pl/math: Add vector/SVE logf

New routine is SVE port of the Neon algorithm in math/, and is
accurate to 3.4 ULP.

8 files changed, 101 insertions, 0 deletions

diff --git a/pl/math/include/mathlib.h b/pl/math/include/mathlib.h
index 0d6077d..1765c9d 100644
--- a/pl/math/include/mathlib.h
+++ b/pl/math/include/mathlib.h
@@ -110,6 +110,7 @@ svfloat64_t __sv_atan_x (svfloat64_t, svbool_t);
 svfloat64_t __sv_atan2_x (svfloat64_t, svfloat64_t, svbool_t);
 svfloat32_t __sv_cosf_x (svfloat32_t, svbool_t);
 svfloat64_t __sv_cos_x (svfloat64_t, svbool_t);
+svfloat32_t __sv_logf_x (svfloat32_t, svbool_t);
 svfloat32_t __sv_log10f_x (svfloat32_t, svbool_t);
 svfloat64_t __sv_log10_x (svfloat64_t, svbool_t);
 svfloat32_t __sv_sinf_x (svfloat32_t, svbool_t);
@@ -121,6 +122,7 @@ svfloat64_t _ZGVsMxv_atan (svfloat64_t, svbool_t);
 svfloat64_t _ZGVsMxvv_atan2 (svfloat64_t, svfloat64_t, svbool_t);
 svfloat32_t _ZGVsMxv_cosf (svfloat32_t, svbool_t);
 svfloat64_t _ZGVsMxv_cos (svfloat64_t, svbool_t);
+svfloat32_t _ZGVsMxv_logf (svfloat32_t, svbool_t);
 svfloat32_t _ZGVsMxv_log10f (svfloat32_t, svbool_t);
 svfloat64_t _ZGVsMxv_log10 (svfloat64_t, svbool_t);
 svfloat32_t _ZGVsMxv_sinf (svfloat32_t, svbool_t);
diff --git a/pl/math/math_config.h b/pl/math/math_config.h
index 10bbd9b..5e22349 100644
--- a/pl/math/math_config.h
+++ b/pl/math/math_config.h
@@ -511,4 +511,7 @@ extern const struct v_log10_data
 #define V_LOG10F_POLY_ORDER 9
 extern const float __v_log10f_poly[V_LOG10F_POLY_ORDER - 1] HIDDEN;
 
+#define SV_LOGF_POLY_ORDER 8
+extern const float __sv_logf_poly[SV_LOGF_POLY_ORDER - 1] HIDDEN;
+
 #endif
diff --git a/pl/math/sv_logf_3u4.c b/pl/math/sv_logf_3u4.c
new file mode 100644
index 0000000..125f806
--- /dev/null
+++ b/pl/math/sv_logf_3u4.c
@@ -0,0 +1,66 @@
+/*
+ * Single-precision vector log function.
+ *
+ * Copyright (c) 2019-2022, Arm Limited.
+ * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
+ */
+
+#include "sv_math.h"
+#if SV_SUPPORTED
+
+#define P(i) __sv_logf_poly[i]
+
+#define Ln2 (0x1.62e43p-1f) /* 0x3f317218 */
+#define Min (0x00800000)
+#define Max (0x7f800000)
+#define Mask (0x007fffff)
+#define Off (0x3f2aaaab) /* 0.666667 */
+
+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)
+{
+  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)
+{
+  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));
+
+  /* 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);
+
+  /* y = log(1+r) + n*ln2.  */
+  sv_f32_t r2 = svmul_f32_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);
+
+  if (unlikely (svptest_any (pg, cmp)))
+    return __sv_logf_specialcase (x, y, cmp);
+  return y;
+}
+
+strong_alias (__sv_logf_x, _ZGVsMxv_logf)
+
+#endif // SV_SUPPORTED
diff --git a/pl/math/sv_logf_data.c b/pl/math/sv_logf_data.c
new file mode 100644
index 0000000..0082ee3
--- /dev/null
+++ b/pl/math/sv_logf_data.c
@@ -0,0 +1,12 @@
+/*
+ * Coefficients for single-precision SVE log function.
+ *
+ * Copyright (c) 2019-2022, Arm Limited.
+ * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
+ */
+
+const float __sv_logf_poly[] = {
+  /* Copied from coeffs for the Neon routine in math/.  */
+  -0x1.3e737cp-3f, 0x1.5a9aa2p-3f, -0x1.4f9934p-3f, 0x1.961348p-3f,
+  -0x1.00187cp-2f, 0x1.555d7cp-2f, -0x1.ffffc8p-2f,
+};
diff --git a/pl/math/test/mathbench_funcs.h b/pl/math/test/mathbench_funcs.h
index 076340e..1ecd1d4 100644
--- a/pl/math/test/mathbench_funcs.h
+++ b/pl/math/test/mathbench_funcs.h
@@ -117,6 +117,9 @@ SVF (_ZGVsMxv_cosf, -3.1, 3.1)
 SVF (__sv_sinf_x, -3.1, 3.1)
 SVF (_ZGVsMxv_sinf, -3.1, 3.1)
 
