about summary refs log tree commit diff
path: root/sysdeps
diff options
context:
space:
mode:
authorJoe Ramsay <Joe.Ramsay@arm.com>2023-10-05 17:10:52 +0100
committerSzabolcs Nagy <szabolcs.nagy@arm.com>2023-10-23 15:00:45 +0100
commit31aaf6fed986fade042f9ffe7535d8b3f2c173a2 (patch)
tree5417d6dffd8eefa74a1506a792ed40330205ce31 /sysdeps
parent067a34156c19fb3c53824e37d70820c0ce5b87b2 (diff)
downloadglibc-31aaf6fed986fade042f9ffe7535d8b3f2c173a2.tar.gz
glibc-31aaf6fed986fade042f9ffe7535d8b3f2c173a2.tar.xz
glibc-31aaf6fed986fade042f9ffe7535d8b3f2c173a2.zip
aarch64: Add vector implementations of exp10 routines
Double-precision routines either reuse the exp table (AdvSIMD) or use
SVE FEXPA intruction.
Diffstat (limited to 'sysdeps')
-rw-r--r--sysdeps/aarch64/fpu/Makefile1
-rw-r--r--sysdeps/aarch64/fpu/Versions4
-rw-r--r--sysdeps/aarch64/fpu/bits/math-vector.h4
-rw-r--r--sysdeps/aarch64/fpu/exp10_advsimd.c145
-rw-r--r--sysdeps/aarch64/fpu/exp10_sve.c127
-rw-r--r--sysdeps/aarch64/fpu/exp10f_advsimd.c140
-rw-r--r--sysdeps/aarch64/fpu/exp10f_sve.c91
-rw-r--r--sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c1
-rw-r--r--sysdeps/aarch64/fpu/test-double-sve-wrappers.c1
-rw-r--r--sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c1
-rw-r--r--sysdeps/aarch64/fpu/test-float-sve-wrappers.c1
-rw-r--r--sysdeps/aarch64/libm-test-ulps8
-rw-r--r--sysdeps/unix/sysv/linux/aarch64/libmvec.abilist4
13 files changed, 528 insertions, 0 deletions
diff --git a/sysdeps/aarch64/fpu/Makefile b/sysdeps/aarch64/fpu/Makefile
index 0a047a150d..1f1ac2a2b8 100644
--- a/sysdeps/aarch64/fpu/Makefile
+++ b/sysdeps/aarch64/fpu/Makefile
@@ -1,5 +1,6 @@
 libmvec-supported-funcs = cos \
                           exp \
+                          exp10 \
                           exp2 \
                           log \
                           log10 \
diff --git a/sysdeps/aarch64/fpu/Versions b/sysdeps/aarch64/fpu/Versions
index 358efed5ee..eb5ad50017 100644
--- a/sysdeps/aarch64/fpu/Versions
+++ b/sysdeps/aarch64/fpu/Versions
@@ -18,6 +18,10 @@ libmvec {
     _ZGVsMxv_sinf;
   }
   GLIBC_2.39 {
+    _ZGVnN4v_exp10f;
+    _ZGVnN2v_exp10;
+    _ZGVsMxv_exp10f;
+    _ZGVsMxv_exp10;
     _ZGVnN4v_exp2f;
     _ZGVnN2v_exp2;
     _ZGVsMxv_exp2f;
diff --git a/sysdeps/aarch64/fpu/bits/math-vector.h b/sysdeps/aarch64/fpu/bits/math-vector.h
index 59f2efa6d7..06587ffa91 100644
--- a/sysdeps/aarch64/fpu/bits/math-vector.h
+++ b/sysdeps/aarch64/fpu/bits/math-vector.h
@@ -51,6 +51,7 @@ typedef __SVBool_t __sv_bool_t;
 
 __vpcs __f32x4_t _ZGVnN4v_cosf (__f32x4_t);
 __vpcs __f32x4_t _ZGVnN4v_expf (__f32x4_t);
+__vpcs __f32x4_t _ZGVnN4v_exp10f (__f32x4_t);
 __vpcs __f32x4_t _ZGVnN4v_exp2f (__f32x4_t);
 __vpcs __f32x4_t _ZGVnN4v_logf (__f32x4_t);
 __vpcs __f32x4_t _ZGVnN4v_log10f (__f32x4_t);
@@ -60,6 +61,7 @@ __vpcs __f32x4_t _ZGVnN4v_tanf (__f32x4_t);
 
 __vpcs __f64x2_t _ZGVnN2v_cos (__f64x2_t);
 __vpcs __f64x2_t _ZGVnN2v_exp (__f64x2_t);
+__vpcs __f64x2_t _ZGVnN2v_exp10 (__f64x2_t);
