freedreno/a6xx: enable LRZ by default

Now that dEQP should be happy, lets flip the switch.

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

freedreno/a6xx: fix LRZ logic

In particular, we need to invalidate the LRZ state when we cannot be
confident in what the Z state would be during rendering:

1) depth test modes not supported by LRZ
2) stencil test, which would require full rasterization and stencil
   test in the binning pass (whereas LRZ normally just needs to
   determine the min and max z value in an 8x8 quad)

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

freedreno/a6xx: fix LRZ layout

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

freedreno/a5xx+a6xx: split LRZ layout to per-gen

Seems to be a bit different for a6xx, so let's split this out.

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

freedreno/a6xx: disable LRZ when blending

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

radeonsi: don't rely on CLEAR_STATE to set PA_SC_GENERIC_SCISSOR_*

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>

radeonsi/gfx10: simplify the tess_turns_off_ngg condition

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>

radeonsi/gfx10: disable vertex grouping

based on PAL.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>

radeonsi: enable NIR by default and document GL 4.6 support

Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>

st/dri: assume external consumers of back buffers can write to the buffers

This was reverted needlessly because if was part of another series.

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-By: Tapani Pälli <tapani.palli@intel.com>

ANV: Stop advertising smoothLines support on gen10+

Reviewed-by: Ivan Briano <ivan.briano@intel.com>

meson/broadcom: libbroadcom_cle needs expat headers

Fixes: 1ae8018a6af81eec4832a57d9d0346aa3dd98d28
       ("meson: Add support for the vc4 driver.")
Reviewed-by: Eric Anholt <eric@anholt.net>

anv: fix incorrect VMA alignment for CCS main surfaces

Maybe finer way of dealing with this requirement would be to increase
the number of pdevice->memory.types[] to add a category for special
alignment cases.

Meanwhile this fixes the problem of CCS surface alignment and it's
probably not going to cause issues given the size of our address
space.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: 6af8a4acc4a4 ("anv: Add aux-map translation for gen12+")
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

anv: fix missing gen12 handling

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: 181be14d4303 ("anv: Build for gen12")
Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

docs: reword a bit and list HTTPS before FTP

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>

meson: drop `intel_` prefix on imgui_core

Again, no real effect, just the name of a temporary build file.

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>

meson: drop duplicate `lib` prefix on libiris_gen*

This has no real effect other than the names of the temporary files in
the build folder.

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>

radv: implement VK_KHR_separate_depth_stencil_layouts

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

radv: initialize HTILE for separate depth/stencil aspects

It either clears the whole HTILE buffer or part of it depending
on the HTILE mask parameter.

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

radv: do not init HTILE as compressed state when dst layout allows it

I don't think this makes much differences and a potential clear
following the initialization will overwrite HTILE anyways.

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

radv: synchronize after performing a separate depth/stencil fast clears

For depth+stencil images, the driver might use an optimized path
if only one aspect is cleared. It either clears the depth or the
stencil part of HTILE. Because the two separate aspects might use
the same HTILE memory we have to synchronize.

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

gitlab-ci: Don't exclude any piglit quick_shader tests

Now that we're running these with process isolation enabled, their
results will hopefully be stable.

Reviewed-by: Eric Anholt <eric@anholt.net>

gallivm: add TGSI bit arithmetic opcodes support

Add TGSI_OPCODE_BFI, TGSI_OPCODE_POPC, TGSI_OPCODE_LSB,
TGSI_OPCODE_IMSB, TGSI_OPCODE_UMSB, TGSI_OPCODE_IBFE,
TGSI_OPCODE_UBFE, TGSI_OPCODE_BREV support.

Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Jan Zielinski <jan.zielinski@intel.com>

radv: fix possibly wrong PA_SC_AA_CONFIG value for conservative rast

PA_SC_AA_CONFIG might be updated when conversative rasterization is
enabled. Because the driver only re-emits the multisample state if
the number of samples is different, that register value might not
be updated correctly.

Found by inspection, doesn't fix anything known.

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

radv: move emission of two PA_SC_* registers to the pipeline CS

They don't have to be updated dynamically.

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

st/dri: use st->flush callback to flush the backbuffer

Previously the flush was done before the call to st->flush but
could lead to problems as FLUSH_VERTICES could push some work
that would change the backbuffer (or modify it).

With this commit, all the backbuffer flushing code is executed
right before the call to st_flush.

Closes: https://gitlab.freedesktop.org/drm/amd/issues/842
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=205049

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

st/mesa: add a notify_before_flush callback param to flush

The new callback is called right before the flush is done to allow
users of st->flush to do some work after all the previous work has
been flushed.

This will be used by dri_flush in the next commit.

