From 1aa7d5922930c4195bc8747a19524a7e5021c4e6 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Samuel=20R=C3=B8dal?= Date: Thu, 17 Dec 2009 15:35:23 +0100 Subject: NEON configure detection and initial blend function implementations. Adds new NEON configure test and -no-neon configure option. NEON implementations can also be turned off by setting the QT_NO_NEON environment variable. Performance improvements (in frames per second): - Blending ARGB32 on RGB32/ARGB32, mostly opaque: 71 % - Blending ARGB32 on RGB32/ARGB32, no opaque pixels: 108 % - Blending ARGB32 on RGB32/ARGB32, with 0.5 opacity: 158 % - Blending RGB32 on RGB32/ARGB32, with 0.5 opacity: 189 % Task-number: QTBUG-6684 Reviewed-by: Gunnar Sletta Reviewed-by: Paul Olav Tvete --- config.tests/unix/neon/neon.cpp | 51 +++++++ config.tests/unix/neon/neon.pro | 2 + configure | 22 ++- src/gui/painting/painting.pri | 7 + src/gui/painting/qblendfunctions.cpp | 2 +- src/gui/painting/qdrawhelper.cpp | 16 ++- src/gui/painting/qdrawhelper_neon.cpp | 260 ++++++++++++++++++++++++++++++++++ src/gui/painting/qdrawhelper_neon_p.h | 76 ++++++++++ 8 files changed, 432 insertions(+), 4 deletions(-) create mode 100644 config.tests/unix/neon/neon.cpp create mode 100644 config.tests/unix/neon/neon.pro create mode 100644 src/gui/painting/qdrawhelper_neon.cpp create mode 100644 src/gui/painting/qdrawhelper_neon_p.h diff --git a/config.tests/unix/neon/neon.cpp b/config.tests/unix/neon/neon.cpp new file mode 100644 index 0000000..c31a9fd --- /dev/null +++ b/config.tests/unix/neon/neon.cpp @@ -0,0 +1,51 @@ +/**************************************************************************** +** +** Copyright (C) 2009 Nokia Corporation and/or its subsidiary(-ies). +** All rights reserved. +** Contact: Nokia Corporation (qt-info@nokia.com) +** +** This file is part of the config.tests of the Qt Toolkit. +** +** $QT_BEGIN_LICENSE:LGPL$ +** No Commercial Usage +** This file contains pre-release code and may not be distributed. +** You may use this file in accordance with the terms and conditions +** contained in the Technology Preview License Agreement accompanying +** this package. +** +** GNU Lesser General Public License Usage +** Alternatively, this file may be used under the terms of the GNU Lesser +** General Public License version 2.1 as published by the Free Software +** Foundation and appearing in the file LICENSE.LGPL included in the +** packaging of this file. Please review the following information to +** ensure the GNU Lesser General Public License version 2.1 requirements +** will be met: http://www.gnu.org/licenses/old-licenses/lgpl-2.1.html. +** +** In addition, as a special exception, Nokia gives you certain additional +** rights. These rights are described in the Nokia Qt LGPL Exception +** version 1.1, included in the file LGPL_EXCEPTION.txt in this package. +** +** If you have questions regarding the use of this file, please contact +** Nokia at qt-info@nokia.com. +** +** +** +** +** +** +** +** +** $QT_END_LICENSE$ +** +****************************************************************************/ + +#include + +int main(int, char**) +{ + int32x4_t null = vdupq_n_s32(0x0); + + int result; + vst1q_lane_s32(&result, null, 0); + return result; +} diff --git a/config.tests/unix/neon/neon.pro b/config.tests/unix/neon/neon.pro new file mode 100644 index 0000000..de20c4e --- /dev/null +++ b/config.tests/unix/neon/neon.pro @@ -0,0 +1,2 @@ +SOURCES = neon.cpp +CONFIG -= x11 qt diff --git a/configure b/configure index 2114863..22e6bd4 100755 --- a/configure +++ b/configure @@ -745,6 +745,7 @@ CFG_HOST_ENDIAN=auto CFG_DOUBLEFORMAT=auto CFG_ARMFPA=auto CFG_IWMMXT=no +CFG_NEON=auto CFG_CLOCK_GETTIME=auto CFG_CLOCK_MONOTONIC=auto CFG_MREMAP=auto @@ -1604,6 +1605,13 @@ while [ "$#" -gt 0 ]; do iwmmxt) CFG_IWMMXT="yes" ;; + neon) + if [ "$VAL" = "no" ]; then + CFG_NEON="$VAL" + else + UNKNOWN_OPT=yes + fi + ;; reduce-relocations) if [ "$VAL" = "yes" ] || [ "$VAL" = "no" ]; then CFG_REDUCE_RELOCATIONS="$VAL" @@ -3889,6 +3897,7 @@ Qt for Embedded Linux only: -iwmmxt ............ Compile using the iWMMXt instruction set (available on some XScale CPUs). + -no-neon ........... Do not compile with use of NEON instructions. EOF fi @@ -4509,6 +4518,15 @@ if [ "$CFG_IWMMXT" = "yes" ]; then fi fi +# detect neon support +if ([ "$CFG_ARCH" = "arm" ] || [ "$CFG_ARCH" = "armv6" ]) && [ "${CFG_NEON}" = "auto" ]; then + if "$unixtests/compile.test" "$XQMAKESPEC" "$QMAKE_CONFIG" $OPT_VERBOSE "$relpath" "$outpath" config.tests/unix/neon "neon" $L_FLAGS $I_FLAGS $l_FLAGS "-mfpu=neon"; then + CFG_NEON=yes + else + CFG_NEON=no + fi +fi + # detect zlib if [ "$CFG_ZLIB" = "no" ]; then # Note: Qt no longer support builds without zlib @@ -6124,6 +6142,7 @@ fi [ "$CFG_SSE" = "yes" ] && QMAKE_CONFIG="$QMAKE_CONFIG sse" [ "$CFG_SSE2" = "yes" ] && QMAKE_CONFIG="$QMAKE_CONFIG sse2" [ "$CFG_IWMMXT" = "yes" ] && QMAKE_CONFIG="$QMAKE_CONFIG iwmmxt" +[ "$CFG_NEON" = "yes" ] && QMAKE_CONFIG="$QMAKE_CONFIG neon" [ "$PLATFORM_MAC" = "yes" ] && QMAKE_CONFIG="$QMAKE_CONFIG $CFG_MAC_ARCHS" if [ "$CFG_IPV6" = "yes" ]; then QT_CONFIG="$QT_CONFIG ipv6" @@ -7416,8 +7435,9 @@ echo "Declarative module .. $CFG_DECLARATIVE" echo "STL support ......... $CFG_STL" echo "PCH support ......... $CFG_PRECOMPILE" echo "MMX/3DNOW/SSE/SSE2.. ${CFG_MMX}/${CFG_3DNOW}/${CFG_SSE}/${CFG_SSE2}" -if [ "${CFG_ARCH}" = "arm" ]; then +if [ "$CFG_ARCH" = "arm" ] || [ "$CFG_ARCH" = "armv6" ]; then echo "iWMMXt support ...... ${CFG_IWMMXT}" + echo "NEON support ........ ${CFG_NEON}" fi [ "${PLATFORM_QWS}" != "yes" ] && echo "Graphics System ..... $CFG_GRAPHICS_SYSTEM" echo "IPv6 support ........ $CFG_IPV6" diff --git a/src/gui/painting/painting.pri b/src/gui/painting/painting.pri index 628a109..0b1e79a 100644 --- a/src/gui/painting/painting.pri +++ b/src/gui/painting/painting.pri @@ -379,6 +379,13 @@ symbian { QMAKE_CXXFLAGS.ARMCC *= -O3 } +neon { + DEFINES += QT_HAVE_NEON + HEADERS += painting/qdrawhelper_neon_p.h + SOURCES += painting/qdrawhelper_neon.cpp + QMAKE.CXXFLAGS *= -mfpu=neon +} + contains(QT_CONFIG, zlib) { INCLUDEPATH += ../3rdparty/zlib } else:!contains(QT_CONFIG, no-zlib) { diff --git a/src/gui/painting/qblendfunctions.cpp b/src/gui/painting/qblendfunctions.cpp index 1d15dac..81d1515 100644 --- a/src/gui/painting/qblendfunctions.cpp +++ b/src/gui/painting/qblendfunctions.cpp @@ -605,7 +605,7 @@ static void qt_blend_argb32_on_argb32(uchar *destPixels, int dbpl, } -static void qt_blend_rgb32_on_rgb32(uchar *destPixels, int dbpl, +void qt_blend_rgb32_on_rgb32(uchar *destPixels, int dbpl, const uchar *srcPixels, int sbpl, int w, int h, int const_alpha) diff --git a/src/gui/painting/qdrawhelper.cpp b/src/gui/painting/qdrawhelper.cpp index 23236ec..84cf5cc 100644 --- a/src/gui/painting/qdrawhelper.cpp +++ b/src/gui/painting/qdrawhelper.cpp @@ -44,6 +44,7 @@ #include #include #include +#include #include #include @@ -7725,7 +7726,8 @@ enum CPUFeatures { SSE = 0x10, SSE2 = 0x20, CMOV = 0x40, - IWMMXT = 0x80 + IWMMXT = 0x80, + NEON = 0x100 }; static uint detectCPUFeatures() @@ -7751,6 +7753,9 @@ static uint detectCPUFeatures() // runtime detection only available when running as a previlegied process static const bool doIWMMXT = !qgetenv("QT_NO_IWMMXT").toInt(); return doIWMMXT ? IWMMXT : 0; +#elif defined(QT_HAVE_NEON) + static const bool doNEON = !qgetenv("QT_NO_NEON").toInt(); + return doNEON ? NEON : 0; #else uint features = 0; #if defined(__x86_64__) || defined(Q_OS_WIN64) @@ -8122,7 +8127,14 @@ void qInitDrawhelperAsm() qBlendFunctions[QImage::Format_ARGB32_Premultiplied][QImage::Format_RGB32] = qt_blend_rgb32_on_rgb32_armv6; qBlendFunctions[QImage::Format_RGB32][QImage::Format_ARGB32_Premultiplied] = qt_blend_argb32_on_argb32_armv6; qBlendFunctions[QImage::Format_ARGB32_Premultiplied][QImage::Format_ARGB32_Premultiplied] = qt_blend_argb32_on_argb32_armv6; -#endif // Q_CC_RVCT && QT_HAVE_ARMV6 +#elif defined(QT_HAVE_NEON) + if (features & NEON) { + qBlendFunctions[QImage::Format_RGB32][QImage::Format_RGB32] = qt_blend_rgb32_on_rgb32_neon; + qBlendFunctions[QImage::Format_ARGB32_Premultiplied][QImage::Format_RGB32] = qt_blend_rgb32_on_rgb32_neon; + qBlendFunctions[QImage::Format_RGB32][QImage::Format_ARGB32_Premultiplied] = qt_blend_argb32_on_argb32_neon; + qBlendFunctions[QImage::Format_ARGB32_Premultiplied][QImage::Format_ARGB32_Premultiplied] = qt_blend_argb32_on_argb32_neon; + } +#endif if (functionForModeSolidAsm) { const int destinationMode = QPainter::CompositionMode_Destination; diff --git a/src/gui/painting/qdrawhelper_neon.cpp b/src/gui/painting/qdrawhelper_neon.cpp new file mode 100644 index 0000000..7fe11bf --- /dev/null +++ b/src/gui/painting/qdrawhelper_neon.cpp @@ -0,0 +1,260 @@ +/**************************************************************************** +** +** Copyright (C) 2009 Nokia Corporation and/or its subsidiary(-ies). +** All rights reserved. +** Contact: Nokia Corporation (qt-info@nokia.com) +** +** This file is part of the QtGui module of the Qt Toolkit. +** +** $QT_BEGIN_LICENSE:LGPL$ +** No Commercial Usage +** This file contains pre-release code and may not be distributed. +** You may use this file in accordance with the terms and conditions +** contained in the Technology Preview License Agreement accompanying +** this package. +** +** GNU Lesser General Public License Usage +** Alternatively, this file may be used under the terms of the GNU Lesser +** General Public License version 2.1 as published by the Free Software +** Foundation and appearing in the file LICENSE.LGPL included in the +** packaging of this file. Please review the following information to +** ensure the GNU Lesser General Public License version 2.1 requirements +** will be met: http://www.gnu.org/licenses/old-licenses/lgpl-2.1.html. +** +** In addition, as a special exception, Nokia gives you certain additional +** rights. These rights are described in the Nokia Qt LGPL Exception +** version 1.1, included in the file LGPL_EXCEPTION.txt in this package. +** +** If you have questions regarding the use of this file, please contact +** Nokia at qt-info@nokia.com. +** +** +** +** +** +** +** +** +** $QT_END_LICENSE$ +** +****************************************************************************/ + +#include + +#ifdef QT_HAVE_NEON + +#include +#include + +QT_BEGIN_NAMESPACE + +static inline int16x8_t qvdiv_255_s16(int16x8_t x, int16x8_t half) +{ + // result = (x + (x >> 8) + 0x80) >> 8 + + const int16x8_t temp = vshrq_n_s16(x, 8); // x >> 8 + const int16x8_t sum_part = vaddq_s16(x, half); // x + 0x80 + const int16x8_t sum = vaddq_s16(temp, sum_part); + + return vreinterpretq_s16_u16(vshrq_n_u16(vreinterpretq_u16_s16(sum), 8)); +} + +static inline int16x8_t qvbyte_mul_s16(int16x8_t x, int16x8_t alpha, int16x8_t half) +{ + // t = qRound(x * alpha / 255.0) + + const int16x8_t t = vmulq_s16(x, alpha); // t + return qvdiv_255_s16(t, half); +} + +static inline int16x8_t qvinterpolate_pixel_255(int16x8_t x, int16x8_t a, int16x8_t y, int16x8_t b, int16x8_t half) +{ + // t = x * a + y * b + + const int16x8_t ta = vmulq_s16(x, a); + const int16x8_t tb = vmulq_s16(y, b); + + return qvdiv_255_s16(vaddq_s16(ta, tb), half); +} + +static inline int16x8_t qvsource_over_s16(int16x8_t src16, int16x8_t dst16, int16x8_t half, int16x8_t full) +{ + const int16x4_t alpha16_high = vdup_lane_s16(vget_high_s16(src16), 3); + const int16x4_t alpha16_low = vdup_lane_s16(vget_low_s16(src16), 3); + + const int16x8_t alpha16 = vsubq_s16(full, vcombine_s16(alpha16_low, alpha16_high)); + + return vaddq_s16(src16, qvbyte_mul_s16(dst16, alpha16, half)); +} + +void qt_blend_argb32_on_argb32_neon(uchar *destPixels, int dbpl, + const uchar *srcPixels, int sbpl, + int w, int h, + int const_alpha) +{ + const uint *src = (const uint *) srcPixels; + uint *dst = (uint *) destPixels; + int16x8_t half = vdupq_n_s16(0x80); + int16x8_t full = vdupq_n_s16(0xff); + if (const_alpha == 256) { + for (int y = 0; y < h; ++y) { + int x = 0; + for (; x < w-3; x += 4) { + int32x4_t src32 = vld1q_s32((int32_t *)&src[x]); + if ((src[x] & src[x+1] & src[x+2] & src[x+3]) >= 0xff000000) { + // all opaque + vst1q_s32((int32_t *)&dst[x], src32); + } else if (src[x] | src[x+1] | src[x+2] | src[x+3]) { + int32x4_t dst32 = vld1q_s32((int32_t *)&dst[x]); + + const uint8x16_t src8 = vreinterpretq_u8_s32(src32); + const uint8x16_t dst8 = vreinterpretq_u8_s32(dst32); + + const uint8x8_t src8_low = vget_low_u8(src8); + const uint8x8_t dst8_low = vget_low_u8(dst8); + + const uint8x8_t src8_high = vget_high_u8(src8); + const uint8x8_t dst8_high = vget_high_u8(dst8); + + const int16x8_t src16_low = vreinterpretq_s16_u16(vmovl_u8(src8_low)); + const int16x8_t dst16_low = vreinterpretq_s16_u16(vmovl_u8(dst8_low)); + + const int16x8_t src16_high = vreinterpretq_s16_u16(vmovl_u8(src8_high)); + const int16x8_t dst16_high = vreinterpretq_s16_u16(vmovl_u8(dst8_high)); + + const int16x8_t result16_low = qvsource_over_s16(src16_low, dst16_low, half, full); + const int16x8_t result16_high = qvsource_over_s16(src16_high, dst16_high, half, full); + + const int32x2_t result32_low = vreinterpret_s32_s8(vmovn_s16(result16_low)); + const int32x2_t result32_high = vreinterpret_s32_s8(vmovn_s16(result16_high)); + + vst1q_s32((int32_t *)&dst[x], vcombine_s32(result32_low, result32_high)); + } + } + for (; x= 0xff000000) + dst[x] = s; + else if (s != 0) + dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s)); + } + dst = (quint32 *)(((uchar *) dst) + dbpl); + src = (const quint32 *)(((const uchar *) src) + sbpl); + } + } else if (const_alpha != 0) { + const_alpha = (const_alpha * 