 __vpcs __f64x2_t _ZGVnN2v_exp2 (__f64x2_t);
 __vpcs __f64x2_t _ZGVnN2v_log (__f64x2_t);
 __vpcs __f64x2_t _ZGVnN2v_log10 (__f64x2_t);
@@ -74,6 +76,7 @@ __vpcs __f64x2_t _ZGVnN2v_tan (__f64x2_t);
 
 __sv_f32_t _ZGVsMxv_cosf (__sv_f32_t, __sv_bool_t);
 __sv_f32_t _ZGVsMxv_expf (__sv_f32_t, __sv_bool_t);
+__sv_f32_t _ZGVsMxv_exp10f (__sv_f32_t, __sv_bool_t);
 __sv_f32_t _ZGVsMxv_exp2f (__sv_f32_t, __sv_bool_t);
 __sv_f32_t _ZGVsMxv_logf (__sv_f32_t, __sv_bool_t);
 __sv_f32_t _ZGVsMxv_log10f (__sv_f32_t, __sv_bool_t);
@@ -83,6 +86,7 @@ __sv_f32_t _ZGVsMxv_tanf (__sv_f32_t, __sv_bool_t);
 
 __sv_f64_t _ZGVsMxv_cos (__sv_f64_t, __sv_bool_t);
 __sv_f64_t _ZGVsMxv_exp (__sv_f64_t, __sv_bool_t);
+__sv_f64_t _ZGVsMxv_exp10 (__sv_f64_t, __sv_bool_t);
 __sv_f64_t _ZGVsMxv_exp2 (__sv_f64_t, __sv_bool_t);
 __sv_f64_t _ZGVsMxv_log (__sv_f64_t, __sv_bool_t);
 __sv_f64_t _ZGVsMxv_log10 (__sv_f64_t, __sv_bool_t);
diff --git a/sysdeps/aarch64/fpu/exp10_advsimd.c b/sysdeps/aarch64/fpu/exp10_advsimd.c
new file mode 100644
index 0000000000..c584e2a270
--- /dev/null
+++ b/sysdeps/aarch64/fpu/exp10_advsimd.c
@@ -0,0 +1,145 @@
+/* Double-precision vector (AdvSIMD) exp10 function.
+
+   Copyright (C) 2023 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"
+
+/* Value of |x| above which scale overflows without special treatment.  */
+#define SpecialBound 306.0 /* floor (log10 (2^1023)) - 1.  */
+/* Value of n above which scale overflows even with special treatment.  */
+#define ScaleBound 163840.0 /* 1280.0 * N.  */
+
+const static struct data
+{
+  float64x2_t poly[4];
+  float64x2_t log10_2, log2_10_hi, log2_10_lo, shift;
+#if !WANT_SIMD_EXCEPT
+  float64x2_t special_bound, scale_thresh;
+#endif
+} data = {
+  /* Coefficients generated using Remez algorithm.
+     rel error: 0x1.5ddf8f28p-54
+     abs error: 0x1.5ed266c8p-54 in [ -log10(2)/256, log10(2)/256 ]
+     maxerr: 1.14432 +0.5 ulp.  */
+  .poly = { V2 (0x1.26bb1bbb5524p1), V2 (0x1.53524c73cecdap1),
+	    V2 (0x1.047060efb781cp1), V2 (0x1.2bd76040f0d16p0) },
+  .log10_2 = V2 (0x1.a934f0979a371p8),	   /* N/log2(10).  */
+  .log2_10_hi = V2 (0x1.34413509f79ffp-9), /* log2(10)/N.  */
+  .log2_10_lo = V2 (-0x1.9dc1da994fd21p-66),
+  .shift = V2 (0x1.8p+52),
+#if !WANT_SIMD_EXCEPT
+  .scale_thresh = V2 (ScaleBound),
+  .special_bound = V2 (SpecialBound),
+#endif
+};
+
+#define N (1 << V_EXP_TABLE_BITS)
+#define IndexMask v_u64 (N - 1)
+
+#if WANT_SIMD_EXCEPT
+
+# define TinyBound v_u64 (0x2000000000000000) /* asuint64 (0x1p-511).  */
+# define BigBound v_u64 (0x4070000000000000)  /* asuint64 (0x1p8).  */
+# define Thres v_u64 (0x2070000000000000)     /* BigBound - TinyBound.  */
+
+static inline float64x2_t VPCS_ATTR
+special_case (float64x2_t x, float64x2_t y, uint64x2_t cmp)
+{
+  /* If fenv exceptions are to be triggered correctly, fall back to the scalar
+     routine for special lanes.  */
+  return v_call_f64 (exp10, x, y, cmp);
+}
+
+#else
+