+SVF (__sv_logf_x, 0.01, 11.1)
+SVF (_ZGVsMxv_logf, 0.01, 11.1)
+
 SVF (__sv_log10f_x, 0.01, 11.1)
 SVF (_ZGVsMxv_log10f, 0.01, 11.1)
 SVD (__sv_log10_x, 0.01, 11.1)
diff --git a/pl/math/test/runulp.sh b/pl/math/test/runulp.sh
index 3475adb..656d7a7 100755
--- a/pl/math/test/runulp.sh
+++ b/pl/math/test/runulp.sh
@@ -332,6 +332,15 @@ range_sve_log10f='
       100       inf  50000
 '
 
+range_sve_logf='
+     -0.0  -0x1p126  100
+ 0x1p-149  0x1p-126  4000
+ 0x1p-126   0x1p-23  50000
+  0x1p-23       1.0  50000
+      1.0       100  50000
+      100       inf  50000
+'
+
 # error limits
 L_erfc=3.7
 L_erfcf=1.0
@@ -359,6 +368,7 @@ L_sve_atan2f=3.0
 L_sve_atan2=2.0
 L_sve_log10=2.5
 L_sve_log10f=3.5
+L_sve_logf=3.5
 
 while read G F R
 do
@@ -446,6 +456,8 @@ sve_atanf    __sv_atanf        $runsv
 sve_atanf    _ZGVsMxv_atanf    $runsv
 sve_log10f   __sv_log10f       $runsv
 sve_log10f   _ZGVsMxv_log10f   $runsv
+sve_logf     __sv_logf         $runsv
+sve_logf     _ZGVsMxv_logf     $runsv
 
 sve_cos    __sv_cos        $runsv
 sve_cos    _ZGVsMxv_cos    $runsv
diff --git a/pl/math/test/ulp_funcs.h b/pl/math/test/ulp_funcs.h
index 78fddac..01d7bd1 100644
--- a/pl/math/test/ulp_funcs.h
+++ b/pl/math/test/ulp_funcs.h
@@ -75,6 +75,8 @@ SVF1 (cos)
 ZSVF1 (cos)
 SVD1 (cos)
 ZSVD1 (cos)
+SVF1 (log)
+ZSVF1 (log)
 SVF1 (log10)
 ZSVF1 (log10)
 SVD1 (log10)
diff --git a/pl/math/test/ulp_wrappers.h b/pl/math/test/ulp_wrappers.h
index d275271..07fa077 100644
--- a/pl/math/test/ulp_wrappers.h
+++ b/pl/math/test/ulp_wrappers.h
@@ -88,6 +88,7 @@ ZVND1_WRAP(log2)
 ZSVNF2_WRAP(atan2)
 ZSVNF1_WRAP(atan)
 ZSVNF1_WRAP(cos)
+ZSVNF1_WRAP(log)
 ZSVNF1_WRAP(log10)
 ZSVNF1_WRAP(sin)