255) >> 8; + int16x8_t const_alpha16 = vdupq_n_s16(const_alpha); + for (int y = 0; y < h; ++y) { + int x = 0; + for (; x < w-3; x += 4) { + if (src[x] | src[x+1] | src[x+2] | src[x+3]) { + int32x4_t src32 = vld1q_s32((int32_t *)&src[x]); + int32x4_t dst32 = vld1q_s32((int32_t *)&dst[x]); + + const uint8x16_t src8 = vreinterpretq_u8_s32(src32); + const uint8x16_t dst8 = vreinterpretq_u8_s32(dst32); + + const uint8x8_t src8_low = vget_low_u8(src8); + const uint8x8_t dst8_low = vget_low_u8(dst8); + + const uint8x8_t src8_high = vget_high_u8(src8); + const uint8x8_t dst8_high = vget_high_u8(dst8); + + const int16x8_t src16_low = vreinterpretq_s16_u16(vmovl_u8(src8_low)); + const int16x8_t dst16_low = vreinterpretq_s16_u16(vmovl_u8(dst8_low)); + + const int16x8_t src16_high = vreinterpretq_s16_u16(vmovl_u8(src8_high)); + const int16x8_t dst16_high = vreinterpretq_s16_u16(vmovl_u8(dst8_high)); + + const int16x8_t srcalpha16_low = qvbyte_mul_s16(src16_low, const_alpha16, half); + const int16x8_t srcalpha16_high = qvbyte_mul_s16(src16_high, const_alpha16, half); + + const int16x8_t result16_low = qvsource_over_s16(srcalpha16_low, dst16_low, half, full); + const int16x8_t result16_high = qvsource_over_s16(srcalpha16_high, dst16_high, half, full); + + const int32x2_t result32_low = vreinterpret_s32_s8(vmovn_s16(result16_low)); + const int32x2_t result32_high = vreinterpret_s32_s8(vmovn_s16(result16_high)); + + vst1q_s32((int32_t *)&dst[x], vcombine_s32(result32_low, result32_high)); + } + } + for (; x> 8; + int one_minus_const_alpha = 255 - const_alpha; + int16x8_t const_alpha16 = vdupq_n_s16(const_alpha); + int16x8_t one_minus_const_alpha16 = vdupq_n_s16(255 - const_alpha); + for (int y = 0; y < h; ++y) { + int x = 0; + for (; x < w-3; x += 4) { + int32x4_t src32 = vld1q_s32((int32_t *)&src[x]); + int32x4_t dst32 = vld1q_s32((int32_t *)&dst[x]); + + const uint8x16_t src8 = vreinterpretq_u8_s32(src32); + const uint8x16_t dst8 = vreinterpretq_u8_s32(dst32); + + const uint8x8_t src8_low = vget_low_u8(src8); + const uint8x8_t dst8_low = vget_low_u8(dst8); + + const uint8x8_t src8_high = vget_high_u8(src8); + const uint8x8_t dst8_high = vget_high_u8(dst8); + + const int16x8_t src16_low = vreinterpretq_s16_u16(vmovl_u8(src8_low)); + const int16x8_t dst16_low = vreinterpretq_s16_u16(vmovl_u8(dst8_low)); + + const int16x8_t src16_high = vreinterpretq_s16_u16(vmovl_u8(src8_high)); + const int16x8_t dst16_high = vreinterpretq_s16_u16(vmovl_u8(dst8_high)); + + const int16x8_t result16_low = qvinterpolate_pixel_255(src16_low, const_alpha16, dst16_low, one_minus_const_alpha16, half); + const int16x8_t result16_high = qvinterpolate_pixel_255(src16_high, const_alpha16, dst16_high, one_minus_const_alpha16, half); + + const int32x2_t result32_low = vreinterpret_s32_s8(vmovn_s16(result16_low)); + const int32x2_t result32_high = vreinterpret_s32_s8(vmovn_s16(result16_high)); + + vst1q_s32((int32_t *)&dst[x], vcombine_s32(result32_low, result32_high)); + } + for (; x + +QT_BEGIN_NAMESPACE + +#ifdef QT_HAVE_NEON + +void qt_blend_argb32_on_argb32_neon(uchar *destPixels, int dbpl, + const uchar *srcPixels, int sbpl, + int w, int h, + int const_alpha); + +void qt_blend_rgb32_on_rgb32_neon(uchar *destPixels, int dbpl, + const uchar *srcPixels, int sbpl, + int w, int h, + int const_alpha); + +#endif // QT_HAVE_NEON + +QT_END_NAMESPACE + +#endif // QDRAWHELPER_NEON_P_H -- cgit v0.12