+# define SpecialOffset v_u64 (0x6000000000000000) /* 0x1p513.  */
+/* SpecialBias1 + SpecialBias1 = asuint(1.0).  */
+# define SpecialBias1 v_u64 (0x7000000000000000)  /* 0x1p769.  */
+# define SpecialBias2 v_u64 (0x3010000000000000)  /* 0x1p-254.  */
+
+static float64x2_t VPCS_ATTR NOINLINE
+special_case (float64x2_t s, float64x2_t y, float64x2_t n,
+	      const struct data *d)
+{
+  /* 2^(n/N) may overflow, break it up into s1*s2.  */
+  uint64x2_t b = vandq_u64 (vcltzq_f64 (n), SpecialOffset);
+  float64x2_t s1 = vreinterpretq_f64_u64 (vsubq_u64 (SpecialBias1, b));
+  float64x2_t s2 = vreinterpretq_f64_u64 (
+      vaddq_u64 (vsubq_u64 (vreinterpretq_u64_f64 (s), SpecialBias2), b));
+  uint64x2_t cmp = vcagtq_f64 (n, d->scale_thresh);
+  float64x2_t r1 = vmulq_f64 (s1, s1);
+  float64x2_t r0 = vmulq_f64 (vfmaq_f64 (s2, y, s2), s1);
+  return vbslq_f64 (cmp, r1, r0);
+}
+
+#endif
+
+/* Fast vector implementation of exp10.
+   Maximum measured error is 1.64 ulp.
+   _ZGVnN2v_exp10(0x1.ccd1c9d82cc8cp+0) got 0x1.f8dab6d7fed0cp+5
+				       want 0x1.f8dab6d7fed0ap+5.  */
+float64x2_t VPCS_ATTR V_NAME_D1 (exp10) (float64x2_t x)
+{
+  const struct data *d = ptr_barrier (&data);
+  uint64x2_t cmp;
+#if WANT_SIMD_EXCEPT
+  /* If any lanes are special, mask them with 1 and retain a copy of x to allow
+     special_case to fix special lanes later. This is only necessary if fenv
+     exceptions are to be triggered correctly.  */
+  float64x2_t xm = x;
+  uint64x2_t iax = vreinterpretq_u64_f64 (vabsq_f64 (x));
+  cmp = vcgeq_u64 (vsubq_u64 (iax, TinyBound), Thres);
+  if (__glibc_unlikely (v_any_u64 (cmp)))
+    x = vbslq_f64 (cmp, v_f64 (1), x);
+#else
+  cmp = vcageq_f64 (x, d->special_bound);
+#endif
+
+  /* n = round(x/(log10(2)/N)).  */
+  float64x2_t z = vfmaq_f64 (d->shift, x, d->log10_2);
+  uint64x2_t u = vreinterpretq_u64_f64 (z);
+  float64x2_t n = vsubq_f64 (z, d->shift);
+
+  /* r = x - n*log10(2)/N.  */
+  float64x2_t r = x;
+  r = vfmsq_f64 (r, d->log2_10_hi, n);
+  r = vfmsq_f64 (r, d->log2_10_lo, n);
+
+  uint64x2_t e = vshlq_n_u64 (u, 52 - V_EXP_TABLE_BITS);
+  uint64x2_t i = vandq_u64 (u, IndexMask);
+
+  /* y = exp10(r) - 1 ~= C0 r + C1 r^2 + C2 r^3 + C3 r^4.  */
+  float64x2_t r2 = vmulq_f64 (r, r);
+  float64x2_t p = vfmaq_f64 (d->poly[0], r, d->poly[1]);
+  float64x2_t y = vfmaq_f64 (d->poly[2], r, d->poly[3]);
+  p = vfmaq_f64 (p, y, r2);
+  y = vmulq_f64 (r, p);
+
+  /* s = 2^(n/N).  */
+  u = v_lookup_u64 (__v_exp_data, i);
+  float64x2_t s = vreinterpretq_f64_u64 (vaddq_u64 (u, e));
+
+  if (__glibc_unlikely (v_any_u64 (cmp)))
+#if WANT_SIMD_EXCEPT
+    return special_case (xm, vfmaq_f64 (s, y, s), cmp);
+#else
+    return special_case (s, y, n, d);
+#endif
+
+  return vfmaq_f64 (s, y, s);
+}
diff --git a/sysdeps/aarch64/fpu/exp10_sve.c b/sysdeps/aarch64/fpu/exp10_sve.c
new file mode 100644
index 0000000000..a8cef7b692
--- /dev/null
+++ b/sysdeps/aarch64/fpu/exp10_sve.c
@@ -0,0 +1,127 @@
+/* Double-precision vector (SVE) exp10 function.
