Source: fftw3 Version: 3.3.3-7 Severity: normal Tags: patch User: debian-...@lists.debian.org Usertag: arm64
In version 3.3.3-6 arm64 was added to the --enable-neon arch list. This is correct in principle but doesn't work because both the enabling compiler option and the compiler define have changed between the 32-bit armv7 (for which this code was written) and armv8 (the 64-bit arch, called 'arm64' in debian and 'aarch64' by ARM and GNU triplets). neon is always enabled on armv8/arm64 unless explictly disabled so the -mfpu=neon option is not understood by the arm64/aarch64 compiler. Also __ARM_NEON__ is not defined by the compiler. So this needs changes in configure.ac and simd-neon.h However, applying this doesn't in fact get a working build as it exposes a compiler ICE in the gcc4.8 we are currently using in unstable. But that should get fixed soon. Note that full support for arm64 will include enabling double-precision, as that is available in armv8 neon, and writing code to support it. Also note that armv8 has fused multiply-accumulate by default, which apparently needs to be recorded for this package. I failed to understand exactly what was going on with that so have not changed anything in this patch, not least because doing so prduced another failure to build, as have_fma and --enable-neon are deemed incompatible. -- System Information: Debian Release: 7.3 APT prefers stable APT policy: (990, 'stable') Architecture: i386 (i686) Kernel: Linux 3.2.0-kvm-i386-20110111 (SMP w/1 CPU core) Locale: LANG=en_GB.UTF-8, LC_CTYPE=en_GB.UTF-8 (charmap=UTF-8) Shell: /bin/sh linked to /bin/bash
diff -Nru fftw3-3.3.3/debian/changelog fftw3-3.3.3/debian/changelog --- fftw3-3.3.3/debian/changelog 2013-10-27 13:32:57.000000000 +0000 +++ fftw3-3.3.3/debian/changelog 2014-01-07 18:04:54.000000000 +0000 @@ -1,3 +1,9 @@ +fftw3 (3.3.3-7arm641) unstable; urgency=low + + * Correct neon arm64 support in configurey + + -- <enevill@apm2> Tue, 07 Jan 2014 18:03:45 +0000 + fftw3 (3.3.3-7) unstable; urgency=low * restrict architectures of long dependency of libfftw3-3 (Closes: #710890) diff -Nru fftw3-3.3.3/debian/patches/arm64-neon-config.patch fftw3-3.3.3/debian/patches/arm64-neon-config.patch --- fftw3-3.3.3/debian/patches/arm64-neon-config.patch 1970-01-01 00:00:00.000000000 +0000 +++ fftw3-3.3.3/debian/patches/arm64-neon-config.patch 2014-01-07 18:43:56.000000000 +0000 @@ -0,0 +1,27 @@ +Index: fftw3-3.3.3/configure.ac +=================================================================== +--- fftw3-3.3.3.orig/configure.ac 2012-11-25 12:33:44.000000000 +0000 ++++ fftw3-3.3.3/configure.ac 2014-01-07 18:25:34.890000000 +0000 +@@ -273,7 +273,8 @@ + [AC_MSG_ERROR([Need a version of gcc with -maltivec])])])]) + fi + +- if test "$have_neon" = "yes" -a "x$NEON_CFLAGS" = x; then ++ # armhf (no flag needed on arm64) ++ if test "$host_cpu" = "arm" -a "$have_neon" = "yes" -a "x$NEON_CFLAGS" = x; then + AX_CHECK_COMPILER_FLAGS(-mfpu=neon, [NEON_CFLAGS="-mfpu=neon"], + [AC_MSG_ERROR([Need a version of gcc with -mfpu=neon])]) + fi +Index: fftw3-3.3.3/simd-support/simd-neon.h +=================================================================== +--- fftw3-3.3.3.orig/simd-support/simd-neon.h 2012-11-25 12:33:44.000000000 +0000 ++++ fftw3-3.3.3/simd-support/simd-neon.h 2014-01-07 18:43:48.500000000 +0000 +@@ -29,7 +29,7 @@ + #define SIMD_VSTRIDE_OKA(x) ((x) == 2) + #define SIMD_STRIDE_OKPAIR SIMD_STRIDE_OK + +-#if defined(__GNUC__) && !defined(__ARM_NEON__) ++#if defined(__GNUC__) && !defined(__ARM_NEON__) && !defined (__aarch64__) + #error "compiling simd-neon.h requires -mfpu=neon or equivalent" + #endif + diff -Nru fftw3-3.3.3/debian/patches/series fftw3-3.3.3/debian/patches/series --- fftw3-3.3.3/debian/patches/series 2013-10-27 13:32:57.000000000 +0000 +++ fftw3-3.3.3/debian/patches/series 2014-01-07 17:38:41.000000000 +0000 @@ -5,3 +5,4 @@ texinfo5-compat.patch wisdom-manpage.patch fix-fftw-wisdom-to-conf.patch +arm64-neon-config.patch
