Android Bitmap加载内存占用彻底分析
Posted
tags:
篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了Android Bitmap加载内存占用彻底分析相关的知识,希望对你有一定的参考价值。
背景
在某个版本应用上线后,偶然测得首页占用的内存非常的大而且一直不能回收掉,经过一轮的排查后最终确定是3张图片引起的!当时每张图片占用了将近20m内存。当时紧急处理好后还一直惦记着此事,后来对android加载Bitmap的内存占用作了彻底的分析,跟踪了相关的源码,在这里总结一下。
图片加载测试
先抛开结论,现在先直观的看一下加载如下一张图片需要多少内存
这里写图片描述
其中图片的宽高都为300像素
计算内存的方法采用 android.graphics.Bitmap#getByteCount
public final int getByteCount() {
// int result permits bitmaps up to 46,340 x 46,340
return getRowBytes() * getHeight();
}
1
2
3
4
预期占用的内存大小为
图片宽*图片高*表示每个像素点的字节数,即
1
这里写图片描述
加载SD卡的图片
加载SD中的图片结果为
这里写图片描述
assets的图片
加载asset目录中的图片结果为
这里写图片描述
加载Resources的图片
drawable目录
这里写图片描述
drawable-mdpi目录
这里写图片描述
drawable-hdpi目录
这里写图片描述
drawable-xhdpi目录
这里写图片描述
drawable-xhhdpi目录
这里写图片描述
drawable-xhhhdpi目录
这里写图片描述
内存占用分析
理论上,300 * 300像素的图片,默认以4byte表示1个像素的情况下,占用的内存为
300 * 300 * 4 = 360000 byte
但是,实际上,只有从SD卡、assets目录、drawable-xhdpi目录下加载图片才等于理论数值,其他数值都不等!
等等!,从图片的大小看,不等于理论值的图片好像被放大或者缩小了?我们可以验证一下,把图片在内存中的实际宽高打印出来
SD卡的
这里写图片描述
drawable-mdpi的
这里写图片描述
发现没有?在drawable-mdpi目录中的图片在加载内存中时的宽高都放大了两倍!!
其实,加载在SD卡和assets目录的图片时,图片的尺寸不会被改变,但是drawable-xxxdpi目录的照片的尺寸会被改变,这里篇幅所限,就不一一截图了,想验证的可以下载demo(文末给出链接)试验一下。至于尺寸改变的原因,下文会讨论,这里卖个关子。
查看源码
正所谓源码面前,了无秘密,欲知原理,还须从源码下手,首先查看BitmapFactory.java文件
BitmapFactory.decodeFile
BitmapFactory.decodeResourceStream
1
2
这两个方法的重载函数最终都会调用到
private static native Bitmap nativeDecode www.yunduanpingtai.cn Stream(InputStream is, byte[] storage,
Rect padding, Options opts);
1
2
这是一个本地方法,其相关实现在
frameworks/base/core/jni/android/graphics/BitmapFactory.cpp
打开文件,找到如下的方法,就是本地方法的实现
static jobject nativeDecodeStream(JNIEnv* env, jobject clazz, jobject is, jbyteArray storage,
jobject padding, jobject options) {
jobject bitmap = NULL;
SkAutoTUnref<SkStream> stream(CreateJavaInputStreamAdaptor(env, is, storage));
if (stream.get()) {
SkAutoTUnref<SkStreamRewindable> bufferedStream(
SkFrontBufferedStream::Create(stream, www.rbuluoyl.cn/BYTES_TO_BUFFER));
SkASSERT(bufferedStream.get() != NULL);
bitmap = doDecode(env, bufferedStream, padding, options);
}
return bitmap;
抓住我们要看的部分,这里还调用了doDecode方法,调到doDecode会发现,bitmap解码的逻辑基本框架都在里面了,分析清楚它的逻辑,我们就能找到答案,方法非常长,有200多行,我把枝干提取出来,并加上注释如下
static jobject doDecode(JNIEnv* env, SkStreamRewindable* stream, jobject padding, jobject options) {
int sampleSize = 1;
SkImageDecoder::Mode decodeMode = SkImageDecoder::kDecodePixels_Mode;
SkColorType prefColorType = kN32_SkColorType;
bool doDither = true;
bool isMutable = false;
float scale = 1.0f;
bool preferQualityOverSpeed = false;
bool requireUnpremultiplied = false;
jobject javaBitmap = NULL;
if (options != NULL) {
//options是BitmapFactory.Options的java对象,这里获取该对象的成员变量值并赋值给本地代码的变量,下面类似格式的方法调用作用相同
sampleSize = env->GetIntField(options, gOptions_sampleSizeFieldID);
if (optionsJustBounds(env, options)) {
decodeMode = SkImageDecoder::kDecodeBounds_Mode;
}
// initialize these, in case we fail later on
env->SetIntField(options, gOptions_widthFieldID, -1);
env->SetIntField(options, gOptions_heightFieldID, -1);
env->SetObjectField(options, gOptions_mimeFieldID, 0);
jobject jconfig = env->GetObjectField(options, gOptions_configFieldID);
prefColorType = GraphicsJNI::getNativeBitmapColorType(env, jconfig);
isMutable = env->GetBooleanField(options, gOptions_mutableFieldID);
