[ZARCH] Fix constraints and source code formatting
This commit is contained in:
@@ -31,53 +31,53 @@ static FLOAT smax_kernel_64(BLASLONG n, FLOAT *x) {
|
||||
FLOAT max;
|
||||
|
||||
__asm__("vl %%v0,0(%[x])\n\t"
|
||||
"srlg %[n],%[n],6\n\t"
|
||||
"xgr %%r1,%%r1\n\t"
|
||||
"0:\n\t"
|
||||
"pfd 1, 1024(%%r1,%[x])\n\t"
|
||||
"vl %%v16,0(%%r1,%[x])\n\t"
|
||||
"vl %%v17,16(%%r1,%[x])\n\t"
|
||||
"vl %%v18,32(%%r1,%[x])\n\t"
|
||||
"vl %%v19,48(%%r1,%[x])\n\t"
|
||||
"vl %%v20,64(%%r1,%[x])\n\t"
|
||||
"vl %%v21,80(%%r1,%[x])\n\t"
|
||||
"vl %%v22,96(%%r1,%[x])\n\t"
|
||||
"vl %%v23,112(%%r1,%[x])\n\t"
|
||||
"vl %%v24,128(%%r1,%[x])\n\t"
|
||||
"vl %%v25,144(%%r1,%[x])\n\t"
|
||||
"vl %%v26,160(%%r1,%[x])\n\t"
|
||||
"vl %%v27,176(%%r1,%[x])\n\t"
|
||||
"vl %%v28,192(%%r1,%[x])\n\t"
|
||||
"vl %%v29,208(%%r1,%[x])\n\t"
|
||||
"vl %%v30,224(%%r1,%[x])\n\t"
|
||||
"vl %%v31,240(%%r1,%[x])\n\t"
|
||||
"vfmaxsb %%v16,%%v16,%%v24,0\n\t"
|
||||
"vfmaxsb %%v17,%%v17,%%v25,0\n\t"
|
||||
"vfmaxsb %%v18,%%v18,%%v26,0\n\t"
|
||||
"vfmaxsb %%v19,%%v19,%%v27,0\n\t"
|
||||
"vfmaxsb %%v20,%%v20,%%v28,0\n\t"
|
||||
"vfmaxsb %%v21,%%v21,%%v29,0\n\t"
|
||||
"vfmaxsb %%v22,%%v22,%%v30,0\n\t"
|
||||
"vfmaxsb %%v23,%%v23,%%v31,0\n\t"
|
||||
"vfmaxsb %%v16,%%v16,%%v20,0\n\t"
|
||||
"vfmaxsb %%v17,%%v17,%%v21,0\n\t"
|
||||
"vfmaxsb %%v18,%%v18,%%v22,0\n\t"
|
||||
"vfmaxsb %%v19,%%v19,%%v23,0\n\t"
|
||||
"vfmaxsb %%v16,%%v16,%%v18,0\n\t"
|
||||
"vfmaxsb %%v17,%%v17,%%v19,0\n\t"
|
||||
"vfmaxsb %%v16,%%v16,%%v17,0\n\t"
|
||||
"vfmaxsb %%v0,%%v0,%%v16,0\n\t"
|
||||
"agfi %%r1, 256\n\t"
|
||||
"brctg %[n], 0b\n\t"
|
||||
"veslg %%v16,%%v0,32\n\t"
|
||||
"vfmaxsb %%v0,%%v0,%%v16,0\n\t"
|
||||
"vrepf %%v16,%%v0,2\n\t"
|
||||
"wfmaxsb %%v0,%%v0,%%v16,0\n\t"
|
||||
"ler %[max],%%f0"
|
||||
: [max] "=f"(max),[n] "+&r"(n)
|
||||
: "m"(*(const FLOAT (*)[n]) x),[x] "a"(x)
|
||||
: "cc", "r1", "v0", "v16", "v17", "v18", "v19", "v20", "v21", "v22",
|
||||
"v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
|
||||
"srlg %[n],%[n],6\n\t"
|
||||
"xgr %%r1,%%r1\n\t"
|
||||
"0:\n\t"
|
||||
"pfd 1, 1024(%%r1,%[x])\n\t"
|
||||
"vl %%v16,0(%%r1,%[x])\n\t"
|
||||
"vl %%v17,16(%%r1,%[x])\n\t"
|
||||
"vl %%v18,32(%%r1,%[x])\n\t"
|
||||
"vl %%v19,48(%%r1,%[x])\n\t"
|
||||
"vl %%v20,64(%%r1,%[x])\n\t"
|
||||
"vl %%v21,80(%%r1,%[x])\n\t"
|
||||
"vl %%v22,96(%%r1,%[x])\n\t"
|
||||
"vl %%v23,112(%%r1,%[x])\n\t"
|
||||
"vl %%v24,128(%%r1,%[x])\n\t"
|
||||
"vl %%v25,144(%%r1,%[x])\n\t"
|
||||
"vl %%v26,160(%%r1,%[x])\n\t"
|
||||
"vl %%v27,176(%%r1,%[x])\n\t"
|
||||
"vl %%v28,192(%%r1,%[x])\n\t"
|
||||
"vl %%v29,208(%%r1,%[x])\n\t"
|
||||
"vl %%v30,224(%%r1,%[x])\n\t"
|
||||
"vl %%v31,240(%%r1,%[x])\n\t"
|
||||
"vfmaxsb %%v16,%%v16,%%v24,0\n\t"
|
||||
"vfmaxsb %%v17,%%v17,%%v25,0\n\t"
|
||||
"vfmaxsb %%v18,%%v18,%%v26,0\n\t"
|
||||
"vfmaxsb %%v19,%%v19,%%v27,0\n\t"
|
||||
"vfmaxsb %%v20,%%v20,%%v28,0\n\t"
|
||||
"vfmaxsb %%v21,%%v21,%%v29,0\n\t"
|
||||
"vfmaxsb %%v22,%%v22,%%v30,0\n\t"
|
||||
"vfmaxsb %%v23,%%v23,%%v31,0\n\t"
|
||||
"vfmaxsb %%v16,%%v16,%%v20,0\n\t"
|
||||
"vfmaxsb %%v17,%%v17,%%v21,0\n\t"
|
||||
"vfmaxsb %%v18,%%v18,%%v22,0\n\t"
|
||||
"vfmaxsb %%v19,%%v19,%%v23,0\n\t"
|
||||
"vfmaxsb %%v16,%%v16,%%v18,0\n\t"
|
||||
"vfmaxsb %%v17,%%v17,%%v19,0\n\t"
|
||||
"vfmaxsb %%v16,%%v16,%%v17,0\n\t"
|
||||
"vfmaxsb %%v0,%%v0,%%v16,0\n\t"
|
||||
"agfi %%r1, 256\n\t"
|
||||
"brctg %[n], 0b\n\t"
|
||||
"veslg %%v16,%%v0,32\n\t"
|
||||
"vfmaxsb %%v0,%%v0,%%v16,0\n\t"
|
||||
"vrepf %%v16,%%v0,2\n\t"
|
||||
"wfmaxsb %%v0,%%v0,%%v16,0\n\t"
|
||||
"ler %[max],%%f0"
|
||||
: [max] "=f"(max),[n] "+&r"(n)
|
||||
: "m"(*(const struct { FLOAT x[n]; } *) x),[x] "a"(x)
|
||||
: "cc", "r1", "v0", "v16", "v17", "v18", "v19", "v20", "v21", "v22",
|
||||
"v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
|
||||
|
||||
return max;
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user