pan/midgard: Handle get/set_swizzle for load/store arguments

Load/store's  main "argument 0" already has its swizzle handled
correctly (for stores, that is). But the tinier arguments, the compact
ones with a component select but not a full swizzle, those are not yet
handled. Let's do something about that!

pan/midgard: Fix block successors

Rather than an ersatz thing that sort of looks like successors but is in
fact just the source order traversal with some backward jumps hacked in
for loops... construct an actual flow graph so we can do analysis
sanely.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Add helper to pack load/store registers

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Decode register/component in load/store argument

3-bits out of 8 down!

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Fix REGISTER_OFFSET

r27 isn't the special one, usually.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Split ld/st unknown to arg_1/arg_2 fields

The 16-bit field can be decomposed to two independent 8-bit fields, each
representing a single (additional) argument to the load/store op,
generally used for encoding registers. Addressable registers here are
substantially limited compared to the main register in a load/store op.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

radv: Expose VK_KHR_imageless_framebuffer.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

radv: Implement VK_KHR_imageless_framebuffer.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

radv: Store image view also outside framebuffer.

So we can use it with imageless framebuffers.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

radv: Store color/depth surface info in attachment info instead of framebuffer.

That way we can use it for imageless framebuffers.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

panfrost: Allocate polygon lists on-demand

Rather than alloacting a huge (64MB) polygon list on context creation
and sharing it across framebuffers, we instead allocate polygon lists as
BOs (which consistently hit the cache) sized appropriately; for about a
month, we've known how to calculate the polygon list size so this has
only recently become possible.

The good news is we can render to truly massive framebuffers without
crashing and, more importantly, we eliminate the 64MB upfront overhead.
If a list that size isn't actually needed, it's not allocated.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

panfrost: Handle the bo == NULL case in panfrost_bo_[un]reference()

Allows us to pass BOs without checking if they're NULL or not.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Get rid of the skippable param in attach_vt_framebuffer()

The only user of this function always passes true.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Don't emit a new FB desc when setting a new FB state

The FB desc will be emitted/attached on the first draw targetting
this new FB.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Bail out early when doing a wallpaper blit

The wallpaper blit is a bit special in that the operation is targetting
the current FB, but the u_blitter logic creates a new surface for it
which makes util_framebuffer_state_equal() return false. In that case
we don't want a new FB descriptor to be emitted/attached, so let's just
copy the new state into ctx->pipe_framebuffer and exit the function.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Bail out early when new and current FB states are equal

If the current FB matches the new one there's nothing to be done in
panfrost_set_framebuffer_state(). By bailing out early in that case we
avoid emitting new FB descriptors (the old ones are still valid).

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Delay FB descriptor allocation

No need to emit SFBD/MFBD at frame invalidation. They can be emitted
when the framebuffer is attached, which saves us a potential FB desc
re-allocation if a new FB is bound after the swap.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Remove job from ctx->jobs at submission time

This guarantees that new draws targetting the same framebuffer will
get a new job instance.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Make ctx->job useful

ctx->job is supposed to serve as a cache to avoid an hash table lookup
everytime we access the job attached to the currently bound FB, except
it was never assigned to anything but NULL.

Fix that by adding the missing assignment in panfrost_get_job_for_fbo().
Also add a missing NULL assignment in the ->set_framebuffer_state()
path.

While at it, add extra assert()s to make sure ctx->job is consistent.

Fixes: 59c9623d0a75 ("panfrost: Import job data structures from v3d")
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

ac/nir,radv: Optimize bounds check for 64 bit CAS.

When the application does not ask for robust buffer access.

Only implemented the check in radv.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

gallivm: fix issue with AtomicCmpXchg wrapper on llvm 3.5-3.8

These versions still need wrapper but already have both success and
failure ordering.
(Compile tested on llvm 3.3, 3.7, 3.8.)

v2: don't duplicate whole function (suggested by Brian).

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111102

Reviewed-by: Charmaine Lee <charmainel@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>

util: Handle differences in pthread_setname_np

There are a lot of unfortunate differences in the implementation of this
function. NetBSD and Mac OS X in particular require different arguments.

https://stackoverflow.com/questions/2369738/how-to-set-the-name-of-a-thread-in-linux-pthreads/7989973#7989973
provides for a good overview of the differences.

