[ZARCH] Fix constraints and source code formatting
This commit is contained in:
maamountki
+230
-222
@@ -31,266 +31,274 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
static void zgemv_kernel_4x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y,
|
||||
FLOAT *alpha) {
|
||||
register FLOAT *ap0 = ap[0];
|
||||
register FLOAT *ap1 = ap[1];
|
||||
register FLOAT *ap2 = ap[2];
|
||||
register FLOAT *ap3 = ap[3];
|
||||
|
||||
__asm__("vzero %%v16\n\t"
|
||||
"vzero %%v17\n\t"
|
||||
"vzero %%v18\n\t"
|
||||
"vzero %%v19\n\t"
|
||||
"vzero %%v20\n\t"
|
||||
"vzero %%v21\n\t"
|
||||
"vzero %%v22\n\t"
|
||||
"vzero %%v23\n\t"
|
||||
"xgr %%r1,%%r1\n\t"
|
||||
"srlg %[n],%[n],1\n\t"
|
||||
"0:\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap0])\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap1])\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap2])\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap3])\n\t"
|
||||
"pfd 1,1024(%%r1,%[x])\n\t"
|
||||
"vl %%v0,0(%%r1,%[x])\n\t"
|
||||
"vzero %%v17\n\t"
|
||||
"vzero %%v18\n\t"
|
||||
"vzero %%v19\n\t"
|
||||
"vzero %%v20\n\t"
|
||||
"vzero %%v21\n\t"
|
||||
"vzero %%v22\n\t"
|
||||
"vzero %%v23\n\t"
|
||||
"xgr %%r1,%%r1\n\t"
|
||||
"srlg %[n],%[n],1\n\t"
|
||||
"0:\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap0])\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap1])\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap2])\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap3])\n\t"
|
||||
"pfd 1,1024(%%r1,%[x])\n\t"
|
||||
"vl %%v0,0(%%r1,%[x])\n\t"
|
||||
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
|
||||
"vleg %%v1,8(%%r1,%[x]),0\n\t"
|
||||
"wflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,0(%%r1,%[x]),1\n\t"
|
||||
"vleg %%v1,8(%%r1,%[x]),0\n\t"
|
||||
"wflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,0(%%r1,%[x]),1\n\t"
|
||||
#else
|
||||
"vleg %%v1,0(%%r1,%[x]),1\n\t"
|
||||
"vflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,8(%%r1,%[x]),0\n\t"
|
||||
"vleg %%v1,0(%%r1,%[x]),1\n\t"
|
||||
"vflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,8(%%r1,%[x]),0\n\t"
|
||||
#endif
|
||||
"vlrepg %%v24,0(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v25,8(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v26,0(%%r1,%[ap1])\n\t"
|
||||
"vlrepg %%v27,8(%%r1,%[ap1])\n\t"
|
||||
"vlrepg %%v28,0(%%r1,%[ap2])\n\t"
|
||||
"vlrepg %%v29,8(%%r1,%[ap2])\n\t"
|
||||
"vlrepg %%v30,0(%%r1,%[ap3])\n\t"
|
||||
