diff mbox series

aarch64: Improve codegen in SVE asinh.

Message ID 20250108093410.3217-1-Luna.lamb@arm.com (mailing list archive)
State New
Headers
Series aarch64: Improve codegen in SVE asinh. |

Checks

Context Check Description
redhat-pt-bot/TryBot-apply_patch success Patch applied to master at the time it was sent
linaro-tcwg-bot/tcwg_glibc_build--master-aarch64 success Build passed
linaro-tcwg-bot/tcwg_glibc_check--master-aarch64 success Test passed
linaro-tcwg-bot/tcwg_glibc_build--master-arm success Build passed
linaro-tcwg-bot/tcwg_glibc_check--master-arm success Test passed
redhat-pt-bot/TryBot-32bit success Build for i686

Commit Message

Luna Lamb Jan. 8, 2025, 9:34 a.m. UTC 
  Use unpredicated muls, use lanewise mla's and improve memory access.
1% regression in throughput microbenchmark on Neoverse V1.
---
OK for master? If so please commit for me as I don't have commit rights.
Thanks,
Luna
 sysdeps/aarch64/fpu/asinh_sve.c | 111 ++++++++++++++++++++++----------
 1 file changed, 77 insertions(+), 34 deletions(-)
diff mbox series

Patch

diff --git a/sysdeps/aarch64/fpu/asinh_sve.c b/sysdeps/aarch64/fpu/asinh_sve.c
index 0889f79dbb..ff6b71390c 100644
--- a/sysdeps/aarch64/fpu/asinh_sve.c
+++ b/sysdeps/aarch64/fpu/asinh_sve.c
@@ -18,36 +18,49 @@ 
    <https://www.gnu.org/licenses/>.  */
 
 #include "sv_math.h"
-#include "poly_sve_f64.h"
 
 #define SignMask (0x8000000000000000)
 #define One (0x3ff0000000000000)
 #define Thres (0x5fe0000000000000) /* asuint64 (0x1p511).  */
+#define IndexMask (((1 << V_LOG_TABLE_BITS) - 1) << 1)
 
 static const struct data
 {
-  double poly[18];
-  double ln2, p3, p1, p4, p0, p2;
-  uint64_t n;
-  uint64_t off;
+  double even_coeffs[9];
+  double ln2, p3, p1, p4, p0, p2, c1, c3, c5, c7, c9, c11, c13, c15, c17;
+  uint64_t off, mask;
 
 } data = {
-  /* Polynomial generated using Remez on [2^-26, 1].  */
-  .poly
-  = { -0x1.55555555554a7p-3, 0x1.3333333326c7p-4, -0x1.6db6db68332e6p-5,
-      0x1.f1c71b26fb40dp-6, -0x1.6e8b8b654a621p-6, 0x1.1c4daa9e67871p-6,
-      -0x1.c9871d10885afp-7, 0x1.7a16e8d9d2ecfp-7, -0x1.3ddca533e9f54p-7,
-      0x1.0becef748dafcp-7, -0x1.b90c7099dd397p-8, 0x1.541f2bb1ffe51p-8,
-      -0x1.d217026a669ecp-9, 0x1.0b5c7977aaf7p-9, -0x1.e0f37daef9127p-11,
-      0x1.388b5fe542a6p-12, -0x1.021a48685e287p-14, 0x1.93d4ba83d34dap-18 },
+   /* Polynomial generated using Remez on [2^-26, 1].  */
+  .even_coeffs ={
+    -0x1.55555555554a7p-3,
+    -0x1.6db6db68332e6p-5,
+    -0x1.6e8b8b654a621p-6,
+    -0x1.c9871d10885afp-7,
+    -0x1.3ddca533e9f54p-7,
+    -0x1.b90c7099dd397p-8,
+    -0x1.d217026a669ecp-9,
+    -0x1.e0f37daef9127p-11,
+    -0x1.021a48685e287p-14, },
+
+  .c1 = 0x1.3333333326c7p-4,
+  .c3 = 0x1.f1c71b26fb40dp-6,
+  .c5 = 0x1.1c4daa9e67871p-6,
+  .c7 = 0x1.7a16e8d9d2ecfp-7,
+  .c9 = 0x1.0becef748dafcp-7,
+  .c11 = 0x1.541f2bb1ffe51p-8,
+  .c13 = 0x1.0b5c7977aaf7p-9,
+  .c15 = 0x1.388b5fe542a6p-12,
+  .c17 = 0x1.93d4ba83d34dap-18,
+
   .ln2 = 0x1.62e42fefa39efp-1,
   .p0 = -0x1.ffffffffffff7p-2,
   .p1 = 0x1.55555555170d4p-2,
   .p2 = -0x1.0000000399c27p-2,
   .p3 = 0x1.999b2e90e94cap-3,
   .p4 = -0x1.554e550bd501ep-3,
-  .n = 1 << V_LOG_TABLE_BITS,
-  .off = 0x3fe6900900000000
+  .off = 0x3fe6900900000000,
+  .mask = 0xfffULL << 52,
 };
 
