anv/query: Use snooping on !LLC platforms

Commit b2c97bc789198427043cd902bc76e194e7e81c7d which made us start
using a busy-wait for individual query results also messed up cache
flushing on !LLC platforms.  For one thing, I forgot the mfence after
the clflush so memory access wasn't properly getting fenced.  More
importantly, however, was that we were clflushing the whole query range
and then waiting for individual queries and then trying to read the
results without clflushing again.  Getting the clflushing both correct
and efficient is very subtle and painful.  Instead, let's side-step the
problem by just snooping.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>

anv: provide anv_gem_busy() stub for the tests

Otherwise linking way fail.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100600
Fixes: f195d40eca4 ("anv/device: Add a helper for querying whether a BO is busy")
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Tested-by: Vinson Lee <vlee@freedesktop.org>

gallium/util: tweak backtrace format with libunwind

To work with addr2line.sh we also need the relative offset within the
DSO.  And addr2line.sh gets confused by the leading stackframe number.

Signed-off-by: Rob Clark <robdclark@gmail.com>

gallium/util: cache symbol lookup with libunwind

Signed-off-by: Rob Clark <robdclark@gmail.com>

gallium/util: fix missing limit check in libunwind backtrace

Fixes: 70c272004f ("gallium/util: libunwind support")
Signed-off-by: Rob Clark <robdclark@gmail.com>

mesa: fix renderbuffer leak

We don't need to call _mesa_reference_renderbuffer() for the first
assignment as refCount starts at 1. For swrast we work around the
fact we will indirectly call _mesa_reference_renderbuffer() by
resetting refCount to 0.

Fixes: 32141e53d1520 (mesa: tidy up renderbuffer RefCount initialisation)
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>

anv/blorp: sample input attachments with resolves on BDW

On Broadwell we still need to do a resolve between the subpass
that writes and the subpass that reads when there is a
self-dependency because HW could not see fast-clears and works
on the render cache as if there was regular non-fast-clear surface.

Fixes 16 tests on BDW:

dEQP-VK.renderpass.formats.*.input.clear.store.self_dep*

Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

radv: don't call radeon_check_space in radv_BindDescriptorSets

This appears to be a leftover from an earlier version of this function.
Nothing is emitted into the CS.

Signed-off-by: Fredrik Höglund <fredrik@kde.org>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: implement VK_KHR_descriptor_update_template

All offsets and strides are precomputed by
radv_CreateDescriptorUpdateTemplateKHR and stored in the template.

v2: Move the new struct declarations from radv_descriptor_set.h
    to radv_private.h (Bas)

Signed-off-by: Fredrik Höglund <fredrik@kde.org>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: implement VK_KHR_push_descriptor

Signed-off-by: Fredrik Höglund <fredrik@kde.org>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: replace an assertion with a conditional

Replace the !binding_layout->immutable_samplers assertion in
radv_update_descriptor_sets with a conditional.

The Vulkan specification does not say that it is illegal to update
a sampler descriptor when it is immutable; only that pImageInfo is
ignored.

This change is also needed for push descriptors, because valid
descriptors must be pushed for all bindings accessed by shaders,
including immutable sampler descriptors.

Signed-off-by: Fredrik Höglund <fredrik@kde.org>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: refactor radv_UpdateDescriptorSets

Move the implementation into a separate function that takes a
cmd_buffer and a dstSetOverride parameter.

When cmd_buffer is not NULL, radv_update_descriptor_sets calls
cs_add_buffer directly instead of updating the buffer list.

This will be used to implement VK_KHR_push_descriptor.

Signed-off-by: Fredrik Höglund <fredrik@kde.org>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

gallium/radeon: fix typo in radeon_winsys.h

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

mesa/main: simplify _mesa_IsRenderbuffer()

_mesa_lookup_renderbuffer() already checks if 'id' is non-zero.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>

mesa: stop abstracting texture object hashtable locking

This doesn't do anything useful so just remove it.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

mesa: stop abstracting buffer object hashtable locking

This doesn't do anything useful so just remove it.

Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

i965/blorp: Bump the batch space estimate

Commit f938354362655a378d474c5f79c52cea9852ab91 recently increased the
alignment on vertex buffer data from 32 to 64.  This caused us to
consume a bit more batch than we were before and we now go over the
estimate by a small amount on certain blits on gen8+.  This commit bumps
then gen8 batch estimate by a bit to compensate.  Haswell and older
still seems to be well within the limit.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100582
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
Cc: "13.0 17.0" <mesa-stable@lists.freedesktop.org>

intel/aubinator: Stop searching after a custom handler is found

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

intel/gen_decoder: return -1 for unknown command formats

Decoding with aubinator encountered a command of 0xffffffff. With the
previous code, it caused aubinator to jump 255 + 2 dwords to start
decoding again.

