mirror of
https://sourceware.org/git/glibc.git
synced 2024-12-11 22:00:08 +00:00
90a6ca8b28
Previously many routines used * to load from vector types stored
in the data table. This is emitted as ldr, which byte-swaps the
entire vector register, and causes bugs for big-endian when not
all lanes contain the same value. When a vector is to be used
this way, it has been replaced with an array and the load with an
explicit ld1 intrinsic, which byte-swaps only within lanes.
As well, many routines previously used non-standard GCC syntax
for vector operations such as indexing into vectors types with []
and assembling vectors using {}. This syntax should not be mixed
with ACLE, as the former does not respect endianness whereas the
latter does. Such examples have been replaced with, for instance,
vcombine_* and vgetq_lane* intrinsics. Helpers which only use the
GCC syntax, such as the v_call helpers, do not need changing as
they do not use intrinsics.
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
112 lines
3.8 KiB
C
112 lines
3.8 KiB
C
/* Double-precision vector (AdvSIMD) exp2 function
|
|
|
|
Copyright (C) 2023-2024 Free Software Foundation, Inc.
|
|
This file is part of the GNU C Library.
|
|
|
|
The GNU C Library is free software; you can redistribute it and/or
|
|
modify it under the terms of the GNU Lesser General Public
|
|
License as published by the Free Software Foundation; either
|
|
version 2.1 of the License, or (at your option) any later version.
|
|
|
|
The GNU C Library is distributed in the hope that it will be useful,
|
|
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
Lesser General Public License for more details.
|
|
|
|
You should have received a copy of the GNU Lesser General Public
|
|
License along with the GNU C Library; if not, see
|
|
<https://www.gnu.org/licenses/>. */
|
|
|
|
#include "v_math.h"
|
|
#include "poly_advsimd_f64.h"
|
|
|
|
#define N (1 << V_LOG2_TABLE_BITS)
|
|
|
|
static const struct data
|
|
{
|
|
uint64x2_t min_norm;
|
|
uint32x4_t special_bound;
|
|
float64x2_t poly[5];
|
|
float64x2_t invln2;
|
|
uint64x2_t sign_exp_mask;
|
|
} data = {
|
|
/* Each coefficient was generated to approximate log(r) for |r| < 0x1.fp-9
|
|
and N = 128, then scaled by log2(e) in extended precision and rounded back
|
|
to double precision. */
|
|
.poly = { V2 (-0x1.71547652b83p-1), V2 (0x1.ec709dc340953p-2),
|
|
V2 (-0x1.71547651c8f35p-2), V2 (0x1.2777ebe12dda5p-2),
|
|
V2 (-0x1.ec738d616fe26p-3) },
|
|
.invln2 = V2 (0x1.71547652b82fep0),
|
|
.min_norm = V2 (0x0010000000000000), /* asuint64(0x1p-1022). */
|
|
.special_bound = V4 (0x7fe00000), /* asuint64(inf) - min_norm. */
|
|
.sign_exp_mask = V2 (0xfff0000000000000),
|
|
};
|
|
|
|
#define Off v_u64 (0x3fe6900900000000)
|
|
#define IndexMask (N - 1)
|
|
|
|
struct entry
|
|
{
|
|
float64x2_t invc;
|
|
float64x2_t log2c;
|
|
};
|
|
|
|
static inline struct entry
|
|
lookup (uint64x2_t i)
|
|
{
|
|
struct entry e;
|
|
uint64_t i0
|
|
= (vgetq_lane_u64 (i, 0) >> (52 - V_LOG2_TABLE_BITS)) & IndexMask;
|
|
uint64_t i1
|
|
= (vgetq_lane_u64 (i, 1) >> (52 - V_LOG2_TABLE_BITS)) & IndexMask;
|
|
float64x2_t e0 = vld1q_f64 (&__v_log2_data.table[i0].invc);
|
|
float64x2_t e1 = vld1q_f64 (&__v_log2_data.table[i1].invc);
|
|
e.invc = vuzp1q_f64 (e0, e1);
|
|
e.log2c = vuzp2q_f64 (e0, e1);
|
|
return e;
|
|
}
|
|
|
|
static float64x2_t VPCS_ATTR NOINLINE
|
|
special_case (float64x2_t x, float64x2_t y, float64x2_t w, float64x2_t r2,
|
|
uint32x2_t special)
|
|
{
|
|
return v_call_f64 (log2, x, vfmaq_f64 (w, r2, y), vmovl_u32 (special));
|
|
}
|
|
|
|
/* Double-precision vector log2 routine. Implements the same algorithm as
|
|
vector log10, with coefficients and table entries scaled in extended
|
|
precision. The maximum observed error is 2.58 ULP:
|
|
_ZGVnN2v_log2(0x1.0b556b093869bp+0) got 0x1.fffb34198d9dap-5
|
|
want 0x1.fffb34198d9ddp-5. */
|
|
float64x2_t VPCS_ATTR V_NAME_D1 (log2) (float64x2_t x)
|
|
{
|
|
const struct data *d = ptr_barrier (&data);
|
|
uint64x2_t ix = vreinterpretq_u64_f64 (x);
|
|
uint32x2_t special = vcge_u32 (vsubhn_u64 (ix, d->min_norm),
|
|
vget_low_u32 (d->special_bound));
|
|
|
|
/* 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. */
|
|
uint64x2_t tmp = vsubq_u64 (ix, Off);
|
|
int64x2_t k = vshrq_n_s64 (vreinterpretq_s64_u64 (tmp), 52);
|
|
uint64x2_t iz = vsubq_u64 (ix, vandq_u64 (tmp, d->sign_exp_mask));
|
|
float64x2_t z = vreinterpretq_f64_u64 (iz);
|
|
|
|
struct entry e = lookup (tmp);
|
|
|
|
/* log2(x) = log1p(z/c-1)/log(2) + log2(c) + k. */
|
|
|
|
float64x2_t r = vfmaq_f64 (v_f64 (-1.0), z, e.invc);
|
|
float64x2_t kd = vcvtq_f64_s64 (k);
|
|
float64x2_t w = vfmaq_f64 (e.log2c, r, d->invln2);
|
|
|
|
float64x2_t r2 = vmulq_f64 (r, r);
|
|
float64x2_t y = v_pw_horner_4_f64 (r, r2, d->poly);
|
|
w = vaddq_f64 (kd, w);
|
|
|
|
if (__glibc_unlikely (v_any_u32h (special)))
|
|
return special_case (x, y, w, r2, special);
|
|
return vfmaq_f64 (w, r2, y);
|
|
}
|