60   LD_SB4(ref, BPS, dest0, dest1, dest2, dest3);  in ITransformOne()
69   ST4x4_UB(res0, res0, 3, 2, 1, 0, dst, BPS);  in ITransformOne()
94   LW4(src, BPS, in0, in1, in2, in3);  in FTransform_MSA()
96   LW4(ref, BPS, in0, in1, in2, in3);  in FTransform_MSA()
183   LW4(in, BPS, in0_m, in1_m, in2_m, in3_m);  in TTransform_MSA()
214   for (y = 0; y < 16 * BPS; y += 4 * BPS) {  in Disto16x16_MSA()
258 #define DST(x, y) dst[(x) + (y) * BPS]
272   SW4(out, out, out, out, dst, BPS);  in VE4()
281   WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(X, I, J));  in HE4()
282   WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(I, J, K));  in HE4()
283   WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(J, K, L));  in HE4()
284   WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L));  in HE4()
293   SW4(dc, dc, dc, dc, dst, BPS);  in DC4()
313   SW4(val3, val2, val1, val0, dst, BPS);  in RD4()
333   SW4(val0, val1, val2, val3, dst, BPS);  in LD4()
427   PCKEV_ST4x4_UB(r0, r1, r2, r3, dst, BPS);  in TM4()
450     ST_UB8(out, out, out, out, out, out, out, out, dst + 0 * BPS, BPS);  \
451     ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS);  \
473       ST_UB4(L0, L1, L2, L3, dst, BPS);  in HorizontalPred16x16()
474       dst += 4 * BPS;  in HorizontalPred16x16()
506         ST_SB4(t0, t1, t2, t3, dst, BPS);  in TrueMotion16x16()
507         dst += 4 * BPS;  in TrueMotion16x16()
573   SD4(out, out, out, out, dst + 0 * BPS, BPS);  \
574   SD4(out, out, out, out, dst + 4 * BPS, BPS);  \
599       SD4(out0, out1, out2, out3, dst, BPS);  in HorizontalPred8x8()
600       dst += 4 * BPS;  in HorizontalPred8x8()
633         SD4(out0, out1, out2, out3, dst, BPS);  in TrueMotion8x8()
634         dst += 4 * BPS;  in TrueMotion8x8()
721   LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7);  in SSE16x16_MSA()
722   LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);  in SSE16x16_MSA()
727   a += 8 * BPS;  in SSE16x16_MSA()
728   b += 8 * BPS;  in SSE16x16_MSA()
729   LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7);  in SSE16x16_MSA()
730   LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);  in SSE16x16_MSA()
748   LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7);  in SSE16x8_MSA()
749   LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);  in SSE16x8_MSA()
768   LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7);  in SSE8x8_MSA()
769   LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);  in SSE8x8_MSA()
788   LW4(a, BPS, src0, src1, src2, src3);  in SSE4x4_MSA()
789   LW4(b, BPS, ref0, ref1, ref2, ref3);  in SSE4x4_MSA()