[Mesa-dev] software implementation of vulkan for gsoc/evoc

[Nicolai Hähnle]

On 14.02.2017 09:58, Jacob Lifshay wrote:
>
>
> On Feb 14, 2017 12:18 AM, "Nicolai Hähnle" <nhaehnle at gmail.com
> <mailto:nhaehnle at gmail.com>> wrote:
>
>     On 13.02.2017 17:54, Jacob Lifshay wrote:
>
>         the algorithm i was going to use would get the union of the sets
>         of live
>         variables at the barriers (union over barriers), create an array of
>         structs that holds them all, then for each barrier, insert the
>         code to
>         store all live variables, then end the for loop over
>         tid_in_workgroup,
>         then run the memory barrier, then start another for loop over
>         tid_in_workgroup, then load all live variables.
>
>
>     Okay, sounds reasonable in theory.
>
>     There are some issues, like: how do you actually determine live
>     variables? If you're working off TGSI like llvmpipe does today,
>     you'd need to write your own analysis for that, but in a structured
>     control flow graph like TGSI has, that shouldn't be too difficult.
>
>
> I was planning on using the spir-v to llvm translator and never using
> tgsi.

Cool, it would be interesting to see how that goes. Mind you, I don't 
think that code is being maintained very well.


> I could implement the pass using llvm coroutines, however, I'd
> need to have several additional passes to convert the output; it might
> not optimize all the way because we would have the switch on the suspend
> point index still left. Also, according to the docs from llvm trunk,
> llvm doesn't support reducing the space required by using the minimum
> size needed to store all the live variables at the suspend point with
> the largest space requirements, instead, it allocates separate space for
> each variable at each suspend
> point: http://llvm.org/docs/Coroutines.html#areas-requiring-attention

Yes, that actually makes sense. About the switches, though, I'm not so 
sure how you can really avoid those. Consider kernel code like this:

void main()
{
    if (cond) {
       ...
       barrier();
       ...
    } else {
       ...
       barrier();
       ...
    }
}

This kernel is perfectly valid and will work as expected if (and only 
if) cond is uniform across the threads of a workgroup.

Consider what you'd want the control flow in the LLVM implementation to 
look like, how you'd handle the fact that the set of live values would 
be different across the different barriers.

As a bonus, perhaps you could set things up so that the user gets a nice 
error message when the kernel is incorrect (i.e., when cond is _not_ 
uniform across a workgroup).

Cheers,
Nicolai

>
>
>     I'd still recommend you to at least seriously read through the LLVM
>     coroutine stuff.
>
>     Cheers,
>     Nicolai
>
>         Jacob Lifshay
>
>         On Feb 13, 2017 08:45, "Nicolai Hähnle" <nhaehnle at gmail.com
>         <mailto:nhaehnle at gmail.com>
>         <mailto:nhaehnle at gmail.com <mailto:nhaehnle at gmail.com>>> wrote:
>
>             [ re-adding mesa-dev on the assumption that it got dropped
>         by accident ]
>
>             On 13.02.2017 17:27, Jacob Lifshay wrote:
>
>                         I would start a thread for each cpu, then have each
>                 thread run the
>                         compute shader a number of times instead of having a
>                 thread per
>                         shader
>                         invocation.
>
>
>                     This will not work.
>
>                     Please, read again what the barrier() instruction
>         does: When the
>                     barrier() call is reached, _all_ threads within the
>                 workgroup are
>                     supposed to be run until they reach that barrier() call.
>
>
>                 to clarify, I had meant that each os thread would run the
>                 sections of
>                 the shader between the barriers for all the shaders in a
>         work group,
>                 then, when it finished the work group, it would go to
>         the next work
>                 group assigned to the os thread.
>
>                 so, if our shader is:
>                 a = b + tid;
>                 barrier();
>                 d = e + f;
>
>                 and our simd width is 4, our work-group size is 128, and
>         we have
>                 16 os
>                 threads, then it will run for each os thread:
>                 for(workgroup = os_thread_index; workgroup <
>         workgroup_count;
>                 workgroup++)
>                 {
>                     for(tid_in_workgroup = 0; tid_in_workgroup < 128;
>                 tid_in_workgroup += 4)
>                     {
>                         ivec4 tid = ivec4(0, 1, 2, 3) +
>         ivec4(tid_in_workgroup +
>                 workgroup * 128);
>                         a[tid_in_workgroup / 4] =
>         ivec_add(b[tid_in_workgroup /
>                 4], tid);
>                     }
>                     memory_fence(); // if needed
>                     for(tid_in_workgroup = 0; tid_in_workgroup < 128;
>                 tid_in_workgroup += 4)
>                     {
>                         d[tid_in_workgroup / 4] =
>         vec_add(e[tid_in_workgroup / 4],
>                 f[tid_in_workgroup / 4]);
>                     }
>                 }
>                 // after this, we run the next rendering or compute job
>
>
>             Okay good, that's the right concept.
>
>             Actually doing that is not at all straightforward though:
>         consider
>             that the barrier() might occur inside a loop in the shader.
>
>             So if you implemented that within the framework of llvmpipe,
>         you'd
>             make a lot of people very happy: it would allow finally adding
>             compute shader support to llvmpipe. Mind you, that in itself
>         would
>             already be a pretty decent-sized project for GSoC!
>
>             Cheers,
>             Nicolai
>
>
>
> Jacob Lifshay