doDither = env->GetBooleanField(options, gOptions_ditherFieldID);
preferQualityOverSpeed = env->GetBooleanField(options,
gOptions_preferQualityOverSpeedFieldID);
requireUnpremultiplied = !env->GetBooleanField(options, gOptions_premultipliedFieldID);
javaBitmap = env->GetObjectField(www.zzktv.cn options, gOptions_bitmapFieldID);
//java里,inScaled默认true,所以这里总是执行,除非手动设置为false
if (env->GetBooleanField(options, gOptions_scaledFieldID)) {
const int density = env->GetIntField(options, gOptions_densityFieldID);
const int targetDensity = env->GetIntField(options, gOptions_targetDensityFieldID);
const int screenDensity = env->GetIntField(options, gOptions_screenDensityFieldID);
//重点就是这里了,density、targetDensity、screenDensity的值决定了是否缩放、以及缩放的倍数
if (density != 0 && targetDensity != 0 && density != screenDensity) {
scale = (float) targetDensity / density;
}
}
}
const bool willScale = scale != 1.0f;
...省略若干行
//真正的decode操作,decodingBitmap是解码的的结果,但如果要缩放,则返回缩放后的bitmap,看后面的代码
SkBitmap decodingBitmap;
if (decoder->decode(stream, &decodingBitmap, prefColorType, decodeMode)
!= SkImageDecoder::kSuccess) {
return nullObjectReturn("decoder->decode returned false");
}
int scaledWidth = decodingBitmap.width();
int scaledHeight = decodingBitmap.height();
if (willScale && decodeMode != SkImageDecoder::kDecodeBounds_Mode) {
scaledWidth = int(scaledWidth * scale + 0.5f);
scaledHeight = int(scaledHeight * scale + 0.5f);
}
// update options (if any)
if (options != NULL) {
jstring mimeType = getMimeTypeString(env, decoder->getFormat());
if (env->ExceptionCheck()) {
return nullObjectReturn("OOM in getMimeTypeString()");
}
env->SetIntField(options, gOptions_widthFieldID, scaledWidth);
env->SetIntField(options, gOptions_heightFieldID, scaledHeight);
env->SetObjectField(options, gOptions_mimeFieldID, mimeType);
}
// if we‘re in justBounds mode, return now (skip the java bitmap)
if (decodeMode == SkImageDecoder::kDecodeBounds_Mode) {
return NULL;
}
...省略若干行
//scale != 1.0f就缩放bitmap,缩放的步骤概扩起来就是申请缩放后的内存,然后把所有的bitmap信息记录复制到outputBitmap变量上;否则直接复制decodingBitmap的内容
if (willScale) {
// This is weird so let me explain: we could use the scale parameter
// directly, but for historical reasons this is how the corresponding
// Dalvik code has always behaved. We simply recreate the behavior here.
// The result is slightly different from simply using scale because of
// the 0.5f rounding bias applied when computing the target image size
const float sx = scaledWidth /www.bomaoyuLe.cn float(decodingBitmap.width());
const float sy = scaledHeight / float(decodingBitmap.height());
// TODO: avoid copying when scaled size equals decodingBitmap size
SkColorType colorType = colorTypeForScaledOutput(decodingBitmap.colorType());
// FIXME: If the alphaType is kUnpremul and the image has alpha, the
// colors may not be correct, since Skia does not yet support drawing
// to/from unpremultiplied bitmaps.
outputBitmap->setInfo(SkImageInfo::Make(scaledWidth, scaledHeight,
colorType, decodingBitmap.alphaType()));
if (!outputBitmap->allocPixels(outputAllocator, NULL)) {
return nullObjectReturn("allocation failed for scaled bitmap");
}
// If outputBitmap‘s pixels are newly allocated by Java, there is no need
// to erase to 0, since the pixels were initialized to 0.
