[Mesa-dev] [PATCH] amdgpu/winsys: adjust IB size based on buffer wait time

[Marek Olšák]

On Fri, Apr 15, 2016 at 7:50 PM, Grigori Goronzy <greg at chown.ath.cx> wrote:
> Small IBs help to reduce stalls for workloads that require a lot of
> synchronization. On the other hand, if there is no notable
> synchronization, we can use a large IB size to slightly improve
> performance in some cases.
>
> This introduces tuning of the IB size based on feedback on the average
> buffer wait time. The average wait time is tracked with exponential
> smoothing.
> ---
>  src/gallium/winsys/amdgpu/drm/amdgpu_bo.c     | 2 ++
>  src/gallium/winsys/amdgpu/drm/amdgpu_cs.c     | 8 ++++++--
>  src/gallium/winsys/amdgpu/drm/amdgpu_winsys.h | 1 +
>  3 files changed, 9 insertions(+), 2 deletions(-)
>
> diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
> index 036301e..1e441e5 100644
> --- a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
> +++ b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
> @@ -195,6 +195,7 @@ static void *amdgpu_bo_map(struct pb_buffer *buf,
>                 return NULL;
>              }
>           }
> +         bo->ws->buffer_wait_time_avg = (3 * bo->ws->buffer_wait_time_avg) / 4;
>        } else {
>           uint64_t time = os_time_get_nano();
>
> @@ -222,6 +223,7 @@ static void *amdgpu_bo_map(struct pb_buffer *buf,
>           }
>
>           bo->ws->buffer_wait_time += os_time_get_nano() - time;
> +         bo->ws->buffer_wait_time_avg = (3 * bo->ws->buffer_wait_time_avg + os_time_get_nano() - time) / 4;
>        }
>     }
>
> diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
> index 3ea0f3d..a9af0ce 100644
> --- a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
> +++ b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
> @@ -201,12 +201,16 @@ amdgpu_ctx_query_reset_status(struct radeon_winsys_ctx *rwctx)
>  static bool amdgpu_get_new_ib(struct radeon_winsys *ws, struct amdgpu_ib *ib,
>                                struct amdgpu_cs_ib_info *info, unsigned ib_type)
>  {
> +   unsigned buffer_size = 128 * 1024 * 4;
> +   unsigned ib_size = 32 * 1024 * 4;
> +
>     /* Small IBs are better than big IBs, because the GPU goes idle quicker
>      * and there is less waiting for buffers and fences. Proof:
>      *   http://www.phoronix.com/scan.php?page=article&item=mesa-111-si&num=1
>      */
> -   unsigned buffer_size = 128 * 1024 * 4;
> -   unsigned ib_size = 20 * 1024 * 4;
> +   uint64_t avg = ((struct amdgpu_winsys *)ws)->buffer_wait_time_avg;
> +   if (avg > 1E4)
> +       ib_size = 10 * 1024 * 4;

Some comment here wouldn't hurt. Also that comparison could use an
integer constant. (1e4 is double I think)

I like the idea.

Marek