arm/insts/fplib.cc

1212     uint16_t a_mnt, b_mnt, x, x_mnt;
1249         x_mnt = a_mnt + b_mnt;
1251         x_mnt = a_mnt - b_mnt;
1254         x_mnt = b_mnt - a_mnt;
1257     if (!x_mnt) {
1262     x_mnt = fp16_normalise(x_mnt, &x_exp);
1264     return fp16_round(x_sgn, x_exp + FP16_EXP_BITS - 3, x_mnt << 1,
1272     uint32_t a_mnt, b_mnt, x, x_mnt;
1309         x_mnt = a_mnt + b_mnt;
1311         x_mnt = a_mnt - b_mnt;
1314         x_mnt = b_mnt - a_mnt;
1317     if (!x_mnt) {
1322     x_mnt = fp32_normalise(x_mnt, &x_exp);
1324     return fp32_round(x_sgn, x_exp + FP32_EXP_BITS - 3, x_mnt << 1,
1332     uint64_t a_mnt, b_mnt, x, x_mnt;
1369         x_mnt = a_mnt + b_mnt;
1371         x_mnt = a_mnt - b_mnt;
1374         x_mnt = b_mnt - a_mnt;
1377     if (!x_mnt) {
1382     x_mnt = fp64_normalise(x_mnt, &x_exp);
1384     return fp64_round(x_sgn, x_exp + FP64_EXP_BITS - 3, x_mnt << 1,
1393     uint32_t x_mnt;
1416     x_mnt = (uint32_t)a_mnt * b_mnt;
1417     x_mnt = fp32_normalise(x_mnt, &x_exp);
1420     x_mnt = lsr32(x_mnt, FP16_BITS - 1) | !!lsl32(x_mnt, FP16_BITS + 1);
1422     return fp16_round(x_sgn, x_exp, x_mnt, mode, flags);
1430     uint64_t x_mnt;
1453     x_mnt = (uint64_t)a_mnt * b_mnt;
1454     x_mnt = fp64_normalise(x_mnt, &x_exp);
1457     x_mnt = lsr64(x_mnt, FP32_BITS - 1) | !!lsl64(x_mnt, FP32_BITS + 1);
1459     return fp32_round(x_sgn, x_exp, x_mnt, mode, flags);
1505     uint32_t x_mnt, y_mnt;
1543     x_mnt = (uint32_t)a_mnt << (FP16_MANT_BITS + 4);
1559         x_mnt = (lsr32(x_mnt, y_exp - x_exp) |
1560                  !!(x_mnt & (lsl32(1, y_exp - x_exp) - 1)));
1564         x_mnt = x_mnt + y_mnt;
1565     } else if (x_mnt >= y_mnt) {
1566         x_mnt = x_mnt - y_mnt;
1569         x_mnt = y_mnt - x_mnt;
1572     if (!x_mnt) {
1578     x_mnt = fp32_normalise(x_mnt, &x_exp);
1579     x_mnt = x_mnt >> (FP16_BITS - 1) | !!(uint16_t)(x_mnt << 1);
1581     return fp16_round(x_sgn, x_exp + scale, x_mnt, mode, flags);
1590     uint64_t x_mnt, y_mnt;
1628     x_mnt = (uint64_t)a_mnt << (FP32_MANT_BITS + 4);
1644         x_mnt = (lsr64(x_mnt, y_exp - x_exp) |
1645                  !!(x_mnt & (lsl64(1, y_exp - x_exp) - 1)));
1649         x_mnt = x_mnt + y_mnt;
1650     } else if (x_mnt >= y_mnt) {
1651         x_mnt = x_mnt - y_mnt;
1654         x_mnt = y_mnt - x_mnt;
1657     if (!x_mnt) {
1663     x_mnt = fp64_normalise(x_mnt, &x_exp);
1664     x_mnt = x_mnt >> (FP32_BITS - 1) | !!(uint32_t)(x_mnt << 1);
1666     return fp32_round(x_sgn, x_exp + scale, x_mnt, mode, flags);
1766     uint32_t x_mnt;
1792     x_mnt = ((uint32_t)a_mnt << (FP16_MANT_BITS - FP16_EXP_BITS + 3)) / b_mnt;
