Re: [Openexr-devel] Efficient read of a single image channel, in a multi-channel image

[Florian Kainz]

There is no particularly efficient way to read a single channel.
The channels in the file are interleaved per scan line or per tile.
This allows efficient random access to individual scan lines or tiles,
as well as sequential reading of the entire image with no seek
operations, but it precludes fast extraction of a single image channel.

If you frequently need to access particular subsets of image channels,
you may want to consider storing those subsets in separate files.

Florian


Jeremy Selan wrote:
> Hello!
>
> I am wondering what the most efficient mechanism is to read out a single 
> image from a multi-channel exr file.  Consider a 12 channel exr file, 
> all 32-bit float, 2Kx2K resolution, no compression. Reading a single 
> channel (code included below), strace reveals:
>
> -- lots of data relating to header access --
> _llseek(3, 17289, [17289], SEEK_SET)    = 0
> -- BEGIN READ PATTERN --
> read(3, "\371\7\0\0\320\177\1\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 
> 8191) = 8191
> read(3, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 
> 90073) = 90073
> futex(0x8073afc, FUTEX_WAKE, 1)         = 0
> -- END READ PATTERN --
>
> for a total of 201180709 bytes read.
>
> ... implying that all 12 channels of data were pulled across the network 
> to read this single channel. Unfortunately, this access pattern limits 
> our ability to store a very large number of data channels in a single 
> file. Is there a more efficient read technique?
>
> All tests were done on Exr 1.4.0.  My apologies if this has changed in 1.5.
>
> Thanks!
> -- Jeremy
>
>
> ============================
> Imf::InputFile sourceFile(argv[argc-1]);
> sourceFile.header().sanityCheck();
>
> Imath::Box2i dw = sourceFile.header().dataWindow();
> unsigned int dataWindow_width = dw.max.x - dw.min.x + 1;
> int dataWindow_ymin = std::min(dw.min.y, dw.max.y);
> int dataWindow_ymax = std::max(dw.min.y, dw.max.y);
>
> std::vector<char> pixelBuffer(dataWindow_width * sizeof(float));
>
> // Read out a single channel
> Imf::ChannelList::ConstIterator it = 
> sourceFile.header().channels().begin();
>
> std::string channelName(it.name());
> std::cout << "[exrchanneltest]: Processing channel: '" << channelName << 
> "'." << std::endl;
>
> // For all scanlines...
> for(int y = dataWindow_ymin; y<=dataWindow_ymax; y++)
> {
>     // Build a decode structure.
>     // This is inefficient to create a new FrameBuffer for each scanline,
>     // but shouldnt matter for this simple test.
>     Imf::FrameBuffer readBuffer;
>     readBuffer.insert(channelName.c_str(), Imf::Slice(Imf::FLOAT,
>                                 &pixelBuffer[0] - dw.min.x 
> *sizeof(float), sizeof(float),
>                                0, 1, 1, 0.0f));
>     sourceFile.setFrameBuffer(readBuffer);
>                   // Decode
>    sourceFile.readPixels(y);
>    float* floatBuffer = reinterpret_cast<float*>(&pixelBuffer[0]);
> }
>
>
>
> _______________________________________________
> Openexr-devel mailing list
> address@hidden
> http://lists.nongnu.org/mailman/listinfo/openexr-devel