"vlrepg %%v31,8(%%r1,%[ap3])\n\t"
|
||||
"vfmadb %%v16,%%v24,%%v0,%%v16\n\t"
|
||||
"vfmadb %%v20,%%v25,%%v1,%%v20\n\t"
|
||||
"vfmadb %%v17,%%v26,%%v0,%%v17\n\t"
|
||||
"vfmadb %%v21,%%v27,%%v1,%%v21\n\t"
|
||||
"vfmadb %%v18,%%v28,%%v0,%%v18\n\t"
|
||||
"vfmadb %%v22,%%v29,%%v1,%%v22\n\t"
|
||||
"vfmadb %%v19,%%v30,%%v0,%%v19\n\t"
|
||||
"vfmadb %%v23,%%v31,%%v1,%%v23\n\t"
|
||||
"vl %%v0,16(%%r1,%[x])\n\t"
|
||||
"vlrepg %%v24,0(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v25,8(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v26,0(%%r1,%[ap1])\n\t"
|
||||
"vlrepg %%v27,8(%%r1,%[ap1])\n\t"
|
||||
"vlrepg %%v28,0(%%r1,%[ap2])\n\t"
|
||||
"vlrepg %%v29,8(%%r1,%[ap2])\n\t"
|
||||
"vlrepg %%v30,0(%%r1,%[ap3])\n\t"
|
||||
"vlrepg %%v31,8(%%r1,%[ap3])\n\t"
|
||||
"vfmadb %%v16,%%v24,%%v0,%%v16\n\t"
|
||||
"vfmadb %%v20,%%v25,%%v1,%%v20\n\t"
|
||||
"vfmadb %%v17,%%v26,%%v0,%%v17\n\t"
|
||||
"vfmadb %%v21,%%v27,%%v1,%%v21\n\t"
|
||||
"vfmadb %%v18,%%v28,%%v0,%%v18\n\t"
|
||||
"vfmadb %%v22,%%v29,%%v1,%%v22\n\t"
|
||||
"vfmadb %%v19,%%v30,%%v0,%%v19\n\t"
|
||||
"vfmadb %%v23,%%v31,%%v1,%%v23\n\t"
|
||||
"vl %%v0,16(%%r1,%[x])\n\t"
|
||||
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
|
||||
"vleg %%v1,24(%%r1,%[x]),0\n\t"
|
||||
"wflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,16(%%r1,%[x]),1\n\t"
|
||||
"vleg %%v1,24(%%r1,%[x]),0\n\t"
|
||||
"wflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,16(%%r1,%[x]),1\n\t"
|
||||
#else
|
||||
"vleg %%v1,16(%%r1,%[x]),1\n\t"
|
||||
"vflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,24(%%r1,%[x]),0\n\t"
|
||||
"vleg %%v1,16(%%r1,%[x]),1\n\t"
|
||||
"vflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,24(%%r1,%[x]),0\n\t"
|
||||
#endif
|
||||
"vlrepg %%v24,16(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v25,24(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v26,16(%%r1,%[ap1])\n\t"
|
||||
"vlrepg %%v27,24(%%r1,%[ap1])\n\t"
|
||||
"vlrepg %%v28,16(%%r1,%[ap2])\n\t"
|
||||
"vlrepg %%v29,24(%%r1,%[ap2])\n\t"
|
||||
"vlrepg %%v30,16(%%r1,%[ap3])\n\t"
|
||||
"vlrepg %%v31,24(%%r1,%[ap3])\n\t"
|
||||
"vfmadb %%v16,%%v24,%%v0,%%v16\n\t"
|
||||
"vfmadb %%v20,%%v25,%%v1,%%v20\n\t"
|
||||
"vfmadb %%v17,%%v26,%%v0,%%v17\n\t"
|
||||
"vfmadb %%v21,%%v27,%%v1,%%v21\n\t"
|
||||
"vfmadb %%v18,%%v28,%%v0,%%v18\n\t"
|
||||
"vfmadb %%v22,%%v29,%%v1,%%v22\n\t"
|
||||
"vfmadb %%v19,%%v30,%%v0,%%v19\n\t"
|
||||
"vfmadb %%v23,%%v31,%%v1,%%v23\n\t"
|
||||
"agfi %%r1,32\n\t"
|
||||
"brctg %[n],0b\n\t"