Instead we can attempt to detect the known instruction formats. If the
format is not recognized, then we can advance just 1 dword.

v2:
 * Update aubinator_error_decode
 * Actually convert the length variable returned into a *signed* integer
   in aubinator.c, intel_batchbuffer.c and aubinator_error_decode.c.

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Acked-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

intel/gen_decoder: Fix length for Media State/Object commands

From BDW PRM, Volume 6: Command Stream Programming, 'Render Command
Header Format'.

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

intel/aubinator_error_decode: Fix structure decode data

The call to gen_print_group should provide a pointer to the beginning
of the the structure data, not the start of the batch data.

Cc: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

st/pbo: select the right swizzle for instance IDs

The system value only has an X component, and radeonsi started
checking that in debug builds.

Reported-by: Michel Dänzer <michel.daenzer@amd.com>
Fixes: 4cf29427770f ("radeonsi: support 64-bit system values")
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

anv/query: Busy-wait for available query entries

Before, we were just looking at whether or not the user wanted us to
wait and waiting on the BO.  Some clients, such as the Serious engine,
use a single query pool for hundreds of individual query results where
the writes for those queries may be split across several command
buffers.  In this scenario, the individual query we're looking for may
become available long before the BO is idle so waiting on the query pool
BO to be finished is wasteful. This commit makes us instead busy-loop on
each query until it's available.

This significantly reduces pipeline bubbles and improves performance of
The Talos Principle on medium settings (where the GPU isn't overloaded
with drawing) by around 20% on my SkyLake gt4.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Tested-by: Eero Tamminen <eero.t.tamminen@intel.com>
Tested-by: Grazvydas Ignotas <notasas@gmail.com>

anv/device: Add a helper for querying whether a BO is busy

This is a bit more efficient than using GEM_WAIT with a timeout of 0.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>

swr: [rasterizer core] SIMD16 Frontend WIP

Implement widened binner for SIMD16

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>

swr: [rasterizer core] Enable 8x2 backend

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>

swr: [rasterizer codegen] remove copy of mako

mako is already a mesa build requirement, extra copy not needed.

Tested building against mesa build baseline (mako-0.8.0).

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>

swr: [rasterizer core/memory] Move intrinics to _simd functions

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>

swr: [rasterizer core] Programmable sample position support

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>

swr: [configure.ac/scons] require c++14

New C++ features used by upcoming swr changes.

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>

swr: [rasterizer core] Fix center sample pattern

Fix long hidden bug in rasterizer handling of center sample pattern.

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>

swr: [rasterizer core/memory] Fix missing avx512 storetile

Fix pre-processor macro handing to eliminate silently missing
implementation for AVX512.

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>

swr: [rasterizer core] SIMD16 Frontend WIP

Implement widened VS output for SIMD16

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>

mesa: use internal function when deleting buffers

This avoids validation and looking up the buffer target for a second time.

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

mesa: rework bind_buffer_object()

This allows internal users to pass buffer objects directly and
allows for KHR_no_error support to be more easily added.

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

mesa: small texstate tidy up

Possibly more efficient, either way it makes the code easier to
follow.

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

mesa: tidy up renderbuffer RefCount initialisation

42aaa548 changed the renderbuffer initialisation of RefCount from
1 to 0.

This is inconsitent with how we use RefCount elsewhere. Also every
driver implementation of NewRenderbuffer() calls
_mesa_init_renderbuffer() so its safe to set it there.

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

Revert "etnaviv: Cannot render to rb-swapped formats"

This reverts commit 658568941d5e232d690e1ffbcddbd6ea9685693a.

With the help of shader variants we can render to rb-swapped
formats now. Fixes about 60 piglits.

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>

etnaviv: add support for rb swap

If we render to rb swapped format we will create a shader variant doing
the involved swizzing in the pixel shader.

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>

etnaviv: adapt shader-db output for variant support

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Lucas Stach <l.stach@pengutronix.de>

etnaviv: bring back shader-db traces

If shader-db run, create a standard variant immediately
(as otherwise nothing will trigger the shader to be
actually compiled).