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

radeonsi: dcc dirty flag

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

radeonsi: fix multi plane buffers creation

When using 3 planes, the sequence produces this chain:
  plane0 -> plane2
This commit fixes this to produce:
  plane0 -> plane1 -> plane2

Fixes: 86e60bc2659 ("radeonsi: remove si_vid_join_surfaces and use combined planar allocations")
Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/2193
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: use gfx9.surf_offset to compute texture offset

Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/2177
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: use compute shader for clear 12-byte buffer

Signed-off-by: Sonny Jiang <sonny.jiang@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

st/mesa: release the draw shader properly to fix driver crashes (iris)

Reviewed-by: Dave Airlie <airlied@redhat.com>

draw, st/mesa: generate TGSI for ffvp/ARB_vp if draw lacks LLVM

Reviewed-by: Roland Scheidegger <sroland@vmware.com>

st/mesa: don't generate VS TGSI if NIR is enabled

it's no longer needed

Reviewed-by: Dave Airlie <airlied@redhat.com>

st/mesa: remove struct st_vp_variant in favor of st_common_variant

Reviewed-by: Dave Airlie <airlied@redhat.com>

st/mesa: remove st_vp_variant::num_inputs

Reviewed-by: Dave Airlie <airlied@redhat.com>

st/mesa: use a separate VS variant for the draw module

instead of keeping the IR indefinitely in st_vp_variant.

This trivially fixes Selection/Feedback/RasterPos for NIR.

Reviewed-by: Dave Airlie <airlied@redhat.com>

st/mesa: support shader images for Selection/Feedback/RasterPos

Reviewed-by: Dave Airlie <airlied@redhat.com>

st/mesa: support SSBOs for Selection/Feedback/RasterPos

Reviewed-by: Dave Airlie <airlied@redhat.com>

st/mesa: support samplers for Selection/Feedback/RasterPos

Reviewed-by: Dave Airlie <airlied@redhat.com>

st/mesa: save currently bound vertex samplers and sampler views in st_context

for st_draw_feedback.c

Reviewed-by: Dave Airlie <airlied@redhat.com>

st/mesa: support UBOs for Selection/Feedback/RasterPos

Reviewed-by: Dave Airlie <airlied@redhat.com>

gallivm: implement LOAD with CONSTBUF but don't enable it for llvmpipe

This is already used in st_draw_feedback.c, because it uses shaders
generated for drivers.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>

llvmpipe: implement TEX_LZ and TXF_LZ opcodes

gallivm receives these opcodes anyway because st_draw_feedback.c uses
shaders that were assembled for drivers, not llvmpipe.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>

drirc: set allow_higher_compat_version for Faster Than Light

With 781a78 ("mesa: enable ARB_direct_state_access in compat for
GL3.1+), it's possible to have DSA with GL3.1+.

FTL creates a GL3.1 compat context, but fails the
_mesa_has_geometry_shaders(..) check in frame_buffer_texture.

Bump the compat version to pass the check.

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

util/atomic: Fix p_atomic_add for unlocked and msvc paths

Braces mismatch (flagged by CI, untested).

Fixes: 385d13f26d2 "util/atomic: Add a _return variant of p_atomic_add"
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

freedreno: Track the set of UBOs to be uploaded in UBO analysis.

We were iterating over the entire 32-entry array each time, when we
can just use a bitset to know that we're only uploading from the first
entry normally.

Knocks ir3_emit_user_consts down from ~.5% of CPU to .1% on WebGL
fishtank.

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

freedreno: Stop forcing ALLOW_MAPPED_BUFFERS_DURING_EXEC off.

The default is to not throw GL errors when drawing with mapped
buffers, but we were forcing it on for unclear reasons.  Internally we
keep all our buffers mapped anyway, so it should be a no-op other than
reducing CPU overhead (.23% in a perf report for WebGL fishtank)

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

freedreno/fdperf: use drmOpen()

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

gallium/util: Support POLYGON in u_stream_outputs_for_vertices

u_decomposed_prims_for_vertices cannot support POLYGON, but POLYGON is
trivial to support as a special case directly (since we have the number
of vertices directly).

Fixes aborts in Panfrost in apps using GL_POLYGON.

Fixes: e881aa8c12c ("gallium/util: Add u_stream_outputs_for_vertices helper")
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Revewied-by: Eric Anholt <eric@anholt.net>

intel: Add pci-ids for Jasper Lake

Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

intel: Add device info for 1x4x6 Jasper Lake

Also removing the FIXME comments after matching the numbers with
updated documentation.

Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

lima: expose tiled format modifier in query_dmabuf_modifiers()

Fixes: 8c12f4e5f24f ("lima: enable tiling")
Reviewed-by: Qiang Yu <yuq825@gmail.com>
Signed-off-by: Vasily Khoruzhick <anarsoul@gmail.com>

lima: handle DRM_FORMAT_MOD_INVALID in resource_from_handle()

Assume that resource is tiled if we get DRM_FORMAT_MOD_INVALID
in resource_from_handle() and we don't have RO.

