if a resource is in use every frame and never goes idle, it becomes
impossible to execute pruning, as there is no tracking for when views
are no longer in use
to avoid eventually ooming in this scenario, add some data to zink_resource_object
which can effectively "queue" pruning of these views if ballooning is
detected at a time when the views are guaranteed to be safe to delete
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19056>
the existing model for descriptor pools works thusly:
* allocate pool
* when possible, reuse existing pool from overflow array
* when pool is full, append to overflow array
* on batch reset, cycle the overflow arrays to enable reuse
the problem with this is it uses two separate arrays for overflow, and these arrays
are cycled but never consolidated, leading to wasted memory allocations for whichever
array is not currently being recycled
to resolve this, make the following changes to batch resets:
* set the recycle index to whichever overflow array is smaller
* copy the elements from the smaller array to the larger array (consolidate)
* set the number of elements in the smaller array to 0
this leads to always appending to an empty array while the array containing all the
"free" pools is always the one that's available for recycled use
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19053>
this mechanism worked off the previous iteration of descriptor updating,
in which pools were stored in a set to the batch state and could be iterated
normally
now, however, they're stored as a sparse array, and so the dynarray util for getting
the number of elements cannot be used
instead, use the calculated size for the array like every other accessor for
the pools to ensure correct indexing
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19053>
in the case where a new descriptor pool cannot be allocated naturally:
* iterate free batch states and try to free up memory
* iterate in-use batch states and try to free up memory
* ???
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19053>
With the latest header update from Khronos, KHRONOS_APICALL (which is then
defined as EGLAPI) is empty, preventing the API symbols to be visible.
This commit adds `PUBLIC` to all the symbols from the EGL API that needs to be
visible.
Signed-off-by: Simon Zeni <simon@bl4ckb0ne.ca>
Reviewed-by: Eric Engestrom <eric@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18848>
Move the assignment of draw->start_index into the num_draws loop
in draw_pt_arrays() so that it's updated for each primitive we're
drawing. Previously, it was only set once to the 0th primitive's
start index.
This fixes incorrect gl_BaseVertex values when using
vkCmdDrawMultiEXT() with more than one VkMultiDrawInfoEXT item.
Signed-off-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19039>
The vertex_id_offset value needs to be used both for the elt and
linear cases. As it was, if a long primitive was split, the vertex id
seen in the VS was also getting to reset to zero for each batch of the
split prim's vertices.
This fixes at least five VMware test cases (dx11-dxsdk-FluidCS11,
dx11-amd-TressFX-v1.0, dx10-win7-dxsdk-samples-nbodygravity,
dx10-win7-dxsdk-samples-gpuboids, dx10-wide-points).
Signed-off-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19039>
this already has locks for async gfx precompiles, but compute shaders
can be compiled on one context and bound on another, which means
all the descriptor stuff they allocate has to be portable:
specifically the pool key ids MUST match across contexts
this is trivially achieved by making pool keys into screen objects
fixes#7434
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19038>
SPIR-V and Vulkan allow implementations to optimize
float ops assuming inputs are not NaN unless SignedZeroInfNanPreserve
execution mode is enabled.
This means isnan which is turned to nir_op_fneu(a,a) and then emitted
as SpvOpFUnordNotEqual could potentially be optimised to a const 0.
isnan is special in this approach and therefore recognise this pattern,
and emit SpvOpIsNan instead of enabling SignedZeroInfNanPreserve which
will hurt performace of some vulkan drivers.
Fixes dEQP-GL45-ES31.functional.shaders.builtin_functions.common.isnan.*
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18910>
Offset decoration in spirv is unsigned and it does not have a notion of an invalid offset.
Unlike NIR which set -1 for invlaid offset. This translates to invalid spirv being produced.
Instead, just don't emit an Offset decoration.
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18910>
Mike Blumenkrantz