The documentation says that if we don't use force_delta, the LOD will be
-infinity for non-active lanes before bias and clamp are applied. This
is not what we want, so let's instead assume all threads are active, and
let helper-invocations do their job to compute correct values.
While this is only needed for the second iteration, let's just leave it
on for both for simplicity.
Fixes: e317136536 ("pan/va: Add support for nir_texop_lod")
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33069>
Applications tend to forget to describe subpass dependencies, especially
when it comes to write -> read dependencies on attachments. The
proprietary driver forces "others" invalidation as a workaround, and this
invalidation even became implicit (done as part of the RUN_FRAGMENT) on
v13+.
We will consider adding a dri-conf hook for this option in the future,
but for now, let's just keep it as an opt-in debug flag.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Reviewed-by: Lars-Ivar Hesselberg Simonsen <lars-ivar.simonsen@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33056>
Right now, we have a problem when we flush draws inside a render pass
and we don't have enough information to re-emit the framebuffer/tiler
descriptors.
Turns out the only situations where this happens is when an occlusion
query end happens, but we shouldn't really flush the draws in that case.
What we should do instead is record the OQ in our command buffer, so we
can signal OQ availability when the fragment job is done.
In order to solve that, we add an OQ chain to the command buffer to
track OQs ending inside the render pass. We then walk this chain at
fragment job emission time to signal the syncobjs attached to each
query.
This also simplifies the whole occlusion query synchronization model:
instead of waiting for each syncobj individually, we now wait on
the iterators to make sure all OQs have landed. Thanks to this new
synchronization, we can batch OQ reset/copy operations and make the
command stream a lot shorter when big query ranges are copied/reset.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Lars-Ivar Hesselberg Simonsen <lars-ivar.simonsen@arm.com>
Reviewed-by: Benjamin Lee <benjamin.lee@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32973>
The SYSTEM scope triggers CPU interrupts we don't really need, so let's
use the CSG scope to avoid those. Note that the scope doesn't encode
the visibility aspect, meaning changes to the sync object with a CSG
scope will still be instantly visible to the CPU, it's just that the
CPU needs to poll the value to detect a change, which is basically what
we're doing for syncobjs attached to events/queries, so we're good.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Lars-Ivar Hesselberg Simonsen <lars-ivar.simonsen@arm.com>
Reviewed-by: Benjamin Lee <benjamin.lee@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32973>
The spec says
vkCmdCopyQueryPoolResults is considered to be a transfer operation,
and its writes to buffer memory must be synchronized using
VK_PIPELINE_STAGE_TRANSFER_BIT and VK_ACCESS_TRANSFER_WRITE_BIT before
using the results.
While STORE_MULTIPLE is not exactly VK_PIPELINE_STAGE_TRANSFER_BIT /
VK_ACCESS_TRANSFER_WRITE_BIT, we can still rely on user barriers to do
the right thing (e.g., flush caches for host access).
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Lars-Ivar Hesselberg Simonsen <lars-ivar.simonsen@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32697>
This feature actually worked, but accidentally triggered some
CTS failures due do an unrelated sample_mask bug. That bug
is fixed in the previous commits in this series, so re-enable
fragmentStoresAndAtomics.
There are two spurious CI failures created by this change; one is in
a ycbcr test (which we're still working on) and the other is a CTS
bug (it's using a feature we don't advertise).
Reviewed-by: Mary Guillemard <mary.guillemard@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32879>
We're about to add more sysvals, and the more we add, the bigger the
sysvals region gets, which increases the amount of memory we have to
allocate when push_uniforms are dirty.
Instead of allocating FAUs for all sysvals/push_constants, track FAU
usage per-shader, and pack those. This implies emitting an FAU buffer
per stage instead of trying to share it, but that's an acceptable
trade-off.
While at it, automate the sysval dirty tracking a bit.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Mary Guillemard <mary.guillemard@collabora.com>
Reviewed-by: Lars-Ivar Hesselberg Simonsen <lars-ivar.simonsen@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32415>
Now that panvk_cmd_buffer.h is accessible from
src/panfrost/vulkan/panvk_vX_xxx.c files, there's no reason to pass
a gazillon arguments to blend_emit_descs(). We can just pass a cmdbuf
and let the helper extract the other parameters from there. It also
allows for extra automation, like dirtying the push_uniform buffer
when the new blend config reads the blend constant.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Mary Guillemard <mary.guillemard@collabora.com>
Reviewed-by: Lars-Ivar Hesselberg Simonsen <lars-ivar.simonsen@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32415>
load_vertex_id_zero_base() is supposed to return the zero-based
vertex ID, which is then offset by load_first_vertex() to get
an absolute vertex ID. At the same time, when we're in a Vulkan
environment, load_first_vertex() also encodes the vertexOffset
passed to the indexed draw.
Midgard/Bifrost have a sligtly different semantics, where
load_first_vertex() returns vertexOffset + minVertexIdInIndexRange,
and load_vertex_id_zero_base() returns an ID that needs to be offset
by this vertexOffset + minVertexIdInIndexRange to get the absolute
vertex ID. Everything works fine as long as all the load_first_vertex()
and load_vertex_id_zero_base() calls are coming from the
load_vertex_id() lowering. But as mentioned above, that's no longer
the case in Vulkan, where gl_BaseVertexARB will be turned into
load_first_vertex() and expect a value of vertexOffset in an
indexed draw context.
We thus need to fix the mismatch by introducing two new
panfrost-specific intrinsic so we can stop abusing load_first_vertex()
and load_vertex_id_zero_base().
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Lars-Ivar Hesselberg Simonsen <lars-ivar.simonsen@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32415>
Benjamin Lee