|
||||
"vfadb %%v16,%%v16,%%v20\n\t"
|
||||
"vfadb %%v17,%%v17,%%v21\n\t"
|
||||
"vfadb %%v18,%%v18,%%v22\n\t"
|
||||
"vfadb %%v19,%%v19,%%v23\n\t"
|
||||
"vpdi %%v20,%%v16,%%v16,4\n\t"
|
||||
"vpdi %%v21,%%v17,%%v17,4\n\t"
|
||||
"vpdi %%v22,%%v18,%%v18,4\n\t"
|
||||
"vpdi %%v23,%%v19,%%v19,4\n\t"
|
||||
"vlrepg %%v24,16(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v25,24(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v26,16(%%r1,%[ap1])\n\t"
|
||||
"vlrepg %%v27,24(%%r1,%[ap1])\n\t"
|
||||
"vlrepg %%v28,16(%%r1,%[ap2])\n\t"
|
||||
"vlrepg %%v29,24(%%r1,%[ap2])\n\t"
|
||||
"vlrepg %%v30,16(%%r1,%[ap3])\n\t"
|
||||
"vlrepg %%v31,24(%%r1,%[ap3])\n\t"
|
||||
"vfmadb %%v16,%%v24,%%v0,%%v16\n\t"
|
||||
"vfmadb %%v20,%%v25,%%v1,%%v20\n\t"
|
||||
"vfmadb %%v17,%%v26,%%v0,%%v17\n\t"
|
||||
"vfmadb %%v21,%%v27,%%v1,%%v21\n\t"
|
||||
"vfmadb %%v18,%%v28,%%v0,%%v18\n\t"
|
||||
"vfmadb %%v22,%%v29,%%v1,%%v22\n\t"
|
||||
"vfmadb %%v19,%%v30,%%v0,%%v19\n\t"
|
||||
"vfmadb %%v23,%%v31,%%v1,%%v23\n\t"
|
||||
"agfi %%r1,32\n\t"
|
||||
"brctg %[n],0b\n\t"
|
||||
"vfadb %%v16,%%v16,%%v20\n\t"
|
||||
"vfadb %%v17,%%v17,%%v21\n\t"
|
||||
"vfadb %%v18,%%v18,%%v22\n\t"
|
||||
"vfadb %%v19,%%v19,%%v23\n\t"
|
||||
"vpdi %%v20,%%v16,%%v16,4\n\t"
|
||||
"vpdi %%v21,%%v17,%%v17,4\n\t"
|
||||
"vpdi %%v22,%%v18,%%v18,4\n\t"
|
||||
"vpdi %%v23,%%v19,%%v19,4\n\t"
|
||||
#if !defined(XCONJ)
|
||||
"vlrepg %%v24,0(%[alpha])\n\t"
|
||||
"vleg %%v25,8(%[alpha]),0\n\t"
|
||||
"wflcdb %%v25,%%v25\n\t"
|
||||
"vleg %%v25,8(%[alpha]),1\n\t"
|
||||
"vlrepg %%v24,0(%[alpha])\n\t"
|
||||
"vleg %%v25,8(%[alpha]),0\n\t"
|
||||
"wflcdb %%v25,%%v25\n\t"
|
||||
"vleg %%v25,8(%[alpha]),1\n\t"
|
||||
#else
|
||||
"vleg %%v24,0(%[alpha]),1\n\t"
|
||||
"vflcdb %%v24,%%v24\n\t"
|
||||
"vleg %%v24,0(%[alpha]),0\n\t"
|
||||
"vlrepg %%v25,8(%[alpha])\n\t"
|
||||
"vleg %%v24,0(%[alpha]),1\n\t"
|
||||
"vflcdb %%v24,%%v24\n\t"
|
||||
"vleg %%v24,0(%[alpha]),0\n\t"
|
||||
"vlrepg %%v25,8(%[alpha])\n\t"
|
||||
#endif
|
||||
"vl %%v26,0(%[y])\n\t"
|
||||
"vl %%v27,16(%[y])\n\t"
|
||||
"vl %%v28,32(%[y])\n\t"
|
||||
"vl %%v29,48(%[y])\n\t"
|
||||
"vfmadb %%v26,%%v16,%%v24,%%v26\n\t"
|
||||
"vfmadb %%v26,%%v20,%%v25,%%v26\n\t"
|
||||
"vfmadb %%v27,%%v17,%%v24,%%v27\n\t"
|
||||
"vfmadb %%v27,%%v21,%%v25,%%v27\n\t"
|
||||
"vfmadb %%v28,%%v18,%%v24,%%v28\n\t"
|
||||
"vfmadb %%v28,%%v22,%%v25,%%v28\n\t"
|
||||
"vfmadb %%v29,%%v19,%%v24,%%v29\n\t"
|
||||
"vfmadb %%v29,%%v23,%%v25,%%v29\n\t"
