Does the OpenCL2.0 SVM support allocate memory for images?

yi_l_ · ‎08-05-2014

We can test the SVM performance like:

d_a=(float*)clSVMAlloc(context,(CL_MEM_READ_WRITE|CL_MEM_SVM_FINE_GRAIN_BUFFER|CL_MEM_SVM_ATOMICS),size,0);
d_b=(float*)clSVMAlloc(context,(CL_MEM_READ_WRITE|CL_MEM_SVM_FINE_GRAIN_BUFFER|CL_MEM_SVM_ATOMICS),size,0);
d_c=(float*)clSVMAlloc(context,(CL_MEM_READ_WRITE|CL_MEM_SVM_FINE_GRAIN_BUFFER|CL_MEM_SVM_ATOMICS),size,0);
for(int i=0;i<size;i++)
{
d_a=Initial_value;
d_b=Initial_value;
}
clSetKernelArgSVMPointer(kernel, 0, d_a);
clSetKernelArgSVMPointer(kernel, 0, d_b);
clSetKernelArgSVMPointer(kernel, 0, d_c);

The SVM performance has been verified good in the example of calculation of a vector,but the clSVMAlloc() cannot allocate memory for images,
how should we do to utilize the SVM for image processing?
Thank you!

The test device is like:

        CL_DEVICE_NAME: Genuine Intel(R) CPU 0000 @ 1.60GHz
        CL_DEVICE_VENDOR:       Intel(R) Corporation
        CL_DRIVER_VERSION:      4.0.0.8228
        CL_DEVICE_PROFILE:      FULL_PROFILE
        CL_DEVICE_VERSION:      OpenCL 2.0 (Build 8228)
        CL_DEVICE_OPENCL_C_VERSION:     OpenCL C 2.0

Aharon_A_Intel · ‎08-06-2014

Hello, Yi.

Although creating an image from an SVM buffer is not explicitly mentioned in the spec (it just specifies creating buffers out of memory allocated by clSVMAlloc), there is no reason why creating an image object with such a pointer and CL_MEM_USE_HOST_PTR shouldn't be valid.