Fixes: 9c411e020d1 ("util: Drop preprocessor guards for glibc-2.12")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111264
Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
[Eric: use DETECT_OS_* instead of PIPE_OS_*]
Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

util/os_time: use detect_os.h to uncouple from gallium

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

util/u_debug: use detect_os.h

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

util/os_misc: use detect_os.h to start uncoupling from gallium

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

util/os_memory: use detect_os.h to uncouple it from gallium

While at it, remove p_compiler.h as well as it is unused.

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

gallium: deduplicate os detection logic by using detect_os.h

This allows us to avoid having to rename all the PIPE_OS_* at once while
still making sure PIPE_OS_* and DETECT_OS_* are always in sync.

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

gallium/utils: drop PIPE_SUBSYSTEM_WINDOWS_USER

This is basically just an alias for PIPE_OS_WINDOWS.

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

scons: rename PIPE_SUBSYSTEM_EMBEDDED to EMBEDDED_DEVICE

It has nothing to do with the PIPE_SUBSYSTEM_* stuff from gallium.

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

gallium: remove never-used PIPE_SUBSYSTEM_DRI

PIPE_SUBSYSTEM_DRI was introduced in dacfef158943665fc0d1 ("gallium: New
configuration header.") 11 years ago, and was never used.

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

util: fix typo in comment

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

util: introduce detect_os.h

Mostly copied from src/gallium/include/pipe/p_config.h, so I kept its
copyright and authorship.

Other than the obvious rename, the big difference is that these are
always defined, to be used as `#if DETECT_OS_LINUX`.

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

freedreno/batch: fix dependency loop detection

We can have a scenario like:

  A -> B
  A -> C -> B

When adding the A->C dependency, it doesn't really matter that C depends
on something that A depends on, that isn't a necessary condition for a
dependency loop.

Instead what we want to know is that nothing C depends on, directly or
indirectly, depends on A.  We can detect this by recursively OR'ing the
dependents_mask of C and all it's dependencies.

Signed-off-by: Rob Clark <robdclark@chromium.org>

freedreno/a6xx: add missing flush/invalidates for blit

Various things we were missing for multiple blits in a single batch.

Signed-off-by: Rob Clark <robdclark@chromium.org>

freedreno/a6xx: skip tiles with no geometry

If no clear, and no geometry according to VSC_STATE[pipe] we can skip
the tile entirely.  If there is a fast-clear, we can't skip restore
(clear) or resolve IBs, but we can still skip draw IB.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>

freedreno/a6xx: VSC overflow detection/handling

Check VSC_SIZE/VSC_SIZE2 regs from cmdstream to detect overflow, and
skip use of VSC visibility stream when overflow is detected, to avoid
GPU hangs.  This is done w/ introduction of some CP_REG_TEST/
CP_COND_REG_EXEC packet pairs.

In addition, eventually (after a frame or two) detect the condition and
resize the VSC buffers until overflow no longer happens.

Note that this significantly reduces the initial size of the VSC
buffers, backing out a previous hack to make them 16x larger than
what should be typically required (the previous "solution" for
VSC overflow).

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>

freedreno/a6xx: remove USE/IGNORE_VISIBILITY draw patching

Seems this isn't needed anymore on a6xx to control whether visibility
stream is used.  And it would be hard to deal with if it was, for
disabling use of VSC stream in draw pass.  So just remove it and
simplify things.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>

freedreno/a6xx: cleanup "blit_mem"

Rename to "control_mem", and switch to using a struct to manage the
layout, rather than just ad-hoc hard-coded offsets.

For recovering from VSC stream overflow, we'll need to add more, but
best to clean it up first.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>

freedreno: refresh tile debug

Fix some #ifdef'd bitrot, and get rid of #ifdef so it doesn't bitrot
again.

