Skip to content

Commit 7a6d6d6

Browse files
NexesenexNexesenex
committed
Revert "ggml-cpu : optimize RVV kernels (ggml-org#15720)"
1 parent 17c358e commit 7a6d6d6

File tree

4 files changed

+106
-272
lines changed

4 files changed

+106
-272
lines changed

‎ggml/src/ggml-cpu/arch/riscv/quants.c‎

Lines changed: 96 additions & 210 deletions
Original file line numberDiff line numberDiff line change
@@ -1270,179 +1270,103 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
12701270
const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
12711271
const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
12721272

1273-
float ftmp, ft2;
1274-
const uint8_t * restrict q40;
1275-
const uint8_t * restrict q41;
1276-
const uint8_t * restrict q42;
1277-
const uint8_t * restrict q43;
1278-
const int8_t * restrict q80;
1279-
const int8_t * restrict q81;
1280-
const int8_t * restrict q82;
1281-
const int8_t * restrict q83;
1282-
int s0, s1, s2, s3;
1283-
1273+
int tmp, tmp2, sumi;
12841274
__asm__ __volatile__(
1285-
"li %[s1], 8\n\t"
1286-
"vsetivli zero, 4, e32, m1, ta, ma\n\t"
1287-
"vle32.v v1, (%[s6b])\n\t"
1288-
"vslide1down.vx v1, v1, zero\n\t"
1289-
"vmv.v.x v16, zero\n\t"
1275+
"vsetivli zero, 12, e8, m1\n\t"
1276+
"vle8.v v1, (%[s6b])\n\t" // {aux[0], aux[1], aux[2]}
1277+
"vsetivli zero, 4, e32, m1\n\t"
12901278
"vslidedown.vi v2, v1, 2\n\t"
12911279
"vmv1r.v v3, v2\n\t"
12921280
"vslideup.vi v2, v3, 1\n\t" // {aux[2], aux[2]}
1293-
"vsetivli zero, 2, e32, m1, ta, ma\n\t"
1281+
"vsetivli zero, 2, e32, m1\n\t"
12941282
"vmv.v.i v4, 4\n\t"
12951283
"vand.vx v8, v1, %[kmask1]\n\t"
12961284
"vslide1up.vx v5, v4, zero\n\t" // {0, 4}
12971285
"vsrl.vi v6, v1, 6\n\t"
12981286
"vsrl.vv v7, v2, v5\n\t"
1299-
"vsse32.v v8, (%[utmp]), %[s1]\n\t"
13001287
"vand.vx v0, v6, %[kmask3]\n\t"
13011288
"vand.vx v2, v7, %[kmask2]\n\t"
13021289
"vsll.vi v6, v0, 4\n\t"
1303-
"addi %[s0], %[utmp], 4\n\t"
1290+
"li %[t2], 8\n\t"
1291+
"addi %[t1], %[utmp], 4\n\t"
13041292
"vor.vv v1, v6, v2\n\t"
1305-
"vsse32.v v1, (%[s0]), %[s1]\n\t"
1306-
"vsetivli zero, 8, e16, m1, ta, ma\n\t"
1293+
"vsse32.v v8, (%[utmp]), %[t2]\n\t"
1294+
"vsse32.v v1, (%[t1]), %[t2]\n\t"
1295+
"vsetivli zero, 8, e16, m1\n\t"
13071296
"vle32.v v2, (%[bsums])\n\t"
13081297
"vnsrl.wi v0, v2, 0\n\t"
13091298
"vnsrl.wi v1, v2, 16\n\t"
13101299
"vadd.vv v2, v0, v1\n\t"
13111300
"vle8.v v3, (%[mins])\n\t"
13121301
"vzext.vf2 v4, v3\n\t"