+
+   Copyright (C) 2023 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 "sv_math.h"
+#include "poly_sve_f64.h"
+
+#define SpecialBound 307.0 /* floor (log10 (2^1023)).  */
+
+static const struct data
+{
+  double poly[5];
+  double shift, log10_2, log2_10_hi, log2_10_lo, scale_thres, special_bound;
+} data = {
+  /* Coefficients generated using Remez algorithm.
+     rel error: 0x1.9fcb9b3p-60
+     abs error: 0x1.a20d9598p-60 in [ -log10(2)/128, log10(2)/128 ]
+     max ulp err 0.52 +0.5.  */
+  .poly = { 0x1.26bb1bbb55516p1, 0x1.53524c73cd32ap1, 0x1.0470591daeafbp1,
+	    0x1.2bd77b1361ef6p0, 0x1.142b5d54e9621p-1 },
+  /* 1.5*2^46+1023. This value is further explained below.  */
+  .shift = 0x1.800000000ffc0p+46,
+  .log10_2 = 0x1.a934f0979a371p1,     /* 1/log2(10).  */
+  .log2_10_hi = 0x1.34413509f79ffp-2, /* log2(10).  */
+  .log2_10_lo = -0x1.9dc1da994fd21p-59,
+  .scale_thres = 1280.0,
+  .special_bound = SpecialBound,
+};
+
+#define SpecialOffset 0x6000000000000000 /* 0x1p513.  */
+/* SpecialBias1 + SpecialBias1 = asuint(1.0).  */
+#define SpecialBias1 0x7000000000000000 /* 0x1p769.  */
+#define SpecialBias2 0x3010000000000000 /* 0x1p-254.  */
+
+/* Update of both special and non-special cases, if any special case is
+   detected.  */
+static inline svfloat64_t
+special_case (svbool_t pg, svfloat64_t s, svfloat64_t y, svfloat64_t n,
+	      const struct data *d)
+{
+  /* s=2^n may overflow, break it up into s=s1*s2,
+     such that exp = s + s*y can be computed as s1*(s2+s2*y)
+     and s1*s1 overflows only if n>0.  */
+
+  /* If n<=0 then set b to 0x6, 0 otherwise.  */
+  svbool_t p_sign = svcmple (pg, n, 0.0); /* n <= 0.  */
+  svuint64_t b = svdup_u64_z (p_sign, SpecialOffset);
+
+  /* Set s1 to generate overflow depending on sign of exponent n.  */
+  svfloat64_t s1 = svreinterpret_f64 (svsubr_x (pg, b, SpecialBias1));
+  /* Offset s to avoid overflow in final result if n is below threshold.  */
+  svfloat64_t s2 = svreinterpret_f64 (
+      svadd_x (pg, svsub_x (pg, svreinterpret_u64 (s), SpecialBias2), b));
+
+  /* |n| > 1280 => 2^(n) overflows.  */
+  svbool_t p_cmp = svacgt (pg, n, d->scale_thres);
+
+  svfloat64_t r1 = svmul_x (pg, s1, s1);
+  svfloat64_t r2 = svmla_x (pg, s2, s2, y);
+  svfloat64_t r0 = svmul_x (pg, r2, s1);
+
+  return svsel (p_cmp, r1, r0);
+}
+
+/* Fast vector implementation of exp10 using FEXPA instruction.
+   Maximum measured error is 1.02 ulp.
