This lets us move fd_batch_update_queries() after resource tracking.
Which will become needed in the next patch which adds validation to
assert needed BOs are attached to the submit.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25465>
If we went the upload_user_buffer() path, cb->buffer would be null,
causing fd_dirty_shader_resource() to be a no-op. What we want to
use is &so->cb[index].
Fixes: b43e5aec0d ("freedreno/batch: Move submit bo tracking to batch")
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25465>
Having this pointer in the key is undesirable since it makes
copying keys difficult and error prone (as seen in previous
patches), also, it is only there for convenience and we don't
strictly need it (in fact the vulkan driver doesn't use it at
all), so let's just get rid of it so our v3d_key is fully
static.
Reviewed-by: Juan A. Suarez <jasuarez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25418>
The RAM shader cache was using the v3d_key for hashes and
comparisons which is not correct. Particularly, this struct
has a void pointer where we store a reference to an uncompiled
shader with the NIR code, and that is of course not accounted
for when hashing and comparing keys, which can lead to bogus
cache hits.
This patch introduces a v3d_cache_key that has both the v3d key
and a sha1 of the uncompiled NIR. Now key hashing and comparison
is done on the static part of the v3d key (that is, excluding the
uncompiled shader pointer, which may be invalid in the cache if
the original shader was deleted) and taking the sha1 from the
uncompiled shader. This also makes sure the shader key we store
in the cache has a NULL shader_state pointer to make it more
clear that this field may not be used at all for caching purposes.
This fixes GPU hangs with some OpenCL tests (through Rusticl)
caused by incorrect RAM cache hits.
Reviewed-by: Juan A. Suarez <jasuarez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25418>
To reduce shader variants, the tilebuffer lowering code does not know the
actual texture targets of the spilled render targets, only whether they are
layered or not. As such, all layered targets (3D, cube map, etc) get written out
uniformly as 2D Arrays. For that to work, the driver needs to do the
corresponding transform.
Regular imageStore() instructions are not affected by any of this.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Do not force a single-layer view, use an actual array attachment when there are
multiple layers, since this corresponds to a layered framebuffer that will write
to an array with the eMRT path.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
We can't technically expose the extension without a higher GL version, but the
implemented subset should work and this lets us test with piglit.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
If we spill render targets with a layered framebuffer, our spilled targets are
assumed to be 2D Arrays (in general). We need to use arrayed image operations to
load/store from these. The layer is given by the layer as read in the fragemnt
shader. This handles the eMRT portion of layered rendering.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Just add a flag for it. We don't care about the actual # of layers when
calculating the layout, only the boolean fact of being layered or not. The
reason we need this at all is because the eMRT implementation needs to
account for layering and that is only keyed off the tilebuffer layout.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
So we get adequate testing of continues, rather than lowering them in GLSL. We
don't really /want/ to see continues but lowering them away will just make them
harder to test... and besides, we should be optimizing them in NIR (not GLSL) so
we can get the win on Vulkan too.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
This makes for cleaner agxdecodes, I think this matches what I've seen on the
macOS side but I might be misremembering. Certainly shouldn't hurt.
This only applies for direct draws.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
In particular, removes the dependency of r600_formats.h on r600_pipe.h so
it can be shared between Gallium and Vulkan.
Reviewed-by: Gert Wollny <gert.wollny@collabora.com>
Signed-off-by: Vitaliy Triang3l Kuzmin <triang3l@yandex.ru>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24513>
The ITEM_CPY macro uses the memcpy function to copy the item variable.
When item is a pointer, the memcpy function will copy the value of the
pointer, not the address that the pointer points to.
Signed-off-by: Cong Liu <liucong2@kylinos.cn>
Reviewerd-by: Feng Jiang <jiangfeng@kylinos.cn>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25453>
We can't skip redundant register updates on compute queues, at least
not all of them.
Fixes: 315231b5a5 - radeonsi: eliminate redundant compute SH register changes
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25390>
This patch resolves the problem by modifying the for loop condition
to ensure that it stays within the bounds of the array (i.e., i < 4)
Signed-off-by: Cong Liu <liucong2@kylinos.cn>
Reviewed-by: Pavel Ondračka <pavel.ondracka@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25419>
We missed out that ICL+ added a programming requiring a CS_STALL.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25426>
Before this commit, zink_format_clamp_channel_color() ignored the format
swizzle, so it was assuming that for emulated formats like alpha,
alpha-luminance etc. that the color had already been swizzled to match
the internal format rather than the emulated one. It's somewhat confusing
that passing in e.g. A8_UNORM actually means R8_UNORM, and led to a bug
when using VK_FORMAT_A8_UNORM for texture border colors because we
didn't swizzle it back. It also wouldn't have worked for media formats
like R10X6G10X6 due to the void channel in the middle.
In order to fix this, we need to untangle the mess in its users.
For convert_color() used when clearing, this means we now need to clamp
and then swizzle instead of swizzle and then clamp, and we can drop the
hack for A8_UNORM.
For texture border colors, the state tracker duplicates colors for the
emulated formats to help drivers, which zink was previously relying on,
but fixing zink_format_clamp_channel_color() breaks this because it
assumes that those duplicated colors are useless and clamps them.
However, because we know the format we can just swizzle the border color
ourself, which convert_color() was already doing. So, we pull that out
into a common zink_convert_color() function that handles both clamping
and format emulation, and have both clearing and border color handling
use it.
This fixes A8_UNORM in turnip+zink once we enable it.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25001>
Solves regression on video players using GPU for
video decoding that just displays the video in green.
Fixes: d13da7782c ("vc4: call blit paths in chain")
Reviewed-by: Juan A. Suarez <jasuarez@igalia.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25412>
Commit 3ec953ed75 left couple of extra things, these changes are
based on a patch from Francisco Jerez.
Signed-off-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25306>
I hit a problem running a set of lvp caselists with
dEQP-VK.transform_feedback.simple.draw_indirect_endqueryindexed_streamid_0_16,Crash
This was crashing due to
assert(setup->viewport_index_slot < 0);
in try_update_scene_state
This was because a previous draw had set viewport index slot to 2,
but a clear then draw sequence never resets it, so the clear gets
a scene, and when the subsequent flush happens for that scene,
the viewport_index_slot is never updated. It only gets updated on
draws.
This just resets it as lp_setup_update_state will always pick up
the correct one for the next draw.
Cc: mesa-stable
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25387>
Even though drivers are required by documentation to handle the samplers
being NULL, st/mesa changed enough so it always sets the pointer.
As being similiar to st/mesa is more important than following docs, we
simply mimic st/mesa here and meet the expectations of drivers.
In the future we want to track the set state similiar to st/mesa, so
`clear_sampler_views` will probably go away and we'll just update what
we'll need.
Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: @LingMan <18294-LingMan@users.noreply.gitlab.freedesktop.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25363>