13131302
"vwmul.vv v6, v4, v2\n\t"
1314-
"vsetivli zero, 4, e32, m1, ta, ma\n\t"
1315-
"vredsum.vs v0, v6, v16\n\t"
1316-
"vredsum.vs v0, v7, v0\n\t"
1317-
"vfcvt.f.x.v v0, v0\n\t"
1318-
"vfmv.f.s %[ftmp], v0\n\t"
1319-
"vsetivli zero, 16, e8, m1, ta, ma\n\t"
1320-
"vle8.v v0, (%[xs])\n\t"
1321-
"fnmsub.s %[sumf], %[dmin], %[ftmp], %[sumf]\n\t"
1322-
"addi %[q40], %[xs], 64\n\t"
1323-
"addi %[q41], %[xs], 16\n\t"
1324-
"addi %[q42], %[xs], 32\n\t"
1325-
"addi %[q43], %[xs], 48\n\t"
1326-
"addi %[q80], %[ys], 64\n\t"
1327-
"vle8.v v1, (%[q41])\n\t"
1328-
"vle8.v v2, (%[q42])\n\t"
1329-
"addi %[q81], %[ys], 16\n\t"
1330-
"addi %[q41], %[q41], 64\n\t"
1331-
"addi %[q82], %[ys], 32\n\t"
1332-
"vle8.v v3, (%[q43])\n\t"
1333-
"vle8.v v8, (%[ys])\n\t"
1334-
"addi %[q42], %[q42], 64\n\t"
1335-
"addi %[q83], %[ys], 48\n\t"
1336-
"addi %[q43], %[q43], 64\n\t"
1337-
"vsrl.vi v4, v0, 4\n\t"
1338-
"vle8.v v9, (%[q81])\n\t"
1339-
"vle8.v v10, (%[q82])\n\t"
1340-
"vand.vi v0, v0, 0xF\n\t"
1341-
"addi %[q81], %[q81], 64\n\t"
1342-
"vsrl.vi v5, v1, 4\n\t"
1343-
"addi %[q82], %[q82], 64\n\t"
1344-
"vle8.v v11, (%[q83])\n\t"
1345-
"vle8.v v12, (%[q80])\n\t"
1346-
"vand.vi v1, v1, 0xF\n\t"
1347-
"addi %[q83], %[q83], 64\n\t"
1348-
"vsrl.vi v6, v2, 4\n\t"
1349-
"addi %[q80], %[q80], 64\n\t"
1350-
"vle8.v v13, (%[q81])\n\t"
1351-
"vle8.v v14, (%[q82])\n\t"
1352-
"vand.vi v2, v2, 0xF\n\t"
1353-
"addi %[q81], %[q81], 64\n\t"
1354-
"vsrl.vi v7, v3, 4\n\t"
1355-
"addi %[q82], %[q82], 64\n\t"
1356-
"vwmul.vv v16, v0, v8\n\t"
1357-
"vle8.v v15, (%[q83])\n\t"
1358-
"vle8.v v0, (%[q40])\n\t"
1359-
"vand.vi v3, v3, 0xF\n\t"
1360-
"addi %[q83], %[q83], 64\n\t"
1361-
"vwmul.vv v24, v2, v12\n\t"
1362-
"vwmul.vv v20, v4, v10\n\t"
1363-
"vwmul.vv v28, v6, v14\n\t"
1364-
"vwmacc.vv v16, v1, v9\n\t"
1365-
"vle8.v v1, (%[q41])\n\t"
1366-
"vle8.v v2, (%[q42])\n\t"
1367-
"vwmacc.vv v24, v3, v13\n\t"
1368-
"vwmacc.vv v20, v5, v11\n\t"
1369-
"vwmacc.vv v28, v7, v15\n\t"
1370-
"addi %[q40], %[q80], 64\n\t"
1371-
"addi %[q41], %[q81], 64\n\t"
1372-
"vle8.v v3, (%[q43])\n\t"
1373-
"vle8.v v8, (%[q80])\n\t"
1374-
"addi %[q42], %[q82], 64\n\t"
1375-
"addi %[q43], %[q83], 64\n\t"
1376-
"vsrl.vi v4, v0, 4\n\t"
1377-
"vle8.v v9, (%[q81])\n\t"
1378-
"vle8.v v10, (%[q82])\n\t"
1379-
"vand.vi v0, v0, 0xF\n\t"
1380-
"vsrl.vi v5, v1, 4\n\t"
1381-
"vsrl.vi v7, v3, 4\n\t"
1382-
"vand.vi v3, v3, 0xF\n\t"
1383-
"vle8.v v11, (%[q83])\n\t"
1384-
