327     f32x2 fp1;  in PSMTX44Copy()  local
330     fp1 = __PSQ_L(src, 0, 0);  in PSMTX44Copy()
333     __PSQ_ST(dst, fp1, 0, 0);  in PSMTX44Copy()
336     fp1 = __PSQ_LX(src, 8, 0, 0);  in PSMTX44Copy()
339     __PSQ_STX(dst, 8, fp1, 0, 0);  in PSMTX44Copy()
342     fp1 = __PSQ_LX(src, 16, 0, 0);  in PSMTX44Copy()
345     __PSQ_STX(dst, 16, fp1, 0, 0);  in PSMTX44Copy()
348     fp1 = __PSQ_LX(src, 24, 0, 0);  in PSMTX44Copy()
351     __PSQ_STX(dst, 24, fp1, 0, 0);  in PSMTX44Copy()
354     fp1 = __PSQ_LX(src, 32, 0, 0);  in PSMTX44Copy()
357     __PSQ_STX(dst, 32, fp1, 0, 0);  in PSMTX44Copy()
360     fp1 = __PSQ_LX(src, 40, 0, 0);  in PSMTX44Copy()
363     __PSQ_STX(dst, 40, fp1, 0, 0);  in PSMTX44Copy()
366     fp1 = __PSQ_LX(src, 48, 0, 0);  in PSMTX44Copy()
369     __PSQ_STX(dst, 48, fp1, 0, 0);  in PSMTX44Copy()
372     fp1 = __PSQ_LX(src, 56, 0, 0);  in PSMTX44Copy()
375     __PSQ_STX(dst, 56, fp1, 0, 0);  in PSMTX44Copy()
454     f32x2 fp0, fp1, fp2, fp3, fp4, fp5, fp6, fp7, fp8, fp9, fp10, fp11, fp12, fp13;  in PSMTX44Concat()  local
485     fp1 = __PSQ_LX(a,  8, 0, 0);  in PSMTX44Concat()
494     fp6 = __PS_MADDS0(fp4, fp1, fp6);  in PSMTX44Concat()
503     fp6 = __PS_MADDS1(fp5, fp1, fp6);  in PSMTX44Concat()
508     fp1 = __PSQ_LX(a,  24, 0, 0);  in PSMTX44Concat()
523     fp8 = __PS_MADDS0(fp4, fp1, fp8);  in PSMTX44Concat()
527     fp8 = __PS_MADDS1(fp5, fp1, fp8);  in PSMTX44Concat()
532     fp1 = __PSQ_LX(a, 40, 0, 0);  in PSMTX44Concat()
547     fp10 = __PS_MADDS0(fp4, fp1, fp10);  in PSMTX44Concat()
551     fp10 = __PS_MADDS1(fp5, fp1, fp10);  in PSMTX44Concat()
556     fp1 = __PSQ_LX(a,  56, 0, 0);  in PSMTX44Concat()
576     fp12 = __PS_MADDS0(fp4, fp1, fp12);  in PSMTX44Concat()
585     fp12 = __PS_MADDS1(fp5, fp1, fp12);  in PSMTX44Concat()
590     fp1 = __PSQ_LX(a,  8, 0, 0);  in PSMTX44Concat()
610     fp7 = __PS_MADDS0(fp4, fp1, fp7);  in PSMTX44Concat()
619     fp7 = __PS_MADDS1(fp5, fp1, fp7);  in PSMTX44Concat()
624     fp1 = __PSQ_LX(a,  24, 0, 0);  in PSMTX44Concat()
644     fp9 = __PS_MADDS0(fp4, fp1, fp9);  in PSMTX44Concat()
653     fp9 = __PS_MADDS1(fp5, fp1, fp9);  in PSMTX44Concat()
658     fp1 = __PSQ_LX(a, 40, 0, 0);  in PSMTX44Concat()
678     fp11 = __PS_MADDS0(fp4, fp1, fp11);  in PSMTX44Concat()
687     fp11 = __PS_MADDS1(fp5, fp1, fp11);  in PSMTX44Concat()
692     fp1 = __PSQ_LX(a, 56, 0, 0);  in PSMTX44Concat()
712     fp13 = __PS_MADDS0(fp4, fp1, fp13);  in PSMTX44Concat()
721     fp13 = __PS_MADDS1(fp5, fp1, fp13);  in PSMTX44Concat()
782     f32x2 fp0, fp1, fp2, fp3, fp4, fp5;  in PSMTX44Transpose()  local
788     fp1 = __PSQ_LX(src, 16, 0, 0);  in PSMTX44Transpose()
791     fp4 = __PS_MERGE00(fp0, fp1);  in PSMTX44Transpose()
800     fp5 = __PS_MERGE11(fp0, fp1);  in PSMTX44Transpose()
821     fp1 = __PSQ_LX(src, 48, 0, 0);  in PSMTX44Transpose()
827     fp4 = __PS_MERGE00(fp0, fp1);  in PSMTX44Transpose()
836     fp5 = __PS_MERGE11(fp0, fp1);  in PSMTX44Transpose()
1839     f32x2 fp1;  in PSMTX34To44()  local
1842     fp1 = __PSQ_L(src, 0, 0);  in PSMTX34To44()
1845     __PSQ_ST(dst, fp1, 0, 0);  in PSMTX34To44()
1848     fp1 = __PSQ_LX(src, 8, 0, 0);  in PSMTX34To44()
1851     __PSQ_STX(dst, 8, fp1, 0, 0);  in PSMTX34To44()
1854     fp1 = __PSQ_LX(src, 16, 0, 0);  in PSMTX34To44()
1857     __PSQ_STX(dst, 16, fp1, 0, 0);  in PSMTX34To44()
1860     fp1 = __PSQ_LX(src, 24, 0, 0);  in PSMTX34To44()
1863     __PSQ_STX(dst, 24, fp1, 0, 0);  in PSMTX34To44()
1866     fp1 = __PSQ_LX(src, 32, 0, 0);  in PSMTX34To44()
1869     __PSQ_STX(dst, 32, fp1, 0, 0);  in PSMTX34To44()
1872     fp1 = __PSQ_LX(src, 40, 0, 0);  in PSMTX34To44()
1875     __PSQ_STX(dst, 40, fp1, 0, 0);  in PSMTX34To44()