data transfer with cl::Buffers

[antibes0415]

Hi, I read the examples and found that data are transfered from host to device. But can I directly transfer a cl::Buffer object to the boost::compute::vector (device to device )or from compute::vector to cl::Buffer ? I think this is really important for high efficiency and interoperability with OpenCL C++ wrapper codes.

[jszuppe]

vector -> buffer: You can get underlying boost::compute::buffer (which is a wrapper for OpenCL buffer object) from boost::compute::vector object using its get_buffer() member function (if you don't want to copy). You can also copy existing vector into created boost::compute::buffer using boost::compute::copy algorithm (you just need to create buffer_iterator using buffer object). boost::compute::buffer can be created from OpenCL buffer object (cl_mem).

buffer -> vector You can't create vector from existing buffer so that buffer becomes vector's underlying buffer. You can copy existing buffer into existing vector using boost::compute::copy , or create new boost::compute::vector object using constructor which takes range [first, last).

[antibes0415]

Hi. Thanks. Could you please explain a bit more about converting compute::vector to cl::buffer ?

Right now I first do the following workflow: cl::buffer -> compute::buffer -> compute::vector -> sort compute::vector codes are:

cl::Buffer m_dev_sampleKeys = cl::Buffer(m_context, CL_MEM_READ_WRITE, m_nbSamples * sizeof(cl_long));
// create data...
boost::compute::buffer bSampleKeys(m_dev_sampleKeys()); // create from cl_mem
boost::compute::vector<cl_long> vSampleKeys(boost::compute::make_buffer_iterator<cl_long>(bSampleKeys, 0),boost::compute::make_buffer_iterator<cl_long>(bSampleKeys, m_nbSamples), m_compute_queue);
boost::compute::sort(vSampleKeys.begin(), vSampleKeys.end(), m_compute_queue);

Now how could I transfer the sorted compute::vector (vSampleKeys) data back to this cl::buffer (m_dev_sampleKeys)? My goal is not to use the host memory.

I tried using boost::compute::copy to copy vSampleKeys buffer to bSampleKeys. But I don't know how to finally copy to m_dev_sampleKeys.

[antibes0415]

By the way, does boost::compute::buffer(cl_mem) always do a copy ? Could it be a reference?

[jszuppe]

Now how could I transfer the sorted compute::vector (vSampleKeys) data back to this cl::buffer (m_dev_sampleKeys)? My goal is not to use the host memory.

I don't know why you want to use cl::Buffer and C++ wrapper from cl.hpp when you can just use Boost.Compute and don't waste any performance on copying. Anyway, to copy back to m_dev_sampleKeys, you just need to do this:

boost::compute::copy(
  vSampleKeys.begin(),
  vSampleKeys.end(), 
  boost::compute::make_buffer_iterator<cl_long>(bSampleKeys, 0),
  m_compute_queue
);

By the way, does boost::compute::buffer(cl_mem) always do a copy ? Could it be a reference?

boost::compute::buffer(cl_mem) constructor does not copy cl_mem object, it just wraps it. That means that when you copy to bSampleKeys you copy to m_dev_sampleKeys.

[jszuppe]

btw. You don't have to create boost::compute::vector to use Boost.Compute algorithms, you can just do:

boost::compute::sort(
  boost::compute::make_buffer_iterator<cl_long>(bSampleKeys, 0),
  boost::compute::make_buffer_iterator<cl_long>(bSampleKeys, m_nbSamples),
  m_compute_queue
);

[antibes0415]

Hi. You help me a lot. Thanks! And my apologies. I'm just a newbie in OpenCL and the original idea is that I want to use a library for something like sort, reduction, scan and count. Beyond that, I thought I would use the OpenCL C++ wrapper. I found that the document of boost.compute is not very detailed and I'm worried I could not make good use of it. The most important part may be the custom kernels. For my program, I need to write lots custom kernels. And I'm not clear how to use boost.compute to write kernels and set NDrange and run them... btw, I'm not clear the efficiency of running kernels compared with OpenCL.