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

[Grigori Goronzy]

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;
 
    if (ib_type == IB_CONST) {
       buffer_size = 512 * 1024 * 4;
diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.h b/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.h
index 91b9be4..56be13e 100644
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.h
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.h
@@ -54,6 +54,7 @@ struct amdgpu_winsys {
    uint64_t allocated_vram;
    uint64_t allocated_gtt;
    uint64_t buffer_wait_time; /* time spent in buffer_wait in ns */
+   uint64_t buffer_wait_time_avg;
    uint64_t num_cs_flushes;
    unsigned gart_page_size;
 
-- 
1.9.1