I'm adding a black (0) padding around Region of interest (center) of NV21 frame got from Android CameraPreview callbacks in a thread.
To avoid overhead of conversion to RGB/Bitmap and reverse, I'm trying to manipulate NV21 byte array directly but this involves nested loops which is also making preview/processing slow.
This is my run() method sending frames to detector after calling method blackNonROI.
public void run() {
Frame outputFrame;
ByteBuffer data;
while (true) {
synchronized (mLock) {
while (mActive && (mPendingFrameData == null))
try{ mLock.wait(); }catch(InterruptedException e){ return; }
if (!mActive) { return; }
// Region of Interest
mPendingFrameData = blackNonROI(mPendingFrameData.array(),mPreviewSize.getWidth(),mPreviewSize.getHeight(),300,300);
outputFrame = new Frame.Builder().setImageData(mPendingFrameData, mPreviewSize.getWidth(),mPreviewSize.getHeight(), ImageFormat.NV21).setId(mPendingFrameId).setTimestampMillis(mPendingTimeMillis).setRotation(mRotation).build();
data = mPendingFrameData;
mPendingFrameData = null;
}
try {
mDetector.receiveFrame(outputFrame);
} catch (Throwable t) {
} finally {
mCamera.addCallbackBuffer(data.array());
}
}
}
Following is the method blackNonROI
private ByteBuffer blackNonROI(byte[] yuvData, int width, int height, int roiWidth, int roiHeight){
int hozMargin = (width - roiWidth) / 2;
int verMargin = (height - roiHeight) / 2;
// top/bottom of center
for(int x=0; x<width; x++){
for(int y=0; y<verMargin; y++)
yuvData[y * width + x] = 0;
for(int y=height-verMargin; y<height; y++)
yuvData[y * width + x] = 0;
}
// left/right of center
for(int y=verMargin; y<height-verMargin; y++){
for (int x = 0; x < hozMargin; x++)
yuvData[y * width + x] = 0;
for (int x = width-hozMargin; x < width; x++)
yuvData[y * width + x] = 0;
}
return ByteBuffer.wrap(yuvData);
}
Example output frame
Note that I'm not cropping the image, just padding black pixels around specified center of image to maintain coordinated for further activities. This works like it should but it's not fast enough and causing lag in preview and frames processing.
- Can I further improve byte array update?
- Is time/place for calling blackNonROI fine?
- Any other way / lib for doing it more efficiently?
- My simple pixel iteration is so slow, how YUV/Bitmap libraries do complex things so fast? do they use GPU?
Edit:
I've replaced both for loops with following code, and it's pretty much fast now (Please refer to greeble31's answer for details):
// full top padding
from = 0;
to = (verMargin-1)*width + width;
Arrays.fill(yuvData,from,to,(byte)1);
// full bottom padding
from = (height-verMargin)*width;
to = (height-1)*width + width;
Arrays.fill(yuvData,from,to,(byte)1);
for(int y=verMargin; y<height-verMargin; y++) {
// left-middle padding
from = y*width;
to = y*width + hozMargin;
Arrays.fill(yuvData,from,to,(byte)1);
// right-middle padding
from = y*width + width-hozMargin;
to = y*width + width;
Arrays.fill(yuvData,from,to,(byte)1);
}
1. Yes. To understand why, let's take a look at the bytecode Android Studio produces for your "left/right of center" nested loop:
(Annotated excerpt from a release build of blackNonROI, AS 3.2.1):
:goto_27
sub-int v2, p2, p4 ;for(int y=verMargin; y<height-verMargin; y++)
if-ge v1, v2, :cond_45
const/4 v2, 0x0
:goto_2c
if-ge v2, p3, :cond_36 ;for (int x = 0; x < hozMargin; x++)
mul-int v3, v1, p1
add-int/2addr v3, v2
.line 759
aput-byte v0, p0, v3
add-int/lit8 v2, v2, 0x1
goto :goto_2c
:cond_36
sub-int v2, p1, p3
:goto_38
if-ge v2, p1, :cond_42 ;for (int x = width-hozMargin; x < width; x++)
mul-int v3, v1, p1
add-int/2addr v3, v2
.line 761
aput-byte v0, p0, v3
add-int/lit8 v2, v2, 0x1
goto :goto_38
:cond_42
add-int/lit8 v1, v1, 0x1
goto :goto_27
.line 764
:cond_45 ;all done with the for loops!
