25 #define IDCT_1D_W(in0, in1, in2, in3, out0, out1, out2, out3) do {  \  argument
41   BUTTERFLY_4(a1_m, b1_m, c1_m, d1_m, out0, out1, out2, out3);      \
82   uint64_t out0, out1, out2, out3;  in FTransform_MSA()  local
130   out3 = __msa_copy_s_d((v2i64)t1, 1);  in FTransform_MSA()
131   SD4(out0, out1, out2, out3, out, 8);  in FTransform_MSA()
598       const uint64_t out3 = __msa_copy_s_d((v2i64)L3, 0);  in HorizontalPred8x8()  local
599       SD4(out0, out1, out2, out3, dst, BPS);  in HorizontalPred8x8()
620         uint64_t out0, out1, out2, out3;  in TrueMotion8x8()  local
632         out3 = __msa_copy_s_d((v2i64)t1, 1);  in TrueMotion8x8()
633         SD4(out0, out1, out2, out3, dst, BPS);  in TrueMotion8x8()
693 #define PACK_DOTP_UB4_SW(in0, in1, in2, in3, out0, out1, out2, out3) do {  \  argument
701   DOTP_SH2_SW(tmp2, tmp3, tmp2, tmp3, out2, out3);                         \
704 #define PACK_DPADD_UB4_SW(in0, in1, in2, in3, out0, out1, out2, out3) do {  \  argument
712   DPADD_SH2_SW(tmp2, tmp3, tmp2, tmp3, out2, out3);                         \
719   v4i32 out0, out1, out2, out3;  in SSE16x16_MSA()  local
723   PACK_DOTP_UB4_SW(src0, ref0, src1, ref1, out0, out1, out2, out3);  in SSE16x16_MSA()
724   PACK_DPADD_UB4_SW(src2, ref2, src3, ref3, out0, out1, out2, out3);  in SSE16x16_MSA()
725   PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3);  in SSE16x16_MSA()
726   PACK_DPADD_UB4_SW(src6, ref6, src7, ref7, out0, out1, out2, out3);  in SSE16x16_MSA()
731   PACK_DPADD_UB4_SW(src0, ref0, src1, ref1, out0, out1, out2, out3);  in SSE16x16_MSA()
732   PACK_DPADD_UB4_SW(src2, ref2, src3, ref3, out0, out1, out2, out3);  in SSE16x16_MSA()
733   PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3);  in SSE16x16_MSA()
734   PACK_DPADD_UB4_SW(src6, ref6, src7, ref7, out0, out1, out2, out3);  in SSE16x16_MSA()
736   out2 += out3;  in SSE16x16_MSA()
746   v4i32 out0, out1, out2, out3;  in SSE16x8_MSA()  local
750   PACK_DOTP_UB4_SW(src0, ref0, src1, ref1, out0, out1, out2, out3);  in SSE16x8_MSA()
751   PACK_DPADD_UB4_SW(src2, ref2, src3, ref3, out0, out1, out2, out3);  in SSE16x8_MSA()
752   PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3);  in SSE16x8_MSA()
753   PACK_DPADD_UB4_SW(src6, ref6, src7, ref7, out0, out1, out2, out3);  in SSE16x8_MSA()
755   out2 += out3;  in SSE16x8_MSA()
766   v4i32 out0, out1, out2, out3;  in SSE8x8_MSA()  local
771   PACK_DOTP_UB4_SW(t0, t2, t1, t3, out0, out1, out2, out3);  in SSE8x8_MSA()
773   PACK_DPADD_UB4_SW(t0, t2, t1, t3, out0, out1, out2, out3);  in SSE8x8_MSA()
775   out2 += out3;  in SSE8x8_MSA()