zink: fix typo that's definitely not at all embarrassing or anything like that

and also adjust some formatting to pad out the diff and really make sure nobody
notices that anything was ever amiss here

Fixes: 787412b7eb8 ("zink: break out region overlap testing function into helper")

Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Joshua Ashton <joshua@froggi.es>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11279>

lavapipe: use cso caching

really this only needs cso_cache api, but it's cumbersome to redo all
the hashing and state creation, so just use the whole context

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11071>

aux/cso: split cso_destroy_context into unbind and a destroy functions

this makes the unbind function reusable

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11071>

aux/cso: add flag to disable vbuf

it may be known in advance that vbuf shouldn't be used, so allow users
to skip this in case caps would otherwise enable it

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11071>

zink: clamp PIPE_CAP_MAX_VIEWPORTS to PIPE_MAX_VIEWPORTS

this is all we need

Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11322>

zink: do not fail when EXT_calibrated_timestamps is unavailable

Fixes: 039078fe ("zink: slight refactor of load_device_extensions()")

Acked-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Reviewed-by: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11314>

zink: use scissor region for discarding clears during blit

ensure that clears aren't being mistakenly discarded or applied due to
scissor region being ignored and full surface geometry being used

Fixes: a8e047e8f4a ("zink: discard pending clears during blit/copy if we'll overwrite the data")

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11307>

ci/lava: Avoid tee as it ruins exit status

I was today years old when I learned this about classic composable UNIX
tools:
    ~/mesa/mesa lava-submitter-overlay * % bash
    [daniels@strictly mesa]$ set -e
    [daniels@strictly mesa]$ false | tee
    [daniels@strictly mesa]$ echo $?
    0

Use tail rather than tee, so it doesn't hide our exit status.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/lava: Dump and artifact YAML again

Now it's safe to do so without leaking JWTs, dump the generated YAML to
make it easier to reproduce things.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/lava: Disable stdout/stderr buffering

Frequency of writes is unlikely to be a performance bottleneck, and
given the number of steps in between execution and the user, less
buffering is gooder.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/lava: Add explicit fatal-error handler

Truth is relative in 2021, and Python's duck-typing means truthiness
isn't what you think it is. Use an explicit fatal-error handler to make
sure we crash out hard on failure, rather than hoping sys.exit() behaves
like you think it does, because it doesn't.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/lava: Remove unused arguments

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/lava: Generate job name from lava-submit.sh

Just use the CI job name rather than open-coding the parameters.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/panfrost: Remove useless variable

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/lava: Use per-job rootfs overlay for environment

Trying to get arbitrary strings suitably quoted for shell, embedded in a
YAML file, processed by Python templating, is like seven bad ideas all
embedded into one big can of bees.

Reuse the same script we use for bare-metal to generate the environment,
tar that up into a per-job overlay which is added to the
inter-pipeline-reusable rootfs built by the container jobs and the
intra-pipeline-reusable overlay built by the build jobs.

@anholt wrote a chunk of this - replacing the $ENV_VARS GitLab CI
variable with a Python loop across the POSIX job environment - in
!11192, but this still had YAML quoting nightmares, and was more
needless duplication between LAVA and bare-metal.

The diff is large and annoying, but is mostly a sed job to get
ENV_VARS="FOO=bar BAZ=quux" into FOO: bar\nBAZ: quux.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Co-authored-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci: Use JOB_ARTIFACTS_BASE for Piglit fails

It's not Piglit-specific per se, it's just another per-job artifact
upload which needs to be made visible through MinIO.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci: Add JOB_ARTIFACTS_BASE variable

Used for both LAVA (uploading results to MinIO because we don't yet have
non-ephemeral NFS storage) and Piglit (for the Tracie dashboard).

