libjpeg-turbo 加速 JPG 有损压缩与解压缩

非典型程序猿
267 阅读9分钟

libjpeg-turbo 是一个 JPEG 图像编解码器，它使用 SIMD 指令加速 x86、x86-64、Arm、PowerPC 和 MIPS 系统上的基本 JPEG 压缩和解压缩，以及 x86、x86-64 和 Arm 系统上的渐进式 JPEG 压缩。在这些系统上其他条件相同的情况下，libjpeg-turbo 通常比 libjpeg 快 2-6 倍。在其他类型的系统上，由于其高度优化的 Huffman 编码算法，libjpeg-turbo 仍能比 libjpeg 有显著的性能提升。在许多情况下，libjpeg-turbo 的性能可与专有的高速 JPEG 编解码器相媲美。

libjpeg-turbo 实现了传统的 libjpeg API 以及功能较弱但更直观的 TurboJPEG API。libjpeg-turbo 最初基于 libjpeg/SIMD，这是 Miyasaka Masaru 开发的一款基于 MMX 加速的 libjpeg v6b 衍生版本。TigerVNC 和 VirtualGL 项目在 2009 年对编解码器进行了多项改进，2010 年初，libjpeg-turbo 成为一个独立的项目，旨在将高速 JPEG 压缩/解压缩技术提供给更广泛的用户和开发人员。libjpeg-turbo 是 JPEG 标准的 ISO/IEC 和 ITU-T 参考实现。

一、编译

项目 Github 地址：github.com/libjpeg-tur…

当前的 release 最新版是 3.0.4，我们基于此版本进行编译。

libjpeg-turbo/BUILDING.md 给出了各类系统的编译方法，比如我们接下来以安卓系统为例编译 armv7&arm64 这两种 so 文件。

Building libjpeg-turbo for Android

Building libjpeg-turbo for Android platforms requires v13b or later of the Android NDK.

Armv7 (32-bit)

NDK r19 or later with Clang recommended

The following is a general recipe script that can be modified for your specific needs.

# Set these variables to suit your needs
NDK_PATH={full path to the NDK directory-- for example,
  /opt/android/android-ndk-r16b}
TOOLCHAIN={"gcc" or "clang"-- "gcc" must be used with NDK r16b and earlier,
  and "clang" must be used with NDK r17c and later}
ANDROID_VERSION={the minimum version of Android to support-- for example,
  "16", "19", etc.}

cd {build_directory}
cmake -G"Unix Makefiles" \
  -DANDROID_ABI=armeabi-v7a \
  -DANDROID_ARM_MODE=arm \
  -DANDROID_PLATFORM=android-${ANDROID_VERSION} \
  -DANDROID_TOOLCHAIN=${TOOLCHAIN} \
  -DCMAKE_ASM_FLAGS="--target=arm-linux-androideabi${ANDROID_VERSION}" \
  -DCMAKE_TOOLCHAIN_FILE=${NDK_PATH}/build/cmake/android.toolchain.cmake \
  [additional CMake flags] {source_directory}
make

Armv8 (64-bit)

Clang recommended

The following is a general recipe script that can be modified for your specific needs.

# Set these variables to suit your needs
NDK_PATH={full path to the NDK directory-- for example,
  /opt/android/android-ndk-r16b}
TOOLCHAIN={"gcc" or "clang"-- "gcc" must be used with NDK r14b and earlier,
  and "clang" must be used with NDK r17c and later}
ANDROID_VERSION={the minimum version of Android to support.  "21" or later
  is required for a 64-bit build.}

cd {build_directory}
cmake -G"Unix Makefiles" \
  -DANDROID_ABI=arm64-v8a \
  -DANDROID_ARM_MODE=arm \
  -DANDROID_PLATFORM=android-${ANDROID_VERSION} \
  -DANDROID_TOOLCHAIN=${TOOLCHAIN} \
  -DCMAKE_ASM_FLAGS="--target=aarch64-linux-android${ANDROID_VERSION}" \
  -DCMAKE_TOOLCHAIN_FILE=${NDK_PATH}/build/cmake/android.toolchain.cmake \
  [additional CMake flags] {source_directory}
make

参考官方给出的编译说明，我们写一个 build_libjpeg_trubo.sh 在 ubuntu 18.04 上编译，重点需要配置 cmake 编译链，由于配置了 CMAKE_INSTALL_PREFIX 为 out 目录，因此编译生成文件位于 out 目录下。

