From 86d35caf94c2dad171dc27f82005131b5afed9ef Mon Sep 17 00:00:00 2001
From: Jim Huang <jserv.tw@gmail.com>
Date: Sat, 29 Jan 2022 22:21:01 +0800
Subject: [PATCH] Rewrite __frsqrte with Arm/NEON counterpart

__frsqrte is the intrinsic for floating reciprocal square root estimate.
In Arm64, we can implement it with NEON intrinsics. Since Armv8.2,
instruction "FRSQRTE" [1] is provided to calculate an approximate square
root for each vector element in the source SIMD and FP register.

With -O3, generated assembly on Apple Silicon M1:

[original]
	fsqrt	s0, s0
	fmov	s1, #1.00000000
	fdiv	s0, s1, s0

[neon]
	dup.4s	v0, v0[0]
	frsqrte.4s	v0, v0

[1] https://developer.arm.com/documentation/100076/0100/a64-instruction-set-reference/a64-simd-scalar-instructions/frsqrte--scalar-
---
 src/Pomme.h | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/src/Pomme.h b/src/Pomme.h
index c3eb024..0ed6823 100644
--- a/src/Pomme.h
+++ b/src/Pomme.h
@@ -16,7 +16,15 @@
 
 #define __fres(x) (1.0f/x)
 #define __fabs(x) fabs(x)
+#if defined(__aarch64__)
+#include <arm_neon.h>
+static inline float __frsqrte(float f)
+{
+    return vrsqrteq_f32(vdupq_n_f32(f))[0];
+}
+#else
 #define __frsqrte(x) (1.0f/sqrtf(x))
+#endif
 
 //-----------------------------------------------------------------------------
 // Source code compat