local memory global used by non-kernel function

[robbert-harms]

When using ROCM with OpenCL 2.0 device side enqueue, I get the following error:

error: <unknown>:0:0: in function __compute_block_invoke void (i8*): local memory global used by non-kernel function

This happens when I try to use a workgroup function in a device side enqueued kernel. For example, the following kernel gives this error:

void compute_sum(float* sum){
    double s = 1;
    *sum = work_group_reduce_add(s);
}

kernel void compute(float* sum){
    enqueue_kernel(
        get_default_queue(), 
        CLK_ENQUEUE_FLAGS_NO_WAIT,
        ndrange_1D(1 * 64, 64), 
        ^{compute_sum(sum);}
    );
}

What I expected was that the device side enqueued kernel would run with a local workgroup size of 64, and as such, the sum variable should return 64.

Do let me know if I misunderstood something from the OpenCL specifications. Otherwise, is there a chance this might be fixed in a future release?

With kind regards,

Robbert

[eladmaimoni]

@robbert-harms I also faced a similar issue. I reported it here.

Did you find a solution?

[robbert-harms]

Unfortunately not, I stopped using kernel side enqueueing for this reason.

A workaround would be to do the program the summation yourself, but I grew rather fond of the work_group_ functions, so I stopped using the kernel side enqueue.

I hope this will be fixed in a future release.

[eladmaimoni]

@robbert-harms does your gpu happens to be from the new navi or rdna architecture?

[robbert-harms]

@elad8a no, I have a Radeon VII (Vega 20) GPU. It seems to be a broader issue than just navi/rdna.

[eladmaimoni]

@robbert-harms , check out the response from AMD, hopefully we will have a fix soon..

https://community.amd.com/thread/251691

[robbert-harms]

Great news! Thanks for notifying me. This issue has been a showstopper for me, I hope the fix works, else workgroup functions are basically useless (in dynamic kernels).