[Mesa-dev] [RFC 1/3] tgsi: Add WORK_DIM System Value

[Hans de Goede]

Hi,

On 28-04-16 10:13, Pierre Moreau wrote:
> On 05:43 PM - Apr 27 2016, Samuel Pitoiset wrote:
>>
>>
>> On 04/27/2016 05:24 PM, Ilia Mirkin wrote:
>>> On Wed, Apr 27, 2016 at 11:19 AM, Hans de Goede <hdegoede at redhat.com> wrote:
>>>> Hi,
>>>>
>>>> On 27-04-16 16:49, Ilia Mirkin wrote:
>>>>>
>>>>> Please add this semantic to src/gallium/docs and explain what it
>>>>> means.
>>>>
>>>>
>>>> Ah, I was under the impression these were not documented, but I
>>>> now see they are, will fix.
>>>>
>>>>> (I'm not even sure what this is, and the easily-found opencl
>>>>>
>>>>> docs helpfully indicate that get_work_dim returns the work_dim...)
>>>>> This wasn't done for some of the other compute-specific semantics.
>>>>
>>>>
>>>> When launching a grid through clEnqueueNDRangeKernel the user
>>>> can tell how many dimensions of the pipe_grid_info->grid /
>>>> pipe_grid_info->block equivalent the user is actually passing in
>>>> this is done through a parameter to that function called work_dim:
>>>>
>>>> https://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clEnqueueNDRangeKernel.html
>>>>
>>>> Clover pads the y and z values (if unused) to 1, so it always
>>>> ends up giving a 3 dimensional grid and block sizes to the driver.
>>>>
>>>> But an opencl kernel can use the opencl builtin get_work_dim
>>>> function, which returns how much dimensions where actually
>>>> passed into clEnqueueNDRangeKernel.
>>>>
>>>> Note that passing in e.g. work_dim = 3 with a grid size of { 4, 1, 1 }
>>>> is perfectly legal and then get_work_dim should actually return 3.
>>>>
>>>> So we need to store the clEnqueueNDRangeKernel workdim parameter
>>>> somewhere, and allow getting it from a kernel. Using a system-value
>>>> seems the best solution for this to me.
>>>>
>>>> I hope this helps to explain things, if not keep asking :)
>>>>
>>>> Also if you've a better solution, I'm all ears.
>>>
>>> Sounds like there's no real way to distinguish it from the kernel, the
>>> parameters given to the hw are all the same, so it's just a thing that
>>> needs to be passed along. I think that what you have is a fine
>>> solution, although I question the usefulness of this being an OpenCL
>>> builtin. But we're not about to change that now.
>>>
>>> An alternative would be for clover to stick it into a uniform and have
>>> the kernel read it out from there. If there are going to be a ton of
>>> these, that might be preferable. If it's just the one (or maybe two),
>>> not worth it.
>>
>> I think your solution is fine too. Storing the number of dimensions passed
>> to clEnqueueNDRangeKernel() as well as the local/global sizes makes sense.
>
> Later versions of OpenCL are adding a few more of those (see [OpenCL 1.0][1]
> compared to [OpenCL 2.0][2]). So putting it in a uniform would be better, but
> we can just switch to a uniform the day we reach OpenCL 2.0.

It looks like most of these can be implemented using libclc wrappers
around the existing workitem functions.

Regards,

Hans