Databases
StackOverflow2013
Postsdbo
PO
 

Note that there are some explanatory texts on larger screens.

plurals
  1. PO
    primarykey
    data
    text
    <p>So, it seems you do have a strange <code>#include</code>, leading to code getting compiled by the wrong compiler. Make a distinction between gpu headers and cpu headers by using .cu.h for cuda headers. Make sure <em>only</em> NVCC compiles <code>.cu</code> and <code>.cu.h</code> files. Cuda files should never be included in cpp files. The kernel and kernel call should be in a <code>.cu</code> or <code>.cu.h</code> files, and those files shouldn't be included anywhere in cpps. </p> <p>Because your <code>.cu</code> is being included in a header which is being compiled by the host compiler, the host compiler ends up hitting the token <code>&lt;&lt;&lt;</code> - which it doesn't recognise. It probably does understand the token <code>&lt;&lt;</code> so it consumes that, leaving an unexpected <code>&lt;</code>.</p> <p>Here's an alternative way of doing things that should work (not tried it but it's similar to code we use)</p> <p><em>(note, this might work but it also might not be the right way to solve the problem. My boss doesn't like it as a solution and would prefer to add an implementation per variation)</em> </p> <p>The underlying problem seems to be lack of distinction between host and device code. I'm leaving the detail out in my solution - things like copying results to and from the device, sum implementation, etc.</p> <p>The problem I'm trying to solve is, given a construct, how can you template it for use both on the host and the device?</p> <p>I'll template <code>Matrix.h</code> on both the type and the implementation detail.</p> <pre><code> template &lt;typename T, typename Implementation&lt;T&gt; &gt; class Matrix { void sum(Matrix&lt;T&gt;&amp; m1, Matrix&lt;T&gt;&amp; m2, Matrix&lt;T&gt;&amp; sum) { Implementation.sumImp(m1, m2, sum); } } </code></pre> <p>The host implementation, <code>HostMatrixSum.h</code> will do things the on the cpu:</p> <pre><code> #include "Matrix.h" template &lt;typename T&gt; struct HostMatrixSum { void sumImp(Matrix&lt;T&gt;&amp; m1, Matrix&lt;T&gt;&amp; m2, Matrix&lt;T&gt;&amp; sum) { ... } } </code></pre> <p>While <code>GpuMatrixSum.cu.h</code> will upload the matrix, do the sum and recover the results:</p> <pre><code> #include "Matrix.h" template &lt;typename T&gt; struct GpuMatrixSum { template&lt;typename T&gt; __global__ void sumKernel(const Matrix&lt;T&gt; m1, const Matrix&lt;T&gt; m2, Matrix&lt;T&gt; sum) { ... } void sumImp(Matrix&lt;T&gt;&amp; m1, Matrix&lt;T&gt;&amp; m2, Matrix&lt;T&gt;&amp; sum) { ... sumKernel&lt;T&gt; &lt;&lt;&lt; dimGrid, dimBlock &gt;&gt;&gt; (m1,m2); ... } } </code></pre> <p>Then when we come to use Matrix from host code we template on the host sum implementation and never need to see any cuda specifics:</p> <pre><code> #include "Matrix.h" #include "HostMatrixSum.h" Matrix&lt;int, HostMatrixSum&gt; m1 = Matrix&lt;int&gt;(...); Matrix&lt;int, HostMatrixSum&gt; m2 = Matrix&lt;int&gt;(...); Matrix&lt;int, HostMatrixSum&gt; result; Matrix.sum(m1,m2,result); </code></pre> <p>And if we're working on the gpu we can use the accelerated gpu implementation of sum:</p> <pre><code> #include "Matrix.h" #include "GpuMatrixSum.cu.h" Matrix&lt;int, GpuMatrixSum&gt; m1 = Matrix&lt;int&gt;(...); Matrix&lt;int, GpuMatrixSum&gt; m2 = Matrix&lt;int&gt;(...); Matrix&lt;int, GpuMatrixSum&gt; result; Matrix.sum(m1,m2,result); </code></pre> <p>Hope that works for you!</p>
    singulars
    1. This table or related slice is empty.
    1. POProblem when calling template CUDA kernel
      singulars
      1. PTQuestion
    1. PTAnswer
    1. USMatt Bond
    1. USMatt Bond
    plurals
    1. This table or related slice is empty.
    1. This table or related slice is empty.
    1. POProblem when calling template CUDA kernel
      singulars
      1. PTQuestion
    1. This table or related slice is empty.
    1. VO
      singulars
      1. PO
      1. This table or related slice is empty.
      1. VTAcceptedByOriginator
    2. VO
      singulars
      1. PO
      1. This table or related slice is empty.
      1. VTUpMod
    3. VO
      singulars
      1. PO
      1. This table or related slice is empty.
      1. VTUpMod
    1. COI'm going to try that. But at the same time it seems strange that in MatrixKernel.h, for instance, the compiler doesn't complain about the __global__ keyword (which can only mean NVCC is compiling it, right?) Another thing: if what you say actually is the problem, where would I implement the sum method? If I don't write '#include "Matrix.cu"' inside "Matrix.h", there will be a link error, for the template must be declared and defined at the same file...
      singulars
      1. PO
      1. USRenan
    2. COI think you were right, I had to forget about implementing Matrix as a template class, cause there was no way I could make it work that way. If I include for instance a .cu file on the header, every other file that includes the Matrix header will also include the .cu, even .cpp ones, which will inevitably give rise to a compilation error. Using template kernels is ok, but making the c++ methods that call them templates too is not possible because of what I just explained. It's a little winding after all...
      singulars
      1. PO
      1. USRenan
    3. COBecause MatrixKernel.h is only included by a cu file, only nvcc ever includes it. If you included MatrixKernel.h in a host cpp file, I suspect it would fall over. I name cuda specific headers .cu.h to clarify that it's only for inclusion from .cu files. I'm working on an answer to the 'how to do it' part of the question - trying to work out an elegant solution :)
      singulars
      1. PO
      1. USMatt Bond
 

Querying!

 
Guidance
A row detail

Detail views are divided into sections. All the information in the data section comes from columns in the selected row. The other sections display data from other, related rows.

Related data can be related in a to-one or a to-many fashion. Captions of data related in a to-many fashion link to a list view showing a filtered view of the table.

Try moving around until you find a non-empty to-many entry and click on the label to get to one. You can move back to the root by clicking on the database name in the header.

SQuiL has stopped working due to an internal error.

If you are curious you may find further information in the browser console, which is accessible through the devtools (F12).

Reload