diff options
Diffstat (limited to 'sysdeps/aarch64/fpu/sv_expf_inline.h')
| -rw-r--r-- | sysdeps/aarch64/fpu/sv_expf_inline.h | 34 |
1 files changed, 17 insertions, 17 deletions
diff --git a/sysdeps/aarch64/fpu/sv_expf_inline.h b/sysdeps/aarch64/fpu/sv_expf_inline.h index 23963b5f8e..6166df6553 100644 --- a/sysdeps/aarch64/fpu/sv_expf_inline.h +++ b/sysdeps/aarch64/fpu/sv_expf_inline.h @@ -24,19 +24,20 @@ struct sv_expf_data { - float poly[5]; - float inv_ln2, ln2_hi, ln2_lo, shift; + float c1, c3, inv_ln2; + float ln2_lo, c0, c2, c4; + float ln2_hi, shift; }; /* Coefficients copied from the polynomial in AdvSIMD variant, reversed for compatibility with polynomial helpers. Shift is 1.5*2^17 + 127. */ #define SV_EXPF_DATA \ { \ - .poly = { 0x1.ffffecp-1f, 0x1.fffdb6p-2f, 0x1.555e66p-3f, 0x1.573e2ep-5f, \ - 0x1.0e4020p-7f }, \ - \ - .inv_ln2 = 0x1.715476p+0f, .ln2_hi = 0x1.62e4p-1f, \ - .ln2_lo = 0x1.7f7d1cp-20f, .shift = 0x1.803f8p17f, \ + /* Coefficients copied from the polynomial in AdvSIMD variant. */ \ + .c0 = 0x1.ffffecp-1f, .c1 = 0x1.fffdb6p-2f, .c2 = 0x1.555e66p-3f, \ + .c3 = 0x1.573e2ep-5f, .c4 = 0x1.0e4020p-7f, .inv_ln2 = 0x1.715476p+0f, \ + .ln2_hi = 0x1.62e4p-1f, .ln2_lo = 0x1.7f7d1cp-20f, \ + .shift = 0x1.803f8p17f, \ } #define C(i) sv_f32 (d->poly[i]) @@ -47,26 +48,25 @@ expf_inline (svfloat32_t x, const svbool_t pg, const struct sv_expf_data *d) /* exp(x) = 2^n (1 + poly(r)), with 1 + poly(r) in [1/sqrt(2),sqrt(2)] x = ln2*n + r, with r in [-ln2/2, ln2/2]. */ - /* Load some constants in quad-word chunks to minimise memory access. */ - svfloat32_t c4_invln2_and_ln2 = svld1rq (svptrue_b32 (), &d->poly[4]); + svfloat32_t lane_consts = svld1rq (svptrue_b32 (), &d->ln2_lo); /* n = round(x/(ln2/N)). */ - svfloat32_t z = svmla_lane (sv_f32 (d->shift), x, c4_invln2_and_ln2, 1); + svfloat32_t z = svmad_x (pg, sv_f32 (d->inv_ln2), x, d->shift); svfloat32_t n = svsub_x (pg, z, d->shift); /* r = x - n*ln2/N. */ - svfloat32_t r = svmls_lane (x, n, c4_invln2_and_ln2, 2); - r = svmls_lane (r, n, c4_invln2_and_ln2, 3); + svfloat32_t r = svmsb_x (pg, sv_f32 (d->ln2_hi), n, x); + r = svmls_lane (r, n, lane_consts, 0); /* scale = 2^(n/N). */ - svfloat32_t scale = svexpa (svreinterpret_u32_f32 (z)); + svfloat32_t scale = svexpa (svreinterpret_u32 (z)); /* y = exp(r) - 1 ~= r + C0 r^2 + C1 r^3 + C2 r^4 + C3 r^5 + C4 r^6. */ - svfloat32_t p12 = svmla_x (pg, C (1), C (2), r); - svfloat32_t p34 = svmla_lane (C (3), r, c4_invln2_and_ln2, 0); - svfloat32_t r2 = svmul_f32_x (pg, r, r); + svfloat32_t p12 = svmla_lane (sv_f32 (d->c1), r, lane_consts, 2); + svfloat32_t p34 = svmla_lane (sv_f32 (d->c3), r, lane_consts, 3); + svfloat32_t r2 = svmul_x (svptrue_b32 (), r, r); svfloat32_t p14 = svmla_x (pg, p12, p34, r2); - svfloat32_t p0 = svmul_f32_x (pg, r, C (0)); + svfloat32_t p0 = svmul_lane (r, lane_consts, 1); svfloat32_t poly = svmla_x (pg, p0, r2, p14); return svmla_x (pg, scale, scale, poly); |