Fix arm32 issues.
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h
index 8fb5b68..d84b1cc 100644
--- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h
+++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h
@@ -582,8 +582,8 @@
// Subtract y * Pi/2 to reduce x to the interval -Pi/4 <= x <= +Pi/4
// using "Extended precision modular arithmetic"
-#if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD)
- // This version requires true FMA for high accuracy
+#if defined(EIGEN_VECTORIZE_FMA)
+ // This version requires true FMA for high accuracy.
// It provides a max error of 1ULP up to (with absolute_error < 5.9605e-08):
const float huge_th = ComputeSine ? 117435.992f : 71476.0625f;
x = pmadd(y, pset1<Packet>(-1.57079601287841796875f), x);
@@ -1181,7 +1181,7 @@
s_lo = psub(y, t);
}
-#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#ifdef EIGEN_VECTORIZE_FMA
// This function implements the extended precision product of
// a pair of floating point numbers. Given {x, y}, it computes the pair
// {p_hi, p_lo} such that x * y = p_hi + p_lo holds exactly and
@@ -1227,7 +1227,7 @@
p_lo = pmadd(x_lo, y_lo, p_lo);
}
-#endif // EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#endif // EIGEN_VECTORIZE_FMA
// This function implements Dekker's algorithm for the addition
// of two double word numbers represented by {x_hi, x_lo} and {y_hi, y_lo}.
diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h
index 4e3a14d..71e5f5f 100644
--- a/Eigen/src/Core/arch/NEON/PacketMath.h
+++ b/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -1271,7 +1271,7 @@
return pset1<Packet2ul>(0ULL);
}
-#ifdef __ARM_FEATURE_FMA
+#ifdef EIGEN_VECTORIZE_FMA
template <>
EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) {
return vfmaq_f32(c, a, b);
@@ -5249,7 +5249,7 @@
return vdivq_f64(a, b);
}
-#ifdef __ARM_FEATURE_FMA
+#ifdef EIGEN_VECTORIZE_FMA
// See bug 936. See above comment about FMA for float.
template <>
EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) {
diff --git a/Eigen/src/Core/util/ConfigureVectorization.h b/Eigen/src/Core/util/ConfigureVectorization.h
index b16952a..e692438 100644
--- a/Eigen/src/Core/util/ConfigureVectorization.h
+++ b/Eigen/src/Core/util/ConfigureVectorization.h
@@ -354,6 +354,7 @@
#define EIGEN_VECTORIZE
#define EIGEN_VECTORIZE_VSX 1
+#define EIGEN_VECTORIZE_FMA
#include <altivec.h>
// We need to #undef all these ugly tokens defined in <altivec.h>
// => use __vector instead of vector
@@ -365,6 +366,7 @@
#define EIGEN_VECTORIZE
#define EIGEN_VECTORIZE_ALTIVEC
+#define EIGEN_VECTORIZE_FMA
#include <altivec.h>
// We need to #undef all these ugly tokens defined in <altivec.h>
// => use __vector instead of vector
@@ -431,6 +433,11 @@
#include <arm_fp16.h>
#endif
+// Enable FMA for ARM.
+#if defined(__ARM_FEATURE_FMA)
+#define EIGEN_VECTORIZE_FMA
+#endif
+
#if defined(__F16C__) && !defined(EIGEN_GPUCC) && (!EIGEN_COMP_CLANG_STRICT || EIGEN_CLANG_STRICT_AT_LEAST(3, 8, 0))
// We can use the optimized fp16 to float and float to fp16 conversion routines
#define EIGEN_HAS_FP16_C
diff --git a/test/array_cwise.cpp b/test/array_cwise.cpp
index 0cfea8b..91db3f9 100644
--- a/test/array_cwise.cpp
+++ b/test/array_cwise.cpp
@@ -98,9 +98,12 @@
Scalar a = actual(i, j);
#if EIGEN_ARCH_ARM
// Work around NEON flush-to-zero mode
- // if ref returns denormalized value and Eigen returns 0, then skip the test
- int ref_fpclass = std::fpclassify(e);
- if (a == Scalar(0) && ref_fpclass == FP_SUBNORMAL) continue;
+ // if ref returns a subnormal value and Eigen returns 0, then skip the test
+ if (a == Scalar(0) &&
+ (e > -(std::numeric_limits<Scalar>::min)() && e < (std::numeric_limits<Scalar>::min)() &&
+ e >= -std::numeric_limits<Scalar>::denorm_min() && e <= std::numeric_limits<Scalar>::denorm_min())) {
+ continue;
+ }
#endif
bool success = (a == e) || ((numext::isfinite)(e) && internal::isApprox(a, e, tol)) ||
((numext::isnan)(a) && (numext::isnan)(e));