Without bothering to decipher this whole thing line-by-line, it is clear that each of your small, inner loops is performing:
- 1 comparison
- 1 integer multiplication
- 1 addition
- 1 store
- 1 goto
That's a lot, when you consider that all that you really need this inner loop to do is set a certain number of successive array elements to 0.
Moreover, some of these bytecodes require multiple machine instructions to implement, so I wouldn't be surprised if you're looking at over 20 cycles, just to do a single iteration of one of the inner loops. (I haven't tested what this code looks like once it's compiled by the Dalvik VM, but I sincerely doubt it is smart enough to optimize the multiplications out of these loops.)
POSSIBLE FIXES
You could improve performance by eliminating some redundant calculations. For example, each inner loop is recalculating y * width each time. Instead, you could pre-calculate that offset, store it in a local variable (in the outer loop), and use that when calculating the indices.
When performance is absolutely critical, I will sometimes do this sort of buffer manipulation in native code. If you can be reasonably certain that mPendingFrameData is a DirectByteBuffer, this is an even more attractive option. The disadvantages are 1.) higher complexity, and 2.) less of a "safety net" if something goes wrong/crashes.
MOST APPROPRIATE FIX
In your case, the most appropriate solution is probably just to use Arrays.fill(), which is more likely to be implemented in an optimized way.
Note that the top and bottom blocks are big, contiguous chunks of memory, and can be handled by one Arrays.fill() each:
Arrays.fill(yuvData, 0, verMargin * width, 0); //top
Arrays.fill(yuvData, width * height - verMargin * width, width * height, 0); //bottom
And then the sides could be handled something like this:
for(int y=verMargin; y<height-verMargin; y++){
int offset = y * width;
Arrays.fill(yuvData, offset, offset + hozMargin, 0); //left
Arrays.fill(yuvData, offset + width, offset + width - hozMargin, 0); //right
}
There are more opportunities for optimization, here, but we're already at the point of diminishing returns. For example, since the end of each row of is adjacent to the start of the next one (in memory), you could actually combine two smaller fill() calls into a larger one that covers both the right side of row N and the left side of row N + 1. And so forth.
2. Not sure. If your preview is displaying without any corruption/tearing, then it's probably a safe place to call the function from (from a thread safety standpoint), and is therefor probably as good a place as any.
3 and 4. There could be libraries for doing this task; I don't know of any offhand, for Java-based NV21 frames. You'd have to do some format conversions, and I don't think it's be worth it. Using a GPU to do this work is excessive over-optimization, in my opinion, but it may be appropriate for some specialized applications. I'd consider going to JNI (native code) before I'd ever consider using the GPU.
I think your choice to do the manipulation directly to the NV21, instead of converting to a bitmap, is a good one (considering your needs and the fact that the task is simple enough to avoid needing a graphics library).
Obviously, the most efficient way to pass image for detection would be to pass the ROI rectangle to detector. All our image processing functions accept bounding box as a parameter.
If the black margin is used for display, consider using a black overlay mask for preview layout instead of pixel manipulation.
If pixel manipulation is inevitable, check if you can limit it to Y OK, you already do this!
If your detector works on a downscaled image (as my face recognition engine does), it may be wise to apply black out to a resized frame.
At any rate, keep your loops clean and tidy, remove all recurring calculations. Using Arrays.fill() operations may help significantly, but not dramatically.
