• State Accepted Answer
  • +1 person also asked this people also asked this
  • Locked Locked
  • Replies 5 replies
  • Subscribers 136 subscribers
  • Views 6749 views
  • Users 0 members are here
Options
Related
How was your experience today?
This discussion has been locked.
You can no longer post new replies to this discussion. If you have a question you can start a new discussion

What's the best way to fill a buffer?

songmao

ARM Graphic expert:

     I have a cl_mem buffer, which will be reused every camera frame,  so it need reset to zero before we can use it.

     for opencl 1.2 we can use clEnqueueFillBuffer, for mali driver, what would be the most efficient way to fill the buffer?

     I have write a kernel and fill the buffer on gpu, but it turn out to be very slow, the kernel:

__kernel void set(__global int* dstptr)

{

    int y = get_global_id(0);

    __global int* dst = dstptr + y*ROW_WIDTH;

    for (int i=0; i<(ROW_WIDTH/4); i++) {

        *dst = 0;

        dst++;

    }  

}

     ROW_WIDTH is a compile time macro, which is set to prime number of the buffer size, for example, for the buffer size 1280*960,  it's 75.

     And it turn out to be very slow, why? and what's can be done?

Parents
  • +1

    Hi Songmao,

    many thanks for your question.

    As you correctly said, clEnqueueFillBuffer is available from OpenCL 1.2.

    Starting from your OpenCL kernel for resetting to 0 a CL buffer, I listed a couple of main strategies for improving its performance

    1. Avoiding the for loop as it may reduce the number of registers the kernel needs allowing more work items to run on the GPU at the same time. The for loop can be replaced by using the second dimension of the GWS.
    2. Exploiting the vector capabilities of ARM GPUs. Mali-T600 and T700 series GPUs natively support all CL data types as have SIMD capabilities. This means that is is possible to compute the same operation on multiple datas allowing to dispatch fewer work-items.

    Using these 2 strategies, a possible OpenCL kernel for resetting to 0 a CL buffer could be:

    __kernel void set(__global int* dstptr)

    { 

        /* Each work item stores 4 int values */

        int x = get_global_id(0) * 4

        int y = get_global_id(1);

        int offset = x + y * ROW_WIDTH;

        vstore4((int4)0, 0, dstptr + offset);

    }

    In order to reproduce the result, could you tell us which ARM GPU you use and if you have disabled the DVFS for your performance evaluation?

    I hope this can help you,

    Thanks,

    Gian Marco

Reply
  • +1

    Hi Songmao,

    many thanks for your question.

    As you correctly said, clEnqueueFillBuffer is available from OpenCL 1.2.

    Starting from your OpenCL kernel for resetting to 0 a CL buffer, I listed a couple of main strategies for improving its performance

    1. Avoiding the for loop as it may reduce the number of registers the kernel needs allowing more work items to run on the GPU at the same time. The for loop can be replaced by using the second dimension of the GWS.
    2. Exploiting the vector capabilities of ARM GPUs. Mali-T600 and T700 series GPUs natively support all CL data types as have SIMD capabilities. This means that is is possible to compute the same operation on multiple datas allowing to dispatch fewer work-items.

    Using these 2 strategies, a possible OpenCL kernel for resetting to 0 a CL buffer could be:

    __kernel void set(__global int* dstptr)

    { 

        /* Each work item stores 4 int values */

        int x = get_global_id(0) * 4

        int y = get_global_id(1);

        int offset = x + y * ROW_WIDTH;

        vstore4((int4)0, 0, dstptr + offset);

    }

    In order to reproduce the result, could you tell us which ARM GPU you use and if you have disabled the DVFS for your performance evaluation?

    I hope this can help you,

    Thanks,

    Gian Marco

Children