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Lucas Stach <l.stach@pengutronix.de>

etnaviv: add etna_shader_key and generate variants if needed

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>

etnaviv: pass a preallocated variant to compiler

In the long run the compiler needs to know the specifc variant
'key' in order to compile appropriate assembly. With this commit
the variant knows its shader and we are able pass the preallocated
variant into etna_compile_shader(..). This saves us from passing
extra ptrs everywhere.

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Lucas Stach <l.stach@pengutronix.de>

etnaviv: make specs const

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>

etnaviv: add struct etna_shader_state

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Lucas Stach <l.stach@pengutronix.de>

etnaviv: add basic shader variant support

This commit adds some basic infrastructure to handle shader
variants. We are still creating exactly one shader variant
for each shader.

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Lucas Stach <l.stach@pengutronix.de>

etnaviv: s/etna_shader/etna_shader_variant

Prep work to add shader variant support.

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>

etnaviv: remove not needed forward declarations

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>

gallium/util: honour LIBUNWIND_CFLAGS

Fixes: 70c272004f72 ("gallium/util: libunwind support")
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>

travis: Add radeonsi to continuous integration

Signed-off-by: Rhys Kidd <rhyskidd@gmail.com>
Acked-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Andres Gomez <agomez@igalia.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>

travis: Add radv vulkan driver to continuous integration

Signed-off-by: Rhys Kidd <rhyskidd@gmail.com>
Reviewed-by: Andres Gomez <agomez@igalia.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>

anv: provide required gem stubs for the tests

Introduce stubs to anv_gem_stub.c that match the anv_gem.c ones.
Otherwise we may get link-time errors, when building the tests.

v2: Introduce all the missing stubs at once.

Cc: Jason Ekstrand <jason@jlekstrand.net>
Cc: Vinson Lee <vlee@freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100574
Fixes: c964f0e485d ("anv: Query the kernel for reset status")
Fixes: 651ec926fc1 ("anv: Add support for 48-bit addresses")
Fixes: 060a6434eca ("anv: Advertise larger heap sizes")
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
---
I've intentionally kept the order the same identical to the anv_gem.c.
This way we can easily grep & diff in the future ;-)

configure.ac: pthread-stubs is not a thing on GNU/kFreeBSD

As mentioned on the xcb mailing list, the platform uses the GLIBC
forwarding mechanism.

https://lists.freedesktop.org/archives/xcb/2016-November/010896.html

Cc: Andreas Boll <andreas.boll.dev@gmail.com>
Reported-by: Andreas Boll <andreas.boll.dev@gmail.com>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>

st/clover: Fix build after shrink of pipe_box

Fixes: 3dfe61e ("gallium: decrease the size of pipe_box - 24 -> 16 bytes")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100569
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Tested-by: Vinson Lee <vlee@freedesktop.org>

radeonsi: add new polaris10 pci id

Reviewed-by: Christian König <christian.koenig@amd.com>
Cc: 13.0 17.0 <mesa-stable@lists.freedesktop.org>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

radeonsi: enable ARB_shader_ballot

Require LLVM 5.0 or later because LLVM 4.0 is easily fooled into
putting the lane select of llvm.amdgcn.readlane into a VGPR and then
fails to continue to compile.

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

radeonsi: optimization barriers to work around LLVM deficiencies

Notably, llvm.amdgcn.readfirstlane and llvm.amdgcn.icmp may be hoisted
out of loops or if/else branches in cases like

  if (cond) {
    v = readFirstInvocationARB(x);
    ... use v ...
  } else {
    v = readFirstInvocationARB(x);
    ... use v ...
  }
===>
  v = readFirstInvocationARB(x);
  if (cond) {
    ... use v ...
  } else {
    ... use v ...
  }

The optimization barrier is a heavy hammer to stop that until LLVM
is taught the semantics of the intrinsic properly.

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

radeonsi: strengthen emit_optimization_barrier

LLVM will lift inline assembly out of if-else-blocks if both paths have
the same inline assembly. Prevent this by adding an irrelevant unique
text to the assembly.

This requires the LLVM assembly parser to be initialized.

Furthermore, allow forcing subsequent computations to happen after the
optimization barrier by defining a data dependency.

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

radeonsi: emit TGSI_OPCODE_READ_*

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

radeonsi: emit TGSI_OPCODE_BALLOT

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

radeonsi: implement TGSI_SEMANTIC_SUBGROUP_*

64-bit system values are stored as v2i32 to simplify the fetch logic.

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

radeonsi: support 64-bit system values

For simplicitly, always store system values as 32-bit values or arrays
of 32-bit values. 64-bit values are unpacked and packed accordingly.

