[Mesa-dev] [PATCH 3/3] glsl: relax loop unrolling restrictions

[Eero Tamminen]

Hi,

On 21.06.2017 13:12, Timothy Arceri wrote:
> The main reason these restriction exist is because glsl the loop
> unrolling pass is super slow with large loops.
>
> be5f27a84d0d fixed things so that expression trees were counted
> against the limit, however it left the limit as
> max_iterations * 5 which is actually fine for most shaders but
> probably over conservative.
>
> This change relaxes the limit to allow more loops to unroll in the
> Unigine Superposition benchmark.
>
> Results from Unigine Superposition @ 1920x1080 - High - Fullscreen
> On radeonsi (RX480)
>
> Before:
>
> Average: 28.20 Frames Per Second
>
> After:
>
> Average: 28.60 Frames Per Second
> ---
>
>  The increase is small but it seems to be consistent, I would be
>  interested in the results if others were interested in testing it.

Did testing on few different Intel machines with larger set of 
benchmarks (older than Superposition).

Didn't see any statistically significant performance changes in either 
direction (due to large number of tests, can't run many rounds so 
variance is fairly high).


Another change you could consider for larger loops is partial unrolling. 
  Unrolling several rounds, and then looping that.


	- Eero

>  src/compiler/glsl/loop_unroll.cpp | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/src/compiler/glsl/loop_unroll.cpp b/src/compiler/glsl/loop_unroll.cpp
> index bc377df..64ebf0f 100644
> --- a/src/compiler/glsl/loop_unroll.cpp
> +++ b/src/compiler/glsl/loop_unroll.cpp
> @@ -350,21 +350,21 @@ loop_unroll_visitor::visit_leave(ir_loop *ir)
>     /* Don't try to unroll loops that have zillions of iterations either.
>      */
>     if (iterations > max_iterations)
>        return visit_continue;
>
>     /* Don't try to unroll nested loops and loops with a huge body.
>      */
>     loop_unroll_count count(&ir->body_instructions, ls, options);
>
>     bool loop_too_large =
> -      count.nested_loop || count.nodes * iterations > max_iterations * 5;
> +      count.nested_loop || count.nodes * iterations > max_iterations * 10;
>
>     if (loop_too_large && !count.unsupported_variable_indexing &&
>         !count.array_indexed_by_induction_var_with_exact_iterations)
>        return visit_continue;
>
>     /* Note: the limiting terminator contributes 1 to ls->num_loop_jumps.
>      * We'll be removing the limiting terminator before we unroll.
>      */
>     assert(ls->num_loop_jumps > 0);
>     unsigned predicted_num_loop_jumps = ls->num_loop_jumps - 1;
>