Commit 361f362258 ("dri: Unify createImage and createImageWithModifiers")
has introduced new behavior for drivers which don't support explicit
format modifiers. Before this commit, INVALID was not special-cased
and any call to dri_create_image() with one or more modifiers returned
NULL. After this commit, INVALID gained a special meaning: it indicates
that the implicit modifier is accepted by the caller. This is surprising
and is an API break.
This causes further API breaks: for instance, before this commit a BO
created via gbm_bo_create_with_modifiers() was guaranteed to always
return a non-INVALID modifier in gbm_bo_get_modifier().
This is inconsistent with gbm_dri_surface_create(): that function
treats INVALID as a bad entry in the modifier list, and fails if
it's the only acceptable modifier.
Additionally, drivers don't special-case INVALID and just ignore it
if they see it in a modifier list. This causes more inconsistencies.
For instance, let's say that a library user passes the modifier list
{ INVALID, FOO } to GBM. If a driver supports explicit modifiers and
doesn't support FOO for scanout, it'll return NULL. If a driver
doesn't support explicit modifiers, the current logic would return
a non-NULL BO with an INVALID modifier. This discrepency makes it
harder to reason about the system: half of the API ignores INVALID,
while the other half assumes INVALID indicates an implicit modifier.
To fix these issues, revert to the behavior before the commit, and
require use of the dedicated API without supplying any modifier for
implicit modifiers.
Signed-off-by: Simon Ser <contact@emersion.fr>
Fixes: 361f362258 ("dri: Unify createImage and createImageWithModifiers")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32396>
We never split live ranges, so we don't need to store the location of
each live value when recording live outs, but the physreg assigned to a
register will still be clobbered when we reload it so we have to record
the original physreg and then make sure to use it when reloading the
live out.
We probably never encountered a case where we needed to reload live outs
in a loop before, but after enabling clustered subgroup reductions
dEQP-VK.subgroups.clustered.compute.subgroupclusteredmin_{i,u}64vec4_requiredsubgroupsize
hits this case and fails in RA validation without this fix.
Fixes: fa22b0901a ("ir3/ra: Add specialized shared register RA/spilling")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32435>
If the caller passed in the same DRM file description, use it for sws->fd
as well. This is simpler than the previously reverted commit and also
fixes https://gitlab.freedesktop.org/mesa/mesa/-/issues/12208.
v2:
* Move fallback sws->fd assignment to proper scope, fixes CI failures.
* Remove close(sws->fd) from amdgpu_winsys_create failure path, it can
never be a valid file descriptor != aws->fd there.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32377>
If the condition of the loop terminator is based on an unsigned value we
can in some cases find the max number of possible loop trips. With the
max loop trips know a complex unroll can unroll the loop.
For example:
uniform uint x;
uint i = x;
while (true) {
if (i >= 4)
break;
i += 6;
}
The above loop can be unrolled even though we don't know the initial
value of the induction variable because it can have at most 1 iteration.
There were no changes with my shader-db collection. Change was inspired
by MR #31312 where builtin shader code failed to unroll.
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31701>
This is mostly adapted from radv's BVH building. This defines a common
"IR" for BVH trees, two algorithms for constructing it, and a callback
that the driver implements for encoding. The framework takes care of
parallelizing the different passes, so the driver just has to split the
encoding process into "stages" and implement just one part for each
stage.
The runtime changes are:
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Konstantin Seurer <konstantin.seurer@gmail.com>
The radv changes are;
Reviewed-by: Friedrich Vock <friedrich.vock@gmx.de>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31433>
This is mostly adapted from radv's BVH building. This defines a common
"IR" for BVH trees, two algorithms for constructing it, and a callback
that the driver implements for encoding. The framework takes care of
parallelizing the different passes, so the driver just has to split the
encoding process into "stages" and implement just one part for each
stage.
The runtime changes are:
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Konstantin Seurer <konstantin.seurer@gmail.com>
The radv changes are;
Reviewed-by: Friedrich Vock <friedrich.vock@gmx.de>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31433>
All of these are functions that could reasonably be incorporated into a
Vulkan extension, but are currently missing. While we could in theory do
BVH building without them, using them simplifies the code significantly
and both radv and turnip can reasonably implement them.
Reviewed-by: Konstantin Seurer <konstantin.seurer@gmail.com>
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31433>
All functions that used to take an ir3_block as argument to append
instructions to now take an ir3_builder as argument.
Add an ir3_builder field to ir3_context and replace all uses of
ir3_context::block for creating instructions with ir3_context::build.
Signed-off-by: Job Noorman <jnoorman@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32088>
During instruction selection, instructions are sometimes emitted to
blocks other than the current one. For example, to predecessor blocks
for phi sources or to the first block for inputs. For those cases, a new
builder is created to emit at the end of the target block. However, if
the target block happens to be the same as the current block, the main
builder would not be updated to point past the new instructions.
Therefore, don't update the cursor when it points to the end of a block
to ensure that new instructions will always be added at the end.
Signed-off-by: Job Noorman <jnoorman@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32088>
Use box with largest ray interval for shadow rays (terminate on first
hit) as it maximizes the probability of finding some object in that box;
for reflection (closest hit) rays, use midpoint instead, which defers
processing of larger boxes the ray origin is in in favor of smaller
boxes closer to origin.
Since the sorting mode must be uniform, when terminate_on_first_hit flag
is divergent, we leave it as closest.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32416>
Alyssa Rosenzweig