Fixes: 8c12f4e5f24f ("lima: enable tiling")
Reviewed-by: Qiang Yu <yuq825@gmail.com>
Signed-off-by: Vasily Khoruzhick <anarsoul@gmail.com>

turnip: add hw binning

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>

radv: do not use VK_TRUE/VK_FALSE

For consistency.

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

gallivm: add bitfield reverse and ufind_msb

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Krzysztof Raszkowski <krzysztof.raszkowski@intel.com>

gallium/scons: fix graw_gdi build

Fixes: 44a6b0107b37 (gallivm: add nir->llvm translation (v2))
Reviewed-by: Dave Airlie <Airlied@redhat.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>

aco: propagate temporaries into expanded vectors

Gives a very slight decrease in code size:
Totals from affected shaders:
Code Size: 1708488 -> 1702768 (-0.33 %) bytes
Max Waves: 2858 -> 2855 (-0.10 %)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: improve readfirstlane after uniform ssbo loads on GFX7

pipeline-db changes for GFX7:

80310 shaders in 40472 tests
Totals:
SGPRS: 3655900 -> 3643916 (-0.33 %)
VGPRS: 2678324 -> 2686324 (0.30 %)
Spilled SGPRs: 1730 -> 1634 (-5.55 %)
Spilled VGPRs: 14 -> 21 (50.00 %)
Scratch size: 15540 -> 15536 (-0.03 %) dwords per thread
Code Size: 136106120 -> 135457616 (-0.48 %) bytes
LDS: 1259 -> 1259 (0.00 %) blocks
Max Waves: 601014 -> 600206 (-0.13 %)

Totals from affected shaders:
SGPRS: 307832 -> 295848 (-3.89 %)
VGPRS: 267864 -> 275864 (2.99 %)
Spilled SGPRs: 770 -> 674 (-12.47 %)
Spilled VGPRs: 14 -> 21 (50.00 %)
Scratch size: 16 -> 12 (-25.00 %) dwords per thread
Code Size: 22007488 -> 21358984 (-2.95 %) bytes
LDS: 65 -> 65 (0.00 %) blocks
Max Waves: 28668 -> 27860 (-2.82 %)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: use soffset for MUBUF instructions on SI/CI

pipeline-db changes for GFX7:

80310 shaders in 40472 tests
Totals:
SGPRS: 3655300 -> 3655900 (0.02 %)
VGPRS: 2677732 -> 2678324 (0.02 %)
Spilled SGPRs: 1730 -> 1730 (0.00 %)
Spilled VGPRs: 14 -> 14 (0.00 %)
Scratch size: 15540 -> 15540 (0.00 %) dwords per thread
Code Size: 136488364 -> 136106120 (-0.28 %) bytes
LDS: 1259 -> 1259 (0.00 %) blocks
Max Waves: 601039 -> 601014 (-0.00 %)

Totals from affected shaders:
SGPRS: 316312 -> 316912 (0.19 %)
VGPRS: 273844 -> 274436 (0.22 %)
Spilled SGPRs: 770 -> 770 (0.00 %)
Spilled VGPRs: 14 -> 14 (0.00 %)
Scratch size: 16 -> 16 (0.00 %) dwords per thread
Code Size: 22724904 -> 22342660 (-1.68 %) bytes
LDS: 114 -> 114 (0.00 %) blocks
Max Waves: 30861 -> 30836 (-0.08 %)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

radv: Enable ACO on GFX7 (Sea Islands)

This patch also disables AMD_shader_ballot on GFX7 by default if ACO is used.
Note that shader_ballot works correctly, but performance seems inferior.
To enable shader_ballot use RADV_PERFTEST=shader_ballot.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: return to loop_active mask at continue_or_break blocks

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

radv: disable Youngblood app profile if ACO is used

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: implement exclusive scan for SI/CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: implement inclusive_scan for SI/CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: implement (clustered) reductions for SI/CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: don't use a scalar temporary for reductions on GFX10

This patch also adds the scalar temporary for scans on SI/CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: flush denorms after fmin/fmax on pre-GFX9

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

radv: only flush scalar cache for SSBO writes with ACO on GFX8+

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: disable disassembly for SI/CI due to lack of support by LLVM

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: implement 64bit ine/ieq for SI/CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: implement 64bit i2b for SI /CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: make 1/2*PI a literal constant on SI/CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: implement 64bit VGPR shifts for SI/CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: split read/writelane opcode into VOP2/VOP3 version for SI/CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: fix disassembly of writelane instructions.