1793     x_mnt |= (x_mnt * b_mnt !=
1797     x_mnt = fp32_normalise(x_mnt, &x_exp);
1798     x_mnt = x_mnt >> (FP16_BITS - 1) | !!(uint16_t)(x_mnt << 1);
1800     return fp16_round(x_sgn, x_exp, x_mnt, mode, flags);
1808     uint64_t x_mnt;
1834     x_mnt = ((uint64_t)a_mnt << (FP32_MANT_BITS - FP32_EXP_BITS + 3)) / b_mnt;
1835     x_mnt |= (x_mnt * b_mnt !=
1839     x_mnt = fp64_normalise(x_mnt, &x_exp);
1840     x_mnt = x_mnt >> (FP32_BITS - 1) | !!(uint32_t)(x_mnt << 1);
1842     return fp32_round(x_sgn, x_exp, x_mnt, mode, flags);
1849     uint64_t a_mnt, b_mnt, x, x_mnt, x0_mnt, x1_mnt;
1874     x_mnt = ~(uint64_t)0 / (b_mnt >> 31);
1875     mul64x32(&x0_mnt, &x1_mnt, b_mnt, x_mnt);
1878     mul64x32(&x0_mnt, &x1_mnt, x0_mnt, x_mnt);
1886     x_mnt = x1_mnt;
1889     mul62x62(&x0_mnt, &x1_mnt, b_mnt >> 2, x_mnt + 1);
1892         ++x_mnt;
1895     x_mnt = fp64_normalise(x_mnt, &x_exp);
1897     return fp64_round(x_sgn, x_exp, x_mnt << 1 | !!c, mode, flags);
2029     uint16_t a_mnt, x_mnt;
2065     x_mnt = ((x - (1 << 18)) >> 19) + 1;
2066     t1 = (uint32_t)x_mnt * x_mnt;
2069         --x_mnt;
2072     x_mnt = fp16_normalise(x_mnt, &x_exp);
2074     return fp16_round(x_sgn, x_exp, x_mnt << 1 | (t1 != t0), mode, flags);
2081     uint32_t a_mnt, x, x_mnt;
2120     x_mnt = ((x - (1 << 5)) >> 6) + 1;
2121     t1 = (uint64_t)x_mnt * x_mnt;
2124         --x_mnt;
2127     x_mnt = fp32_normalise(x_mnt, &x_exp);
2129     return fp32_round(x_sgn, x_exp, x_mnt << 1 | (t1 != t0), mode, flags);
2136     uint64_t a_mnt, x_mnt, r, x0, x1;
2189     x_mnt = x0;
2190     x_mnt = ((x_mnt - (1 << 8)) >> 9) + 1;
2191     mul62x62(&x0, &x1, x_mnt, x_mnt);
2195         --x_mnt;
2197     x_mnt = fp64_normalise(x_mnt, &x_exp);
2199     return fp64_round(x_sgn, x_exp, x_mnt << 1 | !!c, mode, flags);
4833     uint64_t x_mnt = x_sgn ? -a : a;
4836     if (!x_mnt) {
4841     x_mnt = fp64_normalise(x_mnt, &x_exp);
4842     x_mnt = (x_mnt >> (FP64_BITS - FP16_BITS - 1) |
4843              !!(x_mnt & ((1ULL << (FP64_BITS - FP16_BITS - 1)) - 1)));
4845     return fp16_round(x_sgn, x_exp, x_mnt, mode, flags);
4853     uint64_t x_mnt = x_sgn ? -a : a;
4856     if (!x_mnt) {
4861     x_mnt = fp64_normalise(x_mnt, &x_exp);
4862     x_mnt = (x_mnt >> (FP64_BITS - FP32_BITS - 1) |
4863              !!(x_mnt & ((1ULL << (FP64_BITS - FP32_BITS - 1)) - 1)));
4865     return fp32_round(x_sgn, x_exp, x_mnt, mode, flags);
4873     uint64_t x_mnt = x_sgn ? -a : a;
4876     if (!x_mnt) {
4880     x_mnt = fp64_normalise(x_mnt, &x_exp);
4882     return fp64_round(x_sgn, x_exp, x_mnt << 1, mode, flags);