|
||||
"vst %%v26,0(%[y])\n\t"
|
||||
"vst %%v27,16(%[y])\n\t"
|
||||
"vst %%v28,32(%[y])\n\t"
|
||||
"vst %%v29,48(%[y])"
|
||||
: "+m"(*(FLOAT (*)[8]) y),[n] "+&r"(n)
|
||||
: [y] "a"(y), "m"(*(const FLOAT (*)[n * 2]) ap[0]),[ap0] "a"(ap[0]),
|
||||
"m"(*(const FLOAT (*)[n * 2]) ap[1]),[ap1] "a"(ap[1]),
|
||||
"m"(*(const FLOAT (*)[n * 2]) ap[2]),[ap2] "a"(ap[2]),
|
||||
"m"(*(const FLOAT (*)[n * 2]) ap[3]),[ap3] "a"(ap[3]),
|
||||
"m"(*(const FLOAT (*)[n * 2]) x),[x] "a"(x),
|
||||
"m"(*(const FLOAT (*)[2]) alpha),[alpha] "a"(alpha)
|
||||
: "cc", "r1", "v0", "v1", "v16", "v17", "v18", "v19", "v20", "v21",
|
||||
"v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30",
|
||||
"v31");
|
||||
"vl %%v26,0(%[y])\n\t"
|
||||
"vl %%v27,16(%[y])\n\t"
|
||||
"vl %%v28,32(%[y])\n\t"
|
||||
"vl %%v29,48(%[y])\n\t"
|
||||
"vfmadb %%v26,%%v16,%%v24,%%v26\n\t"
|
||||
"vfmadb %%v26,%%v20,%%v25,%%v26\n\t"
|
||||
"vfmadb %%v27,%%v17,%%v24,%%v27\n\t"
|
||||
"vfmadb %%v27,%%v21,%%v25,%%v27\n\t"
|
||||
"vfmadb %%v28,%%v18,%%v24,%%v28\n\t"
|
||||
"vfmadb %%v28,%%v22,%%v25,%%v28\n\t"
|
||||
"vfmadb %%v29,%%v19,%%v24,%%v29\n\t"
|
||||
"vfmadb %%v29,%%v23,%%v25,%%v29\n\t"
|
||||
"vst %%v26,0(%[y])\n\t"
|
||||
"vst %%v27,16(%[y])\n\t"
|
||||
"vst %%v28,32(%[y])\n\t"
|
||||
"vst %%v29,48(%[y])"
|
||||
: "+m"(*(struct { FLOAT x[8]; } *) y),[n] "+&r"(n)
|
||||
: [y] "a"(y), "m"(*(const struct { FLOAT x[n * 2]; } *) ap0),[ap0] "a"(ap0),
|
||||
"m"(*(const struct { FLOAT x[n * 2]; } *) ap1),[ap1] "a"(ap1),
|
||||
"m"(*(const struct { FLOAT x[n * 2]; } *) ap2),[ap2] "a"(ap2),
|
||||
"m"(*(const struct { FLOAT x[n * 2]; } *) ap3),[ap3] "a"(ap3),
|
||||
"m"(*(const struct { FLOAT x[n * 2]; } *) x),[x] "a"(x),
|
||||
"m"(*(const struct { FLOAT x[2]; } *) alpha),[alpha] "a"(alpha)
|
||||
: "cc", "r1", "v0", "v1", "v16", "v17", "v18", "v19", "v20", "v21",
|
||||
"v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30",
|
||||
"v31");
|
||||
}
|
||||
|
||||
static void zgemv_kernel_4x2(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y,
|
||||
FLOAT *alpha) {
|
||||
register FLOAT *ap0 = ap[0];
|
||||
register FLOAT *ap1 = ap[1];
|
||||
|
||||
__asm__("vzero %%v16\n\t"
|
||||
"vzero %%v17\n\t"
|
||||
"vzero %%v18\n\t"
|
||||
"vzero %%v19\n\t"
|
||||
"xgr %%r1,%%r1\n\t"
|
||||
"srlg %[n],%[n],1\n\t"
|
||||
"0:\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap0])\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap1])\n\t"
|
||||
"pfd 1,1024(%%r1,%[x])\n\t"
|
||||