+   SV_NAME_D1 (exp10)(-0x1.2862fec805e58p+2) got 0x1.885a89551d782p-16
+					    want 0x1.885a89551d781p-16.  */
+svfloat64_t SV_NAME_D1 (exp10) (svfloat64_t x, svbool_t pg)
+{
+  const struct data *d = ptr_barrier (&data);
+  svbool_t no_big_scale = svacle (pg, x, d->special_bound);
+  svbool_t special = svnot_z (pg, no_big_scale);
+
+  /* n = round(x/(log10(2)/N)).  */
+  svfloat64_t shift = sv_f64 (d->shift);
+  svfloat64_t z = svmla_x (pg, shift, x, d->log10_2);
+  svfloat64_t n = svsub_x (pg, z, shift);
+
+  /* r = x - n*log10(2)/N.  */
+  svfloat64_t log2_10 = svld1rq (svptrue_b64 (), &d->log2_10_hi);
+  svfloat64_t r = x;
+  r = svmls_lane (r, n, log2_10, 0);
+  r = svmls_lane (r, n, log2_10, 1);
+
+  /* scale = 2^(n/N), computed using FEXPA. FEXPA does not propagate NaNs, so
+     for consistent NaN handling we have to manually propagate them. This
+     comes at significant performance cost.  */
+  svuint64_t u = svreinterpret_u64 (z);
+  svfloat64_t scale = svexpa (u);
+
+  /* Approximate exp10(r) using polynomial.  */
+  svfloat64_t r2 = svmul_x (pg, r, r);
+  svfloat64_t y = svmla_x (pg, svmul_x (pg, r, d->poly[0]), r2,
+			   sv_pairwise_poly_3_f64_x (pg, r, r2, d->poly + 1));
+
+  /* Assemble result as exp10(x) = 2^n * exp10(r).  If |x| > SpecialBound
+     multiplication may overflow, so use special case routine.  */
+  if (__glibc_unlikely (svptest_any (pg, special)))
+    {
+      /* FEXPA zeroes the sign bit, however the sign is meaningful to the
+	 special case function so needs to be copied.
+	 e = sign bit of u << 46.  */
+      svuint64_t e = svand_x (pg, svlsl_x (pg, u, 46), 0x8000000000000000);
+      /* Copy sign to scale.  */
+      scale = svreinterpret_f64 (svadd_x (pg, e, svreinterpret_u64 (scale)));
+      return special_case (pg, scale, y, n, d);
+    }
+
+  /* No special case.  */
+  return svmla_x (pg, scale, scale, y);
+}
diff --git a/sysdeps/aarch64/fpu/exp10f_advsimd.c b/sysdeps/aarch64/fpu/exp10f_advsimd.c
new file mode 100644
index 0000000000..9e754c46fa
--- /dev/null
+++ b/sysdeps/aarch64/fpu/exp10f_advsimd.c
@@ -0,0 +1,140 @@
+/* Single-precision vector (AdvSIMD) exp10 function.
+
+   Copyright (C) 2023 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_f32.h"
+
+#define ScaleBound 192.0f
+
+static const struct data
+{
+  float32x4_t poly[5];
+  float32x4_t shift, log10_2, log2_10_hi, log2_10_lo;
+#if !WANT_SIMD_EXCEPT
+  float32x4_t scale_thresh;
+#endif
+} data = {
+  /* Coefficients generated using Remez algorithm with minimisation of relative
+     error.
+     rel error: 0x1.89dafa3p-24
+     abs error: 0x1.167d55p-23 in [-log10(2)/2, log10(2)/2]
+     maxerr: 1.85943 +0.5 ulp.  */
+  .poly = { V4 (0x1.26bb16p+1f), V4 (0x1.5350d2p+1f), V4 (0x1.04744ap+1f),
+	    V4 (0x1.2d8176p+0f), V4 (0x1.12b41ap-1f) },
+  .shift = V4 (0x1.8p23f),
+  .log10_2 = V4 (0x1.a934fp+1),
+  .log2_10_hi = V4 (0x1.344136p-2),
+  .log2_10_lo = V4 (-0x1.ec10cp-27),
+#if !WANT_SIMD_EXCEPT
+  .scale_thresh = V4 (ScaleBound)
+#endif
+};
+
+#define ExponentBias v_u32 (0x3f800000)
+
+#if WANT_SIMD_EXCEPT
+
+# define SpecialBound 38.0f	       /* rint(log10(2^127)).  */
+# define TinyBound v_u32 (0x20000000) /* asuint (0x1p-63).  */
+# define BigBound v_u32 (0x42180000)  /* asuint (SpecialBound).  */
+# define Thres v_u32 (0x22180000)     /* BigBound - TinyBound.  */
+
+static float32x4_t VPCS_ATTR NOINLINE
+special_case (float32x4_t x, float32x4_t y, uint32x4_t cmp)
+{
+  /* If fenv exceptions are to be triggered correctly, fall back to the scalar
+     routine to special lanes.  */
+  return v_call_f32 (exp10f, x, y, cmp);
+}
+
+#else
+
+# define SpecialBound 126.0f /* rint (log2 (2^127 / (1 + sqrt (2)))).  */
+# define SpecialOffset v_u32 (0x82000000)
+# define SpecialBias v_u32 (0x7f000000)
+
+static float32x4_t VPCS_ATTR NOINLINE
+special_case (float32x4_t poly, float32x4_t n, uint32x4_t e, uint32x4_t cmp1,
+	      float32x4_t scale, const struct data *d)
+{
+  /* 2^n may overflow, break it up into s1*s2.  */
+  uint32x4_t b = vandq_u32 (vclezq_f32 (n), SpecialOffset);
+  float32x4_t s1 = vreinterpretq_f32_u32 (vaddq_u32 (b, SpecialBias));
+  float32x4_t s2 = vreinterpretq_f32_u32 (vsubq_u32 (e, b));
+  uint32x4_t cmp2 = vcagtq_f32 (n, d->scale_thresh);
+  float32x4_t r2 = vmulq_f32 (s1, s1);
+  float32x4_t r1 = vmulq_f32 (vfmaq_f32 (s2, poly, s2), s1);
+  /* Similar to r1 but avoids double rounding in the subnormal range.  */
+  float32x4_t r0 = vfmaq_f32 (scale, poly, scale);
+  float32x4_t r = vbslq_f32 (cmp1, r1, r0);
+  return vbslq_f32 (cmp2, r2, r);
+}
+
+#endif
+
+/* Fast vector implementation of single-precision exp10.