4a8c2274b4ed4ec69c91e888330ca100.png

build_libjpeg_trubo.sh

NDK_PATH=/home/snake/Android/android-ndk-r21e
TOOLCHAIN=clang
ANDROID_VERSION=19

cd libjpeg-turbo-3.0.4

mkdir build
cd build

mkdir armeabi-v7a
cd armeabi-v7a

echo "Build armeabi-v7a arch..."

cmake -G"Unix Makefiles" \
  -DANDROID_ABI=armeabi-v7a \
  -DANDROID_ARM_MODE=arm \
  -DCMAKE_INSTALL_PREFIX=../../../out/armeabi-v7a \
  -DENABLE_STATIC=TRUE \
  -DANDROID_PLATFORM=android-${ANDROID_VERSION} \
  -DANDROID_TOOLCHAIN=${TOOLCHAIN} \
  -DCMAKE_ASM_FLAGS="--target=arm-linux-androideabi${ANDROID_VERSION}" \
  -DCMAKE_TOOLCHAIN_FILE=${NDK_PATH}/build/cmake/android.toolchain.cmake \
  ../../

make -j4
make install

cd ../
mkdir arm64-v8a
cd arm64-v8a

echo "Build arm64-v8a arch..."

cmake -G"Unix Makefiles" \
  -DANDROID_ABI=arm64-v8a \
  -DANDROID_ARM_MODE=arm \
  -DCMAKE_INSTALL_PREFIX=../../../out/arm64-v8a \
  -DENABLE_STATIC=TRUE \
  -DANDROID_PLATFORM=android-${ANDROID_VERSION} \
  -DANDROID_TOOLCHAIN=${TOOLCHAIN} \
  -DCMAKE_ASM_FLAGS="--target=aarch64-linux-android${ANDROID_VERSION}" \
  -DCMAKE_TOOLCHAIN_FILE=${NDK_PATH}/build/cmake/android.toolchain.cmake \
  ../../

make -j4
make install

二、使用

以下代码使用了更直观的 TurboJPEG API，网上这块资料非常少，另外官方 demo 也比较精简，经过摸索 I420 格式的 YUV 是可以正常编码为 JPEG 的，同时将 JPG 转化为了 RGB 数组。

JpegHandler.h

#ifndef JPEGHANDLER_H
#define JPEGHANDLER_H

#include <turbojpeg.h>

static const int HANDLE_JPEG_SUCC = 0;
static const int HANDLE_JPEG_FAIL = -1;

static const int COMPRESS_JPEG_ERR_NONSUPPORT_YUV_TYPE = -100;
static const int COMPRESS_JPEG_ERR_YUV_BUF_SIZE = -101;
static const int COMPRESS_JPEG_ERR_COMPRESS = -101;

static const int DECOMPRESS_JPEG_ERR_HEADER = -200;
static const int DECOMPRESS_JPEG_ERR_DECOMPRESS = -201;

enum YUV_T {
    I420 = 1,
    NV12 = 2,
    NV21 = 3,
};

class JpegHandler {

public:
    JpegHandler();

    ~JpegHandler();

    int compressYUV2Jpeg(const unsigned char *srcYuv, int srcYuvSize, int width, int height,
                         YUV_T yuvT, unsigned char **jpegBuf, unsigned long *jpegSize);

    int decompressJpeg2RGB(const unsigned char *jpegBuf, size_t jpegSize,
                           unsigned char *dstBuf, int *dstWidth, int *dstHeight);

private:
    tjhandle mCompressTjInstance;
    tjhandle mDecompressTjInstance;
};


#endif //JPEGHANDLER_H

以下是 cpp 实现文件，实现了头文件中定义的 compressYUV2Jpeg 压缩 YUV 为 JPEG 函数，和 decompressJpeg2RGB 解压 JPEG 为 RGB 的函数。同时构造函数中调用 tj3Init 初始化了 tjhandle 类型的压缩和解压缩实例，析构函数则调用销毁函数 tj3Destroy 进行“善后”处理。

