Hi,
This is my second attempt to fix the mapping bug adding all the
suggestions that Tom Stellard sent, and, so far, it seems that
it is resolved.
This series changes completely how OpenCL buffers are handled
by the r600g driver. Before this, we would add them directly to
a pool, and this pool would grow whenever we needed more space.
But this process implied destroying the pool and creating a new
one. There could be cases where a buffer would be mapped and
the pool would grow, leaving one side of the mapping pointed
to where the item was. This is the 'mapping bug'
Now, Items will have an intermediate resource, where all mappings
can be done, and when a buffer is going to be used with a kernel
it is promoted to the pool. In the case where a promoted item
is going to be mapped, it is previously demoted, so even if
the pool changes its location due to growing, the map remains
valid. In the case of a buffer mapped for reading, and used
by a kernel to read from it, we will duplicate this buffer,
having the intermediate buffer, where the user has its map, and
an item in the pool, which is the one that the kernel is going
to use.
As a summary for v2:
Patches 1-8: These are the main part of the series, and solve
the mapping bug.
Patches 1 and 7 now use less explicit castings
Patch 2 is new and introduces the 'is_item_in_pool'
function, which is used in patches 3 and 8
Patch 9: Is a complete rewrite of v1 patch 8 using gallium
utils for double lists
Patches 10 and 11: These are just a proof of concept for avoiding
transfers GPU <-> GPU when using all CL Read/Write functions.
They are v1 patch 9 splited in two to separate r600g changes
from clover changes.
Now, in clover's side it introduces and uses
'CLOVER_TRANSFER_MAP_DIRECTLY' so it doesen't collide with
any other OpenCL flag.
Please review and Thanks :)
Bruno Jiménez (11):
r600g/compute: Add an intermediate resource for OpenCL buffers
r600g/compute: Add an util function to know if an item is in the pool
r600g/compute: Add statuses to the compute_memory_items
r600g/compute: divide the item list in two
r600g/compute: Only move to the pool the buffers marked for promoting
r600g/compute: Avoid problems when promoting items mapped for reading
r600g/compute: Implement compute_memory_demote_item
r600g/compute: Map only against intermediate buffers
r600g/compute: Use gallium util functions for double lists
r600g/compute: Map directly the pool in some cases
clover: Use PIPE_TRANSFER_MAP_DIRECTLY when writing/reading buffers
src/gallium/drivers/r600/compute_memory_pool.c | 294 ++++++++++++---------
src/gallium/drivers/r600/compute_memory_pool.h | 31 ++-
src/gallium/drivers/r600/evergreen_compute.c | 38 ++-
src/gallium/state_trackers/clover/api/transfer.cpp | 4 +-
src/gallium/state_trackers/clover/core/object.hpp | 4 +
.../state_trackers/clover/core/resource.cpp | 2 +
6 files changed, 233 insertions(+), 140 deletions(-)
--
2.0.0
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
