diff options
Diffstat (limited to 'pl/math/sv_logf_3u4.c')
| -rw-r--r-- | pl/math/sv_logf_3u4.c | 99 |
1 files changed, 54 insertions, 45 deletions
diff --git a/pl/math/sv_logf_3u4.c b/pl/math/sv_logf_3u4.c index 11f0b8aa12c5..967355247036 100644 --- a/pl/math/sv_logf_3u4.c +++ b/pl/math/sv_logf_3u4.c @@ -9,69 +9,78 @@ #include "pl_sig.h" #include "pl_test.h" -#if SV_SUPPORTED - -#define P(i) __sv_logf_poly[i] +static const struct data +{ + float poly_0135[4]; + float poly_246[3]; + float ln2; +} data = { + .poly_0135 = { + /* Coefficients copied from the AdvSIMD routine in math/, then rearranged so + that coeffs 0, 1, 3 and 5 can be loaded as a single quad-word, hence used + with _lane variant of MLA intrinsic. */ + -0x1.3e737cp-3f, 0x1.5a9aa2p-3f, 0x1.961348p-3f, 0x1.555d7cp-2f + }, + .poly_246 = { -0x1.4f9934p-3f, -0x1.00187cp-2f, -0x1.ffffc8p-2f }, + .ln2 = 0x1.62e43p-1f +}; -#define Ln2 (0x1.62e43p-1f) /* 0x3f317218 */ #define Min (0x00800000) #define Max (0x7f800000) +#define Thresh (0x7f000000) /* Max - Min. */ #define Mask (0x007fffff) -#define Off (0x3f2aaaab) /* 0.666667 */ +#define Off (0x3f2aaaab) /* 0.666667. */ -float -optr_aor_log_f32 (float); +float optr_aor_log_f32 (float); -static NOINLINE sv_f32_t -__sv_logf_specialcase (sv_f32_t x, sv_f32_t y, svbool_t cmp) +static svfloat32_t NOINLINE +special_case (svfloat32_t x, svfloat32_t y, svbool_t cmp) { return sv_call_f32 (optr_aor_log_f32, x, y, cmp); } -/* Optimised implementation of SVE logf, using the same algorithm and polynomial - as the Neon routine in math/. Maximum error is 3.34 ULPs: - __sv_logf(0x1.557298p+0) got 0x1.26edecp-2 - want 0x1.26ede6p-2. */ -sv_f32_t -__sv_logf_x (sv_f32_t x, const svbool_t pg) +/* Optimised implementation of SVE logf, using the same algorithm and + polynomial as the AdvSIMD routine. Maximum error is 3.34 ULPs: + SV_NAME_F1 (log)(0x1.557298p+0) got 0x1.26edecp-2 + want 0x1.26ede6p-2. */ +svfloat32_t SV_NAME_F1 (log) (svfloat32_t x, const svbool_t pg) { - sv_u32_t u = sv_as_u32_f32 (x); - svbool_t cmp - = svcmpge_u32 (pg, svsub_n_u32_x (pg, u, Min), sv_u32 (Max - Min)); + const struct data *d = ptr_barrier (&data); + + svuint32_t u = svreinterpret_u32 (x); + svbool_t cmp = svcmpge (pg, svsub_x (pg, u, Min), Thresh); /* x = 2^n * (1+r), where 2/3 < 1+r < 4/3. */ - u = svsub_n_u32_x (pg, u, Off); - sv_f32_t n = sv_to_f32_s32_x (pg, svasr_n_s32_x (pg, sv_as_s32_u32 (u), - 23)); /* Sign-extend. */ - u = svand_n_u32_x (pg, u, Mask); - u = svadd_n_u32_x (pg, u, Off); - sv_f32_t r = svsub_n_f32_x (pg, sv_as_f32_u32 (u), 1.0f); + u = svsub_x (pg, u, Off); + svfloat32_t n = svcvt_f32_x ( + pg, svasr_x (pg, svreinterpret_s32 (u), 23)); /* Sign-extend. */ + u = svand_x (pg, u, Mask); + u = svadd_x (pg, u, Off); + svfloat32_t r = svsub_x (pg, svreinterpret_f32 (u), 1.0f); /* y = log(1+r) + n*ln2. */ - sv_f32_t r2 = svmul_f32_x (pg, r, r); + svfloat32_t r2 = svmul_x (pg, r, r); /* n*ln2 + r + r2*(P6 + r*P5 + r2*(P4 + r*P3 + r2*(P2 + r*P1 + r2*P0))). */ - sv_f32_t p = sv_fma_n_f32_x (pg, P (1), r, sv_f32 (P (2))); - sv_f32_t q = sv_fma_n_f32_x (pg, P (3), r, sv_f32 (P (4))); - sv_f32_t y = sv_fma_n_f32_x (pg, P (5), r, sv_f32 (P (6))); - p = sv_fma_n_f32_x (pg, P (0), r2, p); - q = sv_fma_f32_x (pg, p, r2, q); - y = sv_fma_f32_x (pg, q, r2, y); - p = sv_fma_n_f32_x (pg, Ln2, n, r); - y = sv_fma_f32_x (pg, y, r2, p); + svfloat32_t p_0135 = svld1rq (svptrue_b32 (), &d->poly_0135[0]); + svfloat32_t p = svmla_lane (sv_f32 (d->poly_246[0]), r, p_0135, 1); + svfloat32_t q = svmla_lane (sv_f32 (d->poly_246[1]), r, p_0135, 2); + svfloat32_t y = svmla_lane (sv_f32 (d->poly_246[2]), r, p_0135, 3); + p = svmla_lane (p, r2, p_0135, 0); + + q = svmla_x (pg, q, r2, p); + y = svmla_x (pg, y, r2, q); + p = svmla_x (pg, r, n, d->ln2); if (unlikely (svptest_any (pg, cmp))) - return __sv_logf_specialcase (x, y, cmp); - return y; + return special_case (x, svmla_x (svnot_z (pg, cmp), p, r2, y), cmp); + return svmla_x (pg, p, r2, y); } -PL_ALIAS (__sv_logf_x, _ZGVsMxv_logf) - PL_SIG (SV, F, 1, log, 0.01, 11.1) -PL_TEST_ULP (__sv_logf, 2.85) -PL_TEST_INTERVAL (__sv_logf, -0.0, -0x1p126, 100) -PL_TEST_INTERVAL (__sv_logf, 0x1p-149, 0x1p-126, 4000) -PL_TEST_INTERVAL (__sv_logf, 0x1p-126, 0x1p-23, 50000) -PL_TEST_INTERVAL (__sv_logf, 0x1p-23, 1.0, 50000) -PL_TEST_INTERVAL (__sv_logf, 1.0, 100, 50000) -PL_TEST_INTERVAL (__sv_logf, 100, inf, 50000) -#endif // SV_SUPPORTED +PL_TEST_ULP (SV_NAME_F1 (log), 2.85) +PL_TEST_INTERVAL (SV_NAME_F1 (log), -0.0, -inf, 100) +PL_TEST_INTERVAL (SV_NAME_F1 (log), 0, 0x1p-126, 100) +PL_TEST_INTERVAL (SV_NAME_F1 (log), 0x1p-126, 0x1p-23, 50000) +PL_TEST_INTERVAL (SV_NAME_F1 (log), 0x1p-23, 1.0, 50000) +PL_TEST_INTERVAL (SV_NAME_F1 (log), 1.0, 100, 50000) +PL_TEST_INTERVAL (SV_NAME_F1 (log), 100, inf, 50000) |