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci: Make PIPELINE_ARTIFACTS_BASE a common variable

$minio/artifacts/$project/$pipeline/ is common between all our CI.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/lava: Clean up variable naming, document them

Our variable names haven't aged very well. Rename them to make them more
clear and straightforward, especially when we bring in a third rootfs
element to download.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/lava: Wrap submission in a shell script

Just do what we're already doing but in a shell script, which will make
it less tedious to expand out later.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/lava: Cosmetic reordering of job init

Split our init up into: base system setup (filesystem mounts, network),
pulling the build artifacts, environment common to us and bare-metal,
bespoke environment, and finally running the tests.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/lava: Pass JWT separately from environment variables

As the JWT is sensitive, we don't want to record or leak it anywhere.
Doing this lets us run --dump-yaml in normal execution so we can
artifact the result, as well as bringing us into line with bare-metal.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/lava: Move LAVA files to lava/

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/bare-metal: Don't leak JWT into logs

The JWT is sensitive - as it can be used to access e.g. private traces -
so we don't want it anywhere in our logs.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

ci/bare-metal: Factor out environment to a separate script

This will let us reuse the same environment generation for both
bare-metal and LAVA.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11309>

meson: kmsro: require dri3 for X11

The current implementation in kmsro relies on buffer sharing using
WINSYS_HANDLE_TYPE_FD, which in x11 is only used by default when dri3
is enabled.
Since the current implementation will not work without it, we can
prevent user error by checking that it is not disabled at configuration
time.

Closes #4861

Signed-off-by: Erico Nunes <nunes.erico@gmail.com>
Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Eric Engestrom <eric@engestrom.ch>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11305>

aco/ra: refactor register assignment for vector operands

No functional changes.

Reviewed-by: Tony Wasserka <tony.wasserka@gmx.de>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8764>

aco/ra: refactor affinity coalescing

Also adds v_interp_p2_f32 to the list of
affinity-related instructions.

Totals from 68 (0.05% of 149839) affected shaders (GFX10.3):
CodeSize: 792928 -> 792056 (-0.11%)
Instrs: 152843 -> 152625 (-0.14%)
Latency: 1235353 -> 1235278 (-0.01%)
InvThroughput: 224087 -> 224049 (-0.02%)
Copies: 9218 -> 9000 (-2.36%)

Reviewed-by: Tony Wasserka <tony.wasserka@gmx.de>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8764>

aco/ra: only create phi-affinities for killed operands

If a phi-operand is not killed, it must be copied anyway.
The additional affinity would only overwrite any potential
better affinity that was already created

Totals from 1067 (0.71% of 149839) affected shaders (GFX10.3):
VGPRs: 68072 -> 68064 (-0.01%)
CodeSize: 8252588 -> 8245220 (-0.09%); split: -0.12%, +0.03%
Instrs: 1596146 -> 1593941 (-0.14%); split: -0.16%, +0.02%
Latency: 18828176 -> 18823914 (-0.02%); split: -0.08%, +0.06%
InvThroughput: 3575063 -> 3574787 (-0.01%); split: -0.05%, +0.04%
VClause: 24345 -> 24325 (-0.08%); split: -0.16%, +0.07%
Copies: 88712 -> 87398 (-1.48%); split: -1.77%, +0.29%
Branches: 52067 -> 51364 (-1.35%); split: -1.38%, +0.03%

Reviewed-by: Tony Wasserka <tony.wasserka@gmx.de>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8764>

zink: limit non-extension version feature to spirv 1.5

In order to use the Vulkan 1.2 core viewport and layer shader outputs,
we need to use SPIR-V 1.5. But we've recently added some preliminary
support to compute the SPIR-V version we're using, with the intention of
adding a SPIR-V version override to work around bugs in tools like
RenderDoc. We haven't implemented the latter yet.

But just to be safe, let's limit this to SPIR-V 1.5. This isn't going to
matter right now, but it might avoid a problem if we decide to finish up
the SPIR-V version overriding.

Reviewed-by: Hoe Hao Cheng <haochengho12907@gmail.com>
Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11259>

zink: reject more illegal blits

Vulkan has some additional restrictions for vkCmdBlitImage that we
weren't testing for. Quting the Vulkan 1.2 spec, section 20.5 "Image
Copies with Scaling", "Valid Usage" subsection:

- If either of srcImage or dstImage was created with a signed integer
  VkFormat, the other must also have been created with a signed integer
  VkFormat
- If either of srcImage or dstImage was created with an unsigned integer
  VkFormat, the other must also have been created with an unsigned
  integer VkFormat.