"vl %%v0,0(%%r1,%[x])\n\t"
|
||||
"vzero %%v17\n\t"
|
||||
"vzero %%v18\n\t"
|
||||
"vzero %%v19\n\t"
|
||||
"xgr %%r1,%%r1\n\t"
|
||||
"srlg %[n],%[n],1\n\t"
|
||||
"0:\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap0])\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap1])\n\t"
|
||||
"pfd 1,1024(%%r1,%[x])\n\t"
|
||||
"vl %%v0,0(%%r1,%[x])\n\t"
|
||||
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
|
||||
"vleg %%v1,8(%%r1,%[x]),0\n\t"
|
||||
"wflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,0(%%r1,%[x]),1\n\t"
|
||||
"vleg %%v1,8(%%r1,%[x]),0\n\t"
|
||||
"wflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,0(%%r1,%[x]),1\n\t"
|
||||
#else
|
||||
"vleg %%v1,0(%%r1,%[x]),1\n\t"
|
||||
"vflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,8(%%r1,%[x]),0\n\t"
|
||||
"vleg %%v1,0(%%r1,%[x]),1\n\t"
|
||||
"vflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,8(%%r1,%[x]),0\n\t"
|
||||
#endif
|
||||
"vlrepg %%v20,0(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v21,8(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v22,0(%%r1,%[ap1])\n\t"
|
||||
"vlrepg %%v23,8(%%r1,%[ap1])\n\t"
|
||||
"vfmadb %%v16,%%v20,%%v0,%%v16\n\t"
|
||||
"vfmadb %%v18,%%v21,%%v1,%%v18\n\t"
|
||||
"vfmadb %%v17,%%v22,%%v0,%%v17\n\t"
|
||||
"vfmadb %%v19,%%v23,%%v1,%%v19\n\t"
|
||||
"vl %%v0,16(%%r1,%[x])\n\t"
|
||||
"vlrepg %%v20,0(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v21,8(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v22,0(%%r1,%[ap1])\n\t"
|
||||
"vlrepg %%v23,8(%%r1,%[ap1])\n\t"
|
||||
"vfmadb %%v16,%%v20,%%v0,%%v16\n\t"
|
||||
"vfmadb %%v18,%%v21,%%v1,%%v18\n\t"
|
||||
"vfmadb %%v17,%%v22,%%v0,%%v17\n\t"
|
||||
"vfmadb %%v19,%%v23,%%v1,%%v19\n\t"
|
||||
"vl %%v0,16(%%r1,%[x])\n\t"
|
||||
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
|
||||
"vleg %%v1,24(%%r1,%[x]),0\n\t"
|
||||
"wflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,16(%%r1,%[x]),1\n\t"
|
||||
"vleg %%v1,24(%%r1,%[x]),0\n\t"
|
||||
"wflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,16(%%r1,%[x]),1\n\t"
|
||||
#else
|
||||
"vleg %%v1,16(%%r1,%[x]),1\n\t"
|
||||
"vflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,24(%%r1,%[x]),0\n\t"
|
||||
"vleg %%v1,16(%%r1,%[x]),1\n\t"
|
||||
"vflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,24(%%r1,%[x]),0\n\t"
|
||||
#endif
|
||||
"vlrepg %%v20,16(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v21,24(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v22,16(%%r1,%[ap1])\n\t"
