diff options
| author | Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com> | 2019-07-11 10:34:40 -0700 |
|---|---|---|
| committer | Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com> | 2019-07-12 15:31:47 -0700 |
| commit | 330cd057adb04a7e579e8e9ba7e3952b448928a8 (patch) | |
| tree | 0a3c664e3df97ed50626f1a689003511d427cd38 /src/gallium/drivers/panfrost/pan_screen.h | |
| parent | 718ebfa2259cb50efc05541af205842b6630cf43 (diff) | |
panfrost: Add panfrost_transient_bo array
We would like transient allocations to occur on the screen (borrowed by
the batch) rather than on the context. Add fields to track this.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Diffstat (limited to 'src/gallium/drivers/panfrost/pan_screen.h')
| -rw-r--r-- | src/gallium/drivers/panfrost/pan_screen.h | 27 |
1 files changed, 27 insertions, 0 deletions
diff --git a/src/gallium/drivers/panfrost/pan_screen.h b/src/gallium/drivers/panfrost/pan_screen.h index bf64b8b40e0..ab39da6d2a1 100644 --- a/src/gallium/drivers/panfrost/pan_screen.h +++ b/src/gallium/drivers/panfrost/pan_screen.h @@ -32,6 +32,8 @@ #include "pipe/p_screen.h" #include "pipe/p_defines.h" #include "renderonly/renderonly.h" +#include "util/u_dynarray.h" +#include "util/bitset.h" #include <panfrost-misc.h> #include "pan_allocate.h" @@ -46,6 +48,22 @@ struct panfrost_screen; #define PAN_ALLOCATE_INVISIBLE (1 << 2) #define PAN_ALLOCATE_COHERENT_LOCAL (1 << 3) +/* Transient slab size. This is a balance between fragmentation against cache + * locality and ease of bookkeeping */ + +#define TRANSIENT_SLAB_PAGES (32) /* 128kb */ +#define TRANSIENT_SLAB_SIZE (4096 * TRANSIENT_SLAB_PAGES) + +/* Maximum number of transient slabs so we don't need dynamic arrays. Most + * interesting Mali boards are 4GB RAM max, so if the entire RAM was filled + * with transient slabs, you could never exceed (4GB / TRANSIENT_SLAB_SIZE) + * allocations anyway. By capping, we can use a fixed-size bitset for tracking + * free slabs, eliminating quite a bit of complexity. We can pack the free + * state of 8 slabs into a single byte, so for 128kb transient slabs the bitset + * occupies a cheap 4kb of memory */ + +#define MAX_TRANSIENT_SLABS (1024*1024 / TRANSIENT_SLAB_PAGES) + struct panfrost_screen { struct pipe_screen base; int fd; @@ -56,6 +74,15 @@ struct panfrost_screen { /* Memory management is based on subdividing slabs with AMD's allocator */ struct pb_slabs slabs; + /* Transient memory management is based on borrowing fixed-size slabs + * off the screen (loaning them out to the batch). Dynamic array + * container of panfrost_bo */ + + struct util_dynarray transient_bo; + + /* Set of free transient BOs */ + BITSET_DECLARE(free_transient, MAX_TRANSIENT_SLABS); + /* While we're busy building up the job for frame N, the GPU is * still busy executing frame N-1. So hold a reference to * yesterjob */ |
