I’m new to threads (don’t kill me for my implementation below 🙂 and I need to do multiple blurring passes of pixels on a separate thread (see below). It’s not the most efficient implementation of box blur (it’s from Gaussian Filter without using ConvolveOp) but the performance spikes don’t occur on the Nexus 7 tablet but they do occur on the Nexus 4 phone.
I’ve posted my testing sample (running on Android 4.2 – see below).
I don’t think it is caused by the GC thrashing the memory (it doesn’t coincide with the spikes).
I think it might be something to do with cache locality or hardware memory thrashing – but I’m not sure.
What would cause the spikes? Sometimes they are sudden onset – e.g. spike of 50%. Sometimes they are slow onset – e.g. spikes increasing/decreasing monotonically, with spikes as follows -> 5%, 10%, 20%, 10%, 5%.
How could I stop them from occurring when doing heavy array processing?
This doesn’t occur on the Nexus 7 tablet which I have also tested (see results below)
Side question: What is the best way to sleep and restart my thread correctly (new to threads)?
MainActivity.java
package com.example.test;
import android.os.Bundle;
import android.app.Activity;
public class MainActivity extends Activity {
private MainThread thread;
@Override
protected void onCreate(Bundle savedInstanceState) {
super.onCreate(savedInstanceState);
thread = new MainThread();
thread.setRunning(true);
thread.start();
setContentView(R.layout.activity_main);
}
@Override
protected void onResume() {
super.onResume();
thread.setRunning(true);
}
@Override
protected void onPause() {
super.onPause();
thread.setRunning(false);
}
}
MainThread.java
package com.example.test;
import android.util.Log;
public class MainThread extends Thread {
int[] pixels;
int kernel_rows = 2;
int kernel_cols = 2;
int width = 512;
int height = 512;
@Override
public void run() {
while (running) {
long start = System.currentTimeMillis();
for (int row = kernel_rows / 2; row < height - kernel_rows / 2; row++) {
for (int col = kernel_cols / 2; col < width - kernel_cols / 2; col++) {
float pixel = 0;
// iterate over each pixel in the kernel
for (int row_offset = 0; row_offset < kernel_rows; row_offset++) {
for (int col_offset = 0; col_offset < kernel_cols; col_offset++) {
// subtract by half the kernel size to center the
// kernel
// on the pixel in question
final int row_index = row + row_offset
- kernel_rows / 2;
final int col_index = col + col_offset
- kernel_cols / 2;
pixel += pixels[row_index * width + col_index] * 1.0f / 4.0f;
}
}
pixels[row * width + col] = (int) pixel;
}
}
long stop = System.currentTimeMillis();
long delta = stop - start;
Log.d("DELTA", Long.toString(delta));
}
}
private boolean running;
public void setRunning(boolean running) {
this.pixels = new int[512 * 512];
this.running = running;
}
}
Logs
Nexus 4 phone (ms):
01-13 10:56:05.663: D/DELTA(13507): 76
01-13 10:56:05.773: D/DELTA(13507): 107
01-13 10:56:05.843: D/DELTA(13507): 77
01-13 10:56:05.923: D/DELTA(13507): 75
01-13 10:56:06.053: D/DELTA(13507): 127
01-13 10:56:06.133: D/DELTA(13507): 78
01-13 10:56:06.213: D/DELTA(13507): 81
01-13 10:56:06.293: D/DELTA(13507): 80
01-13 10:56:06.353: D/DELTA(13507): 77
01-13 10:56:06.433: D/DELTA(13507): 79
01-13 10:56:06.513: D/DELTA(13507): 79
01-13 10:56:06.624: D/DELTA(13507): 106
01-13 10:56:06.694: D/DELTA(13507): 76
Nexus 7 tablet (ms):
01-13 11:01:03.283: D/DELTA(3909): 84
01-13 11:01:03.373: D/DELTA(3909): 85
01-13 11:01:03.453: D/DELTA(3909): 85
01-13 11:01:03.543: D/DELTA(3909): 84
01-13 11:01:03.623: D/DELTA(3909): 85
01-13 11:01:03.703: D/DELTA(3909): 84
01-13 11:01:03.793: D/DELTA(3909): 85
01-13 11:01:03.873: D/DELTA(3909): 84
01-13 11:01:03.963: D/DELTA(3909): 85
01-13 11:01:04.043: D/DELTA(3909): 84
I think I may have somewhat mitigated this effect on the Nexus 4. There is still some variability in calculation consistency but it is bearable – I think – can’t see too many huge spikes – outside of thread startup/shutdown. I’ve done this using the Android NDK and c p_threads to spawn a native thread that is mostly left alone by Java (or so I’m told) until the foreground application is changed or closed.
Here is the code:
MainActivity.java
native.c
(which should be put into a jni folder at the root of the Android project)
Android.mk
(which should be put into a jni folder at the root of the Android project)
Summary
Reduced cost of code by ~20-30% and decreased variability by an order of magnitude.
The code is compiled by executing the
ndk-buildcommand from the NDK library provided by Android at the root folder (found here: http://developer.android.com/tools/sdk/ndk/index.html).Results
Nexus 4 (ms):
Nexus 7 (ms):