Fix rint SSE/NEON again, using optimization barrier.

This is a new version of !423, which failed for MSVC.

Defined `EIGEN_OPTIMIZATION_BARRIER(X)` that uses inline assembly to
prevent operations involving `X` from crossing that barrier. Should
work on most `GNUC` compatible compilers (MSVC doesn't seem to need
this). This is a modified version adapted from what was used in
`psincos_float` and tested on more platforms
(see #1674, https://godbolt.org/z/73ezTG).

Modified `rint` to use the barrier to prevent the add/subtract rounding
trick from being optimized away.

Also fixed an edge case for large inputs that get bumped up a power of two
and ends up rounding away more than just the fractional part.  If we are
over `2^digits` then just return the input.  This edge case was missed in
the test since the test was comparing approximate equality, which was still
satisfied.  Adding a strict equality option catches it.
This commit is contained in:
Antonio Sanchez
2021-03-03 19:22:15 -08:00
parent 5f0b4a4010
commit 82d61af3a4
6 changed files with 172 additions and 50 deletions

View File

@@ -51,7 +51,11 @@
#ifndef EIGEN_STACK_ALLOCATION_LIMIT
// 131072 == 128 KB
#define EIGEN_STACK_ALLOCATION_LIMIT 131072
#if defined(__AVX512F__)
#define EIGEN_STACK_ALLOCATION_LIMIT 0
#else
#define EIGEN_STACK_ALLOCATION_LIMIT 16384
#endif
#endif
//------------------------------------------------------------------------------------------
@@ -1063,6 +1067,64 @@ namespace Eigen {
#endif
// Acts as a barrier preventing operations involving `X` from crossing. This
// occurs, for example, in the fast rounding trick where a magic constant is
// added then subtracted, which is otherwise compiled away with -ffast-math.
//
// See bug 1674
#if !defined(EIGEN_OPTIMIZATION_BARRIER)
#if EIGEN_COMP_GNUC
// According to https://gcc.gnu.org/onlinedocs/gcc/Constraints.html:
// X: Any operand whatsoever.
// r: A register operand is allowed provided that it is in a general
// register.
// g: Any register, memory or immediate integer operand is allowed, except
// for registers that are not general registers.
// w: (AArch32/AArch64) Floating point register, Advanced SIMD vector
// register or SVE vector register.
// x: (SSE) Any SSE register.
// (AArch64) Like w, but restricted to registers 0 to 15 inclusive.
// v: (PowerPC) An Altivec vector register.
// wa:(PowerPC) A VSX register.
//
// "X" (uppercase) should work for all cases, though this seems to fail for
// some versions of GCC for arm/aarch64 with
// "error: inconsistent operand constraints in an 'asm'"
// Clang x86_64/arm/aarch64 seems to require "g" to support both scalars and
// vectors, otherwise
// "error: non-trivial scalar-to-vector conversion, possible invalid
// constraint for vector type"
//
// GCC for ppc64le generates an internal compiler error with x/X/g.
// GCC for AVX generates an internal compiler error with X.
//
// Tested on icc/gcc/clang for sse, avx, avx2, avx512dq
// gcc for arm, aarch64,
// gcc for ppc64le,
// both vectors and scalars.
//
// Note that this is restricted to plain types - this will not work
// directly for std::complex<T>, Eigen::half, Eigen::bfloat16. For these,
// you will need to apply to the underlying POD type.
#if EIGEN_ARCH_PPC
// General, Altivec, VSX.
#define EIGEN_OPTIMIZATION_BARRIER(X) __asm__ ("" : "+r,v,wa" (X));
#elif EIGEN_ARCH_ARM_OR_ARM64
// General, NEON.
#define EIGEN_OPTIMIZATION_BARRIER(X) __asm__ ("" : "+g,w" (X));
#elif EIGEN_ARCH_i386_OR_x86_64
// General, SSE.
#define EIGEN_OPTIMIZATION_BARRIER(X) __asm__ ("" : "+g,x" (X));
#else
// Not implemented for other architectures.
#define EIGEN_OPTIMIZATION_BARRIER(X)
#endif
#else
// Not implemented for other compilers.
#define EIGEN_OPTIMIZATION_BARRIER(X)
#endif
#endif
#if EIGEN_COMP_MSVC
// NOTE MSVC often gives C4127 warnings with compiletime if statements. See bug 1362.
// This workaround is ugly, but it does the job.