some kernels and related api update

example/ck_tile/01_fmha/script/smoke_test_bwd_xqa_ext.sh → example/ck_tile/01_fmha/script/smoke_test_bwd_v3_xqa.sh

@@ -11,13 +11,17 @@ set -x
 for prec in "fp16" "bf16" ; do
 for perm in 0 1 ; do
 for hdim in 128 ; do
+for v3_atomic_fp32 in 0 1 ; do
+for v3_rtz_cvt in 0 1 ; do
 for mask in 0 1 ; do
 
-$EXE -prec=$prec -b=2 -h=4 -h_k=2 -d=$hdim -s=512 -iperm=$perm -operm=$perm -mask=$mask -bwd_v3=1 -v3_atomic_fp32=0 -v3_rtz_cvt=1 -mode=0 -kname=$KNAME $COMMON_ARGS
-$EXE -prec=$prec -b=1 -h=3 -h_k=1 -d=$hdim -s=768 -iperm=$perm -operm=$perm -mask=$mask -bwd_v3=1 -v3_atomic_fp32=0 -v3_rtz_cvt=1 -mode=0 -kname=$KNAME $COMMON_ARGS
+$EXE -prec=$prec -b=2 -h=4 -h_k=2 -d=$hdim -s=512 -iperm=$perm -operm=$perm -mask=$mask -bwd_v3=1 -v3_atomic_fp32=$v3_atomic_fp32 -v3_rtz_cvt=$v3_rtz_cvt -mode=0 -kname=$KNAME $COMMON_ARGS
+$EXE -prec=$prec -b=1 -h=3 -h_k=1 -d=$hdim -s=768 -iperm=$perm -operm=$perm -mask=$mask -bwd_v3=1 -v3_atomic_fp32=$v3_atomic_fp32 -v3_rtz_cvt=$v3_rtz_cvt -mode=0 -kname=$KNAME $COMMON_ARGS
 
 done
 done
 done
 done
+done
+done
 set +x