diff options
Diffstat (limited to 'sysdeps/aarch64/fpu/log_sve.c')
| -rw-r--r-- | sysdeps/aarch64/fpu/log_sve.c | 52 |
1 files changed, 24 insertions, 28 deletions
diff --git a/sysdeps/aarch64/fpu/log_sve.c b/sysdeps/aarch64/fpu/log_sve.c index 93c4f1c018..0c171a4d01 100644 --- a/sysdeps/aarch64/fpu/log_sve.c +++ b/sysdeps/aarch64/fpu/log_sve.c @@ -38,43 +38,39 @@ special_case (svfloat64_t x, svfloat64_t y, svbool_t cmp) want 0x1.ffffff1cca045p-2. */ svfloat64_t SV_NAME_D1 (log) (svfloat64_t x, const svbool_t pg) { - svuint64_t ix = svreinterpret_u64_f64 (x); - svuint64_t top = svlsr_n_u64_x (pg, ix, 52); - svbool_t cmp - = svcmpge_u64 (pg, svsub_n_u64_x (pg, top, MinTop), sv_u64 (ThreshTop)); + svuint64_t ix = svreinterpret_u64 (x); + svuint64_t top = svlsr_x (pg, ix, 52); + svbool_t cmp = svcmpge (pg, svsub_x (pg, top, MinTop), sv_u64 (ThreshTop)); /* x = 2^k z; where z is in range [Off,2*Off) and exact. The range is split into N subintervals. The ith subinterval contains z and c is near its center. */ - svuint64_t tmp = svsub_n_u64_x (pg, ix, Off); - /* Equivalent to (tmp >> (52 - V_LOG_TABLE_BITS)) % N, since N is a power - of 2. */ - svuint64_t i = svand_n_u64_x ( - pg, svlsr_n_u64_x (pg, tmp, (52 - V_LOG_TABLE_BITS)), N - 1); - svint64_t k = svasr_n_s64_x (pg, svreinterpret_s64_u64 (tmp), - 52); /* Arithmetic shift. */ - svuint64_t iz - = svsub_u64_x (pg, ix, svand_n_u64_x (pg, tmp, 0xfffULL << 52)); - svfloat64_t z = svreinterpret_f64_u64 (iz); + svuint64_t tmp = svsub_x (pg, ix, Off); + /* Calculate table index = (tmp >> (52 - V_LOG_TABLE_BITS)) % N. + The actual value of i is double this due to table layout. */ + svuint64_t i + = svand_x (pg, svlsr_x (pg, tmp, (51 - V_LOG_TABLE_BITS)), (N - 1) << 1); + svint64_t k + = svasr_x (pg, svreinterpret_s64 (tmp), 52); /* Arithmetic shift. */ + svuint64_t iz = svsub_x (pg, ix, svand_x (pg, tmp, 0xfffULL << 52)); + svfloat64_t z = svreinterpret_f64 (iz); /* Lookup in 2 global lists (length N). */ - svfloat64_t invc = svld1_gather_u64index_f64 (pg, __v_log_data.invc, i); - svfloat64_t logc = svld1_gather_u64index_f64 (pg, __v_log_data.logc, i); + svfloat64_t invc = svld1_gather_index (pg, &__v_log_data.table[0].invc, i); + svfloat64_t logc = svld1_gather_index (pg, &__v_log_data.table[0].logc, i); /* log(x) = log1p(z/c-1) + log(c) + k*Ln2. */ - svfloat64_t r = svmad_n_f64_x (pg, invc, z, -1); - svfloat64_t kd = svcvt_f64_s64_x (pg, k); + svfloat64_t r = svmad_x (pg, invc, z, -1); + svfloat64_t kd = svcvt_f64_x (pg, k); /* hi = r + log(c) + k*Ln2. */ - svfloat64_t hi - = svmla_n_f64_x (pg, svadd_f64_x (pg, logc, r), kd, __v_log_data.ln2); + svfloat64_t hi = svmla_x (pg, svadd_x (pg, logc, r), kd, __v_log_data.ln2); /* y = r2*(A0 + r*A1 + r2*(A2 + r*A3 + r2*A4)) + hi. */ - svfloat64_t r2 = svmul_f64_x (pg, r, r); - svfloat64_t y = svmla_f64_x (pg, P (2), r, P (3)); - svfloat64_t p = svmla_f64_x (pg, P (0), r, P (1)); - y = svmla_f64_x (pg, y, r2, P (4)); - y = svmla_f64_x (pg, p, r2, y); - y = svmla_f64_x (pg, hi, r2, y); + svfloat64_t r2 = svmul_x (pg, r, r); + svfloat64_t y = svmla_x (pg, P (2), r, P (3)); + svfloat64_t p = svmla_x (pg, P (0), r, P (1)); + y = svmla_x (pg, y, r2, P (4)); + y = svmla_x (pg, p, r2, y); if (__glibc_unlikely (svptest_any (pg, cmp))) - return special_case (x, y, cmp); - return y; + return special_case (x, svmla_x (svnot_z (pg, cmp), hi, r2, y), cmp); + return svmla_x (pg, hi, r2, y); } |