150 #define LW4(psrc, stride, out0, out1, out2, out3) do {  \  argument
158   out3 = LW(ptmp);                                      \
234 #define LD_B4(RTYPE, psrc, stride, out0, out1, out2, out3) do {  \  argument
236   LD_B2(RTYPE, psrc + 2 * stride , stride, out2, out3);          \
242               out0, out1, out2, out3, out4, out5, out6, out7) do {  \  argument
243   LD_B4(RTYPE, psrc, stride, out0, out1, out2, out3);               \
284 #define LD_W4(RTYPE, psrc, stride, out0, out1, out2, out3) do {  \  argument
286   LD_W2(RTYPE, psrc + 2 * stride, stride, out2, out3);           \
646 #define HADD_SH4(RTYPE, in0, in1, in2, in3, out0, out1, out2, out3) do {  \  argument
648   HADD_SH2(RTYPE, in2, in3, out2, out3);                                  \
874                 out0, out1, out2, out3) do {                    \  argument
876   ILVR_B2(RTYPE, in4, in5, in6, in7, out2, out3);               \
900                 out0, out1, out2, out3) do {                    \  argument
902   ILVR_H2(RTYPE, in4, in5, in6, in7, out2, out3);               \
924                 out0, out1, out2, out3) do {                    \  argument
926   ILVR_D2(RTYPE, in4, in5, in6, in7, out2, out3);               \
985                  out0, out1, out2, out3) do {                    \  argument
987   PCKEV_B2(RTYPE, in4, in5, in6, in7, out2, out3);               \
1175              out0, out1, out2, out3) do {             \  argument
1177   ADD2(in4, in5, in6, in7, out2, out3);               \
1198              out0, out1, out2, out3) do {             \  argument
1202   out3 = in6 - in7;                                   \
1230              out0, out1, out2, out3) do {             \  argument
1232   MUL2(in4, in5, in6, in7, out2, out3);               \
1269 #define BUTTERFLY_4(in0, in1, in2, in3, out0, out1, out2, out3) do {  \  argument
1273   out3 = in0 - in3;                                                   \
1284                             out0, out1, out2, out3) do {                   \  argument
1288   ILVEV_D2_UB(tmp2_m, tmp3_m, tmp0_m, tmp1_m, out1, out3);                 \
1292   ILVEV_B2_SD(out1, out3, tmp2_m, tmp3_m, tmp0_m, tmp1_m);                 \
1294   ILVOD_B2_SD(out1, out3, tmp2_m, tmp3_m, tmp0_m, tmp1_m);                 \
1295   ILVEVOD_H2_UB(tmp0_m, tmp1_m, tmp0_m, tmp1_m, out1, out3);               \
1306                             out0, out1, out2, out3, out4, out5,            \  argument
1312   ILVEV_D2_UB(in4, in12, in5, in13, out3, out2);                           \
1316   ILVEV_B2_UB(out3, out2, out1, out0, out5, out7);                         \
1317   ILVOD_B2_SH(out3, out2, out1, out0, tmp6_m, tmp7_m);                     \
1325   ILVEVOD_W2_UB(tmp2_m, tmp3_m, tmp2_m, tmp3_m, out3, out7);               \
1334                        out0, out1, out2, out3) do {                          \  argument
1341   out3 = (RTYPE)__msa_ilvl_d((v2i64)s3_m, (v2i64)s1_m);                      \