+   Algorithm is accurate to 2.36 ULP.
+   _ZGVnN4v_exp10f(0x1.be2b36p+1) got 0x1.7e79c4p+11
+				 want 0x1.7e79cp+11.  */
+float32x4_t VPCS_ATTR V_NAME_F1 (exp10) (float32x4_t x)
+{
+  const struct data *d = ptr_barrier (&data);
+#if WANT_SIMD_EXCEPT
+  /* asuint(x) - TinyBound >= BigBound - TinyBound.  */
+  uint32x4_t cmp = vcgeq_u32 (
+      vsubq_u32 (vandq_u32 (vreinterpretq_u32_f32 (x), v_u32 (0x7fffffff)),
+		 TinyBound),
+      Thres);
+  float32x4_t xm = x;
+  /* If any lanes are special, mask them with 1 and retain a copy of x to allow
+     special case handler to fix special lanes later. This is only necessary if
+     fenv exceptions are to be triggered correctly.  */
+  if (__glibc_unlikely (v_any_u32 (cmp)))
+    x = vbslq_f32 (cmp, v_f32 (1), x);
+#endif
+
+  /* exp10(x) = 2^n * 10^r = 2^n * (1 + poly (r)),
+     with poly(r) in [1/sqrt(2), sqrt(2)] and
+     x = r + n * log10 (2), with r in [-log10(2)/2, log10(2)/2].  */
+  float32x4_t z = vfmaq_f32 (d->shift, x, d->log10_2);
+  float32x4_t n = vsubq_f32 (z, d->shift);
+  float32x4_t r = vfmsq_f32 (x, n, d->log2_10_hi);
+  r = vfmsq_f32 (r, n, d->log2_10_lo);
+  uint32x4_t e = vshlq_n_u32 (vreinterpretq_u32_f32 (z), 23);
+
+  float32x4_t scale = vreinterpretq_f32_u32 (vaddq_u32 (e, ExponentBias));
+
+#if !WANT_SIMD_EXCEPT
+  uint32x4_t cmp = vcagtq_f32 (n, v_f32 (SpecialBound));
+#endif
+
+  float32x4_t r2 = vmulq_f32 (r, r);
+  float32x4_t poly
+      = vfmaq_f32 (vmulq_f32 (r, d->poly[0]),
+		   v_pairwise_poly_3_f32 (r, r2, d->poly + 1), r2);
+
+  if (__glibc_unlikely (v_any_u32 (cmp)))
+#if WANT_SIMD_EXCEPT
+    return special_case (xm, vfmaq_f32 (scale, poly, scale), cmp);
+#else
+    return special_case (poly, n, e, cmp, scale, d);
+#endif
+
+  return vfmaq_f32 (scale, poly, scale);
+}
diff --git a/sysdeps/aarch64/fpu/exp10f_sve.c b/sysdeps/aarch64/fpu/exp10f_sve.c
new file mode 100644
index 0000000000..3df3246e05
--- /dev/null
+++ b/sysdeps/aarch64/fpu/exp10f_sve.c
@@ -0,0 +1,91 @@
+/* Single-precision vector (SVE) exp10 function.