 static svfloat64_t NOINLINE
@@ -64,11 +77,10 @@  __sv_log_inline (svfloat64_t x, const struct data *d, const svbool_t pg)
      of the algorithm used.  */
 
   svuint64_t ix = svreinterpret_u64 (x);
-  svuint64_t tmp = svsub_x (pg, ix, d->off);
-  svuint64_t i = svand_x (pg, svlsr_x (pg, tmp, (51 - V_LOG_TABLE_BITS)),
-			  (d->n - 1) << 1);
-  svint64_t k = svasr_x (pg, svreinterpret_s64 (tmp), 52);
-  svuint64_t iz = svsub_x (pg, ix, svand_x (pg, tmp, 0xfffULL << 52));
+  svuint64_t i_off = svsub_x (pg, ix, d->off);
+  svuint64_t i
+      = svand_x (pg, svlsr_x (pg, i_off, (51 - V_LOG_TABLE_BITS)), IndexMask);
+  svuint64_t iz = svsub_x (pg, ix, svand_x (pg, i_off, d->mask));
   svfloat64_t z = svreinterpret_f64 (iz);
 
   svfloat64_t invc = svld1_gather_index (pg, &__v_log_data.table[0].invc, i);
@@ -78,14 +90,14 @@  __sv_log_inline (svfloat64_t x, const struct data *d, const svbool_t pg)
   svfloat64_t p1_p4 = svld1rq (svptrue_b64 (), &d->p1);
 
   svfloat64_t r = svmla_x (pg, sv_f64 (-1.0), invc, z);
-  svfloat64_t kd = svcvt_f64_x (pg, k);
+  svfloat64_t kd
+      = svcvt_f64_x (pg, svasr_x (pg, svreinterpret_s64 (i_off), 52));
 
   svfloat64_t hi = svmla_lane (svadd_x (pg, logc, r), kd, ln2_p3, 0);
-  svfloat64_t r2 = svmul_x (pg, r, r);
-
+  svfloat64_t r2 = svmul_x (svptrue_b64 (), r, r);
   svfloat64_t y = svmla_lane (sv_f64 (d->p2), r, ln2_p3, 1);
-
   svfloat64_t p = svmla_lane (sv_f64 (d->p0), r, p1_p4, 0);
+
   y = svmla_lane (y, r2, p1_p4, 1);
   y = svmla_x (pg, p, r2, y);
   y = svmla_x (pg, hi, r2, y);
@@ -111,7 +123,6 @@  svfloat64_t SV_NAME_D1 (asinh) (svfloat64_t x, const svbool_t pg)
   svuint64_t iax = svbic_x (pg, ix, SignMask);
   svuint64_t sign = svand_x (pg, ix, SignMask);
   svfloat64_t ax = svreinterpret_f64 (iax);
-
   svbool_t ge1 = svcmpge (pg, iax, One);
   svbool_t special = svcmpge (pg, iax, Thres);
 