JpegHandler.cpp

#include <cstring>
#include "JpegHandler.h"
#include "../logger.h"

static const int YUV_PLANES = 3;

int
JpegHandler::compressYUV2Jpeg(const unsigned char *srcYuv, int srcYuvSize, int width, int height,
                              YUV_T yuvT, unsigned char **jpegBuf, unsigned long *jpegSize) {
    int subSamp = TJSAMP_420;
    switch (yuvT) {
        case I420:
        case NV12:
        case NV21:
            subSamp = TJSAMP_420;
            break;
        default:
            return COMPRESS_JPEG_ERR_NONSUPPORT_YUV_TYPE;
    }

    int align = 1;
    int needSize = tj3YUVBufSize(width, align, height, subSamp);
    if (needSize != srcYuvSize) {
        LOGE("we detect yuv size: %d, but you give: %d, check again.", needSize, srcYuvSize);
        return COMPRESS_JPEG_ERR_YUV_BUF_SIZE;
    }

    int jpegQuality = 90;
    if (tj3Set(mCompressTjInstance, TJPARAM_QUALITY, jpegQuality) == HANDLE_JPEG_FAIL) {
        LOGE("set TJPARAM_QUALITY err");
    }

    if (tj3Set(mCompressTjInstance, TJPARAM_SUBSAMP, subSamp) == HANDLE_JPEG_FAIL) {
        LOGE("set TJPARAM_SUBSAMP err");
    }

    int yLen = width * height;
    int uLen = yLen / 4;
    const unsigned char *const srcPlanes[YUV_PLANES] = {srcYuv,
                                                        srcYuv + yLen,
                                                        srcYuv + yLen + uLen};
    int halfWidth = width / 2;
    const int strides[YUV_PLANES] = {width, halfWidth, halfWidth};

    if (tj3CompressFromYUVPlanes8(mCompressTjInstance,
                                  reinterpret_cast<const unsigned char *const *>(&srcPlanes),
                                  width,
                                  reinterpret_cast<const int *>(&strides),
                                  height,
                                  jpegBuf,
                                  reinterpret_cast<size_t *>(jpegSize)) == HANDLE_JPEG_FAIL) {
        int errCode = tj3GetErrorCode(mCompressTjInstance);
        char *errStr = tj3GetErrorStr(mCompressTjInstance);
        LOGE("Failed to compress YUV. %s(errCode=%d)", errStr, errCode);
        return COMPRESS_JPEG_ERR_COMPRESS;
    }

    return HANDLE_JPEG_SUCC;
}

JpegHandler::JpegHandler() {
    if ((mCompressTjInstance = tj3Init(TJINIT_COMPRESS)) == nullptr) {
        LOGE("Failed to create compress turbojpeg instance.");
    }

    if ((mDecompressTjInstance = tj3Init(TJINIT_DECOMPRESS)) == nullptr) {
        LOGE("Failed to create decompress turbojpeg instance.");
    }
}

JpegHandler::~JpegHandler() {
    if (!mCompressTjInstance) {
        tj3Destroy(mCompressTjInstance);
        mCompressTjInstance = nullptr;
    }

    if (!mDecompressTjInstance) {
        tj3Destroy(mDecompressTjInstance);
        mDecompressTjInstance = nullptr;
    }
}

int JpegHandler::decompressJpeg2RGB(const unsigned char *jpegBuf, size_t jpegSize,
                                    unsigned char *dstBuf, int *dstWidth, int *dstHeight) {

    if (tj3DecompressHeader(mDecompressTjInstance, jpegBuf, jpegSize) == HANDLE_JPEG_FAIL) {
        LOGE("decompress JPEG Header err.");
        return DECOMPRESS_JPEG_ERR_HEADER;
    }

    int jpegWidth = tj3Get(mDecompressTjInstance, TJPARAM_JPEGWIDTH);
    if (jpegWidth == HANDLE_JPEG_FAIL) {
        LOGE("get TJPARAM_JPEGWIDTH err");
    }

    *dstWidth = jpegWidth;

    int jpegHeight = tj3Get(mDecompressTjInstance, TJPARAM_JPEGHEIGHT);
    if (jpegHeight == HANDLE_JPEG_FAIL) {
        LOGE("get TJPARAM_JPEGHEIGHT err");
    }

