From d6b1a36e9a1d9b63d2bb285575201346df47df2a Mon Sep 17 00:00:00 2001
From: Sam Yates <halfflat@gmail.com>
Date: Mon, 24 Feb 2020 10:08:44 +0100
Subject: [PATCH] Replace `isnan()` with fp equality test in neon code (#925)
* Replace serialized isnan(x) calls with cmp_eq(x,x) (changing sense of associated tests accordingly).
---
arbor/include/arbor/simd/neon.hpp | 26 +++++---------------------
1 file changed, 5 insertions(+), 21 deletions(-)
diff --git a/arbor/include/arbor/simd/neon.hpp b/arbor/include/arbor/simd/neon.hpp
index 22ae0e7c..beaa00d1 100644
--- a/arbor/include/arbor/simd/neon.hpp
+++ b/arbor/include/arbor/simd/neon.hpp
@@ -395,12 +395,7 @@ struct neon_double2 : implbase<neon_double2> {
auto is_large = cmp_gt(x, broadcast(exp_maxarg));
auto is_small = cmp_lt(x, broadcast(exp_minarg));
-
- bool a[2];
- a[0] = isnan(vgetq_lane_f64(x, 0)) == 0 ? 0 : 1;
- a[1] = isnan(vgetq_lane_f64(x, 1)) == 0 ? 0 : 1;
-
- auto is_nan = mask_copy_from(a);
+ auto is_not_nan = cmp_eq(x, x);
// Compute n and g.
@@ -430,7 +425,7 @@ struct neon_double2 : implbase<neon_double2> {
return ifelse(is_large, broadcast(HUGE_VAL),
ifelse(is_small, broadcast(0),
- ifelse(is_nan, broadcast(NAN), result)));
+ ifelse(is_not_nan, result, broadcast(NAN))));
}
// Use same rational polynomial expansion as for exp(x), without
@@ -443,12 +438,7 @@ struct neon_double2 : implbase<neon_double2> {
static float64x2_t expm1(const float64x2_t& x) {
auto is_large = cmp_gt(x, broadcast(exp_maxarg));
auto is_small = cmp_lt(x, broadcast(expm1_minarg));
-
- bool a[2];
- a[0] = isnan(vgetq_lane_f64(x, 0)) == 0 ? 0 : 1;
- a[1] = isnan(vgetq_lane_f64(x, 1)) == 0 ? 0 : 1;
-
- auto is_nan = mask_copy_from(a);
+ auto is_not_nan = cmp_eq(x, x);
auto half = broadcast(0.5);
auto one = broadcast(1.);
@@ -484,9 +474,7 @@ struct neon_double2 : implbase<neon_double2> {
return ifelse(is_large, broadcast(HUGE_VAL),
ifelse(is_small, broadcast(-1),
- ifelse(is_nan, broadcast(NAN),
- ifelse(nzero, expgm1, scaled))));
- }
+ ifelse(is_not_nan, ifelse(nzero, expgm1, scaled), broadcast(NAN))));
// Natural logarithm:
//
@@ -514,11 +502,7 @@ struct neon_double2 : implbase<neon_double2> {
auto is_small = cmp_lt(x, broadcast(log_minarg));
auto is_domainerr = cmp_lt(x, broadcast(0));
- bool a[2];
- a[0] = isnan(vgetq_lane_f64(x, 0)) == 0 ? 0 : 1;
- a[1] = isnan(vgetq_lane_f64(x, 0)) == 0 ? 0 : 1;
-
- auto is_nan = mask_copy_from(a);
+ auto is_nan = logical_not(cmp_eq(x, x));
is_domainerr = logical_or(is_nan, is_domainerr);
float64x2_t g = vcvt_f64_f32(vcvt_f32_s32(logb_normal(x)));
--
GitLab