--batch=option_set_bf16
 
 # Separate cases for non-default alpha
 --reset
 --dt=f16
 --alpha=2
+--brgemm-attr=b_is_vnni:0
 --batch=shapes_2d_no_tail_f32
 
+--reset
+--dt=f16
+--alpha=2
+--brgemm-attr=b_is_vnni:1
+--batch=shapes_2d_no_tail_bf16
+
 # Skip-acc feature
 --reset
---brgemm-attr=generate_skip_accumulation:1
+--brgemm-attr=generate_skip_accumulation:1+b_is_vnni:0
 --beta=0,1
 --bia_dt=f32
 --dt=f16
 --batch=option_set_f32
 
+--reset
+--brgemm-attr=generate_skip_accumulation:1+b_is_vnni:1
+--beta=0,1
+--bia_dt=f32
+--dt=f16
+--batch=option_set_bf16
+
 # ukernel wtag support
 --reset
 --wtag=ba
 --dt=f16
+--brgemm-attr=b_is_vnni:0
 --batch=option_set_f32
+
+--reset
+--wtag=ba
+--dt=f16
+--brgemm-attr=b_is_vnni:1
+--batch=option_set_bf16