Archives Discussions

[Ziple]

Hello everyone,

I have a certain kernel, doing some computations on blocks of matrices (I have a block size of 128 matrices).

To process a block, I iterate over all the matrices in a block.

It seems that this loop is too agressively unrolled, and it disminishes the number of in-flight wavefronts from 4 to 2.

As I need a certain number of wavefronts to hide the memory latency, I would like to specify to the OpenCL compiler to NOT unroll this loop, is this possible?

I am currently passing this constant with a kernel argument to avoid the unrolling, but that's not a great solution IMO...

Thank you for your help,

Regards,

Ziple.

[dipak · Accepted solution]

Hi Ziple,

You may try __attribute__((opencl_unroll_hint(1))) for this purpose. As per the OpenCL C Spec:

Section: Specifying Attribute For Unrolling Loops

The __attribute__((opencl_unroll_hint)) and __attribute__((opencl_unroll_hint(n))) attribute qualifiers can be used to specify that a loop (for, while and do loops) can be unrolled. This attribute qualifier can be used to specify full unrolling or partial unrolling by a specified amount. This is a compiler hint and the compiler may ignore this directive.

n is the loop unrolling factor and must be a positive integral compile time constant expression. An unroll factor of 1 disables unrolling. If n is not specified, the compiler determines the unrolling factor for the loop.

NOTE: The __attribute__((opencl_unroll_hint(n))) attribute qualifier must appear immediately before the loop to be affected.

Example:

__attribute__((opencl_unroll_hint(1)))

for (int i=0; i<32; i++)

{

…

}

For details, please refer the section "Specifying Attribute For Unrolling Loops" in OpenCL C Spec.

Regards,

[Ziple]

Thank you for your help, I think that it will do the trick!