diff options
author | Thiago Macieira <thiago.macieira@intel.com> | 2018-05-23 00:54:19 -0300 |
---|---|---|
committer | Thiago Macieira <thiago.macieira@intel.com> | 2018-07-09 00:18:24 +0000 |
commit | c3a4ec5d0bbd5f2710f4fd1d3bd4a2d7f0f507ad (patch) | |
tree | f18c0e54743eb213df0e5374bb5fdf4281aa5c4e /src/corelib | |
parent | 746f15d0c213fef0e46207682815bd839a36ecc6 (diff) |
SIMD: Add a haswell sub-architecture selection to our support
As the comment says, Haswell is a nice divider and is a good
optimization target.
I'm using -march=core-avx2 instead of -march=haswell because the latter
form was only added to GCC 4.9 but we still support 4.7 and that has
support for AVX2.
This commit changes the AVX2-optimized code in QtGui to Haswell-
optimized instead. That means, for example, that qdrawhelper_avx2.cpp
can now use the FMA instructions.
Change-Id: If025d476890745368955fffd153129c1716ba006
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
Diffstat (limited to 'src/corelib')
-rw-r--r-- | src/corelib/tools/qsimd_p.h | 43 |
1 files changed, 42 insertions, 1 deletions
diff --git a/src/corelib/tools/qsimd_p.h b/src/corelib/tools/qsimd_p.h index af262ec88f..9f1321df94 100644 --- a/src/corelib/tools/qsimd_p.h +++ b/src/corelib/tools/qsimd_p.h @@ -1,7 +1,7 @@ /**************************************************************************** ** ** Copyright (C) 2016 The Qt Company Ltd. -** Copyright (C) 2016 Intel Corporation. +** Copyright (C) 2018 Intel Corporation. ** Contact: https://www.qt.io/licensing/ ** ** This file is part of the QtCore module of the Qt Toolkit. @@ -232,8 +232,49 @@ # define __RDRND__ 1 # endif +# if defined(__BMI__) && !defined(__BMI2__) && defined(Q_CC_INTEL) +// BMI2 instructions: +// All processors that support BMI support BMI2 (and AVX2) +// (but neither MSVC nor the Intel compiler define this macro) +# define __BMI2__ 1 +# endif + # include "qsimd_x86_p.h" +// Haswell sub-architecture +// +// The Intel Core 4th generation was codenamed "Haswell" and introduced AVX2, +// BMI1, BMI2, FMA, LZCNT, MOVBE, which makes it a good divider for a +// sub-target for us. The first AMD processor with AVX2 support (Zen) has the +// same features. +// +// macOS's fat binaries support the "x86_64h" sub-architecture and the GNU libc +// ELF loader also supports a "haswell/" subdir (e.g., /usr/lib/haswell). +# define QT_FUNCTION_TARGET_STRING_ARCH_HASWELL "arch=haswell" +# if defined(__AVX2__) && defined(__BMI__) && defined(__BMI2__) && defined(__F16C__) && \ + defined(__FMA__) && defined(__LZCNT__) && defined(__RDRND__) +# define __haswell__ 1 +# endif + +// This constant does not include all CPU features found in a Haswell, only +// those that we'd have optimized code for. +// Note: must use Q_CONSTEXPR here, as this file may be compiled in C mode. +QT_BEGIN_NAMESPACE +static const quint64 CpuFeatureArchHaswell = 0 + | CpuFeatureSSE2 + | CpuFeatureSSE3 + | CpuFeatureSSSE3 + | CpuFeatureSSE4_1 + | CpuFeatureSSE4_2 + | CpuFeatureFMA + | CpuFeaturePOPCNT + | CpuFeatureAVX + | CpuFeatureF16C + | CpuFeatureAVX2 + | CpuFeatureBMI + | CpuFeatureBMI2; +QT_END_NAMESPACE + #endif /* Q_PROCESSOR_X86 */ // Clang compiler fix, see http://lists.llvm.org/pipermail/cfe-commits/Week-of-Mon-20160222/151168.html |