And add a prints for per-tile state.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>

freedreno: update registers

Pull in some updates of VSC regs

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>

freedreno/gmem: small cleanup

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>

freedreno/drm: convert ring_pool to child_pool

Worth another couple percent at driver2

Signed-off-by: Rob Clark <robdclark@chromium.org>

freedreno/drm: remove idx_lock

Since it ends up contended, it is a bit of a bottleneck for workloads
with high driver overhead.  Worth nearly +10% at gfxbench driver2.

Signed-off-by: Rob Clark <robdclark@chromium.org>

freedreno/batch: always update last_fence

Not all flush paths come thru fd_context_flush(), so we should also set
last_fence in the batch flush path.  This avoids some no-op flushes just
to get a fence.  For example when pctx->flush_resource() triggers a
flush.

We should probably keep the last_fence update in fd_context_flush() as
well to handle deferred flush case.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>

freedreno: drop unused fd_fence_ref param

The pscreen param was just there to satisfy pipe_screen::fence_reference

But some of the internal uses passed NULL for screen.  Which is a bit
ugly.  Instead drop the param and add a shim function to plug into the
screen.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>

pan/midgard: Print invert modifier

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Flip conditionals

We would like to flip ops to have a constant in the second place to
enable inlining of the constant.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Add bitwise src/invert fusing

De Morgan's Laws and some special ops basically.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Add .not propagation pass

Essentially .pos propagation but for bitwise.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Fuse invert into bitwise ops

We use the new invert flag to produce ops like inand.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

freedreno: a2xx: implement texture tiling

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Rob Clark <robdclark@chromium.org>

freedreno: a2xx: use nir_lower_alu_to_scalar instead of lowering pass

nir_lower_alu_to_scalar can now be used to only lower certain ops, so we
don't need the custom pass. And we can lower fall_equal/fany_nequal with
lower_vector_cmp instead.

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Rob Clark <robdclark@chromium.org>

freedreno: a2xx: fix HW binning for batches with >256K vertices

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Rob Clark <robdclark@chromium.org>

freedreno: a2xx: fix fneg/fabs/fsat opcodes

Previously we would get a fmov with modifiers, but now that mov has no type
these opcodes need to be supported.

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Rob Clark <robdclark@chromium.org>

freedreno: a2xx: fix order of NIR opts

int_to_float needs to come after bool_to_float, and lower_to_source_mods
needs to come after both, since they don't deal wih source mods.

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Rob Clark <robdclark@chromium.org>

freedreno: a2xx: fix non-etc1 cubemaps

Not sure how this happened, but apparently all cubemaps need swapped XY.

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Rob Clark <robdclark@chromium.org>

freedreno: a2xx: fix fast clear not being used for Z24X8 buffers

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Rob Clark <robdclark@chromium.org>

freedreno: align renderonly scanout buffers

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Rob Clark <robdclark@chromium.org>

gitlab-ci: just build all the tools

This line was mistakenly added while there is already a `-D tools=all`
a few lines below.

Fixes: f60defa72d5d20d99e3a ("gitlab-ci: Add a shader-db run using v3d on drm-shim.")
Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>

i965/clear: clear_value better precision

Test-case with depth-clear 0.5 and format
MESA_FORMAT_Z24_UNORM_X8_UINT fails due inconsistent
clear-value of 0.4999997.
Maybe its better to improve?

CC: Jason Ekstrand <jason.ekstrand@intel.com>
Fixes: 0ae9ce0f29ea (i965/clear: Quantize the depth clear value based on the format)
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111113
Signed-off-by: Sergii Romantsov <sergii.romantsov@globallogic.com>
Signed-off-by: Danylo Piliaiev <danylo.piliaiev@globallogic.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

radv: fix image_has_{cmask,fmask}() helpers

The driver should now rely on cmask_offset because CMASK can be
disabled by the driver for some reasons (eg. mipmaps). Apply the
same change for FMASK, although it should be useless.

