[ZARCH] Set alignment hint for vl/vst
This commit is contained in:
@@ -30,7 +30,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
static BLASLONG idmax_kernel_32(BLASLONG n, FLOAT *x, FLOAT *max) {
|
||||
BLASLONG imax;
|
||||
|
||||
__asm__("vl %%v0,0(%[x])\n\t"
|
||||
__asm__("vl %%v0,0(%[x]),3\n\t"
|
||||
"vleig %%v1,0,0\n\t"
|
||||
"vleig %%v1,1,1\n\t"
|
||||
"vrepig %%v2,16\n\t"
|
||||
@@ -55,14 +55,14 @@ static BLASLONG idmax_kernel_32(BLASLONG n, FLOAT *x, FLOAT *max) {
|
||||
"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 %%v16,0(%%r1,%[x]),3\n\t"
|
||||
"vl %%v17,16(%%r1,%[x]),3\n\t"
|
||||
"vl %%v18,32(%%r1,%[x]),3\n\t"
|
||||
"vl %%v19,48(%%r1,%[x]),3\n\t"
|
||||
"vl %%v20,64(%%r1,%[x]),3\n\t"
|
||||
"vl %%v21,80(%%r1,%[x]),3\n\t"
|
||||
"vl %%v22,96(%%r1,%[x]),3\n\t"
|
||||
"vl %%v23,112(%%r1,%[x]),3\n\t"
|
||||
"vfchedb %%v4,%%v16,%%v17\n\t"
|
||||
"vfchedb %%v5,%%v18,%%v19\n\t"
|
||||
"vfchedb %%v6,%%v20,%%v21\n\t"
|
||||
@@ -89,14 +89,14 @@ static BLASLONG idmax_kernel_32(BLASLONG n, FLOAT *x, FLOAT *max) {
|
||||
"vsel %%v0,%%v0,%%v16,%%v5\n\t"
|
||||
"vsel %%v1,%%v1,%%v4,%%v5\n\t"
|
||||
"vag %%v3,%%v3,%%v2\n\t"
|
||||
"vl %%v16,128(%%r1,%[x])\n\t"
|
||||
"vl %%v17,144(%%r1,%[x])\n\t"
|
||||
"vl %%v18,160(%%r1,%[x])\n\t"
|
||||
"vl %%v19,176(%%r1,%[x])\n\t"
|
||||
"vl %%v20,192(%%r1,%[x])\n\t"
|
||||
"vl %%v21,208(%%r1,%[x])\n\t"
|
||||
"vl %%v22,224(%%r1,%[x])\n\t"
|
||||
"vl %%v23,240(%%r1,%[x])\n\t"
|
||||
"vl %%v16,128(%%r1,%[x]),3\n\t"
|
||||
"vl %%v17,144(%%r1,%[x]),3\n\t"
|
||||
"vl %%v18,160(%%r1,%[x]),3\n\t"
|
||||
"vl %%v19,176(%%r1,%[x]),3\n\t"
|
||||
"vl %%v20,192(%%r1,%[x]),3\n\t"
|
||||
"vl %%v21,208(%%r1,%[x]),3\n\t"
|
||||
"vl %%v22,224(%%r1,%[x]),3\n\t"
|
||||
"vl %%v23,240(%%r1,%[x]),3\n\t"
|
||||
"vfchedb %%v4,%%v16,%%v17\n\t"
|
||||
"vfchedb %%v5,%%v18,%%v19\n\t"
|
||||
"vfchedb %%v6,%%v20,%%v21\n\t"
|
||||
|
||||
Reference in New Issue
Block a user