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

[Jeremy Selan]

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]);
}