if (outputAllocator != &javaAllocator) {
outputBitmap->eraseColor(0);
}
SkPaint paint;
paint.setFilterLevel(SkPaint::kLow_FilterLevel);
SkCanvas canvas(*outputBitmap);
canvas.scale(sx, sy);
canvas.drawBitmap(decodingBitmap, 0.0f, 0.0f, &paint);
} else {
outputBitmap->swap(decodingBitmap);
}
...省略若干行
//后面的部分就是返回bitmap对象给java代码了
if (javaBitmap != NULL) {
bool isPremultiplied = !requireUnpremultiplied;
GraphicsJNI::reinitBitmap(env, javaBitmap, outputBitmap, isPremultiplied);
outputBitmap->notifyPixelsChanged();
// If a java bitmap was passed in for reuse, pass it back
return javaBitmap;
}
int bitmapCreateFlags = 0x0;
if (isMutable) bitmapCreateFlags |= GraphicsJNI::kBitmapCreateFlag_Mutable;
if (!requireUnpremultiplied) bitmapCreateFlags |= GraphicsJNI::kBitmapCreateFlag_Premultiplied;
// now create the java bitmap
return GraphicsJNI::createBitmap(env, outputBitmap, javaAllocator.getStorageObj(),
bitmapCreateFlags, ninePatchChunk, ninePatchInsets, -1);
上面的解析能勾画出大概的逻辑了,其中秘密就在这一小段
//java里,inScaled默认true,所以这里总是执行,除非手动设置为false
if (env->GetBooleanField(options, gOptions_scaledFieldID)) {
const int density = env->GetIntField(options, gOptions_densityFieldID);
const int targetDensity = env->GetIntField(options, gOptions_targetDensityFieldID);
const int screenDensity = env->GetIntField(options, gOptions_screenDensityFieldID);
//重点就是这里了,density、targetDensity、screenDensity的值决定了是否缩放、以及缩放的倍数
if (density != 0 && targetDensity != 0 && density != screenDensity) {
可以看到,BitmapFactory.Options对象的inScaled、inDensity、inTargetDensity、screenDensity四个值共同决定了bitmap是否被缩放以及缩放的倍数。
下面回到java部分的代码继续分析
为什么在drawable文件夹的图片会被缩放而SD卡、assets的图片不会
现在要解决这个问题就是要看BitmapFactory.Options对象的inScaled、inDensity、inTargetDensity、screenDensity四个值是怎样被赋值了
之前提到过,inScaled默认值为true
public Options() {
inDither = false;
inScaled = true;
inPremultiplied = true;
decodeFile方法在调用本地方法前调用会decodeStream和decodeStreamInternal
public static Bitmap decodeStream(InputStream is, Rect outPadding, Options opts) {
// we don‘t throw in this case, thus allowing the caller to only check
// the cache, and not force the image to be decoded.
if (is == null) {
return null;
}
Bitmap bm = null;
Trace.traceBegin(Trace.TRACE_TAG_GRAPHICS, "decodeBitmap");
try {
if (is instanceof AssetManager.AssetInputStream) {
final long asset = ((AssetManager.AssetInputStream) is).getNativeAsset();
bm = nativeDecodeAsset(asset, outPadding, opts);
} else {
bm = decodeStreamInternal(is, outPadding, opts);
}
if (bm == null && opts != null && opts.inBitmap != null) {
throw new IllegalArgumentException("Problem decoding into existing bitmap");
}
setDensityFromOptions(bm, opts);
} finally {
Trace.traceEnd(Trace.TRACE_TAG_GRAPHICS);
}
return bm;
}
private static Bitmap decodeStreamInternal(InputStream is, Rect outPadding, Options opts) {
// ASSERT(is != null);
byte [] tempStorage = null;
if (opts != null) tempStorage = opts.inTempStorage;
if (tempStorage == null) tempStorage = new byte[DECODE_BUFFER_SIZE];
可以看到,如果opts直到调用本地方法之前也没有并没有改变,故加载SD卡的图片和assets的图片并不会被缩放(加载assets的图片对应的本地方法为nativeDecodeAsset,最后都会调用doDecode)
decodeResource方法的调用栈为 decodeResource->decodeResourceStream->decodeStream,后面就跟之前的一样了,其中decodeResourceStream方法如下
/**
* Decode a new Bitmap from an InputStream. This InputStream was obtained from
* resources, which we pass to be able to scale the bitmap accordingly.
*/
public static Bitmap decodeResourceStream(Resources res, TypedValue value,
InputStream is, Rect pad, Options opts) {
if (opts == null) {
opts = new Options();
}
if (opts.inDensity == 0 && value != null) {
final int density = value.density;
if (density == TypedValue.DENSITY_DEFAULT) {
opts.inDensity = DisplayMetrics.DENSITY_DEFAULT;
} else if (density != TypedValue.DENSITY_NONE) {
opts.inDensity = density;
}
}
if (opts.inTargetDensity == 0 && res != null) {
opts.inTargetDensity = res.getDisplayMetrics().densityDpi;
}
return decodeStream(is, pad, opts);
分别是drawable-ldpi、drawable-mdpi、drawable-hdpi、drawable-xhdpi、drawable-xxhdpi、drawable-xxxhdpi目录的dpi值,在这些目录的图片,加载的时候就会被附上对应的值。因为默认的值是DENSITY_MEDIUM,所以drawable目录和drawable-mdpi的图片缩放的大小是一样的
小结
图片被缩放的原因在于资源目录对应着dpi,当加载资源的dpi和屏幕实际的dpi不一样时,进行缩放以使资源显示效果得到优化
图片资源放置选择
前文所述,当我们的图片资源只有一张的时候,该放到哪个目录?放到assets目录似乎是最安全的,不会因图片被放大造成OOM,也不会因图片缩小失真。但是assets目录的资源用起来不方便啊!我认为,在现在屏幕密度基本为720p以上的时代,如果UI设计师只提供了一张图片,就放到xhdpi或者xxhdpi目录吧,不然放在drawable目录会被放大几倍的
以上是关于Android Bitmap加载内存占用彻底分析的主要内容,如果未能解决你的问题,请参考以下文章
ANDROID开发之OOM:一张图片(BitMap)占用内存的计算 图片内存优化
Android 一张图片(BitMap)占用内存的计算 图片内存优化