37 .macro MASK_MATRIX arg0, arg1, arg2, arg3, arg4, arg5, arg6
42     cmhi    \arg4\().16b, \arg5\().16b, \arg4\().16b
46     cmhi    \arg4\().16b, \arg5\().16b, \arg4\().16b
50 .macro DIFF_LUMA_LT4_P1_Q1 arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8, arg9
58     smax    \arg8\().16b, \arg9\().16b, \arg5\().16b
69 .macro DIFF_LUMA_LT4_P0_Q0_1 arg0, arg1, arg2, arg3, arg4, arg5, arg6
70     usubl \arg5\().8h, \arg0\().8b, \arg3\().8b
73     add     \arg5\().8h, \arg5\().8h, \arg6\().8h
74     sqrshrn  \arg4\().8b, \arg5\().8h, #3
77 .macro DIFF_LUMA_LT4_P0_Q0_2 arg0, arg1, arg2, arg3, arg4, arg5, arg6
78     usubl2    \arg5\().8h, \arg0\().16b, \arg3\().16b
81     add     \arg5\().8h, \arg5\().8h, \arg6\().8h
82     sqrshrn2  \arg4\().16b, \arg5\().8h, #3
91 .macro DIFF_LUMA_EQ4_P2P1P0_1 arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8, arg9
103     usubl   \arg8\().8h, \arg5\().8b, \arg1\().8b
106     uaddl \arg8\().8h, \arg2\().8b, \arg5\().8b
115 .macro DIFF_LUMA_EQ4_P2P1P0_2 arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8, arg9
127     usubl2   \arg8\().8h, \arg5\().16b, \arg1\().16b
130     uaddl2    \arg8\().8h, \arg2\().16b, \arg5\().16b
140 .macro DIFF_CHROMA_EQ4_P0Q0_1 arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7
143     usubl \arg5\().8h, \arg1\().8b, \arg3\().8b
144     add   \arg5\().8h, \arg5\().8h, \arg4\().8h
145     rshrn \arg6\().8b, \arg5\().8h, #2
146     usubl \arg5\().8h, \arg2\().8b, \arg0\().8b
147     add   \arg5\().8h, \arg5\().8h, \arg4\().8h
148     rshrn \arg7\().8b, \arg5\().8h, #2
151 .macro DIFF_CHROMA_EQ4_P0Q0_2  arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7
154     usubl2 \arg5\().8h, \arg1\().16b, \arg3\().16b
155     add   \arg5\().8h, \arg5\().8h, \arg4\().8h
156     rshrn2 \arg6\().16b, \arg5\().8h, #2
157     usubl2 \arg5\().8h, \arg2\().16b, \arg0\().16b
158     add   \arg5\().8h, \arg5\().8h, \arg4\().8h
159     rshrn2 \arg7\().16b, \arg5\().8h, #2
167 .macro LOAD_LUMA_DATA_3 arg0, arg1, arg2, arg3, arg4, arg5, arg6
169     ld3   {\arg3\().b, \arg4\().b, \arg5\().b} [\arg6], [x0], x1
172 .macro LOAD_LUMA_DATA_4 arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8
174     ld4   {\arg4\().b, \arg5\().b, \arg6\().b, \arg7\().b} [\arg8], [x0], x1
177 .macro STORE_LUMA_DATA_4 arg0, arg1, arg2, arg3, arg4, arg5  argument
179     st4   {\arg0\().b, \arg1\().b, \arg2\().b, \arg3\().b} [\arg5], [x2], x1
182 .macro STORE_LUMA_DATA_3 arg0, arg1, arg2, arg3, arg4, arg5, arg6
184     st3   {\arg3\().b, \arg4\().b, \arg5\().b} [\arg6], [x0], x1
187 .macro LOAD_CHROMA_DATA_4 arg0, arg1, arg2, arg3, arg4, arg5  argument
188     ld4   {\arg0\().b, \arg1\().b, \arg2\().b, \arg3\().b} [\arg5], [\arg4], x2
238 .macro BS_COMPARE_MV arg0, arg1, arg2, arg3, arg4, arg5  argument
255     addhn  \arg5\().8b, v20.8h, v20.8h
256     addhn2  \arg5\().16b, v21.8h, v21.8h
259 .macro BS_MV_CHECK arg0, arg1, arg2, arg3, arg4, arg5, arg6
283     zip1  \arg5\().4s, v0.4s, v2.4s
287     zip2  v1.4s, \arg5\().4s, v0.4s
288     zip1  v0.4s, \arg5\().4s, v0.4s