+
+   Copyright (C) 2023 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 "sv_math.h"
+#include "poly_sve_f32.h"
+
+/* For x < -SpecialBound, the result is subnormal and not handled correctly by
+   FEXPA.  */
+#define SpecialBound 37.9
+
+static const struct data
+{
+  float poly[5];
+  float shift, log10_2, log2_10_hi, log2_10_lo, special_bound;
+} data = {
+  /* Coefficients generated using Remez algorithm with minimisation of relative
+     error.
+     rel error: 0x1.89dafa3p-24
+     abs error: 0x1.167d55p-23 in [-log10(2)/2, log10(2)/2]
+     maxerr: 0.52 +0.5 ulp.  */
+  .poly = { 0x1.26bb16p+1f, 0x1.5350d2p+1f, 0x1.04744ap+1f, 0x1.2d8176p+0f,
+	    0x1.12b41ap-1f },
+  /* 1.5*2^17 + 127, a shift value suitable for FEXPA.  */
+  .shift = 0x1.903f8p17f,
+  .log10_2 = 0x1.a934fp+1,
+  .log2_10_hi = 0x1.344136p-2,
+  .log2_10_lo = -0x1.ec10cp-27,
+  .special_bound = SpecialBound,
+};
+
+static svfloat32_t NOINLINE
+special_case (svfloat32_t x, svfloat32_t y, svbool_t special)
+{
+  return sv_call_f32 (exp10f, x, y, special);
+}
+
+/* Single-precision SVE exp10f routine. Implements the same algorithm
+   as AdvSIMD exp10f.
+   Worst case error is 1.02 ULPs.
+   _ZGVsMxv_exp10f(-0x1.040488p-4) got 0x1.ba5f9ep-1
+				  want 0x1.ba5f9cp-1.  */
+svfloat32_t SV_NAME_F1 (exp10) (svfloat32_t x, const svbool_t pg)
+{
+  const struct data *d = ptr_barrier (&data);
+  /* exp10(x) = 2^(n/N) * 10^r = 2^n * (1 + poly (r)),
+     with poly(r) in [1/sqrt(2), sqrt(2)] and
+     x = r + n * log10(2) / N, with r in [-log10(2)/2N, log10(2)/2N].  */
+
+  /* Load some constants in quad-word chunks to minimise memory access (last
+     lane is wasted).  */
+  svfloat32_t log10_2_and_inv = svld1rq (svptrue_b32 (), &d->log10_2);
+
+  /* n = round(x/(log10(2)/N)).  */
+  svfloat32_t shift = sv_f32 (d->shift);
+  svfloat32_t z = svmla_lane (shift, x, log10_2_and_inv, 0);
+  svfloat32_t n = svsub_x (pg, z, shift);
+
+  /* r = x - n*log10(2)/N.  */
+  svfloat32_t r = svmls_lane (x, n, log10_2_and_inv, 1);
+  r = svmls_lane (r, n, log10_2_and_inv, 2);
+
+  svbool_t special = svacgt (pg, x, d->special_bound);
+  svfloat32_t scale = svexpa (svreinterpret_u32 (z));
+
+  /* Polynomial evaluation: poly(r) ~ exp10(r)-1.  */
+  svfloat32_t r2 = svmul_x (pg, r, r);
+  svfloat32_t poly
+      = svmla_x (pg, svmul_x (pg, r, d->poly[0]),
+		 sv_pairwise_poly_3_f32_x (pg, r, r2, d->poly + 1), r2);
+
+  if (__glibc_unlikely (svptest_any (pg, special)))
+    return special_case (x, svmla_x (pg, scale, scale, poly), special);
+
+  return svmla_x (pg, scale, scale, poly);
+}
diff --git a/sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c b/sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c
index 8d05498ec9..26d5ecf66f 100644
--- a/sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c
@@ -25,6 +25,7 @@
 
 VPCS_VECTOR_WRAPPER (cos_advsimd, _ZGVnN2v_cos)
 VPCS_VECTOR_WRAPPER (exp_advsimd, _ZGVnN2v_exp)
+VPCS_VECTOR_WRAPPER (exp10_advsimd, _ZGVnN2v_exp10)
 VPCS_VECTOR_WRAPPER (exp2_advsimd, _ZGVnN2v_exp2)
 VPCS_VECTOR_WRAPPER (log_advsimd, _ZGVnN2v_log)