"vle8.v v12, (%[q40])\n\t"
1385-
"vand.vi v1, v1, 0xF\n\t"
1386-
"vsrl.vi v6, v2, 4\n\t"
1387-
"vand.vi v2, v2, 0xF\n\t"
1388-
"vwmul.vv v18, v0, v8\n\t"
1389-
"vle8.v v13, (%[q41])\n\t"
1390-
"vle8.v v14, (%[q42])\n\t"
1391-
"vwmul.vv v26, v2, v12\n\t"
1392-
"vwmul.vv v22, v4, v10\n\t"
1393-
"vwmul.vv v30, v6, v14\n\t"
1394-
"vwmacc.vv v18, v1, v9\n\t"
1395-
"vle8.v v15, (%[q43])\n\t"
1396-
"vwmacc.vv v26, v3, v13\n\t"
1397-
"vwmacc.vv v22, v5, v11\n\t"
1398-
"vwmacc.vv v30, v7, v15\n\t"
13991303
"vmv.v.x v0, zero\n\t"
1400-
"vsetivli zero, 16, e16, m2, ta, ma\n\t"
1401-
"vwredsum.vs v4, v16, v0\n\t"
1402-
"lbu %[s0], 0(%[scale])\n\t"
1403-
"vwredsum.vs v5, v20, v0\n\t"
1404-
"lbu %[s1], 1(%[scale])\n\t"
1405-
"vwredsum.vs v6, v24, v0\n\t"
1406-
"lbu %[s2], 2(%[scale])\n\t"
1407-
"vwredsum.vs v7, v28, v0\n\t"
1408-
"lbu %[s3], 3(%[scale])\n\t"
1409-
"vwredsum.vs v8, v18, v0\n\t"
1410-
"lbu %[q40], 4(%[scale])\n\t"
1411-
"vwredsum.vs v9, v22, v0\n\t"
1412-
"lbu %[q41], 5(%[scale])\n\t"
1413-
"vwredsum.vs v10, v26, v0\n\t"
1414-
"lbu %[q42], 6(%[scale])\n\t"
1415-
"vwredsum.vs v11, v30, v0\n\t"
1416-
"lbu %[q43], 7(%[scale])\n\t"
1417-
"vsetivli zero, 4, e32, m1, ta, ma\n\t"
1418-
"vmul.vx v0, v4, %[s0]\n\t"
1419-
"vmul.vx v1, v8, %[q40]\n\t"
1420-
"vmacc.vx v0, %[s1], v5\n\t"
1421-
"vmacc.vx v1, %[q41], v9\n\t"
1422-
"vmacc.vx v0, %[s2], v6\n\t"
1423-
"vmacc.vx v1, %[q42], v10\n\t"
1424-
"vmacc.vx v0, %[s3], v7\n\t"
1425-
"vmacc.vx v1, %[q43], v11\n\t"
1426-
"vfcvt.f.x.v v0, v0\n\t"
1427-
"vfcvt.f.x.v v1, v1\n\t"
1428-
"vfmv.f.s %[ft2], v0\n\t"
1429-
"vfmv.f.s %[ftmp], v1\n\t"
1430-
"fadd.s %[ft2], %[ft2], %[ftmp]\n\t"
1431-
"fmadd.s %[sumf], %[d], %[ft2], %[sumf]"
1432-
: [ftmp] "=&f" (ftmp), [sumf] "+&f" (sumf), [ft2] "=&f" (ft2)
1433-
, [s0] "=&r" (s0), [s1] "=&r" (s1), [s2] "=&r" (s2), [s3] "=&r" (s3)
1434-
, [q40] "=&r" (q40), [q41] "=&r" (q41), [q42] "=&r" (q42), [q43] "=&r" (q43)
1435-
, [q80] "=&r" (q80), [q81] "=&r" (q81), [q82] "=&r" (q82), [q83] "=&r" (q83)
1436-
: [d] "f" (d), [ys] "r" (y[i].qs), [xs] "r" (x[i].qs), [scale] "r" (scales)
1437-
, [bsums] "r" (y[i].bsums), [mins] "r" (mins), [utmp] "r" (utmp)
1438-
, [s6b] "r" (&x[i]), [kmask1] "r" (kmask1), [dmin] "f" (dmin)
1304+
"vsetivli zero, 8, e32, m2\n\t"
1305+
"vredsum.vs v0, v6, v0\n\t"
1306+
"vmv.x.s %[sumi], v0"
1307+
: [t1] "=&r" (tmp), [t2] "=&r" (tmp2), [sumi] "=&r" (sumi)
1308+
: [bsums] "r" (y[i].bsums), [mins] "r" (mins), [utmp] "r" (utmp)