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

radeonsi: bump RADEON_LLVM_MAX_SYSTEM_VALUES

ARB_shader_ballot introduces 7 new system values that can be used
in all shader stages.

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

st/mesa: enable ARB_shader_ballot

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

st/glsl_to_tgsi: implement ARB_shader_ballot system variables

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

st/glsl_to_tgsi: implement ARB_shader_ballot builtin functions

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

tgsi: add SUBGROUP_* semantics

v2: add documentation (Nicolai)

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

tgsi: add BALLOT/READ_* opcodes

v2 (Nicolai):
- BALLOT isn't per-channel
- expand the documentation (also for VOTE_*)

v3:
- only BALLOT returns a 64-bit lanemask (Boyan)
- relax the requirement on READ_INVOC: the invocation number to read
  from must be uniform within a sub-group. This matches the
  GL_ARB_shader_ballot spect (and the v_readlane instruction of AMD
  GCN)

v4:
- hopefully really fix the doc of VOTE_* returns (Ilia)

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v2)

gallium: add PIPE_CAP_TGSI_BALLOT

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

glsl: add gl_SubGroup*ARB builtins

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

glsl: add ARB_shader_ballot builtin functions

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

glsl: add ARB_shader_ballot operations

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

glsl: add ARB_shader_ballot enable

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

mesa: add GL_ARB_shader_ballot boilerplate

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

swr: automake: add gen_common.py to the tarball

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>

intel: genxml: automake: include gen_bits_header.py in the tarball

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>

intel: genxml: automake: polish automake rules

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>

amd/addrlib: automake: add all headers to the tarball

Fixes: 7f160efcde4 ("amd/addrlib: import gfx9 support")
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>

radeonsi: enable ARB_sparse_buffer

v2:
- fill in DRM version requirement
- disable on SI due to CP DMA faults

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

radeonsi: disable SDMA clears and copies for sparse buffers

VM faults cannot be disabled for SDMA on <= VI.

We could still use SDMA by asking the winsys about which parts of the
buffers are committed. This is left as a potential future improvement.

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

gallium/radeon: implement pipe->resource_commit

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

gallium/radeon: transfers and invalidation for sparse buffers

Sparse buffers can never be mapped by the CPU.

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

gallium/radeon: implement sparse buffer creation

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

winsys/amdgpu: sparse buffer debugging helpers

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

winsys/amdgpu: take fences when freeing a backing buffer

We never add fences to backing buffers during submit. When we free a
backing buffer, it must inherit the sparse buffer's fences, so that it
doesn't get re-used prematurely via the cache.

v2:
- remove pipe_mutex_*

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

winsys/amdgpu: add sparse buffers to CS

... and implement the corresponding fence handling.

v2:
- add missing bit in amdgpu_bo_is_referenced_by_cs_with_usage
- remove pipe_mutex_*

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

winsys/amdgpu: sparse buffer creation / destruction / commitment

This is the bulk of the buffer allocation logic. It is fairly simple and
stupid. We'll probably want to use e.g. interval trees at some point to
keep track of commitments, but Mesa doesn't have an implementation of those
yet.

v2:
- remove pipe_mutex_*
- fix total_backing_pages accounting
- simplify by using the new VA_OP_CLEAR/REPLACE kernel interface

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

winsys/amdgpu: add sparse buffer data structures

v2:
- remove pipe_mutex_*
- use a simple page commitment array

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

winsys/amdgpu: extend amdgpu_add_fence to allow adding multiple fences

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

winsys/amdgpu: build handles and flags list late on submit thread

This probably has only minor performance effects, but it simplifies some
subsequent code slightly.

Ideally, it could also be used to simplify the handling of slab buffers
in the same way, but unfortunately that's not possible as long as we need
indices for relocations.

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

winsys/amdgpu: share common code in amdgpu_add_fence_dependencies

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

winsys/amdgpu: extract amdgpu_do_add_real_buffer

We will use it for delayed adding of sparse buffers' backing buffers.

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

winsys/radeon: sparse buffers will not be supported

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

radeon/winsys: add sparse buffer interface

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

st/mesa: plumbing for sparse buffers

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

st/mesa: enable ARB_sparse_buffer when supported

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

trace: add resource_commit pass-through

v2: fix return type to bool (Marek)

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

ddebug: add resource_commit pass-through

v2: fix return type to bool (Marek)

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

gallium: add sparse buffer interface and capability

v2:
- explain the resource_commit interface in more detail

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

mesa: implement sparse buffer commitment

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>