So let's make sure we reject these illegal blits.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Reviewed-by: Hoe Hao Cheng <haochengho12907@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11217>

zink/ci: re-enable test

Since f34ff037027 ("zink/ci: increase piglit and deqp-runner timeouts"),
we're no longer dangerously close to timing out this test, so it should
be safe to re-enable this.

Acked-by: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11288>

llvmpipe: do not always use pixel-rounded coordinates for points

LLVMpipe always used the bounding-box to rasterize-points, rather than
the actual rasterization-planes. This happened because the primitive was
expanded by one unit outside the bounding box. While this kinda work for
non-multisampled cases, it's not really quite *correct*.

Rasterization of non-legacy points in OpenGL is defined as the
intersection of a the pixel centers with a rectangle of size width and
height, centered around the point in viewport coordinates. This applies
both to multi-sampled and non-multisampled cases.

So let's fix the rasterizer to use the correct definition in both cases.

We leave the legacy case as-is, and just do the inverse adjustment
there so the end result should be the same.

This fixes the following dEQP test-cases:
- dEQP-GLES3.functional.rasterization.fbo.rbo_multisample_4.primitives.points
- dEQP-GLES3.functional.rasterization.fbo.rbo_multisample_max.primitives.points

...as well as this one for Lavapipe:
- dEQP-VK.rasterization.primitives_multisample_4_bit.no_stipple.points

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11183>

llvmpipe: fix multisample lines again

This does a little bit better than what we did in 2c0a078fdb4
("llvmpipe: fix multisample lines."), where parts of the diamond-exit
rule stuff was bypassed. But we should actually bypass *all* of the
diamond-exit rule stuff here instead.

The reason is that multisampled lines have a completely differently
specified set of rasterization rules, as per the OpenGL 4.6 core spec,
section 14.5.4 ("Line Multisample Rasterization").

So let's give multisampled lines their own geometry-generation codepath
instead.

This fixes the following dEQP tests:
- dEQP-GLES3.functional.rasterization.fbo.rbo_multisample_4.primitives.lines
- dEQP-GLES3.functional.rasterization.fbo.rbo_multisample_max.primitives.lines

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11183>

llvmpipe: consistently deal with post-rast state

There's no good reason why we peek into the rasterization state when
dealing with the point_quad_rasterization state, rather than set it
through lp_setup_set_point_state like other point-state.

Let's fix this up, and get rid of a needless NULL-check per primitive.

This makes the code a bit easier to read as well, and will help once
these conditions gets more complicated later on.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11183>

llvmpipe: fix edge-rule logic for lines

In 2737abb44ef, the handling of pixel-offsets and edge rules were
untangled, but one case was missed.

This fixes the following dEQP test-cases on VirGL + LLVMpipe
- dEQP-GLES2.functional.draw.random.10
- dEQP-GLES2.functional.draw.random.42
- dEQP-GLES3.functional.draw.random.105
- dEQP-GLES3.functional.draw.random.114
- dEQP-GLES3.functional.draw.random.135
- dEQP-GLES3.functional.draw.random.144
- dEQP-GLES3.functional.draw.random.155
- dEQP-GLES3.functional.draw.random.174
- dEQP-GLES3.functional.draw.random.206
- dEQP-GLES3.functional.draw.random.31
- dEQP-GLES3.functional.draw.random.43
- dEQP-GLES3.functional.draw.random.84
- dEQP-GLES31.functional.draw_indirect.random.20

...as well as these on Zink + Lavapipe:
- spec@nv_primitive_restart@primitive-restart-disable_vbo
- spec@nv_primitive_restart@primitive-restart-vbo_combined_vertex_and_index
- spec@nv_primitive_restart@primitive-restart-vbo_index_only
- spec@nv_primitive_restart@primitive-restart-vbo_separate_vertex_and_index
- spec@nv_primitive_restart@primitive-restart-vbo_vertex_only

Fixes: 2737abb44ef ("gallium: Replace gl_rasterization_rules with lower_left_origin and half_pixel_center.")
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11183>

virgl: forward the host renderer hardware info

Some game engines rely on the real hardware info to adjust default
graphics quality and other attributes.