Fixes: ad1bc8621df ("radv: remove radv_get_image_fmask_info()")
Fixes: 10d08da52c6 ("radv/gfx10: add missing dcc_tile_swizzle tweak")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: remove radv_get_image_fmask_info()

It's unnecessary to duplicate fields in another struct.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv/gfx10: add missing dcc_tile_swizzle tweak

Fixes: c90f46700dd ("radv/gfx10: mask DCC tile swizzle by alignment")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: remove radv_get_image_cmask_info()

It's unnecessary to duplicate fields in another struct.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: only account for tile_swizzle for color surfaces with DCC

It's 0 for depth surfaces with TC compat HTILE enabled.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: Enable VK_KHR_shader_atomic_int64

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

ac/nir: Implement LLVM9 64-bit buffer compare & exchange.

LLVM 9 does not have a 64-bit buffer compswap intrinsic, so this
extracts the ptr, does a bound check and then uses a cmpxchg LLVM
instruction.

Not ideal, but the earliest release we're going to get a proper
intrinsic is LLVM 10.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

Revert "ac/nir: handle negate modifier"

This reverts commit bfea7e4d2965269bff8f1f6449cb99c312fd7384.

Revert "ac/nir: handle abs modifier"

This reverts commit d3c80733cdfe8552b2f447ec8ed62465d0f2af1a.

These were only appearing due to memory corruption.

iris: bump compat profile support to 4.6

All of the current piglit compat profile tests pass.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

egl: fix OpenGL 3.1 context creation

>From the EGL_KHR_create_context spec:

   "* If OpenGL 3.1 is requested, the context returned may implement
       any of the following versions:

         * Version 3.1. The GL_ARB_compatibility extension may or may
           not be implemented, as determined by the implementation.
         * The core profile of version 3.2 or greater."

Fixes CTS tests:

    dEQP-EGL.functional.create_context_ext.gl_31.rgb888_depth_stencil
    dEQP-EGL.functional.create_context_ext.robust_gl_31.rgb888_depth_stencil
    dEQP-EGL.functional.create_context_ext.gl_31.rgb888_depth_no_stencil
    dEQP-EGL.functional.create_context_ext.robust_gl_31.rgb888_depth_no_stencil
    dEQP-EGL.functional.create_context_ext.gl_31.rgba8888_depth_no_stencil
    dEQP-EGL.functional.create_context_ext.gl_31.rgb888_no_depth_no_stencil
    dEQP-EGL.functional.create_context_ext.robust_gl_31.rgba8888_depth_no_stencil
    dEQP-EGL.functional.create_context_ext.robust_gl_31.rgb888_no_depth_no_stencil
    dEQP-EGL.functional.create_context_ext.gl_31.rgba8888_no_depth_no_stencil
    dEQP-EGL.functional.create_context_ext.robust_gl_31.rgba8888_no_depth_no_stencil
    dEQP-EGL.functional.create_context_ext.gl_31.rgba8888_depth_stencil
    dEQP-EGL.functional.create_context_ext.robust_gl_31.rgba8888_depth_stencil

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

nir/find_array_copies: Reject copies with mismatched type

When we detect a scalar/vector copy through load_deref/store_deref, we
have to be careful since those can bitcast an int to a float and
vice-versa even though copy_deref can't.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111251
Fixes: 156306e5e62 ("nir/find_array_copies: Handle wildcards and overlapping copies")
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

radv: re-apply "Optimize rebinding the same descriptor set."

This makes it cheaper to just change the dynamic offsets with
the same descriptor sets.

This optimization has been reverted a while back because of
random GPU hangs on GFX9, no it looks fine, at least CTS no longer
hangs on GFX9 and it doesn't hang on GFX10 as well.

It fixes a performance problem with Wolfenstein Youngblood.

Suggested-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>

radv/gfx10: use the correct target machine for Wave32

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv/gfx10: add Wave32 support for vertex, tessellation and geometry shaders

It can be enabled with RADV_PERFTEST=gewave32.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv/gfx10: add Wave32 support for fragment shaders

It can be enabled with RADV_PERFTEST=pswave32.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

gallium: Implement GL_EXT_shader_samples_identical via a new capability

This exposes the textureSamplesIdenticalEXT function in GLSL.

We enable it for iris and radeonsi, because their compilers already
have support for this.  Tested on Intel Kabylake and AMD Vega 64.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

intel/tools: Fix aubinator_viewer build.

This functions was recently renamed and not all callers were updated.