1309+
, [s6b] "r" (x[i].scales), [kmask1] "r" (kmask1)
14391310
, [kmask2] "r" (kmask2), [kmask3] "r" (kmask3)
14401311
: "memory"
14411312
, "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
14421313
, "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
14431314
, "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
14441315
, "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
14451316
);
1317+
sumf -= dmin * sumi;
1318+
1319+
const uint8_t * restrict q4 = x[i].qs;
1320+
const int8_t * restrict q8 = y[i].qs;
1321+
1322+
sumi = 0;
1323+
const uint8_t * scale = scales;
1324+
1325+
for (int j = 0; j < QK_K/128; ++j) {
1326+
int vl128 = 128, vl64 = 64, vl32 = 32;
1327+
__asm__ __volatile__(
1328+
"vsetvli zero, %[vl128], e8, m8\n\t"
1329+
"vle8.v v8, (%[q8])\n\t"
1330+
"vsetvli zero, %[vl64], e8, m4\n\t"
1331+
"vle8.v v0, (%[q4])\n\t"
1332+
"vsrl.vi v4, v0, 4\n\t"
1333+
"vand.vi v0, v0, 0xF\n\t"
1334+
"vsetvli zero, %[vl32], e8, m2\n\t"
1335+
"vwmul.vv v28, v6, v14\n\t"
1336+
"vwmul.vv v20, v4, v10\n\t"
1337+
"vwmul.vv v24, v2, v12\n\t"
1338+
"vwmul.vv v16, v0, v8\n\t"
1339+
"vsetivli zero, 4, e32, m1\n\t"
1340+
"vle8.v v2, (%[scale])\n\t"
1341+
"vmv.v.x v0, zero\n\t"
1342+
"vzext.vf4 v1, v2\n\t"
1343+
"vsetvli zero, %[vl32], e16, m4\n\t"
1344+
"vwredsum.vs v6, v24, v0\n\t"
1345+
"vwredsum.vs v7, v28, v0\n\t"
1346+
"vwredsum.vs v4, v16, v0\n\t"
1347+
"vwredsum.vs v5, v20, v0\n\t"
1348+
"vsetivli zero, 4, e32, m1\n\t"
1349+
"vslideup.vi v6, v7, 1\n\t"
1350+
"vslideup.vi v4, v5, 1\n\t"
1351+
"vslideup.vi v4, v6, 2\n\t"
1352+
"vmul.vv v8, v4, v1\n\t"
1353+
"vredsum.vs v0, v8, v0\n\t"
1354+
"vmv.x.s %[tmp], v0\n\t"
1355+
"add %[sumi], %[sumi], %[tmp]"
1356+
: [tmp] "=&r" (tmp), [sumi] "+&r" (sumi)
1357+
: [vl128] "r" (vl128), [vl64] "r" (vl64), [vl32] "r" (vl32)
1358+
, [q4] "r" (q4), [q8] "r" (q8), [scale] "r" (scale)
1359+
: "memory"
1360+
, "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
1361+
, "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
1362+
, "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
1363+
, "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
1364+
);
1365+
1366+
q4 += 64; q8 += 128; scale += 4;
1367+
}
1368+
1369+
sumf += d * sumi;
14461370
}
14471371
break;
14481372
default:
@@ -1769,8 +1693,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
17691693
case 128:
17701694
for (int i = 0; i < nb; ++i) {
17711695

