cuda : generalize pocl_cuda_submit_memfill

[trixirt]

The use of cuMemsetD32|16 is dependent on the alignment and
not as much the size. Check the alignment of the inputs
and use an appropriate cuMemset.

Signed-off-by: Tom Rix trix@redhat.com

[trixirt]

conformance_buffers_fill mostly works now on jetson, before it failed when the size was 8+
failures now are on the flags wanting to use host ptr.

[isuruf]

@trixirt, I don't think this is correct. AFAIK, pattern_size gives the length of the data in pattern, but cuMemsetD32Async considers pattern to be a 32-bit value. Is my understanding incorrect?

[trixirt]

yes i believe so, that cu api is a value not a pointer.
I must have been getting lucky on the conformance test.
I will rework.

[Oblomov]

I'm afraid the only way to properly implement clEnqueueFillBuffer through CUDA is using kernels. This may be optimized for specific alignment/size cases, but a kernel for the generic case is necessary.

[trixirt]

Can you explain why a write and copies would not work ?
The test i am fixing is conformance_buffers_fill

[Oblomov]

If we're only aiming for correctness, then yes, the approach would work. However, the performance is going to be abysmal.

Buffer filling should work at close to peak bandwidth performance. If it's implemented as a doubling copy, you're only going to get < 1/3rd the performance that the hardware could achieve (it's sort of like a “reverse reduction”). Even a simple kernel doing a[i] = pattern would do better in this case.

(That being said, it's obviously OK to strive for correctness now and then aim for performance later.)

[pjaaskel]

What's the plan with this work?

[isuruf]

Now that there are builtin cuda kernel support, it should be easy to write a memfill kernel and plug it to pocl_cuda_submit_memfill