|
||||
"vlrepg %%v23,24(%%r1,%[ap1])\n\t"
|
||||
"vfmadb %%v16,%%v20,%%v0,%%v16\n\t"
|
||||
"vfmadb %%v18,%%v21,%%v1,%%v18\n\t"
|
||||
"vfmadb %%v17,%%v22,%%v0,%%v17\n\t"
|
||||
"vfmadb %%v19,%%v23,%%v1,%%v19\n\t"
|
||||
"agfi %%r1,32\n\t"
|
||||
"brctg %[n],0b\n\t"
|
||||
"vfadb %%v16,%%v16,%%v18\n\t"
|
||||
"vfadb %%v17,%%v17,%%v19\n\t"
|
||||
"vpdi %%v18,%%v16,%%v16,4\n\t"
|
||||
"vpdi %%v19,%%v17,%%v17,4\n\t"
|
||||
"vlrepg %%v20,16(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v21,24(%%r1,%[ap0])\n\t"
|
||||
"vlrepg %%v22,16(%%r1,%[ap1])\n\t"
|
||||
"vlrepg %%v23,24(%%r1,%[ap1])\n\t"
|
||||
"vfmadb %%v16,%%v20,%%v0,%%v16\n\t"
|
||||
"vfmadb %%v18,%%v21,%%v1,%%v18\n\t"
|
||||
"vfmadb %%v17,%%v22,%%v0,%%v17\n\t"
|
||||
"vfmadb %%v19,%%v23,%%v1,%%v19\n\t"
|
||||
"agfi %%r1,32\n\t"
|
||||
"brctg %[n],0b\n\t"
|
||||
"vfadb %%v16,%%v16,%%v18\n\t"
|
||||
"vfadb %%v17,%%v17,%%v19\n\t"
|
||||
"vpdi %%v18,%%v16,%%v16,4\n\t"
|
||||
"vpdi %%v19,%%v17,%%v17,4\n\t"
|
||||
#if !defined(XCONJ)
|
||||
"vlrepg %%v20,0(%[alpha])\n\t"
|
||||
"vleg %%v21,8(%[alpha]),0\n\t"
|
||||
"wflcdb %%v21,%%v21\n\t"
|
||||
"vleg %%v21,8(%[alpha]),1\n\t"
|
||||
"vlrepg %%v20,0(%[alpha])\n\t"
|
||||
"vleg %%v21,8(%[alpha]),0\n\t"
|
||||
"wflcdb %%v21,%%v21\n\t"
|
||||
"vleg %%v21,8(%[alpha]),1\n\t"
|
||||
#else
|
||||
"vleg %%v20,0(%[alpha]),1\n\t"
|
||||
"vflcdb %%v20,%%v20\n\t"
|
||||
"vleg %%v20,0(%[alpha]),0\n\t"
|
||||
"vlrepg %%v21,8(%[alpha])\n\t"
|
||||
"vleg %%v20,0(%[alpha]),1\n\t"
|
||||
"vflcdb %%v20,%%v20\n\t"
|
||||
"vleg %%v20,0(%[alpha]),0\n\t"
|
||||
"vlrepg %%v21,8(%[alpha])\n\t"
|
||||
#endif
|
||||
"vl %%v22,0(%[y])\n\t"
|
||||
"vl %%v23,16(%[y])\n\t"
|
||||
"vfmadb %%v22,%%v16,%%v20,%%v22\n\t"
|
||||
"vfmadb %%v22,%%v18,%%v21,%%v22\n\t"
|
||||
"vfmadb %%v23,%%v17,%%v20,%%v23\n\t"
|
||||
"vfmadb %%v23,%%v19,%%v21,%%v23\n\t"
|
||||
"vst %%v22,0(%[y])\n\t"
|
||||
"vst %%v23,16(%[y])\n\t"
|
||||
: "+m"(*(FLOAT (*)[4]) y),[n] "+&r"(n)
|
||||
: [y] "a"(y), "m"(*(const FLOAT (*)[n * 2]) ap[0]),[ap0] "a"(ap[0]),
|
||||
"m"(*(const FLOAT (*)[n * 2]) ap[1]),[ap1] "a"(ap[1]),
|
||||
"m"(*(const FLOAT (*)[n * 2]) x),[x] "a"(x),
|
||||
"m"(*(const FLOAT (*)[2]) alpha),[alpha] "a"(alpha)
|
||||
: "cc", "r1", "v0", "v1", "v16", "v17", "v18", "v19", "v20", "v21",
|
||||
"v22", "v23");
|
||||
"vl %%v22,0(%[y])\n\t"
|
||||
"vl %%v23,16(%[y])\n\t"
|
||||
"vfmadb %%v22,%%v16,%%v20,%%v22\n\t"
|
||||
"vfmadb %%v22,%%v18,%%v21,%%v22\n\t"
|
||||