1772-
__builtin_prefetch(&x[i + 1].d, 0, 1);
1773-
17741696
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
17751697

17761698
const uint8_t * restrict q6 = x[i].ql;
@@ -1779,59 +1701,23 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
17791701

17801702
const int8_t * restrict scale = x[i].scales;
17811703

1782-
int q6h;
1783-
float ftmp;
1704+
int sum_t = 0;
1705+
int t0;
17841706

17851707
for (int j = 0; j < QK_K/128; ++j) {
17861708
__asm__ __volatile__(
1787-
"addi %[q6h], %[q6], 32\n\t"
1788-
"ld t0, 0(%[scale])\n\t"
1789-
"addi %[scale], %[scale], 8\n\t"
1790-
"slli t6, t0, 1 * 8\n\t"
1791-
"lb zero, 0(%[q6])\n\t"
1792-
"slli t5, t0, 2 * 8\n\t"
1793-
"slli t4, t0, 3 * 8\n\t"
1794-
"lb zero, 0(%[q6h])\n\t"
1795-
"slli t3, t0, 4 * 8\n\t"
1796-
"slli t2, t0, 5 * 8\n\t"
1797-
"lb zero, 0(%[qh])\n\t"
1798-
"lb zero, 31(%[q6h])\n\t"
1799-
"slli t1, t0, 6 * 8\n\t"
1800-
"srai a7, t0, 56\n\t"
18011709
"vsetvli zero, %[vl32], e8, m2\n\t"
1802-
"vle8.v v8, (%[q6])\n\t"
1803-
"srai t6, t6, 56\n\t"
1804-
"srai t5, t5, 56\n\t"
1805-
"srai t4, t4, 56\n\t"
1806-
"srai t3, t3, 56\n\t"
1807-
"vle8.v v10, (%[q6h])\n\t"
1808-
"addi %[q6], %[q6], 64\n\t"
1809-
"slli t0, t0, 7 * 8\n\t"
1810-
"srai t2, t2, 56\n\t"
1811-
"srai t1, t1, 56\n\t"
1812-
"srai t0, t0, 56\n\t"
18131710
"vle8.v v4, (%[qh])\n\t"
1814-
"vsrl.vi v12, v8, 4\n\t"
1815-
"vsrl.vi v14, v10, 4\n\t"
1816-
"lb zero, 0(%[q8])\n\t"
1817-
"vand.vi v8, v8, 0xF\n\t"
1818-
"vand.vi v10, v10, 0xF\n\t"
1819-
"lb zero, 32(%[q8])\n\t"
18201711
"vsll.vi v0, v4, 4\n\t"
18211712
"vsll.vi v2, v4, 2\n\t"
1822-
"lb zero, 64(%[q8])\n\t"
18231713
"vsrl.vi v6, v4, 2\n\t"
1714+
"vsetvli zero, %[vl64], e8, m4\n\t"
1715+
"vle8.v v8, (%[q6])\n\t"
1716+
"vsrl.vi v12, v8, 4\n\t"
1717+
"vand.vi v8, v8, 0xF\n\t"
1718+
"vsetvli zero, %[vl128], e8, m8\n\t"
18241719
"vand.vx v0, v0, %[mask]\n\t"
1825-
"lb zero, 96(%[q8])\n\t"
1826-
"vand.vx v2, v2, %[mask]\n\t"
1827-
"vand.vx v4, v4, %[mask]\n\t"
1828-
"vand.vx v6, v6, %[mask]\n\t"
18291720
"vor.vv v8, v8, v0\n\t"
1830-
"lb zero, 127(%[q8])\n\t"
1831-
"vor.vv v10, v10, v2\n\t"
1832-
"vor.vv v12, v12, v4\n\t"
1833-
"vor.vv v14, v14, v6\n\t"
1834-
"vsetvli zero, %[vl128], e8, m8\n\t"
18351721
"vle8.v v0, (%[q8])\n\t"
18361722
"vsub.vx v8, v8, %[vl32]\n\t"
18371723
"vsetvli zero, %[vl64], e8, m4\n\t"
@@ -1848,34 +1734,34 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
18481734
"vwredsum.vs v13, v28, v0\n\t"
18491735
"vwredsum.vs v14, v30, v0\n\t"
18501736
"vsetivli zero, 4, e32, m1\n\t"
1851-