Prepend "virgl" to avoid app compat issues and to distinguish from
native platforms while giving engines/apps a chance to adjust graphics
defaults.

Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11179>

ac: Enable 32bit predication on gfx9 with fw feature version 52.

Amdvlk does this as well and it passes the vulkan CTS on renoir.

Signed-off-by: Georg Lehmann <dadschoorse@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11297>

ac: Enable 32bit predication on gfx10.

Signed-off-by: Georg Lehmann <dadschoorse@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11297>

ac: Check me_fw_feature for 32bit predication on gfx10.3

Signed-off-by: Georg Lehmann <dadschoorse@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11297>

radv: fix aligning the image offset by using align64()

This doesn't fix anything known. Found by inspection.

Cc: 21.1 mesa-stable
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11302>

util/disk_cache: add nocopy variant of disk cache store function

this is a bit more convenient in some cases

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11304>

egl/android: add aosp_nougat system/window.h back for back compat

Also layer ANativeWindow_* APIs on top of legacy APIs for api level less
than 26 in a new platform_android.h header.

v2: persist frozen system/window.h header

Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Rob Clark <robdclark@chromium.org> (v1)
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11286>

docs: Update freedreno features

Bring features.txt a bit more up to date.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11311>

freedreno: Add missing valid range tracking for SSBOs/images

Normally TC takes care of this for us.  But we might as well not get it
wrong in cases where TC is disabled.

Reported-by: Alyssa Rosenzweig <alyssa@collabora.com>
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11311>

freedreno/registers: add A5XX_RBBM_STATUS3 bit

Same bit as a6xx.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11311>

ci/piglit: Skip WGL on all the Linux runs.

We don't build the tests, since we don't have WGL.

Reviewed-by: Juan A. Suarez <jasuarez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11283>

ci/freedreno: Enable running all of piglit_gl for a530's manual test.

Otherwise the xfails will end up stale after piglit uprevs that change the
test set.

Reviewed-by: Juan A. Suarez <jasuarez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11283>

ci: Update piglit and deqp/piglit-runner.

This brings in some major new features in the runner:

- piglit tests now include subtest reporting
- "-t" support for quick include-filtering of tests.
- piglit tests that crash after their result report are considered crashes.
- throws a nice error if you try to annotate the same failure twice
  (e.g. lvp's dEQP-VK.glsl.builtin.precision.pow.highp.vec2,Fail)

Since the runner catches piglit test bugs where the same subtest is run
twice, we also uprev piglit to pull in the fixes for those.

Reviewed-by: Juan A. Suarez <jasuarez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11283>

ci/lava: Don't overwrite PIGLIT_REPLAY_EXTRA_ARGS

Other Piglit jobs will want different values for this env var.

Signed-off-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11293>

ci/lava: Add --dump-yaml option to submitter

Also useful for local development and testing.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11293>

ci/lava: Add validate-only mode to job submitter

Useful for development.

Signed-off-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11293>

ci/lava: Improve error reporting in lava_job_submitter.py

I'm having trouble figuring out why this is breaking.

Signed-off-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11293>

ci/lava: Finish garbage-collecting the TEST_SUITE variable

We no longer name the template by the test suite being run.

Fixes: 93ec399b2850 ("ci: Use a single template for LAVA jobs")
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Acked-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11293>

iris: don't munmap NULL pointers

This is a regression, the previous commit had this check which was
removed in the patch mentioned below. What happens is that when we
have a buffer that's not mmapped and we try to bo_free it we get some
very funny backtraces. Easily reproducible with fullscreen
gputest.triangle.

