[ZARCH] Set alignment hint for vl/vst
This commit is contained in:
+24
-24
@@ -33,30 +33,30 @@ static void dscal_kernel_16(BLASLONG n, FLOAT da, FLOAT *x) {
|
||||
"xgr %%r1,%%r1\n\t"
|
||||
"0:\n\t"
|
||||
"pfd 2, 1024(%%r1,%[x])\n\t"
|
||||
"vl %%v24,0(%%r1,%[x])\n\t"
|
||||
"vl %%v24,0(%%r1,%[x]),3\n\t"
|
||||
"vfmdb %%v24,%%v24,%%v0\n\t"
|
||||
"vst %%v24,0(%%r1,%[x])\n\t"
|
||||
"vl %%v25,16(%%r1,%[x])\n\t"
|
||||
"vst %%v24,0(%%r1,%[x]),3\n\t"
|
||||
"vl %%v25,16(%%r1,%[x]),3\n\t"
|
||||
"vfmdb %%v25,%%v25,%%v0\n\t"
|
||||
"vst %%v25,16(%%r1,%[x])\n\t"
|
||||
"vl %%v26,32(%%r1,%[x])\n\t"
|
||||
"vst %%v25,16(%%r1,%[x]),3\n\t"
|
||||
"vl %%v26,32(%%r1,%[x]),3\n\t"
|
||||
"vfmdb %%v26,%%v26,%%v0\n\t"
|
||||
"vst %%v26,32(%%r1,%[x])\n\t"
|
||||
"vl %%v27,48(%%r1,%[x])\n\t"
|
||||
"vst %%v26,32(%%r1,%[x]),3\n\t"
|
||||
"vl %%v27,48(%%r1,%[x]),3\n\t"
|
||||
"vfmdb %%v27,%%v27,%%v0\n\t"
|
||||
"vst %%v27,48(%%r1,%[x])\n\t"
|
||||
"vl %%v28,64(%%r1,%[x])\n\t"
|
||||
"vst %%v27,48(%%r1,%[x]),3\n\t"
|
||||
"vl %%v28,64(%%r1,%[x]),3\n\t"
|
||||
"vfmdb %%v28,%%v28,%%v0\n\t"
|
||||
"vst %%v28,64(%%r1,%[x])\n\t"
|
||||
"vl %%v29,80(%%r1,%[x])\n\t"
|
||||
"vst %%v28,64(%%r1,%[x]),3\n\t"
|
||||
"vl %%v29,80(%%r1,%[x]),3\n\t"
|
||||
"vfmdb %%v29,%%v29,%%v0\n\t"
|
||||
"vst %%v29,80(%%r1,%[x])\n\t"
|
||||
"vl %%v30,96(%%r1,%[x])\n\t"
|
||||
"vst %%v29,80(%%r1,%[x]),3\n\t"
|
||||
"vl %%v30,96(%%r1,%[x]),3\n\t"
|
||||
"vfmdb %%v30,%%v30,%%v0\n\t"
|
||||
"vst %%v30,96(%%r1,%[x])\n\t"
|
||||
"vl %%v31,112(%%r1,%[x])\n\t"
|
||||
"vst %%v30,96(%%r1,%[x]),3\n\t"
|
||||
"vl %%v31,112(%%r1,%[x]),3\n\t"
|
||||
"vfmdb %%v31,%%v31,%%v0\n\t"
|
||||
"vst %%v31,112(%%r1,%[x])\n\t"
|
||||
"vst %%v31,112(%%r1,%[x]),3\n\t"
|
||||
"agfi %%r1,128\n\t"
|
||||
"brctg %[n],0b"
|
||||
: "+m"(*(FLOAT (*)[n]) x),[n] "+&r"(n)
|
||||
@@ -71,14 +71,14 @@ static void dscal_kernel_16_zero(BLASLONG n, FLOAT *x) {
|
||||
"xgr %%r1,%%r1\n\t"
|
||||
"0:\n\t"
|
||||
"pfd 2, 1024(%%r1,%[x])\n\t"
|
||||
"vst %%v0,0(%%r1,%[x])\n\t"
|
||||
"vst %%v0,16(%%r1,%[x])\n\t"
|
||||
"vst %%v0,32(%%r1,%[x])\n\t"
|
||||
"vst %%v0,48(%%r1,%[x])\n\t"
|
||||
"vst %%v0,64(%%r1,%[x])\n\t"
|
||||
"vst %%v0,80(%%r1,%[x])\n\t"
|
||||
"vst %%v0,96(%%r1,%[x])\n\t"
|
||||
"vst %%v0,112(%%r1,%[x])\n\t"
|
||||
"vst %%v0,0(%%r1,%[x]),3\n\t"
|
||||
"vst %%v0,16(%%r1,%[x]),3\n\t"
|
||||
"vst %%v0,32(%%r1,%[x]),3\n\t"
|
||||
"vst %%v0,48(%%r1,%[x]),3\n\t"
|
||||
"vst %%v0,64(%%r1,%[x]),3\n\t"
|
||||
"vst %%v0,80(%%r1,%[x]),3\n\t"
|
||||
"vst %%v0,96(%%r1,%[x]),3\n\t"
|
||||
"vst %%v0,112(%%r1,%[x]),3\n\t"
|
||||
"agfi %%r1,128\n\t"
|
||||
"brctg %[n],0b"
|
||||
: "=m"(*(FLOAT (*)[n]) x),[n] "+&r"(n)
|
||||
|
||||
Reference in New Issue
Block a user