    *dstHeight = jpegHeight;

    if (tj3Decompress8(mDecompressTjInstance, jpegBuf, jpegSize,
                       dstBuf, 0, TJCS_RGB) ==
        HANDLE_JPEG_FAIL) {
        int errCode = tj3GetErrorCode(mDecompressTjInstance);
        char *errStr = tj3GetErrorStr(mDecompressTjInstance);
        LOGE("Failed to decompress JPEG. %s(errCode=%d)", errStr, errCode);
        return DECOMPRESS_JPEG_ERR_DECOMPRESS;
    }

    return HANDLE_JPEG_SUCC;
}

其中 compressYUV2Jpeg 函数内，首先调用 tj3YUVBufSize 计算出 YUV 需要的 buffer 大小，接着调用 tj3Set 设置 jpeg Quality（程序中固定为 90），继续调用 tj3Set 设置 YUV 的采样格式，这些参数是调用压缩函数必须配置的。经过前面的两组参数配置，现在就可以调用 tj3CompressFromYUVPlanes8 真正做压缩了，压缩出现异常时可调用 tj3GetErrorCode 和 tj3GetErrorStr 获取相应的错误码和描述。

小技巧：通过错误码和错误描述，一步步得知如何使用此压缩函数正常工作，也是一种调试、摸索使用的手段。

decompressJpeg2RGB 也是类似的，首先调用 tj3DecompressHeader 获取 JPEG 必要的一些参数，比如图像的宽、高等。调用 tj3DecompressHeader 成功后就可以调用 tj3Get 获取各种 jpeg 的参数了，当然包括宽高，此处获取了宽高。最后调用 tj3Decompress8 将 JPEG 转化为了 RGB，当然入参最后一个参数配置为了 TJCS_RGB。

三、实测性能

本人使用华为 Mate60 作为测试设备，进行了几十次测试（压缩和解压基本上都是 10ms 左右可完成），图片分辨率为 1920*1080，将程序打包成 release，压缩和解压缩的速度都是非常快的， debug 版本要慢上不少。

2024-11-22 08:17:59.987 19472-19472/xxx D/MainActivity: compressHandlerJni compress begin
2024-11-22 08:17:59.996 19472-19472/xxx D/MainActivity: compressHandlerJni compress end, jpegLen=291778
2024-11-22 08:17:59.996 19472-19472/xxx D/MainActivity: compressHandlerJni decompress begin
2024-11-22 08:18:00.007 19472-19472/xxx D/MainActivity: compressHandlerJni decompress end, image width=1920, height=1080
2024-11-22 08:18:01.541 19472-19472/xxx D/MainActivity: compressHandlerJni compress begin
2024-11-22 08:18:01.554 19472-19472/xxx D/MainActivity: compressHandlerJni compress end, jpegLen=278227
2024-11-22 08:18:01.554 19472-19472/xxx D/MainActivity: compressHandlerJni decompress begin
2024-11-22 08:18:01.564 19472-19472/xxx D/MainActivity: compressHandlerJni decompress end, image width=1920, height=1080
2024-11-22 08:18:02.407 19472-19472/xxx D/MainActivity: compressHandlerJni compress begin
2024-11-22 08:18:02.419 19472-19472/xxx D/MainActivity: compressHandlerJni compress end, jpegLen=278319
2024-11-22 08:18:02.419 19472-19472/xxx D/MainActivity: compressHandlerJni decompress begin
2024-11-22 08:18:02.429 19472-19472/xxx D/MainActivity: compressHandlerJni decompress end, image width=1920, height=1080
2024-11-22 08:18:32.356 19472-19472/xxx D/MainActivity: compressHandlerJni compress begin
2024-11-22 08:18:32.370 19472-19472/xxx D/MainActivity: compressHandlerJni compress end, jpegLen=337098
2024-11-22 08:18:32.370 19472-19472/xxx D/MainActivity: compressHandlerJni decompress begin
2024-11-22 08:18:32.381 19472-19472/xxx D/MainActivity: compressHandlerJni decompress end, image width=1920, height=1080
2024-11-22 08:18:34.704 19472-19472/xxx D/MainActivity: compressHandlerJni compress begin
2024-11-22 08:18:34.717 19472-19472/xxx D/MainActivity: compressHandlerJni compress end, jpegLen=338297
2024-11-22 08:18:34.717 19472-19472/xxx D/MainActivity: compressHandlerJni decompress begin
2024-11-22 08:18:34.727 19472-19472/xxx D/MainActivity: compressHandlerJni decompress end, image width=1920, height=1080
avatar
文章
阅读
粉丝