diff options
Diffstat (limited to 'pl/math/sv_log10_2u5.c')
| -rw-r--r-- | pl/math/sv_log10_2u5.c | 94 |
1 files changed, 40 insertions, 54 deletions
diff --git a/pl/math/sv_log10_2u5.c b/pl/math/sv_log10_2u5.c index 884e2011d2f8..f55e068fd442 100644 --- a/pl/math/sv_log10_2u5.c +++ b/pl/math/sv_log10_2u5.c @@ -6,84 +6,70 @@ */ #include "sv_math.h" -#include "math_config.h" #include "pl_sig.h" #include "pl_test.h" +#include "poly_sve_f64.h" -#if SV_SUPPORTED - -#define OFF 0x3fe6900900000000 +#define Min 0x0010000000000000 +#define Max 0x7ff0000000000000 +#define Thres 0x7fe0000000000000 /* Max - Min. */ +#define Off 0x3fe6900900000000 #define N (1 << V_LOG10_TABLE_BITS) -#define A(i) __v_log10_data.poly[i] - -static inline sv_f64_t -specialcase (sv_f64_t x, sv_f64_t y, svbool_t special) +static svfloat64_t NOINLINE +special_case (svfloat64_t x, svfloat64_t y, svbool_t special) { return sv_call_f64 (log10, x, y, special); } -/* SVE log10 algorithm. Maximum measured error is 2.46 ulps. - __sv_log10(0x1.131956cd4b627p+0) got 0x1.fffbdf6eaa669p-6 - want 0x1.fffbdf6eaa667p-6. */ -sv_f64_t -__sv_log10_x (sv_f64_t x, const svbool_t pg) +/* SVE log10 algorithm. + Maximum measured error is 2.46 ulps. + SV_NAME_D1 (log10)(0x1.131956cd4b627p+0) got 0x1.fffbdf6eaa669p-6 + want 0x1.fffbdf6eaa667p-6. */ +svfloat64_t SV_NAME_D1 (log10) (svfloat64_t x, const svbool_t pg) { - sv_u64_t ix = sv_as_u64_f64 (x); - sv_u64_t top = svlsr_n_u64_x (pg, ix, 48); - - svbool_t is_special_case - = svcmpge_n_u64 (pg, svsub_n_u64_x (pg, top, 0x0010), 0x07ff0 - 0x0010); + svuint64_t ix = svreinterpret_u64 (x); + svbool_t special = svcmpge (pg, svsub_x (pg, ix, Min), Thres); - /* x = 2^k z; where z is in range [OFF,2*OFF) and exact. + /* 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. */ - sv_u64_t tmp = svsub_n_u64_x (pg, ix, OFF); - sv_u64_t i - = sv_mod_n_u64_x (pg, svlsr_n_u64_x (pg, tmp, 52 - V_LOG10_TABLE_BITS), N); - sv_f64_t k - = sv_to_f64_s64_x (pg, svasr_n_s64_x (pg, sv_as_s64_u64 (tmp), 52)); - sv_f64_t z = sv_as_f64_u64 ( - svsub_u64_x (pg, ix, svand_n_u64_x (pg, tmp, 0xfffULL << 52))); + svuint64_t tmp = svsub_x (pg, ix, Off); + svuint64_t i = svlsr_x (pg, tmp, 51 - V_LOG10_TABLE_BITS); + i = svand_x (pg, i, (N - 1) << 1); + svfloat64_t k = svcvt_f64_x (pg, svasr_x (pg, svreinterpret_s64 (tmp), 52)); + svfloat64_t z = svreinterpret_f64 ( + svsub_x (pg, ix, svand_x (pg, tmp, 0xfffULL << 52))); /* log(x) = k*log(2) + log(c) + log(z/c). */ - - sv_u64_t idx = svmul_n_u64_x (pg, i, 2); - sv_f64_t invc = sv_lookup_f64_x (pg, &__v_log10_data.tab[0].invc, idx); - sv_f64_t logc = sv_lookup_f64_x (pg, &__v_log10_data.tab[0].log10c, idx); + svfloat64_t invc = svld1_gather_index (pg, &__v_log10_data.table[0].invc, i); + svfloat64_t logc + = svld1_gather_index (pg, &__v_log10_data.table[0].log10c, i); /* We approximate log(z/c) with a polynomial P(x) ~= log(x + 1): r = z/c - 1 (we look up precomputed 1/c) log(z/c) ~= P(r). */ - sv_f64_t r = sv_fma_f64_x (pg, z, invc, sv_f64 (-1.0)); + svfloat64_t r = svmad_x (pg, invc, z, -1.0); /* hi = log(c) + k*log(2). */ - sv_f64_t w = sv_fma_n_f64_x (pg, __v_log10_data.invln10, r, logc); - sv_f64_t hi = sv_fma_n_f64_x (pg, __v_log10_data.log10_2, k, w); + svfloat64_t w = svmla_x (pg, logc, r, __v_log10_data.invln10); + svfloat64_t hi = svmla_x (pg, w, k, __v_log10_data.log10_2); /* y = r2*(A0 + r*A1 + r2*(A2 + r*A3 + r2*A4)) + hi. */ - sv_f64_t r2 = svmul_f64_x (pg, r, r); - sv_f64_t y = sv_fma_n_f64_x (pg, A (3), r, sv_f64 (A (2))); - sv_f64_t p = sv_fma_n_f64_x (pg, A (1), r, sv_f64 (A (0))); - y = sv_fma_n_f64_x (pg, A (4), r2, y); - y = sv_fma_f64_x (pg, y, r2, p); - y = sv_fma_f64_x (pg, y, r2, hi); + svfloat64_t r2 = svmul_x (pg, r, r); + svfloat64_t y = sv_pw_horner_4_f64_x (pg, r, r2, __v_log10_data.poly); - if (unlikely (svptest_any (pg, is_special_case))) - { - return specialcase (x, y, is_special_case); - } - return y; + if (unlikely (svptest_any (pg, special))) + return special_case (x, svmla_x (svnot_z (pg, special), hi, r2, y), + special); + return svmla_x (pg, hi, r2, y); } -PL_ALIAS (__sv_log10_x, _ZGVsMxv_log10) - PL_SIG (SV, D, 1, log10, 0.01, 11.1) -PL_TEST_ULP (__sv_log10, 1.97) -PL_TEST_INTERVAL (__sv_log10, -0.0, -0x1p126, 100) -PL_TEST_INTERVAL (__sv_log10, 0x1p-149, 0x1p-126, 4000) -PL_TEST_INTERVAL (__sv_log10, 0x1p-126, 0x1p-23, 50000) -PL_TEST_INTERVAL (__sv_log10, 0x1p-23, 1.0, 50000) -PL_TEST_INTERVAL (__sv_log10, 1.0, 100, 50000) -PL_TEST_INTERVAL (__sv_log10, 100, inf, 50000) -#endif +PL_TEST_ULP (SV_NAME_D1 (log10), 1.97) +PL_TEST_INTERVAL (SV_NAME_D1 (log10), -0.0, -0x1p126, 100) +PL_TEST_INTERVAL (SV_NAME_D1 (log10), 0x1p-149, 0x1p-126, 4000) +PL_TEST_INTERVAL (SV_NAME_D1 (log10), 0x1p-126, 0x1p-23, 50000) +PL_TEST_INTERVAL (SV_NAME_D1 (log10), 0x1p-23, 1.0, 50000) +PL_TEST_INTERVAL (SV_NAME_D1 (log10), 1.0, 100, 50000) +PL_TEST_INTERVAL (SV_NAME_D1 (log10), 100, inf, 50000) |