"vfmadb %%v23,%%v17,%%v20,%%v23\n\t"
|
||||
"vfmadb %%v23,%%v19,%%v21,%%v23\n\t"
|
||||
"vst %%v22,0(%[y])\n\t"
|
||||
"vst %%v23,16(%[y])\n\t"
|
||||
: "+m"(*(struct { FLOAT x[4]; } *) y),[n] "+&r"(n)
|
||||
: [y] "a"(y), "m"(*(const struct { FLOAT x[n * 2]; } *) ap0),[ap0] "a"(ap0),
|
||||
"m"(*(const struct { FLOAT x[n * 2]; } *) ap1),[ap1] "a"(ap1),
|
||||
"m"(*(const struct { FLOAT x[n * 2]; } *) x),[x] "a"(x),
|
||||
"m"(*(const struct { FLOAT x[2]; } *) alpha),[alpha] "a"(alpha)
|
||||
: "cc", "r1", "v0", "v1", "v16", "v17", "v18", "v19", "v20", "v21",
|
||||
"v22", "v23");
|
||||
}
|
||||
|
||||
static void zgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y,
|
||||
FLOAT *alpha) {
|
||||
__asm__("vzero %%v16\n\t"
|
||||
"vzero %%v17\n\t"
|
||||
"xgr %%r1,%%r1\n\t"
|
||||
"srlg %[n],%[n],1\n\t"
|
||||
"0:\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap])\n\t"
|
||||
"pfd 1,1024(%%r1,%[x])\n\t"
|
||||
"vl %%v0,0(%%r1,%[x])\n\t"
|
||||
"vzero %%v17\n\t"
|
||||
"xgr %%r1,%%r1\n\t"
|
||||
"srlg %[n],%[n],1\n\t"
|
||||
"0:\n\t"
|
||||
"pfd 1,1024(%%r1,%[ap])\n\t"
|
||||
"pfd 1,1024(%%r1,%[x])\n\t"
|
||||
"vl %%v0,0(%%r1,%[x])\n\t"
|
||||
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
|
||||
"vleg %%v1,8(%%r1,%[x]),0\n\t"
|
||||
"wflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,0(%%r1,%[x]),1\n\t"
|
||||
"vleg %%v1,8(%%r1,%[x]),0\n\t"
|
||||
"wflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,0(%%r1,%[x]),1\n\t"
|
||||
#else
|
||||
"vleg %%v1,0(%%r1,%[x]),1\n\t"
|
||||
"vflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,8(%%r1,%[x]),0\n\t"
|
||||
"vleg %%v1,0(%%r1,%[x]),1\n\t"
|
||||
"vflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,8(%%r1,%[x]),0\n\t"
|
||||
#endif
|
||||
"vlrepg %%v18,0(%%r1,%[ap])\n\t"
|
||||
"vlrepg %%v19,8(%%r1,%[ap])\n\t"
|
||||
"vfmadb %%v16,%%v18,%%v0,%%v16\n\t"
|
||||
"vfmadb %%v17,%%v19,%%v1,%%v17\n\t"
|
||||
"vl %%v0,16(%%r1,%[x])\n\t"
|
||||
"vlrepg %%v18,0(%%r1,%[ap])\n\t"
|
||||
"vlrepg %%v19,8(%%r1,%[ap])\n\t"
|
||||
"vfmadb %%v16,%%v18,%%v0,%%v16\n\t"
|
||||
"vfmadb %%v17,%%v19,%%v1,%%v17\n\t"
|
||||
"vl %%v0,16(%%r1,%[x])\n\t"
|
||||
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
|
||||
"vleg %%v1,24(%%r1,%[x]),0\n\t"
|
||||
"wflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,16(%%r1,%[x]),1\n\t"
|
||||
"vleg %%v1,24(%%r1,%[x]),0\n\t"
|
||||
"wflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,16(%%r1,%[x]),1\n\t"
|
||||
#else
|
||||
"vleg %%v1,16(%%r1,%[x]),1\n\t"