@@ -120,7 +131,7 @@  svfloat64_t SV_NAME_D1 (asinh) (svfloat64_t x, const svbool_t pg)
   svfloat64_t option_1 = sv_f64 (0);
   if (__glibc_likely (svptest_any (pg, ge1)))
     {
-      svfloat64_t x2 = svmul_x (pg, ax, ax);
+      svfloat64_t x2 = svmul_x (svptrue_b64 (), ax, ax);
       option_1 = __sv_log_inline (
 	  svadd_x (pg, ax, svsqrt_x (pg, svadd_x (pg, x2, 1))), d, pg);
     }
@@ -130,21 +141,53 @@  svfloat64_t SV_NAME_D1 (asinh) (svfloat64_t x, const svbool_t pg)
      The largest observed error in this region is 1.51 ULPs:
      _ZGVsMxv_asinh(0x1.fe12bf8c616a2p-1) got 0x1.c1e649ee2681bp-1
 					 want 0x1.c1e649ee2681dp-1.  */
+
   svfloat64_t option_2 = sv_f64 (0);
   if (__glibc_likely (svptest_any (pg, svnot_z (pg, ge1))))
     {
-      svfloat64_t x2 = svmul_x (pg, ax, ax);
-      svfloat64_t x4 = svmul_x (pg, x2, x2);
-      svfloat64_t p = sv_pw_horner_17_f64_x (pg, x2, x4, d->poly);
-      option_2 = svmla_x (pg, ax, p, svmul_x (pg, x2, ax));
+      svfloat64_t x2 = svmul_x (svptrue_b64 (), ax, ax);
+      svfloat64_t x4 = svmul_x (svptrue_b64 (), x2, x2);
+      /* Order-17 Pairwise Horner scheme.  */
+      svfloat64_t c13 = svld1rq (svptrue_b64 (), &d->c1);
+      svfloat64_t c57 = svld1rq (svptrue_b64 (), &d->c5);
+      svfloat64_t c911 = svld1rq (svptrue_b64 (), &d->c9);
+      svfloat64_t c1315 = svld1rq (svptrue_b64 (), &d->c13);
+
+      svfloat64_t p01 = svmla_lane (sv_f64 (d->even_coeffs[0]), x2, c13, 0);
+      svfloat64_t p23 = svmla_lane (sv_f64 (d->even_coeffs[1]), x2, c13, 1);
+      svfloat64_t p45 = svmla_lane (sv_f64 (d->even_coeffs[2]), x2, c57, 0);
+      svfloat64_t p67 = svmla_lane (sv_f64 (d->even_coeffs[3]), x2, c57, 1);
+      svfloat64_t p89 = svmla_lane (sv_f64 (d->even_coeffs[4]), x2, c911, 0);
+      svfloat64_t p1011 = svmla_lane (sv_f64 (d->even_coeffs[5]), x2, c911, 1);
+      svfloat64_t p1213
+	  = svmla_lane (sv_f64 (d->even_coeffs[6]), x2, c1315, 0);
+      svfloat64_t p1415
+	  = svmla_lane (sv_f64 (d->even_coeffs[7]), x2, c1315, 1);
+      svfloat64_t p1617 = svmla_x (pg, sv_f64 (d->even_coeffs[8]), x2, d->c17);
+
+      svfloat64_t p = svmla_x (pg, p1415, x4, p1617);
+      p = svmla_x (pg, p1213, x4, p);
+      p = svmla_x (pg, p1011, x4, p);
+      p = svmla_x (pg, p89, x4, p);
+
+      p = svmla_x (pg, p67, x4, p);
+      p = svmla_x (pg, p45, x4, p);
+
+      p = svmla_x (pg, p23, x4, p);
+
+      p = svmla_x (pg, p01, x4, p);
+
+      option_2 = svmla_x (pg, ax, p, svmul_x (svptrue_b64 (), x2, ax));
     }
 
-  /* Choose the right option for each lane.  */
-  svfloat64_t y = svsel (ge1, option_1, option_2);
-
   if (__glibc_unlikely (svptest_any (pg, special)))
     return special_case (
-	x, svreinterpret_f64 (sveor_x (pg, svreinterpret_u64 (y), sign)),
+	x,
+	svreinterpret_f64 (sveor_x (
+	    pg, svreinterpret_u64 (svsel (ge1, option_1, option_2)), sign)),
 	special);
+
+  /* Choose the right option for each lane.  */
+  svfloat64_t y = svsel (ge1, option_1, option_2);
   return svreinterpret_f64 (sveor_x (pg, svreinterpret_u64 (y), sign));
 }