 VPCS_VECTOR_WRAPPER (log10_advsimd, _ZGVnN2v_log10)
diff --git a/sysdeps/aarch64/fpu/test-double-sve-wrappers.c b/sysdeps/aarch64/fpu/test-double-sve-wrappers.c
index b65bc6f1e6..86efd60779 100644
--- a/sysdeps/aarch64/fpu/test-double-sve-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-double-sve-wrappers.c
@@ -34,6 +34,7 @@
 
 SVE_VECTOR_WRAPPER (cos_sve, _ZGVsMxv_cos)
 SVE_VECTOR_WRAPPER (exp_sve, _ZGVsMxv_exp)
+SVE_VECTOR_WRAPPER (exp10_sve, _ZGVsMxv_exp10)
 SVE_VECTOR_WRAPPER (exp2_sve, _ZGVsMxv_exp2)
 SVE_VECTOR_WRAPPER (log_sve, _ZGVsMxv_log)
 SVE_VECTOR_WRAPPER (log10_sve, _ZGVsMxv_log10)
diff --git a/sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c b/sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c
index 6ced0d4488..8f7ebea1ac 100644
--- a/sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c
@@ -25,6 +25,7 @@
 
 VPCS_VECTOR_WRAPPER (cosf_advsimd, _ZGVnN4v_cosf)
 VPCS_VECTOR_WRAPPER (expf_advsimd, _ZGVnN4v_expf)
+VPCS_VECTOR_WRAPPER (exp10f_advsimd, _ZGVnN4v_exp10f)
 VPCS_VECTOR_WRAPPER (exp2f_advsimd, _ZGVnN4v_exp2f)
 VPCS_VECTOR_WRAPPER (logf_advsimd, _ZGVnN4v_logf)
 VPCS_VECTOR_WRAPPER (log10f_advsimd, _ZGVnN4v_log10f)
diff --git a/sysdeps/aarch64/fpu/test-float-sve-wrappers.c b/sysdeps/aarch64/fpu/test-float-sve-wrappers.c
index 2ed8d0659a..885e58ac39 100644
--- a/sysdeps/aarch64/fpu/test-float-sve-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-float-sve-wrappers.c
@@ -34,6 +34,7 @@
 
 SVE_VECTOR_WRAPPER (cosf_sve, _ZGVsMxv_cosf)
 SVE_VECTOR_WRAPPER (expf_sve, _ZGVsMxv_expf)
+SVE_VECTOR_WRAPPER (exp10f_sve, _ZGVsMxv_exp10f)
 SVE_VECTOR_WRAPPER (exp2f_sve, _ZGVsMxv_exp2f)
 SVE_VECTOR_WRAPPER (logf_sve, _ZGVsMxv_logf)
 SVE_VECTOR_WRAPPER (log10f_sve, _ZGVsMxv_log10f)
diff --git a/sysdeps/aarch64/libm-test-ulps b/sysdeps/aarch64/libm-test-ulps
index 6641c7fa0b..d117209c06 100644
--- a/sysdeps/aarch64/libm-test-ulps
+++ b/sysdeps/aarch64/libm-test-ulps
@@ -970,11 +970,19 @@ double: 2
 float: 1
 ldouble: 2
 
+Function: "exp10_advsimd":
+double: 1
+float: 2
+
 Function: "exp10_downward":
 double: 2
 float: 1
 ldouble: 3
 
+Function: "exp10_sve":
+double: 1
+float: 1
+
 Function: "exp10_towardzero":
 double: 2
 float: 1
diff --git a/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist b/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist
index f8776a6bea..cad774521a 100644
--- a/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist
+++ b/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist
@@ -14,14 +14,18 @@ GLIBC_2.38 _ZGVsMxv_log F
 GLIBC_2.38 _ZGVsMxv_logf F
 GLIBC_2.38 _ZGVsMxv_sin F
 GLIBC_2.38 _ZGVsMxv_sinf F
+GLIBC_2.39 _ZGVnN2v_exp10 F
 GLIBC_2.39 _ZGVnN2v_exp2 F
 GLIBC_2.39 _ZGVnN2v_log10 F
 GLIBC_2.39 _ZGVnN2v_log2 F
 GLIBC_2.39 _ZGVnN2v_tan F
+GLIBC_2.39 _ZGVnN4v_exp10f F
 GLIBC_2.39 _ZGVnN4v_exp2f F
 GLIBC_2.39 _ZGVnN4v_log10f F
 GLIBC_2.39 _ZGVnN4v_log2f F
 GLIBC_2.39 _ZGVnN4v_tanf F
+GLIBC_2.39 _ZGVsMxv_exp10 F
+GLIBC_2.39 _ZGVsMxv_exp10f F
 GLIBC_2.39 _ZGVsMxv_exp2 F
 GLIBC_2.39 _ZGVsMxv_exp2f F
 GLIBC_2.39 _ZGVsMxv_log10 F