[Mesa-dev] [PATCH 5/5] i965/sched: Change the scheduling heuristics to favor early program termination.

[Jason Ekstrand]

On Tue, Aug 16, 2016 at 1:54 PM, Francisco Jerez <currojerez at riseup.net>
wrote:

> This uses the unblocked time of the exit assigned to each available
> node to attempt to unblock exit nodes as early as possible,
> potentially reducing the runtime of the shader when an exit branch is
> taken.  There is a natural trade-off between terminating the program
> as early as possible and reducing the worst-case latency of the
> program as a whole (since this will typically move exit-unblocking
> nodes closer to its dependencies potentially causing additional stalls
> of the execution pipeline), but in practice the bandwidth and ALU
> cycle savings from terminating the program earlier tend to outweigh
> the slight increase in worst-case program execution latency, so it
> makes sense to prefer nodes likely to unblock an earlier exit
> regardless of the latency benefits of other available nodes.
>
> I haven't observed any benchmark regressions from this change after
> testing on VLV, HSW, BDW, BSW and SKL.  The FPS of the GfxBench
> Manhattan benchmark increases by 10%-20% and the FPS of Unigine Valley
> improves by roughly 5% depending on the platform and settings.
>

Thanks for working on this!  We've known about it for a while and it's nice
to finally get some progress.

Reviewed-by: Jason Ekstrand <jason at jlekstrand.net>


> The change to the register pressure-sensitive heuristic is rather
> conservative and gives precedence to the existing heuristic in order
> to avoid increasing register pressure and causing spill count and SIMD
> width regressions in shader-db.  It may make sense to revisit this
> with additional performance data.
> ---
>  .../drivers/dri/i965/brw_schedule_instructions.cpp    | 19
> ++++++++++++++++---
>  1 file changed, 16 insertions(+), 3 deletions(-)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_schedule_instructions.cpp
> b/src/mesa/drivers/dri/i965/brw_schedule_instructions.cpp
> index 96562cf..dfcaa80 100644
> --- a/src/mesa/drivers/dri/i965/brw_schedule_instructions.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_schedule_instructions.cpp
> @@ -1407,11 +1407,15 @@ fs_instruction_scheduler::choose_instruction_to_
> schedule()
>     if (mode == SCHEDULE_PRE || mode == SCHEDULE_POST) {
>        int chosen_time = 0;
>
> -      /* Of the instructions ready to execute or the closest to
> -       * being ready, choose the oldest one.
> +      /* Of the instructions ready to execute or the closest to being
> ready,
> +       * choose the one most likely to unblock an early program exit, or
> +       * otherwise the oldest one.
>         */
>        foreach_in_list(schedule_node, n, &instructions) {
> -         if (!chosen || n->unblocked_time < chosen_time) {
> +         if (!chosen ||
> +             exit_unblocked_time(n) < exit_unblocked_time(chosen) ||
> +             (exit_unblocked_time(n) == exit_unblocked_time(chosen) &&
> +              n->unblocked_time < chosen_time)) {
>              chosen = n;
>              chosen_time = n->unblocked_time;
>           }
> @@ -1500,6 +1504,15 @@ fs_instruction_scheduler::choose_instruction_to_
> schedule()
>              continue;
>           }
>
> +         /* Prefer the node most likely to unblock an early program exit.
> +          */
> +         if (exit_unblocked_time(n) < exit_unblocked_time(chosen)) {
> +            chosen = n;
> +            continue;
> +         } else if (exit_unblocked_time(n) > exit_unblocked_time(chosen))
> {
> +            continue;
> +         }
> +
>           /* If all other metrics are equal, we prefer the first
> instruction in
>            * the list (program execution).
>            */
> --
> 2.9.0
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20160816/582738d8/attachment.html>