Fixes: f62724ccacff ("iris: Pick a single mmap mode (WB/WC) at BO allocation time")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/4890
Tested-by: Michel Dänzer <mdaenzer@redhat.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11284>

iris: drop unused function declaration

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11306>

zink/codegen: clean the constructor of Extension up

the `functions` parameter is now unused, also rewrote some comments
since they are now outdated

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11036>

zink/codegen: allow conditional enabling of instance extensions

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11036>

zink: use the dispatch tables

- removed all usage of GET_PROC_ADDR
- find-and-replaced all instances of `screen->vk_` to `screen->vk.`

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11036>

zink: slight refactor of load_device_extensions()

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11036>

zink/codegen: add zink_verify_*_extensions()

those are meant to be used with the dispatch tables, by checking whether
the functions added by the enabled extensions are actually loaded

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11036>

zink/codegen: split commands into three groups

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11036>

zink: introduce vk_dispatch_table

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11036>

ir3: Copy propagate immed/const to meta instructions

This is allowed with the new RA, and makes a huge difference in
preventing extra moves when preferential coloring doesn't work.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Insert output collects in the main shader

We were inserting them in what was NIR's end block with the "end"
instruction, which meant that the moves they generated couldn't be
scheduled with the rest of the last block as part of post-RA scheduling.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Add simple CSE pass

RA currently can't handle a live value that's part of a vector and
introduces extra copies. This was espeically a problem for bary.f, where
the bary coords were being split and repeatedly re-collected. But this
could be a problem in other situations as well.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3/sched: Consider unused destinations when computing live effect

If an instruction's destination is unused, then we shouldn't penalize
it. For example, this helps us schedule atomic operations whose results
aren't read. This works around RA failures when CSE is enabled in some
robustness2 tests.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3/sched: Make collects count against tex/sfu limits

In a scenario where there are a lot of texture fetches with constant
coordinates, this prevents the scheduler from scheduling all the setup
instructions after the first group of textures has been scheduled
because they are the only non-syncing thing and scheduling them didn't
decrease tex_delay. Collects with immed/const sources will turn into
moves of those sources, so we should treat them the same.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3/sched: Don't schedule collect early

I don't think there was ever a good reason to do this, but when we start
folding constants/immediates into collect, this can become actively
harmful.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Remove right and left copy prop restrictions

This is leftover from the old RA, and inhibits copy propagation
unnecessarily with the new RA.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3/ra: Add a validation pass

This helps catch tricky-to-debug bugs in RA, or helps rule them out.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Rewrite register allocation

Switch to the new SSA-based register allocator.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Expose occupancy calculation functions

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Add pass to lower arrays to SSA

This will be run right after nir->ir3. Even though we have SSA coming
out of NIR, we still need it for NIR registers, even though we keep the
original array around to insert false dependencies.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Add dominance infrastructure

Mostly lifted from nir.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Remove unused check_src_cond()

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3/postsched: Don't use SSA source information

This was only used for calculating if a source is a tex or SFU
instruction, which is easily replacable. It's going away with the new
RA.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3/delay: Delete pre-RA repeat handling

It looks likely that any implementation of (rptN) in ir3 will have to
actually create (rptN) instructions after RA, which means that this can
be dropped.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Rewrite delay calculation

The old delay calculation relied on the SSA information staying around,
and wouldn't work once we start introducing phi nodes and making
"normal" values defined in multiple blocks not array regs anymore.
What's worse is that properly inserting phi nodes when splitting live
ranges would make that code even more complicated, and this was the last
place post-RA that actually needed that information.

The new version only compares the physical registers of sources and
destinations. It works by going backwards up to a maximum number of
cycles, so it might be slightly slower when the definition is closer but
should be faster when it is farther away.

To avoid complicating the new method, the old method is kept around, but
only for pre-RA scheduling and it can therefore be drastically
simplified as the array case can be dropped.

ir3_delay_calc() is split into a few variants to avoid an explosion of
boolean arguments in users, especially now that merged_regs now has to
be passed to it.

The new method is a little more complicated when it comes to handling
(rptN), because both the assigner and consumer may be (rptN). This adds
some unit tests for those cases, in addition to dropping the to-SSA code
in the test harness since it's no longer needed.