Fixes: 086c486a75f ("intel/device: rename gen_get_device_info")

intel/ir: Fix CFG corruption in opt_predicated_break().

Specifically the optimization of a conditional BREAK + WHILE sequence
into a conditional WHILE seems pretty broken.  The list of successors
of "earlier_block" (where the conditional BREAK was found) is emptied
and then re-created with the same edges for no apparent reason.  On
top of that the list of predecessors of the block immediately after
the WHILE loop is emptied, but only one of the original edges will be
added back, which means that potentially several blocks that still
have it on their list of successors won't be on its list of
predecessors anymore, causing all sorts of hilarity due to the
inconsistency in the control flow graph.

The solution is to remove the code that's removing valid edges from
the CFG.  cfg_t::remove_block() will already clean up after itself.
The assert in bblock_t::combine_with() also needs to be removed since
we will be merging a block with multiple children into the first one
of them.

Found the issue on a hardware enabling branch originally, but
apparently somebody reproduced the same problem independently on
master in the meantime.

Fixes: d13bcdb3a9f ("i965/fs: Extend predicated break pass to predicate WHILE.")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111009
Cc: jiradet.jd@gmail.com
Cc: Sergii Romantsov <sergii.romantsov@globallogic.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: mesa-stable@lists.freedesktop.org
Tested-by: Paul Chelombitko <qamonstergl@gmail.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

intel/device: make internal functions private

The device info initializer makes several fuctions internal:

  - handling of device override
  - updating topology from kernel information

The implementation file is slightly reordered due to the renamed
functions being static.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

intel/device: rename gen_get_device_info

Rename the original device info initialization routine so callers
don't mistakenly call the wrong one:

  gen_get_device_info_from_fd:

      Queries kernel for full device info, including topology
      details.

  gen_get_device_info_from_pci_id:

      Partially initializes device info based on PCI ID lookup, when
      the kernel is not available.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

intel/tools: use device info initializer

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

anv: use initialization routine for gen_device_info

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

iris/screen: use initialization routine for gen_device_info

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

i965: Move device info initialization to common code

With perf queries, initializing the device info is much more complex
than just getting a PCI ID and calling gen_get_device_info.  This commit
adds a new gen_get_device_info_from_fd helper in common code which does
all of the requisite kernel queries to get device info including all of
the topology information.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

i965/perf: verify kernel support before registering OA metrics

When gen_device_info updates the topology in it's initializer, the
kernel queries will fail silently.  Iris and anv have minimum
kernel requirements that support the queries.  i965 must verify kernel
support before reporting OA metrics.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

intel/common: provide common ioctl routine

i965 links against libdrm for drmIoctl, but anv and iris both
re-implement this routine to avoid the dependency.

intel/dev also needs an ioctl wrapper, so lets share the same
implementation everywhere.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

panfrost: Remove unused argument

A relic from when we didn't have an online compiler, hah.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Handle MESA_SHADER_COMPUTE in compile callback

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Use standard list traversal to find initial tag

Fixes a hang (and abort) on empty shaders, which you shouldn't have
anyway but better safe than sorry. DCE going on the fritz is no reason
to freeze the system.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Use gl_shader_stage directly for compiles

No need to add a third set of enums to the mix.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Emit "draw" info for compute jobs

Important fields relating to shader state and UBOs are filled out from
this (misnomer) function.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Feed compute shaders into the compiler

The path for compute shader compiles resembles the graphic shader
compile path, although it is substantially simpler as we don't need any
shader keying.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Expose compute shaders as panfrost_shader_variants

Whether variants are packed by graphics or compute is irrelevant.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Remove shader state *base

It is now unused.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Remove CSO dependency from shader_compile

We want this routine to be generic across graphics and compute, so let
the caller deal with the typing.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Generalize UBO upload for other shader stages

Now that everything is unified, this generalization is nice and easy.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Guard vertex upload by ctx->vertex != NULL

This is irrelevant for graphics but matters for compute workloads.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Generalize vertex shader upload

This allows us to reuse the same code path for compute.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Share gl_enables between VERTEX/COMPUTE

Catch-all for magic bits.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>