Re: [ViennaCL-support] OpenCL porting

[Sajjadul I. <saj...@st...>]

Hi


Thanks for the hint!

I need to use the vectors build on the device to be usable across several
function in the host. For example

//////////////////////////////////////////////////////////////////////////////////////////////////

int main()
{
   .........................

   .........................

    //allocate a host vector with a certain size
    //we are storing the 2D vector field, so each x-component
    //and y-component are stored in a linear manner
    //so the dimension DS is multiplied by 2
    std::vector<cData> hvfield(DS * 2);
    //fill the cpu vector with the default values - 0.0f
    std::fill(hvfield.begin(),hvfield.end(),0.0f);

    //allocate the vector of same size on the device
    viennacl::vector<cData> dvfield(DS * 2);

    //copy the vector from the cpu to the gpu
    copy(hvfield.begin(),hvfield.end(),dvfield.begin());

    viennacl::vector<cData> vxfield(PDS * 2);
    viennacl::vector<cData> vyfield(PDS * 2);

....................................................................

....................................................................

}


void display()
{
    simulateFluids();
}


void simulateFluids(.....)
{
    //i need to use some vectors here that has been allocated in the main
function
}

//////////////////////////////////////////////////////////////////////////////////////////////////


Now if i put the vector allocation at the global scope i get segmentation
fault while running the program. OpenGL's callback display() function do
the necessary calculation for the fluid simulation and i need to access the
vectors i have allocated. Can they (the ones allocated on the device) be
sent as reference in the function parameter ?

I checked the tutorials and none of them matched my requirements.

Any hint?


Regards
Sajjadul

On Wed, Mar 13, 2013 at 4:50 PM, Karl Rupp <ru...@iu...> wrote:

> Hi Sajjadul,
>
> just define a
>  viennacl::vector<float>(2*PDS)**;
> or
>  viennacl::vector<double>(2***PDS);
> and then fill it with the real part on locations [0],[2], etc. and the
> imaginary part on locations [1],[3], etc.
> This is pretty similar to what you actually do with cufft.
>
> The lack of complex_t in current OpenCL is one of the reasons why FFT in
> ViennaCL is still experimental. Thus, you cannot properly define a float2
> on the CPU without running into compiler problems (nvcc can only do this
> because they have control over the full compilation and thus can do any
> non-standard-compliant stuff they want).
>
> Best regards,
> Karli
>
>
>
>
>
> On 03/13/2013 09:43 AM, Sajjadul Islam wrote:
>
>> Hi forum,
>>
>>
>> I am trying to port an existing CUDA application to OpenCL, and the
>> process seems not to be so trivial while working with the custom data
>> type and FFT. Below i am describing some custom data type that has been
>> assigned in CUDA and my effort to convert them to opencl and also the
>> issues that i am currently stuck with:
>>
>>
>> *CUDA*
>> *
>> *
>> *
>>
>> *
>> // Vector data type used to velocity and force fields
>> //typedef is a built-in vector types
>> typedef float2 cData;
>>
>>
>>     //each of the allocation is of size 131584
>>     cudaMalloc((void**)&vxfield, sizeof(cData) * PDS);
>>     cudaMalloc((void**)&vyfield, sizeof(cData) * PDS);
>>
>>
>> ..............................**.........
>> ..............................**........
>>
>>    // Create CUFFT transform plan configuration
>>     // create a simple plan for 2D transformation
>>     cufftPlan2d(&planr2c, DIM, DIM, CUFFT_R2C);
>>     cufftPlan2d(&planc2r, DIM, DIM, CUFFT_C2R);
>>     // TODO: update kernels to use the new unpadded memory layout for perf
>>     // rather than the old FFTW-compatible layout
>>     cufftSetCompatibilityMode(**planr2c, CUFFT_COMPATIBILITY_FFTW_**
>> PADDING);
>>     cufftSetCompatibilityMode(**planc2r, CUFFT_COMPATIBILITY_FFTW_**
>> PADDING);
>>
>>
>> ..............................**.
>> ..............................
>>
>>      // Forward FFT
>>      cufftExecR2C(planr2c, (cufftReal*)vx, (cufftComplex*)vx);
>>      cufftExecR2C(planr2c, (cufftReal*)vy, (cufftComplex*)vy);
>>
>> ..............................**.
>> ..............................
>>
>>      // Inverse FFT
>>      cufftExecC2R(planc2r, (cufftComplex*)vx, (cufftReal*)vx);
>>      cufftExecC2R(planc2r, (cufftComplex*)vy, (cufftReal*)vy);
>>
>>
>>
>> *OpenCL*
>> *
>>
>> *
>> Even if i define custom data type for i am not sure if i can use it
>> along with the ViennaCL as you currently are not supporting any type
>> other than double / float.
>>
>> I need to send data to fft function that is of type cl_float2 as you can
>> imagine from the above CUDA fft statements.
>>
>> Still i believe that  there must be some way to get  around this issue
>> and i need you suggestion.
>>
>>
>> I shall be eagerly waiting for your response over this issue.
>>
>>
>>
>> Regards
>> Sajjad
>>
>>
>> ------------------------------**------------------------------**
>> ------------------
>> Everyone hates slow websites. So do we.
>> Make your web apps faster with AppDynamics
>> Download AppDynamics Lite for free today:
>> http://p.sf.net/sfu/appdyn_**d2d_mar <http://p.sf.net/sfu/appdyn_d2d_mar>
>>
>>
>>
>> ______________________________**_________________
>> ViennaCL-support mailing list
>> ViennaCL-support@lists.**sourceforge.net<Vie...@li...>
>> https://lists.sourceforge.net/**lists/listinfo/viennacl-**support<https://lists.sourceforge.net/lists/listinfo/viennacl-support>
>>
>>
>