Finally, ir3_legalize has to be switched to using physical registers for
the branch condition. This was the one place where IR3_REG_SSA remained
after RA.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Make branch conditions non-SSA

In particular, make sure they have a physreg assigned. This was the last
place after RA where SSA registers were created, which won't work with
the new post-RA delay calculation that relies on the physreg.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Add reg_elems(), reg_elem_size(), and reg_size()

For working with registers in units of half-regs in the new RA.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3/delay: Fix full->half and half->full delay

The current compiler never does this, but the new compiler will start to
in mergeregs mode. There is an extra penalty for this.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Add ir3_register::array.base

There were two different approaches I saw in the post-RA code for
figuring out what regiser range a relative access touched:

1. Use reg->array.offset and reg->array.size. This is wrong in case
   reg->array.offset was non-zero before RA, because array.size is
   the size of the whole array and array.offset has the const offset
   within the array baked in.
2. Lookup the array from the array ID and use the base + range there.
   This is correct, but won't work with the new RA, where an array might
   not always be assigned to the same register.

This replaces both methods with a new ir3_register::array.base field,
and switches all the users I could find to it.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Improve register printing for SSA

Print the ssa name for array destinations, and handle printing undef SSA
sources.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Prepare for instructions with multiple destinations

To simplify the pre-RA merge set code and express the result live-range
splitting in RA, we need to add support for parallel copy instructions,
and for the merge set code these parallel copies need to be in SSA form.
Parallel copies have multiple destinations by necessity, but there was
no way to express this in the existing IR. In particular there was no
support for marking a register as being a destination, and no support
for indicating which destination register out of several an SSA source
refers to. This replaces ir3_register::instr with ir3_register::def and
re-purposes ir3_register::instr. I haven't propagated this into common
helpers, like ssa(), because that would vastly increase the amount of
churn and the number of places that produce such instructions should be
limited -- only RA will create parallel copies and they will be
destroyed right after RA. In the future swz will have multiple
destinations too, but it will only be created after RA via parallel copy
lowering.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Readd support for translating NIR phi nodes

This is roughly based on the support removed a while ago, but it handles
sources better by associating each source with a predecessor block.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Add ir3_start_block()

Name based on nir_start_block(). A number of places were already
open-coding this, convert them.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

ir3: Introduce phi and parallelcopy instructions

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>

docs/panfrost: Update API versions

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>

docs/features: Mark GLES3.1 as done on Panfrost

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>

panfrost/ci: Do fractional dEQP-GLES31 run on Midgard

Drop the skip list and correspondingly populate the fails list.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>

panfrost/ci: Don't skip SSBO tests on G52

These were blocked on failing RA, but that's been resolved now.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>

panfrost/ci: Blank G52 flakes file

Haven't seen these tests flake, and we don't even run dEQP-GLES2 on G52
in CI anymore. (I still do local runs, and I don't see them flake
there.)

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>

pan/decode: Handle cache flush jobs

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>

pan/decode: Fix image attribute counting

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>

panfrost: Advertise GLES3.1

We have CI, we're just a few tests away from conformance on v7, and
Midgard is just a few hundred tests behind. Given the branch point isn't
for another month, I think this is a good time to flip the switch.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>

panfrost: Add "Cache Flush" job XML

Likely useful for efficient memory_barrier and texture_barrier
operations.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>

panfrost: Set vertex job_barrier

Fixes KHR-GLES31.core.vertex_attrib_binding.advanced-iterations which
pingpongs XFB/attributes

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>

panfrost: Flush before compute jobs

Suboptimal but fixes KHR-GLES31.core.compute_shader.pipeline-post-xfb,
which is stubbornly still broken with memory barriers implemented and
cache flush jobs inserted. More investigation needed but probably not
right now.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>

panfrost: Flush everything for glMemoryBarrier

This is inefficient but so far I see the DDK doing the same thing. Fixes
KHR-GLES31.core.shader_storage_buffer_object.advanced-usage-sync-vsfs

In the future we should look into cache flush jobs.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11123>