ACO writes an unused 3rd operand for internal usage
which makes LLVM recoginize it as illegal instruction.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: recognize SI/CI SMRD hazards

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: implement quad swizzles for SI/CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: move buffer_store data to VGPR if needed

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: implement nir_op_isign on SI/CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: only use scalar loads for readonly buffers on SI/CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: implement nir_op_fquantize2f16 for SI/CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: fix SMEM offsets for SI/CI

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: SI/CI - fix sampler aniso

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: handle gfx7 int8/10 clamping on exports

Co-authored-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: Initial GFX7 Support

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

aco: refactor visit_store_fs_output() to use the Builder

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

anv: Re-emit all compute state on pipeline switch

It's a very odd case to hit in the real world.  However, there are some
CTS tests which switch back and forth between dispatch and clear without
changing the pipeline.

Fixes: bc612536eb2f "anv: Emit a dummy MEDIA_VFE_STATE before switching..."
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>

anv: Re-capture all batch and state buffers

When we moved from allocating BOs directly to using the BO cache, we
lost the EXEC_OBJECT_CAPTURE flag on all our state buffers.

Fixes: 3119b96bdf57 "anv: Allocate block pool BOs from the cache"
Fixes: ee77938733cd "anv: Allocate batch and fence buffers from..."
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>

anv: Return VK_ERROR_OUT_OF_DEVICE_MEMORY for too-large buffers

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

freedreno: Enable texture upload memory throttling.

Fixes oom-killer during streaming-texture-upload, which I found while
trying to enable piglit in CI.

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

freedreno: reorder format check

With the addition of the planar formats helper, the
planar formats no longer have a valid block.bits field.
Calling util_format_get_blocksize therefore asserts.

Reorder the check to see if the format is supported
before doing the query to get the blocksize.

Fixes: 20f132e5eff2d ("gallium/util: add planar format layouts and helpers")
Signed-off-by: Fritz Koenig <frkoenig@google.com>
Reviewed-by: Rob Clark <robdclark@chromium.org>

iris: Fix import of multi-planar surfaces with modifiers

Multi-planar surfaces are allowed to have modifiers. Don't require
DRM_FORMAT_MOD_INVALID in order to create a surface for each plane
defined by the format.

Fixes: 246eebba4a8 ("iris: Export and import surfaces with modifiers that have aux data")
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

gallium: Store the image format in winsys_handle

This format will be used to properly handle planar images with modifiers
in iris.

Fixes: 246eebba4a8 ("iris: Export and import surfaces with modifiers that have aux data")
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

gallium/dri2: Fix creation of multi-planar modifier images

The commit noted below assumed and enforced that DRM_MOD_INVALID was the
only valid modifier for multi-planar imported images. Due to that, it
required that modifier on multi-planar images to:

   1. Allow multiple planes.
   2. Perform YUV format lowering and extent adjustments.
   3. Use buffer_index to correctly map the given planes.

Fix these issues by removing or updating the code built on that
assumption.

Fixes: 2066966c106 ("gallium/dri2: Support creating multi-planar modifier images")
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

meson: Include iris in default gallium-drivers for x86/x86_64

We build i965 by default on x86/x86_64 platforms; let's build iris too.

Reviewed-by: Dylan Baker <dylan@pnwbakers.com>

anv: Use BO fences/semaphores for AcquireNextImage

Instead of doing a dummy submit on the command buffer for the fence or a
dummy semaphore and trusting in implicit sync, this commit moves us to
take advantage of implicit sync and just use the WSI image BO as the
fence.  Both semaphores and fences require a tiny bit of extra plumbing
to do this but the result is that we can get rid of a bunch of the extra
synchronization we're doing today.

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

anv: Add a fence_reset_reset_temporary helper

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

anv: Use submit-time implicit sync instead of allocate-time

In 83b943cc2f24, we started making all VkDeviceMemory BOs resident all
the time.  One unfortunate side-effect of this is that every
vkQueueSubmit sets EXEC_OBJECT_WRITE on every WSI memory object which
means that X server or Wayland compositor, instead of waiting on the
last vkQueueSubmit to actually write the buffer, now waits on the last
vkQueueSubmit to from that driver instance relative to whenever the
compositor's GL driver instance calls execbuf.  This potentially leads
to a lot of extra synchronization that we didn't intend to have.

Instead, this commit makes it so that we leave WSI memory objects with
EXEC_OBJECT_ASYNC most of the time and only unset EXEC_OBJECT_ASYNC and
set EXEC_OBJECT_WRITE in the dummy execbuf that we do as part of
vkQueuePresent.  This should hopefully result in tighter integration
with the compositor, lower latency, and better performance.

Testing with DOOM 2016, this seems to reduce latency by at least a frame
if not two and makes the game much more responsive.  Testing was,
however, subjective, so we don't have any hard data on that.

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>