Wavelets

SCU GPU Wavelet Project

To build:

1. You'll need to have CUDA, Java, and Google protocol buffers installed. (See the top of the Makefile)
2. If you're using Linux or Cygwin, just run "make".
3. If you're using Visual Studio, open a visual studio command prompt and run "make". (this is a batch file that calls "nmake" on Makefile.nmake.

I haven't tried integrating it into the Visual Studio IDE yet.

This project uses its own ".cube" file format. It's a container for 3-dimensional chunks of data. One cube file can contain multiple "cubelets". See "cubelet_file.h" for a brief description of the file format. The goal is to split large input file into many cublets, stored them in one cube file, and then process them all in parallel, outputting to another cube file.

Older code (like "haar") used a file format with the descriptive extension ".data". It stores 2-d data in text or binary form. See "data_io.h" for the format.

Here are some of the tools in this project:

1. WaveletSampleImage.java / convert

   This converts between jpeg images and cube files or simple data files. It uses the filename extensions to infer the desired operation.

   It is necessary to set the CLASSPATH to include the code for WaveletSampleImage and protobuf, and this is done automatically by the "convert" wrapper script (which is generated by the Makefile). So the "convert" script calls the Java code in WaveletSampleImage.

   For example, to convert an image to a cube file:

   $ ./convert Images/haleakala_orig.jpg
   2048 x 2048 grayscale data written to "Images/haleakala_orig.cube" in 0.024 sec
   

   By default, to make things easier for the wavelet algorithms, this will crop the image to be square, resize it down to the next smaller power of two, and convert it to grayscale.

   To convert back to an image:

   $ ./convert Images/haleakala_orig.cube Images/haleakala_restored.jpg
   2048 x 2048 image written to "Images/haleakala_restored.jpg".
   

   WaveletSampleImage.java also supports "data" files, which was a simplified text or binary format that was a predecessor to cube files.

2. cube

   This is a utility for manipulating .cube files. For example, to list the contents of a cube file:

   $ ./cube list Images/haleakala_orig.cube
   Cubelet 0,0,0: 2048x2048x1 of UINT8, 4194304 bytes at 91
   

   This can also be used to extract cubelets from a file, generate a cubelet from raw bytes, split a large cubelet into many smaller ones (with optional overlap), or extract part of one cubelet.

3. test_compress_cpu

   Run the compression algorithm on one cubelet using the CPU.

   $ ./cube list Images/haleakala_orig.cube
   Cubelet 0,0,0: 2048x2048x1 of UINT8, 4194304 bytes at 91
   
   $ ./test_compress_cpu Images/haleakala_orig.cube foo.cube
   Read 2048x2048x1 data file: 15.40 ms
   cdf97 wavelet transform (11,11,0 steps): 161.56 ms
   min = -18873.121094, max = 421519.750000, threshold = 1.161753893: 295.86 ms
   Quantize log: 46.40 ms
   Huffman summarize frequencies 19.482 ms, build encoding 0.061 ms
   Huffman encoding: 2118372 bytes, 3.93 bits/pixel, longest encoding = 23 bits
   Write data file: 50.65 ms
   Total: 574.08 ms
   
   $ ./cube list foo.cube
   Cubelet 0,0,0: 1x2048x2048 of INT32, 2118372 compressed bytes at 1174
   

4. test_compress_gpu

   Run the compression algorithm on one cubelet using the GPU.

   $ ./test_compress_gpu Images/haleakala_orig.cube foo.cube
   GPU 0: GeForce GTX 1080
   Read 2048x2048x1 data file: 8.99 ms
   CUDA initialize: 88.48 ms
   Copy to GPU: 3.649 ms
   Wavelet transform: 7.390 ms (transpose 0.308 ms)
   Absolute value: 0.264 ms
   Sort: 2.531 ms
   min = -18873.113281, max = 421519.750000, threshold = 1.161754727
   Quantize: 0.182 ms
   Frequency count: 0.421 ms
   Copy from GPU: 6.881 ms
   Huffman summarize frequencies 18.964 ms, build encoding 0.075 ms
   Huffman encoding: 2118368 bytes, 3.93 bits/pixel, longest encoding = 23 bits
   Write data file: 79.27 ms
   Total: 199.87 ms
   

5. CubeletViewer.java / cubeview

   CubeletViewer is a visualizer to see one cubelet in a cube file. "cubeview" is a Makefile-generated wrapper script that sets the classpath for CubeletViewer.

6. haar.cu

   This is older code that only supports the ".data" 2-d format. This implements the Haar discrete wavelet transform on a 2-d data file. Give it an input data file and an output data file, and it performs the 2-d transform. Add the "-inverse" command line flag and it performs the inverse. By default it will just do a single step of the transform. Add an integer argument after the filenames to specify a different number of steps.

   It does every transform on the CPU and the GPU, and compares the timing and the results. The implementations of each are in dwt_cpu.cc and dwt_gpu.cu. If the CPU and GPU versions are very different, rather than writing the output file, it writes the output of each to "err_cpu.data" and "err_gpu.data".

   Example:

   % java WaveletSampleImage Images/hubble4096.jpg hubble.data
   4096 x 4096 grayscale data written to "hubble.data" in 1.730 sec
   
   % ./haar hubble.data hubble2.data 3
   Reading hubble.data...4096 x 4096
   GPU 0: GeForce GTX 680
   CPU: 593.847 ms
   Tile size 32x32
   Time elapsed creating GPU tasks: 0.164 ms
   Times:
     Copy data to GPU:      10.275 ms
     Transform time:         4.359 ms (6 calls)
     Copy data from GPU:    10.359 ms
   CUDA: 25.003168 ms
   Wrote hubble2.data
   
   % java WaveletSampleImage hubble2.data
   4096 x 4096 image written to "hubble2.jpg".
   
   % ./haar -inverse hubble2.data hubble3.data 3
   Reading hubble2.data...4096 x 4096
   GPU 0: GeForce GTX 680
   CPU: 601.198 ms
   Tile size 32x32
   Time elapsed creating GPU tasks: 0.150 ms
   Times:
     Copy data to GPU:      10.275 ms
     Transform time:         4.290 ms (6 calls)
     Copy data from GPU:    10.153 ms
   CUDA: 24.728704 ms
   Wrote hubble3.data
   
   % java WaveletSampleImage hubble3.data
   4096 x 4096 image written to "hubble3.jpg".