halide / cvpr2015 Goto Github PK

Hello friend, I'm trying to compile a project through Android Studio. But when I run the ndk build command, an exception occurs:

8 warnings generated. [arm64-v8a] SharedLibrary : libFerrugemLib.so C:/Android/Sdk/ndk-bundle/build//../toolchains/aarch64-linux-android-4.9/prebuilt/windows-x86_64/lib/gcc/aarch64-linux-android/4.9.x/../../../../aarch64-linux-android/bin\ld: cannot find -lOpenCL clang++.exe: error: linker command failed with exit code 1 (use -v to see invocation) make: *** [obj/local/arm64-v8a/libFerrugemLib.so] Error 1 make: write error

Could you help me compile the command-lOpenCL

Hello,

I was trying to play with the brighten function and test to see how it performs on my machine. I followed the instructions (downloaded the binary release, untar it and compiled with the provided instruction. I have tested it on a 1920x1080 image by commenting and uncommenting the line brighter.vectorize(x, 8).parallel(y);. The results were similar, sometimes the version with no vectorization and parallelization was actually faster. Also, the first iteration took 550ms, while all the other ones took 0-1ms.

The idea was also to compare it with the PIL Image Enhancement Brightness function to see if I can get better time. So far, the PIL function is by one order of magnitude faster, which is quite surprising to me.

I guess there must be something that I do wrong. I am quite new to this, but eager to learn. I was wondering, would I get better results if I would build Halide from source with LLVM instead of using the binary release? Then again, I suppose the computation is anyway just a simple addition and probably LLVM already does the vectorization quite well, hence that would explain the no difference between the vectorization version and the one without it.

One more thing: By compiling the .cpp file like in the instruction, does it mean that it is actually JIT compiled? Should it be wrapped inside a Generator, so that it would be only AOT compiled?

Thanks!

Update 1 After compiling the brighten function as in Lesson 10, there was a significant speed-up, by a factor of 2. Also, if I now comment the vectorization and parallelization, it is slower. I have also increased the image size (5160x2880).

However, the loading and saving of the image is slower in Halide then in PIL. I'll try to look into that as well.

FYI, I think there is a mistake in your code:

This can be negative,
Shouldn't you add 2 * PI to the value atan result when it's negative and do a modulo pi?

Hello there,

I am trying to customize the way the auto focus work on my camera app and I found your auto focus manager (AutoFocusStateMachine). I have read everything and understand what every function does but I dont see how can I implement that with the Camera2BasicFragment ...

What I meant is the functions describe behaviors like "Once active AF scanning starts AutoFocusStateListener#onAutoFocusScan will be invoked." but I dont know how the CameraDevice is binded to all these functions. I am looking for the bridge between Camera2BasicFragment and AutoFocusStateMachine

Thanks again!

if line 76 change from
cv::Mat input_image = cv::Mat::zeros(in.width(), in.height(), CV_32FC3);
to
cv::Mat input_image = cv::Mat::zeros(in.width(), in.height(), CV_8UC3);

the actual performance comparison result is Opencv's implementation is 2 times faster than Halide, even only one thread is used in the Opencv while Halide use all the threads to compute the blur.