From 9a663973b499a53be3e30eb7521545af820fdfc4 Mon Sep 17 00:00:00 2001
From: Antonio Sánchez <cantonios@google.com>
Date: Wed, 3 Mar 2021 18:51:51 +0000
Subject: Revert "Fix rint for SSE/NEON."

This reverts commit e72dfeb8b9fa5662831b5d0bb9d132521f9173dd
---
 Eigen/src/Core/arch/NEON/PacketMath.h | 34 ++++++++++------------------------
 Eigen/src/Core/arch/SSE/PacketMath.h  | 35 ++++++++++-------------------------
 2 files changed, 20 insertions(+), 49 deletions(-)

(limited to 'Eigen/src')

diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h
index 51cebaf2b..ec6ea90c5 100644
--- a/Eigen/src/Core/arch/NEON/PacketMath.h
+++ b/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -3207,34 +3207,20 @@ template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a)
 
 template<> EIGEN_STRONG_INLINE Packet4f print(const Packet4f& a) {
   // Adds and subtracts signum(a) * 2^23 to force rounding.
-  const Packet4f limit = pset1<Packet4f>(static_cast<float>(1<<23));
-  const Packet4f abs_a = pabs(a);
-  // Inline asm to prevent the compiler from optimizing away the
-  // addition and subtraction.
-  // Packet4f r = psub(padd(abs_a, limit), limit);
-  Packet4f r = abs_a;
-  __asm__ ("vadd.f32 %[r], %[r], %[limit]\n\t"
-           "vsub.f32 %[r], %[r], %[limit]" : [r] "+x" (r) : [limit] "x" (limit));
-  // If greater than limit, simply return a.  Otherwise, account for sign.
-  r = pselect(pcmp_lt(abs_a, limit),
-              pselect(pcmp_lt(a, pzero(a)), pnegate(r), r), a);
-  return r;
+  const Packet4f offset = 
+    pselect(pcmp_lt(a, pzero(a)), 
+      pset1<Packet4f>(-static_cast<float>(1<<23)),
+      pset1<Packet4f>(+static_cast<float>(1<<23)));
+  return psub(padd(a, offset), offset);
 }
 
 template<> EIGEN_STRONG_INLINE Packet2f print(const Packet2f& a) {
   // Adds and subtracts signum(a) * 2^23 to force rounding.
-  const Packet2f limit = pset1<Packet2f>(static_cast<float>(1<<23));
-  const Packet2f abs_a = pabs(a);
-  // Inline asm to prevent the compiler from optimizing away the
-  // addition and subtraction.
-  // Packet4f r = psub(padd(abs_a, limit), limit);
-  Packet2f r = abs_a;
-  __asm__ ("vadd.f32 %[r], %[r], %[limit]\n\t"
-           "vsub.f32 %[r], %[r], %[limit]" : [r] "+x" (r) : [limit] "x" (limit));
-  // If greater than limit, simply return a.  Otherwise, account for sign.
-  r = pselect(pcmp_lt(abs_a, limit),
-              pselect(pcmp_lt(a, pzero(a)), pnegate(r), r), a);
-  return r;
+  const Packet2f offset = 
+    pselect(pcmp_lt(a, pzero(a)), 
+      pset1<Packet2f>(-static_cast<float>(1<<23)),
+      pset1<Packet2f>(+static_cast<float>(1<<23)));
+  return psub(padd(a, offset), offset);
 }
 
 template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a)
diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h
index 652ad1d34..b9821ad80 100755
--- a/Eigen/src/Core/arch/SSE/PacketMath.h
+++ b/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -646,35 +646,20 @@ template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { re
 #else
 template<> EIGEN_STRONG_INLINE Packet4f print(const Packet4f& a) {
   // Adds and subtracts signum(a) * 2^23 to force rounding.
-  const Packet4f limit = pset1<Packet4f>(static_cast<float>(1<<23));
-  const Packet4f abs_a = pabs(a);
-  // Inline asm to prevent the compiler from optimizing away the
-  // addition and subtraction.
-  // Packet4f r = psub(padd(abs_a, limit), limit);
-  Packet4f r = abs_a;
-  __asm__ ("addps %[limit], %[r]\n\t"
-           "subps %[limit], %[r]" : [r] "+x" (r) : [limit] "x" (limit));
-  // If greater than limit, simply return a.  Otherwise, account for sign.
-  r = pselect(pcmp_lt(abs_a, limit),
-              pselect(pcmp_lt(a, pzero(a)), pnegate(r), r), a);
-  return r;
+  const Packet4f offset = 
+    pselect(pcmp_lt(a, pzero(a)), 
+      pset1<Packet4f>(-static_cast<float>(1<<23)),
+      pset1<Packet4f>(+static_cast<float>(1<<23)));
+  return psub(padd(a, offset), offset);
 }
 
 template<> EIGEN_STRONG_INLINE Packet2d print(const Packet2d& a) {
   // Adds and subtracts signum(a) * 2^52 to force rounding.
-  const Packet2d limit = pset1<Packet2d>(static_cast<double>(1ull<<52));
-  const Packet2d abs_a = pabs(a);
-  // Inline asm to prevent the compiler from optimizing away the
-  // addition and subtraction.
-  // Packet2d r = psub(padd(abs_a, limit), limit);
-  Packet2d r = abs_a;
-  asm("addpd %[limit], %[r] \n\t"
-      "subpd %[limit], %[r]" : [r] "+x" (r) : [limit] "x" (limit));
-
-  // If greater than limit, simply return a.  Otherwise, account for sign.
-  r = pselect(pcmp_lt(abs_a, limit),
-              pselect(pcmp_lt(a, pzero(a)), pnegate(r), r), a);
-  return r;
+  const Packet2d offset = 
+    pselect(pcmp_lt(a, pzero(a)), 
+      pset1<Packet2d>(-static_cast<double>(1ull<<52)),
+      pset1<Packet2d>(+static_cast<double>(1ull<<52)));
+  return psub(padd(a, offset), offset);
 }
 
 template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a)
-- 
cgit v1.2.3