"vmul.vx v0, v10, t0\n\t"
1852-
"vmul.vx v1, v9, t1\n\t"
1853-
"vmacc.vx v0, t2, v8\n\t"
1854-
"vmacc.vx v1, t3, v7\n\t"
1855-
"vmacc.vx v0, t4, v11\n\t"
1856-
"vmacc.vx v1, t5, v12\n\t"
1857-
"vmacc.vx v0, t6, v13\n\t"
1858-
"vmacc.vx v1, a7, v14\n\t"
1859-
"vadd.vv v0, v0, v1\n\t"
1860-
"vfcvt.f.x.v v0, v0\n\t"
1861-
"vfmv.f.s %[ftmp], v0\n\t"
1862-
"fmadd.s %[sumf], %[d], %[ftmp], %[sumf]"
1863-
: [q6] "+&r" (q6), [q6h] "=&r" (q6h)
1864-
, [scale] "+&r" (scale)
1865-
, [sumf] "+&f" (sumf), [ftmp] "=&f" (ftmp)
1866-
: [qh] "r" (qh), [q8] "r" (q8)
1737+
"vslideup.vi v10, v9, 1\n\t"
1738+
"vslideup.vi v8, v7, 1\n\t"
1739+
"vslideup.vi v11, v12, 1\n\t"
1740+
"vslideup.vi v13, v14, 1\n\t"
1741+
"vslideup.vi v10, v8, 2\n\t"
1742+
"vslideup.vi v11, v13, 2\n\t"
1743+
"vsetivli zero, 8, e32, m2\n\t"
1744+
"vle8.v v2, (%[scale])\n\t"
1745+
"vsext.vf4 v4, v2\n\t"
1746+
"vmul.vv v2, v4, v10\n\t"
1747+
"vredsum.vs v0, v2, v0\n\t"
1748+
"vmv.x.s %[t0], v0\n\t"
1749+
"add %[sumi], %[sumi], %[t0]"
1750+
: [sumi] "+&r" (sum_t), [t0] "=&r" (t0)
1751+
: [qh] "r" (qh), [q6] "r" (q6), [q8] "r" (q8), [scale] "r" (scale)
18671752
, [vl32] "r" (32), [vl64] "r" (64), [vl128] "r" (128)
1868-
, [mask] "r" (0x30), [d] "f" (d)
1753+
, [mask] "r" (0x30)
18691754
: "memory"
18701755
, "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
18711756
, "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
18721757
, "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
18731758
, "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
1874-
, "t0", "t1", "t2", "t3", "t4", "t5", "t6", "a7"
1875-
, "a6", "a5", "a4", "a3"
18761759
);
1877-
qh += 32; q8 += 128;
1760+
q6 += 64; qh += 32; q8 += 128; scale += 8;
18781761
}
1762+
1763+
sumf += d * sum_t;
1764+
18791765
}
18801766
break;
18811767
default:

‎ggml/src/ggml-cpu/ggml-cpu.c‎

Lines changed: 0 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -4087,13 +4087,6 @@ void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
40874087
uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
40884088
vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
40894089
}
4090-
#elif defined(__riscv_zvfh)
4091-
for (int vl; i < n; i += vl) {
4092-
vl = __riscv_vsetvl_e32m2(n - i);
4093-
vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
4094-
vfloat16m1_t vy = __riscv_vfncvt_f_f_w_f16m1(vx, vl);
4095-
__riscv_vse16_v_f16m1((_Float16 *)&y[i], vy, vl);
4096-
}
40974090
#endif
40984091
for (; i < n; ++i) {
40994092
y[i] = GGML_CPU_FP32_TO_FP16(x[i]);

0 commit comments

Comments
 (0)