On 7/19/2018 2:35 AM, Tvrtko Ursulin wrote:
> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
>
> Request split mode had several bugs, both in the original version and also
> after the recent refactorings.
>
> One big one was that it wasn't considering different submit ports as a
> reason to split execution, and also that it was too time based instead of
> looking at relevant timelines.
>
> In this refactoring we address the former by using the engine timelines
> introduced in the previous patch. Secondary port submissions are moved
> to follow the preceding submission as a first step in the correction
> process.
>
> In the second step, we add context timelines and use then in a similar
> fashion to separate start and end time of coalesced requests. For each
> coalesced request we know its boundaries by looking at the engine
> timeline (via global seqnos), and we know the previous request it should
> only start after, by looking at the context timeline.
>
> v2:
>   * Remove some dead code.
>   * Fix !port0 shifting logic.
>
> v3:
>   * Refactor for less list walking as with incomplete handling.
>
> v4:
>   * Database of context timelines should not contain duplicates!
>     (Converted from array into a hash.)
>
> v5:
>   * Avoid over-accounting runnable time for a coalesced group by recording
>     the time first request entered the GPU and ending the execute delay at
>     that point for the whole group.
>
> v6:
>   * Update for engine class:instance.
>
> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
> Cc: John Harrison <John.C.Harrison@intel.com>
> ---
>   scripts/trace.pl | 138 ++++++++++++++++++++++++++++++++++++-----------
>   1 file changed, 108 insertions(+), 30 deletions(-)
>
>

Reviewed-by: John Harrison <John.C.Harrison@Intel.com>