|
||||
"vflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,24(%%r1,%[x]),0\n\t"
|
||||
"vleg %%v1,16(%%r1,%[x]),1\n\t"
|
||||
"vflcdb %%v1,%%v1\n\t"
|
||||
"vleg %%v1,24(%%r1,%[x]),0\n\t"
|
||||
#endif
|
||||
"vlrepg %%v18,16(%%r1,%[ap])\n\t"
|
||||
"vlrepg %%v19,24(%%r1,%[ap])\n\t"
|
||||
"vfmadb %%v16,%%v18,%%v0,%%v16\n\t"
|
||||
"vfmadb %%v17,%%v19,%%v1,%%v17\n\t"
|
||||
"agfi %%r1,32\n\t"
|
||||
"brctg %[n],0b\n\t"
|
||||
"vfadb %%v16,%%v16,%%v17\n\t"
|
||||
"vpdi %%v17,%%v16,%%v16,4\n\t"
|
||||
"vlrepg %%v18,16(%%r1,%[ap])\n\t"
|
||||
"vlrepg %%v19,24(%%r1,%[ap])\n\t"
|
||||
"vfmadb %%v16,%%v18,%%v0,%%v16\n\t"
|
||||
"vfmadb %%v17,%%v19,%%v1,%%v17\n\t"
|
||||
"agfi %%r1,32\n\t"
|
||||
"brctg %[n],0b\n\t"
|
||||
"vfadb %%v16,%%v16,%%v17\n\t"
|
||||
"vpdi %%v17,%%v16,%%v16,4\n\t"
|
||||
#if !defined(XCONJ)
|
||||
"vlrepg %%v18,0(%[alpha])\n\t"
|
||||
"vleg %%v19,8(%[alpha]),0\n\t"
|
||||
"wflcdb %%v19,%%v19\n\t"
|
||||
"vleg %%v19,8(%[alpha]),1\n\t"
|
||||
"vlrepg %%v18,0(%[alpha])\n\t"
|
||||
"vleg %%v19,8(%[alpha]),0\n\t"
|
||||
"wflcdb %%v19,%%v19\n\t"
|
||||
"vleg %%v19,8(%[alpha]),1\n\t"
|
||||
#else
|
||||
"vleg %%v18,0(%[alpha]),1\n\t"
|
||||
"vflcdb %%v18,%%v18\n\t"
|
||||
"vleg %%v18,0(%[alpha]),0\n\t"
|
||||
"vlrepg %%v19,8(%[alpha])\n\t"
|
||||
"vleg %%v18,0(%[alpha]),1\n\t"
|
||||
"vflcdb %%v18,%%v18\n\t"
|
||||
"vleg %%v18,0(%[alpha]),0\n\t"
|
||||
"vlrepg %%v19,8(%[alpha])\n\t"
|
||||
#endif
|
||||
"vl %%v0,0(%[y])\n\t"
|
||||
"vfmadb %%v0,%%v16,%%v18,%%v0\n\t"
|
||||
"vfmadb %%v0,%%v17,%%v19,%%v0\n\t"
|
||||
"vst %%v0,0(%[y])\n\t"
|
||||
: "+m"(*(FLOAT (*)[2]) y),[n] "+&r"(n)
|
||||
: [y] "a"(y), "m"(*(const FLOAT (*)[n * 2]) ap),[ap] "a"(ap),
|
||||
"m"(*(const FLOAT (*)[n * 2]) x),[x] "a"(x),
|
||||
"m"(*(const FLOAT (*)[2]) alpha),[alpha] "a"(alpha)
|
||||
: "cc", "r1", "v0", "v1", "v16", "v17", "v18", "v19");
|
||||
"vl %%v0,0(%[y])\n\t"
|
||||
"vfmadb %%v0,%%v16,%%v18,%%v0\n\t"
|
||||
"vfmadb %%v0,%%v17,%%v19,%%v0\n\t"
|
||||
"vst %%v0,0(%[y])\n\t"
|
||||
: "+m"(*(struct { FLOAT x[2]; } *) y),[n] "+&r"(n)
|
||||
: [y] "a"(y), "m"(*(const struct { FLOAT x[n * 2]; } *) ap),[ap] "a"(ap),
|
||||
"m"(*(const struct { FLOAT x[n * 2]; } *) x),[x] "a"(x),
|
||||
"m"(*(const struct { FLOAT x[2]; } *) alpha),[alpha] "a"(alpha)
|
||||
: "cc", "r1", "v0", "v1", "v16", "v17", "v18", "v19");
|
||||
}
|
||||
|
||||
static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src) {
|
||||
|
||||
Reference in New Issue
Block a user